1 //===-- NVPTXISelLowering.cpp - NVPTX DAG Lowering Implementation ---------===//
2 //
3 // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
4 // See https://llvm.org/LICENSE.txt for license information.
5 // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
6 //
7 //===----------------------------------------------------------------------===//
8 //
9 // This file defines the interfaces that NVPTX uses to lower LLVM code into a
10 // selection DAG.
11 //
12 //===----------------------------------------------------------------------===//
13
14 #include "NVPTXISelLowering.h"
15 #include "MCTargetDesc/NVPTXBaseInfo.h"
16 #include "NVPTX.h"
17 #include "NVPTXSubtarget.h"
18 #include "NVPTXTargetMachine.h"
19 #include "NVPTXTargetObjectFile.h"
20 #include "NVPTXUtilities.h"
21 #include "llvm/ADT/APInt.h"
22 #include "llvm/ADT/SmallVector.h"
23 #include "llvm/ADT/StringRef.h"
24 #include "llvm/CodeGen/Analysis.h"
25 #include "llvm/CodeGen/MachineFunction.h"
26 #include "llvm/CodeGen/MachineMemOperand.h"
27 #include "llvm/CodeGen/SelectionDAG.h"
28 #include "llvm/CodeGen/SelectionDAGNodes.h"
29 #include "llvm/CodeGen/TargetCallingConv.h"
30 #include "llvm/CodeGen/TargetLowering.h"
31 #include "llvm/CodeGen/ValueTypes.h"
32 #include "llvm/IR/Argument.h"
33 #include "llvm/IR/Attributes.h"
34 #include "llvm/IR/CallSite.h"
35 #include "llvm/IR/Constants.h"
36 #include "llvm/IR/DataLayout.h"
37 #include "llvm/IR/DerivedTypes.h"
38 #include "llvm/IR/Function.h"
39 #include "llvm/IR/GlobalValue.h"
40 #include "llvm/IR/Instruction.h"
41 #include "llvm/IR/Instructions.h"
42 #include "llvm/IR/IntrinsicsNVPTX.h"
43 #include "llvm/IR/Module.h"
44 #include "llvm/IR/Type.h"
45 #include "llvm/IR/Value.h"
46 #include "llvm/Support/Casting.h"
47 #include "llvm/Support/CodeGen.h"
48 #include "llvm/Support/CommandLine.h"
49 #include "llvm/Support/ErrorHandling.h"
50 #include "llvm/Support/MachineValueType.h"
51 #include "llvm/Support/MathExtras.h"
52 #include "llvm/Support/raw_ostream.h"
53 #include "llvm/Target/TargetMachine.h"
54 #include "llvm/Target/TargetOptions.h"
55 #include <algorithm>
56 #include <cassert>
57 #include <cstdint>
58 #include <iterator>
59 #include <sstream>
60 #include <string>
61 #include <utility>
62 #include <vector>
63
64 #define DEBUG_TYPE "nvptx-lower"
65
66 using namespace llvm;
67
68 static unsigned int uniqueCallSite = 0;
69
70 static cl::opt<bool> sched4reg(
71 "nvptx-sched4reg",
72 cl::desc("NVPTX Specific: schedule for register pressue"), cl::init(false));
73
74 static cl::opt<unsigned>
75 FMAContractLevelOpt("nvptx-fma-level", cl::ZeroOrMore, cl::Hidden,
76 cl::desc("NVPTX Specific: FMA contraction (0: don't do it"
77 " 1: do it 2: do it aggressively"),
78 cl::init(2));
79
80 static cl::opt<int> UsePrecDivF32(
81 "nvptx-prec-divf32", cl::ZeroOrMore, cl::Hidden,
82 cl::desc("NVPTX Specifies: 0 use div.approx, 1 use div.full, 2 use"
83 " IEEE Compliant F32 div.rnd if available."),
84 cl::init(2));
85
86 static cl::opt<bool> UsePrecSqrtF32(
87 "nvptx-prec-sqrtf32", cl::Hidden,
88 cl::desc("NVPTX Specific: 0 use sqrt.approx, 1 use sqrt.rn."),
89 cl::init(true));
90
91 static cl::opt<bool> FtzEnabled(
92 "nvptx-f32ftz", cl::ZeroOrMore, cl::Hidden,
93 cl::desc("NVPTX Specific: Flush f32 subnormals to sign-preserving zero."),
94 cl::init(false));
95
getDivF32Level() const96 int NVPTXTargetLowering::getDivF32Level() const {
97 if (UsePrecDivF32.getNumOccurrences() > 0) {
98 // If nvptx-prec-div32=N is used on the command-line, always honor it
99 return UsePrecDivF32;
100 } else {
101 // Otherwise, use div.approx if fast math is enabled
102 if (getTargetMachine().Options.UnsafeFPMath)
103 return 0;
104 else
105 return 2;
106 }
107 }
108
usePrecSqrtF32() const109 bool NVPTXTargetLowering::usePrecSqrtF32() const {
110 if (UsePrecSqrtF32.getNumOccurrences() > 0) {
111 // If nvptx-prec-sqrtf32 is used on the command-line, always honor it
112 return UsePrecSqrtF32;
113 } else {
114 // Otherwise, use sqrt.approx if fast math is enabled
115 return !getTargetMachine().Options.UnsafeFPMath;
116 }
117 }
118
useF32FTZ(const MachineFunction & MF) const119 bool NVPTXTargetLowering::useF32FTZ(const MachineFunction &MF) const {
120 // TODO: Get rid of this flag; there can be only one way to do this.
121 if (FtzEnabled.getNumOccurrences() > 0) {
122 // If nvptx-f32ftz is used on the command-line, always honor it
123 return FtzEnabled;
124 } else {
125 const Function &F = MF.getFunction();
126 // Otherwise, check for an nvptx-f32ftz attribute on the function
127 if (F.hasFnAttribute("nvptx-f32ftz"))
128 return F.getFnAttribute("nvptx-f32ftz").getValueAsString() == "true";
129 else
130 return false;
131 }
132 }
133
IsPTXVectorType(MVT VT)134 static bool IsPTXVectorType(MVT VT) {
135 switch (VT.SimpleTy) {
136 default:
137 return false;
138 case MVT::v2i1:
139 case MVT::v4i1:
140 case MVT::v2i8:
141 case MVT::v4i8:
142 case MVT::v2i16:
143 case MVT::v4i16:
144 case MVT::v2i32:
145 case MVT::v4i32:
146 case MVT::v2i64:
147 case MVT::v2f16:
148 case MVT::v4f16:
149 case MVT::v8f16: // <4 x f16x2>
150 case MVT::v2f32:
151 case MVT::v4f32:
152 case MVT::v2f64:
153 return true;
154 }
155 }
156
157 /// ComputePTXValueVTs - For the given Type \p Ty, returns the set of primitive
158 /// EVTs that compose it. Unlike ComputeValueVTs, this will break apart vectors
159 /// into their primitive components.
160 /// NOTE: This is a band-aid for code that expects ComputeValueVTs to return the
161 /// same number of types as the Ins/Outs arrays in LowerFormalArguments,
162 /// LowerCall, and LowerReturn.
ComputePTXValueVTs(const TargetLowering & TLI,const DataLayout & DL,Type * Ty,SmallVectorImpl<EVT> & ValueVTs,SmallVectorImpl<uint64_t> * Offsets=nullptr,uint64_t StartingOffset=0)163 static void ComputePTXValueVTs(const TargetLowering &TLI, const DataLayout &DL,
164 Type *Ty, SmallVectorImpl<EVT> &ValueVTs,
165 SmallVectorImpl<uint64_t> *Offsets = nullptr,
166 uint64_t StartingOffset = 0) {
167 SmallVector<EVT, 16> TempVTs;
168 SmallVector<uint64_t, 16> TempOffsets;
169
170 // Special case for i128 - decompose to (i64, i64)
171 if (Ty->isIntegerTy(128)) {
172 ValueVTs.push_back(EVT(MVT::i64));
173 ValueVTs.push_back(EVT(MVT::i64));
174
175 if (Offsets) {
176 Offsets->push_back(StartingOffset + 0);
177 Offsets->push_back(StartingOffset + 8);
178 }
179
180 return;
181 }
182
183 // Given a struct type, recursively traverse the elements with custom ComputePTXValueVTs.
184 if (StructType *STy = dyn_cast<StructType>(Ty)) {
185 auto const *SL = DL.getStructLayout(STy);
186 auto ElementNum = 0;
187 for(auto *EI : STy->elements()) {
188 ComputePTXValueVTs(TLI, DL, EI, ValueVTs, Offsets,
189 StartingOffset + SL->getElementOffset(ElementNum));
190 ++ElementNum;
191 }
192 return;
193 }
194
195 ComputeValueVTs(TLI, DL, Ty, TempVTs, &TempOffsets, StartingOffset);
196 for (unsigned i = 0, e = TempVTs.size(); i != e; ++i) {
197 EVT VT = TempVTs[i];
198 uint64_t Off = TempOffsets[i];
199 // Split vectors into individual elements, except for v2f16, which
200 // we will pass as a single scalar.
201 if (VT.isVector()) {
202 unsigned NumElts = VT.getVectorNumElements();
203 EVT EltVT = VT.getVectorElementType();
204 // Vectors with an even number of f16 elements will be passed to
205 // us as an array of v2f16 elements. We must match this so we
206 // stay in sync with Ins/Outs.
207 if (EltVT == MVT::f16 && NumElts % 2 == 0) {
208 EltVT = MVT::v2f16;
209 NumElts /= 2;
210 }
211 for (unsigned j = 0; j != NumElts; ++j) {
212 ValueVTs.push_back(EltVT);
213 if (Offsets)
214 Offsets->push_back(Off + j * EltVT.getStoreSize());
215 }
216 } else {
217 ValueVTs.push_back(VT);
218 if (Offsets)
219 Offsets->push_back(Off);
220 }
221 }
222 }
223
224 // Check whether we can merge loads/stores of some of the pieces of a
225 // flattened function parameter or return value into a single vector
226 // load/store.
227 //
228 // The flattened parameter is represented as a list of EVTs and
229 // offsets, and the whole structure is aligned to ParamAlignment. This
230 // function determines whether we can load/store pieces of the
231 // parameter starting at index Idx using a single vectorized op of
232 // size AccessSize. If so, it returns the number of param pieces
233 // covered by the vector op. Otherwise, it returns 1.
CanMergeParamLoadStoresStartingAt(unsigned Idx,uint32_t AccessSize,const SmallVectorImpl<EVT> & ValueVTs,const SmallVectorImpl<uint64_t> & Offsets,unsigned ParamAlignment)234 static unsigned CanMergeParamLoadStoresStartingAt(
235 unsigned Idx, uint32_t AccessSize, const SmallVectorImpl<EVT> &ValueVTs,
236 const SmallVectorImpl<uint64_t> &Offsets, unsigned ParamAlignment) {
237 assert(isPowerOf2_32(AccessSize) && "must be a power of 2!");
238
239 // Can't vectorize if param alignment is not sufficient.
240 if (AccessSize > ParamAlignment)
241 return 1;
242 // Can't vectorize if offset is not aligned.
243 if (Offsets[Idx] & (AccessSize - 1))
244 return 1;
245
246 EVT EltVT = ValueVTs[Idx];
247 unsigned EltSize = EltVT.getStoreSize();
248
249 // Element is too large to vectorize.
250 if (EltSize >= AccessSize)
251 return 1;
252
253 unsigned NumElts = AccessSize / EltSize;
254 // Can't vectorize if AccessBytes if not a multiple of EltSize.
255 if (AccessSize != EltSize * NumElts)
256 return 1;
257
258 // We don't have enough elements to vectorize.
259 if (Idx + NumElts > ValueVTs.size())
260 return 1;
261
262 // PTX ISA can only deal with 2- and 4-element vector ops.
263 if (NumElts != 4 && NumElts != 2)
264 return 1;
265
266 for (unsigned j = Idx + 1; j < Idx + NumElts; ++j) {
267 // Types do not match.
268 if (ValueVTs[j] != EltVT)
269 return 1;
270
271 // Elements are not contiguous.
272 if (Offsets[j] - Offsets[j - 1] != EltSize)
273 return 1;
274 }
275 // OK. We can vectorize ValueVTs[i..i+NumElts)
276 return NumElts;
277 }
278
279 // Flags for tracking per-element vectorization state of loads/stores
280 // of a flattened function parameter or return value.
281 enum ParamVectorizationFlags {
282 PVF_INNER = 0x0, // Middle elements of a vector.
283 PVF_FIRST = 0x1, // First element of the vector.
284 PVF_LAST = 0x2, // Last element of the vector.
285 // Scalar is effectively a 1-element vector.
286 PVF_SCALAR = PVF_FIRST | PVF_LAST
287 };
288
289 // Computes whether and how we can vectorize the loads/stores of a
290 // flattened function parameter or return value.
291 //
292 // The flattened parameter is represented as the list of ValueVTs and
293 // Offsets, and is aligned to ParamAlignment bytes. We return a vector
294 // of the same size as ValueVTs indicating how each piece should be
295 // loaded/stored (i.e. as a scalar, or as part of a vector
296 // load/store).
297 static SmallVector<ParamVectorizationFlags, 16>
VectorizePTXValueVTs(const SmallVectorImpl<EVT> & ValueVTs,const SmallVectorImpl<uint64_t> & Offsets,unsigned ParamAlignment)298 VectorizePTXValueVTs(const SmallVectorImpl<EVT> &ValueVTs,
299 const SmallVectorImpl<uint64_t> &Offsets,
300 unsigned ParamAlignment) {
301 // Set vector size to match ValueVTs and mark all elements as
302 // scalars by default.
303 SmallVector<ParamVectorizationFlags, 16> VectorInfo;
304 VectorInfo.assign(ValueVTs.size(), PVF_SCALAR);
305
306 // Check what we can vectorize using 128/64/32-bit accesses.
307 for (int I = 0, E = ValueVTs.size(); I != E; ++I) {
308 // Skip elements we've already processed.
309 assert(VectorInfo[I] == PVF_SCALAR && "Unexpected vector info state.");
310 for (unsigned AccessSize : {16, 8, 4, 2}) {
311 unsigned NumElts = CanMergeParamLoadStoresStartingAt(
312 I, AccessSize, ValueVTs, Offsets, ParamAlignment);
313 // Mark vectorized elements.
314 switch (NumElts) {
315 default:
316 llvm_unreachable("Unexpected return value");
317 case 1:
318 // Can't vectorize using this size, try next smaller size.
319 continue;
320 case 2:
321 assert(I + 1 < E && "Not enough elements.");
322 VectorInfo[I] = PVF_FIRST;
323 VectorInfo[I + 1] = PVF_LAST;
324 I += 1;
325 break;
326 case 4:
327 assert(I + 3 < E && "Not enough elements.");
328 VectorInfo[I] = PVF_FIRST;
329 VectorInfo[I + 1] = PVF_INNER;
330 VectorInfo[I + 2] = PVF_INNER;
331 VectorInfo[I + 3] = PVF_LAST;
332 I += 3;
333 break;
334 }
335 // Break out of the inner loop because we've already succeeded
336 // using largest possible AccessSize.
337 break;
338 }
339 }
340 return VectorInfo;
341 }
342
343 // NVPTXTargetLowering Constructor.
NVPTXTargetLowering(const NVPTXTargetMachine & TM,const NVPTXSubtarget & STI)344 NVPTXTargetLowering::NVPTXTargetLowering(const NVPTXTargetMachine &TM,
345 const NVPTXSubtarget &STI)
346 : TargetLowering(TM), nvTM(&TM), STI(STI) {
347 // always lower memset, memcpy, and memmove intrinsics to load/store
348 // instructions, rather
349 // then generating calls to memset, mempcy or memmove.
350 MaxStoresPerMemset = (unsigned) 0xFFFFFFFF;
351 MaxStoresPerMemcpy = (unsigned) 0xFFFFFFFF;
352 MaxStoresPerMemmove = (unsigned) 0xFFFFFFFF;
353
354 setBooleanContents(ZeroOrNegativeOneBooleanContent);
355 setBooleanVectorContents(ZeroOrNegativeOneBooleanContent);
356
357 // Jump is Expensive. Don't create extra control flow for 'and', 'or'
358 // condition branches.
359 setJumpIsExpensive(true);
360
361 // Wide divides are _very_ slow. Try to reduce the width of the divide if
362 // possible.
363 addBypassSlowDiv(64, 32);
364
365 // By default, use the Source scheduling
366 if (sched4reg)
367 setSchedulingPreference(Sched::RegPressure);
368 else
369 setSchedulingPreference(Sched::Source);
370
371 auto setFP16OperationAction = [&](unsigned Op, MVT VT, LegalizeAction Action,
372 LegalizeAction NoF16Action) {
373 setOperationAction(Op, VT, STI.allowFP16Math() ? Action : NoF16Action);
374 };
375
376 addRegisterClass(MVT::i1, &NVPTX::Int1RegsRegClass);
377 addRegisterClass(MVT::i16, &NVPTX::Int16RegsRegClass);
378 addRegisterClass(MVT::i32, &NVPTX::Int32RegsRegClass);
379 addRegisterClass(MVT::i64, &NVPTX::Int64RegsRegClass);
380 addRegisterClass(MVT::f32, &NVPTX::Float32RegsRegClass);
381 addRegisterClass(MVT::f64, &NVPTX::Float64RegsRegClass);
382 addRegisterClass(MVT::f16, &NVPTX::Float16RegsRegClass);
383 addRegisterClass(MVT::v2f16, &NVPTX::Float16x2RegsRegClass);
384
385 // Conversion to/from FP16/FP16x2 is always legal.
386 setOperationAction(ISD::SINT_TO_FP, MVT::f16, Legal);
387 setOperationAction(ISD::FP_TO_SINT, MVT::f16, Legal);
388 setOperationAction(ISD::BUILD_VECTOR, MVT::v2f16, Custom);
389 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v2f16, Custom);
390 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v2f16, Expand);
391 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v2f16, Expand);
392
393 setFP16OperationAction(ISD::SETCC, MVT::f16, Legal, Promote);
394 setFP16OperationAction(ISD::SETCC, MVT::v2f16, Legal, Expand);
395
396 // Operations not directly supported by NVPTX.
397 for (MVT VT : {MVT::f16, MVT::v2f16, MVT::f32, MVT::f64, MVT::i1, MVT::i8,
398 MVT::i16, MVT::i32, MVT::i64}) {
399 setOperationAction(ISD::SELECT_CC, VT, Expand);
400 setOperationAction(ISD::BR_CC, VT, Expand);
401 }
402
403 // Some SIGN_EXTEND_INREG can be done using cvt instruction.
404 // For others we will expand to a SHL/SRA pair.
405 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i64, Legal);
406 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i32, Legal);
407 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i16, Legal);
408 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i8 , Legal);
409 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i1, Expand);
410
411 setOperationAction(ISD::SHL_PARTS, MVT::i32 , Custom);
412 setOperationAction(ISD::SRA_PARTS, MVT::i32 , Custom);
413 setOperationAction(ISD::SRL_PARTS, MVT::i32 , Custom);
414 setOperationAction(ISD::SHL_PARTS, MVT::i64 , Custom);
415 setOperationAction(ISD::SRA_PARTS, MVT::i64 , Custom);
416 setOperationAction(ISD::SRL_PARTS, MVT::i64 , Custom);
417
418 setOperationAction(ISD::BITREVERSE, MVT::i32, Legal);
419 setOperationAction(ISD::BITREVERSE, MVT::i64, Legal);
420
421 // TODO: we may consider expanding ROTL/ROTR on older GPUs. Currently on GPUs
422 // that don't have h/w rotation we lower them to multi-instruction assembly.
423 // See ROT*_sw in NVPTXIntrInfo.td
424 setOperationAction(ISD::ROTL, MVT::i64, Legal);
425 setOperationAction(ISD::ROTR, MVT::i64, Legal);
426 setOperationAction(ISD::ROTL, MVT::i32, Legal);
427 setOperationAction(ISD::ROTR, MVT::i32, Legal);
428
429 setOperationAction(ISD::ROTL, MVT::i16, Expand);
430 setOperationAction(ISD::ROTR, MVT::i16, Expand);
431 setOperationAction(ISD::ROTL, MVT::i8, Expand);
432 setOperationAction(ISD::ROTR, MVT::i8, Expand);
433 setOperationAction(ISD::BSWAP, MVT::i16, Expand);
434 setOperationAction(ISD::BSWAP, MVT::i32, Expand);
435 setOperationAction(ISD::BSWAP, MVT::i64, Expand);
436
437 // Indirect branch is not supported.
438 // This also disables Jump Table creation.
439 setOperationAction(ISD::BR_JT, MVT::Other, Expand);
440 setOperationAction(ISD::BRIND, MVT::Other, Expand);
441
442 setOperationAction(ISD::GlobalAddress, MVT::i32, Custom);
443 setOperationAction(ISD::GlobalAddress, MVT::i64, Custom);
444
445 // We want to legalize constant related memmove and memcopy
446 // intrinsics.
447 setOperationAction(ISD::INTRINSIC_W_CHAIN, MVT::Other, Custom);
448
449 // Turn FP extload into load/fpextend
450 setLoadExtAction(ISD::EXTLOAD, MVT::f32, MVT::f16, Expand);
451 setLoadExtAction(ISD::EXTLOAD, MVT::f64, MVT::f16, Expand);
452 setLoadExtAction(ISD::EXTLOAD, MVT::f64, MVT::f32, Expand);
453 setLoadExtAction(ISD::EXTLOAD, MVT::v2f32, MVT::v2f16, Expand);
454 setLoadExtAction(ISD::EXTLOAD, MVT::v2f64, MVT::v2f16, Expand);
455 setLoadExtAction(ISD::EXTLOAD, MVT::v2f64, MVT::v2f32, Expand);
456 setLoadExtAction(ISD::EXTLOAD, MVT::v4f32, MVT::v4f16, Expand);
457 setLoadExtAction(ISD::EXTLOAD, MVT::v4f64, MVT::v4f16, Expand);
458 setLoadExtAction(ISD::EXTLOAD, MVT::v4f64, MVT::v4f32, Expand);
459 // Turn FP truncstore into trunc + store.
460 // FIXME: vector types should also be expanded
461 setTruncStoreAction(MVT::f32, MVT::f16, Expand);
462 setTruncStoreAction(MVT::f64, MVT::f16, Expand);
463 setTruncStoreAction(MVT::f64, MVT::f32, Expand);
464
465 // PTX does not support load / store predicate registers
466 setOperationAction(ISD::LOAD, MVT::i1, Custom);
467 setOperationAction(ISD::STORE, MVT::i1, Custom);
468
469 for (MVT VT : MVT::integer_valuetypes()) {
470 setLoadExtAction(ISD::SEXTLOAD, VT, MVT::i1, Promote);
471 setLoadExtAction(ISD::ZEXTLOAD, VT, MVT::i1, Promote);
472 setTruncStoreAction(VT, MVT::i1, Expand);
473 }
474
475 // This is legal in NVPTX
476 setOperationAction(ISD::ConstantFP, MVT::f64, Legal);
477 setOperationAction(ISD::ConstantFP, MVT::f32, Legal);
478 setOperationAction(ISD::ConstantFP, MVT::f16, Legal);
479
480 // TRAP can be lowered to PTX trap
481 setOperationAction(ISD::TRAP, MVT::Other, Legal);
482
483 // Register custom handling for vector loads/stores
484 for (MVT VT : MVT::fixedlen_vector_valuetypes()) {
485 if (IsPTXVectorType(VT)) {
486 setOperationAction(ISD::LOAD, VT, Custom);
487 setOperationAction(ISD::STORE, VT, Custom);
488 setOperationAction(ISD::INTRINSIC_W_CHAIN, VT, Custom);
489 }
490 }
491
492 // Custom handling for i8 intrinsics
493 setOperationAction(ISD::INTRINSIC_W_CHAIN, MVT::i8, Custom);
494
495 for (const auto& Ty : {MVT::i16, MVT::i32, MVT::i64}) {
496 setOperationAction(ISD::ABS, Ty, Legal);
497 setOperationAction(ISD::SMIN, Ty, Legal);
498 setOperationAction(ISD::SMAX, Ty, Legal);
499 setOperationAction(ISD::UMIN, Ty, Legal);
500 setOperationAction(ISD::UMAX, Ty, Legal);
501
502 setOperationAction(ISD::CTPOP, Ty, Legal);
503 setOperationAction(ISD::CTLZ, Ty, Legal);
504 }
505
506 setOperationAction(ISD::CTTZ, MVT::i16, Expand);
507 setOperationAction(ISD::CTTZ, MVT::i32, Expand);
508 setOperationAction(ISD::CTTZ, MVT::i64, Expand);
509
510 // PTX does not directly support SELP of i1, so promote to i32 first
511 setOperationAction(ISD::SELECT, MVT::i1, Custom);
512
513 // PTX cannot multiply two i64s in a single instruction.
514 setOperationAction(ISD::SMUL_LOHI, MVT::i64, Expand);
515 setOperationAction(ISD::UMUL_LOHI, MVT::i64, Expand);
516
517 // We have some custom DAG combine patterns for these nodes
518 setTargetDAGCombine(ISD::ADD);
519 setTargetDAGCombine(ISD::AND);
520 setTargetDAGCombine(ISD::FADD);
521 setTargetDAGCombine(ISD::MUL);
522 setTargetDAGCombine(ISD::SHL);
523 setTargetDAGCombine(ISD::SREM);
524 setTargetDAGCombine(ISD::UREM);
525
526 // setcc for f16x2 needs special handling to prevent legalizer's
527 // attempt to scalarize it due to v2i1 not being legal.
528 if (STI.allowFP16Math())
529 setTargetDAGCombine(ISD::SETCC);
530
531 // Promote fp16 arithmetic if fp16 hardware isn't available or the
532 // user passed --nvptx-no-fp16-math. The flag is useful because,
533 // although sm_53+ GPUs have some sort of FP16 support in
534 // hardware, only sm_53 and sm_60 have full implementation. Others
535 // only have token amount of hardware and are likely to run faster
536 // by using fp32 units instead.
537 for (const auto &Op : {ISD::FADD, ISD::FMUL, ISD::FSUB, ISD::FMA}) {
538 setFP16OperationAction(Op, MVT::f16, Legal, Promote);
539 setFP16OperationAction(Op, MVT::v2f16, Legal, Expand);
540 }
541
542 // There's no neg.f16 instruction. Expand to (0-x).
543 setOperationAction(ISD::FNEG, MVT::f16, Expand);
544 setOperationAction(ISD::FNEG, MVT::v2f16, Expand);
545
546 // (would be) Library functions.
547
548 // These map to conversion instructions for scalar FP types.
549 for (const auto &Op : {ISD::FCEIL, ISD::FFLOOR, ISD::FNEARBYINT, ISD::FRINT,
550 ISD::FTRUNC}) {
551 setOperationAction(Op, MVT::f16, Legal);
552 setOperationAction(Op, MVT::f32, Legal);
553 setOperationAction(Op, MVT::f64, Legal);
554 setOperationAction(Op, MVT::v2f16, Expand);
555 }
556
557 setOperationAction(ISD::FROUND, MVT::f16, Promote);
558 setOperationAction(ISD::FROUND, MVT::v2f16, Expand);
559 setOperationAction(ISD::FROUND, MVT::f32, Custom);
560 setOperationAction(ISD::FROUND, MVT::f64, Custom);
561
562
563 // 'Expand' implements FCOPYSIGN without calling an external library.
564 setOperationAction(ISD::FCOPYSIGN, MVT::f16, Expand);
565 setOperationAction(ISD::FCOPYSIGN, MVT::v2f16, Expand);
566 setOperationAction(ISD::FCOPYSIGN, MVT::f32, Expand);
567 setOperationAction(ISD::FCOPYSIGN, MVT::f64, Expand);
568
569 // These map to corresponding instructions for f32/f64. f16 must be
570 // promoted to f32. v2f16 is expanded to f16, which is then promoted
571 // to f32.
572 for (const auto &Op : {ISD::FDIV, ISD::FREM, ISD::FSQRT, ISD::FSIN, ISD::FCOS,
573 ISD::FABS, ISD::FMINNUM, ISD::FMAXNUM}) {
574 setOperationAction(Op, MVT::f16, Promote);
575 setOperationAction(Op, MVT::f32, Legal);
576 setOperationAction(Op, MVT::f64, Legal);
577 setOperationAction(Op, MVT::v2f16, Expand);
578 }
579 setOperationAction(ISD::FMINNUM, MVT::f16, Promote);
580 setOperationAction(ISD::FMAXNUM, MVT::f16, Promote);
581 setOperationAction(ISD::FMINIMUM, MVT::f16, Promote);
582 setOperationAction(ISD::FMAXIMUM, MVT::f16, Promote);
583
584 // No FEXP2, FLOG2. The PTX ex2 and log2 functions are always approximate.
585 // No FPOW or FREM in PTX.
586
587 // Now deduce the information based on the above mentioned
588 // actions
589 computeRegisterProperties(STI.getRegisterInfo());
590 }
591
getTargetNodeName(unsigned Opcode) const592 const char *NVPTXTargetLowering::getTargetNodeName(unsigned Opcode) const {
593 switch ((NVPTXISD::NodeType)Opcode) {
594 case NVPTXISD::FIRST_NUMBER:
595 break;
596 case NVPTXISD::CALL:
597 return "NVPTXISD::CALL";
598 case NVPTXISD::RET_FLAG:
599 return "NVPTXISD::RET_FLAG";
600 case NVPTXISD::LOAD_PARAM:
601 return "NVPTXISD::LOAD_PARAM";
602 case NVPTXISD::Wrapper:
603 return "NVPTXISD::Wrapper";
604 case NVPTXISD::DeclareParam:
605 return "NVPTXISD::DeclareParam";
606 case NVPTXISD::DeclareScalarParam:
607 return "NVPTXISD::DeclareScalarParam";
608 case NVPTXISD::DeclareRet:
609 return "NVPTXISD::DeclareRet";
610 case NVPTXISD::DeclareScalarRet:
611 return "NVPTXISD::DeclareScalarRet";
612 case NVPTXISD::DeclareRetParam:
613 return "NVPTXISD::DeclareRetParam";
614 case NVPTXISD::PrintCall:
615 return "NVPTXISD::PrintCall";
616 case NVPTXISD::PrintConvergentCall:
617 return "NVPTXISD::PrintConvergentCall";
618 case NVPTXISD::PrintCallUni:
619 return "NVPTXISD::PrintCallUni";
620 case NVPTXISD::PrintConvergentCallUni:
621 return "NVPTXISD::PrintConvergentCallUni";
622 case NVPTXISD::LoadParam:
623 return "NVPTXISD::LoadParam";
624 case NVPTXISD::LoadParamV2:
625 return "NVPTXISD::LoadParamV2";
626 case NVPTXISD::LoadParamV4:
627 return "NVPTXISD::LoadParamV4";
628 case NVPTXISD::StoreParam:
629 return "NVPTXISD::StoreParam";
630 case NVPTXISD::StoreParamV2:
631 return "NVPTXISD::StoreParamV2";
632 case NVPTXISD::StoreParamV4:
633 return "NVPTXISD::StoreParamV4";
634 case NVPTXISD::StoreParamS32:
635 return "NVPTXISD::StoreParamS32";
636 case NVPTXISD::StoreParamU32:
637 return "NVPTXISD::StoreParamU32";
638 case NVPTXISD::CallArgBegin:
639 return "NVPTXISD::CallArgBegin";
640 case NVPTXISD::CallArg:
641 return "NVPTXISD::CallArg";
642 case NVPTXISD::LastCallArg:
643 return "NVPTXISD::LastCallArg";
644 case NVPTXISD::CallArgEnd:
645 return "NVPTXISD::CallArgEnd";
646 case NVPTXISD::CallVoid:
647 return "NVPTXISD::CallVoid";
648 case NVPTXISD::CallVal:
649 return "NVPTXISD::CallVal";
650 case NVPTXISD::CallSymbol:
651 return "NVPTXISD::CallSymbol";
652 case NVPTXISD::Prototype:
653 return "NVPTXISD::Prototype";
654 case NVPTXISD::MoveParam:
655 return "NVPTXISD::MoveParam";
656 case NVPTXISD::StoreRetval:
657 return "NVPTXISD::StoreRetval";
658 case NVPTXISD::StoreRetvalV2:
659 return "NVPTXISD::StoreRetvalV2";
660 case NVPTXISD::StoreRetvalV4:
661 return "NVPTXISD::StoreRetvalV4";
662 case NVPTXISD::PseudoUseParam:
663 return "NVPTXISD::PseudoUseParam";
664 case NVPTXISD::RETURN:
665 return "NVPTXISD::RETURN";
666 case NVPTXISD::CallSeqBegin:
667 return "NVPTXISD::CallSeqBegin";
668 case NVPTXISD::CallSeqEnd:
669 return "NVPTXISD::CallSeqEnd";
670 case NVPTXISD::CallPrototype:
671 return "NVPTXISD::CallPrototype";
672 case NVPTXISD::ProxyReg:
673 return "NVPTXISD::ProxyReg";
674 case NVPTXISD::LoadV2:
675 return "NVPTXISD::LoadV2";
676 case NVPTXISD::LoadV4:
677 return "NVPTXISD::LoadV4";
678 case NVPTXISD::LDGV2:
679 return "NVPTXISD::LDGV2";
680 case NVPTXISD::LDGV4:
681 return "NVPTXISD::LDGV4";
682 case NVPTXISD::LDUV2:
683 return "NVPTXISD::LDUV2";
684 case NVPTXISD::LDUV4:
685 return "NVPTXISD::LDUV4";
686 case NVPTXISD::StoreV2:
687 return "NVPTXISD::StoreV2";
688 case NVPTXISD::StoreV4:
689 return "NVPTXISD::StoreV4";
690 case NVPTXISD::FUN_SHFL_CLAMP:
691 return "NVPTXISD::FUN_SHFL_CLAMP";
692 case NVPTXISD::FUN_SHFR_CLAMP:
693 return "NVPTXISD::FUN_SHFR_CLAMP";
694 case NVPTXISD::IMAD:
695 return "NVPTXISD::IMAD";
696 case NVPTXISD::SETP_F16X2:
697 return "NVPTXISD::SETP_F16X2";
698 case NVPTXISD::Dummy:
699 return "NVPTXISD::Dummy";
700 case NVPTXISD::MUL_WIDE_SIGNED:
701 return "NVPTXISD::MUL_WIDE_SIGNED";
702 case NVPTXISD::MUL_WIDE_UNSIGNED:
703 return "NVPTXISD::MUL_WIDE_UNSIGNED";
704 case NVPTXISD::Tex1DFloatS32: return "NVPTXISD::Tex1DFloatS32";
705 case NVPTXISD::Tex1DFloatFloat: return "NVPTXISD::Tex1DFloatFloat";
706 case NVPTXISD::Tex1DFloatFloatLevel:
707 return "NVPTXISD::Tex1DFloatFloatLevel";
708 case NVPTXISD::Tex1DFloatFloatGrad:
709 return "NVPTXISD::Tex1DFloatFloatGrad";
710 case NVPTXISD::Tex1DS32S32: return "NVPTXISD::Tex1DS32S32";
711 case NVPTXISD::Tex1DS32Float: return "NVPTXISD::Tex1DS32Float";
712 case NVPTXISD::Tex1DS32FloatLevel:
713 return "NVPTXISD::Tex1DS32FloatLevel";
714 case NVPTXISD::Tex1DS32FloatGrad:
715 return "NVPTXISD::Tex1DS32FloatGrad";
716 case NVPTXISD::Tex1DU32S32: return "NVPTXISD::Tex1DU32S32";
717 case NVPTXISD::Tex1DU32Float: return "NVPTXISD::Tex1DU32Float";
718 case NVPTXISD::Tex1DU32FloatLevel:
719 return "NVPTXISD::Tex1DU32FloatLevel";
720 case NVPTXISD::Tex1DU32FloatGrad:
721 return "NVPTXISD::Tex1DU32FloatGrad";
722 case NVPTXISD::Tex1DArrayFloatS32: return "NVPTXISD::Tex1DArrayFloatS32";
723 case NVPTXISD::Tex1DArrayFloatFloat: return "NVPTXISD::Tex1DArrayFloatFloat";
724 case NVPTXISD::Tex1DArrayFloatFloatLevel:
725 return "NVPTXISD::Tex1DArrayFloatFloatLevel";
726 case NVPTXISD::Tex1DArrayFloatFloatGrad:
727 return "NVPTXISD::Tex1DArrayFloatFloatGrad";
728 case NVPTXISD::Tex1DArrayS32S32: return "NVPTXISD::Tex1DArrayS32S32";
729 case NVPTXISD::Tex1DArrayS32Float: return "NVPTXISD::Tex1DArrayS32Float";
730 case NVPTXISD::Tex1DArrayS32FloatLevel:
731 return "NVPTXISD::Tex1DArrayS32FloatLevel";
732 case NVPTXISD::Tex1DArrayS32FloatGrad:
733 return "NVPTXISD::Tex1DArrayS32FloatGrad";
734 case NVPTXISD::Tex1DArrayU32S32: return "NVPTXISD::Tex1DArrayU32S32";
735 case NVPTXISD::Tex1DArrayU32Float: return "NVPTXISD::Tex1DArrayU32Float";
736 case NVPTXISD::Tex1DArrayU32FloatLevel:
737 return "NVPTXISD::Tex1DArrayU32FloatLevel";
738 case NVPTXISD::Tex1DArrayU32FloatGrad:
739 return "NVPTXISD::Tex1DArrayU32FloatGrad";
740 case NVPTXISD::Tex2DFloatS32: return "NVPTXISD::Tex2DFloatS32";
741 case NVPTXISD::Tex2DFloatFloat: return "NVPTXISD::Tex2DFloatFloat";
742 case NVPTXISD::Tex2DFloatFloatLevel:
743 return "NVPTXISD::Tex2DFloatFloatLevel";
744 case NVPTXISD::Tex2DFloatFloatGrad:
745 return "NVPTXISD::Tex2DFloatFloatGrad";
746 case NVPTXISD::Tex2DS32S32: return "NVPTXISD::Tex2DS32S32";
747 case NVPTXISD::Tex2DS32Float: return "NVPTXISD::Tex2DS32Float";
748 case NVPTXISD::Tex2DS32FloatLevel:
749 return "NVPTXISD::Tex2DS32FloatLevel";
750 case NVPTXISD::Tex2DS32FloatGrad:
751 return "NVPTXISD::Tex2DS32FloatGrad";
752 case NVPTXISD::Tex2DU32S32: return "NVPTXISD::Tex2DU32S32";
753 case NVPTXISD::Tex2DU32Float: return "NVPTXISD::Tex2DU32Float";
754 case NVPTXISD::Tex2DU32FloatLevel:
755 return "NVPTXISD::Tex2DU32FloatLevel";
756 case NVPTXISD::Tex2DU32FloatGrad:
757 return "NVPTXISD::Tex2DU32FloatGrad";
758 case NVPTXISD::Tex2DArrayFloatS32: return "NVPTXISD::Tex2DArrayFloatS32";
759 case NVPTXISD::Tex2DArrayFloatFloat: return "NVPTXISD::Tex2DArrayFloatFloat";
760 case NVPTXISD::Tex2DArrayFloatFloatLevel:
761 return "NVPTXISD::Tex2DArrayFloatFloatLevel";
762 case NVPTXISD::Tex2DArrayFloatFloatGrad:
763 return "NVPTXISD::Tex2DArrayFloatFloatGrad";
764 case NVPTXISD::Tex2DArrayS32S32: return "NVPTXISD::Tex2DArrayS32S32";
765 case NVPTXISD::Tex2DArrayS32Float: return "NVPTXISD::Tex2DArrayS32Float";
766 case NVPTXISD::Tex2DArrayS32FloatLevel:
767 return "NVPTXISD::Tex2DArrayS32FloatLevel";
768 case NVPTXISD::Tex2DArrayS32FloatGrad:
769 return "NVPTXISD::Tex2DArrayS32FloatGrad";
770 case NVPTXISD::Tex2DArrayU32S32: return "NVPTXISD::Tex2DArrayU32S32";
771 case NVPTXISD::Tex2DArrayU32Float: return "NVPTXISD::Tex2DArrayU32Float";
772 case NVPTXISD::Tex2DArrayU32FloatLevel:
773 return "NVPTXISD::Tex2DArrayU32FloatLevel";
774 case NVPTXISD::Tex2DArrayU32FloatGrad:
775 return "NVPTXISD::Tex2DArrayU32FloatGrad";
776 case NVPTXISD::Tex3DFloatS32: return "NVPTXISD::Tex3DFloatS32";
777 case NVPTXISD::Tex3DFloatFloat: return "NVPTXISD::Tex3DFloatFloat";
778 case NVPTXISD::Tex3DFloatFloatLevel:
779 return "NVPTXISD::Tex3DFloatFloatLevel";
780 case NVPTXISD::Tex3DFloatFloatGrad:
781 return "NVPTXISD::Tex3DFloatFloatGrad";
782 case NVPTXISD::Tex3DS32S32: return "NVPTXISD::Tex3DS32S32";
783 case NVPTXISD::Tex3DS32Float: return "NVPTXISD::Tex3DS32Float";
784 case NVPTXISD::Tex3DS32FloatLevel:
785 return "NVPTXISD::Tex3DS32FloatLevel";
786 case NVPTXISD::Tex3DS32FloatGrad:
787 return "NVPTXISD::Tex3DS32FloatGrad";
788 case NVPTXISD::Tex3DU32S32: return "NVPTXISD::Tex3DU32S32";
789 case NVPTXISD::Tex3DU32Float: return "NVPTXISD::Tex3DU32Float";
790 case NVPTXISD::Tex3DU32FloatLevel:
791 return "NVPTXISD::Tex3DU32FloatLevel";
792 case NVPTXISD::Tex3DU32FloatGrad:
793 return "NVPTXISD::Tex3DU32FloatGrad";
794 case NVPTXISD::TexCubeFloatFloat: return "NVPTXISD::TexCubeFloatFloat";
795 case NVPTXISD::TexCubeFloatFloatLevel:
796 return "NVPTXISD::TexCubeFloatFloatLevel";
797 case NVPTXISD::TexCubeS32Float: return "NVPTXISD::TexCubeS32Float";
798 case NVPTXISD::TexCubeS32FloatLevel:
799 return "NVPTXISD::TexCubeS32FloatLevel";
800 case NVPTXISD::TexCubeU32Float: return "NVPTXISD::TexCubeU32Float";
801 case NVPTXISD::TexCubeU32FloatLevel:
802 return "NVPTXISD::TexCubeU32FloatLevel";
803 case NVPTXISD::TexCubeArrayFloatFloat:
804 return "NVPTXISD::TexCubeArrayFloatFloat";
805 case NVPTXISD::TexCubeArrayFloatFloatLevel:
806 return "NVPTXISD::TexCubeArrayFloatFloatLevel";
807 case NVPTXISD::TexCubeArrayS32Float:
808 return "NVPTXISD::TexCubeArrayS32Float";
809 case NVPTXISD::TexCubeArrayS32FloatLevel:
810 return "NVPTXISD::TexCubeArrayS32FloatLevel";
811 case NVPTXISD::TexCubeArrayU32Float:
812 return "NVPTXISD::TexCubeArrayU32Float";
813 case NVPTXISD::TexCubeArrayU32FloatLevel:
814 return "NVPTXISD::TexCubeArrayU32FloatLevel";
815 case NVPTXISD::Tld4R2DFloatFloat:
816 return "NVPTXISD::Tld4R2DFloatFloat";
817 case NVPTXISD::Tld4G2DFloatFloat:
818 return "NVPTXISD::Tld4G2DFloatFloat";
819 case NVPTXISD::Tld4B2DFloatFloat:
820 return "NVPTXISD::Tld4B2DFloatFloat";
821 case NVPTXISD::Tld4A2DFloatFloat:
822 return "NVPTXISD::Tld4A2DFloatFloat";
823 case NVPTXISD::Tld4R2DS64Float:
824 return "NVPTXISD::Tld4R2DS64Float";
825 case NVPTXISD::Tld4G2DS64Float:
826 return "NVPTXISD::Tld4G2DS64Float";
827 case NVPTXISD::Tld4B2DS64Float:
828 return "NVPTXISD::Tld4B2DS64Float";
829 case NVPTXISD::Tld4A2DS64Float:
830 return "NVPTXISD::Tld4A2DS64Float";
831 case NVPTXISD::Tld4R2DU64Float:
832 return "NVPTXISD::Tld4R2DU64Float";
833 case NVPTXISD::Tld4G2DU64Float:
834 return "NVPTXISD::Tld4G2DU64Float";
835 case NVPTXISD::Tld4B2DU64Float:
836 return "NVPTXISD::Tld4B2DU64Float";
837 case NVPTXISD::Tld4A2DU64Float:
838 return "NVPTXISD::Tld4A2DU64Float";
839
840 case NVPTXISD::TexUnified1DFloatS32:
841 return "NVPTXISD::TexUnified1DFloatS32";
842 case NVPTXISD::TexUnified1DFloatFloat:
843 return "NVPTXISD::TexUnified1DFloatFloat";
844 case NVPTXISD::TexUnified1DFloatFloatLevel:
845 return "NVPTXISD::TexUnified1DFloatFloatLevel";
846 case NVPTXISD::TexUnified1DFloatFloatGrad:
847 return "NVPTXISD::TexUnified1DFloatFloatGrad";
848 case NVPTXISD::TexUnified1DS32S32:
849 return "NVPTXISD::TexUnified1DS32S32";
850 case NVPTXISD::TexUnified1DS32Float:
851 return "NVPTXISD::TexUnified1DS32Float";
852 case NVPTXISD::TexUnified1DS32FloatLevel:
853 return "NVPTXISD::TexUnified1DS32FloatLevel";
854 case NVPTXISD::TexUnified1DS32FloatGrad:
855 return "NVPTXISD::TexUnified1DS32FloatGrad";
856 case NVPTXISD::TexUnified1DU32S32:
857 return "NVPTXISD::TexUnified1DU32S32";
858 case NVPTXISD::TexUnified1DU32Float:
859 return "NVPTXISD::TexUnified1DU32Float";
860 case NVPTXISD::TexUnified1DU32FloatLevel:
861 return "NVPTXISD::TexUnified1DU32FloatLevel";
862 case NVPTXISD::TexUnified1DU32FloatGrad:
863 return "NVPTXISD::TexUnified1DU32FloatGrad";
864 case NVPTXISD::TexUnified1DArrayFloatS32:
865 return "NVPTXISD::TexUnified1DArrayFloatS32";
866 case NVPTXISD::TexUnified1DArrayFloatFloat:
867 return "NVPTXISD::TexUnified1DArrayFloatFloat";
868 case NVPTXISD::TexUnified1DArrayFloatFloatLevel:
869 return "NVPTXISD::TexUnified1DArrayFloatFloatLevel";
870 case NVPTXISD::TexUnified1DArrayFloatFloatGrad:
871 return "NVPTXISD::TexUnified1DArrayFloatFloatGrad";
872 case NVPTXISD::TexUnified1DArrayS32S32:
873 return "NVPTXISD::TexUnified1DArrayS32S32";
874 case NVPTXISD::TexUnified1DArrayS32Float:
875 return "NVPTXISD::TexUnified1DArrayS32Float";
876 case NVPTXISD::TexUnified1DArrayS32FloatLevel:
877 return "NVPTXISD::TexUnified1DArrayS32FloatLevel";
878 case NVPTXISD::TexUnified1DArrayS32FloatGrad:
879 return "NVPTXISD::TexUnified1DArrayS32FloatGrad";
880 case NVPTXISD::TexUnified1DArrayU32S32:
881 return "NVPTXISD::TexUnified1DArrayU32S32";
882 case NVPTXISD::TexUnified1DArrayU32Float:
883 return "NVPTXISD::TexUnified1DArrayU32Float";
884 case NVPTXISD::TexUnified1DArrayU32FloatLevel:
885 return "NVPTXISD::TexUnified1DArrayU32FloatLevel";
886 case NVPTXISD::TexUnified1DArrayU32FloatGrad:
887 return "NVPTXISD::TexUnified1DArrayU32FloatGrad";
888 case NVPTXISD::TexUnified2DFloatS32:
889 return "NVPTXISD::TexUnified2DFloatS32";
890 case NVPTXISD::TexUnified2DFloatFloat:
891 return "NVPTXISD::TexUnified2DFloatFloat";
892 case NVPTXISD::TexUnified2DFloatFloatLevel:
893 return "NVPTXISD::TexUnified2DFloatFloatLevel";
894 case NVPTXISD::TexUnified2DFloatFloatGrad:
895 return "NVPTXISD::TexUnified2DFloatFloatGrad";
896 case NVPTXISD::TexUnified2DS32S32:
897 return "NVPTXISD::TexUnified2DS32S32";
898 case NVPTXISD::TexUnified2DS32Float:
899 return "NVPTXISD::TexUnified2DS32Float";
900 case NVPTXISD::TexUnified2DS32FloatLevel:
901 return "NVPTXISD::TexUnified2DS32FloatLevel";
902 case NVPTXISD::TexUnified2DS32FloatGrad:
903 return "NVPTXISD::TexUnified2DS32FloatGrad";
904 case NVPTXISD::TexUnified2DU32S32:
905 return "NVPTXISD::TexUnified2DU32S32";
906 case NVPTXISD::TexUnified2DU32Float:
907 return "NVPTXISD::TexUnified2DU32Float";
908 case NVPTXISD::TexUnified2DU32FloatLevel:
909 return "NVPTXISD::TexUnified2DU32FloatLevel";
910 case NVPTXISD::TexUnified2DU32FloatGrad:
911 return "NVPTXISD::TexUnified2DU32FloatGrad";
912 case NVPTXISD::TexUnified2DArrayFloatS32:
913 return "NVPTXISD::TexUnified2DArrayFloatS32";
914 case NVPTXISD::TexUnified2DArrayFloatFloat:
915 return "NVPTXISD::TexUnified2DArrayFloatFloat";
916 case NVPTXISD::TexUnified2DArrayFloatFloatLevel:
917 return "NVPTXISD::TexUnified2DArrayFloatFloatLevel";
918 case NVPTXISD::TexUnified2DArrayFloatFloatGrad:
919 return "NVPTXISD::TexUnified2DArrayFloatFloatGrad";
920 case NVPTXISD::TexUnified2DArrayS32S32:
921 return "NVPTXISD::TexUnified2DArrayS32S32";
922 case NVPTXISD::TexUnified2DArrayS32Float:
923 return "NVPTXISD::TexUnified2DArrayS32Float";
924 case NVPTXISD::TexUnified2DArrayS32FloatLevel:
925 return "NVPTXISD::TexUnified2DArrayS32FloatLevel";
926 case NVPTXISD::TexUnified2DArrayS32FloatGrad:
927 return "NVPTXISD::TexUnified2DArrayS32FloatGrad";
928 case NVPTXISD::TexUnified2DArrayU32S32:
929 return "NVPTXISD::TexUnified2DArrayU32S32";
930 case NVPTXISD::TexUnified2DArrayU32Float:
931 return "NVPTXISD::TexUnified2DArrayU32Float";
932 case NVPTXISD::TexUnified2DArrayU32FloatLevel:
933 return "NVPTXISD::TexUnified2DArrayU32FloatLevel";
934 case NVPTXISD::TexUnified2DArrayU32FloatGrad:
935 return "NVPTXISD::TexUnified2DArrayU32FloatGrad";
936 case NVPTXISD::TexUnified3DFloatS32:
937 return "NVPTXISD::TexUnified3DFloatS32";
938 case NVPTXISD::TexUnified3DFloatFloat:
939 return "NVPTXISD::TexUnified3DFloatFloat";
940 case NVPTXISD::TexUnified3DFloatFloatLevel:
941 return "NVPTXISD::TexUnified3DFloatFloatLevel";
942 case NVPTXISD::TexUnified3DFloatFloatGrad:
943 return "NVPTXISD::TexUnified3DFloatFloatGrad";
944 case NVPTXISD::TexUnified3DS32S32:
945 return "NVPTXISD::TexUnified3DS32S32";
946 case NVPTXISD::TexUnified3DS32Float:
947 return "NVPTXISD::TexUnified3DS32Float";
948 case NVPTXISD::TexUnified3DS32FloatLevel:
949 return "NVPTXISD::TexUnified3DS32FloatLevel";
950 case NVPTXISD::TexUnified3DS32FloatGrad:
951 return "NVPTXISD::TexUnified3DS32FloatGrad";
952 case NVPTXISD::TexUnified3DU32S32:
953 return "NVPTXISD::TexUnified3DU32S32";
954 case NVPTXISD::TexUnified3DU32Float:
955 return "NVPTXISD::TexUnified3DU32Float";
956 case NVPTXISD::TexUnified3DU32FloatLevel:
957 return "NVPTXISD::TexUnified3DU32FloatLevel";
958 case NVPTXISD::TexUnified3DU32FloatGrad:
959 return "NVPTXISD::TexUnified3DU32FloatGrad";
960 case NVPTXISD::TexUnifiedCubeFloatFloat:
961 return "NVPTXISD::TexUnifiedCubeFloatFloat";
962 case NVPTXISD::TexUnifiedCubeFloatFloatLevel:
963 return "NVPTXISD::TexUnifiedCubeFloatFloatLevel";
964 case NVPTXISD::TexUnifiedCubeS32Float:
965 return "NVPTXISD::TexUnifiedCubeS32Float";
966 case NVPTXISD::TexUnifiedCubeS32FloatLevel:
967 return "NVPTXISD::TexUnifiedCubeS32FloatLevel";
968 case NVPTXISD::TexUnifiedCubeU32Float:
969 return "NVPTXISD::TexUnifiedCubeU32Float";
970 case NVPTXISD::TexUnifiedCubeU32FloatLevel:
971 return "NVPTXISD::TexUnifiedCubeU32FloatLevel";
972 case NVPTXISD::TexUnifiedCubeArrayFloatFloat:
973 return "NVPTXISD::TexUnifiedCubeArrayFloatFloat";
974 case NVPTXISD::TexUnifiedCubeArrayFloatFloatLevel:
975 return "NVPTXISD::TexUnifiedCubeArrayFloatFloatLevel";
976 case NVPTXISD::TexUnifiedCubeArrayS32Float:
977 return "NVPTXISD::TexUnifiedCubeArrayS32Float";
978 case NVPTXISD::TexUnifiedCubeArrayS32FloatLevel:
979 return "NVPTXISD::TexUnifiedCubeArrayS32FloatLevel";
980 case NVPTXISD::TexUnifiedCubeArrayU32Float:
981 return "NVPTXISD::TexUnifiedCubeArrayU32Float";
982 case NVPTXISD::TexUnifiedCubeArrayU32FloatLevel:
983 return "NVPTXISD::TexUnifiedCubeArrayU32FloatLevel";
984 case NVPTXISD::Tld4UnifiedR2DFloatFloat:
985 return "NVPTXISD::Tld4UnifiedR2DFloatFloat";
986 case NVPTXISD::Tld4UnifiedG2DFloatFloat:
987 return "NVPTXISD::Tld4UnifiedG2DFloatFloat";
988 case NVPTXISD::Tld4UnifiedB2DFloatFloat:
989 return "NVPTXISD::Tld4UnifiedB2DFloatFloat";
990 case NVPTXISD::Tld4UnifiedA2DFloatFloat:
991 return "NVPTXISD::Tld4UnifiedA2DFloatFloat";
992 case NVPTXISD::Tld4UnifiedR2DS64Float:
993 return "NVPTXISD::Tld4UnifiedR2DS64Float";
994 case NVPTXISD::Tld4UnifiedG2DS64Float:
995 return "NVPTXISD::Tld4UnifiedG2DS64Float";
996 case NVPTXISD::Tld4UnifiedB2DS64Float:
997 return "NVPTXISD::Tld4UnifiedB2DS64Float";
998 case NVPTXISD::Tld4UnifiedA2DS64Float:
999 return "NVPTXISD::Tld4UnifiedA2DS64Float";
1000 case NVPTXISD::Tld4UnifiedR2DU64Float:
1001 return "NVPTXISD::Tld4UnifiedR2DU64Float";
1002 case NVPTXISD::Tld4UnifiedG2DU64Float:
1003 return "NVPTXISD::Tld4UnifiedG2DU64Float";
1004 case NVPTXISD::Tld4UnifiedB2DU64Float:
1005 return "NVPTXISD::Tld4UnifiedB2DU64Float";
1006 case NVPTXISD::Tld4UnifiedA2DU64Float:
1007 return "NVPTXISD::Tld4UnifiedA2DU64Float";
1008
1009 case NVPTXISD::Suld1DI8Clamp: return "NVPTXISD::Suld1DI8Clamp";
1010 case NVPTXISD::Suld1DI16Clamp: return "NVPTXISD::Suld1DI16Clamp";
1011 case NVPTXISD::Suld1DI32Clamp: return "NVPTXISD::Suld1DI32Clamp";
1012 case NVPTXISD::Suld1DI64Clamp: return "NVPTXISD::Suld1DI64Clamp";
1013 case NVPTXISD::Suld1DV2I8Clamp: return "NVPTXISD::Suld1DV2I8Clamp";
1014 case NVPTXISD::Suld1DV2I16Clamp: return "NVPTXISD::Suld1DV2I16Clamp";
1015 case NVPTXISD::Suld1DV2I32Clamp: return "NVPTXISD::Suld1DV2I32Clamp";
1016 case NVPTXISD::Suld1DV2I64Clamp: return "NVPTXISD::Suld1DV2I64Clamp";
1017 case NVPTXISD::Suld1DV4I8Clamp: return "NVPTXISD::Suld1DV4I8Clamp";
1018 case NVPTXISD::Suld1DV4I16Clamp: return "NVPTXISD::Suld1DV4I16Clamp";
1019 case NVPTXISD::Suld1DV4I32Clamp: return "NVPTXISD::Suld1DV4I32Clamp";
1020
1021 case NVPTXISD::Suld1DArrayI8Clamp: return "NVPTXISD::Suld1DArrayI8Clamp";
1022 case NVPTXISD::Suld1DArrayI16Clamp: return "NVPTXISD::Suld1DArrayI16Clamp";
1023 case NVPTXISD::Suld1DArrayI32Clamp: return "NVPTXISD::Suld1DArrayI32Clamp";
1024 case NVPTXISD::Suld1DArrayI64Clamp: return "NVPTXISD::Suld1DArrayI64Clamp";
1025 case NVPTXISD::Suld1DArrayV2I8Clamp: return "NVPTXISD::Suld1DArrayV2I8Clamp";
1026 case NVPTXISD::Suld1DArrayV2I16Clamp:return "NVPTXISD::Suld1DArrayV2I16Clamp";
1027 case NVPTXISD::Suld1DArrayV2I32Clamp:return "NVPTXISD::Suld1DArrayV2I32Clamp";
1028 case NVPTXISD::Suld1DArrayV2I64Clamp:return "NVPTXISD::Suld1DArrayV2I64Clamp";
1029 case NVPTXISD::Suld1DArrayV4I8Clamp: return "NVPTXISD::Suld1DArrayV4I8Clamp";
1030 case NVPTXISD::Suld1DArrayV4I16Clamp:return "NVPTXISD::Suld1DArrayV4I16Clamp";
1031 case NVPTXISD::Suld1DArrayV4I32Clamp:return "NVPTXISD::Suld1DArrayV4I32Clamp";
1032
1033 case NVPTXISD::Suld2DI8Clamp: return "NVPTXISD::Suld2DI8Clamp";
1034 case NVPTXISD::Suld2DI16Clamp: return "NVPTXISD::Suld2DI16Clamp";
1035 case NVPTXISD::Suld2DI32Clamp: return "NVPTXISD::Suld2DI32Clamp";
1036 case NVPTXISD::Suld2DI64Clamp: return "NVPTXISD::Suld2DI64Clamp";
1037 case NVPTXISD::Suld2DV2I8Clamp: return "NVPTXISD::Suld2DV2I8Clamp";
1038 case NVPTXISD::Suld2DV2I16Clamp: return "NVPTXISD::Suld2DV2I16Clamp";
1039 case NVPTXISD::Suld2DV2I32Clamp: return "NVPTXISD::Suld2DV2I32Clamp";
1040 case NVPTXISD::Suld2DV2I64Clamp: return "NVPTXISD::Suld2DV2I64Clamp";
1041 case NVPTXISD::Suld2DV4I8Clamp: return "NVPTXISD::Suld2DV4I8Clamp";
1042 case NVPTXISD::Suld2DV4I16Clamp: return "NVPTXISD::Suld2DV4I16Clamp";
1043 case NVPTXISD::Suld2DV4I32Clamp: return "NVPTXISD::Suld2DV4I32Clamp";
1044
1045 case NVPTXISD::Suld2DArrayI8Clamp: return "NVPTXISD::Suld2DArrayI8Clamp";
1046 case NVPTXISD::Suld2DArrayI16Clamp: return "NVPTXISD::Suld2DArrayI16Clamp";
1047 case NVPTXISD::Suld2DArrayI32Clamp: return "NVPTXISD::Suld2DArrayI32Clamp";
1048 case NVPTXISD::Suld2DArrayI64Clamp: return "NVPTXISD::Suld2DArrayI64Clamp";
1049 case NVPTXISD::Suld2DArrayV2I8Clamp: return "NVPTXISD::Suld2DArrayV2I8Clamp";
1050 case NVPTXISD::Suld2DArrayV2I16Clamp:return "NVPTXISD::Suld2DArrayV2I16Clamp";
1051 case NVPTXISD::Suld2DArrayV2I32Clamp:return "NVPTXISD::Suld2DArrayV2I32Clamp";
1052 case NVPTXISD::Suld2DArrayV2I64Clamp:return "NVPTXISD::Suld2DArrayV2I64Clamp";
1053 case NVPTXISD::Suld2DArrayV4I8Clamp: return "NVPTXISD::Suld2DArrayV4I8Clamp";
1054 case NVPTXISD::Suld2DArrayV4I16Clamp:return "NVPTXISD::Suld2DArrayV4I16Clamp";
1055 case NVPTXISD::Suld2DArrayV4I32Clamp:return "NVPTXISD::Suld2DArrayV4I32Clamp";
1056
1057 case NVPTXISD::Suld3DI8Clamp: return "NVPTXISD::Suld3DI8Clamp";
1058 case NVPTXISD::Suld3DI16Clamp: return "NVPTXISD::Suld3DI16Clamp";
1059 case NVPTXISD::Suld3DI32Clamp: return "NVPTXISD::Suld3DI32Clamp";
1060 case NVPTXISD::Suld3DI64Clamp: return "NVPTXISD::Suld3DI64Clamp";
1061 case NVPTXISD::Suld3DV2I8Clamp: return "NVPTXISD::Suld3DV2I8Clamp";
1062 case NVPTXISD::Suld3DV2I16Clamp: return "NVPTXISD::Suld3DV2I16Clamp";
1063 case NVPTXISD::Suld3DV2I32Clamp: return "NVPTXISD::Suld3DV2I32Clamp";
1064 case NVPTXISD::Suld3DV2I64Clamp: return "NVPTXISD::Suld3DV2I64Clamp";
1065 case NVPTXISD::Suld3DV4I8Clamp: return "NVPTXISD::Suld3DV4I8Clamp";
1066 case NVPTXISD::Suld3DV4I16Clamp: return "NVPTXISD::Suld3DV4I16Clamp";
1067 case NVPTXISD::Suld3DV4I32Clamp: return "NVPTXISD::Suld3DV4I32Clamp";
1068
1069 case NVPTXISD::Suld1DI8Trap: return "NVPTXISD::Suld1DI8Trap";
1070 case NVPTXISD::Suld1DI16Trap: return "NVPTXISD::Suld1DI16Trap";
1071 case NVPTXISD::Suld1DI32Trap: return "NVPTXISD::Suld1DI32Trap";
1072 case NVPTXISD::Suld1DI64Trap: return "NVPTXISD::Suld1DI64Trap";
1073 case NVPTXISD::Suld1DV2I8Trap: return "NVPTXISD::Suld1DV2I8Trap";
1074 case NVPTXISD::Suld1DV2I16Trap: return "NVPTXISD::Suld1DV2I16Trap";
1075 case NVPTXISD::Suld1DV2I32Trap: return "NVPTXISD::Suld1DV2I32Trap";
1076 case NVPTXISD::Suld1DV2I64Trap: return "NVPTXISD::Suld1DV2I64Trap";
1077 case NVPTXISD::Suld1DV4I8Trap: return "NVPTXISD::Suld1DV4I8Trap";
1078 case NVPTXISD::Suld1DV4I16Trap: return "NVPTXISD::Suld1DV4I16Trap";
1079 case NVPTXISD::Suld1DV4I32Trap: return "NVPTXISD::Suld1DV4I32Trap";
1080
1081 case NVPTXISD::Suld1DArrayI8Trap: return "NVPTXISD::Suld1DArrayI8Trap";
1082 case NVPTXISD::Suld1DArrayI16Trap: return "NVPTXISD::Suld1DArrayI16Trap";
1083 case NVPTXISD::Suld1DArrayI32Trap: return "NVPTXISD::Suld1DArrayI32Trap";
1084 case NVPTXISD::Suld1DArrayI64Trap: return "NVPTXISD::Suld1DArrayI64Trap";
1085 case NVPTXISD::Suld1DArrayV2I8Trap: return "NVPTXISD::Suld1DArrayV2I8Trap";
1086 case NVPTXISD::Suld1DArrayV2I16Trap: return "NVPTXISD::Suld1DArrayV2I16Trap";
1087 case NVPTXISD::Suld1DArrayV2I32Trap: return "NVPTXISD::Suld1DArrayV2I32Trap";
1088 case NVPTXISD::Suld1DArrayV2I64Trap: return "NVPTXISD::Suld1DArrayV2I64Trap";
1089 case NVPTXISD::Suld1DArrayV4I8Trap: return "NVPTXISD::Suld1DArrayV4I8Trap";
1090 case NVPTXISD::Suld1DArrayV4I16Trap: return "NVPTXISD::Suld1DArrayV4I16Trap";
1091 case NVPTXISD::Suld1DArrayV4I32Trap: return "NVPTXISD::Suld1DArrayV4I32Trap";
1092
1093 case NVPTXISD::Suld2DI8Trap: return "NVPTXISD::Suld2DI8Trap";
1094 case NVPTXISD::Suld2DI16Trap: return "NVPTXISD::Suld2DI16Trap";
1095 case NVPTXISD::Suld2DI32Trap: return "NVPTXISD::Suld2DI32Trap";
1096 case NVPTXISD::Suld2DI64Trap: return "NVPTXISD::Suld2DI64Trap";
1097 case NVPTXISD::Suld2DV2I8Trap: return "NVPTXISD::Suld2DV2I8Trap";
1098 case NVPTXISD::Suld2DV2I16Trap: return "NVPTXISD::Suld2DV2I16Trap";
1099 case NVPTXISD::Suld2DV2I32Trap: return "NVPTXISD::Suld2DV2I32Trap";
1100 case NVPTXISD::Suld2DV2I64Trap: return "NVPTXISD::Suld2DV2I64Trap";
1101 case NVPTXISD::Suld2DV4I8Trap: return "NVPTXISD::Suld2DV4I8Trap";
1102 case NVPTXISD::Suld2DV4I16Trap: return "NVPTXISD::Suld2DV4I16Trap";
1103 case NVPTXISD::Suld2DV4I32Trap: return "NVPTXISD::Suld2DV4I32Trap";
1104
1105 case NVPTXISD::Suld2DArrayI8Trap: return "NVPTXISD::Suld2DArrayI8Trap";
1106 case NVPTXISD::Suld2DArrayI16Trap: return "NVPTXISD::Suld2DArrayI16Trap";
1107 case NVPTXISD::Suld2DArrayI32Trap: return "NVPTXISD::Suld2DArrayI32Trap";
1108 case NVPTXISD::Suld2DArrayI64Trap: return "NVPTXISD::Suld2DArrayI64Trap";
1109 case NVPTXISD::Suld2DArrayV2I8Trap: return "NVPTXISD::Suld2DArrayV2I8Trap";
1110 case NVPTXISD::Suld2DArrayV2I16Trap: return "NVPTXISD::Suld2DArrayV2I16Trap";
1111 case NVPTXISD::Suld2DArrayV2I32Trap: return "NVPTXISD::Suld2DArrayV2I32Trap";
1112 case NVPTXISD::Suld2DArrayV2I64Trap: return "NVPTXISD::Suld2DArrayV2I64Trap";
1113 case NVPTXISD::Suld2DArrayV4I8Trap: return "NVPTXISD::Suld2DArrayV4I8Trap";
1114 case NVPTXISD::Suld2DArrayV4I16Trap: return "NVPTXISD::Suld2DArrayV4I16Trap";
1115 case NVPTXISD::Suld2DArrayV4I32Trap: return "NVPTXISD::Suld2DArrayV4I32Trap";
1116
1117 case NVPTXISD::Suld3DI8Trap: return "NVPTXISD::Suld3DI8Trap";
1118 case NVPTXISD::Suld3DI16Trap: return "NVPTXISD::Suld3DI16Trap";
1119 case NVPTXISD::Suld3DI32Trap: return "NVPTXISD::Suld3DI32Trap";
1120 case NVPTXISD::Suld3DI64Trap: return "NVPTXISD::Suld3DI64Trap";
1121 case NVPTXISD::Suld3DV2I8Trap: return "NVPTXISD::Suld3DV2I8Trap";
1122 case NVPTXISD::Suld3DV2I16Trap: return "NVPTXISD::Suld3DV2I16Trap";
1123 case NVPTXISD::Suld3DV2I32Trap: return "NVPTXISD::Suld3DV2I32Trap";
1124 case NVPTXISD::Suld3DV2I64Trap: return "NVPTXISD::Suld3DV2I64Trap";
1125 case NVPTXISD::Suld3DV4I8Trap: return "NVPTXISD::Suld3DV4I8Trap";
1126 case NVPTXISD::Suld3DV4I16Trap: return "NVPTXISD::Suld3DV4I16Trap";
1127 case NVPTXISD::Suld3DV4I32Trap: return "NVPTXISD::Suld3DV4I32Trap";
1128
1129 case NVPTXISD::Suld1DI8Zero: return "NVPTXISD::Suld1DI8Zero";
1130 case NVPTXISD::Suld1DI16Zero: return "NVPTXISD::Suld1DI16Zero";
1131 case NVPTXISD::Suld1DI32Zero: return "NVPTXISD::Suld1DI32Zero";
1132 case NVPTXISD::Suld1DI64Zero: return "NVPTXISD::Suld1DI64Zero";
1133 case NVPTXISD::Suld1DV2I8Zero: return "NVPTXISD::Suld1DV2I8Zero";
1134 case NVPTXISD::Suld1DV2I16Zero: return "NVPTXISD::Suld1DV2I16Zero";
1135 case NVPTXISD::Suld1DV2I32Zero: return "NVPTXISD::Suld1DV2I32Zero";
1136 case NVPTXISD::Suld1DV2I64Zero: return "NVPTXISD::Suld1DV2I64Zero";
1137 case NVPTXISD::Suld1DV4I8Zero: return "NVPTXISD::Suld1DV4I8Zero";
1138 case NVPTXISD::Suld1DV4I16Zero: return "NVPTXISD::Suld1DV4I16Zero";
1139 case NVPTXISD::Suld1DV4I32Zero: return "NVPTXISD::Suld1DV4I32Zero";
1140
1141 case NVPTXISD::Suld1DArrayI8Zero: return "NVPTXISD::Suld1DArrayI8Zero";
1142 case NVPTXISD::Suld1DArrayI16Zero: return "NVPTXISD::Suld1DArrayI16Zero";
1143 case NVPTXISD::Suld1DArrayI32Zero: return "NVPTXISD::Suld1DArrayI32Zero";
1144 case NVPTXISD::Suld1DArrayI64Zero: return "NVPTXISD::Suld1DArrayI64Zero";
1145 case NVPTXISD::Suld1DArrayV2I8Zero: return "NVPTXISD::Suld1DArrayV2I8Zero";
1146 case NVPTXISD::Suld1DArrayV2I16Zero: return "NVPTXISD::Suld1DArrayV2I16Zero";
1147 case NVPTXISD::Suld1DArrayV2I32Zero: return "NVPTXISD::Suld1DArrayV2I32Zero";
1148 case NVPTXISD::Suld1DArrayV2I64Zero: return "NVPTXISD::Suld1DArrayV2I64Zero";
1149 case NVPTXISD::Suld1DArrayV4I8Zero: return "NVPTXISD::Suld1DArrayV4I8Zero";
1150 case NVPTXISD::Suld1DArrayV4I16Zero: return "NVPTXISD::Suld1DArrayV4I16Zero";
1151 case NVPTXISD::Suld1DArrayV4I32Zero: return "NVPTXISD::Suld1DArrayV4I32Zero";
1152
1153 case NVPTXISD::Suld2DI8Zero: return "NVPTXISD::Suld2DI8Zero";
1154 case NVPTXISD::Suld2DI16Zero: return "NVPTXISD::Suld2DI16Zero";
1155 case NVPTXISD::Suld2DI32Zero: return "NVPTXISD::Suld2DI32Zero";
1156 case NVPTXISD::Suld2DI64Zero: return "NVPTXISD::Suld2DI64Zero";
1157 case NVPTXISD::Suld2DV2I8Zero: return "NVPTXISD::Suld2DV2I8Zero";
1158 case NVPTXISD::Suld2DV2I16Zero: return "NVPTXISD::Suld2DV2I16Zero";
1159 case NVPTXISD::Suld2DV2I32Zero: return "NVPTXISD::Suld2DV2I32Zero";
1160 case NVPTXISD::Suld2DV2I64Zero: return "NVPTXISD::Suld2DV2I64Zero";
1161 case NVPTXISD::Suld2DV4I8Zero: return "NVPTXISD::Suld2DV4I8Zero";
1162 case NVPTXISD::Suld2DV4I16Zero: return "NVPTXISD::Suld2DV4I16Zero";
1163 case NVPTXISD::Suld2DV4I32Zero: return "NVPTXISD::Suld2DV4I32Zero";
1164
1165 case NVPTXISD::Suld2DArrayI8Zero: return "NVPTXISD::Suld2DArrayI8Zero";
1166 case NVPTXISD::Suld2DArrayI16Zero: return "NVPTXISD::Suld2DArrayI16Zero";
1167 case NVPTXISD::Suld2DArrayI32Zero: return "NVPTXISD::Suld2DArrayI32Zero";
1168 case NVPTXISD::Suld2DArrayI64Zero: return "NVPTXISD::Suld2DArrayI64Zero";
1169 case NVPTXISD::Suld2DArrayV2I8Zero: return "NVPTXISD::Suld2DArrayV2I8Zero";
1170 case NVPTXISD::Suld2DArrayV2I16Zero: return "NVPTXISD::Suld2DArrayV2I16Zero";
1171 case NVPTXISD::Suld2DArrayV2I32Zero: return "NVPTXISD::Suld2DArrayV2I32Zero";
1172 case NVPTXISD::Suld2DArrayV2I64Zero: return "NVPTXISD::Suld2DArrayV2I64Zero";
1173 case NVPTXISD::Suld2DArrayV4I8Zero: return "NVPTXISD::Suld2DArrayV4I8Zero";
1174 case NVPTXISD::Suld2DArrayV4I16Zero: return "NVPTXISD::Suld2DArrayV4I16Zero";
1175 case NVPTXISD::Suld2DArrayV4I32Zero: return "NVPTXISD::Suld2DArrayV4I32Zero";
1176
1177 case NVPTXISD::Suld3DI8Zero: return "NVPTXISD::Suld3DI8Zero";
1178 case NVPTXISD::Suld3DI16Zero: return "NVPTXISD::Suld3DI16Zero";
1179 case NVPTXISD::Suld3DI32Zero: return "NVPTXISD::Suld3DI32Zero";
1180 case NVPTXISD::Suld3DI64Zero: return "NVPTXISD::Suld3DI64Zero";
1181 case NVPTXISD::Suld3DV2I8Zero: return "NVPTXISD::Suld3DV2I8Zero";
1182 case NVPTXISD::Suld3DV2I16Zero: return "NVPTXISD::Suld3DV2I16Zero";
1183 case NVPTXISD::Suld3DV2I32Zero: return "NVPTXISD::Suld3DV2I32Zero";
1184 case NVPTXISD::Suld3DV2I64Zero: return "NVPTXISD::Suld3DV2I64Zero";
1185 case NVPTXISD::Suld3DV4I8Zero: return "NVPTXISD::Suld3DV4I8Zero";
1186 case NVPTXISD::Suld3DV4I16Zero: return "NVPTXISD::Suld3DV4I16Zero";
1187 case NVPTXISD::Suld3DV4I32Zero: return "NVPTXISD::Suld3DV4I32Zero";
1188 }
1189 return nullptr;
1190 }
1191
1192 TargetLoweringBase::LegalizeTypeAction
getPreferredVectorAction(MVT VT) const1193 NVPTXTargetLowering::getPreferredVectorAction(MVT VT) const {
1194 if (VT.getVectorNumElements() != 1 && VT.getScalarType() == MVT::i1)
1195 return TypeSplitVector;
1196 if (VT == MVT::v2f16)
1197 return TypeLegal;
1198 return TargetLoweringBase::getPreferredVectorAction(VT);
1199 }
1200
getSqrtEstimate(SDValue Operand,SelectionDAG & DAG,int Enabled,int & ExtraSteps,bool & UseOneConst,bool Reciprocal) const1201 SDValue NVPTXTargetLowering::getSqrtEstimate(SDValue Operand, SelectionDAG &DAG,
1202 int Enabled, int &ExtraSteps,
1203 bool &UseOneConst,
1204 bool Reciprocal) const {
1205 if (!(Enabled == ReciprocalEstimate::Enabled ||
1206 (Enabled == ReciprocalEstimate::Unspecified && !usePrecSqrtF32())))
1207 return SDValue();
1208
1209 if (ExtraSteps == ReciprocalEstimate::Unspecified)
1210 ExtraSteps = 0;
1211
1212 SDLoc DL(Operand);
1213 EVT VT = Operand.getValueType();
1214 bool Ftz = useF32FTZ(DAG.getMachineFunction());
1215
1216 auto MakeIntrinsicCall = [&](Intrinsic::ID IID) {
1217 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
1218 DAG.getConstant(IID, DL, MVT::i32), Operand);
1219 };
1220
1221 // The sqrt and rsqrt refinement processes assume we always start out with an
1222 // approximation of the rsqrt. Therefore, if we're going to do any refinement
1223 // (i.e. ExtraSteps > 0), we must return an rsqrt. But if we're *not* doing
1224 // any refinement, we must return a regular sqrt.
1225 if (Reciprocal || ExtraSteps > 0) {
1226 if (VT == MVT::f32)
1227 return MakeIntrinsicCall(Ftz ? Intrinsic::nvvm_rsqrt_approx_ftz_f
1228 : Intrinsic::nvvm_rsqrt_approx_f);
1229 else if (VT == MVT::f64)
1230 return MakeIntrinsicCall(Intrinsic::nvvm_rsqrt_approx_d);
1231 else
1232 return SDValue();
1233 } else {
1234 if (VT == MVT::f32)
1235 return MakeIntrinsicCall(Ftz ? Intrinsic::nvvm_sqrt_approx_ftz_f
1236 : Intrinsic::nvvm_sqrt_approx_f);
1237 else {
1238 // There's no sqrt.approx.f64 instruction, so we emit
1239 // reciprocal(rsqrt(x)). This is faster than
1240 // select(x == 0, 0, x * rsqrt(x)). (In fact, it's faster than plain
1241 // x * rsqrt(x).)
1242 return DAG.getNode(
1243 ISD::INTRINSIC_WO_CHAIN, DL, VT,
1244 DAG.getConstant(Intrinsic::nvvm_rcp_approx_ftz_d, DL, MVT::i32),
1245 MakeIntrinsicCall(Intrinsic::nvvm_rsqrt_approx_d));
1246 }
1247 }
1248 }
1249
1250 SDValue
LowerGlobalAddress(SDValue Op,SelectionDAG & DAG) const1251 NVPTXTargetLowering::LowerGlobalAddress(SDValue Op, SelectionDAG &DAG) const {
1252 SDLoc dl(Op);
1253 const GlobalAddressSDNode *GAN = cast<GlobalAddressSDNode>(Op);
1254 auto PtrVT = getPointerTy(DAG.getDataLayout(), GAN->getAddressSpace());
1255 Op = DAG.getTargetGlobalAddress(GAN->getGlobal(), dl, PtrVT);
1256 return DAG.getNode(NVPTXISD::Wrapper, dl, PtrVT, Op);
1257 }
1258
getPrototype(const DataLayout & DL,Type * retTy,const ArgListTy & Args,const SmallVectorImpl<ISD::OutputArg> & Outs,unsigned retAlignment,ImmutableCallSite CS) const1259 std::string NVPTXTargetLowering::getPrototype(
1260 const DataLayout &DL, Type *retTy, const ArgListTy &Args,
1261 const SmallVectorImpl<ISD::OutputArg> &Outs, unsigned retAlignment,
1262 ImmutableCallSite CS) const {
1263 auto PtrVT = getPointerTy(DL);
1264
1265 bool isABI = (STI.getSmVersion() >= 20);
1266 assert(isABI && "Non-ABI compilation is not supported");
1267 if (!isABI)
1268 return "";
1269
1270 std::stringstream O;
1271 O << "prototype_" << uniqueCallSite << " : .callprototype ";
1272
1273 if (retTy->getTypeID() == Type::VoidTyID) {
1274 O << "()";
1275 } else {
1276 O << "(";
1277 if (retTy->isFloatingPointTy() || (retTy->isIntegerTy() && !retTy->isIntegerTy(128))) {
1278 unsigned size = 0;
1279 if (auto *ITy = dyn_cast<IntegerType>(retTy)) {
1280 size = ITy->getBitWidth();
1281 } else {
1282 assert(retTy->isFloatingPointTy() &&
1283 "Floating point type expected here");
1284 size = retTy->getPrimitiveSizeInBits();
1285 }
1286 // PTX ABI requires all scalar return values to be at least 32
1287 // bits in size. fp16 normally uses .b16 as its storage type in
1288 // PTX, so its size must be adjusted here, too.
1289 if (size < 32)
1290 size = 32;
1291
1292 O << ".param .b" << size << " _";
1293 } else if (isa<PointerType>(retTy)) {
1294 O << ".param .b" << PtrVT.getSizeInBits() << " _";
1295 } else if (retTy->isAggregateType() || retTy->isVectorTy() ||
1296 retTy->isIntegerTy(128)) {
1297 O << ".param .align " << retAlignment << " .b8 _["
1298 << DL.getTypeAllocSize(retTy) << "]";
1299 } else {
1300 llvm_unreachable("Unknown return type");
1301 }
1302 O << ") ";
1303 }
1304 O << "_ (";
1305
1306 bool first = true;
1307
1308 unsigned OIdx = 0;
1309 for (unsigned i = 0, e = Args.size(); i != e; ++i, ++OIdx) {
1310 Type *Ty = Args[i].Ty;
1311 if (!first) {
1312 O << ", ";
1313 }
1314 first = false;
1315
1316 if (!Outs[OIdx].Flags.isByVal()) {
1317 if (Ty->isAggregateType() || Ty->isVectorTy() || Ty->isIntegerTy(128)) {
1318 unsigned align = 0;
1319 const CallInst *CallI = cast<CallInst>(CS.getInstruction());
1320 // +1 because index 0 is reserved for return type alignment
1321 if (!getAlign(*CallI, i + 1, align))
1322 align = DL.getABITypeAlignment(Ty);
1323 unsigned sz = DL.getTypeAllocSize(Ty);
1324 O << ".param .align " << align << " .b8 ";
1325 O << "_";
1326 O << "[" << sz << "]";
1327 // update the index for Outs
1328 SmallVector<EVT, 16> vtparts;
1329 ComputeValueVTs(*this, DL, Ty, vtparts);
1330 if (unsigned len = vtparts.size())
1331 OIdx += len - 1;
1332 continue;
1333 }
1334 // i8 types in IR will be i16 types in SDAG
1335 assert((getValueType(DL, Ty) == Outs[OIdx].VT ||
1336 (getValueType(DL, Ty) == MVT::i8 && Outs[OIdx].VT == MVT::i16)) &&
1337 "type mismatch between callee prototype and arguments");
1338 // scalar type
1339 unsigned sz = 0;
1340 if (isa<IntegerType>(Ty)) {
1341 sz = cast<IntegerType>(Ty)->getBitWidth();
1342 if (sz < 32)
1343 sz = 32;
1344 } else if (isa<PointerType>(Ty)) {
1345 sz = PtrVT.getSizeInBits();
1346 } else if (Ty->isHalfTy())
1347 // PTX ABI requires all scalar parameters to be at least 32
1348 // bits in size. fp16 normally uses .b16 as its storage type
1349 // in PTX, so its size must be adjusted here, too.
1350 sz = 32;
1351 else
1352 sz = Ty->getPrimitiveSizeInBits();
1353 O << ".param .b" << sz << " ";
1354 O << "_";
1355 continue;
1356 }
1357 auto *PTy = dyn_cast<PointerType>(Ty);
1358 assert(PTy && "Param with byval attribute should be a pointer type");
1359 Type *ETy = PTy->getElementType();
1360
1361 unsigned align = Outs[OIdx].Flags.getByValAlign();
1362 unsigned sz = DL.getTypeAllocSize(ETy);
1363 O << ".param .align " << align << " .b8 ";
1364 O << "_";
1365 O << "[" << sz << "]";
1366 }
1367 O << ");";
1368 return O.str();
1369 }
1370
getArgumentAlignment(SDValue Callee,ImmutableCallSite CS,Type * Ty,unsigned Idx,const DataLayout & DL) const1371 unsigned NVPTXTargetLowering::getArgumentAlignment(SDValue Callee,
1372 ImmutableCallSite CS,
1373 Type *Ty, unsigned Idx,
1374 const DataLayout &DL) const {
1375 if (!CS) {
1376 // CallSite is zero, fallback to ABI type alignment
1377 return DL.getABITypeAlignment(Ty);
1378 }
1379
1380 unsigned Align = 0;
1381 const Value *DirectCallee = CS.getCalledFunction();
1382
1383 if (!DirectCallee) {
1384 // We don't have a direct function symbol, but that may be because of
1385 // constant cast instructions in the call.
1386 const Instruction *CalleeI = CS.getInstruction();
1387 assert(CalleeI && "Call target is not a function or derived value?");
1388
1389 // With bitcast'd call targets, the instruction will be the call
1390 if (isa<CallInst>(CalleeI)) {
1391 // Check if we have call alignment metadata
1392 if (getAlign(*cast<CallInst>(CalleeI), Idx, Align))
1393 return Align;
1394
1395 const Value *CalleeV = cast<CallInst>(CalleeI)->getCalledValue();
1396 // Ignore any bitcast instructions
1397 while (isa<ConstantExpr>(CalleeV)) {
1398 const ConstantExpr *CE = cast<ConstantExpr>(CalleeV);
1399 if (!CE->isCast())
1400 break;
1401 // Look through the bitcast
1402 CalleeV = cast<ConstantExpr>(CalleeV)->getOperand(0);
1403 }
1404
1405 // We have now looked past all of the bitcasts. Do we finally have a
1406 // Function?
1407 if (isa<Function>(CalleeV))
1408 DirectCallee = CalleeV;
1409 }
1410 }
1411
1412 // Check for function alignment information if we found that the
1413 // ultimate target is a Function
1414 if (DirectCallee)
1415 if (getAlign(*cast<Function>(DirectCallee), Idx, Align))
1416 return Align;
1417
1418 // Call is indirect or alignment information is not available, fall back to
1419 // the ABI type alignment
1420 return DL.getABITypeAlignment(Ty);
1421 }
1422
LowerCall(TargetLowering::CallLoweringInfo & CLI,SmallVectorImpl<SDValue> & InVals) const1423 SDValue NVPTXTargetLowering::LowerCall(TargetLowering::CallLoweringInfo &CLI,
1424 SmallVectorImpl<SDValue> &InVals) const {
1425 SelectionDAG &DAG = CLI.DAG;
1426 SDLoc dl = CLI.DL;
1427 SmallVectorImpl<ISD::OutputArg> &Outs = CLI.Outs;
1428 SmallVectorImpl<SDValue> &OutVals = CLI.OutVals;
1429 SmallVectorImpl<ISD::InputArg> &Ins = CLI.Ins;
1430 SDValue Chain = CLI.Chain;
1431 SDValue Callee = CLI.Callee;
1432 bool &isTailCall = CLI.IsTailCall;
1433 ArgListTy &Args = CLI.getArgs();
1434 Type *RetTy = CLI.RetTy;
1435 ImmutableCallSite CS = CLI.CS;
1436 const DataLayout &DL = DAG.getDataLayout();
1437
1438 bool isABI = (STI.getSmVersion() >= 20);
1439 assert(isABI && "Non-ABI compilation is not supported");
1440 if (!isABI)
1441 return Chain;
1442
1443 SDValue tempChain = Chain;
1444 Chain = DAG.getCALLSEQ_START(Chain, uniqueCallSite, 0, dl);
1445 SDValue InFlag = Chain.getValue(1);
1446
1447 unsigned paramCount = 0;
1448 // Args.size() and Outs.size() need not match.
1449 // Outs.size() will be larger
1450 // * if there is an aggregate argument with multiple fields (each field
1451 // showing up separately in Outs)
1452 // * if there is a vector argument with more than typical vector-length
1453 // elements (generally if more than 4) where each vector element is
1454 // individually present in Outs.
1455 // So a different index should be used for indexing into Outs/OutVals.
1456 // See similar issue in LowerFormalArguments.
1457 unsigned OIdx = 0;
1458 // Declare the .params or .reg need to pass values
1459 // to the function
1460 for (unsigned i = 0, e = Args.size(); i != e; ++i, ++OIdx) {
1461 EVT VT = Outs[OIdx].VT;
1462 Type *Ty = Args[i].Ty;
1463
1464 if (!Outs[OIdx].Flags.isByVal()) {
1465 SmallVector<EVT, 16> VTs;
1466 SmallVector<uint64_t, 16> Offsets;
1467 ComputePTXValueVTs(*this, DL, Ty, VTs, &Offsets);
1468 unsigned ArgAlign =
1469 getArgumentAlignment(Callee, CS, Ty, paramCount + 1, DL);
1470 unsigned AllocSize = DL.getTypeAllocSize(Ty);
1471 SDVTList DeclareParamVTs = DAG.getVTList(MVT::Other, MVT::Glue);
1472 bool NeedAlign; // Does argument declaration specify alignment?
1473 if (Ty->isAggregateType() || Ty->isVectorTy() || Ty->isIntegerTy(128)) {
1474 // declare .param .align <align> .b8 .param<n>[<size>];
1475 SDValue DeclareParamOps[] = {
1476 Chain, DAG.getConstant(ArgAlign, dl, MVT::i32),
1477 DAG.getConstant(paramCount, dl, MVT::i32),
1478 DAG.getConstant(AllocSize, dl, MVT::i32), InFlag};
1479 Chain = DAG.getNode(NVPTXISD::DeclareParam, dl, DeclareParamVTs,
1480 DeclareParamOps);
1481 NeedAlign = true;
1482 } else {
1483 // declare .param .b<size> .param<n>;
1484 if ((VT.isInteger() || VT.isFloatingPoint()) && AllocSize < 4) {
1485 // PTX ABI requires integral types to be at least 32 bits in
1486 // size. FP16 is loaded/stored using i16, so it's handled
1487 // here as well.
1488 AllocSize = 4;
1489 }
1490 SDValue DeclareScalarParamOps[] = {
1491 Chain, DAG.getConstant(paramCount, dl, MVT::i32),
1492 DAG.getConstant(AllocSize * 8, dl, MVT::i32),
1493 DAG.getConstant(0, dl, MVT::i32), InFlag};
1494 Chain = DAG.getNode(NVPTXISD::DeclareScalarParam, dl, DeclareParamVTs,
1495 DeclareScalarParamOps);
1496 NeedAlign = false;
1497 }
1498 InFlag = Chain.getValue(1);
1499
1500 // PTX Interoperability Guide 3.3(A): [Integer] Values shorter
1501 // than 32-bits are sign extended or zero extended, depending on
1502 // whether they are signed or unsigned types. This case applies
1503 // only to scalar parameters and not to aggregate values.
1504 bool ExtendIntegerParam =
1505 Ty->isIntegerTy() && DL.getTypeAllocSizeInBits(Ty) < 32;
1506
1507 auto VectorInfo = VectorizePTXValueVTs(VTs, Offsets, ArgAlign);
1508 SmallVector<SDValue, 6> StoreOperands;
1509 for (unsigned j = 0, je = VTs.size(); j != je; ++j) {
1510 // New store.
1511 if (VectorInfo[j] & PVF_FIRST) {
1512 assert(StoreOperands.empty() && "Unfinished preceding store.");
1513 StoreOperands.push_back(Chain);
1514 StoreOperands.push_back(DAG.getConstant(paramCount, dl, MVT::i32));
1515 StoreOperands.push_back(DAG.getConstant(Offsets[j], dl, MVT::i32));
1516 }
1517
1518 EVT EltVT = VTs[j];
1519 SDValue StVal = OutVals[OIdx];
1520 if (ExtendIntegerParam) {
1521 assert(VTs.size() == 1 && "Scalar can't have multiple parts.");
1522 // zext/sext to i32
1523 StVal = DAG.getNode(Outs[OIdx].Flags.isSExt() ? ISD::SIGN_EXTEND
1524 : ISD::ZERO_EXTEND,
1525 dl, MVT::i32, StVal);
1526 } else if (EltVT.getSizeInBits() < 16) {
1527 // Use 16-bit registers for small stores as it's the
1528 // smallest general purpose register size supported by NVPTX.
1529 StVal = DAG.getNode(ISD::ANY_EXTEND, dl, MVT::i16, StVal);
1530 }
1531
1532 // Record the value to store.
1533 StoreOperands.push_back(StVal);
1534
1535 if (VectorInfo[j] & PVF_LAST) {
1536 unsigned NumElts = StoreOperands.size() - 3;
1537 NVPTXISD::NodeType Op;
1538 switch (NumElts) {
1539 case 1:
1540 Op = NVPTXISD::StoreParam;
1541 break;
1542 case 2:
1543 Op = NVPTXISD::StoreParamV2;
1544 break;
1545 case 4:
1546 Op = NVPTXISD::StoreParamV4;
1547 break;
1548 default:
1549 llvm_unreachable("Invalid vector info.");
1550 }
1551
1552 StoreOperands.push_back(InFlag);
1553
1554 // Adjust type of the store op if we've extended the scalar
1555 // return value.
1556 EVT TheStoreType = ExtendIntegerParam ? MVT::i32 : VTs[j];
1557 unsigned EltAlign =
1558 NeedAlign ? GreatestCommonDivisor64(ArgAlign, Offsets[j]) : 0;
1559
1560 Chain = DAG.getMemIntrinsicNode(
1561 Op, dl, DAG.getVTList(MVT::Other, MVT::Glue), StoreOperands,
1562 TheStoreType, MachinePointerInfo(), EltAlign,
1563 MachineMemOperand::MOStore);
1564 InFlag = Chain.getValue(1);
1565
1566 // Cleanup.
1567 StoreOperands.clear();
1568 }
1569 ++OIdx;
1570 }
1571 assert(StoreOperands.empty() && "Unfinished parameter store.");
1572 if (VTs.size() > 0)
1573 --OIdx;
1574 ++paramCount;
1575 continue;
1576 }
1577
1578 // ByVal arguments
1579 SmallVector<EVT, 16> VTs;
1580 SmallVector<uint64_t, 16> Offsets;
1581 auto *PTy = dyn_cast<PointerType>(Args[i].Ty);
1582 assert(PTy && "Type of a byval parameter should be pointer");
1583 ComputePTXValueVTs(*this, DL, PTy->getElementType(), VTs, &Offsets, 0);
1584
1585 // declare .param .align <align> .b8 .param<n>[<size>];
1586 unsigned sz = Outs[OIdx].Flags.getByValSize();
1587 SDVTList DeclareParamVTs = DAG.getVTList(MVT::Other, MVT::Glue);
1588 unsigned ArgAlign = Outs[OIdx].Flags.getByValAlign();
1589 // The ByValAlign in the Outs[OIdx].Flags is alway set at this point,
1590 // so we don't need to worry about natural alignment or not.
1591 // See TargetLowering::LowerCallTo().
1592
1593 // Enforce minumum alignment of 4 to work around ptxas miscompile
1594 // for sm_50+. See corresponding alignment adjustment in
1595 // emitFunctionParamList() for details.
1596 if (ArgAlign < 4)
1597 ArgAlign = 4;
1598 SDValue DeclareParamOps[] = {Chain, DAG.getConstant(ArgAlign, dl, MVT::i32),
1599 DAG.getConstant(paramCount, dl, MVT::i32),
1600 DAG.getConstant(sz, dl, MVT::i32), InFlag};
1601 Chain = DAG.getNode(NVPTXISD::DeclareParam, dl, DeclareParamVTs,
1602 DeclareParamOps);
1603 InFlag = Chain.getValue(1);
1604 for (unsigned j = 0, je = VTs.size(); j != je; ++j) {
1605 EVT elemtype = VTs[j];
1606 int curOffset = Offsets[j];
1607 unsigned PartAlign = GreatestCommonDivisor64(ArgAlign, curOffset);
1608 auto PtrVT = getPointerTy(DL);
1609 SDValue srcAddr = DAG.getNode(ISD::ADD, dl, PtrVT, OutVals[OIdx],
1610 DAG.getConstant(curOffset, dl, PtrVT));
1611 SDValue theVal = DAG.getLoad(elemtype, dl, tempChain, srcAddr,
1612 MachinePointerInfo(), PartAlign);
1613 if (elemtype.getSizeInBits() < 16) {
1614 theVal = DAG.getNode(ISD::ANY_EXTEND, dl, MVT::i16, theVal);
1615 }
1616 SDVTList CopyParamVTs = DAG.getVTList(MVT::Other, MVT::Glue);
1617 SDValue CopyParamOps[] = { Chain,
1618 DAG.getConstant(paramCount, dl, MVT::i32),
1619 DAG.getConstant(curOffset, dl, MVT::i32),
1620 theVal, InFlag };
1621 Chain = DAG.getMemIntrinsicNode(NVPTXISD::StoreParam, dl, CopyParamVTs,
1622 CopyParamOps, elemtype,
1623 MachinePointerInfo(), /* Align */ 0,
1624 MachineMemOperand::MOStore);
1625
1626 InFlag = Chain.getValue(1);
1627 }
1628 ++paramCount;
1629 }
1630
1631 GlobalAddressSDNode *Func = dyn_cast<GlobalAddressSDNode>(Callee.getNode());
1632 unsigned retAlignment = 0;
1633
1634 // Handle Result
1635 if (Ins.size() > 0) {
1636 SmallVector<EVT, 16> resvtparts;
1637 ComputeValueVTs(*this, DL, RetTy, resvtparts);
1638
1639 // Declare
1640 // .param .align 16 .b8 retval0[<size-in-bytes>], or
1641 // .param .b<size-in-bits> retval0
1642 unsigned resultsz = DL.getTypeAllocSizeInBits(RetTy);
1643 // Emit ".param .b<size-in-bits> retval0" instead of byte arrays only for
1644 // these three types to match the logic in
1645 // NVPTXAsmPrinter::printReturnValStr and NVPTXTargetLowering::getPrototype.
1646 // Plus, this behavior is consistent with nvcc's.
1647 if (RetTy->isFloatingPointTy() || RetTy->isPointerTy() ||
1648 (RetTy->isIntegerTy() && !RetTy->isIntegerTy(128))) {
1649 // Scalar needs to be at least 32bit wide
1650 if (resultsz < 32)
1651 resultsz = 32;
1652 SDVTList DeclareRetVTs = DAG.getVTList(MVT::Other, MVT::Glue);
1653 SDValue DeclareRetOps[] = { Chain, DAG.getConstant(1, dl, MVT::i32),
1654 DAG.getConstant(resultsz, dl, MVT::i32),
1655 DAG.getConstant(0, dl, MVT::i32), InFlag };
1656 Chain = DAG.getNode(NVPTXISD::DeclareRet, dl, DeclareRetVTs,
1657 DeclareRetOps);
1658 InFlag = Chain.getValue(1);
1659 } else {
1660 retAlignment = getArgumentAlignment(Callee, CS, RetTy, 0, DL);
1661 SDVTList DeclareRetVTs = DAG.getVTList(MVT::Other, MVT::Glue);
1662 SDValue DeclareRetOps[] = { Chain,
1663 DAG.getConstant(retAlignment, dl, MVT::i32),
1664 DAG.getConstant(resultsz / 8, dl, MVT::i32),
1665 DAG.getConstant(0, dl, MVT::i32), InFlag };
1666 Chain = DAG.getNode(NVPTXISD::DeclareRetParam, dl, DeclareRetVTs,
1667 DeclareRetOps);
1668 InFlag = Chain.getValue(1);
1669 }
1670 }
1671
1672 // Both indirect calls and libcalls have nullptr Func. In order to distinguish
1673 // between them we must rely on the call site value which is valid for
1674 // indirect calls but is always null for libcalls.
1675 bool isIndirectCall = !Func && CS;
1676
1677 if (isa<ExternalSymbolSDNode>(Callee)) {
1678 Function* CalleeFunc = nullptr;
1679
1680 // Try to find the callee in the current module.
1681 Callee = DAG.getSymbolFunctionGlobalAddress(Callee, &CalleeFunc);
1682 assert(CalleeFunc != nullptr && "Libcall callee must be set.");
1683
1684 // Set the "libcall callee" attribute to indicate that the function
1685 // must always have a declaration.
1686 CalleeFunc->addFnAttr("nvptx-libcall-callee", "true");
1687 }
1688
1689 if (isIndirectCall) {
1690 // This is indirect function call case : PTX requires a prototype of the
1691 // form
1692 // proto_0 : .callprototype(.param .b32 _) _ (.param .b32 _);
1693 // to be emitted, and the label has to used as the last arg of call
1694 // instruction.
1695 // The prototype is embedded in a string and put as the operand for a
1696 // CallPrototype SDNode which will print out to the value of the string.
1697 SDVTList ProtoVTs = DAG.getVTList(MVT::Other, MVT::Glue);
1698 std::string Proto = getPrototype(DL, RetTy, Args, Outs, retAlignment, CS);
1699 const char *ProtoStr =
1700 nvTM->getManagedStrPool()->getManagedString(Proto.c_str())->c_str();
1701 SDValue ProtoOps[] = {
1702 Chain, DAG.getTargetExternalSymbol(ProtoStr, MVT::i32), InFlag,
1703 };
1704 Chain = DAG.getNode(NVPTXISD::CallPrototype, dl, ProtoVTs, ProtoOps);
1705 InFlag = Chain.getValue(1);
1706 }
1707 // Op to just print "call"
1708 SDVTList PrintCallVTs = DAG.getVTList(MVT::Other, MVT::Glue);
1709 SDValue PrintCallOps[] = {
1710 Chain, DAG.getConstant((Ins.size() == 0) ? 0 : 1, dl, MVT::i32), InFlag
1711 };
1712 // We model convergent calls as separate opcodes.
1713 unsigned Opcode = isIndirectCall ? NVPTXISD::PrintCall : NVPTXISD::PrintCallUni;
1714 if (CLI.IsConvergent)
1715 Opcode = Opcode == NVPTXISD::PrintCallUni ? NVPTXISD::PrintConvergentCallUni
1716 : NVPTXISD::PrintConvergentCall;
1717 Chain = DAG.getNode(Opcode, dl, PrintCallVTs, PrintCallOps);
1718 InFlag = Chain.getValue(1);
1719
1720 // Ops to print out the function name
1721 SDVTList CallVoidVTs = DAG.getVTList(MVT::Other, MVT::Glue);
1722 SDValue CallVoidOps[] = { Chain, Callee, InFlag };
1723 Chain = DAG.getNode(NVPTXISD::CallVoid, dl, CallVoidVTs, CallVoidOps);
1724 InFlag = Chain.getValue(1);
1725
1726 // Ops to print out the param list
1727 SDVTList CallArgBeginVTs = DAG.getVTList(MVT::Other, MVT::Glue);
1728 SDValue CallArgBeginOps[] = { Chain, InFlag };
1729 Chain = DAG.getNode(NVPTXISD::CallArgBegin, dl, CallArgBeginVTs,
1730 CallArgBeginOps);
1731 InFlag = Chain.getValue(1);
1732
1733 for (unsigned i = 0, e = paramCount; i != e; ++i) {
1734 unsigned opcode;
1735 if (i == (e - 1))
1736 opcode = NVPTXISD::LastCallArg;
1737 else
1738 opcode = NVPTXISD::CallArg;
1739 SDVTList CallArgVTs = DAG.getVTList(MVT::Other, MVT::Glue);
1740 SDValue CallArgOps[] = { Chain, DAG.getConstant(1, dl, MVT::i32),
1741 DAG.getConstant(i, dl, MVT::i32), InFlag };
1742 Chain = DAG.getNode(opcode, dl, CallArgVTs, CallArgOps);
1743 InFlag = Chain.getValue(1);
1744 }
1745 SDVTList CallArgEndVTs = DAG.getVTList(MVT::Other, MVT::Glue);
1746 SDValue CallArgEndOps[] = { Chain,
1747 DAG.getConstant(isIndirectCall ? 0 : 1, dl, MVT::i32),
1748 InFlag };
1749 Chain = DAG.getNode(NVPTXISD::CallArgEnd, dl, CallArgEndVTs, CallArgEndOps);
1750 InFlag = Chain.getValue(1);
1751
1752 if (isIndirectCall) {
1753 SDVTList PrototypeVTs = DAG.getVTList(MVT::Other, MVT::Glue);
1754 SDValue PrototypeOps[] = { Chain,
1755 DAG.getConstant(uniqueCallSite, dl, MVT::i32),
1756 InFlag };
1757 Chain = DAG.getNode(NVPTXISD::Prototype, dl, PrototypeVTs, PrototypeOps);
1758 InFlag = Chain.getValue(1);
1759 }
1760
1761 SmallVector<SDValue, 16> ProxyRegOps;
1762 SmallVector<Optional<MVT>, 16> ProxyRegTruncates;
1763
1764 // Generate loads from param memory/moves from registers for result
1765 if (Ins.size() > 0) {
1766 SmallVector<EVT, 16> VTs;
1767 SmallVector<uint64_t, 16> Offsets;
1768 ComputePTXValueVTs(*this, DL, RetTy, VTs, &Offsets, 0);
1769 assert(VTs.size() == Ins.size() && "Bad value decomposition");
1770
1771 unsigned RetAlign = getArgumentAlignment(Callee, CS, RetTy, 0, DL);
1772 auto VectorInfo = VectorizePTXValueVTs(VTs, Offsets, RetAlign);
1773
1774 SmallVector<EVT, 6> LoadVTs;
1775 int VecIdx = -1; // Index of the first element of the vector.
1776
1777 // PTX Interoperability Guide 3.3(A): [Integer] Values shorter than
1778 // 32-bits are sign extended or zero extended, depending on whether
1779 // they are signed or unsigned types.
1780 bool ExtendIntegerRetVal =
1781 RetTy->isIntegerTy() && DL.getTypeAllocSizeInBits(RetTy) < 32;
1782
1783 for (unsigned i = 0, e = VTs.size(); i != e; ++i) {
1784 bool needTruncate = false;
1785 EVT TheLoadType = VTs[i];
1786 EVT EltType = Ins[i].VT;
1787 unsigned EltAlign = GreatestCommonDivisor64(RetAlign, Offsets[i]);
1788 if (ExtendIntegerRetVal) {
1789 TheLoadType = MVT::i32;
1790 EltType = MVT::i32;
1791 needTruncate = true;
1792 } else if (TheLoadType.getSizeInBits() < 16) {
1793 if (VTs[i].isInteger())
1794 needTruncate = true;
1795 EltType = MVT::i16;
1796 }
1797
1798 // Record index of the very first element of the vector.
1799 if (VectorInfo[i] & PVF_FIRST) {
1800 assert(VecIdx == -1 && LoadVTs.empty() && "Orphaned operand list.");
1801 VecIdx = i;
1802 }
1803
1804 LoadVTs.push_back(EltType);
1805
1806 if (VectorInfo[i] & PVF_LAST) {
1807 unsigned NumElts = LoadVTs.size();
1808 LoadVTs.push_back(MVT::Other);
1809 LoadVTs.push_back(MVT::Glue);
1810 NVPTXISD::NodeType Op;
1811 switch (NumElts) {
1812 case 1:
1813 Op = NVPTXISD::LoadParam;
1814 break;
1815 case 2:
1816 Op = NVPTXISD::LoadParamV2;
1817 break;
1818 case 4:
1819 Op = NVPTXISD::LoadParamV4;
1820 break;
1821 default:
1822 llvm_unreachable("Invalid vector info.");
1823 }
1824
1825 SDValue LoadOperands[] = {
1826 Chain, DAG.getConstant(1, dl, MVT::i32),
1827 DAG.getConstant(Offsets[VecIdx], dl, MVT::i32), InFlag};
1828 SDValue RetVal = DAG.getMemIntrinsicNode(
1829 Op, dl, DAG.getVTList(LoadVTs), LoadOperands, TheLoadType,
1830 MachinePointerInfo(), EltAlign,
1831 MachineMemOperand::MOLoad);
1832
1833 for (unsigned j = 0; j < NumElts; ++j) {
1834 ProxyRegOps.push_back(RetVal.getValue(j));
1835
1836 if (needTruncate)
1837 ProxyRegTruncates.push_back(Optional<MVT>(Ins[VecIdx + j].VT));
1838 else
1839 ProxyRegTruncates.push_back(Optional<MVT>());
1840 }
1841
1842 Chain = RetVal.getValue(NumElts);
1843 InFlag = RetVal.getValue(NumElts + 1);
1844
1845 // Cleanup
1846 VecIdx = -1;
1847 LoadVTs.clear();
1848 }
1849 }
1850 }
1851
1852 Chain = DAG.getCALLSEQ_END(Chain,
1853 DAG.getIntPtrConstant(uniqueCallSite, dl, true),
1854 DAG.getIntPtrConstant(uniqueCallSite + 1, dl,
1855 true),
1856 InFlag, dl);
1857 InFlag = Chain.getValue(1);
1858 uniqueCallSite++;
1859
1860 // Append ProxyReg instructions to the chain to make sure that `callseq_end`
1861 // will not get lost. Otherwise, during libcalls expansion, the nodes can become
1862 // dangling.
1863 for (unsigned i = 0; i < ProxyRegOps.size(); ++i) {
1864 SDValue Ret = DAG.getNode(
1865 NVPTXISD::ProxyReg, dl,
1866 DAG.getVTList(ProxyRegOps[i].getSimpleValueType(), MVT::Other, MVT::Glue),
1867 { Chain, ProxyRegOps[i], InFlag }
1868 );
1869
1870 Chain = Ret.getValue(1);
1871 InFlag = Ret.getValue(2);
1872
1873 if (ProxyRegTruncates[i].hasValue()) {
1874 Ret = DAG.getNode(ISD::TRUNCATE, dl, ProxyRegTruncates[i].getValue(), Ret);
1875 }
1876
1877 InVals.push_back(Ret);
1878 }
1879
1880 // set isTailCall to false for now, until we figure out how to express
1881 // tail call optimization in PTX
1882 isTailCall = false;
1883 return Chain;
1884 }
1885
1886 // By default CONCAT_VECTORS is lowered by ExpandVectorBuildThroughStack()
1887 // (see LegalizeDAG.cpp). This is slow and uses local memory.
1888 // We use extract/insert/build vector just as what LegalizeOp() does in llvm 2.5
1889 SDValue
LowerCONCAT_VECTORS(SDValue Op,SelectionDAG & DAG) const1890 NVPTXTargetLowering::LowerCONCAT_VECTORS(SDValue Op, SelectionDAG &DAG) const {
1891 SDNode *Node = Op.getNode();
1892 SDLoc dl(Node);
1893 SmallVector<SDValue, 8> Ops;
1894 unsigned NumOperands = Node->getNumOperands();
1895 for (unsigned i = 0; i < NumOperands; ++i) {
1896 SDValue SubOp = Node->getOperand(i);
1897 EVT VVT = SubOp.getNode()->getValueType(0);
1898 EVT EltVT = VVT.getVectorElementType();
1899 unsigned NumSubElem = VVT.getVectorNumElements();
1900 for (unsigned j = 0; j < NumSubElem; ++j) {
1901 Ops.push_back(DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, EltVT, SubOp,
1902 DAG.getIntPtrConstant(j, dl)));
1903 }
1904 }
1905 return DAG.getBuildVector(Node->getValueType(0), dl, Ops);
1906 }
1907
1908 // We can init constant f16x2 with a single .b32 move. Normally it
1909 // would get lowered as two constant loads and vector-packing move.
1910 // mov.b16 %h1, 0x4000;
1911 // mov.b16 %h2, 0x3C00;
1912 // mov.b32 %hh2, {%h2, %h1};
1913 // Instead we want just a constant move:
1914 // mov.b32 %hh2, 0x40003C00
1915 //
1916 // This results in better SASS code with CUDA 7.x. Ptxas in CUDA 8.0
1917 // generates good SASS in both cases.
LowerBUILD_VECTOR(SDValue Op,SelectionDAG & DAG) const1918 SDValue NVPTXTargetLowering::LowerBUILD_VECTOR(SDValue Op,
1919 SelectionDAG &DAG) const {
1920 //return Op;
1921 if (!(Op->getValueType(0) == MVT::v2f16 &&
1922 isa<ConstantFPSDNode>(Op->getOperand(0)) &&
1923 isa<ConstantFPSDNode>(Op->getOperand(1))))
1924 return Op;
1925
1926 APInt E0 =
1927 cast<ConstantFPSDNode>(Op->getOperand(0))->getValueAPF().bitcastToAPInt();
1928 APInt E1 =
1929 cast<ConstantFPSDNode>(Op->getOperand(1))->getValueAPF().bitcastToAPInt();
1930 SDValue Const =
1931 DAG.getConstant(E1.zext(32).shl(16) | E0.zext(32), SDLoc(Op), MVT::i32);
1932 return DAG.getNode(ISD::BITCAST, SDLoc(Op), MVT::v2f16, Const);
1933 }
1934
LowerEXTRACT_VECTOR_ELT(SDValue Op,SelectionDAG & DAG) const1935 SDValue NVPTXTargetLowering::LowerEXTRACT_VECTOR_ELT(SDValue Op,
1936 SelectionDAG &DAG) const {
1937 SDValue Index = Op->getOperand(1);
1938 // Constant index will be matched by tablegen.
1939 if (isa<ConstantSDNode>(Index.getNode()))
1940 return Op;
1941
1942 // Extract individual elements and select one of them.
1943 SDValue Vector = Op->getOperand(0);
1944 EVT VectorVT = Vector.getValueType();
1945 assert(VectorVT == MVT::v2f16 && "Unexpected vector type.");
1946 EVT EltVT = VectorVT.getVectorElementType();
1947
1948 SDLoc dl(Op.getNode());
1949 SDValue E0 = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, EltVT, Vector,
1950 DAG.getIntPtrConstant(0, dl));
1951 SDValue E1 = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, EltVT, Vector,
1952 DAG.getIntPtrConstant(1, dl));
1953 return DAG.getSelectCC(dl, Index, DAG.getIntPtrConstant(0, dl), E0, E1,
1954 ISD::CondCode::SETEQ);
1955 }
1956
1957 /// LowerShiftRightParts - Lower SRL_PARTS, SRA_PARTS, which
1958 /// 1) returns two i32 values and take a 2 x i32 value to shift plus a shift
1959 /// amount, or
1960 /// 2) returns two i64 values and take a 2 x i64 value to shift plus a shift
1961 /// amount.
LowerShiftRightParts(SDValue Op,SelectionDAG & DAG) const1962 SDValue NVPTXTargetLowering::LowerShiftRightParts(SDValue Op,
1963 SelectionDAG &DAG) const {
1964 assert(Op.getNumOperands() == 3 && "Not a double-shift!");
1965 assert(Op.getOpcode() == ISD::SRA_PARTS || Op.getOpcode() == ISD::SRL_PARTS);
1966
1967 EVT VT = Op.getValueType();
1968 unsigned VTBits = VT.getSizeInBits();
1969 SDLoc dl(Op);
1970 SDValue ShOpLo = Op.getOperand(0);
1971 SDValue ShOpHi = Op.getOperand(1);
1972 SDValue ShAmt = Op.getOperand(2);
1973 unsigned Opc = (Op.getOpcode() == ISD::SRA_PARTS) ? ISD::SRA : ISD::SRL;
1974
1975 if (VTBits == 32 && STI.getSmVersion() >= 35) {
1976 // For 32bit and sm35, we can use the funnel shift 'shf' instruction.
1977 // {dHi, dLo} = {aHi, aLo} >> Amt
1978 // dHi = aHi >> Amt
1979 // dLo = shf.r.clamp aLo, aHi, Amt
1980
1981 SDValue Hi = DAG.getNode(Opc, dl, VT, ShOpHi, ShAmt);
1982 SDValue Lo = DAG.getNode(NVPTXISD::FUN_SHFR_CLAMP, dl, VT, ShOpLo, ShOpHi,
1983 ShAmt);
1984
1985 SDValue Ops[2] = { Lo, Hi };
1986 return DAG.getMergeValues(Ops, dl);
1987 }
1988 else {
1989 // {dHi, dLo} = {aHi, aLo} >> Amt
1990 // - if (Amt>=size) then
1991 // dLo = aHi >> (Amt-size)
1992 // dHi = aHi >> Amt (this is either all 0 or all 1)
1993 // else
1994 // dLo = (aLo >>logic Amt) | (aHi << (size-Amt))
1995 // dHi = aHi >> Amt
1996
1997 SDValue RevShAmt = DAG.getNode(ISD::SUB, dl, MVT::i32,
1998 DAG.getConstant(VTBits, dl, MVT::i32),
1999 ShAmt);
2000 SDValue Tmp1 = DAG.getNode(ISD::SRL, dl, VT, ShOpLo, ShAmt);
2001 SDValue ExtraShAmt = DAG.getNode(ISD::SUB, dl, MVT::i32, ShAmt,
2002 DAG.getConstant(VTBits, dl, MVT::i32));
2003 SDValue Tmp2 = DAG.getNode(ISD::SHL, dl, VT, ShOpHi, RevShAmt);
2004 SDValue FalseVal = DAG.getNode(ISD::OR, dl, VT, Tmp1, Tmp2);
2005 SDValue TrueVal = DAG.getNode(Opc, dl, VT, ShOpHi, ExtraShAmt);
2006
2007 SDValue Cmp = DAG.getSetCC(dl, MVT::i1, ShAmt,
2008 DAG.getConstant(VTBits, dl, MVT::i32),
2009 ISD::SETGE);
2010 SDValue Hi = DAG.getNode(Opc, dl, VT, ShOpHi, ShAmt);
2011 SDValue Lo = DAG.getNode(ISD::SELECT, dl, VT, Cmp, TrueVal, FalseVal);
2012
2013 SDValue Ops[2] = { Lo, Hi };
2014 return DAG.getMergeValues(Ops, dl);
2015 }
2016 }
2017
2018 /// LowerShiftLeftParts - Lower SHL_PARTS, which
2019 /// 1) returns two i32 values and take a 2 x i32 value to shift plus a shift
2020 /// amount, or
2021 /// 2) returns two i64 values and take a 2 x i64 value to shift plus a shift
2022 /// amount.
LowerShiftLeftParts(SDValue Op,SelectionDAG & DAG) const2023 SDValue NVPTXTargetLowering::LowerShiftLeftParts(SDValue Op,
2024 SelectionDAG &DAG) const {
2025 assert(Op.getNumOperands() == 3 && "Not a double-shift!");
2026 assert(Op.getOpcode() == ISD::SHL_PARTS);
2027
2028 EVT VT = Op.getValueType();
2029 unsigned VTBits = VT.getSizeInBits();
2030 SDLoc dl(Op);
2031 SDValue ShOpLo = Op.getOperand(0);
2032 SDValue ShOpHi = Op.getOperand(1);
2033 SDValue ShAmt = Op.getOperand(2);
2034
2035 if (VTBits == 32 && STI.getSmVersion() >= 35) {
2036 // For 32bit and sm35, we can use the funnel shift 'shf' instruction.
2037 // {dHi, dLo} = {aHi, aLo} << Amt
2038 // dHi = shf.l.clamp aLo, aHi, Amt
2039 // dLo = aLo << Amt
2040
2041 SDValue Hi = DAG.getNode(NVPTXISD::FUN_SHFL_CLAMP, dl, VT, ShOpLo, ShOpHi,
2042 ShAmt);
2043 SDValue Lo = DAG.getNode(ISD::SHL, dl, VT, ShOpLo, ShAmt);
2044
2045 SDValue Ops[2] = { Lo, Hi };
2046 return DAG.getMergeValues(Ops, dl);
2047 }
2048 else {
2049 // {dHi, dLo} = {aHi, aLo} << Amt
2050 // - if (Amt>=size) then
2051 // dLo = aLo << Amt (all 0)
2052 // dLo = aLo << (Amt-size)
2053 // else
2054 // dLo = aLo << Amt
2055 // dHi = (aHi << Amt) | (aLo >> (size-Amt))
2056
2057 SDValue RevShAmt = DAG.getNode(ISD::SUB, dl, MVT::i32,
2058 DAG.getConstant(VTBits, dl, MVT::i32),
2059 ShAmt);
2060 SDValue Tmp1 = DAG.getNode(ISD::SHL, dl, VT, ShOpHi, ShAmt);
2061 SDValue ExtraShAmt = DAG.getNode(ISD::SUB, dl, MVT::i32, ShAmt,
2062 DAG.getConstant(VTBits, dl, MVT::i32));
2063 SDValue Tmp2 = DAG.getNode(ISD::SRL, dl, VT, ShOpLo, RevShAmt);
2064 SDValue FalseVal = DAG.getNode(ISD::OR, dl, VT, Tmp1, Tmp2);
2065 SDValue TrueVal = DAG.getNode(ISD::SHL, dl, VT, ShOpLo, ExtraShAmt);
2066
2067 SDValue Cmp = DAG.getSetCC(dl, MVT::i1, ShAmt,
2068 DAG.getConstant(VTBits, dl, MVT::i32),
2069 ISD::SETGE);
2070 SDValue Lo = DAG.getNode(ISD::SHL, dl, VT, ShOpLo, ShAmt);
2071 SDValue Hi = DAG.getNode(ISD::SELECT, dl, VT, Cmp, TrueVal, FalseVal);
2072
2073 SDValue Ops[2] = { Lo, Hi };
2074 return DAG.getMergeValues(Ops, dl);
2075 }
2076 }
2077
LowerFROUND(SDValue Op,SelectionDAG & DAG) const2078 SDValue NVPTXTargetLowering::LowerFROUND(SDValue Op, SelectionDAG &DAG) const {
2079 EVT VT = Op.getValueType();
2080
2081 if (VT == MVT::f32)
2082 return LowerFROUND32(Op, DAG);
2083
2084 if (VT == MVT::f64)
2085 return LowerFROUND64(Op, DAG);
2086
2087 llvm_unreachable("unhandled type");
2088 }
2089
2090 // This is the the rounding method used in CUDA libdevice in C like code:
2091 // float roundf(float A)
2092 // {
2093 // float RoundedA = (float) (int) ( A > 0 ? (A + 0.5f) : (A - 0.5f));
2094 // RoundedA = abs(A) > 0x1.0p23 ? A : RoundedA;
2095 // return abs(A) < 0.5 ? (float)(int)A : RoundedA;
2096 // }
LowerFROUND32(SDValue Op,SelectionDAG & DAG) const2097 SDValue NVPTXTargetLowering::LowerFROUND32(SDValue Op,
2098 SelectionDAG &DAG) const {
2099 SDLoc SL(Op);
2100 SDValue A = Op.getOperand(0);
2101 EVT VT = Op.getValueType();
2102
2103 SDValue AbsA = DAG.getNode(ISD::FABS, SL, VT, A);
2104
2105 // RoundedA = (float) (int) ( A > 0 ? (A + 0.5f) : (A - 0.5f))
2106 SDValue Bitcast = DAG.getNode(ISD::BITCAST, SL, MVT::i32, A);
2107 const int SignBitMask = 0x80000000;
2108 SDValue Sign = DAG.getNode(ISD::AND, SL, MVT::i32, Bitcast,
2109 DAG.getConstant(SignBitMask, SL, MVT::i32));
2110 const int PointFiveInBits = 0x3F000000;
2111 SDValue PointFiveWithSignRaw =
2112 DAG.getNode(ISD::OR, SL, MVT::i32, Sign,
2113 DAG.getConstant(PointFiveInBits, SL, MVT::i32));
2114 SDValue PointFiveWithSign =
2115 DAG.getNode(ISD::BITCAST, SL, VT, PointFiveWithSignRaw);
2116 SDValue AdjustedA = DAG.getNode(ISD::FADD, SL, VT, A, PointFiveWithSign);
2117 SDValue RoundedA = DAG.getNode(ISD::FTRUNC, SL, VT, AdjustedA);
2118
2119 // RoundedA = abs(A) > 0x1.0p23 ? A : RoundedA;
2120 EVT SetCCVT = getSetCCResultType(DAG.getDataLayout(), *DAG.getContext(), VT);
2121 SDValue IsLarge =
2122 DAG.getSetCC(SL, SetCCVT, AbsA, DAG.getConstantFP(pow(2.0, 23.0), SL, VT),
2123 ISD::SETOGT);
2124 RoundedA = DAG.getNode(ISD::SELECT, SL, VT, IsLarge, A, RoundedA);
2125
2126 // return abs(A) < 0.5 ? (float)(int)A : RoundedA;
2127 SDValue IsSmall =DAG.getSetCC(SL, SetCCVT, AbsA,
2128 DAG.getConstantFP(0.5, SL, VT), ISD::SETOLT);
2129 SDValue RoundedAForSmallA = DAG.getNode(ISD::FTRUNC, SL, VT, A);
2130 return DAG.getNode(ISD::SELECT, SL, VT, IsSmall, RoundedAForSmallA, RoundedA);
2131 }
2132
2133 // The implementation of round(double) is similar to that of round(float) in
2134 // that they both separate the value range into three regions and use a method
2135 // specific to the region to round the values. However, round(double) first
2136 // calculates the round of the absolute value and then adds the sign back while
2137 // round(float) directly rounds the value with sign.
LowerFROUND64(SDValue Op,SelectionDAG & DAG) const2138 SDValue NVPTXTargetLowering::LowerFROUND64(SDValue Op,
2139 SelectionDAG &DAG) const {
2140 SDLoc SL(Op);
2141 SDValue A = Op.getOperand(0);
2142 EVT VT = Op.getValueType();
2143
2144 SDValue AbsA = DAG.getNode(ISD::FABS, SL, VT, A);
2145
2146 // double RoundedA = (double) (int) (abs(A) + 0.5f);
2147 SDValue AdjustedA = DAG.getNode(ISD::FADD, SL, VT, AbsA,
2148 DAG.getConstantFP(0.5, SL, VT));
2149 SDValue RoundedA = DAG.getNode(ISD::FTRUNC, SL, VT, AdjustedA);
2150
2151 // RoundedA = abs(A) < 0.5 ? (double)0 : RoundedA;
2152 EVT SetCCVT = getSetCCResultType(DAG.getDataLayout(), *DAG.getContext(), VT);
2153 SDValue IsSmall =DAG.getSetCC(SL, SetCCVT, AbsA,
2154 DAG.getConstantFP(0.5, SL, VT), ISD::SETOLT);
2155 RoundedA = DAG.getNode(ISD::SELECT, SL, VT, IsSmall,
2156 DAG.getConstantFP(0, SL, VT),
2157 RoundedA);
2158
2159 // Add sign to rounded_A
2160 RoundedA = DAG.getNode(ISD::FCOPYSIGN, SL, VT, RoundedA, A);
2161 DAG.getNode(ISD::FTRUNC, SL, VT, A);
2162
2163 // RoundedA = abs(A) > 0x1.0p52 ? A : RoundedA;
2164 SDValue IsLarge =
2165 DAG.getSetCC(SL, SetCCVT, AbsA, DAG.getConstantFP(pow(2.0, 52.0), SL, VT),
2166 ISD::SETOGT);
2167 return DAG.getNode(ISD::SELECT, SL, VT, IsLarge, A, RoundedA);
2168 }
2169
2170
2171
2172 SDValue
LowerOperation(SDValue Op,SelectionDAG & DAG) const2173 NVPTXTargetLowering::LowerOperation(SDValue Op, SelectionDAG &DAG) const {
2174 switch (Op.getOpcode()) {
2175 case ISD::RETURNADDR:
2176 return SDValue();
2177 case ISD::FRAMEADDR:
2178 return SDValue();
2179 case ISD::GlobalAddress:
2180 return LowerGlobalAddress(Op, DAG);
2181 case ISD::INTRINSIC_W_CHAIN:
2182 return Op;
2183 case ISD::BUILD_VECTOR:
2184 return LowerBUILD_VECTOR(Op, DAG);
2185 case ISD::EXTRACT_SUBVECTOR:
2186 return Op;
2187 case ISD::EXTRACT_VECTOR_ELT:
2188 return LowerEXTRACT_VECTOR_ELT(Op, DAG);
2189 case ISD::CONCAT_VECTORS:
2190 return LowerCONCAT_VECTORS(Op, DAG);
2191 case ISD::STORE:
2192 return LowerSTORE(Op, DAG);
2193 case ISD::LOAD:
2194 return LowerLOAD(Op, DAG);
2195 case ISD::SHL_PARTS:
2196 return LowerShiftLeftParts(Op, DAG);
2197 case ISD::SRA_PARTS:
2198 case ISD::SRL_PARTS:
2199 return LowerShiftRightParts(Op, DAG);
2200 case ISD::SELECT:
2201 return LowerSelect(Op, DAG);
2202 case ISD::FROUND:
2203 return LowerFROUND(Op, DAG);
2204 default:
2205 llvm_unreachable("Custom lowering not defined for operation");
2206 }
2207 }
2208
LowerSelect(SDValue Op,SelectionDAG & DAG) const2209 SDValue NVPTXTargetLowering::LowerSelect(SDValue Op, SelectionDAG &DAG) const {
2210 SDValue Op0 = Op->getOperand(0);
2211 SDValue Op1 = Op->getOperand(1);
2212 SDValue Op2 = Op->getOperand(2);
2213 SDLoc DL(Op.getNode());
2214
2215 assert(Op.getValueType() == MVT::i1 && "Custom lowering enabled only for i1");
2216
2217 Op1 = DAG.getNode(ISD::ANY_EXTEND, DL, MVT::i32, Op1);
2218 Op2 = DAG.getNode(ISD::ANY_EXTEND, DL, MVT::i32, Op2);
2219 SDValue Select = DAG.getNode(ISD::SELECT, DL, MVT::i32, Op0, Op1, Op2);
2220 SDValue Trunc = DAG.getNode(ISD::TRUNCATE, DL, MVT::i1, Select);
2221
2222 return Trunc;
2223 }
2224
LowerLOAD(SDValue Op,SelectionDAG & DAG) const2225 SDValue NVPTXTargetLowering::LowerLOAD(SDValue Op, SelectionDAG &DAG) const {
2226 if (Op.getValueType() == MVT::i1)
2227 return LowerLOADi1(Op, DAG);
2228
2229 // v2f16 is legal, so we can't rely on legalizer to handle unaligned
2230 // loads and have to handle it here.
2231 if (Op.getValueType() == MVT::v2f16) {
2232 LoadSDNode *Load = cast<LoadSDNode>(Op);
2233 EVT MemVT = Load->getMemoryVT();
2234 if (!allowsMemoryAccessForAlignment(*DAG.getContext(), DAG.getDataLayout(),
2235 MemVT, *Load->getMemOperand())) {
2236 SDValue Ops[2];
2237 std::tie(Ops[0], Ops[1]) = expandUnalignedLoad(Load, DAG);
2238 return DAG.getMergeValues(Ops, SDLoc(Op));
2239 }
2240 }
2241
2242 return SDValue();
2243 }
2244
2245 // v = ld i1* addr
2246 // =>
2247 // v1 = ld i8* addr (-> i16)
2248 // v = trunc i16 to i1
LowerLOADi1(SDValue Op,SelectionDAG & DAG) const2249 SDValue NVPTXTargetLowering::LowerLOADi1(SDValue Op, SelectionDAG &DAG) const {
2250 SDNode *Node = Op.getNode();
2251 LoadSDNode *LD = cast<LoadSDNode>(Node);
2252 SDLoc dl(Node);
2253 assert(LD->getExtensionType() == ISD::NON_EXTLOAD);
2254 assert(Node->getValueType(0) == MVT::i1 &&
2255 "Custom lowering for i1 load only");
2256 SDValue newLD = DAG.getLoad(MVT::i16, dl, LD->getChain(), LD->getBasePtr(),
2257 LD->getPointerInfo(), LD->getAlignment(),
2258 LD->getMemOperand()->getFlags());
2259 SDValue result = DAG.getNode(ISD::TRUNCATE, dl, MVT::i1, newLD);
2260 // The legalizer (the caller) is expecting two values from the legalized
2261 // load, so we build a MergeValues node for it. See ExpandUnalignedLoad()
2262 // in LegalizeDAG.cpp which also uses MergeValues.
2263 SDValue Ops[] = { result, LD->getChain() };
2264 return DAG.getMergeValues(Ops, dl);
2265 }
2266
LowerSTORE(SDValue Op,SelectionDAG & DAG) const2267 SDValue NVPTXTargetLowering::LowerSTORE(SDValue Op, SelectionDAG &DAG) const {
2268 StoreSDNode *Store = cast<StoreSDNode>(Op);
2269 EVT VT = Store->getMemoryVT();
2270
2271 if (VT == MVT::i1)
2272 return LowerSTOREi1(Op, DAG);
2273
2274 // v2f16 is legal, so we can't rely on legalizer to handle unaligned
2275 // stores and have to handle it here.
2276 if (VT == MVT::v2f16 &&
2277 !allowsMemoryAccessForAlignment(*DAG.getContext(), DAG.getDataLayout(),
2278 VT, *Store->getMemOperand()))
2279 return expandUnalignedStore(Store, DAG);
2280
2281 if (VT.isVector())
2282 return LowerSTOREVector(Op, DAG);
2283
2284 return SDValue();
2285 }
2286
2287 SDValue
LowerSTOREVector(SDValue Op,SelectionDAG & DAG) const2288 NVPTXTargetLowering::LowerSTOREVector(SDValue Op, SelectionDAG &DAG) const {
2289 SDNode *N = Op.getNode();
2290 SDValue Val = N->getOperand(1);
2291 SDLoc DL(N);
2292 EVT ValVT = Val.getValueType();
2293
2294 if (ValVT.isVector()) {
2295 // We only handle "native" vector sizes for now, e.g. <4 x double> is not
2296 // legal. We can (and should) split that into 2 stores of <2 x double> here
2297 // but I'm leaving that as a TODO for now.
2298 if (!ValVT.isSimple())
2299 return SDValue();
2300 switch (ValVT.getSimpleVT().SimpleTy) {
2301 default:
2302 return SDValue();
2303 case MVT::v2i8:
2304 case MVT::v2i16:
2305 case MVT::v2i32:
2306 case MVT::v2i64:
2307 case MVT::v2f16:
2308 case MVT::v2f32:
2309 case MVT::v2f64:
2310 case MVT::v4i8:
2311 case MVT::v4i16:
2312 case MVT::v4i32:
2313 case MVT::v4f16:
2314 case MVT::v4f32:
2315 case MVT::v8f16: // <4 x f16x2>
2316 // This is a "native" vector type
2317 break;
2318 }
2319
2320 MemSDNode *MemSD = cast<MemSDNode>(N);
2321 const DataLayout &TD = DAG.getDataLayout();
2322
2323 unsigned Align = MemSD->getAlignment();
2324 unsigned PrefAlign =
2325 TD.getPrefTypeAlignment(ValVT.getTypeForEVT(*DAG.getContext()));
2326 if (Align < PrefAlign) {
2327 // This store is not sufficiently aligned, so bail out and let this vector
2328 // store be scalarized. Note that we may still be able to emit smaller
2329 // vector stores. For example, if we are storing a <4 x float> with an
2330 // alignment of 8, this check will fail but the legalizer will try again
2331 // with 2 x <2 x float>, which will succeed with an alignment of 8.
2332 return SDValue();
2333 }
2334
2335 unsigned Opcode = 0;
2336 EVT EltVT = ValVT.getVectorElementType();
2337 unsigned NumElts = ValVT.getVectorNumElements();
2338
2339 // Since StoreV2 is a target node, we cannot rely on DAG type legalization.
2340 // Therefore, we must ensure the type is legal. For i1 and i8, we set the
2341 // stored type to i16 and propagate the "real" type as the memory type.
2342 bool NeedExt = false;
2343 if (EltVT.getSizeInBits() < 16)
2344 NeedExt = true;
2345
2346 bool StoreF16x2 = false;
2347 switch (NumElts) {
2348 default:
2349 return SDValue();
2350 case 2:
2351 Opcode = NVPTXISD::StoreV2;
2352 break;
2353 case 4:
2354 Opcode = NVPTXISD::StoreV4;
2355 break;
2356 case 8:
2357 // v8f16 is a special case. PTX doesn't have st.v8.f16
2358 // instruction. Instead, we split the vector into v2f16 chunks and
2359 // store them with st.v4.b32.
2360 assert(EltVT == MVT::f16 && "Wrong type for the vector.");
2361 Opcode = NVPTXISD::StoreV4;
2362 StoreF16x2 = true;
2363 break;
2364 }
2365
2366 SmallVector<SDValue, 8> Ops;
2367
2368 // First is the chain
2369 Ops.push_back(N->getOperand(0));
2370
2371 if (StoreF16x2) {
2372 // Combine f16,f16 -> v2f16
2373 NumElts /= 2;
2374 for (unsigned i = 0; i < NumElts; ++i) {
2375 SDValue E0 = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, MVT::f16, Val,
2376 DAG.getIntPtrConstant(i * 2, DL));
2377 SDValue E1 = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, MVT::f16, Val,
2378 DAG.getIntPtrConstant(i * 2 + 1, DL));
2379 SDValue V2 = DAG.getNode(ISD::BUILD_VECTOR, DL, MVT::v2f16, E0, E1);
2380 Ops.push_back(V2);
2381 }
2382 } else {
2383 // Then the split values
2384 for (unsigned i = 0; i < NumElts; ++i) {
2385 SDValue ExtVal = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, EltVT, Val,
2386 DAG.getIntPtrConstant(i, DL));
2387 if (NeedExt)
2388 ExtVal = DAG.getNode(ISD::ANY_EXTEND, DL, MVT::i16, ExtVal);
2389 Ops.push_back(ExtVal);
2390 }
2391 }
2392
2393 // Then any remaining arguments
2394 Ops.append(N->op_begin() + 2, N->op_end());
2395
2396 SDValue NewSt =
2397 DAG.getMemIntrinsicNode(Opcode, DL, DAG.getVTList(MVT::Other), Ops,
2398 MemSD->getMemoryVT(), MemSD->getMemOperand());
2399
2400 // return DCI.CombineTo(N, NewSt, true);
2401 return NewSt;
2402 }
2403
2404 return SDValue();
2405 }
2406
2407 // st i1 v, addr
2408 // =>
2409 // v1 = zxt v to i16
2410 // st.u8 i16, addr
LowerSTOREi1(SDValue Op,SelectionDAG & DAG) const2411 SDValue NVPTXTargetLowering::LowerSTOREi1(SDValue Op, SelectionDAG &DAG) const {
2412 SDNode *Node = Op.getNode();
2413 SDLoc dl(Node);
2414 StoreSDNode *ST = cast<StoreSDNode>(Node);
2415 SDValue Tmp1 = ST->getChain();
2416 SDValue Tmp2 = ST->getBasePtr();
2417 SDValue Tmp3 = ST->getValue();
2418 assert(Tmp3.getValueType() == MVT::i1 && "Custom lowering for i1 store only");
2419 Tmp3 = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i16, Tmp3);
2420 SDValue Result =
2421 DAG.getTruncStore(Tmp1, dl, Tmp3, Tmp2, ST->getPointerInfo(), MVT::i8,
2422 ST->getAlignment(), ST->getMemOperand()->getFlags());
2423 return Result;
2424 }
2425
2426 SDValue
getParamSymbol(SelectionDAG & DAG,int idx,EVT v) const2427 NVPTXTargetLowering::getParamSymbol(SelectionDAG &DAG, int idx, EVT v) const {
2428 std::string ParamSym;
2429 raw_string_ostream ParamStr(ParamSym);
2430
2431 ParamStr << DAG.getMachineFunction().getName() << "_param_" << idx;
2432 ParamStr.flush();
2433
2434 std::string *SavedStr =
2435 nvTM->getManagedStrPool()->getManagedString(ParamSym.c_str());
2436 return DAG.getTargetExternalSymbol(SavedStr->c_str(), v);
2437 }
2438
2439 // Check to see if the kernel argument is image*_t or sampler_t
2440
isImageOrSamplerVal(const Value * arg,const Module * context)2441 static bool isImageOrSamplerVal(const Value *arg, const Module *context) {
2442 static const char *const specialTypes[] = { "struct._image2d_t",
2443 "struct._image3d_t",
2444 "struct._sampler_t" };
2445
2446 Type *Ty = arg->getType();
2447 auto *PTy = dyn_cast<PointerType>(Ty);
2448
2449 if (!PTy)
2450 return false;
2451
2452 if (!context)
2453 return false;
2454
2455 auto *STy = dyn_cast<StructType>(PTy->getElementType());
2456 if (!STy || STy->isLiteral())
2457 return false;
2458
2459 return std::find(std::begin(specialTypes), std::end(specialTypes),
2460 STy->getName()) != std::end(specialTypes);
2461 }
2462
LowerFormalArguments(SDValue Chain,CallingConv::ID CallConv,bool isVarArg,const SmallVectorImpl<ISD::InputArg> & Ins,const SDLoc & dl,SelectionDAG & DAG,SmallVectorImpl<SDValue> & InVals) const2463 SDValue NVPTXTargetLowering::LowerFormalArguments(
2464 SDValue Chain, CallingConv::ID CallConv, bool isVarArg,
2465 const SmallVectorImpl<ISD::InputArg> &Ins, const SDLoc &dl,
2466 SelectionDAG &DAG, SmallVectorImpl<SDValue> &InVals) const {
2467 MachineFunction &MF = DAG.getMachineFunction();
2468 const DataLayout &DL = DAG.getDataLayout();
2469 auto PtrVT = getPointerTy(DAG.getDataLayout());
2470
2471 const Function *F = &MF.getFunction();
2472 const AttributeList &PAL = F->getAttributes();
2473 const TargetLowering *TLI = STI.getTargetLowering();
2474
2475 SDValue Root = DAG.getRoot();
2476 std::vector<SDValue> OutChains;
2477
2478 bool isABI = (STI.getSmVersion() >= 20);
2479 assert(isABI && "Non-ABI compilation is not supported");
2480 if (!isABI)
2481 return Chain;
2482
2483 std::vector<Type *> argTypes;
2484 std::vector<const Argument *> theArgs;
2485 for (const Argument &I : F->args()) {
2486 theArgs.push_back(&I);
2487 argTypes.push_back(I.getType());
2488 }
2489 // argTypes.size() (or theArgs.size()) and Ins.size() need not match.
2490 // Ins.size() will be larger
2491 // * if there is an aggregate argument with multiple fields (each field
2492 // showing up separately in Ins)
2493 // * if there is a vector argument with more than typical vector-length
2494 // elements (generally if more than 4) where each vector element is
2495 // individually present in Ins.
2496 // So a different index should be used for indexing into Ins.
2497 // See similar issue in LowerCall.
2498 unsigned InsIdx = 0;
2499
2500 int idx = 0;
2501 for (unsigned i = 0, e = theArgs.size(); i != e; ++i, ++idx, ++InsIdx) {
2502 Type *Ty = argTypes[i];
2503
2504 // If the kernel argument is image*_t or sampler_t, convert it to
2505 // a i32 constant holding the parameter position. This can later
2506 // matched in the AsmPrinter to output the correct mangled name.
2507 if (isImageOrSamplerVal(
2508 theArgs[i],
2509 (theArgs[i]->getParent() ? theArgs[i]->getParent()->getParent()
2510 : nullptr))) {
2511 assert(isKernelFunction(*F) &&
2512 "Only kernels can have image/sampler params");
2513 InVals.push_back(DAG.getConstant(i + 1, dl, MVT::i32));
2514 continue;
2515 }
2516
2517 if (theArgs[i]->use_empty()) {
2518 // argument is dead
2519 if (Ty->isAggregateType() || Ty->isIntegerTy(128)) {
2520 SmallVector<EVT, 16> vtparts;
2521
2522 ComputePTXValueVTs(*this, DAG.getDataLayout(), Ty, vtparts);
2523 assert(vtparts.size() > 0 && "empty aggregate type not expected");
2524 for (unsigned parti = 0, parte = vtparts.size(); parti != parte;
2525 ++parti) {
2526 InVals.push_back(DAG.getNode(ISD::UNDEF, dl, Ins[InsIdx].VT));
2527 ++InsIdx;
2528 }
2529 if (vtparts.size() > 0)
2530 --InsIdx;
2531 continue;
2532 }
2533 if (Ty->isVectorTy()) {
2534 EVT ObjectVT = getValueType(DL, Ty);
2535 unsigned NumRegs = TLI->getNumRegisters(F->getContext(), ObjectVT);
2536 for (unsigned parti = 0; parti < NumRegs; ++parti) {
2537 InVals.push_back(DAG.getNode(ISD::UNDEF, dl, Ins[InsIdx].VT));
2538 ++InsIdx;
2539 }
2540 if (NumRegs > 0)
2541 --InsIdx;
2542 continue;
2543 }
2544 InVals.push_back(DAG.getNode(ISD::UNDEF, dl, Ins[InsIdx].VT));
2545 continue;
2546 }
2547
2548 // In the following cases, assign a node order of "idx+1"
2549 // to newly created nodes. The SDNodes for params have to
2550 // appear in the same order as their order of appearance
2551 // in the original function. "idx+1" holds that order.
2552 if (!PAL.hasParamAttribute(i, Attribute::ByVal)) {
2553 bool aggregateIsPacked = false;
2554 if (StructType *STy = dyn_cast<StructType>(Ty))
2555 aggregateIsPacked = STy->isPacked();
2556
2557 SmallVector<EVT, 16> VTs;
2558 SmallVector<uint64_t, 16> Offsets;
2559 ComputePTXValueVTs(*this, DL, Ty, VTs, &Offsets, 0);
2560 assert(VTs.size() > 0 && "Unexpected empty type.");
2561 auto VectorInfo =
2562 VectorizePTXValueVTs(VTs, Offsets, DL.getABITypeAlignment(Ty));
2563
2564 SDValue Arg = getParamSymbol(DAG, idx, PtrVT);
2565 int VecIdx = -1; // Index of the first element of the current vector.
2566 for (unsigned parti = 0, parte = VTs.size(); parti != parte; ++parti) {
2567 if (VectorInfo[parti] & PVF_FIRST) {
2568 assert(VecIdx == -1 && "Orphaned vector.");
2569 VecIdx = parti;
2570 }
2571
2572 // That's the last element of this store op.
2573 if (VectorInfo[parti] & PVF_LAST) {
2574 unsigned NumElts = parti - VecIdx + 1;
2575 EVT EltVT = VTs[parti];
2576 // i1 is loaded/stored as i8.
2577 EVT LoadVT = EltVT;
2578 if (EltVT == MVT::i1)
2579 LoadVT = MVT::i8;
2580 else if (EltVT == MVT::v2f16)
2581 // getLoad needs a vector type, but it can't handle
2582 // vectors which contain v2f16 elements. So we must load
2583 // using i32 here and then bitcast back.
2584 LoadVT = MVT::i32;
2585
2586 EVT VecVT = EVT::getVectorVT(F->getContext(), LoadVT, NumElts);
2587 SDValue VecAddr =
2588 DAG.getNode(ISD::ADD, dl, PtrVT, Arg,
2589 DAG.getConstant(Offsets[VecIdx], dl, PtrVT));
2590 Value *srcValue = Constant::getNullValue(PointerType::get(
2591 EltVT.getTypeForEVT(F->getContext()), ADDRESS_SPACE_PARAM));
2592 SDValue P =
2593 DAG.getLoad(VecVT, dl, Root, VecAddr,
2594 MachinePointerInfo(srcValue), aggregateIsPacked,
2595 MachineMemOperand::MODereferenceable |
2596 MachineMemOperand::MOInvariant);
2597 if (P.getNode())
2598 P.getNode()->setIROrder(idx + 1);
2599 for (unsigned j = 0; j < NumElts; ++j) {
2600 SDValue Elt = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, LoadVT, P,
2601 DAG.getIntPtrConstant(j, dl));
2602 // We've loaded i1 as an i8 and now must truncate it back to i1
2603 if (EltVT == MVT::i1)
2604 Elt = DAG.getNode(ISD::TRUNCATE, dl, MVT::i1, Elt);
2605 // v2f16 was loaded as an i32. Now we must bitcast it back.
2606 else if (EltVT == MVT::v2f16)
2607 Elt = DAG.getNode(ISD::BITCAST, dl, MVT::v2f16, Elt);
2608 // Extend the element if necessary (e.g. an i8 is loaded
2609 // into an i16 register)
2610 if (Ins[InsIdx].VT.isInteger() &&
2611 Ins[InsIdx].VT.getSizeInBits() > LoadVT.getSizeInBits()) {
2612 unsigned Extend = Ins[InsIdx].Flags.isSExt() ? ISD::SIGN_EXTEND
2613 : ISD::ZERO_EXTEND;
2614 Elt = DAG.getNode(Extend, dl, Ins[InsIdx].VT, Elt);
2615 }
2616 InVals.push_back(Elt);
2617 }
2618
2619 // Reset vector tracking state.
2620 VecIdx = -1;
2621 }
2622 ++InsIdx;
2623 }
2624 if (VTs.size() > 0)
2625 --InsIdx;
2626 continue;
2627 }
2628
2629 // Param has ByVal attribute
2630 // Return MoveParam(param symbol).
2631 // Ideally, the param symbol can be returned directly,
2632 // but when SDNode builder decides to use it in a CopyToReg(),
2633 // machine instruction fails because TargetExternalSymbol
2634 // (not lowered) is target dependent, and CopyToReg assumes
2635 // the source is lowered.
2636 EVT ObjectVT = getValueType(DL, Ty);
2637 assert(ObjectVT == Ins[InsIdx].VT &&
2638 "Ins type did not match function type");
2639 SDValue Arg = getParamSymbol(DAG, idx, PtrVT);
2640 SDValue p = DAG.getNode(NVPTXISD::MoveParam, dl, ObjectVT, Arg);
2641 if (p.getNode())
2642 p.getNode()->setIROrder(idx + 1);
2643 InVals.push_back(p);
2644 }
2645
2646 // Clang will check explicit VarArg and issue error if any. However, Clang
2647 // will let code with
2648 // implicit var arg like f() pass. See bug 617733.
2649 // We treat this case as if the arg list is empty.
2650 // if (F.isVarArg()) {
2651 // assert(0 && "VarArg not supported yet!");
2652 //}
2653
2654 if (!OutChains.empty())
2655 DAG.setRoot(DAG.getNode(ISD::TokenFactor, dl, MVT::Other, OutChains));
2656
2657 return Chain;
2658 }
2659
2660 SDValue
LowerReturn(SDValue Chain,CallingConv::ID CallConv,bool isVarArg,const SmallVectorImpl<ISD::OutputArg> & Outs,const SmallVectorImpl<SDValue> & OutVals,const SDLoc & dl,SelectionDAG & DAG) const2661 NVPTXTargetLowering::LowerReturn(SDValue Chain, CallingConv::ID CallConv,
2662 bool isVarArg,
2663 const SmallVectorImpl<ISD::OutputArg> &Outs,
2664 const SmallVectorImpl<SDValue> &OutVals,
2665 const SDLoc &dl, SelectionDAG &DAG) const {
2666 MachineFunction &MF = DAG.getMachineFunction();
2667 Type *RetTy = MF.getFunction().getReturnType();
2668
2669 bool isABI = (STI.getSmVersion() >= 20);
2670 assert(isABI && "Non-ABI compilation is not supported");
2671 if (!isABI)
2672 return Chain;
2673
2674 const DataLayout DL = DAG.getDataLayout();
2675 SmallVector<EVT, 16> VTs;
2676 SmallVector<uint64_t, 16> Offsets;
2677 ComputePTXValueVTs(*this, DL, RetTy, VTs, &Offsets);
2678 assert(VTs.size() == OutVals.size() && "Bad return value decomposition");
2679
2680 auto VectorInfo = VectorizePTXValueVTs(
2681 VTs, Offsets, RetTy->isSized() ? DL.getABITypeAlignment(RetTy) : 1);
2682
2683 // PTX Interoperability Guide 3.3(A): [Integer] Values shorter than
2684 // 32-bits are sign extended or zero extended, depending on whether
2685 // they are signed or unsigned types.
2686 bool ExtendIntegerRetVal =
2687 RetTy->isIntegerTy() && DL.getTypeAllocSizeInBits(RetTy) < 32;
2688
2689 SmallVector<SDValue, 6> StoreOperands;
2690 for (unsigned i = 0, e = VTs.size(); i != e; ++i) {
2691 // New load/store. Record chain and offset operands.
2692 if (VectorInfo[i] & PVF_FIRST) {
2693 assert(StoreOperands.empty() && "Orphaned operand list.");
2694 StoreOperands.push_back(Chain);
2695 StoreOperands.push_back(DAG.getConstant(Offsets[i], dl, MVT::i32));
2696 }
2697
2698 SDValue RetVal = OutVals[i];
2699 if (ExtendIntegerRetVal) {
2700 RetVal = DAG.getNode(Outs[i].Flags.isSExt() ? ISD::SIGN_EXTEND
2701 : ISD::ZERO_EXTEND,
2702 dl, MVT::i32, RetVal);
2703 } else if (RetVal.getValueSizeInBits() < 16) {
2704 // Use 16-bit registers for small load-stores as it's the
2705 // smallest general purpose register size supported by NVPTX.
2706 RetVal = DAG.getNode(ISD::ANY_EXTEND, dl, MVT::i16, RetVal);
2707 }
2708
2709 // Record the value to return.
2710 StoreOperands.push_back(RetVal);
2711
2712 // That's the last element of this store op.
2713 if (VectorInfo[i] & PVF_LAST) {
2714 NVPTXISD::NodeType Op;
2715 unsigned NumElts = StoreOperands.size() - 2;
2716 switch (NumElts) {
2717 case 1:
2718 Op = NVPTXISD::StoreRetval;
2719 break;
2720 case 2:
2721 Op = NVPTXISD::StoreRetvalV2;
2722 break;
2723 case 4:
2724 Op = NVPTXISD::StoreRetvalV4;
2725 break;
2726 default:
2727 llvm_unreachable("Invalid vector info.");
2728 }
2729
2730 // Adjust type of load/store op if we've extended the scalar
2731 // return value.
2732 EVT TheStoreType = ExtendIntegerRetVal ? MVT::i32 : VTs[i];
2733 Chain = DAG.getMemIntrinsicNode(Op, dl, DAG.getVTList(MVT::Other),
2734 StoreOperands, TheStoreType,
2735 MachinePointerInfo(), /* Align */ 1,
2736 MachineMemOperand::MOStore);
2737 // Cleanup vector state.
2738 StoreOperands.clear();
2739 }
2740 }
2741
2742 return DAG.getNode(NVPTXISD::RET_FLAG, dl, MVT::Other, Chain);
2743 }
2744
LowerAsmOperandForConstraint(SDValue Op,std::string & Constraint,std::vector<SDValue> & Ops,SelectionDAG & DAG) const2745 void NVPTXTargetLowering::LowerAsmOperandForConstraint(
2746 SDValue Op, std::string &Constraint, std::vector<SDValue> &Ops,
2747 SelectionDAG &DAG) const {
2748 if (Constraint.length() > 1)
2749 return;
2750 else
2751 TargetLowering::LowerAsmOperandForConstraint(Op, Constraint, Ops, DAG);
2752 }
2753
getOpcForTextureInstr(unsigned Intrinsic)2754 static unsigned getOpcForTextureInstr(unsigned Intrinsic) {
2755 switch (Intrinsic) {
2756 default:
2757 return 0;
2758
2759 case Intrinsic::nvvm_tex_1d_v4f32_s32:
2760 return NVPTXISD::Tex1DFloatS32;
2761 case Intrinsic::nvvm_tex_1d_v4f32_f32:
2762 return NVPTXISD::Tex1DFloatFloat;
2763 case Intrinsic::nvvm_tex_1d_level_v4f32_f32:
2764 return NVPTXISD::Tex1DFloatFloatLevel;
2765 case Intrinsic::nvvm_tex_1d_grad_v4f32_f32:
2766 return NVPTXISD::Tex1DFloatFloatGrad;
2767 case Intrinsic::nvvm_tex_1d_v4s32_s32:
2768 return NVPTXISD::Tex1DS32S32;
2769 case Intrinsic::nvvm_tex_1d_v4s32_f32:
2770 return NVPTXISD::Tex1DS32Float;
2771 case Intrinsic::nvvm_tex_1d_level_v4s32_f32:
2772 return NVPTXISD::Tex1DS32FloatLevel;
2773 case Intrinsic::nvvm_tex_1d_grad_v4s32_f32:
2774 return NVPTXISD::Tex1DS32FloatGrad;
2775 case Intrinsic::nvvm_tex_1d_v4u32_s32:
2776 return NVPTXISD::Tex1DU32S32;
2777 case Intrinsic::nvvm_tex_1d_v4u32_f32:
2778 return NVPTXISD::Tex1DU32Float;
2779 case Intrinsic::nvvm_tex_1d_level_v4u32_f32:
2780 return NVPTXISD::Tex1DU32FloatLevel;
2781 case Intrinsic::nvvm_tex_1d_grad_v4u32_f32:
2782 return NVPTXISD::Tex1DU32FloatGrad;
2783
2784 case Intrinsic::nvvm_tex_1d_array_v4f32_s32:
2785 return NVPTXISD::Tex1DArrayFloatS32;
2786 case Intrinsic::nvvm_tex_1d_array_v4f32_f32:
2787 return NVPTXISD::Tex1DArrayFloatFloat;
2788 case Intrinsic::nvvm_tex_1d_array_level_v4f32_f32:
2789 return NVPTXISD::Tex1DArrayFloatFloatLevel;
2790 case Intrinsic::nvvm_tex_1d_array_grad_v4f32_f32:
2791 return NVPTXISD::Tex1DArrayFloatFloatGrad;
2792 case Intrinsic::nvvm_tex_1d_array_v4s32_s32:
2793 return NVPTXISD::Tex1DArrayS32S32;
2794 case Intrinsic::nvvm_tex_1d_array_v4s32_f32:
2795 return NVPTXISD::Tex1DArrayS32Float;
2796 case Intrinsic::nvvm_tex_1d_array_level_v4s32_f32:
2797 return NVPTXISD::Tex1DArrayS32FloatLevel;
2798 case Intrinsic::nvvm_tex_1d_array_grad_v4s32_f32:
2799 return NVPTXISD::Tex1DArrayS32FloatGrad;
2800 case Intrinsic::nvvm_tex_1d_array_v4u32_s32:
2801 return NVPTXISD::Tex1DArrayU32S32;
2802 case Intrinsic::nvvm_tex_1d_array_v4u32_f32:
2803 return NVPTXISD::Tex1DArrayU32Float;
2804 case Intrinsic::nvvm_tex_1d_array_level_v4u32_f32:
2805 return NVPTXISD::Tex1DArrayU32FloatLevel;
2806 case Intrinsic::nvvm_tex_1d_array_grad_v4u32_f32:
2807 return NVPTXISD::Tex1DArrayU32FloatGrad;
2808
2809 case Intrinsic::nvvm_tex_2d_v4f32_s32:
2810 return NVPTXISD::Tex2DFloatS32;
2811 case Intrinsic::nvvm_tex_2d_v4f32_f32:
2812 return NVPTXISD::Tex2DFloatFloat;
2813 case Intrinsic::nvvm_tex_2d_level_v4f32_f32:
2814 return NVPTXISD::Tex2DFloatFloatLevel;
2815 case Intrinsic::nvvm_tex_2d_grad_v4f32_f32:
2816 return NVPTXISD::Tex2DFloatFloatGrad;
2817 case Intrinsic::nvvm_tex_2d_v4s32_s32:
2818 return NVPTXISD::Tex2DS32S32;
2819 case Intrinsic::nvvm_tex_2d_v4s32_f32:
2820 return NVPTXISD::Tex2DS32Float;
2821 case Intrinsic::nvvm_tex_2d_level_v4s32_f32:
2822 return NVPTXISD::Tex2DS32FloatLevel;
2823 case Intrinsic::nvvm_tex_2d_grad_v4s32_f32:
2824 return NVPTXISD::Tex2DS32FloatGrad;
2825 case Intrinsic::nvvm_tex_2d_v4u32_s32:
2826 return NVPTXISD::Tex2DU32S32;
2827 case Intrinsic::nvvm_tex_2d_v4u32_f32:
2828 return NVPTXISD::Tex2DU32Float;
2829 case Intrinsic::nvvm_tex_2d_level_v4u32_f32:
2830 return NVPTXISD::Tex2DU32FloatLevel;
2831 case Intrinsic::nvvm_tex_2d_grad_v4u32_f32:
2832 return NVPTXISD::Tex2DU32FloatGrad;
2833
2834 case Intrinsic::nvvm_tex_2d_array_v4f32_s32:
2835 return NVPTXISD::Tex2DArrayFloatS32;
2836 case Intrinsic::nvvm_tex_2d_array_v4f32_f32:
2837 return NVPTXISD::Tex2DArrayFloatFloat;
2838 case Intrinsic::nvvm_tex_2d_array_level_v4f32_f32:
2839 return NVPTXISD::Tex2DArrayFloatFloatLevel;
2840 case Intrinsic::nvvm_tex_2d_array_grad_v4f32_f32:
2841 return NVPTXISD::Tex2DArrayFloatFloatGrad;
2842 case Intrinsic::nvvm_tex_2d_array_v4s32_s32:
2843 return NVPTXISD::Tex2DArrayS32S32;
2844 case Intrinsic::nvvm_tex_2d_array_v4s32_f32:
2845 return NVPTXISD::Tex2DArrayS32Float;
2846 case Intrinsic::nvvm_tex_2d_array_level_v4s32_f32:
2847 return NVPTXISD::Tex2DArrayS32FloatLevel;
2848 case Intrinsic::nvvm_tex_2d_array_grad_v4s32_f32:
2849 return NVPTXISD::Tex2DArrayS32FloatGrad;
2850 case Intrinsic::nvvm_tex_2d_array_v4u32_s32:
2851 return NVPTXISD::Tex2DArrayU32S32;
2852 case Intrinsic::nvvm_tex_2d_array_v4u32_f32:
2853 return NVPTXISD::Tex2DArrayU32Float;
2854 case Intrinsic::nvvm_tex_2d_array_level_v4u32_f32:
2855 return NVPTXISD::Tex2DArrayU32FloatLevel;
2856 case Intrinsic::nvvm_tex_2d_array_grad_v4u32_f32:
2857 return NVPTXISD::Tex2DArrayU32FloatGrad;
2858
2859 case Intrinsic::nvvm_tex_3d_v4f32_s32:
2860 return NVPTXISD::Tex3DFloatS32;
2861 case Intrinsic::nvvm_tex_3d_v4f32_f32:
2862 return NVPTXISD::Tex3DFloatFloat;
2863 case Intrinsic::nvvm_tex_3d_level_v4f32_f32:
2864 return NVPTXISD::Tex3DFloatFloatLevel;
2865 case Intrinsic::nvvm_tex_3d_grad_v4f32_f32:
2866 return NVPTXISD::Tex3DFloatFloatGrad;
2867 case Intrinsic::nvvm_tex_3d_v4s32_s32:
2868 return NVPTXISD::Tex3DS32S32;
2869 case Intrinsic::nvvm_tex_3d_v4s32_f32:
2870 return NVPTXISD::Tex3DS32Float;
2871 case Intrinsic::nvvm_tex_3d_level_v4s32_f32:
2872 return NVPTXISD::Tex3DS32FloatLevel;
2873 case Intrinsic::nvvm_tex_3d_grad_v4s32_f32:
2874 return NVPTXISD::Tex3DS32FloatGrad;
2875 case Intrinsic::nvvm_tex_3d_v4u32_s32:
2876 return NVPTXISD::Tex3DU32S32;
2877 case Intrinsic::nvvm_tex_3d_v4u32_f32:
2878 return NVPTXISD::Tex3DU32Float;
2879 case Intrinsic::nvvm_tex_3d_level_v4u32_f32:
2880 return NVPTXISD::Tex3DU32FloatLevel;
2881 case Intrinsic::nvvm_tex_3d_grad_v4u32_f32:
2882 return NVPTXISD::Tex3DU32FloatGrad;
2883
2884 case Intrinsic::nvvm_tex_cube_v4f32_f32:
2885 return NVPTXISD::TexCubeFloatFloat;
2886 case Intrinsic::nvvm_tex_cube_level_v4f32_f32:
2887 return NVPTXISD::TexCubeFloatFloatLevel;
2888 case Intrinsic::nvvm_tex_cube_v4s32_f32:
2889 return NVPTXISD::TexCubeS32Float;
2890 case Intrinsic::nvvm_tex_cube_level_v4s32_f32:
2891 return NVPTXISD::TexCubeS32FloatLevel;
2892 case Intrinsic::nvvm_tex_cube_v4u32_f32:
2893 return NVPTXISD::TexCubeU32Float;
2894 case Intrinsic::nvvm_tex_cube_level_v4u32_f32:
2895 return NVPTXISD::TexCubeU32FloatLevel;
2896
2897 case Intrinsic::nvvm_tex_cube_array_v4f32_f32:
2898 return NVPTXISD::TexCubeArrayFloatFloat;
2899 case Intrinsic::nvvm_tex_cube_array_level_v4f32_f32:
2900 return NVPTXISD::TexCubeArrayFloatFloatLevel;
2901 case Intrinsic::nvvm_tex_cube_array_v4s32_f32:
2902 return NVPTXISD::TexCubeArrayS32Float;
2903 case Intrinsic::nvvm_tex_cube_array_level_v4s32_f32:
2904 return NVPTXISD::TexCubeArrayS32FloatLevel;
2905 case Intrinsic::nvvm_tex_cube_array_v4u32_f32:
2906 return NVPTXISD::TexCubeArrayU32Float;
2907 case Intrinsic::nvvm_tex_cube_array_level_v4u32_f32:
2908 return NVPTXISD::TexCubeArrayU32FloatLevel;
2909
2910 case Intrinsic::nvvm_tld4_r_2d_v4f32_f32:
2911 return NVPTXISD::Tld4R2DFloatFloat;
2912 case Intrinsic::nvvm_tld4_g_2d_v4f32_f32:
2913 return NVPTXISD::Tld4G2DFloatFloat;
2914 case Intrinsic::nvvm_tld4_b_2d_v4f32_f32:
2915 return NVPTXISD::Tld4B2DFloatFloat;
2916 case Intrinsic::nvvm_tld4_a_2d_v4f32_f32:
2917 return NVPTXISD::Tld4A2DFloatFloat;
2918 case Intrinsic::nvvm_tld4_r_2d_v4s32_f32:
2919 return NVPTXISD::Tld4R2DS64Float;
2920 case Intrinsic::nvvm_tld4_g_2d_v4s32_f32:
2921 return NVPTXISD::Tld4G2DS64Float;
2922 case Intrinsic::nvvm_tld4_b_2d_v4s32_f32:
2923 return NVPTXISD::Tld4B2DS64Float;
2924 case Intrinsic::nvvm_tld4_a_2d_v4s32_f32:
2925 return NVPTXISD::Tld4A2DS64Float;
2926 case Intrinsic::nvvm_tld4_r_2d_v4u32_f32:
2927 return NVPTXISD::Tld4R2DU64Float;
2928 case Intrinsic::nvvm_tld4_g_2d_v4u32_f32:
2929 return NVPTXISD::Tld4G2DU64Float;
2930 case Intrinsic::nvvm_tld4_b_2d_v4u32_f32:
2931 return NVPTXISD::Tld4B2DU64Float;
2932 case Intrinsic::nvvm_tld4_a_2d_v4u32_f32:
2933 return NVPTXISD::Tld4A2DU64Float;
2934
2935 case Intrinsic::nvvm_tex_unified_1d_v4f32_s32:
2936 return NVPTXISD::TexUnified1DFloatS32;
2937 case Intrinsic::nvvm_tex_unified_1d_v4f32_f32:
2938 return NVPTXISD::TexUnified1DFloatFloat;
2939 case Intrinsic::nvvm_tex_unified_1d_level_v4f32_f32:
2940 return NVPTXISD::TexUnified1DFloatFloatLevel;
2941 case Intrinsic::nvvm_tex_unified_1d_grad_v4f32_f32:
2942 return NVPTXISD::TexUnified1DFloatFloatGrad;
2943 case Intrinsic::nvvm_tex_unified_1d_v4s32_s32:
2944 return NVPTXISD::TexUnified1DS32S32;
2945 case Intrinsic::nvvm_tex_unified_1d_v4s32_f32:
2946 return NVPTXISD::TexUnified1DS32Float;
2947 case Intrinsic::nvvm_tex_unified_1d_level_v4s32_f32:
2948 return NVPTXISD::TexUnified1DS32FloatLevel;
2949 case Intrinsic::nvvm_tex_unified_1d_grad_v4s32_f32:
2950 return NVPTXISD::TexUnified1DS32FloatGrad;
2951 case Intrinsic::nvvm_tex_unified_1d_v4u32_s32:
2952 return NVPTXISD::TexUnified1DU32S32;
2953 case Intrinsic::nvvm_tex_unified_1d_v4u32_f32:
2954 return NVPTXISD::TexUnified1DU32Float;
2955 case Intrinsic::nvvm_tex_unified_1d_level_v4u32_f32:
2956 return NVPTXISD::TexUnified1DU32FloatLevel;
2957 case Intrinsic::nvvm_tex_unified_1d_grad_v4u32_f32:
2958 return NVPTXISD::TexUnified1DU32FloatGrad;
2959
2960 case Intrinsic::nvvm_tex_unified_1d_array_v4f32_s32:
2961 return NVPTXISD::TexUnified1DArrayFloatS32;
2962 case Intrinsic::nvvm_tex_unified_1d_array_v4f32_f32:
2963 return NVPTXISD::TexUnified1DArrayFloatFloat;
2964 case Intrinsic::nvvm_tex_unified_1d_array_level_v4f32_f32:
2965 return NVPTXISD::TexUnified1DArrayFloatFloatLevel;
2966 case Intrinsic::nvvm_tex_unified_1d_array_grad_v4f32_f32:
2967 return NVPTXISD::TexUnified1DArrayFloatFloatGrad;
2968 case Intrinsic::nvvm_tex_unified_1d_array_v4s32_s32:
2969 return NVPTXISD::TexUnified1DArrayS32S32;
2970 case Intrinsic::nvvm_tex_unified_1d_array_v4s32_f32:
2971 return NVPTXISD::TexUnified1DArrayS32Float;
2972 case Intrinsic::nvvm_tex_unified_1d_array_level_v4s32_f32:
2973 return NVPTXISD::TexUnified1DArrayS32FloatLevel;
2974 case Intrinsic::nvvm_tex_unified_1d_array_grad_v4s32_f32:
2975 return NVPTXISD::TexUnified1DArrayS32FloatGrad;
2976 case Intrinsic::nvvm_tex_unified_1d_array_v4u32_s32:
2977 return NVPTXISD::TexUnified1DArrayU32S32;
2978 case Intrinsic::nvvm_tex_unified_1d_array_v4u32_f32:
2979 return NVPTXISD::TexUnified1DArrayU32Float;
2980 case Intrinsic::nvvm_tex_unified_1d_array_level_v4u32_f32:
2981 return NVPTXISD::TexUnified1DArrayU32FloatLevel;
2982 case Intrinsic::nvvm_tex_unified_1d_array_grad_v4u32_f32:
2983 return NVPTXISD::TexUnified1DArrayU32FloatGrad;
2984
2985 case Intrinsic::nvvm_tex_unified_2d_v4f32_s32:
2986 return NVPTXISD::TexUnified2DFloatS32;
2987 case Intrinsic::nvvm_tex_unified_2d_v4f32_f32:
2988 return NVPTXISD::TexUnified2DFloatFloat;
2989 case Intrinsic::nvvm_tex_unified_2d_level_v4f32_f32:
2990 return NVPTXISD::TexUnified2DFloatFloatLevel;
2991 case Intrinsic::nvvm_tex_unified_2d_grad_v4f32_f32:
2992 return NVPTXISD::TexUnified2DFloatFloatGrad;
2993 case Intrinsic::nvvm_tex_unified_2d_v4s32_s32:
2994 return NVPTXISD::TexUnified2DS32S32;
2995 case Intrinsic::nvvm_tex_unified_2d_v4s32_f32:
2996 return NVPTXISD::TexUnified2DS32Float;
2997 case Intrinsic::nvvm_tex_unified_2d_level_v4s32_f32:
2998 return NVPTXISD::TexUnified2DS32FloatLevel;
2999 case Intrinsic::nvvm_tex_unified_2d_grad_v4s32_f32:
3000 return NVPTXISD::TexUnified2DS32FloatGrad;
3001 case Intrinsic::nvvm_tex_unified_2d_v4u32_s32:
3002 return NVPTXISD::TexUnified2DU32S32;
3003 case Intrinsic::nvvm_tex_unified_2d_v4u32_f32:
3004 return NVPTXISD::TexUnified2DU32Float;
3005 case Intrinsic::nvvm_tex_unified_2d_level_v4u32_f32:
3006 return NVPTXISD::TexUnified2DU32FloatLevel;
3007 case Intrinsic::nvvm_tex_unified_2d_grad_v4u32_f32:
3008 return NVPTXISD::TexUnified2DU32FloatGrad;
3009
3010 case Intrinsic::nvvm_tex_unified_2d_array_v4f32_s32:
3011 return NVPTXISD::TexUnified2DArrayFloatS32;
3012 case Intrinsic::nvvm_tex_unified_2d_array_v4f32_f32:
3013 return NVPTXISD::TexUnified2DArrayFloatFloat;
3014 case Intrinsic::nvvm_tex_unified_2d_array_level_v4f32_f32:
3015 return NVPTXISD::TexUnified2DArrayFloatFloatLevel;
3016 case Intrinsic::nvvm_tex_unified_2d_array_grad_v4f32_f32:
3017 return NVPTXISD::TexUnified2DArrayFloatFloatGrad;
3018 case Intrinsic::nvvm_tex_unified_2d_array_v4s32_s32:
3019 return NVPTXISD::TexUnified2DArrayS32S32;
3020 case Intrinsic::nvvm_tex_unified_2d_array_v4s32_f32:
3021 return NVPTXISD::TexUnified2DArrayS32Float;
3022 case Intrinsic::nvvm_tex_unified_2d_array_level_v4s32_f32:
3023 return NVPTXISD::TexUnified2DArrayS32FloatLevel;
3024 case Intrinsic::nvvm_tex_unified_2d_array_grad_v4s32_f32:
3025 return NVPTXISD::TexUnified2DArrayS32FloatGrad;
3026 case Intrinsic::nvvm_tex_unified_2d_array_v4u32_s32:
3027 return NVPTXISD::TexUnified2DArrayU32S32;
3028 case Intrinsic::nvvm_tex_unified_2d_array_v4u32_f32:
3029 return NVPTXISD::TexUnified2DArrayU32Float;
3030 case Intrinsic::nvvm_tex_unified_2d_array_level_v4u32_f32:
3031 return NVPTXISD::TexUnified2DArrayU32FloatLevel;
3032 case Intrinsic::nvvm_tex_unified_2d_array_grad_v4u32_f32:
3033 return NVPTXISD::TexUnified2DArrayU32FloatGrad;
3034
3035 case Intrinsic::nvvm_tex_unified_3d_v4f32_s32:
3036 return NVPTXISD::TexUnified3DFloatS32;
3037 case Intrinsic::nvvm_tex_unified_3d_v4f32_f32:
3038 return NVPTXISD::TexUnified3DFloatFloat;
3039 case Intrinsic::nvvm_tex_unified_3d_level_v4f32_f32:
3040 return NVPTXISD::TexUnified3DFloatFloatLevel;
3041 case Intrinsic::nvvm_tex_unified_3d_grad_v4f32_f32:
3042 return NVPTXISD::TexUnified3DFloatFloatGrad;
3043 case Intrinsic::nvvm_tex_unified_3d_v4s32_s32:
3044 return NVPTXISD::TexUnified3DS32S32;
3045 case Intrinsic::nvvm_tex_unified_3d_v4s32_f32:
3046 return NVPTXISD::TexUnified3DS32Float;
3047 case Intrinsic::nvvm_tex_unified_3d_level_v4s32_f32:
3048 return NVPTXISD::TexUnified3DS32FloatLevel;
3049 case Intrinsic::nvvm_tex_unified_3d_grad_v4s32_f32:
3050 return NVPTXISD::TexUnified3DS32FloatGrad;
3051 case Intrinsic::nvvm_tex_unified_3d_v4u32_s32:
3052 return NVPTXISD::TexUnified3DU32S32;
3053 case Intrinsic::nvvm_tex_unified_3d_v4u32_f32:
3054 return NVPTXISD::TexUnified3DU32Float;
3055 case Intrinsic::nvvm_tex_unified_3d_level_v4u32_f32:
3056 return NVPTXISD::TexUnified3DU32FloatLevel;
3057 case Intrinsic::nvvm_tex_unified_3d_grad_v4u32_f32:
3058 return NVPTXISD::TexUnified3DU32FloatGrad;
3059
3060 case Intrinsic::nvvm_tex_unified_cube_v4f32_f32:
3061 return NVPTXISD::TexUnifiedCubeFloatFloat;
3062 case Intrinsic::nvvm_tex_unified_cube_level_v4f32_f32:
3063 return NVPTXISD::TexUnifiedCubeFloatFloatLevel;
3064 case Intrinsic::nvvm_tex_unified_cube_v4s32_f32:
3065 return NVPTXISD::TexUnifiedCubeS32Float;
3066 case Intrinsic::nvvm_tex_unified_cube_level_v4s32_f32:
3067 return NVPTXISD::TexUnifiedCubeS32FloatLevel;
3068 case Intrinsic::nvvm_tex_unified_cube_v4u32_f32:
3069 return NVPTXISD::TexUnifiedCubeU32Float;
3070 case Intrinsic::nvvm_tex_unified_cube_level_v4u32_f32:
3071 return NVPTXISD::TexUnifiedCubeU32FloatLevel;
3072
3073 case Intrinsic::nvvm_tex_unified_cube_array_v4f32_f32:
3074 return NVPTXISD::TexUnifiedCubeArrayFloatFloat;
3075 case Intrinsic::nvvm_tex_unified_cube_array_level_v4f32_f32:
3076 return NVPTXISD::TexUnifiedCubeArrayFloatFloatLevel;
3077 case Intrinsic::nvvm_tex_unified_cube_array_v4s32_f32:
3078 return NVPTXISD::TexUnifiedCubeArrayS32Float;
3079 case Intrinsic::nvvm_tex_unified_cube_array_level_v4s32_f32:
3080 return NVPTXISD::TexUnifiedCubeArrayS32FloatLevel;
3081 case Intrinsic::nvvm_tex_unified_cube_array_v4u32_f32:
3082 return NVPTXISD::TexUnifiedCubeArrayU32Float;
3083 case Intrinsic::nvvm_tex_unified_cube_array_level_v4u32_f32:
3084 return NVPTXISD::TexUnifiedCubeArrayU32FloatLevel;
3085
3086 case Intrinsic::nvvm_tld4_unified_r_2d_v4f32_f32:
3087 return NVPTXISD::Tld4UnifiedR2DFloatFloat;
3088 case Intrinsic::nvvm_tld4_unified_g_2d_v4f32_f32:
3089 return NVPTXISD::Tld4UnifiedG2DFloatFloat;
3090 case Intrinsic::nvvm_tld4_unified_b_2d_v4f32_f32:
3091 return NVPTXISD::Tld4UnifiedB2DFloatFloat;
3092 case Intrinsic::nvvm_tld4_unified_a_2d_v4f32_f32:
3093 return NVPTXISD::Tld4UnifiedA2DFloatFloat;
3094 case Intrinsic::nvvm_tld4_unified_r_2d_v4s32_f32:
3095 return NVPTXISD::Tld4UnifiedR2DS64Float;
3096 case Intrinsic::nvvm_tld4_unified_g_2d_v4s32_f32:
3097 return NVPTXISD::Tld4UnifiedG2DS64Float;
3098 case Intrinsic::nvvm_tld4_unified_b_2d_v4s32_f32:
3099 return NVPTXISD::Tld4UnifiedB2DS64Float;
3100 case Intrinsic::nvvm_tld4_unified_a_2d_v4s32_f32:
3101 return NVPTXISD::Tld4UnifiedA2DS64Float;
3102 case Intrinsic::nvvm_tld4_unified_r_2d_v4u32_f32:
3103 return NVPTXISD::Tld4UnifiedR2DU64Float;
3104 case Intrinsic::nvvm_tld4_unified_g_2d_v4u32_f32:
3105 return NVPTXISD::Tld4UnifiedG2DU64Float;
3106 case Intrinsic::nvvm_tld4_unified_b_2d_v4u32_f32:
3107 return NVPTXISD::Tld4UnifiedB2DU64Float;
3108 case Intrinsic::nvvm_tld4_unified_a_2d_v4u32_f32:
3109 return NVPTXISD::Tld4UnifiedA2DU64Float;
3110 }
3111 }
3112
getOpcForSurfaceInstr(unsigned Intrinsic)3113 static unsigned getOpcForSurfaceInstr(unsigned Intrinsic) {
3114 switch (Intrinsic) {
3115 default:
3116 return 0;
3117 case Intrinsic::nvvm_suld_1d_i8_clamp:
3118 return NVPTXISD::Suld1DI8Clamp;
3119 case Intrinsic::nvvm_suld_1d_i16_clamp:
3120 return NVPTXISD::Suld1DI16Clamp;
3121 case Intrinsic::nvvm_suld_1d_i32_clamp:
3122 return NVPTXISD::Suld1DI32Clamp;
3123 case Intrinsic::nvvm_suld_1d_i64_clamp:
3124 return NVPTXISD::Suld1DI64Clamp;
3125 case Intrinsic::nvvm_suld_1d_v2i8_clamp:
3126 return NVPTXISD::Suld1DV2I8Clamp;
3127 case Intrinsic::nvvm_suld_1d_v2i16_clamp:
3128 return NVPTXISD::Suld1DV2I16Clamp;
3129 case Intrinsic::nvvm_suld_1d_v2i32_clamp:
3130 return NVPTXISD::Suld1DV2I32Clamp;
3131 case Intrinsic::nvvm_suld_1d_v2i64_clamp:
3132 return NVPTXISD::Suld1DV2I64Clamp;
3133 case Intrinsic::nvvm_suld_1d_v4i8_clamp:
3134 return NVPTXISD::Suld1DV4I8Clamp;
3135 case Intrinsic::nvvm_suld_1d_v4i16_clamp:
3136 return NVPTXISD::Suld1DV4I16Clamp;
3137 case Intrinsic::nvvm_suld_1d_v4i32_clamp:
3138 return NVPTXISD::Suld1DV4I32Clamp;
3139 case Intrinsic::nvvm_suld_1d_array_i8_clamp:
3140 return NVPTXISD::Suld1DArrayI8Clamp;
3141 case Intrinsic::nvvm_suld_1d_array_i16_clamp:
3142 return NVPTXISD::Suld1DArrayI16Clamp;
3143 case Intrinsic::nvvm_suld_1d_array_i32_clamp:
3144 return NVPTXISD::Suld1DArrayI32Clamp;
3145 case Intrinsic::nvvm_suld_1d_array_i64_clamp:
3146 return NVPTXISD::Suld1DArrayI64Clamp;
3147 case Intrinsic::nvvm_suld_1d_array_v2i8_clamp:
3148 return NVPTXISD::Suld1DArrayV2I8Clamp;
3149 case Intrinsic::nvvm_suld_1d_array_v2i16_clamp:
3150 return NVPTXISD::Suld1DArrayV2I16Clamp;
3151 case Intrinsic::nvvm_suld_1d_array_v2i32_clamp:
3152 return NVPTXISD::Suld1DArrayV2I32Clamp;
3153 case Intrinsic::nvvm_suld_1d_array_v2i64_clamp:
3154 return NVPTXISD::Suld1DArrayV2I64Clamp;
3155 case Intrinsic::nvvm_suld_1d_array_v4i8_clamp:
3156 return NVPTXISD::Suld1DArrayV4I8Clamp;
3157 case Intrinsic::nvvm_suld_1d_array_v4i16_clamp:
3158 return NVPTXISD::Suld1DArrayV4I16Clamp;
3159 case Intrinsic::nvvm_suld_1d_array_v4i32_clamp:
3160 return NVPTXISD::Suld1DArrayV4I32Clamp;
3161 case Intrinsic::nvvm_suld_2d_i8_clamp:
3162 return NVPTXISD::Suld2DI8Clamp;
3163 case Intrinsic::nvvm_suld_2d_i16_clamp:
3164 return NVPTXISD::Suld2DI16Clamp;
3165 case Intrinsic::nvvm_suld_2d_i32_clamp:
3166 return NVPTXISD::Suld2DI32Clamp;
3167 case Intrinsic::nvvm_suld_2d_i64_clamp:
3168 return NVPTXISD::Suld2DI64Clamp;
3169 case Intrinsic::nvvm_suld_2d_v2i8_clamp:
3170 return NVPTXISD::Suld2DV2I8Clamp;
3171 case Intrinsic::nvvm_suld_2d_v2i16_clamp:
3172 return NVPTXISD::Suld2DV2I16Clamp;
3173 case Intrinsic::nvvm_suld_2d_v2i32_clamp:
3174 return NVPTXISD::Suld2DV2I32Clamp;
3175 case Intrinsic::nvvm_suld_2d_v2i64_clamp:
3176 return NVPTXISD::Suld2DV2I64Clamp;
3177 case Intrinsic::nvvm_suld_2d_v4i8_clamp:
3178 return NVPTXISD::Suld2DV4I8Clamp;
3179 case Intrinsic::nvvm_suld_2d_v4i16_clamp:
3180 return NVPTXISD::Suld2DV4I16Clamp;
3181 case Intrinsic::nvvm_suld_2d_v4i32_clamp:
3182 return NVPTXISD::Suld2DV4I32Clamp;
3183 case Intrinsic::nvvm_suld_2d_array_i8_clamp:
3184 return NVPTXISD::Suld2DArrayI8Clamp;
3185 case Intrinsic::nvvm_suld_2d_array_i16_clamp:
3186 return NVPTXISD::Suld2DArrayI16Clamp;
3187 case Intrinsic::nvvm_suld_2d_array_i32_clamp:
3188 return NVPTXISD::Suld2DArrayI32Clamp;
3189 case Intrinsic::nvvm_suld_2d_array_i64_clamp:
3190 return NVPTXISD::Suld2DArrayI64Clamp;
3191 case Intrinsic::nvvm_suld_2d_array_v2i8_clamp:
3192 return NVPTXISD::Suld2DArrayV2I8Clamp;
3193 case Intrinsic::nvvm_suld_2d_array_v2i16_clamp:
3194 return NVPTXISD::Suld2DArrayV2I16Clamp;
3195 case Intrinsic::nvvm_suld_2d_array_v2i32_clamp:
3196 return NVPTXISD::Suld2DArrayV2I32Clamp;
3197 case Intrinsic::nvvm_suld_2d_array_v2i64_clamp:
3198 return NVPTXISD::Suld2DArrayV2I64Clamp;
3199 case Intrinsic::nvvm_suld_2d_array_v4i8_clamp:
3200 return NVPTXISD::Suld2DArrayV4I8Clamp;
3201 case Intrinsic::nvvm_suld_2d_array_v4i16_clamp:
3202 return NVPTXISD::Suld2DArrayV4I16Clamp;
3203 case Intrinsic::nvvm_suld_2d_array_v4i32_clamp:
3204 return NVPTXISD::Suld2DArrayV4I32Clamp;
3205 case Intrinsic::nvvm_suld_3d_i8_clamp:
3206 return NVPTXISD::Suld3DI8Clamp;
3207 case Intrinsic::nvvm_suld_3d_i16_clamp:
3208 return NVPTXISD::Suld3DI16Clamp;
3209 case Intrinsic::nvvm_suld_3d_i32_clamp:
3210 return NVPTXISD::Suld3DI32Clamp;
3211 case Intrinsic::nvvm_suld_3d_i64_clamp:
3212 return NVPTXISD::Suld3DI64Clamp;
3213 case Intrinsic::nvvm_suld_3d_v2i8_clamp:
3214 return NVPTXISD::Suld3DV2I8Clamp;
3215 case Intrinsic::nvvm_suld_3d_v2i16_clamp:
3216 return NVPTXISD::Suld3DV2I16Clamp;
3217 case Intrinsic::nvvm_suld_3d_v2i32_clamp:
3218 return NVPTXISD::Suld3DV2I32Clamp;
3219 case Intrinsic::nvvm_suld_3d_v2i64_clamp:
3220 return NVPTXISD::Suld3DV2I64Clamp;
3221 case Intrinsic::nvvm_suld_3d_v4i8_clamp:
3222 return NVPTXISD::Suld3DV4I8Clamp;
3223 case Intrinsic::nvvm_suld_3d_v4i16_clamp:
3224 return NVPTXISD::Suld3DV4I16Clamp;
3225 case Intrinsic::nvvm_suld_3d_v4i32_clamp:
3226 return NVPTXISD::Suld3DV4I32Clamp;
3227 case Intrinsic::nvvm_suld_1d_i8_trap:
3228 return NVPTXISD::Suld1DI8Trap;
3229 case Intrinsic::nvvm_suld_1d_i16_trap:
3230 return NVPTXISD::Suld1DI16Trap;
3231 case Intrinsic::nvvm_suld_1d_i32_trap:
3232 return NVPTXISD::Suld1DI32Trap;
3233 case Intrinsic::nvvm_suld_1d_i64_trap:
3234 return NVPTXISD::Suld1DI64Trap;
3235 case Intrinsic::nvvm_suld_1d_v2i8_trap:
3236 return NVPTXISD::Suld1DV2I8Trap;
3237 case Intrinsic::nvvm_suld_1d_v2i16_trap:
3238 return NVPTXISD::Suld1DV2I16Trap;
3239 case Intrinsic::nvvm_suld_1d_v2i32_trap:
3240 return NVPTXISD::Suld1DV2I32Trap;
3241 case Intrinsic::nvvm_suld_1d_v2i64_trap:
3242 return NVPTXISD::Suld1DV2I64Trap;
3243 case Intrinsic::nvvm_suld_1d_v4i8_trap:
3244 return NVPTXISD::Suld1DV4I8Trap;
3245 case Intrinsic::nvvm_suld_1d_v4i16_trap:
3246 return NVPTXISD::Suld1DV4I16Trap;
3247 case Intrinsic::nvvm_suld_1d_v4i32_trap:
3248 return NVPTXISD::Suld1DV4I32Trap;
3249 case Intrinsic::nvvm_suld_1d_array_i8_trap:
3250 return NVPTXISD::Suld1DArrayI8Trap;
3251 case Intrinsic::nvvm_suld_1d_array_i16_trap:
3252 return NVPTXISD::Suld1DArrayI16Trap;
3253 case Intrinsic::nvvm_suld_1d_array_i32_trap:
3254 return NVPTXISD::Suld1DArrayI32Trap;
3255 case Intrinsic::nvvm_suld_1d_array_i64_trap:
3256 return NVPTXISD::Suld1DArrayI64Trap;
3257 case Intrinsic::nvvm_suld_1d_array_v2i8_trap:
3258 return NVPTXISD::Suld1DArrayV2I8Trap;
3259 case Intrinsic::nvvm_suld_1d_array_v2i16_trap:
3260 return NVPTXISD::Suld1DArrayV2I16Trap;
3261 case Intrinsic::nvvm_suld_1d_array_v2i32_trap:
3262 return NVPTXISD::Suld1DArrayV2I32Trap;
3263 case Intrinsic::nvvm_suld_1d_array_v2i64_trap:
3264 return NVPTXISD::Suld1DArrayV2I64Trap;
3265 case Intrinsic::nvvm_suld_1d_array_v4i8_trap:
3266 return NVPTXISD::Suld1DArrayV4I8Trap;
3267 case Intrinsic::nvvm_suld_1d_array_v4i16_trap:
3268 return NVPTXISD::Suld1DArrayV4I16Trap;
3269 case Intrinsic::nvvm_suld_1d_array_v4i32_trap:
3270 return NVPTXISD::Suld1DArrayV4I32Trap;
3271 case Intrinsic::nvvm_suld_2d_i8_trap:
3272 return NVPTXISD::Suld2DI8Trap;
3273 case Intrinsic::nvvm_suld_2d_i16_trap:
3274 return NVPTXISD::Suld2DI16Trap;
3275 case Intrinsic::nvvm_suld_2d_i32_trap:
3276 return NVPTXISD::Suld2DI32Trap;
3277 case Intrinsic::nvvm_suld_2d_i64_trap:
3278 return NVPTXISD::Suld2DI64Trap;
3279 case Intrinsic::nvvm_suld_2d_v2i8_trap:
3280 return NVPTXISD::Suld2DV2I8Trap;
3281 case Intrinsic::nvvm_suld_2d_v2i16_trap:
3282 return NVPTXISD::Suld2DV2I16Trap;
3283 case Intrinsic::nvvm_suld_2d_v2i32_trap:
3284 return NVPTXISD::Suld2DV2I32Trap;
3285 case Intrinsic::nvvm_suld_2d_v2i64_trap:
3286 return NVPTXISD::Suld2DV2I64Trap;
3287 case Intrinsic::nvvm_suld_2d_v4i8_trap:
3288 return NVPTXISD::Suld2DV4I8Trap;
3289 case Intrinsic::nvvm_suld_2d_v4i16_trap:
3290 return NVPTXISD::Suld2DV4I16Trap;
3291 case Intrinsic::nvvm_suld_2d_v4i32_trap:
3292 return NVPTXISD::Suld2DV4I32Trap;
3293 case Intrinsic::nvvm_suld_2d_array_i8_trap:
3294 return NVPTXISD::Suld2DArrayI8Trap;
3295 case Intrinsic::nvvm_suld_2d_array_i16_trap:
3296 return NVPTXISD::Suld2DArrayI16Trap;
3297 case Intrinsic::nvvm_suld_2d_array_i32_trap:
3298 return NVPTXISD::Suld2DArrayI32Trap;
3299 case Intrinsic::nvvm_suld_2d_array_i64_trap:
3300 return NVPTXISD::Suld2DArrayI64Trap;
3301 case Intrinsic::nvvm_suld_2d_array_v2i8_trap:
3302 return NVPTXISD::Suld2DArrayV2I8Trap;
3303 case Intrinsic::nvvm_suld_2d_array_v2i16_trap:
3304 return NVPTXISD::Suld2DArrayV2I16Trap;
3305 case Intrinsic::nvvm_suld_2d_array_v2i32_trap:
3306 return NVPTXISD::Suld2DArrayV2I32Trap;
3307 case Intrinsic::nvvm_suld_2d_array_v2i64_trap:
3308 return NVPTXISD::Suld2DArrayV2I64Trap;
3309 case Intrinsic::nvvm_suld_2d_array_v4i8_trap:
3310 return NVPTXISD::Suld2DArrayV4I8Trap;
3311 case Intrinsic::nvvm_suld_2d_array_v4i16_trap:
3312 return NVPTXISD::Suld2DArrayV4I16Trap;
3313 case Intrinsic::nvvm_suld_2d_array_v4i32_trap:
3314 return NVPTXISD::Suld2DArrayV4I32Trap;
3315 case Intrinsic::nvvm_suld_3d_i8_trap:
3316 return NVPTXISD::Suld3DI8Trap;
3317 case Intrinsic::nvvm_suld_3d_i16_trap:
3318 return NVPTXISD::Suld3DI16Trap;
3319 case Intrinsic::nvvm_suld_3d_i32_trap:
3320 return NVPTXISD::Suld3DI32Trap;
3321 case Intrinsic::nvvm_suld_3d_i64_trap:
3322 return NVPTXISD::Suld3DI64Trap;
3323 case Intrinsic::nvvm_suld_3d_v2i8_trap:
3324 return NVPTXISD::Suld3DV2I8Trap;
3325 case Intrinsic::nvvm_suld_3d_v2i16_trap:
3326 return NVPTXISD::Suld3DV2I16Trap;
3327 case Intrinsic::nvvm_suld_3d_v2i32_trap:
3328 return NVPTXISD::Suld3DV2I32Trap;
3329 case Intrinsic::nvvm_suld_3d_v2i64_trap:
3330 return NVPTXISD::Suld3DV2I64Trap;
3331 case Intrinsic::nvvm_suld_3d_v4i8_trap:
3332 return NVPTXISD::Suld3DV4I8Trap;
3333 case Intrinsic::nvvm_suld_3d_v4i16_trap:
3334 return NVPTXISD::Suld3DV4I16Trap;
3335 case Intrinsic::nvvm_suld_3d_v4i32_trap:
3336 return NVPTXISD::Suld3DV4I32Trap;
3337 case Intrinsic::nvvm_suld_1d_i8_zero:
3338 return NVPTXISD::Suld1DI8Zero;
3339 case Intrinsic::nvvm_suld_1d_i16_zero:
3340 return NVPTXISD::Suld1DI16Zero;
3341 case Intrinsic::nvvm_suld_1d_i32_zero:
3342 return NVPTXISD::Suld1DI32Zero;
3343 case Intrinsic::nvvm_suld_1d_i64_zero:
3344 return NVPTXISD::Suld1DI64Zero;
3345 case Intrinsic::nvvm_suld_1d_v2i8_zero:
3346 return NVPTXISD::Suld1DV2I8Zero;
3347 case Intrinsic::nvvm_suld_1d_v2i16_zero:
3348 return NVPTXISD::Suld1DV2I16Zero;
3349 case Intrinsic::nvvm_suld_1d_v2i32_zero:
3350 return NVPTXISD::Suld1DV2I32Zero;
3351 case Intrinsic::nvvm_suld_1d_v2i64_zero:
3352 return NVPTXISD::Suld1DV2I64Zero;
3353 case Intrinsic::nvvm_suld_1d_v4i8_zero:
3354 return NVPTXISD::Suld1DV4I8Zero;
3355 case Intrinsic::nvvm_suld_1d_v4i16_zero:
3356 return NVPTXISD::Suld1DV4I16Zero;
3357 case Intrinsic::nvvm_suld_1d_v4i32_zero:
3358 return NVPTXISD::Suld1DV4I32Zero;
3359 case Intrinsic::nvvm_suld_1d_array_i8_zero:
3360 return NVPTXISD::Suld1DArrayI8Zero;
3361 case Intrinsic::nvvm_suld_1d_array_i16_zero:
3362 return NVPTXISD::Suld1DArrayI16Zero;
3363 case Intrinsic::nvvm_suld_1d_array_i32_zero:
3364 return NVPTXISD::Suld1DArrayI32Zero;
3365 case Intrinsic::nvvm_suld_1d_array_i64_zero:
3366 return NVPTXISD::Suld1DArrayI64Zero;
3367 case Intrinsic::nvvm_suld_1d_array_v2i8_zero:
3368 return NVPTXISD::Suld1DArrayV2I8Zero;
3369 case Intrinsic::nvvm_suld_1d_array_v2i16_zero:
3370 return NVPTXISD::Suld1DArrayV2I16Zero;
3371 case Intrinsic::nvvm_suld_1d_array_v2i32_zero:
3372 return NVPTXISD::Suld1DArrayV2I32Zero;
3373 case Intrinsic::nvvm_suld_1d_array_v2i64_zero:
3374 return NVPTXISD::Suld1DArrayV2I64Zero;
3375 case Intrinsic::nvvm_suld_1d_array_v4i8_zero:
3376 return NVPTXISD::Suld1DArrayV4I8Zero;
3377 case Intrinsic::nvvm_suld_1d_array_v4i16_zero:
3378 return NVPTXISD::Suld1DArrayV4I16Zero;
3379 case Intrinsic::nvvm_suld_1d_array_v4i32_zero:
3380 return NVPTXISD::Suld1DArrayV4I32Zero;
3381 case Intrinsic::nvvm_suld_2d_i8_zero:
3382 return NVPTXISD::Suld2DI8Zero;
3383 case Intrinsic::nvvm_suld_2d_i16_zero:
3384 return NVPTXISD::Suld2DI16Zero;
3385 case Intrinsic::nvvm_suld_2d_i32_zero:
3386 return NVPTXISD::Suld2DI32Zero;
3387 case Intrinsic::nvvm_suld_2d_i64_zero:
3388 return NVPTXISD::Suld2DI64Zero;
3389 case Intrinsic::nvvm_suld_2d_v2i8_zero:
3390 return NVPTXISD::Suld2DV2I8Zero;
3391 case Intrinsic::nvvm_suld_2d_v2i16_zero:
3392 return NVPTXISD::Suld2DV2I16Zero;
3393 case Intrinsic::nvvm_suld_2d_v2i32_zero:
3394 return NVPTXISD::Suld2DV2I32Zero;
3395 case Intrinsic::nvvm_suld_2d_v2i64_zero:
3396 return NVPTXISD::Suld2DV2I64Zero;
3397 case Intrinsic::nvvm_suld_2d_v4i8_zero:
3398 return NVPTXISD::Suld2DV4I8Zero;
3399 case Intrinsic::nvvm_suld_2d_v4i16_zero:
3400 return NVPTXISD::Suld2DV4I16Zero;
3401 case Intrinsic::nvvm_suld_2d_v4i32_zero:
3402 return NVPTXISD::Suld2DV4I32Zero;
3403 case Intrinsic::nvvm_suld_2d_array_i8_zero:
3404 return NVPTXISD::Suld2DArrayI8Zero;
3405 case Intrinsic::nvvm_suld_2d_array_i16_zero:
3406 return NVPTXISD::Suld2DArrayI16Zero;
3407 case Intrinsic::nvvm_suld_2d_array_i32_zero:
3408 return NVPTXISD::Suld2DArrayI32Zero;
3409 case Intrinsic::nvvm_suld_2d_array_i64_zero:
3410 return NVPTXISD::Suld2DArrayI64Zero;
3411 case Intrinsic::nvvm_suld_2d_array_v2i8_zero:
3412 return NVPTXISD::Suld2DArrayV2I8Zero;
3413 case Intrinsic::nvvm_suld_2d_array_v2i16_zero:
3414 return NVPTXISD::Suld2DArrayV2I16Zero;
3415 case Intrinsic::nvvm_suld_2d_array_v2i32_zero:
3416 return NVPTXISD::Suld2DArrayV2I32Zero;
3417 case Intrinsic::nvvm_suld_2d_array_v2i64_zero:
3418 return NVPTXISD::Suld2DArrayV2I64Zero;
3419 case Intrinsic::nvvm_suld_2d_array_v4i8_zero:
3420 return NVPTXISD::Suld2DArrayV4I8Zero;
3421 case Intrinsic::nvvm_suld_2d_array_v4i16_zero:
3422 return NVPTXISD::Suld2DArrayV4I16Zero;
3423 case Intrinsic::nvvm_suld_2d_array_v4i32_zero:
3424 return NVPTXISD::Suld2DArrayV4I32Zero;
3425 case Intrinsic::nvvm_suld_3d_i8_zero:
3426 return NVPTXISD::Suld3DI8Zero;
3427 case Intrinsic::nvvm_suld_3d_i16_zero:
3428 return NVPTXISD::Suld3DI16Zero;
3429 case Intrinsic::nvvm_suld_3d_i32_zero:
3430 return NVPTXISD::Suld3DI32Zero;
3431 case Intrinsic::nvvm_suld_3d_i64_zero:
3432 return NVPTXISD::Suld3DI64Zero;
3433 case Intrinsic::nvvm_suld_3d_v2i8_zero:
3434 return NVPTXISD::Suld3DV2I8Zero;
3435 case Intrinsic::nvvm_suld_3d_v2i16_zero:
3436 return NVPTXISD::Suld3DV2I16Zero;
3437 case Intrinsic::nvvm_suld_3d_v2i32_zero:
3438 return NVPTXISD::Suld3DV2I32Zero;
3439 case Intrinsic::nvvm_suld_3d_v2i64_zero:
3440 return NVPTXISD::Suld3DV2I64Zero;
3441 case Intrinsic::nvvm_suld_3d_v4i8_zero:
3442 return NVPTXISD::Suld3DV4I8Zero;
3443 case Intrinsic::nvvm_suld_3d_v4i16_zero:
3444 return NVPTXISD::Suld3DV4I16Zero;
3445 case Intrinsic::nvvm_suld_3d_v4i32_zero:
3446 return NVPTXISD::Suld3DV4I32Zero;
3447 }
3448 }
3449
3450 // llvm.ptx.memcpy.const and llvm.ptx.memmove.const need to be modeled as
3451 // TgtMemIntrinsic
3452 // because we need the information that is only available in the "Value" type
3453 // of destination
3454 // pointer. In particular, the address space information.
getTgtMemIntrinsic(IntrinsicInfo & Info,const CallInst & I,MachineFunction & MF,unsigned Intrinsic) const3455 bool NVPTXTargetLowering::getTgtMemIntrinsic(
3456 IntrinsicInfo &Info, const CallInst &I,
3457 MachineFunction &MF, unsigned Intrinsic) const {
3458 switch (Intrinsic) {
3459 default:
3460 return false;
3461 case Intrinsic::nvvm_match_all_sync_i32p:
3462 case Intrinsic::nvvm_match_all_sync_i64p:
3463 Info.opc = ISD::INTRINSIC_W_CHAIN;
3464 // memVT is bogus. These intrinsics have IntrInaccessibleMemOnly attribute
3465 // in order to model data exchange with other threads, but perform no real
3466 // memory accesses.
3467 Info.memVT = MVT::i1;
3468
3469 // Our result depends on both our and other thread's arguments.
3470 Info.flags = MachineMemOperand::MOLoad | MachineMemOperand::MOStore;
3471 return true;
3472 case Intrinsic::nvvm_wmma_m16n16k16_load_a_f16_col:
3473 case Intrinsic::nvvm_wmma_m16n16k16_load_a_f16_row:
3474 case Intrinsic::nvvm_wmma_m16n16k16_load_a_f16_col_stride:
3475 case Intrinsic::nvvm_wmma_m16n16k16_load_a_f16_row_stride:
3476 case Intrinsic::nvvm_wmma_m16n16k16_load_b_f16_col:
3477 case Intrinsic::nvvm_wmma_m16n16k16_load_b_f16_row:
3478 case Intrinsic::nvvm_wmma_m16n16k16_load_b_f16_col_stride:
3479 case Intrinsic::nvvm_wmma_m16n16k16_load_b_f16_row_stride:
3480 case Intrinsic::nvvm_wmma_m32n8k16_load_a_f16_col:
3481 case Intrinsic::nvvm_wmma_m32n8k16_load_a_f16_row:
3482 case Intrinsic::nvvm_wmma_m32n8k16_load_a_f16_col_stride:
3483 case Intrinsic::nvvm_wmma_m32n8k16_load_a_f16_row_stride:
3484 case Intrinsic::nvvm_wmma_m32n8k16_load_b_f16_col:
3485 case Intrinsic::nvvm_wmma_m32n8k16_load_b_f16_row:
3486 case Intrinsic::nvvm_wmma_m32n8k16_load_b_f16_col_stride:
3487 case Intrinsic::nvvm_wmma_m32n8k16_load_b_f16_row_stride:
3488 case Intrinsic::nvvm_wmma_m8n32k16_load_a_f16_col:
3489 case Intrinsic::nvvm_wmma_m8n32k16_load_a_f16_row:
3490 case Intrinsic::nvvm_wmma_m8n32k16_load_a_f16_col_stride:
3491 case Intrinsic::nvvm_wmma_m8n32k16_load_a_f16_row_stride:
3492 case Intrinsic::nvvm_wmma_m8n32k16_load_b_f16_col:
3493 case Intrinsic::nvvm_wmma_m8n32k16_load_b_f16_row:
3494 case Intrinsic::nvvm_wmma_m8n32k16_load_b_f16_col_stride:
3495 case Intrinsic::nvvm_wmma_m8n32k16_load_b_f16_row_stride: {
3496 Info.opc = ISD::INTRINSIC_W_CHAIN;
3497 Info.memVT = MVT::v8f16;
3498 Info.ptrVal = I.getArgOperand(0);
3499 Info.offset = 0;
3500 Info.flags = MachineMemOperand::MOLoad;
3501 Info.align = Align(16);
3502 return true;
3503 }
3504 case Intrinsic::nvvm_wmma_m16n16k16_load_a_s8_col:
3505 case Intrinsic::nvvm_wmma_m16n16k16_load_a_s8_col_stride:
3506 case Intrinsic::nvvm_wmma_m16n16k16_load_a_u8_col_stride:
3507 case Intrinsic::nvvm_wmma_m16n16k16_load_a_u8_col:
3508 case Intrinsic::nvvm_wmma_m16n16k16_load_a_s8_row:
3509 case Intrinsic::nvvm_wmma_m16n16k16_load_a_s8_row_stride:
3510 case Intrinsic::nvvm_wmma_m16n16k16_load_a_u8_row_stride:
3511 case Intrinsic::nvvm_wmma_m16n16k16_load_a_u8_row:
3512 case Intrinsic::nvvm_wmma_m16n16k16_load_b_s8_col:
3513 case Intrinsic::nvvm_wmma_m16n16k16_load_b_s8_col_stride:
3514 case Intrinsic::nvvm_wmma_m16n16k16_load_b_u8_col_stride:
3515 case Intrinsic::nvvm_wmma_m16n16k16_load_b_u8_col:
3516 case Intrinsic::nvvm_wmma_m16n16k16_load_b_s8_row:
3517 case Intrinsic::nvvm_wmma_m16n16k16_load_b_s8_row_stride:
3518 case Intrinsic::nvvm_wmma_m16n16k16_load_b_u8_row_stride:
3519 case Intrinsic::nvvm_wmma_m16n16k16_load_b_u8_row: {
3520 Info.opc = ISD::INTRINSIC_W_CHAIN;
3521 Info.memVT = MVT::v2i32;
3522 Info.ptrVal = I.getArgOperand(0);
3523 Info.offset = 0;
3524 Info.flags = MachineMemOperand::MOLoad;
3525 Info.align = Align(8);
3526 return true;
3527 }
3528
3529 case Intrinsic::nvvm_wmma_m32n8k16_load_a_s8_col:
3530 case Intrinsic::nvvm_wmma_m32n8k16_load_a_s8_col_stride:
3531 case Intrinsic::nvvm_wmma_m32n8k16_load_a_u8_col_stride:
3532 case Intrinsic::nvvm_wmma_m32n8k16_load_a_u8_col:
3533 case Intrinsic::nvvm_wmma_m32n8k16_load_a_s8_row:
3534 case Intrinsic::nvvm_wmma_m32n8k16_load_a_s8_row_stride:
3535 case Intrinsic::nvvm_wmma_m32n8k16_load_a_u8_row_stride:
3536 case Intrinsic::nvvm_wmma_m32n8k16_load_a_u8_row:
3537
3538 case Intrinsic::nvvm_wmma_m8n32k16_load_b_s8_col:
3539 case Intrinsic::nvvm_wmma_m8n32k16_load_b_s8_col_stride:
3540 case Intrinsic::nvvm_wmma_m8n32k16_load_b_u8_col_stride:
3541 case Intrinsic::nvvm_wmma_m8n32k16_load_b_u8_col:
3542 case Intrinsic::nvvm_wmma_m8n32k16_load_b_s8_row:
3543 case Intrinsic::nvvm_wmma_m8n32k16_load_b_s8_row_stride:
3544 case Intrinsic::nvvm_wmma_m8n32k16_load_b_u8_row_stride:
3545 case Intrinsic::nvvm_wmma_m8n32k16_load_b_u8_row: {
3546 Info.opc = ISD::INTRINSIC_W_CHAIN;
3547 Info.memVT = MVT::v4i32;
3548 Info.ptrVal = I.getArgOperand(0);
3549 Info.offset = 0;
3550 Info.flags = MachineMemOperand::MOLoad;
3551 Info.align = Align(16);
3552 return true;
3553 }
3554
3555 case Intrinsic::nvvm_wmma_m32n8k16_load_b_s8_col:
3556 case Intrinsic::nvvm_wmma_m32n8k16_load_b_s8_col_stride:
3557 case Intrinsic::nvvm_wmma_m32n8k16_load_b_u8_col_stride:
3558 case Intrinsic::nvvm_wmma_m32n8k16_load_b_u8_col:
3559 case Intrinsic::nvvm_wmma_m32n8k16_load_b_s8_row:
3560 case Intrinsic::nvvm_wmma_m32n8k16_load_b_s8_row_stride:
3561 case Intrinsic::nvvm_wmma_m32n8k16_load_b_u8_row_stride:
3562 case Intrinsic::nvvm_wmma_m32n8k16_load_b_u8_row:
3563
3564 case Intrinsic::nvvm_wmma_m8n32k16_load_a_s8_col:
3565 case Intrinsic::nvvm_wmma_m8n32k16_load_a_s8_col_stride:
3566 case Intrinsic::nvvm_wmma_m8n32k16_load_a_u8_col_stride:
3567 case Intrinsic::nvvm_wmma_m8n32k16_load_a_u8_col:
3568 case Intrinsic::nvvm_wmma_m8n32k16_load_a_s8_row:
3569 case Intrinsic::nvvm_wmma_m8n32k16_load_a_s8_row_stride:
3570 case Intrinsic::nvvm_wmma_m8n32k16_load_a_u8_row_stride:
3571 case Intrinsic::nvvm_wmma_m8n32k16_load_a_u8_row:
3572 case Intrinsic::nvvm_wmma_m8n8k128_load_a_b1_row:
3573 case Intrinsic::nvvm_wmma_m8n8k128_load_a_b1_row_stride:
3574 case Intrinsic::nvvm_wmma_m8n8k128_load_b_b1_col:
3575 case Intrinsic::nvvm_wmma_m8n8k128_load_b_b1_col_stride:
3576 case Intrinsic::nvvm_wmma_m8n8k32_load_a_s4_row:
3577 case Intrinsic::nvvm_wmma_m8n8k32_load_a_s4_row_stride:
3578 case Intrinsic::nvvm_wmma_m8n8k32_load_a_u4_row_stride:
3579 case Intrinsic::nvvm_wmma_m8n8k32_load_a_u4_row:
3580 case Intrinsic::nvvm_wmma_m8n8k32_load_b_s4_col:
3581 case Intrinsic::nvvm_wmma_m8n8k32_load_b_s4_col_stride:
3582 case Intrinsic::nvvm_wmma_m8n8k32_load_b_u4_col_stride:
3583 case Intrinsic::nvvm_wmma_m8n8k32_load_b_u4_col: {
3584 Info.opc = ISD::INTRINSIC_W_CHAIN;
3585 Info.memVT = MVT::i32;
3586 Info.ptrVal = I.getArgOperand(0);
3587 Info.offset = 0;
3588 Info.flags = MachineMemOperand::MOLoad;
3589 Info.align = Align(4);
3590 return true;
3591 }
3592
3593 case Intrinsic::nvvm_wmma_m16n16k16_load_c_f16_col:
3594 case Intrinsic::nvvm_wmma_m16n16k16_load_c_f16_row:
3595 case Intrinsic::nvvm_wmma_m16n16k16_load_c_f16_col_stride:
3596 case Intrinsic::nvvm_wmma_m16n16k16_load_c_f16_row_stride:
3597 case Intrinsic::nvvm_wmma_m32n8k16_load_c_f16_col:
3598 case Intrinsic::nvvm_wmma_m32n8k16_load_c_f16_row:
3599 case Intrinsic::nvvm_wmma_m32n8k16_load_c_f16_col_stride:
3600 case Intrinsic::nvvm_wmma_m32n8k16_load_c_f16_row_stride:
3601 case Intrinsic::nvvm_wmma_m8n32k16_load_c_f16_col:
3602 case Intrinsic::nvvm_wmma_m8n32k16_load_c_f16_row:
3603 case Intrinsic::nvvm_wmma_m8n32k16_load_c_f16_col_stride:
3604 case Intrinsic::nvvm_wmma_m8n32k16_load_c_f16_row_stride: {
3605 Info.opc = ISD::INTRINSIC_W_CHAIN;
3606 Info.memVT = MVT::v4f16;
3607 Info.ptrVal = I.getArgOperand(0);
3608 Info.offset = 0;
3609 Info.flags = MachineMemOperand::MOLoad;
3610 Info.align = Align(16);
3611 return true;
3612 }
3613
3614 case Intrinsic::nvvm_wmma_m16n16k16_load_c_f32_col:
3615 case Intrinsic::nvvm_wmma_m16n16k16_load_c_f32_row:
3616 case Intrinsic::nvvm_wmma_m16n16k16_load_c_f32_col_stride:
3617 case Intrinsic::nvvm_wmma_m16n16k16_load_c_f32_row_stride:
3618 case Intrinsic::nvvm_wmma_m32n8k16_load_c_f32_col:
3619 case Intrinsic::nvvm_wmma_m32n8k16_load_c_f32_row:
3620 case Intrinsic::nvvm_wmma_m32n8k16_load_c_f32_col_stride:
3621 case Intrinsic::nvvm_wmma_m32n8k16_load_c_f32_row_stride:
3622 case Intrinsic::nvvm_wmma_m8n32k16_load_c_f32_col:
3623 case Intrinsic::nvvm_wmma_m8n32k16_load_c_f32_row:
3624 case Intrinsic::nvvm_wmma_m8n32k16_load_c_f32_col_stride:
3625 case Intrinsic::nvvm_wmma_m8n32k16_load_c_f32_row_stride: {
3626 Info.opc = ISD::INTRINSIC_W_CHAIN;
3627 Info.memVT = MVT::v8f32;
3628 Info.ptrVal = I.getArgOperand(0);
3629 Info.offset = 0;
3630 Info.flags = MachineMemOperand::MOLoad;
3631 Info.align = Align(16);
3632 return true;
3633 }
3634
3635 case Intrinsic::nvvm_wmma_m16n16k16_load_c_s32_col:
3636 case Intrinsic::nvvm_wmma_m16n16k16_load_c_s32_col_stride:
3637 case Intrinsic::nvvm_wmma_m16n16k16_load_c_s32_row:
3638 case Intrinsic::nvvm_wmma_m16n16k16_load_c_s32_row_stride:
3639 case Intrinsic::nvvm_wmma_m32n8k16_load_c_s32_col:
3640 case Intrinsic::nvvm_wmma_m32n8k16_load_c_s32_col_stride:
3641 case Intrinsic::nvvm_wmma_m32n8k16_load_c_s32_row:
3642 case Intrinsic::nvvm_wmma_m32n8k16_load_c_s32_row_stride:
3643 case Intrinsic::nvvm_wmma_m8n32k16_load_c_s32_col:
3644 case Intrinsic::nvvm_wmma_m8n32k16_load_c_s32_col_stride:
3645 case Intrinsic::nvvm_wmma_m8n32k16_load_c_s32_row:
3646 case Intrinsic::nvvm_wmma_m8n32k16_load_c_s32_row_stride: {
3647 Info.opc = ISD::INTRINSIC_W_CHAIN;
3648 Info.memVT = MVT::v8i32;
3649 Info.ptrVal = I.getArgOperand(0);
3650 Info.offset = 0;
3651 Info.flags = MachineMemOperand::MOLoad;
3652 Info.align = Align(16);
3653 return true;
3654 }
3655
3656 case Intrinsic::nvvm_wmma_m8n8k128_load_c_s32_col:
3657 case Intrinsic::nvvm_wmma_m8n8k128_load_c_s32_col_stride:
3658 case Intrinsic::nvvm_wmma_m8n8k128_load_c_s32_row:
3659 case Intrinsic::nvvm_wmma_m8n8k128_load_c_s32_row_stride:
3660 case Intrinsic::nvvm_wmma_m8n8k32_load_c_s32_col:
3661 case Intrinsic::nvvm_wmma_m8n8k32_load_c_s32_col_stride:
3662 case Intrinsic::nvvm_wmma_m8n8k32_load_c_s32_row:
3663 case Intrinsic::nvvm_wmma_m8n8k32_load_c_s32_row_stride: {
3664 Info.opc = ISD::INTRINSIC_W_CHAIN;
3665 Info.memVT = MVT::v2i32;
3666 Info.ptrVal = I.getArgOperand(0);
3667 Info.offset = 0;
3668 Info.flags = MachineMemOperand::MOLoad;
3669 Info.align = Align(8);
3670 return true;
3671 }
3672
3673 case Intrinsic::nvvm_wmma_m16n16k16_store_d_f16_col:
3674 case Intrinsic::nvvm_wmma_m16n16k16_store_d_f16_row:
3675 case Intrinsic::nvvm_wmma_m16n16k16_store_d_f16_col_stride:
3676 case Intrinsic::nvvm_wmma_m16n16k16_store_d_f16_row_stride:
3677 case Intrinsic::nvvm_wmma_m32n8k16_store_d_f16_col:
3678 case Intrinsic::nvvm_wmma_m32n8k16_store_d_f16_row:
3679 case Intrinsic::nvvm_wmma_m32n8k16_store_d_f16_col_stride:
3680 case Intrinsic::nvvm_wmma_m32n8k16_store_d_f16_row_stride:
3681 case Intrinsic::nvvm_wmma_m8n32k16_store_d_f16_col:
3682 case Intrinsic::nvvm_wmma_m8n32k16_store_d_f16_row:
3683 case Intrinsic::nvvm_wmma_m8n32k16_store_d_f16_col_stride:
3684 case Intrinsic::nvvm_wmma_m8n32k16_store_d_f16_row_stride: {
3685 Info.opc = ISD::INTRINSIC_VOID;
3686 Info.memVT = MVT::v4f16;
3687 Info.ptrVal = I.getArgOperand(0);
3688 Info.offset = 0;
3689 Info.flags = MachineMemOperand::MOStore;
3690 Info.align = Align(16);
3691 return true;
3692 }
3693
3694 case Intrinsic::nvvm_wmma_m16n16k16_store_d_f32_col:
3695 case Intrinsic::nvvm_wmma_m16n16k16_store_d_f32_row:
3696 case Intrinsic::nvvm_wmma_m16n16k16_store_d_f32_col_stride:
3697 case Intrinsic::nvvm_wmma_m16n16k16_store_d_f32_row_stride:
3698 case Intrinsic::nvvm_wmma_m32n8k16_store_d_f32_col:
3699 case Intrinsic::nvvm_wmma_m32n8k16_store_d_f32_row:
3700 case Intrinsic::nvvm_wmma_m32n8k16_store_d_f32_col_stride:
3701 case Intrinsic::nvvm_wmma_m32n8k16_store_d_f32_row_stride:
3702 case Intrinsic::nvvm_wmma_m8n32k16_store_d_f32_col:
3703 case Intrinsic::nvvm_wmma_m8n32k16_store_d_f32_row:
3704 case Intrinsic::nvvm_wmma_m8n32k16_store_d_f32_col_stride:
3705 case Intrinsic::nvvm_wmma_m8n32k16_store_d_f32_row_stride: {
3706 Info.opc = ISD::INTRINSIC_VOID;
3707 Info.memVT = MVT::v8f32;
3708 Info.ptrVal = I.getArgOperand(0);
3709 Info.offset = 0;
3710 Info.flags = MachineMemOperand::MOStore;
3711 Info.align = Align(16);
3712 return true;
3713 }
3714
3715 case Intrinsic::nvvm_wmma_m16n16k16_store_d_s32_col:
3716 case Intrinsic::nvvm_wmma_m16n16k16_store_d_s32_col_stride:
3717 case Intrinsic::nvvm_wmma_m16n16k16_store_d_s32_row:
3718 case Intrinsic::nvvm_wmma_m16n16k16_store_d_s32_row_stride:
3719 case Intrinsic::nvvm_wmma_m32n8k16_store_d_s32_col:
3720 case Intrinsic::nvvm_wmma_m32n8k16_store_d_s32_col_stride:
3721 case Intrinsic::nvvm_wmma_m32n8k16_store_d_s32_row:
3722 case Intrinsic::nvvm_wmma_m32n8k16_store_d_s32_row_stride:
3723 case Intrinsic::nvvm_wmma_m8n32k16_store_d_s32_col:
3724 case Intrinsic::nvvm_wmma_m8n32k16_store_d_s32_col_stride:
3725 case Intrinsic::nvvm_wmma_m8n32k16_store_d_s32_row:
3726 case Intrinsic::nvvm_wmma_m8n32k16_store_d_s32_row_stride: {
3727 Info.opc = ISD::INTRINSIC_VOID;
3728 Info.memVT = MVT::v8i32;
3729 Info.ptrVal = I.getArgOperand(0);
3730 Info.offset = 0;
3731 Info.flags = MachineMemOperand::MOStore;
3732 Info.align = Align(16);
3733 return true;
3734 }
3735
3736 case Intrinsic::nvvm_wmma_m8n8k128_store_d_s32_col:
3737 case Intrinsic::nvvm_wmma_m8n8k128_store_d_s32_col_stride:
3738 case Intrinsic::nvvm_wmma_m8n8k128_store_d_s32_row:
3739 case Intrinsic::nvvm_wmma_m8n8k128_store_d_s32_row_stride:
3740 case Intrinsic::nvvm_wmma_m8n8k32_store_d_s32_col:
3741 case Intrinsic::nvvm_wmma_m8n8k32_store_d_s32_col_stride:
3742 case Intrinsic::nvvm_wmma_m8n8k32_store_d_s32_row:
3743 case Intrinsic::nvvm_wmma_m8n8k32_store_d_s32_row_stride: {
3744 Info.opc = ISD::INTRINSIC_VOID;
3745 Info.memVT = MVT::v2i32;
3746 Info.ptrVal = I.getArgOperand(0);
3747 Info.offset = 0;
3748 Info.flags = MachineMemOperand::MOStore;
3749 Info.align = Align(8);
3750 return true;
3751 }
3752
3753 case Intrinsic::nvvm_atomic_load_inc_32:
3754 case Intrinsic::nvvm_atomic_load_dec_32:
3755
3756 case Intrinsic::nvvm_atomic_add_gen_f_cta:
3757 case Intrinsic::nvvm_atomic_add_gen_f_sys:
3758 case Intrinsic::nvvm_atomic_add_gen_i_cta:
3759 case Intrinsic::nvvm_atomic_add_gen_i_sys:
3760 case Intrinsic::nvvm_atomic_and_gen_i_cta:
3761 case Intrinsic::nvvm_atomic_and_gen_i_sys:
3762 case Intrinsic::nvvm_atomic_cas_gen_i_cta:
3763 case Intrinsic::nvvm_atomic_cas_gen_i_sys:
3764 case Intrinsic::nvvm_atomic_dec_gen_i_cta:
3765 case Intrinsic::nvvm_atomic_dec_gen_i_sys:
3766 case Intrinsic::nvvm_atomic_inc_gen_i_cta:
3767 case Intrinsic::nvvm_atomic_inc_gen_i_sys:
3768 case Intrinsic::nvvm_atomic_max_gen_i_cta:
3769 case Intrinsic::nvvm_atomic_max_gen_i_sys:
3770 case Intrinsic::nvvm_atomic_min_gen_i_cta:
3771 case Intrinsic::nvvm_atomic_min_gen_i_sys:
3772 case Intrinsic::nvvm_atomic_or_gen_i_cta:
3773 case Intrinsic::nvvm_atomic_or_gen_i_sys:
3774 case Intrinsic::nvvm_atomic_exch_gen_i_cta:
3775 case Intrinsic::nvvm_atomic_exch_gen_i_sys:
3776 case Intrinsic::nvvm_atomic_xor_gen_i_cta:
3777 case Intrinsic::nvvm_atomic_xor_gen_i_sys: {
3778 auto &DL = I.getModule()->getDataLayout();
3779 Info.opc = ISD::INTRINSIC_W_CHAIN;
3780 Info.memVT = getValueType(DL, I.getType());
3781 Info.ptrVal = I.getArgOperand(0);
3782 Info.offset = 0;
3783 Info.flags = MachineMemOperand::MOLoad | MachineMemOperand::MOStore;
3784 Info.align.reset();
3785 return true;
3786 }
3787
3788 case Intrinsic::nvvm_ldu_global_i:
3789 case Intrinsic::nvvm_ldu_global_f:
3790 case Intrinsic::nvvm_ldu_global_p: {
3791 auto &DL = I.getModule()->getDataLayout();
3792 Info.opc = ISD::INTRINSIC_W_CHAIN;
3793 if (Intrinsic == Intrinsic::nvvm_ldu_global_i)
3794 Info.memVT = getValueType(DL, I.getType());
3795 else if(Intrinsic == Intrinsic::nvvm_ldu_global_p)
3796 Info.memVT = getPointerTy(DL);
3797 else
3798 Info.memVT = getValueType(DL, I.getType());
3799 Info.ptrVal = I.getArgOperand(0);
3800 Info.offset = 0;
3801 Info.flags = MachineMemOperand::MOLoad;
3802 Info.align =
3803 MaybeAlign(cast<ConstantInt>(I.getArgOperand(1))->getZExtValue());
3804
3805 return true;
3806 }
3807 case Intrinsic::nvvm_ldg_global_i:
3808 case Intrinsic::nvvm_ldg_global_f:
3809 case Intrinsic::nvvm_ldg_global_p: {
3810 auto &DL = I.getModule()->getDataLayout();
3811
3812 Info.opc = ISD::INTRINSIC_W_CHAIN;
3813 if (Intrinsic == Intrinsic::nvvm_ldg_global_i)
3814 Info.memVT = getValueType(DL, I.getType());
3815 else if(Intrinsic == Intrinsic::nvvm_ldg_global_p)
3816 Info.memVT = getPointerTy(DL);
3817 else
3818 Info.memVT = getValueType(DL, I.getType());
3819 Info.ptrVal = I.getArgOperand(0);
3820 Info.offset = 0;
3821 Info.flags = MachineMemOperand::MOLoad;
3822 Info.align =
3823 MaybeAlign(cast<ConstantInt>(I.getArgOperand(1))->getZExtValue());
3824
3825 return true;
3826 }
3827
3828 case Intrinsic::nvvm_tex_1d_v4f32_s32:
3829 case Intrinsic::nvvm_tex_1d_v4f32_f32:
3830 case Intrinsic::nvvm_tex_1d_level_v4f32_f32:
3831 case Intrinsic::nvvm_tex_1d_grad_v4f32_f32:
3832 case Intrinsic::nvvm_tex_1d_array_v4f32_s32:
3833 case Intrinsic::nvvm_tex_1d_array_v4f32_f32:
3834 case Intrinsic::nvvm_tex_1d_array_level_v4f32_f32:
3835 case Intrinsic::nvvm_tex_1d_array_grad_v4f32_f32:
3836 case Intrinsic::nvvm_tex_2d_v4f32_s32:
3837 case Intrinsic::nvvm_tex_2d_v4f32_f32:
3838 case Intrinsic::nvvm_tex_2d_level_v4f32_f32:
3839 case Intrinsic::nvvm_tex_2d_grad_v4f32_f32:
3840 case Intrinsic::nvvm_tex_2d_array_v4f32_s32:
3841 case Intrinsic::nvvm_tex_2d_array_v4f32_f32:
3842 case Intrinsic::nvvm_tex_2d_array_level_v4f32_f32:
3843 case Intrinsic::nvvm_tex_2d_array_grad_v4f32_f32:
3844 case Intrinsic::nvvm_tex_3d_v4f32_s32:
3845 case Intrinsic::nvvm_tex_3d_v4f32_f32:
3846 case Intrinsic::nvvm_tex_3d_level_v4f32_f32:
3847 case Intrinsic::nvvm_tex_3d_grad_v4f32_f32:
3848 case Intrinsic::nvvm_tex_cube_v4f32_f32:
3849 case Intrinsic::nvvm_tex_cube_level_v4f32_f32:
3850 case Intrinsic::nvvm_tex_cube_array_v4f32_f32:
3851 case Intrinsic::nvvm_tex_cube_array_level_v4f32_f32:
3852 case Intrinsic::nvvm_tld4_r_2d_v4f32_f32:
3853 case Intrinsic::nvvm_tld4_g_2d_v4f32_f32:
3854 case Intrinsic::nvvm_tld4_b_2d_v4f32_f32:
3855 case Intrinsic::nvvm_tld4_a_2d_v4f32_f32:
3856 case Intrinsic::nvvm_tex_unified_1d_v4f32_s32:
3857 case Intrinsic::nvvm_tex_unified_1d_v4f32_f32:
3858 case Intrinsic::nvvm_tex_unified_1d_level_v4f32_f32:
3859 case Intrinsic::nvvm_tex_unified_1d_grad_v4f32_f32:
3860 case Intrinsic::nvvm_tex_unified_1d_array_v4f32_s32:
3861 case Intrinsic::nvvm_tex_unified_1d_array_v4f32_f32:
3862 case Intrinsic::nvvm_tex_unified_1d_array_level_v4f32_f32:
3863 case Intrinsic::nvvm_tex_unified_1d_array_grad_v4f32_f32:
3864 case Intrinsic::nvvm_tex_unified_2d_v4f32_s32:
3865 case Intrinsic::nvvm_tex_unified_2d_v4f32_f32:
3866 case Intrinsic::nvvm_tex_unified_2d_level_v4f32_f32:
3867 case Intrinsic::nvvm_tex_unified_2d_grad_v4f32_f32:
3868 case Intrinsic::nvvm_tex_unified_2d_array_v4f32_s32:
3869 case Intrinsic::nvvm_tex_unified_2d_array_v4f32_f32:
3870 case Intrinsic::nvvm_tex_unified_2d_array_level_v4f32_f32:
3871 case Intrinsic::nvvm_tex_unified_2d_array_grad_v4f32_f32:
3872 case Intrinsic::nvvm_tex_unified_3d_v4f32_s32:
3873 case Intrinsic::nvvm_tex_unified_3d_v4f32_f32:
3874 case Intrinsic::nvvm_tex_unified_3d_level_v4f32_f32:
3875 case Intrinsic::nvvm_tex_unified_3d_grad_v4f32_f32:
3876 case Intrinsic::nvvm_tex_unified_cube_v4f32_f32:
3877 case Intrinsic::nvvm_tex_unified_cube_level_v4f32_f32:
3878 case Intrinsic::nvvm_tex_unified_cube_array_v4f32_f32:
3879 case Intrinsic::nvvm_tex_unified_cube_array_level_v4f32_f32:
3880 case Intrinsic::nvvm_tld4_unified_r_2d_v4f32_f32:
3881 case Intrinsic::nvvm_tld4_unified_g_2d_v4f32_f32:
3882 case Intrinsic::nvvm_tld4_unified_b_2d_v4f32_f32:
3883 case Intrinsic::nvvm_tld4_unified_a_2d_v4f32_f32:
3884 Info.opc = getOpcForTextureInstr(Intrinsic);
3885 Info.memVT = MVT::v4f32;
3886 Info.ptrVal = nullptr;
3887 Info.offset = 0;
3888 Info.flags = MachineMemOperand::MOLoad;
3889 Info.align = Align(16);
3890 return true;
3891
3892 case Intrinsic::nvvm_tex_1d_v4s32_s32:
3893 case Intrinsic::nvvm_tex_1d_v4s32_f32:
3894 case Intrinsic::nvvm_tex_1d_level_v4s32_f32:
3895 case Intrinsic::nvvm_tex_1d_grad_v4s32_f32:
3896 case Intrinsic::nvvm_tex_1d_array_v4s32_s32:
3897 case Intrinsic::nvvm_tex_1d_array_v4s32_f32:
3898 case Intrinsic::nvvm_tex_1d_array_level_v4s32_f32:
3899 case Intrinsic::nvvm_tex_1d_array_grad_v4s32_f32:
3900 case Intrinsic::nvvm_tex_2d_v4s32_s32:
3901 case Intrinsic::nvvm_tex_2d_v4s32_f32:
3902 case Intrinsic::nvvm_tex_2d_level_v4s32_f32:
3903 case Intrinsic::nvvm_tex_2d_grad_v4s32_f32:
3904 case Intrinsic::nvvm_tex_2d_array_v4s32_s32:
3905 case Intrinsic::nvvm_tex_2d_array_v4s32_f32:
3906 case Intrinsic::nvvm_tex_2d_array_level_v4s32_f32:
3907 case Intrinsic::nvvm_tex_2d_array_grad_v4s32_f32:
3908 case Intrinsic::nvvm_tex_3d_v4s32_s32:
3909 case Intrinsic::nvvm_tex_3d_v4s32_f32:
3910 case Intrinsic::nvvm_tex_3d_level_v4s32_f32:
3911 case Intrinsic::nvvm_tex_3d_grad_v4s32_f32:
3912 case Intrinsic::nvvm_tex_cube_v4s32_f32:
3913 case Intrinsic::nvvm_tex_cube_level_v4s32_f32:
3914 case Intrinsic::nvvm_tex_cube_array_v4s32_f32:
3915 case Intrinsic::nvvm_tex_cube_array_level_v4s32_f32:
3916 case Intrinsic::nvvm_tex_cube_v4u32_f32:
3917 case Intrinsic::nvvm_tex_cube_level_v4u32_f32:
3918 case Intrinsic::nvvm_tex_cube_array_v4u32_f32:
3919 case Intrinsic::nvvm_tex_cube_array_level_v4u32_f32:
3920 case Intrinsic::nvvm_tex_1d_v4u32_s32:
3921 case Intrinsic::nvvm_tex_1d_v4u32_f32:
3922 case Intrinsic::nvvm_tex_1d_level_v4u32_f32:
3923 case Intrinsic::nvvm_tex_1d_grad_v4u32_f32:
3924 case Intrinsic::nvvm_tex_1d_array_v4u32_s32:
3925 case Intrinsic::nvvm_tex_1d_array_v4u32_f32:
3926 case Intrinsic::nvvm_tex_1d_array_level_v4u32_f32:
3927 case Intrinsic::nvvm_tex_1d_array_grad_v4u32_f32:
3928 case Intrinsic::nvvm_tex_2d_v4u32_s32:
3929 case Intrinsic::nvvm_tex_2d_v4u32_f32:
3930 case Intrinsic::nvvm_tex_2d_level_v4u32_f32:
3931 case Intrinsic::nvvm_tex_2d_grad_v4u32_f32:
3932 case Intrinsic::nvvm_tex_2d_array_v4u32_s32:
3933 case Intrinsic::nvvm_tex_2d_array_v4u32_f32:
3934 case Intrinsic::nvvm_tex_2d_array_level_v4u32_f32:
3935 case Intrinsic::nvvm_tex_2d_array_grad_v4u32_f32:
3936 case Intrinsic::nvvm_tex_3d_v4u32_s32:
3937 case Intrinsic::nvvm_tex_3d_v4u32_f32:
3938 case Intrinsic::nvvm_tex_3d_level_v4u32_f32:
3939 case Intrinsic::nvvm_tex_3d_grad_v4u32_f32:
3940 case Intrinsic::nvvm_tld4_r_2d_v4s32_f32:
3941 case Intrinsic::nvvm_tld4_g_2d_v4s32_f32:
3942 case Intrinsic::nvvm_tld4_b_2d_v4s32_f32:
3943 case Intrinsic::nvvm_tld4_a_2d_v4s32_f32:
3944 case Intrinsic::nvvm_tld4_r_2d_v4u32_f32:
3945 case Intrinsic::nvvm_tld4_g_2d_v4u32_f32:
3946 case Intrinsic::nvvm_tld4_b_2d_v4u32_f32:
3947 case Intrinsic::nvvm_tld4_a_2d_v4u32_f32:
3948 case Intrinsic::nvvm_tex_unified_1d_v4s32_s32:
3949 case Intrinsic::nvvm_tex_unified_1d_v4s32_f32:
3950 case Intrinsic::nvvm_tex_unified_1d_level_v4s32_f32:
3951 case Intrinsic::nvvm_tex_unified_1d_grad_v4s32_f32:
3952 case Intrinsic::nvvm_tex_unified_1d_array_v4s32_s32:
3953 case Intrinsic::nvvm_tex_unified_1d_array_v4s32_f32:
3954 case Intrinsic::nvvm_tex_unified_1d_array_level_v4s32_f32:
3955 case Intrinsic::nvvm_tex_unified_1d_array_grad_v4s32_f32:
3956 case Intrinsic::nvvm_tex_unified_2d_v4s32_s32:
3957 case Intrinsic::nvvm_tex_unified_2d_v4s32_f32:
3958 case Intrinsic::nvvm_tex_unified_2d_level_v4s32_f32:
3959 case Intrinsic::nvvm_tex_unified_2d_grad_v4s32_f32:
3960 case Intrinsic::nvvm_tex_unified_2d_array_v4s32_s32:
3961 case Intrinsic::nvvm_tex_unified_2d_array_v4s32_f32:
3962 case Intrinsic::nvvm_tex_unified_2d_array_level_v4s32_f32:
3963 case Intrinsic::nvvm_tex_unified_2d_array_grad_v4s32_f32:
3964 case Intrinsic::nvvm_tex_unified_3d_v4s32_s32:
3965 case Intrinsic::nvvm_tex_unified_3d_v4s32_f32:
3966 case Intrinsic::nvvm_tex_unified_3d_level_v4s32_f32:
3967 case Intrinsic::nvvm_tex_unified_3d_grad_v4s32_f32:
3968 case Intrinsic::nvvm_tex_unified_1d_v4u32_s32:
3969 case Intrinsic::nvvm_tex_unified_1d_v4u32_f32:
3970 case Intrinsic::nvvm_tex_unified_1d_level_v4u32_f32:
3971 case Intrinsic::nvvm_tex_unified_1d_grad_v4u32_f32:
3972 case Intrinsic::nvvm_tex_unified_1d_array_v4u32_s32:
3973 case Intrinsic::nvvm_tex_unified_1d_array_v4u32_f32:
3974 case Intrinsic::nvvm_tex_unified_1d_array_level_v4u32_f32:
3975 case Intrinsic::nvvm_tex_unified_1d_array_grad_v4u32_f32:
3976 case Intrinsic::nvvm_tex_unified_2d_v4u32_s32:
3977 case Intrinsic::nvvm_tex_unified_2d_v4u32_f32:
3978 case Intrinsic::nvvm_tex_unified_2d_level_v4u32_f32:
3979 case Intrinsic::nvvm_tex_unified_2d_grad_v4u32_f32:
3980 case Intrinsic::nvvm_tex_unified_2d_array_v4u32_s32:
3981 case Intrinsic::nvvm_tex_unified_2d_array_v4u32_f32:
3982 case Intrinsic::nvvm_tex_unified_2d_array_level_v4u32_f32:
3983 case Intrinsic::nvvm_tex_unified_2d_array_grad_v4u32_f32:
3984 case Intrinsic::nvvm_tex_unified_3d_v4u32_s32:
3985 case Intrinsic::nvvm_tex_unified_3d_v4u32_f32:
3986 case Intrinsic::nvvm_tex_unified_3d_level_v4u32_f32:
3987 case Intrinsic::nvvm_tex_unified_3d_grad_v4u32_f32:
3988 case Intrinsic::nvvm_tex_unified_cube_v4s32_f32:
3989 case Intrinsic::nvvm_tex_unified_cube_level_v4s32_f32:
3990 case Intrinsic::nvvm_tex_unified_cube_array_v4s32_f32:
3991 case Intrinsic::nvvm_tex_unified_cube_array_level_v4s32_f32:
3992 case Intrinsic::nvvm_tex_unified_cube_v4u32_f32:
3993 case Intrinsic::nvvm_tex_unified_cube_level_v4u32_f32:
3994 case Intrinsic::nvvm_tex_unified_cube_array_v4u32_f32:
3995 case Intrinsic::nvvm_tex_unified_cube_array_level_v4u32_f32:
3996 case Intrinsic::nvvm_tld4_unified_r_2d_v4s32_f32:
3997 case Intrinsic::nvvm_tld4_unified_g_2d_v4s32_f32:
3998 case Intrinsic::nvvm_tld4_unified_b_2d_v4s32_f32:
3999 case Intrinsic::nvvm_tld4_unified_a_2d_v4s32_f32:
4000 case Intrinsic::nvvm_tld4_unified_r_2d_v4u32_f32:
4001 case Intrinsic::nvvm_tld4_unified_g_2d_v4u32_f32:
4002 case Intrinsic::nvvm_tld4_unified_b_2d_v4u32_f32:
4003 case Intrinsic::nvvm_tld4_unified_a_2d_v4u32_f32:
4004 Info.opc = getOpcForTextureInstr(Intrinsic);
4005 Info.memVT = MVT::v4i32;
4006 Info.ptrVal = nullptr;
4007 Info.offset = 0;
4008 Info.flags = MachineMemOperand::MOLoad;
4009 Info.align = Align(16);
4010 return true;
4011
4012 case Intrinsic::nvvm_suld_1d_i8_clamp:
4013 case Intrinsic::nvvm_suld_1d_v2i8_clamp:
4014 case Intrinsic::nvvm_suld_1d_v4i8_clamp:
4015 case Intrinsic::nvvm_suld_1d_array_i8_clamp:
4016 case Intrinsic::nvvm_suld_1d_array_v2i8_clamp:
4017 case Intrinsic::nvvm_suld_1d_array_v4i8_clamp:
4018 case Intrinsic::nvvm_suld_2d_i8_clamp:
4019 case Intrinsic::nvvm_suld_2d_v2i8_clamp:
4020 case Intrinsic::nvvm_suld_2d_v4i8_clamp:
4021 case Intrinsic::nvvm_suld_2d_array_i8_clamp:
4022 case Intrinsic::nvvm_suld_2d_array_v2i8_clamp:
4023 case Intrinsic::nvvm_suld_2d_array_v4i8_clamp:
4024 case Intrinsic::nvvm_suld_3d_i8_clamp:
4025 case Intrinsic::nvvm_suld_3d_v2i8_clamp:
4026 case Intrinsic::nvvm_suld_3d_v4i8_clamp:
4027 case Intrinsic::nvvm_suld_1d_i8_trap:
4028 case Intrinsic::nvvm_suld_1d_v2i8_trap:
4029 case Intrinsic::nvvm_suld_1d_v4i8_trap:
4030 case Intrinsic::nvvm_suld_1d_array_i8_trap:
4031 case Intrinsic::nvvm_suld_1d_array_v2i8_trap:
4032 case Intrinsic::nvvm_suld_1d_array_v4i8_trap:
4033 case Intrinsic::nvvm_suld_2d_i8_trap:
4034 case Intrinsic::nvvm_suld_2d_v2i8_trap:
4035 case Intrinsic::nvvm_suld_2d_v4i8_trap:
4036 case Intrinsic::nvvm_suld_2d_array_i8_trap:
4037 case Intrinsic::nvvm_suld_2d_array_v2i8_trap:
4038 case Intrinsic::nvvm_suld_2d_array_v4i8_trap:
4039 case Intrinsic::nvvm_suld_3d_i8_trap:
4040 case Intrinsic::nvvm_suld_3d_v2i8_trap:
4041 case Intrinsic::nvvm_suld_3d_v4i8_trap:
4042 case Intrinsic::nvvm_suld_1d_i8_zero:
4043 case Intrinsic::nvvm_suld_1d_v2i8_zero:
4044 case Intrinsic::nvvm_suld_1d_v4i8_zero:
4045 case Intrinsic::nvvm_suld_1d_array_i8_zero:
4046 case Intrinsic::nvvm_suld_1d_array_v2i8_zero:
4047 case Intrinsic::nvvm_suld_1d_array_v4i8_zero:
4048 case Intrinsic::nvvm_suld_2d_i8_zero:
4049 case Intrinsic::nvvm_suld_2d_v2i8_zero:
4050 case Intrinsic::nvvm_suld_2d_v4i8_zero:
4051 case Intrinsic::nvvm_suld_2d_array_i8_zero:
4052 case Intrinsic::nvvm_suld_2d_array_v2i8_zero:
4053 case Intrinsic::nvvm_suld_2d_array_v4i8_zero:
4054 case Intrinsic::nvvm_suld_3d_i8_zero:
4055 case Intrinsic::nvvm_suld_3d_v2i8_zero:
4056 case Intrinsic::nvvm_suld_3d_v4i8_zero:
4057 Info.opc = getOpcForSurfaceInstr(Intrinsic);
4058 Info.memVT = MVT::i8;
4059 Info.ptrVal = nullptr;
4060 Info.offset = 0;
4061 Info.flags = MachineMemOperand::MOLoad;
4062 Info.align = Align(16);
4063 return true;
4064
4065 case Intrinsic::nvvm_suld_1d_i16_clamp:
4066 case Intrinsic::nvvm_suld_1d_v2i16_clamp:
4067 case Intrinsic::nvvm_suld_1d_v4i16_clamp:
4068 case Intrinsic::nvvm_suld_1d_array_i16_clamp:
4069 case Intrinsic::nvvm_suld_1d_array_v2i16_clamp:
4070 case Intrinsic::nvvm_suld_1d_array_v4i16_clamp:
4071 case Intrinsic::nvvm_suld_2d_i16_clamp:
4072 case Intrinsic::nvvm_suld_2d_v2i16_clamp:
4073 case Intrinsic::nvvm_suld_2d_v4i16_clamp:
4074 case Intrinsic::nvvm_suld_2d_array_i16_clamp:
4075 case Intrinsic::nvvm_suld_2d_array_v2i16_clamp:
4076 case Intrinsic::nvvm_suld_2d_array_v4i16_clamp:
4077 case Intrinsic::nvvm_suld_3d_i16_clamp:
4078 case Intrinsic::nvvm_suld_3d_v2i16_clamp:
4079 case Intrinsic::nvvm_suld_3d_v4i16_clamp:
4080 case Intrinsic::nvvm_suld_1d_i16_trap:
4081 case Intrinsic::nvvm_suld_1d_v2i16_trap:
4082 case Intrinsic::nvvm_suld_1d_v4i16_trap:
4083 case Intrinsic::nvvm_suld_1d_array_i16_trap:
4084 case Intrinsic::nvvm_suld_1d_array_v2i16_trap:
4085 case Intrinsic::nvvm_suld_1d_array_v4i16_trap:
4086 case Intrinsic::nvvm_suld_2d_i16_trap:
4087 case Intrinsic::nvvm_suld_2d_v2i16_trap:
4088 case Intrinsic::nvvm_suld_2d_v4i16_trap:
4089 case Intrinsic::nvvm_suld_2d_array_i16_trap:
4090 case Intrinsic::nvvm_suld_2d_array_v2i16_trap:
4091 case Intrinsic::nvvm_suld_2d_array_v4i16_trap:
4092 case Intrinsic::nvvm_suld_3d_i16_trap:
4093 case Intrinsic::nvvm_suld_3d_v2i16_trap:
4094 case Intrinsic::nvvm_suld_3d_v4i16_trap:
4095 case Intrinsic::nvvm_suld_1d_i16_zero:
4096 case Intrinsic::nvvm_suld_1d_v2i16_zero:
4097 case Intrinsic::nvvm_suld_1d_v4i16_zero:
4098 case Intrinsic::nvvm_suld_1d_array_i16_zero:
4099 case Intrinsic::nvvm_suld_1d_array_v2i16_zero:
4100 case Intrinsic::nvvm_suld_1d_array_v4i16_zero:
4101 case Intrinsic::nvvm_suld_2d_i16_zero:
4102 case Intrinsic::nvvm_suld_2d_v2i16_zero:
4103 case Intrinsic::nvvm_suld_2d_v4i16_zero:
4104 case Intrinsic::nvvm_suld_2d_array_i16_zero:
4105 case Intrinsic::nvvm_suld_2d_array_v2i16_zero:
4106 case Intrinsic::nvvm_suld_2d_array_v4i16_zero:
4107 case Intrinsic::nvvm_suld_3d_i16_zero:
4108 case Intrinsic::nvvm_suld_3d_v2i16_zero:
4109 case Intrinsic::nvvm_suld_3d_v4i16_zero:
4110 Info.opc = getOpcForSurfaceInstr(Intrinsic);
4111 Info.memVT = MVT::i16;
4112 Info.ptrVal = nullptr;
4113 Info.offset = 0;
4114 Info.flags = MachineMemOperand::MOLoad;
4115 Info.align = Align(16);
4116 return true;
4117
4118 case Intrinsic::nvvm_suld_1d_i32_clamp:
4119 case Intrinsic::nvvm_suld_1d_v2i32_clamp:
4120 case Intrinsic::nvvm_suld_1d_v4i32_clamp:
4121 case Intrinsic::nvvm_suld_1d_array_i32_clamp:
4122 case Intrinsic::nvvm_suld_1d_array_v2i32_clamp:
4123 case Intrinsic::nvvm_suld_1d_array_v4i32_clamp:
4124 case Intrinsic::nvvm_suld_2d_i32_clamp:
4125 case Intrinsic::nvvm_suld_2d_v2i32_clamp:
4126 case Intrinsic::nvvm_suld_2d_v4i32_clamp:
4127 case Intrinsic::nvvm_suld_2d_array_i32_clamp:
4128 case Intrinsic::nvvm_suld_2d_array_v2i32_clamp:
4129 case Intrinsic::nvvm_suld_2d_array_v4i32_clamp:
4130 case Intrinsic::nvvm_suld_3d_i32_clamp:
4131 case Intrinsic::nvvm_suld_3d_v2i32_clamp:
4132 case Intrinsic::nvvm_suld_3d_v4i32_clamp:
4133 case Intrinsic::nvvm_suld_1d_i32_trap:
4134 case Intrinsic::nvvm_suld_1d_v2i32_trap:
4135 case Intrinsic::nvvm_suld_1d_v4i32_trap:
4136 case Intrinsic::nvvm_suld_1d_array_i32_trap:
4137 case Intrinsic::nvvm_suld_1d_array_v2i32_trap:
4138 case Intrinsic::nvvm_suld_1d_array_v4i32_trap:
4139 case Intrinsic::nvvm_suld_2d_i32_trap:
4140 case Intrinsic::nvvm_suld_2d_v2i32_trap:
4141 case Intrinsic::nvvm_suld_2d_v4i32_trap:
4142 case Intrinsic::nvvm_suld_2d_array_i32_trap:
4143 case Intrinsic::nvvm_suld_2d_array_v2i32_trap:
4144 case Intrinsic::nvvm_suld_2d_array_v4i32_trap:
4145 case Intrinsic::nvvm_suld_3d_i32_trap:
4146 case Intrinsic::nvvm_suld_3d_v2i32_trap:
4147 case Intrinsic::nvvm_suld_3d_v4i32_trap:
4148 case Intrinsic::nvvm_suld_1d_i32_zero:
4149 case Intrinsic::nvvm_suld_1d_v2i32_zero:
4150 case Intrinsic::nvvm_suld_1d_v4i32_zero:
4151 case Intrinsic::nvvm_suld_1d_array_i32_zero:
4152 case Intrinsic::nvvm_suld_1d_array_v2i32_zero:
4153 case Intrinsic::nvvm_suld_1d_array_v4i32_zero:
4154 case Intrinsic::nvvm_suld_2d_i32_zero:
4155 case Intrinsic::nvvm_suld_2d_v2i32_zero:
4156 case Intrinsic::nvvm_suld_2d_v4i32_zero:
4157 case Intrinsic::nvvm_suld_2d_array_i32_zero:
4158 case Intrinsic::nvvm_suld_2d_array_v2i32_zero:
4159 case Intrinsic::nvvm_suld_2d_array_v4i32_zero:
4160 case Intrinsic::nvvm_suld_3d_i32_zero:
4161 case Intrinsic::nvvm_suld_3d_v2i32_zero:
4162 case Intrinsic::nvvm_suld_3d_v4i32_zero:
4163 Info.opc = getOpcForSurfaceInstr(Intrinsic);
4164 Info.memVT = MVT::i32;
4165 Info.ptrVal = nullptr;
4166 Info.offset = 0;
4167 Info.flags = MachineMemOperand::MOLoad;
4168 Info.align = Align(16);
4169 return true;
4170
4171 case Intrinsic::nvvm_suld_1d_i64_clamp:
4172 case Intrinsic::nvvm_suld_1d_v2i64_clamp:
4173 case Intrinsic::nvvm_suld_1d_array_i64_clamp:
4174 case Intrinsic::nvvm_suld_1d_array_v2i64_clamp:
4175 case Intrinsic::nvvm_suld_2d_i64_clamp:
4176 case Intrinsic::nvvm_suld_2d_v2i64_clamp:
4177 case Intrinsic::nvvm_suld_2d_array_i64_clamp:
4178 case Intrinsic::nvvm_suld_2d_array_v2i64_clamp:
4179 case Intrinsic::nvvm_suld_3d_i64_clamp:
4180 case Intrinsic::nvvm_suld_3d_v2i64_clamp:
4181 case Intrinsic::nvvm_suld_1d_i64_trap:
4182 case Intrinsic::nvvm_suld_1d_v2i64_trap:
4183 case Intrinsic::nvvm_suld_1d_array_i64_trap:
4184 case Intrinsic::nvvm_suld_1d_array_v2i64_trap:
4185 case Intrinsic::nvvm_suld_2d_i64_trap:
4186 case Intrinsic::nvvm_suld_2d_v2i64_trap:
4187 case Intrinsic::nvvm_suld_2d_array_i64_trap:
4188 case Intrinsic::nvvm_suld_2d_array_v2i64_trap:
4189 case Intrinsic::nvvm_suld_3d_i64_trap:
4190 case Intrinsic::nvvm_suld_3d_v2i64_trap:
4191 case Intrinsic::nvvm_suld_1d_i64_zero:
4192 case Intrinsic::nvvm_suld_1d_v2i64_zero:
4193 case Intrinsic::nvvm_suld_1d_array_i64_zero:
4194 case Intrinsic::nvvm_suld_1d_array_v2i64_zero:
4195 case Intrinsic::nvvm_suld_2d_i64_zero:
4196 case Intrinsic::nvvm_suld_2d_v2i64_zero:
4197 case Intrinsic::nvvm_suld_2d_array_i64_zero:
4198 case Intrinsic::nvvm_suld_2d_array_v2i64_zero:
4199 case Intrinsic::nvvm_suld_3d_i64_zero:
4200 case Intrinsic::nvvm_suld_3d_v2i64_zero:
4201 Info.opc = getOpcForSurfaceInstr(Intrinsic);
4202 Info.memVT = MVT::i64;
4203 Info.ptrVal = nullptr;
4204 Info.offset = 0;
4205 Info.flags = MachineMemOperand::MOLoad;
4206 Info.align = Align(16);
4207 return true;
4208 }
4209 return false;
4210 }
4211
4212 /// isLegalAddressingMode - Return true if the addressing mode represented
4213 /// by AM is legal for this target, for a load/store of the specified type.
4214 /// Used to guide target specific optimizations, like loop strength reduction
4215 /// (LoopStrengthReduce.cpp) and memory optimization for address mode
4216 /// (CodeGenPrepare.cpp)
isLegalAddressingMode(const DataLayout & DL,const AddrMode & AM,Type * Ty,unsigned AS,Instruction * I) const4217 bool NVPTXTargetLowering::isLegalAddressingMode(const DataLayout &DL,
4218 const AddrMode &AM, Type *Ty,
4219 unsigned AS, Instruction *I) const {
4220 // AddrMode - This represents an addressing mode of:
4221 // BaseGV + BaseOffs + BaseReg + Scale*ScaleReg
4222 //
4223 // The legal address modes are
4224 // - [avar]
4225 // - [areg]
4226 // - [areg+immoff]
4227 // - [immAddr]
4228
4229 if (AM.BaseGV) {
4230 return !AM.BaseOffs && !AM.HasBaseReg && !AM.Scale;
4231 }
4232
4233 switch (AM.Scale) {
4234 case 0: // "r", "r+i" or "i" is allowed
4235 break;
4236 case 1:
4237 if (AM.HasBaseReg) // "r+r+i" or "r+r" is not allowed.
4238 return false;
4239 // Otherwise we have r+i.
4240 break;
4241 default:
4242 // No scale > 1 is allowed
4243 return false;
4244 }
4245 return true;
4246 }
4247
4248 //===----------------------------------------------------------------------===//
4249 // NVPTX Inline Assembly Support
4250 //===----------------------------------------------------------------------===//
4251
4252 /// getConstraintType - Given a constraint letter, return the type of
4253 /// constraint it is for this target.
4254 NVPTXTargetLowering::ConstraintType
getConstraintType(StringRef Constraint) const4255 NVPTXTargetLowering::getConstraintType(StringRef Constraint) const {
4256 if (Constraint.size() == 1) {
4257 switch (Constraint[0]) {
4258 default:
4259 break;
4260 case 'b':
4261 case 'r':
4262 case 'h':
4263 case 'c':
4264 case 'l':
4265 case 'f':
4266 case 'd':
4267 case '0':
4268 case 'N':
4269 return C_RegisterClass;
4270 }
4271 }
4272 return TargetLowering::getConstraintType(Constraint);
4273 }
4274
4275 std::pair<unsigned, const TargetRegisterClass *>
getRegForInlineAsmConstraint(const TargetRegisterInfo * TRI,StringRef Constraint,MVT VT) const4276 NVPTXTargetLowering::getRegForInlineAsmConstraint(const TargetRegisterInfo *TRI,
4277 StringRef Constraint,
4278 MVT VT) const {
4279 if (Constraint.size() == 1) {
4280 switch (Constraint[0]) {
4281 case 'b':
4282 return std::make_pair(0U, &NVPTX::Int1RegsRegClass);
4283 case 'c':
4284 return std::make_pair(0U, &NVPTX::Int16RegsRegClass);
4285 case 'h':
4286 return std::make_pair(0U, &NVPTX::Int16RegsRegClass);
4287 case 'r':
4288 return std::make_pair(0U, &NVPTX::Int32RegsRegClass);
4289 case 'l':
4290 case 'N':
4291 return std::make_pair(0U, &NVPTX::Int64RegsRegClass);
4292 case 'f':
4293 return std::make_pair(0U, &NVPTX::Float32RegsRegClass);
4294 case 'd':
4295 return std::make_pair(0U, &NVPTX::Float64RegsRegClass);
4296 }
4297 }
4298 return TargetLowering::getRegForInlineAsmConstraint(TRI, Constraint, VT);
4299 }
4300
4301 //===----------------------------------------------------------------------===//
4302 // NVPTX DAG Combining
4303 //===----------------------------------------------------------------------===//
4304
allowFMA(MachineFunction & MF,CodeGenOpt::Level OptLevel) const4305 bool NVPTXTargetLowering::allowFMA(MachineFunction &MF,
4306 CodeGenOpt::Level OptLevel) const {
4307 // Always honor command-line argument
4308 if (FMAContractLevelOpt.getNumOccurrences() > 0)
4309 return FMAContractLevelOpt > 0;
4310
4311 // Do not contract if we're not optimizing the code.
4312 if (OptLevel == 0)
4313 return false;
4314
4315 // Honor TargetOptions flags that explicitly say fusion is okay.
4316 if (MF.getTarget().Options.AllowFPOpFusion == FPOpFusion::Fast)
4317 return true;
4318
4319 return allowUnsafeFPMath(MF);
4320 }
4321
allowUnsafeFPMath(MachineFunction & MF) const4322 bool NVPTXTargetLowering::allowUnsafeFPMath(MachineFunction &MF) const {
4323 // Honor TargetOptions flags that explicitly say unsafe math is okay.
4324 if (MF.getTarget().Options.UnsafeFPMath)
4325 return true;
4326
4327 // Allow unsafe math if unsafe-fp-math attribute explicitly says so.
4328 const Function &F = MF.getFunction();
4329 if (F.hasFnAttribute("unsafe-fp-math")) {
4330 Attribute Attr = F.getFnAttribute("unsafe-fp-math");
4331 StringRef Val = Attr.getValueAsString();
4332 if (Val == "true")
4333 return true;
4334 }
4335
4336 return false;
4337 }
4338
4339 /// PerformADDCombineWithOperands - Try DAG combinations for an ADD with
4340 /// operands N0 and N1. This is a helper for PerformADDCombine that is
4341 /// called with the default operands, and if that fails, with commuted
4342 /// operands.
PerformADDCombineWithOperands(SDNode * N,SDValue N0,SDValue N1,TargetLowering::DAGCombinerInfo & DCI,const NVPTXSubtarget & Subtarget,CodeGenOpt::Level OptLevel)4343 static SDValue PerformADDCombineWithOperands(SDNode *N, SDValue N0, SDValue N1,
4344 TargetLowering::DAGCombinerInfo &DCI,
4345 const NVPTXSubtarget &Subtarget,
4346 CodeGenOpt::Level OptLevel) {
4347 SelectionDAG &DAG = DCI.DAG;
4348 // Skip non-integer, non-scalar case
4349 EVT VT=N0.getValueType();
4350 if (VT.isVector())
4351 return SDValue();
4352
4353 // fold (add (mul a, b), c) -> (mad a, b, c)
4354 //
4355 if (N0.getOpcode() == ISD::MUL) {
4356 assert (VT.isInteger());
4357 // For integer:
4358 // Since integer multiply-add costs the same as integer multiply
4359 // but is more costly than integer add, do the fusion only when
4360 // the mul is only used in the add.
4361 if (OptLevel==CodeGenOpt::None || VT != MVT::i32 ||
4362 !N0.getNode()->hasOneUse())
4363 return SDValue();
4364
4365 // Do the folding
4366 return DAG.getNode(NVPTXISD::IMAD, SDLoc(N), VT,
4367 N0.getOperand(0), N0.getOperand(1), N1);
4368 }
4369 else if (N0.getOpcode() == ISD::FMUL) {
4370 if (VT == MVT::f32 || VT == MVT::f64) {
4371 const auto *TLI = static_cast<const NVPTXTargetLowering *>(
4372 &DAG.getTargetLoweringInfo());
4373 if (!TLI->allowFMA(DAG.getMachineFunction(), OptLevel))
4374 return SDValue();
4375
4376 // For floating point:
4377 // Do the fusion only when the mul has less than 5 uses and all
4378 // are add.
4379 // The heuristic is that if a use is not an add, then that use
4380 // cannot be fused into fma, therefore mul is still needed anyway.
4381 // If there are more than 4 uses, even if they are all add, fusing
4382 // them will increase register pressue.
4383 //
4384 int numUses = 0;
4385 int nonAddCount = 0;
4386 for (SDNode::use_iterator UI = N0.getNode()->use_begin(),
4387 UE = N0.getNode()->use_end();
4388 UI != UE; ++UI) {
4389 numUses++;
4390 SDNode *User = *UI;
4391 if (User->getOpcode() != ISD::FADD)
4392 ++nonAddCount;
4393 }
4394 if (numUses >= 5)
4395 return SDValue();
4396 if (nonAddCount) {
4397 int orderNo = N->getIROrder();
4398 int orderNo2 = N0.getNode()->getIROrder();
4399 // simple heuristics here for considering potential register
4400 // pressure, the logics here is that the differnce are used
4401 // to measure the distance between def and use, the longer distance
4402 // more likely cause register pressure.
4403 if (orderNo - orderNo2 < 500)
4404 return SDValue();
4405
4406 // Now, check if at least one of the FMUL's operands is live beyond the node N,
4407 // which guarantees that the FMA will not increase register pressure at node N.
4408 bool opIsLive = false;
4409 const SDNode *left = N0.getOperand(0).getNode();
4410 const SDNode *right = N0.getOperand(1).getNode();
4411
4412 if (isa<ConstantSDNode>(left) || isa<ConstantSDNode>(right))
4413 opIsLive = true;
4414
4415 if (!opIsLive)
4416 for (SDNode::use_iterator UI = left->use_begin(), UE = left->use_end(); UI != UE; ++UI) {
4417 SDNode *User = *UI;
4418 int orderNo3 = User->getIROrder();
4419 if (orderNo3 > orderNo) {
4420 opIsLive = true;
4421 break;
4422 }
4423 }
4424
4425 if (!opIsLive)
4426 for (SDNode::use_iterator UI = right->use_begin(), UE = right->use_end(); UI != UE; ++UI) {
4427 SDNode *User = *UI;
4428 int orderNo3 = User->getIROrder();
4429 if (orderNo3 > orderNo) {
4430 opIsLive = true;
4431 break;
4432 }
4433 }
4434
4435 if (!opIsLive)
4436 return SDValue();
4437 }
4438
4439 return DAG.getNode(ISD::FMA, SDLoc(N), VT,
4440 N0.getOperand(0), N0.getOperand(1), N1);
4441 }
4442 }
4443
4444 return SDValue();
4445 }
4446
4447 /// PerformADDCombine - Target-specific dag combine xforms for ISD::ADD.
4448 ///
PerformADDCombine(SDNode * N,TargetLowering::DAGCombinerInfo & DCI,const NVPTXSubtarget & Subtarget,CodeGenOpt::Level OptLevel)4449 static SDValue PerformADDCombine(SDNode *N,
4450 TargetLowering::DAGCombinerInfo &DCI,
4451 const NVPTXSubtarget &Subtarget,
4452 CodeGenOpt::Level OptLevel) {
4453 SDValue N0 = N->getOperand(0);
4454 SDValue N1 = N->getOperand(1);
4455
4456 // First try with the default operand order.
4457 if (SDValue Result =
4458 PerformADDCombineWithOperands(N, N0, N1, DCI, Subtarget, OptLevel))
4459 return Result;
4460
4461 // If that didn't work, try again with the operands commuted.
4462 return PerformADDCombineWithOperands(N, N1, N0, DCI, Subtarget, OptLevel);
4463 }
4464
PerformANDCombine(SDNode * N,TargetLowering::DAGCombinerInfo & DCI)4465 static SDValue PerformANDCombine(SDNode *N,
4466 TargetLowering::DAGCombinerInfo &DCI) {
4467 // The type legalizer turns a vector load of i8 values into a zextload to i16
4468 // registers, optionally ANY_EXTENDs it (if target type is integer),
4469 // and ANDs off the high 8 bits. Since we turn this load into a
4470 // target-specific DAG node, the DAG combiner fails to eliminate these AND
4471 // nodes. Do that here.
4472 SDValue Val = N->getOperand(0);
4473 SDValue Mask = N->getOperand(1);
4474
4475 if (isa<ConstantSDNode>(Val)) {
4476 std::swap(Val, Mask);
4477 }
4478
4479 SDValue AExt;
4480 // Generally, we will see zextload -> IMOV16rr -> ANY_EXTEND -> and
4481 if (Val.getOpcode() == ISD::ANY_EXTEND) {
4482 AExt = Val;
4483 Val = Val->getOperand(0);
4484 }
4485
4486 if (Val->isMachineOpcode() && Val->getMachineOpcode() == NVPTX::IMOV16rr) {
4487 Val = Val->getOperand(0);
4488 }
4489
4490 if (Val->getOpcode() == NVPTXISD::LoadV2 ||
4491 Val->getOpcode() == NVPTXISD::LoadV4) {
4492 ConstantSDNode *MaskCnst = dyn_cast<ConstantSDNode>(Mask);
4493 if (!MaskCnst) {
4494 // Not an AND with a constant
4495 return SDValue();
4496 }
4497
4498 uint64_t MaskVal = MaskCnst->getZExtValue();
4499 if (MaskVal != 0xff) {
4500 // Not an AND that chops off top 8 bits
4501 return SDValue();
4502 }
4503
4504 MemSDNode *Mem = dyn_cast<MemSDNode>(Val);
4505 if (!Mem) {
4506 // Not a MemSDNode?!?
4507 return SDValue();
4508 }
4509
4510 EVT MemVT = Mem->getMemoryVT();
4511 if (MemVT != MVT::v2i8 && MemVT != MVT::v4i8) {
4512 // We only handle the i8 case
4513 return SDValue();
4514 }
4515
4516 unsigned ExtType =
4517 cast<ConstantSDNode>(Val->getOperand(Val->getNumOperands()-1))->
4518 getZExtValue();
4519 if (ExtType == ISD::SEXTLOAD) {
4520 // If for some reason the load is a sextload, the and is needed to zero
4521 // out the high 8 bits
4522 return SDValue();
4523 }
4524
4525 bool AddTo = false;
4526 if (AExt.getNode() != nullptr) {
4527 // Re-insert the ext as a zext.
4528 Val = DCI.DAG.getNode(ISD::ZERO_EXTEND, SDLoc(N),
4529 AExt.getValueType(), Val);
4530 AddTo = true;
4531 }
4532
4533 // If we get here, the AND is unnecessary. Just replace it with the load
4534 DCI.CombineTo(N, Val, AddTo);
4535 }
4536
4537 return SDValue();
4538 }
4539
PerformREMCombine(SDNode * N,TargetLowering::DAGCombinerInfo & DCI,CodeGenOpt::Level OptLevel)4540 static SDValue PerformREMCombine(SDNode *N,
4541 TargetLowering::DAGCombinerInfo &DCI,
4542 CodeGenOpt::Level OptLevel) {
4543 assert(N->getOpcode() == ISD::SREM || N->getOpcode() == ISD::UREM);
4544
4545 // Don't do anything at less than -O2.
4546 if (OptLevel < CodeGenOpt::Default)
4547 return SDValue();
4548
4549 SelectionDAG &DAG = DCI.DAG;
4550 SDLoc DL(N);
4551 EVT VT = N->getValueType(0);
4552 bool IsSigned = N->getOpcode() == ISD::SREM;
4553 unsigned DivOpc = IsSigned ? ISD::SDIV : ISD::UDIV;
4554
4555 const SDValue &Num = N->getOperand(0);
4556 const SDValue &Den = N->getOperand(1);
4557
4558 for (const SDNode *U : Num->uses()) {
4559 if (U->getOpcode() == DivOpc && U->getOperand(0) == Num &&
4560 U->getOperand(1) == Den) {
4561 // Num % Den -> Num - (Num / Den) * Den
4562 return DAG.getNode(ISD::SUB, DL, VT, Num,
4563 DAG.getNode(ISD::MUL, DL, VT,
4564 DAG.getNode(DivOpc, DL, VT, Num, Den),
4565 Den));
4566 }
4567 }
4568 return SDValue();
4569 }
4570
4571 enum OperandSignedness {
4572 Signed = 0,
4573 Unsigned,
4574 Unknown
4575 };
4576
4577 /// IsMulWideOperandDemotable - Checks if the provided DAG node is an operand
4578 /// that can be demoted to \p OptSize bits without loss of information. The
4579 /// signedness of the operand, if determinable, is placed in \p S.
IsMulWideOperandDemotable(SDValue Op,unsigned OptSize,OperandSignedness & S)4580 static bool IsMulWideOperandDemotable(SDValue Op,
4581 unsigned OptSize,
4582 OperandSignedness &S) {
4583 S = Unknown;
4584
4585 if (Op.getOpcode() == ISD::SIGN_EXTEND ||
4586 Op.getOpcode() == ISD::SIGN_EXTEND_INREG) {
4587 EVT OrigVT = Op.getOperand(0).getValueType();
4588 if (OrigVT.getSizeInBits() <= OptSize) {
4589 S = Signed;
4590 return true;
4591 }
4592 } else if (Op.getOpcode() == ISD::ZERO_EXTEND) {
4593 EVT OrigVT = Op.getOperand(0).getValueType();
4594 if (OrigVT.getSizeInBits() <= OptSize) {
4595 S = Unsigned;
4596 return true;
4597 }
4598 }
4599
4600 return false;
4601 }
4602
4603 /// AreMulWideOperandsDemotable - Checks if the given LHS and RHS operands can
4604 /// be demoted to \p OptSize bits without loss of information. If the operands
4605 /// contain a constant, it should appear as the RHS operand. The signedness of
4606 /// the operands is placed in \p IsSigned.
AreMulWideOperandsDemotable(SDValue LHS,SDValue RHS,unsigned OptSize,bool & IsSigned)4607 static bool AreMulWideOperandsDemotable(SDValue LHS, SDValue RHS,
4608 unsigned OptSize,
4609 bool &IsSigned) {
4610 OperandSignedness LHSSign;
4611
4612 // The LHS operand must be a demotable op
4613 if (!IsMulWideOperandDemotable(LHS, OptSize, LHSSign))
4614 return false;
4615
4616 // We should have been able to determine the signedness from the LHS
4617 if (LHSSign == Unknown)
4618 return false;
4619
4620 IsSigned = (LHSSign == Signed);
4621
4622 // The RHS can be a demotable op or a constant
4623 if (ConstantSDNode *CI = dyn_cast<ConstantSDNode>(RHS)) {
4624 const APInt &Val = CI->getAPIntValue();
4625 if (LHSSign == Unsigned) {
4626 return Val.isIntN(OptSize);
4627 } else {
4628 return Val.isSignedIntN(OptSize);
4629 }
4630 } else {
4631 OperandSignedness RHSSign;
4632 if (!IsMulWideOperandDemotable(RHS, OptSize, RHSSign))
4633 return false;
4634
4635 return LHSSign == RHSSign;
4636 }
4637 }
4638
4639 /// TryMULWIDECombine - Attempt to replace a multiply of M bits with a multiply
4640 /// of M/2 bits that produces an M-bit result (i.e. mul.wide). This transform
4641 /// works on both multiply DAG nodes and SHL DAG nodes with a constant shift
4642 /// amount.
TryMULWIDECombine(SDNode * N,TargetLowering::DAGCombinerInfo & DCI)4643 static SDValue TryMULWIDECombine(SDNode *N,
4644 TargetLowering::DAGCombinerInfo &DCI) {
4645 EVT MulType = N->getValueType(0);
4646 if (MulType != MVT::i32 && MulType != MVT::i64) {
4647 return SDValue();
4648 }
4649
4650 SDLoc DL(N);
4651 unsigned OptSize = MulType.getSizeInBits() >> 1;
4652 SDValue LHS = N->getOperand(0);
4653 SDValue RHS = N->getOperand(1);
4654
4655 // Canonicalize the multiply so the constant (if any) is on the right
4656 if (N->getOpcode() == ISD::MUL) {
4657 if (isa<ConstantSDNode>(LHS)) {
4658 std::swap(LHS, RHS);
4659 }
4660 }
4661
4662 // If we have a SHL, determine the actual multiply amount
4663 if (N->getOpcode() == ISD::SHL) {
4664 ConstantSDNode *ShlRHS = dyn_cast<ConstantSDNode>(RHS);
4665 if (!ShlRHS) {
4666 return SDValue();
4667 }
4668
4669 APInt ShiftAmt = ShlRHS->getAPIntValue();
4670 unsigned BitWidth = MulType.getSizeInBits();
4671 if (ShiftAmt.sge(0) && ShiftAmt.slt(BitWidth)) {
4672 APInt MulVal = APInt(BitWidth, 1) << ShiftAmt;
4673 RHS = DCI.DAG.getConstant(MulVal, DL, MulType);
4674 } else {
4675 return SDValue();
4676 }
4677 }
4678
4679 bool Signed;
4680 // Verify that our operands are demotable
4681 if (!AreMulWideOperandsDemotable(LHS, RHS, OptSize, Signed)) {
4682 return SDValue();
4683 }
4684
4685 EVT DemotedVT;
4686 if (MulType == MVT::i32) {
4687 DemotedVT = MVT::i16;
4688 } else {
4689 DemotedVT = MVT::i32;
4690 }
4691
4692 // Truncate the operands to the correct size. Note that these are just for
4693 // type consistency and will (likely) be eliminated in later phases.
4694 SDValue TruncLHS =
4695 DCI.DAG.getNode(ISD::TRUNCATE, DL, DemotedVT, LHS);
4696 SDValue TruncRHS =
4697 DCI.DAG.getNode(ISD::TRUNCATE, DL, DemotedVT, RHS);
4698
4699 unsigned Opc;
4700 if (Signed) {
4701 Opc = NVPTXISD::MUL_WIDE_SIGNED;
4702 } else {
4703 Opc = NVPTXISD::MUL_WIDE_UNSIGNED;
4704 }
4705
4706 return DCI.DAG.getNode(Opc, DL, MulType, TruncLHS, TruncRHS);
4707 }
4708
4709 /// PerformMULCombine - Runs PTX-specific DAG combine patterns on MUL nodes.
PerformMULCombine(SDNode * N,TargetLowering::DAGCombinerInfo & DCI,CodeGenOpt::Level OptLevel)4710 static SDValue PerformMULCombine(SDNode *N,
4711 TargetLowering::DAGCombinerInfo &DCI,
4712 CodeGenOpt::Level OptLevel) {
4713 if (OptLevel > 0) {
4714 // Try mul.wide combining at OptLevel > 0
4715 if (SDValue Ret = TryMULWIDECombine(N, DCI))
4716 return Ret;
4717 }
4718
4719 return SDValue();
4720 }
4721
4722 /// PerformSHLCombine - Runs PTX-specific DAG combine patterns on SHL nodes.
PerformSHLCombine(SDNode * N,TargetLowering::DAGCombinerInfo & DCI,CodeGenOpt::Level OptLevel)4723 static SDValue PerformSHLCombine(SDNode *N,
4724 TargetLowering::DAGCombinerInfo &DCI,
4725 CodeGenOpt::Level OptLevel) {
4726 if (OptLevel > 0) {
4727 // Try mul.wide combining at OptLevel > 0
4728 if (SDValue Ret = TryMULWIDECombine(N, DCI))
4729 return Ret;
4730 }
4731
4732 return SDValue();
4733 }
4734
PerformSETCCCombine(SDNode * N,TargetLowering::DAGCombinerInfo & DCI)4735 static SDValue PerformSETCCCombine(SDNode *N,
4736 TargetLowering::DAGCombinerInfo &DCI) {
4737 EVT CCType = N->getValueType(0);
4738 SDValue A = N->getOperand(0);
4739 SDValue B = N->getOperand(1);
4740
4741 if (CCType != MVT::v2i1 || A.getValueType() != MVT::v2f16)
4742 return SDValue();
4743
4744 SDLoc DL(N);
4745 // setp.f16x2 returns two scalar predicates, which we need to
4746 // convert back to v2i1. The returned result will be scalarized by
4747 // the legalizer, but the comparison will remain a single vector
4748 // instruction.
4749 SDValue CCNode = DCI.DAG.getNode(NVPTXISD::SETP_F16X2, DL,
4750 DCI.DAG.getVTList(MVT::i1, MVT::i1),
4751 {A, B, N->getOperand(2)});
4752 return DCI.DAG.getNode(ISD::BUILD_VECTOR, DL, CCType, CCNode.getValue(0),
4753 CCNode.getValue(1));
4754 }
4755
PerformDAGCombine(SDNode * N,DAGCombinerInfo & DCI) const4756 SDValue NVPTXTargetLowering::PerformDAGCombine(SDNode *N,
4757 DAGCombinerInfo &DCI) const {
4758 CodeGenOpt::Level OptLevel = getTargetMachine().getOptLevel();
4759 switch (N->getOpcode()) {
4760 default: break;
4761 case ISD::ADD:
4762 case ISD::FADD:
4763 return PerformADDCombine(N, DCI, STI, OptLevel);
4764 case ISD::MUL:
4765 return PerformMULCombine(N, DCI, OptLevel);
4766 case ISD::SHL:
4767 return PerformSHLCombine(N, DCI, OptLevel);
4768 case ISD::AND:
4769 return PerformANDCombine(N, DCI);
4770 case ISD::UREM:
4771 case ISD::SREM:
4772 return PerformREMCombine(N, DCI, OptLevel);
4773 case ISD::SETCC:
4774 return PerformSETCCCombine(N, DCI);
4775 }
4776 return SDValue();
4777 }
4778
4779 /// ReplaceVectorLoad - Convert vector loads into multi-output scalar loads.
ReplaceLoadVector(SDNode * N,SelectionDAG & DAG,SmallVectorImpl<SDValue> & Results)4780 static void ReplaceLoadVector(SDNode *N, SelectionDAG &DAG,
4781 SmallVectorImpl<SDValue> &Results) {
4782 EVT ResVT = N->getValueType(0);
4783 SDLoc DL(N);
4784
4785 assert(ResVT.isVector() && "Vector load must have vector type");
4786
4787 // We only handle "native" vector sizes for now, e.g. <4 x double> is not
4788 // legal. We can (and should) split that into 2 loads of <2 x double> here
4789 // but I'm leaving that as a TODO for now.
4790 assert(ResVT.isSimple() && "Can only handle simple types");
4791 switch (ResVT.getSimpleVT().SimpleTy) {
4792 default:
4793 return;
4794 case MVT::v2i8:
4795 case MVT::v2i16:
4796 case MVT::v2i32:
4797 case MVT::v2i64:
4798 case MVT::v2f16:
4799 case MVT::v2f32:
4800 case MVT::v2f64:
4801 case MVT::v4i8:
4802 case MVT::v4i16:
4803 case MVT::v4i32:
4804 case MVT::v4f16:
4805 case MVT::v4f32:
4806 case MVT::v8f16: // <4 x f16x2>
4807 // This is a "native" vector type
4808 break;
4809 }
4810
4811 LoadSDNode *LD = cast<LoadSDNode>(N);
4812
4813 unsigned Align = LD->getAlignment();
4814 auto &TD = DAG.getDataLayout();
4815 unsigned PrefAlign =
4816 TD.getPrefTypeAlignment(ResVT.getTypeForEVT(*DAG.getContext()));
4817 if (Align < PrefAlign) {
4818 // This load is not sufficiently aligned, so bail out and let this vector
4819 // load be scalarized. Note that we may still be able to emit smaller
4820 // vector loads. For example, if we are loading a <4 x float> with an
4821 // alignment of 8, this check will fail but the legalizer will try again
4822 // with 2 x <2 x float>, which will succeed with an alignment of 8.
4823 return;
4824 }
4825
4826 EVT EltVT = ResVT.getVectorElementType();
4827 unsigned NumElts = ResVT.getVectorNumElements();
4828
4829 // Since LoadV2 is a target node, we cannot rely on DAG type legalization.
4830 // Therefore, we must ensure the type is legal. For i1 and i8, we set the
4831 // loaded type to i16 and propagate the "real" type as the memory type.
4832 bool NeedTrunc = false;
4833 if (EltVT.getSizeInBits() < 16) {
4834 EltVT = MVT::i16;
4835 NeedTrunc = true;
4836 }
4837
4838 unsigned Opcode = 0;
4839 SDVTList LdResVTs;
4840 bool LoadF16x2 = false;
4841
4842 switch (NumElts) {
4843 default:
4844 return;
4845 case 2:
4846 Opcode = NVPTXISD::LoadV2;
4847 LdResVTs = DAG.getVTList(EltVT, EltVT, MVT::Other);
4848 break;
4849 case 4: {
4850 Opcode = NVPTXISD::LoadV4;
4851 EVT ListVTs[] = { EltVT, EltVT, EltVT, EltVT, MVT::Other };
4852 LdResVTs = DAG.getVTList(ListVTs);
4853 break;
4854 }
4855 case 8: {
4856 // v8f16 is a special case. PTX doesn't have ld.v8.f16
4857 // instruction. Instead, we split the vector into v2f16 chunks and
4858 // load them with ld.v4.b32.
4859 assert(EltVT == MVT::f16 && "Unsupported v8 vector type.");
4860 LoadF16x2 = true;
4861 Opcode = NVPTXISD::LoadV4;
4862 EVT ListVTs[] = {MVT::v2f16, MVT::v2f16, MVT::v2f16, MVT::v2f16,
4863 MVT::Other};
4864 LdResVTs = DAG.getVTList(ListVTs);
4865 break;
4866 }
4867 }
4868
4869 // Copy regular operands
4870 SmallVector<SDValue, 8> OtherOps(N->op_begin(), N->op_end());
4871
4872 // The select routine does not have access to the LoadSDNode instance, so
4873 // pass along the extension information
4874 OtherOps.push_back(DAG.getIntPtrConstant(LD->getExtensionType(), DL));
4875
4876 SDValue NewLD = DAG.getMemIntrinsicNode(Opcode, DL, LdResVTs, OtherOps,
4877 LD->getMemoryVT(),
4878 LD->getMemOperand());
4879
4880 SmallVector<SDValue, 8> ScalarRes;
4881 if (LoadF16x2) {
4882 // Split v2f16 subvectors back into individual elements.
4883 NumElts /= 2;
4884 for (unsigned i = 0; i < NumElts; ++i) {
4885 SDValue SubVector = NewLD.getValue(i);
4886 SDValue E0 = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, EltVT, SubVector,
4887 DAG.getIntPtrConstant(0, DL));
4888 SDValue E1 = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, EltVT, SubVector,
4889 DAG.getIntPtrConstant(1, DL));
4890 ScalarRes.push_back(E0);
4891 ScalarRes.push_back(E1);
4892 }
4893 } else {
4894 for (unsigned i = 0; i < NumElts; ++i) {
4895 SDValue Res = NewLD.getValue(i);
4896 if (NeedTrunc)
4897 Res = DAG.getNode(ISD::TRUNCATE, DL, ResVT.getVectorElementType(), Res);
4898 ScalarRes.push_back(Res);
4899 }
4900 }
4901
4902 SDValue LoadChain = NewLD.getValue(NumElts);
4903
4904 SDValue BuildVec = DAG.getBuildVector(ResVT, DL, ScalarRes);
4905
4906 Results.push_back(BuildVec);
4907 Results.push_back(LoadChain);
4908 }
4909
ReplaceINTRINSIC_W_CHAIN(SDNode * N,SelectionDAG & DAG,SmallVectorImpl<SDValue> & Results)4910 static void ReplaceINTRINSIC_W_CHAIN(SDNode *N, SelectionDAG &DAG,
4911 SmallVectorImpl<SDValue> &Results) {
4912 SDValue Chain = N->getOperand(0);
4913 SDValue Intrin = N->getOperand(1);
4914 SDLoc DL(N);
4915
4916 // Get the intrinsic ID
4917 unsigned IntrinNo = cast<ConstantSDNode>(Intrin.getNode())->getZExtValue();
4918 switch (IntrinNo) {
4919 default:
4920 return;
4921 case Intrinsic::nvvm_ldg_global_i:
4922 case Intrinsic::nvvm_ldg_global_f:
4923 case Intrinsic::nvvm_ldg_global_p:
4924 case Intrinsic::nvvm_ldu_global_i:
4925 case Intrinsic::nvvm_ldu_global_f:
4926 case Intrinsic::nvvm_ldu_global_p: {
4927 EVT ResVT = N->getValueType(0);
4928
4929 if (ResVT.isVector()) {
4930 // Vector LDG/LDU
4931
4932 unsigned NumElts = ResVT.getVectorNumElements();
4933 EVT EltVT = ResVT.getVectorElementType();
4934
4935 // Since LDU/LDG are target nodes, we cannot rely on DAG type
4936 // legalization.
4937 // Therefore, we must ensure the type is legal. For i1 and i8, we set the
4938 // loaded type to i16 and propagate the "real" type as the memory type.
4939 bool NeedTrunc = false;
4940 if (EltVT.getSizeInBits() < 16) {
4941 EltVT = MVT::i16;
4942 NeedTrunc = true;
4943 }
4944
4945 unsigned Opcode = 0;
4946 SDVTList LdResVTs;
4947
4948 switch (NumElts) {
4949 default:
4950 return;
4951 case 2:
4952 switch (IntrinNo) {
4953 default:
4954 return;
4955 case Intrinsic::nvvm_ldg_global_i:
4956 case Intrinsic::nvvm_ldg_global_f:
4957 case Intrinsic::nvvm_ldg_global_p:
4958 Opcode = NVPTXISD::LDGV2;
4959 break;
4960 case Intrinsic::nvvm_ldu_global_i:
4961 case Intrinsic::nvvm_ldu_global_f:
4962 case Intrinsic::nvvm_ldu_global_p:
4963 Opcode = NVPTXISD::LDUV2;
4964 break;
4965 }
4966 LdResVTs = DAG.getVTList(EltVT, EltVT, MVT::Other);
4967 break;
4968 case 4: {
4969 switch (IntrinNo) {
4970 default:
4971 return;
4972 case Intrinsic::nvvm_ldg_global_i:
4973 case Intrinsic::nvvm_ldg_global_f:
4974 case Intrinsic::nvvm_ldg_global_p:
4975 Opcode = NVPTXISD::LDGV4;
4976 break;
4977 case Intrinsic::nvvm_ldu_global_i:
4978 case Intrinsic::nvvm_ldu_global_f:
4979 case Intrinsic::nvvm_ldu_global_p:
4980 Opcode = NVPTXISD::LDUV4;
4981 break;
4982 }
4983 EVT ListVTs[] = { EltVT, EltVT, EltVT, EltVT, MVT::Other };
4984 LdResVTs = DAG.getVTList(ListVTs);
4985 break;
4986 }
4987 }
4988
4989 SmallVector<SDValue, 8> OtherOps;
4990
4991 // Copy regular operands
4992
4993 OtherOps.push_back(Chain); // Chain
4994 // Skip operand 1 (intrinsic ID)
4995 // Others
4996 OtherOps.append(N->op_begin() + 2, N->op_end());
4997
4998 MemIntrinsicSDNode *MemSD = cast<MemIntrinsicSDNode>(N);
4999
5000 SDValue NewLD = DAG.getMemIntrinsicNode(Opcode, DL, LdResVTs, OtherOps,
5001 MemSD->getMemoryVT(),
5002 MemSD->getMemOperand());
5003
5004 SmallVector<SDValue, 4> ScalarRes;
5005
5006 for (unsigned i = 0; i < NumElts; ++i) {
5007 SDValue Res = NewLD.getValue(i);
5008 if (NeedTrunc)
5009 Res =
5010 DAG.getNode(ISD::TRUNCATE, DL, ResVT.getVectorElementType(), Res);
5011 ScalarRes.push_back(Res);
5012 }
5013
5014 SDValue LoadChain = NewLD.getValue(NumElts);
5015
5016 SDValue BuildVec =
5017 DAG.getBuildVector(ResVT, DL, ScalarRes);
5018
5019 Results.push_back(BuildVec);
5020 Results.push_back(LoadChain);
5021 } else {
5022 // i8 LDG/LDU
5023 assert(ResVT.isSimple() && ResVT.getSimpleVT().SimpleTy == MVT::i8 &&
5024 "Custom handling of non-i8 ldu/ldg?");
5025
5026 // Just copy all operands as-is
5027 SmallVector<SDValue, 4> Ops(N->op_begin(), N->op_end());
5028
5029 // Force output to i16
5030 SDVTList LdResVTs = DAG.getVTList(MVT::i16, MVT::Other);
5031
5032 MemIntrinsicSDNode *MemSD = cast<MemIntrinsicSDNode>(N);
5033
5034 // We make sure the memory type is i8, which will be used during isel
5035 // to select the proper instruction.
5036 SDValue NewLD =
5037 DAG.getMemIntrinsicNode(ISD::INTRINSIC_W_CHAIN, DL, LdResVTs, Ops,
5038 MVT::i8, MemSD->getMemOperand());
5039
5040 Results.push_back(DAG.getNode(ISD::TRUNCATE, DL, MVT::i8,
5041 NewLD.getValue(0)));
5042 Results.push_back(NewLD.getValue(1));
5043 }
5044 }
5045 }
5046 }
5047
ReplaceNodeResults(SDNode * N,SmallVectorImpl<SDValue> & Results,SelectionDAG & DAG) const5048 void NVPTXTargetLowering::ReplaceNodeResults(
5049 SDNode *N, SmallVectorImpl<SDValue> &Results, SelectionDAG &DAG) const {
5050 switch (N->getOpcode()) {
5051 default:
5052 report_fatal_error("Unhandled custom legalization");
5053 case ISD::LOAD:
5054 ReplaceLoadVector(N, DAG, Results);
5055 return;
5056 case ISD::INTRINSIC_W_CHAIN:
5057 ReplaceINTRINSIC_W_CHAIN(N, DAG, Results);
5058 return;
5059 }
5060 }
5061
5062 // Pin NVPTXTargetObjectFile's vtables to this file.
~NVPTXTargetObjectFile()5063 NVPTXTargetObjectFile::~NVPTXTargetObjectFile() {}
5064
SelectSectionForGlobal(const GlobalObject * GO,SectionKind Kind,const TargetMachine & TM) const5065 MCSection *NVPTXTargetObjectFile::SelectSectionForGlobal(
5066 const GlobalObject *GO, SectionKind Kind, const TargetMachine &TM) const {
5067 return getDataSection();
5068 }
5069