1 //===- TargetLoweringBase.cpp - Implement the TargetLoweringBase class ----===//
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 implements the TargetLoweringBase class.
10 //
11 //===----------------------------------------------------------------------===//
12 
13 #include "llvm/ADT/BitVector.h"
14 #include "llvm/ADT/STLExtras.h"
15 #include "llvm/ADT/SmallVector.h"
16 #include "llvm/ADT/StringExtras.h"
17 #include "llvm/ADT/StringRef.h"
18 #include "llvm/ADT/Triple.h"
19 #include "llvm/ADT/Twine.h"
20 #include "llvm/Analysis/Loads.h"
21 #include "llvm/Analysis/TargetTransformInfo.h"
22 #include "llvm/CodeGen/Analysis.h"
23 #include "llvm/CodeGen/ISDOpcodes.h"
24 #include "llvm/CodeGen/MachineBasicBlock.h"
25 #include "llvm/CodeGen/MachineFrameInfo.h"
26 #include "llvm/CodeGen/MachineFunction.h"
27 #include "llvm/CodeGen/MachineInstr.h"
28 #include "llvm/CodeGen/MachineInstrBuilder.h"
29 #include "llvm/CodeGen/MachineMemOperand.h"
30 #include "llvm/CodeGen/MachineOperand.h"
31 #include "llvm/CodeGen/MachineRegisterInfo.h"
32 #include "llvm/CodeGen/RuntimeLibcalls.h"
33 #include "llvm/CodeGen/StackMaps.h"
34 #include "llvm/CodeGen/TargetLowering.h"
35 #include "llvm/CodeGen/TargetOpcodes.h"
36 #include "llvm/CodeGen/TargetRegisterInfo.h"
37 #include "llvm/CodeGen/ValueTypes.h"
38 #include "llvm/IR/Attributes.h"
39 #include "llvm/IR/CallingConv.h"
40 #include "llvm/IR/DataLayout.h"
41 #include "llvm/IR/DerivedTypes.h"
42 #include "llvm/IR/Function.h"
43 #include "llvm/IR/GlobalValue.h"
44 #include "llvm/IR/GlobalVariable.h"
45 #include "llvm/IR/IRBuilder.h"
46 #include "llvm/IR/Module.h"
47 #include "llvm/IR/Type.h"
48 #include "llvm/Support/Casting.h"
49 #include "llvm/Support/CommandLine.h"
50 #include "llvm/Support/Compiler.h"
51 #include "llvm/Support/ErrorHandling.h"
52 #include "llvm/Support/MachineValueType.h"
53 #include "llvm/Support/MathExtras.h"
54 #include "llvm/Target/TargetMachine.h"
55 #include "llvm/Target/TargetOptions.h"
56 #include "llvm/Transforms/Utils/SizeOpts.h"
57 #include <algorithm>
58 #include <cassert>
59 #include <cstdint>
60 #include <cstring>
61 #include <iterator>
62 #include <string>
63 #include <tuple>
64 #include <utility>
65 
66 using namespace llvm;
67 
68 static cl::opt<bool> JumpIsExpensiveOverride(
69     "jump-is-expensive", cl::init(false),
70     cl::desc("Do not create extra branches to split comparison logic."),
71     cl::Hidden);
72 
73 static cl::opt<unsigned> MinimumJumpTableEntries
74   ("min-jump-table-entries", cl::init(4), cl::Hidden,
75    cl::desc("Set minimum number of entries to use a jump table."));
76 
77 static cl::opt<unsigned> MaximumJumpTableSize
78   ("max-jump-table-size", cl::init(UINT_MAX), cl::Hidden,
79    cl::desc("Set maximum size of jump tables."));
80 
81 /// Minimum jump table density for normal functions.
82 static cl::opt<unsigned>
83     JumpTableDensity("jump-table-density", cl::init(10), cl::Hidden,
84                      cl::desc("Minimum density for building a jump table in "
85                               "a normal function"));
86 
87 /// Minimum jump table density for -Os or -Oz functions.
88 static cl::opt<unsigned> OptsizeJumpTableDensity(
89     "optsize-jump-table-density", cl::init(40), cl::Hidden,
90     cl::desc("Minimum density for building a jump table in "
91              "an optsize function"));
92 
93 // FIXME: This option is only to test if the strict fp operation processed
94 // correctly by preventing mutating strict fp operation to normal fp operation
95 // during development. When the backend supports strict float operation, this
96 // option will be meaningless.
97 static cl::opt<bool> DisableStrictNodeMutation("disable-strictnode-mutation",
98        cl::desc("Don't mutate strict-float node to a legalize node"),
99        cl::init(false), cl::Hidden);
100 
101 static bool darwinHasSinCos(const Triple &TT) {
102   assert(TT.isOSDarwin() && "should be called with darwin triple");
103   // Don't bother with 32 bit x86.
104   if (TT.getArch() == Triple::x86)
105     return false;
106   // Macos < 10.9 has no sincos_stret.
107   if (TT.isMacOSX())
108     return !TT.isMacOSXVersionLT(10, 9) && TT.isArch64Bit();
109   // iOS < 7.0 has no sincos_stret.
110   if (TT.isiOS())
111     return !TT.isOSVersionLT(7, 0);
112   // Any other darwin such as WatchOS/TvOS is new enough.
113   return true;
114 }
115 
116 void TargetLoweringBase::InitLibcalls(const Triple &TT) {
117 #define HANDLE_LIBCALL(code, name) \
118   setLibcallName(RTLIB::code, name);
119 #include "llvm/IR/RuntimeLibcalls.def"
120 #undef HANDLE_LIBCALL
121   // Initialize calling conventions to their default.
122   for (int LC = 0; LC < RTLIB::UNKNOWN_LIBCALL; ++LC)
123     setLibcallCallingConv((RTLIB::Libcall)LC, CallingConv::C);
124 
125   // For IEEE quad-precision libcall names, PPC uses "kf" instead of "tf".
126   if (TT.isPPC()) {
127     setLibcallName(RTLIB::ADD_F128, "__addkf3");
128     setLibcallName(RTLIB::SUB_F128, "__subkf3");
129     setLibcallName(RTLIB::MUL_F128, "__mulkf3");
130     setLibcallName(RTLIB::DIV_F128, "__divkf3");
131     setLibcallName(RTLIB::POWI_F128, "__powikf2");
132     setLibcallName(RTLIB::FPEXT_F32_F128, "__extendsfkf2");
133     setLibcallName(RTLIB::FPEXT_F64_F128, "__extenddfkf2");
134     setLibcallName(RTLIB::FPROUND_F128_F32, "__trunckfsf2");
135     setLibcallName(RTLIB::FPROUND_F128_F64, "__trunckfdf2");
136     setLibcallName(RTLIB::FPTOSINT_F128_I32, "__fixkfsi");
137     setLibcallName(RTLIB::FPTOSINT_F128_I64, "__fixkfdi");
138     setLibcallName(RTLIB::FPTOSINT_F128_I128, "__fixkfti");
139     setLibcallName(RTLIB::FPTOUINT_F128_I32, "__fixunskfsi");
140     setLibcallName(RTLIB::FPTOUINT_F128_I64, "__fixunskfdi");
141     setLibcallName(RTLIB::FPTOUINT_F128_I128, "__fixunskfti");
142     setLibcallName(RTLIB::SINTTOFP_I32_F128, "__floatsikf");
143     setLibcallName(RTLIB::SINTTOFP_I64_F128, "__floatdikf");
144     setLibcallName(RTLIB::SINTTOFP_I128_F128, "__floattikf");
145     setLibcallName(RTLIB::UINTTOFP_I32_F128, "__floatunsikf");
146     setLibcallName(RTLIB::UINTTOFP_I64_F128, "__floatundikf");
147     setLibcallName(RTLIB::UINTTOFP_I128_F128, "__floatuntikf");
148     setLibcallName(RTLIB::OEQ_F128, "__eqkf2");
149     setLibcallName(RTLIB::UNE_F128, "__nekf2");
150     setLibcallName(RTLIB::OGE_F128, "__gekf2");
151     setLibcallName(RTLIB::OLT_F128, "__ltkf2");
152     setLibcallName(RTLIB::OLE_F128, "__lekf2");
153     setLibcallName(RTLIB::OGT_F128, "__gtkf2");
154     setLibcallName(RTLIB::UO_F128, "__unordkf2");
155   }
156 
157   // A few names are different on particular architectures or environments.
158   if (TT.isOSDarwin()) {
159     // For f16/f32 conversions, Darwin uses the standard naming scheme, instead
160     // of the gnueabi-style __gnu_*_ieee.
161     // FIXME: What about other targets?
162     setLibcallName(RTLIB::FPEXT_F16_F32, "__extendhfsf2");
163     setLibcallName(RTLIB::FPROUND_F32_F16, "__truncsfhf2");
164 
165     // Some darwins have an optimized __bzero/bzero function.
166     switch (TT.getArch()) {
167     case Triple::x86:
168     case Triple::x86_64:
169       if (TT.isMacOSX() && !TT.isMacOSXVersionLT(10, 6))
170         setLibcallName(RTLIB::BZERO, "__bzero");
171       break;
172     case Triple::aarch64:
173     case Triple::aarch64_32:
174       setLibcallName(RTLIB::BZERO, "bzero");
175       break;
176     default:
177       break;
178     }
179 
180     if (darwinHasSinCos(TT)) {
181       setLibcallName(RTLIB::SINCOS_STRET_F32, "__sincosf_stret");
182       setLibcallName(RTLIB::SINCOS_STRET_F64, "__sincos_stret");
183       if (TT.isWatchABI()) {
184         setLibcallCallingConv(RTLIB::SINCOS_STRET_F32,
185                               CallingConv::ARM_AAPCS_VFP);
186         setLibcallCallingConv(RTLIB::SINCOS_STRET_F64,
187                               CallingConv::ARM_AAPCS_VFP);
188       }
189     }
190   } else {
191     setLibcallName(RTLIB::FPEXT_F16_F32, "__gnu_h2f_ieee");
192     setLibcallName(RTLIB::FPROUND_F32_F16, "__gnu_f2h_ieee");
193   }
194 
195   if (TT.isGNUEnvironment() || TT.isOSFuchsia() ||
196       (TT.isAndroid() && !TT.isAndroidVersionLT(9))) {
197     setLibcallName(RTLIB::SINCOS_F32, "sincosf");
198     setLibcallName(RTLIB::SINCOS_F64, "sincos");
199     setLibcallName(RTLIB::SINCOS_F80, "sincosl");
200     setLibcallName(RTLIB::SINCOS_F128, "sincosl");
201     setLibcallName(RTLIB::SINCOS_PPCF128, "sincosl");
202   }
203 
204   if (TT.isPS()) {
205     setLibcallName(RTLIB::SINCOS_F32, "sincosf");
206     setLibcallName(RTLIB::SINCOS_F64, "sincos");
207   }
208 
209   if (TT.isOSOpenBSD()) {
210     setLibcallName(RTLIB::STACKPROTECTOR_CHECK_FAIL, nullptr);
211   }
212 }
213 
214 /// GetFPLibCall - Helper to return the right libcall for the given floating
215 /// point type, or UNKNOWN_LIBCALL if there is none.
216 RTLIB::Libcall RTLIB::getFPLibCall(EVT VT,
217                                    RTLIB::Libcall Call_F32,
218                                    RTLIB::Libcall Call_F64,
219                                    RTLIB::Libcall Call_F80,
220                                    RTLIB::Libcall Call_F128,
221                                    RTLIB::Libcall Call_PPCF128) {
222   return
223     VT == MVT::f32 ? Call_F32 :
224     VT == MVT::f64 ? Call_F64 :
225     VT == MVT::f80 ? Call_F80 :
226     VT == MVT::f128 ? Call_F128 :
227     VT == MVT::ppcf128 ? Call_PPCF128 :
228     RTLIB::UNKNOWN_LIBCALL;
229 }
230 
231 /// getFPEXT - Return the FPEXT_*_* value for the given types, or
232 /// UNKNOWN_LIBCALL if there is none.
233 RTLIB::Libcall RTLIB::getFPEXT(EVT OpVT, EVT RetVT) {
234   if (OpVT == MVT::f16) {
235     if (RetVT == MVT::f32)
236       return FPEXT_F16_F32;
237     if (RetVT == MVT::f64)
238       return FPEXT_F16_F64;
239     if (RetVT == MVT::f80)
240       return FPEXT_F16_F80;
241     if (RetVT == MVT::f128)
242       return FPEXT_F16_F128;
243   } else if (OpVT == MVT::f32) {
244     if (RetVT == MVT::f64)
245       return FPEXT_F32_F64;
246     if (RetVT == MVT::f128)
247       return FPEXT_F32_F128;
248     if (RetVT == MVT::ppcf128)
249       return FPEXT_F32_PPCF128;
250   } else if (OpVT == MVT::f64) {
251     if (RetVT == MVT::f128)
252       return FPEXT_F64_F128;
253     else if (RetVT == MVT::ppcf128)
254       return FPEXT_F64_PPCF128;
255   } else if (OpVT == MVT::f80) {
256     if (RetVT == MVT::f128)
257       return FPEXT_F80_F128;
258   }
259 
260   return UNKNOWN_LIBCALL;
261 }
262 
263 /// getFPROUND - Return the FPROUND_*_* value for the given types, or
264 /// UNKNOWN_LIBCALL if there is none.
265 RTLIB::Libcall RTLIB::getFPROUND(EVT OpVT, EVT RetVT) {
266   if (RetVT == MVT::f16) {
267     if (OpVT == MVT::f32)
268       return FPROUND_F32_F16;
269     if (OpVT == MVT::f64)
270       return FPROUND_F64_F16;
271     if (OpVT == MVT::f80)
272       return FPROUND_F80_F16;
273     if (OpVT == MVT::f128)
274       return FPROUND_F128_F16;
275     if (OpVT == MVT::ppcf128)
276       return FPROUND_PPCF128_F16;
277   } else if (RetVT == MVT::bf16) {
278     if (OpVT == MVT::f32)
279       return FPROUND_F32_BF16;
280     if (OpVT == MVT::f64)
281       return FPROUND_F64_BF16;
282   } else if (RetVT == MVT::f32) {
283     if (OpVT == MVT::f64)
284       return FPROUND_F64_F32;
285     if (OpVT == MVT::f80)
286       return FPROUND_F80_F32;
287     if (OpVT == MVT::f128)
288       return FPROUND_F128_F32;
289     if (OpVT == MVT::ppcf128)
290       return FPROUND_PPCF128_F32;
291   } else if (RetVT == MVT::f64) {
292     if (OpVT == MVT::f80)
293       return FPROUND_F80_F64;
294     if (OpVT == MVT::f128)
295       return FPROUND_F128_F64;
296     if (OpVT == MVT::ppcf128)
297       return FPROUND_PPCF128_F64;
298   } else if (RetVT == MVT::f80) {
299     if (OpVT == MVT::f128)
300       return FPROUND_F128_F80;
301   }
302 
303   return UNKNOWN_LIBCALL;
304 }
305 
306 /// getFPTOSINT - Return the FPTOSINT_*_* value for the given types, or
307 /// UNKNOWN_LIBCALL if there is none.
308 RTLIB::Libcall RTLIB::getFPTOSINT(EVT OpVT, EVT RetVT) {
309   if (OpVT == MVT::f16) {
310     if (RetVT == MVT::i32)
311       return FPTOSINT_F16_I32;
312     if (RetVT == MVT::i64)
313       return FPTOSINT_F16_I64;
314     if (RetVT == MVT::i128)
315       return FPTOSINT_F16_I128;
316   } else if (OpVT == MVT::f32) {
317     if (RetVT == MVT::i32)
318       return FPTOSINT_F32_I32;
319     if (RetVT == MVT::i64)
320       return FPTOSINT_F32_I64;
321     if (RetVT == MVT::i128)
322       return FPTOSINT_F32_I128;
323   } else if (OpVT == MVT::f64) {
324     if (RetVT == MVT::i32)
325       return FPTOSINT_F64_I32;
326     if (RetVT == MVT::i64)
327       return FPTOSINT_F64_I64;
328     if (RetVT == MVT::i128)
329       return FPTOSINT_F64_I128;
330   } else if (OpVT == MVT::f80) {
331     if (RetVT == MVT::i32)
332       return FPTOSINT_F80_I32;
333     if (RetVT == MVT::i64)
334       return FPTOSINT_F80_I64;
335     if (RetVT == MVT::i128)
336       return FPTOSINT_F80_I128;
337   } else if (OpVT == MVT::f128) {
338     if (RetVT == MVT::i32)
339       return FPTOSINT_F128_I32;
340     if (RetVT == MVT::i64)
341       return FPTOSINT_F128_I64;
342     if (RetVT == MVT::i128)
343       return FPTOSINT_F128_I128;
344   } else if (OpVT == MVT::ppcf128) {
345     if (RetVT == MVT::i32)
346       return FPTOSINT_PPCF128_I32;
347     if (RetVT == MVT::i64)
348       return FPTOSINT_PPCF128_I64;
349     if (RetVT == MVT::i128)
350       return FPTOSINT_PPCF128_I128;
351   }
352   return UNKNOWN_LIBCALL;
353 }
354 
355 /// getFPTOUINT - Return the FPTOUINT_*_* value for the given types, or
356 /// UNKNOWN_LIBCALL if there is none.
357 RTLIB::Libcall RTLIB::getFPTOUINT(EVT OpVT, EVT RetVT) {
358   if (OpVT == MVT::f16) {
359     if (RetVT == MVT::i32)
360       return FPTOUINT_F16_I32;
361     if (RetVT == MVT::i64)
362       return FPTOUINT_F16_I64;
363     if (RetVT == MVT::i128)
364       return FPTOUINT_F16_I128;
365   } else if (OpVT == MVT::f32) {
366     if (RetVT == MVT::i32)
367       return FPTOUINT_F32_I32;
368     if (RetVT == MVT::i64)
369       return FPTOUINT_F32_I64;
370     if (RetVT == MVT::i128)
371       return FPTOUINT_F32_I128;
372   } else if (OpVT == MVT::f64) {
373     if (RetVT == MVT::i32)
374       return FPTOUINT_F64_I32;
375     if (RetVT == MVT::i64)
376       return FPTOUINT_F64_I64;
377     if (RetVT == MVT::i128)
378       return FPTOUINT_F64_I128;
379   } else if (OpVT == MVT::f80) {
380     if (RetVT == MVT::i32)
381       return FPTOUINT_F80_I32;
382     if (RetVT == MVT::i64)
383       return FPTOUINT_F80_I64;
384     if (RetVT == MVT::i128)
385       return FPTOUINT_F80_I128;
386   } else if (OpVT == MVT::f128) {
387     if (RetVT == MVT::i32)
388       return FPTOUINT_F128_I32;
389     if (RetVT == MVT::i64)
390       return FPTOUINT_F128_I64;
391     if (RetVT == MVT::i128)
392       return FPTOUINT_F128_I128;
393   } else if (OpVT == MVT::ppcf128) {
394     if (RetVT == MVT::i32)
395       return FPTOUINT_PPCF128_I32;
396     if (RetVT == MVT::i64)
397       return FPTOUINT_PPCF128_I64;
398     if (RetVT == MVT::i128)
399       return FPTOUINT_PPCF128_I128;
400   }
401   return UNKNOWN_LIBCALL;
402 }
403 
404 /// getSINTTOFP - Return the SINTTOFP_*_* value for the given types, or
405 /// UNKNOWN_LIBCALL if there is none.
406 RTLIB::Libcall RTLIB::getSINTTOFP(EVT OpVT, EVT RetVT) {
407   if (OpVT == MVT::i32) {
408     if (RetVT == MVT::f16)
409       return SINTTOFP_I32_F16;
410     if (RetVT == MVT::f32)
411       return SINTTOFP_I32_F32;
412     if (RetVT == MVT::f64)
413       return SINTTOFP_I32_F64;
414     if (RetVT == MVT::f80)
415       return SINTTOFP_I32_F80;
416     if (RetVT == MVT::f128)
417       return SINTTOFP_I32_F128;
418     if (RetVT == MVT::ppcf128)
419       return SINTTOFP_I32_PPCF128;
420   } else if (OpVT == MVT::i64) {
421     if (RetVT == MVT::f16)
422       return SINTTOFP_I64_F16;
423     if (RetVT == MVT::f32)
424       return SINTTOFP_I64_F32;
425     if (RetVT == MVT::f64)
426       return SINTTOFP_I64_F64;
427     if (RetVT == MVT::f80)
428       return SINTTOFP_I64_F80;
429     if (RetVT == MVT::f128)
430       return SINTTOFP_I64_F128;
431     if (RetVT == MVT::ppcf128)
432       return SINTTOFP_I64_PPCF128;
433   } else if (OpVT == MVT::i128) {
434     if (RetVT == MVT::f16)
435       return SINTTOFP_I128_F16;
436     if (RetVT == MVT::f32)
437       return SINTTOFP_I128_F32;
438     if (RetVT == MVT::f64)
439       return SINTTOFP_I128_F64;
440     if (RetVT == MVT::f80)
441       return SINTTOFP_I128_F80;
442     if (RetVT == MVT::f128)
443       return SINTTOFP_I128_F128;
444     if (RetVT == MVT::ppcf128)
445       return SINTTOFP_I128_PPCF128;
446   }
447   return UNKNOWN_LIBCALL;
448 }
449 
450 /// getUINTTOFP - Return the UINTTOFP_*_* value for the given types, or
451 /// UNKNOWN_LIBCALL if there is none.
452 RTLIB::Libcall RTLIB::getUINTTOFP(EVT OpVT, EVT RetVT) {
453   if (OpVT == MVT::i32) {
454     if (RetVT == MVT::f16)
455       return UINTTOFP_I32_F16;
456     if (RetVT == MVT::f32)
457       return UINTTOFP_I32_F32;
458     if (RetVT == MVT::f64)
459       return UINTTOFP_I32_F64;
460     if (RetVT == MVT::f80)
461       return UINTTOFP_I32_F80;
462     if (RetVT == MVT::f128)
463       return UINTTOFP_I32_F128;
464     if (RetVT == MVT::ppcf128)
465       return UINTTOFP_I32_PPCF128;
466   } else if (OpVT == MVT::i64) {
467     if (RetVT == MVT::f16)
468       return UINTTOFP_I64_F16;
469     if (RetVT == MVT::f32)
470       return UINTTOFP_I64_F32;
471     if (RetVT == MVT::f64)
472       return UINTTOFP_I64_F64;
473     if (RetVT == MVT::f80)
474       return UINTTOFP_I64_F80;
475     if (RetVT == MVT::f128)
476       return UINTTOFP_I64_F128;
477     if (RetVT == MVT::ppcf128)
478       return UINTTOFP_I64_PPCF128;
479   } else if (OpVT == MVT::i128) {
480     if (RetVT == MVT::f16)
481       return UINTTOFP_I128_F16;
482     if (RetVT == MVT::f32)
483       return UINTTOFP_I128_F32;
484     if (RetVT == MVT::f64)
485       return UINTTOFP_I128_F64;
486     if (RetVT == MVT::f80)
487       return UINTTOFP_I128_F80;
488     if (RetVT == MVT::f128)
489       return UINTTOFP_I128_F128;
490     if (RetVT == MVT::ppcf128)
491       return UINTTOFP_I128_PPCF128;
492   }
493   return UNKNOWN_LIBCALL;
494 }
495 
496 RTLIB::Libcall RTLIB::getPOWI(EVT RetVT) {
497   return getFPLibCall(RetVT, POWI_F32, POWI_F64, POWI_F80, POWI_F128,
498                       POWI_PPCF128);
499 }
500 
501 RTLIB::Libcall RTLIB::getOUTLINE_ATOMIC(unsigned Opc, AtomicOrdering Order,
502                                         MVT VT) {
503   unsigned ModeN, ModelN;
504   switch (VT.SimpleTy) {
505   case MVT::i8:
506     ModeN = 0;
507     break;
508   case MVT::i16:
509     ModeN = 1;
510     break;
511   case MVT::i32:
512     ModeN = 2;
513     break;
514   case MVT::i64:
515     ModeN = 3;
516     break;
517   case MVT::i128:
518     ModeN = 4;
519     break;
520   default:
521     return UNKNOWN_LIBCALL;
522   }
523 
524   switch (Order) {
525   case AtomicOrdering::Monotonic:
526     ModelN = 0;
527     break;
528   case AtomicOrdering::Acquire:
529     ModelN = 1;
530     break;
531   case AtomicOrdering::Release:
532     ModelN = 2;
533     break;
534   case AtomicOrdering::AcquireRelease:
535   case AtomicOrdering::SequentiallyConsistent:
536     ModelN = 3;
537     break;
538   default:
539     return UNKNOWN_LIBCALL;
540   }
541 
542 #define LCALLS(A, B)                                                           \
543   { A##B##_RELAX, A##B##_ACQ, A##B##_REL, A##B##_ACQ_REL }
544 #define LCALL5(A)                                                              \
545   LCALLS(A, 1), LCALLS(A, 2), LCALLS(A, 4), LCALLS(A, 8), LCALLS(A, 16)
546   switch (Opc) {
547   case ISD::ATOMIC_CMP_SWAP: {
548     const Libcall LC[5][4] = {LCALL5(OUTLINE_ATOMIC_CAS)};
549     return LC[ModeN][ModelN];
550   }
551   case ISD::ATOMIC_SWAP: {
552     const Libcall LC[5][4] = {LCALL5(OUTLINE_ATOMIC_SWP)};
553     return LC[ModeN][ModelN];
554   }
555   case ISD::ATOMIC_LOAD_ADD: {
556     const Libcall LC[5][4] = {LCALL5(OUTLINE_ATOMIC_LDADD)};
557     return LC[ModeN][ModelN];
558   }
559   case ISD::ATOMIC_LOAD_OR: {
560     const Libcall LC[5][4] = {LCALL5(OUTLINE_ATOMIC_LDSET)};
561     return LC[ModeN][ModelN];
562   }
563   case ISD::ATOMIC_LOAD_CLR: {
564     const Libcall LC[5][4] = {LCALL5(OUTLINE_ATOMIC_LDCLR)};
565     return LC[ModeN][ModelN];
566   }
567   case ISD::ATOMIC_LOAD_XOR: {
568     const Libcall LC[5][4] = {LCALL5(OUTLINE_ATOMIC_LDEOR)};
569     return LC[ModeN][ModelN];
570   }
571   default:
572     return UNKNOWN_LIBCALL;
573   }
574 #undef LCALLS
575 #undef LCALL5
576 }
577 
578 RTLIB::Libcall RTLIB::getSYNC(unsigned Opc, MVT VT) {
579 #define OP_TO_LIBCALL(Name, Enum)                                              \
580   case Name:                                                                   \
581     switch (VT.SimpleTy) {                                                     \
582     default:                                                                   \
583       return UNKNOWN_LIBCALL;                                                  \
584     case MVT::i8:                                                              \
585       return Enum##_1;                                                         \
586     case MVT::i16:                                                             \
587       return Enum##_2;                                                         \
588     case MVT::i32:                                                             \
589       return Enum##_4;                                                         \
590     case MVT::i64:                                                             \
591       return Enum##_8;                                                         \
592     case MVT::i128:                                                            \
593       return Enum##_16;                                                        \
594     }
595 
596   switch (Opc) {
597     OP_TO_LIBCALL(ISD::ATOMIC_SWAP, SYNC_LOCK_TEST_AND_SET)
598     OP_TO_LIBCALL(ISD::ATOMIC_CMP_SWAP, SYNC_VAL_COMPARE_AND_SWAP)
599     OP_TO_LIBCALL(ISD::ATOMIC_LOAD_ADD, SYNC_FETCH_AND_ADD)
600     OP_TO_LIBCALL(ISD::ATOMIC_LOAD_SUB, SYNC_FETCH_AND_SUB)
601     OP_TO_LIBCALL(ISD::ATOMIC_LOAD_AND, SYNC_FETCH_AND_AND)
602     OP_TO_LIBCALL(ISD::ATOMIC_LOAD_OR, SYNC_FETCH_AND_OR)
603     OP_TO_LIBCALL(ISD::ATOMIC_LOAD_XOR, SYNC_FETCH_AND_XOR)
604     OP_TO_LIBCALL(ISD::ATOMIC_LOAD_NAND, SYNC_FETCH_AND_NAND)
605     OP_TO_LIBCALL(ISD::ATOMIC_LOAD_MAX, SYNC_FETCH_AND_MAX)
606     OP_TO_LIBCALL(ISD::ATOMIC_LOAD_UMAX, SYNC_FETCH_AND_UMAX)
607     OP_TO_LIBCALL(ISD::ATOMIC_LOAD_MIN, SYNC_FETCH_AND_MIN)
608     OP_TO_LIBCALL(ISD::ATOMIC_LOAD_UMIN, SYNC_FETCH_AND_UMIN)
609   }
610 
611 #undef OP_TO_LIBCALL
612 
613   return UNKNOWN_LIBCALL;
614 }
615 
616 RTLIB::Libcall RTLIB::getMEMCPY_ELEMENT_UNORDERED_ATOMIC(uint64_t ElementSize) {
617   switch (ElementSize) {
618   case 1:
619     return MEMCPY_ELEMENT_UNORDERED_ATOMIC_1;
620   case 2:
621     return MEMCPY_ELEMENT_UNORDERED_ATOMIC_2;
622   case 4:
623     return MEMCPY_ELEMENT_UNORDERED_ATOMIC_4;
624   case 8:
625     return MEMCPY_ELEMENT_UNORDERED_ATOMIC_8;
626   case 16:
627     return MEMCPY_ELEMENT_UNORDERED_ATOMIC_16;
628   default:
629     return UNKNOWN_LIBCALL;
630   }
631 }
632 
633 RTLIB::Libcall RTLIB::getMEMMOVE_ELEMENT_UNORDERED_ATOMIC(uint64_t ElementSize) {
634   switch (ElementSize) {
635   case 1:
636     return MEMMOVE_ELEMENT_UNORDERED_ATOMIC_1;
637   case 2:
638     return MEMMOVE_ELEMENT_UNORDERED_ATOMIC_2;
639   case 4:
640     return MEMMOVE_ELEMENT_UNORDERED_ATOMIC_4;
641   case 8:
642     return MEMMOVE_ELEMENT_UNORDERED_ATOMIC_8;
643   case 16:
644     return MEMMOVE_ELEMENT_UNORDERED_ATOMIC_16;
645   default:
646     return UNKNOWN_LIBCALL;
647   }
648 }
649 
650 RTLIB::Libcall RTLIB::getMEMSET_ELEMENT_UNORDERED_ATOMIC(uint64_t ElementSize) {
651   switch (ElementSize) {
652   case 1:
653     return MEMSET_ELEMENT_UNORDERED_ATOMIC_1;
654   case 2:
655     return MEMSET_ELEMENT_UNORDERED_ATOMIC_2;
656   case 4:
657     return MEMSET_ELEMENT_UNORDERED_ATOMIC_4;
658   case 8:
659     return MEMSET_ELEMENT_UNORDERED_ATOMIC_8;
660   case 16:
661     return MEMSET_ELEMENT_UNORDERED_ATOMIC_16;
662   default:
663     return UNKNOWN_LIBCALL;
664   }
665 }
666 
667 /// InitCmpLibcallCCs - Set default comparison libcall CC.
668 static void InitCmpLibcallCCs(ISD::CondCode *CCs) {
669   std::fill(CCs, CCs + RTLIB::UNKNOWN_LIBCALL, ISD::SETCC_INVALID);
670   CCs[RTLIB::OEQ_F32] = ISD::SETEQ;
671   CCs[RTLIB::OEQ_F64] = ISD::SETEQ;
672   CCs[RTLIB::OEQ_F128] = ISD::SETEQ;
673   CCs[RTLIB::OEQ_PPCF128] = ISD::SETEQ;
674   CCs[RTLIB::UNE_F32] = ISD::SETNE;
675   CCs[RTLIB::UNE_F64] = ISD::SETNE;
676   CCs[RTLIB::UNE_F128] = ISD::SETNE;
677   CCs[RTLIB::UNE_PPCF128] = ISD::SETNE;
678   CCs[RTLIB::OGE_F32] = ISD::SETGE;
679   CCs[RTLIB::OGE_F64] = ISD::SETGE;
680   CCs[RTLIB::OGE_F128] = ISD::SETGE;
681   CCs[RTLIB::OGE_PPCF128] = ISD::SETGE;
682   CCs[RTLIB::OLT_F32] = ISD::SETLT;
683   CCs[RTLIB::OLT_F64] = ISD::SETLT;
684   CCs[RTLIB::OLT_F128] = ISD::SETLT;
685   CCs[RTLIB::OLT_PPCF128] = ISD::SETLT;
686   CCs[RTLIB::OLE_F32] = ISD::SETLE;
687   CCs[RTLIB::OLE_F64] = ISD::SETLE;
688   CCs[RTLIB::OLE_F128] = ISD::SETLE;
689   CCs[RTLIB::OLE_PPCF128] = ISD::SETLE;
690   CCs[RTLIB::OGT_F32] = ISD::SETGT;
691   CCs[RTLIB::OGT_F64] = ISD::SETGT;
692   CCs[RTLIB::OGT_F128] = ISD::SETGT;
693   CCs[RTLIB::OGT_PPCF128] = ISD::SETGT;
694   CCs[RTLIB::UO_F32] = ISD::SETNE;
695   CCs[RTLIB::UO_F64] = ISD::SETNE;
696   CCs[RTLIB::UO_F128] = ISD::SETNE;
697   CCs[RTLIB::UO_PPCF128] = ISD::SETNE;
698 }
699 
700 /// NOTE: The TargetMachine owns TLOF.
701 TargetLoweringBase::TargetLoweringBase(const TargetMachine &tm) : TM(tm) {
702   initActions();
703 
704   // Perform these initializations only once.
705   MaxStoresPerMemset = MaxStoresPerMemcpy = MaxStoresPerMemmove =
706       MaxLoadsPerMemcmp = 8;
707   MaxGluedStoresPerMemcpy = 0;
708   MaxStoresPerMemsetOptSize = MaxStoresPerMemcpyOptSize =
709       MaxStoresPerMemmoveOptSize = MaxLoadsPerMemcmpOptSize = 4;
710   HasMultipleConditionRegisters = false;
711   HasExtractBitsInsn = false;
712   JumpIsExpensive = JumpIsExpensiveOverride;
713   PredictableSelectIsExpensive = false;
714   EnableExtLdPromotion = false;
715   StackPointerRegisterToSaveRestore = 0;
716   BooleanContents = UndefinedBooleanContent;
717   BooleanFloatContents = UndefinedBooleanContent;
718   BooleanVectorContents = UndefinedBooleanContent;
719   SchedPreferenceInfo = Sched::ILP;
720   GatherAllAliasesMaxDepth = 18;
721   IsStrictFPEnabled = DisableStrictNodeMutation;
722   MaxBytesForAlignment = 0;
723   // TODO: the default will be switched to 0 in the next commit, along
724   // with the Target-specific changes necessary.
725   MaxAtomicSizeInBitsSupported = 1024;
726 
727   MinCmpXchgSizeInBits = 0;
728   SupportsUnalignedAtomics = false;
729 
730   std::fill(std::begin(LibcallRoutineNames), std::end(LibcallRoutineNames), nullptr);
731 
732   InitLibcalls(TM.getTargetTriple());
733   InitCmpLibcallCCs(CmpLibcallCCs);
734 }
735 
736 void TargetLoweringBase::initActions() {
737   // All operations default to being supported.
738   memset(OpActions, 0, sizeof(OpActions));
739   memset(LoadExtActions, 0, sizeof(LoadExtActions));
740   memset(TruncStoreActions, 0, sizeof(TruncStoreActions));
741   memset(IndexedModeActions, 0, sizeof(IndexedModeActions));
742   memset(CondCodeActions, 0, sizeof(CondCodeActions));
743   std::fill(std::begin(RegClassForVT), std::end(RegClassForVT), nullptr);
744   std::fill(std::begin(TargetDAGCombineArray),
745             std::end(TargetDAGCombineArray), 0);
746 
747   // We're somewhat special casing MVT::i2 and MVT::i4. Ideally we want to
748   // remove this and targets should individually set these types if not legal.
749   for (ISD::NodeType NT : enum_seq(ISD::DELETED_NODE, ISD::BUILTIN_OP_END,
750                                    force_iteration_on_noniterable_enum)) {
751     for (MVT VT : {MVT::i2, MVT::i4})
752       OpActions[(unsigned)VT.SimpleTy][NT] = Expand;
753   }
754   for (MVT AVT : MVT::all_valuetypes()) {
755     for (MVT VT : {MVT::i2, MVT::i4, MVT::v128i2, MVT::v64i4}) {
756       setTruncStoreAction(AVT, VT, Expand);
757       setLoadExtAction(ISD::EXTLOAD, AVT, VT, Expand);
758       setLoadExtAction(ISD::ZEXTLOAD, AVT, VT, Expand);
759     }
760   }
761   for (unsigned IM = (unsigned)ISD::PRE_INC;
762        IM != (unsigned)ISD::LAST_INDEXED_MODE; ++IM) {
763     for (MVT VT : {MVT::i2, MVT::i4}) {
764       setIndexedLoadAction(IM, VT, Expand);
765       setIndexedStoreAction(IM, VT, Expand);
766       setIndexedMaskedLoadAction(IM, VT, Expand);
767       setIndexedMaskedStoreAction(IM, VT, Expand);
768     }
769   }
770 
771   for (MVT VT : MVT::fp_valuetypes()) {
772     MVT IntVT = MVT::getIntegerVT(VT.getFixedSizeInBits());
773     if (IntVT.isValid()) {
774       setOperationAction(ISD::ATOMIC_SWAP, VT, Promote);
775       AddPromotedToType(ISD::ATOMIC_SWAP, VT, IntVT);
776     }
777   }
778 
779   // Set default actions for various operations.
780   for (MVT VT : MVT::all_valuetypes()) {
781     // Default all indexed load / store to expand.
782     for (unsigned IM = (unsigned)ISD::PRE_INC;
783          IM != (unsigned)ISD::LAST_INDEXED_MODE; ++IM) {
784       setIndexedLoadAction(IM, VT, Expand);
785       setIndexedStoreAction(IM, VT, Expand);
786       setIndexedMaskedLoadAction(IM, VT, Expand);
787       setIndexedMaskedStoreAction(IM, VT, Expand);
788     }
789 
790     // Most backends expect to see the node which just returns the value loaded.
791     setOperationAction(ISD::ATOMIC_CMP_SWAP_WITH_SUCCESS, VT, Expand);
792 
793     // These operations default to expand.
794     setOperationAction({ISD::FGETSIGN,       ISD::CONCAT_VECTORS,
795                         ISD::FMINNUM,        ISD::FMAXNUM,
796                         ISD::FMINNUM_IEEE,   ISD::FMAXNUM_IEEE,
797                         ISD::FMINIMUM,       ISD::FMAXIMUM,
798                         ISD::FMAD,           ISD::SMIN,
799                         ISD::SMAX,           ISD::UMIN,
800                         ISD::UMAX,           ISD::ABS,
801                         ISD::FSHL,           ISD::FSHR,
802                         ISD::SADDSAT,        ISD::UADDSAT,
803                         ISD::SSUBSAT,        ISD::USUBSAT,
804                         ISD::SSHLSAT,        ISD::USHLSAT,
805                         ISD::SMULFIX,        ISD::SMULFIXSAT,
806                         ISD::UMULFIX,        ISD::UMULFIXSAT,
807                         ISD::SDIVFIX,        ISD::SDIVFIXSAT,
808                         ISD::UDIVFIX,        ISD::UDIVFIXSAT,
809                         ISD::FP_TO_SINT_SAT, ISD::FP_TO_UINT_SAT,
810                         ISD::IS_FPCLASS},
811                        VT, Expand);
812 
813     // Overflow operations default to expand
814     setOperationAction({ISD::SADDO, ISD::SSUBO, ISD::UADDO, ISD::USUBO,
815                         ISD::SMULO, ISD::UMULO},
816                        VT, Expand);
817 
818     // ADDCARRY operations default to expand
819     setOperationAction({ISD::ADDCARRY, ISD::SUBCARRY, ISD::SETCCCARRY,
820                         ISD::SADDO_CARRY, ISD::SSUBO_CARRY},
821                        VT, Expand);
822 
823     // ADDC/ADDE/SUBC/SUBE default to expand.
824     setOperationAction({ISD::ADDC, ISD::ADDE, ISD::SUBC, ISD::SUBE}, VT,
825                        Expand);
826 
827     // Halving adds
828     setOperationAction(
829         {ISD::AVGFLOORS, ISD::AVGFLOORU, ISD::AVGCEILS, ISD::AVGCEILU}, VT,
830         Expand);
831 
832     // Absolute difference
833     setOperationAction({ISD::ABDS, ISD::ABDU}, VT, Expand);
834 
835     // These default to Expand so they will be expanded to CTLZ/CTTZ by default.
836     setOperationAction({ISD::CTLZ_ZERO_UNDEF, ISD::CTTZ_ZERO_UNDEF}, VT,
837                        Expand);
838 
839     setOperationAction({ISD::BITREVERSE, ISD::PARITY}, VT, Expand);
840 
841     // These library functions default to expand.
842     setOperationAction({ISD::FROUND, ISD::FROUNDEVEN, ISD::FPOWI}, VT, Expand);
843 
844     // These operations default to expand for vector types.
845     if (VT.isVector())
846       setOperationAction({ISD::FCOPYSIGN, ISD::SIGN_EXTEND_INREG,
847                           ISD::ANY_EXTEND_VECTOR_INREG,
848                           ISD::SIGN_EXTEND_VECTOR_INREG,
849                           ISD::ZERO_EXTEND_VECTOR_INREG, ISD::SPLAT_VECTOR},
850                          VT, Expand);
851 
852     // Constrained floating-point operations default to expand.
853 #define DAG_INSTRUCTION(NAME, NARG, ROUND_MODE, INTRINSIC, DAGN)               \
854     setOperationAction(ISD::STRICT_##DAGN, VT, Expand);
855 #include "llvm/IR/ConstrainedOps.def"
856 
857     // For most targets @llvm.get.dynamic.area.offset just returns 0.
858     setOperationAction(ISD::GET_DYNAMIC_AREA_OFFSET, VT, Expand);
859 
860     // Vector reduction default to expand.
861     setOperationAction(
862         {ISD::VECREDUCE_FADD, ISD::VECREDUCE_FMUL, ISD::VECREDUCE_ADD,
863          ISD::VECREDUCE_MUL, ISD::VECREDUCE_AND, ISD::VECREDUCE_OR,
864          ISD::VECREDUCE_XOR, ISD::VECREDUCE_SMAX, ISD::VECREDUCE_SMIN,
865          ISD::VECREDUCE_UMAX, ISD::VECREDUCE_UMIN, ISD::VECREDUCE_FMAX,
866          ISD::VECREDUCE_FMIN, ISD::VECREDUCE_SEQ_FADD, ISD::VECREDUCE_SEQ_FMUL},
867         VT, Expand);
868 
869     // Named vector shuffles default to expand.
870     setOperationAction(ISD::VECTOR_SPLICE, VT, Expand);
871   }
872 
873   // Most targets ignore the @llvm.prefetch intrinsic.
874   setOperationAction(ISD::PREFETCH, MVT::Other, Expand);
875 
876   // Most targets also ignore the @llvm.readcyclecounter intrinsic.
877   setOperationAction(ISD::READCYCLECOUNTER, MVT::i64, Expand);
878 
879   // ConstantFP nodes default to expand.  Targets can either change this to
880   // Legal, in which case all fp constants are legal, or use isFPImmLegal()
881   // to optimize expansions for certain constants.
882   setOperationAction(ISD::ConstantFP,
883                      {MVT::f16, MVT::f32, MVT::f64, MVT::f80, MVT::f128},
884                      Expand);
885 
886   // These library functions default to expand.
887   setOperationAction({ISD::FCBRT, ISD::FLOG, ISD::FLOG2, ISD::FLOG10, ISD::FEXP,
888                       ISD::FEXP2, ISD::FFLOOR, ISD::FNEARBYINT, ISD::FCEIL,
889                       ISD::FRINT, ISD::FTRUNC, ISD::LROUND, ISD::LLROUND,
890                       ISD::LRINT, ISD::LLRINT},
891                      {MVT::f32, MVT::f64, MVT::f128}, Expand);
892 
893   // Default ISD::TRAP to expand (which turns it into abort).
894   setOperationAction(ISD::TRAP, MVT::Other, Expand);
895 
896   // On most systems, DEBUGTRAP and TRAP have no difference. The "Expand"
897   // here is to inform DAG Legalizer to replace DEBUGTRAP with TRAP.
898   setOperationAction(ISD::DEBUGTRAP, MVT::Other, Expand);
899 
900   setOperationAction(ISD::UBSANTRAP, MVT::Other, Expand);
901 }
902 
903 MVT TargetLoweringBase::getScalarShiftAmountTy(const DataLayout &DL,
904                                                EVT) const {
905   return MVT::getIntegerVT(DL.getPointerSizeInBits(0));
906 }
907 
908 EVT TargetLoweringBase::getShiftAmountTy(EVT LHSTy, const DataLayout &DL,
909                                          bool LegalTypes) const {
910   assert(LHSTy.isInteger() && "Shift amount is not an integer type!");
911   if (LHSTy.isVector())
912     return LHSTy;
913   MVT ShiftVT =
914       LegalTypes ? getScalarShiftAmountTy(DL, LHSTy) : getPointerTy(DL);
915   // If any possible shift value won't fit in the prefered type, just use
916   // something safe. Assume it will be legalized when the shift is expanded.
917   if (ShiftVT.getSizeInBits() < Log2_32_Ceil(LHSTy.getSizeInBits()))
918     ShiftVT = MVT::i32;
919   assert(ShiftVT.getSizeInBits() >= Log2_32_Ceil(LHSTy.getSizeInBits()) &&
920          "ShiftVT is still too small!");
921   return ShiftVT;
922 }
923 
924 bool TargetLoweringBase::canOpTrap(unsigned Op, EVT VT) const {
925   assert(isTypeLegal(VT));
926   switch (Op) {
927   default:
928     return false;
929   case ISD::SDIV:
930   case ISD::UDIV:
931   case ISD::SREM:
932   case ISD::UREM:
933     return true;
934   }
935 }
936 
937 bool TargetLoweringBase::isFreeAddrSpaceCast(unsigned SrcAS,
938                                              unsigned DestAS) const {
939   return TM.isNoopAddrSpaceCast(SrcAS, DestAS);
940 }
941 
942 void TargetLoweringBase::setJumpIsExpensive(bool isExpensive) {
943   // If the command-line option was specified, ignore this request.
944   if (!JumpIsExpensiveOverride.getNumOccurrences())
945     JumpIsExpensive = isExpensive;
946 }
947 
948 TargetLoweringBase::LegalizeKind
949 TargetLoweringBase::getTypeConversion(LLVMContext &Context, EVT VT) const {
950   // If this is a simple type, use the ComputeRegisterProp mechanism.
951   if (VT.isSimple()) {
952     MVT SVT = VT.getSimpleVT();
953     assert((unsigned)SVT.SimpleTy < array_lengthof(TransformToType));
954     MVT NVT = TransformToType[SVT.SimpleTy];
955     LegalizeTypeAction LA = ValueTypeActions.getTypeAction(SVT);
956 
957     assert((LA == TypeLegal || LA == TypeSoftenFloat ||
958             LA == TypeSoftPromoteHalf ||
959             (NVT.isVector() ||
960              ValueTypeActions.getTypeAction(NVT) != TypePromoteInteger)) &&
961            "Promote may not follow Expand or Promote");
962 
963     if (LA == TypeSplitVector)
964       return LegalizeKind(LA, EVT(SVT).getHalfNumVectorElementsVT(Context));
965     if (LA == TypeScalarizeVector)
966       return LegalizeKind(LA, SVT.getVectorElementType());
967     return LegalizeKind(LA, NVT);
968   }
969 
970   // Handle Extended Scalar Types.
971   if (!VT.isVector()) {
972     assert(VT.isInteger() && "Float types must be simple");
973     unsigned BitSize = VT.getSizeInBits();
974     // First promote to a power-of-two size, then expand if necessary.
975     if (BitSize < 8 || !isPowerOf2_32(BitSize)) {
976       EVT NVT = VT.getRoundIntegerType(Context);
977       assert(NVT != VT && "Unable to round integer VT");
978       LegalizeKind NextStep = getTypeConversion(Context, NVT);
979       // Avoid multi-step promotion.
980       if (NextStep.first == TypePromoteInteger)
981         return NextStep;
982       // Return rounded integer type.
983       return LegalizeKind(TypePromoteInteger, NVT);
984     }
985 
986     return LegalizeKind(TypeExpandInteger,
987                         EVT::getIntegerVT(Context, VT.getSizeInBits() / 2));
988   }
989 
990   // Handle vector types.
991   ElementCount NumElts = VT.getVectorElementCount();
992   EVT EltVT = VT.getVectorElementType();
993 
994   // Vectors with only one element are always scalarized.
995   if (NumElts.isScalar())
996     return LegalizeKind(TypeScalarizeVector, EltVT);
997 
998   // Try to widen vector elements until the element type is a power of two and
999   // promote it to a legal type later on, for example:
1000   // <3 x i8> -> <4 x i8> -> <4 x i32>
1001   if (EltVT.isInteger()) {
1002     // Vectors with a number of elements that is not a power of two are always
1003     // widened, for example <3 x i8> -> <4 x i8>.
1004     if (!VT.isPow2VectorType()) {
1005       NumElts = NumElts.coefficientNextPowerOf2();
1006       EVT NVT = EVT::getVectorVT(Context, EltVT, NumElts);
1007       return LegalizeKind(TypeWidenVector, NVT);
1008     }
1009 
1010     // Examine the element type.
1011     LegalizeKind LK = getTypeConversion(Context, EltVT);
1012 
1013     // If type is to be expanded, split the vector.
1014     //  <4 x i140> -> <2 x i140>
1015     if (LK.first == TypeExpandInteger) {
1016       if (VT.getVectorElementCount().isScalable())
1017         return LegalizeKind(TypeScalarizeScalableVector, EltVT);
1018       return LegalizeKind(TypeSplitVector,
1019                           VT.getHalfNumVectorElementsVT(Context));
1020     }
1021 
1022     // Promote the integer element types until a legal vector type is found
1023     // or until the element integer type is too big. If a legal type was not
1024     // found, fallback to the usual mechanism of widening/splitting the
1025     // vector.
1026     EVT OldEltVT = EltVT;
1027     while (true) {
1028       // Increase the bitwidth of the element to the next pow-of-two
1029       // (which is greater than 8 bits).
1030       EltVT = EVT::getIntegerVT(Context, 1 + EltVT.getSizeInBits())
1031                   .getRoundIntegerType(Context);
1032 
1033       // Stop trying when getting a non-simple element type.
1034       // Note that vector elements may be greater than legal vector element
1035       // types. Example: X86 XMM registers hold 64bit element on 32bit
1036       // systems.
1037       if (!EltVT.isSimple())
1038         break;
1039 
1040       // Build a new vector type and check if it is legal.
1041       MVT NVT = MVT::getVectorVT(EltVT.getSimpleVT(), NumElts);
1042       // Found a legal promoted vector type.
1043       if (NVT != MVT() && ValueTypeActions.getTypeAction(NVT) == TypeLegal)
1044         return LegalizeKind(TypePromoteInteger,
1045                             EVT::getVectorVT(Context, EltVT, NumElts));
1046     }
1047 
1048     // Reset the type to the unexpanded type if we did not find a legal vector
1049     // type with a promoted vector element type.
1050     EltVT = OldEltVT;
1051   }
1052 
1053   // Try to widen the vector until a legal type is found.
1054   // If there is no wider legal type, split the vector.
1055   while (true) {
1056     // Round up to the next power of 2.
1057     NumElts = NumElts.coefficientNextPowerOf2();
1058 
1059     // If there is no simple vector type with this many elements then there
1060     // cannot be a larger legal vector type.  Note that this assumes that
1061     // there are no skipped intermediate vector types in the simple types.
1062     if (!EltVT.isSimple())
1063       break;
1064     MVT LargerVector = MVT::getVectorVT(EltVT.getSimpleVT(), NumElts);
1065     if (LargerVector == MVT())
1066       break;
1067 
1068     // If this type is legal then widen the vector.
1069     if (ValueTypeActions.getTypeAction(LargerVector) == TypeLegal)
1070       return LegalizeKind(TypeWidenVector, LargerVector);
1071   }
1072 
1073   // Widen odd vectors to next power of two.
1074   if (!VT.isPow2VectorType()) {
1075     EVT NVT = VT.getPow2VectorType(Context);
1076     return LegalizeKind(TypeWidenVector, NVT);
1077   }
1078 
1079   if (VT.getVectorElementCount() == ElementCount::getScalable(1))
1080     return LegalizeKind(TypeScalarizeScalableVector, EltVT);
1081 
1082   // Vectors with illegal element types are expanded.
1083   EVT NVT = EVT::getVectorVT(Context, EltVT,
1084                              VT.getVectorElementCount().divideCoefficientBy(2));
1085   return LegalizeKind(TypeSplitVector, NVT);
1086 }
1087 
1088 static unsigned getVectorTypeBreakdownMVT(MVT VT, MVT &IntermediateVT,
1089                                           unsigned &NumIntermediates,
1090                                           MVT &RegisterVT,
1091                                           TargetLoweringBase *TLI) {
1092   // Figure out the right, legal destination reg to copy into.
1093   ElementCount EC = VT.getVectorElementCount();
1094   MVT EltTy = VT.getVectorElementType();
1095 
1096   unsigned NumVectorRegs = 1;
1097 
1098   // Scalable vectors cannot be scalarized, so splitting or widening is
1099   // required.
1100   if (VT.isScalableVector() && !isPowerOf2_32(EC.getKnownMinValue()))
1101     llvm_unreachable(
1102         "Splitting or widening of non-power-of-2 MVTs is not implemented.");
1103 
1104   // FIXME: We don't support non-power-of-2-sized vectors for now.
1105   // Ideally we could break down into LHS/RHS like LegalizeDAG does.
1106   if (!isPowerOf2_32(EC.getKnownMinValue())) {
1107     // Split EC to unit size (scalable property is preserved).
1108     NumVectorRegs = EC.getKnownMinValue();
1109     EC = ElementCount::getFixed(1);
1110   }
1111 
1112   // Divide the input until we get to a supported size. This will
1113   // always end up with an EC that represent a scalar or a scalable
1114   // scalar.
1115   while (EC.getKnownMinValue() > 1 &&
1116          !TLI->isTypeLegal(MVT::getVectorVT(EltTy, EC))) {
1117     EC = EC.divideCoefficientBy(2);
1118     NumVectorRegs <<= 1;
1119   }
1120 
1121   NumIntermediates = NumVectorRegs;
1122 
1123   MVT NewVT = MVT::getVectorVT(EltTy, EC);
1124   if (!TLI->isTypeLegal(NewVT))
1125     NewVT = EltTy;
1126   IntermediateVT = NewVT;
1127 
1128   unsigned LaneSizeInBits = NewVT.getScalarSizeInBits();
1129 
1130   // Convert sizes such as i33 to i64.
1131   if (!isPowerOf2_32(LaneSizeInBits))
1132     LaneSizeInBits = NextPowerOf2(LaneSizeInBits);
1133 
1134   MVT DestVT = TLI->getRegisterType(NewVT);
1135   RegisterVT = DestVT;
1136   if (EVT(DestVT).bitsLT(NewVT))    // Value is expanded, e.g. i64 -> i16.
1137     return NumVectorRegs * (LaneSizeInBits / DestVT.getScalarSizeInBits());
1138 
1139   // Otherwise, promotion or legal types use the same number of registers as
1140   // the vector decimated to the appropriate level.
1141   return NumVectorRegs;
1142 }
1143 
1144 /// isLegalRC - Return true if the value types that can be represented by the
1145 /// specified register class are all legal.
1146 bool TargetLoweringBase::isLegalRC(const TargetRegisterInfo &TRI,
1147                                    const TargetRegisterClass &RC) const {
1148   for (auto I = TRI.legalclasstypes_begin(RC); *I != MVT::Other; ++I)
1149     if (isTypeLegal(*I))
1150       return true;
1151   return false;
1152 }
1153 
1154 /// Replace/modify any TargetFrameIndex operands with a targte-dependent
1155 /// sequence of memory operands that is recognized by PrologEpilogInserter.
1156 MachineBasicBlock *
1157 TargetLoweringBase::emitPatchPoint(MachineInstr &InitialMI,
1158                                    MachineBasicBlock *MBB) const {
1159   MachineInstr *MI = &InitialMI;
1160   MachineFunction &MF = *MI->getMF();
1161   MachineFrameInfo &MFI = MF.getFrameInfo();
1162 
1163   // We're handling multiple types of operands here:
1164   // PATCHPOINT MetaArgs - live-in, read only, direct
1165   // STATEPOINT Deopt Spill - live-through, read only, indirect
1166   // STATEPOINT Deopt Alloca - live-through, read only, direct
1167   // (We're currently conservative and mark the deopt slots read/write in
1168   // practice.)
1169   // STATEPOINT GC Spill - live-through, read/write, indirect
1170   // STATEPOINT GC Alloca - live-through, read/write, direct
1171   // The live-in vs live-through is handled already (the live through ones are
1172   // all stack slots), but we need to handle the different type of stackmap
1173   // operands and memory effects here.
1174 
1175   if (llvm::none_of(MI->operands(),
1176                     [](MachineOperand &Operand) { return Operand.isFI(); }))
1177     return MBB;
1178 
1179   MachineInstrBuilder MIB = BuildMI(MF, MI->getDebugLoc(), MI->getDesc());
1180 
1181   // Inherit previous memory operands.
1182   MIB.cloneMemRefs(*MI);
1183 
1184   for (unsigned i = 0; i < MI->getNumOperands(); ++i) {
1185     MachineOperand &MO = MI->getOperand(i);
1186     if (!MO.isFI()) {
1187       // Index of Def operand this Use it tied to.
1188       // Since Defs are coming before Uses, if Use is tied, then
1189       // index of Def must be smaller that index of that Use.
1190       // Also, Defs preserve their position in new MI.
1191       unsigned TiedTo = i;
1192       if (MO.isReg() && MO.isTied())
1193         TiedTo = MI->findTiedOperandIdx(i);
1194       MIB.add(MO);
1195       if (TiedTo < i)
1196         MIB->tieOperands(TiedTo, MIB->getNumOperands() - 1);
1197       continue;
1198     }
1199 
1200     // foldMemoryOperand builds a new MI after replacing a single FI operand
1201     // with the canonical set of five x86 addressing-mode operands.
1202     int FI = MO.getIndex();
1203 
1204     // Add frame index operands recognized by stackmaps.cpp
1205     if (MFI.isStatepointSpillSlotObjectIndex(FI)) {
1206       // indirect-mem-ref tag, size, #FI, offset.
1207       // Used for spills inserted by StatepointLowering.  This codepath is not
1208       // used for patchpoints/stackmaps at all, for these spilling is done via
1209       // foldMemoryOperand callback only.
1210       assert(MI->getOpcode() == TargetOpcode::STATEPOINT && "sanity");
1211       MIB.addImm(StackMaps::IndirectMemRefOp);
1212       MIB.addImm(MFI.getObjectSize(FI));
1213       MIB.add(MO);
1214       MIB.addImm(0);
1215     } else {
1216       // direct-mem-ref tag, #FI, offset.
1217       // Used by patchpoint, and direct alloca arguments to statepoints
1218       MIB.addImm(StackMaps::DirectMemRefOp);
1219       MIB.add(MO);
1220       MIB.addImm(0);
1221     }
1222 
1223     assert(MIB->mayLoad() && "Folded a stackmap use to a non-load!");
1224 
1225     // Add a new memory operand for this FI.
1226     assert(MFI.getObjectOffset(FI) != -1);
1227 
1228     // Note: STATEPOINT MMOs are added during SelectionDAG.  STACKMAP, and
1229     // PATCHPOINT should be updated to do the same. (TODO)
1230     if (MI->getOpcode() != TargetOpcode::STATEPOINT) {
1231       auto Flags = MachineMemOperand::MOLoad;
1232       MachineMemOperand *MMO = MF.getMachineMemOperand(
1233           MachinePointerInfo::getFixedStack(MF, FI), Flags,
1234           MF.getDataLayout().getPointerSize(), MFI.getObjectAlign(FI));
1235       MIB->addMemOperand(MF, MMO);
1236     }
1237   }
1238   MBB->insert(MachineBasicBlock::iterator(MI), MIB);
1239   MI->eraseFromParent();
1240   return MBB;
1241 }
1242 
1243 /// findRepresentativeClass - Return the largest legal super-reg register class
1244 /// of the register class for the specified type and its associated "cost".
1245 // This function is in TargetLowering because it uses RegClassForVT which would
1246 // need to be moved to TargetRegisterInfo and would necessitate moving
1247 // isTypeLegal over as well - a massive change that would just require
1248 // TargetLowering having a TargetRegisterInfo class member that it would use.
1249 std::pair<const TargetRegisterClass *, uint8_t>
1250 TargetLoweringBase::findRepresentativeClass(const TargetRegisterInfo *TRI,
1251                                             MVT VT) const {
1252   const TargetRegisterClass *RC = RegClassForVT[VT.SimpleTy];
1253   if (!RC)
1254     return std::make_pair(RC, 0);
1255 
1256   // Compute the set of all super-register classes.
1257   BitVector SuperRegRC(TRI->getNumRegClasses());
1258   for (SuperRegClassIterator RCI(RC, TRI); RCI.isValid(); ++RCI)
1259     SuperRegRC.setBitsInMask(RCI.getMask());
1260 
1261   // Find the first legal register class with the largest spill size.
1262   const TargetRegisterClass *BestRC = RC;
1263   for (unsigned i : SuperRegRC.set_bits()) {
1264     const TargetRegisterClass *SuperRC = TRI->getRegClass(i);
1265     // We want the largest possible spill size.
1266     if (TRI->getSpillSize(*SuperRC) <= TRI->getSpillSize(*BestRC))
1267       continue;
1268     if (!isLegalRC(*TRI, *SuperRC))
1269       continue;
1270     BestRC = SuperRC;
1271   }
1272   return std::make_pair(BestRC, 1);
1273 }
1274 
1275 /// computeRegisterProperties - Once all of the register classes are added,
1276 /// this allows us to compute derived properties we expose.
1277 void TargetLoweringBase::computeRegisterProperties(
1278     const TargetRegisterInfo *TRI) {
1279   static_assert(MVT::VALUETYPE_SIZE <= MVT::MAX_ALLOWED_VALUETYPE,
1280                 "Too many value types for ValueTypeActions to hold!");
1281 
1282   // Everything defaults to needing one register.
1283   for (unsigned i = 0; i != MVT::VALUETYPE_SIZE; ++i) {
1284     NumRegistersForVT[i] = 1;
1285     RegisterTypeForVT[i] = TransformToType[i] = (MVT::SimpleValueType)i;
1286   }
1287   // ...except isVoid, which doesn't need any registers.
1288   NumRegistersForVT[MVT::isVoid] = 0;
1289 
1290   // Find the largest integer register class.
1291   unsigned LargestIntReg = MVT::LAST_INTEGER_VALUETYPE;
1292   for (; RegClassForVT[LargestIntReg] == nullptr; --LargestIntReg)
1293     assert(LargestIntReg != MVT::i1 && "No integer registers defined!");
1294 
1295   // Every integer value type larger than this largest register takes twice as
1296   // many registers to represent as the previous ValueType.
1297   for (unsigned ExpandedReg = LargestIntReg + 1;
1298        ExpandedReg <= MVT::LAST_INTEGER_VALUETYPE; ++ExpandedReg) {
1299     NumRegistersForVT[ExpandedReg] = 2*NumRegistersForVT[ExpandedReg-1];
1300     RegisterTypeForVT[ExpandedReg] = (MVT::SimpleValueType)LargestIntReg;
1301     TransformToType[ExpandedReg] = (MVT::SimpleValueType)(ExpandedReg - 1);
1302     ValueTypeActions.setTypeAction((MVT::SimpleValueType)ExpandedReg,
1303                                    TypeExpandInteger);
1304   }
1305 
1306   // Inspect all of the ValueType's smaller than the largest integer
1307   // register to see which ones need promotion.
1308   unsigned LegalIntReg = LargestIntReg;
1309   for (unsigned IntReg = LargestIntReg - 1;
1310        IntReg >= (unsigned)MVT::i1; --IntReg) {
1311     MVT IVT = (MVT::SimpleValueType)IntReg;
1312     if (isTypeLegal(IVT)) {
1313       LegalIntReg = IntReg;
1314     } else {
1315       RegisterTypeForVT[IntReg] = TransformToType[IntReg] =
1316         (MVT::SimpleValueType)LegalIntReg;
1317       ValueTypeActions.setTypeAction(IVT, TypePromoteInteger);
1318     }
1319   }
1320 
1321   // ppcf128 type is really two f64's.
1322   if (!isTypeLegal(MVT::ppcf128)) {
1323     if (isTypeLegal(MVT::f64)) {
1324       NumRegistersForVT[MVT::ppcf128] = 2*NumRegistersForVT[MVT::f64];
1325       RegisterTypeForVT[MVT::ppcf128] = MVT::f64;
1326       TransformToType[MVT::ppcf128] = MVT::f64;
1327       ValueTypeActions.setTypeAction(MVT::ppcf128, TypeExpandFloat);
1328     } else {
1329       NumRegistersForVT[MVT::ppcf128] = NumRegistersForVT[MVT::i128];
1330       RegisterTypeForVT[MVT::ppcf128] = RegisterTypeForVT[MVT::i128];
1331       TransformToType[MVT::ppcf128] = MVT::i128;
1332       ValueTypeActions.setTypeAction(MVT::ppcf128, TypeSoftenFloat);
1333     }
1334   }
1335 
1336   // Decide how to handle f128. If the target does not have native f128 support,
1337   // expand it to i128 and we will be generating soft float library calls.
1338   if (!isTypeLegal(MVT::f128)) {
1339     NumRegistersForVT[MVT::f128] = NumRegistersForVT[MVT::i128];
1340     RegisterTypeForVT[MVT::f128] = RegisterTypeForVT[MVT::i128];
1341     TransformToType[MVT::f128] = MVT::i128;
1342     ValueTypeActions.setTypeAction(MVT::f128, TypeSoftenFloat);
1343   }
1344 
1345   // Decide how to handle f64. If the target does not have native f64 support,
1346   // expand it to i64 and we will be generating soft float library calls.
1347   if (!isTypeLegal(MVT::f64)) {
1348     NumRegistersForVT[MVT::f64] = NumRegistersForVT[MVT::i64];
1349     RegisterTypeForVT[MVT::f64] = RegisterTypeForVT[MVT::i64];
1350     TransformToType[MVT::f64] = MVT::i64;
1351     ValueTypeActions.setTypeAction(MVT::f64, TypeSoftenFloat);
1352   }
1353 
1354   // Decide how to handle f32. If the target does not have native f32 support,
1355   // expand it to i32 and we will be generating soft float library calls.
1356   if (!isTypeLegal(MVT::f32)) {
1357     NumRegistersForVT[MVT::f32] = NumRegistersForVT[MVT::i32];
1358     RegisterTypeForVT[MVT::f32] = RegisterTypeForVT[MVT::i32];
1359     TransformToType[MVT::f32] = MVT::i32;
1360     ValueTypeActions.setTypeAction(MVT::f32, TypeSoftenFloat);
1361   }
1362 
1363   // Decide how to handle f16. If the target does not have native f16 support,
1364   // promote it to f32, because there are no f16 library calls (except for
1365   // conversions).
1366   if (!isTypeLegal(MVT::f16)) {
1367     // Allow targets to control how we legalize half.
1368     if (softPromoteHalfType()) {
1369       NumRegistersForVT[MVT::f16] = NumRegistersForVT[MVT::i16];
1370       RegisterTypeForVT[MVT::f16] = RegisterTypeForVT[MVT::i16];
1371       TransformToType[MVT::f16] = MVT::f32;
1372       ValueTypeActions.setTypeAction(MVT::f16, TypeSoftPromoteHalf);
1373     } else {
1374       NumRegistersForVT[MVT::f16] = NumRegistersForVT[MVT::f32];
1375       RegisterTypeForVT[MVT::f16] = RegisterTypeForVT[MVT::f32];
1376       TransformToType[MVT::f16] = MVT::f32;
1377       ValueTypeActions.setTypeAction(MVT::f16, TypePromoteFloat);
1378     }
1379   }
1380 
1381   // Decide how to handle bf16. If the target does not have native bf16 support,
1382   // promote it to f32, because there are no bf16 library calls (except for
1383   // converting from f32 to bf16).
1384   if (!isTypeLegal(MVT::bf16)) {
1385     NumRegistersForVT[MVT::bf16] = NumRegistersForVT[MVT::f32];
1386     RegisterTypeForVT[MVT::bf16] = RegisterTypeForVT[MVT::f32];
1387     TransformToType[MVT::bf16] = MVT::f32;
1388     ValueTypeActions.setTypeAction(MVT::bf16, TypePromoteFloat);
1389   }
1390 
1391   // Loop over all of the vector value types to see which need transformations.
1392   for (unsigned i = MVT::FIRST_VECTOR_VALUETYPE;
1393        i <= (unsigned)MVT::LAST_VECTOR_VALUETYPE; ++i) {
1394     MVT VT = (MVT::SimpleValueType) i;
1395     if (isTypeLegal(VT))
1396       continue;
1397 
1398     MVT EltVT = VT.getVectorElementType();
1399     ElementCount EC = VT.getVectorElementCount();
1400     bool IsLegalWiderType = false;
1401     bool IsScalable = VT.isScalableVector();
1402     LegalizeTypeAction PreferredAction = getPreferredVectorAction(VT);
1403     switch (PreferredAction) {
1404     case TypePromoteInteger: {
1405       MVT::SimpleValueType EndVT = IsScalable ?
1406                                    MVT::LAST_INTEGER_SCALABLE_VECTOR_VALUETYPE :
1407                                    MVT::LAST_INTEGER_FIXEDLEN_VECTOR_VALUETYPE;
1408       // Try to promote the elements of integer vectors. If no legal
1409       // promotion was found, fall through to the widen-vector method.
1410       for (unsigned nVT = i + 1;
1411            (MVT::SimpleValueType)nVT <= EndVT; ++nVT) {
1412         MVT SVT = (MVT::SimpleValueType) nVT;
1413         // Promote vectors of integers to vectors with the same number
1414         // of elements, with a wider element type.
1415         if (SVT.getScalarSizeInBits() > EltVT.getFixedSizeInBits() &&
1416             SVT.getVectorElementCount() == EC && isTypeLegal(SVT)) {
1417           TransformToType[i] = SVT;
1418           RegisterTypeForVT[i] = SVT;
1419           NumRegistersForVT[i] = 1;
1420           ValueTypeActions.setTypeAction(VT, TypePromoteInteger);
1421           IsLegalWiderType = true;
1422           break;
1423         }
1424       }
1425       if (IsLegalWiderType)
1426         break;
1427       LLVM_FALLTHROUGH;
1428     }
1429 
1430     case TypeWidenVector:
1431       if (isPowerOf2_32(EC.getKnownMinValue())) {
1432         // Try to widen the vector.
1433         for (unsigned nVT = i + 1; nVT <= MVT::LAST_VECTOR_VALUETYPE; ++nVT) {
1434           MVT SVT = (MVT::SimpleValueType) nVT;
1435           if (SVT.getVectorElementType() == EltVT &&
1436               SVT.isScalableVector() == IsScalable &&
1437               SVT.getVectorElementCount().getKnownMinValue() >
1438                   EC.getKnownMinValue() &&
1439               isTypeLegal(SVT)) {
1440             TransformToType[i] = SVT;
1441             RegisterTypeForVT[i] = SVT;
1442             NumRegistersForVT[i] = 1;
1443             ValueTypeActions.setTypeAction(VT, TypeWidenVector);
1444             IsLegalWiderType = true;
1445             break;
1446           }
1447         }
1448         if (IsLegalWiderType)
1449           break;
1450       } else {
1451         // Only widen to the next power of 2 to keep consistency with EVT.
1452         MVT NVT = VT.getPow2VectorType();
1453         if (isTypeLegal(NVT)) {
1454           TransformToType[i] = NVT;
1455           ValueTypeActions.setTypeAction(VT, TypeWidenVector);
1456           RegisterTypeForVT[i] = NVT;
1457           NumRegistersForVT[i] = 1;
1458           break;
1459         }
1460       }
1461       LLVM_FALLTHROUGH;
1462 
1463     case TypeSplitVector:
1464     case TypeScalarizeVector: {
1465       MVT IntermediateVT;
1466       MVT RegisterVT;
1467       unsigned NumIntermediates;
1468       unsigned NumRegisters = getVectorTypeBreakdownMVT(VT, IntermediateVT,
1469           NumIntermediates, RegisterVT, this);
1470       NumRegistersForVT[i] = NumRegisters;
1471       assert(NumRegistersForVT[i] == NumRegisters &&
1472              "NumRegistersForVT size cannot represent NumRegisters!");
1473       RegisterTypeForVT[i] = RegisterVT;
1474 
1475       MVT NVT = VT.getPow2VectorType();
1476       if (NVT == VT) {
1477         // Type is already a power of 2.  The default action is to split.
1478         TransformToType[i] = MVT::Other;
1479         if (PreferredAction == TypeScalarizeVector)
1480           ValueTypeActions.setTypeAction(VT, TypeScalarizeVector);
1481         else if (PreferredAction == TypeSplitVector)
1482           ValueTypeActions.setTypeAction(VT, TypeSplitVector);
1483         else if (EC.getKnownMinValue() > 1)
1484           ValueTypeActions.setTypeAction(VT, TypeSplitVector);
1485         else
1486           ValueTypeActions.setTypeAction(VT, EC.isScalable()
1487                                                  ? TypeScalarizeScalableVector
1488                                                  : TypeScalarizeVector);
1489       } else {
1490         TransformToType[i] = NVT;
1491         ValueTypeActions.setTypeAction(VT, TypeWidenVector);
1492       }
1493       break;
1494     }
1495     default:
1496       llvm_unreachable("Unknown vector legalization action!");
1497     }
1498   }
1499 
1500   // Determine the 'representative' register class for each value type.
1501   // An representative register class is the largest (meaning one which is
1502   // not a sub-register class / subreg register class) legal register class for
1503   // a group of value types. For example, on i386, i8, i16, and i32
1504   // representative would be GR32; while on x86_64 it's GR64.
1505   for (unsigned i = 0; i != MVT::VALUETYPE_SIZE; ++i) {
1506     const TargetRegisterClass* RRC;
1507     uint8_t Cost;
1508     std::tie(RRC, Cost) = findRepresentativeClass(TRI, (MVT::SimpleValueType)i);
1509     RepRegClassForVT[i] = RRC;
1510     RepRegClassCostForVT[i] = Cost;
1511   }
1512 }
1513 
1514 EVT TargetLoweringBase::getSetCCResultType(const DataLayout &DL, LLVMContext &,
1515                                            EVT VT) const {
1516   assert(!VT.isVector() && "No default SetCC type for vectors!");
1517   return getPointerTy(DL).SimpleTy;
1518 }
1519 
1520 MVT::SimpleValueType TargetLoweringBase::getCmpLibcallReturnType() const {
1521   return MVT::i32; // return the default value
1522 }
1523 
1524 /// getVectorTypeBreakdown - Vector types are broken down into some number of
1525 /// legal first class types.  For example, MVT::v8f32 maps to 2 MVT::v4f32
1526 /// with Altivec or SSE1, or 8 promoted MVT::f64 values with the X86 FP stack.
1527 /// Similarly, MVT::v2i64 turns into 4 MVT::i32 values with both PPC and X86.
1528 ///
1529 /// This method returns the number of registers needed, and the VT for each
1530 /// register.  It also returns the VT and quantity of the intermediate values
1531 /// before they are promoted/expanded.
1532 unsigned TargetLoweringBase::getVectorTypeBreakdown(LLVMContext &Context,
1533                                                     EVT VT, EVT &IntermediateVT,
1534                                                     unsigned &NumIntermediates,
1535                                                     MVT &RegisterVT) const {
1536   ElementCount EltCnt = VT.getVectorElementCount();
1537 
1538   // If there is a wider vector type with the same element type as this one,
1539   // or a promoted vector type that has the same number of elements which
1540   // are wider, then we should convert to that legal vector type.
1541   // This handles things like <2 x float> -> <4 x float> and
1542   // <4 x i1> -> <4 x i32>.
1543   LegalizeTypeAction TA = getTypeAction(Context, VT);
1544   if (!EltCnt.isScalar() &&
1545       (TA == TypeWidenVector || TA == TypePromoteInteger)) {
1546     EVT RegisterEVT = getTypeToTransformTo(Context, VT);
1547     if (isTypeLegal(RegisterEVT)) {
1548       IntermediateVT = RegisterEVT;
1549       RegisterVT = RegisterEVT.getSimpleVT();
1550       NumIntermediates = 1;
1551       return 1;
1552     }
1553   }
1554 
1555   // Figure out the right, legal destination reg to copy into.
1556   EVT EltTy = VT.getVectorElementType();
1557 
1558   unsigned NumVectorRegs = 1;
1559 
1560   // Scalable vectors cannot be scalarized, so handle the legalisation of the
1561   // types like done elsewhere in SelectionDAG.
1562   if (EltCnt.isScalable()) {
1563     LegalizeKind LK;
1564     EVT PartVT = VT;
1565     do {
1566       // Iterate until we've found a legal (part) type to hold VT.
1567       LK = getTypeConversion(Context, PartVT);
1568       PartVT = LK.second;
1569     } while (LK.first != TypeLegal);
1570 
1571     if (!PartVT.isVector()) {
1572       report_fatal_error(
1573           "Don't know how to legalize this scalable vector type");
1574     }
1575 
1576     NumIntermediates =
1577         divideCeil(VT.getVectorElementCount().getKnownMinValue(),
1578                    PartVT.getVectorElementCount().getKnownMinValue());
1579     IntermediateVT = PartVT;
1580     RegisterVT = getRegisterType(Context, IntermediateVT);
1581     return NumIntermediates;
1582   }
1583 
1584   // FIXME: We don't support non-power-of-2-sized vectors for now.  Ideally
1585   // we could break down into LHS/RHS like LegalizeDAG does.
1586   if (!isPowerOf2_32(EltCnt.getKnownMinValue())) {
1587     NumVectorRegs = EltCnt.getKnownMinValue();
1588     EltCnt = ElementCount::getFixed(1);
1589   }
1590 
1591   // Divide the input until we get to a supported size.  This will always
1592   // end with a scalar if the target doesn't support vectors.
1593   while (EltCnt.getKnownMinValue() > 1 &&
1594          !isTypeLegal(EVT::getVectorVT(Context, EltTy, EltCnt))) {
1595     EltCnt = EltCnt.divideCoefficientBy(2);
1596     NumVectorRegs <<= 1;
1597   }
1598 
1599   NumIntermediates = NumVectorRegs;
1600 
1601   EVT NewVT = EVT::getVectorVT(Context, EltTy, EltCnt);
1602   if (!isTypeLegal(NewVT))
1603     NewVT = EltTy;
1604   IntermediateVT = NewVT;
1605 
1606   MVT DestVT = getRegisterType(Context, NewVT);
1607   RegisterVT = DestVT;
1608 
1609   if (EVT(DestVT).bitsLT(NewVT)) {  // Value is expanded, e.g. i64 -> i16.
1610     TypeSize NewVTSize = NewVT.getSizeInBits();
1611     // Convert sizes such as i33 to i64.
1612     if (!isPowerOf2_32(NewVTSize.getKnownMinSize()))
1613       NewVTSize = NewVTSize.coefficientNextPowerOf2();
1614     return NumVectorRegs*(NewVTSize/DestVT.getSizeInBits());
1615   }
1616 
1617   // Otherwise, promotion or legal types use the same number of registers as
1618   // the vector decimated to the appropriate level.
1619   return NumVectorRegs;
1620 }
1621 
1622 bool TargetLoweringBase::isSuitableForJumpTable(const SwitchInst *SI,
1623                                                 uint64_t NumCases,
1624                                                 uint64_t Range,
1625                                                 ProfileSummaryInfo *PSI,
1626                                                 BlockFrequencyInfo *BFI) const {
1627   // FIXME: This function check the maximum table size and density, but the
1628   // minimum size is not checked. It would be nice if the minimum size is
1629   // also combined within this function. Currently, the minimum size check is
1630   // performed in findJumpTable() in SelectionDAGBuiler and
1631   // getEstimatedNumberOfCaseClusters() in BasicTTIImpl.
1632   const bool OptForSize =
1633       SI->getParent()->getParent()->hasOptSize() ||
1634       llvm::shouldOptimizeForSize(SI->getParent(), PSI, BFI);
1635   const unsigned MinDensity = getMinimumJumpTableDensity(OptForSize);
1636   const unsigned MaxJumpTableSize = getMaximumJumpTableSize();
1637 
1638   // Check whether the number of cases is small enough and
1639   // the range is dense enough for a jump table.
1640   return (OptForSize || Range <= MaxJumpTableSize) &&
1641          (NumCases * 100 >= Range * MinDensity);
1642 }
1643 
1644 MVT TargetLoweringBase::getPreferredSwitchConditionType(LLVMContext &Context,
1645                                                         EVT ConditionVT) const {
1646   return getRegisterType(Context, ConditionVT);
1647 }
1648 
1649 /// Get the EVTs and ArgFlags collections that represent the legalized return
1650 /// type of the given function.  This does not require a DAG or a return value,
1651 /// and is suitable for use before any DAGs for the function are constructed.
1652 /// TODO: Move this out of TargetLowering.cpp.
1653 void llvm::GetReturnInfo(CallingConv::ID CC, Type *ReturnType,
1654                          AttributeList attr,
1655                          SmallVectorImpl<ISD::OutputArg> &Outs,
1656                          const TargetLowering &TLI, const DataLayout &DL) {
1657   SmallVector<EVT, 4> ValueVTs;
1658   ComputeValueVTs(TLI, DL, ReturnType, ValueVTs);
1659   unsigned NumValues = ValueVTs.size();
1660   if (NumValues == 0) return;
1661 
1662   for (unsigned j = 0, f = NumValues; j != f; ++j) {
1663     EVT VT = ValueVTs[j];
1664     ISD::NodeType ExtendKind = ISD::ANY_EXTEND;
1665 
1666     if (attr.hasRetAttr(Attribute::SExt))
1667       ExtendKind = ISD::SIGN_EXTEND;
1668     else if (attr.hasRetAttr(Attribute::ZExt))
1669       ExtendKind = ISD::ZERO_EXTEND;
1670 
1671     // FIXME: C calling convention requires the return type to be promoted to
1672     // at least 32-bit. But this is not necessary for non-C calling
1673     // conventions. The frontend should mark functions whose return values
1674     // require promoting with signext or zeroext attributes.
1675     if (ExtendKind != ISD::ANY_EXTEND && VT.isInteger()) {
1676       MVT MinVT = TLI.getRegisterType(ReturnType->getContext(), MVT::i32);
1677       if (VT.bitsLT(MinVT))
1678         VT = MinVT;
1679     }
1680 
1681     unsigned NumParts =
1682         TLI.getNumRegistersForCallingConv(ReturnType->getContext(), CC, VT);
1683     MVT PartVT =
1684         TLI.getRegisterTypeForCallingConv(ReturnType->getContext(), CC, VT);
1685 
1686     // 'inreg' on function refers to return value
1687     ISD::ArgFlagsTy Flags = ISD::ArgFlagsTy();
1688     if (attr.hasRetAttr(Attribute::InReg))
1689       Flags.setInReg();
1690 
1691     // Propagate extension type if any
1692     if (attr.hasRetAttr(Attribute::SExt))
1693       Flags.setSExt();
1694     else if (attr.hasRetAttr(Attribute::ZExt))
1695       Flags.setZExt();
1696 
1697     for (unsigned i = 0; i < NumParts; ++i)
1698       Outs.push_back(ISD::OutputArg(Flags, PartVT, VT, /*isfixed=*/true, 0, 0));
1699   }
1700 }
1701 
1702 /// getByValTypeAlignment - Return the desired alignment for ByVal aggregate
1703 /// function arguments in the caller parameter area.  This is the actual
1704 /// alignment, not its logarithm.
1705 uint64_t TargetLoweringBase::getByValTypeAlignment(Type *Ty,
1706                                                    const DataLayout &DL) const {
1707   return DL.getABITypeAlign(Ty).value();
1708 }
1709 
1710 bool TargetLoweringBase::allowsMemoryAccessForAlignment(
1711     LLVMContext &Context, const DataLayout &DL, EVT VT, unsigned AddrSpace,
1712     Align Alignment, MachineMemOperand::Flags Flags, bool *Fast) const {
1713   // Check if the specified alignment is sufficient based on the data layout.
1714   // TODO: While using the data layout works in practice, a better solution
1715   // would be to implement this check directly (make this a virtual function).
1716   // For example, the ABI alignment may change based on software platform while
1717   // this function should only be affected by hardware implementation.
1718   Type *Ty = VT.getTypeForEVT(Context);
1719   if (VT.isZeroSized() || Alignment >= DL.getABITypeAlign(Ty)) {
1720     // Assume that an access that meets the ABI-specified alignment is fast.
1721     if (Fast != nullptr)
1722       *Fast = true;
1723     return true;
1724   }
1725 
1726   // This is a misaligned access.
1727   return allowsMisalignedMemoryAccesses(VT, AddrSpace, Alignment, Flags, Fast);
1728 }
1729 
1730 bool TargetLoweringBase::allowsMemoryAccessForAlignment(
1731     LLVMContext &Context, const DataLayout &DL, EVT VT,
1732     const MachineMemOperand &MMO, bool *Fast) const {
1733   return allowsMemoryAccessForAlignment(Context, DL, VT, MMO.getAddrSpace(),
1734                                         MMO.getAlign(), MMO.getFlags(), Fast);
1735 }
1736 
1737 bool TargetLoweringBase::allowsMemoryAccess(LLVMContext &Context,
1738                                             const DataLayout &DL, EVT VT,
1739                                             unsigned AddrSpace, Align Alignment,
1740                                             MachineMemOperand::Flags Flags,
1741                                             bool *Fast) const {
1742   return allowsMemoryAccessForAlignment(Context, DL, VT, AddrSpace, Alignment,
1743                                         Flags, Fast);
1744 }
1745 
1746 bool TargetLoweringBase::allowsMemoryAccess(LLVMContext &Context,
1747                                             const DataLayout &DL, EVT VT,
1748                                             const MachineMemOperand &MMO,
1749                                             bool *Fast) const {
1750   return allowsMemoryAccess(Context, DL, VT, MMO.getAddrSpace(), MMO.getAlign(),
1751                             MMO.getFlags(), Fast);
1752 }
1753 
1754 bool TargetLoweringBase::allowsMemoryAccess(LLVMContext &Context,
1755                                             const DataLayout &DL, LLT Ty,
1756                                             const MachineMemOperand &MMO,
1757                                             bool *Fast) const {
1758   EVT VT = getApproximateEVTForLLT(Ty, DL, Context);
1759   return allowsMemoryAccess(Context, DL, VT, MMO.getAddrSpace(), MMO.getAlign(),
1760                             MMO.getFlags(), Fast);
1761 }
1762 
1763 //===----------------------------------------------------------------------===//
1764 //  TargetTransformInfo Helpers
1765 //===----------------------------------------------------------------------===//
1766 
1767 int TargetLoweringBase::InstructionOpcodeToISD(unsigned Opcode) const {
1768   enum InstructionOpcodes {
1769 #define HANDLE_INST(NUM, OPCODE, CLASS) OPCODE = NUM,
1770 #define LAST_OTHER_INST(NUM) InstructionOpcodesCount = NUM
1771 #include "llvm/IR/Instruction.def"
1772   };
1773   switch (static_cast<InstructionOpcodes>(Opcode)) {
1774   case Ret:            return 0;
1775   case Br:             return 0;
1776   case Switch:         return 0;
1777   case IndirectBr:     return 0;
1778   case Invoke:         return 0;
1779   case CallBr:         return 0;
1780   case Resume:         return 0;
1781   case Unreachable:    return 0;
1782   case CleanupRet:     return 0;
1783   case CatchRet:       return 0;
1784   case CatchPad:       return 0;
1785   case CatchSwitch:    return 0;
1786   case CleanupPad:     return 0;
1787   case FNeg:           return ISD::FNEG;
1788   case Add:            return ISD::ADD;
1789   case FAdd:           return ISD::FADD;
1790   case Sub:            return ISD::SUB;
1791   case FSub:           return ISD::FSUB;
1792   case Mul:            return ISD::MUL;
1793   case FMul:           return ISD::FMUL;
1794   case UDiv:           return ISD::UDIV;
1795   case SDiv:           return ISD::SDIV;
1796   case FDiv:           return ISD::FDIV;
1797   case URem:           return ISD::UREM;
1798   case SRem:           return ISD::SREM;
1799   case FRem:           return ISD::FREM;
1800   case Shl:            return ISD::SHL;
1801   case LShr:           return ISD::SRL;
1802   case AShr:           return ISD::SRA;
1803   case And:            return ISD::AND;
1804   case Or:             return ISD::OR;
1805   case Xor:            return ISD::XOR;
1806   case Alloca:         return 0;
1807   case Load:           return ISD::LOAD;
1808   case Store:          return ISD::STORE;
1809   case GetElementPtr:  return 0;
1810   case Fence:          return 0;
1811   case AtomicCmpXchg:  return 0;
1812   case AtomicRMW:      return 0;
1813   case Trunc:          return ISD::TRUNCATE;
1814   case ZExt:           return ISD::ZERO_EXTEND;
1815   case SExt:           return ISD::SIGN_EXTEND;
1816   case FPToUI:         return ISD::FP_TO_UINT;
1817   case FPToSI:         return ISD::FP_TO_SINT;
1818   case UIToFP:         return ISD::UINT_TO_FP;
1819   case SIToFP:         return ISD::SINT_TO_FP;
1820   case FPTrunc:        return ISD::FP_ROUND;
1821   case FPExt:          return ISD::FP_EXTEND;
1822   case PtrToInt:       return ISD::BITCAST;
1823   case IntToPtr:       return ISD::BITCAST;
1824   case BitCast:        return ISD::BITCAST;
1825   case AddrSpaceCast:  return ISD::ADDRSPACECAST;
1826   case ICmp:           return ISD::SETCC;
1827   case FCmp:           return ISD::SETCC;
1828   case PHI:            return 0;
1829   case Call:           return 0;
1830   case Select:         return ISD::SELECT;
1831   case UserOp1:        return 0;
1832   case UserOp2:        return 0;
1833   case VAArg:          return 0;
1834   case ExtractElement: return ISD::EXTRACT_VECTOR_ELT;
1835   case InsertElement:  return ISD::INSERT_VECTOR_ELT;
1836   case ShuffleVector:  return ISD::VECTOR_SHUFFLE;
1837   case ExtractValue:   return ISD::MERGE_VALUES;
1838   case InsertValue:    return ISD::MERGE_VALUES;
1839   case LandingPad:     return 0;
1840   case Freeze:         return ISD::FREEZE;
1841   }
1842 
1843   llvm_unreachable("Unknown instruction type encountered!");
1844 }
1845 
1846 std::pair<InstructionCost, MVT>
1847 TargetLoweringBase::getTypeLegalizationCost(const DataLayout &DL,
1848                                             Type *Ty) const {
1849   LLVMContext &C = Ty->getContext();
1850   EVT MTy = getValueType(DL, Ty);
1851 
1852   InstructionCost Cost = 1;
1853   // We keep legalizing the type until we find a legal kind. We assume that
1854   // the only operation that costs anything is the split. After splitting
1855   // we need to handle two types.
1856   while (true) {
1857     LegalizeKind LK = getTypeConversion(C, MTy);
1858 
1859     if (LK.first == TypeScalarizeScalableVector) {
1860       // Ensure we return a sensible simple VT here, since many callers of this
1861       // function require it.
1862       MVT VT = MTy.isSimple() ? MTy.getSimpleVT() : MVT::i64;
1863       return std::make_pair(InstructionCost::getInvalid(), VT);
1864     }
1865 
1866     if (LK.first == TypeLegal)
1867       return std::make_pair(Cost, MTy.getSimpleVT());
1868 
1869     if (LK.first == TypeSplitVector || LK.first == TypeExpandInteger)
1870       Cost *= 2;
1871 
1872     // Do not loop with f128 type.
1873     if (MTy == LK.second)
1874       return std::make_pair(Cost, MTy.getSimpleVT());
1875 
1876     // Keep legalizing the type.
1877     MTy = LK.second;
1878   }
1879 }
1880 
1881 Value *
1882 TargetLoweringBase::getDefaultSafeStackPointerLocation(IRBuilderBase &IRB,
1883                                                        bool UseTLS) const {
1884   // compiler-rt provides a variable with a magic name.  Targets that do not
1885   // link with compiler-rt may also provide such a variable.
1886   Module *M = IRB.GetInsertBlock()->getParent()->getParent();
1887   const char *UnsafeStackPtrVar = "__safestack_unsafe_stack_ptr";
1888   auto UnsafeStackPtr =
1889       dyn_cast_or_null<GlobalVariable>(M->getNamedValue(UnsafeStackPtrVar));
1890 
1891   Type *StackPtrTy = Type::getInt8PtrTy(M->getContext());
1892 
1893   if (!UnsafeStackPtr) {
1894     auto TLSModel = UseTLS ?
1895         GlobalValue::InitialExecTLSModel :
1896         GlobalValue::NotThreadLocal;
1897     // The global variable is not defined yet, define it ourselves.
1898     // We use the initial-exec TLS model because we do not support the
1899     // variable living anywhere other than in the main executable.
1900     UnsafeStackPtr = new GlobalVariable(
1901         *M, StackPtrTy, false, GlobalValue::ExternalLinkage, nullptr,
1902         UnsafeStackPtrVar, nullptr, TLSModel);
1903   } else {
1904     // The variable exists, check its type and attributes.
1905     if (UnsafeStackPtr->getValueType() != StackPtrTy)
1906       report_fatal_error(Twine(UnsafeStackPtrVar) + " must have void* type");
1907     if (UseTLS != UnsafeStackPtr->isThreadLocal())
1908       report_fatal_error(Twine(UnsafeStackPtrVar) + " must " +
1909                          (UseTLS ? "" : "not ") + "be thread-local");
1910   }
1911   return UnsafeStackPtr;
1912 }
1913 
1914 Value *
1915 TargetLoweringBase::getSafeStackPointerLocation(IRBuilderBase &IRB) const {
1916   if (!TM.getTargetTriple().isAndroid())
1917     return getDefaultSafeStackPointerLocation(IRB, true);
1918 
1919   // Android provides a libc function to retrieve the address of the current
1920   // thread's unsafe stack pointer.
1921   Module *M = IRB.GetInsertBlock()->getParent()->getParent();
1922   Type *StackPtrTy = Type::getInt8PtrTy(M->getContext());
1923   FunctionCallee Fn = M->getOrInsertFunction("__safestack_pointer_address",
1924                                              StackPtrTy->getPointerTo(0));
1925   return IRB.CreateCall(Fn);
1926 }
1927 
1928 //===----------------------------------------------------------------------===//
1929 //  Loop Strength Reduction hooks
1930 //===----------------------------------------------------------------------===//
1931 
1932 /// isLegalAddressingMode - Return true if the addressing mode represented
1933 /// by AM is legal for this target, for a load/store of the specified type.
1934 bool TargetLoweringBase::isLegalAddressingMode(const DataLayout &DL,
1935                                                const AddrMode &AM, Type *Ty,
1936                                                unsigned AS, Instruction *I) const {
1937   // The default implementation of this implements a conservative RISCy, r+r and
1938   // r+i addr mode.
1939 
1940   // Allows a sign-extended 16-bit immediate field.
1941   if (AM.BaseOffs <= -(1LL << 16) || AM.BaseOffs >= (1LL << 16)-1)
1942     return false;
1943 
1944   // No global is ever allowed as a base.
1945   if (AM.BaseGV)
1946     return false;
1947 
1948   // Only support r+r,
1949   switch (AM.Scale) {
1950   case 0:  // "r+i" or just "i", depending on HasBaseReg.
1951     break;
1952   case 1:
1953     if (AM.HasBaseReg && AM.BaseOffs)  // "r+r+i" is not allowed.
1954       return false;
1955     // Otherwise we have r+r or r+i.
1956     break;
1957   case 2:
1958     if (AM.HasBaseReg || AM.BaseOffs)  // 2*r+r  or  2*r+i is not allowed.
1959       return false;
1960     // Allow 2*r as r+r.
1961     break;
1962   default: // Don't allow n * r
1963     return false;
1964   }
1965 
1966   return true;
1967 }
1968 
1969 //===----------------------------------------------------------------------===//
1970 //  Stack Protector
1971 //===----------------------------------------------------------------------===//
1972 
1973 // For OpenBSD return its special guard variable. Otherwise return nullptr,
1974 // so that SelectionDAG handle SSP.
1975 Value *TargetLoweringBase::getIRStackGuard(IRBuilderBase &IRB) const {
1976   if (getTargetMachine().getTargetTriple().isOSOpenBSD()) {
1977     Module &M = *IRB.GetInsertBlock()->getParent()->getParent();
1978     PointerType *PtrTy = Type::getInt8PtrTy(M.getContext());
1979     Constant *C = M.getOrInsertGlobal("__guard_local", PtrTy);
1980     if (GlobalVariable *G = dyn_cast_or_null<GlobalVariable>(C))
1981       G->setVisibility(GlobalValue::HiddenVisibility);
1982     return C;
1983   }
1984   return nullptr;
1985 }
1986 
1987 // Currently only support "standard" __stack_chk_guard.
1988 // TODO: add LOAD_STACK_GUARD support.
1989 void TargetLoweringBase::insertSSPDeclarations(Module &M) const {
1990   if (!M.getNamedValue("__stack_chk_guard")) {
1991     auto *GV = new GlobalVariable(M, Type::getInt8PtrTy(M.getContext()), false,
1992                                   GlobalVariable::ExternalLinkage, nullptr,
1993                                   "__stack_chk_guard");
1994 
1995     // FreeBSD has "__stack_chk_guard" defined externally on libc.so
1996     if (TM.getRelocationModel() == Reloc::Static &&
1997         !TM.getTargetTriple().isWindowsGNUEnvironment() &&
1998         !(TM.getTargetTriple().isPPC64() && TM.getTargetTriple().isOSFreeBSD()))
1999       GV->setDSOLocal(true);
2000   }
2001 }
2002 
2003 // Currently only support "standard" __stack_chk_guard.
2004 // TODO: add LOAD_STACK_GUARD support.
2005 Value *TargetLoweringBase::getSDagStackGuard(const Module &M) const {
2006   return M.getNamedValue("__stack_chk_guard");
2007 }
2008 
2009 Function *TargetLoweringBase::getSSPStackGuardCheck(const Module &M) const {
2010   return nullptr;
2011 }
2012 
2013 unsigned TargetLoweringBase::getMinimumJumpTableEntries() const {
2014   return MinimumJumpTableEntries;
2015 }
2016 
2017 void TargetLoweringBase::setMinimumJumpTableEntries(unsigned Val) {
2018   MinimumJumpTableEntries = Val;
2019 }
2020 
2021 unsigned TargetLoweringBase::getMinimumJumpTableDensity(bool OptForSize) const {
2022   return OptForSize ? OptsizeJumpTableDensity : JumpTableDensity;
2023 }
2024 
2025 unsigned TargetLoweringBase::getMaximumJumpTableSize() const {
2026   return MaximumJumpTableSize;
2027 }
2028 
2029 void TargetLoweringBase::setMaximumJumpTableSize(unsigned Val) {
2030   MaximumJumpTableSize = Val;
2031 }
2032 
2033 bool TargetLoweringBase::isJumpTableRelative() const {
2034   return getTargetMachine().isPositionIndependent();
2035 }
2036 
2037 Align TargetLoweringBase::getPrefLoopAlignment(MachineLoop *ML) const {
2038   if (TM.Options.LoopAlignment)
2039     return Align(TM.Options.LoopAlignment);
2040   return PrefLoopAlignment;
2041 }
2042 
2043 unsigned TargetLoweringBase::getMaxPermittedBytesForAlignment(
2044     MachineBasicBlock *MBB) const {
2045   return MaxBytesForAlignment;
2046 }
2047 
2048 //===----------------------------------------------------------------------===//
2049 //  Reciprocal Estimates
2050 //===----------------------------------------------------------------------===//
2051 
2052 /// Get the reciprocal estimate attribute string for a function that will
2053 /// override the target defaults.
2054 static StringRef getRecipEstimateForFunc(MachineFunction &MF) {
2055   const Function &F = MF.getFunction();
2056   return F.getFnAttribute("reciprocal-estimates").getValueAsString();
2057 }
2058 
2059 /// Construct a string for the given reciprocal operation of the given type.
2060 /// This string should match the corresponding option to the front-end's
2061 /// "-mrecip" flag assuming those strings have been passed through in an
2062 /// attribute string. For example, "vec-divf" for a division of a vXf32.
2063 static std::string getReciprocalOpName(bool IsSqrt, EVT VT) {
2064   std::string Name = VT.isVector() ? "vec-" : "";
2065 
2066   Name += IsSqrt ? "sqrt" : "div";
2067 
2068   // TODO: Handle other float types?
2069   if (VT.getScalarType() == MVT::f64) {
2070     Name += "d";
2071   } else if (VT.getScalarType() == MVT::f16) {
2072     Name += "h";
2073   } else {
2074     assert(VT.getScalarType() == MVT::f32 &&
2075            "Unexpected FP type for reciprocal estimate");
2076     Name += "f";
2077   }
2078 
2079   return Name;
2080 }
2081 
2082 /// Return the character position and value (a single numeric character) of a
2083 /// customized refinement operation in the input string if it exists. Return
2084 /// false if there is no customized refinement step count.
2085 static bool parseRefinementStep(StringRef In, size_t &Position,
2086                                 uint8_t &Value) {
2087   const char RefStepToken = ':';
2088   Position = In.find(RefStepToken);
2089   if (Position == StringRef::npos)
2090     return false;
2091 
2092   StringRef RefStepString = In.substr(Position + 1);
2093   // Allow exactly one numeric character for the additional refinement
2094   // step parameter.
2095   if (RefStepString.size() == 1) {
2096     char RefStepChar = RefStepString[0];
2097     if (isDigit(RefStepChar)) {
2098       Value = RefStepChar - '0';
2099       return true;
2100     }
2101   }
2102   report_fatal_error("Invalid refinement step for -recip.");
2103 }
2104 
2105 /// For the input attribute string, return one of the ReciprocalEstimate enum
2106 /// status values (enabled, disabled, or not specified) for this operation on
2107 /// the specified data type.
2108 static int getOpEnabled(bool IsSqrt, EVT VT, StringRef Override) {
2109   if (Override.empty())
2110     return TargetLoweringBase::ReciprocalEstimate::Unspecified;
2111 
2112   SmallVector<StringRef, 4> OverrideVector;
2113   Override.split(OverrideVector, ',');
2114   unsigned NumArgs = OverrideVector.size();
2115 
2116   // Check if "all", "none", or "default" was specified.
2117   if (NumArgs == 1) {
2118     // Look for an optional setting of the number of refinement steps needed
2119     // for this type of reciprocal operation.
2120     size_t RefPos;
2121     uint8_t RefSteps;
2122     if (parseRefinementStep(Override, RefPos, RefSteps)) {
2123       // Split the string for further processing.
2124       Override = Override.substr(0, RefPos);
2125     }
2126 
2127     // All reciprocal types are enabled.
2128     if (Override == "all")
2129       return TargetLoweringBase::ReciprocalEstimate::Enabled;
2130 
2131     // All reciprocal types are disabled.
2132     if (Override == "none")
2133       return TargetLoweringBase::ReciprocalEstimate::Disabled;
2134 
2135     // Target defaults for enablement are used.
2136     if (Override == "default")
2137       return TargetLoweringBase::ReciprocalEstimate::Unspecified;
2138   }
2139 
2140   // The attribute string may omit the size suffix ('f'/'d').
2141   std::string VTName = getReciprocalOpName(IsSqrt, VT);
2142   std::string VTNameNoSize = VTName;
2143   VTNameNoSize.pop_back();
2144   static const char DisabledPrefix = '!';
2145 
2146   for (StringRef RecipType : OverrideVector) {
2147     size_t RefPos;
2148     uint8_t RefSteps;
2149     if (parseRefinementStep(RecipType, RefPos, RefSteps))
2150       RecipType = RecipType.substr(0, RefPos);
2151 
2152     // Ignore the disablement token for string matching.
2153     bool IsDisabled = RecipType[0] == DisabledPrefix;
2154     if (IsDisabled)
2155       RecipType = RecipType.substr(1);
2156 
2157     if (RecipType.equals(VTName) || RecipType.equals(VTNameNoSize))
2158       return IsDisabled ? TargetLoweringBase::ReciprocalEstimate::Disabled
2159                         : TargetLoweringBase::ReciprocalEstimate::Enabled;
2160   }
2161 
2162   return TargetLoweringBase::ReciprocalEstimate::Unspecified;
2163 }
2164 
2165 /// For the input attribute string, return the customized refinement step count
2166 /// for this operation on the specified data type. If the step count does not
2167 /// exist, return the ReciprocalEstimate enum value for unspecified.
2168 static int getOpRefinementSteps(bool IsSqrt, EVT VT, StringRef Override) {
2169   if (Override.empty())
2170     return TargetLoweringBase::ReciprocalEstimate::Unspecified;
2171 
2172   SmallVector<StringRef, 4> OverrideVector;
2173   Override.split(OverrideVector, ',');
2174   unsigned NumArgs = OverrideVector.size();
2175 
2176   // Check if "all", "default", or "none" was specified.
2177   if (NumArgs == 1) {
2178     // Look for an optional setting of the number of refinement steps needed
2179     // for this type of reciprocal operation.
2180     size_t RefPos;
2181     uint8_t RefSteps;
2182     if (!parseRefinementStep(Override, RefPos, RefSteps))
2183       return TargetLoweringBase::ReciprocalEstimate::Unspecified;
2184 
2185     // Split the string for further processing.
2186     Override = Override.substr(0, RefPos);
2187     assert(Override != "none" &&
2188            "Disabled reciprocals, but specifed refinement steps?");
2189 
2190     // If this is a general override, return the specified number of steps.
2191     if (Override == "all" || Override == "default")
2192       return RefSteps;
2193   }
2194 
2195   // The attribute string may omit the size suffix ('f'/'d').
2196   std::string VTName = getReciprocalOpName(IsSqrt, VT);
2197   std::string VTNameNoSize = VTName;
2198   VTNameNoSize.pop_back();
2199 
2200   for (StringRef RecipType : OverrideVector) {
2201     size_t RefPos;
2202     uint8_t RefSteps;
2203     if (!parseRefinementStep(RecipType, RefPos, RefSteps))
2204       continue;
2205 
2206     RecipType = RecipType.substr(0, RefPos);
2207     if (RecipType.equals(VTName) || RecipType.equals(VTNameNoSize))
2208       return RefSteps;
2209   }
2210 
2211   return TargetLoweringBase::ReciprocalEstimate::Unspecified;
2212 }
2213 
2214 int TargetLoweringBase::getRecipEstimateSqrtEnabled(EVT VT,
2215                                                     MachineFunction &MF) const {
2216   return getOpEnabled(true, VT, getRecipEstimateForFunc(MF));
2217 }
2218 
2219 int TargetLoweringBase::getRecipEstimateDivEnabled(EVT VT,
2220                                                    MachineFunction &MF) const {
2221   return getOpEnabled(false, VT, getRecipEstimateForFunc(MF));
2222 }
2223 
2224 int TargetLoweringBase::getSqrtRefinementSteps(EVT VT,
2225                                                MachineFunction &MF) const {
2226   return getOpRefinementSteps(true, VT, getRecipEstimateForFunc(MF));
2227 }
2228 
2229 int TargetLoweringBase::getDivRefinementSteps(EVT VT,
2230                                               MachineFunction &MF) const {
2231   return getOpRefinementSteps(false, VT, getRecipEstimateForFunc(MF));
2232 }
2233 
2234 void TargetLoweringBase::finalizeLowering(MachineFunction &MF) const {
2235   MF.getRegInfo().freezeReservedRegs(MF);
2236 }
2237 
2238 MachineMemOperand::Flags
2239 TargetLoweringBase::getLoadMemOperandFlags(const LoadInst &LI,
2240                                            const DataLayout &DL) const {
2241   MachineMemOperand::Flags Flags = MachineMemOperand::MOLoad;
2242   if (LI.isVolatile())
2243     Flags |= MachineMemOperand::MOVolatile;
2244 
2245   if (LI.hasMetadata(LLVMContext::MD_nontemporal))
2246     Flags |= MachineMemOperand::MONonTemporal;
2247 
2248   if (LI.hasMetadata(LLVMContext::MD_invariant_load))
2249     Flags |= MachineMemOperand::MOInvariant;
2250 
2251   if (isDereferenceablePointer(LI.getPointerOperand(), LI.getType(), DL))
2252     Flags |= MachineMemOperand::MODereferenceable;
2253 
2254   Flags |= getTargetMMOFlags(LI);
2255   return Flags;
2256 }
2257 
2258 MachineMemOperand::Flags
2259 TargetLoweringBase::getStoreMemOperandFlags(const StoreInst &SI,
2260                                             const DataLayout &DL) const {
2261   MachineMemOperand::Flags Flags = MachineMemOperand::MOStore;
2262 
2263   if (SI.isVolatile())
2264     Flags |= MachineMemOperand::MOVolatile;
2265 
2266   if (SI.hasMetadata(LLVMContext::MD_nontemporal))
2267     Flags |= MachineMemOperand::MONonTemporal;
2268 
2269   // FIXME: Not preserving dereferenceable
2270   Flags |= getTargetMMOFlags(SI);
2271   return Flags;
2272 }
2273 
2274 MachineMemOperand::Flags
2275 TargetLoweringBase::getAtomicMemOperandFlags(const Instruction &AI,
2276                                              const DataLayout &DL) const {
2277   auto Flags = MachineMemOperand::MOLoad | MachineMemOperand::MOStore;
2278 
2279   if (const AtomicRMWInst *RMW = dyn_cast<AtomicRMWInst>(&AI)) {
2280     if (RMW->isVolatile())
2281       Flags |= MachineMemOperand::MOVolatile;
2282   } else if (const AtomicCmpXchgInst *CmpX = dyn_cast<AtomicCmpXchgInst>(&AI)) {
2283     if (CmpX->isVolatile())
2284       Flags |= MachineMemOperand::MOVolatile;
2285   } else
2286     llvm_unreachable("not an atomic instruction");
2287 
2288   // FIXME: Not preserving dereferenceable
2289   Flags |= getTargetMMOFlags(AI);
2290   return Flags;
2291 }
2292 
2293 Instruction *TargetLoweringBase::emitLeadingFence(IRBuilderBase &Builder,
2294                                                   Instruction *Inst,
2295                                                   AtomicOrdering Ord) const {
2296   if (isReleaseOrStronger(Ord) && Inst->hasAtomicStore())
2297     return Builder.CreateFence(Ord);
2298   else
2299     return nullptr;
2300 }
2301 
2302 Instruction *TargetLoweringBase::emitTrailingFence(IRBuilderBase &Builder,
2303                                                    Instruction *Inst,
2304                                                    AtomicOrdering Ord) const {
2305   if (isAcquireOrStronger(Ord))
2306     return Builder.CreateFence(Ord);
2307   else
2308     return nullptr;
2309 }
2310 
2311 //===----------------------------------------------------------------------===//
2312 //  GlobalISel Hooks
2313 //===----------------------------------------------------------------------===//
2314 
2315 bool TargetLoweringBase::shouldLocalize(const MachineInstr &MI,
2316                                         const TargetTransformInfo *TTI) const {
2317   auto &MF = *MI.getMF();
2318   auto &MRI = MF.getRegInfo();
2319   // Assuming a spill and reload of a value has a cost of 1 instruction each,
2320   // this helper function computes the maximum number of uses we should consider
2321   // for remat. E.g. on arm64 global addresses take 2 insts to materialize. We
2322   // break even in terms of code size when the original MI has 2 users vs
2323   // choosing to potentially spill. Any more than 2 users we we have a net code
2324   // size increase. This doesn't take into account register pressure though.
2325   auto maxUses = [](unsigned RematCost) {
2326     // A cost of 1 means remats are basically free.
2327     if (RematCost == 1)
2328       return UINT_MAX;
2329     if (RematCost == 2)
2330       return 2U;
2331 
2332     // Remat is too expensive, only sink if there's one user.
2333     if (RematCost > 2)
2334       return 1U;
2335     llvm_unreachable("Unexpected remat cost");
2336   };
2337 
2338   // Helper to walk through uses and terminate if we've reached a limit. Saves
2339   // us spending time traversing uses if all we want to know is if it's >= min.
2340   auto isUsesAtMost = [&](unsigned Reg, unsigned MaxUses) {
2341     unsigned NumUses = 0;
2342     auto UI = MRI.use_instr_nodbg_begin(Reg), UE = MRI.use_instr_nodbg_end();
2343     for (; UI != UE && NumUses < MaxUses; ++UI) {
2344       NumUses++;
2345     }
2346     // If we haven't reached the end yet then there are more than MaxUses users.
2347     return UI == UE;
2348   };
2349 
2350   switch (MI.getOpcode()) {
2351   default:
2352     return false;
2353   // Constants-like instructions should be close to their users.
2354   // We don't want long live-ranges for them.
2355   case TargetOpcode::G_CONSTANT:
2356   case TargetOpcode::G_FCONSTANT:
2357   case TargetOpcode::G_FRAME_INDEX:
2358   case TargetOpcode::G_INTTOPTR:
2359     return true;
2360   case TargetOpcode::G_GLOBAL_VALUE: {
2361     unsigned RematCost = TTI->getGISelRematGlobalCost();
2362     Register Reg = MI.getOperand(0).getReg();
2363     unsigned MaxUses = maxUses(RematCost);
2364     if (MaxUses == UINT_MAX)
2365       return true; // Remats are "free" so always localize.
2366     bool B = isUsesAtMost(Reg, MaxUses);
2367     return B;
2368   }
2369   }
2370 }
2371