1 //===-- RISCVISelLowering.cpp - RISCV DAG Lowering Implementation  --------===//
2 //
3 // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
4 // See https://llvm.org/LICENSE.txt for license information.
5 // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
6 //
7 //===----------------------------------------------------------------------===//
8 //
9 // This file defines the interfaces that RISCV uses to lower LLVM code into a
10 // selection DAG.
11 //
12 //===----------------------------------------------------------------------===//
13 
14 #include "RISCVISelLowering.h"
15 #include "MCTargetDesc/RISCVMatInt.h"
16 #include "RISCV.h"
17 #include "RISCVMachineFunctionInfo.h"
18 #include "RISCVRegisterInfo.h"
19 #include "RISCVSubtarget.h"
20 #include "RISCVTargetMachine.h"
21 #include "llvm/ADT/SmallSet.h"
22 #include "llvm/ADT/Statistic.h"
23 #include "llvm/Analysis/MemoryLocation.h"
24 #include "llvm/CodeGen/MachineFrameInfo.h"
25 #include "llvm/CodeGen/MachineFunction.h"
26 #include "llvm/CodeGen/MachineInstrBuilder.h"
27 #include "llvm/CodeGen/MachineJumpTableInfo.h"
28 #include "llvm/CodeGen/MachineRegisterInfo.h"
29 #include "llvm/CodeGen/TargetLoweringObjectFileImpl.h"
30 #include "llvm/CodeGen/ValueTypes.h"
31 #include "llvm/IR/DiagnosticInfo.h"
32 #include "llvm/IR/DiagnosticPrinter.h"
33 #include "llvm/IR/IRBuilder.h"
34 #include "llvm/IR/IntrinsicsRISCV.h"
35 #include "llvm/IR/PatternMatch.h"
36 #include "llvm/Support/CommandLine.h"
37 #include "llvm/Support/Debug.h"
38 #include "llvm/Support/ErrorHandling.h"
39 #include "llvm/Support/KnownBits.h"
40 #include "llvm/Support/MathExtras.h"
41 #include "llvm/Support/raw_ostream.h"
42 #include <optional>
43 
44 using namespace llvm;
45 
46 #define DEBUG_TYPE "riscv-lower"
47 
48 STATISTIC(NumTailCalls, "Number of tail calls");
49 
50 static cl::opt<unsigned> ExtensionMaxWebSize(
51     DEBUG_TYPE "-ext-max-web-size", cl::Hidden,
52     cl::desc("Give the maximum size (in number of nodes) of the web of "
53              "instructions that we will consider for VW expansion"),
54     cl::init(18));
55 
56 static cl::opt<bool>
57     AllowSplatInVW_W(DEBUG_TYPE "-form-vw-w-with-splat", cl::Hidden,
58                      cl::desc("Allow the formation of VW_W operations (e.g., "
59                               "VWADD_W) with splat constants"),
60                      cl::init(false));
61 
62 static cl::opt<unsigned> NumRepeatedDivisors(
63     DEBUG_TYPE "-fp-repeated-divisors", cl::Hidden,
64     cl::desc("Set the minimum number of repetitions of a divisor to allow "
65              "transformation to multiplications by the reciprocal"),
66     cl::init(2));
67 
RISCVTargetLowering(const TargetMachine & TM,const RISCVSubtarget & STI)68 RISCVTargetLowering::RISCVTargetLowering(const TargetMachine &TM,
69                                          const RISCVSubtarget &STI)
70     : TargetLowering(TM), Subtarget(STI) {
71 
72   if (Subtarget.isRV32E())
73     report_fatal_error("Codegen not yet implemented for RV32E");
74 
75   RISCVABI::ABI ABI = Subtarget.getTargetABI();
76   assert(ABI != RISCVABI::ABI_Unknown && "Improperly initialised target ABI");
77 
78   if ((ABI == RISCVABI::ABI_ILP32F || ABI == RISCVABI::ABI_LP64F) &&
79       !Subtarget.hasStdExtF()) {
80     errs() << "Hard-float 'f' ABI can't be used for a target that "
81                 "doesn't support the F instruction set extension (ignoring "
82                           "target-abi)\n";
83     ABI = Subtarget.is64Bit() ? RISCVABI::ABI_LP64 : RISCVABI::ABI_ILP32;
84   } else if ((ABI == RISCVABI::ABI_ILP32D || ABI == RISCVABI::ABI_LP64D) &&
85              !Subtarget.hasStdExtD()) {
86     errs() << "Hard-float 'd' ABI can't be used for a target that "
87               "doesn't support the D instruction set extension (ignoring "
88               "target-abi)\n";
89     ABI = Subtarget.is64Bit() ? RISCVABI::ABI_LP64 : RISCVABI::ABI_ILP32;
90   }
91 
92   switch (ABI) {
93   default:
94     report_fatal_error("Don't know how to lower this ABI");
95   case RISCVABI::ABI_ILP32:
96   case RISCVABI::ABI_ILP32F:
97   case RISCVABI::ABI_ILP32D:
98   case RISCVABI::ABI_LP64:
99   case RISCVABI::ABI_LP64F:
100   case RISCVABI::ABI_LP64D:
101     break;
102   }
103 
104   MVT XLenVT = Subtarget.getXLenVT();
105 
106   // Set up the register classes.
107   addRegisterClass(XLenVT, &RISCV::GPRRegClass);
108 
109   if (Subtarget.hasStdExtZfhOrZfhmin())
110     addRegisterClass(MVT::f16, &RISCV::FPR16RegClass);
111   if (Subtarget.hasStdExtF())
112     addRegisterClass(MVT::f32, &RISCV::FPR32RegClass);
113   if (Subtarget.hasStdExtD())
114     addRegisterClass(MVT::f64, &RISCV::FPR64RegClass);
115 
116   static const MVT::SimpleValueType BoolVecVTs[] = {
117       MVT::nxv1i1,  MVT::nxv2i1,  MVT::nxv4i1, MVT::nxv8i1,
118       MVT::nxv16i1, MVT::nxv32i1, MVT::nxv64i1};
119   static const MVT::SimpleValueType IntVecVTs[] = {
120       MVT::nxv1i8,  MVT::nxv2i8,   MVT::nxv4i8,   MVT::nxv8i8,  MVT::nxv16i8,
121       MVT::nxv32i8, MVT::nxv64i8,  MVT::nxv1i16,  MVT::nxv2i16, MVT::nxv4i16,
122       MVT::nxv8i16, MVT::nxv16i16, MVT::nxv32i16, MVT::nxv1i32, MVT::nxv2i32,
123       MVT::nxv4i32, MVT::nxv8i32,  MVT::nxv16i32, MVT::nxv1i64, MVT::nxv2i64,
124       MVT::nxv4i64, MVT::nxv8i64};
125   static const MVT::SimpleValueType F16VecVTs[] = {
126       MVT::nxv1f16, MVT::nxv2f16,  MVT::nxv4f16,
127       MVT::nxv8f16, MVT::nxv16f16, MVT::nxv32f16};
128   static const MVT::SimpleValueType F32VecVTs[] = {
129       MVT::nxv1f32, MVT::nxv2f32, MVT::nxv4f32, MVT::nxv8f32, MVT::nxv16f32};
130   static const MVT::SimpleValueType F64VecVTs[] = {
131       MVT::nxv1f64, MVT::nxv2f64, MVT::nxv4f64, MVT::nxv8f64};
132 
133   if (Subtarget.hasVInstructions()) {
134     auto addRegClassForRVV = [this](MVT VT) {
135       // Disable the smallest fractional LMUL types if ELEN is less than
136       // RVVBitsPerBlock.
137       unsigned MinElts = RISCV::RVVBitsPerBlock / Subtarget.getELEN();
138       if (VT.getVectorMinNumElements() < MinElts)
139         return;
140 
141       unsigned Size = VT.getSizeInBits().getKnownMinValue();
142       const TargetRegisterClass *RC;
143       if (Size <= RISCV::RVVBitsPerBlock)
144         RC = &RISCV::VRRegClass;
145       else if (Size == 2 * RISCV::RVVBitsPerBlock)
146         RC = &RISCV::VRM2RegClass;
147       else if (Size == 4 * RISCV::RVVBitsPerBlock)
148         RC = &RISCV::VRM4RegClass;
149       else if (Size == 8 * RISCV::RVVBitsPerBlock)
150         RC = &RISCV::VRM8RegClass;
151       else
152         llvm_unreachable("Unexpected size");
153 
154       addRegisterClass(VT, RC);
155     };
156 
157     for (MVT VT : BoolVecVTs)
158       addRegClassForRVV(VT);
159     for (MVT VT : IntVecVTs) {
160       if (VT.getVectorElementType() == MVT::i64 &&
161           !Subtarget.hasVInstructionsI64())
162         continue;
163       addRegClassForRVV(VT);
164     }
165 
166     if (Subtarget.hasVInstructionsF16())
167       for (MVT VT : F16VecVTs)
168         addRegClassForRVV(VT);
169 
170     if (Subtarget.hasVInstructionsF32())
171       for (MVT VT : F32VecVTs)
172         addRegClassForRVV(VT);
173 
174     if (Subtarget.hasVInstructionsF64())
175       for (MVT VT : F64VecVTs)
176         addRegClassForRVV(VT);
177 
178     if (Subtarget.useRVVForFixedLengthVectors()) {
179       auto addRegClassForFixedVectors = [this](MVT VT) {
180         MVT ContainerVT = getContainerForFixedLengthVector(VT);
181         unsigned RCID = getRegClassIDForVecVT(ContainerVT);
182         const RISCVRegisterInfo &TRI = *Subtarget.getRegisterInfo();
183         addRegisterClass(VT, TRI.getRegClass(RCID));
184       };
185       for (MVT VT : MVT::integer_fixedlen_vector_valuetypes())
186         if (useRVVForFixedLengthVectorVT(VT))
187           addRegClassForFixedVectors(VT);
188 
189       for (MVT VT : MVT::fp_fixedlen_vector_valuetypes())
190         if (useRVVForFixedLengthVectorVT(VT))
191           addRegClassForFixedVectors(VT);
192     }
193   }
194 
195   // Compute derived properties from the register classes.
196   computeRegisterProperties(STI.getRegisterInfo());
197 
198   setStackPointerRegisterToSaveRestore(RISCV::X2);
199 
200   setLoadExtAction({ISD::EXTLOAD, ISD::SEXTLOAD, ISD::ZEXTLOAD}, XLenVT,
201                    MVT::i1, Promote);
202   // DAGCombiner can call isLoadExtLegal for types that aren't legal.
203   setLoadExtAction({ISD::EXTLOAD, ISD::SEXTLOAD, ISD::ZEXTLOAD}, MVT::i32,
204                    MVT::i1, Promote);
205 
206   // TODO: add all necessary setOperationAction calls.
207   setOperationAction(ISD::DYNAMIC_STACKALLOC, XLenVT, Expand);
208 
209   setOperationAction(ISD::BR_JT, MVT::Other, Expand);
210   setOperationAction(ISD::BR_CC, XLenVT, Expand);
211   setOperationAction(ISD::BRCOND, MVT::Other, Custom);
212   setOperationAction(ISD::SELECT_CC, XLenVT, Expand);
213 
214   setCondCodeAction(ISD::SETLE, XLenVT, Expand);
215   setCondCodeAction(ISD::SETGT, XLenVT, Custom);
216   setCondCodeAction(ISD::SETGE, XLenVT, Expand);
217   setCondCodeAction(ISD::SETULE, XLenVT, Expand);
218   setCondCodeAction(ISD::SETUGT, XLenVT, Custom);
219   setCondCodeAction(ISD::SETUGE, XLenVT, Expand);
220 
221   setOperationAction({ISD::STACKSAVE, ISD::STACKRESTORE}, MVT::Other, Expand);
222 
223   setOperationAction(ISD::VASTART, MVT::Other, Custom);
224   setOperationAction({ISD::VAARG, ISD::VACOPY, ISD::VAEND}, MVT::Other, Expand);
225 
226   setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i1, Expand);
227 
228   setOperationAction(ISD::EH_DWARF_CFA, MVT::i32, Custom);
229 
230   if (!Subtarget.hasStdExtZbb())
231     setOperationAction(ISD::SIGN_EXTEND_INREG, {MVT::i8, MVT::i16}, Expand);
232 
233   if (Subtarget.is64Bit()) {
234     setOperationAction(ISD::EH_DWARF_CFA, MVT::i64, Custom);
235 
236     setOperationAction(ISD::LOAD, MVT::i32, Custom);
237 
238     setOperationAction({ISD::ADD, ISD::SUB, ISD::SHL, ISD::SRA, ISD::SRL},
239                        MVT::i32, Custom);
240 
241     setOperationAction({ISD::UADDO, ISD::USUBO, ISD::UADDSAT, ISD::USUBSAT},
242                        MVT::i32, Custom);
243   } else {
244     setLibcallName(
245         {RTLIB::SHL_I128, RTLIB::SRL_I128, RTLIB::SRA_I128, RTLIB::MUL_I128},
246         nullptr);
247     setLibcallName(RTLIB::MULO_I64, nullptr);
248   }
249 
250   if (!Subtarget.hasStdExtM() && !Subtarget.hasStdExtZmmul()) {
251     setOperationAction({ISD::MUL, ISD::MULHS, ISD::MULHU}, XLenVT, Expand);
252   } else {
253     if (Subtarget.is64Bit()) {
254       setOperationAction(ISD::MUL, {MVT::i32, MVT::i128}, Custom);
255     } else {
256       setOperationAction(ISD::MUL, MVT::i64, Custom);
257     }
258   }
259 
260   if (!Subtarget.hasStdExtM()) {
261     setOperationAction({ISD::SDIV, ISD::UDIV, ISD::SREM, ISD::UREM},
262                        XLenVT, Expand);
263   } else {
264     if (Subtarget.is64Bit()) {
265       setOperationAction({ISD::SDIV, ISD::UDIV, ISD::UREM},
266                           {MVT::i8, MVT::i16, MVT::i32}, Custom);
267     }
268   }
269 
270   setOperationAction(
271       {ISD::SDIVREM, ISD::UDIVREM, ISD::SMUL_LOHI, ISD::UMUL_LOHI}, XLenVT,
272       Expand);
273 
274   setOperationAction({ISD::SHL_PARTS, ISD::SRL_PARTS, ISD::SRA_PARTS}, XLenVT,
275                      Custom);
276 
277   if (Subtarget.hasStdExtZbb() || Subtarget.hasStdExtZbkb()) {
278     if (Subtarget.is64Bit())
279       setOperationAction({ISD::ROTL, ISD::ROTR}, MVT::i32, Custom);
280   } else {
281     setOperationAction({ISD::ROTL, ISD::ROTR}, XLenVT, Expand);
282   }
283 
284   // With Zbb we have an XLen rev8 instruction, but not GREVI. So we'll
285   // pattern match it directly in isel.
286   setOperationAction(ISD::BSWAP, XLenVT,
287                      (Subtarget.hasStdExtZbb() || Subtarget.hasStdExtZbkb())
288                          ? Legal
289                          : Expand);
290   // Zbkb can use rev8+brev8 to implement bitreverse.
291   setOperationAction(ISD::BITREVERSE, XLenVT,
292                      Subtarget.hasStdExtZbkb() ? Custom : Expand);
293 
294   if (Subtarget.hasStdExtZbb()) {
295     setOperationAction({ISD::SMIN, ISD::SMAX, ISD::UMIN, ISD::UMAX}, XLenVT,
296                        Legal);
297 
298     if (Subtarget.is64Bit())
299       setOperationAction(
300           {ISD::CTTZ, ISD::CTTZ_ZERO_UNDEF, ISD::CTLZ, ISD::CTLZ_ZERO_UNDEF},
301           MVT::i32, Custom);
302   } else {
303     setOperationAction({ISD::CTTZ, ISD::CTLZ, ISD::CTPOP}, XLenVT, Expand);
304   }
305 
306   if (Subtarget.is64Bit())
307     setOperationAction(ISD::ABS, MVT::i32, Custom);
308 
309   if (!Subtarget.hasVendorXVentanaCondOps())
310     setOperationAction(ISD::SELECT, XLenVT, Custom);
311 
312   static const unsigned FPLegalNodeTypes[] = {
313       ISD::FMINNUM,        ISD::FMAXNUM,       ISD::LRINT,
314       ISD::LLRINT,         ISD::LROUND,        ISD::LLROUND,
315       ISD::STRICT_LRINT,   ISD::STRICT_LLRINT, ISD::STRICT_LROUND,
316       ISD::STRICT_LLROUND, ISD::STRICT_FMA,    ISD::STRICT_FADD,
317       ISD::STRICT_FSUB,    ISD::STRICT_FMUL,   ISD::STRICT_FDIV,
318       ISD::STRICT_FSQRT,   ISD::STRICT_FSETCC, ISD::STRICT_FSETCCS};
319 
320   static const ISD::CondCode FPCCToExpand[] = {
321       ISD::SETOGT, ISD::SETOGE, ISD::SETONE, ISD::SETUEQ, ISD::SETUGT,
322       ISD::SETUGE, ISD::SETULT, ISD::SETULE, ISD::SETUNE, ISD::SETGT,
323       ISD::SETGE,  ISD::SETNE,  ISD::SETO,   ISD::SETUO};
324 
325   static const unsigned FPOpToExpand[] = {
326       ISD::FSIN, ISD::FCOS,       ISD::FSINCOS,   ISD::FPOW,
327       ISD::FREM, ISD::FP16_TO_FP, ISD::FP_TO_FP16};
328 
329   static const unsigned FPRndMode[] = {
330       ISD::FCEIL, ISD::FFLOOR, ISD::FTRUNC, ISD::FRINT, ISD::FROUND,
331       ISD::FROUNDEVEN};
332 
333   if (Subtarget.hasStdExtZfhOrZfhmin())
334     setOperationAction(ISD::BITCAST, MVT::i16, Custom);
335 
336   if (Subtarget.hasStdExtZfhOrZfhmin()) {
337     if (Subtarget.hasStdExtZfh()) {
338       setOperationAction(FPLegalNodeTypes, MVT::f16, Legal);
339       setOperationAction(FPRndMode, MVT::f16, Custom);
340       setOperationAction(ISD::SELECT, MVT::f16, Custom);
341     } else {
342       static const unsigned ZfhminPromoteOps[] = {
343           ISD::FMINNUM,      ISD::FMAXNUM,       ISD::FADD,
344           ISD::FSUB,         ISD::FMUL,          ISD::FMA,
345           ISD::FDIV,         ISD::FSQRT,         ISD::FABS,
346           ISD::FNEG,         ISD::STRICT_FMA,    ISD::STRICT_FADD,
347           ISD::STRICT_FSUB,  ISD::STRICT_FMUL,   ISD::STRICT_FDIV,
348           ISD::STRICT_FSQRT, ISD::STRICT_FSETCC, ISD::STRICT_FSETCCS,
349           ISD::SETCC,        ISD::FCEIL,         ISD::FFLOOR,
350           ISD::FTRUNC,       ISD::FRINT,         ISD::FROUND,
351           ISD::FROUNDEVEN,   ISD::SELECT};
352 
353       setOperationAction(ZfhminPromoteOps, MVT::f16, Promote);
354       setOperationAction({ISD::STRICT_LRINT, ISD::STRICT_LLRINT,
355                           ISD::STRICT_LROUND, ISD::STRICT_LLROUND},
356                          MVT::f16, Legal);
357       // FIXME: Need to promote f16 FCOPYSIGN to f32, but the
358       // DAGCombiner::visitFP_ROUND probably needs improvements first.
359       setOperationAction(ISD::FCOPYSIGN, MVT::f16, Expand);
360     }
361 
362     setOperationAction(ISD::STRICT_FP_ROUND, MVT::f16, Legal);
363     setOperationAction(ISD::STRICT_FP_EXTEND, MVT::f32, Legal);
364     setCondCodeAction(FPCCToExpand, MVT::f16, Expand);
365     setOperationAction(ISD::SELECT_CC, MVT::f16, Expand);
366     setOperationAction(ISD::BR_CC, MVT::f16, Expand);
367 
368     setOperationAction({ISD::FREM, ISD::FNEARBYINT, ISD::FPOW, ISD::FPOWI,
369                         ISD::FCOS, ISD::FSIN, ISD::FSINCOS, ISD::FEXP,
370                         ISD::FEXP2, ISD::FLOG, ISD::FLOG2, ISD::FLOG10},
371                        MVT::f16, Promote);
372 
373     // FIXME: Need to promote f16 STRICT_* to f32 libcalls, but we don't have
374     // complete support for all operations in LegalizeDAG.
375     setOperationAction({ISD::STRICT_FCEIL, ISD::STRICT_FFLOOR,
376                         ISD::STRICT_FNEARBYINT, ISD::STRICT_FRINT,
377                         ISD::STRICT_FROUND, ISD::STRICT_FROUNDEVEN,
378                         ISD::STRICT_FTRUNC},
379                        MVT::f16, Promote);
380 
381     // We need to custom promote this.
382     if (Subtarget.is64Bit())
383       setOperationAction(ISD::FPOWI, MVT::i32, Custom);
384   }
385 
386   if (Subtarget.hasStdExtF()) {
387     setOperationAction(FPLegalNodeTypes, MVT::f32, Legal);
388     setOperationAction(FPRndMode, MVT::f32, Custom);
389     setCondCodeAction(FPCCToExpand, MVT::f32, Expand);
390     setOperationAction(ISD::SELECT_CC, MVT::f32, Expand);
391     setOperationAction(ISD::SELECT, MVT::f32, Custom);
392     setOperationAction(ISD::BR_CC, MVT::f32, Expand);
393     setOperationAction(FPOpToExpand, MVT::f32, Expand);
394     setLoadExtAction(ISD::EXTLOAD, MVT::f32, MVT::f16, Expand);
395     setTruncStoreAction(MVT::f32, MVT::f16, Expand);
396   }
397 
398   if (Subtarget.hasStdExtF() && Subtarget.is64Bit())
399     setOperationAction(ISD::BITCAST, MVT::i32, Custom);
400 
401   if (Subtarget.hasStdExtD()) {
402     setOperationAction(FPLegalNodeTypes, MVT::f64, Legal);
403     if (Subtarget.is64Bit()) {
404       setOperationAction(FPRndMode, MVT::f64, Custom);
405     }
406     setOperationAction(ISD::STRICT_FP_ROUND, MVT::f32, Legal);
407     setOperationAction(ISD::STRICT_FP_EXTEND, MVT::f64, Legal);
408     setCondCodeAction(FPCCToExpand, MVT::f64, Expand);
409     setOperationAction(ISD::SELECT_CC, MVT::f64, Expand);
410     setOperationAction(ISD::SELECT, MVT::f64, Custom);
411     setOperationAction(ISD::BR_CC, MVT::f64, Expand);
412     setLoadExtAction(ISD::EXTLOAD, MVT::f64, MVT::f32, Expand);
413     setTruncStoreAction(MVT::f64, MVT::f32, Expand);
414     setOperationAction(FPOpToExpand, MVT::f64, Expand);
415     setLoadExtAction(ISD::EXTLOAD, MVT::f64, MVT::f16, Expand);
416     setTruncStoreAction(MVT::f64, MVT::f16, Expand);
417   }
418 
419   if (Subtarget.is64Bit())
420     setOperationAction({ISD::FP_TO_UINT, ISD::FP_TO_SINT,
421                         ISD::STRICT_FP_TO_UINT, ISD::STRICT_FP_TO_SINT},
422                        MVT::i32, Custom);
423 
424   if (Subtarget.hasStdExtF()) {
425     setOperationAction({ISD::FP_TO_UINT_SAT, ISD::FP_TO_SINT_SAT}, XLenVT,
426                        Custom);
427 
428     setOperationAction({ISD::STRICT_FP_TO_UINT, ISD::STRICT_FP_TO_SINT,
429                         ISD::STRICT_UINT_TO_FP, ISD::STRICT_SINT_TO_FP},
430                        XLenVT, Legal);
431 
432     setOperationAction(ISD::GET_ROUNDING, XLenVT, Custom);
433     setOperationAction(ISD::SET_ROUNDING, MVT::Other, Custom);
434   }
435 
436   setOperationAction({ISD::GlobalAddress, ISD::BlockAddress, ISD::ConstantPool,
437                       ISD::JumpTable},
438                      XLenVT, Custom);
439 
440   setOperationAction(ISD::GlobalTLSAddress, XLenVT, Custom);
441 
442   if (Subtarget.is64Bit())
443     setOperationAction(ISD::Constant, MVT::i64, Custom);
444 
445   // TODO: On M-mode only targets, the cycle[h] CSR may not be present.
446   // Unfortunately this can't be determined just from the ISA naming string.
447   setOperationAction(ISD::READCYCLECOUNTER, MVT::i64,
448                      Subtarget.is64Bit() ? Legal : Custom);
449 
450   setOperationAction({ISD::TRAP, ISD::DEBUGTRAP}, MVT::Other, Legal);
451   setOperationAction(ISD::INTRINSIC_WO_CHAIN, MVT::Other, Custom);
452   if (Subtarget.is64Bit())
453     setOperationAction(ISD::INTRINSIC_WO_CHAIN, MVT::i32, Custom);
454 
455   if (Subtarget.hasStdExtA()) {
456     setMaxAtomicSizeInBitsSupported(Subtarget.getXLen());
457     setMinCmpXchgSizeInBits(32);
458   } else if (Subtarget.hasForcedAtomics()) {
459     setMaxAtomicSizeInBitsSupported(Subtarget.getXLen());
460   } else {
461     setMaxAtomicSizeInBitsSupported(0);
462   }
463 
464   setOperationAction(ISD::ATOMIC_FENCE, MVT::Other, Custom);
465 
466   setBooleanContents(ZeroOrOneBooleanContent);
467 
468   if (Subtarget.hasVInstructions()) {
469     setBooleanVectorContents(ZeroOrOneBooleanContent);
470 
471     setOperationAction(ISD::VSCALE, XLenVT, Custom);
472 
473     // RVV intrinsics may have illegal operands.
474     // We also need to custom legalize vmv.x.s.
475     setOperationAction({ISD::INTRINSIC_WO_CHAIN, ISD::INTRINSIC_W_CHAIN},
476                        {MVT::i8, MVT::i16}, Custom);
477     if (Subtarget.is64Bit())
478       setOperationAction(ISD::INTRINSIC_W_CHAIN, MVT::i32, Custom);
479     else
480       setOperationAction({ISD::INTRINSIC_WO_CHAIN, ISD::INTRINSIC_W_CHAIN},
481                          MVT::i64, Custom);
482 
483     setOperationAction({ISD::INTRINSIC_W_CHAIN, ISD::INTRINSIC_VOID},
484                        MVT::Other, Custom);
485 
486     static const unsigned IntegerVPOps[] = {
487         ISD::VP_ADD,         ISD::VP_SUB,         ISD::VP_MUL,
488         ISD::VP_SDIV,        ISD::VP_UDIV,        ISD::VP_SREM,
489         ISD::VP_UREM,        ISD::VP_AND,         ISD::VP_OR,
490         ISD::VP_XOR,         ISD::VP_ASHR,        ISD::VP_LSHR,
491         ISD::VP_SHL,         ISD::VP_REDUCE_ADD,  ISD::VP_REDUCE_AND,
492         ISD::VP_REDUCE_OR,   ISD::VP_REDUCE_XOR,  ISD::VP_REDUCE_SMAX,
493         ISD::VP_REDUCE_SMIN, ISD::VP_REDUCE_UMAX, ISD::VP_REDUCE_UMIN,
494         ISD::VP_MERGE,       ISD::VP_SELECT,      ISD::VP_FP_TO_SINT,
495         ISD::VP_FP_TO_UINT,  ISD::VP_SETCC,       ISD::VP_SIGN_EXTEND,
496         ISD::VP_ZERO_EXTEND, ISD::VP_TRUNCATE,    ISD::VP_SMIN,
497         ISD::VP_SMAX,        ISD::VP_UMIN,        ISD::VP_UMAX,
498         ISD::VP_ABS};
499 
500     static const unsigned FloatingPointVPOps[] = {
501         ISD::VP_FADD,        ISD::VP_FSUB,        ISD::VP_FMUL,
502         ISD::VP_FDIV,        ISD::VP_FNEG,        ISD::VP_FABS,
503         ISD::VP_FMA,         ISD::VP_REDUCE_FADD, ISD::VP_REDUCE_SEQ_FADD,
504         ISD::VP_REDUCE_FMIN, ISD::VP_REDUCE_FMAX, ISD::VP_MERGE,
505         ISD::VP_SELECT,      ISD::VP_SINT_TO_FP,  ISD::VP_UINT_TO_FP,
506         ISD::VP_SETCC,       ISD::VP_FP_ROUND,    ISD::VP_FP_EXTEND,
507         ISD::VP_SQRT,        ISD::VP_FMINNUM,     ISD::VP_FMAXNUM,
508         ISD::VP_FCEIL,       ISD::VP_FFLOOR,      ISD::VP_FROUND,
509         ISD::VP_FROUNDEVEN,  ISD::VP_FCOPYSIGN,   ISD::VP_FROUNDTOZERO,
510         ISD::VP_FRINT,       ISD::VP_FNEARBYINT};
511 
512     static const unsigned IntegerVecReduceOps[] = {
513         ISD::VECREDUCE_ADD,  ISD::VECREDUCE_AND,  ISD::VECREDUCE_OR,
514         ISD::VECREDUCE_XOR,  ISD::VECREDUCE_SMAX, ISD::VECREDUCE_SMIN,
515         ISD::VECREDUCE_UMAX, ISD::VECREDUCE_UMIN};
516 
517     static const unsigned FloatingPointVecReduceOps[] = {
518         ISD::VECREDUCE_FADD, ISD::VECREDUCE_SEQ_FADD, ISD::VECREDUCE_FMIN,
519         ISD::VECREDUCE_FMAX};
520 
521     if (!Subtarget.is64Bit()) {
522       // We must custom-lower certain vXi64 operations on RV32 due to the vector
523       // element type being illegal.
524       setOperationAction({ISD::INSERT_VECTOR_ELT, ISD::EXTRACT_VECTOR_ELT},
525                          MVT::i64, Custom);
526 
527       setOperationAction(IntegerVecReduceOps, MVT::i64, Custom);
528 
529       setOperationAction({ISD::VP_REDUCE_ADD, ISD::VP_REDUCE_AND,
530                           ISD::VP_REDUCE_OR, ISD::VP_REDUCE_XOR,
531                           ISD::VP_REDUCE_SMAX, ISD::VP_REDUCE_SMIN,
532                           ISD::VP_REDUCE_UMAX, ISD::VP_REDUCE_UMIN},
533                          MVT::i64, Custom);
534     }
535 
536     for (MVT VT : BoolVecVTs) {
537       if (!isTypeLegal(VT))
538         continue;
539 
540       setOperationAction(ISD::SPLAT_VECTOR, VT, Custom);
541 
542       // Mask VTs are custom-expanded into a series of standard nodes
543       setOperationAction({ISD::TRUNCATE, ISD::CONCAT_VECTORS,
544                           ISD::INSERT_SUBVECTOR, ISD::EXTRACT_SUBVECTOR},
545                          VT, Custom);
546 
547       setOperationAction({ISD::INSERT_VECTOR_ELT, ISD::EXTRACT_VECTOR_ELT}, VT,
548                          Custom);
549 
550       setOperationAction(ISD::SELECT, VT, Custom);
551       setOperationAction(
552           {ISD::SELECT_CC, ISD::VSELECT, ISD::VP_MERGE, ISD::VP_SELECT}, VT,
553           Expand);
554 
555       setOperationAction({ISD::VP_AND, ISD::VP_OR, ISD::VP_XOR}, VT, Custom);
556 
557       setOperationAction(
558           {ISD::VECREDUCE_AND, ISD::VECREDUCE_OR, ISD::VECREDUCE_XOR}, VT,
559           Custom);
560 
561       setOperationAction(
562           {ISD::VP_REDUCE_AND, ISD::VP_REDUCE_OR, ISD::VP_REDUCE_XOR}, VT,
563           Custom);
564 
565       // RVV has native int->float & float->int conversions where the
566       // element type sizes are within one power-of-two of each other. Any
567       // wider distances between type sizes have to be lowered as sequences
568       // which progressively narrow the gap in stages.
569       setOperationAction(
570           {ISD::SINT_TO_FP, ISD::UINT_TO_FP, ISD::FP_TO_SINT, ISD::FP_TO_UINT},
571           VT, Custom);
572       setOperationAction({ISD::FP_TO_SINT_SAT, ISD::FP_TO_UINT_SAT}, VT,
573                          Custom);
574 
575       // Expand all extending loads to types larger than this, and truncating
576       // stores from types larger than this.
577       for (MVT OtherVT : MVT::integer_scalable_vector_valuetypes()) {
578         setTruncStoreAction(OtherVT, VT, Expand);
579         setLoadExtAction({ISD::EXTLOAD, ISD::SEXTLOAD, ISD::ZEXTLOAD}, OtherVT,
580                          VT, Expand);
581       }
582 
583       setOperationAction({ISD::VP_FP_TO_SINT, ISD::VP_FP_TO_UINT,
584                           ISD::VP_TRUNCATE, ISD::VP_SETCC},
585                          VT, Custom);
586       setOperationAction(ISD::VECTOR_REVERSE, VT, Custom);
587 
588       setOperationPromotedToType(
589           ISD::VECTOR_SPLICE, VT,
590           MVT::getVectorVT(MVT::i8, VT.getVectorElementCount()));
591     }
592 
593     for (MVT VT : IntVecVTs) {
594       if (!isTypeLegal(VT))
595         continue;
596 
597       setOperationAction(ISD::SPLAT_VECTOR, VT, Legal);
598       setOperationAction(ISD::SPLAT_VECTOR_PARTS, VT, Custom);
599 
600       // Vectors implement MULHS/MULHU.
601       setOperationAction({ISD::SMUL_LOHI, ISD::UMUL_LOHI}, VT, Expand);
602 
603       // nxvXi64 MULHS/MULHU requires the V extension instead of Zve64*.
604       if (VT.getVectorElementType() == MVT::i64 && !Subtarget.hasStdExtV())
605         setOperationAction({ISD::MULHU, ISD::MULHS}, VT, Expand);
606 
607       setOperationAction({ISD::SMIN, ISD::SMAX, ISD::UMIN, ISD::UMAX}, VT,
608                          Legal);
609 
610       setOperationAction({ISD::ROTL, ISD::ROTR}, VT, Expand);
611 
612       setOperationAction({ISD::CTTZ, ISD::CTLZ, ISD::CTPOP}, VT, Expand);
613 
614       setOperationAction(ISD::BSWAP, VT, Expand);
615       setOperationAction({ISD::VP_BSWAP, ISD::VP_BITREVERSE}, VT, Expand);
616       setOperationAction({ISD::VP_FSHL, ISD::VP_FSHR}, VT, Expand);
617       setOperationAction({ISD::VP_CTLZ, ISD::VP_CTLZ_ZERO_UNDEF, ISD::VP_CTTZ,
618                           ISD::VP_CTTZ_ZERO_UNDEF, ISD::VP_CTPOP},
619                          VT, Expand);
620 
621       // Custom-lower extensions and truncations from/to mask types.
622       setOperationAction({ISD::ANY_EXTEND, ISD::SIGN_EXTEND, ISD::ZERO_EXTEND},
623                          VT, Custom);
624 
625       // RVV has native int->float & float->int conversions where the
626       // element type sizes are within one power-of-two of each other. Any
627       // wider distances between type sizes have to be lowered as sequences
628       // which progressively narrow the gap in stages.
629       setOperationAction(
630           {ISD::SINT_TO_FP, ISD::UINT_TO_FP, ISD::FP_TO_SINT, ISD::FP_TO_UINT},
631           VT, Custom);
632       setOperationAction({ISD::FP_TO_SINT_SAT, ISD::FP_TO_UINT_SAT}, VT,
633                          Custom);
634 
635       setOperationAction(
636           {ISD::SADDSAT, ISD::UADDSAT, ISD::SSUBSAT, ISD::USUBSAT}, VT, Legal);
637 
638       // Integer VTs are lowered as a series of "RISCVISD::TRUNCATE_VECTOR_VL"
639       // nodes which truncate by one power of two at a time.
640       setOperationAction(ISD::TRUNCATE, VT, Custom);
641 
642       // Custom-lower insert/extract operations to simplify patterns.
643       setOperationAction({ISD::INSERT_VECTOR_ELT, ISD::EXTRACT_VECTOR_ELT}, VT,
644                          Custom);
645 
646       // Custom-lower reduction operations to set up the corresponding custom
647       // nodes' operands.
648       setOperationAction(IntegerVecReduceOps, VT, Custom);
649 
650       setOperationAction(IntegerVPOps, VT, Custom);
651 
652       setOperationAction({ISD::LOAD, ISD::STORE}, VT, Custom);
653 
654       setOperationAction({ISD::MLOAD, ISD::MSTORE, ISD::MGATHER, ISD::MSCATTER},
655                          VT, Custom);
656 
657       setOperationAction(
658           {ISD::VP_LOAD, ISD::VP_STORE, ISD::EXPERIMENTAL_VP_STRIDED_LOAD,
659            ISD::EXPERIMENTAL_VP_STRIDED_STORE, ISD::VP_GATHER, ISD::VP_SCATTER},
660           VT, Custom);
661 
662       setOperationAction(
663           {ISD::CONCAT_VECTORS, ISD::INSERT_SUBVECTOR, ISD::EXTRACT_SUBVECTOR},
664           VT, Custom);
665 
666       setOperationAction(ISD::SELECT, VT, Custom);
667       setOperationAction(ISD::SELECT_CC, VT, Expand);
668 
669       setOperationAction({ISD::STEP_VECTOR, ISD::VECTOR_REVERSE}, VT, Custom);
670 
671       for (MVT OtherVT : MVT::integer_scalable_vector_valuetypes()) {
672         setTruncStoreAction(VT, OtherVT, Expand);
673         setLoadExtAction({ISD::EXTLOAD, ISD::SEXTLOAD, ISD::ZEXTLOAD}, OtherVT,
674                          VT, Expand);
675       }
676 
677       // Splice
678       setOperationAction(ISD::VECTOR_SPLICE, VT, Custom);
679 
680       // Lower CTLZ_ZERO_UNDEF and CTTZ_ZERO_UNDEF if element of VT in the range
681       // of f32.
682       EVT FloatVT = MVT::getVectorVT(MVT::f32, VT.getVectorElementCount());
683       if (isTypeLegal(FloatVT)) {
684         setOperationAction(
685             {ISD::CTLZ, ISD::CTLZ_ZERO_UNDEF, ISD::CTTZ_ZERO_UNDEF}, VT,
686             Custom);
687       }
688     }
689 
690     // Expand various CCs to best match the RVV ISA, which natively supports UNE
691     // but no other unordered comparisons, and supports all ordered comparisons
692     // except ONE. Additionally, we expand GT,OGT,GE,OGE for optimization
693     // purposes; they are expanded to their swapped-operand CCs (LT,OLT,LE,OLE),
694     // and we pattern-match those back to the "original", swapping operands once
695     // more. This way we catch both operations and both "vf" and "fv" forms with
696     // fewer patterns.
697     static const ISD::CondCode VFPCCToExpand[] = {
698         ISD::SETO,   ISD::SETONE, ISD::SETUEQ, ISD::SETUGT,
699         ISD::SETUGE, ISD::SETULT, ISD::SETULE, ISD::SETUO,
700         ISD::SETGT,  ISD::SETOGT, ISD::SETGE,  ISD::SETOGE,
701     };
702 
703     // Sets common operation actions on RVV floating-point vector types.
704     const auto SetCommonVFPActions = [&](MVT VT) {
705       setOperationAction(ISD::SPLAT_VECTOR, VT, Legal);
706       // RVV has native FP_ROUND & FP_EXTEND conversions where the element type
707       // sizes are within one power-of-two of each other. Therefore conversions
708       // between vXf16 and vXf64 must be lowered as sequences which convert via
709       // vXf32.
710       setOperationAction({ISD::FP_ROUND, ISD::FP_EXTEND}, VT, Custom);
711       // Custom-lower insert/extract operations to simplify patterns.
712       setOperationAction({ISD::INSERT_VECTOR_ELT, ISD::EXTRACT_VECTOR_ELT}, VT,
713                          Custom);
714       // Expand various condition codes (explained above).
715       setCondCodeAction(VFPCCToExpand, VT, Expand);
716 
717       setOperationAction({ISD::FMINNUM, ISD::FMAXNUM}, VT, Legal);
718 
719       setOperationAction(
720           {ISD::FTRUNC, ISD::FCEIL, ISD::FFLOOR, ISD::FROUND, ISD::FROUNDEVEN},
721           VT, Custom);
722 
723       setOperationAction(FloatingPointVecReduceOps, VT, Custom);
724 
725       // Expand FP operations that need libcalls.
726       setOperationAction(ISD::FREM, VT, Expand);
727       setOperationAction(ISD::FPOW, VT, Expand);
728       setOperationAction(ISD::FCOS, VT, Expand);
729       setOperationAction(ISD::FSIN, VT, Expand);
730       setOperationAction(ISD::FSINCOS, VT, Expand);
731       setOperationAction(ISD::FEXP, VT, Expand);
732       setOperationAction(ISD::FEXP2, VT, Expand);
733       setOperationAction(ISD::FLOG, VT, Expand);
734       setOperationAction(ISD::FLOG2, VT, Expand);
735       setOperationAction(ISD::FLOG10, VT, Expand);
736       setOperationAction(ISD::FRINT, VT, Expand);
737       setOperationAction(ISD::FNEARBYINT, VT, Expand);
738 
739       setOperationAction(ISD::FCOPYSIGN, VT, Legal);
740 
741       setOperationAction({ISD::LOAD, ISD::STORE}, VT, Custom);
742 
743       setOperationAction({ISD::MLOAD, ISD::MSTORE, ISD::MGATHER, ISD::MSCATTER},
744                          VT, Custom);
745 
746       setOperationAction(
747           {ISD::VP_LOAD, ISD::VP_STORE, ISD::EXPERIMENTAL_VP_STRIDED_LOAD,
748            ISD::EXPERIMENTAL_VP_STRIDED_STORE, ISD::VP_GATHER, ISD::VP_SCATTER},
749           VT, Custom);
750 
751       setOperationAction(ISD::SELECT, VT, Custom);
752       setOperationAction(ISD::SELECT_CC, VT, Expand);
753 
754       setOperationAction(
755           {ISD::CONCAT_VECTORS, ISD::INSERT_SUBVECTOR, ISD::EXTRACT_SUBVECTOR},
756           VT, Custom);
757 
758       setOperationAction({ISD::VECTOR_REVERSE, ISD::VECTOR_SPLICE}, VT, Custom);
759 
760       setOperationAction(FloatingPointVPOps, VT, Custom);
761     };
762 
763     // Sets common extload/truncstore actions on RVV floating-point vector
764     // types.
765     const auto SetCommonVFPExtLoadTruncStoreActions =
766         [&](MVT VT, ArrayRef<MVT::SimpleValueType> SmallerVTs) {
767           for (auto SmallVT : SmallerVTs) {
768             setTruncStoreAction(VT, SmallVT, Expand);
769             setLoadExtAction(ISD::EXTLOAD, VT, SmallVT, Expand);
770           }
771         };
772 
773     if (Subtarget.hasVInstructionsF16()) {
774       for (MVT VT : F16VecVTs) {
775         if (!isTypeLegal(VT))
776           continue;
777         SetCommonVFPActions(VT);
778       }
779     }
780 
781     if (Subtarget.hasVInstructionsF32()) {
782       for (MVT VT : F32VecVTs) {
783         if (!isTypeLegal(VT))
784           continue;
785         SetCommonVFPActions(VT);
786         SetCommonVFPExtLoadTruncStoreActions(VT, F16VecVTs);
787       }
788     }
789 
790     if (Subtarget.hasVInstructionsF64()) {
791       for (MVT VT : F64VecVTs) {
792         if (!isTypeLegal(VT))
793           continue;
794         SetCommonVFPActions(VT);
795         SetCommonVFPExtLoadTruncStoreActions(VT, F16VecVTs);
796         SetCommonVFPExtLoadTruncStoreActions(VT, F32VecVTs);
797       }
798     }
799 
800     if (Subtarget.useRVVForFixedLengthVectors()) {
801       for (MVT VT : MVT::integer_fixedlen_vector_valuetypes()) {
802         if (!useRVVForFixedLengthVectorVT(VT))
803           continue;
804 
805         // By default everything must be expanded.
806         for (unsigned Op = 0; Op < ISD::BUILTIN_OP_END; ++Op)
807           setOperationAction(Op, VT, Expand);
808         for (MVT OtherVT : MVT::integer_fixedlen_vector_valuetypes()) {
809           setTruncStoreAction(VT, OtherVT, Expand);
810           setLoadExtAction({ISD::EXTLOAD, ISD::SEXTLOAD, ISD::ZEXTLOAD},
811                            OtherVT, VT, Expand);
812         }
813 
814         // We use EXTRACT_SUBVECTOR as a "cast" from scalable to fixed.
815         setOperationAction({ISD::INSERT_SUBVECTOR, ISD::EXTRACT_SUBVECTOR}, VT,
816                            Custom);
817 
818         setOperationAction({ISD::BUILD_VECTOR, ISD::CONCAT_VECTORS}, VT,
819                            Custom);
820 
821         setOperationAction({ISD::INSERT_VECTOR_ELT, ISD::EXTRACT_VECTOR_ELT},
822                            VT, Custom);
823 
824         setOperationAction({ISD::LOAD, ISD::STORE}, VT, Custom);
825 
826         setOperationAction(ISD::SETCC, VT, Custom);
827 
828         setOperationAction(ISD::SELECT, VT, Custom);
829 
830         setOperationAction(ISD::TRUNCATE, VT, Custom);
831 
832         setOperationAction(ISD::BITCAST, VT, Custom);
833 
834         setOperationAction(
835             {ISD::VECREDUCE_AND, ISD::VECREDUCE_OR, ISD::VECREDUCE_XOR}, VT,
836             Custom);
837 
838         setOperationAction(
839             {ISD::VP_REDUCE_AND, ISD::VP_REDUCE_OR, ISD::VP_REDUCE_XOR}, VT,
840             Custom);
841 
842         setOperationAction({ISD::SINT_TO_FP, ISD::UINT_TO_FP, ISD::FP_TO_SINT,
843                             ISD::FP_TO_UINT},
844                            VT, Custom);
845         setOperationAction({ISD::FP_TO_SINT_SAT, ISD::FP_TO_UINT_SAT}, VT,
846                            Custom);
847 
848         // Operations below are different for between masks and other vectors.
849         if (VT.getVectorElementType() == MVT::i1) {
850           setOperationAction({ISD::VP_AND, ISD::VP_OR, ISD::VP_XOR, ISD::AND,
851                               ISD::OR, ISD::XOR},
852                              VT, Custom);
853 
854           setOperationAction({ISD::VP_FP_TO_SINT, ISD::VP_FP_TO_UINT,
855                               ISD::VP_SETCC, ISD::VP_TRUNCATE},
856                              VT, Custom);
857           continue;
858         }
859 
860         // Make SPLAT_VECTOR Legal so DAGCombine will convert splat vectors to
861         // it before type legalization for i64 vectors on RV32. It will then be
862         // type legalized to SPLAT_VECTOR_PARTS which we need to Custom handle.
863         // FIXME: Use SPLAT_VECTOR for all types? DAGCombine probably needs
864         // improvements first.
865         if (!Subtarget.is64Bit() && VT.getVectorElementType() == MVT::i64) {
866           setOperationAction(ISD::SPLAT_VECTOR, VT, Legal);
867           setOperationAction(ISD::SPLAT_VECTOR_PARTS, VT, Custom);
868         }
869 
870         setOperationAction(ISD::VECTOR_SHUFFLE, VT, Custom);
871 
872         setOperationAction(
873             {ISD::MLOAD, ISD::MSTORE, ISD::MGATHER, ISD::MSCATTER}, VT, Custom);
874 
875         setOperationAction({ISD::VP_LOAD, ISD::VP_STORE,
876                             ISD::EXPERIMENTAL_VP_STRIDED_LOAD,
877                             ISD::EXPERIMENTAL_VP_STRIDED_STORE, ISD::VP_GATHER,
878                             ISD::VP_SCATTER},
879                            VT, Custom);
880 
881         setOperationAction({ISD::ADD, ISD::MUL, ISD::SUB, ISD::AND, ISD::OR,
882                             ISD::XOR, ISD::SDIV, ISD::SREM, ISD::UDIV,
883                             ISD::UREM, ISD::SHL, ISD::SRA, ISD::SRL},
884                            VT, Custom);
885 
886         setOperationAction(
887             {ISD::SMIN, ISD::SMAX, ISD::UMIN, ISD::UMAX, ISD::ABS}, VT, Custom);
888 
889         // vXi64 MULHS/MULHU requires the V extension instead of Zve64*.
890         if (VT.getVectorElementType() != MVT::i64 || Subtarget.hasStdExtV())
891           setOperationAction({ISD::MULHS, ISD::MULHU}, VT, Custom);
892 
893         setOperationAction(
894             {ISD::SADDSAT, ISD::UADDSAT, ISD::SSUBSAT, ISD::USUBSAT}, VT,
895             Custom);
896 
897         setOperationAction(ISD::VSELECT, VT, Custom);
898         setOperationAction(ISD::SELECT_CC, VT, Expand);
899 
900         setOperationAction(
901             {ISD::ANY_EXTEND, ISD::SIGN_EXTEND, ISD::ZERO_EXTEND}, VT, Custom);
902 
903         // Custom-lower reduction operations to set up the corresponding custom
904         // nodes' operands.
905         setOperationAction({ISD::VECREDUCE_ADD, ISD::VECREDUCE_SMAX,
906                             ISD::VECREDUCE_SMIN, ISD::VECREDUCE_UMAX,
907                             ISD::VECREDUCE_UMIN},
908                            VT, Custom);
909 
910         setOperationAction(IntegerVPOps, VT, Custom);
911 
912         // Lower CTLZ_ZERO_UNDEF and CTTZ_ZERO_UNDEF if element of VT in the
913         // range of f32.
914         EVT FloatVT = MVT::getVectorVT(MVT::f32, VT.getVectorElementCount());
915         if (isTypeLegal(FloatVT))
916           setOperationAction(
917               {ISD::CTLZ, ISD::CTLZ_ZERO_UNDEF, ISD::CTTZ_ZERO_UNDEF}, VT,
918               Custom);
919       }
920 
921       for (MVT VT : MVT::fp_fixedlen_vector_valuetypes()) {
922         if (!useRVVForFixedLengthVectorVT(VT))
923           continue;
924 
925         // By default everything must be expanded.
926         for (unsigned Op = 0; Op < ISD::BUILTIN_OP_END; ++Op)
927           setOperationAction(Op, VT, Expand);
928         for (MVT OtherVT : MVT::fp_fixedlen_vector_valuetypes()) {
929           setLoadExtAction(ISD::EXTLOAD, OtherVT, VT, Expand);
930           setTruncStoreAction(VT, OtherVT, Expand);
931         }
932 
933         // We use EXTRACT_SUBVECTOR as a "cast" from scalable to fixed.
934         setOperationAction({ISD::INSERT_SUBVECTOR, ISD::EXTRACT_SUBVECTOR}, VT,
935                            Custom);
936 
937         setOperationAction({ISD::BUILD_VECTOR, ISD::CONCAT_VECTORS,
938                             ISD::VECTOR_SHUFFLE, ISD::INSERT_VECTOR_ELT,
939                             ISD::EXTRACT_VECTOR_ELT},
940                            VT, Custom);
941 
942         setOperationAction({ISD::LOAD, ISD::STORE, ISD::MLOAD, ISD::MSTORE,
943                             ISD::MGATHER, ISD::MSCATTER},
944                            VT, Custom);
945 
946         setOperationAction({ISD::VP_LOAD, ISD::VP_STORE,
947                             ISD::EXPERIMENTAL_VP_STRIDED_LOAD,
948                             ISD::EXPERIMENTAL_VP_STRIDED_STORE, ISD::VP_GATHER,
949                             ISD::VP_SCATTER},
950                            VT, Custom);
951 
952         setOperationAction({ISD::FADD, ISD::FSUB, ISD::FMUL, ISD::FDIV,
953                             ISD::FNEG, ISD::FABS, ISD::FCOPYSIGN, ISD::FSQRT,
954                             ISD::FMA, ISD::FMINNUM, ISD::FMAXNUM},
955                            VT, Custom);
956 
957         setOperationAction({ISD::FP_ROUND, ISD::FP_EXTEND}, VT, Custom);
958 
959         setOperationAction({ISD::FTRUNC, ISD::FCEIL, ISD::FFLOOR, ISD::FROUND,
960                             ISD::FROUNDEVEN},
961                            VT, Custom);
962 
963         setCondCodeAction(VFPCCToExpand, VT, Expand);
964 
965         setOperationAction({ISD::VSELECT, ISD::SELECT}, VT, Custom);
966         setOperationAction(ISD::SELECT_CC, VT, Expand);
967 
968         setOperationAction(ISD::BITCAST, VT, Custom);
969 
970         setOperationAction(FloatingPointVecReduceOps, VT, Custom);
971 
972         setOperationAction(FloatingPointVPOps, VT, Custom);
973       }
974 
975       // Custom-legalize bitcasts from fixed-length vectors to scalar types.
976       setOperationAction(ISD::BITCAST, {MVT::i8, MVT::i16, MVT::i32, MVT::i64},
977                          Custom);
978       if (Subtarget.hasStdExtZfhOrZfhmin())
979         setOperationAction(ISD::BITCAST, MVT::f16, Custom);
980       if (Subtarget.hasStdExtF())
981         setOperationAction(ISD::BITCAST, MVT::f32, Custom);
982       if (Subtarget.hasStdExtD())
983         setOperationAction(ISD::BITCAST, MVT::f64, Custom);
984     }
985   }
986 
987   if (Subtarget.hasForcedAtomics()) {
988     // Set atomic rmw/cas operations to expand to force __sync libcalls.
989     setOperationAction(
990         {ISD::ATOMIC_CMP_SWAP, ISD::ATOMIC_SWAP, ISD::ATOMIC_LOAD_ADD,
991          ISD::ATOMIC_LOAD_SUB, ISD::ATOMIC_LOAD_AND, ISD::ATOMIC_LOAD_OR,
992          ISD::ATOMIC_LOAD_XOR, ISD::ATOMIC_LOAD_NAND, ISD::ATOMIC_LOAD_MIN,
993          ISD::ATOMIC_LOAD_MAX, ISD::ATOMIC_LOAD_UMIN, ISD::ATOMIC_LOAD_UMAX},
994         XLenVT, Expand);
995   }
996 
997   // Function alignments.
998   const Align FunctionAlignment(Subtarget.hasStdExtCOrZca() ? 2 : 4);
999   setMinFunctionAlignment(FunctionAlignment);
1000   setPrefFunctionAlignment(FunctionAlignment);
1001 
1002   setMinimumJumpTableEntries(5);
1003 
1004   // Jumps are expensive, compared to logic
1005   setJumpIsExpensive();
1006 
1007   setTargetDAGCombine({ISD::INTRINSIC_WO_CHAIN, ISD::ADD, ISD::SUB, ISD::AND,
1008                        ISD::OR, ISD::XOR, ISD::SETCC, ISD::SELECT});
1009   if (Subtarget.is64Bit())
1010     setTargetDAGCombine(ISD::SRA);
1011 
1012   if (Subtarget.hasStdExtF())
1013     setTargetDAGCombine({ISD::FADD, ISD::FMAXNUM, ISD::FMINNUM});
1014 
1015   if (Subtarget.hasStdExtZbb())
1016     setTargetDAGCombine({ISD::UMAX, ISD::UMIN, ISD::SMAX, ISD::SMIN});
1017 
1018   if (Subtarget.hasStdExtZbs() && Subtarget.is64Bit())
1019     setTargetDAGCombine(ISD::TRUNCATE);
1020 
1021   if (Subtarget.hasStdExtZbkb())
1022     setTargetDAGCombine(ISD::BITREVERSE);
1023   if (Subtarget.hasStdExtZfhOrZfhmin())
1024     setTargetDAGCombine(ISD::SIGN_EXTEND_INREG);
1025   if (Subtarget.hasStdExtF())
1026     setTargetDAGCombine({ISD::ZERO_EXTEND, ISD::FP_TO_SINT, ISD::FP_TO_UINT,
1027                          ISD::FP_TO_SINT_SAT, ISD::FP_TO_UINT_SAT});
1028   if (Subtarget.hasVInstructions())
1029     setTargetDAGCombine({ISD::FCOPYSIGN, ISD::MGATHER, ISD::MSCATTER,
1030                          ISD::VP_GATHER, ISD::VP_SCATTER, ISD::SRA, ISD::SRL,
1031                          ISD::SHL, ISD::STORE, ISD::SPLAT_VECTOR});
1032   if (Subtarget.useRVVForFixedLengthVectors())
1033     setTargetDAGCombine(ISD::BITCAST);
1034 
1035   setLibcallName(RTLIB::FPEXT_F16_F32, "__extendhfsf2");
1036   setLibcallName(RTLIB::FPROUND_F32_F16, "__truncsfhf2");
1037 }
1038 
getSetCCResultType(const DataLayout & DL,LLVMContext & Context,EVT VT) const1039 EVT RISCVTargetLowering::getSetCCResultType(const DataLayout &DL,
1040                                             LLVMContext &Context,
1041                                             EVT VT) const {
1042   if (!VT.isVector())
1043     return getPointerTy(DL);
1044   if (Subtarget.hasVInstructions() &&
1045       (VT.isScalableVector() || Subtarget.useRVVForFixedLengthVectors()))
1046     return EVT::getVectorVT(Context, MVT::i1, VT.getVectorElementCount());
1047   return VT.changeVectorElementTypeToInteger();
1048 }
1049 
getVPExplicitVectorLengthTy() const1050 MVT RISCVTargetLowering::getVPExplicitVectorLengthTy() const {
1051   return Subtarget.getXLenVT();
1052 }
1053 
getTgtMemIntrinsic(IntrinsicInfo & Info,const CallInst & I,MachineFunction & MF,unsigned Intrinsic) const1054 bool RISCVTargetLowering::getTgtMemIntrinsic(IntrinsicInfo &Info,
1055                                              const CallInst &I,
1056                                              MachineFunction &MF,
1057                                              unsigned Intrinsic) const {
1058   auto &DL = I.getModule()->getDataLayout();
1059   switch (Intrinsic) {
1060   default:
1061     return false;
1062   case Intrinsic::riscv_masked_atomicrmw_xchg_i32:
1063   case Intrinsic::riscv_masked_atomicrmw_add_i32:
1064   case Intrinsic::riscv_masked_atomicrmw_sub_i32:
1065   case Intrinsic::riscv_masked_atomicrmw_nand_i32:
1066   case Intrinsic::riscv_masked_atomicrmw_max_i32:
1067   case Intrinsic::riscv_masked_atomicrmw_min_i32:
1068   case Intrinsic::riscv_masked_atomicrmw_umax_i32:
1069   case Intrinsic::riscv_masked_atomicrmw_umin_i32:
1070   case Intrinsic::riscv_masked_cmpxchg_i32:
1071     Info.opc = ISD::INTRINSIC_W_CHAIN;
1072     Info.memVT = MVT::i32;
1073     Info.ptrVal = I.getArgOperand(0);
1074     Info.offset = 0;
1075     Info.align = Align(4);
1076     Info.flags = MachineMemOperand::MOLoad | MachineMemOperand::MOStore |
1077                  MachineMemOperand::MOVolatile;
1078     return true;
1079   case Intrinsic::riscv_masked_strided_load:
1080     Info.opc = ISD::INTRINSIC_W_CHAIN;
1081     Info.ptrVal = I.getArgOperand(1);
1082     Info.memVT = getValueType(DL, I.getType()->getScalarType());
1083     Info.align = Align(DL.getTypeSizeInBits(I.getType()->getScalarType()) / 8);
1084     Info.size = MemoryLocation::UnknownSize;
1085     Info.flags |= MachineMemOperand::MOLoad;
1086     return true;
1087   case Intrinsic::riscv_masked_strided_store:
1088     Info.opc = ISD::INTRINSIC_VOID;
1089     Info.ptrVal = I.getArgOperand(1);
1090     Info.memVT =
1091         getValueType(DL, I.getArgOperand(0)->getType()->getScalarType());
1092     Info.align = Align(
1093         DL.getTypeSizeInBits(I.getArgOperand(0)->getType()->getScalarType()) /
1094         8);
1095     Info.size = MemoryLocation::UnknownSize;
1096     Info.flags |= MachineMemOperand::MOStore;
1097     return true;
1098   case Intrinsic::riscv_seg2_load:
1099   case Intrinsic::riscv_seg3_load:
1100   case Intrinsic::riscv_seg4_load:
1101   case Intrinsic::riscv_seg5_load:
1102   case Intrinsic::riscv_seg6_load:
1103   case Intrinsic::riscv_seg7_load:
1104   case Intrinsic::riscv_seg8_load:
1105     Info.opc = ISD::INTRINSIC_W_CHAIN;
1106     Info.ptrVal = I.getArgOperand(0);
1107     Info.memVT =
1108         getValueType(DL, I.getType()->getStructElementType(0)->getScalarType());
1109     Info.align =
1110         Align(DL.getTypeSizeInBits(
1111                   I.getType()->getStructElementType(0)->getScalarType()) /
1112               8);
1113     Info.size = MemoryLocation::UnknownSize;
1114     Info.flags |= MachineMemOperand::MOLoad;
1115     return true;
1116   }
1117 }
1118 
isLegalAddressingMode(const DataLayout & DL,const AddrMode & AM,Type * Ty,unsigned AS,Instruction * I) const1119 bool RISCVTargetLowering::isLegalAddressingMode(const DataLayout &DL,
1120                                                 const AddrMode &AM, Type *Ty,
1121                                                 unsigned AS,
1122                                                 Instruction *I) const {
1123   // No global is ever allowed as a base.
1124   if (AM.BaseGV)
1125     return false;
1126 
1127   // RVV instructions only support register addressing.
1128   if (Subtarget.hasVInstructions() && isa<VectorType>(Ty))
1129     return AM.HasBaseReg && AM.Scale == 0 && !AM.BaseOffs;
1130 
1131   // Require a 12-bit signed offset.
1132   if (!isInt<12>(AM.BaseOffs))
1133     return false;
1134 
1135   switch (AM.Scale) {
1136   case 0: // "r+i" or just "i", depending on HasBaseReg.
1137     break;
1138   case 1:
1139     if (!AM.HasBaseReg) // allow "r+i".
1140       break;
1141     return false; // disallow "r+r" or "r+r+i".
1142   default:
1143     return false;
1144   }
1145 
1146   return true;
1147 }
1148 
isLegalICmpImmediate(int64_t Imm) const1149 bool RISCVTargetLowering::isLegalICmpImmediate(int64_t Imm) const {
1150   return isInt<12>(Imm);
1151 }
1152 
isLegalAddImmediate(int64_t Imm) const1153 bool RISCVTargetLowering::isLegalAddImmediate(int64_t Imm) const {
1154   return isInt<12>(Imm);
1155 }
1156 
1157 // On RV32, 64-bit integers are split into their high and low parts and held
1158 // in two different registers, so the trunc is free since the low register can
1159 // just be used.
1160 // FIXME: Should we consider i64->i32 free on RV64 to match the EVT version of
1161 // isTruncateFree?
isTruncateFree(Type * SrcTy,Type * DstTy) const1162 bool RISCVTargetLowering::isTruncateFree(Type *SrcTy, Type *DstTy) const {
1163   if (Subtarget.is64Bit() || !SrcTy->isIntegerTy() || !DstTy->isIntegerTy())
1164     return false;
1165   unsigned SrcBits = SrcTy->getPrimitiveSizeInBits();
1166   unsigned DestBits = DstTy->getPrimitiveSizeInBits();
1167   return (SrcBits == 64 && DestBits == 32);
1168 }
1169 
isTruncateFree(EVT SrcVT,EVT DstVT) const1170 bool RISCVTargetLowering::isTruncateFree(EVT SrcVT, EVT DstVT) const {
1171   // We consider i64->i32 free on RV64 since we have good selection of W
1172   // instructions that make promoting operations back to i64 free in many cases.
1173   if (SrcVT.isVector() || DstVT.isVector() || !SrcVT.isInteger() ||
1174       !DstVT.isInteger())
1175     return false;
1176   unsigned SrcBits = SrcVT.getSizeInBits();
1177   unsigned DestBits = DstVT.getSizeInBits();
1178   return (SrcBits == 64 && DestBits == 32);
1179 }
1180 
isZExtFree(SDValue Val,EVT VT2) const1181 bool RISCVTargetLowering::isZExtFree(SDValue Val, EVT VT2) const {
1182   // Zexts are free if they can be combined with a load.
1183   // Don't advertise i32->i64 zextload as being free for RV64. It interacts
1184   // poorly with type legalization of compares preferring sext.
1185   if (auto *LD = dyn_cast<LoadSDNode>(Val)) {
1186     EVT MemVT = LD->getMemoryVT();
1187     if ((MemVT == MVT::i8 || MemVT == MVT::i16) &&
1188         (LD->getExtensionType() == ISD::NON_EXTLOAD ||
1189          LD->getExtensionType() == ISD::ZEXTLOAD))
1190       return true;
1191   }
1192 
1193   return TargetLowering::isZExtFree(Val, VT2);
1194 }
1195 
isSExtCheaperThanZExt(EVT SrcVT,EVT DstVT) const1196 bool RISCVTargetLowering::isSExtCheaperThanZExt(EVT SrcVT, EVT DstVT) const {
1197   return Subtarget.is64Bit() && SrcVT == MVT::i32 && DstVT == MVT::i64;
1198 }
1199 
signExtendConstant(const ConstantInt * CI) const1200 bool RISCVTargetLowering::signExtendConstant(const ConstantInt *CI) const {
1201   return Subtarget.is64Bit() && CI->getType()->isIntegerTy(32);
1202 }
1203 
isCheapToSpeculateCttz(Type * Ty) const1204 bool RISCVTargetLowering::isCheapToSpeculateCttz(Type *Ty) const {
1205   return Subtarget.hasStdExtZbb();
1206 }
1207 
isCheapToSpeculateCtlz(Type * Ty) const1208 bool RISCVTargetLowering::isCheapToSpeculateCtlz(Type *Ty) const {
1209   return Subtarget.hasStdExtZbb();
1210 }
1211 
isMaskAndCmp0FoldingBeneficial(const Instruction & AndI) const1212 bool RISCVTargetLowering::isMaskAndCmp0FoldingBeneficial(
1213     const Instruction &AndI) const {
1214   // We expect to be able to match a bit extraction instruction if the Zbs
1215   // extension is supported and the mask is a power of two. However, we
1216   // conservatively return false if the mask would fit in an ANDI instruction,
1217   // on the basis that it's possible the sinking+duplication of the AND in
1218   // CodeGenPrepare triggered by this hook wouldn't decrease the instruction
1219   // count and would increase code size (e.g. ANDI+BNEZ => BEXTI+BNEZ).
1220   if (!Subtarget.hasStdExtZbs())
1221     return false;
1222   ConstantInt *Mask = dyn_cast<ConstantInt>(AndI.getOperand(1));
1223   if (!Mask)
1224     return false;
1225   return !Mask->getValue().isSignedIntN(12) && Mask->getValue().isPowerOf2();
1226 }
1227 
hasAndNotCompare(SDValue Y) const1228 bool RISCVTargetLowering::hasAndNotCompare(SDValue Y) const {
1229   EVT VT = Y.getValueType();
1230 
1231   // FIXME: Support vectors once we have tests.
1232   if (VT.isVector())
1233     return false;
1234 
1235   return (Subtarget.hasStdExtZbb() || Subtarget.hasStdExtZbkb()) &&
1236          !isa<ConstantSDNode>(Y);
1237 }
1238 
hasBitTest(SDValue X,SDValue Y) const1239 bool RISCVTargetLowering::hasBitTest(SDValue X, SDValue Y) const {
1240   // Zbs provides BEXT[_I], which can be used with SEQZ/SNEZ as a bit test.
1241   if (Subtarget.hasStdExtZbs())
1242     return X.getValueType().isScalarInteger();
1243   // We can use ANDI+SEQZ/SNEZ as a bit test. Y contains the bit position.
1244   auto *C = dyn_cast<ConstantSDNode>(Y);
1245   return C && C->getAPIntValue().ule(10);
1246 }
1247 
shouldFoldSelectWithIdentityConstant(unsigned Opcode,EVT VT) const1248 bool RISCVTargetLowering::shouldFoldSelectWithIdentityConstant(unsigned Opcode,
1249                                                                EVT VT) const {
1250   // Only enable for rvv.
1251   if (!VT.isVector() || !Subtarget.hasVInstructions())
1252     return false;
1253 
1254   if (VT.isFixedLengthVector() && !isTypeLegal(VT))
1255     return false;
1256 
1257   return true;
1258 }
1259 
shouldConvertConstantLoadToIntImm(const APInt & Imm,Type * Ty) const1260 bool RISCVTargetLowering::shouldConvertConstantLoadToIntImm(const APInt &Imm,
1261                                                             Type *Ty) const {
1262   assert(Ty->isIntegerTy());
1263 
1264   unsigned BitSize = Ty->getIntegerBitWidth();
1265   if (BitSize > Subtarget.getXLen())
1266     return false;
1267 
1268   // Fast path, assume 32-bit immediates are cheap.
1269   int64_t Val = Imm.getSExtValue();
1270   if (isInt<32>(Val))
1271     return true;
1272 
1273   // A constant pool entry may be more aligned thant he load we're trying to
1274   // replace. If we don't support unaligned scalar mem, prefer the constant
1275   // pool.
1276   // TODO: Can the caller pass down the alignment?
1277   if (!Subtarget.enableUnalignedScalarMem())
1278     return true;
1279 
1280   // Prefer to keep the load if it would require many instructions.
1281   // This uses the same threshold we use for constant pools but doesn't
1282   // check useConstantPoolForLargeInts.
1283   // TODO: Should we keep the load only when we're definitely going to emit a
1284   // constant pool?
1285 
1286   RISCVMatInt::InstSeq Seq =
1287       RISCVMatInt::generateInstSeq(Val, Subtarget.getFeatureBits());
1288   return Seq.size() <= Subtarget.getMaxBuildIntsCost();
1289 }
1290 
1291 bool RISCVTargetLowering::
shouldProduceAndByConstByHoistingConstFromShiftsLHSOfAnd(SDValue X,ConstantSDNode * XC,ConstantSDNode * CC,SDValue Y,unsigned OldShiftOpcode,unsigned NewShiftOpcode,SelectionDAG & DAG) const1292     shouldProduceAndByConstByHoistingConstFromShiftsLHSOfAnd(
1293         SDValue X, ConstantSDNode *XC, ConstantSDNode *CC, SDValue Y,
1294         unsigned OldShiftOpcode, unsigned NewShiftOpcode,
1295         SelectionDAG &DAG) const {
1296   // One interesting pattern that we'd want to form is 'bit extract':
1297   //   ((1 >> Y) & 1) ==/!= 0
1298   // But we also need to be careful not to try to reverse that fold.
1299 
1300   // Is this '((1 >> Y) & 1)'?
1301   if (XC && OldShiftOpcode == ISD::SRL && XC->isOne())
1302     return false; // Keep the 'bit extract' pattern.
1303 
1304   // Will this be '((1 >> Y) & 1)' after the transform?
1305   if (NewShiftOpcode == ISD::SRL && CC->isOne())
1306     return true; // Do form the 'bit extract' pattern.
1307 
1308   // If 'X' is a constant, and we transform, then we will immediately
1309   // try to undo the fold, thus causing endless combine loop.
1310   // So only do the transform if X is not a constant. This matches the default
1311   // implementation of this function.
1312   return !XC;
1313 }
1314 
canSplatOperand(unsigned Opcode,int Operand) const1315 bool RISCVTargetLowering::canSplatOperand(unsigned Opcode, int Operand) const {
1316   switch (Opcode) {
1317   case Instruction::Add:
1318   case Instruction::Sub:
1319   case Instruction::Mul:
1320   case Instruction::And:
1321   case Instruction::Or:
1322   case Instruction::Xor:
1323   case Instruction::FAdd:
1324   case Instruction::FSub:
1325   case Instruction::FMul:
1326   case Instruction::FDiv:
1327   case Instruction::ICmp:
1328   case Instruction::FCmp:
1329     return true;
1330   case Instruction::Shl:
1331   case Instruction::LShr:
1332   case Instruction::AShr:
1333   case Instruction::UDiv:
1334   case Instruction::SDiv:
1335   case Instruction::URem:
1336   case Instruction::SRem:
1337     return Operand == 1;
1338   default:
1339     return false;
1340   }
1341 }
1342 
1343 
canSplatOperand(Instruction * I,int Operand) const1344 bool RISCVTargetLowering::canSplatOperand(Instruction *I, int Operand) const {
1345   if (!I->getType()->isVectorTy() || !Subtarget.hasVInstructions())
1346     return false;
1347 
1348   if (canSplatOperand(I->getOpcode(), Operand))
1349     return true;
1350 
1351   auto *II = dyn_cast<IntrinsicInst>(I);
1352   if (!II)
1353     return false;
1354 
1355   switch (II->getIntrinsicID()) {
1356   case Intrinsic::fma:
1357   case Intrinsic::vp_fma:
1358     return Operand == 0 || Operand == 1;
1359   case Intrinsic::vp_shl:
1360   case Intrinsic::vp_lshr:
1361   case Intrinsic::vp_ashr:
1362   case Intrinsic::vp_udiv:
1363   case Intrinsic::vp_sdiv:
1364   case Intrinsic::vp_urem:
1365   case Intrinsic::vp_srem:
1366     return Operand == 1;
1367     // These intrinsics are commutative.
1368   case Intrinsic::vp_add:
1369   case Intrinsic::vp_mul:
1370   case Intrinsic::vp_and:
1371   case Intrinsic::vp_or:
1372   case Intrinsic::vp_xor:
1373   case Intrinsic::vp_fadd:
1374   case Intrinsic::vp_fmul:
1375     // These intrinsics have 'vr' versions.
1376   case Intrinsic::vp_sub:
1377   case Intrinsic::vp_fsub:
1378   case Intrinsic::vp_fdiv:
1379     return Operand == 0 || Operand == 1;
1380   default:
1381     return false;
1382   }
1383 }
1384 
1385 /// Check if sinking \p I's operands to I's basic block is profitable, because
1386 /// the operands can be folded into a target instruction, e.g.
1387 /// splats of scalars can fold into vector instructions.
shouldSinkOperands(Instruction * I,SmallVectorImpl<Use * > & Ops) const1388 bool RISCVTargetLowering::shouldSinkOperands(
1389     Instruction *I, SmallVectorImpl<Use *> &Ops) const {
1390   using namespace llvm::PatternMatch;
1391 
1392   if (!I->getType()->isVectorTy() || !Subtarget.hasVInstructions())
1393     return false;
1394 
1395   for (auto OpIdx : enumerate(I->operands())) {
1396     if (!canSplatOperand(I, OpIdx.index()))
1397       continue;
1398 
1399     Instruction *Op = dyn_cast<Instruction>(OpIdx.value().get());
1400     // Make sure we are not already sinking this operand
1401     if (!Op || any_of(Ops, [&](Use *U) { return U->get() == Op; }))
1402       continue;
1403 
1404     // We are looking for a splat that can be sunk.
1405     if (!match(Op, m_Shuffle(m_InsertElt(m_Undef(), m_Value(), m_ZeroInt()),
1406                              m_Undef(), m_ZeroMask())))
1407       continue;
1408 
1409     // All uses of the shuffle should be sunk to avoid duplicating it across gpr
1410     // and vector registers
1411     for (Use &U : Op->uses()) {
1412       Instruction *Insn = cast<Instruction>(U.getUser());
1413       if (!canSplatOperand(Insn, U.getOperandNo()))
1414         return false;
1415     }
1416 
1417     Ops.push_back(&Op->getOperandUse(0));
1418     Ops.push_back(&OpIdx.value());
1419   }
1420   return true;
1421 }
1422 
shouldScalarizeBinop(SDValue VecOp) const1423 bool RISCVTargetLowering::shouldScalarizeBinop(SDValue VecOp) const {
1424   unsigned Opc = VecOp.getOpcode();
1425 
1426   // Assume target opcodes can't be scalarized.
1427   // TODO - do we have any exceptions?
1428   if (Opc >= ISD::BUILTIN_OP_END)
1429     return false;
1430 
1431   // If the vector op is not supported, try to convert to scalar.
1432   EVT VecVT = VecOp.getValueType();
1433   if (!isOperationLegalOrCustomOrPromote(Opc, VecVT))
1434     return true;
1435 
1436   // If the vector op is supported, but the scalar op is not, the transform may
1437   // not be worthwhile.
1438   EVT ScalarVT = VecVT.getScalarType();
1439   return isOperationLegalOrCustomOrPromote(Opc, ScalarVT);
1440 }
1441 
isOffsetFoldingLegal(const GlobalAddressSDNode * GA) const1442 bool RISCVTargetLowering::isOffsetFoldingLegal(
1443     const GlobalAddressSDNode *GA) const {
1444   // In order to maximise the opportunity for common subexpression elimination,
1445   // keep a separate ADD node for the global address offset instead of folding
1446   // it in the global address node. Later peephole optimisations may choose to
1447   // fold it back in when profitable.
1448   return false;
1449 }
1450 
isFPImmLegal(const APFloat & Imm,EVT VT,bool ForCodeSize) const1451 bool RISCVTargetLowering::isFPImmLegal(const APFloat &Imm, EVT VT,
1452                                        bool ForCodeSize) const {
1453   if (VT == MVT::f16 && !Subtarget.hasStdExtZfhOrZfhmin())
1454     return false;
1455   if (VT == MVT::f32 && !Subtarget.hasStdExtF())
1456     return false;
1457   if (VT == MVT::f64 && !Subtarget.hasStdExtD())
1458     return false;
1459   return Imm.isZero();
1460 }
1461 
1462 // TODO: This is very conservative.
isExtractSubvectorCheap(EVT ResVT,EVT SrcVT,unsigned Index) const1463 bool RISCVTargetLowering::isExtractSubvectorCheap(EVT ResVT, EVT SrcVT,
1464                                                   unsigned Index) const {
1465   if (!isOperationLegalOrCustom(ISD::EXTRACT_SUBVECTOR, ResVT))
1466     return false;
1467 
1468   // Only support extracting a fixed from a fixed vector for now.
1469   if (ResVT.isScalableVector() || SrcVT.isScalableVector())
1470     return false;
1471 
1472   unsigned ResElts = ResVT.getVectorNumElements();
1473   unsigned SrcElts = SrcVT.getVectorNumElements();
1474 
1475   // Convervatively only handle extracting half of a vector.
1476   // TODO: Relax this.
1477   if ((ResElts * 2) != SrcElts)
1478     return false;
1479 
1480   // The smallest type we can slide is i8.
1481   // TODO: We can extract index 0 from a mask vector without a slide.
1482   if (ResVT.getVectorElementType() == MVT::i1)
1483     return false;
1484 
1485   // Slide can support arbitrary index, but we only treat vslidedown.vi as
1486   // cheap.
1487   if (Index >= 32)
1488     return false;
1489 
1490   // TODO: We can do arbitrary slidedowns, but for now only support extracting
1491   // the upper half of a vector until we have more test coverage.
1492   return Index == 0 || Index == ResElts;
1493 }
1494 
hasBitPreservingFPLogic(EVT VT) const1495 bool RISCVTargetLowering::hasBitPreservingFPLogic(EVT VT) const {
1496   return (VT == MVT::f16 && Subtarget.hasStdExtZfhOrZfhmin()) ||
1497          (VT == MVT::f32 && Subtarget.hasStdExtF()) ||
1498          (VT == MVT::f64 && Subtarget.hasStdExtD());
1499 }
1500 
getRegisterTypeForCallingConv(LLVMContext & Context,CallingConv::ID CC,EVT VT) const1501 MVT RISCVTargetLowering::getRegisterTypeForCallingConv(LLVMContext &Context,
1502                                                       CallingConv::ID CC,
1503                                                       EVT VT) const {
1504   // Use f32 to pass f16 if it is legal and Zfh/Zfhmin is not enabled.
1505   // We might still end up using a GPR but that will be decided based on ABI.
1506   if (VT == MVT::f16 && Subtarget.hasStdExtF() &&
1507       !Subtarget.hasStdExtZfhOrZfhmin())
1508     return MVT::f32;
1509 
1510   return TargetLowering::getRegisterTypeForCallingConv(Context, CC, VT);
1511 }
1512 
getNumRegistersForCallingConv(LLVMContext & Context,CallingConv::ID CC,EVT VT) const1513 unsigned RISCVTargetLowering::getNumRegistersForCallingConv(LLVMContext &Context,
1514                                                            CallingConv::ID CC,
1515                                                            EVT VT) const {
1516   // Use f32 to pass f16 if it is legal and Zfh/Zfhmin is not enabled.
1517   // We might still end up using a GPR but that will be decided based on ABI.
1518   if (VT == MVT::f16 && Subtarget.hasStdExtF() &&
1519       !Subtarget.hasStdExtZfhOrZfhmin())
1520     return 1;
1521 
1522   return TargetLowering::getNumRegistersForCallingConv(Context, CC, VT);
1523 }
1524 
1525 // Changes the condition code and swaps operands if necessary, so the SetCC
1526 // operation matches one of the comparisons supported directly by branches
1527 // in the RISC-V ISA. May adjust compares to favor compare with 0 over compare
1528 // with 1/-1.
translateSetCCForBranch(const SDLoc & DL,SDValue & LHS,SDValue & RHS,ISD::CondCode & CC,SelectionDAG & DAG)1529 static void translateSetCCForBranch(const SDLoc &DL, SDValue &LHS, SDValue &RHS,
1530                                     ISD::CondCode &CC, SelectionDAG &DAG) {
1531   // If this is a single bit test that can't be handled by ANDI, shift the
1532   // bit to be tested to the MSB and perform a signed compare with 0.
1533   if (isIntEqualitySetCC(CC) && isNullConstant(RHS) &&
1534       LHS.getOpcode() == ISD::AND && LHS.hasOneUse() &&
1535       isa<ConstantSDNode>(LHS.getOperand(1))) {
1536     uint64_t Mask = LHS.getConstantOperandVal(1);
1537     if ((isPowerOf2_64(Mask) || isMask_64(Mask)) && !isInt<12>(Mask)) {
1538       unsigned ShAmt = 0;
1539       if (isPowerOf2_64(Mask)) {
1540         CC = CC == ISD::SETEQ ? ISD::SETGE : ISD::SETLT;
1541         ShAmt = LHS.getValueSizeInBits() - 1 - Log2_64(Mask);
1542       } else {
1543         ShAmt = LHS.getValueSizeInBits() - llvm::bit_width(Mask);
1544       }
1545 
1546       LHS = LHS.getOperand(0);
1547       if (ShAmt != 0)
1548         LHS = DAG.getNode(ISD::SHL, DL, LHS.getValueType(), LHS,
1549                           DAG.getConstant(ShAmt, DL, LHS.getValueType()));
1550       return;
1551     }
1552   }
1553 
1554   if (auto *RHSC = dyn_cast<ConstantSDNode>(RHS)) {
1555     int64_t C = RHSC->getSExtValue();
1556     switch (CC) {
1557     default: break;
1558     case ISD::SETGT:
1559       // Convert X > -1 to X >= 0.
1560       if (C == -1) {
1561         RHS = DAG.getConstant(0, DL, RHS.getValueType());
1562         CC = ISD::SETGE;
1563         return;
1564       }
1565       break;
1566     case ISD::SETLT:
1567       // Convert X < 1 to 0 <= X.
1568       if (C == 1) {
1569         RHS = LHS;
1570         LHS = DAG.getConstant(0, DL, RHS.getValueType());
1571         CC = ISD::SETGE;
1572         return;
1573       }
1574       break;
1575     }
1576   }
1577 
1578   switch (CC) {
1579   default:
1580     break;
1581   case ISD::SETGT:
1582   case ISD::SETLE:
1583   case ISD::SETUGT:
1584   case ISD::SETULE:
1585     CC = ISD::getSetCCSwappedOperands(CC);
1586     std::swap(LHS, RHS);
1587     break;
1588   }
1589 }
1590 
getLMUL(MVT VT)1591 RISCVII::VLMUL RISCVTargetLowering::getLMUL(MVT VT) {
1592   assert(VT.isScalableVector() && "Expecting a scalable vector type");
1593   unsigned KnownSize = VT.getSizeInBits().getKnownMinValue();
1594   if (VT.getVectorElementType() == MVT::i1)
1595     KnownSize *= 8;
1596 
1597   switch (KnownSize) {
1598   default:
1599     llvm_unreachable("Invalid LMUL.");
1600   case 8:
1601     return RISCVII::VLMUL::LMUL_F8;
1602   case 16:
1603     return RISCVII::VLMUL::LMUL_F4;
1604   case 32:
1605     return RISCVII::VLMUL::LMUL_F2;
1606   case 64:
1607     return RISCVII::VLMUL::LMUL_1;
1608   case 128:
1609     return RISCVII::VLMUL::LMUL_2;
1610   case 256:
1611     return RISCVII::VLMUL::LMUL_4;
1612   case 512:
1613     return RISCVII::VLMUL::LMUL_8;
1614   }
1615 }
1616 
getRegClassIDForLMUL(RISCVII::VLMUL LMul)1617 unsigned RISCVTargetLowering::getRegClassIDForLMUL(RISCVII::VLMUL LMul) {
1618   switch (LMul) {
1619   default:
1620     llvm_unreachable("Invalid LMUL.");
1621   case RISCVII::VLMUL::LMUL_F8:
1622   case RISCVII::VLMUL::LMUL_F4:
1623   case RISCVII::VLMUL::LMUL_F2:
1624   case RISCVII::VLMUL::LMUL_1:
1625     return RISCV::VRRegClassID;
1626   case RISCVII::VLMUL::LMUL_2:
1627     return RISCV::VRM2RegClassID;
1628   case RISCVII::VLMUL::LMUL_4:
1629     return RISCV::VRM4RegClassID;
1630   case RISCVII::VLMUL::LMUL_8:
1631     return RISCV::VRM8RegClassID;
1632   }
1633 }
1634 
getSubregIndexByMVT(MVT VT,unsigned Index)1635 unsigned RISCVTargetLowering::getSubregIndexByMVT(MVT VT, unsigned Index) {
1636   RISCVII::VLMUL LMUL = getLMUL(VT);
1637   if (LMUL == RISCVII::VLMUL::LMUL_F8 ||
1638       LMUL == RISCVII::VLMUL::LMUL_F4 ||
1639       LMUL == RISCVII::VLMUL::LMUL_F2 ||
1640       LMUL == RISCVII::VLMUL::LMUL_1) {
1641     static_assert(RISCV::sub_vrm1_7 == RISCV::sub_vrm1_0 + 7,
1642                   "Unexpected subreg numbering");
1643     return RISCV::sub_vrm1_0 + Index;
1644   }
1645   if (LMUL == RISCVII::VLMUL::LMUL_2) {
1646     static_assert(RISCV::sub_vrm2_3 == RISCV::sub_vrm2_0 + 3,
1647                   "Unexpected subreg numbering");
1648     return RISCV::sub_vrm2_0 + Index;
1649   }
1650   if (LMUL == RISCVII::VLMUL::LMUL_4) {
1651     static_assert(RISCV::sub_vrm4_1 == RISCV::sub_vrm4_0 + 1,
1652                   "Unexpected subreg numbering");
1653     return RISCV::sub_vrm4_0 + Index;
1654   }
1655   llvm_unreachable("Invalid vector type.");
1656 }
1657 
getRegClassIDForVecVT(MVT VT)1658 unsigned RISCVTargetLowering::getRegClassIDForVecVT(MVT VT) {
1659   if (VT.getVectorElementType() == MVT::i1)
1660     return RISCV::VRRegClassID;
1661   return getRegClassIDForLMUL(getLMUL(VT));
1662 }
1663 
1664 // Attempt to decompose a subvector insert/extract between VecVT and
1665 // SubVecVT via subregister indices. Returns the subregister index that
1666 // can perform the subvector insert/extract with the given element index, as
1667 // well as the index corresponding to any leftover subvectors that must be
1668 // further inserted/extracted within the register class for SubVecVT.
1669 std::pair<unsigned, unsigned>
decomposeSubvectorInsertExtractToSubRegs(MVT VecVT,MVT SubVecVT,unsigned InsertExtractIdx,const RISCVRegisterInfo * TRI)1670 RISCVTargetLowering::decomposeSubvectorInsertExtractToSubRegs(
1671     MVT VecVT, MVT SubVecVT, unsigned InsertExtractIdx,
1672     const RISCVRegisterInfo *TRI) {
1673   static_assert((RISCV::VRM8RegClassID > RISCV::VRM4RegClassID &&
1674                  RISCV::VRM4RegClassID > RISCV::VRM2RegClassID &&
1675                  RISCV::VRM2RegClassID > RISCV::VRRegClassID),
1676                 "Register classes not ordered");
1677   unsigned VecRegClassID = getRegClassIDForVecVT(VecVT);
1678   unsigned SubRegClassID = getRegClassIDForVecVT(SubVecVT);
1679   // Try to compose a subregister index that takes us from the incoming
1680   // LMUL>1 register class down to the outgoing one. At each step we half
1681   // the LMUL:
1682   //   nxv16i32@12 -> nxv2i32: sub_vrm4_1_then_sub_vrm2_1_then_sub_vrm1_0
1683   // Note that this is not guaranteed to find a subregister index, such as
1684   // when we are extracting from one VR type to another.
1685   unsigned SubRegIdx = RISCV::NoSubRegister;
1686   for (const unsigned RCID :
1687        {RISCV::VRM4RegClassID, RISCV::VRM2RegClassID, RISCV::VRRegClassID})
1688     if (VecRegClassID > RCID && SubRegClassID <= RCID) {
1689       VecVT = VecVT.getHalfNumVectorElementsVT();
1690       bool IsHi =
1691           InsertExtractIdx >= VecVT.getVectorElementCount().getKnownMinValue();
1692       SubRegIdx = TRI->composeSubRegIndices(SubRegIdx,
1693                                             getSubregIndexByMVT(VecVT, IsHi));
1694       if (IsHi)
1695         InsertExtractIdx -= VecVT.getVectorElementCount().getKnownMinValue();
1696     }
1697   return {SubRegIdx, InsertExtractIdx};
1698 }
1699 
1700 // Permit combining of mask vectors as BUILD_VECTOR never expands to scalar
1701 // stores for those types.
mergeStoresAfterLegalization(EVT VT) const1702 bool RISCVTargetLowering::mergeStoresAfterLegalization(EVT VT) const {
1703   return !Subtarget.useRVVForFixedLengthVectors() ||
1704          (VT.isFixedLengthVector() && VT.getVectorElementType() == MVT::i1);
1705 }
1706 
isLegalElementTypeForRVV(Type * ScalarTy) const1707 bool RISCVTargetLowering::isLegalElementTypeForRVV(Type *ScalarTy) const {
1708   if (ScalarTy->isPointerTy())
1709     return true;
1710 
1711   if (ScalarTy->isIntegerTy(8) || ScalarTy->isIntegerTy(16) ||
1712       ScalarTy->isIntegerTy(32))
1713     return true;
1714 
1715   if (ScalarTy->isIntegerTy(64))
1716     return Subtarget.hasVInstructionsI64();
1717 
1718   if (ScalarTy->isHalfTy())
1719     return Subtarget.hasVInstructionsF16();
1720   if (ScalarTy->isFloatTy())
1721     return Subtarget.hasVInstructionsF32();
1722   if (ScalarTy->isDoubleTy())
1723     return Subtarget.hasVInstructionsF64();
1724 
1725   return false;
1726 }
1727 
combineRepeatedFPDivisors() const1728 unsigned RISCVTargetLowering::combineRepeatedFPDivisors() const {
1729   return NumRepeatedDivisors;
1730 }
1731 
getVLOperand(SDValue Op)1732 static SDValue getVLOperand(SDValue Op) {
1733   assert((Op.getOpcode() == ISD::INTRINSIC_WO_CHAIN ||
1734           Op.getOpcode() == ISD::INTRINSIC_W_CHAIN) &&
1735          "Unexpected opcode");
1736   bool HasChain = Op.getOpcode() == ISD::INTRINSIC_W_CHAIN;
1737   unsigned IntNo = Op.getConstantOperandVal(HasChain ? 1 : 0);
1738   const RISCVVIntrinsicsTable::RISCVVIntrinsicInfo *II =
1739       RISCVVIntrinsicsTable::getRISCVVIntrinsicInfo(IntNo);
1740   if (!II)
1741     return SDValue();
1742   return Op.getOperand(II->VLOperand + 1 + HasChain);
1743 }
1744 
useRVVForFixedLengthVectorVT(MVT VT,const RISCVSubtarget & Subtarget)1745 static bool useRVVForFixedLengthVectorVT(MVT VT,
1746                                          const RISCVSubtarget &Subtarget) {
1747   assert(VT.isFixedLengthVector() && "Expected a fixed length vector type!");
1748   if (!Subtarget.useRVVForFixedLengthVectors())
1749     return false;
1750 
1751   // We only support a set of vector types with a consistent maximum fixed size
1752   // across all supported vector element types to avoid legalization issues.
1753   // Therefore -- since the largest is v1024i8/v512i16/etc -- the largest
1754   // fixed-length vector type we support is 1024 bytes.
1755   if (VT.getFixedSizeInBits() > 1024 * 8)
1756     return false;
1757 
1758   unsigned MinVLen = Subtarget.getRealMinVLen();
1759 
1760   MVT EltVT = VT.getVectorElementType();
1761 
1762   // Don't use RVV for vectors we cannot scalarize if required.
1763   switch (EltVT.SimpleTy) {
1764   // i1 is supported but has different rules.
1765   default:
1766     return false;
1767   case MVT::i1:
1768     // Masks can only use a single register.
1769     if (VT.getVectorNumElements() > MinVLen)
1770       return false;
1771     MinVLen /= 8;
1772     break;
1773   case MVT::i8:
1774   case MVT::i16:
1775   case MVT::i32:
1776     break;
1777   case MVT::i64:
1778     if (!Subtarget.hasVInstructionsI64())
1779       return false;
1780     break;
1781   case MVT::f16:
1782     if (!Subtarget.hasVInstructionsF16())
1783       return false;
1784     break;
1785   case MVT::f32:
1786     if (!Subtarget.hasVInstructionsF32())
1787       return false;
1788     break;
1789   case MVT::f64:
1790     if (!Subtarget.hasVInstructionsF64())
1791       return false;
1792     break;
1793   }
1794 
1795   // Reject elements larger than ELEN.
1796   if (EltVT.getSizeInBits() > Subtarget.getELEN())
1797     return false;
1798 
1799   unsigned LMul = divideCeil(VT.getSizeInBits(), MinVLen);
1800   // Don't use RVV for types that don't fit.
1801   if (LMul > Subtarget.getMaxLMULForFixedLengthVectors())
1802     return false;
1803 
1804   // TODO: Perhaps an artificial restriction, but worth having whilst getting
1805   // the base fixed length RVV support in place.
1806   if (!VT.isPow2VectorType())
1807     return false;
1808 
1809   return true;
1810 }
1811 
useRVVForFixedLengthVectorVT(MVT VT) const1812 bool RISCVTargetLowering::useRVVForFixedLengthVectorVT(MVT VT) const {
1813   return ::useRVVForFixedLengthVectorVT(VT, Subtarget);
1814 }
1815 
1816 // Return the largest legal scalable vector type that matches VT's element type.
getContainerForFixedLengthVector(const TargetLowering & TLI,MVT VT,const RISCVSubtarget & Subtarget)1817 static MVT getContainerForFixedLengthVector(const TargetLowering &TLI, MVT VT,
1818                                             const RISCVSubtarget &Subtarget) {
1819   // This may be called before legal types are setup.
1820   assert(((VT.isFixedLengthVector() && TLI.isTypeLegal(VT)) ||
1821           useRVVForFixedLengthVectorVT(VT, Subtarget)) &&
1822          "Expected legal fixed length vector!");
1823 
1824   unsigned MinVLen = Subtarget.getRealMinVLen();
1825   unsigned MaxELen = Subtarget.getELEN();
1826 
1827   MVT EltVT = VT.getVectorElementType();
1828   switch (EltVT.SimpleTy) {
1829   default:
1830     llvm_unreachable("unexpected element type for RVV container");
1831   case MVT::i1:
1832   case MVT::i8:
1833   case MVT::i16:
1834   case MVT::i32:
1835   case MVT::i64:
1836   case MVT::f16:
1837   case MVT::f32:
1838   case MVT::f64: {
1839     // We prefer to use LMUL=1 for VLEN sized types. Use fractional lmuls for
1840     // narrower types. The smallest fractional LMUL we support is 8/ELEN. Within
1841     // each fractional LMUL we support SEW between 8 and LMUL*ELEN.
1842     unsigned NumElts =
1843         (VT.getVectorNumElements() * RISCV::RVVBitsPerBlock) / MinVLen;
1844     NumElts = std::max(NumElts, RISCV::RVVBitsPerBlock / MaxELen);
1845     assert(isPowerOf2_32(NumElts) && "Expected power of 2 NumElts");
1846     return MVT::getScalableVectorVT(EltVT, NumElts);
1847   }
1848   }
1849 }
1850 
getContainerForFixedLengthVector(SelectionDAG & DAG,MVT VT,const RISCVSubtarget & Subtarget)1851 static MVT getContainerForFixedLengthVector(SelectionDAG &DAG, MVT VT,
1852                                             const RISCVSubtarget &Subtarget) {
1853   return getContainerForFixedLengthVector(DAG.getTargetLoweringInfo(), VT,
1854                                           Subtarget);
1855 }
1856 
getContainerForFixedLengthVector(MVT VT) const1857 MVT RISCVTargetLowering::getContainerForFixedLengthVector(MVT VT) const {
1858   return ::getContainerForFixedLengthVector(*this, VT, getSubtarget());
1859 }
1860 
1861 // Grow V to consume an entire RVV register.
convertToScalableVector(EVT VT,SDValue V,SelectionDAG & DAG,const RISCVSubtarget & Subtarget)1862 static SDValue convertToScalableVector(EVT VT, SDValue V, SelectionDAG &DAG,
1863                                        const RISCVSubtarget &Subtarget) {
1864   assert(VT.isScalableVector() &&
1865          "Expected to convert into a scalable vector!");
1866   assert(V.getValueType().isFixedLengthVector() &&
1867          "Expected a fixed length vector operand!");
1868   SDLoc DL(V);
1869   SDValue Zero = DAG.getConstant(0, DL, Subtarget.getXLenVT());
1870   return DAG.getNode(ISD::INSERT_SUBVECTOR, DL, VT, DAG.getUNDEF(VT), V, Zero);
1871 }
1872 
1873 // Shrink V so it's just big enough to maintain a VT's worth of data.
convertFromScalableVector(EVT VT,SDValue V,SelectionDAG & DAG,const RISCVSubtarget & Subtarget)1874 static SDValue convertFromScalableVector(EVT VT, SDValue V, SelectionDAG &DAG,
1875                                          const RISCVSubtarget &Subtarget) {
1876   assert(VT.isFixedLengthVector() &&
1877          "Expected to convert into a fixed length vector!");
1878   assert(V.getValueType().isScalableVector() &&
1879          "Expected a scalable vector operand!");
1880   SDLoc DL(V);
1881   SDValue Zero = DAG.getConstant(0, DL, Subtarget.getXLenVT());
1882   return DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, VT, V, Zero);
1883 }
1884 
1885 /// Return the type of the mask type suitable for masking the provided
1886 /// vector type.  This is simply an i1 element type vector of the same
1887 /// (possibly scalable) length.
getMaskTypeFor(MVT VecVT)1888 static MVT getMaskTypeFor(MVT VecVT) {
1889   assert(VecVT.isVector());
1890   ElementCount EC = VecVT.getVectorElementCount();
1891   return MVT::getVectorVT(MVT::i1, EC);
1892 }
1893 
1894 /// Creates an all ones mask suitable for masking a vector of type VecTy with
1895 /// vector length VL.  .
getAllOnesMask(MVT VecVT,SDValue VL,SDLoc DL,SelectionDAG & DAG)1896 static SDValue getAllOnesMask(MVT VecVT, SDValue VL, SDLoc DL,
1897                               SelectionDAG &DAG) {
1898   MVT MaskVT = getMaskTypeFor(VecVT);
1899   return DAG.getNode(RISCVISD::VMSET_VL, DL, MaskVT, VL);
1900 }
1901 
getVLOp(uint64_t NumElts,SDLoc DL,SelectionDAG & DAG,const RISCVSubtarget & Subtarget)1902 static SDValue getVLOp(uint64_t NumElts, SDLoc DL, SelectionDAG &DAG,
1903                        const RISCVSubtarget &Subtarget) {
1904   return DAG.getConstant(NumElts, DL, Subtarget.getXLenVT());
1905 }
1906 
1907 static std::pair<SDValue, SDValue>
getDefaultVLOps(uint64_t NumElts,MVT ContainerVT,SDLoc DL,SelectionDAG & DAG,const RISCVSubtarget & Subtarget)1908 getDefaultVLOps(uint64_t NumElts, MVT ContainerVT, SDLoc DL, SelectionDAG &DAG,
1909                 const RISCVSubtarget &Subtarget) {
1910   assert(ContainerVT.isScalableVector() && "Expecting scalable container type");
1911   SDValue VL = getVLOp(NumElts, DL, DAG, Subtarget);
1912   SDValue Mask = getAllOnesMask(ContainerVT, VL, DL, DAG);
1913   return {Mask, VL};
1914 }
1915 
1916 // Gets the two common "VL" operands: an all-ones mask and the vector length.
1917 // VecVT is a vector type, either fixed-length or scalable, and ContainerVT is
1918 // the vector type that the fixed-length vector is contained in. Otherwise if
1919 // VecVT is scalable, then ContainerVT should be the same as VecVT.
1920 static std::pair<SDValue, SDValue>
getDefaultVLOps(MVT VecVT,MVT ContainerVT,SDLoc DL,SelectionDAG & DAG,const RISCVSubtarget & Subtarget)1921 getDefaultVLOps(MVT VecVT, MVT ContainerVT, SDLoc DL, SelectionDAG &DAG,
1922                 const RISCVSubtarget &Subtarget) {
1923   if (VecVT.isFixedLengthVector())
1924     return getDefaultVLOps(VecVT.getVectorNumElements(), ContainerVT, DL, DAG,
1925                            Subtarget);
1926   assert(ContainerVT.isScalableVector() && "Expecting scalable container type");
1927   MVT XLenVT = Subtarget.getXLenVT();
1928   SDValue VL = DAG.getRegister(RISCV::X0, XLenVT);
1929   SDValue Mask = getAllOnesMask(ContainerVT, VL, DL, DAG);
1930   return {Mask, VL};
1931 }
1932 
1933 // As above but assuming the given type is a scalable vector type.
1934 static std::pair<SDValue, SDValue>
getDefaultScalableVLOps(MVT VecVT,SDLoc DL,SelectionDAG & DAG,const RISCVSubtarget & Subtarget)1935 getDefaultScalableVLOps(MVT VecVT, SDLoc DL, SelectionDAG &DAG,
1936                         const RISCVSubtarget &Subtarget) {
1937   assert(VecVT.isScalableVector() && "Expecting a scalable vector");
1938   return getDefaultVLOps(VecVT, VecVT, DL, DAG, Subtarget);
1939 }
1940 
1941 // The state of RVV BUILD_VECTOR and VECTOR_SHUFFLE lowering is that very few
1942 // of either is (currently) supported. This can get us into an infinite loop
1943 // where we try to lower a BUILD_VECTOR as a VECTOR_SHUFFLE as a BUILD_VECTOR
1944 // as a ..., etc.
1945 // Until either (or both) of these can reliably lower any node, reporting that
1946 // we don't want to expand BUILD_VECTORs via VECTOR_SHUFFLEs at least breaks
1947 // the infinite loop. Note that this lowers BUILD_VECTOR through the stack,
1948 // which is not desirable.
shouldExpandBuildVectorWithShuffles(EVT VT,unsigned DefinedValues) const1949 bool RISCVTargetLowering::shouldExpandBuildVectorWithShuffles(
1950     EVT VT, unsigned DefinedValues) const {
1951   return false;
1952 }
1953 
lowerFP_TO_INT_SAT(SDValue Op,SelectionDAG & DAG,const RISCVSubtarget & Subtarget)1954 static SDValue lowerFP_TO_INT_SAT(SDValue Op, SelectionDAG &DAG,
1955                                   const RISCVSubtarget &Subtarget) {
1956   // RISCV FP-to-int conversions saturate to the destination register size, but
1957   // don't produce 0 for nan. We can use a conversion instruction and fix the
1958   // nan case with a compare and a select.
1959   SDValue Src = Op.getOperand(0);
1960 
1961   MVT DstVT = Op.getSimpleValueType();
1962   EVT SatVT = cast<VTSDNode>(Op.getOperand(1))->getVT();
1963 
1964   bool IsSigned = Op.getOpcode() == ISD::FP_TO_SINT_SAT;
1965 
1966   if (!DstVT.isVector()) {
1967     // In absense of Zfh, promote f16 to f32, then saturate the result.
1968     if (Src.getSimpleValueType() == MVT::f16 && !Subtarget.hasStdExtZfh()) {
1969       Src = DAG.getNode(ISD::FP_EXTEND, SDLoc(Op), MVT::f32, Src);
1970     }
1971 
1972     unsigned Opc;
1973     if (SatVT == DstVT)
1974       Opc = IsSigned ? RISCVISD::FCVT_X : RISCVISD::FCVT_XU;
1975     else if (DstVT == MVT::i64 && SatVT == MVT::i32)
1976       Opc = IsSigned ? RISCVISD::FCVT_W_RV64 : RISCVISD::FCVT_WU_RV64;
1977     else
1978       return SDValue();
1979     // FIXME: Support other SatVTs by clamping before or after the conversion.
1980 
1981     SDLoc DL(Op);
1982     SDValue FpToInt = DAG.getNode(
1983         Opc, DL, DstVT, Src,
1984         DAG.getTargetConstant(RISCVFPRndMode::RTZ, DL, Subtarget.getXLenVT()));
1985 
1986     if (Opc == RISCVISD::FCVT_WU_RV64)
1987       FpToInt = DAG.getZeroExtendInReg(FpToInt, DL, MVT::i32);
1988 
1989     SDValue ZeroInt = DAG.getConstant(0, DL, DstVT);
1990     return DAG.getSelectCC(DL, Src, Src, ZeroInt, FpToInt,
1991                            ISD::CondCode::SETUO);
1992   }
1993 
1994   // Vectors.
1995 
1996   MVT DstEltVT = DstVT.getVectorElementType();
1997   MVT SrcVT = Src.getSimpleValueType();
1998   MVT SrcEltVT = SrcVT.getVectorElementType();
1999   unsigned SrcEltSize = SrcEltVT.getSizeInBits();
2000   unsigned DstEltSize = DstEltVT.getSizeInBits();
2001 
2002   // Only handle saturating to the destination type.
2003   if (SatVT != DstEltVT)
2004     return SDValue();
2005 
2006   // FIXME: Don't support narrowing by more than 1 steps for now.
2007   if (SrcEltSize > (2 * DstEltSize))
2008     return SDValue();
2009 
2010   MVT DstContainerVT = DstVT;
2011   MVT SrcContainerVT = SrcVT;
2012   if (DstVT.isFixedLengthVector()) {
2013     DstContainerVT = getContainerForFixedLengthVector(DAG, DstVT, Subtarget);
2014     SrcContainerVT = getContainerForFixedLengthVector(DAG, SrcVT, Subtarget);
2015     assert(DstContainerVT.getVectorElementCount() ==
2016                SrcContainerVT.getVectorElementCount() &&
2017            "Expected same element count");
2018     Src = convertToScalableVector(SrcContainerVT, Src, DAG, Subtarget);
2019   }
2020 
2021   SDLoc DL(Op);
2022 
2023   auto [Mask, VL] = getDefaultVLOps(DstVT, DstContainerVT, DL, DAG, Subtarget);
2024 
2025   SDValue IsNan = DAG.getNode(RISCVISD::SETCC_VL, DL, Mask.getValueType(),
2026                               {Src, Src, DAG.getCondCode(ISD::SETNE),
2027                                DAG.getUNDEF(Mask.getValueType()), Mask, VL});
2028 
2029   // Need to widen by more than 1 step, promote the FP type, then do a widening
2030   // convert.
2031   if (DstEltSize > (2 * SrcEltSize)) {
2032     assert(SrcContainerVT.getVectorElementType() == MVT::f16 && "Unexpected VT!");
2033     MVT InterVT = SrcContainerVT.changeVectorElementType(MVT::f32);
2034     Src = DAG.getNode(RISCVISD::FP_EXTEND_VL, DL, InterVT, Src, Mask, VL);
2035   }
2036 
2037   unsigned RVVOpc =
2038       IsSigned ? RISCVISD::VFCVT_RTZ_X_F_VL : RISCVISD::VFCVT_RTZ_XU_F_VL;
2039   SDValue Res = DAG.getNode(RVVOpc, DL, DstContainerVT, Src, Mask, VL);
2040 
2041   SDValue SplatZero = DAG.getNode(
2042       RISCVISD::VMV_V_X_VL, DL, DstContainerVT, DAG.getUNDEF(DstContainerVT),
2043       DAG.getConstant(0, DL, Subtarget.getXLenVT()), VL);
2044   Res = DAG.getNode(RISCVISD::VSELECT_VL, DL, DstContainerVT, IsNan, SplatZero,
2045                     Res, VL);
2046 
2047   if (DstVT.isFixedLengthVector())
2048     Res = convertFromScalableVector(DstVT, Res, DAG, Subtarget);
2049 
2050   return Res;
2051 }
2052 
matchRoundingOp(unsigned Opc)2053 static RISCVFPRndMode::RoundingMode matchRoundingOp(unsigned Opc) {
2054   switch (Opc) {
2055   case ISD::FROUNDEVEN:
2056   case ISD::VP_FROUNDEVEN:
2057     return RISCVFPRndMode::RNE;
2058   case ISD::FTRUNC:
2059   case ISD::VP_FROUNDTOZERO:
2060     return RISCVFPRndMode::RTZ;
2061   case ISD::FFLOOR:
2062   case ISD::VP_FFLOOR:
2063     return RISCVFPRndMode::RDN;
2064   case ISD::FCEIL:
2065   case ISD::VP_FCEIL:
2066     return RISCVFPRndMode::RUP;
2067   case ISD::FROUND:
2068   case ISD::VP_FROUND:
2069     return RISCVFPRndMode::RMM;
2070   case ISD::FRINT:
2071     return RISCVFPRndMode::DYN;
2072   }
2073 
2074   return RISCVFPRndMode::Invalid;
2075 }
2076 
2077 // Expand vector FTRUNC, FCEIL, FFLOOR, FROUND, VP_FCEIL, VP_FFLOOR, VP_FROUND
2078 // VP_FROUNDEVEN, VP_FROUNDTOZERO, VP_FRINT and VP_FNEARBYINT by converting to
2079 // the integer domain and back. Taking care to avoid converting values that are
2080 // nan or already correct.
2081 static SDValue
lowerVectorFTRUNC_FCEIL_FFLOOR_FROUND(SDValue Op,SelectionDAG & DAG,const RISCVSubtarget & Subtarget)2082 lowerVectorFTRUNC_FCEIL_FFLOOR_FROUND(SDValue Op, SelectionDAG &DAG,
2083                                       const RISCVSubtarget &Subtarget) {
2084   MVT VT = Op.getSimpleValueType();
2085   assert(VT.isVector() && "Unexpected type");
2086 
2087   SDLoc DL(Op);
2088 
2089   SDValue Src = Op.getOperand(0);
2090 
2091   MVT ContainerVT = VT;
2092   if (VT.isFixedLengthVector()) {
2093     ContainerVT = getContainerForFixedLengthVector(DAG, VT, Subtarget);
2094     Src = convertToScalableVector(ContainerVT, Src, DAG, Subtarget);
2095   }
2096 
2097   SDValue Mask, VL;
2098   if (Op->isVPOpcode()) {
2099     Mask = Op.getOperand(1);
2100     VL = Op.getOperand(2);
2101   } else {
2102     std::tie(Mask, VL) = getDefaultVLOps(VT, ContainerVT, DL, DAG, Subtarget);
2103   }
2104 
2105   // Freeze the source since we are increasing the number of uses.
2106   Src = DAG.getFreeze(Src);
2107 
2108   // We do the conversion on the absolute value and fix the sign at the end.
2109   SDValue Abs = DAG.getNode(RISCVISD::FABS_VL, DL, ContainerVT, Src, Mask, VL);
2110 
2111   // Determine the largest integer that can be represented exactly. This and
2112   // values larger than it don't have any fractional bits so don't need to
2113   // be converted.
2114   const fltSemantics &FltSem = DAG.EVTToAPFloatSemantics(ContainerVT);
2115   unsigned Precision = APFloat::semanticsPrecision(FltSem);
2116   APFloat MaxVal = APFloat(FltSem);
2117   MaxVal.convertFromAPInt(APInt::getOneBitSet(Precision, Precision - 1),
2118                           /*IsSigned*/ false, APFloat::rmNearestTiesToEven);
2119   SDValue MaxValNode =
2120       DAG.getConstantFP(MaxVal, DL, ContainerVT.getVectorElementType());
2121   SDValue MaxValSplat = DAG.getNode(RISCVISD::VFMV_V_F_VL, DL, ContainerVT,
2122                                     DAG.getUNDEF(ContainerVT), MaxValNode, VL);
2123 
2124   // If abs(Src) was larger than MaxVal or nan, keep it.
2125   MVT SetccVT = MVT::getVectorVT(MVT::i1, ContainerVT.getVectorElementCount());
2126   Mask =
2127       DAG.getNode(RISCVISD::SETCC_VL, DL, SetccVT,
2128                   {Abs, MaxValSplat, DAG.getCondCode(ISD::SETOLT),
2129                    Mask, Mask, VL});
2130 
2131   // Truncate to integer and convert back to FP.
2132   MVT IntVT = ContainerVT.changeVectorElementTypeToInteger();
2133   MVT XLenVT = Subtarget.getXLenVT();
2134   SDValue Truncated;
2135 
2136   switch (Op.getOpcode()) {
2137   default:
2138     llvm_unreachable("Unexpected opcode");
2139   case ISD::FCEIL:
2140   case ISD::VP_FCEIL:
2141   case ISD::FFLOOR:
2142   case ISD::VP_FFLOOR:
2143   case ISD::FROUND:
2144   case ISD::FROUNDEVEN:
2145   case ISD::VP_FROUND:
2146   case ISD::VP_FROUNDEVEN:
2147   case ISD::VP_FROUNDTOZERO: {
2148     RISCVFPRndMode::RoundingMode FRM = matchRoundingOp(Op.getOpcode());
2149     assert(FRM != RISCVFPRndMode::Invalid);
2150     Truncated = DAG.getNode(RISCVISD::VFCVT_RM_X_F_VL, DL, IntVT, Src, Mask,
2151                             DAG.getTargetConstant(FRM, DL, XLenVT), VL);
2152     break;
2153   }
2154   case ISD::FTRUNC:
2155     Truncated = DAG.getNode(RISCVISD::VFCVT_RTZ_X_F_VL, DL, IntVT, Src,
2156                             Mask, VL);
2157     break;
2158   case ISD::VP_FRINT:
2159     Truncated = DAG.getNode(RISCVISD::VFCVT_X_F_VL, DL, IntVT, Src, Mask, VL);
2160     break;
2161   case ISD::VP_FNEARBYINT:
2162     Truncated = DAG.getNode(RISCVISD::VFROUND_NOEXCEPT_VL, DL, ContainerVT, Src,
2163                             Mask, VL);
2164     break;
2165   }
2166 
2167   // VFROUND_NOEXCEPT_VL includes SINT_TO_FP_VL.
2168   if (Op.getOpcode() != ISD::VP_FNEARBYINT)
2169     Truncated = DAG.getNode(RISCVISD::SINT_TO_FP_VL, DL, ContainerVT, Truncated,
2170                             Mask, VL);
2171 
2172   // Restore the original sign so that -0.0 is preserved.
2173   Truncated = DAG.getNode(RISCVISD::FCOPYSIGN_VL, DL, ContainerVT, Truncated,
2174                           Src, Src, Mask, VL);
2175 
2176   if (!VT.isFixedLengthVector())
2177     return Truncated;
2178 
2179   return convertFromScalableVector(VT, Truncated, DAG, Subtarget);
2180 }
2181 
2182 static SDValue
lowerFTRUNC_FCEIL_FFLOOR_FROUND(SDValue Op,SelectionDAG & DAG,const RISCVSubtarget & Subtarget)2183 lowerFTRUNC_FCEIL_FFLOOR_FROUND(SDValue Op, SelectionDAG &DAG,
2184                                 const RISCVSubtarget &Subtarget) {
2185   MVT VT = Op.getSimpleValueType();
2186   if (VT.isVector())
2187     return lowerVectorFTRUNC_FCEIL_FFLOOR_FROUND(Op, DAG, Subtarget);
2188 
2189   if (DAG.shouldOptForSize())
2190     return SDValue();
2191 
2192   SDLoc DL(Op);
2193   SDValue Src = Op.getOperand(0);
2194 
2195   // Create an integer the size of the mantissa with the MSB set. This and all
2196   // values larger than it don't have any fractional bits so don't need to be
2197   // converted.
2198   const fltSemantics &FltSem = DAG.EVTToAPFloatSemantics(VT);
2199   unsigned Precision = APFloat::semanticsPrecision(FltSem);
2200   APFloat MaxVal = APFloat(FltSem);
2201   MaxVal.convertFromAPInt(APInt::getOneBitSet(Precision, Precision - 1),
2202                           /*IsSigned*/ false, APFloat::rmNearestTiesToEven);
2203   SDValue MaxValNode = DAG.getConstantFP(MaxVal, DL, VT);
2204 
2205   RISCVFPRndMode::RoundingMode FRM = matchRoundingOp(Op.getOpcode());
2206   return DAG.getNode(RISCVISD::FROUND, DL, VT, Src, MaxValNode,
2207                      DAG.getTargetConstant(FRM, DL, Subtarget.getXLenVT()));
2208 }
2209 
2210 struct VIDSequence {
2211   int64_t StepNumerator;
2212   unsigned StepDenominator;
2213   int64_t Addend;
2214 };
2215 
getExactInteger(const APFloat & APF,uint32_t BitWidth)2216 static std::optional<uint64_t> getExactInteger(const APFloat &APF,
2217                                                uint32_t BitWidth) {
2218   APSInt ValInt(BitWidth, !APF.isNegative());
2219   // We use an arbitrary rounding mode here. If a floating-point is an exact
2220   // integer (e.g., 1.0), the rounding mode does not affect the output value. If
2221   // the rounding mode changes the output value, then it is not an exact
2222   // integer.
2223   RoundingMode ArbitraryRM = RoundingMode::TowardZero;
2224   bool IsExact;
2225   // If it is out of signed integer range, it will return an invalid operation.
2226   // If it is not an exact integer, IsExact is false.
2227   if ((APF.convertToInteger(ValInt, ArbitraryRM, &IsExact) ==
2228        APFloatBase::opInvalidOp) ||
2229       !IsExact)
2230     return std::nullopt;
2231   return ValInt.extractBitsAsZExtValue(BitWidth, 0);
2232 }
2233 
2234 // Try to match an arithmetic-sequence BUILD_VECTOR [X,X+S,X+2*S,...,X+(N-1)*S]
2235 // to the (non-zero) step S and start value X. This can be then lowered as the
2236 // RVV sequence (VID * S) + X, for example.
2237 // The step S is represented as an integer numerator divided by a positive
2238 // denominator. Note that the implementation currently only identifies
2239 // sequences in which either the numerator is +/- 1 or the denominator is 1. It
2240 // cannot detect 2/3, for example.
2241 // Note that this method will also match potentially unappealing index
2242 // sequences, like <i32 0, i32 50939494>, however it is left to the caller to
2243 // determine whether this is worth generating code for.
isSimpleVIDSequence(SDValue Op)2244 static std::optional<VIDSequence> isSimpleVIDSequence(SDValue Op) {
2245   unsigned NumElts = Op.getNumOperands();
2246   assert(Op.getOpcode() == ISD::BUILD_VECTOR && "Unexpected BUILD_VECTOR");
2247   bool IsInteger = Op.getValueType().isInteger();
2248 
2249   std::optional<unsigned> SeqStepDenom;
2250   std::optional<int64_t> SeqStepNum, SeqAddend;
2251   std::optional<std::pair<uint64_t, unsigned>> PrevElt;
2252   unsigned EltSizeInBits = Op.getValueType().getScalarSizeInBits();
2253   for (unsigned Idx = 0; Idx < NumElts; Idx++) {
2254     // Assume undef elements match the sequence; we just have to be careful
2255     // when interpolating across them.
2256     if (Op.getOperand(Idx).isUndef())
2257       continue;
2258 
2259     uint64_t Val;
2260     if (IsInteger) {
2261       // The BUILD_VECTOR must be all constants.
2262       if (!isa<ConstantSDNode>(Op.getOperand(Idx)))
2263         return std::nullopt;
2264       Val = Op.getConstantOperandVal(Idx) &
2265             maskTrailingOnes<uint64_t>(EltSizeInBits);
2266     } else {
2267       // The BUILD_VECTOR must be all constants.
2268       if (!isa<ConstantFPSDNode>(Op.getOperand(Idx)))
2269         return std::nullopt;
2270       if (auto ExactInteger = getExactInteger(
2271               cast<ConstantFPSDNode>(Op.getOperand(Idx))->getValueAPF(),
2272               EltSizeInBits))
2273         Val = *ExactInteger;
2274       else
2275         return std::nullopt;
2276     }
2277 
2278     if (PrevElt) {
2279       // Calculate the step since the last non-undef element, and ensure
2280       // it's consistent across the entire sequence.
2281       unsigned IdxDiff = Idx - PrevElt->second;
2282       int64_t ValDiff = SignExtend64(Val - PrevElt->first, EltSizeInBits);
2283 
2284       // A zero-value value difference means that we're somewhere in the middle
2285       // of a fractional step, e.g. <0,0,0*,0,1,1,1,1>. Wait until we notice a
2286       // step change before evaluating the sequence.
2287       if (ValDiff == 0)
2288         continue;
2289 
2290       int64_t Remainder = ValDiff % IdxDiff;
2291       // Normalize the step if it's greater than 1.
2292       if (Remainder != ValDiff) {
2293         // The difference must cleanly divide the element span.
2294         if (Remainder != 0)
2295           return std::nullopt;
2296         ValDiff /= IdxDiff;
2297         IdxDiff = 1;
2298       }
2299 
2300       if (!SeqStepNum)
2301         SeqStepNum = ValDiff;
2302       else if (ValDiff != SeqStepNum)
2303         return std::nullopt;
2304 
2305       if (!SeqStepDenom)
2306         SeqStepDenom = IdxDiff;
2307       else if (IdxDiff != *SeqStepDenom)
2308         return std::nullopt;
2309     }
2310 
2311     // Record this non-undef element for later.
2312     if (!PrevElt || PrevElt->first != Val)
2313       PrevElt = std::make_pair(Val, Idx);
2314   }
2315 
2316   // We need to have logged a step for this to count as a legal index sequence.
2317   if (!SeqStepNum || !SeqStepDenom)
2318     return std::nullopt;
2319 
2320   // Loop back through the sequence and validate elements we might have skipped
2321   // while waiting for a valid step. While doing this, log any sequence addend.
2322   for (unsigned Idx = 0; Idx < NumElts; Idx++) {
2323     if (Op.getOperand(Idx).isUndef())
2324       continue;
2325     uint64_t Val;
2326     if (IsInteger) {
2327       Val = Op.getConstantOperandVal(Idx) &
2328             maskTrailingOnes<uint64_t>(EltSizeInBits);
2329     } else {
2330       Val = *getExactInteger(
2331           cast<ConstantFPSDNode>(Op.getOperand(Idx))->getValueAPF(),
2332           EltSizeInBits);
2333     }
2334     uint64_t ExpectedVal =
2335         (int64_t)(Idx * (uint64_t)*SeqStepNum) / *SeqStepDenom;
2336     int64_t Addend = SignExtend64(Val - ExpectedVal, EltSizeInBits);
2337     if (!SeqAddend)
2338       SeqAddend = Addend;
2339     else if (Addend != SeqAddend)
2340       return std::nullopt;
2341   }
2342 
2343   assert(SeqAddend && "Must have an addend if we have a step");
2344 
2345   return VIDSequence{*SeqStepNum, *SeqStepDenom, *SeqAddend};
2346 }
2347 
2348 // Match a splatted value (SPLAT_VECTOR/BUILD_VECTOR) of an EXTRACT_VECTOR_ELT
2349 // and lower it as a VRGATHER_VX_VL from the source vector.
matchSplatAsGather(SDValue SplatVal,MVT VT,const SDLoc & DL,SelectionDAG & DAG,const RISCVSubtarget & Subtarget)2350 static SDValue matchSplatAsGather(SDValue SplatVal, MVT VT, const SDLoc &DL,
2351                                   SelectionDAG &DAG,
2352                                   const RISCVSubtarget &Subtarget) {
2353   if (SplatVal.getOpcode() != ISD::EXTRACT_VECTOR_ELT)
2354     return SDValue();
2355   SDValue Vec = SplatVal.getOperand(0);
2356   // Only perform this optimization on vectors of the same size for simplicity.
2357   // Don't perform this optimization for i1 vectors.
2358   // FIXME: Support i1 vectors, maybe by promoting to i8?
2359   if (Vec.getValueType() != VT || VT.getVectorElementType() == MVT::i1)
2360     return SDValue();
2361   SDValue Idx = SplatVal.getOperand(1);
2362   // The index must be a legal type.
2363   if (Idx.getValueType() != Subtarget.getXLenVT())
2364     return SDValue();
2365 
2366   MVT ContainerVT = VT;
2367   if (VT.isFixedLengthVector()) {
2368     ContainerVT = getContainerForFixedLengthVector(DAG, VT, Subtarget);
2369     Vec = convertToScalableVector(ContainerVT, Vec, DAG, Subtarget);
2370   }
2371 
2372   auto [Mask, VL] = getDefaultVLOps(VT, ContainerVT, DL, DAG, Subtarget);
2373 
2374   SDValue Gather = DAG.getNode(RISCVISD::VRGATHER_VX_VL, DL, ContainerVT, Vec,
2375                                Idx, DAG.getUNDEF(ContainerVT), Mask, VL);
2376 
2377   if (!VT.isFixedLengthVector())
2378     return Gather;
2379 
2380   return convertFromScalableVector(VT, Gather, DAG, Subtarget);
2381 }
2382 
lowerBUILD_VECTOR(SDValue Op,SelectionDAG & DAG,const RISCVSubtarget & Subtarget)2383 static SDValue lowerBUILD_VECTOR(SDValue Op, SelectionDAG &DAG,
2384                                  const RISCVSubtarget &Subtarget) {
2385   MVT VT = Op.getSimpleValueType();
2386   assert(VT.isFixedLengthVector() && "Unexpected vector!");
2387 
2388   MVT ContainerVT = getContainerForFixedLengthVector(DAG, VT, Subtarget);
2389 
2390   SDLoc DL(Op);
2391   auto [Mask, VL] = getDefaultVLOps(VT, ContainerVT, DL, DAG, Subtarget);
2392 
2393   MVT XLenVT = Subtarget.getXLenVT();
2394   unsigned NumElts = Op.getNumOperands();
2395 
2396   if (VT.getVectorElementType() == MVT::i1) {
2397     if (ISD::isBuildVectorAllZeros(Op.getNode())) {
2398       SDValue VMClr = DAG.getNode(RISCVISD::VMCLR_VL, DL, ContainerVT, VL);
2399       return convertFromScalableVector(VT, VMClr, DAG, Subtarget);
2400     }
2401 
2402     if (ISD::isBuildVectorAllOnes(Op.getNode())) {
2403       SDValue VMSet = DAG.getNode(RISCVISD::VMSET_VL, DL, ContainerVT, VL);
2404       return convertFromScalableVector(VT, VMSet, DAG, Subtarget);
2405     }
2406 
2407     // Lower constant mask BUILD_VECTORs via an integer vector type, in
2408     // scalar integer chunks whose bit-width depends on the number of mask
2409     // bits and XLEN.
2410     // First, determine the most appropriate scalar integer type to use. This
2411     // is at most XLenVT, but may be shrunk to a smaller vector element type
2412     // according to the size of the final vector - use i8 chunks rather than
2413     // XLenVT if we're producing a v8i1. This results in more consistent
2414     // codegen across RV32 and RV64.
2415     unsigned NumViaIntegerBits =
2416         std::min(std::max(NumElts, 8u), Subtarget.getXLen());
2417     NumViaIntegerBits = std::min(NumViaIntegerBits, Subtarget.getELEN());
2418     if (ISD::isBuildVectorOfConstantSDNodes(Op.getNode())) {
2419       // If we have to use more than one INSERT_VECTOR_ELT then this
2420       // optimization is likely to increase code size; avoid peforming it in
2421       // such a case. We can use a load from a constant pool in this case.
2422       if (DAG.shouldOptForSize() && NumElts > NumViaIntegerBits)
2423         return SDValue();
2424       // Now we can create our integer vector type. Note that it may be larger
2425       // than the resulting mask type: v4i1 would use v1i8 as its integer type.
2426       MVT IntegerViaVecVT =
2427           MVT::getVectorVT(MVT::getIntegerVT(NumViaIntegerBits),
2428                            divideCeil(NumElts, NumViaIntegerBits));
2429 
2430       uint64_t Bits = 0;
2431       unsigned BitPos = 0, IntegerEltIdx = 0;
2432       SDValue Vec = DAG.getUNDEF(IntegerViaVecVT);
2433 
2434       for (unsigned I = 0; I < NumElts; I++, BitPos++) {
2435         // Once we accumulate enough bits to fill our scalar type, insert into
2436         // our vector and clear our accumulated data.
2437         if (I != 0 && I % NumViaIntegerBits == 0) {
2438           if (NumViaIntegerBits <= 32)
2439             Bits = SignExtend64<32>(Bits);
2440           SDValue Elt = DAG.getConstant(Bits, DL, XLenVT);
2441           Vec = DAG.getNode(ISD::INSERT_VECTOR_ELT, DL, IntegerViaVecVT, Vec,
2442                             Elt, DAG.getConstant(IntegerEltIdx, DL, XLenVT));
2443           Bits = 0;
2444           BitPos = 0;
2445           IntegerEltIdx++;
2446         }
2447         SDValue V = Op.getOperand(I);
2448         bool BitValue = !V.isUndef() && cast<ConstantSDNode>(V)->getZExtValue();
2449         Bits |= ((uint64_t)BitValue << BitPos);
2450       }
2451 
2452       // Insert the (remaining) scalar value into position in our integer
2453       // vector type.
2454       if (NumViaIntegerBits <= 32)
2455         Bits = SignExtend64<32>(Bits);
2456       SDValue Elt = DAG.getConstant(Bits, DL, XLenVT);
2457       Vec = DAG.getNode(ISD::INSERT_VECTOR_ELT, DL, IntegerViaVecVT, Vec, Elt,
2458                         DAG.getConstant(IntegerEltIdx, DL, XLenVT));
2459 
2460       if (NumElts < NumViaIntegerBits) {
2461         // If we're producing a smaller vector than our minimum legal integer
2462         // type, bitcast to the equivalent (known-legal) mask type, and extract
2463         // our final mask.
2464         assert(IntegerViaVecVT == MVT::v1i8 && "Unexpected mask vector type");
2465         Vec = DAG.getBitcast(MVT::v8i1, Vec);
2466         Vec = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, VT, Vec,
2467                           DAG.getConstant(0, DL, XLenVT));
2468       } else {
2469         // Else we must have produced an integer type with the same size as the
2470         // mask type; bitcast for the final result.
2471         assert(VT.getSizeInBits() == IntegerViaVecVT.getSizeInBits());
2472         Vec = DAG.getBitcast(VT, Vec);
2473       }
2474 
2475       return Vec;
2476     }
2477 
2478     // A BUILD_VECTOR can be lowered as a SETCC. For each fixed-length mask
2479     // vector type, we have a legal equivalently-sized i8 type, so we can use
2480     // that.
2481     MVT WideVecVT = VT.changeVectorElementType(MVT::i8);
2482     SDValue VecZero = DAG.getConstant(0, DL, WideVecVT);
2483 
2484     SDValue WideVec;
2485     if (SDValue Splat = cast<BuildVectorSDNode>(Op)->getSplatValue()) {
2486       // For a splat, perform a scalar truncate before creating the wider
2487       // vector.
2488       assert(Splat.getValueType() == XLenVT &&
2489              "Unexpected type for i1 splat value");
2490       Splat = DAG.getNode(ISD::AND, DL, XLenVT, Splat,
2491                           DAG.getConstant(1, DL, XLenVT));
2492       WideVec = DAG.getSplatBuildVector(WideVecVT, DL, Splat);
2493     } else {
2494       SmallVector<SDValue, 8> Ops(Op->op_values());
2495       WideVec = DAG.getBuildVector(WideVecVT, DL, Ops);
2496       SDValue VecOne = DAG.getConstant(1, DL, WideVecVT);
2497       WideVec = DAG.getNode(ISD::AND, DL, WideVecVT, WideVec, VecOne);
2498     }
2499 
2500     return DAG.getSetCC(DL, VT, WideVec, VecZero, ISD::SETNE);
2501   }
2502 
2503   if (SDValue Splat = cast<BuildVectorSDNode>(Op)->getSplatValue()) {
2504     if (auto Gather = matchSplatAsGather(Splat, VT, DL, DAG, Subtarget))
2505       return Gather;
2506     unsigned Opc = VT.isFloatingPoint() ? RISCVISD::VFMV_V_F_VL
2507                                         : RISCVISD::VMV_V_X_VL;
2508     Splat =
2509         DAG.getNode(Opc, DL, ContainerVT, DAG.getUNDEF(ContainerVT), Splat, VL);
2510     return convertFromScalableVector(VT, Splat, DAG, Subtarget);
2511   }
2512 
2513   // Try and match index sequences, which we can lower to the vid instruction
2514   // with optional modifications. An all-undef vector is matched by
2515   // getSplatValue, above.
2516   if (auto SimpleVID = isSimpleVIDSequence(Op)) {
2517     int64_t StepNumerator = SimpleVID->StepNumerator;
2518     unsigned StepDenominator = SimpleVID->StepDenominator;
2519     int64_t Addend = SimpleVID->Addend;
2520 
2521     assert(StepNumerator != 0 && "Invalid step");
2522     bool Negate = false;
2523     int64_t SplatStepVal = StepNumerator;
2524     unsigned StepOpcode = ISD::MUL;
2525     if (StepNumerator != 1) {
2526       if (isPowerOf2_64(std::abs(StepNumerator))) {
2527         Negate = StepNumerator < 0;
2528         StepOpcode = ISD::SHL;
2529         SplatStepVal = Log2_64(std::abs(StepNumerator));
2530       }
2531     }
2532 
2533     // Only emit VIDs with suitably-small steps/addends. We use imm5 is a
2534     // threshold since it's the immediate value many RVV instructions accept.
2535     // There is no vmul.vi instruction so ensure multiply constant can fit in
2536     // a single addi instruction.
2537     if (((StepOpcode == ISD::MUL && isInt<12>(SplatStepVal)) ||
2538          (StepOpcode == ISD::SHL && isUInt<5>(SplatStepVal))) &&
2539         isPowerOf2_32(StepDenominator) &&
2540         (SplatStepVal >= 0 || StepDenominator == 1) && isInt<5>(Addend)) {
2541       MVT VIDVT =
2542           VT.isFloatingPoint() ? VT.changeVectorElementTypeToInteger() : VT;
2543       MVT VIDContainerVT =
2544           getContainerForFixedLengthVector(DAG, VIDVT, Subtarget);
2545       SDValue VID = DAG.getNode(RISCVISD::VID_VL, DL, VIDContainerVT, Mask, VL);
2546       // Convert right out of the scalable type so we can use standard ISD
2547       // nodes for the rest of the computation. If we used scalable types with
2548       // these, we'd lose the fixed-length vector info and generate worse
2549       // vsetvli code.
2550       VID = convertFromScalableVector(VIDVT, VID, DAG, Subtarget);
2551       if ((StepOpcode == ISD::MUL && SplatStepVal != 1) ||
2552           (StepOpcode == ISD::SHL && SplatStepVal != 0)) {
2553         SDValue SplatStep = DAG.getSplatBuildVector(
2554             VIDVT, DL, DAG.getConstant(SplatStepVal, DL, XLenVT));
2555         VID = DAG.getNode(StepOpcode, DL, VIDVT, VID, SplatStep);
2556       }
2557       if (StepDenominator != 1) {
2558         SDValue SplatStep = DAG.getSplatBuildVector(
2559             VIDVT, DL, DAG.getConstant(Log2_64(StepDenominator), DL, XLenVT));
2560         VID = DAG.getNode(ISD::SRL, DL, VIDVT, VID, SplatStep);
2561       }
2562       if (Addend != 0 || Negate) {
2563         SDValue SplatAddend = DAG.getSplatBuildVector(
2564             VIDVT, DL, DAG.getConstant(Addend, DL, XLenVT));
2565         VID = DAG.getNode(Negate ? ISD::SUB : ISD::ADD, DL, VIDVT, SplatAddend,
2566                           VID);
2567       }
2568       if (VT.isFloatingPoint()) {
2569         // TODO: Use vfwcvt to reduce register pressure.
2570         VID = DAG.getNode(ISD::SINT_TO_FP, DL, VT, VID);
2571       }
2572       return VID;
2573     }
2574   }
2575 
2576   // Attempt to detect "hidden" splats, which only reveal themselves as splats
2577   // when re-interpreted as a vector with a larger element type. For example,
2578   //   v4i16 = build_vector i16 0, i16 1, i16 0, i16 1
2579   // could be instead splat as
2580   //   v2i32 = build_vector i32 0x00010000, i32 0x00010000
2581   // TODO: This optimization could also work on non-constant splats, but it
2582   // would require bit-manipulation instructions to construct the splat value.
2583   SmallVector<SDValue> Sequence;
2584   unsigned EltBitSize = VT.getScalarSizeInBits();
2585   const auto *BV = cast<BuildVectorSDNode>(Op);
2586   if (VT.isInteger() && EltBitSize < 64 &&
2587       ISD::isBuildVectorOfConstantSDNodes(Op.getNode()) &&
2588       BV->getRepeatedSequence(Sequence) &&
2589       (Sequence.size() * EltBitSize) <= 64) {
2590     unsigned SeqLen = Sequence.size();
2591     MVT ViaIntVT = MVT::getIntegerVT(EltBitSize * SeqLen);
2592     MVT ViaVecVT = MVT::getVectorVT(ViaIntVT, NumElts / SeqLen);
2593     assert((ViaIntVT == MVT::i16 || ViaIntVT == MVT::i32 ||
2594             ViaIntVT == MVT::i64) &&
2595            "Unexpected sequence type");
2596 
2597     unsigned EltIdx = 0;
2598     uint64_t EltMask = maskTrailingOnes<uint64_t>(EltBitSize);
2599     uint64_t SplatValue = 0;
2600     // Construct the amalgamated value which can be splatted as this larger
2601     // vector type.
2602     for (const auto &SeqV : Sequence) {
2603       if (!SeqV.isUndef())
2604         SplatValue |= ((cast<ConstantSDNode>(SeqV)->getZExtValue() & EltMask)
2605                        << (EltIdx * EltBitSize));
2606       EltIdx++;
2607     }
2608 
2609     // On RV64, sign-extend from 32 to 64 bits where possible in order to
2610     // achieve better constant materializion.
2611     if (Subtarget.is64Bit() && ViaIntVT == MVT::i32)
2612       SplatValue = SignExtend64<32>(SplatValue);
2613 
2614     // Since we can't introduce illegal i64 types at this stage, we can only
2615     // perform an i64 splat on RV32 if it is its own sign-extended value. That
2616     // way we can use RVV instructions to splat.
2617     assert((ViaIntVT.bitsLE(XLenVT) ||
2618             (!Subtarget.is64Bit() && ViaIntVT == MVT::i64)) &&
2619            "Unexpected bitcast sequence");
2620     if (ViaIntVT.bitsLE(XLenVT) || isInt<32>(SplatValue)) {
2621       SDValue ViaVL =
2622           DAG.getConstant(ViaVecVT.getVectorNumElements(), DL, XLenVT);
2623       MVT ViaContainerVT =
2624           getContainerForFixedLengthVector(DAG, ViaVecVT, Subtarget);
2625       SDValue Splat =
2626           DAG.getNode(RISCVISD::VMV_V_X_VL, DL, ViaContainerVT,
2627                       DAG.getUNDEF(ViaContainerVT),
2628                       DAG.getConstant(SplatValue, DL, XLenVT), ViaVL);
2629       Splat = convertFromScalableVector(ViaVecVT, Splat, DAG, Subtarget);
2630       return DAG.getBitcast(VT, Splat);
2631     }
2632   }
2633 
2634   // Try and optimize BUILD_VECTORs with "dominant values" - these are values
2635   // which constitute a large proportion of the elements. In such cases we can
2636   // splat a vector with the dominant element and make up the shortfall with
2637   // INSERT_VECTOR_ELTs.
2638   // Note that this includes vectors of 2 elements by association. The
2639   // upper-most element is the "dominant" one, allowing us to use a splat to
2640   // "insert" the upper element, and an insert of the lower element at position
2641   // 0, which improves codegen.
2642   SDValue DominantValue;
2643   unsigned MostCommonCount = 0;
2644   DenseMap<SDValue, unsigned> ValueCounts;
2645   unsigned NumUndefElts =
2646       count_if(Op->op_values(), [](const SDValue &V) { return V.isUndef(); });
2647 
2648   // Track the number of scalar loads we know we'd be inserting, estimated as
2649   // any non-zero floating-point constant. Other kinds of element are either
2650   // already in registers or are materialized on demand. The threshold at which
2651   // a vector load is more desirable than several scalar materializion and
2652   // vector-insertion instructions is not known.
2653   unsigned NumScalarLoads = 0;
2654 
2655   for (SDValue V : Op->op_values()) {
2656     if (V.isUndef())
2657       continue;
2658 
2659     ValueCounts.insert(std::make_pair(V, 0));
2660     unsigned &Count = ValueCounts[V];
2661 
2662     if (auto *CFP = dyn_cast<ConstantFPSDNode>(V))
2663       NumScalarLoads += !CFP->isExactlyValue(+0.0);
2664 
2665     // Is this value dominant? In case of a tie, prefer the highest element as
2666     // it's cheaper to insert near the beginning of a vector than it is at the
2667     // end.
2668     if (++Count >= MostCommonCount) {
2669       DominantValue = V;
2670       MostCommonCount = Count;
2671     }
2672   }
2673 
2674   assert(DominantValue && "Not expecting an all-undef BUILD_VECTOR");
2675   unsigned NumDefElts = NumElts - NumUndefElts;
2676   unsigned DominantValueCountThreshold = NumDefElts <= 2 ? 0 : NumDefElts - 2;
2677 
2678   // Don't perform this optimization when optimizing for size, since
2679   // materializing elements and inserting them tends to cause code bloat.
2680   if (!DAG.shouldOptForSize() && NumScalarLoads < NumElts &&
2681       ((MostCommonCount > DominantValueCountThreshold) ||
2682        (ValueCounts.size() <= Log2_32(NumDefElts)))) {
2683     // Start by splatting the most common element.
2684     SDValue Vec = DAG.getSplatBuildVector(VT, DL, DominantValue);
2685 
2686     DenseSet<SDValue> Processed{DominantValue};
2687     MVT SelMaskTy = VT.changeVectorElementType(MVT::i1);
2688     for (const auto &OpIdx : enumerate(Op->ops())) {
2689       const SDValue &V = OpIdx.value();
2690       if (V.isUndef() || !Processed.insert(V).second)
2691         continue;
2692       if (ValueCounts[V] == 1) {
2693         Vec = DAG.getNode(ISD::INSERT_VECTOR_ELT, DL, VT, Vec, V,
2694                           DAG.getConstant(OpIdx.index(), DL, XLenVT));
2695       } else {
2696         // Blend in all instances of this value using a VSELECT, using a
2697         // mask where each bit signals whether that element is the one
2698         // we're after.
2699         SmallVector<SDValue> Ops;
2700         transform(Op->op_values(), std::back_inserter(Ops), [&](SDValue V1) {
2701           return DAG.getConstant(V == V1, DL, XLenVT);
2702         });
2703         Vec = DAG.getNode(ISD::VSELECT, DL, VT,
2704                           DAG.getBuildVector(SelMaskTy, DL, Ops),
2705                           DAG.getSplatBuildVector(VT, DL, V), Vec);
2706       }
2707     }
2708 
2709     return Vec;
2710   }
2711 
2712   return SDValue();
2713 }
2714 
splatPartsI64WithVL(const SDLoc & DL,MVT VT,SDValue Passthru,SDValue Lo,SDValue Hi,SDValue VL,SelectionDAG & DAG)2715 static SDValue splatPartsI64WithVL(const SDLoc &DL, MVT VT, SDValue Passthru,
2716                                    SDValue Lo, SDValue Hi, SDValue VL,
2717                                    SelectionDAG &DAG) {
2718   if (!Passthru)
2719     Passthru = DAG.getUNDEF(VT);
2720   if (isa<ConstantSDNode>(Lo) && isa<ConstantSDNode>(Hi)) {
2721     int32_t LoC = cast<ConstantSDNode>(Lo)->getSExtValue();
2722     int32_t HiC = cast<ConstantSDNode>(Hi)->getSExtValue();
2723     // If Hi constant is all the same sign bit as Lo, lower this as a custom
2724     // node in order to try and match RVV vector/scalar instructions.
2725     if ((LoC >> 31) == HiC)
2726       return DAG.getNode(RISCVISD::VMV_V_X_VL, DL, VT, Passthru, Lo, VL);
2727 
2728     // If vl is equal to XLEN_MAX and Hi constant is equal to Lo, we could use
2729     // vmv.v.x whose EEW = 32 to lower it.
2730     auto *Const = dyn_cast<ConstantSDNode>(VL);
2731     if (LoC == HiC && Const && Const->isAllOnesValue()) {
2732       MVT InterVT = MVT::getVectorVT(MVT::i32, VT.getVectorElementCount() * 2);
2733       // TODO: if vl <= min(VLMAX), we can also do this. But we could not
2734       // access the subtarget here now.
2735       auto InterVec = DAG.getNode(
2736           RISCVISD::VMV_V_X_VL, DL, InterVT, DAG.getUNDEF(InterVT), Lo,
2737                                   DAG.getRegister(RISCV::X0, MVT::i32));
2738       return DAG.getNode(ISD::BITCAST, DL, VT, InterVec);
2739     }
2740   }
2741 
2742   // Fall back to a stack store and stride x0 vector load.
2743   return DAG.getNode(RISCVISD::SPLAT_VECTOR_SPLIT_I64_VL, DL, VT, Passthru, Lo,
2744                      Hi, VL);
2745 }
2746 
2747 // Called by type legalization to handle splat of i64 on RV32.
2748 // FIXME: We can optimize this when the type has sign or zero bits in one
2749 // of the halves.
splatSplitI64WithVL(const SDLoc & DL,MVT VT,SDValue Passthru,SDValue Scalar,SDValue VL,SelectionDAG & DAG)2750 static SDValue splatSplitI64WithVL(const SDLoc &DL, MVT VT, SDValue Passthru,
2751                                    SDValue Scalar, SDValue VL,
2752                                    SelectionDAG &DAG) {
2753   assert(Scalar.getValueType() == MVT::i64 && "Unexpected VT!");
2754   SDValue Lo = DAG.getNode(ISD::EXTRACT_ELEMENT, DL, MVT::i32, Scalar,
2755                            DAG.getConstant(0, DL, MVT::i32));
2756   SDValue Hi = DAG.getNode(ISD::EXTRACT_ELEMENT, DL, MVT::i32, Scalar,
2757                            DAG.getConstant(1, DL, MVT::i32));
2758   return splatPartsI64WithVL(DL, VT, Passthru, Lo, Hi, VL, DAG);
2759 }
2760 
2761 // This function lowers a splat of a scalar operand Splat with the vector
2762 // length VL. It ensures the final sequence is type legal, which is useful when
2763 // lowering a splat after type legalization.
lowerScalarSplat(SDValue Passthru,SDValue Scalar,SDValue VL,MVT VT,SDLoc DL,SelectionDAG & DAG,const RISCVSubtarget & Subtarget)2764 static SDValue lowerScalarSplat(SDValue Passthru, SDValue Scalar, SDValue VL,
2765                                 MVT VT, SDLoc DL, SelectionDAG &DAG,
2766                                 const RISCVSubtarget &Subtarget) {
2767   bool HasPassthru = Passthru && !Passthru.isUndef();
2768   if (!HasPassthru && !Passthru)
2769     Passthru = DAG.getUNDEF(VT);
2770   if (VT.isFloatingPoint()) {
2771     // If VL is 1, we could use vfmv.s.f.
2772     if (isOneConstant(VL))
2773       return DAG.getNode(RISCVISD::VFMV_S_F_VL, DL, VT, Passthru, Scalar, VL);
2774     return DAG.getNode(RISCVISD::VFMV_V_F_VL, DL, VT, Passthru, Scalar, VL);
2775   }
2776 
2777   MVT XLenVT = Subtarget.getXLenVT();
2778 
2779   // Simplest case is that the operand needs to be promoted to XLenVT.
2780   if (Scalar.getValueType().bitsLE(XLenVT)) {
2781     // If the operand is a constant, sign extend to increase our chances
2782     // of being able to use a .vi instruction. ANY_EXTEND would become a
2783     // a zero extend and the simm5 check in isel would fail.
2784     // FIXME: Should we ignore the upper bits in isel instead?
2785     unsigned ExtOpc =
2786         isa<ConstantSDNode>(Scalar) ? ISD::SIGN_EXTEND : ISD::ANY_EXTEND;
2787     Scalar = DAG.getNode(ExtOpc, DL, XLenVT, Scalar);
2788     ConstantSDNode *Const = dyn_cast<ConstantSDNode>(Scalar);
2789     // If VL is 1 and the scalar value won't benefit from immediate, we could
2790     // use vmv.s.x.
2791     if (isOneConstant(VL) &&
2792         (!Const || isNullConstant(Scalar) || !isInt<5>(Const->getSExtValue())))
2793       return DAG.getNode(RISCVISD::VMV_S_X_VL, DL, VT, Passthru, Scalar, VL);
2794     return DAG.getNode(RISCVISD::VMV_V_X_VL, DL, VT, Passthru, Scalar, VL);
2795   }
2796 
2797   assert(XLenVT == MVT::i32 && Scalar.getValueType() == MVT::i64 &&
2798          "Unexpected scalar for splat lowering!");
2799 
2800   if (isOneConstant(VL) && isNullConstant(Scalar))
2801     return DAG.getNode(RISCVISD::VMV_S_X_VL, DL, VT, Passthru,
2802                        DAG.getConstant(0, DL, XLenVT), VL);
2803 
2804   // Otherwise use the more complicated splatting algorithm.
2805   return splatSplitI64WithVL(DL, VT, Passthru, Scalar, VL, DAG);
2806 }
2807 
getLMUL1VT(MVT VT)2808 static MVT getLMUL1VT(MVT VT) {
2809   assert(VT.getVectorElementType().getSizeInBits() <= 64 &&
2810          "Unexpected vector MVT");
2811   return MVT::getScalableVectorVT(
2812       VT.getVectorElementType(),
2813       RISCV::RVVBitsPerBlock / VT.getVectorElementType().getSizeInBits());
2814 }
2815 
2816 // This function lowers an insert of a scalar operand Scalar into lane
2817 // 0 of the vector regardless of the value of VL.  The contents of the
2818 // remaining lanes of the result vector are unspecified.  VL is assumed
2819 // to be non-zero.
lowerScalarInsert(SDValue Scalar,SDValue VL,MVT VT,SDLoc DL,SelectionDAG & DAG,const RISCVSubtarget & Subtarget)2820 static SDValue lowerScalarInsert(SDValue Scalar, SDValue VL,
2821                                  MVT VT, SDLoc DL, SelectionDAG &DAG,
2822                                  const RISCVSubtarget &Subtarget) {
2823   const MVT XLenVT = Subtarget.getXLenVT();
2824 
2825   SDValue Passthru = DAG.getUNDEF(VT);
2826   if (VT.isFloatingPoint()) {
2827     // TODO: Use vmv.v.i for appropriate constants
2828     // Use M1 or smaller to avoid over constraining register allocation
2829     const MVT M1VT = getLMUL1VT(VT);
2830     auto InnerVT = VT.bitsLE(M1VT) ? VT : M1VT;
2831     SDValue Result = DAG.getNode(RISCVISD::VFMV_S_F_VL, DL, InnerVT,
2832                                  DAG.getUNDEF(InnerVT), Scalar, VL);
2833     if (VT != InnerVT)
2834       Result = DAG.getNode(ISD::INSERT_SUBVECTOR, DL, VT,
2835                            DAG.getUNDEF(VT),
2836                            Result, DAG.getConstant(0, DL, XLenVT));
2837     return Result;
2838   }
2839 
2840 
2841   // Avoid the tricky legalization cases by falling back to using the
2842   // splat code which already handles it gracefully.
2843   if (!Scalar.getValueType().bitsLE(XLenVT))
2844     return lowerScalarSplat(DAG.getUNDEF(VT), Scalar,
2845                             DAG.getConstant(1, DL, XLenVT),
2846                             VT, DL, DAG, Subtarget);
2847 
2848   // If the operand is a constant, sign extend to increase our chances
2849   // of being able to use a .vi instruction. ANY_EXTEND would become a
2850   // a zero extend and the simm5 check in isel would fail.
2851   // FIXME: Should we ignore the upper bits in isel instead?
2852   unsigned ExtOpc =
2853     isa<ConstantSDNode>(Scalar) ? ISD::SIGN_EXTEND : ISD::ANY_EXTEND;
2854   Scalar = DAG.getNode(ExtOpc, DL, XLenVT, Scalar);
2855   // We use a vmv.v.i if possible.  We limit this to LMUL1.  LMUL2 or
2856   // higher would involve overly constraining the register allocator for
2857   // no purpose.
2858   if (ConstantSDNode *Const = dyn_cast<ConstantSDNode>(Scalar)) {
2859     if (!isNullConstant(Scalar) && isInt<5>(Const->getSExtValue()) &&
2860         VT.bitsLE(getLMUL1VT(VT)))
2861       return DAG.getNode(RISCVISD::VMV_V_X_VL, DL, VT, Passthru, Scalar, VL);
2862   }
2863   // Use M1 or smaller to avoid over constraining register allocation
2864   const MVT M1VT = getLMUL1VT(VT);
2865   auto InnerVT = VT.bitsLE(M1VT) ? VT : M1VT;
2866   SDValue Result = DAG.getNode(RISCVISD::VMV_S_X_VL, DL, InnerVT,
2867                                DAG.getUNDEF(InnerVT), Scalar, VL);
2868   if (VT != InnerVT)
2869     Result = DAG.getNode(ISD::INSERT_SUBVECTOR, DL, VT,
2870                          DAG.getUNDEF(VT),
2871                          Result, DAG.getConstant(0, DL, XLenVT));
2872   return Result;
2873 
2874 }
2875 
isInterleaveShuffle(ArrayRef<int> Mask,MVT VT,bool & SwapSources,const RISCVSubtarget & Subtarget)2876 static bool isInterleaveShuffle(ArrayRef<int> Mask, MVT VT, bool &SwapSources,
2877                                 const RISCVSubtarget &Subtarget) {
2878   // We need to be able to widen elements to the next larger integer type.
2879   if (VT.getScalarSizeInBits() >= Subtarget.getELEN())
2880     return false;
2881 
2882   int Size = Mask.size();
2883   assert(Size == (int)VT.getVectorNumElements() && "Unexpected mask size");
2884 
2885   int Srcs[] = {-1, -1};
2886   for (int i = 0; i != Size; ++i) {
2887     // Ignore undef elements.
2888     if (Mask[i] < 0)
2889       continue;
2890 
2891     // Is this an even or odd element.
2892     int Pol = i % 2;
2893 
2894     // Ensure we consistently use the same source for this element polarity.
2895     int Src = Mask[i] / Size;
2896     if (Srcs[Pol] < 0)
2897       Srcs[Pol] = Src;
2898     if (Srcs[Pol] != Src)
2899       return false;
2900 
2901     // Make sure the element within the source is appropriate for this element
2902     // in the destination.
2903     int Elt = Mask[i] % Size;
2904     if (Elt != i / 2)
2905       return false;
2906   }
2907 
2908   // We need to find a source for each polarity and they can't be the same.
2909   if (Srcs[0] < 0 || Srcs[1] < 0 || Srcs[0] == Srcs[1])
2910     return false;
2911 
2912   // Swap the sources if the second source was in the even polarity.
2913   SwapSources = Srcs[0] > Srcs[1];
2914 
2915   return true;
2916 }
2917 
2918 /// Match shuffles that concatenate two vectors, rotate the concatenation,
2919 /// and then extract the original number of elements from the rotated result.
2920 /// This is equivalent to vector.splice or X86's PALIGNR instruction. The
2921 /// returned rotation amount is for a rotate right, where elements move from
2922 /// higher elements to lower elements. \p LoSrc indicates the first source
2923 /// vector of the rotate or -1 for undef. \p HiSrc indicates the second vector
2924 /// of the rotate or -1 for undef. At least one of \p LoSrc and \p HiSrc will be
2925 /// 0 or 1 if a rotation is found.
2926 ///
2927 /// NOTE: We talk about rotate to the right which matches how bit shift and
2928 /// rotate instructions are described where LSBs are on the right, but LLVM IR
2929 /// and the table below write vectors with the lowest elements on the left.
isElementRotate(int & LoSrc,int & HiSrc,ArrayRef<int> Mask)2930 static int isElementRotate(int &LoSrc, int &HiSrc, ArrayRef<int> Mask) {
2931   int Size = Mask.size();
2932 
2933   // We need to detect various ways of spelling a rotation:
2934   //   [11, 12, 13, 14, 15,  0,  1,  2]
2935   //   [-1, 12, 13, 14, -1, -1,  1, -1]
2936   //   [-1, -1, -1, -1, -1, -1,  1,  2]
2937   //   [ 3,  4,  5,  6,  7,  8,  9, 10]
2938   //   [-1,  4,  5,  6, -1, -1,  9, -1]
2939   //   [-1,  4,  5,  6, -1, -1, -1, -1]
2940   int Rotation = 0;
2941   LoSrc = -1;
2942   HiSrc = -1;
2943   for (int i = 0; i != Size; ++i) {
2944     int M = Mask[i];
2945     if (M < 0)
2946       continue;
2947 
2948     // Determine where a rotate vector would have started.
2949     int StartIdx = i - (M % Size);
2950     // The identity rotation isn't interesting, stop.
2951     if (StartIdx == 0)
2952       return -1;
2953 
2954     // If we found the tail of a vector the rotation must be the missing
2955     // front. If we found the head of a vector, it must be how much of the
2956     // head.
2957     int CandidateRotation = StartIdx < 0 ? -StartIdx : Size - StartIdx;
2958 
2959     if (Rotation == 0)
2960       Rotation = CandidateRotation;
2961     else if (Rotation != CandidateRotation)
2962       // The rotations don't match, so we can't match this mask.
2963       return -1;
2964 
2965     // Compute which value this mask is pointing at.
2966     int MaskSrc = M < Size ? 0 : 1;
2967 
2968     // Compute which of the two target values this index should be assigned to.
2969     // This reflects whether the high elements are remaining or the low elemnts
2970     // are remaining.
2971     int &TargetSrc = StartIdx < 0 ? HiSrc : LoSrc;
2972 
2973     // Either set up this value if we've not encountered it before, or check
2974     // that it remains consistent.
2975     if (TargetSrc < 0)
2976       TargetSrc = MaskSrc;
2977     else if (TargetSrc != MaskSrc)
2978       // This may be a rotation, but it pulls from the inputs in some
2979       // unsupported interleaving.
2980       return -1;
2981   }
2982 
2983   // Check that we successfully analyzed the mask, and normalize the results.
2984   assert(Rotation != 0 && "Failed to locate a viable rotation!");
2985   assert((LoSrc >= 0 || HiSrc >= 0) &&
2986          "Failed to find a rotated input vector!");
2987 
2988   return Rotation;
2989 }
2990 
2991 // Lower the following shuffles to vnsrl.
2992 // t34: v8i8 = extract_subvector t11, Constant:i64<0>
2993 // t33: v8i8 = extract_subvector t11, Constant:i64<8>
2994 // a) t35: v8i8 = vector_shuffle<0,2,4,6,8,10,12,14> t34, t33
2995 // b) t35: v8i8 = vector_shuffle<1,3,5,7,9,11,13,15> t34, t33
lowerVECTOR_SHUFFLEAsVNSRL(const SDLoc & DL,MVT VT,MVT ContainerVT,SDValue V1,SDValue V2,SDValue TrueMask,SDValue VL,ArrayRef<int> Mask,const RISCVSubtarget & Subtarget,SelectionDAG & DAG)2996 static SDValue lowerVECTOR_SHUFFLEAsVNSRL(const SDLoc &DL, MVT VT,
2997                                           MVT ContainerVT, SDValue V1,
2998                                           SDValue V2, SDValue TrueMask,
2999                                           SDValue VL, ArrayRef<int> Mask,
3000                                           const RISCVSubtarget &Subtarget,
3001                                           SelectionDAG &DAG) {
3002   // Need to be able to widen the vector.
3003   if (VT.getScalarSizeInBits() >= Subtarget.getELEN())
3004     return SDValue();
3005 
3006   // Both input must be extracts.
3007   if (V1.getOpcode() != ISD::EXTRACT_SUBVECTOR ||
3008       V2.getOpcode() != ISD::EXTRACT_SUBVECTOR)
3009     return SDValue();
3010 
3011   // Extracting from the same source.
3012   SDValue Src = V1.getOperand(0);
3013   if (Src != V2.getOperand(0))
3014     return SDValue();
3015 
3016   // Src needs to have twice the number of elements.
3017   if (Src.getValueType().getVectorNumElements() != (Mask.size() * 2))
3018     return SDValue();
3019 
3020   // The extracts must extract the two halves of the source.
3021   if (V1.getConstantOperandVal(1) != 0 ||
3022       V2.getConstantOperandVal(1) != Mask.size())
3023     return SDValue();
3024 
3025   // First index must be the first even or odd element from V1.
3026   if (Mask[0] != 0 && Mask[0] != 1)
3027     return SDValue();
3028 
3029   // The others must increase by 2 each time.
3030   // TODO: Support undef elements?
3031   for (unsigned i = 1; i != Mask.size(); ++i)
3032     if (Mask[i] != Mask[i - 1] + 2)
3033       return SDValue();
3034 
3035   // Convert the source using a container type with twice the elements. Since
3036   // source VT is legal and twice this VT, we know VT isn't LMUL=8 so it is
3037   // safe to double.
3038   MVT DoubleContainerVT =
3039       MVT::getVectorVT(ContainerVT.getVectorElementType(),
3040                        ContainerVT.getVectorElementCount() * 2);
3041   Src = convertToScalableVector(DoubleContainerVT, Src, DAG, Subtarget);
3042 
3043   // Convert the vector to a wider integer type with the original element
3044   // count. This also converts FP to int.
3045   unsigned EltBits = ContainerVT.getScalarSizeInBits();
3046   MVT WideIntEltVT = MVT::getIntegerVT(EltBits * 2);
3047   MVT WideIntContainerVT =
3048       MVT::getVectorVT(WideIntEltVT, ContainerVT.getVectorElementCount());
3049   Src = DAG.getBitcast(WideIntContainerVT, Src);
3050 
3051   // Convert to the integer version of the container type.
3052   MVT IntEltVT = MVT::getIntegerVT(EltBits);
3053   MVT IntContainerVT =
3054       MVT::getVectorVT(IntEltVT, ContainerVT.getVectorElementCount());
3055 
3056   // If we want even elements, then the shift amount is 0. Otherwise, shift by
3057   // the original element size.
3058   unsigned Shift = Mask[0] == 0 ? 0 : EltBits;
3059   SDValue SplatShift = DAG.getNode(
3060       RISCVISD::VMV_V_X_VL, DL, IntContainerVT, DAG.getUNDEF(ContainerVT),
3061       DAG.getConstant(Shift, DL, Subtarget.getXLenVT()), VL);
3062   SDValue Res =
3063       DAG.getNode(RISCVISD::VNSRL_VL, DL, IntContainerVT, Src, SplatShift,
3064                   DAG.getUNDEF(IntContainerVT), TrueMask, VL);
3065   // Cast back to FP if needed.
3066   Res = DAG.getBitcast(ContainerVT, Res);
3067 
3068   return convertFromScalableVector(VT, Res, DAG, Subtarget);
3069 }
3070 
3071 static SDValue
getVSlidedown(SelectionDAG & DAG,const RISCVSubtarget & Subtarget,SDLoc DL,EVT VT,SDValue Merge,SDValue Op,SDValue Offset,SDValue Mask,SDValue VL,unsigned Policy=RISCVII::TAIL_UNDISTURBED_MASK_UNDISTURBED)3072 getVSlidedown(SelectionDAG &DAG, const RISCVSubtarget &Subtarget, SDLoc DL,
3073               EVT VT, SDValue Merge, SDValue Op, SDValue Offset, SDValue Mask,
3074               SDValue VL,
3075               unsigned Policy = RISCVII::TAIL_UNDISTURBED_MASK_UNDISTURBED) {
3076   if (Merge.isUndef())
3077     Policy = RISCVII::TAIL_AGNOSTIC | RISCVII::MASK_AGNOSTIC;
3078   SDValue PolicyOp = DAG.getTargetConstant(Policy, DL, Subtarget.getXLenVT());
3079   SDValue Ops[] = {Merge, Op, Offset, Mask, VL, PolicyOp};
3080   return DAG.getNode(RISCVISD::VSLIDEDOWN_VL, DL, VT, Ops);
3081 }
3082 
3083 static SDValue
getVSlideup(SelectionDAG & DAG,const RISCVSubtarget & Subtarget,SDLoc DL,EVT VT,SDValue Merge,SDValue Op,SDValue Offset,SDValue Mask,SDValue VL,unsigned Policy=RISCVII::TAIL_UNDISTURBED_MASK_UNDISTURBED)3084 getVSlideup(SelectionDAG &DAG, const RISCVSubtarget &Subtarget, SDLoc DL,
3085             EVT VT, SDValue Merge, SDValue Op, SDValue Offset, SDValue Mask,
3086             SDValue VL,
3087             unsigned Policy = RISCVII::TAIL_UNDISTURBED_MASK_UNDISTURBED) {
3088   if (Merge.isUndef())
3089     Policy = RISCVII::TAIL_AGNOSTIC | RISCVII::MASK_AGNOSTIC;
3090   SDValue PolicyOp = DAG.getTargetConstant(Policy, DL, Subtarget.getXLenVT());
3091   SDValue Ops[] = {Merge, Op, Offset, Mask, VL, PolicyOp};
3092   return DAG.getNode(RISCVISD::VSLIDEUP_VL, DL, VT, Ops);
3093 }
3094 
3095 // Lower the following shuffle to vslidedown.
3096 // a)
3097 // t49: v8i8 = extract_subvector t13, Constant:i64<0>
3098 // t109: v8i8 = extract_subvector t13, Constant:i64<8>
3099 // t108: v8i8 = vector_shuffle<1,2,3,4,5,6,7,8> t49, t106
3100 // b)
3101 // t69: v16i16 = extract_subvector t68, Constant:i64<0>
3102 // t23: v8i16 = extract_subvector t69, Constant:i64<0>
3103 // t29: v4i16 = extract_subvector t23, Constant:i64<4>
3104 // t26: v8i16 = extract_subvector t69, Constant:i64<8>
3105 // t30: v4i16 = extract_subvector t26, Constant:i64<0>
3106 // t54: v4i16 = vector_shuffle<1,2,3,4> t29, t30
lowerVECTOR_SHUFFLEAsVSlidedown(const SDLoc & DL,MVT VT,SDValue V1,SDValue V2,ArrayRef<int> Mask,const RISCVSubtarget & Subtarget,SelectionDAG & DAG)3107 static SDValue lowerVECTOR_SHUFFLEAsVSlidedown(const SDLoc &DL, MVT VT,
3108                                                SDValue V1, SDValue V2,
3109                                                ArrayRef<int> Mask,
3110                                                const RISCVSubtarget &Subtarget,
3111                                                SelectionDAG &DAG) {
3112   auto findNonEXTRACT_SUBVECTORParent =
3113       [](SDValue Parent) -> std::pair<SDValue, uint64_t> {
3114     uint64_t Offset = 0;
3115     while (Parent.getOpcode() == ISD::EXTRACT_SUBVECTOR &&
3116            // EXTRACT_SUBVECTOR can be used to extract a fixed-width vector from
3117            // a scalable vector. But we don't want to match the case.
3118            Parent.getOperand(0).getSimpleValueType().isFixedLengthVector()) {
3119       Offset += Parent.getConstantOperandVal(1);
3120       Parent = Parent.getOperand(0);
3121     }
3122     return std::make_pair(Parent, Offset);
3123   };
3124 
3125   auto [V1Src, V1IndexOffset] = findNonEXTRACT_SUBVECTORParent(V1);
3126   auto [V2Src, V2IndexOffset] = findNonEXTRACT_SUBVECTORParent(V2);
3127 
3128   // Extracting from the same source.
3129   SDValue Src = V1Src;
3130   if (Src != V2Src)
3131     return SDValue();
3132 
3133   // Rebuild mask because Src may be from multiple EXTRACT_SUBVECTORs.
3134   SmallVector<int, 16> NewMask(Mask);
3135   for (size_t i = 0; i != NewMask.size(); ++i) {
3136     if (NewMask[i] == -1)
3137       continue;
3138 
3139     if (static_cast<size_t>(NewMask[i]) < NewMask.size()) {
3140       NewMask[i] = NewMask[i] + V1IndexOffset;
3141     } else {
3142       // Minus NewMask.size() is needed. Otherwise, the b case would be
3143       // <5,6,7,12> instead of <5,6,7,8>.
3144       NewMask[i] = NewMask[i] - NewMask.size() + V2IndexOffset;
3145     }
3146   }
3147 
3148   // First index must be known and non-zero. It will be used as the slidedown
3149   // amount.
3150   if (NewMask[0] <= 0)
3151     return SDValue();
3152 
3153   // NewMask is also continuous.
3154   for (unsigned i = 1; i != NewMask.size(); ++i)
3155     if (NewMask[i - 1] + 1 != NewMask[i])
3156       return SDValue();
3157 
3158   MVT XLenVT = Subtarget.getXLenVT();
3159   MVT SrcVT = Src.getSimpleValueType();
3160   MVT ContainerVT = getContainerForFixedLengthVector(DAG, SrcVT, Subtarget);
3161   auto [TrueMask, VL] = getDefaultVLOps(SrcVT, ContainerVT, DL, DAG, Subtarget);
3162   SDValue Slidedown =
3163       getVSlidedown(DAG, Subtarget, DL, ContainerVT, DAG.getUNDEF(ContainerVT),
3164                     convertToScalableVector(ContainerVT, Src, DAG, Subtarget),
3165                     DAG.getConstant(NewMask[0], DL, XLenVT), TrueMask, VL);
3166   return DAG.getNode(
3167       ISD::EXTRACT_SUBVECTOR, DL, VT,
3168       convertFromScalableVector(SrcVT, Slidedown, DAG, Subtarget),
3169       DAG.getConstant(0, DL, XLenVT));
3170 }
3171 
lowerVECTOR_SHUFFLE(SDValue Op,SelectionDAG & DAG,const RISCVSubtarget & Subtarget)3172 static SDValue lowerVECTOR_SHUFFLE(SDValue Op, SelectionDAG &DAG,
3173                                    const RISCVSubtarget &Subtarget) {
3174   SDValue V1 = Op.getOperand(0);
3175   SDValue V2 = Op.getOperand(1);
3176   SDLoc DL(Op);
3177   MVT XLenVT = Subtarget.getXLenVT();
3178   MVT VT = Op.getSimpleValueType();
3179   unsigned NumElts = VT.getVectorNumElements();
3180   ShuffleVectorSDNode *SVN = cast<ShuffleVectorSDNode>(Op.getNode());
3181 
3182   MVT ContainerVT = getContainerForFixedLengthVector(DAG, VT, Subtarget);
3183 
3184   auto [TrueMask, VL] = getDefaultVLOps(VT, ContainerVT, DL, DAG, Subtarget);
3185 
3186   if (SVN->isSplat()) {
3187     const int Lane = SVN->getSplatIndex();
3188     if (Lane >= 0) {
3189       MVT SVT = VT.getVectorElementType();
3190 
3191       // Turn splatted vector load into a strided load with an X0 stride.
3192       SDValue V = V1;
3193       // Peek through CONCAT_VECTORS as VectorCombine can concat a vector
3194       // with undef.
3195       // FIXME: Peek through INSERT_SUBVECTOR, EXTRACT_SUBVECTOR, bitcasts?
3196       int Offset = Lane;
3197       if (V.getOpcode() == ISD::CONCAT_VECTORS) {
3198         int OpElements =
3199             V.getOperand(0).getSimpleValueType().getVectorNumElements();
3200         V = V.getOperand(Offset / OpElements);
3201         Offset %= OpElements;
3202       }
3203 
3204       // We need to ensure the load isn't atomic or volatile.
3205       if (ISD::isNormalLoad(V.getNode()) && cast<LoadSDNode>(V)->isSimple()) {
3206         auto *Ld = cast<LoadSDNode>(V);
3207         Offset *= SVT.getStoreSize();
3208         SDValue NewAddr = DAG.getMemBasePlusOffset(Ld->getBasePtr(),
3209                                                    TypeSize::Fixed(Offset), DL);
3210 
3211         // If this is SEW=64 on RV32, use a strided load with a stride of x0.
3212         if (SVT.isInteger() && SVT.bitsGT(XLenVT)) {
3213           SDVTList VTs = DAG.getVTList({ContainerVT, MVT::Other});
3214           SDValue IntID =
3215               DAG.getTargetConstant(Intrinsic::riscv_vlse, DL, XLenVT);
3216           SDValue Ops[] = {Ld->getChain(),
3217                            IntID,
3218                            DAG.getUNDEF(ContainerVT),
3219                            NewAddr,
3220                            DAG.getRegister(RISCV::X0, XLenVT),
3221                            VL};
3222           SDValue NewLoad = DAG.getMemIntrinsicNode(
3223               ISD::INTRINSIC_W_CHAIN, DL, VTs, Ops, SVT,
3224               DAG.getMachineFunction().getMachineMemOperand(
3225                   Ld->getMemOperand(), Offset, SVT.getStoreSize()));
3226           DAG.makeEquivalentMemoryOrdering(Ld, NewLoad);
3227           return convertFromScalableVector(VT, NewLoad, DAG, Subtarget);
3228         }
3229 
3230         // Otherwise use a scalar load and splat. This will give the best
3231         // opportunity to fold a splat into the operation. ISel can turn it into
3232         // the x0 strided load if we aren't able to fold away the select.
3233         if (SVT.isFloatingPoint())
3234           V = DAG.getLoad(SVT, DL, Ld->getChain(), NewAddr,
3235                           Ld->getPointerInfo().getWithOffset(Offset),
3236                           Ld->getOriginalAlign(),
3237                           Ld->getMemOperand()->getFlags());
3238         else
3239           V = DAG.getExtLoad(ISD::SEXTLOAD, DL, XLenVT, Ld->getChain(), NewAddr,
3240                              Ld->getPointerInfo().getWithOffset(Offset), SVT,
3241                              Ld->getOriginalAlign(),
3242                              Ld->getMemOperand()->getFlags());
3243         DAG.makeEquivalentMemoryOrdering(Ld, V);
3244 
3245         unsigned Opc =
3246             VT.isFloatingPoint() ? RISCVISD::VFMV_V_F_VL : RISCVISD::VMV_V_X_VL;
3247         SDValue Splat =
3248             DAG.getNode(Opc, DL, ContainerVT, DAG.getUNDEF(ContainerVT), V, VL);
3249         return convertFromScalableVector(VT, Splat, DAG, Subtarget);
3250       }
3251 
3252       V1 = convertToScalableVector(ContainerVT, V1, DAG, Subtarget);
3253       assert(Lane < (int)NumElts && "Unexpected lane!");
3254       SDValue Gather = DAG.getNode(RISCVISD::VRGATHER_VX_VL, DL, ContainerVT,
3255                                    V1, DAG.getConstant(Lane, DL, XLenVT),
3256                                    DAG.getUNDEF(ContainerVT), TrueMask, VL);
3257       return convertFromScalableVector(VT, Gather, DAG, Subtarget);
3258     }
3259   }
3260 
3261   ArrayRef<int> Mask = SVN->getMask();
3262 
3263   if (SDValue V =
3264           lowerVECTOR_SHUFFLEAsVSlidedown(DL, VT, V1, V2, Mask, Subtarget, DAG))
3265     return V;
3266 
3267   // Lower rotations to a SLIDEDOWN and a SLIDEUP. One of the source vectors may
3268   // be undef which can be handled with a single SLIDEDOWN/UP.
3269   int LoSrc, HiSrc;
3270   int Rotation = isElementRotate(LoSrc, HiSrc, Mask);
3271   if (Rotation > 0) {
3272     SDValue LoV, HiV;
3273     if (LoSrc >= 0) {
3274       LoV = LoSrc == 0 ? V1 : V2;
3275       LoV = convertToScalableVector(ContainerVT, LoV, DAG, Subtarget);
3276     }
3277     if (HiSrc >= 0) {
3278       HiV = HiSrc == 0 ? V1 : V2;
3279       HiV = convertToScalableVector(ContainerVT, HiV, DAG, Subtarget);
3280     }
3281 
3282     // We found a rotation. We need to slide HiV down by Rotation. Then we need
3283     // to slide LoV up by (NumElts - Rotation).
3284     unsigned InvRotate = NumElts - Rotation;
3285 
3286     SDValue Res = DAG.getUNDEF(ContainerVT);
3287     if (HiV) {
3288       // If we are doing a SLIDEDOWN+SLIDEUP, reduce the VL for the SLIDEDOWN.
3289       // FIXME: If we are only doing a SLIDEDOWN, don't reduce the VL as it
3290       // causes multiple vsetvlis in some test cases such as lowering
3291       // reduce.mul
3292       SDValue DownVL = VL;
3293       if (LoV)
3294         DownVL = DAG.getConstant(InvRotate, DL, XLenVT);
3295       Res = getVSlidedown(DAG, Subtarget, DL, ContainerVT, Res, HiV,
3296                           DAG.getConstant(Rotation, DL, XLenVT), TrueMask,
3297                           DownVL);
3298     }
3299     if (LoV)
3300       Res = getVSlideup(DAG, Subtarget, DL, ContainerVT, Res, LoV,
3301                         DAG.getConstant(InvRotate, DL, XLenVT), TrueMask, VL,
3302                         RISCVII::TAIL_AGNOSTIC);
3303 
3304     return convertFromScalableVector(VT, Res, DAG, Subtarget);
3305   }
3306 
3307   if (SDValue V = lowerVECTOR_SHUFFLEAsVNSRL(
3308           DL, VT, ContainerVT, V1, V2, TrueMask, VL, Mask, Subtarget, DAG))
3309     return V;
3310 
3311   // Detect an interleave shuffle and lower to
3312   // (vmaccu.vx (vwaddu.vx lohalf(V1), lohalf(V2)), lohalf(V2), (2^eltbits - 1))
3313   bool SwapSources;
3314   if (isInterleaveShuffle(Mask, VT, SwapSources, Subtarget)) {
3315     // Swap sources if needed.
3316     if (SwapSources)
3317       std::swap(V1, V2);
3318 
3319     // Extract the lower half of the vectors.
3320     MVT HalfVT = VT.getHalfNumVectorElementsVT();
3321     V1 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, HalfVT, V1,
3322                      DAG.getConstant(0, DL, XLenVT));
3323     V2 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, HalfVT, V2,
3324                      DAG.getConstant(0, DL, XLenVT));
3325 
3326     // Double the element width and halve the number of elements in an int type.
3327     unsigned EltBits = VT.getScalarSizeInBits();
3328     MVT WideIntEltVT = MVT::getIntegerVT(EltBits * 2);
3329     MVT WideIntVT =
3330         MVT::getVectorVT(WideIntEltVT, VT.getVectorNumElements() / 2);
3331     // Convert this to a scalable vector. We need to base this on the
3332     // destination size to ensure there's always a type with a smaller LMUL.
3333     MVT WideIntContainerVT =
3334         getContainerForFixedLengthVector(DAG, WideIntVT, Subtarget);
3335 
3336     // Convert sources to scalable vectors with the same element count as the
3337     // larger type.
3338     MVT HalfContainerVT = MVT::getVectorVT(
3339         VT.getVectorElementType(), WideIntContainerVT.getVectorElementCount());
3340     V1 = convertToScalableVector(HalfContainerVT, V1, DAG, Subtarget);
3341     V2 = convertToScalableVector(HalfContainerVT, V2, DAG, Subtarget);
3342 
3343     // Cast sources to integer.
3344     MVT IntEltVT = MVT::getIntegerVT(EltBits);
3345     MVT IntHalfVT =
3346         MVT::getVectorVT(IntEltVT, HalfContainerVT.getVectorElementCount());
3347     V1 = DAG.getBitcast(IntHalfVT, V1);
3348     V2 = DAG.getBitcast(IntHalfVT, V2);
3349 
3350     // Freeze V2 since we use it twice and we need to be sure that the add and
3351     // multiply see the same value.
3352     V2 = DAG.getFreeze(V2);
3353 
3354     // Recreate TrueMask using the widened type's element count.
3355     TrueMask = getAllOnesMask(HalfContainerVT, VL, DL, DAG);
3356 
3357     // Widen V1 and V2 with 0s and add one copy of V2 to V1.
3358     SDValue Add =
3359         DAG.getNode(RISCVISD::VWADDU_VL, DL, WideIntContainerVT, V1, V2,
3360                     DAG.getUNDEF(WideIntContainerVT), TrueMask, VL);
3361     // Create 2^eltbits - 1 copies of V2 by multiplying by the largest integer.
3362     SDValue Multiplier = DAG.getNode(RISCVISD::VMV_V_X_VL, DL, IntHalfVT,
3363                                      DAG.getUNDEF(IntHalfVT),
3364                                      DAG.getAllOnesConstant(DL, XLenVT), VL);
3365     SDValue WidenMul =
3366         DAG.getNode(RISCVISD::VWMULU_VL, DL, WideIntContainerVT, V2, Multiplier,
3367                     DAG.getUNDEF(WideIntContainerVT), TrueMask, VL);
3368     // Add the new copies to our previous addition giving us 2^eltbits copies of
3369     // V2. This is equivalent to shifting V2 left by eltbits. This should
3370     // combine with the vwmulu.vv above to form vwmaccu.vv.
3371     Add = DAG.getNode(RISCVISD::ADD_VL, DL, WideIntContainerVT, Add, WidenMul,
3372                       DAG.getUNDEF(WideIntContainerVT), TrueMask, VL);
3373     // Cast back to ContainerVT. We need to re-create a new ContainerVT in case
3374     // WideIntContainerVT is a larger fractional LMUL than implied by the fixed
3375     // vector VT.
3376     ContainerVT =
3377         MVT::getVectorVT(VT.getVectorElementType(),
3378                          WideIntContainerVT.getVectorElementCount() * 2);
3379     Add = DAG.getBitcast(ContainerVT, Add);
3380     return convertFromScalableVector(VT, Add, DAG, Subtarget);
3381   }
3382 
3383   // Detect shuffles which can be re-expressed as vector selects; these are
3384   // shuffles in which each element in the destination is taken from an element
3385   // at the corresponding index in either source vectors.
3386   bool IsSelect = all_of(enumerate(Mask), [&](const auto &MaskIdx) {
3387     int MaskIndex = MaskIdx.value();
3388     return MaskIndex < 0 || MaskIdx.index() == (unsigned)MaskIndex % NumElts;
3389   });
3390 
3391   assert(!V1.isUndef() && "Unexpected shuffle canonicalization");
3392 
3393   SmallVector<SDValue> MaskVals;
3394   // As a backup, shuffles can be lowered via a vrgather instruction, possibly
3395   // merged with a second vrgather.
3396   SmallVector<SDValue> GatherIndicesLHS, GatherIndicesRHS;
3397 
3398   // By default we preserve the original operand order, and use a mask to
3399   // select LHS as true and RHS as false. However, since RVV vector selects may
3400   // feature splats but only on the LHS, we may choose to invert our mask and
3401   // instead select between RHS and LHS.
3402   bool SwapOps = DAG.isSplatValue(V2) && !DAG.isSplatValue(V1);
3403   bool InvertMask = IsSelect == SwapOps;
3404 
3405   // Keep a track of which non-undef indices are used by each LHS/RHS shuffle
3406   // half.
3407   DenseMap<int, unsigned> LHSIndexCounts, RHSIndexCounts;
3408 
3409   // Now construct the mask that will be used by the vselect or blended
3410   // vrgather operation. For vrgathers, construct the appropriate indices into
3411   // each vector.
3412   for (int MaskIndex : Mask) {
3413     bool SelectMaskVal = (MaskIndex < (int)NumElts) ^ InvertMask;
3414     MaskVals.push_back(DAG.getConstant(SelectMaskVal, DL, XLenVT));
3415     if (!IsSelect) {
3416       bool IsLHSOrUndefIndex = MaskIndex < (int)NumElts;
3417       GatherIndicesLHS.push_back(IsLHSOrUndefIndex && MaskIndex >= 0
3418                                      ? DAG.getConstant(MaskIndex, DL, XLenVT)
3419                                      : DAG.getUNDEF(XLenVT));
3420       GatherIndicesRHS.push_back(
3421           IsLHSOrUndefIndex ? DAG.getUNDEF(XLenVT)
3422                             : DAG.getConstant(MaskIndex - NumElts, DL, XLenVT));
3423       if (IsLHSOrUndefIndex && MaskIndex >= 0)
3424         ++LHSIndexCounts[MaskIndex];
3425       if (!IsLHSOrUndefIndex)
3426         ++RHSIndexCounts[MaskIndex - NumElts];
3427     }
3428   }
3429 
3430   if (SwapOps) {
3431     std::swap(V1, V2);
3432     std::swap(GatherIndicesLHS, GatherIndicesRHS);
3433   }
3434 
3435   assert(MaskVals.size() == NumElts && "Unexpected select-like shuffle");
3436   MVT MaskVT = MVT::getVectorVT(MVT::i1, NumElts);
3437   SDValue SelectMask = DAG.getBuildVector(MaskVT, DL, MaskVals);
3438 
3439   if (IsSelect)
3440     return DAG.getNode(ISD::VSELECT, DL, VT, SelectMask, V1, V2);
3441 
3442   if (VT.getScalarSizeInBits() == 8 && VT.getVectorNumElements() > 256) {
3443     // On such a large vector we're unable to use i8 as the index type.
3444     // FIXME: We could promote the index to i16 and use vrgatherei16, but that
3445     // may involve vector splitting if we're already at LMUL=8, or our
3446     // user-supplied maximum fixed-length LMUL.
3447     return SDValue();
3448   }
3449 
3450   unsigned GatherVXOpc = RISCVISD::VRGATHER_VX_VL;
3451   unsigned GatherVVOpc = RISCVISD::VRGATHER_VV_VL;
3452   MVT IndexVT = VT.changeTypeToInteger();
3453   // Since we can't introduce illegal index types at this stage, use i16 and
3454   // vrgatherei16 if the corresponding index type for plain vrgather is greater
3455   // than XLenVT.
3456   if (IndexVT.getScalarType().bitsGT(XLenVT)) {
3457     GatherVVOpc = RISCVISD::VRGATHEREI16_VV_VL;
3458     IndexVT = IndexVT.changeVectorElementType(MVT::i16);
3459   }
3460 
3461   MVT IndexContainerVT =
3462       ContainerVT.changeVectorElementType(IndexVT.getScalarType());
3463 
3464   SDValue Gather;
3465   // TODO: This doesn't trigger for i64 vectors on RV32, since there we
3466   // encounter a bitcasted BUILD_VECTOR with low/high i32 values.
3467   if (SDValue SplatValue = DAG.getSplatValue(V1, /*LegalTypes*/ true)) {
3468     Gather = lowerScalarSplat(SDValue(), SplatValue, VL, ContainerVT, DL, DAG,
3469                               Subtarget);
3470   } else {
3471     V1 = convertToScalableVector(ContainerVT, V1, DAG, Subtarget);
3472     // If only one index is used, we can use a "splat" vrgather.
3473     // TODO: We can splat the most-common index and fix-up any stragglers, if
3474     // that's beneficial.
3475     if (LHSIndexCounts.size() == 1) {
3476       int SplatIndex = LHSIndexCounts.begin()->getFirst();
3477       Gather = DAG.getNode(GatherVXOpc, DL, ContainerVT, V1,
3478                            DAG.getConstant(SplatIndex, DL, XLenVT),
3479                            DAG.getUNDEF(ContainerVT), TrueMask, VL);
3480     } else {
3481       SDValue LHSIndices = DAG.getBuildVector(IndexVT, DL, GatherIndicesLHS);
3482       LHSIndices =
3483           convertToScalableVector(IndexContainerVT, LHSIndices, DAG, Subtarget);
3484 
3485       Gather = DAG.getNode(GatherVVOpc, DL, ContainerVT, V1, LHSIndices,
3486                            DAG.getUNDEF(ContainerVT), TrueMask, VL);
3487     }
3488   }
3489 
3490   // If a second vector operand is used by this shuffle, blend it in with an
3491   // additional vrgather.
3492   if (!V2.isUndef()) {
3493     V2 = convertToScalableVector(ContainerVT, V2, DAG, Subtarget);
3494 
3495     MVT MaskContainerVT = ContainerVT.changeVectorElementType(MVT::i1);
3496     SelectMask =
3497         convertToScalableVector(MaskContainerVT, SelectMask, DAG, Subtarget);
3498 
3499     // If only one index is used, we can use a "splat" vrgather.
3500     // TODO: We can splat the most-common index and fix-up any stragglers, if
3501     // that's beneficial.
3502     if (RHSIndexCounts.size() == 1) {
3503       int SplatIndex = RHSIndexCounts.begin()->getFirst();
3504       Gather = DAG.getNode(GatherVXOpc, DL, ContainerVT, V2,
3505                            DAG.getConstant(SplatIndex, DL, XLenVT), Gather,
3506                            SelectMask, VL);
3507     } else {
3508       SDValue RHSIndices = DAG.getBuildVector(IndexVT, DL, GatherIndicesRHS);
3509       RHSIndices =
3510           convertToScalableVector(IndexContainerVT, RHSIndices, DAG, Subtarget);
3511       Gather = DAG.getNode(GatherVVOpc, DL, ContainerVT, V2, RHSIndices, Gather,
3512                            SelectMask, VL);
3513     }
3514   }
3515 
3516   return convertFromScalableVector(VT, Gather, DAG, Subtarget);
3517 }
3518 
isShuffleMaskLegal(ArrayRef<int> M,EVT VT) const3519 bool RISCVTargetLowering::isShuffleMaskLegal(ArrayRef<int> M, EVT VT) const {
3520   // Support splats for any type. These should type legalize well.
3521   if (ShuffleVectorSDNode::isSplatMask(M.data(), VT))
3522     return true;
3523 
3524   // Only support legal VTs for other shuffles for now.
3525   if (!isTypeLegal(VT))
3526     return false;
3527 
3528   MVT SVT = VT.getSimpleVT();
3529 
3530   bool SwapSources;
3531   int LoSrc, HiSrc;
3532   return (isElementRotate(LoSrc, HiSrc, M) > 0) ||
3533          isInterleaveShuffle(M, SVT, SwapSources, Subtarget);
3534 }
3535 
3536 // Lower CTLZ_ZERO_UNDEF or CTTZ_ZERO_UNDEF by converting to FP and extracting
3537 // the exponent.
3538 SDValue
lowerCTLZ_CTTZ_ZERO_UNDEF(SDValue Op,SelectionDAG & DAG) const3539 RISCVTargetLowering::lowerCTLZ_CTTZ_ZERO_UNDEF(SDValue Op,
3540                                                SelectionDAG &DAG) const {
3541   MVT VT = Op.getSimpleValueType();
3542   unsigned EltSize = VT.getScalarSizeInBits();
3543   SDValue Src = Op.getOperand(0);
3544   SDLoc DL(Op);
3545 
3546   // We choose FP type that can represent the value if possible. Otherwise, we
3547   // use rounding to zero conversion for correct exponent of the result.
3548   // TODO: Use f16 for i8 when possible?
3549   MVT FloatEltVT = (EltSize >= 32) ? MVT::f64 : MVT::f32;
3550   if (!isTypeLegal(MVT::getVectorVT(FloatEltVT, VT.getVectorElementCount())))
3551     FloatEltVT = MVT::f32;
3552   MVT FloatVT = MVT::getVectorVT(FloatEltVT, VT.getVectorElementCount());
3553 
3554   // Legal types should have been checked in the RISCVTargetLowering
3555   // constructor.
3556   // TODO: Splitting may make sense in some cases.
3557   assert(DAG.getTargetLoweringInfo().isTypeLegal(FloatVT) &&
3558          "Expected legal float type!");
3559 
3560   // For CTTZ_ZERO_UNDEF, we need to extract the lowest set bit using X & -X.
3561   // The trailing zero count is equal to log2 of this single bit value.
3562   if (Op.getOpcode() == ISD::CTTZ_ZERO_UNDEF) {
3563     SDValue Neg = DAG.getNegative(Src, DL, VT);
3564     Src = DAG.getNode(ISD::AND, DL, VT, Src, Neg);
3565   }
3566 
3567   // We have a legal FP type, convert to it.
3568   SDValue FloatVal;
3569   if (FloatVT.bitsGT(VT)) {
3570     FloatVal = DAG.getNode(ISD::UINT_TO_FP, DL, FloatVT, Src);
3571   } else {
3572     // Use RTZ to avoid rounding influencing exponent of FloatVal.
3573     MVT ContainerVT = VT;
3574     if (VT.isFixedLengthVector()) {
3575       ContainerVT = getContainerForFixedLengthVector(VT);
3576       Src = convertToScalableVector(ContainerVT, Src, DAG, Subtarget);
3577     }
3578 
3579     auto [Mask, VL] = getDefaultVLOps(VT, ContainerVT, DL, DAG, Subtarget);
3580     SDValue RTZRM =
3581         DAG.getTargetConstant(RISCVFPRndMode::RTZ, DL, Subtarget.getXLenVT());
3582     MVT ContainerFloatVT =
3583         MVT::getVectorVT(FloatEltVT, ContainerVT.getVectorElementCount());
3584     FloatVal = DAG.getNode(RISCVISD::VFCVT_RM_F_XU_VL, DL, ContainerFloatVT,
3585                            Src, Mask, RTZRM, VL);
3586     if (VT.isFixedLengthVector())
3587       FloatVal = convertFromScalableVector(FloatVT, FloatVal, DAG, Subtarget);
3588   }
3589   // Bitcast to integer and shift the exponent to the LSB.
3590   EVT IntVT = FloatVT.changeVectorElementTypeToInteger();
3591   SDValue Bitcast = DAG.getBitcast(IntVT, FloatVal);
3592   unsigned ShiftAmt = FloatEltVT == MVT::f64 ? 52 : 23;
3593   SDValue Exp = DAG.getNode(ISD::SRL, DL, IntVT, Bitcast,
3594                             DAG.getConstant(ShiftAmt, DL, IntVT));
3595   // Restore back to original type. Truncation after SRL is to generate vnsrl.
3596   if (IntVT.bitsLT(VT))
3597     Exp = DAG.getNode(ISD::ZERO_EXTEND, DL, VT, Exp);
3598   else if (IntVT.bitsGT(VT))
3599     Exp = DAG.getNode(ISD::TRUNCATE, DL, VT, Exp);
3600   // The exponent contains log2 of the value in biased form.
3601   unsigned ExponentBias = FloatEltVT == MVT::f64 ? 1023 : 127;
3602 
3603   // For trailing zeros, we just need to subtract the bias.
3604   if (Op.getOpcode() == ISD::CTTZ_ZERO_UNDEF)
3605     return DAG.getNode(ISD::SUB, DL, VT, Exp,
3606                        DAG.getConstant(ExponentBias, DL, VT));
3607 
3608   // For leading zeros, we need to remove the bias and convert from log2 to
3609   // leading zeros. We can do this by subtracting from (Bias + (EltSize - 1)).
3610   unsigned Adjust = ExponentBias + (EltSize - 1);
3611   SDValue Res =
3612       DAG.getNode(ISD::SUB, DL, VT, DAG.getConstant(Adjust, DL, VT), Exp);
3613   // The above result with zero input equals to Adjust which is greater than
3614   // EltSize. Hence, we can do min(Res, EltSize) for CTLZ.
3615   if (Op.getOpcode() == ISD::CTLZ)
3616     Res = DAG.getNode(ISD::UMIN, DL, VT, Res, DAG.getConstant(EltSize, DL, VT));
3617   return Res;
3618 }
3619 
3620 // While RVV has alignment restrictions, we should always be able to load as a
3621 // legal equivalently-sized byte-typed vector instead. This method is
3622 // responsible for re-expressing a ISD::LOAD via a correctly-aligned type. If
3623 // the load is already correctly-aligned, it returns SDValue().
expandUnalignedRVVLoad(SDValue Op,SelectionDAG & DAG) const3624 SDValue RISCVTargetLowering::expandUnalignedRVVLoad(SDValue Op,
3625                                                     SelectionDAG &DAG) const {
3626   auto *Load = cast<LoadSDNode>(Op);
3627   assert(Load && Load->getMemoryVT().isVector() && "Expected vector load");
3628 
3629   if (allowsMemoryAccessForAlignment(*DAG.getContext(), DAG.getDataLayout(),
3630                                      Load->getMemoryVT(),
3631                                      *Load->getMemOperand()))
3632     return SDValue();
3633 
3634   SDLoc DL(Op);
3635   MVT VT = Op.getSimpleValueType();
3636   unsigned EltSizeBits = VT.getScalarSizeInBits();
3637   assert((EltSizeBits == 16 || EltSizeBits == 32 || EltSizeBits == 64) &&
3638          "Unexpected unaligned RVV load type");
3639   MVT NewVT =
3640       MVT::getVectorVT(MVT::i8, VT.getVectorElementCount() * (EltSizeBits / 8));
3641   assert(NewVT.isValid() &&
3642          "Expecting equally-sized RVV vector types to be legal");
3643   SDValue L = DAG.getLoad(NewVT, DL, Load->getChain(), Load->getBasePtr(),
3644                           Load->getPointerInfo(), Load->getOriginalAlign(),
3645                           Load->getMemOperand()->getFlags());
3646   return DAG.getMergeValues({DAG.getBitcast(VT, L), L.getValue(1)}, DL);
3647 }
3648 
3649 // While RVV has alignment restrictions, we should always be able to store as a
3650 // legal equivalently-sized byte-typed vector instead. This method is
3651 // responsible for re-expressing a ISD::STORE via a correctly-aligned type. It
3652 // returns SDValue() if the store is already correctly aligned.
expandUnalignedRVVStore(SDValue Op,SelectionDAG & DAG) const3653 SDValue RISCVTargetLowering::expandUnalignedRVVStore(SDValue Op,
3654                                                      SelectionDAG &DAG) const {
3655   auto *Store = cast<StoreSDNode>(Op);
3656   assert(Store && Store->getValue().getValueType().isVector() &&
3657          "Expected vector store");
3658 
3659   if (allowsMemoryAccessForAlignment(*DAG.getContext(), DAG.getDataLayout(),
3660                                      Store->getMemoryVT(),
3661                                      *Store->getMemOperand()))
3662     return SDValue();
3663 
3664   SDLoc DL(Op);
3665   SDValue StoredVal = Store->getValue();
3666   MVT VT = StoredVal.getSimpleValueType();
3667   unsigned EltSizeBits = VT.getScalarSizeInBits();
3668   assert((EltSizeBits == 16 || EltSizeBits == 32 || EltSizeBits == 64) &&
3669          "Unexpected unaligned RVV store type");
3670   MVT NewVT =
3671       MVT::getVectorVT(MVT::i8, VT.getVectorElementCount() * (EltSizeBits / 8));
3672   assert(NewVT.isValid() &&
3673          "Expecting equally-sized RVV vector types to be legal");
3674   StoredVal = DAG.getBitcast(NewVT, StoredVal);
3675   return DAG.getStore(Store->getChain(), DL, StoredVal, Store->getBasePtr(),
3676                       Store->getPointerInfo(), Store->getOriginalAlign(),
3677                       Store->getMemOperand()->getFlags());
3678 }
3679 
lowerConstant(SDValue Op,SelectionDAG & DAG,const RISCVSubtarget & Subtarget)3680 static SDValue lowerConstant(SDValue Op, SelectionDAG &DAG,
3681                              const RISCVSubtarget &Subtarget) {
3682   assert(Op.getValueType() == MVT::i64 && "Unexpected VT");
3683 
3684   int64_t Imm = cast<ConstantSDNode>(Op)->getSExtValue();
3685 
3686   // All simm32 constants should be handled by isel.
3687   // NOTE: The getMaxBuildIntsCost call below should return a value >= 2 making
3688   // this check redundant, but small immediates are common so this check
3689   // should have better compile time.
3690   if (isInt<32>(Imm))
3691     return Op;
3692 
3693   // We only need to cost the immediate, if constant pool lowering is enabled.
3694   if (!Subtarget.useConstantPoolForLargeInts())
3695     return Op;
3696 
3697   RISCVMatInt::InstSeq Seq =
3698       RISCVMatInt::generateInstSeq(Imm, Subtarget.getFeatureBits());
3699   if (Seq.size() <= Subtarget.getMaxBuildIntsCost())
3700     return Op;
3701 
3702   // Expand to a constant pool using the default expansion code.
3703   return SDValue();
3704 }
3705 
LowerATOMIC_FENCE(SDValue Op,SelectionDAG & DAG)3706 static SDValue LowerATOMIC_FENCE(SDValue Op, SelectionDAG &DAG) {
3707   SDLoc dl(Op);
3708   SyncScope::ID FenceSSID =
3709       static_cast<SyncScope::ID>(Op.getConstantOperandVal(2));
3710 
3711   // singlethread fences only synchronize with signal handlers on the same
3712   // thread and thus only need to preserve instruction order, not actually
3713   // enforce memory ordering.
3714   if (FenceSSID == SyncScope::SingleThread)
3715     // MEMBARRIER is a compiler barrier; it codegens to a no-op.
3716     return DAG.getNode(ISD::MEMBARRIER, dl, MVT::Other, Op.getOperand(0));
3717 
3718   return Op;
3719 }
3720 
LowerOperation(SDValue Op,SelectionDAG & DAG) const3721 SDValue RISCVTargetLowering::LowerOperation(SDValue Op,
3722                                             SelectionDAG &DAG) const {
3723   switch (Op.getOpcode()) {
3724   default:
3725     report_fatal_error("unimplemented operand");
3726   case ISD::ATOMIC_FENCE:
3727     return LowerATOMIC_FENCE(Op, DAG);
3728   case ISD::GlobalAddress:
3729     return lowerGlobalAddress(Op, DAG);
3730   case ISD::BlockAddress:
3731     return lowerBlockAddress(Op, DAG);
3732   case ISD::ConstantPool:
3733     return lowerConstantPool(Op, DAG);
3734   case ISD::JumpTable:
3735     return lowerJumpTable(Op, DAG);
3736   case ISD::GlobalTLSAddress:
3737     return lowerGlobalTLSAddress(Op, DAG);
3738   case ISD::Constant:
3739     return lowerConstant(Op, DAG, Subtarget);
3740   case ISD::SELECT:
3741     return lowerSELECT(Op, DAG);
3742   case ISD::BRCOND:
3743     return lowerBRCOND(Op, DAG);
3744   case ISD::VASTART:
3745     return lowerVASTART(Op, DAG);
3746   case ISD::FRAMEADDR:
3747     return lowerFRAMEADDR(Op, DAG);
3748   case ISD::RETURNADDR:
3749     return lowerRETURNADDR(Op, DAG);
3750   case ISD::SHL_PARTS:
3751     return lowerShiftLeftParts(Op, DAG);
3752   case ISD::SRA_PARTS:
3753     return lowerShiftRightParts(Op, DAG, true);
3754   case ISD::SRL_PARTS:
3755     return lowerShiftRightParts(Op, DAG, false);
3756   case ISD::BITCAST: {
3757     SDLoc DL(Op);
3758     EVT VT = Op.getValueType();
3759     SDValue Op0 = Op.getOperand(0);
3760     EVT Op0VT = Op0.getValueType();
3761     MVT XLenVT = Subtarget.getXLenVT();
3762     if (VT == MVT::f16 && Op0VT == MVT::i16 &&
3763         Subtarget.hasStdExtZfhOrZfhmin()) {
3764       SDValue NewOp0 = DAG.getNode(ISD::ANY_EXTEND, DL, XLenVT, Op0);
3765       SDValue FPConv = DAG.getNode(RISCVISD::FMV_H_X, DL, MVT::f16, NewOp0);
3766       return FPConv;
3767     }
3768     if (VT == MVT::f32 && Op0VT == MVT::i32 && Subtarget.is64Bit() &&
3769         Subtarget.hasStdExtF()) {
3770       SDValue NewOp0 = DAG.getNode(ISD::ANY_EXTEND, DL, MVT::i64, Op0);
3771       SDValue FPConv =
3772           DAG.getNode(RISCVISD::FMV_W_X_RV64, DL, MVT::f32, NewOp0);
3773       return FPConv;
3774     }
3775 
3776     // Consider other scalar<->scalar casts as legal if the types are legal.
3777     // Otherwise expand them.
3778     if (!VT.isVector() && !Op0VT.isVector()) {
3779       if (isTypeLegal(VT) && isTypeLegal(Op0VT))
3780         return Op;
3781       return SDValue();
3782     }
3783 
3784     assert(!VT.isScalableVector() && !Op0VT.isScalableVector() &&
3785            "Unexpected types");
3786 
3787     if (VT.isFixedLengthVector()) {
3788       // We can handle fixed length vector bitcasts with a simple replacement
3789       // in isel.
3790       if (Op0VT.isFixedLengthVector())
3791         return Op;
3792       // When bitcasting from scalar to fixed-length vector, insert the scalar
3793       // into a one-element vector of the result type, and perform a vector
3794       // bitcast.
3795       if (!Op0VT.isVector()) {
3796         EVT BVT = EVT::getVectorVT(*DAG.getContext(), Op0VT, 1);
3797         if (!isTypeLegal(BVT))
3798           return SDValue();
3799         return DAG.getBitcast(VT, DAG.getNode(ISD::INSERT_VECTOR_ELT, DL, BVT,
3800                                               DAG.getUNDEF(BVT), Op0,
3801                                               DAG.getConstant(0, DL, XLenVT)));
3802       }
3803       return SDValue();
3804     }
3805     // Custom-legalize bitcasts from fixed-length vector types to scalar types
3806     // thus: bitcast the vector to a one-element vector type whose element type
3807     // is the same as the result type, and extract the first element.
3808     if (!VT.isVector() && Op0VT.isFixedLengthVector()) {
3809       EVT BVT = EVT::getVectorVT(*DAG.getContext(), VT, 1);
3810       if (!isTypeLegal(BVT))
3811         return SDValue();
3812       SDValue BVec = DAG.getBitcast(BVT, Op0);
3813       return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, VT, BVec,
3814                          DAG.getConstant(0, DL, XLenVT));
3815     }
3816     return SDValue();
3817   }
3818   case ISD::INTRINSIC_WO_CHAIN:
3819     return LowerINTRINSIC_WO_CHAIN(Op, DAG);
3820   case ISD::INTRINSIC_W_CHAIN:
3821     return LowerINTRINSIC_W_CHAIN(Op, DAG);
3822   case ISD::INTRINSIC_VOID:
3823     return LowerINTRINSIC_VOID(Op, DAG);
3824   case ISD::BITREVERSE: {
3825     MVT VT = Op.getSimpleValueType();
3826     SDLoc DL(Op);
3827     assert(Subtarget.hasStdExtZbkb() && "Unexpected custom legalization");
3828     assert(Op.getOpcode() == ISD::BITREVERSE && "Unexpected opcode");
3829     // Expand bitreverse to a bswap(rev8) followed by brev8.
3830     SDValue BSwap = DAG.getNode(ISD::BSWAP, DL, VT, Op.getOperand(0));
3831     return DAG.getNode(RISCVISD::BREV8, DL, VT, BSwap);
3832   }
3833   case ISD::TRUNCATE:
3834     // Only custom-lower vector truncates
3835     if (!Op.getSimpleValueType().isVector())
3836       return Op;
3837     return lowerVectorTruncLike(Op, DAG);
3838   case ISD::ANY_EXTEND:
3839   case ISD::ZERO_EXTEND:
3840     if (Op.getOperand(0).getValueType().isVector() &&
3841         Op.getOperand(0).getValueType().getVectorElementType() == MVT::i1)
3842       return lowerVectorMaskExt(Op, DAG, /*ExtVal*/ 1);
3843     return lowerFixedLengthVectorExtendToRVV(Op, DAG, RISCVISD::VZEXT_VL);
3844   case ISD::SIGN_EXTEND:
3845     if (Op.getOperand(0).getValueType().isVector() &&
3846         Op.getOperand(0).getValueType().getVectorElementType() == MVT::i1)
3847       return lowerVectorMaskExt(Op, DAG, /*ExtVal*/ -1);
3848     return lowerFixedLengthVectorExtendToRVV(Op, DAG, RISCVISD::VSEXT_VL);
3849   case ISD::SPLAT_VECTOR_PARTS:
3850     return lowerSPLAT_VECTOR_PARTS(Op, DAG);
3851   case ISD::INSERT_VECTOR_ELT:
3852     return lowerINSERT_VECTOR_ELT(Op, DAG);
3853   case ISD::EXTRACT_VECTOR_ELT:
3854     return lowerEXTRACT_VECTOR_ELT(Op, DAG);
3855   case ISD::VSCALE: {
3856     MVT VT = Op.getSimpleValueType();
3857     SDLoc DL(Op);
3858     SDValue VLENB = DAG.getNode(RISCVISD::READ_VLENB, DL, VT);
3859     // We define our scalable vector types for lmul=1 to use a 64 bit known
3860     // minimum size. e.g. <vscale x 2 x i32>. VLENB is in bytes so we calculate
3861     // vscale as VLENB / 8.
3862     static_assert(RISCV::RVVBitsPerBlock == 64, "Unexpected bits per block!");
3863     if (Subtarget.getRealMinVLen() < RISCV::RVVBitsPerBlock)
3864       report_fatal_error("Support for VLEN==32 is incomplete.");
3865     // We assume VLENB is a multiple of 8. We manually choose the best shift
3866     // here because SimplifyDemandedBits isn't always able to simplify it.
3867     uint64_t Val = Op.getConstantOperandVal(0);
3868     if (isPowerOf2_64(Val)) {
3869       uint64_t Log2 = Log2_64(Val);
3870       if (Log2 < 3)
3871         return DAG.getNode(ISD::SRL, DL, VT, VLENB,
3872                            DAG.getConstant(3 - Log2, DL, VT));
3873       if (Log2 > 3)
3874         return DAG.getNode(ISD::SHL, DL, VT, VLENB,
3875                            DAG.getConstant(Log2 - 3, DL, VT));
3876       return VLENB;
3877     }
3878     // If the multiplier is a multiple of 8, scale it down to avoid needing
3879     // to shift the VLENB value.
3880     if ((Val % 8) == 0)
3881       return DAG.getNode(ISD::MUL, DL, VT, VLENB,
3882                          DAG.getConstant(Val / 8, DL, VT));
3883 
3884     SDValue VScale = DAG.getNode(ISD::SRL, DL, VT, VLENB,
3885                                  DAG.getConstant(3, DL, VT));
3886     return DAG.getNode(ISD::MUL, DL, VT, VScale, Op.getOperand(0));
3887   }
3888   case ISD::FPOWI: {
3889     // Custom promote f16 powi with illegal i32 integer type on RV64. Once
3890     // promoted this will be legalized into a libcall by LegalizeIntegerTypes.
3891     if (Op.getValueType() == MVT::f16 && Subtarget.is64Bit() &&
3892         Op.getOperand(1).getValueType() == MVT::i32) {
3893       SDLoc DL(Op);
3894       SDValue Op0 = DAG.getNode(ISD::FP_EXTEND, DL, MVT::f32, Op.getOperand(0));
3895       SDValue Powi =
3896           DAG.getNode(ISD::FPOWI, DL, MVT::f32, Op0, Op.getOperand(1));
3897       return DAG.getNode(ISD::FP_ROUND, DL, MVT::f16, Powi,
3898                          DAG.getIntPtrConstant(0, DL, /*isTarget=*/true));
3899     }
3900     return SDValue();
3901   }
3902   case ISD::FP_EXTEND:
3903   case ISD::FP_ROUND:
3904     if (!Op.getValueType().isVector())
3905       return Op;
3906     return lowerVectorFPExtendOrRoundLike(Op, DAG);
3907   case ISD::FP_TO_SINT:
3908   case ISD::FP_TO_UINT:
3909   case ISD::SINT_TO_FP:
3910   case ISD::UINT_TO_FP: {
3911     // RVV can only do fp<->int conversions to types half/double the size as
3912     // the source. We custom-lower any conversions that do two hops into
3913     // sequences.
3914     MVT VT = Op.getSimpleValueType();
3915     if (!VT.isVector())
3916       return Op;
3917     SDLoc DL(Op);
3918     SDValue Src = Op.getOperand(0);
3919     MVT EltVT = VT.getVectorElementType();
3920     MVT SrcVT = Src.getSimpleValueType();
3921     MVT SrcEltVT = SrcVT.getVectorElementType();
3922     unsigned EltSize = EltVT.getSizeInBits();
3923     unsigned SrcEltSize = SrcEltVT.getSizeInBits();
3924     assert(isPowerOf2_32(EltSize) && isPowerOf2_32(SrcEltSize) &&
3925            "Unexpected vector element types");
3926 
3927     bool IsInt2FP = SrcEltVT.isInteger();
3928     // Widening conversions
3929     if (EltSize > (2 * SrcEltSize)) {
3930       if (IsInt2FP) {
3931         // Do a regular integer sign/zero extension then convert to float.
3932         MVT IVecVT = MVT::getVectorVT(MVT::getIntegerVT(EltSize / 2),
3933                                       VT.getVectorElementCount());
3934         unsigned ExtOpcode = Op.getOpcode() == ISD::UINT_TO_FP
3935                                  ? ISD::ZERO_EXTEND
3936                                  : ISD::SIGN_EXTEND;
3937         SDValue Ext = DAG.getNode(ExtOpcode, DL, IVecVT, Src);
3938         return DAG.getNode(Op.getOpcode(), DL, VT, Ext);
3939       }
3940       // FP2Int
3941       assert(SrcEltVT == MVT::f16 && "Unexpected FP_TO_[US]INT lowering");
3942       // Do one doubling fp_extend then complete the operation by converting
3943       // to int.
3944       MVT InterimFVT = MVT::getVectorVT(MVT::f32, VT.getVectorElementCount());
3945       SDValue FExt = DAG.getFPExtendOrRound(Src, DL, InterimFVT);
3946       return DAG.getNode(Op.getOpcode(), DL, VT, FExt);
3947     }
3948 
3949     // Narrowing conversions
3950     if (SrcEltSize > (2 * EltSize)) {
3951       if (IsInt2FP) {
3952         // One narrowing int_to_fp, then an fp_round.
3953         assert(EltVT == MVT::f16 && "Unexpected [US]_TO_FP lowering");
3954         MVT InterimFVT = MVT::getVectorVT(MVT::f32, VT.getVectorElementCount());
3955         SDValue Int2FP = DAG.getNode(Op.getOpcode(), DL, InterimFVT, Src);
3956         return DAG.getFPExtendOrRound(Int2FP, DL, VT);
3957       }
3958       // FP2Int
3959       // One narrowing fp_to_int, then truncate the integer. If the float isn't
3960       // representable by the integer, the result is poison.
3961       MVT IVecVT = MVT::getVectorVT(MVT::getIntegerVT(SrcEltSize / 2),
3962                                     VT.getVectorElementCount());
3963       SDValue FP2Int = DAG.getNode(Op.getOpcode(), DL, IVecVT, Src);
3964       return DAG.getNode(ISD::TRUNCATE, DL, VT, FP2Int);
3965     }
3966 
3967     // Scalable vectors can exit here. Patterns will handle equally-sized
3968     // conversions halving/doubling ones.
3969     if (!VT.isFixedLengthVector())
3970       return Op;
3971 
3972     // For fixed-length vectors we lower to a custom "VL" node.
3973     unsigned RVVOpc = 0;
3974     switch (Op.getOpcode()) {
3975     default:
3976       llvm_unreachable("Impossible opcode");
3977     case ISD::FP_TO_SINT:
3978       RVVOpc = RISCVISD::VFCVT_RTZ_X_F_VL;
3979       break;
3980     case ISD::FP_TO_UINT:
3981       RVVOpc = RISCVISD::VFCVT_RTZ_XU_F_VL;
3982       break;
3983     case ISD::SINT_TO_FP:
3984       RVVOpc = RISCVISD::SINT_TO_FP_VL;
3985       break;
3986     case ISD::UINT_TO_FP:
3987       RVVOpc = RISCVISD::UINT_TO_FP_VL;
3988       break;
3989     }
3990 
3991     MVT ContainerVT = getContainerForFixedLengthVector(VT);
3992     MVT SrcContainerVT = getContainerForFixedLengthVector(SrcVT);
3993     assert(ContainerVT.getVectorElementCount() == SrcContainerVT.getVectorElementCount() &&
3994            "Expected same element count");
3995 
3996     auto [Mask, VL] = getDefaultVLOps(VT, ContainerVT, DL, DAG, Subtarget);
3997 
3998     Src = convertToScalableVector(SrcContainerVT, Src, DAG, Subtarget);
3999     Src = DAG.getNode(RVVOpc, DL, ContainerVT, Src, Mask, VL);
4000     return convertFromScalableVector(VT, Src, DAG, Subtarget);
4001   }
4002   case ISD::FP_TO_SINT_SAT:
4003   case ISD::FP_TO_UINT_SAT:
4004     return lowerFP_TO_INT_SAT(Op, DAG, Subtarget);
4005   case ISD::FTRUNC:
4006   case ISD::FCEIL:
4007   case ISD::FFLOOR:
4008   case ISD::FRINT:
4009   case ISD::FROUND:
4010   case ISD::FROUNDEVEN:
4011     return lowerFTRUNC_FCEIL_FFLOOR_FROUND(Op, DAG, Subtarget);
4012   case ISD::VECREDUCE_ADD:
4013   case ISD::VECREDUCE_UMAX:
4014   case ISD::VECREDUCE_SMAX:
4015   case ISD::VECREDUCE_UMIN:
4016   case ISD::VECREDUCE_SMIN:
4017     return lowerVECREDUCE(Op, DAG);
4018   case ISD::VECREDUCE_AND:
4019   case ISD::VECREDUCE_OR:
4020   case ISD::VECREDUCE_XOR:
4021     if (Op.getOperand(0).getValueType().getVectorElementType() == MVT::i1)
4022       return lowerVectorMaskVecReduction(Op, DAG, /*IsVP*/ false);
4023     return lowerVECREDUCE(Op, DAG);
4024   case ISD::VECREDUCE_FADD:
4025   case ISD::VECREDUCE_SEQ_FADD:
4026   case ISD::VECREDUCE_FMIN:
4027   case ISD::VECREDUCE_FMAX:
4028     return lowerFPVECREDUCE(Op, DAG);
4029   case ISD::VP_REDUCE_ADD:
4030   case ISD::VP_REDUCE_UMAX:
4031   case ISD::VP_REDUCE_SMAX:
4032   case ISD::VP_REDUCE_UMIN:
4033   case ISD::VP_REDUCE_SMIN:
4034   case ISD::VP_REDUCE_FADD:
4035   case ISD::VP_REDUCE_SEQ_FADD:
4036   case ISD::VP_REDUCE_FMIN:
4037   case ISD::VP_REDUCE_FMAX:
4038     return lowerVPREDUCE(Op, DAG);
4039   case ISD::VP_REDUCE_AND:
4040   case ISD::VP_REDUCE_OR:
4041   case ISD::VP_REDUCE_XOR:
4042     if (Op.getOperand(1).getValueType().getVectorElementType() == MVT::i1)
4043       return lowerVectorMaskVecReduction(Op, DAG, /*IsVP*/ true);
4044     return lowerVPREDUCE(Op, DAG);
4045   case ISD::INSERT_SUBVECTOR:
4046     return lowerINSERT_SUBVECTOR(Op, DAG);
4047   case ISD::EXTRACT_SUBVECTOR:
4048     return lowerEXTRACT_SUBVECTOR(Op, DAG);
4049   case ISD::STEP_VECTOR:
4050     return lowerSTEP_VECTOR(Op, DAG);
4051   case ISD::VECTOR_REVERSE:
4052     return lowerVECTOR_REVERSE(Op, DAG);
4053   case ISD::VECTOR_SPLICE:
4054     return lowerVECTOR_SPLICE(Op, DAG);
4055   case ISD::BUILD_VECTOR:
4056     return lowerBUILD_VECTOR(Op, DAG, Subtarget);
4057   case ISD::SPLAT_VECTOR:
4058     if (Op.getValueType().getVectorElementType() == MVT::i1)
4059       return lowerVectorMaskSplat(Op, DAG);
4060     return SDValue();
4061   case ISD::VECTOR_SHUFFLE:
4062     return lowerVECTOR_SHUFFLE(Op, DAG, Subtarget);
4063   case ISD::CONCAT_VECTORS: {
4064     // Split CONCAT_VECTORS into a series of INSERT_SUBVECTOR nodes. This is
4065     // better than going through the stack, as the default expansion does.
4066     SDLoc DL(Op);
4067     MVT VT = Op.getSimpleValueType();
4068     unsigned NumOpElts =
4069         Op.getOperand(0).getSimpleValueType().getVectorMinNumElements();
4070     SDValue Vec = DAG.getUNDEF(VT);
4071     for (const auto &OpIdx : enumerate(Op->ops())) {
4072       SDValue SubVec = OpIdx.value();
4073       // Don't insert undef subvectors.
4074       if (SubVec.isUndef())
4075         continue;
4076       Vec = DAG.getNode(ISD::INSERT_SUBVECTOR, DL, VT, Vec, SubVec,
4077                         DAG.getIntPtrConstant(OpIdx.index() * NumOpElts, DL));
4078     }
4079     return Vec;
4080   }
4081   case ISD::LOAD:
4082     if (auto V = expandUnalignedRVVLoad(Op, DAG))
4083       return V;
4084     if (Op.getValueType().isFixedLengthVector())
4085       return lowerFixedLengthVectorLoadToRVV(Op, DAG);
4086     return Op;
4087   case ISD::STORE:
4088     if (auto V = expandUnalignedRVVStore(Op, DAG))
4089       return V;
4090     if (Op.getOperand(1).getValueType().isFixedLengthVector())
4091       return lowerFixedLengthVectorStoreToRVV(Op, DAG);
4092     return Op;
4093   case ISD::MLOAD:
4094   case ISD::VP_LOAD:
4095     return lowerMaskedLoad(Op, DAG);
4096   case ISD::MSTORE:
4097   case ISD::VP_STORE:
4098     return lowerMaskedStore(Op, DAG);
4099   case ISD::SELECT_CC: {
4100     // This occurs because we custom legalize SETGT and SETUGT for setcc. That
4101     // causes LegalizeDAG to think we need to custom legalize select_cc. Expand
4102     // into separate SETCC+SELECT_CC just like LegalizeDAG.
4103     SDValue Tmp1 = Op.getOperand(0);
4104     SDValue Tmp2 = Op.getOperand(1);
4105     SDValue True = Op.getOperand(2);
4106     SDValue False = Op.getOperand(3);
4107     EVT VT = Op.getValueType();
4108     SDValue CC = Op.getOperand(4);
4109     EVT CmpVT = Tmp1.getValueType();
4110     EVT CCVT =
4111         getSetCCResultType(DAG.getDataLayout(), *DAG.getContext(), CmpVT);
4112     SDLoc DL(Op);
4113     SDValue Cond =
4114         DAG.getNode(ISD::SETCC, DL, CCVT, Tmp1, Tmp2, CC, Op->getFlags());
4115     return DAG.getSelect(DL, VT, Cond, True, False);
4116   }
4117   case ISD::SETCC: {
4118     MVT OpVT = Op.getOperand(0).getSimpleValueType();
4119     if (OpVT.isScalarInteger()) {
4120       MVT VT = Op.getSimpleValueType();
4121       SDValue LHS = Op.getOperand(0);
4122       SDValue RHS = Op.getOperand(1);
4123       ISD::CondCode CCVal = cast<CondCodeSDNode>(Op.getOperand(2))->get();
4124       assert((CCVal == ISD::SETGT || CCVal == ISD::SETUGT) &&
4125              "Unexpected CondCode");
4126 
4127       SDLoc DL(Op);
4128 
4129       // If the RHS is a constant in the range [-2049, 0) or (0, 2046], we can
4130       // convert this to the equivalent of (set(u)ge X, C+1) by using
4131       // (xori (slti(u) X, C+1), 1). This avoids materializing a small constant
4132       // in a register.
4133       if (isa<ConstantSDNode>(RHS)) {
4134         int64_t Imm = cast<ConstantSDNode>(RHS)->getSExtValue();
4135         if (Imm != 0 && isInt<12>((uint64_t)Imm + 1)) {
4136           // X > -1 should have been replaced with false.
4137           assert((CCVal != ISD::SETUGT || Imm != -1) &&
4138                  "Missing canonicalization");
4139           // Using getSetCCSwappedOperands will convert SET(U)GT->SET(U)LT.
4140           CCVal = ISD::getSetCCSwappedOperands(CCVal);
4141           SDValue SetCC = DAG.getSetCC(
4142               DL, VT, LHS, DAG.getConstant(Imm + 1, DL, OpVT), CCVal);
4143           return DAG.getLogicalNOT(DL, SetCC, VT);
4144         }
4145       }
4146 
4147       // Not a constant we could handle, swap the operands and condition code to
4148       // SETLT/SETULT.
4149       CCVal = ISD::getSetCCSwappedOperands(CCVal);
4150       return DAG.getSetCC(DL, VT, RHS, LHS, CCVal);
4151     }
4152 
4153     return lowerFixedLengthVectorSetccToRVV(Op, DAG);
4154   }
4155   case ISD::ADD:
4156     return lowerToScalableOp(Op, DAG, RISCVISD::ADD_VL, /*HasMergeOp*/ true);
4157   case ISD::SUB:
4158     return lowerToScalableOp(Op, DAG, RISCVISD::SUB_VL, /*HasMergeOp*/ true);
4159   case ISD::MUL:
4160     return lowerToScalableOp(Op, DAG, RISCVISD::MUL_VL, /*HasMergeOp*/ true);
4161   case ISD::MULHS:
4162     return lowerToScalableOp(Op, DAG, RISCVISD::MULHS_VL, /*HasMergeOp*/ true);
4163   case ISD::MULHU:
4164     return lowerToScalableOp(Op, DAG, RISCVISD::MULHU_VL, /*HasMergeOp*/ true);
4165   case ISD::AND:
4166     return lowerFixedLengthVectorLogicOpToRVV(Op, DAG, RISCVISD::VMAND_VL,
4167                                               RISCVISD::AND_VL);
4168   case ISD::OR:
4169     return lowerFixedLengthVectorLogicOpToRVV(Op, DAG, RISCVISD::VMOR_VL,
4170                                               RISCVISD::OR_VL);
4171   case ISD::XOR:
4172     return lowerFixedLengthVectorLogicOpToRVV(Op, DAG, RISCVISD::VMXOR_VL,
4173                                               RISCVISD::XOR_VL);
4174   case ISD::SDIV:
4175     return lowerToScalableOp(Op, DAG, RISCVISD::SDIV_VL, /*HasMergeOp*/ true);
4176   case ISD::SREM:
4177     return lowerToScalableOp(Op, DAG, RISCVISD::SREM_VL, /*HasMergeOp*/ true);
4178   case ISD::UDIV:
4179     return lowerToScalableOp(Op, DAG, RISCVISD::UDIV_VL, /*HasMergeOp*/ true);
4180   case ISD::UREM:
4181     return lowerToScalableOp(Op, DAG, RISCVISD::UREM_VL, /*HasMergeOp*/ true);
4182   case ISD::SHL:
4183   case ISD::SRA:
4184   case ISD::SRL:
4185     if (Op.getSimpleValueType().isFixedLengthVector())
4186       return lowerFixedLengthVectorShiftToRVV(Op, DAG);
4187     // This can be called for an i32 shift amount that needs to be promoted.
4188     assert(Op.getOperand(1).getValueType() == MVT::i32 && Subtarget.is64Bit() &&
4189            "Unexpected custom legalisation");
4190     return SDValue();
4191   case ISD::SADDSAT:
4192     return lowerToScalableOp(Op, DAG, RISCVISD::SADDSAT_VL,
4193                              /*HasMergeOp*/ true);
4194   case ISD::UADDSAT:
4195     return lowerToScalableOp(Op, DAG, RISCVISD::UADDSAT_VL,
4196                              /*HasMergeOp*/ true);
4197   case ISD::SSUBSAT:
4198     return lowerToScalableOp(Op, DAG, RISCVISD::SSUBSAT_VL,
4199                              /*HasMergeOp*/ true);
4200   case ISD::USUBSAT:
4201     return lowerToScalableOp(Op, DAG, RISCVISD::USUBSAT_VL,
4202                              /*HasMergeOp*/ true);
4203   case ISD::FADD:
4204     return lowerToScalableOp(Op, DAG, RISCVISD::FADD_VL, /*HasMergeOp*/ true);
4205   case ISD::FSUB:
4206     return lowerToScalableOp(Op, DAG, RISCVISD::FSUB_VL, /*HasMergeOp*/ true);
4207   case ISD::FMUL:
4208     return lowerToScalableOp(Op, DAG, RISCVISD::FMUL_VL, /*HasMergeOp*/ true);
4209   case ISD::FDIV:
4210     return lowerToScalableOp(Op, DAG, RISCVISD::FDIV_VL, /*HasMergeOp*/ true);
4211   case ISD::FNEG:
4212     return lowerToScalableOp(Op, DAG, RISCVISD::FNEG_VL);
4213   case ISD::FABS:
4214     return lowerToScalableOp(Op, DAG, RISCVISD::FABS_VL);
4215   case ISD::FSQRT:
4216     return lowerToScalableOp(Op, DAG, RISCVISD::FSQRT_VL);
4217   case ISD::FMA:
4218     return lowerToScalableOp(Op, DAG, RISCVISD::VFMADD_VL);
4219   case ISD::SMIN:
4220     return lowerToScalableOp(Op, DAG, RISCVISD::SMIN_VL, /*HasMergeOp*/ true);
4221   case ISD::SMAX:
4222     return lowerToScalableOp(Op, DAG, RISCVISD::SMAX_VL, /*HasMergeOp*/ true);
4223   case ISD::UMIN:
4224     return lowerToScalableOp(Op, DAG, RISCVISD::UMIN_VL, /*HasMergeOp*/ true);
4225   case ISD::UMAX:
4226     return lowerToScalableOp(Op, DAG, RISCVISD::UMAX_VL, /*HasMergeOp*/ true);
4227   case ISD::FMINNUM:
4228     return lowerToScalableOp(Op, DAG, RISCVISD::FMINNUM_VL,
4229                              /*HasMergeOp*/ true);
4230   case ISD::FMAXNUM:
4231     return lowerToScalableOp(Op, DAG, RISCVISD::FMAXNUM_VL,
4232                              /*HasMergeOp*/ true);
4233   case ISD::ABS:
4234   case ISD::VP_ABS:
4235     return lowerABS(Op, DAG);
4236   case ISD::CTLZ:
4237   case ISD::CTLZ_ZERO_UNDEF:
4238   case ISD::CTTZ_ZERO_UNDEF:
4239     return lowerCTLZ_CTTZ_ZERO_UNDEF(Op, DAG);
4240   case ISD::VSELECT:
4241     return lowerFixedLengthVectorSelectToRVV(Op, DAG);
4242   case ISD::FCOPYSIGN:
4243     return lowerFixedLengthVectorFCOPYSIGNToRVV(Op, DAG);
4244   case ISD::MGATHER:
4245   case ISD::VP_GATHER:
4246     return lowerMaskedGather(Op, DAG);
4247   case ISD::MSCATTER:
4248   case ISD::VP_SCATTER:
4249     return lowerMaskedScatter(Op, DAG);
4250   case ISD::GET_ROUNDING:
4251     return lowerGET_ROUNDING(Op, DAG);
4252   case ISD::SET_ROUNDING:
4253     return lowerSET_ROUNDING(Op, DAG);
4254   case ISD::EH_DWARF_CFA:
4255     return lowerEH_DWARF_CFA(Op, DAG);
4256   case ISD::VP_SELECT:
4257     return lowerVPOp(Op, DAG, RISCVISD::VSELECT_VL);
4258   case ISD::VP_MERGE:
4259     return lowerVPOp(Op, DAG, RISCVISD::VP_MERGE_VL);
4260   case ISD::VP_ADD:
4261     return lowerVPOp(Op, DAG, RISCVISD::ADD_VL, /*HasMergeOp*/ true);
4262   case ISD::VP_SUB:
4263     return lowerVPOp(Op, DAG, RISCVISD::SUB_VL, /*HasMergeOp*/ true);
4264   case ISD::VP_MUL:
4265     return lowerVPOp(Op, DAG, RISCVISD::MUL_VL, /*HasMergeOp*/ true);
4266   case ISD::VP_SDIV:
4267     return lowerVPOp(Op, DAG, RISCVISD::SDIV_VL, /*HasMergeOp*/ true);
4268   case ISD::VP_UDIV:
4269     return lowerVPOp(Op, DAG, RISCVISD::UDIV_VL, /*HasMergeOp*/ true);
4270   case ISD::VP_SREM:
4271     return lowerVPOp(Op, DAG, RISCVISD::SREM_VL, /*HasMergeOp*/ true);
4272   case ISD::VP_UREM:
4273     return lowerVPOp(Op, DAG, RISCVISD::UREM_VL, /*HasMergeOp*/ true);
4274   case ISD::VP_AND:
4275     return lowerLogicVPOp(Op, DAG, RISCVISD::VMAND_VL, RISCVISD::AND_VL);
4276   case ISD::VP_OR:
4277     return lowerLogicVPOp(Op, DAG, RISCVISD::VMOR_VL, RISCVISD::OR_VL);
4278   case ISD::VP_XOR:
4279     return lowerLogicVPOp(Op, DAG, RISCVISD::VMXOR_VL, RISCVISD::XOR_VL);
4280   case ISD::VP_ASHR:
4281     return lowerVPOp(Op, DAG, RISCVISD::SRA_VL, /*HasMergeOp*/ true);
4282   case ISD::VP_LSHR:
4283     return lowerVPOp(Op, DAG, RISCVISD::SRL_VL, /*HasMergeOp*/ true);
4284   case ISD::VP_SHL:
4285     return lowerVPOp(Op, DAG, RISCVISD::SHL_VL, /*HasMergeOp*/ true);
4286   case ISD::VP_FADD:
4287     return lowerVPOp(Op, DAG, RISCVISD::FADD_VL, /*HasMergeOp*/ true);
4288   case ISD::VP_FSUB:
4289     return lowerVPOp(Op, DAG, RISCVISD::FSUB_VL, /*HasMergeOp*/ true);
4290   case ISD::VP_FMUL:
4291     return lowerVPOp(Op, DAG, RISCVISD::FMUL_VL, /*HasMergeOp*/ true);
4292   case ISD::VP_FDIV:
4293     return lowerVPOp(Op, DAG, RISCVISD::FDIV_VL, /*HasMergeOp*/ true);
4294   case ISD::VP_FNEG:
4295     return lowerVPOp(Op, DAG, RISCVISD::FNEG_VL);
4296   case ISD::VP_FABS:
4297     return lowerVPOp(Op, DAG, RISCVISD::FABS_VL);
4298   case ISD::VP_SQRT:
4299     return lowerVPOp(Op, DAG, RISCVISD::FSQRT_VL);
4300   case ISD::VP_FMA:
4301     return lowerVPOp(Op, DAG, RISCVISD::VFMADD_VL);
4302   case ISD::VP_FMINNUM:
4303     return lowerVPOp(Op, DAG, RISCVISD::FMINNUM_VL, /*HasMergeOp*/ true);
4304   case ISD::VP_FMAXNUM:
4305     return lowerVPOp(Op, DAG, RISCVISD::FMAXNUM_VL, /*HasMergeOp*/ true);
4306   case ISD::VP_FCOPYSIGN:
4307     return lowerVPOp(Op, DAG, RISCVISD::FCOPYSIGN_VL, /*HasMergeOp*/ true);
4308   case ISD::VP_SIGN_EXTEND:
4309   case ISD::VP_ZERO_EXTEND:
4310     if (Op.getOperand(0).getSimpleValueType().getVectorElementType() == MVT::i1)
4311       return lowerVPExtMaskOp(Op, DAG);
4312     return lowerVPOp(Op, DAG,
4313                      Op.getOpcode() == ISD::VP_SIGN_EXTEND
4314                          ? RISCVISD::VSEXT_VL
4315                          : RISCVISD::VZEXT_VL);
4316   case ISD::VP_TRUNCATE:
4317     return lowerVectorTruncLike(Op, DAG);
4318   case ISD::VP_FP_EXTEND:
4319   case ISD::VP_FP_ROUND:
4320     return lowerVectorFPExtendOrRoundLike(Op, DAG);
4321   case ISD::VP_FP_TO_SINT:
4322     return lowerVPFPIntConvOp(Op, DAG, RISCVISD::VFCVT_RTZ_X_F_VL);
4323   case ISD::VP_FP_TO_UINT:
4324     return lowerVPFPIntConvOp(Op, DAG, RISCVISD::VFCVT_RTZ_XU_F_VL);
4325   case ISD::VP_SINT_TO_FP:
4326     return lowerVPFPIntConvOp(Op, DAG, RISCVISD::SINT_TO_FP_VL);
4327   case ISD::VP_UINT_TO_FP:
4328     return lowerVPFPIntConvOp(Op, DAG, RISCVISD::UINT_TO_FP_VL);
4329   case ISD::VP_SETCC:
4330     if (Op.getOperand(0).getSimpleValueType().getVectorElementType() == MVT::i1)
4331       return lowerVPSetCCMaskOp(Op, DAG);
4332     return lowerVPOp(Op, DAG, RISCVISD::SETCC_VL, /*HasMergeOp*/ true);
4333   case ISD::VP_SMIN:
4334     return lowerVPOp(Op, DAG, RISCVISD::SMIN_VL, /*HasMergeOp*/ true);
4335   case ISD::VP_SMAX:
4336     return lowerVPOp(Op, DAG, RISCVISD::SMAX_VL, /*HasMergeOp*/ true);
4337   case ISD::VP_UMIN:
4338     return lowerVPOp(Op, DAG, RISCVISD::UMIN_VL, /*HasMergeOp*/ true);
4339   case ISD::VP_UMAX:
4340     return lowerVPOp(Op, DAG, RISCVISD::UMAX_VL, /*HasMergeOp*/ true);
4341   case ISD::EXPERIMENTAL_VP_STRIDED_LOAD:
4342     return lowerVPStridedLoad(Op, DAG);
4343   case ISD::EXPERIMENTAL_VP_STRIDED_STORE:
4344     return lowerVPStridedStore(Op, DAG);
4345   case ISD::VP_FCEIL:
4346   case ISD::VP_FFLOOR:
4347   case ISD::VP_FRINT:
4348   case ISD::VP_FNEARBYINT:
4349   case ISD::VP_FROUND:
4350   case ISD::VP_FROUNDEVEN:
4351   case ISD::VP_FROUNDTOZERO:
4352     return lowerVectorFTRUNC_FCEIL_FFLOOR_FROUND(Op, DAG, Subtarget);
4353   }
4354 }
4355 
getTargetNode(GlobalAddressSDNode * N,SDLoc DL,EVT Ty,SelectionDAG & DAG,unsigned Flags)4356 static SDValue getTargetNode(GlobalAddressSDNode *N, SDLoc DL, EVT Ty,
4357                              SelectionDAG &DAG, unsigned Flags) {
4358   return DAG.getTargetGlobalAddress(N->getGlobal(), DL, Ty, 0, Flags);
4359 }
4360 
getTargetNode(BlockAddressSDNode * N,SDLoc DL,EVT Ty,SelectionDAG & DAG,unsigned Flags)4361 static SDValue getTargetNode(BlockAddressSDNode *N, SDLoc DL, EVT Ty,
4362                              SelectionDAG &DAG, unsigned Flags) {
4363   return DAG.getTargetBlockAddress(N->getBlockAddress(), Ty, N->getOffset(),
4364                                    Flags);
4365 }
4366 
getTargetNode(ConstantPoolSDNode * N,SDLoc DL,EVT Ty,SelectionDAG & DAG,unsigned Flags)4367 static SDValue getTargetNode(ConstantPoolSDNode *N, SDLoc DL, EVT Ty,
4368                              SelectionDAG &DAG, unsigned Flags) {
4369   return DAG.getTargetConstantPool(N->getConstVal(), Ty, N->getAlign(),
4370                                    N->getOffset(), Flags);
4371 }
4372 
getTargetNode(JumpTableSDNode * N,SDLoc DL,EVT Ty,SelectionDAG & DAG,unsigned Flags)4373 static SDValue getTargetNode(JumpTableSDNode *N, SDLoc DL, EVT Ty,
4374                              SelectionDAG &DAG, unsigned Flags) {
4375   return DAG.getTargetJumpTable(N->getIndex(), Ty, Flags);
4376 }
4377 
4378 template <class NodeTy>
getAddr(NodeTy * N,SelectionDAG & DAG,bool IsLocal) const4379 SDValue RISCVTargetLowering::getAddr(NodeTy *N, SelectionDAG &DAG,
4380                                      bool IsLocal) const {
4381   SDLoc DL(N);
4382   EVT Ty = getPointerTy(DAG.getDataLayout());
4383 
4384   // When HWASAN is used and tagging of global variables is enabled
4385   // they should be accessed via the GOT, since the tagged address of a global
4386   // is incompatible with existing code models. This also applies to non-pic
4387   // mode.
4388   if (isPositionIndependent() || Subtarget.allowTaggedGlobals()) {
4389     SDValue Addr = getTargetNode(N, DL, Ty, DAG, 0);
4390     if (IsLocal && !Subtarget.allowTaggedGlobals())
4391       // Use PC-relative addressing to access the symbol. This generates the
4392       // pattern (PseudoLLA sym), which expands to (addi (auipc %pcrel_hi(sym))
4393       // %pcrel_lo(auipc)).
4394       return DAG.getNode(RISCVISD::LLA, DL, Ty, Addr);
4395 
4396     // Use PC-relative addressing to access the GOT for this symbol, then load
4397     // the address from the GOT. This generates the pattern (PseudoLA sym),
4398     // which expands to (ld (addi (auipc %got_pcrel_hi(sym)) %pcrel_lo(auipc))).
4399     MachineFunction &MF = DAG.getMachineFunction();
4400     MachineMemOperand *MemOp = MF.getMachineMemOperand(
4401         MachinePointerInfo::getGOT(MF),
4402         MachineMemOperand::MOLoad | MachineMemOperand::MODereferenceable |
4403             MachineMemOperand::MOInvariant,
4404         LLT(Ty.getSimpleVT()), Align(Ty.getFixedSizeInBits() / 8));
4405     SDValue Load =
4406         DAG.getMemIntrinsicNode(RISCVISD::LA, DL, DAG.getVTList(Ty, MVT::Other),
4407                                 {DAG.getEntryNode(), Addr}, Ty, MemOp);
4408     return Load;
4409   }
4410 
4411   switch (getTargetMachine().getCodeModel()) {
4412   default:
4413     report_fatal_error("Unsupported code model for lowering");
4414   case CodeModel::Small: {
4415     // Generate a sequence for accessing addresses within the first 2 GiB of
4416     // address space. This generates the pattern (addi (lui %hi(sym)) %lo(sym)).
4417     SDValue AddrHi = getTargetNode(N, DL, Ty, DAG, RISCVII::MO_HI);
4418     SDValue AddrLo = getTargetNode(N, DL, Ty, DAG, RISCVII::MO_LO);
4419     SDValue MNHi = DAG.getNode(RISCVISD::HI, DL, Ty, AddrHi);
4420     return DAG.getNode(RISCVISD::ADD_LO, DL, Ty, MNHi, AddrLo);
4421   }
4422   case CodeModel::Medium: {
4423     // Generate a sequence for accessing addresses within any 2GiB range within
4424     // the address space. This generates the pattern (PseudoLLA sym), which
4425     // expands to (addi (auipc %pcrel_hi(sym)) %pcrel_lo(auipc)).
4426     SDValue Addr = getTargetNode(N, DL, Ty, DAG, 0);
4427     return DAG.getNode(RISCVISD::LLA, DL, Ty, Addr);
4428   }
4429   }
4430 }
4431 
lowerGlobalAddress(SDValue Op,SelectionDAG & DAG) const4432 SDValue RISCVTargetLowering::lowerGlobalAddress(SDValue Op,
4433                                                 SelectionDAG &DAG) const {
4434   GlobalAddressSDNode *N = cast<GlobalAddressSDNode>(Op);
4435   assert(N->getOffset() == 0 && "unexpected offset in global node");
4436   return getAddr(N, DAG, N->getGlobal()->isDSOLocal());
4437 }
4438 
lowerBlockAddress(SDValue Op,SelectionDAG & DAG) const4439 SDValue RISCVTargetLowering::lowerBlockAddress(SDValue Op,
4440                                                SelectionDAG &DAG) const {
4441   BlockAddressSDNode *N = cast<BlockAddressSDNode>(Op);
4442 
4443   return getAddr(N, DAG);
4444 }
4445 
lowerConstantPool(SDValue Op,SelectionDAG & DAG) const4446 SDValue RISCVTargetLowering::lowerConstantPool(SDValue Op,
4447                                                SelectionDAG &DAG) const {
4448   ConstantPoolSDNode *N = cast<ConstantPoolSDNode>(Op);
4449 
4450   return getAddr(N, DAG);
4451 }
4452 
lowerJumpTable(SDValue Op,SelectionDAG & DAG) const4453 SDValue RISCVTargetLowering::lowerJumpTable(SDValue Op,
4454                                             SelectionDAG &DAG) const {
4455   JumpTableSDNode *N = cast<JumpTableSDNode>(Op);
4456 
4457   return getAddr(N, DAG);
4458 }
4459 
getStaticTLSAddr(GlobalAddressSDNode * N,SelectionDAG & DAG,bool UseGOT) const4460 SDValue RISCVTargetLowering::getStaticTLSAddr(GlobalAddressSDNode *N,
4461                                               SelectionDAG &DAG,
4462                                               bool UseGOT) const {
4463   SDLoc DL(N);
4464   EVT Ty = getPointerTy(DAG.getDataLayout());
4465   const GlobalValue *GV = N->getGlobal();
4466   MVT XLenVT = Subtarget.getXLenVT();
4467 
4468   if (UseGOT) {
4469     // Use PC-relative addressing to access the GOT for this TLS symbol, then
4470     // load the address from the GOT and add the thread pointer. This generates
4471     // the pattern (PseudoLA_TLS_IE sym), which expands to
4472     // (ld (auipc %tls_ie_pcrel_hi(sym)) %pcrel_lo(auipc)).
4473     SDValue Addr = DAG.getTargetGlobalAddress(GV, DL, Ty, 0, 0);
4474     MachineFunction &MF = DAG.getMachineFunction();
4475     MachineMemOperand *MemOp = MF.getMachineMemOperand(
4476         MachinePointerInfo::getGOT(MF),
4477         MachineMemOperand::MOLoad | MachineMemOperand::MODereferenceable |
4478             MachineMemOperand::MOInvariant,
4479         LLT(Ty.getSimpleVT()), Align(Ty.getFixedSizeInBits() / 8));
4480     SDValue Load = DAG.getMemIntrinsicNode(
4481         RISCVISD::LA_TLS_IE, DL, DAG.getVTList(Ty, MVT::Other),
4482         {DAG.getEntryNode(), Addr}, Ty, MemOp);
4483 
4484     // Add the thread pointer.
4485     SDValue TPReg = DAG.getRegister(RISCV::X4, XLenVT);
4486     return DAG.getNode(ISD::ADD, DL, Ty, Load, TPReg);
4487   }
4488 
4489   // Generate a sequence for accessing the address relative to the thread
4490   // pointer, with the appropriate adjustment for the thread pointer offset.
4491   // This generates the pattern
4492   // (add (add_tprel (lui %tprel_hi(sym)) tp %tprel_add(sym)) %tprel_lo(sym))
4493   SDValue AddrHi =
4494       DAG.getTargetGlobalAddress(GV, DL, Ty, 0, RISCVII::MO_TPREL_HI);
4495   SDValue AddrAdd =
4496       DAG.getTargetGlobalAddress(GV, DL, Ty, 0, RISCVII::MO_TPREL_ADD);
4497   SDValue AddrLo =
4498       DAG.getTargetGlobalAddress(GV, DL, Ty, 0, RISCVII::MO_TPREL_LO);
4499 
4500   SDValue MNHi = DAG.getNode(RISCVISD::HI, DL, Ty, AddrHi);
4501   SDValue TPReg = DAG.getRegister(RISCV::X4, XLenVT);
4502   SDValue MNAdd =
4503       DAG.getNode(RISCVISD::ADD_TPREL, DL, Ty, MNHi, TPReg, AddrAdd);
4504   return DAG.getNode(RISCVISD::ADD_LO, DL, Ty, MNAdd, AddrLo);
4505 }
4506 
getDynamicTLSAddr(GlobalAddressSDNode * N,SelectionDAG & DAG) const4507 SDValue RISCVTargetLowering::getDynamicTLSAddr(GlobalAddressSDNode *N,
4508                                                SelectionDAG &DAG) const {
4509   SDLoc DL(N);
4510   EVT Ty = getPointerTy(DAG.getDataLayout());
4511   IntegerType *CallTy = Type::getIntNTy(*DAG.getContext(), Ty.getSizeInBits());
4512   const GlobalValue *GV = N->getGlobal();
4513 
4514   // Use a PC-relative addressing mode to access the global dynamic GOT address.
4515   // This generates the pattern (PseudoLA_TLS_GD sym), which expands to
4516   // (addi (auipc %tls_gd_pcrel_hi(sym)) %pcrel_lo(auipc)).
4517   SDValue Addr = DAG.getTargetGlobalAddress(GV, DL, Ty, 0, 0);
4518   SDValue Load = DAG.getNode(RISCVISD::LA_TLS_GD, DL, Ty, Addr);
4519 
4520   // Prepare argument list to generate call.
4521   ArgListTy Args;
4522   ArgListEntry Entry;
4523   Entry.Node = Load;
4524   Entry.Ty = CallTy;
4525   Args.push_back(Entry);
4526 
4527   // Setup call to __tls_get_addr.
4528   TargetLowering::CallLoweringInfo CLI(DAG);
4529   CLI.setDebugLoc(DL)
4530       .setChain(DAG.getEntryNode())
4531       .setLibCallee(CallingConv::C, CallTy,
4532                     DAG.getExternalSymbol("__tls_get_addr", Ty),
4533                     std::move(Args));
4534 
4535   return LowerCallTo(CLI).first;
4536 }
4537 
lowerGlobalTLSAddress(SDValue Op,SelectionDAG & DAG) const4538 SDValue RISCVTargetLowering::lowerGlobalTLSAddress(SDValue Op,
4539                                                    SelectionDAG &DAG) const {
4540   const GlobalAddressSDNode *GA = cast<GlobalAddressSDNode>(Op);
4541   if (DAG.getTarget().useEmulatedTLS())
4542     return LowerToTLSEmulatedModel(GA, DAG);
4543 
4544   GlobalAddressSDNode *N = cast<GlobalAddressSDNode>(Op);
4545   assert(N->getOffset() == 0 && "unexpected offset in global node");
4546 
4547   if (DAG.getTarget().useEmulatedTLS())
4548     return LowerToTLSEmulatedModel(N, DAG);
4549 
4550   TLSModel::Model Model = getTargetMachine().getTLSModel(N->getGlobal());
4551 
4552   if (DAG.getMachineFunction().getFunction().getCallingConv() ==
4553       CallingConv::GHC)
4554     report_fatal_error("In GHC calling convention TLS is not supported");
4555 
4556   SDValue Addr;
4557   switch (Model) {
4558   case TLSModel::LocalExec:
4559     Addr = getStaticTLSAddr(N, DAG, /*UseGOT=*/false);
4560     break;
4561   case TLSModel::InitialExec:
4562     Addr = getStaticTLSAddr(N, DAG, /*UseGOT=*/true);
4563     break;
4564   case TLSModel::LocalDynamic:
4565   case TLSModel::GeneralDynamic:
4566     Addr = getDynamicTLSAddr(N, DAG);
4567     break;
4568   }
4569 
4570   return Addr;
4571 }
4572 
lowerSELECT(SDValue Op,SelectionDAG & DAG) const4573 SDValue RISCVTargetLowering::lowerSELECT(SDValue Op, SelectionDAG &DAG) const {
4574   SDValue CondV = Op.getOperand(0);
4575   SDValue TrueV = Op.getOperand(1);
4576   SDValue FalseV = Op.getOperand(2);
4577   SDLoc DL(Op);
4578   MVT VT = Op.getSimpleValueType();
4579   MVT XLenVT = Subtarget.getXLenVT();
4580 
4581   // Lower vector SELECTs to VSELECTs by splatting the condition.
4582   if (VT.isVector()) {
4583     MVT SplatCondVT = VT.changeVectorElementType(MVT::i1);
4584     SDValue CondSplat = DAG.getSplat(SplatCondVT, DL, CondV);
4585     return DAG.getNode(ISD::VSELECT, DL, VT, CondSplat, TrueV, FalseV);
4586   }
4587 
4588   if (!Subtarget.hasShortForwardBranchOpt()) {
4589     // (select c, -1, y) -> -c | y
4590     if (isAllOnesConstant(TrueV)) {
4591       SDValue Neg = DAG.getNegative(CondV, DL, VT);
4592       return DAG.getNode(ISD::OR, DL, VT, Neg, FalseV);
4593     }
4594     // (select c, y, -1) -> (c-1) | y
4595     if (isAllOnesConstant(FalseV)) {
4596       SDValue Neg = DAG.getNode(ISD::ADD, DL, VT, CondV,
4597                                 DAG.getAllOnesConstant(DL, VT));
4598       return DAG.getNode(ISD::OR, DL, VT, Neg, TrueV);
4599     }
4600 
4601     // (select c, 0, y) -> (c-1) & y
4602     if (isNullConstant(TrueV)) {
4603       SDValue Neg = DAG.getNode(ISD::ADD, DL, VT, CondV,
4604                                 DAG.getAllOnesConstant(DL, VT));
4605       return DAG.getNode(ISD::AND, DL, VT, Neg, FalseV);
4606     }
4607     // (select c, y, 0) -> -c & y
4608     if (isNullConstant(FalseV)) {
4609       SDValue Neg = DAG.getNegative(CondV, DL, VT);
4610       return DAG.getNode(ISD::AND, DL, VT, Neg, TrueV);
4611     }
4612   }
4613 
4614   // If the condition is not an integer SETCC which operates on XLenVT, we need
4615   // to emit a RISCVISD::SELECT_CC comparing the condition to zero. i.e.:
4616   // (select condv, truev, falsev)
4617   // -> (riscvisd::select_cc condv, zero, setne, truev, falsev)
4618   if (CondV.getOpcode() != ISD::SETCC ||
4619       CondV.getOperand(0).getSimpleValueType() != XLenVT) {
4620     SDValue Zero = DAG.getConstant(0, DL, XLenVT);
4621     SDValue SetNE = DAG.getCondCode(ISD::SETNE);
4622 
4623     SDValue Ops[] = {CondV, Zero, SetNE, TrueV, FalseV};
4624 
4625     return DAG.getNode(RISCVISD::SELECT_CC, DL, VT, Ops);
4626   }
4627 
4628   // If the CondV is the output of a SETCC node which operates on XLenVT inputs,
4629   // then merge the SETCC node into the lowered RISCVISD::SELECT_CC to take
4630   // advantage of the integer compare+branch instructions. i.e.:
4631   // (select (setcc lhs, rhs, cc), truev, falsev)
4632   // -> (riscvisd::select_cc lhs, rhs, cc, truev, falsev)
4633   SDValue LHS = CondV.getOperand(0);
4634   SDValue RHS = CondV.getOperand(1);
4635   ISD::CondCode CCVal = cast<CondCodeSDNode>(CondV.getOperand(2))->get();
4636 
4637   // Special case for a select of 2 constants that have a diffence of 1.
4638   // Normally this is done by DAGCombine, but if the select is introduced by
4639   // type legalization or op legalization, we miss it. Restricting to SETLT
4640   // case for now because that is what signed saturating add/sub need.
4641   // FIXME: We don't need the condition to be SETLT or even a SETCC,
4642   // but we would probably want to swap the true/false values if the condition
4643   // is SETGE/SETLE to avoid an XORI.
4644   if (isa<ConstantSDNode>(TrueV) && isa<ConstantSDNode>(FalseV) &&
4645       CCVal == ISD::SETLT) {
4646     const APInt &TrueVal = cast<ConstantSDNode>(TrueV)->getAPIntValue();
4647     const APInt &FalseVal = cast<ConstantSDNode>(FalseV)->getAPIntValue();
4648     if (TrueVal - 1 == FalseVal)
4649       return DAG.getNode(ISD::ADD, DL, VT, CondV, FalseV);
4650     if (TrueVal + 1 == FalseVal)
4651       return DAG.getNode(ISD::SUB, DL, VT, FalseV, CondV);
4652   }
4653 
4654   translateSetCCForBranch(DL, LHS, RHS, CCVal, DAG);
4655   // 1 < x ? x : 1 -> 0 < x ? x : 1
4656   if (isOneConstant(LHS) && (CCVal == ISD::SETLT || CCVal == ISD::SETULT) &&
4657       RHS == TrueV && LHS == FalseV) {
4658     LHS = DAG.getConstant(0, DL, VT);
4659     // 0 <u x is the same as x != 0.
4660     if (CCVal == ISD::SETULT) {
4661       std::swap(LHS, RHS);
4662       CCVal = ISD::SETNE;
4663     }
4664   }
4665 
4666   // x <s -1 ? x : -1 -> x <s 0 ? x : -1
4667   if (isAllOnesConstant(RHS) && CCVal == ISD::SETLT && LHS == TrueV &&
4668       RHS == FalseV) {
4669     RHS = DAG.getConstant(0, DL, VT);
4670   }
4671 
4672   SDValue TargetCC = DAG.getCondCode(CCVal);
4673 
4674   if (isa<ConstantSDNode>(TrueV) && !isa<ConstantSDNode>(FalseV)) {
4675     // (select (setcc lhs, rhs, CC), constant, falsev)
4676     // -> (select (setcc lhs, rhs, InverseCC), falsev, constant)
4677     std::swap(TrueV, FalseV);
4678     TargetCC = DAG.getCondCode(ISD::getSetCCInverse(CCVal, LHS.getValueType()));
4679   }
4680 
4681   SDValue Ops[] = {LHS, RHS, TargetCC, TrueV, FalseV};
4682   return DAG.getNode(RISCVISD::SELECT_CC, DL, VT, Ops);
4683 }
4684 
lowerBRCOND(SDValue Op,SelectionDAG & DAG) const4685 SDValue RISCVTargetLowering::lowerBRCOND(SDValue Op, SelectionDAG &DAG) const {
4686   SDValue CondV = Op.getOperand(1);
4687   SDLoc DL(Op);
4688   MVT XLenVT = Subtarget.getXLenVT();
4689 
4690   if (CondV.getOpcode() == ISD::SETCC &&
4691       CondV.getOperand(0).getValueType() == XLenVT) {
4692     SDValue LHS = CondV.getOperand(0);
4693     SDValue RHS = CondV.getOperand(1);
4694     ISD::CondCode CCVal = cast<CondCodeSDNode>(CondV.getOperand(2))->get();
4695 
4696     translateSetCCForBranch(DL, LHS, RHS, CCVal, DAG);
4697 
4698     SDValue TargetCC = DAG.getCondCode(CCVal);
4699     return DAG.getNode(RISCVISD::BR_CC, DL, Op.getValueType(), Op.getOperand(0),
4700                        LHS, RHS, TargetCC, Op.getOperand(2));
4701   }
4702 
4703   return DAG.getNode(RISCVISD::BR_CC, DL, Op.getValueType(), Op.getOperand(0),
4704                      CondV, DAG.getConstant(0, DL, XLenVT),
4705                      DAG.getCondCode(ISD::SETNE), Op.getOperand(2));
4706 }
4707 
lowerVASTART(SDValue Op,SelectionDAG & DAG) const4708 SDValue RISCVTargetLowering::lowerVASTART(SDValue Op, SelectionDAG &DAG) const {
4709   MachineFunction &MF = DAG.getMachineFunction();
4710   RISCVMachineFunctionInfo *FuncInfo = MF.getInfo<RISCVMachineFunctionInfo>();
4711 
4712   SDLoc DL(Op);
4713   SDValue FI = DAG.getFrameIndex(FuncInfo->getVarArgsFrameIndex(),
4714                                  getPointerTy(MF.getDataLayout()));
4715 
4716   // vastart just stores the address of the VarArgsFrameIndex slot into the
4717   // memory location argument.
4718   const Value *SV = cast<SrcValueSDNode>(Op.getOperand(2))->getValue();
4719   return DAG.getStore(Op.getOperand(0), DL, FI, Op.getOperand(1),
4720                       MachinePointerInfo(SV));
4721 }
4722 
lowerFRAMEADDR(SDValue Op,SelectionDAG & DAG) const4723 SDValue RISCVTargetLowering::lowerFRAMEADDR(SDValue Op,
4724                                             SelectionDAG &DAG) const {
4725   const RISCVRegisterInfo &RI = *Subtarget.getRegisterInfo();
4726   MachineFunction &MF = DAG.getMachineFunction();
4727   MachineFrameInfo &MFI = MF.getFrameInfo();
4728   MFI.setFrameAddressIsTaken(true);
4729   Register FrameReg = RI.getFrameRegister(MF);
4730   int XLenInBytes = Subtarget.getXLen() / 8;
4731 
4732   EVT VT = Op.getValueType();
4733   SDLoc DL(Op);
4734   SDValue FrameAddr = DAG.getCopyFromReg(DAG.getEntryNode(), DL, FrameReg, VT);
4735   unsigned Depth = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
4736   while (Depth--) {
4737     int Offset = -(XLenInBytes * 2);
4738     SDValue Ptr = DAG.getNode(ISD::ADD, DL, VT, FrameAddr,
4739                               DAG.getIntPtrConstant(Offset, DL));
4740     FrameAddr =
4741         DAG.getLoad(VT, DL, DAG.getEntryNode(), Ptr, MachinePointerInfo());
4742   }
4743   return FrameAddr;
4744 }
4745 
lowerRETURNADDR(SDValue Op,SelectionDAG & DAG) const4746 SDValue RISCVTargetLowering::lowerRETURNADDR(SDValue Op,
4747                                              SelectionDAG &DAG) const {
4748   const RISCVRegisterInfo &RI = *Subtarget.getRegisterInfo();
4749   MachineFunction &MF = DAG.getMachineFunction();
4750   MachineFrameInfo &MFI = MF.getFrameInfo();
4751   MFI.setReturnAddressIsTaken(true);
4752   MVT XLenVT = Subtarget.getXLenVT();
4753   int XLenInBytes = Subtarget.getXLen() / 8;
4754 
4755   if (verifyReturnAddressArgumentIsConstant(Op, DAG))
4756     return SDValue();
4757 
4758   EVT VT = Op.getValueType();
4759   SDLoc DL(Op);
4760   unsigned Depth = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
4761   if (Depth) {
4762     int Off = -XLenInBytes;
4763     SDValue FrameAddr = lowerFRAMEADDR(Op, DAG);
4764     SDValue Offset = DAG.getConstant(Off, DL, VT);
4765     return DAG.getLoad(VT, DL, DAG.getEntryNode(),
4766                        DAG.getNode(ISD::ADD, DL, VT, FrameAddr, Offset),
4767                        MachinePointerInfo());
4768   }
4769 
4770   // Return the value of the return address register, marking it an implicit
4771   // live-in.
4772   Register Reg = MF.addLiveIn(RI.getRARegister(), getRegClassFor(XLenVT));
4773   return DAG.getCopyFromReg(DAG.getEntryNode(), DL, Reg, XLenVT);
4774 }
4775 
lowerShiftLeftParts(SDValue Op,SelectionDAG & DAG) const4776 SDValue RISCVTargetLowering::lowerShiftLeftParts(SDValue Op,
4777                                                  SelectionDAG &DAG) const {
4778   SDLoc DL(Op);
4779   SDValue Lo = Op.getOperand(0);
4780   SDValue Hi = Op.getOperand(1);
4781   SDValue Shamt = Op.getOperand(2);
4782   EVT VT = Lo.getValueType();
4783 
4784   // if Shamt-XLEN < 0: // Shamt < XLEN
4785   //   Lo = Lo << Shamt
4786   //   Hi = (Hi << Shamt) | ((Lo >>u 1) >>u (XLEN-1 ^ Shamt))
4787   // else:
4788   //   Lo = 0
4789   //   Hi = Lo << (Shamt-XLEN)
4790 
4791   SDValue Zero = DAG.getConstant(0, DL, VT);
4792   SDValue One = DAG.getConstant(1, DL, VT);
4793   SDValue MinusXLen = DAG.getConstant(-(int)Subtarget.getXLen(), DL, VT);
4794   SDValue XLenMinus1 = DAG.getConstant(Subtarget.getXLen() - 1, DL, VT);
4795   SDValue ShamtMinusXLen = DAG.getNode(ISD::ADD, DL, VT, Shamt, MinusXLen);
4796   SDValue XLenMinus1Shamt = DAG.getNode(ISD::SUB, DL, VT, XLenMinus1, Shamt);
4797 
4798   SDValue LoTrue = DAG.getNode(ISD::SHL, DL, VT, Lo, Shamt);
4799   SDValue ShiftRight1Lo = DAG.getNode(ISD::SRL, DL, VT, Lo, One);
4800   SDValue ShiftRightLo =
4801       DAG.getNode(ISD::SRL, DL, VT, ShiftRight1Lo, XLenMinus1Shamt);
4802   SDValue ShiftLeftHi = DAG.getNode(ISD::SHL, DL, VT, Hi, Shamt);
4803   SDValue HiTrue = DAG.getNode(ISD::OR, DL, VT, ShiftLeftHi, ShiftRightLo);
4804   SDValue HiFalse = DAG.getNode(ISD::SHL, DL, VT, Lo, ShamtMinusXLen);
4805 
4806   SDValue CC = DAG.getSetCC(DL, VT, ShamtMinusXLen, Zero, ISD::SETLT);
4807 
4808   Lo = DAG.getNode(ISD::SELECT, DL, VT, CC, LoTrue, Zero);
4809   Hi = DAG.getNode(ISD::SELECT, DL, VT, CC, HiTrue, HiFalse);
4810 
4811   SDValue Parts[2] = {Lo, Hi};
4812   return DAG.getMergeValues(Parts, DL);
4813 }
4814 
lowerShiftRightParts(SDValue Op,SelectionDAG & DAG,bool IsSRA) const4815 SDValue RISCVTargetLowering::lowerShiftRightParts(SDValue Op, SelectionDAG &DAG,
4816                                                   bool IsSRA) const {
4817   SDLoc DL(Op);
4818   SDValue Lo = Op.getOperand(0);
4819   SDValue Hi = Op.getOperand(1);
4820   SDValue Shamt = Op.getOperand(2);
4821   EVT VT = Lo.getValueType();
4822 
4823   // SRA expansion:
4824   //   if Shamt-XLEN < 0: // Shamt < XLEN
4825   //     Lo = (Lo >>u Shamt) | ((Hi << 1) << (ShAmt ^ XLEN-1))
4826   //     Hi = Hi >>s Shamt
4827   //   else:
4828   //     Lo = Hi >>s (Shamt-XLEN);
4829   //     Hi = Hi >>s (XLEN-1)
4830   //
4831   // SRL expansion:
4832   //   if Shamt-XLEN < 0: // Shamt < XLEN
4833   //     Lo = (Lo >>u Shamt) | ((Hi << 1) << (ShAmt ^ XLEN-1))
4834   //     Hi = Hi >>u Shamt
4835   //   else:
4836   //     Lo = Hi >>u (Shamt-XLEN);
4837   //     Hi = 0;
4838 
4839   unsigned ShiftRightOp = IsSRA ? ISD::SRA : ISD::SRL;
4840 
4841   SDValue Zero = DAG.getConstant(0, DL, VT);
4842   SDValue One = DAG.getConstant(1, DL, VT);
4843   SDValue MinusXLen = DAG.getConstant(-(int)Subtarget.getXLen(), DL, VT);
4844   SDValue XLenMinus1 = DAG.getConstant(Subtarget.getXLen() - 1, DL, VT);
4845   SDValue ShamtMinusXLen = DAG.getNode(ISD::ADD, DL, VT, Shamt, MinusXLen);
4846   SDValue XLenMinus1Shamt = DAG.getNode(ISD::SUB, DL, VT, XLenMinus1, Shamt);
4847 
4848   SDValue ShiftRightLo = DAG.getNode(ISD::SRL, DL, VT, Lo, Shamt);
4849   SDValue ShiftLeftHi1 = DAG.getNode(ISD::SHL, DL, VT, Hi, One);
4850   SDValue ShiftLeftHi =
4851       DAG.getNode(ISD::SHL, DL, VT, ShiftLeftHi1, XLenMinus1Shamt);
4852   SDValue LoTrue = DAG.getNode(ISD::OR, DL, VT, ShiftRightLo, ShiftLeftHi);
4853   SDValue HiTrue = DAG.getNode(ShiftRightOp, DL, VT, Hi, Shamt);
4854   SDValue LoFalse = DAG.getNode(ShiftRightOp, DL, VT, Hi, ShamtMinusXLen);
4855   SDValue HiFalse =
4856       IsSRA ? DAG.getNode(ISD::SRA, DL, VT, Hi, XLenMinus1) : Zero;
4857 
4858   SDValue CC = DAG.getSetCC(DL, VT, ShamtMinusXLen, Zero, ISD::SETLT);
4859 
4860   Lo = DAG.getNode(ISD::SELECT, DL, VT, CC, LoTrue, LoFalse);
4861   Hi = DAG.getNode(ISD::SELECT, DL, VT, CC, HiTrue, HiFalse);
4862 
4863   SDValue Parts[2] = {Lo, Hi};
4864   return DAG.getMergeValues(Parts, DL);
4865 }
4866 
4867 // Lower splats of i1 types to SETCC. For each mask vector type, we have a
4868 // legal equivalently-sized i8 type, so we can use that as a go-between.
lowerVectorMaskSplat(SDValue Op,SelectionDAG & DAG) const4869 SDValue RISCVTargetLowering::lowerVectorMaskSplat(SDValue Op,
4870                                                   SelectionDAG &DAG) const {
4871   SDLoc DL(Op);
4872   MVT VT = Op.getSimpleValueType();
4873   SDValue SplatVal = Op.getOperand(0);
4874   // All-zeros or all-ones splats are handled specially.
4875   if (ISD::isConstantSplatVectorAllOnes(Op.getNode())) {
4876     SDValue VL = getDefaultScalableVLOps(VT, DL, DAG, Subtarget).second;
4877     return DAG.getNode(RISCVISD::VMSET_VL, DL, VT, VL);
4878   }
4879   if (ISD::isConstantSplatVectorAllZeros(Op.getNode())) {
4880     SDValue VL = getDefaultScalableVLOps(VT, DL, DAG, Subtarget).second;
4881     return DAG.getNode(RISCVISD::VMCLR_VL, DL, VT, VL);
4882   }
4883   MVT XLenVT = Subtarget.getXLenVT();
4884   assert(SplatVal.getValueType() == XLenVT &&
4885          "Unexpected type for i1 splat value");
4886   MVT InterVT = VT.changeVectorElementType(MVT::i8);
4887   SplatVal = DAG.getNode(ISD::AND, DL, XLenVT, SplatVal,
4888                          DAG.getConstant(1, DL, XLenVT));
4889   SDValue LHS = DAG.getSplatVector(InterVT, DL, SplatVal);
4890   SDValue Zero = DAG.getConstant(0, DL, InterVT);
4891   return DAG.getSetCC(DL, VT, LHS, Zero, ISD::SETNE);
4892 }
4893 
4894 // Custom-lower a SPLAT_VECTOR_PARTS where XLEN<SEW, as the SEW element type is
4895 // illegal (currently only vXi64 RV32).
4896 // FIXME: We could also catch non-constant sign-extended i32 values and lower
4897 // them to VMV_V_X_VL.
lowerSPLAT_VECTOR_PARTS(SDValue Op,SelectionDAG & DAG) const4898 SDValue RISCVTargetLowering::lowerSPLAT_VECTOR_PARTS(SDValue Op,
4899                                                      SelectionDAG &DAG) const {
4900   SDLoc DL(Op);
4901   MVT VecVT = Op.getSimpleValueType();
4902   assert(!Subtarget.is64Bit() && VecVT.getVectorElementType() == MVT::i64 &&
4903          "Unexpected SPLAT_VECTOR_PARTS lowering");
4904 
4905   assert(Op.getNumOperands() == 2 && "Unexpected number of operands!");
4906   SDValue Lo = Op.getOperand(0);
4907   SDValue Hi = Op.getOperand(1);
4908 
4909   if (VecVT.isFixedLengthVector()) {
4910     MVT ContainerVT = getContainerForFixedLengthVector(VecVT);
4911     SDLoc DL(Op);
4912     auto VL = getDefaultVLOps(VecVT, ContainerVT, DL, DAG, Subtarget).second;
4913 
4914     SDValue Res =
4915         splatPartsI64WithVL(DL, ContainerVT, SDValue(), Lo, Hi, VL, DAG);
4916     return convertFromScalableVector(VecVT, Res, DAG, Subtarget);
4917   }
4918 
4919   if (isa<ConstantSDNode>(Lo) && isa<ConstantSDNode>(Hi)) {
4920     int32_t LoC = cast<ConstantSDNode>(Lo)->getSExtValue();
4921     int32_t HiC = cast<ConstantSDNode>(Hi)->getSExtValue();
4922     // If Hi constant is all the same sign bit as Lo, lower this as a custom
4923     // node in order to try and match RVV vector/scalar instructions.
4924     if ((LoC >> 31) == HiC)
4925       return DAG.getNode(RISCVISD::VMV_V_X_VL, DL, VecVT, DAG.getUNDEF(VecVT),
4926                          Lo, DAG.getRegister(RISCV::X0, MVT::i32));
4927   }
4928 
4929   // Detect cases where Hi is (SRA Lo, 31) which means Hi is Lo sign extended.
4930   if (Hi.getOpcode() == ISD::SRA && Hi.getOperand(0) == Lo &&
4931       isa<ConstantSDNode>(Hi.getOperand(1)) &&
4932       Hi.getConstantOperandVal(1) == 31)
4933     return DAG.getNode(RISCVISD::VMV_V_X_VL, DL, VecVT, DAG.getUNDEF(VecVT), Lo,
4934                        DAG.getRegister(RISCV::X0, MVT::i32));
4935 
4936   // Fall back to use a stack store and stride x0 vector load. Use X0 as VL.
4937   return DAG.getNode(RISCVISD::SPLAT_VECTOR_SPLIT_I64_VL, DL, VecVT,
4938                      DAG.getUNDEF(VecVT), Lo, Hi,
4939                      DAG.getRegister(RISCV::X0, MVT::i32));
4940 }
4941 
4942 // Custom-lower extensions from mask vectors by using a vselect either with 1
4943 // for zero/any-extension or -1 for sign-extension:
4944 //   (vXiN = (s|z)ext vXi1:vmask) -> (vXiN = vselect vmask, (-1 or 1), 0)
4945 // Note that any-extension is lowered identically to zero-extension.
lowerVectorMaskExt(SDValue Op,SelectionDAG & DAG,int64_t ExtTrueVal) const4946 SDValue RISCVTargetLowering::lowerVectorMaskExt(SDValue Op, SelectionDAG &DAG,
4947                                                 int64_t ExtTrueVal) const {
4948   SDLoc DL(Op);
4949   MVT VecVT = Op.getSimpleValueType();
4950   SDValue Src = Op.getOperand(0);
4951   // Only custom-lower extensions from mask types
4952   assert(Src.getValueType().isVector() &&
4953          Src.getValueType().getVectorElementType() == MVT::i1);
4954 
4955   if (VecVT.isScalableVector()) {
4956     SDValue SplatZero = DAG.getConstant(0, DL, VecVT);
4957     SDValue SplatTrueVal = DAG.getConstant(ExtTrueVal, DL, VecVT);
4958     return DAG.getNode(ISD::VSELECT, DL, VecVT, Src, SplatTrueVal, SplatZero);
4959   }
4960 
4961   MVT ContainerVT = getContainerForFixedLengthVector(VecVT);
4962   MVT I1ContainerVT =
4963       MVT::getVectorVT(MVT::i1, ContainerVT.getVectorElementCount());
4964 
4965   SDValue CC = convertToScalableVector(I1ContainerVT, Src, DAG, Subtarget);
4966 
4967   SDValue VL = getDefaultVLOps(VecVT, ContainerVT, DL, DAG, Subtarget).second;
4968 
4969   MVT XLenVT = Subtarget.getXLenVT();
4970   SDValue SplatZero = DAG.getConstant(0, DL, XLenVT);
4971   SDValue SplatTrueVal = DAG.getConstant(ExtTrueVal, DL, XLenVT);
4972 
4973   SplatZero = DAG.getNode(RISCVISD::VMV_V_X_VL, DL, ContainerVT,
4974                           DAG.getUNDEF(ContainerVT), SplatZero, VL);
4975   SplatTrueVal = DAG.getNode(RISCVISD::VMV_V_X_VL, DL, ContainerVT,
4976                              DAG.getUNDEF(ContainerVT), SplatTrueVal, VL);
4977   SDValue Select = DAG.getNode(RISCVISD::VSELECT_VL, DL, ContainerVT, CC,
4978                                SplatTrueVal, SplatZero, VL);
4979 
4980   return convertFromScalableVector(VecVT, Select, DAG, Subtarget);
4981 }
4982 
lowerFixedLengthVectorExtendToRVV(SDValue Op,SelectionDAG & DAG,unsigned ExtendOpc) const4983 SDValue RISCVTargetLowering::lowerFixedLengthVectorExtendToRVV(
4984     SDValue Op, SelectionDAG &DAG, unsigned ExtendOpc) const {
4985   MVT ExtVT = Op.getSimpleValueType();
4986   // Only custom-lower extensions from fixed-length vector types.
4987   if (!ExtVT.isFixedLengthVector())
4988     return Op;
4989   MVT VT = Op.getOperand(0).getSimpleValueType();
4990   // Grab the canonical container type for the extended type. Infer the smaller
4991   // type from that to ensure the same number of vector elements, as we know
4992   // the LMUL will be sufficient to hold the smaller type.
4993   MVT ContainerExtVT = getContainerForFixedLengthVector(ExtVT);
4994   // Get the extended container type manually to ensure the same number of
4995   // vector elements between source and dest.
4996   MVT ContainerVT = MVT::getVectorVT(VT.getVectorElementType(),
4997                                      ContainerExtVT.getVectorElementCount());
4998 
4999   SDValue Op1 =
5000       convertToScalableVector(ContainerVT, Op.getOperand(0), DAG, Subtarget);
5001 
5002   SDLoc DL(Op);
5003   auto [Mask, VL] = getDefaultVLOps(VT, ContainerVT, DL, DAG, Subtarget);
5004 
5005   SDValue Ext = DAG.getNode(ExtendOpc, DL, ContainerExtVT, Op1, Mask, VL);
5006 
5007   return convertFromScalableVector(ExtVT, Ext, DAG, Subtarget);
5008 }
5009 
5010 // Custom-lower truncations from vectors to mask vectors by using a mask and a
5011 // setcc operation:
5012 //   (vXi1 = trunc vXiN vec) -> (vXi1 = setcc (and vec, 1), 0, ne)
lowerVectorMaskTruncLike(SDValue Op,SelectionDAG & DAG) const5013 SDValue RISCVTargetLowering::lowerVectorMaskTruncLike(SDValue Op,
5014                                                       SelectionDAG &DAG) const {
5015   bool IsVPTrunc = Op.getOpcode() == ISD::VP_TRUNCATE;
5016   SDLoc DL(Op);
5017   EVT MaskVT = Op.getValueType();
5018   // Only expect to custom-lower truncations to mask types
5019   assert(MaskVT.isVector() && MaskVT.getVectorElementType() == MVT::i1 &&
5020          "Unexpected type for vector mask lowering");
5021   SDValue Src = Op.getOperand(0);
5022   MVT VecVT = Src.getSimpleValueType();
5023   SDValue Mask, VL;
5024   if (IsVPTrunc) {
5025     Mask = Op.getOperand(1);
5026     VL = Op.getOperand(2);
5027   }
5028   // If this is a fixed vector, we need to convert it to a scalable vector.
5029   MVT ContainerVT = VecVT;
5030 
5031   if (VecVT.isFixedLengthVector()) {
5032     ContainerVT = getContainerForFixedLengthVector(VecVT);
5033     Src = convertToScalableVector(ContainerVT, Src, DAG, Subtarget);
5034     if (IsVPTrunc) {
5035       MVT MaskContainerVT =
5036           getContainerForFixedLengthVector(Mask.getSimpleValueType());
5037       Mask = convertToScalableVector(MaskContainerVT, Mask, DAG, Subtarget);
5038     }
5039   }
5040 
5041   if (!IsVPTrunc) {
5042     std::tie(Mask, VL) =
5043         getDefaultVLOps(VecVT, ContainerVT, DL, DAG, Subtarget);
5044   }
5045 
5046   SDValue SplatOne = DAG.getConstant(1, DL, Subtarget.getXLenVT());
5047   SDValue SplatZero = DAG.getConstant(0, DL, Subtarget.getXLenVT());
5048 
5049   SplatOne = DAG.getNode(RISCVISD::VMV_V_X_VL, DL, ContainerVT,
5050                          DAG.getUNDEF(ContainerVT), SplatOne, VL);
5051   SplatZero = DAG.getNode(RISCVISD::VMV_V_X_VL, DL, ContainerVT,
5052                           DAG.getUNDEF(ContainerVT), SplatZero, VL);
5053 
5054   MVT MaskContainerVT = ContainerVT.changeVectorElementType(MVT::i1);
5055   SDValue Trunc = DAG.getNode(RISCVISD::AND_VL, DL, ContainerVT, Src, SplatOne,
5056                               DAG.getUNDEF(ContainerVT), Mask, VL);
5057   Trunc = DAG.getNode(RISCVISD::SETCC_VL, DL, MaskContainerVT,
5058                       {Trunc, SplatZero, DAG.getCondCode(ISD::SETNE),
5059                        DAG.getUNDEF(MaskContainerVT), Mask, VL});
5060   if (MaskVT.isFixedLengthVector())
5061     Trunc = convertFromScalableVector(MaskVT, Trunc, DAG, Subtarget);
5062   return Trunc;
5063 }
5064 
lowerVectorTruncLike(SDValue Op,SelectionDAG & DAG) const5065 SDValue RISCVTargetLowering::lowerVectorTruncLike(SDValue Op,
5066                                                   SelectionDAG &DAG) const {
5067   bool IsVPTrunc = Op.getOpcode() == ISD::VP_TRUNCATE;
5068   SDLoc DL(Op);
5069 
5070   MVT VT = Op.getSimpleValueType();
5071   // Only custom-lower vector truncates
5072   assert(VT.isVector() && "Unexpected type for vector truncate lowering");
5073 
5074   // Truncates to mask types are handled differently
5075   if (VT.getVectorElementType() == MVT::i1)
5076     return lowerVectorMaskTruncLike(Op, DAG);
5077 
5078   // RVV only has truncates which operate from SEW*2->SEW, so lower arbitrary
5079   // truncates as a series of "RISCVISD::TRUNCATE_VECTOR_VL" nodes which
5080   // truncate by one power of two at a time.
5081   MVT DstEltVT = VT.getVectorElementType();
5082 
5083   SDValue Src = Op.getOperand(0);
5084   MVT SrcVT = Src.getSimpleValueType();
5085   MVT SrcEltVT = SrcVT.getVectorElementType();
5086 
5087   assert(DstEltVT.bitsLT(SrcEltVT) && isPowerOf2_64(DstEltVT.getSizeInBits()) &&
5088          isPowerOf2_64(SrcEltVT.getSizeInBits()) &&
5089          "Unexpected vector truncate lowering");
5090 
5091   MVT ContainerVT = SrcVT;
5092   SDValue Mask, VL;
5093   if (IsVPTrunc) {
5094     Mask = Op.getOperand(1);
5095     VL = Op.getOperand(2);
5096   }
5097   if (SrcVT.isFixedLengthVector()) {
5098     ContainerVT = getContainerForFixedLengthVector(SrcVT);
5099     Src = convertToScalableVector(ContainerVT, Src, DAG, Subtarget);
5100     if (IsVPTrunc) {
5101       MVT MaskVT = getMaskTypeFor(ContainerVT);
5102       Mask = convertToScalableVector(MaskVT, Mask, DAG, Subtarget);
5103     }
5104   }
5105 
5106   SDValue Result = Src;
5107   if (!IsVPTrunc) {
5108     std::tie(Mask, VL) =
5109         getDefaultVLOps(SrcVT, ContainerVT, DL, DAG, Subtarget);
5110   }
5111 
5112   LLVMContext &Context = *DAG.getContext();
5113   const ElementCount Count = ContainerVT.getVectorElementCount();
5114   do {
5115     SrcEltVT = MVT::getIntegerVT(SrcEltVT.getSizeInBits() / 2);
5116     EVT ResultVT = EVT::getVectorVT(Context, SrcEltVT, Count);
5117     Result = DAG.getNode(RISCVISD::TRUNCATE_VECTOR_VL, DL, ResultVT, Result,
5118                          Mask, VL);
5119   } while (SrcEltVT != DstEltVT);
5120 
5121   if (SrcVT.isFixedLengthVector())
5122     Result = convertFromScalableVector(VT, Result, DAG, Subtarget);
5123 
5124   return Result;
5125 }
5126 
5127 SDValue
lowerVectorFPExtendOrRoundLike(SDValue Op,SelectionDAG & DAG) const5128 RISCVTargetLowering::lowerVectorFPExtendOrRoundLike(SDValue Op,
5129                                                     SelectionDAG &DAG) const {
5130   bool IsVP =
5131       Op.getOpcode() == ISD::VP_FP_ROUND || Op.getOpcode() == ISD::VP_FP_EXTEND;
5132   bool IsExtend =
5133       Op.getOpcode() == ISD::VP_FP_EXTEND || Op.getOpcode() == ISD::FP_EXTEND;
5134   // RVV can only do truncate fp to types half the size as the source. We
5135   // custom-lower f64->f16 rounds via RVV's round-to-odd float
5136   // conversion instruction.
5137   SDLoc DL(Op);
5138   MVT VT = Op.getSimpleValueType();
5139 
5140   assert(VT.isVector() && "Unexpected type for vector truncate lowering");
5141 
5142   SDValue Src = Op.getOperand(0);
5143   MVT SrcVT = Src.getSimpleValueType();
5144 
5145   bool IsDirectExtend = IsExtend && (VT.getVectorElementType() != MVT::f64 ||
5146                                      SrcVT.getVectorElementType() != MVT::f16);
5147   bool IsDirectTrunc = !IsExtend && (VT.getVectorElementType() != MVT::f16 ||
5148                                      SrcVT.getVectorElementType() != MVT::f64);
5149 
5150   bool IsDirectConv = IsDirectExtend || IsDirectTrunc;
5151 
5152   // Prepare any fixed-length vector operands.
5153   MVT ContainerVT = VT;
5154   SDValue Mask, VL;
5155   if (IsVP) {
5156     Mask = Op.getOperand(1);
5157     VL = Op.getOperand(2);
5158   }
5159   if (VT.isFixedLengthVector()) {
5160     MVT SrcContainerVT = getContainerForFixedLengthVector(SrcVT);
5161     ContainerVT =
5162         SrcContainerVT.changeVectorElementType(VT.getVectorElementType());
5163     Src = convertToScalableVector(SrcContainerVT, Src, DAG, Subtarget);
5164     if (IsVP) {
5165       MVT MaskVT = getMaskTypeFor(ContainerVT);
5166       Mask = convertToScalableVector(MaskVT, Mask, DAG, Subtarget);
5167     }
5168   }
5169 
5170   if (!IsVP)
5171     std::tie(Mask, VL) =
5172         getDefaultVLOps(SrcVT, ContainerVT, DL, DAG, Subtarget);
5173 
5174   unsigned ConvOpc = IsExtend ? RISCVISD::FP_EXTEND_VL : RISCVISD::FP_ROUND_VL;
5175 
5176   if (IsDirectConv) {
5177     Src = DAG.getNode(ConvOpc, DL, ContainerVT, Src, Mask, VL);
5178     if (VT.isFixedLengthVector())
5179       Src = convertFromScalableVector(VT, Src, DAG, Subtarget);
5180     return Src;
5181   }
5182 
5183   unsigned InterConvOpc =
5184       IsExtend ? RISCVISD::FP_EXTEND_VL : RISCVISD::VFNCVT_ROD_VL;
5185 
5186   MVT InterVT = ContainerVT.changeVectorElementType(MVT::f32);
5187   SDValue IntermediateConv =
5188       DAG.getNode(InterConvOpc, DL, InterVT, Src, Mask, VL);
5189   SDValue Result =
5190       DAG.getNode(ConvOpc, DL, ContainerVT, IntermediateConv, Mask, VL);
5191   if (VT.isFixedLengthVector())
5192     return convertFromScalableVector(VT, Result, DAG, Subtarget);
5193   return Result;
5194 }
5195 
5196 // Custom-legalize INSERT_VECTOR_ELT so that the value is inserted into the
5197 // first position of a vector, and that vector is slid up to the insert index.
5198 // By limiting the active vector length to index+1 and merging with the
5199 // original vector (with an undisturbed tail policy for elements >= VL), we
5200 // achieve the desired result of leaving all elements untouched except the one
5201 // at VL-1, which is replaced with the desired value.
lowerINSERT_VECTOR_ELT(SDValue Op,SelectionDAG & DAG) const5202 SDValue RISCVTargetLowering::lowerINSERT_VECTOR_ELT(SDValue Op,
5203                                                     SelectionDAG &DAG) const {
5204   SDLoc DL(Op);
5205   MVT VecVT = Op.getSimpleValueType();
5206   SDValue Vec = Op.getOperand(0);
5207   SDValue Val = Op.getOperand(1);
5208   SDValue Idx = Op.getOperand(2);
5209 
5210   if (VecVT.getVectorElementType() == MVT::i1) {
5211     // FIXME: For now we just promote to an i8 vector and insert into that,
5212     // but this is probably not optimal.
5213     MVT WideVT = MVT::getVectorVT(MVT::i8, VecVT.getVectorElementCount());
5214     Vec = DAG.getNode(ISD::ZERO_EXTEND, DL, WideVT, Vec);
5215     Vec = DAG.getNode(ISD::INSERT_VECTOR_ELT, DL, WideVT, Vec, Val, Idx);
5216     return DAG.getNode(ISD::TRUNCATE, DL, VecVT, Vec);
5217   }
5218 
5219   MVT ContainerVT = VecVT;
5220   // If the operand is a fixed-length vector, convert to a scalable one.
5221   if (VecVT.isFixedLengthVector()) {
5222     ContainerVT = getContainerForFixedLengthVector(VecVT);
5223     Vec = convertToScalableVector(ContainerVT, Vec, DAG, Subtarget);
5224   }
5225 
5226   MVT XLenVT = Subtarget.getXLenVT();
5227 
5228   SDValue Zero = DAG.getConstant(0, DL, XLenVT);
5229   bool IsLegalInsert = Subtarget.is64Bit() || Val.getValueType() != MVT::i64;
5230   // Even i64-element vectors on RV32 can be lowered without scalar
5231   // legalization if the most-significant 32 bits of the value are not affected
5232   // by the sign-extension of the lower 32 bits.
5233   // TODO: We could also catch sign extensions of a 32-bit value.
5234   if (!IsLegalInsert && isa<ConstantSDNode>(Val)) {
5235     const auto *CVal = cast<ConstantSDNode>(Val);
5236     if (isInt<32>(CVal->getSExtValue())) {
5237       IsLegalInsert = true;
5238       Val = DAG.getConstant(CVal->getSExtValue(), DL, MVT::i32);
5239     }
5240   }
5241 
5242   auto [Mask, VL] = getDefaultVLOps(VecVT, ContainerVT, DL, DAG, Subtarget);
5243 
5244   SDValue ValInVec;
5245 
5246   if (IsLegalInsert) {
5247     unsigned Opc =
5248         VecVT.isFloatingPoint() ? RISCVISD::VFMV_S_F_VL : RISCVISD::VMV_S_X_VL;
5249     if (isNullConstant(Idx)) {
5250       Vec = DAG.getNode(Opc, DL, ContainerVT, Vec, Val, VL);
5251       if (!VecVT.isFixedLengthVector())
5252         return Vec;
5253       return convertFromScalableVector(VecVT, Vec, DAG, Subtarget);
5254     }
5255     ValInVec = lowerScalarInsert(Val, VL, ContainerVT, DL, DAG, Subtarget);
5256   } else {
5257     // On RV32, i64-element vectors must be specially handled to place the
5258     // value at element 0, by using two vslide1down instructions in sequence on
5259     // the i32 split lo/hi value. Use an equivalently-sized i32 vector for
5260     // this.
5261     SDValue One = DAG.getConstant(1, DL, XLenVT);
5262     SDValue ValLo = DAG.getNode(ISD::EXTRACT_ELEMENT, DL, MVT::i32, Val, Zero);
5263     SDValue ValHi = DAG.getNode(ISD::EXTRACT_ELEMENT, DL, MVT::i32, Val, One);
5264     MVT I32ContainerVT =
5265         MVT::getVectorVT(MVT::i32, ContainerVT.getVectorElementCount() * 2);
5266     SDValue I32Mask =
5267         getDefaultScalableVLOps(I32ContainerVT, DL, DAG, Subtarget).first;
5268     // Limit the active VL to two.
5269     SDValue InsertI64VL = DAG.getConstant(2, DL, XLenVT);
5270     // If the Idx is 0 we can insert directly into the vector.
5271     if (isNullConstant(Idx)) {
5272       // First slide in the lo value, then the hi in above it. We use slide1down
5273       // to avoid the register group overlap constraint of vslide1up.
5274       ValInVec = DAG.getNode(RISCVISD::VSLIDE1DOWN_VL, DL, I32ContainerVT,
5275                              Vec, Vec, ValLo, I32Mask, InsertI64VL);
5276       // If the source vector is undef don't pass along the tail elements from
5277       // the previous slide1down.
5278       SDValue Tail = Vec.isUndef() ? Vec : ValInVec;
5279       ValInVec = DAG.getNode(RISCVISD::VSLIDE1DOWN_VL, DL, I32ContainerVT,
5280                              Tail, ValInVec, ValHi, I32Mask, InsertI64VL);
5281       // Bitcast back to the right container type.
5282       ValInVec = DAG.getBitcast(ContainerVT, ValInVec);
5283 
5284       if (!VecVT.isFixedLengthVector())
5285         return ValInVec;
5286       return convertFromScalableVector(VecVT, ValInVec, DAG, Subtarget);
5287     }
5288 
5289     // First slide in the lo value, then the hi in above it. We use slide1down
5290     // to avoid the register group overlap constraint of vslide1up.
5291     ValInVec = DAG.getNode(RISCVISD::VSLIDE1DOWN_VL, DL, I32ContainerVT,
5292                            DAG.getUNDEF(I32ContainerVT),
5293                            DAG.getUNDEF(I32ContainerVT), ValLo,
5294                            I32Mask, InsertI64VL);
5295     ValInVec = DAG.getNode(RISCVISD::VSLIDE1DOWN_VL, DL, I32ContainerVT,
5296                            DAG.getUNDEF(I32ContainerVT), ValInVec, ValHi,
5297                            I32Mask, InsertI64VL);
5298     // Bitcast back to the right container type.
5299     ValInVec = DAG.getBitcast(ContainerVT, ValInVec);
5300   }
5301 
5302   // Now that the value is in a vector, slide it into position.
5303   SDValue InsertVL =
5304       DAG.getNode(ISD::ADD, DL, XLenVT, Idx, DAG.getConstant(1, DL, XLenVT));
5305 
5306   // Use tail agnostic policy if Idx is the last index of Vec.
5307   unsigned Policy = RISCVII::TAIL_UNDISTURBED_MASK_UNDISTURBED;
5308   if (VecVT.isFixedLengthVector() && isa<ConstantSDNode>(Idx) &&
5309       cast<ConstantSDNode>(Idx)->getZExtValue() + 1 ==
5310           VecVT.getVectorNumElements())
5311     Policy = RISCVII::TAIL_AGNOSTIC;
5312   SDValue Slideup = getVSlideup(DAG, Subtarget, DL, ContainerVT, Vec, ValInVec,
5313                                 Idx, Mask, InsertVL, Policy);
5314   if (!VecVT.isFixedLengthVector())
5315     return Slideup;
5316   return convertFromScalableVector(VecVT, Slideup, DAG, Subtarget);
5317 }
5318 
5319 // Custom-lower EXTRACT_VECTOR_ELT operations to slide the vector down, then
5320 // extract the first element: (extractelt (slidedown vec, idx), 0). For integer
5321 // types this is done using VMV_X_S to allow us to glean information about the
5322 // sign bits of the result.
lowerEXTRACT_VECTOR_ELT(SDValue Op,SelectionDAG & DAG) const5323 SDValue RISCVTargetLowering::lowerEXTRACT_VECTOR_ELT(SDValue Op,
5324                                                      SelectionDAG &DAG) const {
5325   SDLoc DL(Op);
5326   SDValue Idx = Op.getOperand(1);
5327   SDValue Vec = Op.getOperand(0);
5328   EVT EltVT = Op.getValueType();
5329   MVT VecVT = Vec.getSimpleValueType();
5330   MVT XLenVT = Subtarget.getXLenVT();
5331 
5332   if (VecVT.getVectorElementType() == MVT::i1) {
5333     // Use vfirst.m to extract the first bit.
5334     if (isNullConstant(Idx)) {
5335       MVT ContainerVT = VecVT;
5336       if (VecVT.isFixedLengthVector()) {
5337         ContainerVT = getContainerForFixedLengthVector(VecVT);
5338         Vec = convertToScalableVector(ContainerVT, Vec, DAG, Subtarget);
5339       }
5340       auto [Mask, VL] = getDefaultVLOps(VecVT, ContainerVT, DL, DAG, Subtarget);
5341       SDValue Vfirst =
5342           DAG.getNode(RISCVISD::VFIRST_VL, DL, XLenVT, Vec, Mask, VL);
5343       return DAG.getSetCC(DL, XLenVT, Vfirst, DAG.getConstant(0, DL, XLenVT),
5344                           ISD::SETEQ);
5345     }
5346     if (VecVT.isFixedLengthVector()) {
5347       unsigned NumElts = VecVT.getVectorNumElements();
5348       if (NumElts >= 8) {
5349         MVT WideEltVT;
5350         unsigned WidenVecLen;
5351         SDValue ExtractElementIdx;
5352         SDValue ExtractBitIdx;
5353         unsigned MaxEEW = Subtarget.getELEN();
5354         MVT LargestEltVT = MVT::getIntegerVT(
5355             std::min(MaxEEW, unsigned(XLenVT.getSizeInBits())));
5356         if (NumElts <= LargestEltVT.getSizeInBits()) {
5357           assert(isPowerOf2_32(NumElts) &&
5358                  "the number of elements should be power of 2");
5359           WideEltVT = MVT::getIntegerVT(NumElts);
5360           WidenVecLen = 1;
5361           ExtractElementIdx = DAG.getConstant(0, DL, XLenVT);
5362           ExtractBitIdx = Idx;
5363         } else {
5364           WideEltVT = LargestEltVT;
5365           WidenVecLen = NumElts / WideEltVT.getSizeInBits();
5366           // extract element index = index / element width
5367           ExtractElementIdx = DAG.getNode(
5368               ISD::SRL, DL, XLenVT, Idx,
5369               DAG.getConstant(Log2_64(WideEltVT.getSizeInBits()), DL, XLenVT));
5370           // mask bit index = index % element width
5371           ExtractBitIdx = DAG.getNode(
5372               ISD::AND, DL, XLenVT, Idx,
5373               DAG.getConstant(WideEltVT.getSizeInBits() - 1, DL, XLenVT));
5374         }
5375         MVT WideVT = MVT::getVectorVT(WideEltVT, WidenVecLen);
5376         Vec = DAG.getNode(ISD::BITCAST, DL, WideVT, Vec);
5377         SDValue ExtractElt = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, XLenVT,
5378                                          Vec, ExtractElementIdx);
5379         // Extract the bit from GPR.
5380         SDValue ShiftRight =
5381             DAG.getNode(ISD::SRL, DL, XLenVT, ExtractElt, ExtractBitIdx);
5382         return DAG.getNode(ISD::AND, DL, XLenVT, ShiftRight,
5383                            DAG.getConstant(1, DL, XLenVT));
5384       }
5385     }
5386     // Otherwise, promote to an i8 vector and extract from that.
5387     MVT WideVT = MVT::getVectorVT(MVT::i8, VecVT.getVectorElementCount());
5388     Vec = DAG.getNode(ISD::ZERO_EXTEND, DL, WideVT, Vec);
5389     return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, EltVT, Vec, Idx);
5390   }
5391 
5392   // If this is a fixed vector, we need to convert it to a scalable vector.
5393   MVT ContainerVT = VecVT;
5394   if (VecVT.isFixedLengthVector()) {
5395     ContainerVT = getContainerForFixedLengthVector(VecVT);
5396     Vec = convertToScalableVector(ContainerVT, Vec, DAG, Subtarget);
5397   }
5398 
5399   // If the index is 0, the vector is already in the right position.
5400   if (!isNullConstant(Idx)) {
5401     // Use a VL of 1 to avoid processing more elements than we need.
5402     auto [Mask, VL] = getDefaultVLOps(1, ContainerVT, DL, DAG, Subtarget);
5403     Vec = getVSlidedown(DAG, Subtarget, DL, ContainerVT,
5404                         DAG.getUNDEF(ContainerVT), Vec, Idx, Mask, VL);
5405   }
5406 
5407   if (!EltVT.isInteger()) {
5408     // Floating-point extracts are handled in TableGen.
5409     return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, EltVT, Vec,
5410                        DAG.getConstant(0, DL, XLenVT));
5411   }
5412 
5413   SDValue Elt0 = DAG.getNode(RISCVISD::VMV_X_S, DL, XLenVT, Vec);
5414   return DAG.getNode(ISD::TRUNCATE, DL, EltVT, Elt0);
5415 }
5416 
5417 // Some RVV intrinsics may claim that they want an integer operand to be
5418 // promoted or expanded.
lowerVectorIntrinsicScalars(SDValue Op,SelectionDAG & DAG,const RISCVSubtarget & Subtarget)5419 static SDValue lowerVectorIntrinsicScalars(SDValue Op, SelectionDAG &DAG,
5420                                            const RISCVSubtarget &Subtarget) {
5421   assert((Op.getOpcode() == ISD::INTRINSIC_WO_CHAIN ||
5422           Op.getOpcode() == ISD::INTRINSIC_W_CHAIN) &&
5423          "Unexpected opcode");
5424 
5425   if (!Subtarget.hasVInstructions())
5426     return SDValue();
5427 
5428   bool HasChain = Op.getOpcode() == ISD::INTRINSIC_W_CHAIN;
5429   unsigned IntNo = Op.getConstantOperandVal(HasChain ? 1 : 0);
5430   SDLoc DL(Op);
5431 
5432   const RISCVVIntrinsicsTable::RISCVVIntrinsicInfo *II =
5433       RISCVVIntrinsicsTable::getRISCVVIntrinsicInfo(IntNo);
5434   if (!II || !II->hasScalarOperand())
5435     return SDValue();
5436 
5437   unsigned SplatOp = II->ScalarOperand + 1 + HasChain;
5438   assert(SplatOp < Op.getNumOperands());
5439 
5440   SmallVector<SDValue, 8> Operands(Op->op_begin(), Op->op_end());
5441   SDValue &ScalarOp = Operands[SplatOp];
5442   MVT OpVT = ScalarOp.getSimpleValueType();
5443   MVT XLenVT = Subtarget.getXLenVT();
5444 
5445   // If this isn't a scalar, or its type is XLenVT we're done.
5446   if (!OpVT.isScalarInteger() || OpVT == XLenVT)
5447     return SDValue();
5448 
5449   // Simplest case is that the operand needs to be promoted to XLenVT.
5450   if (OpVT.bitsLT(XLenVT)) {
5451     // If the operand is a constant, sign extend to increase our chances
5452     // of being able to use a .vi instruction. ANY_EXTEND would become a
5453     // a zero extend and the simm5 check in isel would fail.
5454     // FIXME: Should we ignore the upper bits in isel instead?
5455     unsigned ExtOpc =
5456         isa<ConstantSDNode>(ScalarOp) ? ISD::SIGN_EXTEND : ISD::ANY_EXTEND;
5457     ScalarOp = DAG.getNode(ExtOpc, DL, XLenVT, ScalarOp);
5458     return DAG.getNode(Op->getOpcode(), DL, Op->getVTList(), Operands);
5459   }
5460 
5461   // Use the previous operand to get the vXi64 VT. The result might be a mask
5462   // VT for compares. Using the previous operand assumes that the previous
5463   // operand will never have a smaller element size than a scalar operand and
5464   // that a widening operation never uses SEW=64.
5465   // NOTE: If this fails the below assert, we can probably just find the
5466   // element count from any operand or result and use it to construct the VT.
5467   assert(II->ScalarOperand > 0 && "Unexpected splat operand!");
5468   MVT VT = Op.getOperand(SplatOp - 1).getSimpleValueType();
5469 
5470   // The more complex case is when the scalar is larger than XLenVT.
5471   assert(XLenVT == MVT::i32 && OpVT == MVT::i64 &&
5472          VT.getVectorElementType() == MVT::i64 && "Unexpected VTs!");
5473 
5474   // If this is a sign-extended 32-bit value, we can truncate it and rely on the
5475   // instruction to sign-extend since SEW>XLEN.
5476   if (DAG.ComputeNumSignBits(ScalarOp) > 32) {
5477     ScalarOp = DAG.getNode(ISD::TRUNCATE, DL, MVT::i32, ScalarOp);
5478     return DAG.getNode(Op->getOpcode(), DL, Op->getVTList(), Operands);
5479   }
5480 
5481   switch (IntNo) {
5482   case Intrinsic::riscv_vslide1up:
5483   case Intrinsic::riscv_vslide1down:
5484   case Intrinsic::riscv_vslide1up_mask:
5485   case Intrinsic::riscv_vslide1down_mask: {
5486     // We need to special case these when the scalar is larger than XLen.
5487     unsigned NumOps = Op.getNumOperands();
5488     bool IsMasked = NumOps == 7;
5489 
5490     // Convert the vector source to the equivalent nxvXi32 vector.
5491     MVT I32VT = MVT::getVectorVT(MVT::i32, VT.getVectorElementCount() * 2);
5492     SDValue Vec = DAG.getBitcast(I32VT, Operands[2]);
5493 
5494     SDValue ScalarLo = DAG.getNode(ISD::EXTRACT_ELEMENT, DL, MVT::i32, ScalarOp,
5495                                    DAG.getConstant(0, DL, XLenVT));
5496     SDValue ScalarHi = DAG.getNode(ISD::EXTRACT_ELEMENT, DL, MVT::i32, ScalarOp,
5497                                    DAG.getConstant(1, DL, XLenVT));
5498 
5499     // Double the VL since we halved SEW.
5500     SDValue AVL = getVLOperand(Op);
5501     SDValue I32VL;
5502 
5503     // Optimize for constant AVL
5504     if (isa<ConstantSDNode>(AVL)) {
5505       unsigned EltSize = VT.getScalarSizeInBits();
5506       unsigned MinSize = VT.getSizeInBits().getKnownMinValue();
5507 
5508       unsigned VectorBitsMax = Subtarget.getRealMaxVLen();
5509       unsigned MaxVLMAX =
5510           RISCVTargetLowering::computeVLMAX(VectorBitsMax, EltSize, MinSize);
5511 
5512       unsigned VectorBitsMin = Subtarget.getRealMinVLen();
5513       unsigned MinVLMAX =
5514           RISCVTargetLowering::computeVLMAX(VectorBitsMin, EltSize, MinSize);
5515 
5516       uint64_t AVLInt = cast<ConstantSDNode>(AVL)->getZExtValue();
5517       if (AVLInt <= MinVLMAX) {
5518         I32VL = DAG.getConstant(2 * AVLInt, DL, XLenVT);
5519       } else if (AVLInt >= 2 * MaxVLMAX) {
5520         // Just set vl to VLMAX in this situation
5521         RISCVII::VLMUL Lmul = RISCVTargetLowering::getLMUL(I32VT);
5522         SDValue LMUL = DAG.getConstant(Lmul, DL, XLenVT);
5523         unsigned Sew = RISCVVType::encodeSEW(I32VT.getScalarSizeInBits());
5524         SDValue SEW = DAG.getConstant(Sew, DL, XLenVT);
5525         SDValue SETVLMAX = DAG.getTargetConstant(
5526             Intrinsic::riscv_vsetvlimax_opt, DL, MVT::i32);
5527         I32VL = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, XLenVT, SETVLMAX, SEW,
5528                             LMUL);
5529       } else {
5530         // For AVL between (MinVLMAX, 2 * MaxVLMAX), the actual working vl
5531         // is related to the hardware implementation.
5532         // So let the following code handle
5533       }
5534     }
5535     if (!I32VL) {
5536       RISCVII::VLMUL Lmul = RISCVTargetLowering::getLMUL(VT);
5537       SDValue LMUL = DAG.getConstant(Lmul, DL, XLenVT);
5538       unsigned Sew = RISCVVType::encodeSEW(VT.getScalarSizeInBits());
5539       SDValue SEW = DAG.getConstant(Sew, DL, XLenVT);
5540       SDValue SETVL =
5541           DAG.getTargetConstant(Intrinsic::riscv_vsetvli_opt, DL, MVT::i32);
5542       // Using vsetvli instruction to get actually used length which related to
5543       // the hardware implementation
5544       SDValue VL = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, XLenVT, SETVL, AVL,
5545                                SEW, LMUL);
5546       I32VL =
5547           DAG.getNode(ISD::SHL, DL, XLenVT, VL, DAG.getConstant(1, DL, XLenVT));
5548     }
5549 
5550     SDValue I32Mask = getAllOnesMask(I32VT, I32VL, DL, DAG);
5551 
5552     // Shift the two scalar parts in using SEW=32 slide1up/slide1down
5553     // instructions.
5554     SDValue Passthru;
5555     if (IsMasked)
5556       Passthru = DAG.getUNDEF(I32VT);
5557     else
5558       Passthru = DAG.getBitcast(I32VT, Operands[1]);
5559 
5560     if (IntNo == Intrinsic::riscv_vslide1up ||
5561         IntNo == Intrinsic::riscv_vslide1up_mask) {
5562       Vec = DAG.getNode(RISCVISD::VSLIDE1UP_VL, DL, I32VT, Passthru, Vec,
5563                         ScalarHi, I32Mask, I32VL);
5564       Vec = DAG.getNode(RISCVISD::VSLIDE1UP_VL, DL, I32VT, Passthru, Vec,
5565                         ScalarLo, I32Mask, I32VL);
5566     } else {
5567       Vec = DAG.getNode(RISCVISD::VSLIDE1DOWN_VL, DL, I32VT, Passthru, Vec,
5568                         ScalarLo, I32Mask, I32VL);
5569       Vec = DAG.getNode(RISCVISD::VSLIDE1DOWN_VL, DL, I32VT, Passthru, Vec,
5570                         ScalarHi, I32Mask, I32VL);
5571     }
5572 
5573     // Convert back to nxvXi64.
5574     Vec = DAG.getBitcast(VT, Vec);
5575 
5576     if (!IsMasked)
5577       return Vec;
5578     // Apply mask after the operation.
5579     SDValue Mask = Operands[NumOps - 3];
5580     SDValue MaskedOff = Operands[1];
5581     // Assume Policy operand is the last operand.
5582     uint64_t Policy =
5583         cast<ConstantSDNode>(Operands[NumOps - 1])->getZExtValue();
5584     // We don't need to select maskedoff if it's undef.
5585     if (MaskedOff.isUndef())
5586       return Vec;
5587     // TAMU
5588     if (Policy == RISCVII::TAIL_AGNOSTIC)
5589       return DAG.getNode(RISCVISD::VSELECT_VL, DL, VT, Mask, Vec, MaskedOff,
5590                          AVL);
5591     // TUMA or TUMU: Currently we always emit tumu policy regardless of tuma.
5592     // It's fine because vmerge does not care mask policy.
5593     return DAG.getNode(RISCVISD::VP_MERGE_VL, DL, VT, Mask, Vec, MaskedOff,
5594                        AVL);
5595   }
5596   }
5597 
5598   // We need to convert the scalar to a splat vector.
5599   SDValue VL = getVLOperand(Op);
5600   assert(VL.getValueType() == XLenVT);
5601   ScalarOp = splatSplitI64WithVL(DL, VT, SDValue(), ScalarOp, VL, DAG);
5602   return DAG.getNode(Op->getOpcode(), DL, Op->getVTList(), Operands);
5603 }
5604 
LowerINTRINSIC_WO_CHAIN(SDValue Op,SelectionDAG & DAG) const5605 SDValue RISCVTargetLowering::LowerINTRINSIC_WO_CHAIN(SDValue Op,
5606                                                      SelectionDAG &DAG) const {
5607   unsigned IntNo = Op.getConstantOperandVal(0);
5608   SDLoc DL(Op);
5609   MVT XLenVT = Subtarget.getXLenVT();
5610 
5611   switch (IntNo) {
5612   default:
5613     break; // Don't custom lower most intrinsics.
5614   case Intrinsic::thread_pointer: {
5615     EVT PtrVT = getPointerTy(DAG.getDataLayout());
5616     return DAG.getRegister(RISCV::X4, PtrVT);
5617   }
5618   case Intrinsic::riscv_orc_b:
5619   case Intrinsic::riscv_brev8: {
5620     unsigned Opc =
5621         IntNo == Intrinsic::riscv_brev8 ? RISCVISD::BREV8 : RISCVISD::ORC_B;
5622     return DAG.getNode(Opc, DL, XLenVT, Op.getOperand(1));
5623   }
5624   case Intrinsic::riscv_zip:
5625   case Intrinsic::riscv_unzip: {
5626     unsigned Opc =
5627         IntNo == Intrinsic::riscv_zip ? RISCVISD::ZIP : RISCVISD::UNZIP;
5628     return DAG.getNode(Opc, DL, XLenVT, Op.getOperand(1));
5629   }
5630   case Intrinsic::riscv_vmv_x_s:
5631     assert(Op.getValueType() == XLenVT && "Unexpected VT!");
5632     return DAG.getNode(RISCVISD::VMV_X_S, DL, Op.getValueType(),
5633                        Op.getOperand(1));
5634   case Intrinsic::riscv_vfmv_f_s:
5635     return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, Op.getValueType(),
5636                        Op.getOperand(1), DAG.getConstant(0, DL, XLenVT));
5637   case Intrinsic::riscv_vmv_v_x:
5638     return lowerScalarSplat(Op.getOperand(1), Op.getOperand(2),
5639                             Op.getOperand(3), Op.getSimpleValueType(), DL, DAG,
5640                             Subtarget);
5641   case Intrinsic::riscv_vfmv_v_f:
5642     return DAG.getNode(RISCVISD::VFMV_V_F_VL, DL, Op.getValueType(),
5643                        Op.getOperand(1), Op.getOperand(2), Op.getOperand(3));
5644   case Intrinsic::riscv_vmv_s_x: {
5645     SDValue Scalar = Op.getOperand(2);
5646 
5647     if (Scalar.getValueType().bitsLE(XLenVT)) {
5648       Scalar = DAG.getNode(ISD::ANY_EXTEND, DL, XLenVT, Scalar);
5649       return DAG.getNode(RISCVISD::VMV_S_X_VL, DL, Op.getValueType(),
5650                          Op.getOperand(1), Scalar, Op.getOperand(3));
5651     }
5652 
5653     assert(Scalar.getValueType() == MVT::i64 && "Unexpected scalar VT!");
5654 
5655     // This is an i64 value that lives in two scalar registers. We have to
5656     // insert this in a convoluted way. First we build vXi64 splat containing
5657     // the two values that we assemble using some bit math. Next we'll use
5658     // vid.v and vmseq to build a mask with bit 0 set. Then we'll use that mask
5659     // to merge element 0 from our splat into the source vector.
5660     // FIXME: This is probably not the best way to do this, but it is
5661     // consistent with INSERT_VECTOR_ELT lowering so it is a good starting
5662     // point.
5663     //   sw lo, (a0)
5664     //   sw hi, 4(a0)
5665     //   vlse vX, (a0)
5666     //
5667     //   vid.v      vVid
5668     //   vmseq.vx   mMask, vVid, 0
5669     //   vmerge.vvm vDest, vSrc, vVal, mMask
5670     MVT VT = Op.getSimpleValueType();
5671     SDValue Vec = Op.getOperand(1);
5672     SDValue VL = getVLOperand(Op);
5673 
5674     SDValue SplattedVal = splatSplitI64WithVL(DL, VT, SDValue(), Scalar, VL, DAG);
5675     if (Op.getOperand(1).isUndef())
5676       return SplattedVal;
5677     SDValue SplattedIdx =
5678         DAG.getNode(RISCVISD::VMV_V_X_VL, DL, VT, DAG.getUNDEF(VT),
5679                     DAG.getConstant(0, DL, MVT::i32), VL);
5680 
5681     MVT MaskVT = getMaskTypeFor(VT);
5682     SDValue Mask = getAllOnesMask(VT, VL, DL, DAG);
5683     SDValue VID = DAG.getNode(RISCVISD::VID_VL, DL, VT, Mask, VL);
5684     SDValue SelectCond =
5685         DAG.getNode(RISCVISD::SETCC_VL, DL, MaskVT,
5686                     {VID, SplattedIdx, DAG.getCondCode(ISD::SETEQ),
5687                      DAG.getUNDEF(MaskVT), Mask, VL});
5688     return DAG.getNode(RISCVISD::VSELECT_VL, DL, VT, SelectCond, SplattedVal,
5689                        Vec, VL);
5690   }
5691   }
5692 
5693   return lowerVectorIntrinsicScalars(Op, DAG, Subtarget);
5694 }
5695 
LowerINTRINSIC_W_CHAIN(SDValue Op,SelectionDAG & DAG) const5696 SDValue RISCVTargetLowering::LowerINTRINSIC_W_CHAIN(SDValue Op,
5697                                                     SelectionDAG &DAG) const {
5698   unsigned IntNo = Op.getConstantOperandVal(1);
5699   switch (IntNo) {
5700   default:
5701     break;
5702   case Intrinsic::riscv_masked_strided_load: {
5703     SDLoc DL(Op);
5704     MVT XLenVT = Subtarget.getXLenVT();
5705 
5706     // If the mask is known to be all ones, optimize to an unmasked intrinsic;
5707     // the selection of the masked intrinsics doesn't do this for us.
5708     SDValue Mask = Op.getOperand(5);
5709     bool IsUnmasked = ISD::isConstantSplatVectorAllOnes(Mask.getNode());
5710 
5711     MVT VT = Op->getSimpleValueType(0);
5712     MVT ContainerVT = VT;
5713     if (VT.isFixedLengthVector())
5714       ContainerVT = getContainerForFixedLengthVector(VT);
5715 
5716     SDValue PassThru = Op.getOperand(2);
5717     if (!IsUnmasked) {
5718       MVT MaskVT = getMaskTypeFor(ContainerVT);
5719       if (VT.isFixedLengthVector()) {
5720         Mask = convertToScalableVector(MaskVT, Mask, DAG, Subtarget);
5721         PassThru = convertToScalableVector(ContainerVT, PassThru, DAG, Subtarget);
5722       }
5723     }
5724 
5725     auto *Load = cast<MemIntrinsicSDNode>(Op);
5726     SDValue VL = getDefaultVLOps(VT, ContainerVT, DL, DAG, Subtarget).second;
5727     SDValue Ptr = Op.getOperand(3);
5728     SDValue Stride = Op.getOperand(4);
5729     SDValue Result, Chain;
5730 
5731     // TODO: We restrict this to unmasked loads currently in consideration of
5732     // the complexity of hanlding all falses masks.
5733     if (IsUnmasked && isNullConstant(Stride)) {
5734       MVT ScalarVT = ContainerVT.getVectorElementType();
5735       SDValue ScalarLoad =
5736           DAG.getExtLoad(ISD::ZEXTLOAD, DL, XLenVT, Load->getChain(), Ptr,
5737                          ScalarVT, Load->getMemOperand());
5738       Chain = ScalarLoad.getValue(1);
5739       Result = lowerScalarSplat(SDValue(), ScalarLoad, VL, ContainerVT, DL, DAG,
5740                                 Subtarget);
5741     } else {
5742       SDValue IntID = DAG.getTargetConstant(
5743           IsUnmasked ? Intrinsic::riscv_vlse : Intrinsic::riscv_vlse_mask, DL,
5744           XLenVT);
5745 
5746       SmallVector<SDValue, 8> Ops{Load->getChain(), IntID};
5747       if (IsUnmasked)
5748         Ops.push_back(DAG.getUNDEF(ContainerVT));
5749       else
5750         Ops.push_back(PassThru);
5751       Ops.push_back(Ptr);
5752       Ops.push_back(Stride);
5753       if (!IsUnmasked)
5754         Ops.push_back(Mask);
5755       Ops.push_back(VL);
5756       if (!IsUnmasked) {
5757         SDValue Policy =
5758             DAG.getTargetConstant(RISCVII::TAIL_AGNOSTIC, DL, XLenVT);
5759         Ops.push_back(Policy);
5760       }
5761 
5762       SDVTList VTs = DAG.getVTList({ContainerVT, MVT::Other});
5763       Result =
5764           DAG.getMemIntrinsicNode(ISD::INTRINSIC_W_CHAIN, DL, VTs, Ops,
5765                                   Load->getMemoryVT(), Load->getMemOperand());
5766       Chain = Result.getValue(1);
5767     }
5768     if (VT.isFixedLengthVector())
5769       Result = convertFromScalableVector(VT, Result, DAG, Subtarget);
5770     return DAG.getMergeValues({Result, Chain}, DL);
5771   }
5772   case Intrinsic::riscv_seg2_load:
5773   case Intrinsic::riscv_seg3_load:
5774   case Intrinsic::riscv_seg4_load:
5775   case Intrinsic::riscv_seg5_load:
5776   case Intrinsic::riscv_seg6_load:
5777   case Intrinsic::riscv_seg7_load:
5778   case Intrinsic::riscv_seg8_load: {
5779     SDLoc DL(Op);
5780     static const Intrinsic::ID VlsegInts[7] = {
5781         Intrinsic::riscv_vlseg2, Intrinsic::riscv_vlseg3,
5782         Intrinsic::riscv_vlseg4, Intrinsic::riscv_vlseg5,
5783         Intrinsic::riscv_vlseg6, Intrinsic::riscv_vlseg7,
5784         Intrinsic::riscv_vlseg8};
5785     unsigned NF = Op->getNumValues() - 1;
5786     assert(NF >= 2 && NF <= 8 && "Unexpected seg number");
5787     MVT XLenVT = Subtarget.getXLenVT();
5788     MVT VT = Op->getSimpleValueType(0);
5789     MVT ContainerVT = getContainerForFixedLengthVector(VT);
5790 
5791     SDValue VL = getVLOp(VT.getVectorNumElements(), DL, DAG, Subtarget);
5792     SDValue IntID = DAG.getTargetConstant(VlsegInts[NF - 2], DL, XLenVT);
5793     auto *Load = cast<MemIntrinsicSDNode>(Op);
5794     SmallVector<EVT, 9> ContainerVTs(NF, ContainerVT);
5795     ContainerVTs.push_back(MVT::Other);
5796     SDVTList VTs = DAG.getVTList(ContainerVTs);
5797     SmallVector<SDValue, 12> Ops = {Load->getChain(), IntID};
5798     Ops.insert(Ops.end(), NF, DAG.getUNDEF(ContainerVT));
5799     Ops.push_back(Op.getOperand(2));
5800     Ops.push_back(VL);
5801     SDValue Result =
5802         DAG.getMemIntrinsicNode(ISD::INTRINSIC_W_CHAIN, DL, VTs, Ops,
5803                                 Load->getMemoryVT(), Load->getMemOperand());
5804     SmallVector<SDValue, 9> Results;
5805     for (unsigned int RetIdx = 0; RetIdx < NF; RetIdx++)
5806       Results.push_back(convertFromScalableVector(VT, Result.getValue(RetIdx),
5807                                                   DAG, Subtarget));
5808     Results.push_back(Result.getValue(NF));
5809     return DAG.getMergeValues(Results, DL);
5810   }
5811   }
5812 
5813   return lowerVectorIntrinsicScalars(Op, DAG, Subtarget);
5814 }
5815 
LowerINTRINSIC_VOID(SDValue Op,SelectionDAG & DAG) const5816 SDValue RISCVTargetLowering::LowerINTRINSIC_VOID(SDValue Op,
5817                                                  SelectionDAG &DAG) const {
5818   unsigned IntNo = Op.getConstantOperandVal(1);
5819   switch (IntNo) {
5820   default:
5821     break;
5822   case Intrinsic::riscv_masked_strided_store: {
5823     SDLoc DL(Op);
5824     MVT XLenVT = Subtarget.getXLenVT();
5825 
5826     // If the mask is known to be all ones, optimize to an unmasked intrinsic;
5827     // the selection of the masked intrinsics doesn't do this for us.
5828     SDValue Mask = Op.getOperand(5);
5829     bool IsUnmasked = ISD::isConstantSplatVectorAllOnes(Mask.getNode());
5830 
5831     SDValue Val = Op.getOperand(2);
5832     MVT VT = Val.getSimpleValueType();
5833     MVT ContainerVT = VT;
5834     if (VT.isFixedLengthVector()) {
5835       ContainerVT = getContainerForFixedLengthVector(VT);
5836       Val = convertToScalableVector(ContainerVT, Val, DAG, Subtarget);
5837     }
5838     if (!IsUnmasked) {
5839       MVT MaskVT = getMaskTypeFor(ContainerVT);
5840       if (VT.isFixedLengthVector())
5841         Mask = convertToScalableVector(MaskVT, Mask, DAG, Subtarget);
5842     }
5843 
5844     SDValue VL = getDefaultVLOps(VT, ContainerVT, DL, DAG, Subtarget).second;
5845 
5846     SDValue IntID = DAG.getTargetConstant(
5847         IsUnmasked ? Intrinsic::riscv_vsse : Intrinsic::riscv_vsse_mask, DL,
5848         XLenVT);
5849 
5850     auto *Store = cast<MemIntrinsicSDNode>(Op);
5851     SmallVector<SDValue, 8> Ops{Store->getChain(), IntID};
5852     Ops.push_back(Val);
5853     Ops.push_back(Op.getOperand(3)); // Ptr
5854     Ops.push_back(Op.getOperand(4)); // Stride
5855     if (!IsUnmasked)
5856       Ops.push_back(Mask);
5857     Ops.push_back(VL);
5858 
5859     return DAG.getMemIntrinsicNode(ISD::INTRINSIC_VOID, DL, Store->getVTList(),
5860                                    Ops, Store->getMemoryVT(),
5861                                    Store->getMemOperand());
5862   }
5863   }
5864 
5865   return SDValue();
5866 }
5867 
getRVVReductionOp(unsigned ISDOpcode)5868 static unsigned getRVVReductionOp(unsigned ISDOpcode) {
5869   switch (ISDOpcode) {
5870   default:
5871     llvm_unreachable("Unhandled reduction");
5872   case ISD::VECREDUCE_ADD:
5873     return RISCVISD::VECREDUCE_ADD_VL;
5874   case ISD::VECREDUCE_UMAX:
5875     return RISCVISD::VECREDUCE_UMAX_VL;
5876   case ISD::VECREDUCE_SMAX:
5877     return RISCVISD::VECREDUCE_SMAX_VL;
5878   case ISD::VECREDUCE_UMIN:
5879     return RISCVISD::VECREDUCE_UMIN_VL;
5880   case ISD::VECREDUCE_SMIN:
5881     return RISCVISD::VECREDUCE_SMIN_VL;
5882   case ISD::VECREDUCE_AND:
5883     return RISCVISD::VECREDUCE_AND_VL;
5884   case ISD::VECREDUCE_OR:
5885     return RISCVISD::VECREDUCE_OR_VL;
5886   case ISD::VECREDUCE_XOR:
5887     return RISCVISD::VECREDUCE_XOR_VL;
5888   }
5889 }
5890 
lowerVectorMaskVecReduction(SDValue Op,SelectionDAG & DAG,bool IsVP) const5891 SDValue RISCVTargetLowering::lowerVectorMaskVecReduction(SDValue Op,
5892                                                          SelectionDAG &DAG,
5893                                                          bool IsVP) const {
5894   SDLoc DL(Op);
5895   SDValue Vec = Op.getOperand(IsVP ? 1 : 0);
5896   MVT VecVT = Vec.getSimpleValueType();
5897   assert((Op.getOpcode() == ISD::VECREDUCE_AND ||
5898           Op.getOpcode() == ISD::VECREDUCE_OR ||
5899           Op.getOpcode() == ISD::VECREDUCE_XOR ||
5900           Op.getOpcode() == ISD::VP_REDUCE_AND ||
5901           Op.getOpcode() == ISD::VP_REDUCE_OR ||
5902           Op.getOpcode() == ISD::VP_REDUCE_XOR) &&
5903          "Unexpected reduction lowering");
5904 
5905   MVT XLenVT = Subtarget.getXLenVT();
5906   assert(Op.getValueType() == XLenVT &&
5907          "Expected reduction output to be legalized to XLenVT");
5908 
5909   MVT ContainerVT = VecVT;
5910   if (VecVT.isFixedLengthVector()) {
5911     ContainerVT = getContainerForFixedLengthVector(VecVT);
5912     Vec = convertToScalableVector(ContainerVT, Vec, DAG, Subtarget);
5913   }
5914 
5915   SDValue Mask, VL;
5916   if (IsVP) {
5917     Mask = Op.getOperand(2);
5918     VL = Op.getOperand(3);
5919   } else {
5920     std::tie(Mask, VL) =
5921         getDefaultVLOps(VecVT, ContainerVT, DL, DAG, Subtarget);
5922   }
5923 
5924   unsigned BaseOpc;
5925   ISD::CondCode CC;
5926   SDValue Zero = DAG.getConstant(0, DL, XLenVT);
5927 
5928   switch (Op.getOpcode()) {
5929   default:
5930     llvm_unreachable("Unhandled reduction");
5931   case ISD::VECREDUCE_AND:
5932   case ISD::VP_REDUCE_AND: {
5933     // vcpop ~x == 0
5934     SDValue TrueMask = DAG.getNode(RISCVISD::VMSET_VL, DL, ContainerVT, VL);
5935     Vec = DAG.getNode(RISCVISD::VMXOR_VL, DL, ContainerVT, Vec, TrueMask, VL);
5936     Vec = DAG.getNode(RISCVISD::VCPOP_VL, DL, XLenVT, Vec, Mask, VL);
5937     CC = ISD::SETEQ;
5938     BaseOpc = ISD::AND;
5939     break;
5940   }
5941   case ISD::VECREDUCE_OR:
5942   case ISD::VP_REDUCE_OR:
5943     // vcpop x != 0
5944     Vec = DAG.getNode(RISCVISD::VCPOP_VL, DL, XLenVT, Vec, Mask, VL);
5945     CC = ISD::SETNE;
5946     BaseOpc = ISD::OR;
5947     break;
5948   case ISD::VECREDUCE_XOR:
5949   case ISD::VP_REDUCE_XOR: {
5950     // ((vcpop x) & 1) != 0
5951     SDValue One = DAG.getConstant(1, DL, XLenVT);
5952     Vec = DAG.getNode(RISCVISD::VCPOP_VL, DL, XLenVT, Vec, Mask, VL);
5953     Vec = DAG.getNode(ISD::AND, DL, XLenVT, Vec, One);
5954     CC = ISD::SETNE;
5955     BaseOpc = ISD::XOR;
5956     break;
5957   }
5958   }
5959 
5960   SDValue SetCC = DAG.getSetCC(DL, XLenVT, Vec, Zero, CC);
5961 
5962   if (!IsVP)
5963     return SetCC;
5964 
5965   // Now include the start value in the operation.
5966   // Note that we must return the start value when no elements are operated
5967   // upon. The vcpop instructions we've emitted in each case above will return
5968   // 0 for an inactive vector, and so we've already received the neutral value:
5969   // AND gives us (0 == 0) -> 1 and OR/XOR give us (0 != 0) -> 0. Therefore we
5970   // can simply include the start value.
5971   return DAG.getNode(BaseOpc, DL, XLenVT, SetCC, Op.getOperand(0));
5972 }
5973 
hasNonZeroAVL(SDValue AVL)5974 static bool hasNonZeroAVL(SDValue AVL) {
5975   auto *RegisterAVL = dyn_cast<RegisterSDNode>(AVL);
5976   auto *ImmAVL = dyn_cast<ConstantSDNode>(AVL);
5977   return (RegisterAVL && RegisterAVL->getReg() == RISCV::X0) ||
5978          (ImmAVL && ImmAVL->getZExtValue() >= 1);
5979 }
5980 
5981 /// Helper to lower a reduction sequence of the form:
5982 /// scalar = reduce_op vec, scalar_start
lowerReductionSeq(unsigned RVVOpcode,MVT ResVT,SDValue StartValue,SDValue Vec,SDValue Mask,SDValue VL,SDLoc DL,SelectionDAG & DAG,const RISCVSubtarget & Subtarget)5983 static SDValue lowerReductionSeq(unsigned RVVOpcode, MVT ResVT,
5984                                  SDValue StartValue, SDValue Vec, SDValue Mask,
5985                                  SDValue VL, SDLoc DL, SelectionDAG &DAG,
5986                                  const RISCVSubtarget &Subtarget) {
5987   const MVT VecVT = Vec.getSimpleValueType();
5988   const MVT M1VT = getLMUL1VT(VecVT);
5989   const MVT XLenVT = Subtarget.getXLenVT();
5990   const bool NonZeroAVL = hasNonZeroAVL(VL);
5991 
5992   // The reduction needs an LMUL1 input; do the splat at either LMUL1
5993   // or the original VT if fractional.
5994   auto InnerVT = VecVT.bitsLE(M1VT) ? VecVT : M1VT;
5995   // We reuse the VL of the reduction to reduce vsetvli toggles if we can
5996   // prove it is non-zero.  For the AVL=0 case, we need the scalar to
5997   // be the result of the reduction operation.
5998   auto InnerVL = NonZeroAVL ? VL : DAG.getConstant(1, DL, XLenVT);
5999   SDValue InitialValue = lowerScalarInsert(StartValue, InnerVL, InnerVT, DL,
6000                                            DAG, Subtarget);
6001   if (M1VT != InnerVT)
6002     InitialValue = DAG.getNode(ISD::INSERT_SUBVECTOR, DL, M1VT,
6003                                DAG.getUNDEF(M1VT),
6004                                InitialValue, DAG.getConstant(0, DL, XLenVT));
6005   SDValue PassThru = NonZeroAVL ? DAG.getUNDEF(M1VT) : InitialValue;
6006   SDValue Reduction = DAG.getNode(RVVOpcode, DL, M1VT, PassThru, Vec,
6007                                   InitialValue, Mask, VL);
6008   return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, ResVT, Reduction,
6009                      DAG.getConstant(0, DL, XLenVT));
6010 }
6011 
lowerVECREDUCE(SDValue Op,SelectionDAG & DAG) const6012 SDValue RISCVTargetLowering::lowerVECREDUCE(SDValue Op,
6013                                             SelectionDAG &DAG) const {
6014   SDLoc DL(Op);
6015   SDValue Vec = Op.getOperand(0);
6016   EVT VecEVT = Vec.getValueType();
6017 
6018   unsigned BaseOpc = ISD::getVecReduceBaseOpcode(Op.getOpcode());
6019 
6020   // Due to ordering in legalize types we may have a vector type that needs to
6021   // be split. Do that manually so we can get down to a legal type.
6022   while (getTypeAction(*DAG.getContext(), VecEVT) ==
6023          TargetLowering::TypeSplitVector) {
6024     auto [Lo, Hi] = DAG.SplitVector(Vec, DL);
6025     VecEVT = Lo.getValueType();
6026     Vec = DAG.getNode(BaseOpc, DL, VecEVT, Lo, Hi);
6027   }
6028 
6029   // TODO: The type may need to be widened rather than split. Or widened before
6030   // it can be split.
6031   if (!isTypeLegal(VecEVT))
6032     return SDValue();
6033 
6034   MVT VecVT = VecEVT.getSimpleVT();
6035   MVT VecEltVT = VecVT.getVectorElementType();
6036   unsigned RVVOpcode = getRVVReductionOp(Op.getOpcode());
6037 
6038   MVT ContainerVT = VecVT;
6039   if (VecVT.isFixedLengthVector()) {
6040     ContainerVT = getContainerForFixedLengthVector(VecVT);
6041     Vec = convertToScalableVector(ContainerVT, Vec, DAG, Subtarget);
6042   }
6043 
6044   auto [Mask, VL] = getDefaultVLOps(VecVT, ContainerVT, DL, DAG, Subtarget);
6045 
6046   SDValue NeutralElem =
6047       DAG.getNeutralElement(BaseOpc, DL, VecEltVT, SDNodeFlags());
6048   return lowerReductionSeq(RVVOpcode, Op.getSimpleValueType(), NeutralElem, Vec,
6049                            Mask, VL, DL, DAG, Subtarget);
6050 }
6051 
6052 // Given a reduction op, this function returns the matching reduction opcode,
6053 // the vector SDValue and the scalar SDValue required to lower this to a
6054 // RISCVISD node.
6055 static std::tuple<unsigned, SDValue, SDValue>
getRVVFPReductionOpAndOperands(SDValue Op,SelectionDAG & DAG,EVT EltVT)6056 getRVVFPReductionOpAndOperands(SDValue Op, SelectionDAG &DAG, EVT EltVT) {
6057   SDLoc DL(Op);
6058   auto Flags = Op->getFlags();
6059   unsigned Opcode = Op.getOpcode();
6060   unsigned BaseOpcode = ISD::getVecReduceBaseOpcode(Opcode);
6061   switch (Opcode) {
6062   default:
6063     llvm_unreachable("Unhandled reduction");
6064   case ISD::VECREDUCE_FADD: {
6065     // Use positive zero if we can. It is cheaper to materialize.
6066     SDValue Zero =
6067         DAG.getConstantFP(Flags.hasNoSignedZeros() ? 0.0 : -0.0, DL, EltVT);
6068     return std::make_tuple(RISCVISD::VECREDUCE_FADD_VL, Op.getOperand(0), Zero);
6069   }
6070   case ISD::VECREDUCE_SEQ_FADD:
6071     return std::make_tuple(RISCVISD::VECREDUCE_SEQ_FADD_VL, Op.getOperand(1),
6072                            Op.getOperand(0));
6073   case ISD::VECREDUCE_FMIN:
6074     return std::make_tuple(RISCVISD::VECREDUCE_FMIN_VL, Op.getOperand(0),
6075                            DAG.getNeutralElement(BaseOpcode, DL, EltVT, Flags));
6076   case ISD::VECREDUCE_FMAX:
6077     return std::make_tuple(RISCVISD::VECREDUCE_FMAX_VL, Op.getOperand(0),
6078                            DAG.getNeutralElement(BaseOpcode, DL, EltVT, Flags));
6079   }
6080 }
6081 
lowerFPVECREDUCE(SDValue Op,SelectionDAG & DAG) const6082 SDValue RISCVTargetLowering::lowerFPVECREDUCE(SDValue Op,
6083                                               SelectionDAG &DAG) const {
6084   SDLoc DL(Op);
6085   MVT VecEltVT = Op.getSimpleValueType();
6086 
6087   unsigned RVVOpcode;
6088   SDValue VectorVal, ScalarVal;
6089   std::tie(RVVOpcode, VectorVal, ScalarVal) =
6090       getRVVFPReductionOpAndOperands(Op, DAG, VecEltVT);
6091   MVT VecVT = VectorVal.getSimpleValueType();
6092 
6093   MVT ContainerVT = VecVT;
6094   if (VecVT.isFixedLengthVector()) {
6095     ContainerVT = getContainerForFixedLengthVector(VecVT);
6096     VectorVal = convertToScalableVector(ContainerVT, VectorVal, DAG, Subtarget);
6097   }
6098 
6099   auto [Mask, VL] = getDefaultVLOps(VecVT, ContainerVT, DL, DAG, Subtarget);
6100   return lowerReductionSeq(RVVOpcode, Op.getSimpleValueType(), ScalarVal,
6101                            VectorVal, Mask, VL, DL, DAG, Subtarget);
6102 }
6103 
getRVVVPReductionOp(unsigned ISDOpcode)6104 static unsigned getRVVVPReductionOp(unsigned ISDOpcode) {
6105   switch (ISDOpcode) {
6106   default:
6107     llvm_unreachable("Unhandled reduction");
6108   case ISD::VP_REDUCE_ADD:
6109     return RISCVISD::VECREDUCE_ADD_VL;
6110   case ISD::VP_REDUCE_UMAX:
6111     return RISCVISD::VECREDUCE_UMAX_VL;
6112   case ISD::VP_REDUCE_SMAX:
6113     return RISCVISD::VECREDUCE_SMAX_VL;
6114   case ISD::VP_REDUCE_UMIN:
6115     return RISCVISD::VECREDUCE_UMIN_VL;
6116   case ISD::VP_REDUCE_SMIN:
6117     return RISCVISD::VECREDUCE_SMIN_VL;
6118   case ISD::VP_REDUCE_AND:
6119     return RISCVISD::VECREDUCE_AND_VL;
6120   case ISD::VP_REDUCE_OR:
6121     return RISCVISD::VECREDUCE_OR_VL;
6122   case ISD::VP_REDUCE_XOR:
6123     return RISCVISD::VECREDUCE_XOR_VL;
6124   case ISD::VP_REDUCE_FADD:
6125     return RISCVISD::VECREDUCE_FADD_VL;
6126   case ISD::VP_REDUCE_SEQ_FADD:
6127     return RISCVISD::VECREDUCE_SEQ_FADD_VL;
6128   case ISD::VP_REDUCE_FMAX:
6129     return RISCVISD::VECREDUCE_FMAX_VL;
6130   case ISD::VP_REDUCE_FMIN:
6131     return RISCVISD::VECREDUCE_FMIN_VL;
6132   }
6133 }
6134 
lowerVPREDUCE(SDValue Op,SelectionDAG & DAG) const6135 SDValue RISCVTargetLowering::lowerVPREDUCE(SDValue Op,
6136                                            SelectionDAG &DAG) const {
6137   SDLoc DL(Op);
6138   SDValue Vec = Op.getOperand(1);
6139   EVT VecEVT = Vec.getValueType();
6140 
6141   // TODO: The type may need to be widened rather than split. Or widened before
6142   // it can be split.
6143   if (!isTypeLegal(VecEVT))
6144     return SDValue();
6145 
6146   MVT VecVT = VecEVT.getSimpleVT();
6147   unsigned RVVOpcode = getRVVVPReductionOp(Op.getOpcode());
6148 
6149   if (VecVT.isFixedLengthVector()) {
6150     auto ContainerVT = getContainerForFixedLengthVector(VecVT);
6151     Vec = convertToScalableVector(ContainerVT, Vec, DAG, Subtarget);
6152   }
6153 
6154   SDValue VL = Op.getOperand(3);
6155   SDValue Mask = Op.getOperand(2);
6156   return lowerReductionSeq(RVVOpcode, Op.getSimpleValueType(), Op.getOperand(0),
6157                            Vec, Mask, VL, DL, DAG, Subtarget);
6158 }
6159 
lowerINSERT_SUBVECTOR(SDValue Op,SelectionDAG & DAG) const6160 SDValue RISCVTargetLowering::lowerINSERT_SUBVECTOR(SDValue Op,
6161                                                    SelectionDAG &DAG) const {
6162   SDValue Vec = Op.getOperand(0);
6163   SDValue SubVec = Op.getOperand(1);
6164   MVT VecVT = Vec.getSimpleValueType();
6165   MVT SubVecVT = SubVec.getSimpleValueType();
6166 
6167   SDLoc DL(Op);
6168   MVT XLenVT = Subtarget.getXLenVT();
6169   unsigned OrigIdx = Op.getConstantOperandVal(2);
6170   const RISCVRegisterInfo *TRI = Subtarget.getRegisterInfo();
6171 
6172   // We don't have the ability to slide mask vectors up indexed by their i1
6173   // elements; the smallest we can do is i8. Often we are able to bitcast to
6174   // equivalent i8 vectors. Note that when inserting a fixed-length vector
6175   // into a scalable one, we might not necessarily have enough scalable
6176   // elements to safely divide by 8: nxv1i1 = insert nxv1i1, v4i1 is valid.
6177   if (SubVecVT.getVectorElementType() == MVT::i1 &&
6178       (OrigIdx != 0 || !Vec.isUndef())) {
6179     if (VecVT.getVectorMinNumElements() >= 8 &&
6180         SubVecVT.getVectorMinNumElements() >= 8) {
6181       assert(OrigIdx % 8 == 0 && "Invalid index");
6182       assert(VecVT.getVectorMinNumElements() % 8 == 0 &&
6183              SubVecVT.getVectorMinNumElements() % 8 == 0 &&
6184              "Unexpected mask vector lowering");
6185       OrigIdx /= 8;
6186       SubVecVT =
6187           MVT::getVectorVT(MVT::i8, SubVecVT.getVectorMinNumElements() / 8,
6188                            SubVecVT.isScalableVector());
6189       VecVT = MVT::getVectorVT(MVT::i8, VecVT.getVectorMinNumElements() / 8,
6190                                VecVT.isScalableVector());
6191       Vec = DAG.getBitcast(VecVT, Vec);
6192       SubVec = DAG.getBitcast(SubVecVT, SubVec);
6193     } else {
6194       // We can't slide this mask vector up indexed by its i1 elements.
6195       // This poses a problem when we wish to insert a scalable vector which
6196       // can't be re-expressed as a larger type. Just choose the slow path and
6197       // extend to a larger type, then truncate back down.
6198       MVT ExtVecVT = VecVT.changeVectorElementType(MVT::i8);
6199       MVT ExtSubVecVT = SubVecVT.changeVectorElementType(MVT::i8);
6200       Vec = DAG.getNode(ISD::ZERO_EXTEND, DL, ExtVecVT, Vec);
6201       SubVec = DAG.getNode(ISD::ZERO_EXTEND, DL, ExtSubVecVT, SubVec);
6202       Vec = DAG.getNode(ISD::INSERT_SUBVECTOR, DL, ExtVecVT, Vec, SubVec,
6203                         Op.getOperand(2));
6204       SDValue SplatZero = DAG.getConstant(0, DL, ExtVecVT);
6205       return DAG.getSetCC(DL, VecVT, Vec, SplatZero, ISD::SETNE);
6206     }
6207   }
6208 
6209   // If the subvector vector is a fixed-length type, we cannot use subregister
6210   // manipulation to simplify the codegen; we don't know which register of a
6211   // LMUL group contains the specific subvector as we only know the minimum
6212   // register size. Therefore we must slide the vector group up the full
6213   // amount.
6214   if (SubVecVT.isFixedLengthVector()) {
6215     if (OrigIdx == 0 && Vec.isUndef() && !VecVT.isFixedLengthVector())
6216       return Op;
6217     MVT ContainerVT = VecVT;
6218     if (VecVT.isFixedLengthVector()) {
6219       ContainerVT = getContainerForFixedLengthVector(VecVT);
6220       Vec = convertToScalableVector(ContainerVT, Vec, DAG, Subtarget);
6221     }
6222     SubVec = DAG.getNode(ISD::INSERT_SUBVECTOR, DL, ContainerVT,
6223                          DAG.getUNDEF(ContainerVT), SubVec,
6224                          DAG.getConstant(0, DL, XLenVT));
6225     if (OrigIdx == 0 && Vec.isUndef() && VecVT.isFixedLengthVector()) {
6226       SubVec = convertFromScalableVector(VecVT, SubVec, DAG, Subtarget);
6227       return DAG.getBitcast(Op.getValueType(), SubVec);
6228     }
6229     SDValue Mask =
6230         getDefaultVLOps(VecVT, ContainerVT, DL, DAG, Subtarget).first;
6231     // Set the vector length to only the number of elements we care about. Note
6232     // that for slideup this includes the offset.
6233     SDValue VL =
6234         getVLOp(OrigIdx + SubVecVT.getVectorNumElements(), DL, DAG, Subtarget);
6235     SDValue SlideupAmt = DAG.getConstant(OrigIdx, DL, XLenVT);
6236 
6237     // Use tail agnostic policy if OrigIdx is the last index of Vec.
6238     unsigned Policy = RISCVII::TAIL_UNDISTURBED_MASK_UNDISTURBED;
6239     if (VecVT.isFixedLengthVector() &&
6240         OrigIdx + 1 == VecVT.getVectorNumElements())
6241       Policy = RISCVII::TAIL_AGNOSTIC;
6242     SDValue Slideup = getVSlideup(DAG, Subtarget, DL, ContainerVT, Vec, SubVec,
6243                                   SlideupAmt, Mask, VL, Policy);
6244     if (VecVT.isFixedLengthVector())
6245       Slideup = convertFromScalableVector(VecVT, Slideup, DAG, Subtarget);
6246     return DAG.getBitcast(Op.getValueType(), Slideup);
6247   }
6248 
6249   unsigned SubRegIdx, RemIdx;
6250   std::tie(SubRegIdx, RemIdx) =
6251       RISCVTargetLowering::decomposeSubvectorInsertExtractToSubRegs(
6252           VecVT, SubVecVT, OrigIdx, TRI);
6253 
6254   RISCVII::VLMUL SubVecLMUL = RISCVTargetLowering::getLMUL(SubVecVT);
6255   bool IsSubVecPartReg = SubVecLMUL == RISCVII::VLMUL::LMUL_F2 ||
6256                          SubVecLMUL == RISCVII::VLMUL::LMUL_F4 ||
6257                          SubVecLMUL == RISCVII::VLMUL::LMUL_F8;
6258 
6259   // 1. If the Idx has been completely eliminated and this subvector's size is
6260   // a vector register or a multiple thereof, or the surrounding elements are
6261   // undef, then this is a subvector insert which naturally aligns to a vector
6262   // register. These can easily be handled using subregister manipulation.
6263   // 2. If the subvector is smaller than a vector register, then the insertion
6264   // must preserve the undisturbed elements of the register. We do this by
6265   // lowering to an EXTRACT_SUBVECTOR grabbing the nearest LMUL=1 vector type
6266   // (which resolves to a subregister copy), performing a VSLIDEUP to place the
6267   // subvector within the vector register, and an INSERT_SUBVECTOR of that
6268   // LMUL=1 type back into the larger vector (resolving to another subregister
6269   // operation). See below for how our VSLIDEUP works. We go via a LMUL=1 type
6270   // to avoid allocating a large register group to hold our subvector.
6271   if (RemIdx == 0 && (!IsSubVecPartReg || Vec.isUndef()))
6272     return Op;
6273 
6274   // VSLIDEUP works by leaving elements 0<i<OFFSET undisturbed, elements
6275   // OFFSET<=i<VL set to the "subvector" and vl<=i<VLMAX set to the tail policy
6276   // (in our case undisturbed). This means we can set up a subvector insertion
6277   // where OFFSET is the insertion offset, and the VL is the OFFSET plus the
6278   // size of the subvector.
6279   MVT InterSubVT = VecVT;
6280   SDValue AlignedExtract = Vec;
6281   unsigned AlignedIdx = OrigIdx - RemIdx;
6282   if (VecVT.bitsGT(getLMUL1VT(VecVT))) {
6283     InterSubVT = getLMUL1VT(VecVT);
6284     // Extract a subvector equal to the nearest full vector register type. This
6285     // should resolve to a EXTRACT_SUBREG instruction.
6286     AlignedExtract = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, InterSubVT, Vec,
6287                                  DAG.getConstant(AlignedIdx, DL, XLenVT));
6288   }
6289 
6290   SDValue SlideupAmt = DAG.getConstant(RemIdx, DL, XLenVT);
6291   // For scalable vectors this must be further multiplied by vscale.
6292   SlideupAmt = DAG.getNode(ISD::VSCALE, DL, XLenVT, SlideupAmt);
6293 
6294   auto [Mask, VL] = getDefaultScalableVLOps(VecVT, DL, DAG, Subtarget);
6295 
6296   // Construct the vector length corresponding to RemIdx + length(SubVecVT).
6297   VL = DAG.getConstant(SubVecVT.getVectorMinNumElements(), DL, XLenVT);
6298   VL = DAG.getNode(ISD::VSCALE, DL, XLenVT, VL);
6299   VL = DAG.getNode(ISD::ADD, DL, XLenVT, SlideupAmt, VL);
6300 
6301   SubVec = DAG.getNode(ISD::INSERT_SUBVECTOR, DL, InterSubVT,
6302                        DAG.getUNDEF(InterSubVT), SubVec,
6303                        DAG.getConstant(0, DL, XLenVT));
6304 
6305   SDValue Slideup = getVSlideup(DAG, Subtarget, DL, InterSubVT, AlignedExtract,
6306                                 SubVec, SlideupAmt, Mask, VL);
6307 
6308   // If required, insert this subvector back into the correct vector register.
6309   // This should resolve to an INSERT_SUBREG instruction.
6310   if (VecVT.bitsGT(InterSubVT))
6311     Slideup = DAG.getNode(ISD::INSERT_SUBVECTOR, DL, VecVT, Vec, Slideup,
6312                           DAG.getConstant(AlignedIdx, DL, XLenVT));
6313 
6314   // We might have bitcast from a mask type: cast back to the original type if
6315   // required.
6316   return DAG.getBitcast(Op.getSimpleValueType(), Slideup);
6317 }
6318 
lowerEXTRACT_SUBVECTOR(SDValue Op,SelectionDAG & DAG) const6319 SDValue RISCVTargetLowering::lowerEXTRACT_SUBVECTOR(SDValue Op,
6320                                                     SelectionDAG &DAG) const {
6321   SDValue Vec = Op.getOperand(0);
6322   MVT SubVecVT = Op.getSimpleValueType();
6323   MVT VecVT = Vec.getSimpleValueType();
6324 
6325   SDLoc DL(Op);
6326   MVT XLenVT = Subtarget.getXLenVT();
6327   unsigned OrigIdx = Op.getConstantOperandVal(1);
6328   const RISCVRegisterInfo *TRI = Subtarget.getRegisterInfo();
6329 
6330   // We don't have the ability to slide mask vectors down indexed by their i1
6331   // elements; the smallest we can do is i8. Often we are able to bitcast to
6332   // equivalent i8 vectors. Note that when extracting a fixed-length vector
6333   // from a scalable one, we might not necessarily have enough scalable
6334   // elements to safely divide by 8: v8i1 = extract nxv1i1 is valid.
6335   if (SubVecVT.getVectorElementType() == MVT::i1 && OrigIdx != 0) {
6336     if (VecVT.getVectorMinNumElements() >= 8 &&
6337         SubVecVT.getVectorMinNumElements() >= 8) {
6338       assert(OrigIdx % 8 == 0 && "Invalid index");
6339       assert(VecVT.getVectorMinNumElements() % 8 == 0 &&
6340              SubVecVT.getVectorMinNumElements() % 8 == 0 &&
6341              "Unexpected mask vector lowering");
6342       OrigIdx /= 8;
6343       SubVecVT =
6344           MVT::getVectorVT(MVT::i8, SubVecVT.getVectorMinNumElements() / 8,
6345                            SubVecVT.isScalableVector());
6346       VecVT = MVT::getVectorVT(MVT::i8, VecVT.getVectorMinNumElements() / 8,
6347                                VecVT.isScalableVector());
6348       Vec = DAG.getBitcast(VecVT, Vec);
6349     } else {
6350       // We can't slide this mask vector down, indexed by its i1 elements.
6351       // This poses a problem when we wish to extract a scalable vector which
6352       // can't be re-expressed as a larger type. Just choose the slow path and
6353       // extend to a larger type, then truncate back down.
6354       // TODO: We could probably improve this when extracting certain fixed
6355       // from fixed, where we can extract as i8 and shift the correct element
6356       // right to reach the desired subvector?
6357       MVT ExtVecVT = VecVT.changeVectorElementType(MVT::i8);
6358       MVT ExtSubVecVT = SubVecVT.changeVectorElementType(MVT::i8);
6359       Vec = DAG.getNode(ISD::ZERO_EXTEND, DL, ExtVecVT, Vec);
6360       Vec = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, ExtSubVecVT, Vec,
6361                         Op.getOperand(1));
6362       SDValue SplatZero = DAG.getConstant(0, DL, ExtSubVecVT);
6363       return DAG.getSetCC(DL, SubVecVT, Vec, SplatZero, ISD::SETNE);
6364     }
6365   }
6366 
6367   // If the subvector vector is a fixed-length type, we cannot use subregister
6368   // manipulation to simplify the codegen; we don't know which register of a
6369   // LMUL group contains the specific subvector as we only know the minimum
6370   // register size. Therefore we must slide the vector group down the full
6371   // amount.
6372   if (SubVecVT.isFixedLengthVector()) {
6373     // With an index of 0 this is a cast-like subvector, which can be performed
6374     // with subregister operations.
6375     if (OrigIdx == 0)
6376       return Op;
6377     MVT ContainerVT = VecVT;
6378     if (VecVT.isFixedLengthVector()) {
6379       ContainerVT = getContainerForFixedLengthVector(VecVT);
6380       Vec = convertToScalableVector(ContainerVT, Vec, DAG, Subtarget);
6381     }
6382     SDValue Mask =
6383         getDefaultVLOps(VecVT, ContainerVT, DL, DAG, Subtarget).first;
6384     // Set the vector length to only the number of elements we care about. This
6385     // avoids sliding down elements we're going to discard straight away.
6386     SDValue VL = getVLOp(SubVecVT.getVectorNumElements(), DL, DAG, Subtarget);
6387     SDValue SlidedownAmt = DAG.getConstant(OrigIdx, DL, XLenVT);
6388     SDValue Slidedown =
6389         getVSlidedown(DAG, Subtarget, DL, ContainerVT,
6390                       DAG.getUNDEF(ContainerVT), Vec, SlidedownAmt, Mask, VL);
6391     // Now we can use a cast-like subvector extract to get the result.
6392     Slidedown = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, SubVecVT, Slidedown,
6393                             DAG.getConstant(0, DL, XLenVT));
6394     return DAG.getBitcast(Op.getValueType(), Slidedown);
6395   }
6396 
6397   unsigned SubRegIdx, RemIdx;
6398   std::tie(SubRegIdx, RemIdx) =
6399       RISCVTargetLowering::decomposeSubvectorInsertExtractToSubRegs(
6400           VecVT, SubVecVT, OrigIdx, TRI);
6401 
6402   // If the Idx has been completely eliminated then this is a subvector extract
6403   // which naturally aligns to a vector register. These can easily be handled
6404   // using subregister manipulation.
6405   if (RemIdx == 0)
6406     return Op;
6407 
6408   // Else we must shift our vector register directly to extract the subvector.
6409   // Do this using VSLIDEDOWN.
6410 
6411   // If the vector type is an LMUL-group type, extract a subvector equal to the
6412   // nearest full vector register type. This should resolve to a EXTRACT_SUBREG
6413   // instruction.
6414   MVT InterSubVT = VecVT;
6415   if (VecVT.bitsGT(getLMUL1VT(VecVT))) {
6416     InterSubVT = getLMUL1VT(VecVT);
6417     Vec = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, InterSubVT, Vec,
6418                       DAG.getConstant(OrigIdx - RemIdx, DL, XLenVT));
6419   }
6420 
6421   // Slide this vector register down by the desired number of elements in order
6422   // to place the desired subvector starting at element 0.
6423   SDValue SlidedownAmt = DAG.getConstant(RemIdx, DL, XLenVT);
6424   // For scalable vectors this must be further multiplied by vscale.
6425   SlidedownAmt = DAG.getNode(ISD::VSCALE, DL, XLenVT, SlidedownAmt);
6426 
6427   auto [Mask, VL] = getDefaultScalableVLOps(InterSubVT, DL, DAG, Subtarget);
6428   SDValue Slidedown =
6429       getVSlidedown(DAG, Subtarget, DL, InterSubVT, DAG.getUNDEF(InterSubVT),
6430                     Vec, SlidedownAmt, Mask, VL);
6431 
6432   // Now the vector is in the right position, extract our final subvector. This
6433   // should resolve to a COPY.
6434   Slidedown = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, SubVecVT, Slidedown,
6435                           DAG.getConstant(0, DL, XLenVT));
6436 
6437   // We might have bitcast from a mask type: cast back to the original type if
6438   // required.
6439   return DAG.getBitcast(Op.getSimpleValueType(), Slidedown);
6440 }
6441 
6442 // Lower step_vector to the vid instruction. Any non-identity step value must
6443 // be accounted for my manual expansion.
lowerSTEP_VECTOR(SDValue Op,SelectionDAG & DAG) const6444 SDValue RISCVTargetLowering::lowerSTEP_VECTOR(SDValue Op,
6445                                               SelectionDAG &DAG) const {
6446   SDLoc DL(Op);
6447   MVT VT = Op.getSimpleValueType();
6448   assert(VT.isScalableVector() && "Expected scalable vector");
6449   MVT XLenVT = Subtarget.getXLenVT();
6450   auto [Mask, VL] = getDefaultScalableVLOps(VT, DL, DAG, Subtarget);
6451   SDValue StepVec = DAG.getNode(RISCVISD::VID_VL, DL, VT, Mask, VL);
6452   uint64_t StepValImm = Op.getConstantOperandVal(0);
6453   if (StepValImm != 1) {
6454     if (isPowerOf2_64(StepValImm)) {
6455       SDValue StepVal =
6456           DAG.getNode(RISCVISD::VMV_V_X_VL, DL, VT, DAG.getUNDEF(VT),
6457                       DAG.getConstant(Log2_64(StepValImm), DL, XLenVT), VL);
6458       StepVec = DAG.getNode(ISD::SHL, DL, VT, StepVec, StepVal);
6459     } else {
6460       SDValue StepVal = lowerScalarSplat(
6461           SDValue(), DAG.getConstant(StepValImm, DL, VT.getVectorElementType()),
6462           VL, VT, DL, DAG, Subtarget);
6463       StepVec = DAG.getNode(ISD::MUL, DL, VT, StepVec, StepVal);
6464     }
6465   }
6466   return StepVec;
6467 }
6468 
6469 // Implement vector_reverse using vrgather.vv with indices determined by
6470 // subtracting the id of each element from (VLMAX-1). This will convert
6471 // the indices like so:
6472 // (0, 1,..., VLMAX-2, VLMAX-1) -> (VLMAX-1, VLMAX-2,..., 1, 0).
6473 // TODO: This code assumes VLMAX <= 65536 for LMUL=8 SEW=16.
lowerVECTOR_REVERSE(SDValue Op,SelectionDAG & DAG) const6474 SDValue RISCVTargetLowering::lowerVECTOR_REVERSE(SDValue Op,
6475                                                  SelectionDAG &DAG) const {
6476   SDLoc DL(Op);
6477   MVT VecVT = Op.getSimpleValueType();
6478   if (VecVT.getVectorElementType() == MVT::i1) {
6479     MVT WidenVT = MVT::getVectorVT(MVT::i8, VecVT.getVectorElementCount());
6480     SDValue Op1 = DAG.getNode(ISD::ZERO_EXTEND, DL, WidenVT, Op.getOperand(0));
6481     SDValue Op2 = DAG.getNode(ISD::VECTOR_REVERSE, DL, WidenVT, Op1);
6482     return DAG.getNode(ISD::TRUNCATE, DL, VecVT, Op2);
6483   }
6484   unsigned EltSize = VecVT.getScalarSizeInBits();
6485   unsigned MinSize = VecVT.getSizeInBits().getKnownMinValue();
6486   unsigned VectorBitsMax = Subtarget.getRealMaxVLen();
6487   unsigned MaxVLMAX =
6488     RISCVTargetLowering::computeVLMAX(VectorBitsMax, EltSize, MinSize);
6489 
6490   unsigned GatherOpc = RISCVISD::VRGATHER_VV_VL;
6491   MVT IntVT = VecVT.changeVectorElementTypeToInteger();
6492 
6493   // If this is SEW=8 and VLMAX is potentially more than 256, we need
6494   // to use vrgatherei16.vv.
6495   // TODO: It's also possible to use vrgatherei16.vv for other types to
6496   // decrease register width for the index calculation.
6497   if (MaxVLMAX > 256 && EltSize == 8) {
6498     // If this is LMUL=8, we have to split before can use vrgatherei16.vv.
6499     // Reverse each half, then reassemble them in reverse order.
6500     // NOTE: It's also possible that after splitting that VLMAX no longer
6501     // requires vrgatherei16.vv.
6502     if (MinSize == (8 * RISCV::RVVBitsPerBlock)) {
6503       auto [Lo, Hi] = DAG.SplitVectorOperand(Op.getNode(), 0);
6504       auto [LoVT, HiVT] = DAG.GetSplitDestVTs(VecVT);
6505       Lo = DAG.getNode(ISD::VECTOR_REVERSE, DL, LoVT, Lo);
6506       Hi = DAG.getNode(ISD::VECTOR_REVERSE, DL, HiVT, Hi);
6507       // Reassemble the low and high pieces reversed.
6508       // FIXME: This is a CONCAT_VECTORS.
6509       SDValue Res =
6510           DAG.getNode(ISD::INSERT_SUBVECTOR, DL, VecVT, DAG.getUNDEF(VecVT), Hi,
6511                       DAG.getIntPtrConstant(0, DL));
6512       return DAG.getNode(
6513           ISD::INSERT_SUBVECTOR, DL, VecVT, Res, Lo,
6514           DAG.getIntPtrConstant(LoVT.getVectorMinNumElements(), DL));
6515     }
6516 
6517     // Just promote the int type to i16 which will double the LMUL.
6518     IntVT = MVT::getVectorVT(MVT::i16, VecVT.getVectorElementCount());
6519     GatherOpc = RISCVISD::VRGATHEREI16_VV_VL;
6520   }
6521 
6522   MVT XLenVT = Subtarget.getXLenVT();
6523   auto [Mask, VL] = getDefaultScalableVLOps(VecVT, DL, DAG, Subtarget);
6524 
6525   // Calculate VLMAX-1 for the desired SEW.
6526   unsigned MinElts = VecVT.getVectorMinNumElements();
6527   SDValue VLMax = DAG.getNode(ISD::VSCALE, DL, XLenVT,
6528                               getVLOp(MinElts, DL, DAG, Subtarget));
6529   SDValue VLMinus1 =
6530       DAG.getNode(ISD::SUB, DL, XLenVT, VLMax, DAG.getConstant(1, DL, XLenVT));
6531 
6532   // Splat VLMAX-1 taking care to handle SEW==64 on RV32.
6533   bool IsRV32E64 =
6534       !Subtarget.is64Bit() && IntVT.getVectorElementType() == MVT::i64;
6535   SDValue SplatVL;
6536   if (!IsRV32E64)
6537     SplatVL = DAG.getSplatVector(IntVT, DL, VLMinus1);
6538   else
6539     SplatVL = DAG.getNode(RISCVISD::VMV_V_X_VL, DL, IntVT, DAG.getUNDEF(IntVT),
6540                           VLMinus1, DAG.getRegister(RISCV::X0, XLenVT));
6541 
6542   SDValue VID = DAG.getNode(RISCVISD::VID_VL, DL, IntVT, Mask, VL);
6543   SDValue Indices = DAG.getNode(RISCVISD::SUB_VL, DL, IntVT, SplatVL, VID,
6544                                 DAG.getUNDEF(IntVT), Mask, VL);
6545 
6546   return DAG.getNode(GatherOpc, DL, VecVT, Op.getOperand(0), Indices,
6547                      DAG.getUNDEF(VecVT), Mask, VL);
6548 }
6549 
lowerVECTOR_SPLICE(SDValue Op,SelectionDAG & DAG) const6550 SDValue RISCVTargetLowering::lowerVECTOR_SPLICE(SDValue Op,
6551                                                 SelectionDAG &DAG) const {
6552   SDLoc DL(Op);
6553   SDValue V1 = Op.getOperand(0);
6554   SDValue V2 = Op.getOperand(1);
6555   MVT XLenVT = Subtarget.getXLenVT();
6556   MVT VecVT = Op.getSimpleValueType();
6557 
6558   unsigned MinElts = VecVT.getVectorMinNumElements();
6559   SDValue VLMax = DAG.getNode(ISD::VSCALE, DL, XLenVT,
6560                               getVLOp(MinElts, DL, DAG, Subtarget));
6561 
6562   int64_t ImmValue = cast<ConstantSDNode>(Op.getOperand(2))->getSExtValue();
6563   SDValue DownOffset, UpOffset;
6564   if (ImmValue >= 0) {
6565     // The operand is a TargetConstant, we need to rebuild it as a regular
6566     // constant.
6567     DownOffset = DAG.getConstant(ImmValue, DL, XLenVT);
6568     UpOffset = DAG.getNode(ISD::SUB, DL, XLenVT, VLMax, DownOffset);
6569   } else {
6570     // The operand is a TargetConstant, we need to rebuild it as a regular
6571     // constant rather than negating the original operand.
6572     UpOffset = DAG.getConstant(-ImmValue, DL, XLenVT);
6573     DownOffset = DAG.getNode(ISD::SUB, DL, XLenVT, VLMax, UpOffset);
6574   }
6575 
6576   SDValue TrueMask = getAllOnesMask(VecVT, VLMax, DL, DAG);
6577 
6578   SDValue SlideDown =
6579       getVSlidedown(DAG, Subtarget, DL, VecVT, DAG.getUNDEF(VecVT), V1,
6580                     DownOffset, TrueMask, UpOffset);
6581   return getVSlideup(DAG, Subtarget, DL, VecVT, SlideDown, V2, UpOffset,
6582                      TrueMask, DAG.getRegister(RISCV::X0, XLenVT),
6583                      RISCVII::TAIL_AGNOSTIC);
6584 }
6585 
6586 SDValue
lowerFixedLengthVectorLoadToRVV(SDValue Op,SelectionDAG & DAG) const6587 RISCVTargetLowering::lowerFixedLengthVectorLoadToRVV(SDValue Op,
6588                                                      SelectionDAG &DAG) const {
6589   SDLoc DL(Op);
6590   auto *Load = cast<LoadSDNode>(Op);
6591 
6592   assert(allowsMemoryAccessForAlignment(*DAG.getContext(), DAG.getDataLayout(),
6593                                         Load->getMemoryVT(),
6594                                         *Load->getMemOperand()) &&
6595          "Expecting a correctly-aligned load");
6596 
6597   MVT VT = Op.getSimpleValueType();
6598   MVT XLenVT = Subtarget.getXLenVT();
6599   MVT ContainerVT = getContainerForFixedLengthVector(VT);
6600 
6601   SDValue VL = getVLOp(VT.getVectorNumElements(), DL, DAG, Subtarget);
6602 
6603   bool IsMaskOp = VT.getVectorElementType() == MVT::i1;
6604   SDValue IntID = DAG.getTargetConstant(
6605       IsMaskOp ? Intrinsic::riscv_vlm : Intrinsic::riscv_vle, DL, XLenVT);
6606   SmallVector<SDValue, 4> Ops{Load->getChain(), IntID};
6607   if (!IsMaskOp)
6608     Ops.push_back(DAG.getUNDEF(ContainerVT));
6609   Ops.push_back(Load->getBasePtr());
6610   Ops.push_back(VL);
6611   SDVTList VTs = DAG.getVTList({ContainerVT, MVT::Other});
6612   SDValue NewLoad =
6613       DAG.getMemIntrinsicNode(ISD::INTRINSIC_W_CHAIN, DL, VTs, Ops,
6614                               Load->getMemoryVT(), Load->getMemOperand());
6615 
6616   SDValue Result = convertFromScalableVector(VT, NewLoad, DAG, Subtarget);
6617   return DAG.getMergeValues({Result, NewLoad.getValue(1)}, DL);
6618 }
6619 
6620 SDValue
lowerFixedLengthVectorStoreToRVV(SDValue Op,SelectionDAG & DAG) const6621 RISCVTargetLowering::lowerFixedLengthVectorStoreToRVV(SDValue Op,
6622                                                       SelectionDAG &DAG) const {
6623   SDLoc DL(Op);
6624   auto *Store = cast<StoreSDNode>(Op);
6625 
6626   assert(allowsMemoryAccessForAlignment(*DAG.getContext(), DAG.getDataLayout(),
6627                                         Store->getMemoryVT(),
6628                                         *Store->getMemOperand()) &&
6629          "Expecting a correctly-aligned store");
6630 
6631   SDValue StoreVal = Store->getValue();
6632   MVT VT = StoreVal.getSimpleValueType();
6633   MVT XLenVT = Subtarget.getXLenVT();
6634 
6635   // If the size less than a byte, we need to pad with zeros to make a byte.
6636   if (VT.getVectorElementType() == MVT::i1 && VT.getVectorNumElements() < 8) {
6637     VT = MVT::v8i1;
6638     StoreVal = DAG.getNode(ISD::INSERT_SUBVECTOR, DL, VT,
6639                            DAG.getConstant(0, DL, VT), StoreVal,
6640                            DAG.getIntPtrConstant(0, DL));
6641   }
6642 
6643   MVT ContainerVT = getContainerForFixedLengthVector(VT);
6644 
6645   SDValue VL = getVLOp(VT.getVectorNumElements(), DL, DAG, Subtarget);
6646 
6647   SDValue NewValue =
6648       convertToScalableVector(ContainerVT, StoreVal, DAG, Subtarget);
6649 
6650   bool IsMaskOp = VT.getVectorElementType() == MVT::i1;
6651   SDValue IntID = DAG.getTargetConstant(
6652       IsMaskOp ? Intrinsic::riscv_vsm : Intrinsic::riscv_vse, DL, XLenVT);
6653   return DAG.getMemIntrinsicNode(
6654       ISD::INTRINSIC_VOID, DL, DAG.getVTList(MVT::Other),
6655       {Store->getChain(), IntID, NewValue, Store->getBasePtr(), VL},
6656       Store->getMemoryVT(), Store->getMemOperand());
6657 }
6658 
lowerMaskedLoad(SDValue Op,SelectionDAG & DAG) const6659 SDValue RISCVTargetLowering::lowerMaskedLoad(SDValue Op,
6660                                              SelectionDAG &DAG) const {
6661   SDLoc DL(Op);
6662   MVT VT = Op.getSimpleValueType();
6663 
6664   const auto *MemSD = cast<MemSDNode>(Op);
6665   EVT MemVT = MemSD->getMemoryVT();
6666   MachineMemOperand *MMO = MemSD->getMemOperand();
6667   SDValue Chain = MemSD->getChain();
6668   SDValue BasePtr = MemSD->getBasePtr();
6669 
6670   SDValue Mask, PassThru, VL;
6671   if (const auto *VPLoad = dyn_cast<VPLoadSDNode>(Op)) {
6672     Mask = VPLoad->getMask();
6673     PassThru = DAG.getUNDEF(VT);
6674     VL = VPLoad->getVectorLength();
6675   } else {
6676     const auto *MLoad = cast<MaskedLoadSDNode>(Op);
6677     Mask = MLoad->getMask();
6678     PassThru = MLoad->getPassThru();
6679   }
6680 
6681   bool IsUnmasked = ISD::isConstantSplatVectorAllOnes(Mask.getNode());
6682 
6683   MVT XLenVT = Subtarget.getXLenVT();
6684 
6685   MVT ContainerVT = VT;
6686   if (VT.isFixedLengthVector()) {
6687     ContainerVT = getContainerForFixedLengthVector(VT);
6688     PassThru = convertToScalableVector(ContainerVT, PassThru, DAG, Subtarget);
6689     if (!IsUnmasked) {
6690       MVT MaskVT = getMaskTypeFor(ContainerVT);
6691       Mask = convertToScalableVector(MaskVT, Mask, DAG, Subtarget);
6692     }
6693   }
6694 
6695   if (!VL)
6696     VL = getDefaultVLOps(VT, ContainerVT, DL, DAG, Subtarget).second;
6697 
6698   unsigned IntID =
6699       IsUnmasked ? Intrinsic::riscv_vle : Intrinsic::riscv_vle_mask;
6700   SmallVector<SDValue, 8> Ops{Chain, DAG.getTargetConstant(IntID, DL, XLenVT)};
6701   if (IsUnmasked)
6702     Ops.push_back(DAG.getUNDEF(ContainerVT));
6703   else
6704     Ops.push_back(PassThru);
6705   Ops.push_back(BasePtr);
6706   if (!IsUnmasked)
6707     Ops.push_back(Mask);
6708   Ops.push_back(VL);
6709   if (!IsUnmasked)
6710     Ops.push_back(DAG.getTargetConstant(RISCVII::TAIL_AGNOSTIC, DL, XLenVT));
6711 
6712   SDVTList VTs = DAG.getVTList({ContainerVT, MVT::Other});
6713 
6714   SDValue Result =
6715       DAG.getMemIntrinsicNode(ISD::INTRINSIC_W_CHAIN, DL, VTs, Ops, MemVT, MMO);
6716   Chain = Result.getValue(1);
6717 
6718   if (VT.isFixedLengthVector())
6719     Result = convertFromScalableVector(VT, Result, DAG, Subtarget);
6720 
6721   return DAG.getMergeValues({Result, Chain}, DL);
6722 }
6723 
lowerMaskedStore(SDValue Op,SelectionDAG & DAG) const6724 SDValue RISCVTargetLowering::lowerMaskedStore(SDValue Op,
6725                                               SelectionDAG &DAG) const {
6726   SDLoc DL(Op);
6727 
6728   const auto *MemSD = cast<MemSDNode>(Op);
6729   EVT MemVT = MemSD->getMemoryVT();
6730   MachineMemOperand *MMO = MemSD->getMemOperand();
6731   SDValue Chain = MemSD->getChain();
6732   SDValue BasePtr = MemSD->getBasePtr();
6733   SDValue Val, Mask, VL;
6734 
6735   if (const auto *VPStore = dyn_cast<VPStoreSDNode>(Op)) {
6736     Val = VPStore->getValue();
6737     Mask = VPStore->getMask();
6738     VL = VPStore->getVectorLength();
6739   } else {
6740     const auto *MStore = cast<MaskedStoreSDNode>(Op);
6741     Val = MStore->getValue();
6742     Mask = MStore->getMask();
6743   }
6744 
6745   bool IsUnmasked = ISD::isConstantSplatVectorAllOnes(Mask.getNode());
6746 
6747   MVT VT = Val.getSimpleValueType();
6748   MVT XLenVT = Subtarget.getXLenVT();
6749 
6750   MVT ContainerVT = VT;
6751   if (VT.isFixedLengthVector()) {
6752     ContainerVT = getContainerForFixedLengthVector(VT);
6753 
6754     Val = convertToScalableVector(ContainerVT, Val, DAG, Subtarget);
6755     if (!IsUnmasked) {
6756       MVT MaskVT = getMaskTypeFor(ContainerVT);
6757       Mask = convertToScalableVector(MaskVT, Mask, DAG, Subtarget);
6758     }
6759   }
6760 
6761   if (!VL)
6762     VL = getDefaultVLOps(VT, ContainerVT, DL, DAG, Subtarget).second;
6763 
6764   unsigned IntID =
6765       IsUnmasked ? Intrinsic::riscv_vse : Intrinsic::riscv_vse_mask;
6766   SmallVector<SDValue, 8> Ops{Chain, DAG.getTargetConstant(IntID, DL, XLenVT)};
6767   Ops.push_back(Val);
6768   Ops.push_back(BasePtr);
6769   if (!IsUnmasked)
6770     Ops.push_back(Mask);
6771   Ops.push_back(VL);
6772 
6773   return DAG.getMemIntrinsicNode(ISD::INTRINSIC_VOID, DL,
6774                                  DAG.getVTList(MVT::Other), Ops, MemVT, MMO);
6775 }
6776 
6777 SDValue
lowerFixedLengthVectorSetccToRVV(SDValue Op,SelectionDAG & DAG) const6778 RISCVTargetLowering::lowerFixedLengthVectorSetccToRVV(SDValue Op,
6779                                                       SelectionDAG &DAG) const {
6780   MVT InVT = Op.getOperand(0).getSimpleValueType();
6781   MVT ContainerVT = getContainerForFixedLengthVector(InVT);
6782 
6783   MVT VT = Op.getSimpleValueType();
6784 
6785   SDValue Op1 =
6786       convertToScalableVector(ContainerVT, Op.getOperand(0), DAG, Subtarget);
6787   SDValue Op2 =
6788       convertToScalableVector(ContainerVT, Op.getOperand(1), DAG, Subtarget);
6789 
6790   SDLoc DL(Op);
6791   auto [Mask, VL] = getDefaultVLOps(VT.getVectorNumElements(), ContainerVT, DL,
6792                                     DAG, Subtarget);
6793   MVT MaskVT = getMaskTypeFor(ContainerVT);
6794 
6795   SDValue Cmp =
6796       DAG.getNode(RISCVISD::SETCC_VL, DL, MaskVT,
6797                   {Op1, Op2, Op.getOperand(2), DAG.getUNDEF(MaskVT), Mask, VL});
6798 
6799   return convertFromScalableVector(VT, Cmp, DAG, Subtarget);
6800 }
6801 
lowerFixedLengthVectorLogicOpToRVV(SDValue Op,SelectionDAG & DAG,unsigned MaskOpc,unsigned VecOpc) const6802 SDValue RISCVTargetLowering::lowerFixedLengthVectorLogicOpToRVV(
6803     SDValue Op, SelectionDAG &DAG, unsigned MaskOpc, unsigned VecOpc) const {
6804   MVT VT = Op.getSimpleValueType();
6805 
6806   if (VT.getVectorElementType() == MVT::i1)
6807     return lowerToScalableOp(Op, DAG, MaskOpc, /*HasMergeOp*/ false,
6808                              /*HasMask*/ false);
6809 
6810   return lowerToScalableOp(Op, DAG, VecOpc, /*HasMergeOp*/ true);
6811 }
6812 
6813 SDValue
lowerFixedLengthVectorShiftToRVV(SDValue Op,SelectionDAG & DAG) const6814 RISCVTargetLowering::lowerFixedLengthVectorShiftToRVV(SDValue Op,
6815                                                       SelectionDAG &DAG) const {
6816   unsigned Opc;
6817   switch (Op.getOpcode()) {
6818   default: llvm_unreachable("Unexpected opcode!");
6819   case ISD::SHL: Opc = RISCVISD::SHL_VL; break;
6820   case ISD::SRA: Opc = RISCVISD::SRA_VL; break;
6821   case ISD::SRL: Opc = RISCVISD::SRL_VL; break;
6822   }
6823 
6824   return lowerToScalableOp(Op, DAG, Opc, /*HasMergeOp*/ true);
6825 }
6826 
6827 // Lower vector ABS to smax(X, sub(0, X)).
lowerABS(SDValue Op,SelectionDAG & DAG) const6828 SDValue RISCVTargetLowering::lowerABS(SDValue Op, SelectionDAG &DAG) const {
6829   SDLoc DL(Op);
6830   MVT VT = Op.getSimpleValueType();
6831   SDValue X = Op.getOperand(0);
6832 
6833   assert((Op.getOpcode() == ISD::VP_ABS || VT.isFixedLengthVector()) &&
6834          "Unexpected type for ISD::ABS");
6835 
6836   MVT ContainerVT = VT;
6837   if (VT.isFixedLengthVector()) {
6838     ContainerVT = getContainerForFixedLengthVector(VT);
6839     X = convertToScalableVector(ContainerVT, X, DAG, Subtarget);
6840   }
6841 
6842   SDValue Mask, VL;
6843   if (Op->getOpcode() == ISD::VP_ABS) {
6844     Mask = Op->getOperand(1);
6845     VL = Op->getOperand(2);
6846   } else
6847     std::tie(Mask, VL) = getDefaultVLOps(VT, ContainerVT, DL, DAG, Subtarget);
6848 
6849   SDValue SplatZero = DAG.getNode(
6850       RISCVISD::VMV_V_X_VL, DL, ContainerVT, DAG.getUNDEF(ContainerVT),
6851       DAG.getConstant(0, DL, Subtarget.getXLenVT()), VL);
6852   SDValue NegX = DAG.getNode(RISCVISD::SUB_VL, DL, ContainerVT, SplatZero, X,
6853                              DAG.getUNDEF(ContainerVT), Mask, VL);
6854   SDValue Max = DAG.getNode(RISCVISD::SMAX_VL, DL, ContainerVT, X, NegX,
6855                             DAG.getUNDEF(ContainerVT), Mask, VL);
6856 
6857   if (VT.isFixedLengthVector())
6858     Max = convertFromScalableVector(VT, Max, DAG, Subtarget);
6859   return Max;
6860 }
6861 
lowerFixedLengthVectorFCOPYSIGNToRVV(SDValue Op,SelectionDAG & DAG) const6862 SDValue RISCVTargetLowering::lowerFixedLengthVectorFCOPYSIGNToRVV(
6863     SDValue Op, SelectionDAG &DAG) const {
6864   SDLoc DL(Op);
6865   MVT VT = Op.getSimpleValueType();
6866   SDValue Mag = Op.getOperand(0);
6867   SDValue Sign = Op.getOperand(1);
6868   assert(Mag.getValueType() == Sign.getValueType() &&
6869          "Can only handle COPYSIGN with matching types.");
6870 
6871   MVT ContainerVT = getContainerForFixedLengthVector(VT);
6872   Mag = convertToScalableVector(ContainerVT, Mag, DAG, Subtarget);
6873   Sign = convertToScalableVector(ContainerVT, Sign, DAG, Subtarget);
6874 
6875   auto [Mask, VL] = getDefaultVLOps(VT, ContainerVT, DL, DAG, Subtarget);
6876 
6877   SDValue CopySign = DAG.getNode(RISCVISD::FCOPYSIGN_VL, DL, ContainerVT, Mag,
6878                                  Sign, DAG.getUNDEF(ContainerVT), Mask, VL);
6879 
6880   return convertFromScalableVector(VT, CopySign, DAG, Subtarget);
6881 }
6882 
lowerFixedLengthVectorSelectToRVV(SDValue Op,SelectionDAG & DAG) const6883 SDValue RISCVTargetLowering::lowerFixedLengthVectorSelectToRVV(
6884     SDValue Op, SelectionDAG &DAG) const {
6885   MVT VT = Op.getSimpleValueType();
6886   MVT ContainerVT = getContainerForFixedLengthVector(VT);
6887 
6888   MVT I1ContainerVT =
6889       MVT::getVectorVT(MVT::i1, ContainerVT.getVectorElementCount());
6890 
6891   SDValue CC =
6892       convertToScalableVector(I1ContainerVT, Op.getOperand(0), DAG, Subtarget);
6893   SDValue Op1 =
6894       convertToScalableVector(ContainerVT, Op.getOperand(1), DAG, Subtarget);
6895   SDValue Op2 =
6896       convertToScalableVector(ContainerVT, Op.getOperand(2), DAG, Subtarget);
6897 
6898   SDLoc DL(Op);
6899   SDValue VL = getDefaultVLOps(VT, ContainerVT, DL, DAG, Subtarget).second;
6900 
6901   SDValue Select =
6902       DAG.getNode(RISCVISD::VSELECT_VL, DL, ContainerVT, CC, Op1, Op2, VL);
6903 
6904   return convertFromScalableVector(VT, Select, DAG, Subtarget);
6905 }
6906 
lowerToScalableOp(SDValue Op,SelectionDAG & DAG,unsigned NewOpc,bool HasMergeOp,bool HasMask) const6907 SDValue RISCVTargetLowering::lowerToScalableOp(SDValue Op, SelectionDAG &DAG,
6908                                                unsigned NewOpc, bool HasMergeOp,
6909                                                bool HasMask) const {
6910   MVT VT = Op.getSimpleValueType();
6911   MVT ContainerVT = getContainerForFixedLengthVector(VT);
6912 
6913   // Create list of operands by converting existing ones to scalable types.
6914   SmallVector<SDValue, 6> Ops;
6915   for (const SDValue &V : Op->op_values()) {
6916     assert(!isa<VTSDNode>(V) && "Unexpected VTSDNode node!");
6917 
6918     // Pass through non-vector operands.
6919     if (!V.getValueType().isVector()) {
6920       Ops.push_back(V);
6921       continue;
6922     }
6923 
6924     // "cast" fixed length vector to a scalable vector.
6925     assert(useRVVForFixedLengthVectorVT(V.getSimpleValueType()) &&
6926            "Only fixed length vectors are supported!");
6927     Ops.push_back(convertToScalableVector(ContainerVT, V, DAG, Subtarget));
6928   }
6929 
6930   SDLoc DL(Op);
6931   auto [Mask, VL] = getDefaultVLOps(VT, ContainerVT, DL, DAG, Subtarget);
6932   if (HasMergeOp)
6933     Ops.push_back(DAG.getUNDEF(ContainerVT));
6934   if (HasMask)
6935     Ops.push_back(Mask);
6936   Ops.push_back(VL);
6937 
6938   SDValue ScalableRes =
6939       DAG.getNode(NewOpc, DL, ContainerVT, Ops, Op->getFlags());
6940   return convertFromScalableVector(VT, ScalableRes, DAG, Subtarget);
6941 }
6942 
6943 // Lower a VP_* ISD node to the corresponding RISCVISD::*_VL node:
6944 // * Operands of each node are assumed to be in the same order.
6945 // * The EVL operand is promoted from i32 to i64 on RV64.
6946 // * Fixed-length vectors are converted to their scalable-vector container
6947 //   types.
lowerVPOp(SDValue Op,SelectionDAG & DAG,unsigned RISCVISDOpc,bool HasMergeOp) const6948 SDValue RISCVTargetLowering::lowerVPOp(SDValue Op, SelectionDAG &DAG,
6949                                        unsigned RISCVISDOpc,
6950                                        bool HasMergeOp) const {
6951   SDLoc DL(Op);
6952   MVT VT = Op.getSimpleValueType();
6953   SmallVector<SDValue, 4> Ops;
6954 
6955   MVT ContainerVT = VT;
6956   if (VT.isFixedLengthVector())
6957     ContainerVT = getContainerForFixedLengthVector(VT);
6958 
6959   for (const auto &OpIdx : enumerate(Op->ops())) {
6960     SDValue V = OpIdx.value();
6961     assert(!isa<VTSDNode>(V) && "Unexpected VTSDNode node!");
6962     // Add dummy merge value before the mask.
6963     if (HasMergeOp && *ISD::getVPMaskIdx(Op.getOpcode()) == OpIdx.index())
6964       Ops.push_back(DAG.getUNDEF(ContainerVT));
6965     // Pass through operands which aren't fixed-length vectors.
6966     if (!V.getValueType().isFixedLengthVector()) {
6967       Ops.push_back(V);
6968       continue;
6969     }
6970     // "cast" fixed length vector to a scalable vector.
6971     MVT OpVT = V.getSimpleValueType();
6972     MVT ContainerVT = getContainerForFixedLengthVector(OpVT);
6973     assert(useRVVForFixedLengthVectorVT(OpVT) &&
6974            "Only fixed length vectors are supported!");
6975     Ops.push_back(convertToScalableVector(ContainerVT, V, DAG, Subtarget));
6976   }
6977 
6978   if (!VT.isFixedLengthVector())
6979     return DAG.getNode(RISCVISDOpc, DL, VT, Ops, Op->getFlags());
6980 
6981   SDValue VPOp = DAG.getNode(RISCVISDOpc, DL, ContainerVT, Ops, Op->getFlags());
6982 
6983   return convertFromScalableVector(VT, VPOp, DAG, Subtarget);
6984 }
6985 
lowerVPExtMaskOp(SDValue Op,SelectionDAG & DAG) const6986 SDValue RISCVTargetLowering::lowerVPExtMaskOp(SDValue Op,
6987                                               SelectionDAG &DAG) const {
6988   SDLoc DL(Op);
6989   MVT VT = Op.getSimpleValueType();
6990 
6991   SDValue Src = Op.getOperand(0);
6992   // NOTE: Mask is dropped.
6993   SDValue VL = Op.getOperand(2);
6994 
6995   MVT ContainerVT = VT;
6996   if (VT.isFixedLengthVector()) {
6997     ContainerVT = getContainerForFixedLengthVector(VT);
6998     MVT SrcVT = MVT::getVectorVT(MVT::i1, ContainerVT.getVectorElementCount());
6999     Src = convertToScalableVector(SrcVT, Src, DAG, Subtarget);
7000   }
7001 
7002   MVT XLenVT = Subtarget.getXLenVT();
7003   SDValue Zero = DAG.getConstant(0, DL, XLenVT);
7004   SDValue ZeroSplat = DAG.getNode(RISCVISD::VMV_V_X_VL, DL, ContainerVT,
7005                                   DAG.getUNDEF(ContainerVT), Zero, VL);
7006 
7007   SDValue SplatValue = DAG.getConstant(
7008       Op.getOpcode() == ISD::VP_ZERO_EXTEND ? 1 : -1, DL, XLenVT);
7009   SDValue Splat = DAG.getNode(RISCVISD::VMV_V_X_VL, DL, ContainerVT,
7010                               DAG.getUNDEF(ContainerVT), SplatValue, VL);
7011 
7012   SDValue Result = DAG.getNode(RISCVISD::VSELECT_VL, DL, ContainerVT, Src,
7013                                Splat, ZeroSplat, VL);
7014   if (!VT.isFixedLengthVector())
7015     return Result;
7016   return convertFromScalableVector(VT, Result, DAG, Subtarget);
7017 }
7018 
lowerVPSetCCMaskOp(SDValue Op,SelectionDAG & DAG) const7019 SDValue RISCVTargetLowering::lowerVPSetCCMaskOp(SDValue Op,
7020                                                 SelectionDAG &DAG) const {
7021   SDLoc DL(Op);
7022   MVT VT = Op.getSimpleValueType();
7023 
7024   SDValue Op1 = Op.getOperand(0);
7025   SDValue Op2 = Op.getOperand(1);
7026   ISD::CondCode Condition = cast<CondCodeSDNode>(Op.getOperand(2))->get();
7027   // NOTE: Mask is dropped.
7028   SDValue VL = Op.getOperand(4);
7029 
7030   MVT ContainerVT = VT;
7031   if (VT.isFixedLengthVector()) {
7032     ContainerVT = getContainerForFixedLengthVector(VT);
7033     Op1 = convertToScalableVector(ContainerVT, Op1, DAG, Subtarget);
7034     Op2 = convertToScalableVector(ContainerVT, Op2, DAG, Subtarget);
7035   }
7036 
7037   SDValue Result;
7038   SDValue AllOneMask = DAG.getNode(RISCVISD::VMSET_VL, DL, ContainerVT, VL);
7039 
7040   switch (Condition) {
7041   default:
7042     break;
7043   // X != Y  --> (X^Y)
7044   case ISD::SETNE:
7045     Result = DAG.getNode(RISCVISD::VMXOR_VL, DL, ContainerVT, Op1, Op2, VL);
7046     break;
7047   // X == Y  --> ~(X^Y)
7048   case ISD::SETEQ: {
7049     SDValue Temp =
7050         DAG.getNode(RISCVISD::VMXOR_VL, DL, ContainerVT, Op1, Op2, VL);
7051     Result =
7052         DAG.getNode(RISCVISD::VMXOR_VL, DL, ContainerVT, Temp, AllOneMask, VL);
7053     break;
7054   }
7055   // X >s Y   -->  X == 0 & Y == 1  -->  ~X & Y
7056   // X <u Y   -->  X == 0 & Y == 1  -->  ~X & Y
7057   case ISD::SETGT:
7058   case ISD::SETULT: {
7059     SDValue Temp =
7060         DAG.getNode(RISCVISD::VMXOR_VL, DL, ContainerVT, Op1, AllOneMask, VL);
7061     Result = DAG.getNode(RISCVISD::VMAND_VL, DL, ContainerVT, Temp, Op2, VL);
7062     break;
7063   }
7064   // X <s Y   --> X == 1 & Y == 0  -->  ~Y & X
7065   // X >u Y   --> X == 1 & Y == 0  -->  ~Y & X
7066   case ISD::SETLT:
7067   case ISD::SETUGT: {
7068     SDValue Temp =
7069         DAG.getNode(RISCVISD::VMXOR_VL, DL, ContainerVT, Op2, AllOneMask, VL);
7070     Result = DAG.getNode(RISCVISD::VMAND_VL, DL, ContainerVT, Op1, Temp, VL);
7071     break;
7072   }
7073   // X >=s Y  --> X == 0 | Y == 1  -->  ~X | Y
7074   // X <=u Y  --> X == 0 | Y == 1  -->  ~X | Y
7075   case ISD::SETGE:
7076   case ISD::SETULE: {
7077     SDValue Temp =
7078         DAG.getNode(RISCVISD::VMXOR_VL, DL, ContainerVT, Op1, AllOneMask, VL);
7079     Result = DAG.getNode(RISCVISD::VMXOR_VL, DL, ContainerVT, Temp, Op2, VL);
7080     break;
7081   }
7082   // X <=s Y  --> X == 1 | Y == 0  -->  ~Y | X
7083   // X >=u Y  --> X == 1 | Y == 0  -->  ~Y | X
7084   case ISD::SETLE:
7085   case ISD::SETUGE: {
7086     SDValue Temp =
7087         DAG.getNode(RISCVISD::VMXOR_VL, DL, ContainerVT, Op2, AllOneMask, VL);
7088     Result = DAG.getNode(RISCVISD::VMXOR_VL, DL, ContainerVT, Temp, Op1, VL);
7089     break;
7090   }
7091   }
7092 
7093   if (!VT.isFixedLengthVector())
7094     return Result;
7095   return convertFromScalableVector(VT, Result, DAG, Subtarget);
7096 }
7097 
7098 // Lower Floating-Point/Integer Type-Convert VP SDNodes
lowerVPFPIntConvOp(SDValue Op,SelectionDAG & DAG,unsigned RISCVISDOpc) const7099 SDValue RISCVTargetLowering::lowerVPFPIntConvOp(SDValue Op, SelectionDAG &DAG,
7100                                                 unsigned RISCVISDOpc) const {
7101   SDLoc DL(Op);
7102 
7103   SDValue Src = Op.getOperand(0);
7104   SDValue Mask = Op.getOperand(1);
7105   SDValue VL = Op.getOperand(2);
7106 
7107   MVT DstVT = Op.getSimpleValueType();
7108   MVT SrcVT = Src.getSimpleValueType();
7109   if (DstVT.isFixedLengthVector()) {
7110     DstVT = getContainerForFixedLengthVector(DstVT);
7111     SrcVT = getContainerForFixedLengthVector(SrcVT);
7112     Src = convertToScalableVector(SrcVT, Src, DAG, Subtarget);
7113     MVT MaskVT = getMaskTypeFor(DstVT);
7114     Mask = convertToScalableVector(MaskVT, Mask, DAG, Subtarget);
7115   }
7116 
7117   unsigned DstEltSize = DstVT.getScalarSizeInBits();
7118   unsigned SrcEltSize = SrcVT.getScalarSizeInBits();
7119 
7120   SDValue Result;
7121   if (DstEltSize >= SrcEltSize) { // Single-width and widening conversion.
7122     if (SrcVT.isInteger()) {
7123       assert(DstVT.isFloatingPoint() && "Wrong input/output vector types");
7124 
7125       unsigned RISCVISDExtOpc = RISCVISDOpc == RISCVISD::SINT_TO_FP_VL
7126                                     ? RISCVISD::VSEXT_VL
7127                                     : RISCVISD::VZEXT_VL;
7128 
7129       // Do we need to do any pre-widening before converting?
7130       if (SrcEltSize == 1) {
7131         MVT IntVT = DstVT.changeVectorElementTypeToInteger();
7132         MVT XLenVT = Subtarget.getXLenVT();
7133         SDValue Zero = DAG.getConstant(0, DL, XLenVT);
7134         SDValue ZeroSplat = DAG.getNode(RISCVISD::VMV_V_X_VL, DL, IntVT,
7135                                         DAG.getUNDEF(IntVT), Zero, VL);
7136         SDValue One = DAG.getConstant(
7137             RISCVISDExtOpc == RISCVISD::VZEXT_VL ? 1 : -1, DL, XLenVT);
7138         SDValue OneSplat = DAG.getNode(RISCVISD::VMV_V_X_VL, DL, IntVT,
7139                                        DAG.getUNDEF(IntVT), One, VL);
7140         Src = DAG.getNode(RISCVISD::VSELECT_VL, DL, IntVT, Src, OneSplat,
7141                           ZeroSplat, VL);
7142       } else if (DstEltSize > (2 * SrcEltSize)) {
7143         // Widen before converting.
7144         MVT IntVT = MVT::getVectorVT(MVT::getIntegerVT(DstEltSize / 2),
7145                                      DstVT.getVectorElementCount());
7146         Src = DAG.getNode(RISCVISDExtOpc, DL, IntVT, Src, Mask, VL);
7147       }
7148 
7149       Result = DAG.getNode(RISCVISDOpc, DL, DstVT, Src, Mask, VL);
7150     } else {
7151       assert(SrcVT.isFloatingPoint() && DstVT.isInteger() &&
7152              "Wrong input/output vector types");
7153 
7154       // Convert f16 to f32 then convert f32 to i64.
7155       if (DstEltSize > (2 * SrcEltSize)) {
7156         assert(SrcVT.getVectorElementType() == MVT::f16 && "Unexpected type!");
7157         MVT InterimFVT =
7158             MVT::getVectorVT(MVT::f32, DstVT.getVectorElementCount());
7159         Src =
7160             DAG.getNode(RISCVISD::FP_EXTEND_VL, DL, InterimFVT, Src, Mask, VL);
7161       }
7162 
7163       Result = DAG.getNode(RISCVISDOpc, DL, DstVT, Src, Mask, VL);
7164     }
7165   } else { // Narrowing + Conversion
7166     if (SrcVT.isInteger()) {
7167       assert(DstVT.isFloatingPoint() && "Wrong input/output vector types");
7168       // First do a narrowing convert to an FP type half the size, then round
7169       // the FP type to a small FP type if needed.
7170 
7171       MVT InterimFVT = DstVT;
7172       if (SrcEltSize > (2 * DstEltSize)) {
7173         assert(SrcEltSize == (4 * DstEltSize) && "Unexpected types!");
7174         assert(DstVT.getVectorElementType() == MVT::f16 && "Unexpected type!");
7175         InterimFVT = MVT::getVectorVT(MVT::f32, DstVT.getVectorElementCount());
7176       }
7177 
7178       Result = DAG.getNode(RISCVISDOpc, DL, InterimFVT, Src, Mask, VL);
7179 
7180       if (InterimFVT != DstVT) {
7181         Src = Result;
7182         Result = DAG.getNode(RISCVISD::FP_ROUND_VL, DL, DstVT, Src, Mask, VL);
7183       }
7184     } else {
7185       assert(SrcVT.isFloatingPoint() && DstVT.isInteger() &&
7186              "Wrong input/output vector types");
7187       // First do a narrowing conversion to an integer half the size, then
7188       // truncate if needed.
7189 
7190       if (DstEltSize == 1) {
7191         // First convert to the same size integer, then convert to mask using
7192         // setcc.
7193         assert(SrcEltSize >= 16 && "Unexpected FP type!");
7194         MVT InterimIVT = MVT::getVectorVT(MVT::getIntegerVT(SrcEltSize),
7195                                           DstVT.getVectorElementCount());
7196         Result = DAG.getNode(RISCVISDOpc, DL, InterimIVT, Src, Mask, VL);
7197 
7198         // Compare the integer result to 0. The integer should be 0 or 1/-1,
7199         // otherwise the conversion was undefined.
7200         MVT XLenVT = Subtarget.getXLenVT();
7201         SDValue SplatZero = DAG.getConstant(0, DL, XLenVT);
7202         SplatZero = DAG.getNode(RISCVISD::VMV_V_X_VL, DL, InterimIVT,
7203                                 DAG.getUNDEF(InterimIVT), SplatZero, VL);
7204         Result = DAG.getNode(RISCVISD::SETCC_VL, DL, DstVT,
7205                              {Result, SplatZero, DAG.getCondCode(ISD::SETNE),
7206                               DAG.getUNDEF(DstVT), Mask, VL});
7207       } else {
7208         MVT InterimIVT = MVT::getVectorVT(MVT::getIntegerVT(SrcEltSize / 2),
7209                                           DstVT.getVectorElementCount());
7210 
7211         Result = DAG.getNode(RISCVISDOpc, DL, InterimIVT, Src, Mask, VL);
7212 
7213         while (InterimIVT != DstVT) {
7214           SrcEltSize /= 2;
7215           Src = Result;
7216           InterimIVT = MVT::getVectorVT(MVT::getIntegerVT(SrcEltSize / 2),
7217                                         DstVT.getVectorElementCount());
7218           Result = DAG.getNode(RISCVISD::TRUNCATE_VECTOR_VL, DL, InterimIVT,
7219                                Src, Mask, VL);
7220         }
7221       }
7222     }
7223   }
7224 
7225   MVT VT = Op.getSimpleValueType();
7226   if (!VT.isFixedLengthVector())
7227     return Result;
7228   return convertFromScalableVector(VT, Result, DAG, Subtarget);
7229 }
7230 
lowerLogicVPOp(SDValue Op,SelectionDAG & DAG,unsigned MaskOpc,unsigned VecOpc) const7231 SDValue RISCVTargetLowering::lowerLogicVPOp(SDValue Op, SelectionDAG &DAG,
7232                                             unsigned MaskOpc,
7233                                             unsigned VecOpc) const {
7234   MVT VT = Op.getSimpleValueType();
7235   if (VT.getVectorElementType() != MVT::i1)
7236     return lowerVPOp(Op, DAG, VecOpc, true);
7237 
7238   // It is safe to drop mask parameter as masked-off elements are undef.
7239   SDValue Op1 = Op->getOperand(0);
7240   SDValue Op2 = Op->getOperand(1);
7241   SDValue VL = Op->getOperand(3);
7242 
7243   MVT ContainerVT = VT;
7244   const bool IsFixed = VT.isFixedLengthVector();
7245   if (IsFixed) {
7246     ContainerVT = getContainerForFixedLengthVector(VT);
7247     Op1 = convertToScalableVector(ContainerVT, Op1, DAG, Subtarget);
7248     Op2 = convertToScalableVector(ContainerVT, Op2, DAG, Subtarget);
7249   }
7250 
7251   SDLoc DL(Op);
7252   SDValue Val = DAG.getNode(MaskOpc, DL, ContainerVT, Op1, Op2, VL);
7253   if (!IsFixed)
7254     return Val;
7255   return convertFromScalableVector(VT, Val, DAG, Subtarget);
7256 }
7257 
lowerVPStridedLoad(SDValue Op,SelectionDAG & DAG) const7258 SDValue RISCVTargetLowering::lowerVPStridedLoad(SDValue Op,
7259                                                 SelectionDAG &DAG) const {
7260   SDLoc DL(Op);
7261   MVT XLenVT = Subtarget.getXLenVT();
7262   MVT VT = Op.getSimpleValueType();
7263   MVT ContainerVT = VT;
7264   if (VT.isFixedLengthVector())
7265     ContainerVT = getContainerForFixedLengthVector(VT);
7266 
7267   SDVTList VTs = DAG.getVTList({ContainerVT, MVT::Other});
7268 
7269   auto *VPNode = cast<VPStridedLoadSDNode>(Op);
7270   // Check if the mask is known to be all ones
7271   SDValue Mask = VPNode->getMask();
7272   bool IsUnmasked = ISD::isConstantSplatVectorAllOnes(Mask.getNode());
7273 
7274   SDValue IntID = DAG.getTargetConstant(IsUnmasked ? Intrinsic::riscv_vlse
7275                                                    : Intrinsic::riscv_vlse_mask,
7276                                         DL, XLenVT);
7277   SmallVector<SDValue, 8> Ops{VPNode->getChain(), IntID,
7278                               DAG.getUNDEF(ContainerVT), VPNode->getBasePtr(),
7279                               VPNode->getStride()};
7280   if (!IsUnmasked) {
7281     if (VT.isFixedLengthVector()) {
7282       MVT MaskVT = ContainerVT.changeVectorElementType(MVT::i1);
7283       Mask = convertToScalableVector(MaskVT, Mask, DAG, Subtarget);
7284     }
7285     Ops.push_back(Mask);
7286   }
7287   Ops.push_back(VPNode->getVectorLength());
7288   if (!IsUnmasked) {
7289     SDValue Policy = DAG.getTargetConstant(RISCVII::TAIL_AGNOSTIC, DL, XLenVT);
7290     Ops.push_back(Policy);
7291   }
7292 
7293   SDValue Result =
7294       DAG.getMemIntrinsicNode(ISD::INTRINSIC_W_CHAIN, DL, VTs, Ops,
7295                               VPNode->getMemoryVT(), VPNode->getMemOperand());
7296   SDValue Chain = Result.getValue(1);
7297 
7298   if (VT.isFixedLengthVector())
7299     Result = convertFromScalableVector(VT, Result, DAG, Subtarget);
7300 
7301   return DAG.getMergeValues({Result, Chain}, DL);
7302 }
7303 
lowerVPStridedStore(SDValue Op,SelectionDAG & DAG) const7304 SDValue RISCVTargetLowering::lowerVPStridedStore(SDValue Op,
7305                                                  SelectionDAG &DAG) const {
7306   SDLoc DL(Op);
7307   MVT XLenVT = Subtarget.getXLenVT();
7308 
7309   auto *VPNode = cast<VPStridedStoreSDNode>(Op);
7310   SDValue StoreVal = VPNode->getValue();
7311   MVT VT = StoreVal.getSimpleValueType();
7312   MVT ContainerVT = VT;
7313   if (VT.isFixedLengthVector()) {
7314     ContainerVT = getContainerForFixedLengthVector(VT);
7315     StoreVal = convertToScalableVector(ContainerVT, StoreVal, DAG, Subtarget);
7316   }
7317 
7318   // Check if the mask is known to be all ones
7319   SDValue Mask = VPNode->getMask();
7320   bool IsUnmasked = ISD::isConstantSplatVectorAllOnes(Mask.getNode());
7321 
7322   SDValue IntID = DAG.getTargetConstant(IsUnmasked ? Intrinsic::riscv_vsse
7323                                                    : Intrinsic::riscv_vsse_mask,
7324                                         DL, XLenVT);
7325   SmallVector<SDValue, 8> Ops{VPNode->getChain(), IntID, StoreVal,
7326                               VPNode->getBasePtr(), VPNode->getStride()};
7327   if (!IsUnmasked) {
7328     if (VT.isFixedLengthVector()) {
7329       MVT MaskVT = ContainerVT.changeVectorElementType(MVT::i1);
7330       Mask = convertToScalableVector(MaskVT, Mask, DAG, Subtarget);
7331     }
7332     Ops.push_back(Mask);
7333   }
7334   Ops.push_back(VPNode->getVectorLength());
7335 
7336   return DAG.getMemIntrinsicNode(ISD::INTRINSIC_VOID, DL, VPNode->getVTList(),
7337                                  Ops, VPNode->getMemoryVT(),
7338                                  VPNode->getMemOperand());
7339 }
7340 
7341 // Custom lower MGATHER/VP_GATHER to a legalized form for RVV. It will then be
7342 // matched to a RVV indexed load. The RVV indexed load instructions only
7343 // support the "unsigned unscaled" addressing mode; indices are implicitly
7344 // zero-extended or truncated to XLEN and are treated as byte offsets. Any
7345 // signed or scaled indexing is extended to the XLEN value type and scaled
7346 // accordingly.
lowerMaskedGather(SDValue Op,SelectionDAG & DAG) const7347 SDValue RISCVTargetLowering::lowerMaskedGather(SDValue Op,
7348                                                SelectionDAG &DAG) const {
7349   SDLoc DL(Op);
7350   MVT VT = Op.getSimpleValueType();
7351 
7352   const auto *MemSD = cast<MemSDNode>(Op.getNode());
7353   EVT MemVT = MemSD->getMemoryVT();
7354   MachineMemOperand *MMO = MemSD->getMemOperand();
7355   SDValue Chain = MemSD->getChain();
7356   SDValue BasePtr = MemSD->getBasePtr();
7357 
7358   ISD::LoadExtType LoadExtType;
7359   SDValue Index, Mask, PassThru, VL;
7360 
7361   if (auto *VPGN = dyn_cast<VPGatherSDNode>(Op.getNode())) {
7362     Index = VPGN->getIndex();
7363     Mask = VPGN->getMask();
7364     PassThru = DAG.getUNDEF(VT);
7365     VL = VPGN->getVectorLength();
7366     // VP doesn't support extending loads.
7367     LoadExtType = ISD::NON_EXTLOAD;
7368   } else {
7369     // Else it must be a MGATHER.
7370     auto *MGN = cast<MaskedGatherSDNode>(Op.getNode());
7371     Index = MGN->getIndex();
7372     Mask = MGN->getMask();
7373     PassThru = MGN->getPassThru();
7374     LoadExtType = MGN->getExtensionType();
7375   }
7376 
7377   MVT IndexVT = Index.getSimpleValueType();
7378   MVT XLenVT = Subtarget.getXLenVT();
7379 
7380   assert(VT.getVectorElementCount() == IndexVT.getVectorElementCount() &&
7381          "Unexpected VTs!");
7382   assert(BasePtr.getSimpleValueType() == XLenVT && "Unexpected pointer type");
7383   // Targets have to explicitly opt-in for extending vector loads.
7384   assert(LoadExtType == ISD::NON_EXTLOAD &&
7385          "Unexpected extending MGATHER/VP_GATHER");
7386   (void)LoadExtType;
7387 
7388   // If the mask is known to be all ones, optimize to an unmasked intrinsic;
7389   // the selection of the masked intrinsics doesn't do this for us.
7390   bool IsUnmasked = ISD::isConstantSplatVectorAllOnes(Mask.getNode());
7391 
7392   MVT ContainerVT = VT;
7393   if (VT.isFixedLengthVector()) {
7394     ContainerVT = getContainerForFixedLengthVector(VT);
7395     IndexVT = MVT::getVectorVT(IndexVT.getVectorElementType(),
7396                                ContainerVT.getVectorElementCount());
7397 
7398     Index = convertToScalableVector(IndexVT, Index, DAG, Subtarget);
7399 
7400     if (!IsUnmasked) {
7401       MVT MaskVT = getMaskTypeFor(ContainerVT);
7402       Mask = convertToScalableVector(MaskVT, Mask, DAG, Subtarget);
7403       PassThru = convertToScalableVector(ContainerVT, PassThru, DAG, Subtarget);
7404     }
7405   }
7406 
7407   if (!VL)
7408     VL = getDefaultVLOps(VT, ContainerVT, DL, DAG, Subtarget).second;
7409 
7410   if (XLenVT == MVT::i32 && IndexVT.getVectorElementType().bitsGT(XLenVT)) {
7411     IndexVT = IndexVT.changeVectorElementType(XLenVT);
7412     SDValue TrueMask = DAG.getNode(RISCVISD::VMSET_VL, DL, Mask.getValueType(),
7413                                    VL);
7414     Index = DAG.getNode(RISCVISD::TRUNCATE_VECTOR_VL, DL, IndexVT, Index,
7415                         TrueMask, VL);
7416   }
7417 
7418   unsigned IntID =
7419       IsUnmasked ? Intrinsic::riscv_vluxei : Intrinsic::riscv_vluxei_mask;
7420   SmallVector<SDValue, 8> Ops{Chain, DAG.getTargetConstant(IntID, DL, XLenVT)};
7421   if (IsUnmasked)
7422     Ops.push_back(DAG.getUNDEF(ContainerVT));
7423   else
7424     Ops.push_back(PassThru);
7425   Ops.push_back(BasePtr);
7426   Ops.push_back(Index);
7427   if (!IsUnmasked)
7428     Ops.push_back(Mask);
7429   Ops.push_back(VL);
7430   if (!IsUnmasked)
7431     Ops.push_back(DAG.getTargetConstant(RISCVII::TAIL_AGNOSTIC, DL, XLenVT));
7432 
7433   SDVTList VTs = DAG.getVTList({ContainerVT, MVT::Other});
7434   SDValue Result =
7435       DAG.getMemIntrinsicNode(ISD::INTRINSIC_W_CHAIN, DL, VTs, Ops, MemVT, MMO);
7436   Chain = Result.getValue(1);
7437 
7438   if (VT.isFixedLengthVector())
7439     Result = convertFromScalableVector(VT, Result, DAG, Subtarget);
7440 
7441   return DAG.getMergeValues({Result, Chain}, DL);
7442 }
7443 
7444 // Custom lower MSCATTER/VP_SCATTER to a legalized form for RVV. It will then be
7445 // matched to a RVV indexed store. The RVV indexed store instructions only
7446 // support the "unsigned unscaled" addressing mode; indices are implicitly
7447 // zero-extended or truncated to XLEN and are treated as byte offsets. Any
7448 // signed or scaled indexing is extended to the XLEN value type and scaled
7449 // accordingly.
lowerMaskedScatter(SDValue Op,SelectionDAG & DAG) const7450 SDValue RISCVTargetLowering::lowerMaskedScatter(SDValue Op,
7451                                                 SelectionDAG &DAG) const {
7452   SDLoc DL(Op);
7453   const auto *MemSD = cast<MemSDNode>(Op.getNode());
7454   EVT MemVT = MemSD->getMemoryVT();
7455   MachineMemOperand *MMO = MemSD->getMemOperand();
7456   SDValue Chain = MemSD->getChain();
7457   SDValue BasePtr = MemSD->getBasePtr();
7458 
7459   bool IsTruncatingStore = false;
7460   SDValue Index, Mask, Val, VL;
7461 
7462   if (auto *VPSN = dyn_cast<VPScatterSDNode>(Op.getNode())) {
7463     Index = VPSN->getIndex();
7464     Mask = VPSN->getMask();
7465     Val = VPSN->getValue();
7466     VL = VPSN->getVectorLength();
7467     // VP doesn't support truncating stores.
7468     IsTruncatingStore = false;
7469   } else {
7470     // Else it must be a MSCATTER.
7471     auto *MSN = cast<MaskedScatterSDNode>(Op.getNode());
7472     Index = MSN->getIndex();
7473     Mask = MSN->getMask();
7474     Val = MSN->getValue();
7475     IsTruncatingStore = MSN->isTruncatingStore();
7476   }
7477 
7478   MVT VT = Val.getSimpleValueType();
7479   MVT IndexVT = Index.getSimpleValueType();
7480   MVT XLenVT = Subtarget.getXLenVT();
7481 
7482   assert(VT.getVectorElementCount() == IndexVT.getVectorElementCount() &&
7483          "Unexpected VTs!");
7484   assert(BasePtr.getSimpleValueType() == XLenVT && "Unexpected pointer type");
7485   // Targets have to explicitly opt-in for extending vector loads and
7486   // truncating vector stores.
7487   assert(!IsTruncatingStore && "Unexpected truncating MSCATTER/VP_SCATTER");
7488   (void)IsTruncatingStore;
7489 
7490   // If the mask is known to be all ones, optimize to an unmasked intrinsic;
7491   // the selection of the masked intrinsics doesn't do this for us.
7492   bool IsUnmasked = ISD::isConstantSplatVectorAllOnes(Mask.getNode());
7493 
7494   MVT ContainerVT = VT;
7495   if (VT.isFixedLengthVector()) {
7496     ContainerVT = getContainerForFixedLengthVector(VT);
7497     IndexVT = MVT::getVectorVT(IndexVT.getVectorElementType(),
7498                                ContainerVT.getVectorElementCount());
7499 
7500     Index = convertToScalableVector(IndexVT, Index, DAG, Subtarget);
7501     Val = convertToScalableVector(ContainerVT, Val, DAG, Subtarget);
7502 
7503     if (!IsUnmasked) {
7504       MVT MaskVT = getMaskTypeFor(ContainerVT);
7505       Mask = convertToScalableVector(MaskVT, Mask, DAG, Subtarget);
7506     }
7507   }
7508 
7509   if (!VL)
7510     VL = getDefaultVLOps(VT, ContainerVT, DL, DAG, Subtarget).second;
7511 
7512   if (XLenVT == MVT::i32 && IndexVT.getVectorElementType().bitsGT(XLenVT)) {
7513     IndexVT = IndexVT.changeVectorElementType(XLenVT);
7514     SDValue TrueMask = DAG.getNode(RISCVISD::VMSET_VL, DL, Mask.getValueType(),
7515                                    VL);
7516     Index = DAG.getNode(RISCVISD::TRUNCATE_VECTOR_VL, DL, IndexVT, Index,
7517                         TrueMask, VL);
7518   }
7519 
7520   unsigned IntID =
7521       IsUnmasked ? Intrinsic::riscv_vsoxei : Intrinsic::riscv_vsoxei_mask;
7522   SmallVector<SDValue, 8> Ops{Chain, DAG.getTargetConstant(IntID, DL, XLenVT)};
7523   Ops.push_back(Val);
7524   Ops.push_back(BasePtr);
7525   Ops.push_back(Index);
7526   if (!IsUnmasked)
7527     Ops.push_back(Mask);
7528   Ops.push_back(VL);
7529 
7530   return DAG.getMemIntrinsicNode(ISD::INTRINSIC_VOID, DL,
7531                                  DAG.getVTList(MVT::Other), Ops, MemVT, MMO);
7532 }
7533 
lowerGET_ROUNDING(SDValue Op,SelectionDAG & DAG) const7534 SDValue RISCVTargetLowering::lowerGET_ROUNDING(SDValue Op,
7535                                                SelectionDAG &DAG) const {
7536   const MVT XLenVT = Subtarget.getXLenVT();
7537   SDLoc DL(Op);
7538   SDValue Chain = Op->getOperand(0);
7539   SDValue SysRegNo = DAG.getTargetConstant(
7540       RISCVSysReg::lookupSysRegByName("FRM")->Encoding, DL, XLenVT);
7541   SDVTList VTs = DAG.getVTList(XLenVT, MVT::Other);
7542   SDValue RM = DAG.getNode(RISCVISD::READ_CSR, DL, VTs, Chain, SysRegNo);
7543 
7544   // Encoding used for rounding mode in RISCV differs from that used in
7545   // FLT_ROUNDS. To convert it the RISCV rounding mode is used as an index in a
7546   // table, which consists of a sequence of 4-bit fields, each representing
7547   // corresponding FLT_ROUNDS mode.
7548   static const int Table =
7549       (int(RoundingMode::NearestTiesToEven) << 4 * RISCVFPRndMode::RNE) |
7550       (int(RoundingMode::TowardZero) << 4 * RISCVFPRndMode::RTZ) |
7551       (int(RoundingMode::TowardNegative) << 4 * RISCVFPRndMode::RDN) |
7552       (int(RoundingMode::TowardPositive) << 4 * RISCVFPRndMode::RUP) |
7553       (int(RoundingMode::NearestTiesToAway) << 4 * RISCVFPRndMode::RMM);
7554 
7555   SDValue Shift =
7556       DAG.getNode(ISD::SHL, DL, XLenVT, RM, DAG.getConstant(2, DL, XLenVT));
7557   SDValue Shifted = DAG.getNode(ISD::SRL, DL, XLenVT,
7558                                 DAG.getConstant(Table, DL, XLenVT), Shift);
7559   SDValue Masked = DAG.getNode(ISD::AND, DL, XLenVT, Shifted,
7560                                DAG.getConstant(7, DL, XLenVT));
7561 
7562   return DAG.getMergeValues({Masked, Chain}, DL);
7563 }
7564 
lowerSET_ROUNDING(SDValue Op,SelectionDAG & DAG) const7565 SDValue RISCVTargetLowering::lowerSET_ROUNDING(SDValue Op,
7566                                                SelectionDAG &DAG) const {
7567   const MVT XLenVT = Subtarget.getXLenVT();
7568   SDLoc DL(Op);
7569   SDValue Chain = Op->getOperand(0);
7570   SDValue RMValue = Op->getOperand(1);
7571   SDValue SysRegNo = DAG.getTargetConstant(
7572       RISCVSysReg::lookupSysRegByName("FRM")->Encoding, DL, XLenVT);
7573 
7574   // Encoding used for rounding mode in RISCV differs from that used in
7575   // FLT_ROUNDS. To convert it the C rounding mode is used as an index in
7576   // a table, which consists of a sequence of 4-bit fields, each representing
7577   // corresponding RISCV mode.
7578   static const unsigned Table =
7579       (RISCVFPRndMode::RNE << 4 * int(RoundingMode::NearestTiesToEven)) |
7580       (RISCVFPRndMode::RTZ << 4 * int(RoundingMode::TowardZero)) |
7581       (RISCVFPRndMode::RDN << 4 * int(RoundingMode::TowardNegative)) |
7582       (RISCVFPRndMode::RUP << 4 * int(RoundingMode::TowardPositive)) |
7583       (RISCVFPRndMode::RMM << 4 * int(RoundingMode::NearestTiesToAway));
7584 
7585   SDValue Shift = DAG.getNode(ISD::SHL, DL, XLenVT, RMValue,
7586                               DAG.getConstant(2, DL, XLenVT));
7587   SDValue Shifted = DAG.getNode(ISD::SRL, DL, XLenVT,
7588                                 DAG.getConstant(Table, DL, XLenVT), Shift);
7589   RMValue = DAG.getNode(ISD::AND, DL, XLenVT, Shifted,
7590                         DAG.getConstant(0x7, DL, XLenVT));
7591   return DAG.getNode(RISCVISD::WRITE_CSR, DL, MVT::Other, Chain, SysRegNo,
7592                      RMValue);
7593 }
7594 
lowerEH_DWARF_CFA(SDValue Op,SelectionDAG & DAG) const7595 SDValue RISCVTargetLowering::lowerEH_DWARF_CFA(SDValue Op,
7596                                                SelectionDAG &DAG) const {
7597   MachineFunction &MF = DAG.getMachineFunction();
7598 
7599   bool isRISCV64 = Subtarget.is64Bit();
7600   EVT PtrVT = getPointerTy(DAG.getDataLayout());
7601 
7602   int FI = MF.getFrameInfo().CreateFixedObject(isRISCV64 ? 8 : 4, 0, false);
7603   return DAG.getFrameIndex(FI, PtrVT);
7604 }
7605 
7606 // Returns the opcode of the target-specific SDNode that implements the 32-bit
7607 // form of the given Opcode.
getRISCVWOpcode(unsigned Opcode)7608 static RISCVISD::NodeType getRISCVWOpcode(unsigned Opcode) {
7609   switch (Opcode) {
7610   default:
7611     llvm_unreachable("Unexpected opcode");
7612   case ISD::SHL:
7613     return RISCVISD::SLLW;
7614   case ISD::SRA:
7615     return RISCVISD::SRAW;
7616   case ISD::SRL:
7617     return RISCVISD::SRLW;
7618   case ISD::SDIV:
7619     return RISCVISD::DIVW;
7620   case ISD::UDIV:
7621     return RISCVISD::DIVUW;
7622   case ISD::UREM:
7623     return RISCVISD::REMUW;
7624   case ISD::ROTL:
7625     return RISCVISD::ROLW;
7626   case ISD::ROTR:
7627     return RISCVISD::RORW;
7628   }
7629 }
7630 
7631 // Converts the given i8/i16/i32 operation to a target-specific SelectionDAG
7632 // node. Because i8/i16/i32 isn't a legal type for RV64, these operations would
7633 // otherwise be promoted to i64, making it difficult to select the
7634 // SLLW/DIVUW/.../*W later one because the fact the operation was originally of
7635 // type i8/i16/i32 is lost.
customLegalizeToWOp(SDNode * N,SelectionDAG & DAG,unsigned ExtOpc=ISD::ANY_EXTEND)7636 static SDValue customLegalizeToWOp(SDNode *N, SelectionDAG &DAG,
7637                                    unsigned ExtOpc = ISD::ANY_EXTEND) {
7638   SDLoc DL(N);
7639   RISCVISD::NodeType WOpcode = getRISCVWOpcode(N->getOpcode());
7640   SDValue NewOp0 = DAG.getNode(ExtOpc, DL, MVT::i64, N->getOperand(0));
7641   SDValue NewOp1 = DAG.getNode(ExtOpc, DL, MVT::i64, N->getOperand(1));
7642   SDValue NewRes = DAG.getNode(WOpcode, DL, MVT::i64, NewOp0, NewOp1);
7643   // ReplaceNodeResults requires we maintain the same type for the return value.
7644   return DAG.getNode(ISD::TRUNCATE, DL, N->getValueType(0), NewRes);
7645 }
7646 
7647 // Converts the given 32-bit operation to a i64 operation with signed extension
7648 // semantic to reduce the signed extension instructions.
customLegalizeToWOpWithSExt(SDNode * N,SelectionDAG & DAG)7649 static SDValue customLegalizeToWOpWithSExt(SDNode *N, SelectionDAG &DAG) {
7650   SDLoc DL(N);
7651   SDValue NewOp0 = DAG.getNode(ISD::ANY_EXTEND, DL, MVT::i64, N->getOperand(0));
7652   SDValue NewOp1 = DAG.getNode(ISD::ANY_EXTEND, DL, MVT::i64, N->getOperand(1));
7653   SDValue NewWOp = DAG.getNode(N->getOpcode(), DL, MVT::i64, NewOp0, NewOp1);
7654   SDValue NewRes = DAG.getNode(ISD::SIGN_EXTEND_INREG, DL, MVT::i64, NewWOp,
7655                                DAG.getValueType(MVT::i32));
7656   return DAG.getNode(ISD::TRUNCATE, DL, MVT::i32, NewRes);
7657 }
7658 
ReplaceNodeResults(SDNode * N,SmallVectorImpl<SDValue> & Results,SelectionDAG & DAG) const7659 void RISCVTargetLowering::ReplaceNodeResults(SDNode *N,
7660                                              SmallVectorImpl<SDValue> &Results,
7661                                              SelectionDAG &DAG) const {
7662   SDLoc DL(N);
7663   switch (N->getOpcode()) {
7664   default:
7665     llvm_unreachable("Don't know how to custom type legalize this operation!");
7666   case ISD::STRICT_FP_TO_SINT:
7667   case ISD::STRICT_FP_TO_UINT:
7668   case ISD::FP_TO_SINT:
7669   case ISD::FP_TO_UINT: {
7670     assert(N->getValueType(0) == MVT::i32 && Subtarget.is64Bit() &&
7671            "Unexpected custom legalisation");
7672     bool IsStrict = N->isStrictFPOpcode();
7673     bool IsSigned = N->getOpcode() == ISD::FP_TO_SINT ||
7674                     N->getOpcode() == ISD::STRICT_FP_TO_SINT;
7675     SDValue Op0 = IsStrict ? N->getOperand(1) : N->getOperand(0);
7676     if (getTypeAction(*DAG.getContext(), Op0.getValueType()) !=
7677         TargetLowering::TypeSoftenFloat) {
7678       if (!isTypeLegal(Op0.getValueType()))
7679         return;
7680       if (IsStrict) {
7681         SDValue Chain = N->getOperand(0);
7682         // In absense of Zfh, promote f16 to f32, then convert.
7683         if (Op0.getValueType() == MVT::f16 && !Subtarget.hasStdExtZfh()) {
7684           Op0 = DAG.getNode(ISD::STRICT_FP_EXTEND, DL, {MVT::f32, MVT::Other},
7685                             {Chain, Op0});
7686           Chain = Op0.getValue(1);
7687         }
7688         unsigned Opc = IsSigned ? RISCVISD::STRICT_FCVT_W_RV64
7689                                 : RISCVISD::STRICT_FCVT_WU_RV64;
7690         SDVTList VTs = DAG.getVTList(MVT::i64, MVT::Other);
7691         SDValue Res = DAG.getNode(
7692             Opc, DL, VTs, Chain, Op0,
7693             DAG.getTargetConstant(RISCVFPRndMode::RTZ, DL, MVT::i64));
7694         Results.push_back(DAG.getNode(ISD::TRUNCATE, DL, MVT::i32, Res));
7695         Results.push_back(Res.getValue(1));
7696         return;
7697       }
7698       // In absense of Zfh, promote f16 to f32, then convert.
7699       if (Op0.getValueType() == MVT::f16 && !Subtarget.hasStdExtZfh())
7700         Op0 = DAG.getNode(ISD::FP_EXTEND, DL, MVT::f32, Op0);
7701 
7702       unsigned Opc = IsSigned ? RISCVISD::FCVT_W_RV64 : RISCVISD::FCVT_WU_RV64;
7703       SDValue Res =
7704           DAG.getNode(Opc, DL, MVT::i64, Op0,
7705                       DAG.getTargetConstant(RISCVFPRndMode::RTZ, DL, MVT::i64));
7706       Results.push_back(DAG.getNode(ISD::TRUNCATE, DL, MVT::i32, Res));
7707       return;
7708     }
7709     // If the FP type needs to be softened, emit a library call using the 'si'
7710     // version. If we left it to default legalization we'd end up with 'di'. If
7711     // the FP type doesn't need to be softened just let generic type
7712     // legalization promote the result type.
7713     RTLIB::Libcall LC;
7714     if (IsSigned)
7715       LC = RTLIB::getFPTOSINT(Op0.getValueType(), N->getValueType(0));
7716     else
7717       LC = RTLIB::getFPTOUINT(Op0.getValueType(), N->getValueType(0));
7718     MakeLibCallOptions CallOptions;
7719     EVT OpVT = Op0.getValueType();
7720     CallOptions.setTypeListBeforeSoften(OpVT, N->getValueType(0), true);
7721     SDValue Chain = IsStrict ? N->getOperand(0) : SDValue();
7722     SDValue Result;
7723     std::tie(Result, Chain) =
7724         makeLibCall(DAG, LC, N->getValueType(0), Op0, CallOptions, DL, Chain);
7725     Results.push_back(Result);
7726     if (IsStrict)
7727       Results.push_back(Chain);
7728     break;
7729   }
7730   case ISD::READCYCLECOUNTER: {
7731     assert(!Subtarget.is64Bit() &&
7732            "READCYCLECOUNTER only has custom type legalization on riscv32");
7733 
7734     SDVTList VTs = DAG.getVTList(MVT::i32, MVT::i32, MVT::Other);
7735     SDValue RCW =
7736         DAG.getNode(RISCVISD::READ_CYCLE_WIDE, DL, VTs, N->getOperand(0));
7737 
7738     Results.push_back(
7739         DAG.getNode(ISD::BUILD_PAIR, DL, MVT::i64, RCW, RCW.getValue(1)));
7740     Results.push_back(RCW.getValue(2));
7741     break;
7742   }
7743   case ISD::LOAD: {
7744     if (!ISD::isNON_EXTLoad(N))
7745       return;
7746 
7747     // Use a SEXTLOAD instead of the default EXTLOAD. Similar to the
7748     // sext_inreg we emit for ADD/SUB/MUL/SLLI.
7749     LoadSDNode *Ld = cast<LoadSDNode>(N);
7750 
7751     SDLoc dl(N);
7752     SDValue Res = DAG.getExtLoad(ISD::SEXTLOAD, dl, MVT::i64, Ld->getChain(),
7753                                  Ld->getBasePtr(), Ld->getMemoryVT(),
7754                                  Ld->getMemOperand());
7755     Results.push_back(DAG.getNode(ISD::TRUNCATE, dl, MVT::i32, Res));
7756     Results.push_back(Res.getValue(1));
7757     return;
7758   }
7759   case ISD::MUL: {
7760     unsigned Size = N->getSimpleValueType(0).getSizeInBits();
7761     unsigned XLen = Subtarget.getXLen();
7762     // This multiply needs to be expanded, try to use MULHSU+MUL if possible.
7763     if (Size > XLen) {
7764       assert(Size == (XLen * 2) && "Unexpected custom legalisation");
7765       SDValue LHS = N->getOperand(0);
7766       SDValue RHS = N->getOperand(1);
7767       APInt HighMask = APInt::getHighBitsSet(Size, XLen);
7768 
7769       bool LHSIsU = DAG.MaskedValueIsZero(LHS, HighMask);
7770       bool RHSIsU = DAG.MaskedValueIsZero(RHS, HighMask);
7771       // We need exactly one side to be unsigned.
7772       if (LHSIsU == RHSIsU)
7773         return;
7774 
7775       auto MakeMULPair = [&](SDValue S, SDValue U) {
7776         MVT XLenVT = Subtarget.getXLenVT();
7777         S = DAG.getNode(ISD::TRUNCATE, DL, XLenVT, S);
7778         U = DAG.getNode(ISD::TRUNCATE, DL, XLenVT, U);
7779         SDValue Lo = DAG.getNode(ISD::MUL, DL, XLenVT, S, U);
7780         SDValue Hi = DAG.getNode(RISCVISD::MULHSU, DL, XLenVT, S, U);
7781         return DAG.getNode(ISD::BUILD_PAIR, DL, N->getValueType(0), Lo, Hi);
7782       };
7783 
7784       bool LHSIsS = DAG.ComputeNumSignBits(LHS) > XLen;
7785       bool RHSIsS = DAG.ComputeNumSignBits(RHS) > XLen;
7786 
7787       // The other operand should be signed, but still prefer MULH when
7788       // possible.
7789       if (RHSIsU && LHSIsS && !RHSIsS)
7790         Results.push_back(MakeMULPair(LHS, RHS));
7791       else if (LHSIsU && RHSIsS && !LHSIsS)
7792         Results.push_back(MakeMULPair(RHS, LHS));
7793 
7794       return;
7795     }
7796     [[fallthrough]];
7797   }
7798   case ISD::ADD:
7799   case ISD::SUB:
7800     assert(N->getValueType(0) == MVT::i32 && Subtarget.is64Bit() &&
7801            "Unexpected custom legalisation");
7802     Results.push_back(customLegalizeToWOpWithSExt(N, DAG));
7803     break;
7804   case ISD::SHL:
7805   case ISD::SRA:
7806   case ISD::SRL:
7807     assert(N->getValueType(0) == MVT::i32 && Subtarget.is64Bit() &&
7808            "Unexpected custom legalisation");
7809     if (N->getOperand(1).getOpcode() != ISD::Constant) {
7810       // If we can use a BSET instruction, allow default promotion to apply.
7811       if (N->getOpcode() == ISD::SHL && Subtarget.hasStdExtZbs() &&
7812           isOneConstant(N->getOperand(0)))
7813         break;
7814       Results.push_back(customLegalizeToWOp(N, DAG));
7815       break;
7816     }
7817 
7818     // Custom legalize ISD::SHL by placing a SIGN_EXTEND_INREG after. This is
7819     // similar to customLegalizeToWOpWithSExt, but we must zero_extend the
7820     // shift amount.
7821     if (N->getOpcode() == ISD::SHL) {
7822       SDLoc DL(N);
7823       SDValue NewOp0 =
7824           DAG.getNode(ISD::ANY_EXTEND, DL, MVT::i64, N->getOperand(0));
7825       SDValue NewOp1 =
7826           DAG.getNode(ISD::ZERO_EXTEND, DL, MVT::i64, N->getOperand(1));
7827       SDValue NewWOp = DAG.getNode(ISD::SHL, DL, MVT::i64, NewOp0, NewOp1);
7828       SDValue NewRes = DAG.getNode(ISD::SIGN_EXTEND_INREG, DL, MVT::i64, NewWOp,
7829                                    DAG.getValueType(MVT::i32));
7830       Results.push_back(DAG.getNode(ISD::TRUNCATE, DL, MVT::i32, NewRes));
7831     }
7832 
7833     break;
7834   case ISD::ROTL:
7835   case ISD::ROTR:
7836     assert(N->getValueType(0) == MVT::i32 && Subtarget.is64Bit() &&
7837            "Unexpected custom legalisation");
7838     Results.push_back(customLegalizeToWOp(N, DAG));
7839     break;
7840   case ISD::CTTZ:
7841   case ISD::CTTZ_ZERO_UNDEF:
7842   case ISD::CTLZ:
7843   case ISD::CTLZ_ZERO_UNDEF: {
7844     assert(N->getValueType(0) == MVT::i32 && Subtarget.is64Bit() &&
7845            "Unexpected custom legalisation");
7846 
7847     SDValue NewOp0 =
7848         DAG.getNode(ISD::ANY_EXTEND, DL, MVT::i64, N->getOperand(0));
7849     bool IsCTZ =
7850         N->getOpcode() == ISD::CTTZ || N->getOpcode() == ISD::CTTZ_ZERO_UNDEF;
7851     unsigned Opc = IsCTZ ? RISCVISD::CTZW : RISCVISD::CLZW;
7852     SDValue Res = DAG.getNode(Opc, DL, MVT::i64, NewOp0);
7853     Results.push_back(DAG.getNode(ISD::TRUNCATE, DL, MVT::i32, Res));
7854     return;
7855   }
7856   case ISD::SDIV:
7857   case ISD::UDIV:
7858   case ISD::UREM: {
7859     MVT VT = N->getSimpleValueType(0);
7860     assert((VT == MVT::i8 || VT == MVT::i16 || VT == MVT::i32) &&
7861            Subtarget.is64Bit() && Subtarget.hasStdExtM() &&
7862            "Unexpected custom legalisation");
7863     // Don't promote division/remainder by constant since we should expand those
7864     // to multiply by magic constant.
7865     AttributeList Attr = DAG.getMachineFunction().getFunction().getAttributes();
7866     if (N->getOperand(1).getOpcode() == ISD::Constant &&
7867         !isIntDivCheap(N->getValueType(0), Attr))
7868       return;
7869 
7870     // If the input is i32, use ANY_EXTEND since the W instructions don't read
7871     // the upper 32 bits. For other types we need to sign or zero extend
7872     // based on the opcode.
7873     unsigned ExtOpc = ISD::ANY_EXTEND;
7874     if (VT != MVT::i32)
7875       ExtOpc = N->getOpcode() == ISD::SDIV ? ISD::SIGN_EXTEND
7876                                            : ISD::ZERO_EXTEND;
7877 
7878     Results.push_back(customLegalizeToWOp(N, DAG, ExtOpc));
7879     break;
7880   }
7881   case ISD::UADDO:
7882   case ISD::USUBO: {
7883     assert(N->getValueType(0) == MVT::i32 && Subtarget.is64Bit() &&
7884            "Unexpected custom legalisation");
7885     bool IsAdd = N->getOpcode() == ISD::UADDO;
7886     // Create an ADDW or SUBW.
7887     SDValue LHS = DAG.getNode(ISD::ANY_EXTEND, DL, MVT::i64, N->getOperand(0));
7888     SDValue RHS = DAG.getNode(ISD::ANY_EXTEND, DL, MVT::i64, N->getOperand(1));
7889     SDValue Res =
7890         DAG.getNode(IsAdd ? ISD::ADD : ISD::SUB, DL, MVT::i64, LHS, RHS);
7891     Res = DAG.getNode(ISD::SIGN_EXTEND_INREG, DL, MVT::i64, Res,
7892                       DAG.getValueType(MVT::i32));
7893 
7894     SDValue Overflow;
7895     if (IsAdd && isOneConstant(RHS)) {
7896       // Special case uaddo X, 1 overflowed if the addition result is 0.
7897       // The general case (X + C) < C is not necessarily beneficial. Although we
7898       // reduce the live range of X, we may introduce the materialization of
7899       // constant C, especially when the setcc result is used by branch. We have
7900       // no compare with constant and branch instructions.
7901       Overflow = DAG.getSetCC(DL, N->getValueType(1), Res,
7902                               DAG.getConstant(0, DL, MVT::i64), ISD::SETEQ);
7903     } else {
7904       // Sign extend the LHS and perform an unsigned compare with the ADDW
7905       // result. Since the inputs are sign extended from i32, this is equivalent
7906       // to comparing the lower 32 bits.
7907       LHS = DAG.getNode(ISD::SIGN_EXTEND, DL, MVT::i64, N->getOperand(0));
7908       Overflow = DAG.getSetCC(DL, N->getValueType(1), Res, LHS,
7909                               IsAdd ? ISD::SETULT : ISD::SETUGT);
7910     }
7911 
7912     Results.push_back(DAG.getNode(ISD::TRUNCATE, DL, MVT::i32, Res));
7913     Results.push_back(Overflow);
7914     return;
7915   }
7916   case ISD::UADDSAT:
7917   case ISD::USUBSAT: {
7918     assert(N->getValueType(0) == MVT::i32 && Subtarget.is64Bit() &&
7919            "Unexpected custom legalisation");
7920     if (Subtarget.hasStdExtZbb()) {
7921       // With Zbb we can sign extend and let LegalizeDAG use minu/maxu. Using
7922       // sign extend allows overflow of the lower 32 bits to be detected on
7923       // the promoted size.
7924       SDValue LHS =
7925           DAG.getNode(ISD::SIGN_EXTEND, DL, MVT::i64, N->getOperand(0));
7926       SDValue RHS =
7927           DAG.getNode(ISD::SIGN_EXTEND, DL, MVT::i64, N->getOperand(1));
7928       SDValue Res = DAG.getNode(N->getOpcode(), DL, MVT::i64, LHS, RHS);
7929       Results.push_back(DAG.getNode(ISD::TRUNCATE, DL, MVT::i32, Res));
7930       return;
7931     }
7932 
7933     // Without Zbb, expand to UADDO/USUBO+select which will trigger our custom
7934     // promotion for UADDO/USUBO.
7935     Results.push_back(expandAddSubSat(N, DAG));
7936     return;
7937   }
7938   case ISD::ABS: {
7939     assert(N->getValueType(0) == MVT::i32 && Subtarget.is64Bit() &&
7940            "Unexpected custom legalisation");
7941 
7942     if (Subtarget.hasStdExtZbb()) {
7943       // Emit a special ABSW node that will be expanded to NEGW+MAX at isel.
7944       // This allows us to remember that the result is sign extended. Expanding
7945       // to NEGW+MAX here requires a Freeze which breaks ComputeNumSignBits.
7946       SDValue Src = DAG.getNode(ISD::SIGN_EXTEND, DL, MVT::i64,
7947                                 N->getOperand(0));
7948       SDValue Abs = DAG.getNode(RISCVISD::ABSW, DL, MVT::i64, Src);
7949       Results.push_back(DAG.getNode(ISD::TRUNCATE, DL, MVT::i32, Abs));
7950       return;
7951     }
7952 
7953     // Expand abs to Y = (sraiw X, 31); subw(xor(X, Y), Y)
7954     SDValue Src = DAG.getNode(ISD::ANY_EXTEND, DL, MVT::i64, N->getOperand(0));
7955 
7956     // Freeze the source so we can increase it's use count.
7957     Src = DAG.getFreeze(Src);
7958 
7959     // Copy sign bit to all bits using the sraiw pattern.
7960     SDValue SignFill = DAG.getNode(ISD::SIGN_EXTEND_INREG, DL, MVT::i64, Src,
7961                                    DAG.getValueType(MVT::i32));
7962     SignFill = DAG.getNode(ISD::SRA, DL, MVT::i64, SignFill,
7963                            DAG.getConstant(31, DL, MVT::i64));
7964 
7965     SDValue NewRes = DAG.getNode(ISD::XOR, DL, MVT::i64, Src, SignFill);
7966     NewRes = DAG.getNode(ISD::SUB, DL, MVT::i64, NewRes, SignFill);
7967 
7968     // NOTE: The result is only required to be anyextended, but sext is
7969     // consistent with type legalization of sub.
7970     NewRes = DAG.getNode(ISD::SIGN_EXTEND_INREG, DL, MVT::i64, NewRes,
7971                          DAG.getValueType(MVT::i32));
7972     Results.push_back(DAG.getNode(ISD::TRUNCATE, DL, MVT::i32, NewRes));
7973     return;
7974   }
7975   case ISD::BITCAST: {
7976     EVT VT = N->getValueType(0);
7977     assert(VT.isInteger() && !VT.isVector() && "Unexpected VT!");
7978     SDValue Op0 = N->getOperand(0);
7979     EVT Op0VT = Op0.getValueType();
7980     MVT XLenVT = Subtarget.getXLenVT();
7981     if (VT == MVT::i16 && Op0VT == MVT::f16 &&
7982         Subtarget.hasStdExtZfhOrZfhmin()) {
7983       SDValue FPConv = DAG.getNode(RISCVISD::FMV_X_ANYEXTH, DL, XLenVT, Op0);
7984       Results.push_back(DAG.getNode(ISD::TRUNCATE, DL, MVT::i16, FPConv));
7985     } else if (VT == MVT::i32 && Op0VT == MVT::f32 && Subtarget.is64Bit() &&
7986                Subtarget.hasStdExtF()) {
7987       SDValue FPConv =
7988           DAG.getNode(RISCVISD::FMV_X_ANYEXTW_RV64, DL, MVT::i64, Op0);
7989       Results.push_back(DAG.getNode(ISD::TRUNCATE, DL, MVT::i32, FPConv));
7990     } else if (!VT.isVector() && Op0VT.isFixedLengthVector() &&
7991                isTypeLegal(Op0VT)) {
7992       // Custom-legalize bitcasts from fixed-length vector types to illegal
7993       // scalar types in order to improve codegen. Bitcast the vector to a
7994       // one-element vector type whose element type is the same as the result
7995       // type, and extract the first element.
7996       EVT BVT = EVT::getVectorVT(*DAG.getContext(), VT, 1);
7997       if (isTypeLegal(BVT)) {
7998         SDValue BVec = DAG.getBitcast(BVT, Op0);
7999         Results.push_back(DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, VT, BVec,
8000                                       DAG.getConstant(0, DL, XLenVT)));
8001       }
8002     }
8003     break;
8004   }
8005   case RISCVISD::BREV8: {
8006     MVT VT = N->getSimpleValueType(0);
8007     MVT XLenVT = Subtarget.getXLenVT();
8008     assert((VT == MVT::i16 || (VT == MVT::i32 && Subtarget.is64Bit())) &&
8009            "Unexpected custom legalisation");
8010     assert(Subtarget.hasStdExtZbkb() && "Unexpected extension");
8011     SDValue NewOp = DAG.getNode(ISD::ANY_EXTEND, DL, XLenVT, N->getOperand(0));
8012     SDValue NewRes = DAG.getNode(N->getOpcode(), DL, XLenVT, NewOp);
8013     // ReplaceNodeResults requires we maintain the same type for the return
8014     // value.
8015     Results.push_back(DAG.getNode(ISD::TRUNCATE, DL, VT, NewRes));
8016     break;
8017   }
8018   case ISD::EXTRACT_VECTOR_ELT: {
8019     // Custom-legalize an EXTRACT_VECTOR_ELT where XLEN<SEW, as the SEW element
8020     // type is illegal (currently only vXi64 RV32).
8021     // With vmv.x.s, when SEW > XLEN, only the least-significant XLEN bits are
8022     // transferred to the destination register. We issue two of these from the
8023     // upper- and lower- halves of the SEW-bit vector element, slid down to the
8024     // first element.
8025     SDValue Vec = N->getOperand(0);
8026     SDValue Idx = N->getOperand(1);
8027 
8028     // The vector type hasn't been legalized yet so we can't issue target
8029     // specific nodes if it needs legalization.
8030     // FIXME: We would manually legalize if it's important.
8031     if (!isTypeLegal(Vec.getValueType()))
8032       return;
8033 
8034     MVT VecVT = Vec.getSimpleValueType();
8035 
8036     assert(!Subtarget.is64Bit() && N->getValueType(0) == MVT::i64 &&
8037            VecVT.getVectorElementType() == MVT::i64 &&
8038            "Unexpected EXTRACT_VECTOR_ELT legalization");
8039 
8040     // If this is a fixed vector, we need to convert it to a scalable vector.
8041     MVT ContainerVT = VecVT;
8042     if (VecVT.isFixedLengthVector()) {
8043       ContainerVT = getContainerForFixedLengthVector(VecVT);
8044       Vec = convertToScalableVector(ContainerVT, Vec, DAG, Subtarget);
8045     }
8046 
8047     MVT XLenVT = Subtarget.getXLenVT();
8048 
8049     // Use a VL of 1 to avoid processing more elements than we need.
8050     auto [Mask, VL] = getDefaultVLOps(1, ContainerVT, DL, DAG, Subtarget);
8051 
8052     // Unless the index is known to be 0, we must slide the vector down to get
8053     // the desired element into index 0.
8054     if (!isNullConstant(Idx)) {
8055       Vec = getVSlidedown(DAG, Subtarget, DL, ContainerVT,
8056                           DAG.getUNDEF(ContainerVT), Vec, Idx, Mask, VL);
8057     }
8058 
8059     // Extract the lower XLEN bits of the correct vector element.
8060     SDValue EltLo = DAG.getNode(RISCVISD::VMV_X_S, DL, XLenVT, Vec);
8061 
8062     // To extract the upper XLEN bits of the vector element, shift the first
8063     // element right by 32 bits and re-extract the lower XLEN bits.
8064     SDValue ThirtyTwoV = DAG.getNode(RISCVISD::VMV_V_X_VL, DL, ContainerVT,
8065                                      DAG.getUNDEF(ContainerVT),
8066                                      DAG.getConstant(32, DL, XLenVT), VL);
8067     SDValue LShr32 =
8068         DAG.getNode(RISCVISD::SRL_VL, DL, ContainerVT, Vec, ThirtyTwoV,
8069                     DAG.getUNDEF(ContainerVT), Mask, VL);
8070 
8071     SDValue EltHi = DAG.getNode(RISCVISD::VMV_X_S, DL, XLenVT, LShr32);
8072 
8073     Results.push_back(DAG.getNode(ISD::BUILD_PAIR, DL, MVT::i64, EltLo, EltHi));
8074     break;
8075   }
8076   case ISD::INTRINSIC_WO_CHAIN: {
8077     unsigned IntNo = cast<ConstantSDNode>(N->getOperand(0))->getZExtValue();
8078     switch (IntNo) {
8079     default:
8080       llvm_unreachable(
8081           "Don't know how to custom type legalize this intrinsic!");
8082     case Intrinsic::riscv_orc_b: {
8083       SDValue NewOp =
8084           DAG.getNode(ISD::ANY_EXTEND, DL, MVT::i64, N->getOperand(1));
8085       SDValue Res = DAG.getNode(RISCVISD::ORC_B, DL, MVT::i64, NewOp);
8086       Results.push_back(DAG.getNode(ISD::TRUNCATE, DL, MVT::i32, Res));
8087       return;
8088     }
8089     case Intrinsic::riscv_vmv_x_s: {
8090       EVT VT = N->getValueType(0);
8091       MVT XLenVT = Subtarget.getXLenVT();
8092       if (VT.bitsLT(XLenVT)) {
8093         // Simple case just extract using vmv.x.s and truncate.
8094         SDValue Extract = DAG.getNode(RISCVISD::VMV_X_S, DL,
8095                                       Subtarget.getXLenVT(), N->getOperand(1));
8096         Results.push_back(DAG.getNode(ISD::TRUNCATE, DL, VT, Extract));
8097         return;
8098       }
8099 
8100       assert(VT == MVT::i64 && !Subtarget.is64Bit() &&
8101              "Unexpected custom legalization");
8102 
8103       // We need to do the move in two steps.
8104       SDValue Vec = N->getOperand(1);
8105       MVT VecVT = Vec.getSimpleValueType();
8106 
8107       // First extract the lower XLEN bits of the element.
8108       SDValue EltLo = DAG.getNode(RISCVISD::VMV_X_S, DL, XLenVT, Vec);
8109 
8110       // To extract the upper XLEN bits of the vector element, shift the first
8111       // element right by 32 bits and re-extract the lower XLEN bits.
8112       auto [Mask, VL] = getDefaultVLOps(1, VecVT, DL, DAG, Subtarget);
8113 
8114       SDValue ThirtyTwoV =
8115           DAG.getNode(RISCVISD::VMV_V_X_VL, DL, VecVT, DAG.getUNDEF(VecVT),
8116                       DAG.getConstant(32, DL, XLenVT), VL);
8117       SDValue LShr32 = DAG.getNode(RISCVISD::SRL_VL, DL, VecVT, Vec, ThirtyTwoV,
8118                                    DAG.getUNDEF(VecVT), Mask, VL);
8119       SDValue EltHi = DAG.getNode(RISCVISD::VMV_X_S, DL, XLenVT, LShr32);
8120 
8121       Results.push_back(
8122           DAG.getNode(ISD::BUILD_PAIR, DL, MVT::i64, EltLo, EltHi));
8123       break;
8124     }
8125     }
8126     break;
8127   }
8128   case ISD::VECREDUCE_ADD:
8129   case ISD::VECREDUCE_AND:
8130   case ISD::VECREDUCE_OR:
8131   case ISD::VECREDUCE_XOR:
8132   case ISD::VECREDUCE_SMAX:
8133   case ISD::VECREDUCE_UMAX:
8134   case ISD::VECREDUCE_SMIN:
8135   case ISD::VECREDUCE_UMIN:
8136     if (SDValue V = lowerVECREDUCE(SDValue(N, 0), DAG))
8137       Results.push_back(V);
8138     break;
8139   case ISD::VP_REDUCE_ADD:
8140   case ISD::VP_REDUCE_AND:
8141   case ISD::VP_REDUCE_OR:
8142   case ISD::VP_REDUCE_XOR:
8143   case ISD::VP_REDUCE_SMAX:
8144   case ISD::VP_REDUCE_UMAX:
8145   case ISD::VP_REDUCE_SMIN:
8146   case ISD::VP_REDUCE_UMIN:
8147     if (SDValue V = lowerVPREDUCE(SDValue(N, 0), DAG))
8148       Results.push_back(V);
8149     break;
8150   case ISD::GET_ROUNDING: {
8151     SDVTList VTs = DAG.getVTList(Subtarget.getXLenVT(), MVT::Other);
8152     SDValue Res = DAG.getNode(ISD::GET_ROUNDING, DL, VTs, N->getOperand(0));
8153     Results.push_back(Res.getValue(0));
8154     Results.push_back(Res.getValue(1));
8155     break;
8156   }
8157   }
8158 }
8159 
8160 // Try to fold (<bop> x, (reduction.<bop> vec, start))
combineBinOpToReduce(SDNode * N,SelectionDAG & DAG,const RISCVSubtarget & Subtarget)8161 static SDValue combineBinOpToReduce(SDNode *N, SelectionDAG &DAG,
8162                                     const RISCVSubtarget &Subtarget) {
8163   auto BinOpToRVVReduce = [](unsigned Opc) {
8164     switch (Opc) {
8165     default:
8166       llvm_unreachable("Unhandled binary to transfrom reduction");
8167     case ISD::ADD:
8168       return RISCVISD::VECREDUCE_ADD_VL;
8169     case ISD::UMAX:
8170       return RISCVISD::VECREDUCE_UMAX_VL;
8171     case ISD::SMAX:
8172       return RISCVISD::VECREDUCE_SMAX_VL;
8173     case ISD::UMIN:
8174       return RISCVISD::VECREDUCE_UMIN_VL;
8175     case ISD::SMIN:
8176       return RISCVISD::VECREDUCE_SMIN_VL;
8177     case ISD::AND:
8178       return RISCVISD::VECREDUCE_AND_VL;
8179     case ISD::OR:
8180       return RISCVISD::VECREDUCE_OR_VL;
8181     case ISD::XOR:
8182       return RISCVISD::VECREDUCE_XOR_VL;
8183     case ISD::FADD:
8184       return RISCVISD::VECREDUCE_FADD_VL;
8185     case ISD::FMAXNUM:
8186       return RISCVISD::VECREDUCE_FMAX_VL;
8187     case ISD::FMINNUM:
8188       return RISCVISD::VECREDUCE_FMIN_VL;
8189     }
8190   };
8191 
8192   auto IsReduction = [&BinOpToRVVReduce](SDValue V, unsigned Opc) {
8193     return V.getOpcode() == ISD::EXTRACT_VECTOR_ELT &&
8194            isNullConstant(V.getOperand(1)) &&
8195            V.getOperand(0).getOpcode() == BinOpToRVVReduce(Opc);
8196   };
8197 
8198   unsigned Opc = N->getOpcode();
8199   unsigned ReduceIdx;
8200   if (IsReduction(N->getOperand(0), Opc))
8201     ReduceIdx = 0;
8202   else if (IsReduction(N->getOperand(1), Opc))
8203     ReduceIdx = 1;
8204   else
8205     return SDValue();
8206 
8207   // Skip if FADD disallows reassociation but the combiner needs.
8208   if (Opc == ISD::FADD && !N->getFlags().hasAllowReassociation())
8209     return SDValue();
8210 
8211   SDValue Extract = N->getOperand(ReduceIdx);
8212   SDValue Reduce = Extract.getOperand(0);
8213   if (!Reduce.hasOneUse())
8214     return SDValue();
8215 
8216   SDValue ScalarV = Reduce.getOperand(2);
8217   EVT ScalarVT = ScalarV.getValueType();
8218   if (ScalarV.getOpcode() == ISD::INSERT_SUBVECTOR &&
8219       ScalarV.getOperand(0)->isUndef())
8220     ScalarV = ScalarV.getOperand(1);
8221 
8222   // Make sure that ScalarV is a splat with VL=1.
8223   if (ScalarV.getOpcode() != RISCVISD::VFMV_S_F_VL &&
8224       ScalarV.getOpcode() != RISCVISD::VMV_S_X_VL &&
8225       ScalarV.getOpcode() != RISCVISD::VMV_V_X_VL)
8226     return SDValue();
8227 
8228   if (!hasNonZeroAVL(ScalarV.getOperand(2)))
8229     return SDValue();
8230 
8231   // Check the scalar of ScalarV is neutral element
8232   // TODO: Deal with value other than neutral element.
8233   if (!isNeutralConstant(N->getOpcode(), N->getFlags(), ScalarV.getOperand(1),
8234                          0))
8235     return SDValue();
8236 
8237   if (!ScalarV.hasOneUse())
8238     return SDValue();
8239 
8240   SDValue NewStart = N->getOperand(1 - ReduceIdx);
8241 
8242   SDLoc DL(N);
8243   SDValue NewScalarV =
8244       lowerScalarInsert(NewStart, ScalarV.getOperand(2),
8245                         ScalarV.getSimpleValueType(), DL, DAG, Subtarget);
8246 
8247   // If we looked through an INSERT_SUBVECTOR we need to restore it.
8248   if (ScalarVT != ScalarV.getValueType())
8249     NewScalarV =
8250         DAG.getNode(ISD::INSERT_SUBVECTOR, DL, ScalarVT, DAG.getUNDEF(ScalarVT),
8251                     NewScalarV, DAG.getConstant(0, DL, Subtarget.getXLenVT()));
8252 
8253   SDValue NewReduce =
8254       DAG.getNode(Reduce.getOpcode(), DL, Reduce.getValueType(),
8255                   Reduce.getOperand(0), Reduce.getOperand(1), NewScalarV,
8256                   Reduce.getOperand(3), Reduce.getOperand(4));
8257   return DAG.getNode(Extract.getOpcode(), DL, Extract.getValueType(), NewReduce,
8258                      Extract.getOperand(1));
8259 }
8260 
8261 // Optimize (add (shl x, c0), (shl y, c1)) ->
8262 //          (SLLI (SH*ADD x, y), c0), if c1-c0 equals to [1|2|3].
transformAddShlImm(SDNode * N,SelectionDAG & DAG,const RISCVSubtarget & Subtarget)8263 static SDValue transformAddShlImm(SDNode *N, SelectionDAG &DAG,
8264                                   const RISCVSubtarget &Subtarget) {
8265   // Perform this optimization only in the zba extension.
8266   if (!Subtarget.hasStdExtZba())
8267     return SDValue();
8268 
8269   // Skip for vector types and larger types.
8270   EVT VT = N->getValueType(0);
8271   if (VT.isVector() || VT.getSizeInBits() > Subtarget.getXLen())
8272     return SDValue();
8273 
8274   // The two operand nodes must be SHL and have no other use.
8275   SDValue N0 = N->getOperand(0);
8276   SDValue N1 = N->getOperand(1);
8277   if (N0->getOpcode() != ISD::SHL || N1->getOpcode() != ISD::SHL ||
8278       !N0->hasOneUse() || !N1->hasOneUse())
8279     return SDValue();
8280 
8281   // Check c0 and c1.
8282   auto *N0C = dyn_cast<ConstantSDNode>(N0->getOperand(1));
8283   auto *N1C = dyn_cast<ConstantSDNode>(N1->getOperand(1));
8284   if (!N0C || !N1C)
8285     return SDValue();
8286   int64_t C0 = N0C->getSExtValue();
8287   int64_t C1 = N1C->getSExtValue();
8288   if (C0 <= 0 || C1 <= 0)
8289     return SDValue();
8290 
8291   // Skip if SH1ADD/SH2ADD/SH3ADD are not applicable.
8292   int64_t Bits = std::min(C0, C1);
8293   int64_t Diff = std::abs(C0 - C1);
8294   if (Diff != 1 && Diff != 2 && Diff != 3)
8295     return SDValue();
8296 
8297   // Build nodes.
8298   SDLoc DL(N);
8299   SDValue NS = (C0 < C1) ? N0->getOperand(0) : N1->getOperand(0);
8300   SDValue NL = (C0 > C1) ? N0->getOperand(0) : N1->getOperand(0);
8301   SDValue NA0 =
8302       DAG.getNode(ISD::SHL, DL, VT, NL, DAG.getConstant(Diff, DL, VT));
8303   SDValue NA1 = DAG.getNode(ISD::ADD, DL, VT, NA0, NS);
8304   return DAG.getNode(ISD::SHL, DL, VT, NA1, DAG.getConstant(Bits, DL, VT));
8305 }
8306 
8307 // Combine a constant select operand into its use:
8308 //
8309 // (and (select cond, -1, c), x)
8310 //   -> (select cond, x, (and x, c))  [AllOnes=1]
8311 // (or  (select cond, 0, c), x)
8312 //   -> (select cond, x, (or x, c))  [AllOnes=0]
8313 // (xor (select cond, 0, c), x)
8314 //   -> (select cond, x, (xor x, c))  [AllOnes=0]
8315 // (add (select cond, 0, c), x)
8316 //   -> (select cond, x, (add x, c))  [AllOnes=0]
8317 // (sub x, (select cond, 0, c))
8318 //   -> (select cond, x, (sub x, c))  [AllOnes=0]
combineSelectAndUse(SDNode * N,SDValue Slct,SDValue OtherOp,SelectionDAG & DAG,bool AllOnes,const RISCVSubtarget & Subtarget)8319 static SDValue combineSelectAndUse(SDNode *N, SDValue Slct, SDValue OtherOp,
8320                                    SelectionDAG &DAG, bool AllOnes,
8321                                    const RISCVSubtarget &Subtarget) {
8322   EVT VT = N->getValueType(0);
8323 
8324   // Skip vectors.
8325   if (VT.isVector())
8326     return SDValue();
8327 
8328   if (!Subtarget.hasShortForwardBranchOpt() ||
8329       (Slct.getOpcode() != ISD::SELECT &&
8330        Slct.getOpcode() != RISCVISD::SELECT_CC) ||
8331       !Slct.hasOneUse())
8332     return SDValue();
8333 
8334   auto isZeroOrAllOnes = [](SDValue N, bool AllOnes) {
8335     return AllOnes ? isAllOnesConstant(N) : isNullConstant(N);
8336   };
8337 
8338   bool SwapSelectOps;
8339   unsigned OpOffset = Slct.getOpcode() == RISCVISD::SELECT_CC ? 2 : 0;
8340   SDValue TrueVal = Slct.getOperand(1 + OpOffset);
8341   SDValue FalseVal = Slct.getOperand(2 + OpOffset);
8342   SDValue NonConstantVal;
8343   if (isZeroOrAllOnes(TrueVal, AllOnes)) {
8344     SwapSelectOps = false;
8345     NonConstantVal = FalseVal;
8346   } else if (isZeroOrAllOnes(FalseVal, AllOnes)) {
8347     SwapSelectOps = true;
8348     NonConstantVal = TrueVal;
8349   } else
8350     return SDValue();
8351 
8352   // Slct is now know to be the desired identity constant when CC is true.
8353   TrueVal = OtherOp;
8354   FalseVal = DAG.getNode(N->getOpcode(), SDLoc(N), VT, OtherOp, NonConstantVal);
8355   // Unless SwapSelectOps says the condition should be false.
8356   if (SwapSelectOps)
8357     std::swap(TrueVal, FalseVal);
8358 
8359   if (Slct.getOpcode() == RISCVISD::SELECT_CC)
8360     return DAG.getNode(RISCVISD::SELECT_CC, SDLoc(N), VT,
8361                        {Slct.getOperand(0), Slct.getOperand(1),
8362                         Slct.getOperand(2), TrueVal, FalseVal});
8363 
8364   return DAG.getNode(ISD::SELECT, SDLoc(N), VT,
8365                      {Slct.getOperand(0), TrueVal, FalseVal});
8366 }
8367 
8368 // Attempt combineSelectAndUse on each operand of a commutative operator N.
combineSelectAndUseCommutative(SDNode * N,SelectionDAG & DAG,bool AllOnes,const RISCVSubtarget & Subtarget)8369 static SDValue combineSelectAndUseCommutative(SDNode *N, SelectionDAG &DAG,
8370                                               bool AllOnes,
8371                                               const RISCVSubtarget &Subtarget) {
8372   SDValue N0 = N->getOperand(0);
8373   SDValue N1 = N->getOperand(1);
8374   if (SDValue Result = combineSelectAndUse(N, N0, N1, DAG, AllOnes, Subtarget))
8375     return Result;
8376   if (SDValue Result = combineSelectAndUse(N, N1, N0, DAG, AllOnes, Subtarget))
8377     return Result;
8378   return SDValue();
8379 }
8380 
8381 // Transform (add (mul x, c0), c1) ->
8382 //           (add (mul (add x, c1/c0), c0), c1%c0).
8383 // if c1/c0 and c1%c0 are simm12, while c1 is not. A special corner case
8384 // that should be excluded is when c0*(c1/c0) is simm12, which will lead
8385 // to an infinite loop in DAGCombine if transformed.
8386 // Or transform (add (mul x, c0), c1) ->
8387 //              (add (mul (add x, c1/c0+1), c0), c1%c0-c0),
8388 // if c1/c0+1 and c1%c0-c0 are simm12, while c1 is not. A special corner
8389 // case that should be excluded is when c0*(c1/c0+1) is simm12, which will
8390 // lead to an infinite loop in DAGCombine if transformed.
8391 // Or transform (add (mul x, c0), c1) ->
8392 //              (add (mul (add x, c1/c0-1), c0), c1%c0+c0),
8393 // if c1/c0-1 and c1%c0+c0 are simm12, while c1 is not. A special corner
8394 // case that should be excluded is when c0*(c1/c0-1) is simm12, which will
8395 // lead to an infinite loop in DAGCombine if transformed.
8396 // Or transform (add (mul x, c0), c1) ->
8397 //              (mul (add x, c1/c0), c0).
8398 // if c1%c0 is zero, and c1/c0 is simm12 while c1 is not.
transformAddImmMulImm(SDNode * N,SelectionDAG & DAG,const RISCVSubtarget & Subtarget)8399 static SDValue transformAddImmMulImm(SDNode *N, SelectionDAG &DAG,
8400                                      const RISCVSubtarget &Subtarget) {
8401   // Skip for vector types and larger types.
8402   EVT VT = N->getValueType(0);
8403   if (VT.isVector() || VT.getSizeInBits() > Subtarget.getXLen())
8404     return SDValue();
8405   // The first operand node must be a MUL and has no other use.
8406   SDValue N0 = N->getOperand(0);
8407   if (!N0->hasOneUse() || N0->getOpcode() != ISD::MUL)
8408     return SDValue();
8409   // Check if c0 and c1 match above conditions.
8410   auto *N0C = dyn_cast<ConstantSDNode>(N0->getOperand(1));
8411   auto *N1C = dyn_cast<ConstantSDNode>(N->getOperand(1));
8412   if (!N0C || !N1C)
8413     return SDValue();
8414   // If N0C has multiple uses it's possible one of the cases in
8415   // DAGCombiner::isMulAddWithConstProfitable will be true, which would result
8416   // in an infinite loop.
8417   if (!N0C->hasOneUse())
8418     return SDValue();
8419   int64_t C0 = N0C->getSExtValue();
8420   int64_t C1 = N1C->getSExtValue();
8421   int64_t CA, CB;
8422   if (C0 == -1 || C0 == 0 || C0 == 1 || isInt<12>(C1))
8423     return SDValue();
8424   // Search for proper CA (non-zero) and CB that both are simm12.
8425   if ((C1 / C0) != 0 && isInt<12>(C1 / C0) && isInt<12>(C1 % C0) &&
8426       !isInt<12>(C0 * (C1 / C0))) {
8427     CA = C1 / C0;
8428     CB = C1 % C0;
8429   } else if ((C1 / C0 + 1) != 0 && isInt<12>(C1 / C0 + 1) &&
8430              isInt<12>(C1 % C0 - C0) && !isInt<12>(C0 * (C1 / C0 + 1))) {
8431     CA = C1 / C0 + 1;
8432     CB = C1 % C0 - C0;
8433   } else if ((C1 / C0 - 1) != 0 && isInt<12>(C1 / C0 - 1) &&
8434              isInt<12>(C1 % C0 + C0) && !isInt<12>(C0 * (C1 / C0 - 1))) {
8435     CA = C1 / C0 - 1;
8436     CB = C1 % C0 + C0;
8437   } else
8438     return SDValue();
8439   // Build new nodes (add (mul (add x, c1/c0), c0), c1%c0).
8440   SDLoc DL(N);
8441   SDValue New0 = DAG.getNode(ISD::ADD, DL, VT, N0->getOperand(0),
8442                              DAG.getConstant(CA, DL, VT));
8443   SDValue New1 =
8444       DAG.getNode(ISD::MUL, DL, VT, New0, DAG.getConstant(C0, DL, VT));
8445   return DAG.getNode(ISD::ADD, DL, VT, New1, DAG.getConstant(CB, DL, VT));
8446 }
8447 
performADDCombine(SDNode * N,SelectionDAG & DAG,const RISCVSubtarget & Subtarget)8448 static SDValue performADDCombine(SDNode *N, SelectionDAG &DAG,
8449                                  const RISCVSubtarget &Subtarget) {
8450   if (SDValue V = transformAddImmMulImm(N, DAG, Subtarget))
8451     return V;
8452   if (SDValue V = transformAddShlImm(N, DAG, Subtarget))
8453     return V;
8454   if (SDValue V = combineBinOpToReduce(N, DAG, Subtarget))
8455     return V;
8456   // fold (add (select lhs, rhs, cc, 0, y), x) ->
8457   //      (select lhs, rhs, cc, x, (add x, y))
8458   return combineSelectAndUseCommutative(N, DAG, /*AllOnes*/ false, Subtarget);
8459 }
8460 
8461 // Try to turn a sub boolean RHS and constant LHS into an addi.
combineSubOfBoolean(SDNode * N,SelectionDAG & DAG)8462 static SDValue combineSubOfBoolean(SDNode *N, SelectionDAG &DAG) {
8463   SDValue N0 = N->getOperand(0);
8464   SDValue N1 = N->getOperand(1);
8465   EVT VT = N->getValueType(0);
8466   SDLoc DL(N);
8467 
8468   // Require a constant LHS.
8469   auto *N0C = dyn_cast<ConstantSDNode>(N0);
8470   if (!N0C)
8471     return SDValue();
8472 
8473   // All our optimizations involve subtracting 1 from the immediate and forming
8474   // an ADDI. Make sure the new immediate is valid for an ADDI.
8475   APInt ImmValMinus1 = N0C->getAPIntValue() - 1;
8476   if (!ImmValMinus1.isSignedIntN(12))
8477     return SDValue();
8478 
8479   SDValue NewLHS;
8480   if (N1.getOpcode() == ISD::SETCC && N1.hasOneUse()) {
8481     // (sub constant, (setcc x, y, eq/neq)) ->
8482     // (add (setcc x, y, neq/eq), constant - 1)
8483     ISD::CondCode CCVal = cast<CondCodeSDNode>(N1.getOperand(2))->get();
8484     EVT SetCCOpVT = N1.getOperand(0).getValueType();
8485     if (!isIntEqualitySetCC(CCVal) || !SetCCOpVT.isInteger())
8486       return SDValue();
8487     CCVal = ISD::getSetCCInverse(CCVal, SetCCOpVT);
8488     NewLHS =
8489         DAG.getSetCC(SDLoc(N1), VT, N1.getOperand(0), N1.getOperand(1), CCVal);
8490   } else if (N1.getOpcode() == ISD::XOR && isOneConstant(N1.getOperand(1)) &&
8491              N1.getOperand(0).getOpcode() == ISD::SETCC) {
8492     // (sub C, (xor (setcc), 1)) -> (add (setcc), C-1).
8493     // Since setcc returns a bool the xor is equivalent to 1-setcc.
8494     NewLHS = N1.getOperand(0);
8495   } else
8496     return SDValue();
8497 
8498   SDValue NewRHS = DAG.getConstant(ImmValMinus1, DL, VT);
8499   return DAG.getNode(ISD::ADD, DL, VT, NewLHS, NewRHS);
8500 }
8501 
performSUBCombine(SDNode * N,SelectionDAG & DAG,const RISCVSubtarget & Subtarget)8502 static SDValue performSUBCombine(SDNode *N, SelectionDAG &DAG,
8503                                  const RISCVSubtarget &Subtarget) {
8504   if (SDValue V = combineSubOfBoolean(N, DAG))
8505     return V;
8506 
8507   // fold (sub x, (select lhs, rhs, cc, 0, y)) ->
8508   //      (select lhs, rhs, cc, x, (sub x, y))
8509   SDValue N0 = N->getOperand(0);
8510   SDValue N1 = N->getOperand(1);
8511   return combineSelectAndUse(N, N1, N0, DAG, /*AllOnes*/ false, Subtarget);
8512 }
8513 
8514 // Apply DeMorgan's law to (and/or (xor X, 1), (xor Y, 1)) if X and Y are 0/1.
8515 // Legalizing setcc can introduce xors like this. Doing this transform reduces
8516 // the number of xors and may allow the xor to fold into a branch condition.
combineDeMorganOfBoolean(SDNode * N,SelectionDAG & DAG)8517 static SDValue combineDeMorganOfBoolean(SDNode *N, SelectionDAG &DAG) {
8518   SDValue N0 = N->getOperand(0);
8519   SDValue N1 = N->getOperand(1);
8520   bool IsAnd = N->getOpcode() == ISD::AND;
8521 
8522   if (N0.getOpcode() != ISD::XOR || N1.getOpcode() != ISD::XOR)
8523     return SDValue();
8524 
8525   if (!N0.hasOneUse() || !N1.hasOneUse())
8526     return SDValue();
8527 
8528   SDValue N01 = N0.getOperand(1);
8529   SDValue N11 = N1.getOperand(1);
8530 
8531   // For AND, SimplifyDemandedBits may have turned one of the (xor X, 1) into
8532   // (xor X, -1) based on the upper bits of the other operand being 0. If the
8533   // operation is And, allow one of the Xors to use -1.
8534   if (isOneConstant(N01)) {
8535     if (!isOneConstant(N11) && !(IsAnd && isAllOnesConstant(N11)))
8536       return SDValue();
8537   } else if (isOneConstant(N11)) {
8538     // N01 and N11 being 1 was already handled. Handle N11==1 and N01==-1.
8539     if (!(IsAnd && isAllOnesConstant(N01)))
8540       return SDValue();
8541   } else
8542     return SDValue();
8543 
8544   EVT VT = N->getValueType(0);
8545 
8546   SDValue N00 = N0.getOperand(0);
8547   SDValue N10 = N1.getOperand(0);
8548 
8549   // The LHS of the xors needs to be 0/1.
8550   APInt Mask = APInt::getBitsSetFrom(VT.getSizeInBits(), 1);
8551   if (!DAG.MaskedValueIsZero(N00, Mask) || !DAG.MaskedValueIsZero(N10, Mask))
8552     return SDValue();
8553 
8554   // Invert the opcode and insert a new xor.
8555   SDLoc DL(N);
8556   unsigned Opc = IsAnd ? ISD::OR : ISD::AND;
8557   SDValue Logic = DAG.getNode(Opc, DL, VT, N00, N10);
8558   return DAG.getNode(ISD::XOR, DL, VT, Logic, DAG.getConstant(1, DL, VT));
8559 }
8560 
performTRUNCATECombine(SDNode * N,SelectionDAG & DAG,const RISCVSubtarget & Subtarget)8561 static SDValue performTRUNCATECombine(SDNode *N, SelectionDAG &DAG,
8562                                       const RISCVSubtarget &Subtarget) {
8563   SDValue N0 = N->getOperand(0);
8564   EVT VT = N->getValueType(0);
8565 
8566   // Pre-promote (i1 (truncate (srl X, Y))) on RV64 with Zbs without zero
8567   // extending X. This is safe since we only need the LSB after the shift and
8568   // shift amounts larger than 31 would produce poison. If we wait until
8569   // type legalization, we'll create RISCVISD::SRLW and we can't recover it
8570   // to use a BEXT instruction.
8571   if (Subtarget.is64Bit() && Subtarget.hasStdExtZbs() && VT == MVT::i1 &&
8572       N0.getValueType() == MVT::i32 && N0.getOpcode() == ISD::SRL &&
8573       !isa<ConstantSDNode>(N0.getOperand(1)) && N0.hasOneUse()) {
8574     SDLoc DL(N0);
8575     SDValue Op0 = DAG.getNode(ISD::ANY_EXTEND, DL, MVT::i64, N0.getOperand(0));
8576     SDValue Op1 = DAG.getNode(ISD::ZERO_EXTEND, DL, MVT::i64, N0.getOperand(1));
8577     SDValue Srl = DAG.getNode(ISD::SRL, DL, MVT::i64, Op0, Op1);
8578     return DAG.getNode(ISD::TRUNCATE, SDLoc(N), VT, Srl);
8579   }
8580 
8581   return SDValue();
8582 }
8583 
8584 namespace {
8585 // Helper class contains information about comparison operation.
8586 // The first two operands of this operation are compared values and the
8587 // last one is the operation.
8588 // Compared values are stored in Ops.
8589 // Comparison operation is stored in CCode.
8590 class CmpOpInfo {
8591   static unsigned constexpr Size = 2u;
8592 
8593   // Type for storing operands of compare operation.
8594   using OpsArray = std::array<SDValue, Size>;
8595   OpsArray Ops;
8596 
8597   using const_iterator = OpsArray::const_iterator;
begin() const8598   const_iterator begin() const { return Ops.begin(); }
end() const8599   const_iterator end() const { return Ops.end(); }
8600 
8601   ISD::CondCode CCode;
8602 
8603   unsigned CommonPos{Size};
8604   unsigned DifferPos{Size};
8605 
8606   // Sets CommonPos and DifferPos based on incoming position
8607   // of common operand CPos.
setPositions(const_iterator CPos)8608   void setPositions(const_iterator CPos) {
8609     assert(CPos != Ops.end() && "Common operand has to be in OpsArray.\n");
8610     CommonPos = CPos == Ops.begin() ? 0 : 1;
8611     DifferPos = 1 - CommonPos;
8612     assert((DifferPos == 0 || DifferPos == 1) &&
8613            "Positions can be only 0 or 1.");
8614   }
8615 
8616   // Private constructor of comparison info based on comparison operator.
8617   // It is private because CmpOpInfo only reasonable relative to other
8618   // comparison operator. Therefore, infos about comparison operation
8619   // have to be collected simultaneously via CmpOpInfo::getInfoAbout().
CmpOpInfo(const SDValue & CmpOp)8620   CmpOpInfo(const SDValue &CmpOp)
8621       : Ops{CmpOp.getOperand(0), CmpOp.getOperand(1)},
8622         CCode{cast<CondCodeSDNode>(CmpOp.getOperand(2))->get()} {}
8623 
8624   // Finds common operand of Op1 and Op2 and finishes filling CmpOpInfos.
8625   // Returns true if common operand is found. Otherwise - false.
establishCorrespondence(CmpOpInfo & Op1,CmpOpInfo & Op2)8626   static bool establishCorrespondence(CmpOpInfo &Op1, CmpOpInfo &Op2) {
8627     const auto CommonOpIt1 =
8628         std::find_first_of(Op1.begin(), Op1.end(), Op2.begin(), Op2.end());
8629     if (CommonOpIt1 == Op1.end())
8630       return false;
8631 
8632     const auto CommonOpIt2 = std::find(Op2.begin(), Op2.end(), *CommonOpIt1);
8633     assert(CommonOpIt2 != Op2.end() &&
8634            "Cannot find common operand in the second comparison operation.");
8635 
8636     Op1.setPositions(CommonOpIt1);
8637     Op2.setPositions(CommonOpIt2);
8638 
8639     return true;
8640   }
8641 
8642 public:
8643   CmpOpInfo(const CmpOpInfo &) = default;
8644   CmpOpInfo(CmpOpInfo &&) = default;
8645 
operator [](unsigned Pos) const8646   SDValue const &operator[](unsigned Pos) const {
8647     assert(Pos < Size && "Out of range\n");
8648     return Ops[Pos];
8649   }
8650 
8651   // Creates infos about comparison operations CmpOp0 and CmpOp1.
8652   // If there is no common operand returns None. Otherwise, returns
8653   // correspondence info about comparison operations.
8654   static std::optional<std::pair<CmpOpInfo, CmpOpInfo>>
getInfoAbout(SDValue const & CmpOp0,SDValue const & CmpOp1)8655   getInfoAbout(SDValue const &CmpOp0, SDValue const &CmpOp1) {
8656     CmpOpInfo Op0{CmpOp0};
8657     CmpOpInfo Op1{CmpOp1};
8658     if (!establishCorrespondence(Op0, Op1))
8659       return std::nullopt;
8660     return std::make_pair(Op0, Op1);
8661   }
8662 
8663   // Returns position of common operand.
getCPos() const8664   unsigned getCPos() const { return CommonPos; }
8665 
8666   // Returns position of differ operand.
getDPos() const8667   unsigned getDPos() const { return DifferPos; }
8668 
8669   // Returns common operand.
getCOp() const8670   SDValue const &getCOp() const { return operator[](CommonPos); }
8671 
8672   // Returns differ operand.
getDOp() const8673   SDValue const &getDOp() const { return operator[](DifferPos); }
8674 
8675   // Returns consition code of comparison operation.
getCondCode() const8676   ISD::CondCode getCondCode() const { return CCode; }
8677 };
8678 } // namespace
8679 
8680 // Verifies conditions to apply an optimization.
8681 // Returns Reference comparison code and three operands A, B, C.
8682 // Conditions for optimization:
8683 //   One operand of the compasions has to be common.
8684 //   This operand is written to C.
8685 //   Two others operands are differend. They are written to A and B.
8686 //   Comparisons has to be similar with respect to common operand C.
8687 //     e.g. A < C; C > B are similar
8688 //      but A < C; B > C are not.
8689 //   Reference comparison code is the comparison code if
8690 //   common operand is right placed.
8691 //     e.g. C > A will be swapped to A < C.
8692 static std::optional<std::tuple<ISD::CondCode, SDValue, SDValue, SDValue>>
verifyCompareConds(SDNode * N,SelectionDAG & DAG)8693 verifyCompareConds(SDNode *N, SelectionDAG &DAG) {
8694   LLVM_DEBUG(
8695       dbgs() << "Checking conditions for comparison operation combining.\n";);
8696 
8697   SDValue V0 = N->getOperand(0);
8698   SDValue V1 = N->getOperand(1);
8699   assert(V0.getValueType() == V1.getValueType() &&
8700          "Operations must have the same value type.");
8701 
8702   // Condition 1. Operations have to be used only in logic operation.
8703   if (!V0.hasOneUse() || !V1.hasOneUse())
8704     return std::nullopt;
8705 
8706   // Condition 2. Operands have to be comparison operations.
8707   if (V0.getOpcode() != ISD::SETCC || V1.getOpcode() != ISD::SETCC)
8708     return std::nullopt;
8709 
8710   // Condition 3.1. Operations only with integers.
8711   if (!V0.getOperand(0).getValueType().isInteger())
8712     return std::nullopt;
8713 
8714   const auto ComparisonInfo = CmpOpInfo::getInfoAbout(V0, V1);
8715   // Condition 3.2. Common operand has to be in comparison.
8716   if (!ComparisonInfo)
8717     return std::nullopt;
8718 
8719   const auto [Op0, Op1] = ComparisonInfo.value();
8720 
8721   LLVM_DEBUG(dbgs() << "Shared operands are on positions: " << Op0.getCPos()
8722                     << " and " << Op1.getCPos() << '\n';);
8723   // If common operand at the first position then swap operation to convert to
8724   // strict pattern. Common operand has to be right hand side.
8725   ISD::CondCode RefCond = Op0.getCondCode();
8726   ISD::CondCode AssistCode = Op1.getCondCode();
8727   if (!Op0.getCPos())
8728     RefCond = ISD::getSetCCSwappedOperands(RefCond);
8729   if (!Op1.getCPos())
8730     AssistCode = ISD::getSetCCSwappedOperands(AssistCode);
8731   LLVM_DEBUG(dbgs() << "Reference condition is: " << RefCond << '\n';);
8732   // If there are different comparison operations then do not perform an
8733   // optimization. a < c; c < b -> will be changed to b > c.
8734   if (RefCond != AssistCode)
8735     return std::nullopt;
8736 
8737   // Conditions can be only similar to Less or Greater. (>, >=, <, <=)
8738   // Applying this mask to the operation will determine Less and Greater
8739   // operations.
8740   const unsigned CmpMask = 0b110;
8741   const unsigned MaskedOpcode = CmpMask & RefCond;
8742   // If masking gave 0b110, then this is an operation NE, O or TRUE.
8743   if (MaskedOpcode == CmpMask)
8744     return std::nullopt;
8745   // If masking gave 00000, then this is an operation E, O or FALSE.
8746   if (MaskedOpcode == 0)
8747     return std::nullopt;
8748   // Everything else is similar to Less or Greater.
8749 
8750   SDValue A = Op0.getDOp();
8751   SDValue B = Op1.getDOp();
8752   SDValue C = Op0.getCOp();
8753 
8754   LLVM_DEBUG(
8755       dbgs() << "The conditions for combining comparisons are satisfied.\n";);
8756   return std::make_tuple(RefCond, A, B, C);
8757 }
8758 
getSelectionCode(bool IsUnsigned,bool IsAnd,bool IsGreaterOp)8759 static ISD::NodeType getSelectionCode(bool IsUnsigned, bool IsAnd,
8760                                       bool IsGreaterOp) {
8761   // Codes of selection operation. The first index selects signed or unsigned,
8762   // the second index selects MIN/MAX.
8763   static constexpr ISD::NodeType SelectionCodes[2][2] = {
8764       {ISD::SMIN, ISD::SMAX}, {ISD::UMIN, ISD::UMAX}};
8765   const bool ChooseSelCode = IsAnd ^ IsGreaterOp;
8766   return SelectionCodes[IsUnsigned][ChooseSelCode];
8767 }
8768 
8769 // Combines two comparison operation and logic operation to one selection
8770 // operation(min, max) and logic operation. Returns new constructed Node if
8771 // conditions for optimization are satisfied.
combineCmpOp(SDNode * N,SelectionDAG & DAG,const RISCVSubtarget & Subtarget)8772 static SDValue combineCmpOp(SDNode *N, SelectionDAG &DAG,
8773                             const RISCVSubtarget &Subtarget) {
8774   if (!Subtarget.hasStdExtZbb())
8775     return SDValue();
8776 
8777   const unsigned BitOpcode = N->getOpcode();
8778   assert((BitOpcode == ISD::AND || BitOpcode == ISD::OR) &&
8779          "This optimization can be used only with AND/OR operations");
8780 
8781   const auto Props = verifyCompareConds(N, DAG);
8782   // If conditions are invalidated then do not perform an optimization.
8783   if (!Props)
8784     return SDValue();
8785 
8786   const auto [RefOpcode, A, B, C] = Props.value();
8787   const EVT CmpOpVT = A.getValueType();
8788 
8789   const bool IsGreaterOp = RefOpcode & 0b10;
8790   const bool IsUnsigned = ISD::isUnsignedIntSetCC(RefOpcode);
8791   assert((IsUnsigned || ISD::isSignedIntSetCC(RefOpcode)) &&
8792          "Operation neither with signed or unsigned integers.");
8793 
8794   const bool IsAnd = BitOpcode == ISD::AND;
8795   const ISD::NodeType PickCode =
8796       getSelectionCode(IsUnsigned, IsAnd, IsGreaterOp);
8797 
8798   SDLoc DL(N);
8799   SDValue Pick = DAG.getNode(PickCode, DL, CmpOpVT, A, B);
8800   SDValue Cmp =
8801       DAG.getSetCC(DL, N->getOperand(0).getValueType(), Pick, C, RefOpcode);
8802 
8803   return Cmp;
8804 }
8805 
performANDCombine(SDNode * N,TargetLowering::DAGCombinerInfo & DCI,const RISCVSubtarget & Subtarget)8806 static SDValue performANDCombine(SDNode *N,
8807                                  TargetLowering::DAGCombinerInfo &DCI,
8808                                  const RISCVSubtarget &Subtarget) {
8809   SelectionDAG &DAG = DCI.DAG;
8810 
8811   SDValue N0 = N->getOperand(0);
8812   // Pre-promote (i32 (and (srl X, Y), 1)) on RV64 with Zbs without zero
8813   // extending X. This is safe since we only need the LSB after the shift and
8814   // shift amounts larger than 31 would produce poison. If we wait until
8815   // type legalization, we'll create RISCVISD::SRLW and we can't recover it
8816   // to use a BEXT instruction.
8817   if (Subtarget.is64Bit() && Subtarget.hasStdExtZbs() &&
8818       N->getValueType(0) == MVT::i32 && isOneConstant(N->getOperand(1)) &&
8819       N0.getOpcode() == ISD::SRL && !isa<ConstantSDNode>(N0.getOperand(1)) &&
8820       N0.hasOneUse()) {
8821     SDLoc DL(N);
8822     SDValue Op0 = DAG.getNode(ISD::ANY_EXTEND, DL, MVT::i64, N0.getOperand(0));
8823     SDValue Op1 = DAG.getNode(ISD::ZERO_EXTEND, DL, MVT::i64, N0.getOperand(1));
8824     SDValue Srl = DAG.getNode(ISD::SRL, DL, MVT::i64, Op0, Op1);
8825     SDValue And = DAG.getNode(ISD::AND, DL, MVT::i64, Srl,
8826                               DAG.getConstant(1, DL, MVT::i64));
8827     return DAG.getNode(ISD::TRUNCATE, DL, MVT::i32, And);
8828   }
8829 
8830   if (SDValue V = combineCmpOp(N, DAG, Subtarget))
8831     return V;
8832 
8833   if (SDValue V = combineBinOpToReduce(N, DAG, Subtarget))
8834     return V;
8835 
8836   if (DCI.isAfterLegalizeDAG())
8837     if (SDValue V = combineDeMorganOfBoolean(N, DAG))
8838       return V;
8839 
8840   // fold (and (select lhs, rhs, cc, -1, y), x) ->
8841   //      (select lhs, rhs, cc, x, (and x, y))
8842   return combineSelectAndUseCommutative(N, DAG, /*AllOnes*/ true, Subtarget);
8843 }
8844 
performORCombine(SDNode * N,TargetLowering::DAGCombinerInfo & DCI,const RISCVSubtarget & Subtarget)8845 static SDValue performORCombine(SDNode *N, TargetLowering::DAGCombinerInfo &DCI,
8846                                 const RISCVSubtarget &Subtarget) {
8847   SelectionDAG &DAG = DCI.DAG;
8848 
8849   if (SDValue V = combineCmpOp(N, DAG, Subtarget))
8850     return V;
8851 
8852   if (SDValue V = combineBinOpToReduce(N, DAG, Subtarget))
8853     return V;
8854 
8855   if (DCI.isAfterLegalizeDAG())
8856     if (SDValue V = combineDeMorganOfBoolean(N, DAG))
8857       return V;
8858 
8859   // fold (or (select cond, 0, y), x) ->
8860   //      (select cond, x, (or x, y))
8861   return combineSelectAndUseCommutative(N, DAG, /*AllOnes*/ false, Subtarget);
8862 }
8863 
performXORCombine(SDNode * N,SelectionDAG & DAG,const RISCVSubtarget & Subtarget)8864 static SDValue performXORCombine(SDNode *N, SelectionDAG &DAG,
8865                                  const RISCVSubtarget &Subtarget) {
8866   SDValue N0 = N->getOperand(0);
8867   SDValue N1 = N->getOperand(1);
8868 
8869   // fold (xor (sllw 1, x), -1) -> (rolw ~1, x)
8870   // NOTE: Assumes ROL being legal means ROLW is legal.
8871   const TargetLowering &TLI = DAG.getTargetLoweringInfo();
8872   if (N0.getOpcode() == RISCVISD::SLLW &&
8873       isAllOnesConstant(N1) && isOneConstant(N0.getOperand(0)) &&
8874       TLI.isOperationLegal(ISD::ROTL, MVT::i64)) {
8875     SDLoc DL(N);
8876     return DAG.getNode(RISCVISD::ROLW, DL, MVT::i64,
8877                        DAG.getConstant(~1, DL, MVT::i64), N0.getOperand(1));
8878   }
8879 
8880   if (SDValue V = combineBinOpToReduce(N, DAG, Subtarget))
8881     return V;
8882   // fold (xor (select cond, 0, y), x) ->
8883   //      (select cond, x, (xor x, y))
8884   return combineSelectAndUseCommutative(N, DAG, /*AllOnes*/ false, Subtarget);
8885 }
8886 
8887 // Replace (seteq (i64 (and X, 0xffffffff)), C1) with
8888 // (seteq (i64 (sext_inreg (X, i32)), C1')) where C1' is C1 sign extended from
8889 // bit 31. Same for setne. C1' may be cheaper to materialize and the sext_inreg
8890 // can become a sext.w instead of a shift pair.
performSETCCCombine(SDNode * N,SelectionDAG & DAG,const RISCVSubtarget & Subtarget)8891 static SDValue performSETCCCombine(SDNode *N, SelectionDAG &DAG,
8892                                    const RISCVSubtarget &Subtarget) {
8893   SDValue N0 = N->getOperand(0);
8894   SDValue N1 = N->getOperand(1);
8895   EVT VT = N->getValueType(0);
8896   EVT OpVT = N0.getValueType();
8897 
8898   if (OpVT != MVT::i64 || !Subtarget.is64Bit())
8899     return SDValue();
8900 
8901   // RHS needs to be a constant.
8902   auto *N1C = dyn_cast<ConstantSDNode>(N1);
8903   if (!N1C)
8904     return SDValue();
8905 
8906   // LHS needs to be (and X, 0xffffffff).
8907   if (N0.getOpcode() != ISD::AND || !N0.hasOneUse() ||
8908       !isa<ConstantSDNode>(N0.getOperand(1)) ||
8909       N0.getConstantOperandVal(1) != UINT64_C(0xffffffff))
8910     return SDValue();
8911 
8912   // Looking for an equality compare.
8913   ISD::CondCode Cond = cast<CondCodeSDNode>(N->getOperand(2))->get();
8914   if (!isIntEqualitySetCC(Cond))
8915     return SDValue();
8916 
8917   // Don't do this if the sign bit is provably zero, it will be turned back into
8918   // an AND.
8919   APInt SignMask = APInt::getOneBitSet(64, 31);
8920   if (DAG.MaskedValueIsZero(N0.getOperand(0), SignMask))
8921     return SDValue();
8922 
8923   const APInt &C1 = N1C->getAPIntValue();
8924 
8925   SDLoc dl(N);
8926   // If the constant is larger than 2^32 - 1 it is impossible for both sides
8927   // to be equal.
8928   if (C1.getActiveBits() > 32)
8929     return DAG.getBoolConstant(Cond == ISD::SETNE, dl, VT, OpVT);
8930 
8931   SDValue SExtOp = DAG.getNode(ISD::SIGN_EXTEND_INREG, N, OpVT,
8932                                N0.getOperand(0), DAG.getValueType(MVT::i32));
8933   return DAG.getSetCC(dl, VT, SExtOp, DAG.getConstant(C1.trunc(32).sext(64),
8934                                                       dl, OpVT), Cond);
8935 }
8936 
8937 static SDValue
performSIGN_EXTEND_INREGCombine(SDNode * N,SelectionDAG & DAG,const RISCVSubtarget & Subtarget)8938 performSIGN_EXTEND_INREGCombine(SDNode *N, SelectionDAG &DAG,
8939                                 const RISCVSubtarget &Subtarget) {
8940   SDValue Src = N->getOperand(0);
8941   EVT VT = N->getValueType(0);
8942 
8943   // Fold (sext_inreg (fmv_x_anyexth X), i16) -> (fmv_x_signexth X)
8944   if (Src.getOpcode() == RISCVISD::FMV_X_ANYEXTH &&
8945       cast<VTSDNode>(N->getOperand(1))->getVT().bitsGE(MVT::i16))
8946     return DAG.getNode(RISCVISD::FMV_X_SIGNEXTH, SDLoc(N), VT,
8947                        Src.getOperand(0));
8948 
8949   return SDValue();
8950 }
8951 
8952 namespace {
8953 // Forward declaration of the structure holding the necessary information to
8954 // apply a combine.
8955 struct CombineResult;
8956 
8957 /// Helper class for folding sign/zero extensions.
8958 /// In particular, this class is used for the following combines:
8959 /// add_vl -> vwadd(u) | vwadd(u)_w
8960 /// sub_vl -> vwsub(u) | vwsub(u)_w
8961 /// mul_vl -> vwmul(u) | vwmul_su
8962 ///
8963 /// An object of this class represents an operand of the operation we want to
8964 /// combine.
8965 /// E.g., when trying to combine `mul_vl a, b`, we will have one instance of
8966 /// NodeExtensionHelper for `a` and one for `b`.
8967 ///
8968 /// This class abstracts away how the extension is materialized and
8969 /// how its Mask, VL, number of users affect the combines.
8970 ///
8971 /// In particular:
8972 /// - VWADD_W is conceptually == add(op0, sext(op1))
8973 /// - VWADDU_W == add(op0, zext(op1))
8974 /// - VWSUB_W == sub(op0, sext(op1))
8975 /// - VWSUBU_W == sub(op0, zext(op1))
8976 ///
8977 /// And VMV_V_X_VL, depending on the value, is conceptually equivalent to
8978 /// zext|sext(smaller_value).
8979 struct NodeExtensionHelper {
8980   /// Records if this operand is like being zero extended.
8981   bool SupportsZExt;
8982   /// Records if this operand is like being sign extended.
8983   /// Note: SupportsZExt and SupportsSExt are not mutually exclusive. For
8984   /// instance, a splat constant (e.g., 3), would support being both sign and
8985   /// zero extended.
8986   bool SupportsSExt;
8987   /// This boolean captures whether we care if this operand would still be
8988   /// around after the folding happens.
8989   bool EnforceOneUse;
8990   /// Records if this operand's mask needs to match the mask of the operation
8991   /// that it will fold into.
8992   bool CheckMask;
8993   /// Value of the Mask for this operand.
8994   /// It may be SDValue().
8995   SDValue Mask;
8996   /// Value of the vector length operand.
8997   /// It may be SDValue().
8998   SDValue VL;
8999   /// Original value that this NodeExtensionHelper represents.
9000   SDValue OrigOperand;
9001 
9002   /// Get the value feeding the extension or the value itself.
9003   /// E.g., for zext(a), this would return a.
getSource__anona89954840f11::NodeExtensionHelper9004   SDValue getSource() const {
9005     switch (OrigOperand.getOpcode()) {
9006     case RISCVISD::VSEXT_VL:
9007     case RISCVISD::VZEXT_VL:
9008       return OrigOperand.getOperand(0);
9009     default:
9010       return OrigOperand;
9011     }
9012   }
9013 
9014   /// Check if this instance represents a splat.
isSplat__anona89954840f11::NodeExtensionHelper9015   bool isSplat() const {
9016     return OrigOperand.getOpcode() == RISCVISD::VMV_V_X_VL;
9017   }
9018 
9019   /// Get or create a value that can feed \p Root with the given extension \p
9020   /// SExt. If \p SExt is None, this returns the source of this operand.
9021   /// \see ::getSource().
getOrCreateExtendedOp__anona89954840f11::NodeExtensionHelper9022   SDValue getOrCreateExtendedOp(const SDNode *Root, SelectionDAG &DAG,
9023                                 std::optional<bool> SExt) const {
9024     if (!SExt.has_value())
9025       return OrigOperand;
9026 
9027     MVT NarrowVT = getNarrowType(Root);
9028 
9029     SDValue Source = getSource();
9030     if (Source.getValueType() == NarrowVT)
9031       return Source;
9032 
9033     unsigned ExtOpc = *SExt ? RISCVISD::VSEXT_VL : RISCVISD::VZEXT_VL;
9034 
9035     // If we need an extension, we should be changing the type.
9036     SDLoc DL(Root);
9037     auto [Mask, VL] = getMaskAndVL(Root);
9038     switch (OrigOperand.getOpcode()) {
9039     case RISCVISD::VSEXT_VL:
9040     case RISCVISD::VZEXT_VL:
9041       return DAG.getNode(ExtOpc, DL, NarrowVT, Source, Mask, VL);
9042     case RISCVISD::VMV_V_X_VL:
9043       return DAG.getNode(RISCVISD::VMV_V_X_VL, DL, NarrowVT,
9044                          DAG.getUNDEF(NarrowVT), Source.getOperand(1), VL);
9045     default:
9046       // Other opcodes can only come from the original LHS of VW(ADD|SUB)_W_VL
9047       // and that operand should already have the right NarrowVT so no
9048       // extension should be required at this point.
9049       llvm_unreachable("Unsupported opcode");
9050     }
9051   }
9052 
9053   /// Helper function to get the narrow type for \p Root.
9054   /// The narrow type is the type of \p Root where we divided the size of each
9055   /// element by 2. E.g., if Root's type <2xi16> -> narrow type <2xi8>.
9056   /// \pre The size of the type of the elements of Root must be a multiple of 2
9057   /// and be greater than 16.
getNarrowType__anona89954840f11::NodeExtensionHelper9058   static MVT getNarrowType(const SDNode *Root) {
9059     MVT VT = Root->getSimpleValueType(0);
9060 
9061     // Determine the narrow size.
9062     unsigned NarrowSize = VT.getScalarSizeInBits() / 2;
9063     assert(NarrowSize >= 8 && "Trying to extend something we can't represent");
9064     MVT NarrowVT = MVT::getVectorVT(MVT::getIntegerVT(NarrowSize),
9065                                     VT.getVectorElementCount());
9066     return NarrowVT;
9067   }
9068 
9069   /// Return the opcode required to materialize the folding of the sign
9070   /// extensions (\p IsSExt == true) or zero extensions (IsSExt == false) for
9071   /// both operands for \p Opcode.
9072   /// Put differently, get the opcode to materialize:
9073   /// - ISExt == true: \p Opcode(sext(a), sext(b)) -> newOpcode(a, b)
9074   /// - ISExt == false: \p Opcode(zext(a), zext(b)) -> newOpcode(a, b)
9075   /// \pre \p Opcode represents a supported root (\see ::isSupportedRoot()).
getSameExtensionOpcode__anona89954840f11::NodeExtensionHelper9076   static unsigned getSameExtensionOpcode(unsigned Opcode, bool IsSExt) {
9077     switch (Opcode) {
9078     case RISCVISD::ADD_VL:
9079     case RISCVISD::VWADD_W_VL:
9080     case RISCVISD::VWADDU_W_VL:
9081       return IsSExt ? RISCVISD::VWADD_VL : RISCVISD::VWADDU_VL;
9082     case RISCVISD::MUL_VL:
9083       return IsSExt ? RISCVISD::VWMUL_VL : RISCVISD::VWMULU_VL;
9084     case RISCVISD::SUB_VL:
9085     case RISCVISD::VWSUB_W_VL:
9086     case RISCVISD::VWSUBU_W_VL:
9087       return IsSExt ? RISCVISD::VWSUB_VL : RISCVISD::VWSUBU_VL;
9088     default:
9089       llvm_unreachable("Unexpected opcode");
9090     }
9091   }
9092 
9093   /// Get the opcode to materialize \p Opcode(sext(a), zext(b)) ->
9094   /// newOpcode(a, b).
getSUOpcode__anona89954840f11::NodeExtensionHelper9095   static unsigned getSUOpcode(unsigned Opcode) {
9096     assert(Opcode == RISCVISD::MUL_VL && "SU is only supported for MUL");
9097     return RISCVISD::VWMULSU_VL;
9098   }
9099 
9100   /// Get the opcode to materialize \p Opcode(a, s|zext(b)) ->
9101   /// newOpcode(a, b).
getWOpcode__anona89954840f11::NodeExtensionHelper9102   static unsigned getWOpcode(unsigned Opcode, bool IsSExt) {
9103     switch (Opcode) {
9104     case RISCVISD::ADD_VL:
9105       return IsSExt ? RISCVISD::VWADD_W_VL : RISCVISD::VWADDU_W_VL;
9106     case RISCVISD::SUB_VL:
9107       return IsSExt ? RISCVISD::VWSUB_W_VL : RISCVISD::VWSUBU_W_VL;
9108     default:
9109       llvm_unreachable("Unexpected opcode");
9110     }
9111   }
9112 
9113   using CombineToTry = std::function<std::optional<CombineResult>(
9114       SDNode * /*Root*/, const NodeExtensionHelper & /*LHS*/,
9115       const NodeExtensionHelper & /*RHS*/)>;
9116 
9117   /// Check if this node needs to be fully folded or extended for all users.
needToPromoteOtherUsers__anona89954840f11::NodeExtensionHelper9118   bool needToPromoteOtherUsers() const { return EnforceOneUse; }
9119 
9120   /// Helper method to set the various fields of this struct based on the
9121   /// type of \p Root.
fillUpExtensionSupport__anona89954840f11::NodeExtensionHelper9122   void fillUpExtensionSupport(SDNode *Root, SelectionDAG &DAG) {
9123     SupportsZExt = false;
9124     SupportsSExt = false;
9125     EnforceOneUse = true;
9126     CheckMask = true;
9127     switch (OrigOperand.getOpcode()) {
9128     case RISCVISD::VZEXT_VL:
9129       SupportsZExt = true;
9130       Mask = OrigOperand.getOperand(1);
9131       VL = OrigOperand.getOperand(2);
9132       break;
9133     case RISCVISD::VSEXT_VL:
9134       SupportsSExt = true;
9135       Mask = OrigOperand.getOperand(1);
9136       VL = OrigOperand.getOperand(2);
9137       break;
9138     case RISCVISD::VMV_V_X_VL: {
9139       // Historically, we didn't care about splat values not disappearing during
9140       // combines.
9141       EnforceOneUse = false;
9142       CheckMask = false;
9143       VL = OrigOperand.getOperand(2);
9144 
9145       // The operand is a splat of a scalar.
9146 
9147       // The pasthru must be undef for tail agnostic.
9148       if (!OrigOperand.getOperand(0).isUndef())
9149         break;
9150 
9151       // Get the scalar value.
9152       SDValue Op = OrigOperand.getOperand(1);
9153 
9154       // See if we have enough sign bits or zero bits in the scalar to use a
9155       // widening opcode by splatting to smaller element size.
9156       MVT VT = Root->getSimpleValueType(0);
9157       unsigned EltBits = VT.getScalarSizeInBits();
9158       unsigned ScalarBits = Op.getValueSizeInBits();
9159       // Make sure we're getting all element bits from the scalar register.
9160       // FIXME: Support implicit sign extension of vmv.v.x?
9161       if (ScalarBits < EltBits)
9162         break;
9163 
9164       unsigned NarrowSize = VT.getScalarSizeInBits() / 2;
9165       // If the narrow type cannot be expressed with a legal VMV,
9166       // this is not a valid candidate.
9167       if (NarrowSize < 8)
9168         break;
9169 
9170       if (DAG.ComputeMaxSignificantBits(Op) <= NarrowSize)
9171         SupportsSExt = true;
9172       if (DAG.MaskedValueIsZero(Op,
9173                                 APInt::getBitsSetFrom(ScalarBits, NarrowSize)))
9174         SupportsZExt = true;
9175       break;
9176     }
9177     default:
9178       break;
9179     }
9180   }
9181 
9182   /// Check if \p Root supports any extension folding combines.
isSupportedRoot__anona89954840f11::NodeExtensionHelper9183   static bool isSupportedRoot(const SDNode *Root) {
9184     switch (Root->getOpcode()) {
9185     case RISCVISD::ADD_VL:
9186     case RISCVISD::MUL_VL:
9187     case RISCVISD::VWADD_W_VL:
9188     case RISCVISD::VWADDU_W_VL:
9189     case RISCVISD::SUB_VL:
9190     case RISCVISD::VWSUB_W_VL:
9191     case RISCVISD::VWSUBU_W_VL:
9192       return true;
9193     default:
9194       return false;
9195     }
9196   }
9197 
9198   /// Build a NodeExtensionHelper for \p Root.getOperand(\p OperandIdx).
NodeExtensionHelper__anona89954840f11::NodeExtensionHelper9199   NodeExtensionHelper(SDNode *Root, unsigned OperandIdx, SelectionDAG &DAG) {
9200     assert(isSupportedRoot(Root) && "Trying to build an helper with an "
9201                                     "unsupported root");
9202     assert(OperandIdx < 2 && "Requesting something else than LHS or RHS");
9203     OrigOperand = Root->getOperand(OperandIdx);
9204 
9205     unsigned Opc = Root->getOpcode();
9206     switch (Opc) {
9207     // We consider VW<ADD|SUB>(U)_W(LHS, RHS) as if they were
9208     // <ADD|SUB>(LHS, S|ZEXT(RHS))
9209     case RISCVISD::VWADD_W_VL:
9210     case RISCVISD::VWADDU_W_VL:
9211     case RISCVISD::VWSUB_W_VL:
9212     case RISCVISD::VWSUBU_W_VL:
9213       if (OperandIdx == 1) {
9214         SupportsZExt =
9215             Opc == RISCVISD::VWADDU_W_VL || Opc == RISCVISD::VWSUBU_W_VL;
9216         SupportsSExt = !SupportsZExt;
9217         std::tie(Mask, VL) = getMaskAndVL(Root);
9218         CheckMask = true;
9219         // There's no existing extension here, so we don't have to worry about
9220         // making sure it gets removed.
9221         EnforceOneUse = false;
9222         break;
9223       }
9224       [[fallthrough]];
9225     default:
9226       fillUpExtensionSupport(Root, DAG);
9227       break;
9228     }
9229   }
9230 
9231   /// Check if this operand is compatible with the given vector length \p VL.
isVLCompatible__anona89954840f11::NodeExtensionHelper9232   bool isVLCompatible(SDValue VL) const {
9233     return this->VL != SDValue() && this->VL == VL;
9234   }
9235 
9236   /// Check if this operand is compatible with the given \p Mask.
isMaskCompatible__anona89954840f11::NodeExtensionHelper9237   bool isMaskCompatible(SDValue Mask) const {
9238     return !CheckMask || (this->Mask != SDValue() && this->Mask == Mask);
9239   }
9240 
9241   /// Helper function to get the Mask and VL from \p Root.
getMaskAndVL__anona89954840f11::NodeExtensionHelper9242   static std::pair<SDValue, SDValue> getMaskAndVL(const SDNode *Root) {
9243     assert(isSupportedRoot(Root) && "Unexpected root");
9244     return std::make_pair(Root->getOperand(3), Root->getOperand(4));
9245   }
9246 
9247   /// Check if the Mask and VL of this operand are compatible with \p Root.
areVLAndMaskCompatible__anona89954840f11::NodeExtensionHelper9248   bool areVLAndMaskCompatible(const SDNode *Root) const {
9249     auto [Mask, VL] = getMaskAndVL(Root);
9250     return isMaskCompatible(Mask) && isVLCompatible(VL);
9251   }
9252 
9253   /// Helper function to check if \p N is commutative with respect to the
9254   /// foldings that are supported by this class.
isCommutative__anona89954840f11::NodeExtensionHelper9255   static bool isCommutative(const SDNode *N) {
9256     switch (N->getOpcode()) {
9257     case RISCVISD::ADD_VL:
9258     case RISCVISD::MUL_VL:
9259     case RISCVISD::VWADD_W_VL:
9260     case RISCVISD::VWADDU_W_VL:
9261       return true;
9262     case RISCVISD::SUB_VL:
9263     case RISCVISD::VWSUB_W_VL:
9264     case RISCVISD::VWSUBU_W_VL:
9265       return false;
9266     default:
9267       llvm_unreachable("Unexpected opcode");
9268     }
9269   }
9270 
9271   /// Get a list of combine to try for folding extensions in \p Root.
9272   /// Note that each returned CombineToTry function doesn't actually modify
9273   /// anything. Instead they produce an optional CombineResult that if not None,
9274   /// need to be materialized for the combine to be applied.
9275   /// \see CombineResult::materialize.
9276   /// If the related CombineToTry function returns std::nullopt, that means the
9277   /// combine didn't match.
9278   static SmallVector<CombineToTry> getSupportedFoldings(const SDNode *Root);
9279 };
9280 
9281 /// Helper structure that holds all the necessary information to materialize a
9282 /// combine that does some extension folding.
9283 struct CombineResult {
9284   /// Opcode to be generated when materializing the combine.
9285   unsigned TargetOpcode;
9286   // No value means no extension is needed. If extension is needed, the value
9287   // indicates if it needs to be sign extended.
9288   std::optional<bool> SExtLHS;
9289   std::optional<bool> SExtRHS;
9290   /// Root of the combine.
9291   SDNode *Root;
9292   /// LHS of the TargetOpcode.
9293   NodeExtensionHelper LHS;
9294   /// RHS of the TargetOpcode.
9295   NodeExtensionHelper RHS;
9296 
CombineResult__anona89954840f11::CombineResult9297   CombineResult(unsigned TargetOpcode, SDNode *Root,
9298                 const NodeExtensionHelper &LHS, std::optional<bool> SExtLHS,
9299                 const NodeExtensionHelper &RHS, std::optional<bool> SExtRHS)
9300       : TargetOpcode(TargetOpcode), SExtLHS(SExtLHS), SExtRHS(SExtRHS),
9301         Root(Root), LHS(LHS), RHS(RHS) {}
9302 
9303   /// Return a value that uses TargetOpcode and that can be used to replace
9304   /// Root.
9305   /// The actual replacement is *not* done in that method.
materialize__anona89954840f11::CombineResult9306   SDValue materialize(SelectionDAG &DAG) const {
9307     SDValue Mask, VL, Merge;
9308     std::tie(Mask, VL) = NodeExtensionHelper::getMaskAndVL(Root);
9309     Merge = Root->getOperand(2);
9310     return DAG.getNode(TargetOpcode, SDLoc(Root), Root->getValueType(0),
9311                        LHS.getOrCreateExtendedOp(Root, DAG, SExtLHS),
9312                        RHS.getOrCreateExtendedOp(Root, DAG, SExtRHS), Merge,
9313                        Mask, VL);
9314   }
9315 };
9316 
9317 /// Check if \p Root follows a pattern Root(ext(LHS), ext(RHS))
9318 /// where `ext` is the same for both LHS and RHS (i.e., both are sext or both
9319 /// are zext) and LHS and RHS can be folded into Root.
9320 /// AllowSExt and AllozZExt define which form `ext` can take in this pattern.
9321 ///
9322 /// \note If the pattern can match with both zext and sext, the returned
9323 /// CombineResult will feature the zext result.
9324 ///
9325 /// \returns std::nullopt if the pattern doesn't match or a CombineResult that
9326 /// can be used to apply the pattern.
9327 static std::optional<CombineResult>
canFoldToVWWithSameExtensionImpl(SDNode * Root,const NodeExtensionHelper & LHS,const NodeExtensionHelper & RHS,bool AllowSExt,bool AllowZExt)9328 canFoldToVWWithSameExtensionImpl(SDNode *Root, const NodeExtensionHelper &LHS,
9329                                  const NodeExtensionHelper &RHS, bool AllowSExt,
9330                                  bool AllowZExt) {
9331   assert((AllowSExt || AllowZExt) && "Forgot to set what you want?");
9332   if (!LHS.areVLAndMaskCompatible(Root) || !RHS.areVLAndMaskCompatible(Root))
9333     return std::nullopt;
9334   if (AllowZExt && LHS.SupportsZExt && RHS.SupportsZExt)
9335     return CombineResult(NodeExtensionHelper::getSameExtensionOpcode(
9336                              Root->getOpcode(), /*IsSExt=*/false),
9337                          Root, LHS, /*SExtLHS=*/false, RHS,
9338                          /*SExtRHS=*/false);
9339   if (AllowSExt && LHS.SupportsSExt && RHS.SupportsSExt)
9340     return CombineResult(NodeExtensionHelper::getSameExtensionOpcode(
9341                              Root->getOpcode(), /*IsSExt=*/true),
9342                          Root, LHS, /*SExtLHS=*/true, RHS,
9343                          /*SExtRHS=*/true);
9344   return std::nullopt;
9345 }
9346 
9347 /// Check if \p Root follows a pattern Root(ext(LHS), ext(RHS))
9348 /// where `ext` is the same for both LHS and RHS (i.e., both are sext or both
9349 /// are zext) and LHS and RHS can be folded into Root.
9350 ///
9351 /// \returns std::nullopt if the pattern doesn't match or a CombineResult that
9352 /// can be used to apply the pattern.
9353 static std::optional<CombineResult>
canFoldToVWWithSameExtension(SDNode * Root,const NodeExtensionHelper & LHS,const NodeExtensionHelper & RHS)9354 canFoldToVWWithSameExtension(SDNode *Root, const NodeExtensionHelper &LHS,
9355                              const NodeExtensionHelper &RHS) {
9356   return canFoldToVWWithSameExtensionImpl(Root, LHS, RHS, /*AllowSExt=*/true,
9357                                           /*AllowZExt=*/true);
9358 }
9359 
9360 /// Check if \p Root follows a pattern Root(LHS, ext(RHS))
9361 ///
9362 /// \returns std::nullopt if the pattern doesn't match or a CombineResult that
9363 /// can be used to apply the pattern.
9364 static std::optional<CombineResult>
canFoldToVW_W(SDNode * Root,const NodeExtensionHelper & LHS,const NodeExtensionHelper & RHS)9365 canFoldToVW_W(SDNode *Root, const NodeExtensionHelper &LHS,
9366               const NodeExtensionHelper &RHS) {
9367   if (!RHS.areVLAndMaskCompatible(Root))
9368     return std::nullopt;
9369 
9370   // FIXME: Is it useful to form a vwadd.wx or vwsub.wx if it removes a scalar
9371   // sext/zext?
9372   // Control this behavior behind an option (AllowSplatInVW_W) for testing
9373   // purposes.
9374   if (RHS.SupportsZExt && (!RHS.isSplat() || AllowSplatInVW_W))
9375     return CombineResult(
9376         NodeExtensionHelper::getWOpcode(Root->getOpcode(), /*IsSExt=*/false),
9377         Root, LHS, /*SExtLHS=*/std::nullopt, RHS, /*SExtRHS=*/false);
9378   if (RHS.SupportsSExt && (!RHS.isSplat() || AllowSplatInVW_W))
9379     return CombineResult(
9380         NodeExtensionHelper::getWOpcode(Root->getOpcode(), /*IsSExt=*/true),
9381         Root, LHS, /*SExtLHS=*/std::nullopt, RHS, /*SExtRHS=*/true);
9382   return std::nullopt;
9383 }
9384 
9385 /// Check if \p Root follows a pattern Root(sext(LHS), sext(RHS))
9386 ///
9387 /// \returns std::nullopt if the pattern doesn't match or a CombineResult that
9388 /// can be used to apply the pattern.
9389 static std::optional<CombineResult>
canFoldToVWWithSEXT(SDNode * Root,const NodeExtensionHelper & LHS,const NodeExtensionHelper & RHS)9390 canFoldToVWWithSEXT(SDNode *Root, const NodeExtensionHelper &LHS,
9391                     const NodeExtensionHelper &RHS) {
9392   return canFoldToVWWithSameExtensionImpl(Root, LHS, RHS, /*AllowSExt=*/true,
9393                                           /*AllowZExt=*/false);
9394 }
9395 
9396 /// Check if \p Root follows a pattern Root(zext(LHS), zext(RHS))
9397 ///
9398 /// \returns std::nullopt if the pattern doesn't match or a CombineResult that
9399 /// can be used to apply the pattern.
9400 static std::optional<CombineResult>
canFoldToVWWithZEXT(SDNode * Root,const NodeExtensionHelper & LHS,const NodeExtensionHelper & RHS)9401 canFoldToVWWithZEXT(SDNode *Root, const NodeExtensionHelper &LHS,
9402                     const NodeExtensionHelper &RHS) {
9403   return canFoldToVWWithSameExtensionImpl(Root, LHS, RHS, /*AllowSExt=*/false,
9404                                           /*AllowZExt=*/true);
9405 }
9406 
9407 /// Check if \p Root follows a pattern Root(sext(LHS), zext(RHS))
9408 ///
9409 /// \returns std::nullopt if the pattern doesn't match or a CombineResult that
9410 /// can be used to apply the pattern.
9411 static std::optional<CombineResult>
canFoldToVW_SU(SDNode * Root,const NodeExtensionHelper & LHS,const NodeExtensionHelper & RHS)9412 canFoldToVW_SU(SDNode *Root, const NodeExtensionHelper &LHS,
9413                const NodeExtensionHelper &RHS) {
9414   if (!LHS.SupportsSExt || !RHS.SupportsZExt)
9415     return std::nullopt;
9416   if (!LHS.areVLAndMaskCompatible(Root) || !RHS.areVLAndMaskCompatible(Root))
9417     return std::nullopt;
9418   return CombineResult(NodeExtensionHelper::getSUOpcode(Root->getOpcode()),
9419                        Root, LHS, /*SExtLHS=*/true, RHS, /*SExtRHS=*/false);
9420 }
9421 
9422 SmallVector<NodeExtensionHelper::CombineToTry>
getSupportedFoldings(const SDNode * Root)9423 NodeExtensionHelper::getSupportedFoldings(const SDNode *Root) {
9424   SmallVector<CombineToTry> Strategies;
9425   switch (Root->getOpcode()) {
9426   case RISCVISD::ADD_VL:
9427   case RISCVISD::SUB_VL:
9428     // add|sub -> vwadd(u)|vwsub(u)
9429     Strategies.push_back(canFoldToVWWithSameExtension);
9430     // add|sub -> vwadd(u)_w|vwsub(u)_w
9431     Strategies.push_back(canFoldToVW_W);
9432     break;
9433   case RISCVISD::MUL_VL:
9434     // mul -> vwmul(u)
9435     Strategies.push_back(canFoldToVWWithSameExtension);
9436     // mul -> vwmulsu
9437     Strategies.push_back(canFoldToVW_SU);
9438     break;
9439   case RISCVISD::VWADD_W_VL:
9440   case RISCVISD::VWSUB_W_VL:
9441     // vwadd_w|vwsub_w -> vwadd|vwsub
9442     Strategies.push_back(canFoldToVWWithSEXT);
9443     break;
9444   case RISCVISD::VWADDU_W_VL:
9445   case RISCVISD::VWSUBU_W_VL:
9446     // vwaddu_w|vwsubu_w -> vwaddu|vwsubu
9447     Strategies.push_back(canFoldToVWWithZEXT);
9448     break;
9449   default:
9450     llvm_unreachable("Unexpected opcode");
9451   }
9452   return Strategies;
9453 }
9454 } // End anonymous namespace.
9455 
9456 /// Combine a binary operation to its equivalent VW or VW_W form.
9457 /// The supported combines are:
9458 /// add_vl -> vwadd(u) | vwadd(u)_w
9459 /// sub_vl -> vwsub(u) | vwsub(u)_w
9460 /// mul_vl -> vwmul(u) | vwmul_su
9461 /// vwadd_w(u) -> vwadd(u)
9462 /// vwub_w(u) -> vwadd(u)
9463 static SDValue
combineBinOp_VLToVWBinOp_VL(SDNode * N,TargetLowering::DAGCombinerInfo & DCI)9464 combineBinOp_VLToVWBinOp_VL(SDNode *N, TargetLowering::DAGCombinerInfo &DCI) {
9465   SelectionDAG &DAG = DCI.DAG;
9466 
9467   assert(NodeExtensionHelper::isSupportedRoot(N) &&
9468          "Shouldn't have called this method");
9469   SmallVector<SDNode *> Worklist;
9470   SmallSet<SDNode *, 8> Inserted;
9471   Worklist.push_back(N);
9472   Inserted.insert(N);
9473   SmallVector<CombineResult> CombinesToApply;
9474 
9475   while (!Worklist.empty()) {
9476     SDNode *Root = Worklist.pop_back_val();
9477     if (!NodeExtensionHelper::isSupportedRoot(Root))
9478       return SDValue();
9479 
9480     NodeExtensionHelper LHS(N, 0, DAG);
9481     NodeExtensionHelper RHS(N, 1, DAG);
9482     auto AppendUsersIfNeeded = [&Worklist,
9483                                 &Inserted](const NodeExtensionHelper &Op) {
9484       if (Op.needToPromoteOtherUsers()) {
9485         for (SDNode *TheUse : Op.OrigOperand->uses()) {
9486           if (Inserted.insert(TheUse).second)
9487             Worklist.push_back(TheUse);
9488         }
9489       }
9490     };
9491 
9492     // Control the compile time by limiting the number of node we look at in
9493     // total.
9494     if (Inserted.size() > ExtensionMaxWebSize)
9495       return SDValue();
9496 
9497     SmallVector<NodeExtensionHelper::CombineToTry> FoldingStrategies =
9498         NodeExtensionHelper::getSupportedFoldings(N);
9499 
9500     assert(!FoldingStrategies.empty() && "Nothing to be folded");
9501     bool Matched = false;
9502     for (int Attempt = 0;
9503          (Attempt != 1 + NodeExtensionHelper::isCommutative(N)) && !Matched;
9504          ++Attempt) {
9505 
9506       for (NodeExtensionHelper::CombineToTry FoldingStrategy :
9507            FoldingStrategies) {
9508         std::optional<CombineResult> Res = FoldingStrategy(N, LHS, RHS);
9509         if (Res) {
9510           Matched = true;
9511           CombinesToApply.push_back(*Res);
9512           // All the inputs that are extended need to be folded, otherwise
9513           // we would be leaving the old input (since it is may still be used),
9514           // and the new one.
9515           if (Res->SExtLHS.has_value())
9516             AppendUsersIfNeeded(LHS);
9517           if (Res->SExtRHS.has_value())
9518             AppendUsersIfNeeded(RHS);
9519           break;
9520         }
9521       }
9522       std::swap(LHS, RHS);
9523     }
9524     // Right now we do an all or nothing approach.
9525     if (!Matched)
9526       return SDValue();
9527   }
9528   // Store the value for the replacement of the input node separately.
9529   SDValue InputRootReplacement;
9530   // We do the RAUW after we materialize all the combines, because some replaced
9531   // nodes may be feeding some of the yet-to-be-replaced nodes. Put differently,
9532   // some of these nodes may appear in the NodeExtensionHelpers of some of the
9533   // yet-to-be-visited CombinesToApply roots.
9534   SmallVector<std::pair<SDValue, SDValue>> ValuesToReplace;
9535   ValuesToReplace.reserve(CombinesToApply.size());
9536   for (CombineResult Res : CombinesToApply) {
9537     SDValue NewValue = Res.materialize(DAG);
9538     if (!InputRootReplacement) {
9539       assert(Res.Root == N &&
9540              "First element is expected to be the current node");
9541       InputRootReplacement = NewValue;
9542     } else {
9543       ValuesToReplace.emplace_back(SDValue(Res.Root, 0), NewValue);
9544     }
9545   }
9546   for (std::pair<SDValue, SDValue> OldNewValues : ValuesToReplace) {
9547     DAG.ReplaceAllUsesOfValueWith(OldNewValues.first, OldNewValues.second);
9548     DCI.AddToWorklist(OldNewValues.second.getNode());
9549   }
9550   return InputRootReplacement;
9551 }
9552 
9553 // Fold
9554 //   (fp_to_int (froundeven X)) -> fcvt X, rne
9555 //   (fp_to_int (ftrunc X))     -> fcvt X, rtz
9556 //   (fp_to_int (ffloor X))     -> fcvt X, rdn
9557 //   (fp_to_int (fceil X))      -> fcvt X, rup
9558 //   (fp_to_int (fround X))     -> fcvt X, rmm
performFP_TO_INTCombine(SDNode * N,TargetLowering::DAGCombinerInfo & DCI,const RISCVSubtarget & Subtarget)9559 static SDValue performFP_TO_INTCombine(SDNode *N,
9560                                        TargetLowering::DAGCombinerInfo &DCI,
9561                                        const RISCVSubtarget &Subtarget) {
9562   SelectionDAG &DAG = DCI.DAG;
9563   const TargetLowering &TLI = DAG.getTargetLoweringInfo();
9564   MVT XLenVT = Subtarget.getXLenVT();
9565 
9566   SDValue Src = N->getOperand(0);
9567 
9568   // Ensure the FP type is legal.
9569   if (!TLI.isTypeLegal(Src.getValueType()))
9570     return SDValue();
9571 
9572   // Don't do this for f16 with Zfhmin and not Zfh.
9573   if (Src.getValueType() == MVT::f16 && !Subtarget.hasStdExtZfh())
9574     return SDValue();
9575 
9576   RISCVFPRndMode::RoundingMode FRM = matchRoundingOp(Src.getOpcode());
9577   if (FRM == RISCVFPRndMode::Invalid)
9578     return SDValue();
9579 
9580   SDLoc DL(N);
9581   bool IsSigned = N->getOpcode() == ISD::FP_TO_SINT;
9582   EVT VT = N->getValueType(0);
9583 
9584   if (VT.isVector() && TLI.isTypeLegal(VT)) {
9585     MVT SrcVT = Src.getSimpleValueType();
9586     MVT SrcContainerVT = SrcVT;
9587     MVT ContainerVT = VT.getSimpleVT();
9588     SDValue XVal = Src.getOperand(0);
9589 
9590     // For widening and narrowing conversions we just combine it into a
9591     // VFCVT_..._VL node, as there are no specific VFWCVT/VFNCVT VL nodes. They
9592     // end up getting lowered to their appropriate pseudo instructions based on
9593     // their operand types
9594     if (VT.getScalarSizeInBits() > SrcVT.getScalarSizeInBits() * 2 ||
9595         VT.getScalarSizeInBits() * 2 < SrcVT.getScalarSizeInBits())
9596       return SDValue();
9597 
9598     // Make fixed-length vectors scalable first
9599     if (SrcVT.isFixedLengthVector()) {
9600       SrcContainerVT = getContainerForFixedLengthVector(DAG, SrcVT, Subtarget);
9601       XVal = convertToScalableVector(SrcContainerVT, XVal, DAG, Subtarget);
9602       ContainerVT =
9603           getContainerForFixedLengthVector(DAG, ContainerVT, Subtarget);
9604     }
9605 
9606     auto [Mask, VL] =
9607         getDefaultVLOps(SrcVT, SrcContainerVT, DL, DAG, Subtarget);
9608 
9609     SDValue FpToInt;
9610     if (FRM == RISCVFPRndMode::RTZ) {
9611       // Use the dedicated trunc static rounding mode if we're truncating so we
9612       // don't need to generate calls to fsrmi/fsrm
9613       unsigned Opc =
9614           IsSigned ? RISCVISD::VFCVT_RTZ_X_F_VL : RISCVISD::VFCVT_RTZ_XU_F_VL;
9615       FpToInt = DAG.getNode(Opc, DL, ContainerVT, XVal, Mask, VL);
9616     } else {
9617       unsigned Opc =
9618           IsSigned ? RISCVISD::VFCVT_RM_X_F_VL : RISCVISD::VFCVT_RM_XU_F_VL;
9619       FpToInt = DAG.getNode(Opc, DL, ContainerVT, XVal, Mask,
9620                             DAG.getTargetConstant(FRM, DL, XLenVT), VL);
9621     }
9622 
9623     // If converted from fixed-length to scalable, convert back
9624     if (VT.isFixedLengthVector())
9625       FpToInt = convertFromScalableVector(VT, FpToInt, DAG, Subtarget);
9626 
9627     return FpToInt;
9628   }
9629 
9630   // Only handle XLen or i32 types. Other types narrower than XLen will
9631   // eventually be legalized to XLenVT.
9632   if (VT != MVT::i32 && VT != XLenVT)
9633     return SDValue();
9634 
9635   unsigned Opc;
9636   if (VT == XLenVT)
9637     Opc = IsSigned ? RISCVISD::FCVT_X : RISCVISD::FCVT_XU;
9638   else
9639     Opc = IsSigned ? RISCVISD::FCVT_W_RV64 : RISCVISD::FCVT_WU_RV64;
9640 
9641   SDValue FpToInt = DAG.getNode(Opc, DL, XLenVT, Src.getOperand(0),
9642                                 DAG.getTargetConstant(FRM, DL, XLenVT));
9643   return DAG.getNode(ISD::TRUNCATE, DL, VT, FpToInt);
9644 }
9645 
9646 // Fold
9647 //   (fp_to_int_sat (froundeven X)) -> (select X == nan, 0, (fcvt X, rne))
9648 //   (fp_to_int_sat (ftrunc X))     -> (select X == nan, 0, (fcvt X, rtz))
9649 //   (fp_to_int_sat (ffloor X))     -> (select X == nan, 0, (fcvt X, rdn))
9650 //   (fp_to_int_sat (fceil X))      -> (select X == nan, 0, (fcvt X, rup))
9651 //   (fp_to_int_sat (fround X))     -> (select X == nan, 0, (fcvt X, rmm))
performFP_TO_INT_SATCombine(SDNode * N,TargetLowering::DAGCombinerInfo & DCI,const RISCVSubtarget & Subtarget)9652 static SDValue performFP_TO_INT_SATCombine(SDNode *N,
9653                                        TargetLowering::DAGCombinerInfo &DCI,
9654                                        const RISCVSubtarget &Subtarget) {
9655   SelectionDAG &DAG = DCI.DAG;
9656   const TargetLowering &TLI = DAG.getTargetLoweringInfo();
9657   MVT XLenVT = Subtarget.getXLenVT();
9658 
9659   // Only handle XLen types. Other types narrower than XLen will eventually be
9660   // legalized to XLenVT.
9661   EVT DstVT = N->getValueType(0);
9662   if (DstVT != XLenVT)
9663     return SDValue();
9664 
9665   SDValue Src = N->getOperand(0);
9666 
9667   // Ensure the FP type is also legal.
9668   if (!TLI.isTypeLegal(Src.getValueType()))
9669     return SDValue();
9670 
9671   // Don't do this for f16 with Zfhmin and not Zfh.
9672   if (Src.getValueType() == MVT::f16 && !Subtarget.hasStdExtZfh())
9673     return SDValue();
9674 
9675   EVT SatVT = cast<VTSDNode>(N->getOperand(1))->getVT();
9676 
9677   RISCVFPRndMode::RoundingMode FRM = matchRoundingOp(Src.getOpcode());
9678   if (FRM == RISCVFPRndMode::Invalid)
9679     return SDValue();
9680 
9681   bool IsSigned = N->getOpcode() == ISD::FP_TO_SINT_SAT;
9682 
9683   unsigned Opc;
9684   if (SatVT == DstVT)
9685     Opc = IsSigned ? RISCVISD::FCVT_X : RISCVISD::FCVT_XU;
9686   else if (DstVT == MVT::i64 && SatVT == MVT::i32)
9687     Opc = IsSigned ? RISCVISD::FCVT_W_RV64 : RISCVISD::FCVT_WU_RV64;
9688   else
9689     return SDValue();
9690   // FIXME: Support other SatVTs by clamping before or after the conversion.
9691 
9692   Src = Src.getOperand(0);
9693 
9694   SDLoc DL(N);
9695   SDValue FpToInt = DAG.getNode(Opc, DL, XLenVT, Src,
9696                                 DAG.getTargetConstant(FRM, DL, XLenVT));
9697 
9698   // fcvt.wu.* sign extends bit 31 on RV64. FP_TO_UINT_SAT expects to zero
9699   // extend.
9700   if (Opc == RISCVISD::FCVT_WU_RV64)
9701     FpToInt = DAG.getZeroExtendInReg(FpToInt, DL, MVT::i32);
9702 
9703   // RISCV FP-to-int conversions saturate to the destination register size, but
9704   // don't produce 0 for nan.
9705   SDValue ZeroInt = DAG.getConstant(0, DL, DstVT);
9706   return DAG.getSelectCC(DL, Src, Src, ZeroInt, FpToInt, ISD::CondCode::SETUO);
9707 }
9708 
9709 // Combine (bitreverse (bswap X)) to the BREV8 GREVI encoding if the type is
9710 // smaller than XLenVT.
performBITREVERSECombine(SDNode * N,SelectionDAG & DAG,const RISCVSubtarget & Subtarget)9711 static SDValue performBITREVERSECombine(SDNode *N, SelectionDAG &DAG,
9712                                         const RISCVSubtarget &Subtarget) {
9713   assert(Subtarget.hasStdExtZbkb() && "Unexpected extension");
9714 
9715   SDValue Src = N->getOperand(0);
9716   if (Src.getOpcode() != ISD::BSWAP)
9717     return SDValue();
9718 
9719   EVT VT = N->getValueType(0);
9720   if (!VT.isScalarInteger() || VT.getSizeInBits() >= Subtarget.getXLen() ||
9721       !isPowerOf2_32(VT.getSizeInBits()))
9722     return SDValue();
9723 
9724   SDLoc DL(N);
9725   return DAG.getNode(RISCVISD::BREV8, DL, VT, Src.getOperand(0));
9726 }
9727 
9728 // Convert from one FMA opcode to another based on whether we are negating the
9729 // multiply result and/or the accumulator.
9730 // NOTE: Only supports RVV operations with VL.
negateFMAOpcode(unsigned Opcode,bool NegMul,bool NegAcc)9731 static unsigned negateFMAOpcode(unsigned Opcode, bool NegMul, bool NegAcc) {
9732   assert((NegMul || NegAcc) && "Not negating anything?");
9733 
9734   // Negating the multiply result changes ADD<->SUB and toggles 'N'.
9735   if (NegMul) {
9736     // clang-format off
9737     switch (Opcode) {
9738     default: llvm_unreachable("Unexpected opcode");
9739     case RISCVISD::VFMADD_VL:  Opcode = RISCVISD::VFNMSUB_VL; break;
9740     case RISCVISD::VFNMSUB_VL: Opcode = RISCVISD::VFMADD_VL;  break;
9741     case RISCVISD::VFNMADD_VL: Opcode = RISCVISD::VFMSUB_VL;  break;
9742     case RISCVISD::VFMSUB_VL:  Opcode = RISCVISD::VFNMADD_VL; break;
9743     }
9744     // clang-format on
9745   }
9746 
9747   // Negating the accumulator changes ADD<->SUB.
9748   if (NegAcc) {
9749     // clang-format off
9750     switch (Opcode) {
9751     default: llvm_unreachable("Unexpected opcode");
9752     case RISCVISD::VFMADD_VL:  Opcode = RISCVISD::VFMSUB_VL;  break;
9753     case RISCVISD::VFMSUB_VL:  Opcode = RISCVISD::VFMADD_VL;  break;
9754     case RISCVISD::VFNMADD_VL: Opcode = RISCVISD::VFNMSUB_VL; break;
9755     case RISCVISD::VFNMSUB_VL: Opcode = RISCVISD::VFNMADD_VL; break;
9756     }
9757     // clang-format on
9758   }
9759 
9760   return Opcode;
9761 }
9762 
performSRACombine(SDNode * N,SelectionDAG & DAG,const RISCVSubtarget & Subtarget)9763 static SDValue performSRACombine(SDNode *N, SelectionDAG &DAG,
9764                                  const RISCVSubtarget &Subtarget) {
9765   assert(N->getOpcode() == ISD::SRA && "Unexpected opcode");
9766 
9767   if (N->getValueType(0) != MVT::i64 || !Subtarget.is64Bit())
9768     return SDValue();
9769 
9770   if (!isa<ConstantSDNode>(N->getOperand(1)))
9771     return SDValue();
9772   uint64_t ShAmt = N->getConstantOperandVal(1);
9773   if (ShAmt > 32)
9774     return SDValue();
9775 
9776   SDValue N0 = N->getOperand(0);
9777 
9778   // Combine (sra (sext_inreg (shl X, C1), i32), C2) ->
9779   // (sra (shl X, C1+32), C2+32) so it gets selected as SLLI+SRAI instead of
9780   // SLLIW+SRAIW. SLLI+SRAI have compressed forms.
9781   if (ShAmt < 32 &&
9782       N0.getOpcode() == ISD::SIGN_EXTEND_INREG && N0.hasOneUse() &&
9783       cast<VTSDNode>(N0.getOperand(1))->getVT() == MVT::i32 &&
9784       N0.getOperand(0).getOpcode() == ISD::SHL && N0.getOperand(0).hasOneUse() &&
9785       isa<ConstantSDNode>(N0.getOperand(0).getOperand(1))) {
9786     uint64_t LShAmt = N0.getOperand(0).getConstantOperandVal(1);
9787     if (LShAmt < 32) {
9788       SDLoc ShlDL(N0.getOperand(0));
9789       SDValue Shl = DAG.getNode(ISD::SHL, ShlDL, MVT::i64,
9790                                 N0.getOperand(0).getOperand(0),
9791                                 DAG.getConstant(LShAmt + 32, ShlDL, MVT::i64));
9792       SDLoc DL(N);
9793       return DAG.getNode(ISD::SRA, DL, MVT::i64, Shl,
9794                          DAG.getConstant(ShAmt + 32, DL, MVT::i64));
9795     }
9796   }
9797 
9798   // Combine (sra (shl X, 32), 32 - C) -> (shl (sext_inreg X, i32), C)
9799   // FIXME: Should this be a generic combine? There's a similar combine on X86.
9800   //
9801   // Also try these folds where an add or sub is in the middle.
9802   // (sra (add (shl X, 32), C1), 32 - C) -> (shl (sext_inreg (add X, C1), C)
9803   // (sra (sub C1, (shl X, 32)), 32 - C) -> (shl (sext_inreg (sub C1, X), C)
9804   SDValue Shl;
9805   ConstantSDNode *AddC = nullptr;
9806 
9807   // We might have an ADD or SUB between the SRA and SHL.
9808   bool IsAdd = N0.getOpcode() == ISD::ADD;
9809   if ((IsAdd || N0.getOpcode() == ISD::SUB)) {
9810     // Other operand needs to be a constant we can modify.
9811     AddC = dyn_cast<ConstantSDNode>(N0.getOperand(IsAdd ? 1 : 0));
9812     if (!AddC)
9813       return SDValue();
9814 
9815     // AddC needs to have at least 32 trailing zeros.
9816     if (AddC->getAPIntValue().countTrailingZeros() < 32)
9817       return SDValue();
9818 
9819     // All users should be a shift by constant less than or equal to 32. This
9820     // ensures we'll do this optimization for each of them to produce an
9821     // add/sub+sext_inreg they can all share.
9822     for (SDNode *U : N0->uses()) {
9823       if (U->getOpcode() != ISD::SRA ||
9824           !isa<ConstantSDNode>(U->getOperand(1)) ||
9825           cast<ConstantSDNode>(U->getOperand(1))->getZExtValue() > 32)
9826         return SDValue();
9827     }
9828 
9829     Shl = N0.getOperand(IsAdd ? 0 : 1);
9830   } else {
9831     // Not an ADD or SUB.
9832     Shl = N0;
9833   }
9834 
9835   // Look for a shift left by 32.
9836   if (Shl.getOpcode() != ISD::SHL || !isa<ConstantSDNode>(Shl.getOperand(1)) ||
9837       Shl.getConstantOperandVal(1) != 32)
9838     return SDValue();
9839 
9840   // We if we didn't look through an add/sub, then the shl should have one use.
9841   // If we did look through an add/sub, the sext_inreg we create is free so
9842   // we're only creating 2 new instructions. It's enough to only remove the
9843   // original sra+add/sub.
9844   if (!AddC && !Shl.hasOneUse())
9845     return SDValue();
9846 
9847   SDLoc DL(N);
9848   SDValue In = Shl.getOperand(0);
9849 
9850   // If we looked through an ADD or SUB, we need to rebuild it with the shifted
9851   // constant.
9852   if (AddC) {
9853     SDValue ShiftedAddC =
9854         DAG.getConstant(AddC->getAPIntValue().lshr(32), DL, MVT::i64);
9855     if (IsAdd)
9856       In = DAG.getNode(ISD::ADD, DL, MVT::i64, In, ShiftedAddC);
9857     else
9858       In = DAG.getNode(ISD::SUB, DL, MVT::i64, ShiftedAddC, In);
9859   }
9860 
9861   SDValue SExt = DAG.getNode(ISD::SIGN_EXTEND_INREG, DL, MVT::i64, In,
9862                              DAG.getValueType(MVT::i32));
9863   if (ShAmt == 32)
9864     return SExt;
9865 
9866   return DAG.getNode(
9867       ISD::SHL, DL, MVT::i64, SExt,
9868       DAG.getConstant(32 - ShAmt, DL, MVT::i64));
9869 }
9870 
9871 // Invert (and/or (set cc X, Y), (xor Z, 1)) to (or/and (set !cc X, Y)), Z) if
9872 // the result is used as the conditon of a br_cc or select_cc we can invert,
9873 // inverting the setcc is free, and Z is 0/1. Caller will invert the
9874 // br_cc/select_cc.
tryDemorganOfBooleanCondition(SDValue Cond,SelectionDAG & DAG)9875 static SDValue tryDemorganOfBooleanCondition(SDValue Cond, SelectionDAG &DAG) {
9876   bool IsAnd = Cond.getOpcode() == ISD::AND;
9877   if (!IsAnd && Cond.getOpcode() != ISD::OR)
9878     return SDValue();
9879 
9880   if (!Cond.hasOneUse())
9881     return SDValue();
9882 
9883   SDValue Setcc = Cond.getOperand(0);
9884   SDValue Xor = Cond.getOperand(1);
9885   // Canonicalize setcc to LHS.
9886   if (Setcc.getOpcode() != ISD::SETCC)
9887     std::swap(Setcc, Xor);
9888   // LHS should be a setcc and RHS should be an xor.
9889   if (Setcc.getOpcode() != ISD::SETCC || !Setcc.hasOneUse() ||
9890       Xor.getOpcode() != ISD::XOR || !Xor.hasOneUse())
9891     return SDValue();
9892 
9893   // If the condition is an And, SimplifyDemandedBits may have changed
9894   // (xor Z, 1) to (not Z).
9895   SDValue Xor1 = Xor.getOperand(1);
9896   if (!isOneConstant(Xor1) && !(IsAnd && isAllOnesConstant(Xor1)))
9897     return SDValue();
9898 
9899   EVT VT = Cond.getValueType();
9900   SDValue Xor0 = Xor.getOperand(0);
9901 
9902   // The LHS of the xor needs to be 0/1.
9903   APInt Mask = APInt::getBitsSetFrom(VT.getSizeInBits(), 1);
9904   if (!DAG.MaskedValueIsZero(Xor0, Mask))
9905     return SDValue();
9906 
9907   // We can only invert integer setccs.
9908   EVT SetCCOpVT = Setcc.getOperand(0).getValueType();
9909   if (!SetCCOpVT.isScalarInteger())
9910     return SDValue();
9911 
9912   ISD::CondCode CCVal = cast<CondCodeSDNode>(Setcc.getOperand(2))->get();
9913   if (ISD::isIntEqualitySetCC(CCVal)) {
9914     CCVal = ISD::getSetCCInverse(CCVal, SetCCOpVT);
9915     Setcc = DAG.getSetCC(SDLoc(Setcc), VT, Setcc.getOperand(0),
9916                          Setcc.getOperand(1), CCVal);
9917   } else if (CCVal == ISD::SETLT && isNullConstant(Setcc.getOperand(0))) {
9918     // Invert (setlt 0, X) by converting to (setlt X, 1).
9919     Setcc = DAG.getSetCC(SDLoc(Setcc), VT, Setcc.getOperand(1),
9920                          DAG.getConstant(1, SDLoc(Setcc), VT), CCVal);
9921   } else if (CCVal == ISD::SETLT && isOneConstant(Setcc.getOperand(1))) {
9922     // (setlt X, 1) by converting to (setlt 0, X).
9923     Setcc = DAG.getSetCC(SDLoc(Setcc), VT,
9924                          DAG.getConstant(0, SDLoc(Setcc), VT),
9925                          Setcc.getOperand(0), CCVal);
9926   } else
9927     return SDValue();
9928 
9929   unsigned Opc = IsAnd ? ISD::OR : ISD::AND;
9930   return DAG.getNode(Opc, SDLoc(Cond), VT, Setcc, Xor.getOperand(0));
9931 }
9932 
9933 // Perform common combines for BR_CC and SELECT_CC condtions.
combine_CC(SDValue & LHS,SDValue & RHS,SDValue & CC,const SDLoc & DL,SelectionDAG & DAG,const RISCVSubtarget & Subtarget)9934 static bool combine_CC(SDValue &LHS, SDValue &RHS, SDValue &CC, const SDLoc &DL,
9935                        SelectionDAG &DAG, const RISCVSubtarget &Subtarget) {
9936   ISD::CondCode CCVal = cast<CondCodeSDNode>(CC)->get();
9937 
9938   // As far as arithmetic right shift always saves the sign,
9939   // shift can be omitted.
9940   // Fold setlt (sra X, N), 0 -> setlt X, 0 and
9941   // setge (sra X, N), 0 -> setge X, 0
9942   if (auto *RHSConst = dyn_cast<ConstantSDNode>(RHS.getNode())) {
9943     if ((CCVal == ISD::SETGE || CCVal == ISD::SETLT) &&
9944         LHS.getOpcode() == ISD::SRA && RHSConst->isZero()) {
9945       LHS = LHS.getOperand(0);
9946       return true;
9947     }
9948   }
9949 
9950   if (!ISD::isIntEqualitySetCC(CCVal))
9951     return false;
9952 
9953   // Fold ((setlt X, Y), 0, ne) -> (X, Y, lt)
9954   // Sometimes the setcc is introduced after br_cc/select_cc has been formed.
9955   if (LHS.getOpcode() == ISD::SETCC && isNullConstant(RHS) &&
9956       LHS.getOperand(0).getValueType() == Subtarget.getXLenVT()) {
9957     // If we're looking for eq 0 instead of ne 0, we need to invert the
9958     // condition.
9959     bool Invert = CCVal == ISD::SETEQ;
9960     CCVal = cast<CondCodeSDNode>(LHS.getOperand(2))->get();
9961     if (Invert)
9962       CCVal = ISD::getSetCCInverse(CCVal, LHS.getValueType());
9963 
9964     RHS = LHS.getOperand(1);
9965     LHS = LHS.getOperand(0);
9966     translateSetCCForBranch(DL, LHS, RHS, CCVal, DAG);
9967 
9968     CC = DAG.getCondCode(CCVal);
9969     return true;
9970   }
9971 
9972   // Fold ((xor X, Y), 0, eq/ne) -> (X, Y, eq/ne)
9973   if (LHS.getOpcode() == ISD::XOR && isNullConstant(RHS)) {
9974     RHS = LHS.getOperand(1);
9975     LHS = LHS.getOperand(0);
9976     return true;
9977   }
9978 
9979   // Fold ((srl (and X, 1<<C), C), 0, eq/ne) -> ((shl X, XLen-1-C), 0, ge/lt)
9980   if (isNullConstant(RHS) && LHS.getOpcode() == ISD::SRL && LHS.hasOneUse() &&
9981       LHS.getOperand(1).getOpcode() == ISD::Constant) {
9982     SDValue LHS0 = LHS.getOperand(0);
9983     if (LHS0.getOpcode() == ISD::AND &&
9984         LHS0.getOperand(1).getOpcode() == ISD::Constant) {
9985       uint64_t Mask = LHS0.getConstantOperandVal(1);
9986       uint64_t ShAmt = LHS.getConstantOperandVal(1);
9987       if (isPowerOf2_64(Mask) && Log2_64(Mask) == ShAmt) {
9988         CCVal = CCVal == ISD::SETEQ ? ISD::SETGE : ISD::SETLT;
9989         CC = DAG.getCondCode(CCVal);
9990 
9991         ShAmt = LHS.getValueSizeInBits() - 1 - ShAmt;
9992         LHS = LHS0.getOperand(0);
9993         if (ShAmt != 0)
9994           LHS =
9995               DAG.getNode(ISD::SHL, DL, LHS.getValueType(), LHS0.getOperand(0),
9996                           DAG.getConstant(ShAmt, DL, LHS.getValueType()));
9997         return true;
9998       }
9999     }
10000   }
10001 
10002   // (X, 1, setne) -> // (X, 0, seteq) if we can prove X is 0/1.
10003   // This can occur when legalizing some floating point comparisons.
10004   APInt Mask = APInt::getBitsSetFrom(LHS.getValueSizeInBits(), 1);
10005   if (isOneConstant(RHS) && DAG.MaskedValueIsZero(LHS, Mask)) {
10006     CCVal = ISD::getSetCCInverse(CCVal, LHS.getValueType());
10007     CC = DAG.getCondCode(CCVal);
10008     RHS = DAG.getConstant(0, DL, LHS.getValueType());
10009     return true;
10010   }
10011 
10012   if (isNullConstant(RHS)) {
10013     if (SDValue NewCond = tryDemorganOfBooleanCondition(LHS, DAG)) {
10014       CCVal = ISD::getSetCCInverse(CCVal, LHS.getValueType());
10015       CC = DAG.getCondCode(CCVal);
10016       LHS = NewCond;
10017       return true;
10018     }
10019   }
10020 
10021   return false;
10022 }
10023 
10024 // Fold
10025 // (select C, (add Y, X), Y) -> (add Y, (select C, X, 0)).
10026 // (select C, (sub Y, X), Y) -> (sub Y, (select C, X, 0)).
10027 // (select C, (or Y, X), Y)  -> (or Y, (select C, X, 0)).
10028 // (select C, (xor Y, X), Y) -> (xor Y, (select C, X, 0)).
tryFoldSelectIntoOp(SDNode * N,SelectionDAG & DAG,SDValue TrueVal,SDValue FalseVal,bool Swapped)10029 static SDValue tryFoldSelectIntoOp(SDNode *N, SelectionDAG &DAG,
10030                                    SDValue TrueVal, SDValue FalseVal,
10031                                    bool Swapped) {
10032   bool Commutative = true;
10033   switch (TrueVal.getOpcode()) {
10034   default:
10035     return SDValue();
10036   case ISD::SUB:
10037     Commutative = false;
10038     break;
10039   case ISD::ADD:
10040   case ISD::OR:
10041   case ISD::XOR:
10042     break;
10043   }
10044 
10045   if (!TrueVal.hasOneUse() || isa<ConstantSDNode>(FalseVal))
10046     return SDValue();
10047 
10048   unsigned OpToFold;
10049   if (FalseVal == TrueVal.getOperand(0))
10050     OpToFold = 0;
10051   else if (Commutative && FalseVal == TrueVal.getOperand(1))
10052     OpToFold = 1;
10053   else
10054     return SDValue();
10055 
10056   EVT VT = N->getValueType(0);
10057   SDLoc DL(N);
10058   SDValue Zero = DAG.getConstant(0, DL, VT);
10059   SDValue OtherOp = TrueVal.getOperand(1 - OpToFold);
10060 
10061   if (Swapped)
10062     std::swap(OtherOp, Zero);
10063   SDValue NewSel = DAG.getSelect(DL, VT, N->getOperand(0), OtherOp, Zero);
10064   return DAG.getNode(TrueVal.getOpcode(), DL, VT, FalseVal, NewSel);
10065 }
10066 
performSELECTCombine(SDNode * N,SelectionDAG & DAG,const RISCVSubtarget & Subtarget)10067 static SDValue performSELECTCombine(SDNode *N, SelectionDAG &DAG,
10068                                     const RISCVSubtarget &Subtarget) {
10069   if (Subtarget.hasShortForwardBranchOpt())
10070     return SDValue();
10071 
10072   // Only support XLenVT.
10073   if (N->getValueType(0) != Subtarget.getXLenVT())
10074     return SDValue();
10075 
10076   SDValue TrueVal = N->getOperand(1);
10077   SDValue FalseVal = N->getOperand(2);
10078   if (SDValue V = tryFoldSelectIntoOp(N, DAG, TrueVal, FalseVal, /*Swapped*/false))
10079     return V;
10080   return tryFoldSelectIntoOp(N, DAG, FalseVal, TrueVal, /*Swapped*/true);
10081 }
10082 
PerformDAGCombine(SDNode * N,DAGCombinerInfo & DCI) const10083 SDValue RISCVTargetLowering::PerformDAGCombine(SDNode *N,
10084                                                DAGCombinerInfo &DCI) const {
10085   SelectionDAG &DAG = DCI.DAG;
10086 
10087   // Helper to call SimplifyDemandedBits on an operand of N where only some low
10088   // bits are demanded. N will be added to the Worklist if it was not deleted.
10089   // Caller should return SDValue(N, 0) if this returns true.
10090   auto SimplifyDemandedLowBitsHelper = [&](unsigned OpNo, unsigned LowBits) {
10091     SDValue Op = N->getOperand(OpNo);
10092     APInt Mask = APInt::getLowBitsSet(Op.getValueSizeInBits(), LowBits);
10093     if (!SimplifyDemandedBits(Op, Mask, DCI))
10094       return false;
10095 
10096     if (N->getOpcode() != ISD::DELETED_NODE)
10097       DCI.AddToWorklist(N);
10098     return true;
10099   };
10100 
10101   switch (N->getOpcode()) {
10102   default:
10103     break;
10104   case RISCVISD::SplitF64: {
10105     SDValue Op0 = N->getOperand(0);
10106     // If the input to SplitF64 is just BuildPairF64 then the operation is
10107     // redundant. Instead, use BuildPairF64's operands directly.
10108     if (Op0->getOpcode() == RISCVISD::BuildPairF64)
10109       return DCI.CombineTo(N, Op0.getOperand(0), Op0.getOperand(1));
10110 
10111     if (Op0->isUndef()) {
10112       SDValue Lo = DAG.getUNDEF(MVT::i32);
10113       SDValue Hi = DAG.getUNDEF(MVT::i32);
10114       return DCI.CombineTo(N, Lo, Hi);
10115     }
10116 
10117     SDLoc DL(N);
10118 
10119     // It's cheaper to materialise two 32-bit integers than to load a double
10120     // from the constant pool and transfer it to integer registers through the
10121     // stack.
10122     if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(Op0)) {
10123       APInt V = C->getValueAPF().bitcastToAPInt();
10124       SDValue Lo = DAG.getConstant(V.trunc(32), DL, MVT::i32);
10125       SDValue Hi = DAG.getConstant(V.lshr(32).trunc(32), DL, MVT::i32);
10126       return DCI.CombineTo(N, Lo, Hi);
10127     }
10128 
10129     // This is a target-specific version of a DAGCombine performed in
10130     // DAGCombiner::visitBITCAST. It performs the equivalent of:
10131     // fold (bitconvert (fneg x)) -> (xor (bitconvert x), signbit)
10132     // fold (bitconvert (fabs x)) -> (and (bitconvert x), (not signbit))
10133     if (!(Op0.getOpcode() == ISD::FNEG || Op0.getOpcode() == ISD::FABS) ||
10134         !Op0.getNode()->hasOneUse())
10135       break;
10136     SDValue NewSplitF64 =
10137         DAG.getNode(RISCVISD::SplitF64, DL, DAG.getVTList(MVT::i32, MVT::i32),
10138                     Op0.getOperand(0));
10139     SDValue Lo = NewSplitF64.getValue(0);
10140     SDValue Hi = NewSplitF64.getValue(1);
10141     APInt SignBit = APInt::getSignMask(32);
10142     if (Op0.getOpcode() == ISD::FNEG) {
10143       SDValue NewHi = DAG.getNode(ISD::XOR, DL, MVT::i32, Hi,
10144                                   DAG.getConstant(SignBit, DL, MVT::i32));
10145       return DCI.CombineTo(N, Lo, NewHi);
10146     }
10147     assert(Op0.getOpcode() == ISD::FABS);
10148     SDValue NewHi = DAG.getNode(ISD::AND, DL, MVT::i32, Hi,
10149                                 DAG.getConstant(~SignBit, DL, MVT::i32));
10150     return DCI.CombineTo(N, Lo, NewHi);
10151   }
10152   case RISCVISD::SLLW:
10153   case RISCVISD::SRAW:
10154   case RISCVISD::SRLW:
10155   case RISCVISD::RORW:
10156   case RISCVISD::ROLW: {
10157     // Only the lower 32 bits of LHS and lower 5 bits of RHS are read.
10158     if (SimplifyDemandedLowBitsHelper(0, 32) ||
10159         SimplifyDemandedLowBitsHelper(1, 5))
10160       return SDValue(N, 0);
10161 
10162     break;
10163   }
10164   case RISCVISD::CLZW:
10165   case RISCVISD::CTZW: {
10166     // Only the lower 32 bits of the first operand are read
10167     if (SimplifyDemandedLowBitsHelper(0, 32))
10168       return SDValue(N, 0);
10169     break;
10170   }
10171   case RISCVISD::FMV_X_ANYEXTH:
10172   case RISCVISD::FMV_X_ANYEXTW_RV64: {
10173     SDLoc DL(N);
10174     SDValue Op0 = N->getOperand(0);
10175     MVT VT = N->getSimpleValueType(0);
10176     // If the input to FMV_X_ANYEXTW_RV64 is just FMV_W_X_RV64 then the
10177     // conversion is unnecessary and can be replaced with the FMV_W_X_RV64
10178     // operand. Similar for FMV_X_ANYEXTH and FMV_H_X.
10179     if ((N->getOpcode() == RISCVISD::FMV_X_ANYEXTW_RV64 &&
10180          Op0->getOpcode() == RISCVISD::FMV_W_X_RV64) ||
10181         (N->getOpcode() == RISCVISD::FMV_X_ANYEXTH &&
10182          Op0->getOpcode() == RISCVISD::FMV_H_X)) {
10183       assert(Op0.getOperand(0).getValueType() == VT &&
10184              "Unexpected value type!");
10185       return Op0.getOperand(0);
10186     }
10187 
10188     // This is a target-specific version of a DAGCombine performed in
10189     // DAGCombiner::visitBITCAST. It performs the equivalent of:
10190     // fold (bitconvert (fneg x)) -> (xor (bitconvert x), signbit)
10191     // fold (bitconvert (fabs x)) -> (and (bitconvert x), (not signbit))
10192     if (!(Op0.getOpcode() == ISD::FNEG || Op0.getOpcode() == ISD::FABS) ||
10193         !Op0.getNode()->hasOneUse())
10194       break;
10195     SDValue NewFMV = DAG.getNode(N->getOpcode(), DL, VT, Op0.getOperand(0));
10196     unsigned FPBits = N->getOpcode() == RISCVISD::FMV_X_ANYEXTW_RV64 ? 32 : 16;
10197     APInt SignBit = APInt::getSignMask(FPBits).sext(VT.getSizeInBits());
10198     if (Op0.getOpcode() == ISD::FNEG)
10199       return DAG.getNode(ISD::XOR, DL, VT, NewFMV,
10200                          DAG.getConstant(SignBit, DL, VT));
10201 
10202     assert(Op0.getOpcode() == ISD::FABS);
10203     return DAG.getNode(ISD::AND, DL, VT, NewFMV,
10204                        DAG.getConstant(~SignBit, DL, VT));
10205   }
10206   case ISD::ADD:
10207     return performADDCombine(N, DAG, Subtarget);
10208   case ISD::SUB:
10209     return performSUBCombine(N, DAG, Subtarget);
10210   case ISD::AND:
10211     return performANDCombine(N, DCI, Subtarget);
10212   case ISD::OR:
10213     return performORCombine(N, DCI, Subtarget);
10214   case ISD::XOR:
10215     return performXORCombine(N, DAG, Subtarget);
10216   case ISD::FADD:
10217   case ISD::UMAX:
10218   case ISD::UMIN:
10219   case ISD::SMAX:
10220   case ISD::SMIN:
10221   case ISD::FMAXNUM:
10222   case ISD::FMINNUM:
10223     return combineBinOpToReduce(N, DAG, Subtarget);
10224   case ISD::SETCC:
10225     return performSETCCCombine(N, DAG, Subtarget);
10226   case ISD::SIGN_EXTEND_INREG:
10227     return performSIGN_EXTEND_INREGCombine(N, DAG, Subtarget);
10228   case ISD::ZERO_EXTEND:
10229     // Fold (zero_extend (fp_to_uint X)) to prevent forming fcvt+zexti32 during
10230     // type legalization. This is safe because fp_to_uint produces poison if
10231     // it overflows.
10232     if (N->getValueType(0) == MVT::i64 && Subtarget.is64Bit()) {
10233       SDValue Src = N->getOperand(0);
10234       if (Src.getOpcode() == ISD::FP_TO_UINT &&
10235           isTypeLegal(Src.getOperand(0).getValueType()))
10236         return DAG.getNode(ISD::FP_TO_UINT, SDLoc(N), MVT::i64,
10237                            Src.getOperand(0));
10238       if (Src.getOpcode() == ISD::STRICT_FP_TO_UINT && Src.hasOneUse() &&
10239           isTypeLegal(Src.getOperand(1).getValueType())) {
10240         SDVTList VTs = DAG.getVTList(MVT::i64, MVT::Other);
10241         SDValue Res = DAG.getNode(ISD::STRICT_FP_TO_UINT, SDLoc(N), VTs,
10242                                   Src.getOperand(0), Src.getOperand(1));
10243         DCI.CombineTo(N, Res);
10244         DAG.ReplaceAllUsesOfValueWith(Src.getValue(1), Res.getValue(1));
10245         DCI.recursivelyDeleteUnusedNodes(Src.getNode());
10246         return SDValue(N, 0); // Return N so it doesn't get rechecked.
10247       }
10248     }
10249     return SDValue();
10250   case ISD::TRUNCATE:
10251     return performTRUNCATECombine(N, DAG, Subtarget);
10252   case ISD::SELECT:
10253     return performSELECTCombine(N, DAG, Subtarget);
10254   case RISCVISD::SELECT_CC: {
10255     // Transform
10256     SDValue LHS = N->getOperand(0);
10257     SDValue RHS = N->getOperand(1);
10258     SDValue CC = N->getOperand(2);
10259     ISD::CondCode CCVal = cast<CondCodeSDNode>(CC)->get();
10260     SDValue TrueV = N->getOperand(3);
10261     SDValue FalseV = N->getOperand(4);
10262     SDLoc DL(N);
10263     EVT VT = N->getValueType(0);
10264 
10265     // If the True and False values are the same, we don't need a select_cc.
10266     if (TrueV == FalseV)
10267       return TrueV;
10268 
10269     // (select (x < 0), y, z)  -> x >> (XLEN - 1) & (y - z) + z
10270     // (select (x >= 0), y, z) -> x >> (XLEN - 1) & (z - y) + y
10271     if (!Subtarget.hasShortForwardBranchOpt() && isa<ConstantSDNode>(TrueV) &&
10272         isa<ConstantSDNode>(FalseV) && isNullConstant(RHS) &&
10273         (CCVal == ISD::CondCode::SETLT || CCVal == ISD::CondCode::SETGE)) {
10274       if (CCVal == ISD::CondCode::SETGE)
10275         std::swap(TrueV, FalseV);
10276 
10277       int64_t TrueSImm = cast<ConstantSDNode>(TrueV)->getSExtValue();
10278       int64_t FalseSImm = cast<ConstantSDNode>(FalseV)->getSExtValue();
10279       // Only handle simm12, if it is not in this range, it can be considered as
10280       // register.
10281       if (isInt<12>(TrueSImm) && isInt<12>(FalseSImm) &&
10282           isInt<12>(TrueSImm - FalseSImm)) {
10283         SDValue SRA =
10284             DAG.getNode(ISD::SRA, DL, VT, LHS,
10285                         DAG.getConstant(Subtarget.getXLen() - 1, DL, VT));
10286         SDValue AND =
10287             DAG.getNode(ISD::AND, DL, VT, SRA,
10288                         DAG.getConstant(TrueSImm - FalseSImm, DL, VT));
10289         return DAG.getNode(ISD::ADD, DL, VT, AND, FalseV);
10290       }
10291 
10292       if (CCVal == ISD::CondCode::SETGE)
10293         std::swap(TrueV, FalseV);
10294     }
10295 
10296     if (combine_CC(LHS, RHS, CC, DL, DAG, Subtarget))
10297       return DAG.getNode(RISCVISD::SELECT_CC, DL, N->getValueType(0),
10298                          {LHS, RHS, CC, TrueV, FalseV});
10299 
10300     if (!Subtarget.hasShortForwardBranchOpt()) {
10301       // (select c, -1, y) -> -c | y
10302       if (isAllOnesConstant(TrueV)) {
10303         SDValue C = DAG.getSetCC(DL, VT, LHS, RHS, CCVal);
10304         SDValue Neg = DAG.getNegative(C, DL, VT);
10305         return DAG.getNode(ISD::OR, DL, VT, Neg, FalseV);
10306       }
10307       // (select c, y, -1) -> -!c | y
10308       if (isAllOnesConstant(FalseV)) {
10309         SDValue C =
10310             DAG.getSetCC(DL, VT, LHS, RHS, ISD::getSetCCInverse(CCVal, VT));
10311         SDValue Neg = DAG.getNegative(C, DL, VT);
10312         return DAG.getNode(ISD::OR, DL, VT, Neg, TrueV);
10313       }
10314 
10315       // (select c, 0, y) -> -!c & y
10316       if (isNullConstant(TrueV)) {
10317         SDValue C =
10318             DAG.getSetCC(DL, VT, LHS, RHS, ISD::getSetCCInverse(CCVal, VT));
10319         SDValue Neg = DAG.getNegative(C, DL, VT);
10320         return DAG.getNode(ISD::AND, DL, VT, Neg, FalseV);
10321       }
10322       // (select c, y, 0) -> -c & y
10323       if (isNullConstant(FalseV)) {
10324         SDValue C = DAG.getSetCC(DL, VT, LHS, RHS, CCVal);
10325         SDValue Neg = DAG.getNegative(C, DL, VT);
10326         return DAG.getNode(ISD::AND, DL, VT, Neg, TrueV);
10327       }
10328     }
10329 
10330     return SDValue();
10331   }
10332   case RISCVISD::BR_CC: {
10333     SDValue LHS = N->getOperand(1);
10334     SDValue RHS = N->getOperand(2);
10335     SDValue CC = N->getOperand(3);
10336     SDLoc DL(N);
10337 
10338     if (combine_CC(LHS, RHS, CC, DL, DAG, Subtarget))
10339       return DAG.getNode(RISCVISD::BR_CC, DL, N->getValueType(0),
10340                          N->getOperand(0), LHS, RHS, CC, N->getOperand(4));
10341 
10342     return SDValue();
10343   }
10344   case ISD::BITREVERSE:
10345     return performBITREVERSECombine(N, DAG, Subtarget);
10346   case ISD::FP_TO_SINT:
10347   case ISD::FP_TO_UINT:
10348     return performFP_TO_INTCombine(N, DCI, Subtarget);
10349   case ISD::FP_TO_SINT_SAT:
10350   case ISD::FP_TO_UINT_SAT:
10351     return performFP_TO_INT_SATCombine(N, DCI, Subtarget);
10352   case ISD::FCOPYSIGN: {
10353     EVT VT = N->getValueType(0);
10354     if (!VT.isVector())
10355       break;
10356     // There is a form of VFSGNJ which injects the negated sign of its second
10357     // operand. Try and bubble any FNEG up after the extend/round to produce
10358     // this optimized pattern. Avoid modifying cases where FP_ROUND and
10359     // TRUNC=1.
10360     SDValue In2 = N->getOperand(1);
10361     // Avoid cases where the extend/round has multiple uses, as duplicating
10362     // those is typically more expensive than removing a fneg.
10363     if (!In2.hasOneUse())
10364       break;
10365     if (In2.getOpcode() != ISD::FP_EXTEND &&
10366         (In2.getOpcode() != ISD::FP_ROUND || In2.getConstantOperandVal(1) != 0))
10367       break;
10368     In2 = In2.getOperand(0);
10369     if (In2.getOpcode() != ISD::FNEG)
10370       break;
10371     SDLoc DL(N);
10372     SDValue NewFPExtRound = DAG.getFPExtendOrRound(In2.getOperand(0), DL, VT);
10373     return DAG.getNode(ISD::FCOPYSIGN, DL, VT, N->getOperand(0),
10374                        DAG.getNode(ISD::FNEG, DL, VT, NewFPExtRound));
10375   }
10376   case ISD::MGATHER:
10377   case ISD::MSCATTER:
10378   case ISD::VP_GATHER:
10379   case ISD::VP_SCATTER: {
10380     if (!DCI.isBeforeLegalize())
10381       break;
10382     SDValue Index, ScaleOp;
10383     bool IsIndexSigned = false;
10384     if (const auto *VPGSN = dyn_cast<VPGatherScatterSDNode>(N)) {
10385       Index = VPGSN->getIndex();
10386       ScaleOp = VPGSN->getScale();
10387       IsIndexSigned = VPGSN->isIndexSigned();
10388       assert(!VPGSN->isIndexScaled() &&
10389              "Scaled gather/scatter should not be formed");
10390     } else {
10391       const auto *MGSN = cast<MaskedGatherScatterSDNode>(N);
10392       Index = MGSN->getIndex();
10393       ScaleOp = MGSN->getScale();
10394       IsIndexSigned = MGSN->isIndexSigned();
10395       assert(!MGSN->isIndexScaled() &&
10396              "Scaled gather/scatter should not be formed");
10397 
10398     }
10399     EVT IndexVT = Index.getValueType();
10400     MVT XLenVT = Subtarget.getXLenVT();
10401     // RISCV indexed loads only support the "unsigned unscaled" addressing
10402     // mode, so anything else must be manually legalized.
10403     bool NeedsIdxLegalization =
10404         (IsIndexSigned && IndexVT.getVectorElementType().bitsLT(XLenVT));
10405     if (!NeedsIdxLegalization)
10406       break;
10407 
10408     SDLoc DL(N);
10409 
10410     // Any index legalization should first promote to XLenVT, so we don't lose
10411     // bits when scaling. This may create an illegal index type so we let
10412     // LLVM's legalization take care of the splitting.
10413     // FIXME: LLVM can't split VP_GATHER or VP_SCATTER yet.
10414     if (IndexVT.getVectorElementType().bitsLT(XLenVT)) {
10415       IndexVT = IndexVT.changeVectorElementType(XLenVT);
10416       Index = DAG.getNode(IsIndexSigned ? ISD::SIGN_EXTEND : ISD::ZERO_EXTEND,
10417                           DL, IndexVT, Index);
10418     }
10419 
10420     ISD::MemIndexType NewIndexTy = ISD::UNSIGNED_SCALED;
10421     if (const auto *VPGN = dyn_cast<VPGatherSDNode>(N))
10422       return DAG.getGatherVP(N->getVTList(), VPGN->getMemoryVT(), DL,
10423                              {VPGN->getChain(), VPGN->getBasePtr(), Index,
10424                               ScaleOp, VPGN->getMask(),
10425                               VPGN->getVectorLength()},
10426                              VPGN->getMemOperand(), NewIndexTy);
10427     if (const auto *VPSN = dyn_cast<VPScatterSDNode>(N))
10428       return DAG.getScatterVP(N->getVTList(), VPSN->getMemoryVT(), DL,
10429                               {VPSN->getChain(), VPSN->getValue(),
10430                                VPSN->getBasePtr(), Index, ScaleOp,
10431                                VPSN->getMask(), VPSN->getVectorLength()},
10432                               VPSN->getMemOperand(), NewIndexTy);
10433     if (const auto *MGN = dyn_cast<MaskedGatherSDNode>(N))
10434       return DAG.getMaskedGather(
10435           N->getVTList(), MGN->getMemoryVT(), DL,
10436           {MGN->getChain(), MGN->getPassThru(), MGN->getMask(),
10437            MGN->getBasePtr(), Index, ScaleOp},
10438           MGN->getMemOperand(), NewIndexTy, MGN->getExtensionType());
10439     const auto *MSN = cast<MaskedScatterSDNode>(N);
10440     return DAG.getMaskedScatter(
10441         N->getVTList(), MSN->getMemoryVT(), DL,
10442         {MSN->getChain(), MSN->getValue(), MSN->getMask(), MSN->getBasePtr(),
10443          Index, ScaleOp},
10444         MSN->getMemOperand(), NewIndexTy, MSN->isTruncatingStore());
10445   }
10446   case RISCVISD::SRA_VL:
10447   case RISCVISD::SRL_VL:
10448   case RISCVISD::SHL_VL: {
10449     SDValue ShAmt = N->getOperand(1);
10450     if (ShAmt.getOpcode() == RISCVISD::SPLAT_VECTOR_SPLIT_I64_VL) {
10451       // We don't need the upper 32 bits of a 64-bit element for a shift amount.
10452       SDLoc DL(N);
10453       SDValue VL = N->getOperand(3);
10454       EVT VT = N->getValueType(0);
10455       ShAmt = DAG.getNode(RISCVISD::VMV_V_X_VL, DL, VT, DAG.getUNDEF(VT),
10456                           ShAmt.getOperand(1), VL);
10457       return DAG.getNode(N->getOpcode(), DL, VT, N->getOperand(0), ShAmt,
10458                          N->getOperand(2), N->getOperand(3), N->getOperand(4));
10459     }
10460     break;
10461   }
10462   case ISD::SRA:
10463     if (SDValue V = performSRACombine(N, DAG, Subtarget))
10464       return V;
10465     [[fallthrough]];
10466   case ISD::SRL:
10467   case ISD::SHL: {
10468     SDValue ShAmt = N->getOperand(1);
10469     if (ShAmt.getOpcode() == RISCVISD::SPLAT_VECTOR_SPLIT_I64_VL) {
10470       // We don't need the upper 32 bits of a 64-bit element for a shift amount.
10471       SDLoc DL(N);
10472       EVT VT = N->getValueType(0);
10473       ShAmt = DAG.getNode(RISCVISD::VMV_V_X_VL, DL, VT, DAG.getUNDEF(VT),
10474                           ShAmt.getOperand(1),
10475                           DAG.getRegister(RISCV::X0, Subtarget.getXLenVT()));
10476       return DAG.getNode(N->getOpcode(), DL, VT, N->getOperand(0), ShAmt);
10477     }
10478     break;
10479   }
10480   case RISCVISD::ADD_VL:
10481   case RISCVISD::SUB_VL:
10482   case RISCVISD::VWADD_W_VL:
10483   case RISCVISD::VWADDU_W_VL:
10484   case RISCVISD::VWSUB_W_VL:
10485   case RISCVISD::VWSUBU_W_VL:
10486   case RISCVISD::MUL_VL:
10487     return combineBinOp_VLToVWBinOp_VL(N, DCI);
10488   case RISCVISD::VFMADD_VL:
10489   case RISCVISD::VFNMADD_VL:
10490   case RISCVISD::VFMSUB_VL:
10491   case RISCVISD::VFNMSUB_VL: {
10492     // Fold FNEG_VL into FMA opcodes.
10493     SDValue A = N->getOperand(0);
10494     SDValue B = N->getOperand(1);
10495     SDValue C = N->getOperand(2);
10496     SDValue Mask = N->getOperand(3);
10497     SDValue VL = N->getOperand(4);
10498 
10499     auto invertIfNegative = [&Mask, &VL](SDValue &V) {
10500       if (V.getOpcode() == RISCVISD::FNEG_VL && V.getOperand(1) == Mask &&
10501           V.getOperand(2) == VL) {
10502         // Return the negated input.
10503         V = V.getOperand(0);
10504         return true;
10505       }
10506 
10507       return false;
10508     };
10509 
10510     bool NegA = invertIfNegative(A);
10511     bool NegB = invertIfNegative(B);
10512     bool NegC = invertIfNegative(C);
10513 
10514     // If no operands are negated, we're done.
10515     if (!NegA && !NegB && !NegC)
10516       return SDValue();
10517 
10518     unsigned NewOpcode = negateFMAOpcode(N->getOpcode(), NegA != NegB, NegC);
10519     return DAG.getNode(NewOpcode, SDLoc(N), N->getValueType(0), A, B, C, Mask,
10520                        VL);
10521   }
10522   case ISD::STORE: {
10523     auto *Store = cast<StoreSDNode>(N);
10524     SDValue Val = Store->getValue();
10525     // Combine store of vmv.x.s/vfmv.f.s to vse with VL of 1.
10526     // vfmv.f.s is represented as extract element from 0. Match it late to avoid
10527     // any illegal types.
10528     if (Val.getOpcode() == RISCVISD::VMV_X_S ||
10529         (DCI.isAfterLegalizeDAG() &&
10530          Val.getOpcode() == ISD::EXTRACT_VECTOR_ELT &&
10531          isNullConstant(Val.getOperand(1)))) {
10532       SDValue Src = Val.getOperand(0);
10533       MVT VecVT = Src.getSimpleValueType();
10534       EVT MemVT = Store->getMemoryVT();
10535       // VecVT should be scalable and memory VT should match the element type.
10536       if (VecVT.isScalableVector() &&
10537           MemVT == VecVT.getVectorElementType()) {
10538         SDLoc DL(N);
10539         MVT MaskVT = getMaskTypeFor(VecVT);
10540         return DAG.getStoreVP(
10541             Store->getChain(), DL, Src, Store->getBasePtr(), Store->getOffset(),
10542             DAG.getConstant(1, DL, MaskVT),
10543             DAG.getConstant(1, DL, Subtarget.getXLenVT()), MemVT,
10544             Store->getMemOperand(), Store->getAddressingMode(),
10545             Store->isTruncatingStore(), /*IsCompress*/ false);
10546       }
10547     }
10548 
10549     break;
10550   }
10551   case ISD::SPLAT_VECTOR: {
10552     EVT VT = N->getValueType(0);
10553     // Only perform this combine on legal MVT types.
10554     if (!isTypeLegal(VT))
10555       break;
10556     if (auto Gather = matchSplatAsGather(N->getOperand(0), VT.getSimpleVT(), N,
10557                                          DAG, Subtarget))
10558       return Gather;
10559     break;
10560   }
10561   case RISCVISD::VMV_V_X_VL: {
10562     // Tail agnostic VMV.V.X only demands the vector element bitwidth from the
10563     // scalar input.
10564     unsigned ScalarSize = N->getOperand(1).getValueSizeInBits();
10565     unsigned EltWidth = N->getValueType(0).getScalarSizeInBits();
10566     if (ScalarSize > EltWidth && N->getOperand(0).isUndef())
10567       if (SimplifyDemandedLowBitsHelper(1, EltWidth))
10568         return SDValue(N, 0);
10569 
10570     break;
10571   }
10572   case RISCVISD::VFMV_S_F_VL: {
10573     SDValue Src = N->getOperand(1);
10574     // Try to remove vector->scalar->vector if the scalar->vector is inserting
10575     // into an undef vector.
10576     // TODO: Could use a vslide or vmv.v.v for non-undef.
10577     if (N->getOperand(0).isUndef() &&
10578         Src.getOpcode() == ISD::EXTRACT_VECTOR_ELT &&
10579         isNullConstant(Src.getOperand(1)) &&
10580         Src.getOperand(0).getValueType().isScalableVector()) {
10581       EVT VT = N->getValueType(0);
10582       EVT SrcVT = Src.getOperand(0).getValueType();
10583       assert(SrcVT.getVectorElementType() == VT.getVectorElementType());
10584       // Widths match, just return the original vector.
10585       if (SrcVT == VT)
10586         return Src.getOperand(0);
10587       // TODO: Use insert_subvector/extract_subvector to change widen/narrow?
10588     }
10589     break;
10590   }
10591   case ISD::INTRINSIC_WO_CHAIN: {
10592     unsigned IntNo = N->getConstantOperandVal(0);
10593     switch (IntNo) {
10594       // By default we do not combine any intrinsic.
10595     default:
10596       return SDValue();
10597     case Intrinsic::riscv_vcpop:
10598     case Intrinsic::riscv_vcpop_mask:
10599     case Intrinsic::riscv_vfirst:
10600     case Intrinsic::riscv_vfirst_mask: {
10601       SDValue VL = N->getOperand(2);
10602       if (IntNo == Intrinsic::riscv_vcpop_mask ||
10603           IntNo == Intrinsic::riscv_vfirst_mask)
10604         VL = N->getOperand(3);
10605       if (!isNullConstant(VL))
10606         return SDValue();
10607       // If VL is 0, vcpop -> li 0, vfirst -> li -1.
10608       SDLoc DL(N);
10609       EVT VT = N->getValueType(0);
10610       if (IntNo == Intrinsic::riscv_vfirst ||
10611           IntNo == Intrinsic::riscv_vfirst_mask)
10612         return DAG.getConstant(-1, DL, VT);
10613       return DAG.getConstant(0, DL, VT);
10614     }
10615     }
10616   }
10617   case ISD::BITCAST: {
10618     assert(Subtarget.useRVVForFixedLengthVectors());
10619     SDValue N0 = N->getOperand(0);
10620     EVT VT = N->getValueType(0);
10621     EVT SrcVT = N0.getValueType();
10622     // If this is a bitcast between a MVT::v4i1/v2i1/v1i1 and an illegal integer
10623     // type, widen both sides to avoid a trip through memory.
10624     if ((SrcVT == MVT::v1i1 || SrcVT == MVT::v2i1 || SrcVT == MVT::v4i1) &&
10625         VT.isScalarInteger()) {
10626       unsigned NumConcats = 8 / SrcVT.getVectorNumElements();
10627       SmallVector<SDValue, 4> Ops(NumConcats, DAG.getUNDEF(SrcVT));
10628       Ops[0] = N0;
10629       SDLoc DL(N);
10630       N0 = DAG.getNode(ISD::CONCAT_VECTORS, DL, MVT::v8i1, Ops);
10631       N0 = DAG.getBitcast(MVT::i8, N0);
10632       return DAG.getNode(ISD::TRUNCATE, DL, VT, N0);
10633     }
10634 
10635     return SDValue();
10636   }
10637   }
10638 
10639   return SDValue();
10640 }
10641 
isDesirableToCommuteWithShift(const SDNode * N,CombineLevel Level) const10642 bool RISCVTargetLowering::isDesirableToCommuteWithShift(
10643     const SDNode *N, CombineLevel Level) const {
10644   assert((N->getOpcode() == ISD::SHL || N->getOpcode() == ISD::SRA ||
10645           N->getOpcode() == ISD::SRL) &&
10646          "Expected shift op");
10647 
10648   // The following folds are only desirable if `(OP _, c1 << c2)` can be
10649   // materialised in fewer instructions than `(OP _, c1)`:
10650   //
10651   //   (shl (add x, c1), c2) -> (add (shl x, c2), c1 << c2)
10652   //   (shl (or x, c1), c2) -> (or (shl x, c2), c1 << c2)
10653   SDValue N0 = N->getOperand(0);
10654   EVT Ty = N0.getValueType();
10655   if (Ty.isScalarInteger() &&
10656       (N0.getOpcode() == ISD::ADD || N0.getOpcode() == ISD::OR)) {
10657     auto *C1 = dyn_cast<ConstantSDNode>(N0->getOperand(1));
10658     auto *C2 = dyn_cast<ConstantSDNode>(N->getOperand(1));
10659     if (C1 && C2) {
10660       const APInt &C1Int = C1->getAPIntValue();
10661       APInt ShiftedC1Int = C1Int << C2->getAPIntValue();
10662 
10663       // We can materialise `c1 << c2` into an add immediate, so it's "free",
10664       // and the combine should happen, to potentially allow further combines
10665       // later.
10666       if (ShiftedC1Int.getMinSignedBits() <= 64 &&
10667           isLegalAddImmediate(ShiftedC1Int.getSExtValue()))
10668         return true;
10669 
10670       // We can materialise `c1` in an add immediate, so it's "free", and the
10671       // combine should be prevented.
10672       if (C1Int.getMinSignedBits() <= 64 &&
10673           isLegalAddImmediate(C1Int.getSExtValue()))
10674         return false;
10675 
10676       // Neither constant will fit into an immediate, so find materialisation
10677       // costs.
10678       int C1Cost = RISCVMatInt::getIntMatCost(C1Int, Ty.getSizeInBits(),
10679                                               Subtarget.getFeatureBits(),
10680                                               /*CompressionCost*/true);
10681       int ShiftedC1Cost = RISCVMatInt::getIntMatCost(
10682           ShiftedC1Int, Ty.getSizeInBits(), Subtarget.getFeatureBits(),
10683           /*CompressionCost*/true);
10684 
10685       // Materialising `c1` is cheaper than materialising `c1 << c2`, so the
10686       // combine should be prevented.
10687       if (C1Cost < ShiftedC1Cost)
10688         return false;
10689     }
10690   }
10691   return true;
10692 }
10693 
targetShrinkDemandedConstant(SDValue Op,const APInt & DemandedBits,const APInt & DemandedElts,TargetLoweringOpt & TLO) const10694 bool RISCVTargetLowering::targetShrinkDemandedConstant(
10695     SDValue Op, const APInt &DemandedBits, const APInt &DemandedElts,
10696     TargetLoweringOpt &TLO) const {
10697   // Delay this optimization as late as possible.
10698   if (!TLO.LegalOps)
10699     return false;
10700 
10701   EVT VT = Op.getValueType();
10702   if (VT.isVector())
10703     return false;
10704 
10705   unsigned Opcode = Op.getOpcode();
10706   if (Opcode != ISD::AND && Opcode != ISD::OR && Opcode != ISD::XOR)
10707     return false;
10708 
10709   ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op.getOperand(1));
10710   if (!C)
10711     return false;
10712 
10713   const APInt &Mask = C->getAPIntValue();
10714 
10715   // Clear all non-demanded bits initially.
10716   APInt ShrunkMask = Mask & DemandedBits;
10717 
10718   // Try to make a smaller immediate by setting undemanded bits.
10719 
10720   APInt ExpandedMask = Mask | ~DemandedBits;
10721 
10722   auto IsLegalMask = [ShrunkMask, ExpandedMask](const APInt &Mask) -> bool {
10723     return ShrunkMask.isSubsetOf(Mask) && Mask.isSubsetOf(ExpandedMask);
10724   };
10725   auto UseMask = [Mask, Op, &TLO](const APInt &NewMask) -> bool {
10726     if (NewMask == Mask)
10727       return true;
10728     SDLoc DL(Op);
10729     SDValue NewC = TLO.DAG.getConstant(NewMask, DL, Op.getValueType());
10730     SDValue NewOp = TLO.DAG.getNode(Op.getOpcode(), DL, Op.getValueType(),
10731                                     Op.getOperand(0), NewC);
10732     return TLO.CombineTo(Op, NewOp);
10733   };
10734 
10735   // If the shrunk mask fits in sign extended 12 bits, let the target
10736   // independent code apply it.
10737   if (ShrunkMask.isSignedIntN(12))
10738     return false;
10739 
10740   // And has a few special cases for zext.
10741   if (Opcode == ISD::AND) {
10742     // Preserve (and X, 0xffff), if zext.h exists use zext.h,
10743     // otherwise use SLLI + SRLI.
10744     APInt NewMask = APInt(Mask.getBitWidth(), 0xffff);
10745     if (IsLegalMask(NewMask))
10746       return UseMask(NewMask);
10747 
10748     // Try to preserve (and X, 0xffffffff), the (zext_inreg X, i32) pattern.
10749     if (VT == MVT::i64) {
10750       APInt NewMask = APInt(64, 0xffffffff);
10751       if (IsLegalMask(NewMask))
10752         return UseMask(NewMask);
10753     }
10754   }
10755 
10756   // For the remaining optimizations, we need to be able to make a negative
10757   // number through a combination of mask and undemanded bits.
10758   if (!ExpandedMask.isNegative())
10759     return false;
10760 
10761   // What is the fewest number of bits we need to represent the negative number.
10762   unsigned MinSignedBits = ExpandedMask.getMinSignedBits();
10763 
10764   // Try to make a 12 bit negative immediate. If that fails try to make a 32
10765   // bit negative immediate unless the shrunk immediate already fits in 32 bits.
10766   // If we can't create a simm12, we shouldn't change opaque constants.
10767   APInt NewMask = ShrunkMask;
10768   if (MinSignedBits <= 12)
10769     NewMask.setBitsFrom(11);
10770   else if (!C->isOpaque() && MinSignedBits <= 32 && !ShrunkMask.isSignedIntN(32))
10771     NewMask.setBitsFrom(31);
10772   else
10773     return false;
10774 
10775   // Check that our new mask is a subset of the demanded mask.
10776   assert(IsLegalMask(NewMask));
10777   return UseMask(NewMask);
10778 }
10779 
computeGREVOrGORC(uint64_t x,unsigned ShAmt,bool IsGORC)10780 static uint64_t computeGREVOrGORC(uint64_t x, unsigned ShAmt, bool IsGORC) {
10781   static const uint64_t GREVMasks[] = {
10782       0x5555555555555555ULL, 0x3333333333333333ULL, 0x0F0F0F0F0F0F0F0FULL,
10783       0x00FF00FF00FF00FFULL, 0x0000FFFF0000FFFFULL, 0x00000000FFFFFFFFULL};
10784 
10785   for (unsigned Stage = 0; Stage != 6; ++Stage) {
10786     unsigned Shift = 1 << Stage;
10787     if (ShAmt & Shift) {
10788       uint64_t Mask = GREVMasks[Stage];
10789       uint64_t Res = ((x & Mask) << Shift) | ((x >> Shift) & Mask);
10790       if (IsGORC)
10791         Res |= x;
10792       x = Res;
10793     }
10794   }
10795 
10796   return x;
10797 }
10798 
computeKnownBitsForTargetNode(const SDValue Op,KnownBits & Known,const APInt & DemandedElts,const SelectionDAG & DAG,unsigned Depth) const10799 void RISCVTargetLowering::computeKnownBitsForTargetNode(const SDValue Op,
10800                                                         KnownBits &Known,
10801                                                         const APInt &DemandedElts,
10802                                                         const SelectionDAG &DAG,
10803                                                         unsigned Depth) const {
10804   unsigned BitWidth = Known.getBitWidth();
10805   unsigned Opc = Op.getOpcode();
10806   assert((Opc >= ISD::BUILTIN_OP_END ||
10807           Opc == ISD::INTRINSIC_WO_CHAIN ||
10808           Opc == ISD::INTRINSIC_W_CHAIN ||
10809           Opc == ISD::INTRINSIC_VOID) &&
10810          "Should use MaskedValueIsZero if you don't know whether Op"
10811          " is a target node!");
10812 
10813   Known.resetAll();
10814   switch (Opc) {
10815   default: break;
10816   case RISCVISD::SELECT_CC: {
10817     Known = DAG.computeKnownBits(Op.getOperand(4), Depth + 1);
10818     // If we don't know any bits, early out.
10819     if (Known.isUnknown())
10820       break;
10821     KnownBits Known2 = DAG.computeKnownBits(Op.getOperand(3), Depth + 1);
10822 
10823     // Only known if known in both the LHS and RHS.
10824     Known = KnownBits::commonBits(Known, Known2);
10825     break;
10826   }
10827   case RISCVISD::REMUW: {
10828     KnownBits Known2;
10829     Known = DAG.computeKnownBits(Op.getOperand(0), DemandedElts, Depth + 1);
10830     Known2 = DAG.computeKnownBits(Op.getOperand(1), DemandedElts, Depth + 1);
10831     // We only care about the lower 32 bits.
10832     Known = KnownBits::urem(Known.trunc(32), Known2.trunc(32));
10833     // Restore the original width by sign extending.
10834     Known = Known.sext(BitWidth);
10835     break;
10836   }
10837   case RISCVISD::DIVUW: {
10838     KnownBits Known2;
10839     Known = DAG.computeKnownBits(Op.getOperand(0), DemandedElts, Depth + 1);
10840     Known2 = DAG.computeKnownBits(Op.getOperand(1), DemandedElts, Depth + 1);
10841     // We only care about the lower 32 bits.
10842     Known = KnownBits::udiv(Known.trunc(32), Known2.trunc(32));
10843     // Restore the original width by sign extending.
10844     Known = Known.sext(BitWidth);
10845     break;
10846   }
10847   case RISCVISD::CTZW: {
10848     KnownBits Known2 = DAG.computeKnownBits(Op.getOperand(0), Depth + 1);
10849     unsigned PossibleTZ = Known2.trunc(32).countMaxTrailingZeros();
10850     unsigned LowBits = llvm::bit_width(PossibleTZ);
10851     Known.Zero.setBitsFrom(LowBits);
10852     break;
10853   }
10854   case RISCVISD::CLZW: {
10855     KnownBits Known2 = DAG.computeKnownBits(Op.getOperand(0), Depth + 1);
10856     unsigned PossibleLZ = Known2.trunc(32).countMaxLeadingZeros();
10857     unsigned LowBits = llvm::bit_width(PossibleLZ);
10858     Known.Zero.setBitsFrom(LowBits);
10859     break;
10860   }
10861   case RISCVISD::BREV8:
10862   case RISCVISD::ORC_B: {
10863     // FIXME: This is based on the non-ratified Zbp GREV and GORC where a
10864     // control value of 7 is equivalent to brev8 and orc.b.
10865     Known = DAG.computeKnownBits(Op.getOperand(0), Depth + 1);
10866     bool IsGORC = Op.getOpcode() == RISCVISD::ORC_B;
10867     // To compute zeros, we need to invert the value and invert it back after.
10868     Known.Zero =
10869         ~computeGREVOrGORC(~Known.Zero.getZExtValue(), 7, IsGORC);
10870     Known.One = computeGREVOrGORC(Known.One.getZExtValue(), 7, IsGORC);
10871     break;
10872   }
10873   case RISCVISD::READ_VLENB: {
10874     // We can use the minimum and maximum VLEN values to bound VLENB.  We
10875     // know VLEN must be a power of two.
10876     const unsigned MinVLenB = Subtarget.getRealMinVLen() / 8;
10877     const unsigned MaxVLenB = Subtarget.getRealMaxVLen() / 8;
10878     assert(MinVLenB > 0 && "READ_VLENB without vector extension enabled?");
10879     Known.Zero.setLowBits(Log2_32(MinVLenB));
10880     Known.Zero.setBitsFrom(Log2_32(MaxVLenB)+1);
10881     if (MaxVLenB == MinVLenB)
10882       Known.One.setBit(Log2_32(MinVLenB));
10883     break;
10884   }
10885   case ISD::INTRINSIC_W_CHAIN:
10886   case ISD::INTRINSIC_WO_CHAIN: {
10887     unsigned IntNo =
10888         Op.getConstantOperandVal(Opc == ISD::INTRINSIC_WO_CHAIN ? 0 : 1);
10889     switch (IntNo) {
10890     default:
10891       // We can't do anything for most intrinsics.
10892       break;
10893     case Intrinsic::riscv_vsetvli:
10894     case Intrinsic::riscv_vsetvlimax:
10895     case Intrinsic::riscv_vsetvli_opt:
10896     case Intrinsic::riscv_vsetvlimax_opt:
10897       // Assume that VL output is positive and would fit in an int32_t.
10898       // TODO: VLEN might be capped at 16 bits in a future V spec update.
10899       if (BitWidth >= 32)
10900         Known.Zero.setBitsFrom(31);
10901       break;
10902     }
10903     break;
10904   }
10905   }
10906 }
10907 
ComputeNumSignBitsForTargetNode(SDValue Op,const APInt & DemandedElts,const SelectionDAG & DAG,unsigned Depth) const10908 unsigned RISCVTargetLowering::ComputeNumSignBitsForTargetNode(
10909     SDValue Op, const APInt &DemandedElts, const SelectionDAG &DAG,
10910     unsigned Depth) const {
10911   switch (Op.getOpcode()) {
10912   default:
10913     break;
10914   case RISCVISD::SELECT_CC: {
10915     unsigned Tmp =
10916         DAG.ComputeNumSignBits(Op.getOperand(3), DemandedElts, Depth + 1);
10917     if (Tmp == 1) return 1;  // Early out.
10918     unsigned Tmp2 =
10919         DAG.ComputeNumSignBits(Op.getOperand(4), DemandedElts, Depth + 1);
10920     return std::min(Tmp, Tmp2);
10921   }
10922   case RISCVISD::ABSW: {
10923     // We expand this at isel to negw+max. The result will have 33 sign bits
10924     // if the input has at least 33 sign bits.
10925     unsigned Tmp =
10926         DAG.ComputeNumSignBits(Op.getOperand(0), DemandedElts, Depth + 1);
10927     if (Tmp < 33) return 1;
10928     return 33;
10929   }
10930   case RISCVISD::SLLW:
10931   case RISCVISD::SRAW:
10932   case RISCVISD::SRLW:
10933   case RISCVISD::DIVW:
10934   case RISCVISD::DIVUW:
10935   case RISCVISD::REMUW:
10936   case RISCVISD::ROLW:
10937   case RISCVISD::RORW:
10938   case RISCVISD::FCVT_W_RV64:
10939   case RISCVISD::FCVT_WU_RV64:
10940   case RISCVISD::STRICT_FCVT_W_RV64:
10941   case RISCVISD::STRICT_FCVT_WU_RV64:
10942     // TODO: As the result is sign-extended, this is conservatively correct. A
10943     // more precise answer could be calculated for SRAW depending on known
10944     // bits in the shift amount.
10945     return 33;
10946   case RISCVISD::VMV_X_S: {
10947     // The number of sign bits of the scalar result is computed by obtaining the
10948     // element type of the input vector operand, subtracting its width from the
10949     // XLEN, and then adding one (sign bit within the element type). If the
10950     // element type is wider than XLen, the least-significant XLEN bits are
10951     // taken.
10952     unsigned XLen = Subtarget.getXLen();
10953     unsigned EltBits = Op.getOperand(0).getScalarValueSizeInBits();
10954     if (EltBits <= XLen)
10955       return XLen - EltBits + 1;
10956     break;
10957   }
10958   case ISD::INTRINSIC_W_CHAIN: {
10959     unsigned IntNo = Op.getConstantOperandVal(1);
10960     switch (IntNo) {
10961     default:
10962       break;
10963     case Intrinsic::riscv_masked_atomicrmw_xchg_i64:
10964     case Intrinsic::riscv_masked_atomicrmw_add_i64:
10965     case Intrinsic::riscv_masked_atomicrmw_sub_i64:
10966     case Intrinsic::riscv_masked_atomicrmw_nand_i64:
10967     case Intrinsic::riscv_masked_atomicrmw_max_i64:
10968     case Intrinsic::riscv_masked_atomicrmw_min_i64:
10969     case Intrinsic::riscv_masked_atomicrmw_umax_i64:
10970     case Intrinsic::riscv_masked_atomicrmw_umin_i64:
10971     case Intrinsic::riscv_masked_cmpxchg_i64:
10972       // riscv_masked_{atomicrmw_*,cmpxchg} intrinsics represent an emulated
10973       // narrow atomic operation. These are implemented using atomic
10974       // operations at the minimum supported atomicrmw/cmpxchg width whose
10975       // result is then sign extended to XLEN. With +A, the minimum width is
10976       // 32 for both 64 and 32.
10977       assert(Subtarget.getXLen() == 64);
10978       assert(getMinCmpXchgSizeInBits() == 32);
10979       assert(Subtarget.hasStdExtA());
10980       return 33;
10981     }
10982   }
10983   }
10984 
10985   return 1;
10986 }
10987 
10988 const Constant *
getTargetConstantFromLoad(LoadSDNode * Ld) const10989 RISCVTargetLowering::getTargetConstantFromLoad(LoadSDNode *Ld) const {
10990   assert(Ld && "Unexpected null LoadSDNode");
10991   if (!ISD::isNormalLoad(Ld))
10992     return nullptr;
10993 
10994   SDValue Ptr = Ld->getBasePtr();
10995 
10996   // Only constant pools with no offset are supported.
10997   auto GetSupportedConstantPool = [](SDValue Ptr) -> ConstantPoolSDNode * {
10998     auto *CNode = dyn_cast<ConstantPoolSDNode>(Ptr);
10999     if (!CNode || CNode->isMachineConstantPoolEntry() ||
11000         CNode->getOffset() != 0)
11001       return nullptr;
11002 
11003     return CNode;
11004   };
11005 
11006   // Simple case, LLA.
11007   if (Ptr.getOpcode() == RISCVISD::LLA) {
11008     auto *CNode = GetSupportedConstantPool(Ptr);
11009     if (!CNode || CNode->getTargetFlags() != 0)
11010       return nullptr;
11011 
11012     return CNode->getConstVal();
11013   }
11014 
11015   // Look for a HI and ADD_LO pair.
11016   if (Ptr.getOpcode() != RISCVISD::ADD_LO ||
11017       Ptr.getOperand(0).getOpcode() != RISCVISD::HI)
11018     return nullptr;
11019 
11020   auto *CNodeLo = GetSupportedConstantPool(Ptr.getOperand(1));
11021   auto *CNodeHi = GetSupportedConstantPool(Ptr.getOperand(0).getOperand(0));
11022 
11023   if (!CNodeLo || CNodeLo->getTargetFlags() != RISCVII::MO_LO ||
11024       !CNodeHi || CNodeHi->getTargetFlags() != RISCVII::MO_HI)
11025     return nullptr;
11026 
11027   if (CNodeLo->getConstVal() != CNodeHi->getConstVal())
11028     return nullptr;
11029 
11030   return CNodeLo->getConstVal();
11031 }
11032 
emitReadCycleWidePseudo(MachineInstr & MI,MachineBasicBlock * BB)11033 static MachineBasicBlock *emitReadCycleWidePseudo(MachineInstr &MI,
11034                                                   MachineBasicBlock *BB) {
11035   assert(MI.getOpcode() == RISCV::ReadCycleWide && "Unexpected instruction");
11036 
11037   // To read the 64-bit cycle CSR on a 32-bit target, we read the two halves.
11038   // Should the count have wrapped while it was being read, we need to try
11039   // again.
11040   // ...
11041   // read:
11042   // rdcycleh x3 # load high word of cycle
11043   // rdcycle  x2 # load low word of cycle
11044   // rdcycleh x4 # load high word of cycle
11045   // bne x3, x4, read # check if high word reads match, otherwise try again
11046   // ...
11047 
11048   MachineFunction &MF = *BB->getParent();
11049   const BasicBlock *LLVM_BB = BB->getBasicBlock();
11050   MachineFunction::iterator It = ++BB->getIterator();
11051 
11052   MachineBasicBlock *LoopMBB = MF.CreateMachineBasicBlock(LLVM_BB);
11053   MF.insert(It, LoopMBB);
11054 
11055   MachineBasicBlock *DoneMBB = MF.CreateMachineBasicBlock(LLVM_BB);
11056   MF.insert(It, DoneMBB);
11057 
11058   // Transfer the remainder of BB and its successor edges to DoneMBB.
11059   DoneMBB->splice(DoneMBB->begin(), BB,
11060                   std::next(MachineBasicBlock::iterator(MI)), BB->end());
11061   DoneMBB->transferSuccessorsAndUpdatePHIs(BB);
11062 
11063   BB->addSuccessor(LoopMBB);
11064 
11065   MachineRegisterInfo &RegInfo = MF.getRegInfo();
11066   Register ReadAgainReg = RegInfo.createVirtualRegister(&RISCV::GPRRegClass);
11067   Register LoReg = MI.getOperand(0).getReg();
11068   Register HiReg = MI.getOperand(1).getReg();
11069   DebugLoc DL = MI.getDebugLoc();
11070 
11071   const TargetInstrInfo *TII = MF.getSubtarget().getInstrInfo();
11072   BuildMI(LoopMBB, DL, TII->get(RISCV::CSRRS), HiReg)
11073       .addImm(RISCVSysReg::lookupSysRegByName("CYCLEH")->Encoding)
11074       .addReg(RISCV::X0);
11075   BuildMI(LoopMBB, DL, TII->get(RISCV::CSRRS), LoReg)
11076       .addImm(RISCVSysReg::lookupSysRegByName("CYCLE")->Encoding)
11077       .addReg(RISCV::X0);
11078   BuildMI(LoopMBB, DL, TII->get(RISCV::CSRRS), ReadAgainReg)
11079       .addImm(RISCVSysReg::lookupSysRegByName("CYCLEH")->Encoding)
11080       .addReg(RISCV::X0);
11081 
11082   BuildMI(LoopMBB, DL, TII->get(RISCV::BNE))
11083       .addReg(HiReg)
11084       .addReg(ReadAgainReg)
11085       .addMBB(LoopMBB);
11086 
11087   LoopMBB->addSuccessor(LoopMBB);
11088   LoopMBB->addSuccessor(DoneMBB);
11089 
11090   MI.eraseFromParent();
11091 
11092   return DoneMBB;
11093 }
11094 
emitSplitF64Pseudo(MachineInstr & MI,MachineBasicBlock * BB)11095 static MachineBasicBlock *emitSplitF64Pseudo(MachineInstr &MI,
11096                                              MachineBasicBlock *BB) {
11097   assert(MI.getOpcode() == RISCV::SplitF64Pseudo && "Unexpected instruction");
11098 
11099   MachineFunction &MF = *BB->getParent();
11100   DebugLoc DL = MI.getDebugLoc();
11101   const TargetInstrInfo &TII = *MF.getSubtarget().getInstrInfo();
11102   const TargetRegisterInfo *RI = MF.getSubtarget().getRegisterInfo();
11103   Register LoReg = MI.getOperand(0).getReg();
11104   Register HiReg = MI.getOperand(1).getReg();
11105   Register SrcReg = MI.getOperand(2).getReg();
11106   const TargetRegisterClass *SrcRC = &RISCV::FPR64RegClass;
11107   int FI = MF.getInfo<RISCVMachineFunctionInfo>()->getMoveF64FrameIndex(MF);
11108 
11109   TII.storeRegToStackSlot(*BB, MI, SrcReg, MI.getOperand(2).isKill(), FI, SrcRC,
11110                           RI, Register());
11111   MachinePointerInfo MPI = MachinePointerInfo::getFixedStack(MF, FI);
11112   MachineMemOperand *MMOLo =
11113       MF.getMachineMemOperand(MPI, MachineMemOperand::MOLoad, 4, Align(8));
11114   MachineMemOperand *MMOHi = MF.getMachineMemOperand(
11115       MPI.getWithOffset(4), MachineMemOperand::MOLoad, 4, Align(8));
11116   BuildMI(*BB, MI, DL, TII.get(RISCV::LW), LoReg)
11117       .addFrameIndex(FI)
11118       .addImm(0)
11119       .addMemOperand(MMOLo);
11120   BuildMI(*BB, MI, DL, TII.get(RISCV::LW), HiReg)
11121       .addFrameIndex(FI)
11122       .addImm(4)
11123       .addMemOperand(MMOHi);
11124   MI.eraseFromParent(); // The pseudo instruction is gone now.
11125   return BB;
11126 }
11127 
emitBuildPairF64Pseudo(MachineInstr & MI,MachineBasicBlock * BB)11128 static MachineBasicBlock *emitBuildPairF64Pseudo(MachineInstr &MI,
11129                                                  MachineBasicBlock *BB) {
11130   assert(MI.getOpcode() == RISCV::BuildPairF64Pseudo &&
11131          "Unexpected instruction");
11132 
11133   MachineFunction &MF = *BB->getParent();
11134   DebugLoc DL = MI.getDebugLoc();
11135   const TargetInstrInfo &TII = *MF.getSubtarget().getInstrInfo();
11136   const TargetRegisterInfo *RI = MF.getSubtarget().getRegisterInfo();
11137   Register DstReg = MI.getOperand(0).getReg();
11138   Register LoReg = MI.getOperand(1).getReg();
11139   Register HiReg = MI.getOperand(2).getReg();
11140   const TargetRegisterClass *DstRC = &RISCV::FPR64RegClass;
11141   int FI = MF.getInfo<RISCVMachineFunctionInfo>()->getMoveF64FrameIndex(MF);
11142 
11143   MachinePointerInfo MPI = MachinePointerInfo::getFixedStack(MF, FI);
11144   MachineMemOperand *MMOLo =
11145       MF.getMachineMemOperand(MPI, MachineMemOperand::MOStore, 4, Align(8));
11146   MachineMemOperand *MMOHi = MF.getMachineMemOperand(
11147       MPI.getWithOffset(4), MachineMemOperand::MOStore, 4, Align(8));
11148   BuildMI(*BB, MI, DL, TII.get(RISCV::SW))
11149       .addReg(LoReg, getKillRegState(MI.getOperand(1).isKill()))
11150       .addFrameIndex(FI)
11151       .addImm(0)
11152       .addMemOperand(MMOLo);
11153   BuildMI(*BB, MI, DL, TII.get(RISCV::SW))
11154       .addReg(HiReg, getKillRegState(MI.getOperand(2).isKill()))
11155       .addFrameIndex(FI)
11156       .addImm(4)
11157       .addMemOperand(MMOHi);
11158   TII.loadRegFromStackSlot(*BB, MI, DstReg, FI, DstRC, RI, Register());
11159   MI.eraseFromParent(); // The pseudo instruction is gone now.
11160   return BB;
11161 }
11162 
isSelectPseudo(MachineInstr & MI)11163 static bool isSelectPseudo(MachineInstr &MI) {
11164   switch (MI.getOpcode()) {
11165   default:
11166     return false;
11167   case RISCV::Select_GPR_Using_CC_GPR:
11168   case RISCV::Select_FPR16_Using_CC_GPR:
11169   case RISCV::Select_FPR32_Using_CC_GPR:
11170   case RISCV::Select_FPR64_Using_CC_GPR:
11171     return true;
11172   }
11173 }
11174 
emitQuietFCMP(MachineInstr & MI,MachineBasicBlock * BB,unsigned RelOpcode,unsigned EqOpcode,const RISCVSubtarget & Subtarget)11175 static MachineBasicBlock *emitQuietFCMP(MachineInstr &MI, MachineBasicBlock *BB,
11176                                         unsigned RelOpcode, unsigned EqOpcode,
11177                                         const RISCVSubtarget &Subtarget) {
11178   DebugLoc DL = MI.getDebugLoc();
11179   Register DstReg = MI.getOperand(0).getReg();
11180   Register Src1Reg = MI.getOperand(1).getReg();
11181   Register Src2Reg = MI.getOperand(2).getReg();
11182   MachineRegisterInfo &MRI = BB->getParent()->getRegInfo();
11183   Register SavedFFlags = MRI.createVirtualRegister(&RISCV::GPRRegClass);
11184   const TargetInstrInfo &TII = *BB->getParent()->getSubtarget().getInstrInfo();
11185 
11186   // Save the current FFLAGS.
11187   BuildMI(*BB, MI, DL, TII.get(RISCV::ReadFFLAGS), SavedFFlags);
11188 
11189   auto MIB = BuildMI(*BB, MI, DL, TII.get(RelOpcode), DstReg)
11190                  .addReg(Src1Reg)
11191                  .addReg(Src2Reg);
11192   if (MI.getFlag(MachineInstr::MIFlag::NoFPExcept))
11193     MIB->setFlag(MachineInstr::MIFlag::NoFPExcept);
11194 
11195   // Restore the FFLAGS.
11196   BuildMI(*BB, MI, DL, TII.get(RISCV::WriteFFLAGS))
11197       .addReg(SavedFFlags, RegState::Kill);
11198 
11199   // Issue a dummy FEQ opcode to raise exception for signaling NaNs.
11200   auto MIB2 = BuildMI(*BB, MI, DL, TII.get(EqOpcode), RISCV::X0)
11201                   .addReg(Src1Reg, getKillRegState(MI.getOperand(1).isKill()))
11202                   .addReg(Src2Reg, getKillRegState(MI.getOperand(2).isKill()));
11203   if (MI.getFlag(MachineInstr::MIFlag::NoFPExcept))
11204     MIB2->setFlag(MachineInstr::MIFlag::NoFPExcept);
11205 
11206   // Erase the pseudoinstruction.
11207   MI.eraseFromParent();
11208   return BB;
11209 }
11210 
11211 static MachineBasicBlock *
EmitLoweredCascadedSelect(MachineInstr & First,MachineInstr & Second,MachineBasicBlock * ThisMBB,const RISCVSubtarget & Subtarget)11212 EmitLoweredCascadedSelect(MachineInstr &First, MachineInstr &Second,
11213                           MachineBasicBlock *ThisMBB,
11214                           const RISCVSubtarget &Subtarget) {
11215   // Select_FPRX_ (rs1, rs2, imm, rs4, (Select_FPRX_ rs1, rs2, imm, rs4, rs5)
11216   // Without this, custom-inserter would have generated:
11217   //
11218   //   A
11219   //   | \
11220   //   |  B
11221   //   | /
11222   //   C
11223   //   | \
11224   //   |  D
11225   //   | /
11226   //   E
11227   //
11228   // A: X = ...; Y = ...
11229   // B: empty
11230   // C: Z = PHI [X, A], [Y, B]
11231   // D: empty
11232   // E: PHI [X, C], [Z, D]
11233   //
11234   // If we lower both Select_FPRX_ in a single step, we can instead generate:
11235   //
11236   //   A
11237   //   | \
11238   //   |  C
11239   //   | /|
11240   //   |/ |
11241   //   |  |
11242   //   |  D
11243   //   | /
11244   //   E
11245   //
11246   // A: X = ...; Y = ...
11247   // D: empty
11248   // E: PHI [X, A], [X, C], [Y, D]
11249 
11250   const RISCVInstrInfo &TII = *Subtarget.getInstrInfo();
11251   const DebugLoc &DL = First.getDebugLoc();
11252   const BasicBlock *LLVM_BB = ThisMBB->getBasicBlock();
11253   MachineFunction *F = ThisMBB->getParent();
11254   MachineBasicBlock *FirstMBB = F->CreateMachineBasicBlock(LLVM_BB);
11255   MachineBasicBlock *SecondMBB = F->CreateMachineBasicBlock(LLVM_BB);
11256   MachineBasicBlock *SinkMBB = F->CreateMachineBasicBlock(LLVM_BB);
11257   MachineFunction::iterator It = ++ThisMBB->getIterator();
11258   F->insert(It, FirstMBB);
11259   F->insert(It, SecondMBB);
11260   F->insert(It, SinkMBB);
11261 
11262   // Transfer the remainder of ThisMBB and its successor edges to SinkMBB.
11263   SinkMBB->splice(SinkMBB->begin(), ThisMBB,
11264                   std::next(MachineBasicBlock::iterator(First)),
11265                   ThisMBB->end());
11266   SinkMBB->transferSuccessorsAndUpdatePHIs(ThisMBB);
11267 
11268   // Fallthrough block for ThisMBB.
11269   ThisMBB->addSuccessor(FirstMBB);
11270   // Fallthrough block for FirstMBB.
11271   FirstMBB->addSuccessor(SecondMBB);
11272   ThisMBB->addSuccessor(SinkMBB);
11273   FirstMBB->addSuccessor(SinkMBB);
11274   // This is fallthrough.
11275   SecondMBB->addSuccessor(SinkMBB);
11276 
11277   auto FirstCC = static_cast<RISCVCC::CondCode>(First.getOperand(3).getImm());
11278   Register FLHS = First.getOperand(1).getReg();
11279   Register FRHS = First.getOperand(2).getReg();
11280   // Insert appropriate branch.
11281   BuildMI(FirstMBB, DL, TII.getBrCond(FirstCC))
11282       .addReg(FLHS)
11283       .addReg(FRHS)
11284       .addMBB(SinkMBB);
11285 
11286   Register SLHS = Second.getOperand(1).getReg();
11287   Register SRHS = Second.getOperand(2).getReg();
11288   Register Op1Reg4 = First.getOperand(4).getReg();
11289   Register Op1Reg5 = First.getOperand(5).getReg();
11290 
11291   auto SecondCC = static_cast<RISCVCC::CondCode>(Second.getOperand(3).getImm());
11292   // Insert appropriate branch.
11293   BuildMI(ThisMBB, DL, TII.getBrCond(SecondCC))
11294       .addReg(SLHS)
11295       .addReg(SRHS)
11296       .addMBB(SinkMBB);
11297 
11298   Register DestReg = Second.getOperand(0).getReg();
11299   Register Op2Reg4 = Second.getOperand(4).getReg();
11300   BuildMI(*SinkMBB, SinkMBB->begin(), DL, TII.get(RISCV::PHI), DestReg)
11301       .addReg(Op2Reg4)
11302       .addMBB(ThisMBB)
11303       .addReg(Op1Reg4)
11304       .addMBB(FirstMBB)
11305       .addReg(Op1Reg5)
11306       .addMBB(SecondMBB);
11307 
11308   // Now remove the Select_FPRX_s.
11309   First.eraseFromParent();
11310   Second.eraseFromParent();
11311   return SinkMBB;
11312 }
11313 
emitSelectPseudo(MachineInstr & MI,MachineBasicBlock * BB,const RISCVSubtarget & Subtarget)11314 static MachineBasicBlock *emitSelectPseudo(MachineInstr &MI,
11315                                            MachineBasicBlock *BB,
11316                                            const RISCVSubtarget &Subtarget) {
11317   // To "insert" Select_* instructions, we actually have to insert the triangle
11318   // control-flow pattern.  The incoming instructions know the destination vreg
11319   // to set, the condition code register to branch on, the true/false values to
11320   // select between, and the condcode to use to select the appropriate branch.
11321   //
11322   // We produce the following control flow:
11323   //     HeadMBB
11324   //     |  \
11325   //     |  IfFalseMBB
11326   //     | /
11327   //    TailMBB
11328   //
11329   // When we find a sequence of selects we attempt to optimize their emission
11330   // by sharing the control flow. Currently we only handle cases where we have
11331   // multiple selects with the exact same condition (same LHS, RHS and CC).
11332   // The selects may be interleaved with other instructions if the other
11333   // instructions meet some requirements we deem safe:
11334   // - They are not pseudo instructions.
11335   // - They are debug instructions. Otherwise,
11336   // - They do not have side-effects, do not access memory and their inputs do
11337   //   not depend on the results of the select pseudo-instructions.
11338   // The TrueV/FalseV operands of the selects cannot depend on the result of
11339   // previous selects in the sequence.
11340   // These conditions could be further relaxed. See the X86 target for a
11341   // related approach and more information.
11342   //
11343   // Select_FPRX_ (rs1, rs2, imm, rs4, (Select_FPRX_ rs1, rs2, imm, rs4, rs5))
11344   // is checked here and handled by a separate function -
11345   // EmitLoweredCascadedSelect.
11346   Register LHS = MI.getOperand(1).getReg();
11347   Register RHS = MI.getOperand(2).getReg();
11348   auto CC = static_cast<RISCVCC::CondCode>(MI.getOperand(3).getImm());
11349 
11350   SmallVector<MachineInstr *, 4> SelectDebugValues;
11351   SmallSet<Register, 4> SelectDests;
11352   SelectDests.insert(MI.getOperand(0).getReg());
11353 
11354   MachineInstr *LastSelectPseudo = &MI;
11355   auto Next = next_nodbg(MI.getIterator(), BB->instr_end());
11356   if (MI.getOpcode() != RISCV::Select_GPR_Using_CC_GPR && Next != BB->end() &&
11357       Next->getOpcode() == MI.getOpcode() &&
11358       Next->getOperand(5).getReg() == MI.getOperand(0).getReg() &&
11359       Next->getOperand(5).isKill()) {
11360     return EmitLoweredCascadedSelect(MI, *Next, BB, Subtarget);
11361   }
11362 
11363   for (auto E = BB->end(), SequenceMBBI = MachineBasicBlock::iterator(MI);
11364        SequenceMBBI != E; ++SequenceMBBI) {
11365     if (SequenceMBBI->isDebugInstr())
11366       continue;
11367     if (isSelectPseudo(*SequenceMBBI)) {
11368       if (SequenceMBBI->getOperand(1).getReg() != LHS ||
11369           SequenceMBBI->getOperand(2).getReg() != RHS ||
11370           SequenceMBBI->getOperand(3).getImm() != CC ||
11371           SelectDests.count(SequenceMBBI->getOperand(4).getReg()) ||
11372           SelectDests.count(SequenceMBBI->getOperand(5).getReg()))
11373         break;
11374       LastSelectPseudo = &*SequenceMBBI;
11375       SequenceMBBI->collectDebugValues(SelectDebugValues);
11376       SelectDests.insert(SequenceMBBI->getOperand(0).getReg());
11377       continue;
11378     }
11379     if (SequenceMBBI->hasUnmodeledSideEffects() ||
11380         SequenceMBBI->mayLoadOrStore() ||
11381         SequenceMBBI->usesCustomInsertionHook())
11382       break;
11383     if (llvm::any_of(SequenceMBBI->operands(), [&](MachineOperand &MO) {
11384           return MO.isReg() && MO.isUse() && SelectDests.count(MO.getReg());
11385         }))
11386       break;
11387   }
11388 
11389   const RISCVInstrInfo &TII = *Subtarget.getInstrInfo();
11390   const BasicBlock *LLVM_BB = BB->getBasicBlock();
11391   DebugLoc DL = MI.getDebugLoc();
11392   MachineFunction::iterator I = ++BB->getIterator();
11393 
11394   MachineBasicBlock *HeadMBB = BB;
11395   MachineFunction *F = BB->getParent();
11396   MachineBasicBlock *TailMBB = F->CreateMachineBasicBlock(LLVM_BB);
11397   MachineBasicBlock *IfFalseMBB = F->CreateMachineBasicBlock(LLVM_BB);
11398 
11399   F->insert(I, IfFalseMBB);
11400   F->insert(I, TailMBB);
11401 
11402   // Transfer debug instructions associated with the selects to TailMBB.
11403   for (MachineInstr *DebugInstr : SelectDebugValues) {
11404     TailMBB->push_back(DebugInstr->removeFromParent());
11405   }
11406 
11407   // Move all instructions after the sequence to TailMBB.
11408   TailMBB->splice(TailMBB->end(), HeadMBB,
11409                   std::next(LastSelectPseudo->getIterator()), HeadMBB->end());
11410   // Update machine-CFG edges by transferring all successors of the current
11411   // block to the new block which will contain the Phi nodes for the selects.
11412   TailMBB->transferSuccessorsAndUpdatePHIs(HeadMBB);
11413   // Set the successors for HeadMBB.
11414   HeadMBB->addSuccessor(IfFalseMBB);
11415   HeadMBB->addSuccessor(TailMBB);
11416 
11417   // Insert appropriate branch.
11418   BuildMI(HeadMBB, DL, TII.getBrCond(CC))
11419     .addReg(LHS)
11420     .addReg(RHS)
11421     .addMBB(TailMBB);
11422 
11423   // IfFalseMBB just falls through to TailMBB.
11424   IfFalseMBB->addSuccessor(TailMBB);
11425 
11426   // Create PHIs for all of the select pseudo-instructions.
11427   auto SelectMBBI = MI.getIterator();
11428   auto SelectEnd = std::next(LastSelectPseudo->getIterator());
11429   auto InsertionPoint = TailMBB->begin();
11430   while (SelectMBBI != SelectEnd) {
11431     auto Next = std::next(SelectMBBI);
11432     if (isSelectPseudo(*SelectMBBI)) {
11433       // %Result = phi [ %TrueValue, HeadMBB ], [ %FalseValue, IfFalseMBB ]
11434       BuildMI(*TailMBB, InsertionPoint, SelectMBBI->getDebugLoc(),
11435               TII.get(RISCV::PHI), SelectMBBI->getOperand(0).getReg())
11436           .addReg(SelectMBBI->getOperand(4).getReg())
11437           .addMBB(HeadMBB)
11438           .addReg(SelectMBBI->getOperand(5).getReg())
11439           .addMBB(IfFalseMBB);
11440       SelectMBBI->eraseFromParent();
11441     }
11442     SelectMBBI = Next;
11443   }
11444 
11445   F->getProperties().reset(MachineFunctionProperties::Property::NoPHIs);
11446   return TailMBB;
11447 }
11448 
11449 static MachineBasicBlock *
emitVFCVT_RM_MASK(MachineInstr & MI,MachineBasicBlock * BB,unsigned Opcode)11450 emitVFCVT_RM_MASK(MachineInstr &MI, MachineBasicBlock *BB, unsigned Opcode) {
11451   DebugLoc DL = MI.getDebugLoc();
11452 
11453   const TargetInstrInfo &TII = *BB->getParent()->getSubtarget().getInstrInfo();
11454 
11455   MachineRegisterInfo &MRI = BB->getParent()->getRegInfo();
11456   Register SavedFRM = MRI.createVirtualRegister(&RISCV::GPRRegClass);
11457 
11458   // Update FRM and save the old value.
11459   BuildMI(*BB, MI, DL, TII.get(RISCV::SwapFRMImm), SavedFRM)
11460       .addImm(MI.getOperand(4).getImm());
11461 
11462   // Emit an VFCVT without the FRM operand.
11463   assert(MI.getNumOperands() == 8);
11464   auto MIB = BuildMI(*BB, MI, DL, TII.get(Opcode))
11465                  .add(MI.getOperand(0))
11466                  .add(MI.getOperand(1))
11467                  .add(MI.getOperand(2))
11468                  .add(MI.getOperand(3))
11469                  .add(MI.getOperand(5))
11470                  .add(MI.getOperand(6))
11471                  .add(MI.getOperand(7));
11472   if (MI.getFlag(MachineInstr::MIFlag::NoFPExcept))
11473     MIB->setFlag(MachineInstr::MIFlag::NoFPExcept);
11474 
11475   // Restore FRM.
11476   BuildMI(*BB, MI, DL, TII.get(RISCV::WriteFRM))
11477       .addReg(SavedFRM, RegState::Kill);
11478 
11479   // Erase the pseudoinstruction.
11480   MI.eraseFromParent();
11481   return BB;
11482 }
11483 
emitVFROUND_NOEXCEPT_MASK(MachineInstr & MI,MachineBasicBlock * BB,unsigned CVTXOpc,unsigned CVTFOpc)11484 static MachineBasicBlock *emitVFROUND_NOEXCEPT_MASK(MachineInstr &MI,
11485                                                     MachineBasicBlock *BB,
11486                                                     unsigned CVTXOpc,
11487                                                     unsigned CVTFOpc) {
11488   DebugLoc DL = MI.getDebugLoc();
11489 
11490   const TargetInstrInfo &TII = *BB->getParent()->getSubtarget().getInstrInfo();
11491 
11492   MachineRegisterInfo &MRI = BB->getParent()->getRegInfo();
11493   Register SavedFFLAGS = MRI.createVirtualRegister(&RISCV::GPRRegClass);
11494 
11495   // Save the old value of FFLAGS.
11496   BuildMI(*BB, MI, DL, TII.get(RISCV::ReadFFLAGS), SavedFFLAGS);
11497 
11498   assert(MI.getNumOperands() == 7);
11499 
11500   // Emit a VFCVT_X_F
11501   const TargetRegisterInfo *TRI =
11502       BB->getParent()->getSubtarget().getRegisterInfo();
11503   const TargetRegisterClass *RC = MI.getRegClassConstraint(0, &TII, TRI);
11504   Register Tmp = MRI.createVirtualRegister(RC);
11505   BuildMI(*BB, MI, DL, TII.get(CVTXOpc), Tmp)
11506       .add(MI.getOperand(1))
11507       .add(MI.getOperand(2))
11508       .add(MI.getOperand(3))
11509       .add(MI.getOperand(4))
11510       .add(MI.getOperand(5))
11511       .add(MI.getOperand(6));
11512 
11513   // Emit a VFCVT_F_X
11514   BuildMI(*BB, MI, DL, TII.get(CVTFOpc))
11515       .add(MI.getOperand(0))
11516       .add(MI.getOperand(1))
11517       .addReg(Tmp)
11518       .add(MI.getOperand(3))
11519       .add(MI.getOperand(4))
11520       .add(MI.getOperand(5))
11521       .add(MI.getOperand(6));
11522 
11523   // Restore FFLAGS.
11524   BuildMI(*BB, MI, DL, TII.get(RISCV::WriteFFLAGS))
11525       .addReg(SavedFFLAGS, RegState::Kill);
11526 
11527   // Erase the pseudoinstruction.
11528   MI.eraseFromParent();
11529   return BB;
11530 }
11531 
emitFROUND(MachineInstr & MI,MachineBasicBlock * MBB,const RISCVSubtarget & Subtarget)11532 static MachineBasicBlock *emitFROUND(MachineInstr &MI, MachineBasicBlock *MBB,
11533                                      const RISCVSubtarget &Subtarget) {
11534   unsigned CmpOpc, F2IOpc, I2FOpc, FSGNJOpc, FSGNJXOpc;
11535   const TargetRegisterClass *RC;
11536   switch (MI.getOpcode()) {
11537   default:
11538     llvm_unreachable("Unexpected opcode");
11539   case RISCV::PseudoFROUND_H:
11540     CmpOpc = RISCV::FLT_H;
11541     F2IOpc = RISCV::FCVT_W_H;
11542     I2FOpc = RISCV::FCVT_H_W;
11543     FSGNJOpc = RISCV::FSGNJ_H;
11544     FSGNJXOpc = RISCV::FSGNJX_H;
11545     RC = &RISCV::FPR16RegClass;
11546     break;
11547   case RISCV::PseudoFROUND_S:
11548     CmpOpc = RISCV::FLT_S;
11549     F2IOpc = RISCV::FCVT_W_S;
11550     I2FOpc = RISCV::FCVT_S_W;
11551     FSGNJOpc = RISCV::FSGNJ_S;
11552     FSGNJXOpc = RISCV::FSGNJX_S;
11553     RC = &RISCV::FPR32RegClass;
11554     break;
11555   case RISCV::PseudoFROUND_D:
11556     assert(Subtarget.is64Bit() && "Expected 64-bit GPR.");
11557     CmpOpc = RISCV::FLT_D;
11558     F2IOpc = RISCV::FCVT_L_D;
11559     I2FOpc = RISCV::FCVT_D_L;
11560     FSGNJOpc = RISCV::FSGNJ_D;
11561     FSGNJXOpc = RISCV::FSGNJX_D;
11562     RC = &RISCV::FPR64RegClass;
11563     break;
11564   }
11565 
11566   const BasicBlock *BB = MBB->getBasicBlock();
11567   DebugLoc DL = MI.getDebugLoc();
11568   MachineFunction::iterator I = ++MBB->getIterator();
11569 
11570   MachineFunction *F = MBB->getParent();
11571   MachineBasicBlock *CvtMBB = F->CreateMachineBasicBlock(BB);
11572   MachineBasicBlock *DoneMBB = F->CreateMachineBasicBlock(BB);
11573 
11574   F->insert(I, CvtMBB);
11575   F->insert(I, DoneMBB);
11576   // Move all instructions after the sequence to DoneMBB.
11577   DoneMBB->splice(DoneMBB->end(), MBB, MachineBasicBlock::iterator(MI),
11578                   MBB->end());
11579   // Update machine-CFG edges by transferring all successors of the current
11580   // block to the new block which will contain the Phi nodes for the selects.
11581   DoneMBB->transferSuccessorsAndUpdatePHIs(MBB);
11582   // Set the successors for MBB.
11583   MBB->addSuccessor(CvtMBB);
11584   MBB->addSuccessor(DoneMBB);
11585 
11586   Register DstReg = MI.getOperand(0).getReg();
11587   Register SrcReg = MI.getOperand(1).getReg();
11588   Register MaxReg = MI.getOperand(2).getReg();
11589   int64_t FRM = MI.getOperand(3).getImm();
11590 
11591   const RISCVInstrInfo &TII = *Subtarget.getInstrInfo();
11592   MachineRegisterInfo &MRI = MBB->getParent()->getRegInfo();
11593 
11594   Register FabsReg = MRI.createVirtualRegister(RC);
11595   BuildMI(MBB, DL, TII.get(FSGNJXOpc), FabsReg).addReg(SrcReg).addReg(SrcReg);
11596 
11597   // Compare the FP value to the max value.
11598   Register CmpReg = MRI.createVirtualRegister(&RISCV::GPRRegClass);
11599   auto MIB =
11600       BuildMI(MBB, DL, TII.get(CmpOpc), CmpReg).addReg(FabsReg).addReg(MaxReg);
11601   if (MI.getFlag(MachineInstr::MIFlag::NoFPExcept))
11602     MIB->setFlag(MachineInstr::MIFlag::NoFPExcept);
11603 
11604   // Insert branch.
11605   BuildMI(MBB, DL, TII.get(RISCV::BEQ))
11606       .addReg(CmpReg)
11607       .addReg(RISCV::X0)
11608       .addMBB(DoneMBB);
11609 
11610   CvtMBB->addSuccessor(DoneMBB);
11611 
11612   // Convert to integer.
11613   Register F2IReg = MRI.createVirtualRegister(&RISCV::GPRRegClass);
11614   MIB = BuildMI(CvtMBB, DL, TII.get(F2IOpc), F2IReg).addReg(SrcReg).addImm(FRM);
11615   if (MI.getFlag(MachineInstr::MIFlag::NoFPExcept))
11616     MIB->setFlag(MachineInstr::MIFlag::NoFPExcept);
11617 
11618   // Convert back to FP.
11619   Register I2FReg = MRI.createVirtualRegister(RC);
11620   MIB = BuildMI(CvtMBB, DL, TII.get(I2FOpc), I2FReg).addReg(F2IReg).addImm(FRM);
11621   if (MI.getFlag(MachineInstr::MIFlag::NoFPExcept))
11622     MIB->setFlag(MachineInstr::MIFlag::NoFPExcept);
11623 
11624   // Restore the sign bit.
11625   Register CvtReg = MRI.createVirtualRegister(RC);
11626   BuildMI(CvtMBB, DL, TII.get(FSGNJOpc), CvtReg).addReg(I2FReg).addReg(SrcReg);
11627 
11628   // Merge the results.
11629   BuildMI(*DoneMBB, DoneMBB->begin(), DL, TII.get(RISCV::PHI), DstReg)
11630       .addReg(SrcReg)
11631       .addMBB(MBB)
11632       .addReg(CvtReg)
11633       .addMBB(CvtMBB);
11634 
11635   MI.eraseFromParent();
11636   return DoneMBB;
11637 }
11638 
11639 MachineBasicBlock *
EmitInstrWithCustomInserter(MachineInstr & MI,MachineBasicBlock * BB) const11640 RISCVTargetLowering::EmitInstrWithCustomInserter(MachineInstr &MI,
11641                                                  MachineBasicBlock *BB) const {
11642   switch (MI.getOpcode()) {
11643   default:
11644     llvm_unreachable("Unexpected instr type to insert");
11645   case RISCV::ReadCycleWide:
11646     assert(!Subtarget.is64Bit() &&
11647            "ReadCycleWrite is only to be used on riscv32");
11648     return emitReadCycleWidePseudo(MI, BB);
11649   case RISCV::Select_GPR_Using_CC_GPR:
11650   case RISCV::Select_FPR16_Using_CC_GPR:
11651   case RISCV::Select_FPR32_Using_CC_GPR:
11652   case RISCV::Select_FPR64_Using_CC_GPR:
11653     return emitSelectPseudo(MI, BB, Subtarget);
11654   case RISCV::BuildPairF64Pseudo:
11655     return emitBuildPairF64Pseudo(MI, BB);
11656   case RISCV::SplitF64Pseudo:
11657     return emitSplitF64Pseudo(MI, BB);
11658   case RISCV::PseudoQuietFLE_H:
11659     return emitQuietFCMP(MI, BB, RISCV::FLE_H, RISCV::FEQ_H, Subtarget);
11660   case RISCV::PseudoQuietFLT_H:
11661     return emitQuietFCMP(MI, BB, RISCV::FLT_H, RISCV::FEQ_H, Subtarget);
11662   case RISCV::PseudoQuietFLE_S:
11663     return emitQuietFCMP(MI, BB, RISCV::FLE_S, RISCV::FEQ_S, Subtarget);
11664   case RISCV::PseudoQuietFLT_S:
11665     return emitQuietFCMP(MI, BB, RISCV::FLT_S, RISCV::FEQ_S, Subtarget);
11666   case RISCV::PseudoQuietFLE_D:
11667     return emitQuietFCMP(MI, BB, RISCV::FLE_D, RISCV::FEQ_D, Subtarget);
11668   case RISCV::PseudoQuietFLT_D:
11669     return emitQuietFCMP(MI, BB, RISCV::FLT_D, RISCV::FEQ_D, Subtarget);
11670 
11671     // =========================================================================
11672     // VFCVT
11673     // =========================================================================
11674 
11675   case RISCV::PseudoVFCVT_RM_X_F_V_M1_MASK:
11676     return emitVFCVT_RM_MASK(MI, BB, RISCV::PseudoVFCVT_X_F_V_M1_MASK);
11677   case RISCV::PseudoVFCVT_RM_X_F_V_M2_MASK:
11678     return emitVFCVT_RM_MASK(MI, BB, RISCV::PseudoVFCVT_X_F_V_M2_MASK);
11679   case RISCV::PseudoVFCVT_RM_X_F_V_M4_MASK:
11680     return emitVFCVT_RM_MASK(MI, BB, RISCV::PseudoVFCVT_X_F_V_M4_MASK);
11681   case RISCV::PseudoVFCVT_RM_X_F_V_M8_MASK:
11682     return emitVFCVT_RM_MASK(MI, BB, RISCV::PseudoVFCVT_X_F_V_M8_MASK);
11683   case RISCV::PseudoVFCVT_RM_X_F_V_MF2_MASK:
11684     return emitVFCVT_RM_MASK(MI, BB, RISCV::PseudoVFCVT_X_F_V_MF2_MASK);
11685   case RISCV::PseudoVFCVT_RM_X_F_V_MF4_MASK:
11686     return emitVFCVT_RM_MASK(MI, BB, RISCV::PseudoVFCVT_X_F_V_MF4_MASK);
11687 
11688   case RISCV::PseudoVFCVT_RM_XU_F_V_M1_MASK:
11689     return emitVFCVT_RM_MASK(MI, BB, RISCV::PseudoVFCVT_XU_F_V_M1_MASK);
11690   case RISCV::PseudoVFCVT_RM_XU_F_V_M2_MASK:
11691     return emitVFCVT_RM_MASK(MI, BB, RISCV::PseudoVFCVT_XU_F_V_M2_MASK);
11692   case RISCV::PseudoVFCVT_RM_XU_F_V_M4_MASK:
11693     return emitVFCVT_RM_MASK(MI, BB, RISCV::PseudoVFCVT_XU_F_V_M4_MASK);
11694   case RISCV::PseudoVFCVT_RM_XU_F_V_M8_MASK:
11695     return emitVFCVT_RM_MASK(MI, BB, RISCV::PseudoVFCVT_XU_F_V_M8_MASK);
11696   case RISCV::PseudoVFCVT_RM_XU_F_V_MF2_MASK:
11697     return emitVFCVT_RM_MASK(MI, BB, RISCV::PseudoVFCVT_XU_F_V_MF2_MASK);
11698   case RISCV::PseudoVFCVT_RM_XU_F_V_MF4_MASK:
11699     return emitVFCVT_RM_MASK(MI, BB, RISCV::PseudoVFCVT_XU_F_V_MF4_MASK);
11700 
11701   case RISCV::PseudoVFCVT_RM_F_XU_V_M1_MASK:
11702     return emitVFCVT_RM_MASK(MI, BB, RISCV::PseudoVFCVT_F_XU_V_M1_MASK);
11703   case RISCV::PseudoVFCVT_RM_F_XU_V_M2_MASK:
11704     return emitVFCVT_RM_MASK(MI, BB, RISCV::PseudoVFCVT_F_XU_V_M2_MASK);
11705   case RISCV::PseudoVFCVT_RM_F_XU_V_M4_MASK:
11706     return emitVFCVT_RM_MASK(MI, BB, RISCV::PseudoVFCVT_F_XU_V_M4_MASK);
11707   case RISCV::PseudoVFCVT_RM_F_XU_V_M8_MASK:
11708     return emitVFCVT_RM_MASK(MI, BB, RISCV::PseudoVFCVT_F_XU_V_M8_MASK);
11709   case RISCV::PseudoVFCVT_RM_F_XU_V_MF2_MASK:
11710     return emitVFCVT_RM_MASK(MI, BB, RISCV::PseudoVFCVT_F_XU_V_MF2_MASK);
11711   case RISCV::PseudoVFCVT_RM_F_XU_V_MF4_MASK:
11712     return emitVFCVT_RM_MASK(MI, BB, RISCV::PseudoVFCVT_F_XU_V_MF4_MASK);
11713 
11714   case RISCV::PseudoVFCVT_RM_F_X_V_M1_MASK:
11715     return emitVFCVT_RM_MASK(MI, BB, RISCV::PseudoVFCVT_F_X_V_M1_MASK);
11716   case RISCV::PseudoVFCVT_RM_F_X_V_M2_MASK:
11717     return emitVFCVT_RM_MASK(MI, BB, RISCV::PseudoVFCVT_F_X_V_M2_MASK);
11718   case RISCV::PseudoVFCVT_RM_F_X_V_M4_MASK:
11719     return emitVFCVT_RM_MASK(MI, BB, RISCV::PseudoVFCVT_F_X_V_M4_MASK);
11720   case RISCV::PseudoVFCVT_RM_F_X_V_M8_MASK:
11721     return emitVFCVT_RM_MASK(MI, BB, RISCV::PseudoVFCVT_F_X_V_M8_MASK);
11722   case RISCV::PseudoVFCVT_RM_F_X_V_MF2_MASK:
11723     return emitVFCVT_RM_MASK(MI, BB, RISCV::PseudoVFCVT_F_X_V_MF2_MASK);
11724   case RISCV::PseudoVFCVT_RM_F_X_V_MF4_MASK:
11725     return emitVFCVT_RM_MASK(MI, BB, RISCV::PseudoVFCVT_F_X_V_MF4_MASK);
11726 
11727     // =========================================================================
11728     // VFWCVT
11729     // =========================================================================
11730 
11731   case RISCV::PseudoVFWCVT_RM_XU_F_V_M1_MASK:
11732     return emitVFCVT_RM_MASK(MI, BB, RISCV::PseudoVFWCVT_X_F_V_M1_MASK);
11733   case RISCV::PseudoVFWCVT_RM_XU_F_V_M2_MASK:
11734     return emitVFCVT_RM_MASK(MI, BB, RISCV::PseudoVFWCVT_X_F_V_M2_MASK);
11735   case RISCV::PseudoVFWCVT_RM_XU_F_V_M4_MASK:
11736     return emitVFCVT_RM_MASK(MI, BB, RISCV::PseudoVFWCVT_X_F_V_M4_MASK);
11737   case RISCV::PseudoVFWCVT_RM_XU_F_V_MF2_MASK:
11738     return emitVFCVT_RM_MASK(MI, BB, RISCV::PseudoVFWCVT_X_F_V_MF2_MASK);
11739   case RISCV::PseudoVFWCVT_RM_XU_F_V_MF4_MASK:
11740     return emitVFCVT_RM_MASK(MI, BB, RISCV::PseudoVFWCVT_X_F_V_MF4_MASK);
11741 
11742   case RISCV::PseudoVFWCVT_RM_X_F_V_M1_MASK:
11743     return emitVFCVT_RM_MASK(MI, BB, RISCV::PseudoVFWCVT_X_F_V_M1_MASK);
11744   case RISCV::PseudoVFWCVT_RM_X_F_V_M2_MASK:
11745     return emitVFCVT_RM_MASK(MI, BB, RISCV::PseudoVFWCVT_X_F_V_M2_MASK);
11746   case RISCV::PseudoVFWCVT_RM_X_F_V_M4_MASK:
11747     return emitVFCVT_RM_MASK(MI, BB, RISCV::PseudoVFWCVT_X_F_V_M4_MASK);
11748   case RISCV::PseudoVFWCVT_RM_X_F_V_MF2_MASK:
11749     return emitVFCVT_RM_MASK(MI, BB, RISCV::PseudoVFWCVT_X_F_V_MF2_MASK);
11750   case RISCV::PseudoVFWCVT_RM_X_F_V_MF4_MASK:
11751     return emitVFCVT_RM_MASK(MI, BB, RISCV::PseudoVFWCVT_X_F_V_MF4_MASK);
11752 
11753   case RISCV::PseudoVFWCVT_RM_F_XU_V_M1_MASK:
11754     return emitVFCVT_RM_MASK(MI, BB, RISCV::PseudoVFWCVT_F_XU_V_M1_MASK);
11755   case RISCV::PseudoVFWCVT_RM_F_XU_V_M2_MASK:
11756     return emitVFCVT_RM_MASK(MI, BB, RISCV::PseudoVFWCVT_F_XU_V_M2_MASK);
11757   case RISCV::PseudoVFWCVT_RM_F_XU_V_M4_MASK:
11758     return emitVFCVT_RM_MASK(MI, BB, RISCV::PseudoVFWCVT_F_XU_V_M4_MASK);
11759   case RISCV::PseudoVFWCVT_RM_F_XU_V_MF2_MASK:
11760     return emitVFCVT_RM_MASK(MI, BB, RISCV::PseudoVFWCVT_F_XU_V_MF2_MASK);
11761   case RISCV::PseudoVFWCVT_RM_F_XU_V_MF4_MASK:
11762     return emitVFCVT_RM_MASK(MI, BB, RISCV::PseudoVFWCVT_F_XU_V_MF4_MASK);
11763   case RISCV::PseudoVFWCVT_RM_F_XU_V_MF8_MASK:
11764     return emitVFCVT_RM_MASK(MI, BB, RISCV::PseudoVFWCVT_F_XU_V_MF8_MASK);
11765 
11766   case RISCV::PseudoVFWCVT_RM_F_X_V_M1_MASK:
11767     return emitVFCVT_RM_MASK(MI, BB, RISCV::PseudoVFWCVT_F_XU_V_M1_MASK);
11768   case RISCV::PseudoVFWCVT_RM_F_X_V_M2_MASK:
11769     return emitVFCVT_RM_MASK(MI, BB, RISCV::PseudoVFWCVT_F_XU_V_M2_MASK);
11770   case RISCV::PseudoVFWCVT_RM_F_X_V_M4_MASK:
11771     return emitVFCVT_RM_MASK(MI, BB, RISCV::PseudoVFWCVT_F_XU_V_M4_MASK);
11772   case RISCV::PseudoVFWCVT_RM_F_X_V_MF2_MASK:
11773     return emitVFCVT_RM_MASK(MI, BB, RISCV::PseudoVFWCVT_F_XU_V_MF2_MASK);
11774   case RISCV::PseudoVFWCVT_RM_F_X_V_MF4_MASK:
11775     return emitVFCVT_RM_MASK(MI, BB, RISCV::PseudoVFWCVT_F_XU_V_MF4_MASK);
11776   case RISCV::PseudoVFWCVT_RM_F_X_V_MF8_MASK:
11777     return emitVFCVT_RM_MASK(MI, BB, RISCV::PseudoVFWCVT_F_XU_V_MF8_MASK);
11778 
11779     // =========================================================================
11780     // VFNCVT
11781     // =========================================================================
11782 
11783   case RISCV::PseudoVFNCVT_RM_XU_F_W_M1_MASK:
11784     return emitVFCVT_RM_MASK(MI, BB, RISCV::PseudoVFNCVT_X_F_W_M1_MASK);
11785   case RISCV::PseudoVFNCVT_RM_XU_F_W_M2_MASK:
11786     return emitVFCVT_RM_MASK(MI, BB, RISCV::PseudoVFNCVT_X_F_W_M2_MASK);
11787   case RISCV::PseudoVFNCVT_RM_XU_F_W_M4_MASK:
11788     return emitVFCVT_RM_MASK(MI, BB, RISCV::PseudoVFNCVT_X_F_W_M4_MASK);
11789   case RISCV::PseudoVFNCVT_RM_XU_F_W_MF2_MASK:
11790     return emitVFCVT_RM_MASK(MI, BB, RISCV::PseudoVFNCVT_X_F_W_MF2_MASK);
11791   case RISCV::PseudoVFNCVT_RM_XU_F_W_MF4_MASK:
11792     return emitVFCVT_RM_MASK(MI, BB, RISCV::PseudoVFNCVT_X_F_W_MF4_MASK);
11793   case RISCV::PseudoVFNCVT_RM_XU_F_W_MF8_MASK:
11794     return emitVFCVT_RM_MASK(MI, BB, RISCV::PseudoVFNCVT_XU_F_W_MF8_MASK);
11795 
11796   case RISCV::PseudoVFNCVT_RM_X_F_W_M1_MASK:
11797     return emitVFCVT_RM_MASK(MI, BB, RISCV::PseudoVFNCVT_X_F_W_M1_MASK);
11798   case RISCV::PseudoVFNCVT_RM_X_F_W_M2_MASK:
11799     return emitVFCVT_RM_MASK(MI, BB, RISCV::PseudoVFNCVT_X_F_W_M2_MASK);
11800   case RISCV::PseudoVFNCVT_RM_X_F_W_M4_MASK:
11801     return emitVFCVT_RM_MASK(MI, BB, RISCV::PseudoVFNCVT_X_F_W_M4_MASK);
11802   case RISCV::PseudoVFNCVT_RM_X_F_W_MF2_MASK:
11803     return emitVFCVT_RM_MASK(MI, BB, RISCV::PseudoVFNCVT_X_F_W_MF2_MASK);
11804   case RISCV::PseudoVFNCVT_RM_X_F_W_MF4_MASK:
11805     return emitVFCVT_RM_MASK(MI, BB, RISCV::PseudoVFNCVT_X_F_W_MF4_MASK);
11806   case RISCV::PseudoVFNCVT_RM_X_F_W_MF8_MASK:
11807     return emitVFCVT_RM_MASK(MI, BB, RISCV::PseudoVFNCVT_X_F_W_MF8_MASK);
11808 
11809   case RISCV::PseudoVFNCVT_RM_F_XU_W_M1_MASK:
11810     return emitVFCVT_RM_MASK(MI, BB, RISCV::PseudoVFNCVT_F_XU_W_M1_MASK);
11811   case RISCV::PseudoVFNCVT_RM_F_XU_W_M2_MASK:
11812     return emitVFCVT_RM_MASK(MI, BB, RISCV::PseudoVFNCVT_F_XU_W_M2_MASK);
11813   case RISCV::PseudoVFNCVT_RM_F_XU_W_M4_MASK:
11814     return emitVFCVT_RM_MASK(MI, BB, RISCV::PseudoVFNCVT_F_XU_W_M4_MASK);
11815   case RISCV::PseudoVFNCVT_RM_F_XU_W_MF2_MASK:
11816     return emitVFCVT_RM_MASK(MI, BB, RISCV::PseudoVFNCVT_F_XU_W_MF2_MASK);
11817   case RISCV::PseudoVFNCVT_RM_F_XU_W_MF4_MASK:
11818     return emitVFCVT_RM_MASK(MI, BB, RISCV::PseudoVFNCVT_F_XU_W_MF4_MASK);
11819 
11820   case RISCV::PseudoVFNCVT_RM_F_X_W_M1_MASK:
11821     return emitVFCVT_RM_MASK(MI, BB, RISCV::PseudoVFNCVT_F_XU_W_M1_MASK);
11822   case RISCV::PseudoVFNCVT_RM_F_X_W_M2_MASK:
11823     return emitVFCVT_RM_MASK(MI, BB, RISCV::PseudoVFNCVT_F_XU_W_M2_MASK);
11824   case RISCV::PseudoVFNCVT_RM_F_X_W_M4_MASK:
11825     return emitVFCVT_RM_MASK(MI, BB, RISCV::PseudoVFNCVT_F_XU_W_M4_MASK);
11826   case RISCV::PseudoVFNCVT_RM_F_X_W_MF2_MASK:
11827     return emitVFCVT_RM_MASK(MI, BB, RISCV::PseudoVFNCVT_F_XU_W_MF2_MASK);
11828   case RISCV::PseudoVFNCVT_RM_F_X_W_MF4_MASK:
11829     return emitVFCVT_RM_MASK(MI, BB, RISCV::PseudoVFNCVT_F_XU_W_MF4_MASK);
11830 
11831   case RISCV::PseudoVFROUND_NOEXCEPT_V_M1_MASK:
11832     return emitVFROUND_NOEXCEPT_MASK(MI, BB, RISCV::PseudoVFCVT_X_F_V_M1_MASK,
11833                                      RISCV::PseudoVFCVT_F_X_V_M1_MASK);
11834   case RISCV::PseudoVFROUND_NOEXCEPT_V_M2_MASK:
11835     return emitVFROUND_NOEXCEPT_MASK(MI, BB, RISCV::PseudoVFCVT_X_F_V_M2_MASK,
11836                                      RISCV::PseudoVFCVT_F_X_V_M2_MASK);
11837   case RISCV::PseudoVFROUND_NOEXCEPT_V_M4_MASK:
11838     return emitVFROUND_NOEXCEPT_MASK(MI, BB, RISCV::PseudoVFCVT_X_F_V_M4_MASK,
11839                                      RISCV::PseudoVFCVT_F_X_V_M4_MASK);
11840   case RISCV::PseudoVFROUND_NOEXCEPT_V_M8_MASK:
11841     return emitVFROUND_NOEXCEPT_MASK(MI, BB, RISCV::PseudoVFCVT_X_F_V_M8_MASK,
11842                                      RISCV::PseudoVFCVT_F_X_V_M8_MASK);
11843   case RISCV::PseudoVFROUND_NOEXCEPT_V_MF2_MASK:
11844     return emitVFROUND_NOEXCEPT_MASK(MI, BB, RISCV::PseudoVFCVT_X_F_V_MF2_MASK,
11845                                      RISCV::PseudoVFCVT_F_X_V_MF2_MASK);
11846   case RISCV::PseudoVFROUND_NOEXCEPT_V_MF4_MASK:
11847     return emitVFROUND_NOEXCEPT_MASK(MI, BB, RISCV::PseudoVFCVT_X_F_V_MF4_MASK,
11848                                      RISCV::PseudoVFCVT_F_X_V_MF4_MASK);
11849   case RISCV::PseudoFROUND_H:
11850   case RISCV::PseudoFROUND_S:
11851   case RISCV::PseudoFROUND_D:
11852     return emitFROUND(MI, BB, Subtarget);
11853   }
11854 }
11855 
AdjustInstrPostInstrSelection(MachineInstr & MI,SDNode * Node) const11856 void RISCVTargetLowering::AdjustInstrPostInstrSelection(MachineInstr &MI,
11857                                                         SDNode *Node) const {
11858   // Add FRM dependency to any instructions with dynamic rounding mode.
11859   unsigned Opc = MI.getOpcode();
11860   auto Idx = RISCV::getNamedOperandIdx(Opc, RISCV::OpName::frm);
11861   if (Idx < 0)
11862     return;
11863   if (MI.getOperand(Idx).getImm() != RISCVFPRndMode::DYN)
11864     return;
11865   // If the instruction already reads FRM, don't add another read.
11866   if (MI.readsRegister(RISCV::FRM))
11867     return;
11868   MI.addOperand(
11869       MachineOperand::CreateReg(RISCV::FRM, /*isDef*/ false, /*isImp*/ true));
11870 }
11871 
11872 // Calling Convention Implementation.
11873 // The expectations for frontend ABI lowering vary from target to target.
11874 // Ideally, an LLVM frontend would be able to avoid worrying about many ABI
11875 // details, but this is a longer term goal. For now, we simply try to keep the
11876 // role of the frontend as simple and well-defined as possible. The rules can
11877 // be summarised as:
11878 // * Never split up large scalar arguments. We handle them here.
11879 // * If a hardfloat calling convention is being used, and the struct may be
11880 // passed in a pair of registers (fp+fp, int+fp), and both registers are
11881 // available, then pass as two separate arguments. If either the GPRs or FPRs
11882 // are exhausted, then pass according to the rule below.
11883 // * If a struct could never be passed in registers or directly in a stack
11884 // slot (as it is larger than 2*XLEN and the floating point rules don't
11885 // apply), then pass it using a pointer with the byval attribute.
11886 // * If a struct is less than 2*XLEN, then coerce to either a two-element
11887 // word-sized array or a 2*XLEN scalar (depending on alignment).
11888 // * The frontend can determine whether a struct is returned by reference or
11889 // not based on its size and fields. If it will be returned by reference, the
11890 // frontend must modify the prototype so a pointer with the sret annotation is
11891 // passed as the first argument. This is not necessary for large scalar
11892 // returns.
11893 // * Struct return values and varargs should be coerced to structs containing
11894 // register-size fields in the same situations they would be for fixed
11895 // arguments.
11896 
11897 static const MCPhysReg ArgGPRs[] = {
11898   RISCV::X10, RISCV::X11, RISCV::X12, RISCV::X13,
11899   RISCV::X14, RISCV::X15, RISCV::X16, RISCV::X17
11900 };
11901 static const MCPhysReg ArgFPR16s[] = {
11902   RISCV::F10_H, RISCV::F11_H, RISCV::F12_H, RISCV::F13_H,
11903   RISCV::F14_H, RISCV::F15_H, RISCV::F16_H, RISCV::F17_H
11904 };
11905 static const MCPhysReg ArgFPR32s[] = {
11906   RISCV::F10_F, RISCV::F11_F, RISCV::F12_F, RISCV::F13_F,
11907   RISCV::F14_F, RISCV::F15_F, RISCV::F16_F, RISCV::F17_F
11908 };
11909 static const MCPhysReg ArgFPR64s[] = {
11910   RISCV::F10_D, RISCV::F11_D, RISCV::F12_D, RISCV::F13_D,
11911   RISCV::F14_D, RISCV::F15_D, RISCV::F16_D, RISCV::F17_D
11912 };
11913 // This is an interim calling convention and it may be changed in the future.
11914 static const MCPhysReg ArgVRs[] = {
11915     RISCV::V8,  RISCV::V9,  RISCV::V10, RISCV::V11, RISCV::V12, RISCV::V13,
11916     RISCV::V14, RISCV::V15, RISCV::V16, RISCV::V17, RISCV::V18, RISCV::V19,
11917     RISCV::V20, RISCV::V21, RISCV::V22, RISCV::V23};
11918 static const MCPhysReg ArgVRM2s[] = {RISCV::V8M2,  RISCV::V10M2, RISCV::V12M2,
11919                                      RISCV::V14M2, RISCV::V16M2, RISCV::V18M2,
11920                                      RISCV::V20M2, RISCV::V22M2};
11921 static const MCPhysReg ArgVRM4s[] = {RISCV::V8M4, RISCV::V12M4, RISCV::V16M4,
11922                                      RISCV::V20M4};
11923 static const MCPhysReg ArgVRM8s[] = {RISCV::V8M8, RISCV::V16M8};
11924 
11925 // Pass a 2*XLEN argument that has been split into two XLEN values through
11926 // registers or the stack as necessary.
CC_RISCVAssign2XLen(unsigned XLen,CCState & State,CCValAssign VA1,ISD::ArgFlagsTy ArgFlags1,unsigned ValNo2,MVT ValVT2,MVT LocVT2,ISD::ArgFlagsTy ArgFlags2)11927 static bool CC_RISCVAssign2XLen(unsigned XLen, CCState &State, CCValAssign VA1,
11928                                 ISD::ArgFlagsTy ArgFlags1, unsigned ValNo2,
11929                                 MVT ValVT2, MVT LocVT2,
11930                                 ISD::ArgFlagsTy ArgFlags2) {
11931   unsigned XLenInBytes = XLen / 8;
11932   if (Register Reg = State.AllocateReg(ArgGPRs)) {
11933     // At least one half can be passed via register.
11934     State.addLoc(CCValAssign::getReg(VA1.getValNo(), VA1.getValVT(), Reg,
11935                                      VA1.getLocVT(), CCValAssign::Full));
11936   } else {
11937     // Both halves must be passed on the stack, with proper alignment.
11938     Align StackAlign =
11939         std::max(Align(XLenInBytes), ArgFlags1.getNonZeroOrigAlign());
11940     State.addLoc(
11941         CCValAssign::getMem(VA1.getValNo(), VA1.getValVT(),
11942                             State.AllocateStack(XLenInBytes, StackAlign),
11943                             VA1.getLocVT(), CCValAssign::Full));
11944     State.addLoc(CCValAssign::getMem(
11945         ValNo2, ValVT2, State.AllocateStack(XLenInBytes, Align(XLenInBytes)),
11946         LocVT2, CCValAssign::Full));
11947     return false;
11948   }
11949 
11950   if (Register Reg = State.AllocateReg(ArgGPRs)) {
11951     // The second half can also be passed via register.
11952     State.addLoc(
11953         CCValAssign::getReg(ValNo2, ValVT2, Reg, LocVT2, CCValAssign::Full));
11954   } else {
11955     // The second half is passed via the stack, without additional alignment.
11956     State.addLoc(CCValAssign::getMem(
11957         ValNo2, ValVT2, State.AllocateStack(XLenInBytes, Align(XLenInBytes)),
11958         LocVT2, CCValAssign::Full));
11959   }
11960 
11961   return false;
11962 }
11963 
allocateRVVReg(MVT ValVT,unsigned ValNo,std::optional<unsigned> FirstMaskArgument,CCState & State,const RISCVTargetLowering & TLI)11964 static unsigned allocateRVVReg(MVT ValVT, unsigned ValNo,
11965                                std::optional<unsigned> FirstMaskArgument,
11966                                CCState &State, const RISCVTargetLowering &TLI) {
11967   const TargetRegisterClass *RC = TLI.getRegClassFor(ValVT);
11968   if (RC == &RISCV::VRRegClass) {
11969     // Assign the first mask argument to V0.
11970     // This is an interim calling convention and it may be changed in the
11971     // future.
11972     if (FirstMaskArgument && ValNo == *FirstMaskArgument)
11973       return State.AllocateReg(RISCV::V0);
11974     return State.AllocateReg(ArgVRs);
11975   }
11976   if (RC == &RISCV::VRM2RegClass)
11977     return State.AllocateReg(ArgVRM2s);
11978   if (RC == &RISCV::VRM4RegClass)
11979     return State.AllocateReg(ArgVRM4s);
11980   if (RC == &RISCV::VRM8RegClass)
11981     return State.AllocateReg(ArgVRM8s);
11982   llvm_unreachable("Unhandled register class for ValueType");
11983 }
11984 
11985 // Implements the RISC-V calling convention. Returns true upon failure.
CC_RISCV(const DataLayout & DL,RISCVABI::ABI ABI,unsigned ValNo,MVT ValVT,MVT LocVT,CCValAssign::LocInfo LocInfo,ISD::ArgFlagsTy ArgFlags,CCState & State,bool IsFixed,bool IsRet,Type * OrigTy,const RISCVTargetLowering & TLI,std::optional<unsigned> FirstMaskArgument)11986 static bool CC_RISCV(const DataLayout &DL, RISCVABI::ABI ABI, unsigned ValNo,
11987                      MVT ValVT, MVT LocVT, CCValAssign::LocInfo LocInfo,
11988                      ISD::ArgFlagsTy ArgFlags, CCState &State, bool IsFixed,
11989                      bool IsRet, Type *OrigTy, const RISCVTargetLowering &TLI,
11990                      std::optional<unsigned> FirstMaskArgument) {
11991   unsigned XLen = DL.getLargestLegalIntTypeSizeInBits();
11992   assert(XLen == 32 || XLen == 64);
11993   MVT XLenVT = XLen == 32 ? MVT::i32 : MVT::i64;
11994 
11995   // Static chain parameter must not be passed in normal argument registers,
11996   // so we assign t2 for it as done in GCC's __builtin_call_with_static_chain
11997   if (ArgFlags.isNest()) {
11998     if (unsigned Reg = State.AllocateReg(RISCV::X7)) {
11999       State.addLoc(CCValAssign::getReg(ValNo, ValVT, Reg, LocVT, LocInfo));
12000       return false;
12001     }
12002   }
12003 
12004   // Any return value split in to more than two values can't be returned
12005   // directly. Vectors are returned via the available vector registers.
12006   if (!LocVT.isVector() && IsRet && ValNo > 1)
12007     return true;
12008 
12009   // UseGPRForF16_F32 if targeting one of the soft-float ABIs, if passing a
12010   // variadic argument, or if no F16/F32 argument registers are available.
12011   bool UseGPRForF16_F32 = true;
12012   // UseGPRForF64 if targeting soft-float ABIs or an FLEN=32 ABI, if passing a
12013   // variadic argument, or if no F64 argument registers are available.
12014   bool UseGPRForF64 = true;
12015 
12016   switch (ABI) {
12017   default:
12018     llvm_unreachable("Unexpected ABI");
12019   case RISCVABI::ABI_ILP32:
12020   case RISCVABI::ABI_LP64:
12021     break;
12022   case RISCVABI::ABI_ILP32F:
12023   case RISCVABI::ABI_LP64F:
12024     UseGPRForF16_F32 = !IsFixed;
12025     break;
12026   case RISCVABI::ABI_ILP32D:
12027   case RISCVABI::ABI_LP64D:
12028     UseGPRForF16_F32 = !IsFixed;
12029     UseGPRForF64 = !IsFixed;
12030     break;
12031   }
12032 
12033   // FPR16, FPR32, and FPR64 alias each other.
12034   if (State.getFirstUnallocated(ArgFPR32s) == std::size(ArgFPR32s)) {
12035     UseGPRForF16_F32 = true;
12036     UseGPRForF64 = true;
12037   }
12038 
12039   // From this point on, rely on UseGPRForF16_F32, UseGPRForF64 and
12040   // similar local variables rather than directly checking against the target
12041   // ABI.
12042 
12043   if (UseGPRForF16_F32 && (ValVT == MVT::f16 || ValVT == MVT::f32)) {
12044     LocVT = XLenVT;
12045     LocInfo = CCValAssign::BCvt;
12046   } else if (UseGPRForF64 && XLen == 64 && ValVT == MVT::f64) {
12047     LocVT = MVT::i64;
12048     LocInfo = CCValAssign::BCvt;
12049   }
12050 
12051   // If this is a variadic argument, the RISC-V calling convention requires
12052   // that it is assigned an 'even' or 'aligned' register if it has 8-byte
12053   // alignment (RV32) or 16-byte alignment (RV64). An aligned register should
12054   // be used regardless of whether the original argument was split during
12055   // legalisation or not. The argument will not be passed by registers if the
12056   // original type is larger than 2*XLEN, so the register alignment rule does
12057   // not apply.
12058   unsigned TwoXLenInBytes = (2 * XLen) / 8;
12059   if (!IsFixed && ArgFlags.getNonZeroOrigAlign() == TwoXLenInBytes &&
12060       DL.getTypeAllocSize(OrigTy) == TwoXLenInBytes) {
12061     unsigned RegIdx = State.getFirstUnallocated(ArgGPRs);
12062     // Skip 'odd' register if necessary.
12063     if (RegIdx != std::size(ArgGPRs) && RegIdx % 2 == 1)
12064       State.AllocateReg(ArgGPRs);
12065   }
12066 
12067   SmallVectorImpl<CCValAssign> &PendingLocs = State.getPendingLocs();
12068   SmallVectorImpl<ISD::ArgFlagsTy> &PendingArgFlags =
12069       State.getPendingArgFlags();
12070 
12071   assert(PendingLocs.size() == PendingArgFlags.size() &&
12072          "PendingLocs and PendingArgFlags out of sync");
12073 
12074   // Handle passing f64 on RV32D with a soft float ABI or when floating point
12075   // registers are exhausted.
12076   if (UseGPRForF64 && XLen == 32 && ValVT == MVT::f64) {
12077     assert(!ArgFlags.isSplit() && PendingLocs.empty() &&
12078            "Can't lower f64 if it is split");
12079     // Depending on available argument GPRS, f64 may be passed in a pair of
12080     // GPRs, split between a GPR and the stack, or passed completely on the
12081     // stack. LowerCall/LowerFormalArguments/LowerReturn must recognise these
12082     // cases.
12083     Register Reg = State.AllocateReg(ArgGPRs);
12084     LocVT = MVT::i32;
12085     if (!Reg) {
12086       unsigned StackOffset = State.AllocateStack(8, Align(8));
12087       State.addLoc(
12088           CCValAssign::getMem(ValNo, ValVT, StackOffset, LocVT, LocInfo));
12089       return false;
12090     }
12091     if (!State.AllocateReg(ArgGPRs))
12092       State.AllocateStack(4, Align(4));
12093     State.addLoc(CCValAssign::getReg(ValNo, ValVT, Reg, LocVT, LocInfo));
12094     return false;
12095   }
12096 
12097   // Fixed-length vectors are located in the corresponding scalable-vector
12098   // container types.
12099   if (ValVT.isFixedLengthVector())
12100     LocVT = TLI.getContainerForFixedLengthVector(LocVT);
12101 
12102   // Split arguments might be passed indirectly, so keep track of the pending
12103   // values. Split vectors are passed via a mix of registers and indirectly, so
12104   // treat them as we would any other argument.
12105   if (ValVT.isScalarInteger() && (ArgFlags.isSplit() || !PendingLocs.empty())) {
12106     LocVT = XLenVT;
12107     LocInfo = CCValAssign::Indirect;
12108     PendingLocs.push_back(
12109         CCValAssign::getPending(ValNo, ValVT, LocVT, LocInfo));
12110     PendingArgFlags.push_back(ArgFlags);
12111     if (!ArgFlags.isSplitEnd()) {
12112       return false;
12113     }
12114   }
12115 
12116   // If the split argument only had two elements, it should be passed directly
12117   // in registers or on the stack.
12118   if (ValVT.isScalarInteger() && ArgFlags.isSplitEnd() &&
12119       PendingLocs.size() <= 2) {
12120     assert(PendingLocs.size() == 2 && "Unexpected PendingLocs.size()");
12121     // Apply the normal calling convention rules to the first half of the
12122     // split argument.
12123     CCValAssign VA = PendingLocs[0];
12124     ISD::ArgFlagsTy AF = PendingArgFlags[0];
12125     PendingLocs.clear();
12126     PendingArgFlags.clear();
12127     return CC_RISCVAssign2XLen(XLen, State, VA, AF, ValNo, ValVT, LocVT,
12128                                ArgFlags);
12129   }
12130 
12131   // Allocate to a register if possible, or else a stack slot.
12132   Register Reg;
12133   unsigned StoreSizeBytes = XLen / 8;
12134   Align StackAlign = Align(XLen / 8);
12135 
12136   if (ValVT == MVT::f16 && !UseGPRForF16_F32)
12137     Reg = State.AllocateReg(ArgFPR16s);
12138   else if (ValVT == MVT::f32 && !UseGPRForF16_F32)
12139     Reg = State.AllocateReg(ArgFPR32s);
12140   else if (ValVT == MVT::f64 && !UseGPRForF64)
12141     Reg = State.AllocateReg(ArgFPR64s);
12142   else if (ValVT.isVector()) {
12143     Reg = allocateRVVReg(ValVT, ValNo, FirstMaskArgument, State, TLI);
12144     if (!Reg) {
12145       // For return values, the vector must be passed fully via registers or
12146       // via the stack.
12147       // FIXME: The proposed vector ABI only mandates v8-v15 for return values,
12148       // but we're using all of them.
12149       if (IsRet)
12150         return true;
12151       // Try using a GPR to pass the address
12152       if ((Reg = State.AllocateReg(ArgGPRs))) {
12153         LocVT = XLenVT;
12154         LocInfo = CCValAssign::Indirect;
12155       } else if (ValVT.isScalableVector()) {
12156         LocVT = XLenVT;
12157         LocInfo = CCValAssign::Indirect;
12158       } else {
12159         // Pass fixed-length vectors on the stack.
12160         LocVT = ValVT;
12161         StoreSizeBytes = ValVT.getStoreSize();
12162         // Align vectors to their element sizes, being careful for vXi1
12163         // vectors.
12164         StackAlign = MaybeAlign(ValVT.getScalarSizeInBits() / 8).valueOrOne();
12165       }
12166     }
12167   } else {
12168     Reg = State.AllocateReg(ArgGPRs);
12169   }
12170 
12171   unsigned StackOffset =
12172       Reg ? 0 : State.AllocateStack(StoreSizeBytes, StackAlign);
12173 
12174   // If we reach this point and PendingLocs is non-empty, we must be at the
12175   // end of a split argument that must be passed indirectly.
12176   if (!PendingLocs.empty()) {
12177     assert(ArgFlags.isSplitEnd() && "Expected ArgFlags.isSplitEnd()");
12178     assert(PendingLocs.size() > 2 && "Unexpected PendingLocs.size()");
12179 
12180     for (auto &It : PendingLocs) {
12181       if (Reg)
12182         It.convertToReg(Reg);
12183       else
12184         It.convertToMem(StackOffset);
12185       State.addLoc(It);
12186     }
12187     PendingLocs.clear();
12188     PendingArgFlags.clear();
12189     return false;
12190   }
12191 
12192   assert((!UseGPRForF16_F32 || !UseGPRForF64 || LocVT == XLenVT ||
12193           (TLI.getSubtarget().hasVInstructions() && ValVT.isVector())) &&
12194          "Expected an XLenVT or vector types at this stage");
12195 
12196   if (Reg) {
12197     State.addLoc(CCValAssign::getReg(ValNo, ValVT, Reg, LocVT, LocInfo));
12198     return false;
12199   }
12200 
12201   // When a floating-point value is passed on the stack, no bit-conversion is
12202   // needed.
12203   if (ValVT.isFloatingPoint()) {
12204     LocVT = ValVT;
12205     LocInfo = CCValAssign::Full;
12206   }
12207   State.addLoc(CCValAssign::getMem(ValNo, ValVT, StackOffset, LocVT, LocInfo));
12208   return false;
12209 }
12210 
12211 template <typename ArgTy>
preAssignMask(const ArgTy & Args)12212 static std::optional<unsigned> preAssignMask(const ArgTy &Args) {
12213   for (const auto &ArgIdx : enumerate(Args)) {
12214     MVT ArgVT = ArgIdx.value().VT;
12215     if (ArgVT.isVector() && ArgVT.getVectorElementType() == MVT::i1)
12216       return ArgIdx.index();
12217   }
12218   return std::nullopt;
12219 }
12220 
analyzeInputArgs(MachineFunction & MF,CCState & CCInfo,const SmallVectorImpl<ISD::InputArg> & Ins,bool IsRet,RISCVCCAssignFn Fn) const12221 void RISCVTargetLowering::analyzeInputArgs(
12222     MachineFunction &MF, CCState &CCInfo,
12223     const SmallVectorImpl<ISD::InputArg> &Ins, bool IsRet,
12224     RISCVCCAssignFn Fn) const {
12225   unsigned NumArgs = Ins.size();
12226   FunctionType *FType = MF.getFunction().getFunctionType();
12227 
12228   std::optional<unsigned> FirstMaskArgument;
12229   if (Subtarget.hasVInstructions())
12230     FirstMaskArgument = preAssignMask(Ins);
12231 
12232   for (unsigned i = 0; i != NumArgs; ++i) {
12233     MVT ArgVT = Ins[i].VT;
12234     ISD::ArgFlagsTy ArgFlags = Ins[i].Flags;
12235 
12236     Type *ArgTy = nullptr;
12237     if (IsRet)
12238       ArgTy = FType->getReturnType();
12239     else if (Ins[i].isOrigArg())
12240       ArgTy = FType->getParamType(Ins[i].getOrigArgIndex());
12241 
12242     RISCVABI::ABI ABI = MF.getSubtarget<RISCVSubtarget>().getTargetABI();
12243     if (Fn(MF.getDataLayout(), ABI, i, ArgVT, ArgVT, CCValAssign::Full,
12244            ArgFlags, CCInfo, /*IsFixed=*/true, IsRet, ArgTy, *this,
12245            FirstMaskArgument)) {
12246       LLVM_DEBUG(dbgs() << "InputArg #" << i << " has unhandled type "
12247                         << EVT(ArgVT).getEVTString() << '\n');
12248       llvm_unreachable(nullptr);
12249     }
12250   }
12251 }
12252 
analyzeOutputArgs(MachineFunction & MF,CCState & CCInfo,const SmallVectorImpl<ISD::OutputArg> & Outs,bool IsRet,CallLoweringInfo * CLI,RISCVCCAssignFn Fn) const12253 void RISCVTargetLowering::analyzeOutputArgs(
12254     MachineFunction &MF, CCState &CCInfo,
12255     const SmallVectorImpl<ISD::OutputArg> &Outs, bool IsRet,
12256     CallLoweringInfo *CLI, RISCVCCAssignFn Fn) const {
12257   unsigned NumArgs = Outs.size();
12258 
12259   std::optional<unsigned> FirstMaskArgument;
12260   if (Subtarget.hasVInstructions())
12261     FirstMaskArgument = preAssignMask(Outs);
12262 
12263   for (unsigned i = 0; i != NumArgs; i++) {
12264     MVT ArgVT = Outs[i].VT;
12265     ISD::ArgFlagsTy ArgFlags = Outs[i].Flags;
12266     Type *OrigTy = CLI ? CLI->getArgs()[Outs[i].OrigArgIndex].Ty : nullptr;
12267 
12268     RISCVABI::ABI ABI = MF.getSubtarget<RISCVSubtarget>().getTargetABI();
12269     if (Fn(MF.getDataLayout(), ABI, i, ArgVT, ArgVT, CCValAssign::Full,
12270            ArgFlags, CCInfo, Outs[i].IsFixed, IsRet, OrigTy, *this,
12271            FirstMaskArgument)) {
12272       LLVM_DEBUG(dbgs() << "OutputArg #" << i << " has unhandled type "
12273                         << EVT(ArgVT).getEVTString() << "\n");
12274       llvm_unreachable(nullptr);
12275     }
12276   }
12277 }
12278 
12279 // Convert Val to a ValVT. Should not be called for CCValAssign::Indirect
12280 // values.
convertLocVTToValVT(SelectionDAG & DAG,SDValue Val,const CCValAssign & VA,const SDLoc & DL,const RISCVSubtarget & Subtarget)12281 static SDValue convertLocVTToValVT(SelectionDAG &DAG, SDValue Val,
12282                                    const CCValAssign &VA, const SDLoc &DL,
12283                                    const RISCVSubtarget &Subtarget) {
12284   switch (VA.getLocInfo()) {
12285   default:
12286     llvm_unreachable("Unexpected CCValAssign::LocInfo");
12287   case CCValAssign::Full:
12288     if (VA.getValVT().isFixedLengthVector() && VA.getLocVT().isScalableVector())
12289       Val = convertFromScalableVector(VA.getValVT(), Val, DAG, Subtarget);
12290     break;
12291   case CCValAssign::BCvt:
12292     if (VA.getLocVT().isInteger() && VA.getValVT() == MVT::f16)
12293       Val = DAG.getNode(RISCVISD::FMV_H_X, DL, MVT::f16, Val);
12294     else if (VA.getLocVT() == MVT::i64 && VA.getValVT() == MVT::f32)
12295       Val = DAG.getNode(RISCVISD::FMV_W_X_RV64, DL, MVT::f32, Val);
12296     else
12297       Val = DAG.getNode(ISD::BITCAST, DL, VA.getValVT(), Val);
12298     break;
12299   }
12300   return Val;
12301 }
12302 
12303 // The caller is responsible for loading the full value if the argument is
12304 // passed with CCValAssign::Indirect.
unpackFromRegLoc(SelectionDAG & DAG,SDValue Chain,const CCValAssign & VA,const SDLoc & DL,const ISD::InputArg & In,const RISCVTargetLowering & TLI)12305 static SDValue unpackFromRegLoc(SelectionDAG &DAG, SDValue Chain,
12306                                 const CCValAssign &VA, const SDLoc &DL,
12307                                 const ISD::InputArg &In,
12308                                 const RISCVTargetLowering &TLI) {
12309   MachineFunction &MF = DAG.getMachineFunction();
12310   MachineRegisterInfo &RegInfo = MF.getRegInfo();
12311   EVT LocVT = VA.getLocVT();
12312   SDValue Val;
12313   const TargetRegisterClass *RC = TLI.getRegClassFor(LocVT.getSimpleVT());
12314   Register VReg = RegInfo.createVirtualRegister(RC);
12315   RegInfo.addLiveIn(VA.getLocReg(), VReg);
12316   Val = DAG.getCopyFromReg(Chain, DL, VReg, LocVT);
12317 
12318   // If input is sign extended from 32 bits, note it for the SExtWRemoval pass.
12319   if (In.isOrigArg()) {
12320     Argument *OrigArg = MF.getFunction().getArg(In.getOrigArgIndex());
12321     if (OrigArg->getType()->isIntegerTy()) {
12322       unsigned BitWidth = OrigArg->getType()->getIntegerBitWidth();
12323       // An input zero extended from i31 can also be considered sign extended.
12324       if ((BitWidth <= 32 && In.Flags.isSExt()) ||
12325           (BitWidth < 32 && In.Flags.isZExt())) {
12326         RISCVMachineFunctionInfo *RVFI = MF.getInfo<RISCVMachineFunctionInfo>();
12327         RVFI->addSExt32Register(VReg);
12328       }
12329     }
12330   }
12331 
12332   if (VA.getLocInfo() == CCValAssign::Indirect)
12333     return Val;
12334 
12335   return convertLocVTToValVT(DAG, Val, VA, DL, TLI.getSubtarget());
12336 }
12337 
convertValVTToLocVT(SelectionDAG & DAG,SDValue Val,const CCValAssign & VA,const SDLoc & DL,const RISCVSubtarget & Subtarget)12338 static SDValue convertValVTToLocVT(SelectionDAG &DAG, SDValue Val,
12339                                    const CCValAssign &VA, const SDLoc &DL,
12340                                    const RISCVSubtarget &Subtarget) {
12341   EVT LocVT = VA.getLocVT();
12342 
12343   switch (VA.getLocInfo()) {
12344   default:
12345     llvm_unreachable("Unexpected CCValAssign::LocInfo");
12346   case CCValAssign::Full:
12347     if (VA.getValVT().isFixedLengthVector() && LocVT.isScalableVector())
12348       Val = convertToScalableVector(LocVT, Val, DAG, Subtarget);
12349     break;
12350   case CCValAssign::BCvt:
12351     if (VA.getLocVT().isInteger() && VA.getValVT() == MVT::f16)
12352       Val = DAG.getNode(RISCVISD::FMV_X_ANYEXTH, DL, VA.getLocVT(), Val);
12353     else if (VA.getLocVT() == MVT::i64 && VA.getValVT() == MVT::f32)
12354       Val = DAG.getNode(RISCVISD::FMV_X_ANYEXTW_RV64, DL, MVT::i64, Val);
12355     else
12356       Val = DAG.getNode(ISD::BITCAST, DL, LocVT, Val);
12357     break;
12358   }
12359   return Val;
12360 }
12361 
12362 // The caller is responsible for loading the full value if the argument is
12363 // passed with CCValAssign::Indirect.
unpackFromMemLoc(SelectionDAG & DAG,SDValue Chain,const CCValAssign & VA,const SDLoc & DL)12364 static SDValue unpackFromMemLoc(SelectionDAG &DAG, SDValue Chain,
12365                                 const CCValAssign &VA, const SDLoc &DL) {
12366   MachineFunction &MF = DAG.getMachineFunction();
12367   MachineFrameInfo &MFI = MF.getFrameInfo();
12368   EVT LocVT = VA.getLocVT();
12369   EVT ValVT = VA.getValVT();
12370   EVT PtrVT = MVT::getIntegerVT(DAG.getDataLayout().getPointerSizeInBits(0));
12371   if (ValVT.isScalableVector()) {
12372     // When the value is a scalable vector, we save the pointer which points to
12373     // the scalable vector value in the stack. The ValVT will be the pointer
12374     // type, instead of the scalable vector type.
12375     ValVT = LocVT;
12376   }
12377   int FI = MFI.CreateFixedObject(ValVT.getStoreSize(), VA.getLocMemOffset(),
12378                                  /*IsImmutable=*/true);
12379   SDValue FIN = DAG.getFrameIndex(FI, PtrVT);
12380   SDValue Val;
12381 
12382   ISD::LoadExtType ExtType;
12383   switch (VA.getLocInfo()) {
12384   default:
12385     llvm_unreachable("Unexpected CCValAssign::LocInfo");
12386   case CCValAssign::Full:
12387   case CCValAssign::Indirect:
12388   case CCValAssign::BCvt:
12389     ExtType = ISD::NON_EXTLOAD;
12390     break;
12391   }
12392   Val = DAG.getExtLoad(
12393       ExtType, DL, LocVT, Chain, FIN,
12394       MachinePointerInfo::getFixedStack(DAG.getMachineFunction(), FI), ValVT);
12395   return Val;
12396 }
12397 
unpackF64OnRV32DSoftABI(SelectionDAG & DAG,SDValue Chain,const CCValAssign & VA,const SDLoc & DL)12398 static SDValue unpackF64OnRV32DSoftABI(SelectionDAG &DAG, SDValue Chain,
12399                                        const CCValAssign &VA, const SDLoc &DL) {
12400   assert(VA.getLocVT() == MVT::i32 && VA.getValVT() == MVT::f64 &&
12401          "Unexpected VA");
12402   MachineFunction &MF = DAG.getMachineFunction();
12403   MachineFrameInfo &MFI = MF.getFrameInfo();
12404   MachineRegisterInfo &RegInfo = MF.getRegInfo();
12405 
12406   if (VA.isMemLoc()) {
12407     // f64 is passed on the stack.
12408     int FI =
12409         MFI.CreateFixedObject(8, VA.getLocMemOffset(), /*IsImmutable=*/true);
12410     SDValue FIN = DAG.getFrameIndex(FI, MVT::i32);
12411     return DAG.getLoad(MVT::f64, DL, Chain, FIN,
12412                        MachinePointerInfo::getFixedStack(MF, FI));
12413   }
12414 
12415   assert(VA.isRegLoc() && "Expected register VA assignment");
12416 
12417   Register LoVReg = RegInfo.createVirtualRegister(&RISCV::GPRRegClass);
12418   RegInfo.addLiveIn(VA.getLocReg(), LoVReg);
12419   SDValue Lo = DAG.getCopyFromReg(Chain, DL, LoVReg, MVT::i32);
12420   SDValue Hi;
12421   if (VA.getLocReg() == RISCV::X17) {
12422     // Second half of f64 is passed on the stack.
12423     int FI = MFI.CreateFixedObject(4, 0, /*IsImmutable=*/true);
12424     SDValue FIN = DAG.getFrameIndex(FI, MVT::i32);
12425     Hi = DAG.getLoad(MVT::i32, DL, Chain, FIN,
12426                      MachinePointerInfo::getFixedStack(MF, FI));
12427   } else {
12428     // Second half of f64 is passed in another GPR.
12429     Register HiVReg = RegInfo.createVirtualRegister(&RISCV::GPRRegClass);
12430     RegInfo.addLiveIn(VA.getLocReg() + 1, HiVReg);
12431     Hi = DAG.getCopyFromReg(Chain, DL, HiVReg, MVT::i32);
12432   }
12433   return DAG.getNode(RISCVISD::BuildPairF64, DL, MVT::f64, Lo, Hi);
12434 }
12435 
12436 // FastCC has less than 1% performance improvement for some particular
12437 // benchmark. But theoretically, it may has benenfit for some cases.
CC_RISCV_FastCC(const DataLayout & DL,RISCVABI::ABI ABI,unsigned ValNo,MVT ValVT,MVT LocVT,CCValAssign::LocInfo LocInfo,ISD::ArgFlagsTy ArgFlags,CCState & State,bool IsFixed,bool IsRet,Type * OrigTy,const RISCVTargetLowering & TLI,std::optional<unsigned> FirstMaskArgument)12438 static bool CC_RISCV_FastCC(const DataLayout &DL, RISCVABI::ABI ABI,
12439                             unsigned ValNo, MVT ValVT, MVT LocVT,
12440                             CCValAssign::LocInfo LocInfo,
12441                             ISD::ArgFlagsTy ArgFlags, CCState &State,
12442                             bool IsFixed, bool IsRet, Type *OrigTy,
12443                             const RISCVTargetLowering &TLI,
12444                             std::optional<unsigned> FirstMaskArgument) {
12445 
12446   // X5 and X6 might be used for save-restore libcall.
12447   static const MCPhysReg GPRList[] = {
12448       RISCV::X10, RISCV::X11, RISCV::X12, RISCV::X13, RISCV::X14,
12449       RISCV::X15, RISCV::X16, RISCV::X17, RISCV::X7,  RISCV::X28,
12450       RISCV::X29, RISCV::X30, RISCV::X31};
12451 
12452   if (LocVT == MVT::i32 || LocVT == MVT::i64) {
12453     if (unsigned Reg = State.AllocateReg(GPRList)) {
12454       State.addLoc(CCValAssign::getReg(ValNo, ValVT, Reg, LocVT, LocInfo));
12455       return false;
12456     }
12457   }
12458 
12459   if (LocVT == MVT::f16) {
12460     static const MCPhysReg FPR16List[] = {
12461         RISCV::F10_H, RISCV::F11_H, RISCV::F12_H, RISCV::F13_H, RISCV::F14_H,
12462         RISCV::F15_H, RISCV::F16_H, RISCV::F17_H, RISCV::F0_H,  RISCV::F1_H,
12463         RISCV::F2_H,  RISCV::F3_H,  RISCV::F4_H,  RISCV::F5_H,  RISCV::F6_H,
12464         RISCV::F7_H,  RISCV::F28_H, RISCV::F29_H, RISCV::F30_H, RISCV::F31_H};
12465     if (unsigned Reg = State.AllocateReg(FPR16List)) {
12466       State.addLoc(CCValAssign::getReg(ValNo, ValVT, Reg, LocVT, LocInfo));
12467       return false;
12468     }
12469   }
12470 
12471   if (LocVT == MVT::f32) {
12472     static const MCPhysReg FPR32List[] = {
12473         RISCV::F10_F, RISCV::F11_F, RISCV::F12_F, RISCV::F13_F, RISCV::F14_F,
12474         RISCV::F15_F, RISCV::F16_F, RISCV::F17_F, RISCV::F0_F,  RISCV::F1_F,
12475         RISCV::F2_F,  RISCV::F3_F,  RISCV::F4_F,  RISCV::F5_F,  RISCV::F6_F,
12476         RISCV::F7_F,  RISCV::F28_F, RISCV::F29_F, RISCV::F30_F, RISCV::F31_F};
12477     if (unsigned Reg = State.AllocateReg(FPR32List)) {
12478       State.addLoc(CCValAssign::getReg(ValNo, ValVT, Reg, LocVT, LocInfo));
12479       return false;
12480     }
12481   }
12482 
12483   if (LocVT == MVT::f64) {
12484     static const MCPhysReg FPR64List[] = {
12485         RISCV::F10_D, RISCV::F11_D, RISCV::F12_D, RISCV::F13_D, RISCV::F14_D,
12486         RISCV::F15_D, RISCV::F16_D, RISCV::F17_D, RISCV::F0_D,  RISCV::F1_D,
12487         RISCV::F2_D,  RISCV::F3_D,  RISCV::F4_D,  RISCV::F5_D,  RISCV::F6_D,
12488         RISCV::F7_D,  RISCV::F28_D, RISCV::F29_D, RISCV::F30_D, RISCV::F31_D};
12489     if (unsigned Reg = State.AllocateReg(FPR64List)) {
12490       State.addLoc(CCValAssign::getReg(ValNo, ValVT, Reg, LocVT, LocInfo));
12491       return false;
12492     }
12493   }
12494 
12495   if (LocVT == MVT::i32 || LocVT == MVT::f32) {
12496     unsigned Offset4 = State.AllocateStack(4, Align(4));
12497     State.addLoc(CCValAssign::getMem(ValNo, ValVT, Offset4, LocVT, LocInfo));
12498     return false;
12499   }
12500 
12501   if (LocVT == MVT::i64 || LocVT == MVT::f64) {
12502     unsigned Offset5 = State.AllocateStack(8, Align(8));
12503     State.addLoc(CCValAssign::getMem(ValNo, ValVT, Offset5, LocVT, LocInfo));
12504     return false;
12505   }
12506 
12507   if (LocVT.isVector()) {
12508     if (unsigned Reg =
12509             allocateRVVReg(ValVT, ValNo, FirstMaskArgument, State, TLI)) {
12510       // Fixed-length vectors are located in the corresponding scalable-vector
12511       // container types.
12512       if (ValVT.isFixedLengthVector())
12513         LocVT = TLI.getContainerForFixedLengthVector(LocVT);
12514       State.addLoc(CCValAssign::getReg(ValNo, ValVT, Reg, LocVT, LocInfo));
12515     } else {
12516       // Try and pass the address via a "fast" GPR.
12517       if (unsigned GPRReg = State.AllocateReg(GPRList)) {
12518         LocInfo = CCValAssign::Indirect;
12519         LocVT = TLI.getSubtarget().getXLenVT();
12520         State.addLoc(CCValAssign::getReg(ValNo, ValVT, GPRReg, LocVT, LocInfo));
12521       } else if (ValVT.isFixedLengthVector()) {
12522         auto StackAlign =
12523             MaybeAlign(ValVT.getScalarSizeInBits() / 8).valueOrOne();
12524         unsigned StackOffset =
12525             State.AllocateStack(ValVT.getStoreSize(), StackAlign);
12526         State.addLoc(
12527             CCValAssign::getMem(ValNo, ValVT, StackOffset, LocVT, LocInfo));
12528       } else {
12529         // Can't pass scalable vectors on the stack.
12530         return true;
12531       }
12532     }
12533 
12534     return false;
12535   }
12536 
12537   return true; // CC didn't match.
12538 }
12539 
CC_RISCV_GHC(unsigned ValNo,MVT ValVT,MVT LocVT,CCValAssign::LocInfo LocInfo,ISD::ArgFlagsTy ArgFlags,CCState & State)12540 static bool CC_RISCV_GHC(unsigned ValNo, MVT ValVT, MVT LocVT,
12541                          CCValAssign::LocInfo LocInfo,
12542                          ISD::ArgFlagsTy ArgFlags, CCState &State) {
12543 
12544   if (ArgFlags.isNest()) {
12545     report_fatal_error(
12546         "Attribute 'nest' is not supported in GHC calling convention");
12547   }
12548 
12549   if (LocVT == MVT::i32 || LocVT == MVT::i64) {
12550     // Pass in STG registers: Base, Sp, Hp, R1, R2, R3, R4, R5, R6, R7, SpLim
12551     //                        s1    s2  s3  s4  s5  s6  s7  s8  s9  s10 s11
12552     static const MCPhysReg GPRList[] = {
12553         RISCV::X9, RISCV::X18, RISCV::X19, RISCV::X20, RISCV::X21, RISCV::X22,
12554         RISCV::X23, RISCV::X24, RISCV::X25, RISCV::X26, RISCV::X27};
12555     if (unsigned Reg = State.AllocateReg(GPRList)) {
12556       State.addLoc(CCValAssign::getReg(ValNo, ValVT, Reg, LocVT, LocInfo));
12557       return false;
12558     }
12559   }
12560 
12561   if (LocVT == MVT::f32) {
12562     // Pass in STG registers: F1, ..., F6
12563     //                        fs0 ... fs5
12564     static const MCPhysReg FPR32List[] = {RISCV::F8_F, RISCV::F9_F,
12565                                           RISCV::F18_F, RISCV::F19_F,
12566                                           RISCV::F20_F, RISCV::F21_F};
12567     if (unsigned Reg = State.AllocateReg(FPR32List)) {
12568       State.addLoc(CCValAssign::getReg(ValNo, ValVT, Reg, LocVT, LocInfo));
12569       return false;
12570     }
12571   }
12572 
12573   if (LocVT == MVT::f64) {
12574     // Pass in STG registers: D1, ..., D6
12575     //                        fs6 ... fs11
12576     static const MCPhysReg FPR64List[] = {RISCV::F22_D, RISCV::F23_D,
12577                                           RISCV::F24_D, RISCV::F25_D,
12578                                           RISCV::F26_D, RISCV::F27_D};
12579     if (unsigned Reg = State.AllocateReg(FPR64List)) {
12580       State.addLoc(CCValAssign::getReg(ValNo, ValVT, Reg, LocVT, LocInfo));
12581       return false;
12582     }
12583   }
12584 
12585   report_fatal_error("No registers left in GHC calling convention");
12586   return true;
12587 }
12588 
12589 // Transform physical registers into virtual registers.
LowerFormalArguments(SDValue Chain,CallingConv::ID CallConv,bool IsVarArg,const SmallVectorImpl<ISD::InputArg> & Ins,const SDLoc & DL,SelectionDAG & DAG,SmallVectorImpl<SDValue> & InVals) const12590 SDValue RISCVTargetLowering::LowerFormalArguments(
12591     SDValue Chain, CallingConv::ID CallConv, bool IsVarArg,
12592     const SmallVectorImpl<ISD::InputArg> &Ins, const SDLoc &DL,
12593     SelectionDAG &DAG, SmallVectorImpl<SDValue> &InVals) const {
12594 
12595   MachineFunction &MF = DAG.getMachineFunction();
12596 
12597   switch (CallConv) {
12598   default:
12599     report_fatal_error("Unsupported calling convention");
12600   case CallingConv::C:
12601   case CallingConv::Fast:
12602     break;
12603   case CallingConv::GHC:
12604     if (!MF.getSubtarget().getFeatureBits()[RISCV::FeatureStdExtF] ||
12605         !MF.getSubtarget().getFeatureBits()[RISCV::FeatureStdExtD])
12606       report_fatal_error(
12607         "GHC calling convention requires the F and D instruction set extensions");
12608   }
12609 
12610   const Function &Func = MF.getFunction();
12611   if (Func.hasFnAttribute("interrupt")) {
12612     if (!Func.arg_empty())
12613       report_fatal_error(
12614         "Functions with the interrupt attribute cannot have arguments!");
12615 
12616     StringRef Kind =
12617       MF.getFunction().getFnAttribute("interrupt").getValueAsString();
12618 
12619     if (!(Kind == "user" || Kind == "supervisor" || Kind == "machine"))
12620       report_fatal_error(
12621         "Function interrupt attribute argument not supported!");
12622   }
12623 
12624   EVT PtrVT = getPointerTy(DAG.getDataLayout());
12625   MVT XLenVT = Subtarget.getXLenVT();
12626   unsigned XLenInBytes = Subtarget.getXLen() / 8;
12627   // Used with vargs to acumulate store chains.
12628   std::vector<SDValue> OutChains;
12629 
12630   // Assign locations to all of the incoming arguments.
12631   SmallVector<CCValAssign, 16> ArgLocs;
12632   CCState CCInfo(CallConv, IsVarArg, MF, ArgLocs, *DAG.getContext());
12633 
12634   if (CallConv == CallingConv::GHC)
12635     CCInfo.AnalyzeFormalArguments(Ins, CC_RISCV_GHC);
12636   else
12637     analyzeInputArgs(MF, CCInfo, Ins, /*IsRet=*/false,
12638                      CallConv == CallingConv::Fast ? CC_RISCV_FastCC
12639                                                    : CC_RISCV);
12640 
12641   for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
12642     CCValAssign &VA = ArgLocs[i];
12643     SDValue ArgValue;
12644     // Passing f64 on RV32D with a soft float ABI must be handled as a special
12645     // case.
12646     if (VA.getLocVT() == MVT::i32 && VA.getValVT() == MVT::f64)
12647       ArgValue = unpackF64OnRV32DSoftABI(DAG, Chain, VA, DL);
12648     else if (VA.isRegLoc())
12649       ArgValue = unpackFromRegLoc(DAG, Chain, VA, DL, Ins[i], *this);
12650     else
12651       ArgValue = unpackFromMemLoc(DAG, Chain, VA, DL);
12652 
12653     if (VA.getLocInfo() == CCValAssign::Indirect) {
12654       // If the original argument was split and passed by reference (e.g. i128
12655       // on RV32), we need to load all parts of it here (using the same
12656       // address). Vectors may be partly split to registers and partly to the
12657       // stack, in which case the base address is partly offset and subsequent
12658       // stores are relative to that.
12659       InVals.push_back(DAG.getLoad(VA.getValVT(), DL, Chain, ArgValue,
12660                                    MachinePointerInfo()));
12661       unsigned ArgIndex = Ins[i].OrigArgIndex;
12662       unsigned ArgPartOffset = Ins[i].PartOffset;
12663       assert(VA.getValVT().isVector() || ArgPartOffset == 0);
12664       while (i + 1 != e && Ins[i + 1].OrigArgIndex == ArgIndex) {
12665         CCValAssign &PartVA = ArgLocs[i + 1];
12666         unsigned PartOffset = Ins[i + 1].PartOffset - ArgPartOffset;
12667         SDValue Offset = DAG.getIntPtrConstant(PartOffset, DL);
12668         if (PartVA.getValVT().isScalableVector())
12669           Offset = DAG.getNode(ISD::VSCALE, DL, XLenVT, Offset);
12670         SDValue Address = DAG.getNode(ISD::ADD, DL, PtrVT, ArgValue, Offset);
12671         InVals.push_back(DAG.getLoad(PartVA.getValVT(), DL, Chain, Address,
12672                                      MachinePointerInfo()));
12673         ++i;
12674       }
12675       continue;
12676     }
12677     InVals.push_back(ArgValue);
12678   }
12679 
12680   if (any_of(ArgLocs,
12681              [](CCValAssign &VA) { return VA.getLocVT().isScalableVector(); }))
12682     MF.getInfo<RISCVMachineFunctionInfo>()->setIsVectorCall();
12683 
12684   if (IsVarArg) {
12685     ArrayRef<MCPhysReg> ArgRegs = ArrayRef(ArgGPRs);
12686     unsigned Idx = CCInfo.getFirstUnallocated(ArgRegs);
12687     const TargetRegisterClass *RC = &RISCV::GPRRegClass;
12688     MachineFrameInfo &MFI = MF.getFrameInfo();
12689     MachineRegisterInfo &RegInfo = MF.getRegInfo();
12690     RISCVMachineFunctionInfo *RVFI = MF.getInfo<RISCVMachineFunctionInfo>();
12691 
12692     // Offset of the first variable argument from stack pointer, and size of
12693     // the vararg save area. For now, the varargs save area is either zero or
12694     // large enough to hold a0-a7.
12695     int VaArgOffset, VarArgsSaveSize;
12696 
12697     // If all registers are allocated, then all varargs must be passed on the
12698     // stack and we don't need to save any argregs.
12699     if (ArgRegs.size() == Idx) {
12700       VaArgOffset = CCInfo.getNextStackOffset();
12701       VarArgsSaveSize = 0;
12702     } else {
12703       VarArgsSaveSize = XLenInBytes * (ArgRegs.size() - Idx);
12704       VaArgOffset = -VarArgsSaveSize;
12705     }
12706 
12707     // Record the frame index of the first variable argument
12708     // which is a value necessary to VASTART.
12709     int FI = MFI.CreateFixedObject(XLenInBytes, VaArgOffset, true);
12710     RVFI->setVarArgsFrameIndex(FI);
12711 
12712     // If saving an odd number of registers then create an extra stack slot to
12713     // ensure that the frame pointer is 2*XLEN-aligned, which in turn ensures
12714     // offsets to even-numbered registered remain 2*XLEN-aligned.
12715     if (Idx % 2) {
12716       MFI.CreateFixedObject(XLenInBytes, VaArgOffset - (int)XLenInBytes, true);
12717       VarArgsSaveSize += XLenInBytes;
12718     }
12719 
12720     // Copy the integer registers that may have been used for passing varargs
12721     // to the vararg save area.
12722     for (unsigned I = Idx; I < ArgRegs.size();
12723          ++I, VaArgOffset += XLenInBytes) {
12724       const Register Reg = RegInfo.createVirtualRegister(RC);
12725       RegInfo.addLiveIn(ArgRegs[I], Reg);
12726       SDValue ArgValue = DAG.getCopyFromReg(Chain, DL, Reg, XLenVT);
12727       FI = MFI.CreateFixedObject(XLenInBytes, VaArgOffset, true);
12728       SDValue PtrOff = DAG.getFrameIndex(FI, getPointerTy(DAG.getDataLayout()));
12729       SDValue Store = DAG.getStore(Chain, DL, ArgValue, PtrOff,
12730                                    MachinePointerInfo::getFixedStack(MF, FI));
12731       cast<StoreSDNode>(Store.getNode())
12732           ->getMemOperand()
12733           ->setValue((Value *)nullptr);
12734       OutChains.push_back(Store);
12735     }
12736     RVFI->setVarArgsSaveSize(VarArgsSaveSize);
12737   }
12738 
12739   // All stores are grouped in one node to allow the matching between
12740   // the size of Ins and InVals. This only happens for vararg functions.
12741   if (!OutChains.empty()) {
12742     OutChains.push_back(Chain);
12743     Chain = DAG.getNode(ISD::TokenFactor, DL, MVT::Other, OutChains);
12744   }
12745 
12746   return Chain;
12747 }
12748 
12749 /// isEligibleForTailCallOptimization - Check whether the call is eligible
12750 /// for tail call optimization.
12751 /// Note: This is modelled after ARM's IsEligibleForTailCallOptimization.
isEligibleForTailCallOptimization(CCState & CCInfo,CallLoweringInfo & CLI,MachineFunction & MF,const SmallVector<CCValAssign,16> & ArgLocs) const12752 bool RISCVTargetLowering::isEligibleForTailCallOptimization(
12753     CCState &CCInfo, CallLoweringInfo &CLI, MachineFunction &MF,
12754     const SmallVector<CCValAssign, 16> &ArgLocs) const {
12755 
12756   auto &Callee = CLI.Callee;
12757   auto CalleeCC = CLI.CallConv;
12758   auto &Outs = CLI.Outs;
12759   auto &Caller = MF.getFunction();
12760   auto CallerCC = Caller.getCallingConv();
12761 
12762   // Exception-handling functions need a special set of instructions to
12763   // indicate a return to the hardware. Tail-calling another function would
12764   // probably break this.
12765   // TODO: The "interrupt" attribute isn't currently defined by RISC-V. This
12766   // should be expanded as new function attributes are introduced.
12767   if (Caller.hasFnAttribute("interrupt"))
12768     return false;
12769 
12770   // Do not tail call opt if the stack is used to pass parameters.
12771   if (CCInfo.getNextStackOffset() != 0)
12772     return false;
12773 
12774   // Do not tail call opt if any parameters need to be passed indirectly.
12775   // Since long doubles (fp128) and i128 are larger than 2*XLEN, they are
12776   // passed indirectly. So the address of the value will be passed in a
12777   // register, or if not available, then the address is put on the stack. In
12778   // order to pass indirectly, space on the stack often needs to be allocated
12779   // in order to store the value. In this case the CCInfo.getNextStackOffset()
12780   // != 0 check is not enough and we need to check if any CCValAssign ArgsLocs
12781   // are passed CCValAssign::Indirect.
12782   for (auto &VA : ArgLocs)
12783     if (VA.getLocInfo() == CCValAssign::Indirect)
12784       return false;
12785 
12786   // Do not tail call opt if either caller or callee uses struct return
12787   // semantics.
12788   auto IsCallerStructRet = Caller.hasStructRetAttr();
12789   auto IsCalleeStructRet = Outs.empty() ? false : Outs[0].Flags.isSRet();
12790   if (IsCallerStructRet || IsCalleeStructRet)
12791     return false;
12792 
12793   // Externally-defined functions with weak linkage should not be
12794   // tail-called. The behaviour of branch instructions in this situation (as
12795   // used for tail calls) is implementation-defined, so we cannot rely on the
12796   // linker replacing the tail call with a return.
12797   if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) {
12798     const GlobalValue *GV = G->getGlobal();
12799     if (GV->hasExternalWeakLinkage())
12800       return false;
12801   }
12802 
12803   // The callee has to preserve all registers the caller needs to preserve.
12804   const RISCVRegisterInfo *TRI = Subtarget.getRegisterInfo();
12805   const uint32_t *CallerPreserved = TRI->getCallPreservedMask(MF, CallerCC);
12806   if (CalleeCC != CallerCC) {
12807     const uint32_t *CalleePreserved = TRI->getCallPreservedMask(MF, CalleeCC);
12808     if (!TRI->regmaskSubsetEqual(CallerPreserved, CalleePreserved))
12809       return false;
12810   }
12811 
12812   // Byval parameters hand the function a pointer directly into the stack area
12813   // we want to reuse during a tail call. Working around this *is* possible
12814   // but less efficient and uglier in LowerCall.
12815   for (auto &Arg : Outs)
12816     if (Arg.Flags.isByVal())
12817       return false;
12818 
12819   return true;
12820 }
12821 
getPrefTypeAlign(EVT VT,SelectionDAG & DAG)12822 static Align getPrefTypeAlign(EVT VT, SelectionDAG &DAG) {
12823   return DAG.getDataLayout().getPrefTypeAlign(
12824       VT.getTypeForEVT(*DAG.getContext()));
12825 }
12826 
12827 // Lower a call to a callseq_start + CALL + callseq_end chain, and add input
12828 // and output parameter nodes.
LowerCall(CallLoweringInfo & CLI,SmallVectorImpl<SDValue> & InVals) const12829 SDValue RISCVTargetLowering::LowerCall(CallLoweringInfo &CLI,
12830                                        SmallVectorImpl<SDValue> &InVals) const {
12831   SelectionDAG &DAG = CLI.DAG;
12832   SDLoc &DL = CLI.DL;
12833   SmallVectorImpl<ISD::OutputArg> &Outs = CLI.Outs;
12834   SmallVectorImpl<SDValue> &OutVals = CLI.OutVals;
12835   SmallVectorImpl<ISD::InputArg> &Ins = CLI.Ins;
12836   SDValue Chain = CLI.Chain;
12837   SDValue Callee = CLI.Callee;
12838   bool &IsTailCall = CLI.IsTailCall;
12839   CallingConv::ID CallConv = CLI.CallConv;
12840   bool IsVarArg = CLI.IsVarArg;
12841   EVT PtrVT = getPointerTy(DAG.getDataLayout());
12842   MVT XLenVT = Subtarget.getXLenVT();
12843 
12844   MachineFunction &MF = DAG.getMachineFunction();
12845 
12846   // Analyze the operands of the call, assigning locations to each operand.
12847   SmallVector<CCValAssign, 16> ArgLocs;
12848   CCState ArgCCInfo(CallConv, IsVarArg, MF, ArgLocs, *DAG.getContext());
12849 
12850   if (CallConv == CallingConv::GHC)
12851     ArgCCInfo.AnalyzeCallOperands(Outs, CC_RISCV_GHC);
12852   else
12853     analyzeOutputArgs(MF, ArgCCInfo, Outs, /*IsRet=*/false, &CLI,
12854                       CallConv == CallingConv::Fast ? CC_RISCV_FastCC
12855                                                     : CC_RISCV);
12856 
12857   // Check if it's really possible to do a tail call.
12858   if (IsTailCall)
12859     IsTailCall = isEligibleForTailCallOptimization(ArgCCInfo, CLI, MF, ArgLocs);
12860 
12861   if (IsTailCall)
12862     ++NumTailCalls;
12863   else if (CLI.CB && CLI.CB->isMustTailCall())
12864     report_fatal_error("failed to perform tail call elimination on a call "
12865                        "site marked musttail");
12866 
12867   // Get a count of how many bytes are to be pushed on the stack.
12868   unsigned NumBytes = ArgCCInfo.getNextStackOffset();
12869 
12870   // Create local copies for byval args
12871   SmallVector<SDValue, 8> ByValArgs;
12872   for (unsigned i = 0, e = Outs.size(); i != e; ++i) {
12873     ISD::ArgFlagsTy Flags = Outs[i].Flags;
12874     if (!Flags.isByVal())
12875       continue;
12876 
12877     SDValue Arg = OutVals[i];
12878     unsigned Size = Flags.getByValSize();
12879     Align Alignment = Flags.getNonZeroByValAlign();
12880 
12881     int FI =
12882         MF.getFrameInfo().CreateStackObject(Size, Alignment, /*isSS=*/false);
12883     SDValue FIPtr = DAG.getFrameIndex(FI, getPointerTy(DAG.getDataLayout()));
12884     SDValue SizeNode = DAG.getConstant(Size, DL, XLenVT);
12885 
12886     Chain = DAG.getMemcpy(Chain, DL, FIPtr, Arg, SizeNode, Alignment,
12887                           /*IsVolatile=*/false,
12888                           /*AlwaysInline=*/false, IsTailCall,
12889                           MachinePointerInfo(), MachinePointerInfo());
12890     ByValArgs.push_back(FIPtr);
12891   }
12892 
12893   if (!IsTailCall)
12894     Chain = DAG.getCALLSEQ_START(Chain, NumBytes, 0, CLI.DL);
12895 
12896   // Copy argument values to their designated locations.
12897   SmallVector<std::pair<Register, SDValue>, 8> RegsToPass;
12898   SmallVector<SDValue, 8> MemOpChains;
12899   SDValue StackPtr;
12900   for (unsigned i = 0, j = 0, e = ArgLocs.size(); i != e; ++i) {
12901     CCValAssign &VA = ArgLocs[i];
12902     SDValue ArgValue = OutVals[i];
12903     ISD::ArgFlagsTy Flags = Outs[i].Flags;
12904 
12905     // Handle passing f64 on RV32D with a soft float ABI as a special case.
12906     bool IsF64OnRV32DSoftABI =
12907         VA.getLocVT() == MVT::i32 && VA.getValVT() == MVT::f64;
12908     if (IsF64OnRV32DSoftABI && VA.isRegLoc()) {
12909       SDValue SplitF64 = DAG.getNode(
12910           RISCVISD::SplitF64, DL, DAG.getVTList(MVT::i32, MVT::i32), ArgValue);
12911       SDValue Lo = SplitF64.getValue(0);
12912       SDValue Hi = SplitF64.getValue(1);
12913 
12914       Register RegLo = VA.getLocReg();
12915       RegsToPass.push_back(std::make_pair(RegLo, Lo));
12916 
12917       if (RegLo == RISCV::X17) {
12918         // Second half of f64 is passed on the stack.
12919         // Work out the address of the stack slot.
12920         if (!StackPtr.getNode())
12921           StackPtr = DAG.getCopyFromReg(Chain, DL, RISCV::X2, PtrVT);
12922         // Emit the store.
12923         MemOpChains.push_back(
12924             DAG.getStore(Chain, DL, Hi, StackPtr, MachinePointerInfo()));
12925       } else {
12926         // Second half of f64 is passed in another GPR.
12927         assert(RegLo < RISCV::X31 && "Invalid register pair");
12928         Register RegHigh = RegLo + 1;
12929         RegsToPass.push_back(std::make_pair(RegHigh, Hi));
12930       }
12931       continue;
12932     }
12933 
12934     // IsF64OnRV32DSoftABI && VA.isMemLoc() is handled below in the same way
12935     // as any other MemLoc.
12936 
12937     // Promote the value if needed.
12938     // For now, only handle fully promoted and indirect arguments.
12939     if (VA.getLocInfo() == CCValAssign::Indirect) {
12940       // Store the argument in a stack slot and pass its address.
12941       Align StackAlign =
12942           std::max(getPrefTypeAlign(Outs[i].ArgVT, DAG),
12943                    getPrefTypeAlign(ArgValue.getValueType(), DAG));
12944       TypeSize StoredSize = ArgValue.getValueType().getStoreSize();
12945       // If the original argument was split (e.g. i128), we need
12946       // to store the required parts of it here (and pass just one address).
12947       // Vectors may be partly split to registers and partly to the stack, in
12948       // which case the base address is partly offset and subsequent stores are
12949       // relative to that.
12950       unsigned ArgIndex = Outs[i].OrigArgIndex;
12951       unsigned ArgPartOffset = Outs[i].PartOffset;
12952       assert(VA.getValVT().isVector() || ArgPartOffset == 0);
12953       // Calculate the total size to store. We don't have access to what we're
12954       // actually storing other than performing the loop and collecting the
12955       // info.
12956       SmallVector<std::pair<SDValue, SDValue>> Parts;
12957       while (i + 1 != e && Outs[i + 1].OrigArgIndex == ArgIndex) {
12958         SDValue PartValue = OutVals[i + 1];
12959         unsigned PartOffset = Outs[i + 1].PartOffset - ArgPartOffset;
12960         SDValue Offset = DAG.getIntPtrConstant(PartOffset, DL);
12961         EVT PartVT = PartValue.getValueType();
12962         if (PartVT.isScalableVector())
12963           Offset = DAG.getNode(ISD::VSCALE, DL, XLenVT, Offset);
12964         StoredSize += PartVT.getStoreSize();
12965         StackAlign = std::max(StackAlign, getPrefTypeAlign(PartVT, DAG));
12966         Parts.push_back(std::make_pair(PartValue, Offset));
12967         ++i;
12968       }
12969       SDValue SpillSlot = DAG.CreateStackTemporary(StoredSize, StackAlign);
12970       int FI = cast<FrameIndexSDNode>(SpillSlot)->getIndex();
12971       MemOpChains.push_back(
12972           DAG.getStore(Chain, DL, ArgValue, SpillSlot,
12973                        MachinePointerInfo::getFixedStack(MF, FI)));
12974       for (const auto &Part : Parts) {
12975         SDValue PartValue = Part.first;
12976         SDValue PartOffset = Part.second;
12977         SDValue Address =
12978             DAG.getNode(ISD::ADD, DL, PtrVT, SpillSlot, PartOffset);
12979         MemOpChains.push_back(
12980             DAG.getStore(Chain, DL, PartValue, Address,
12981                          MachinePointerInfo::getFixedStack(MF, FI)));
12982       }
12983       ArgValue = SpillSlot;
12984     } else {
12985       ArgValue = convertValVTToLocVT(DAG, ArgValue, VA, DL, Subtarget);
12986     }
12987 
12988     // Use local copy if it is a byval arg.
12989     if (Flags.isByVal())
12990       ArgValue = ByValArgs[j++];
12991 
12992     if (VA.isRegLoc()) {
12993       // Queue up the argument copies and emit them at the end.
12994       RegsToPass.push_back(std::make_pair(VA.getLocReg(), ArgValue));
12995     } else {
12996       assert(VA.isMemLoc() && "Argument not register or memory");
12997       assert(!IsTailCall && "Tail call not allowed if stack is used "
12998                             "for passing parameters");
12999 
13000       // Work out the address of the stack slot.
13001       if (!StackPtr.getNode())
13002         StackPtr = DAG.getCopyFromReg(Chain, DL, RISCV::X2, PtrVT);
13003       SDValue Address =
13004           DAG.getNode(ISD::ADD, DL, PtrVT, StackPtr,
13005                       DAG.getIntPtrConstant(VA.getLocMemOffset(), DL));
13006 
13007       // Emit the store.
13008       MemOpChains.push_back(
13009           DAG.getStore(Chain, DL, ArgValue, Address, MachinePointerInfo()));
13010     }
13011   }
13012 
13013   // Join the stores, which are independent of one another.
13014   if (!MemOpChains.empty())
13015     Chain = DAG.getNode(ISD::TokenFactor, DL, MVT::Other, MemOpChains);
13016 
13017   SDValue Glue;
13018 
13019   // Build a sequence of copy-to-reg nodes, chained and glued together.
13020   for (auto &Reg : RegsToPass) {
13021     Chain = DAG.getCopyToReg(Chain, DL, Reg.first, Reg.second, Glue);
13022     Glue = Chain.getValue(1);
13023   }
13024 
13025   // Validate that none of the argument registers have been marked as
13026   // reserved, if so report an error. Do the same for the return address if this
13027   // is not a tailcall.
13028   validateCCReservedRegs(RegsToPass, MF);
13029   if (!IsTailCall &&
13030       MF.getSubtarget<RISCVSubtarget>().isRegisterReservedByUser(RISCV::X1))
13031     MF.getFunction().getContext().diagnose(DiagnosticInfoUnsupported{
13032         MF.getFunction(),
13033         "Return address register required, but has been reserved."});
13034 
13035   // If the callee is a GlobalAddress/ExternalSymbol node, turn it into a
13036   // TargetGlobalAddress/TargetExternalSymbol node so that legalize won't
13037   // split it and then direct call can be matched by PseudoCALL.
13038   if (GlobalAddressSDNode *S = dyn_cast<GlobalAddressSDNode>(Callee)) {
13039     const GlobalValue *GV = S->getGlobal();
13040 
13041     unsigned OpFlags = RISCVII::MO_CALL;
13042     if (!getTargetMachine().shouldAssumeDSOLocal(*GV->getParent(), GV))
13043       OpFlags = RISCVII::MO_PLT;
13044 
13045     Callee = DAG.getTargetGlobalAddress(GV, DL, PtrVT, 0, OpFlags);
13046   } else if (ExternalSymbolSDNode *S = dyn_cast<ExternalSymbolSDNode>(Callee)) {
13047     unsigned OpFlags = RISCVII::MO_CALL;
13048 
13049     if (!getTargetMachine().shouldAssumeDSOLocal(*MF.getFunction().getParent(),
13050                                                  nullptr))
13051       OpFlags = RISCVII::MO_PLT;
13052 
13053     Callee = DAG.getTargetExternalSymbol(S->getSymbol(), PtrVT, OpFlags);
13054   }
13055 
13056   // The first call operand is the chain and the second is the target address.
13057   SmallVector<SDValue, 8> Ops;
13058   Ops.push_back(Chain);
13059   Ops.push_back(Callee);
13060 
13061   // Add argument registers to the end of the list so that they are
13062   // known live into the call.
13063   for (auto &Reg : RegsToPass)
13064     Ops.push_back(DAG.getRegister(Reg.first, Reg.second.getValueType()));
13065 
13066   if (!IsTailCall) {
13067     // Add a register mask operand representing the call-preserved registers.
13068     const TargetRegisterInfo *TRI = Subtarget.getRegisterInfo();
13069     const uint32_t *Mask = TRI->getCallPreservedMask(MF, CallConv);
13070     assert(Mask && "Missing call preserved mask for calling convention");
13071     Ops.push_back(DAG.getRegisterMask(Mask));
13072   }
13073 
13074   // Glue the call to the argument copies, if any.
13075   if (Glue.getNode())
13076     Ops.push_back(Glue);
13077 
13078   // Emit the call.
13079   SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Glue);
13080 
13081   if (IsTailCall) {
13082     MF.getFrameInfo().setHasTailCall();
13083     return DAG.getNode(RISCVISD::TAIL, DL, NodeTys, Ops);
13084   }
13085 
13086   Chain = DAG.getNode(RISCVISD::CALL, DL, NodeTys, Ops);
13087   DAG.addNoMergeSiteInfo(Chain.getNode(), CLI.NoMerge);
13088   Glue = Chain.getValue(1);
13089 
13090   // Mark the end of the call, which is glued to the call itself.
13091   Chain = DAG.getCALLSEQ_END(Chain, NumBytes, 0, Glue, DL);
13092   Glue = Chain.getValue(1);
13093 
13094   // Assign locations to each value returned by this call.
13095   SmallVector<CCValAssign, 16> RVLocs;
13096   CCState RetCCInfo(CallConv, IsVarArg, MF, RVLocs, *DAG.getContext());
13097   analyzeInputArgs(MF, RetCCInfo, Ins, /*IsRet=*/true, CC_RISCV);
13098 
13099   // Copy all of the result registers out of their specified physreg.
13100   for (auto &VA : RVLocs) {
13101     // Copy the value out
13102     SDValue RetValue =
13103         DAG.getCopyFromReg(Chain, DL, VA.getLocReg(), VA.getLocVT(), Glue);
13104     // Glue the RetValue to the end of the call sequence
13105     Chain = RetValue.getValue(1);
13106     Glue = RetValue.getValue(2);
13107 
13108     if (VA.getLocVT() == MVT::i32 && VA.getValVT() == MVT::f64) {
13109       assert(VA.getLocReg() == ArgGPRs[0] && "Unexpected reg assignment");
13110       SDValue RetValue2 =
13111           DAG.getCopyFromReg(Chain, DL, ArgGPRs[1], MVT::i32, Glue);
13112       Chain = RetValue2.getValue(1);
13113       Glue = RetValue2.getValue(2);
13114       RetValue = DAG.getNode(RISCVISD::BuildPairF64, DL, MVT::f64, RetValue,
13115                              RetValue2);
13116     }
13117 
13118     RetValue = convertLocVTToValVT(DAG, RetValue, VA, DL, Subtarget);
13119 
13120     InVals.push_back(RetValue);
13121   }
13122 
13123   return Chain;
13124 }
13125 
CanLowerReturn(CallingConv::ID CallConv,MachineFunction & MF,bool IsVarArg,const SmallVectorImpl<ISD::OutputArg> & Outs,LLVMContext & Context) const13126 bool RISCVTargetLowering::CanLowerReturn(
13127     CallingConv::ID CallConv, MachineFunction &MF, bool IsVarArg,
13128     const SmallVectorImpl<ISD::OutputArg> &Outs, LLVMContext &Context) const {
13129   SmallVector<CCValAssign, 16> RVLocs;
13130   CCState CCInfo(CallConv, IsVarArg, MF, RVLocs, Context);
13131 
13132   std::optional<unsigned> FirstMaskArgument;
13133   if (Subtarget.hasVInstructions())
13134     FirstMaskArgument = preAssignMask(Outs);
13135 
13136   for (unsigned i = 0, e = Outs.size(); i != e; ++i) {
13137     MVT VT = Outs[i].VT;
13138     ISD::ArgFlagsTy ArgFlags = Outs[i].Flags;
13139     RISCVABI::ABI ABI = MF.getSubtarget<RISCVSubtarget>().getTargetABI();
13140     if (CC_RISCV(MF.getDataLayout(), ABI, i, VT, VT, CCValAssign::Full,
13141                  ArgFlags, CCInfo, /*IsFixed=*/true, /*IsRet=*/true, nullptr,
13142                  *this, FirstMaskArgument))
13143       return false;
13144   }
13145   return true;
13146 }
13147 
13148 SDValue
LowerReturn(SDValue Chain,CallingConv::ID CallConv,bool IsVarArg,const SmallVectorImpl<ISD::OutputArg> & Outs,const SmallVectorImpl<SDValue> & OutVals,const SDLoc & DL,SelectionDAG & DAG) const13149 RISCVTargetLowering::LowerReturn(SDValue Chain, CallingConv::ID CallConv,
13150                                  bool IsVarArg,
13151                                  const SmallVectorImpl<ISD::OutputArg> &Outs,
13152                                  const SmallVectorImpl<SDValue> &OutVals,
13153                                  const SDLoc &DL, SelectionDAG &DAG) const {
13154   MachineFunction &MF = DAG.getMachineFunction();
13155   const RISCVSubtarget &STI = MF.getSubtarget<RISCVSubtarget>();
13156 
13157   // Stores the assignment of the return value to a location.
13158   SmallVector<CCValAssign, 16> RVLocs;
13159 
13160   // Info about the registers and stack slot.
13161   CCState CCInfo(CallConv, IsVarArg, DAG.getMachineFunction(), RVLocs,
13162                  *DAG.getContext());
13163 
13164   analyzeOutputArgs(DAG.getMachineFunction(), CCInfo, Outs, /*IsRet=*/true,
13165                     nullptr, CC_RISCV);
13166 
13167   if (CallConv == CallingConv::GHC && !RVLocs.empty())
13168     report_fatal_error("GHC functions return void only");
13169 
13170   SDValue Glue;
13171   SmallVector<SDValue, 4> RetOps(1, Chain);
13172 
13173   // Copy the result values into the output registers.
13174   for (unsigned i = 0, e = RVLocs.size(); i < e; ++i) {
13175     SDValue Val = OutVals[i];
13176     CCValAssign &VA = RVLocs[i];
13177     assert(VA.isRegLoc() && "Can only return in registers!");
13178 
13179     if (VA.getLocVT() == MVT::i32 && VA.getValVT() == MVT::f64) {
13180       // Handle returning f64 on RV32D with a soft float ABI.
13181       assert(VA.isRegLoc() && "Expected return via registers");
13182       SDValue SplitF64 = DAG.getNode(RISCVISD::SplitF64, DL,
13183                                      DAG.getVTList(MVT::i32, MVT::i32), Val);
13184       SDValue Lo = SplitF64.getValue(0);
13185       SDValue Hi = SplitF64.getValue(1);
13186       Register RegLo = VA.getLocReg();
13187       assert(RegLo < RISCV::X31 && "Invalid register pair");
13188       Register RegHi = RegLo + 1;
13189 
13190       if (STI.isRegisterReservedByUser(RegLo) ||
13191           STI.isRegisterReservedByUser(RegHi))
13192         MF.getFunction().getContext().diagnose(DiagnosticInfoUnsupported{
13193             MF.getFunction(),
13194             "Return value register required, but has been reserved."});
13195 
13196       Chain = DAG.getCopyToReg(Chain, DL, RegLo, Lo, Glue);
13197       Glue = Chain.getValue(1);
13198       RetOps.push_back(DAG.getRegister(RegLo, MVT::i32));
13199       Chain = DAG.getCopyToReg(Chain, DL, RegHi, Hi, Glue);
13200       Glue = Chain.getValue(1);
13201       RetOps.push_back(DAG.getRegister(RegHi, MVT::i32));
13202     } else {
13203       // Handle a 'normal' return.
13204       Val = convertValVTToLocVT(DAG, Val, VA, DL, Subtarget);
13205       Chain = DAG.getCopyToReg(Chain, DL, VA.getLocReg(), Val, Glue);
13206 
13207       if (STI.isRegisterReservedByUser(VA.getLocReg()))
13208         MF.getFunction().getContext().diagnose(DiagnosticInfoUnsupported{
13209             MF.getFunction(),
13210             "Return value register required, but has been reserved."});
13211 
13212       // Guarantee that all emitted copies are stuck together.
13213       Glue = Chain.getValue(1);
13214       RetOps.push_back(DAG.getRegister(VA.getLocReg(), VA.getLocVT()));
13215     }
13216   }
13217 
13218   RetOps[0] = Chain; // Update chain.
13219 
13220   // Add the glue node if we have it.
13221   if (Glue.getNode()) {
13222     RetOps.push_back(Glue);
13223   }
13224 
13225   if (any_of(RVLocs,
13226              [](CCValAssign &VA) { return VA.getLocVT().isScalableVector(); }))
13227     MF.getInfo<RISCVMachineFunctionInfo>()->setIsVectorCall();
13228 
13229   unsigned RetOpc = RISCVISD::RET_FLAG;
13230   // Interrupt service routines use different return instructions.
13231   const Function &Func = DAG.getMachineFunction().getFunction();
13232   if (Func.hasFnAttribute("interrupt")) {
13233     if (!Func.getReturnType()->isVoidTy())
13234       report_fatal_error(
13235           "Functions with the interrupt attribute must have void return type!");
13236 
13237     MachineFunction &MF = DAG.getMachineFunction();
13238     StringRef Kind =
13239       MF.getFunction().getFnAttribute("interrupt").getValueAsString();
13240 
13241     if (Kind == "user")
13242       RetOpc = RISCVISD::URET_FLAG;
13243     else if (Kind == "supervisor")
13244       RetOpc = RISCVISD::SRET_FLAG;
13245     else
13246       RetOpc = RISCVISD::MRET_FLAG;
13247   }
13248 
13249   return DAG.getNode(RetOpc, DL, MVT::Other, RetOps);
13250 }
13251 
validateCCReservedRegs(const SmallVectorImpl<std::pair<llvm::Register,llvm::SDValue>> & Regs,MachineFunction & MF) const13252 void RISCVTargetLowering::validateCCReservedRegs(
13253     const SmallVectorImpl<std::pair<llvm::Register, llvm::SDValue>> &Regs,
13254     MachineFunction &MF) const {
13255   const Function &F = MF.getFunction();
13256   const RISCVSubtarget &STI = MF.getSubtarget<RISCVSubtarget>();
13257 
13258   if (llvm::any_of(Regs, [&STI](auto Reg) {
13259         return STI.isRegisterReservedByUser(Reg.first);
13260       }))
13261     F.getContext().diagnose(DiagnosticInfoUnsupported{
13262         F, "Argument register required, but has been reserved."});
13263 }
13264 
13265 // Check if the result of the node is only used as a return value, as
13266 // otherwise we can't perform a tail-call.
isUsedByReturnOnly(SDNode * N,SDValue & Chain) const13267 bool RISCVTargetLowering::isUsedByReturnOnly(SDNode *N, SDValue &Chain) const {
13268   if (N->getNumValues() != 1)
13269     return false;
13270   if (!N->hasNUsesOfValue(1, 0))
13271     return false;
13272 
13273   SDNode *Copy = *N->use_begin();
13274   // TODO: Handle additional opcodes in order to support tail-calling libcalls
13275   // with soft float ABIs.
13276   if (Copy->getOpcode() != ISD::CopyToReg) {
13277     return false;
13278   }
13279 
13280   // If the ISD::CopyToReg has a glue operand, we conservatively assume it
13281   // isn't safe to perform a tail call.
13282   if (Copy->getOperand(Copy->getNumOperands() - 1).getValueType() == MVT::Glue)
13283     return false;
13284 
13285   // The copy must be used by a RISCVISD::RET_FLAG, and nothing else.
13286   bool HasRet = false;
13287   for (SDNode *Node : Copy->uses()) {
13288     if (Node->getOpcode() != RISCVISD::RET_FLAG)
13289       return false;
13290     HasRet = true;
13291   }
13292   if (!HasRet)
13293     return false;
13294 
13295   Chain = Copy->getOperand(0);
13296   return true;
13297 }
13298 
mayBeEmittedAsTailCall(const CallInst * CI) const13299 bool RISCVTargetLowering::mayBeEmittedAsTailCall(const CallInst *CI) const {
13300   return CI->isTailCall();
13301 }
13302 
getTargetNodeName(unsigned Opcode) const13303 const char *RISCVTargetLowering::getTargetNodeName(unsigned Opcode) const {
13304 #define NODE_NAME_CASE(NODE)                                                   \
13305   case RISCVISD::NODE:                                                         \
13306     return "RISCVISD::" #NODE;
13307   // clang-format off
13308   switch ((RISCVISD::NodeType)Opcode) {
13309   case RISCVISD::FIRST_NUMBER:
13310     break;
13311   NODE_NAME_CASE(RET_FLAG)
13312   NODE_NAME_CASE(URET_FLAG)
13313   NODE_NAME_CASE(SRET_FLAG)
13314   NODE_NAME_CASE(MRET_FLAG)
13315   NODE_NAME_CASE(CALL)
13316   NODE_NAME_CASE(SELECT_CC)
13317   NODE_NAME_CASE(BR_CC)
13318   NODE_NAME_CASE(BuildPairF64)
13319   NODE_NAME_CASE(SplitF64)
13320   NODE_NAME_CASE(TAIL)
13321   NODE_NAME_CASE(ADD_LO)
13322   NODE_NAME_CASE(HI)
13323   NODE_NAME_CASE(LLA)
13324   NODE_NAME_CASE(ADD_TPREL)
13325   NODE_NAME_CASE(LA)
13326   NODE_NAME_CASE(LA_TLS_IE)
13327   NODE_NAME_CASE(LA_TLS_GD)
13328   NODE_NAME_CASE(MULHSU)
13329   NODE_NAME_CASE(SLLW)
13330   NODE_NAME_CASE(SRAW)
13331   NODE_NAME_CASE(SRLW)
13332   NODE_NAME_CASE(DIVW)
13333   NODE_NAME_CASE(DIVUW)
13334   NODE_NAME_CASE(REMUW)
13335   NODE_NAME_CASE(ROLW)
13336   NODE_NAME_CASE(RORW)
13337   NODE_NAME_CASE(CLZW)
13338   NODE_NAME_CASE(CTZW)
13339   NODE_NAME_CASE(ABSW)
13340   NODE_NAME_CASE(FMV_H_X)
13341   NODE_NAME_CASE(FMV_X_ANYEXTH)
13342   NODE_NAME_CASE(FMV_X_SIGNEXTH)
13343   NODE_NAME_CASE(FMV_W_X_RV64)
13344   NODE_NAME_CASE(FMV_X_ANYEXTW_RV64)
13345   NODE_NAME_CASE(FCVT_X)
13346   NODE_NAME_CASE(FCVT_XU)
13347   NODE_NAME_CASE(FCVT_W_RV64)
13348   NODE_NAME_CASE(FCVT_WU_RV64)
13349   NODE_NAME_CASE(STRICT_FCVT_W_RV64)
13350   NODE_NAME_CASE(STRICT_FCVT_WU_RV64)
13351   NODE_NAME_CASE(FROUND)
13352   NODE_NAME_CASE(READ_CYCLE_WIDE)
13353   NODE_NAME_CASE(BREV8)
13354   NODE_NAME_CASE(ORC_B)
13355   NODE_NAME_CASE(ZIP)
13356   NODE_NAME_CASE(UNZIP)
13357   NODE_NAME_CASE(VMV_V_X_VL)
13358   NODE_NAME_CASE(VFMV_V_F_VL)
13359   NODE_NAME_CASE(VMV_X_S)
13360   NODE_NAME_CASE(VMV_S_X_VL)
13361   NODE_NAME_CASE(VFMV_S_F_VL)
13362   NODE_NAME_CASE(SPLAT_VECTOR_SPLIT_I64_VL)
13363   NODE_NAME_CASE(READ_VLENB)
13364   NODE_NAME_CASE(TRUNCATE_VECTOR_VL)
13365   NODE_NAME_CASE(VSLIDEUP_VL)
13366   NODE_NAME_CASE(VSLIDE1UP_VL)
13367   NODE_NAME_CASE(VSLIDEDOWN_VL)
13368   NODE_NAME_CASE(VSLIDE1DOWN_VL)
13369   NODE_NAME_CASE(VID_VL)
13370   NODE_NAME_CASE(VFNCVT_ROD_VL)
13371   NODE_NAME_CASE(VECREDUCE_ADD_VL)
13372   NODE_NAME_CASE(VECREDUCE_UMAX_VL)
13373   NODE_NAME_CASE(VECREDUCE_SMAX_VL)
13374   NODE_NAME_CASE(VECREDUCE_UMIN_VL)
13375   NODE_NAME_CASE(VECREDUCE_SMIN_VL)
13376   NODE_NAME_CASE(VECREDUCE_AND_VL)
13377   NODE_NAME_CASE(VECREDUCE_OR_VL)
13378   NODE_NAME_CASE(VECREDUCE_XOR_VL)
13379   NODE_NAME_CASE(VECREDUCE_FADD_VL)
13380   NODE_NAME_CASE(VECREDUCE_SEQ_FADD_VL)
13381   NODE_NAME_CASE(VECREDUCE_FMIN_VL)
13382   NODE_NAME_CASE(VECREDUCE_FMAX_VL)
13383   NODE_NAME_CASE(ADD_VL)
13384   NODE_NAME_CASE(AND_VL)
13385   NODE_NAME_CASE(MUL_VL)
13386   NODE_NAME_CASE(OR_VL)
13387   NODE_NAME_CASE(SDIV_VL)
13388   NODE_NAME_CASE(SHL_VL)
13389   NODE_NAME_CASE(SREM_VL)
13390   NODE_NAME_CASE(SRA_VL)
13391   NODE_NAME_CASE(SRL_VL)
13392   NODE_NAME_CASE(SUB_VL)
13393   NODE_NAME_CASE(UDIV_VL)
13394   NODE_NAME_CASE(UREM_VL)
13395   NODE_NAME_CASE(XOR_VL)
13396   NODE_NAME_CASE(SADDSAT_VL)
13397   NODE_NAME_CASE(UADDSAT_VL)
13398   NODE_NAME_CASE(SSUBSAT_VL)
13399   NODE_NAME_CASE(USUBSAT_VL)
13400   NODE_NAME_CASE(FADD_VL)
13401   NODE_NAME_CASE(FSUB_VL)
13402   NODE_NAME_CASE(FMUL_VL)
13403   NODE_NAME_CASE(FDIV_VL)
13404   NODE_NAME_CASE(FNEG_VL)
13405   NODE_NAME_CASE(FABS_VL)
13406   NODE_NAME_CASE(FSQRT_VL)
13407   NODE_NAME_CASE(VFMADD_VL)
13408   NODE_NAME_CASE(VFNMADD_VL)
13409   NODE_NAME_CASE(VFMSUB_VL)
13410   NODE_NAME_CASE(VFNMSUB_VL)
13411   NODE_NAME_CASE(FCOPYSIGN_VL)
13412   NODE_NAME_CASE(SMIN_VL)
13413   NODE_NAME_CASE(SMAX_VL)
13414   NODE_NAME_CASE(UMIN_VL)
13415   NODE_NAME_CASE(UMAX_VL)
13416   NODE_NAME_CASE(FMINNUM_VL)
13417   NODE_NAME_CASE(FMAXNUM_VL)
13418   NODE_NAME_CASE(MULHS_VL)
13419   NODE_NAME_CASE(MULHU_VL)
13420   NODE_NAME_CASE(VFCVT_RTZ_X_F_VL)
13421   NODE_NAME_CASE(VFCVT_RTZ_XU_F_VL)
13422   NODE_NAME_CASE(VFCVT_RM_X_F_VL)
13423   NODE_NAME_CASE(VFCVT_RM_XU_F_VL)
13424   NODE_NAME_CASE(VFCVT_X_F_VL)
13425   NODE_NAME_CASE(VFCVT_XU_F_VL)
13426   NODE_NAME_CASE(VFROUND_NOEXCEPT_VL)
13427   NODE_NAME_CASE(SINT_TO_FP_VL)
13428   NODE_NAME_CASE(UINT_TO_FP_VL)
13429   NODE_NAME_CASE(VFCVT_RM_F_XU_VL)
13430   NODE_NAME_CASE(VFCVT_RM_F_X_VL)
13431   NODE_NAME_CASE(FP_EXTEND_VL)
13432   NODE_NAME_CASE(FP_ROUND_VL)
13433   NODE_NAME_CASE(VWMUL_VL)
13434   NODE_NAME_CASE(VWMULU_VL)
13435   NODE_NAME_CASE(VWMULSU_VL)
13436   NODE_NAME_CASE(VWADD_VL)
13437   NODE_NAME_CASE(VWADDU_VL)
13438   NODE_NAME_CASE(VWSUB_VL)
13439   NODE_NAME_CASE(VWSUBU_VL)
13440   NODE_NAME_CASE(VWADD_W_VL)
13441   NODE_NAME_CASE(VWADDU_W_VL)
13442   NODE_NAME_CASE(VWSUB_W_VL)
13443   NODE_NAME_CASE(VWSUBU_W_VL)
13444   NODE_NAME_CASE(VNSRL_VL)
13445   NODE_NAME_CASE(SETCC_VL)
13446   NODE_NAME_CASE(VSELECT_VL)
13447   NODE_NAME_CASE(VP_MERGE_VL)
13448   NODE_NAME_CASE(VMAND_VL)
13449   NODE_NAME_CASE(VMOR_VL)
13450   NODE_NAME_CASE(VMXOR_VL)
13451   NODE_NAME_CASE(VMCLR_VL)
13452   NODE_NAME_CASE(VMSET_VL)
13453   NODE_NAME_CASE(VRGATHER_VX_VL)
13454   NODE_NAME_CASE(VRGATHER_VV_VL)
13455   NODE_NAME_CASE(VRGATHEREI16_VV_VL)
13456   NODE_NAME_CASE(VSEXT_VL)
13457   NODE_NAME_CASE(VZEXT_VL)
13458   NODE_NAME_CASE(VCPOP_VL)
13459   NODE_NAME_CASE(VFIRST_VL)
13460   NODE_NAME_CASE(READ_CSR)
13461   NODE_NAME_CASE(WRITE_CSR)
13462   NODE_NAME_CASE(SWAP_CSR)
13463   }
13464   // clang-format on
13465   return nullptr;
13466 #undef NODE_NAME_CASE
13467 }
13468 
13469 /// getConstraintType - Given a constraint letter, return the type of
13470 /// constraint it is for this target.
13471 RISCVTargetLowering::ConstraintType
getConstraintType(StringRef Constraint) const13472 RISCVTargetLowering::getConstraintType(StringRef Constraint) const {
13473   if (Constraint.size() == 1) {
13474     switch (Constraint[0]) {
13475     default:
13476       break;
13477     case 'f':
13478       return C_RegisterClass;
13479     case 'I':
13480     case 'J':
13481     case 'K':
13482       return C_Immediate;
13483     case 'A':
13484       return C_Memory;
13485     case 'S': // A symbolic address
13486       return C_Other;
13487     }
13488   } else {
13489     if (Constraint == "vr" || Constraint == "vm")
13490       return C_RegisterClass;
13491   }
13492   return TargetLowering::getConstraintType(Constraint);
13493 }
13494 
13495 std::pair<unsigned, const TargetRegisterClass *>
getRegForInlineAsmConstraint(const TargetRegisterInfo * TRI,StringRef Constraint,MVT VT) const13496 RISCVTargetLowering::getRegForInlineAsmConstraint(const TargetRegisterInfo *TRI,
13497                                                   StringRef Constraint,
13498                                                   MVT VT) const {
13499   // First, see if this is a constraint that directly corresponds to a
13500   // RISCV register class.
13501   if (Constraint.size() == 1) {
13502     switch (Constraint[0]) {
13503     case 'r':
13504       // TODO: Support fixed vectors up to XLen for P extension?
13505       if (VT.isVector())
13506         break;
13507       return std::make_pair(0U, &RISCV::GPRRegClass);
13508     case 'f':
13509       if (Subtarget.hasStdExtZfhOrZfhmin() && VT == MVT::f16)
13510         return std::make_pair(0U, &RISCV::FPR16RegClass);
13511       if (Subtarget.hasStdExtF() && VT == MVT::f32)
13512         return std::make_pair(0U, &RISCV::FPR32RegClass);
13513       if (Subtarget.hasStdExtD() && VT == MVT::f64)
13514         return std::make_pair(0U, &RISCV::FPR64RegClass);
13515       break;
13516     default:
13517       break;
13518     }
13519   } else if (Constraint == "vr") {
13520     for (const auto *RC : {&RISCV::VRRegClass, &RISCV::VRM2RegClass,
13521                            &RISCV::VRM4RegClass, &RISCV::VRM8RegClass}) {
13522       if (TRI->isTypeLegalForClass(*RC, VT.SimpleTy))
13523         return std::make_pair(0U, RC);
13524     }
13525   } else if (Constraint == "vm") {
13526     if (TRI->isTypeLegalForClass(RISCV::VMV0RegClass, VT.SimpleTy))
13527       return std::make_pair(0U, &RISCV::VMV0RegClass);
13528   }
13529 
13530   // Clang will correctly decode the usage of register name aliases into their
13531   // official names. However, other frontends like `rustc` do not. This allows
13532   // users of these frontends to use the ABI names for registers in LLVM-style
13533   // register constraints.
13534   unsigned XRegFromAlias = StringSwitch<unsigned>(Constraint.lower())
13535                                .Case("{zero}", RISCV::X0)
13536                                .Case("{ra}", RISCV::X1)
13537                                .Case("{sp}", RISCV::X2)
13538                                .Case("{gp}", RISCV::X3)
13539                                .Case("{tp}", RISCV::X4)
13540                                .Case("{t0}", RISCV::X5)
13541                                .Case("{t1}", RISCV::X6)
13542                                .Case("{t2}", RISCV::X7)
13543                                .Cases("{s0}", "{fp}", RISCV::X8)
13544                                .Case("{s1}", RISCV::X9)
13545                                .Case("{a0}", RISCV::X10)
13546                                .Case("{a1}", RISCV::X11)
13547                                .Case("{a2}", RISCV::X12)
13548                                .Case("{a3}", RISCV::X13)
13549                                .Case("{a4}", RISCV::X14)
13550                                .Case("{a5}", RISCV::X15)
13551                                .Case("{a6}", RISCV::X16)
13552                                .Case("{a7}", RISCV::X17)
13553                                .Case("{s2}", RISCV::X18)
13554                                .Case("{s3}", RISCV::X19)
13555                                .Case("{s4}", RISCV::X20)
13556                                .Case("{s5}", RISCV::X21)
13557                                .Case("{s6}", RISCV::X22)
13558                                .Case("{s7}", RISCV::X23)
13559                                .Case("{s8}", RISCV::X24)
13560                                .Case("{s9}", RISCV::X25)
13561                                .Case("{s10}", RISCV::X26)
13562                                .Case("{s11}", RISCV::X27)
13563                                .Case("{t3}", RISCV::X28)
13564                                .Case("{t4}", RISCV::X29)
13565                                .Case("{t5}", RISCV::X30)
13566                                .Case("{t6}", RISCV::X31)
13567                                .Default(RISCV::NoRegister);
13568   if (XRegFromAlias != RISCV::NoRegister)
13569     return std::make_pair(XRegFromAlias, &RISCV::GPRRegClass);
13570 
13571   // Since TargetLowering::getRegForInlineAsmConstraint uses the name of the
13572   // TableGen record rather than the AsmName to choose registers for InlineAsm
13573   // constraints, plus we want to match those names to the widest floating point
13574   // register type available, manually select floating point registers here.
13575   //
13576   // The second case is the ABI name of the register, so that frontends can also
13577   // use the ABI names in register constraint lists.
13578   if (Subtarget.hasStdExtF()) {
13579     unsigned FReg = StringSwitch<unsigned>(Constraint.lower())
13580                         .Cases("{f0}", "{ft0}", RISCV::F0_F)
13581                         .Cases("{f1}", "{ft1}", RISCV::F1_F)
13582                         .Cases("{f2}", "{ft2}", RISCV::F2_F)
13583                         .Cases("{f3}", "{ft3}", RISCV::F3_F)
13584                         .Cases("{f4}", "{ft4}", RISCV::F4_F)
13585                         .Cases("{f5}", "{ft5}", RISCV::F5_F)
13586                         .Cases("{f6}", "{ft6}", RISCV::F6_F)
13587                         .Cases("{f7}", "{ft7}", RISCV::F7_F)
13588                         .Cases("{f8}", "{fs0}", RISCV::F8_F)
13589                         .Cases("{f9}", "{fs1}", RISCV::F9_F)
13590                         .Cases("{f10}", "{fa0}", RISCV::F10_F)
13591                         .Cases("{f11}", "{fa1}", RISCV::F11_F)
13592                         .Cases("{f12}", "{fa2}", RISCV::F12_F)
13593                         .Cases("{f13}", "{fa3}", RISCV::F13_F)
13594                         .Cases("{f14}", "{fa4}", RISCV::F14_F)
13595                         .Cases("{f15}", "{fa5}", RISCV::F15_F)
13596                         .Cases("{f16}", "{fa6}", RISCV::F16_F)
13597                         .Cases("{f17}", "{fa7}", RISCV::F17_F)
13598                         .Cases("{f18}", "{fs2}", RISCV::F18_F)
13599                         .Cases("{f19}", "{fs3}", RISCV::F19_F)
13600                         .Cases("{f20}", "{fs4}", RISCV::F20_F)
13601                         .Cases("{f21}", "{fs5}", RISCV::F21_F)
13602                         .Cases("{f22}", "{fs6}", RISCV::F22_F)
13603                         .Cases("{f23}", "{fs7}", RISCV::F23_F)
13604                         .Cases("{f24}", "{fs8}", RISCV::F24_F)
13605                         .Cases("{f25}", "{fs9}", RISCV::F25_F)
13606                         .Cases("{f26}", "{fs10}", RISCV::F26_F)
13607                         .Cases("{f27}", "{fs11}", RISCV::F27_F)
13608                         .Cases("{f28}", "{ft8}", RISCV::F28_F)
13609                         .Cases("{f29}", "{ft9}", RISCV::F29_F)
13610                         .Cases("{f30}", "{ft10}", RISCV::F30_F)
13611                         .Cases("{f31}", "{ft11}", RISCV::F31_F)
13612                         .Default(RISCV::NoRegister);
13613     if (FReg != RISCV::NoRegister) {
13614       assert(RISCV::F0_F <= FReg && FReg <= RISCV::F31_F && "Unknown fp-reg");
13615       if (Subtarget.hasStdExtD() && (VT == MVT::f64 || VT == MVT::Other)) {
13616         unsigned RegNo = FReg - RISCV::F0_F;
13617         unsigned DReg = RISCV::F0_D + RegNo;
13618         return std::make_pair(DReg, &RISCV::FPR64RegClass);
13619       }
13620       if (VT == MVT::f32 || VT == MVT::Other)
13621         return std::make_pair(FReg, &RISCV::FPR32RegClass);
13622       if (Subtarget.hasStdExtZfhOrZfhmin() && VT == MVT::f16) {
13623         unsigned RegNo = FReg - RISCV::F0_F;
13624         unsigned HReg = RISCV::F0_H + RegNo;
13625         return std::make_pair(HReg, &RISCV::FPR16RegClass);
13626       }
13627     }
13628   }
13629 
13630   if (Subtarget.hasVInstructions()) {
13631     Register VReg = StringSwitch<Register>(Constraint.lower())
13632                         .Case("{v0}", RISCV::V0)
13633                         .Case("{v1}", RISCV::V1)
13634                         .Case("{v2}", RISCV::V2)
13635                         .Case("{v3}", RISCV::V3)
13636                         .Case("{v4}", RISCV::V4)
13637                         .Case("{v5}", RISCV::V5)
13638                         .Case("{v6}", RISCV::V6)
13639                         .Case("{v7}", RISCV::V7)
13640                         .Case("{v8}", RISCV::V8)
13641                         .Case("{v9}", RISCV::V9)
13642                         .Case("{v10}", RISCV::V10)
13643                         .Case("{v11}", RISCV::V11)
13644                         .Case("{v12}", RISCV::V12)
13645                         .Case("{v13}", RISCV::V13)
13646                         .Case("{v14}", RISCV::V14)
13647                         .Case("{v15}", RISCV::V15)
13648                         .Case("{v16}", RISCV::V16)
13649                         .Case("{v17}", RISCV::V17)
13650                         .Case("{v18}", RISCV::V18)
13651                         .Case("{v19}", RISCV::V19)
13652                         .Case("{v20}", RISCV::V20)
13653                         .Case("{v21}", RISCV::V21)
13654                         .Case("{v22}", RISCV::V22)
13655                         .Case("{v23}", RISCV::V23)
13656                         .Case("{v24}", RISCV::V24)
13657                         .Case("{v25}", RISCV::V25)
13658                         .Case("{v26}", RISCV::V26)
13659                         .Case("{v27}", RISCV::V27)
13660                         .Case("{v28}", RISCV::V28)
13661                         .Case("{v29}", RISCV::V29)
13662                         .Case("{v30}", RISCV::V30)
13663                         .Case("{v31}", RISCV::V31)
13664                         .Default(RISCV::NoRegister);
13665     if (VReg != RISCV::NoRegister) {
13666       if (TRI->isTypeLegalForClass(RISCV::VMRegClass, VT.SimpleTy))
13667         return std::make_pair(VReg, &RISCV::VMRegClass);
13668       if (TRI->isTypeLegalForClass(RISCV::VRRegClass, VT.SimpleTy))
13669         return std::make_pair(VReg, &RISCV::VRRegClass);
13670       for (const auto *RC :
13671            {&RISCV::VRM2RegClass, &RISCV::VRM4RegClass, &RISCV::VRM8RegClass}) {
13672         if (TRI->isTypeLegalForClass(*RC, VT.SimpleTy)) {
13673           VReg = TRI->getMatchingSuperReg(VReg, RISCV::sub_vrm1_0, RC);
13674           return std::make_pair(VReg, RC);
13675         }
13676       }
13677     }
13678   }
13679 
13680   std::pair<Register, const TargetRegisterClass *> Res =
13681       TargetLowering::getRegForInlineAsmConstraint(TRI, Constraint, VT);
13682 
13683   // If we picked one of the Zfinx register classes, remap it to the GPR class.
13684   // FIXME: When Zfinx is supported in CodeGen this will need to take the
13685   // Subtarget into account.
13686   if (Res.second == &RISCV::GPRF16RegClass ||
13687       Res.second == &RISCV::GPRF32RegClass ||
13688       Res.second == &RISCV::GPRF64RegClass)
13689     return std::make_pair(Res.first, &RISCV::GPRRegClass);
13690 
13691   return Res;
13692 }
13693 
13694 unsigned
getInlineAsmMemConstraint(StringRef ConstraintCode) const13695 RISCVTargetLowering::getInlineAsmMemConstraint(StringRef ConstraintCode) const {
13696   // Currently only support length 1 constraints.
13697   if (ConstraintCode.size() == 1) {
13698     switch (ConstraintCode[0]) {
13699     case 'A':
13700       return InlineAsm::Constraint_A;
13701     default:
13702       break;
13703     }
13704   }
13705 
13706   return TargetLowering::getInlineAsmMemConstraint(ConstraintCode);
13707 }
13708 
LowerAsmOperandForConstraint(SDValue Op,std::string & Constraint,std::vector<SDValue> & Ops,SelectionDAG & DAG) const13709 void RISCVTargetLowering::LowerAsmOperandForConstraint(
13710     SDValue Op, std::string &Constraint, std::vector<SDValue> &Ops,
13711     SelectionDAG &DAG) const {
13712   // Currently only support length 1 constraints.
13713   if (Constraint.length() == 1) {
13714     switch (Constraint[0]) {
13715     case 'I':
13716       // Validate & create a 12-bit signed immediate operand.
13717       if (auto *C = dyn_cast<ConstantSDNode>(Op)) {
13718         uint64_t CVal = C->getSExtValue();
13719         if (isInt<12>(CVal))
13720           Ops.push_back(
13721               DAG.getTargetConstant(CVal, SDLoc(Op), Subtarget.getXLenVT()));
13722       }
13723       return;
13724     case 'J':
13725       // Validate & create an integer zero operand.
13726       if (auto *C = dyn_cast<ConstantSDNode>(Op))
13727         if (C->getZExtValue() == 0)
13728           Ops.push_back(
13729               DAG.getTargetConstant(0, SDLoc(Op), Subtarget.getXLenVT()));
13730       return;
13731     case 'K':
13732       // Validate & create a 5-bit unsigned immediate operand.
13733       if (auto *C = dyn_cast<ConstantSDNode>(Op)) {
13734         uint64_t CVal = C->getZExtValue();
13735         if (isUInt<5>(CVal))
13736           Ops.push_back(
13737               DAG.getTargetConstant(CVal, SDLoc(Op), Subtarget.getXLenVT()));
13738       }
13739       return;
13740     case 'S':
13741       if (const auto *GA = dyn_cast<GlobalAddressSDNode>(Op)) {
13742         Ops.push_back(DAG.getTargetGlobalAddress(GA->getGlobal(), SDLoc(Op),
13743                                                  GA->getValueType(0)));
13744       } else if (const auto *BA = dyn_cast<BlockAddressSDNode>(Op)) {
13745         Ops.push_back(DAG.getTargetBlockAddress(BA->getBlockAddress(),
13746                                                 BA->getValueType(0)));
13747       }
13748       return;
13749     default:
13750       break;
13751     }
13752   }
13753   TargetLowering::LowerAsmOperandForConstraint(Op, Constraint, Ops, DAG);
13754 }
13755 
emitLeadingFence(IRBuilderBase & Builder,Instruction * Inst,AtomicOrdering Ord) const13756 Instruction *RISCVTargetLowering::emitLeadingFence(IRBuilderBase &Builder,
13757                                                    Instruction *Inst,
13758                                                    AtomicOrdering Ord) const {
13759   if (isa<LoadInst>(Inst) && Ord == AtomicOrdering::SequentiallyConsistent)
13760     return Builder.CreateFence(Ord);
13761   if (isa<StoreInst>(Inst) && isReleaseOrStronger(Ord))
13762     return Builder.CreateFence(AtomicOrdering::Release);
13763   return nullptr;
13764 }
13765 
emitTrailingFence(IRBuilderBase & Builder,Instruction * Inst,AtomicOrdering Ord) const13766 Instruction *RISCVTargetLowering::emitTrailingFence(IRBuilderBase &Builder,
13767                                                     Instruction *Inst,
13768                                                     AtomicOrdering Ord) const {
13769   if (isa<LoadInst>(Inst) && isAcquireOrStronger(Ord))
13770     return Builder.CreateFence(AtomicOrdering::Acquire);
13771   return nullptr;
13772 }
13773 
13774 TargetLowering::AtomicExpansionKind
shouldExpandAtomicRMWInIR(AtomicRMWInst * AI) const13775 RISCVTargetLowering::shouldExpandAtomicRMWInIR(AtomicRMWInst *AI) const {
13776   // atomicrmw {fadd,fsub} must be expanded to use compare-exchange, as floating
13777   // point operations can't be used in an lr/sc sequence without breaking the
13778   // forward-progress guarantee.
13779   if (AI->isFloatingPointOperation() ||
13780       AI->getOperation() == AtomicRMWInst::UIncWrap ||
13781       AI->getOperation() == AtomicRMWInst::UDecWrap)
13782     return AtomicExpansionKind::CmpXChg;
13783 
13784   // Don't expand forced atomics, we want to have __sync libcalls instead.
13785   if (Subtarget.hasForcedAtomics())
13786     return AtomicExpansionKind::None;
13787 
13788   unsigned Size = AI->getType()->getPrimitiveSizeInBits();
13789   if (Size == 8 || Size == 16)
13790     return AtomicExpansionKind::MaskedIntrinsic;
13791   return AtomicExpansionKind::None;
13792 }
13793 
13794 static Intrinsic::ID
getIntrinsicForMaskedAtomicRMWBinOp(unsigned XLen,AtomicRMWInst::BinOp BinOp)13795 getIntrinsicForMaskedAtomicRMWBinOp(unsigned XLen, AtomicRMWInst::BinOp BinOp) {
13796   if (XLen == 32) {
13797     switch (BinOp) {
13798     default:
13799       llvm_unreachable("Unexpected AtomicRMW BinOp");
13800     case AtomicRMWInst::Xchg:
13801       return Intrinsic::riscv_masked_atomicrmw_xchg_i32;
13802     case AtomicRMWInst::Add:
13803       return Intrinsic::riscv_masked_atomicrmw_add_i32;
13804     case AtomicRMWInst::Sub:
13805       return Intrinsic::riscv_masked_atomicrmw_sub_i32;
13806     case AtomicRMWInst::Nand:
13807       return Intrinsic::riscv_masked_atomicrmw_nand_i32;
13808     case AtomicRMWInst::Max:
13809       return Intrinsic::riscv_masked_atomicrmw_max_i32;
13810     case AtomicRMWInst::Min:
13811       return Intrinsic::riscv_masked_atomicrmw_min_i32;
13812     case AtomicRMWInst::UMax:
13813       return Intrinsic::riscv_masked_atomicrmw_umax_i32;
13814     case AtomicRMWInst::UMin:
13815       return Intrinsic::riscv_masked_atomicrmw_umin_i32;
13816     }
13817   }
13818 
13819   if (XLen == 64) {
13820     switch (BinOp) {
13821     default:
13822       llvm_unreachable("Unexpected AtomicRMW BinOp");
13823     case AtomicRMWInst::Xchg:
13824       return Intrinsic::riscv_masked_atomicrmw_xchg_i64;
13825     case AtomicRMWInst::Add:
13826       return Intrinsic::riscv_masked_atomicrmw_add_i64;
13827     case AtomicRMWInst::Sub:
13828       return Intrinsic::riscv_masked_atomicrmw_sub_i64;
13829     case AtomicRMWInst::Nand:
13830       return Intrinsic::riscv_masked_atomicrmw_nand_i64;
13831     case AtomicRMWInst::Max:
13832       return Intrinsic::riscv_masked_atomicrmw_max_i64;
13833     case AtomicRMWInst::Min:
13834       return Intrinsic::riscv_masked_atomicrmw_min_i64;
13835     case AtomicRMWInst::UMax:
13836       return Intrinsic::riscv_masked_atomicrmw_umax_i64;
13837     case AtomicRMWInst::UMin:
13838       return Intrinsic::riscv_masked_atomicrmw_umin_i64;
13839     }
13840   }
13841 
13842   llvm_unreachable("Unexpected XLen\n");
13843 }
13844 
emitMaskedAtomicRMWIntrinsic(IRBuilderBase & Builder,AtomicRMWInst * AI,Value * AlignedAddr,Value * Incr,Value * Mask,Value * ShiftAmt,AtomicOrdering Ord) const13845 Value *RISCVTargetLowering::emitMaskedAtomicRMWIntrinsic(
13846     IRBuilderBase &Builder, AtomicRMWInst *AI, Value *AlignedAddr, Value *Incr,
13847     Value *Mask, Value *ShiftAmt, AtomicOrdering Ord) const {
13848   unsigned XLen = Subtarget.getXLen();
13849   Value *Ordering =
13850       Builder.getIntN(XLen, static_cast<uint64_t>(AI->getOrdering()));
13851   Type *Tys[] = {AlignedAddr->getType()};
13852   Function *LrwOpScwLoop = Intrinsic::getDeclaration(
13853       AI->getModule(),
13854       getIntrinsicForMaskedAtomicRMWBinOp(XLen, AI->getOperation()), Tys);
13855 
13856   if (XLen == 64) {
13857     Incr = Builder.CreateSExt(Incr, Builder.getInt64Ty());
13858     Mask = Builder.CreateSExt(Mask, Builder.getInt64Ty());
13859     ShiftAmt = Builder.CreateSExt(ShiftAmt, Builder.getInt64Ty());
13860   }
13861 
13862   Value *Result;
13863 
13864   // Must pass the shift amount needed to sign extend the loaded value prior
13865   // to performing a signed comparison for min/max. ShiftAmt is the number of
13866   // bits to shift the value into position. Pass XLen-ShiftAmt-ValWidth, which
13867   // is the number of bits to left+right shift the value in order to
13868   // sign-extend.
13869   if (AI->getOperation() == AtomicRMWInst::Min ||
13870       AI->getOperation() == AtomicRMWInst::Max) {
13871     const DataLayout &DL = AI->getModule()->getDataLayout();
13872     unsigned ValWidth =
13873         DL.getTypeStoreSizeInBits(AI->getValOperand()->getType());
13874     Value *SextShamt =
13875         Builder.CreateSub(Builder.getIntN(XLen, XLen - ValWidth), ShiftAmt);
13876     Result = Builder.CreateCall(LrwOpScwLoop,
13877                                 {AlignedAddr, Incr, Mask, SextShamt, Ordering});
13878   } else {
13879     Result =
13880         Builder.CreateCall(LrwOpScwLoop, {AlignedAddr, Incr, Mask, Ordering});
13881   }
13882 
13883   if (XLen == 64)
13884     Result = Builder.CreateTrunc(Result, Builder.getInt32Ty());
13885   return Result;
13886 }
13887 
13888 TargetLowering::AtomicExpansionKind
shouldExpandAtomicCmpXchgInIR(AtomicCmpXchgInst * CI) const13889 RISCVTargetLowering::shouldExpandAtomicCmpXchgInIR(
13890     AtomicCmpXchgInst *CI) const {
13891   // Don't expand forced atomics, we want to have __sync libcalls instead.
13892   if (Subtarget.hasForcedAtomics())
13893     return AtomicExpansionKind::None;
13894 
13895   unsigned Size = CI->getCompareOperand()->getType()->getPrimitiveSizeInBits();
13896   if (Size == 8 || Size == 16)
13897     return AtomicExpansionKind::MaskedIntrinsic;
13898   return AtomicExpansionKind::None;
13899 }
13900 
emitMaskedAtomicCmpXchgIntrinsic(IRBuilderBase & Builder,AtomicCmpXchgInst * CI,Value * AlignedAddr,Value * CmpVal,Value * NewVal,Value * Mask,AtomicOrdering Ord) const13901 Value *RISCVTargetLowering::emitMaskedAtomicCmpXchgIntrinsic(
13902     IRBuilderBase &Builder, AtomicCmpXchgInst *CI, Value *AlignedAddr,
13903     Value *CmpVal, Value *NewVal, Value *Mask, AtomicOrdering Ord) const {
13904   unsigned XLen = Subtarget.getXLen();
13905   Value *Ordering = Builder.getIntN(XLen, static_cast<uint64_t>(Ord));
13906   Intrinsic::ID CmpXchgIntrID = Intrinsic::riscv_masked_cmpxchg_i32;
13907   if (XLen == 64) {
13908     CmpVal = Builder.CreateSExt(CmpVal, Builder.getInt64Ty());
13909     NewVal = Builder.CreateSExt(NewVal, Builder.getInt64Ty());
13910     Mask = Builder.CreateSExt(Mask, Builder.getInt64Ty());
13911     CmpXchgIntrID = Intrinsic::riscv_masked_cmpxchg_i64;
13912   }
13913   Type *Tys[] = {AlignedAddr->getType()};
13914   Function *MaskedCmpXchg =
13915       Intrinsic::getDeclaration(CI->getModule(), CmpXchgIntrID, Tys);
13916   Value *Result = Builder.CreateCall(
13917       MaskedCmpXchg, {AlignedAddr, CmpVal, NewVal, Mask, Ordering});
13918   if (XLen == 64)
13919     Result = Builder.CreateTrunc(Result, Builder.getInt32Ty());
13920   return Result;
13921 }
13922 
shouldRemoveExtendFromGSIndex(EVT IndexVT,EVT DataVT) const13923 bool RISCVTargetLowering::shouldRemoveExtendFromGSIndex(EVT IndexVT,
13924                                                         EVT DataVT) const {
13925   return false;
13926 }
13927 
shouldConvertFpToSat(unsigned Op,EVT FPVT,EVT VT) const13928 bool RISCVTargetLowering::shouldConvertFpToSat(unsigned Op, EVT FPVT,
13929                                                EVT VT) const {
13930   if (!isOperationLegalOrCustom(Op, VT) || !FPVT.isSimple())
13931     return false;
13932 
13933   switch (FPVT.getSimpleVT().SimpleTy) {
13934   case MVT::f16:
13935     return Subtarget.hasStdExtZfhOrZfhmin();
13936   case MVT::f32:
13937     return Subtarget.hasStdExtF();
13938   case MVT::f64:
13939     return Subtarget.hasStdExtD();
13940   default:
13941     return false;
13942   }
13943 }
13944 
getJumpTableEncoding() const13945 unsigned RISCVTargetLowering::getJumpTableEncoding() const {
13946   // If we are using the small code model, we can reduce size of jump table
13947   // entry to 4 bytes.
13948   if (Subtarget.is64Bit() && !isPositionIndependent() &&
13949       getTargetMachine().getCodeModel() == CodeModel::Small) {
13950     return MachineJumpTableInfo::EK_Custom32;
13951   }
13952   return TargetLowering::getJumpTableEncoding();
13953 }
13954 
LowerCustomJumpTableEntry(const MachineJumpTableInfo * MJTI,const MachineBasicBlock * MBB,unsigned uid,MCContext & Ctx) const13955 const MCExpr *RISCVTargetLowering::LowerCustomJumpTableEntry(
13956     const MachineJumpTableInfo *MJTI, const MachineBasicBlock *MBB,
13957     unsigned uid, MCContext &Ctx) const {
13958   assert(Subtarget.is64Bit() && !isPositionIndependent() &&
13959          getTargetMachine().getCodeModel() == CodeModel::Small);
13960   return MCSymbolRefExpr::create(MBB->getSymbol(), Ctx);
13961 }
13962 
isVScaleKnownToBeAPowerOfTwo() const13963 bool RISCVTargetLowering::isVScaleKnownToBeAPowerOfTwo() const {
13964   // We define vscale to be VLEN/RVVBitsPerBlock.  VLEN is always a power
13965   // of two >= 64, and RVVBitsPerBlock is 64.  Thus, vscale must be
13966   // a power of two as well.
13967   // FIXME: This doesn't work for zve32, but that's already broken
13968   // elsewhere for the same reason.
13969   assert(Subtarget.getRealMinVLen() >= 64 && "zve32* unsupported");
13970   static_assert(RISCV::RVVBitsPerBlock == 64,
13971                 "RVVBitsPerBlock changed, audit needed");
13972   return true;
13973 }
13974 
isFMAFasterThanFMulAndFAdd(const MachineFunction & MF,EVT VT) const13975 bool RISCVTargetLowering::isFMAFasterThanFMulAndFAdd(const MachineFunction &MF,
13976                                                      EVT VT) const {
13977   EVT SVT = VT.getScalarType();
13978 
13979   if (!SVT.isSimple())
13980     return false;
13981 
13982   switch (SVT.getSimpleVT().SimpleTy) {
13983   case MVT::f16:
13984     return VT.isVector() ? Subtarget.hasVInstructionsF16()
13985                          : Subtarget.hasStdExtZfh();
13986   case MVT::f32:
13987     return Subtarget.hasStdExtF();
13988   case MVT::f64:
13989     return Subtarget.hasStdExtD();
13990   default:
13991     break;
13992   }
13993 
13994   return false;
13995 }
13996 
getExceptionPointerRegister(const Constant * PersonalityFn) const13997 Register RISCVTargetLowering::getExceptionPointerRegister(
13998     const Constant *PersonalityFn) const {
13999   return RISCV::X10;
14000 }
14001 
getExceptionSelectorRegister(const Constant * PersonalityFn) const14002 Register RISCVTargetLowering::getExceptionSelectorRegister(
14003     const Constant *PersonalityFn) const {
14004   return RISCV::X11;
14005 }
14006 
shouldExtendTypeInLibCall(EVT Type) const14007 bool RISCVTargetLowering::shouldExtendTypeInLibCall(EVT Type) const {
14008   // Return false to suppress the unnecessary extensions if the LibCall
14009   // arguments or return value is f32 type for LP64 ABI.
14010   RISCVABI::ABI ABI = Subtarget.getTargetABI();
14011   if (ABI == RISCVABI::ABI_LP64 && (Type == MVT::f32))
14012     return false;
14013 
14014   return true;
14015 }
14016 
shouldSignExtendTypeInLibCall(EVT Type,bool IsSigned) const14017 bool RISCVTargetLowering::shouldSignExtendTypeInLibCall(EVT Type, bool IsSigned) const {
14018   if (Subtarget.is64Bit() && Type == MVT::i32)
14019     return true;
14020 
14021   return IsSigned;
14022 }
14023 
decomposeMulByConstant(LLVMContext & Context,EVT VT,SDValue C) const14024 bool RISCVTargetLowering::decomposeMulByConstant(LLVMContext &Context, EVT VT,
14025                                                  SDValue C) const {
14026   // Check integral scalar types.
14027   const bool HasExtMOrZmmul =
14028       Subtarget.hasStdExtM() || Subtarget.hasStdExtZmmul();
14029   if (VT.isScalarInteger()) {
14030     // Omit the optimization if the sub target has the M extension and the data
14031     // size exceeds XLen.
14032     if (HasExtMOrZmmul && VT.getSizeInBits() > Subtarget.getXLen())
14033       return false;
14034     if (auto *ConstNode = dyn_cast<ConstantSDNode>(C.getNode())) {
14035       // Break the MUL to a SLLI and an ADD/SUB.
14036       const APInt &Imm = ConstNode->getAPIntValue();
14037       if ((Imm + 1).isPowerOf2() || (Imm - 1).isPowerOf2() ||
14038           (1 - Imm).isPowerOf2() || (-1 - Imm).isPowerOf2())
14039         return true;
14040       // Optimize the MUL to (SH*ADD x, (SLLI x, bits)) if Imm is not simm12.
14041       if (Subtarget.hasStdExtZba() && !Imm.isSignedIntN(12) &&
14042           ((Imm - 2).isPowerOf2() || (Imm - 4).isPowerOf2() ||
14043            (Imm - 8).isPowerOf2()))
14044         return true;
14045       // Omit the following optimization if the sub target has the M extension
14046       // and the data size >= XLen.
14047       if (HasExtMOrZmmul && VT.getSizeInBits() >= Subtarget.getXLen())
14048         return false;
14049       // Break the MUL to two SLLI instructions and an ADD/SUB, if Imm needs
14050       // a pair of LUI/ADDI.
14051       if (!Imm.isSignedIntN(12) && Imm.countTrailingZeros() < 12) {
14052         APInt ImmS = Imm.ashr(Imm.countTrailingZeros());
14053         if ((ImmS + 1).isPowerOf2() || (ImmS - 1).isPowerOf2() ||
14054             (1 - ImmS).isPowerOf2())
14055           return true;
14056       }
14057     }
14058   }
14059 
14060   return false;
14061 }
14062 
isMulAddWithConstProfitable(SDValue AddNode,SDValue ConstNode) const14063 bool RISCVTargetLowering::isMulAddWithConstProfitable(SDValue AddNode,
14064                                                       SDValue ConstNode) const {
14065   // Let the DAGCombiner decide for vectors.
14066   EVT VT = AddNode.getValueType();
14067   if (VT.isVector())
14068     return true;
14069 
14070   // Let the DAGCombiner decide for larger types.
14071   if (VT.getScalarSizeInBits() > Subtarget.getXLen())
14072     return true;
14073 
14074   // It is worse if c1 is simm12 while c1*c2 is not.
14075   ConstantSDNode *C1Node = cast<ConstantSDNode>(AddNode.getOperand(1));
14076   ConstantSDNode *C2Node = cast<ConstantSDNode>(ConstNode);
14077   const APInt &C1 = C1Node->getAPIntValue();
14078   const APInt &C2 = C2Node->getAPIntValue();
14079   if (C1.isSignedIntN(12) && !(C1 * C2).isSignedIntN(12))
14080     return false;
14081 
14082   // Default to true and let the DAGCombiner decide.
14083   return true;
14084 }
14085 
allowsMisalignedMemoryAccesses(EVT VT,unsigned AddrSpace,Align Alignment,MachineMemOperand::Flags Flags,unsigned * Fast) const14086 bool RISCVTargetLowering::allowsMisalignedMemoryAccesses(
14087     EVT VT, unsigned AddrSpace, Align Alignment, MachineMemOperand::Flags Flags,
14088     unsigned *Fast) const {
14089   if (!VT.isVector()) {
14090     if (Fast)
14091       *Fast = 0;
14092     return Subtarget.enableUnalignedScalarMem();
14093   }
14094 
14095   // All vector implementations must support element alignment
14096   EVT ElemVT = VT.getVectorElementType();
14097   if (Alignment >= ElemVT.getStoreSize()) {
14098     if (Fast)
14099       *Fast = 1;
14100     return true;
14101   }
14102 
14103   return false;
14104 }
14105 
splitValueIntoRegisterParts(SelectionDAG & DAG,const SDLoc & DL,SDValue Val,SDValue * Parts,unsigned NumParts,MVT PartVT,std::optional<CallingConv::ID> CC) const14106 bool RISCVTargetLowering::splitValueIntoRegisterParts(
14107     SelectionDAG &DAG, const SDLoc &DL, SDValue Val, SDValue *Parts,
14108     unsigned NumParts, MVT PartVT, std::optional<CallingConv::ID> CC) const {
14109   bool IsABIRegCopy = CC.has_value();
14110   EVT ValueVT = Val.getValueType();
14111   if (IsABIRegCopy && ValueVT == MVT::f16 && PartVT == MVT::f32) {
14112     // Cast the f16 to i16, extend to i32, pad with ones to make a float nan,
14113     // and cast to f32.
14114     Val = DAG.getNode(ISD::BITCAST, DL, MVT::i16, Val);
14115     Val = DAG.getNode(ISD::ANY_EXTEND, DL, MVT::i32, Val);
14116     Val = DAG.getNode(ISD::OR, DL, MVT::i32, Val,
14117                       DAG.getConstant(0xFFFF0000, DL, MVT::i32));
14118     Val = DAG.getNode(ISD::BITCAST, DL, MVT::f32, Val);
14119     Parts[0] = Val;
14120     return true;
14121   }
14122 
14123   if (ValueVT.isScalableVector() && PartVT.isScalableVector()) {
14124     LLVMContext &Context = *DAG.getContext();
14125     EVT ValueEltVT = ValueVT.getVectorElementType();
14126     EVT PartEltVT = PartVT.getVectorElementType();
14127     unsigned ValueVTBitSize = ValueVT.getSizeInBits().getKnownMinValue();
14128     unsigned PartVTBitSize = PartVT.getSizeInBits().getKnownMinValue();
14129     if (PartVTBitSize % ValueVTBitSize == 0) {
14130       assert(PartVTBitSize >= ValueVTBitSize);
14131       // If the element types are different, bitcast to the same element type of
14132       // PartVT first.
14133       // Give an example here, we want copy a <vscale x 1 x i8> value to
14134       // <vscale x 4 x i16>.
14135       // We need to convert <vscale x 1 x i8> to <vscale x 8 x i8> by insert
14136       // subvector, then we can bitcast to <vscale x 4 x i16>.
14137       if (ValueEltVT != PartEltVT) {
14138         if (PartVTBitSize > ValueVTBitSize) {
14139           unsigned Count = PartVTBitSize / ValueEltVT.getFixedSizeInBits();
14140           assert(Count != 0 && "The number of element should not be zero.");
14141           EVT SameEltTypeVT =
14142               EVT::getVectorVT(Context, ValueEltVT, Count, /*IsScalable=*/true);
14143           Val = DAG.getNode(ISD::INSERT_SUBVECTOR, DL, SameEltTypeVT,
14144                             DAG.getUNDEF(SameEltTypeVT), Val,
14145                             DAG.getVectorIdxConstant(0, DL));
14146         }
14147         Val = DAG.getNode(ISD::BITCAST, DL, PartVT, Val);
14148       } else {
14149         Val =
14150             DAG.getNode(ISD::INSERT_SUBVECTOR, DL, PartVT, DAG.getUNDEF(PartVT),
14151                         Val, DAG.getVectorIdxConstant(0, DL));
14152       }
14153       Parts[0] = Val;
14154       return true;
14155     }
14156   }
14157   return false;
14158 }
14159 
joinRegisterPartsIntoValue(SelectionDAG & DAG,const SDLoc & DL,const SDValue * Parts,unsigned NumParts,MVT PartVT,EVT ValueVT,std::optional<CallingConv::ID> CC) const14160 SDValue RISCVTargetLowering::joinRegisterPartsIntoValue(
14161     SelectionDAG &DAG, const SDLoc &DL, const SDValue *Parts, unsigned NumParts,
14162     MVT PartVT, EVT ValueVT, std::optional<CallingConv::ID> CC) const {
14163   bool IsABIRegCopy = CC.has_value();
14164   if (IsABIRegCopy && ValueVT == MVT::f16 && PartVT == MVT::f32) {
14165     SDValue Val = Parts[0];
14166 
14167     // Cast the f32 to i32, truncate to i16, and cast back to f16.
14168     Val = DAG.getNode(ISD::BITCAST, DL, MVT::i32, Val);
14169     Val = DAG.getNode(ISD::TRUNCATE, DL, MVT::i16, Val);
14170     Val = DAG.getNode(ISD::BITCAST, DL, MVT::f16, Val);
14171     return Val;
14172   }
14173 
14174   if (ValueVT.isScalableVector() && PartVT.isScalableVector()) {
14175     LLVMContext &Context = *DAG.getContext();
14176     SDValue Val = Parts[0];
14177     EVT ValueEltVT = ValueVT.getVectorElementType();
14178     EVT PartEltVT = PartVT.getVectorElementType();
14179     unsigned ValueVTBitSize = ValueVT.getSizeInBits().getKnownMinValue();
14180     unsigned PartVTBitSize = PartVT.getSizeInBits().getKnownMinValue();
14181     if (PartVTBitSize % ValueVTBitSize == 0) {
14182       assert(PartVTBitSize >= ValueVTBitSize);
14183       EVT SameEltTypeVT = ValueVT;
14184       // If the element types are different, convert it to the same element type
14185       // of PartVT.
14186       // Give an example here, we want copy a <vscale x 1 x i8> value from
14187       // <vscale x 4 x i16>.
14188       // We need to convert <vscale x 4 x i16> to <vscale x 8 x i8> first,
14189       // then we can extract <vscale x 1 x i8>.
14190       if (ValueEltVT != PartEltVT) {
14191         unsigned Count = PartVTBitSize / ValueEltVT.getFixedSizeInBits();
14192         assert(Count != 0 && "The number of element should not be zero.");
14193         SameEltTypeVT =
14194             EVT::getVectorVT(Context, ValueEltVT, Count, /*IsScalable=*/true);
14195         Val = DAG.getNode(ISD::BITCAST, DL, SameEltTypeVT, Val);
14196       }
14197       Val = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, ValueVT, Val,
14198                         DAG.getVectorIdxConstant(0, DL));
14199       return Val;
14200     }
14201   }
14202   return SDValue();
14203 }
14204 
isIntDivCheap(EVT VT,AttributeList Attr) const14205 bool RISCVTargetLowering::isIntDivCheap(EVT VT, AttributeList Attr) const {
14206   // When aggressively optimizing for code size, we prefer to use a div
14207   // instruction, as it is usually smaller than the alternative sequence.
14208   // TODO: Add vector division?
14209   bool OptSize = Attr.hasFnAttr(Attribute::MinSize);
14210   return OptSize && !VT.isVector();
14211 }
14212 
preferScalarizeSplat(unsigned Opc) const14213 bool RISCVTargetLowering::preferScalarizeSplat(unsigned Opc) const {
14214   // Scalarize zero_ext and sign_ext might stop match to widening instruction in
14215   // some situation.
14216   if (Opc == ISD::ZERO_EXTEND || Opc == ISD::SIGN_EXTEND)
14217     return false;
14218   return true;
14219 }
14220 
useTpOffset(IRBuilderBase & IRB,unsigned Offset)14221 static Value *useTpOffset(IRBuilderBase &IRB, unsigned Offset) {
14222   Module *M = IRB.GetInsertBlock()->getParent()->getParent();
14223   Function *ThreadPointerFunc =
14224       Intrinsic::getDeclaration(M, Intrinsic::thread_pointer);
14225   return IRB.CreatePointerCast(
14226       IRB.CreateConstGEP1_32(IRB.getInt8Ty(),
14227                              IRB.CreateCall(ThreadPointerFunc), Offset),
14228       IRB.getInt8PtrTy()->getPointerTo(0));
14229 }
14230 
getIRStackGuard(IRBuilderBase & IRB) const14231 Value *RISCVTargetLowering::getIRStackGuard(IRBuilderBase &IRB) const {
14232   // Fuchsia provides a fixed TLS slot for the stack cookie.
14233   // <zircon/tls.h> defines ZX_TLS_STACK_GUARD_OFFSET with this value.
14234   if (Subtarget.isTargetFuchsia())
14235     return useTpOffset(IRB, -0x10);
14236 
14237   return TargetLowering::getIRStackGuard(IRB);
14238 }
14239 
14240 #define GET_REGISTER_MATCHER
14241 #include "RISCVGenAsmMatcher.inc"
14242 
14243 Register
getRegisterByName(const char * RegName,LLT VT,const MachineFunction & MF) const14244 RISCVTargetLowering::getRegisterByName(const char *RegName, LLT VT,
14245                                        const MachineFunction &MF) const {
14246   Register Reg = MatchRegisterAltName(RegName);
14247   if (Reg == RISCV::NoRegister)
14248     Reg = MatchRegisterName(RegName);
14249   if (Reg == RISCV::NoRegister)
14250     report_fatal_error(
14251         Twine("Invalid register name \"" + StringRef(RegName) + "\"."));
14252   BitVector ReservedRegs = Subtarget.getRegisterInfo()->getReservedRegs(MF);
14253   if (!ReservedRegs.test(Reg) && !Subtarget.isRegisterReservedByUser(Reg))
14254     report_fatal_error(Twine("Trying to obtain non-reserved register \"" +
14255                              StringRef(RegName) + "\"."));
14256   return Reg;
14257 }
14258 
14259 namespace llvm::RISCVVIntrinsicsTable {
14260 
14261 #define GET_RISCVVIntrinsicsTable_IMPL
14262 #include "RISCVGenSearchableTables.inc"
14263 
14264 } // namespace llvm::RISCVVIntrinsicsTable
14265