LLVM 23.0.0git
SPIRVInstructionSelector.cpp
Go to the documentation of this file.
1//===- SPIRVInstructionSelector.cpp ------------------------------*- C++ -*-==//
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 implements the targeting of the InstructionSelector class for
10// SPIRV.
11// TODO: This should be generated by TableGen.
12//
13//===----------------------------------------------------------------------===//
14
15#include "MCTargetDesc/SPIRVBaseInfo.h"
16#include "MCTargetDesc/SPIRVMCTargetDesc.h"
17#include "SPIRV.h"
18#include "SPIRVGlobalRegistry.h"
19#include "SPIRVInstrInfo.h"
20#include "SPIRVRegisterInfo.h"
21#include "SPIRVTargetMachine.h"
22#include "SPIRVTypeInst.h"
23#include "SPIRVUtils.h"
24#include "llvm/ADT/APFloat.h"
25#include "llvm/ADT/StringExtras.h"
26#include "llvm/CodeGen/GlobalISel/GIMatchTableExecutorImpl.h"
27#include "llvm/CodeGen/GlobalISel/GenericMachineInstrs.h"
28#include "llvm/CodeGen/GlobalISel/InstructionSelector.h"
29#include "llvm/CodeGen/MachineInstrBuilder.h"
30#include "llvm/CodeGen/MachineRegisterInfo.h"
31#include "llvm/CodeGen/Register.h"
32#include "llvm/CodeGen/TargetOpcodes.h"
33#include "llvm/IR/DiagnosticInfo.h"
34#include "llvm/IR/IntrinsicsSPIRV.h"
35#include "llvm/Support/Debug.h"
36#include "llvm/Support/ErrorHandling.h"
37#include <functional>
38#include <optional>
39
40#define DEBUG_TYPE "spirv-isel"
41
42using namespace llvm;
43namespace CL = SPIRV::OpenCLExtInst;
44namespace GL = SPIRV::GLSLExtInst;
45
46using ExtInstList =
47 std::vector<std::pair<SPIRV::InstructionSet::InstructionSet, uint32_t>>;
48
49namespace {
50
51struct ImageOperands {
52 std::optional<Register> Bias;
53 std::optional<Register> Offset;
54 std::optional<Register> MinLod;
55 std::optional<Register> GradX;
56 std::optional<Register> GradY;
57 std::optional<Register> Lod;
58 std::optional<Register> Compare;
59};
60
61struct SplitParts {
62 SPIRVTypeInst Type = nullptr;
63 Register High;
64 Register Low;
65 bool IsScalar = false;
66};
67
68llvm::SPIRV::SelectionControl::SelectionControl
69getSelectionOperandForImm(int Imm) {
70 if (Imm == 2)
71 return SPIRV::SelectionControl::Flatten;
72 if (Imm == 1)
73 return SPIRV::SelectionControl::DontFlatten;
74 if (Imm == 0)
75 return SPIRV::SelectionControl::None;
76 llvm_unreachable("Invalid immediate");
77}
78
79#define GET_GLOBALISEL_PREDICATE_BITSET
80#include "SPIRVGenGlobalISel.inc"
81#undef GET_GLOBALISEL_PREDICATE_BITSET
82
83class SPIRVInstructionSelector : public InstructionSelector {
84 const SPIRVSubtarget &STI;
85 const SPIRVInstrInfo &TII;
86 const SPIRVRegisterInfo &TRI;
87 const RegisterBankInfo &RBI;
88 SPIRVGlobalRegistry &GR;
89 MachineRegisterInfo *MRI;
90 MachineFunction *HasVRegsReset = nullptr;
91
92 /// We need to keep track of the number we give to anonymous global values to
93 /// generate the same name every time when this is needed.
94 mutable DenseMap<const GlobalValue *, unsigned> UnnamedGlobalIDs;
95 SmallPtrSet<MachineInstr *, 8> DeadMIs;
96
97public:
98 SPIRVInstructionSelector(const SPIRVTargetMachine &TM,
99 const SPIRVSubtarget &ST,
100 const RegisterBankInfo &RBI);
101 void setupMF(MachineFunction &MF, GISelValueTracking *VT,
102 CodeGenCoverage *CoverageInfo, ProfileSummaryInfo *PSI,
103 BlockFrequencyInfo *BFI) override;
104 // Common selection code. Instruction-specific selection occurs in spvSelect.
105 bool select(MachineInstr &I) override;
106 static const char *getName() { return DEBUG_TYPE; }
107
108#define GET_GLOBALISEL_PREDICATES_DECL
109#include "SPIRVGenGlobalISel.inc"
110#undef GET_GLOBALISEL_PREDICATES_DECL
111
112#define GET_GLOBALISEL_TEMPORARIES_DECL
113#include "SPIRVGenGlobalISel.inc"
114#undef GET_GLOBALISEL_TEMPORARIES_DECL
115
116private:
117 void resetVRegsType(MachineFunction &MF);
118 void removeDeadInstruction(MachineInstr &MI) const;
119 void removeOpNamesForDeadMI(MachineInstr &MI) const;
120
121 // tblgen-erated 'select' implementation, used as the initial selector for
122 // the patterns that don't require complex C++.
123 bool selectImpl(MachineInstr &I, CodeGenCoverage &CoverageInfo) const;
124
125 // All instruction-specific selection that didn't happen in "select()".
126 // Is basically a large Switch/Case delegating to all other select method.
127 bool spvSelect(Register ResVReg, SPIRVTypeInst ResType,
128 MachineInstr &I) const;
129
130 bool selectFirstBitHigh(Register ResVReg, SPIRVTypeInst ResType,
131 MachineInstr &I, bool IsSigned) const;
132
133 bool selectFirstBitLow(Register ResVReg, SPIRVTypeInst ResType,
134 MachineInstr &I) const;
135
136 bool selectFirstBitSet16(Register ResVReg, SPIRVTypeInst ResType,
137 MachineInstr &I, unsigned ExtendOpcode,
138 unsigned BitSetOpcode) const;
139
140 bool selectFirstBitSet32(Register ResVReg, SPIRVTypeInst ResType,
141 MachineInstr &I, Register SrcReg,
142 unsigned BitSetOpcode) const;
143
144 bool selectFirstBitSet64(Register ResVReg, SPIRVTypeInst ResType,
145 MachineInstr &I, Register SrcReg,
146 unsigned BitSetOpcode, bool SwapPrimarySide) const;
147
148 bool selectGlobalValue(Register ResVReg, MachineInstr &I,
149 const MachineInstr *Init = nullptr) const;
150
151 bool selectOpWithSrcs(Register ResVReg, SPIRVTypeInst ResType,
152 MachineInstr &I, std::vector<Register> SrcRegs,
153 unsigned Opcode) const;
154
155 bool selectUnOp(Register ResVReg, SPIRVTypeInst ResType, MachineInstr &I,
156 unsigned Opcode) const;
157
158 bool selectBitcast(Register ResVReg, SPIRVTypeInst ResType,
159 MachineInstr &I) const;
160
161 bool selectLoad(Register ResVReg, SPIRVTypeInst ResType,
162 MachineInstr &I) const;
163 bool selectAtomicLoad(Register ResVReg, SPIRVTypeInst ResType,
164 MachineInstr &I) const;
165 bool selectStore(MachineInstr &I) const;
166 bool selectAtomicStore(MachineInstr &I) const;
167
168 bool selectStackSave(Register ResVReg, SPIRVTypeInst ResType,
169 MachineInstr &I) const;
170 bool selectStackRestore(MachineInstr &I) const;
171
172 bool selectMemOperation(Register ResVReg, MachineInstr &I) const;
173 Register getOrCreateMemSetGlobal(MachineInstr &I) const;
174 bool selectCopyMemory(MachineInstr &I, Register SrcReg) const;
175 bool selectCopyMemorySized(MachineInstr &I, Register SrcReg) const;
176
177 bool selectAtomicRMW(Register ResVReg, SPIRVTypeInst ResType, MachineInstr &I,
178 unsigned NewOpcode, unsigned NegateOpcode = 0) const;
179
180 bool selectAtomicCmpXchg(Register ResVReg, SPIRVTypeInst ResType,
181 MachineInstr &I) const;
182
183 bool selectFence(MachineInstr &I) const;
184
185 bool selectAddrSpaceCast(Register ResVReg, SPIRVTypeInst ResType,
186 MachineInstr &I) const;
187
188 bool selectAnyOrAll(Register ResVReg, SPIRVTypeInst ResType, MachineInstr &I,
189 unsigned OpType) const;
190
191 bool selectAll(Register ResVReg, SPIRVTypeInst ResType,
192 MachineInstr &I) const;
193
194 bool selectAny(Register ResVReg, SPIRVTypeInst ResType,
195 MachineInstr &I) const;
196
197 bool selectBitreverse(Register ResVReg, SPIRVTypeInst ResType,
198 MachineInstr &I) const;
199
200 bool selectBitreverse16(Register ResVReg, SPIRVTypeInst ResType,
201 MachineInstr &I, Register Op) const;
202
203 bool selectBitreverse64(Register ResVReg, SPIRVTypeInst ResType,
204 MachineInstr &I, Register SrcReg) const;
205
206 bool selectBitreverseNative(Register ResVReg, SPIRVTypeInst ResType,
207 MachineInstr &I, Register Op) const;
208
209 bool selectBuildVector(Register ResVReg, SPIRVTypeInst ResType,
210 MachineInstr &I) const;
211 bool selectSplatVector(Register ResVReg, SPIRVTypeInst ResType,
212 MachineInstr &I) const;
213
214 bool selectCmp(Register ResVReg, SPIRVTypeInst ResType,
215 unsigned comparisonOpcode, MachineInstr &I) const;
216 bool selectDiscard(Register ResVReg, SPIRVTypeInst ResType,
217 MachineInstr &I) const;
218
219 bool selectICmp(Register ResVReg, SPIRVTypeInst ResType,
220 MachineInstr &I) const;
221 bool selectFCmp(Register ResVReg, SPIRVTypeInst ResType,
222 MachineInstr &I) const;
223
224 bool selectSign(Register ResVReg, SPIRVTypeInst ResType,
225 MachineInstr &I) const;
226
227 bool selectFloatDot(Register ResVReg, SPIRVTypeInst ResType,
228 MachineInstr &I) const;
229
230 bool selectOverflowArith(Register ResVReg, SPIRVTypeInst ResType,
231 MachineInstr &I, unsigned Opcode) const;
232 bool selectDebugTrap(Register ResVReg, SPIRVTypeInst ResType,
233 MachineInstr &I) const;
234
235 bool selectIntegerDot(Register ResVReg, SPIRVTypeInst ResType,
236 MachineInstr &I, bool Signed) const;
237
238 bool selectIntegerDotExpansion(Register ResVReg, SPIRVTypeInst ResType,
239 MachineInstr &I) const;
240
241 bool selectOpIsInf(Register ResVReg, SPIRVTypeInst ResType,
242 MachineInstr &I) const;
243
244 bool selectOpIsNan(Register ResVReg, SPIRVTypeInst ResType,
245 MachineInstr &I) const;
246
247 bool selectPopCount(Register ResVReg, SPIRVTypeInst ResType,
248 MachineInstr &I, unsigned Opcode) const;
249
250 bool selectPopCount16(Register ResVReg, SPIRVTypeInst ResType,
251 MachineInstr &I, unsigned ExtOpcode,
252 unsigned Opcode) const;
253
254 bool selectPopCount32(Register ResVReg, SPIRVTypeInst ResType,
255 MachineInstr &I, Register SrcReg,
256 unsigned Opcode) const;
257
258 bool selectPopCount64(Register ResVReg, SPIRVTypeInst ResType,
259 MachineInstr &I, Register SrcReg,
260 unsigned Opcode) const;
261
262 template <bool Signed>
263 bool selectDot4AddPacked(Register ResVReg, SPIRVTypeInst ResType,
264 MachineInstr &I) const;
265 template <bool Signed>
266 bool selectDot4AddPackedExpansion(Register ResVReg, SPIRVTypeInst ResType,
267 MachineInstr &I) const;
268
269 bool selectWavePrefixBitCount(Register ResVReg, SPIRVTypeInst ResType,
270 MachineInstr &I) const;
271
272 template <typename PickOpcodeFn>
273 bool selectWaveReduce(Register ResVReg, SPIRVTypeInst ResType,
274 MachineInstr &I, bool IsUnsigned,
275 PickOpcodeFn &&PickOpcode) const;
276
277 bool selectWaveReduceOp(Register ResVReg, SPIRVTypeInst ResType,
278 MachineInstr &I, unsigned Opcode) const;
279
280 bool selectWaveReduceMax(Register ResVReg, SPIRVTypeInst ResType,
281 MachineInstr &I, bool IsUnsigned) const;
282
283 bool selectWaveReduceMin(Register ResVReg, SPIRVTypeInst ResType,
284 MachineInstr &I, bool IsUnsigned) const;
285
286 bool selectWaveReduceSum(Register ResVReg, SPIRVTypeInst ResType,
287 MachineInstr &I) const;
288
289 bool selectWaveReduceProduct(Register ResVReg, const SPIRVTypeInst ResType,
290 MachineInstr &I) const;
291
292 template <typename PickOpcodeFn>
293 bool selectWaveExclusiveScan(Register ResVReg, SPIRVTypeInst ResType,
294 MachineInstr &I, bool IsUnsigned,
295 PickOpcodeFn &&PickOpcode) const;
296
297 bool selectWaveExclusiveScanSum(Register ResVReg, SPIRVTypeInst ResType,
298 MachineInstr &I) const;
299
300 bool selectWaveExclusiveScanProduct(Register ResVReg, SPIRVTypeInst ResType,
301 MachineInstr &I) const;
302
303 bool selectQuadSwap(Register ResVReg, SPIRVTypeInst ResType, MachineInstr &I,
304 unsigned Direction) const;
305
306 bool selectConst(Register ResVReg, SPIRVTypeInst ResType,
307 MachineInstr &I) const;
308
309 bool selectSelect(Register ResVReg, SPIRVTypeInst ResType,
310 MachineInstr &I) const;
311 bool selectBoolToInt(Register ResVReg, SPIRVTypeInst ResType,
312 Register BooleanVReg, MachineInstr &InsertAt,
313 bool IsSigned) const;
314 bool selectIToF(Register ResVReg, SPIRVTypeInst ResType, MachineInstr &I,
315 bool IsSigned, unsigned Opcode) const;
316 bool selectExt(Register ResVReg, SPIRVTypeInst ResType, MachineInstr &I,
317 bool IsSigned) const;
318
319 bool selectTrunc(Register ResVReg, SPIRVTypeInst ResType,
320 MachineInstr &I) const;
321
322 bool selectSUCmp(Register ResVReg, SPIRVTypeInst ResType, MachineInstr &I,
323 bool IsSigned) const;
324
325 bool selectIntToBool(Register IntReg, Register ResVReg, MachineInstr &I,
326 SPIRVTypeInst intTy, SPIRVTypeInst boolTy) const;
327
328 bool selectOpUndef(Register ResVReg, SPIRVTypeInst ResType,
329 MachineInstr &I) const;
330 bool selectFreeze(Register ResVReg, SPIRVTypeInst ResType,
331 MachineInstr &I) const;
332 bool selectIntrinsic(Register ResVReg, SPIRVTypeInst ResType,
333 MachineInstr &I) const;
334 bool selectExtractVal(Register ResVReg, SPIRVTypeInst ResType,
335 MachineInstr &I) const;
336 bool selectInsertVal(Register ResVReg, SPIRVTypeInst ResType,
337 MachineInstr &I) const;
338 bool selectExtractElt(Register ResVReg, SPIRVTypeInst ResType,
339 MachineInstr &I) const;
340 bool selectInsertElt(Register ResVReg, SPIRVTypeInst ResType,
341 MachineInstr &I) const;
342 bool selectGEP(Register ResVReg, SPIRVTypeInst ResType,
343 MachineInstr &I) const;
344
345 bool selectMaskedGather(Register ResVReg, SPIRVTypeInst ResType,
346 MachineInstr &I) const;
347 bool selectMaskedScatter(MachineInstr &I) const;
348
349 bool diagnoseUnsupported(const MachineInstr &I, const Twine &Msg) const;
350
351 bool selectFrameIndex(Register ResVReg, SPIRVTypeInst ResType,
352 MachineInstr &I) const;
353 bool selectAllocaArray(Register ResVReg, SPIRVTypeInst ResType,
354 MachineInstr &I) const;
355
356 bool selectBranch(MachineInstr &I) const;
357 bool selectBranchCond(MachineInstr &I) const;
358
359 bool selectPhi(Register ResVReg, MachineInstr &I) const;
360
361 bool selectExtInst(Register ResVReg, SPIRVTypeInst RestType, MachineInstr &I,
362 GL::GLSLExtInst GLInst, bool setMIFlags = true,
363 bool useMISrc = true,
364 ArrayRef<Register> SrcRegs = {}) const;
365 bool selectExtInst(Register ResVReg, SPIRVTypeInst ResType, MachineInstr &I,
366 CL::OpenCLExtInst CLInst, bool setMIFlags = true,
367 bool useMISrc = true,
368 ArrayRef<Register> SrcRegs = {}) const;
369 bool selectExtInst(Register ResVReg, SPIRVTypeInst ResType, MachineInstr &I,
370 CL::OpenCLExtInst CLInst, GL::GLSLExtInst GLInst,
371 bool setMIFlags = true, bool useMISrc = true,
372 ArrayRef<Register> SrcRegs = {}) const;
373 bool selectExtInst(Register ResVReg, SPIRVTypeInst ResType, MachineInstr &I,
374 const ExtInstList &ExtInsts, bool setMIFlags = true,
375 bool useMISrc = true,
376 ArrayRef<Register> SrcRegs = {}) const;
377
378 bool selectLog10(Register ResVReg, SPIRVTypeInst ResType,
379 MachineInstr &I) const;
380
381 bool selectFpowi(Register ResVReg, SPIRVTypeInst ResType,
382 MachineInstr &I) const;
383
384 bool selectSaturate(Register ResVReg, SPIRVTypeInst ResType,
385 MachineInstr &I) const;
386
387 bool selectWaveOpInst(Register ResVReg, SPIRVTypeInst ResType,
388 MachineInstr &I, unsigned Opcode) const;
389
390 bool selectBarrierInst(MachineInstr &I, unsigned Scope, unsigned MemSem,
391 bool WithGroupSync) const;
392
393 bool selectWaveActiveCountBits(Register ResVReg, SPIRVTypeInst ResType,
394 MachineInstr &I) const;
395
396 bool selectWaveActiveAllEqual(Register ResVReg, SPIRVTypeInst ResType,
397 MachineInstr &I) const;
398
399 bool selectUnmergeValues(MachineInstr &I) const;
400
401 bool selectHandleFromBinding(Register &ResVReg, SPIRVTypeInst ResType,
402 MachineInstr &I) const;
403
404 bool selectCounterHandleFromBinding(Register &ResVReg, SPIRVTypeInst ResType,
405 MachineInstr &I) const;
406
407 bool selectReadImageIntrinsic(Register &ResVReg, SPIRVTypeInst ResType,
408 MachineInstr &I) const;
409 bool selectGetDimensionsIntrinsic(Register &ResVReg, SPIRVTypeInst ResType,
410 MachineInstr &I) const;
411 bool selectGetDimensionsLevelsIntrinsic(Register &ResVReg,
412 SPIRVTypeInst ResType,
413 MachineInstr &I) const;
414 bool selectGetDimensionsMSIntrinsic(Register &ResVReg, SPIRVTypeInst ResType,
415 MachineInstr &I) const;
416 bool
417 selectImageQuerySize(Register ImageReg, Register &ResVReg, MachineInstr &I,
418 std::optional<Register> LodReg = std::nullopt) const;
419 bool selectSampleBasicIntrinsic(Register &ResVReg, SPIRVTypeInst ResType,
420 MachineInstr &I) const;
421 bool selectCalculateLodIntrinsic(Register &ResVReg, SPIRVTypeInst ResType,
422 MachineInstr &I) const;
423 bool selectSampleBiasIntrinsic(Register &ResVReg, SPIRVTypeInst ResType,
424 MachineInstr &I) const;
425 bool selectSampleGradIntrinsic(Register &ResVReg, SPIRVTypeInst ResType,
426 MachineInstr &I) const;
427 bool selectSampleLevelIntrinsic(Register &ResVReg, SPIRVTypeInst ResType,
428 MachineInstr &I) const;
429 bool selectLoadLevelIntrinsic(Register &ResVReg, SPIRVTypeInst ResType,
430 MachineInstr &I) const;
431 bool selectSampleCmpIntrinsic(Register &ResVReg, SPIRVTypeInst ResType,
432 MachineInstr &I) const;
433 bool selectSampleCmpLevelZeroIntrinsic(Register &ResVReg,
434 SPIRVTypeInst ResType,
435 MachineInstr &I) const;
436 bool selectGatherIntrinsic(Register &ResVReg, SPIRVTypeInst ResType,
437 MachineInstr &I) const;
438 bool selectImageWriteIntrinsic(MachineInstr &I) const;
439 bool selectResourceGetPointer(Register &ResVReg, SPIRVTypeInst ResType,
440 MachineInstr &I) const;
441 bool selectPushConstantGetPointer(Register &ResVReg, SPIRVTypeInst ResType,
442 MachineInstr &I) const;
443 bool selectResourceNonUniformIndex(Register &ResVReg, SPIRVTypeInst ResType,
444 MachineInstr &I) const;
445 bool selectModf(Register ResVReg, SPIRVTypeInst ResType,
446 MachineInstr &I) const;
447 bool selectUpdateCounter(Register &ResVReg, SPIRVTypeInst ResType,
448 MachineInstr &I) const;
449 bool selectFrexp(Register ResVReg, SPIRVTypeInst ResType,
450 MachineInstr &I) const;
451 bool selectSincos(Register ResVReg, SPIRVTypeInst ResType,
452 MachineInstr &I) const;
453 bool selectExp10(Register ResVReg, SPIRVTypeInst ResType,
454 MachineInstr &I) const;
455 bool selectDerivativeInst(Register ResVReg, SPIRVTypeInst ResType,
456 MachineInstr &I, const unsigned DPdOpCode) const;
457 // Utilities
458 Register buildI32Constant(uint32_t Val, MachineInstr &I,
459 SPIRVTypeInst ResType = nullptr) const;
460
461 Register buildZerosVal(SPIRVTypeInst ResType, MachineInstr &I) const;
462 bool isScalarOrVectorIntConstantZero(Register Reg) const;
463 Register buildZerosValF(SPIRVTypeInst ResType, MachineInstr &I) const;
464 Register buildOnesVal(bool AllOnes, SPIRVTypeInst ResType,
465 MachineInstr &I) const;
466 Register buildOnesValF(SPIRVTypeInst ResType, MachineInstr &I) const;
467
468 bool wrapIntoSpecConstantOp(MachineInstr &I,
469 SmallVector<Register> &CompositeArgs) const;
470
471 Register getUcharPtrTypeReg(MachineInstr &I,
472 SPIRV::StorageClass::StorageClass SC) const;
473 MachineInstrBuilder buildSpecConstantOp(MachineInstr &I, Register Dest,
474 Register Src, Register DestType,
475 uint32_t Opcode) const;
476 MachineInstrBuilder buildConstGenericPtr(MachineInstr &I, Register SrcPtr,
477 SPIRVTypeInst SrcPtrTy) const;
478 Register buildPointerToResource(SPIRVTypeInst ResType,
479 SPIRV::StorageClass::StorageClass SC,
480 uint32_t Set, uint32_t Binding,
481 uint32_t ArraySize, Register IndexReg,
482 StringRef Name,
483 MachineIRBuilder MIRBuilder) const;
484 SPIRVTypeInst widenTypeToVec4(SPIRVTypeInst Type, MachineInstr &I) const;
485 bool extractSubvector(Register &ResVReg, SPIRVTypeInst ResType,
486 Register &ReadReg, MachineInstr &InsertionPoint) const;
487 bool generateImageReadOrFetch(Register &ResVReg, SPIRVTypeInst ResType,
488 Register ImageReg, Register IdxReg,
489 DebugLoc Loc, MachineInstr &Pos,
490 const ImageOperands *ImOps = nullptr) const;
491 bool generateSampleImage(Register ResVReg, SPIRVTypeInst ResType,
492 Register ImageReg, Register SamplerReg,
493 Register CoordinateReg, const ImageOperands &ImOps,
494 DebugLoc Loc, MachineInstr &I) const;
495 bool BuildCOPY(Register DestReg, Register SrcReg, MachineInstr &I) const;
496 bool loadVec3BuiltinInputID(SPIRV::BuiltIn::BuiltIn BuiltInValue,
497 Register ResVReg, SPIRVTypeInst ResType,
498 MachineInstr &I) const;
499 bool loadBuiltinInputID(SPIRV::BuiltIn::BuiltIn BuiltInValue,
500 Register ResVReg, SPIRVTypeInst ResType,
501 MachineInstr &I) const;
502 bool loadHandleBeforePosition(Register &HandleReg, SPIRVTypeInst ResType,
503 GIntrinsic &HandleDef, MachineInstr &Pos) const;
504 void decorateUsesAsNonUniform(Register &NonUniformReg) const;
505 void errorIfInstrOutsideShader(MachineInstr &I) const;
506
507 std::optional<SplitParts> splitEvenOddLanes(Register PopCountReg,
508 unsigned ComponentCount,
509 MachineInstr &I,
510 SPIRVTypeInst I32Type) const;
511
512 bool
513 handle64BitOverflow(Register ResVReg, SPIRVTypeInst ResType, MachineInstr &I,
514 Register SrcReg, unsigned int Opcode,
515 std::function<bool(Register, SPIRVTypeInst,
516 MachineInstr &, Register, unsigned)>
517 CallbackFunction) const;
518};
519
520bool sampledTypeIsSignedInteger(const llvm::Type *HandleType) {
521 const TargetExtType *TET = cast<TargetExtType>(HandleType);
522 if (TET->getTargetExtName() == "spirv.Image") {
523 return false;
524 }
525 assert(TET->getTargetExtName() == "spirv.SignedImage");
526 return TET->getTypeParameter(0)->isIntegerTy();
527}
528} // end anonymous namespace
529
530#define GET_GLOBALISEL_IMPL
531#include "SPIRVGenGlobalISel.inc"
532#undef GET_GLOBALISEL_IMPL
533
534SPIRVInstructionSelector::SPIRVInstructionSelector(const SPIRVTargetMachine &TM,
535 const SPIRVSubtarget &ST,
536 const RegisterBankInfo &RBI)
537 : InstructionSelector(), STI(ST), TII(*ST.getInstrInfo()),
538 TRI(*ST.getRegisterInfo()), RBI(RBI), GR(*ST.getSPIRVGlobalRegistry()),
539 MRI(nullptr),
540#define GET_GLOBALISEL_PREDICATES_INIT
541#include "SPIRVGenGlobalISel.inc"
542#undef GET_GLOBALISEL_PREDICATES_INIT
543#define GET_GLOBALISEL_TEMPORARIES_INIT
544#include "SPIRVGenGlobalISel.inc"
545#undef GET_GLOBALISEL_TEMPORARIES_INIT
546{
547}
548
549void SPIRVInstructionSelector::setupMF(MachineFunction &MF,
550 GISelValueTracking *VT,
551 CodeGenCoverage *CoverageInfo,
552 ProfileSummaryInfo *PSI,
553 BlockFrequencyInfo *BFI) {
554 MRI = &MF.getRegInfo();
555 GR.setCurrentFunc(MF);
556 InstructionSelector::setupMF(MF, VT, CoverageInfo, PSI, BFI);
557}
558
559// Ensure that register classes correspond to pattern matching rules.
560void SPIRVInstructionSelector::resetVRegsType(MachineFunction &MF) {
561 if (HasVRegsReset == &MF)
562 return;
563 HasVRegsReset = &MF;
564
565 MachineRegisterInfo &MRI = MF.getRegInfo();
566 for (unsigned I = 0, E = MRI.getNumVirtRegs(); I != E; ++I) {
567 Register Reg = Register::index2VirtReg(I);
568 LLT RegType = MRI.getType(Reg);
569 if (RegType.isScalar())
570 MRI.setType(Reg, LLT::scalar(64));
571 else if (RegType.isPointer())
572 MRI.setType(Reg, LLT::pointer(0, 64));
573 else if (RegType.isVector())
574 MRI.setType(Reg, LLT::fixed_vector(2, LLT::scalar(64)));
575 }
576 for (const auto &MBB : MF) {
577 for (const auto &MI : MBB) {
578 if (isPreISelGenericOpcode(MI.getOpcode()))
579 GR.erase(&MI);
580 if (MI.getOpcode() != SPIRV::ASSIGN_TYPE)
581 continue;
582
583 Register DstReg = MI.getOperand(0).getReg();
584 LLT DstType = MRI.getType(DstReg);
585 Register SrcReg = MI.getOperand(1).getReg();
586 LLT SrcType = MRI.getType(SrcReg);
587 if (DstType != SrcType)
588 MRI.setType(DstReg, MRI.getType(SrcReg));
589
590 const TargetRegisterClass *DstRC = MRI.getRegClassOrNull(DstReg);
591 const TargetRegisterClass *SrcRC = MRI.getRegClassOrNull(SrcReg);
592 if (DstRC != SrcRC && SrcRC)
593 MRI.setRegClass(DstReg, SrcRC);
594 }
595 }
596}
597
598// Return true if the MachineInstr represents a constant register
599static bool isConstReg(MachineRegisterInfo *MRI, MachineInstr *OpDef) {
600
601 SmallVector<MachineInstr *> Stack = {OpDef};
602 SmallPtrSet<MachineInstr *, 4> Visited;
603
604 while (!Stack.empty()) {
605 MachineInstr *MI = Stack.pop_back_val();
606 MI = passCopy(MI, MRI);
607 if (!Visited.insert(MI).second)
608 continue;
609 switch (MI->getOpcode()) {
610 case TargetOpcode::G_INTRINSIC:
611 case TargetOpcode::G_INTRINSIC_W_SIDE_EFFECTS:
612 case TargetOpcode::G_INTRINSIC_CONVERGENT_W_SIDE_EFFECTS: {
613 GIntrinsic *GIntr = cast<GIntrinsic>(MI);
614 unsigned IntrID = GIntr->getIntrinsicID();
615 if (IntrID != Intrinsic::spv_const_composite &&
616 IntrID != Intrinsic::spv_undef)
617 return false;
618 continue;
619 }
620 case TargetOpcode::G_BUILD_VECTOR:
621 case TargetOpcode::G_SPLAT_VECTOR:
622 for (unsigned i = OpDef->getNumExplicitDefs();
623 i < OpDef->getNumOperands(); i++) {
624 if (!OpDef->getOperand(i).isReg())
625 continue;
626 MachineInstr *OpNestedDef =
627 MRI->getVRegDef(OpDef->getOperand(i).getReg());
628 Stack.push_back(OpNestedDef);
629 }
630 continue;
631 case TargetOpcode::G_CONSTANT:
632 case TargetOpcode::G_FCONSTANT:
633 case TargetOpcode::G_IMPLICIT_DEF:
634 case SPIRV::OpConstantTrue:
635 case SPIRV::OpConstantFalse:
636 case SPIRV::OpConstantI:
637 case SPIRV::OpConstantF:
638 case SPIRV::OpConstantComposite:
639 case SPIRV::OpConstantCompositeContinuedINTEL:
640 case SPIRV::OpConstantSampler:
641 case SPIRV::OpConstantNull:
642 case SPIRV::OpUndef:
643 case SPIRV::OpConstantFunctionPointerINTEL:
644 continue;
645 default:
646 return false;
647 }
648 }
649 return true;
650}
651
652// Return true if the virtual register represents a constant
653static bool isConstReg(MachineRegisterInfo *MRI, Register OpReg) {
654 if (MachineInstr *OpDef = MRI->getVRegDef(OpReg))
655 return isConstReg(MRI, OpDef);
656 return false;
657}
658
659// TODO(168736): We should make this either a flag in tabelgen
660// or reduce our dependence on the global registry, so we can remove this
661// function. It can easily be missed when new intrinsics are added.
662
663// Most SPIR-V intrinsics are considered to have side-effects in their tablegen
664// definition because they are referenced in the global registry. This is a list
665// of intrinsics that have no side effects other than their references in the
666// global registry.
667static bool intrinsicHasSideEffects(Intrinsic::ID ID) {
668 switch (ID) {
669 // This is not an exhaustive list and may need to be updated.
670 case Intrinsic::spv_all:
671 case Intrinsic::spv_alloca:
672 case Intrinsic::spv_any:
673 case Intrinsic::spv_bitcast:
674 case Intrinsic::spv_const_composite:
675 case Intrinsic::spv_cross:
676 case Intrinsic::spv_degrees:
677 case Intrinsic::spv_distance:
678 case Intrinsic::spv_extractelt:
679 case Intrinsic::spv_extractv:
680 case Intrinsic::spv_faceforward:
681 case Intrinsic::spv_fdot:
682 case Intrinsic::spv_firstbitlow:
683 case Intrinsic::spv_firstbitshigh:
684 case Intrinsic::spv_firstbituhigh:
685 case Intrinsic::spv_frac:
686 case Intrinsic::spv_gep:
687 case Intrinsic::spv_global_offset:
688 case Intrinsic::spv_global_size:
689 case Intrinsic::spv_group_id:
690 case Intrinsic::spv_insertelt:
691 case Intrinsic::spv_insertv:
692 case Intrinsic::spv_isinf:
693 case Intrinsic::spv_isnan:
694 case Intrinsic::spv_lerp:
695 case Intrinsic::spv_length:
696 case Intrinsic::spv_normalize:
697 case Intrinsic::spv_num_subgroups:
698 case Intrinsic::spv_num_workgroups:
699 case Intrinsic::spv_ptrcast:
700 case Intrinsic::spv_radians:
701 case Intrinsic::spv_reflect:
702 case Intrinsic::spv_refract:
703 case Intrinsic::spv_resource_getpointer:
704 case Intrinsic::spv_resource_handlefrombinding:
705 case Intrinsic::spv_resource_handlefromimplicitbinding:
706 case Intrinsic::spv_resource_nonuniformindex:
707 case Intrinsic::spv_resource_sample:
708 case Intrinsic::spv_rsqrt:
709 case Intrinsic::spv_saturate:
710 case Intrinsic::spv_sdot:
711 case Intrinsic::spv_sign:
712 case Intrinsic::spv_smoothstep:
713 case Intrinsic::spv_step:
714 case Intrinsic::spv_subgroup_id:
715 case Intrinsic::spv_subgroup_local_invocation_id:
716 case Intrinsic::spv_subgroup_max_size:
717 case Intrinsic::spv_subgroup_size:
718 case Intrinsic::spv_thread_id:
719 case Intrinsic::spv_thread_id_in_group:
720 case Intrinsic::spv_udot:
721 case Intrinsic::spv_undef:
722 case Intrinsic::spv_value_md:
723 case Intrinsic::spv_workgroup_size:
724 return false;
725 default:
726 return true;
727 }
728}
729
730// TODO(168736): We should make this either a flag in tabelgen
731// or reduce our dependence on the global registry, so we can remove this
732// function. It can easily be missed when new intrinsics are added.
733static bool isOpcodeWithNoSideEffects(unsigned Opcode) {
734 switch (Opcode) {
735 case SPIRV::OpTypeVoid:
736 case SPIRV::OpTypeBool:
737 case SPIRV::OpTypeInt:
738 case SPIRV::OpTypeFloat:
739 case SPIRV::OpTypeVector:
740 case SPIRV::OpTypeMatrix:
741 case SPIRV::OpTypeImage:
742 case SPIRV::OpTypeSampler:
743 case SPIRV::OpTypeSampledImage:
744 case SPIRV::OpTypeArray:
745 case SPIRV::OpTypeRuntimeArray:
746 case SPIRV::OpTypeStruct:
747 case SPIRV::OpTypeOpaque:
748 case SPIRV::OpTypePointer:
749 case SPIRV::OpTypeFunction:
750 case SPIRV::OpTypeEvent:
751 case SPIRV::OpTypeDeviceEvent:
752 case SPIRV::OpTypeReserveId:
753 case SPIRV::OpTypeQueue:
754 case SPIRV::OpTypePipe:
755 case SPIRV::OpTypeForwardPointer:
756 case SPIRV::OpTypePipeStorage:
757 case SPIRV::OpTypeNamedBarrier:
758 case SPIRV::OpTypeAccelerationStructureNV:
759 case SPIRV::OpTypeCooperativeMatrixNV:
760 case SPIRV::OpTypeCooperativeMatrixKHR:
761 return true;
762 default:
763 return false;
764 }
765}
766
767bool isDead(const MachineInstr &MI, const MachineRegisterInfo &MRI) {
768 // If there are no definitions, then assume there is some other
769 // side-effect that makes this instruction live.
770 if (MI.getNumDefs() == 0)
771 return false;
772
773 for (const auto &MO : MI.all_defs()) {
774 Register Reg = MO.getReg();
775 if (Reg.isPhysical()) {
776 LLVM_DEBUG(dbgs() << "Not dead: def of physical register " << Reg);
777 return false;
778 }
779 for (const auto &UseMI : MRI.use_nodbg_instructions(Reg)) {
780 if (UseMI.getOpcode() != SPIRV::OpName) {
781 LLVM_DEBUG(dbgs() << "Not dead: def " << MO << " has use in " << UseMI);
782 return false;
783 }
784 }
785 }
786
787 if (MI.getOpcode() == TargetOpcode::LOCAL_ESCAPE || MI.isFakeUse() ||
788 MI.isLifetimeMarker()) {
789 LLVM_DEBUG(
790 dbgs()
791 << "Not dead: Opcode is LOCAL_ESCAPE, fake use, or lifetime marker.\n");
792 return false;
793 }
794 if (MI.isPHI()) {
795 LLVM_DEBUG(dbgs() << "Dead: Phi instruction with no uses.\n");
796 return true;
797 }
798
799 // It is possible that the only side effect is that the instruction is
800 // referenced in the global registry. If that is the only side effect, the
801 // intrinsic is dead.
802 if (MI.getOpcode() == TargetOpcode::G_INTRINSIC_W_SIDE_EFFECTS ||
803 MI.getOpcode() == TargetOpcode::G_INTRINSIC_CONVERGENT_W_SIDE_EFFECTS) {
804 const auto &Intr = cast<GIntrinsic>(MI);
805 if (!intrinsicHasSideEffects(Intr.getIntrinsicID())) {
806 LLVM_DEBUG(dbgs() << "Dead: Intrinsic with no real side effects.\n");
807 return true;
808 }
809 }
810
811 if (MI.mayStore() || MI.isCall() ||
812 (MI.mayLoad() && MI.hasOrderedMemoryRef()) || MI.isPosition() ||
813 MI.isDebugInstr() || MI.isTerminator() || MI.isJumpTableDebugInfo()) {
814 LLVM_DEBUG(dbgs() << "Not dead: instruction has side effects.\n");
815 return false;
816 }
817
818 if (isPreISelGenericOpcode(MI.getOpcode())) {
819 // TODO: Is there a generic way to check if the opcode has side effects?
820 LLVM_DEBUG(dbgs() << "Dead: Generic opcode with no uses.\n");
821 return true;
822 }
823
824 if (isOpcodeWithNoSideEffects(MI.getOpcode())) {
825 LLVM_DEBUG(dbgs() << "Dead: known opcode with no side effects\n");
826 return true;
827 }
828
829 return false;
830}
831
832void SPIRVInstructionSelector::removeOpNamesForDeadMI(MachineInstr &MI) const {
833 // Delete the OpName that uses the result if there is one.
834 for (const auto &MO : MI.all_defs()) {
835 Register Reg = MO.getReg();
836 if (Reg.isPhysical())
837 continue;
838 SmallVector<MachineInstr *, 4> UselessOpNames;
839 for (MachineInstr &UseMI : MRI->use_nodbg_instructions(Reg)) {
840 assert(UseMI.getOpcode() == SPIRV::OpName &&
841 "There is still a use of the dead function.");
842 UselessOpNames.push_back(&UseMI);
843 }
844 for (MachineInstr *OpNameMI : UselessOpNames) {
845 GR.invalidateMachineInstr(OpNameMI);
846 OpNameMI->eraseFromParent();
847 }
848 }
849}
850
851void SPIRVInstructionSelector::removeDeadInstruction(MachineInstr &MI) const {
852 salvageDebugInfo(*MRI, MI);
853 GR.invalidateMachineInstr(&MI);
854 removeOpNamesForDeadMI(MI);
855 MI.eraseFromParent();
856}
857
858bool SPIRVInstructionSelector::select(MachineInstr &I) {
859 resetVRegsType(*I.getParent()->getParent());
860
861 assert(I.getParent() && "Instruction should be in a basic block!");
862 assert(I.getParent()->getParent() && "Instruction should be in a function!");
863
864 LLVM_DEBUG(dbgs() << "Checking if instruction is dead: " << I;);
865 if (isDead(I, *MRI)) {
866 LLVM_DEBUG(dbgs() << "Instruction is dead.\n");
867 removeDeadInstruction(I);
868 return true;
869 }
870
871 Register Opcode = I.getOpcode();
872 // If it's not a GMIR instruction, we've selected it already.
873 if (!isPreISelGenericOpcode(Opcode)) {
874 if (Opcode == SPIRV::ASSIGN_TYPE) { // These pseudos aren't needed any more.
875 Register DstReg = I.getOperand(0).getReg();
876 Register SrcReg = I.getOperand(1).getReg();
877 auto *Def = MRI->getVRegDef(SrcReg);
878 if (isTypeFoldingSupported(Def->getOpcode()) &&
879 Def->getOpcode() != TargetOpcode::G_CONSTANT &&
880 Def->getOpcode() != TargetOpcode::G_FCONSTANT) {
881 if (Def->getOpcode() == TargetOpcode::G_SELECT) {
882 Register SelectDstReg = Def->getOperand(0).getReg();
883 bool SuccessToSelectSelect [[maybe_unused]] = selectSelect(
884 SelectDstReg, GR.getSPIRVTypeForVReg(SelectDstReg), *Def);
885 assert(SuccessToSelectSelect);
886 GR.invalidateMachineInstr(Def);
887 Def->eraseFromParent();
888 MRI->replaceRegWith(DstReg, SelectDstReg);
889 GR.invalidateMachineInstr(&I);
890 I.eraseFromParent();
891 return true;
892 }
893
894 bool Res = selectImpl(I, *CoverageInfo);
895 LLVM_DEBUG({
896 if (!Res && Def->getOpcode() != TargetOpcode::G_CONSTANT) {
897 dbgs() << "Unexpected pattern in ASSIGN_TYPE.\nInstruction: ";
898 I.print(dbgs());
899 }
900 });
901 assert(Res || Def->getOpcode() == TargetOpcode::G_CONSTANT);
902 if (Res) {
903 if (!isTriviallyDead(*Def, *MRI) && isDead(*Def, *MRI))
904 DeadMIs.insert(Def);
905 return Res;
906 }
907 }
908 MRI->setRegClass(SrcReg, MRI->getRegClass(DstReg));
909 MRI->replaceRegWith(SrcReg, DstReg);
910 GR.invalidateMachineInstr(&I);
911 I.eraseFromParent();
912 return true;
913 } else if (I.getNumDefs() == 1) {
914 // Make all vregs 64 bits (for SPIR-V IDs).
915 MRI->setType(I.getOperand(0).getReg(), LLT::scalar(64));
916 }
917 constrainSelectedInstRegOperands(I, TII, TRI, RBI);
918 return true;
919 }
920
921 if (DeadMIs.contains(&I)) {
922 // if the instruction has been already made dead by folding it away
923 // erase it
924 LLVM_DEBUG(dbgs() << "Instruction is folded and dead.\n");
925 removeDeadInstruction(I);
926 DeadMIs.erase(&I);
927 return true;
928 }
929
930 if (I.getNumOperands() != I.getNumExplicitOperands()) {
931 LLVM_DEBUG(errs() << "Generic instr has unexpected implicit operands\n");
932 return false;
933 }
934
935 // Common code for getting return reg+type, and removing selected instr
936 // from parent occurs here. Instr-specific selection happens in spvSelect().
937 bool HasDefs = I.getNumDefs() > 0;
938 Register ResVReg = HasDefs ? I.getOperand(0).getReg() : Register(0);
939 SPIRVTypeInst ResType = HasDefs ? GR.getSPIRVTypeForVReg(ResVReg) : nullptr;
940 assert(!HasDefs || ResType || I.getOpcode() == TargetOpcode::G_GLOBAL_VALUE ||
941 I.getOpcode() == TargetOpcode::G_IMPLICIT_DEF);
942 if (spvSelect(ResVReg, ResType, I)) {
943 if (HasDefs) // Make all vregs 64 bits (for SPIR-V IDs).
944 for (unsigned i = 0; i < I.getNumDefs(); ++i)
945 MRI->setType(I.getOperand(i).getReg(), LLT::scalar(64));
946 GR.invalidateMachineInstr(&I);
947 I.eraseFromParent();
948 return true;
949 }
950 return false;
951}
952
953static bool mayApplyGenericSelection(unsigned Opcode) {
954 switch (Opcode) {
955 case TargetOpcode::G_CONSTANT:
956 case TargetOpcode::G_FCONSTANT:
957 return false;
958 }
959 return isTypeFoldingSupported(Opcode);
960}
961
962bool SPIRVInstructionSelector::BuildCOPY(Register DestReg, Register SrcReg,
963 MachineInstr &I) const {
964 const TargetRegisterClass *DstRC = MRI->getRegClassOrNull(DestReg);
965 const TargetRegisterClass *SrcRC = MRI->getRegClassOrNull(SrcReg);
966 if (DstRC != SrcRC && SrcRC)
967 MRI->setRegClass(DestReg, SrcRC);
968 BuildMI(*I.getParent(), I, I.getDebugLoc(), TII.get(TargetOpcode::COPY))
969 .addDef(DestReg)
970 .addUse(SrcReg)
971 .constrainAllUses(TII, TRI, RBI);
972 return true;
973}
974
975bool SPIRVInstructionSelector::spvSelect(Register ResVReg,
976 SPIRVTypeInst ResType,
977 MachineInstr &I) const {
978 const unsigned Opcode = I.getOpcode();
979 if (mayApplyGenericSelection(Opcode))
980 return selectImpl(I, *CoverageInfo);
981 switch (Opcode) {
982 case TargetOpcode::G_CONSTANT:
983 case TargetOpcode::G_FCONSTANT:
984 return selectConst(ResVReg, ResType, I);
985 case TargetOpcode::G_GLOBAL_VALUE:
986 return selectGlobalValue(ResVReg, I);
987 case TargetOpcode::G_IMPLICIT_DEF:
988 return selectOpUndef(ResVReg, ResType, I);
989 case TargetOpcode::G_FREEZE:
990 return selectFreeze(ResVReg, ResType, I);
991
992 case TargetOpcode::G_INTRINSIC:
993 case TargetOpcode::G_INTRINSIC_W_SIDE_EFFECTS:
994 case TargetOpcode::G_INTRINSIC_CONVERGENT:
995 case TargetOpcode::G_INTRINSIC_CONVERGENT_W_SIDE_EFFECTS:
996 return selectIntrinsic(ResVReg, ResType, I);
997 case TargetOpcode::G_BITREVERSE:
998 return selectBitreverse(ResVReg, ResType, I);
999
1000 case TargetOpcode::G_BUILD_VECTOR:
1001 return selectBuildVector(ResVReg, ResType, I);
1002 case TargetOpcode::G_SPLAT_VECTOR:
1003 return selectSplatVector(ResVReg, ResType, I);
1004
1005 case TargetOpcode::G_SHUFFLE_VECTOR: {
1006 MachineBasicBlock &BB = *I.getParent();
1007 auto MIB = BuildMI(BB, I, I.getDebugLoc(), TII.get(SPIRV::OpVectorShuffle))
1008 .addDef(ResVReg)
1009 .addUse(GR.getSPIRVTypeID(ResType))
1010 .addUse(I.getOperand(1).getReg())
1011 .addUse(I.getOperand(2).getReg());
1012 for (auto V : I.getOperand(3).getShuffleMask())
1013 MIB.addImm(V);
1014 MIB.constrainAllUses(TII, TRI, RBI);
1015 return true;
1016 }
1017 case TargetOpcode::G_MEMMOVE:
1018 case TargetOpcode::G_MEMCPY:
1019 case TargetOpcode::G_MEMSET:
1020 return selectMemOperation(ResVReg, I);
1021
1022 case TargetOpcode::G_ICMP:
1023 return selectICmp(ResVReg, ResType, I);
1024 case TargetOpcode::G_FCMP:
1025 return selectFCmp(ResVReg, ResType, I);
1026
1027 case TargetOpcode::G_FRAME_INDEX:
1028 return selectFrameIndex(ResVReg, ResType, I);
1029
1030 case TargetOpcode::G_LOAD:
1031 return selectLoad(ResVReg, ResType, I);
1032 case TargetOpcode::G_STORE:
1033 return selectStore(I);
1034
1035 case TargetOpcode::G_BR:
1036 return selectBranch(I);
1037 case TargetOpcode::G_BRCOND:
1038 return selectBranchCond(I);
1039
1040 case TargetOpcode::G_PHI:
1041 return selectPhi(ResVReg, I);
1042
1043 case TargetOpcode::G_FPTOSI:
1044 return selectUnOp(ResVReg, ResType, I, SPIRV::OpConvertFToS);
1045 case TargetOpcode::G_FPTOUI:
1046 return selectUnOp(ResVReg, ResType, I, SPIRV::OpConvertFToU);
1047
1048 case TargetOpcode::G_FPTOSI_SAT:
1049 return selectUnOp(ResVReg, ResType, I, SPIRV::OpConvertFToS);
1050 case TargetOpcode::G_FPTOUI_SAT:
1051 return selectUnOp(ResVReg, ResType, I, SPIRV::OpConvertFToU);
1052
1053 case TargetOpcode::G_SITOFP:
1054 return selectIToF(ResVReg, ResType, I, true, SPIRV::OpConvertSToF);
1055 case TargetOpcode::G_UITOFP:
1056 return selectIToF(ResVReg, ResType, I, false, SPIRV::OpConvertUToF);
1057
1058 case TargetOpcode::G_CTPOP:
1059 return selectPopCount(ResVReg, ResType, I, SPIRV::OpBitCount);
1060 case TargetOpcode::G_SMIN:
1061 return selectExtInst(ResVReg, ResType, I, CL::s_min, GL::SMin);
1062 case TargetOpcode::G_UMIN:
1063 return selectExtInst(ResVReg, ResType, I, CL::u_min, GL::UMin);
1064
1065 case TargetOpcode::G_SMAX:
1066 return selectExtInst(ResVReg, ResType, I, CL::s_max, GL::SMax);
1067 case TargetOpcode::G_UMAX:
1068 return selectExtInst(ResVReg, ResType, I, CL::u_max, GL::UMax);
1069
1070 case TargetOpcode::G_SCMP:
1071 return selectSUCmp(ResVReg, ResType, I, true);
1072 case TargetOpcode::G_UCMP:
1073 return selectSUCmp(ResVReg, ResType, I, false);
1074 case TargetOpcode::G_LROUND:
1075 case TargetOpcode::G_LLROUND: {
1076 Register regForLround =
1077 MRI->createVirtualRegister(MRI->getRegClass(ResVReg), "lround");
1078 MRI->setRegClass(regForLround, &SPIRV::iIDRegClass);
1079 GR.assignSPIRVTypeToVReg(GR.getSPIRVTypeForVReg(I.getOperand(1).getReg()),
1080 regForLround, *(I.getParent()->getParent()));
1081 selectExtInst(regForLround, GR.getSPIRVTypeForVReg(regForLround), I,
1082 CL::round, GL::Round, /* setMIFlags */ false);
1083 MachineBasicBlock &BB = *I.getParent();
1084 auto MIB = BuildMI(BB, I, I.getDebugLoc(), TII.get(SPIRV::OpConvertFToS))
1085 .addDef(ResVReg)
1086 .addUse(GR.getSPIRVTypeID(ResType))
1087 .addUse(regForLround);
1088 MIB.constrainAllUses(TII, TRI, RBI);
1089 return true;
1090 }
1091 case TargetOpcode::G_STRICT_FMA:
1092 case TargetOpcode::G_FMA: {
1093 if (STI.canUseExtension(SPIRV::Extension::SPV_KHR_fma)) {
1094 MachineBasicBlock &BB = *I.getParent();
1095 auto MIB = BuildMI(BB, I, I.getDebugLoc(), TII.get(SPIRV::OpFmaKHR))
1096 .addDef(ResVReg)
1097 .addUse(GR.getSPIRVTypeID(ResType))
1098 .addUse(I.getOperand(1).getReg())
1099 .addUse(I.getOperand(2).getReg())
1100 .addUse(I.getOperand(3).getReg())
1101 .setMIFlags(I.getFlags());
1102 MIB.constrainAllUses(TII, TRI, RBI);
1103 return true;
1104 }
1105 return selectExtInst(ResVReg, ResType, I, CL::fma, GL::Fma);
1106 }
1107
1108 case TargetOpcode::G_STRICT_FLDEXP:
1109 return selectExtInst(ResVReg, ResType, I, CL::ldexp);
1110
1111 case TargetOpcode::G_FPOW:
1112 return selectExtInst(ResVReg, ResType, I, CL::pow, GL::Pow);
1113 case TargetOpcode::G_FPOWI:
1114 return selectFpowi(ResVReg, ResType, I);
1115
1116 case TargetOpcode::G_FEXP:
1117 return selectExtInst(ResVReg, ResType, I, CL::exp, GL::Exp);
1118 case TargetOpcode::G_FEXP2:
1119 return selectExtInst(ResVReg, ResType, I, CL::exp2, GL::Exp2);
1120 case TargetOpcode::G_FEXP10:
1121 return selectExp10(ResVReg, ResType, I);
1122
1123 case TargetOpcode::G_FMODF:
1124 return selectModf(ResVReg, ResType, I);
1125 case TargetOpcode::G_FSINCOS:
1126 return selectSincos(ResVReg, ResType, I);
1127
1128 case TargetOpcode::G_FLOG:
1129 return selectExtInst(ResVReg, ResType, I, CL::log, GL::Log);
1130 case TargetOpcode::G_FLOG2:
1131 return selectExtInst(ResVReg, ResType, I, CL::log2, GL::Log2);
1132 case TargetOpcode::G_FLOG10:
1133 return selectLog10(ResVReg, ResType, I);
1134
1135 case TargetOpcode::G_FABS:
1136 return selectExtInst(ResVReg, ResType, I, CL::fabs, GL::FAbs);
1137 case TargetOpcode::G_ABS:
1138 return selectExtInst(ResVReg, ResType, I, CL::s_abs, GL::SAbs);
1139
1140 case TargetOpcode::G_FMINNUM:
1141 case TargetOpcode::G_FMINIMUM:
1142 return selectExtInst(ResVReg, ResType, I, CL::fmin, GL::NMin);
1143 case TargetOpcode::G_FMAXNUM:
1144 case TargetOpcode::G_FMAXIMUM:
1145 return selectExtInst(ResVReg, ResType, I, CL::fmax, GL::NMax);
1146
1147 case TargetOpcode::G_FCOPYSIGN:
1148 return selectExtInst(ResVReg, ResType, I, CL::copysign);
1149
1150 case TargetOpcode::G_FCEIL:
1151 return selectExtInst(ResVReg, ResType, I, CL::ceil, GL::Ceil);
1152 case TargetOpcode::G_FFLOOR:
1153 return selectExtInst(ResVReg, ResType, I, CL::floor, GL::Floor);
1154
1155 case TargetOpcode::G_FCOS:
1156 return selectExtInst(ResVReg, ResType, I, CL::cos, GL::Cos);
1157 case TargetOpcode::G_FSIN:
1158 return selectExtInst(ResVReg, ResType, I, CL::sin, GL::Sin);
1159 case TargetOpcode::G_FTAN:
1160 return selectExtInst(ResVReg, ResType, I, CL::tan, GL::Tan);
1161 case TargetOpcode::G_FACOS:
1162 return selectExtInst(ResVReg, ResType, I, CL::acos, GL::Acos);
1163 case TargetOpcode::G_FASIN:
1164 return selectExtInst(ResVReg, ResType, I, CL::asin, GL::Asin);
1165 case TargetOpcode::G_FATAN:
1166 return selectExtInst(ResVReg, ResType, I, CL::atan, GL::Atan);
1167 case TargetOpcode::G_FATAN2:
1168 return selectExtInst(ResVReg, ResType, I, CL::atan2, GL::Atan2);
1169 case TargetOpcode::G_FCOSH:
1170 return selectExtInst(ResVReg, ResType, I, CL::cosh, GL::Cosh);
1171 case TargetOpcode::G_FSINH:
1172 return selectExtInst(ResVReg, ResType, I, CL::sinh, GL::Sinh);
1173 case TargetOpcode::G_FTANH:
1174 return selectExtInst(ResVReg, ResType, I, CL::tanh, GL::Tanh);
1175
1176 case TargetOpcode::G_STRICT_FSQRT:
1177 case TargetOpcode::G_FSQRT:
1178 return selectExtInst(ResVReg, ResType, I, CL::sqrt, GL::Sqrt);
1179
1180 case TargetOpcode::G_CTTZ:
1181 case TargetOpcode::G_CTTZ_ZERO_UNDEF:
1182 return selectExtInst(ResVReg, ResType, I, CL::ctz);
1183 case TargetOpcode::G_CTLZ:
1184 case TargetOpcode::G_CTLZ_ZERO_UNDEF:
1185 return selectExtInst(ResVReg, ResType, I, CL::clz);
1186
1187 case TargetOpcode::G_INTRINSIC_ROUND:
1188 return selectExtInst(ResVReg, ResType, I, CL::round, GL::Round);
1189 case TargetOpcode::G_INTRINSIC_ROUNDEVEN:
1190 return selectExtInst(ResVReg, ResType, I, CL::rint, GL::RoundEven);
1191 case TargetOpcode::G_INTRINSIC_TRUNC:
1192 return selectExtInst(ResVReg, ResType, I, CL::trunc, GL::Trunc);
1193 case TargetOpcode::G_FRINT:
1194 case TargetOpcode::G_FNEARBYINT:
1195 return selectExtInst(ResVReg, ResType, I, CL::rint, GL::RoundEven);
1196
1197 case TargetOpcode::G_SMULH:
1198 return selectExtInst(ResVReg, ResType, I, CL::s_mul_hi);
1199 case TargetOpcode::G_UMULH:
1200 return selectExtInst(ResVReg, ResType, I, CL::u_mul_hi);
1201
1202 case TargetOpcode::G_SADDSAT:
1203 return selectExtInst(ResVReg, ResType, I, CL::s_add_sat);
1204 case TargetOpcode::G_UADDSAT:
1205 return selectExtInst(ResVReg, ResType, I, CL::u_add_sat);
1206 case TargetOpcode::G_SSUBSAT:
1207 return selectExtInst(ResVReg, ResType, I, CL::s_sub_sat);
1208 case TargetOpcode::G_USUBSAT:
1209 return selectExtInst(ResVReg, ResType, I, CL::u_sub_sat);
1210
1211 case TargetOpcode::G_FFREXP:
1212 return selectFrexp(ResVReg, ResType, I);
1213
1214 case TargetOpcode::G_UADDO:
1215 return selectOverflowArith(ResVReg, ResType, I,
1216 ResType->getOpcode() == SPIRV::OpTypeVector
1217 ? SPIRV::OpIAddCarryV
1218 : SPIRV::OpIAddCarryS);
1219 case TargetOpcode::G_USUBO:
1220 return selectOverflowArith(ResVReg, ResType, I,
1221 ResType->getOpcode() == SPIRV::OpTypeVector
1222 ? SPIRV::OpISubBorrowV
1223 : SPIRV::OpISubBorrowS);
1224 case TargetOpcode::G_UMULO:
1225 return selectOverflowArith(ResVReg, ResType, I, SPIRV::OpUMulExtended);
1226 case TargetOpcode::G_SMULO:
1227 return selectOverflowArith(ResVReg, ResType, I, SPIRV::OpSMulExtended);
1228
1229 case TargetOpcode::G_SEXT:
1230 return selectExt(ResVReg, ResType, I, true);
1231 case TargetOpcode::G_ANYEXT:
1232 case TargetOpcode::G_ZEXT:
1233 return selectExt(ResVReg, ResType, I, false);
1234 case TargetOpcode::G_TRUNC:
1235 return selectTrunc(ResVReg, ResType, I);
1236 case TargetOpcode::G_FPTRUNC:
1237 case TargetOpcode::G_FPEXT:
1238 return selectUnOp(ResVReg, ResType, I, SPIRV::OpFConvert);
1239
1240 case TargetOpcode::G_PTRTOINT:
1241 return selectUnOp(ResVReg, ResType, I, SPIRV::OpConvertPtrToU);
1242 case TargetOpcode::G_INTTOPTR:
1243 return selectUnOp(ResVReg, ResType, I, SPIRV::OpConvertUToPtr);
1244 case TargetOpcode::G_BITCAST:
1245 return selectBitcast(ResVReg, ResType, I);
1246 case TargetOpcode::G_ADDRSPACE_CAST:
1247 return selectAddrSpaceCast(ResVReg, ResType, I);
1248 case TargetOpcode::G_PTR_ADD: {
1249 // Currently, we get G_PTR_ADD only applied to global variables.
1250 assert(I.getOperand(1).isReg() && I.getOperand(2).isReg());
1251 Register GV = I.getOperand(1).getReg();
1252 MachineRegisterInfo::def_instr_iterator II = MRI->def_instr_begin(GV);
1253 (void)II;
1254 assert(((*II).getOpcode() == TargetOpcode::G_GLOBAL_VALUE ||
1255 (*II).getOpcode() == TargetOpcode::COPY ||
1256 (*II).getOpcode() == SPIRV::OpVariable) &&
1257 getImm(I.getOperand(2), MRI));
1258 // It may be the initialization of a global variable.
1259 bool IsGVInit = false;
1260 for (MachineRegisterInfo::use_instr_iterator
1261 UseIt = MRI->use_instr_begin(I.getOperand(0).getReg()),
1262 UseEnd = MRI->use_instr_end();
1263 UseIt != UseEnd; UseIt = std::next(UseIt)) {
1264 if ((*UseIt).getOpcode() == TargetOpcode::G_GLOBAL_VALUE ||
1265 (*UseIt).getOpcode() == SPIRV::OpSpecConstantOp ||
1266 (*UseIt).getOpcode() == SPIRV::OpVariable) {
1267 IsGVInit = true;
1268 break;
1269 }
1270 }
1271 MachineBasicBlock &BB = *I.getParent();
1272 if (!IsGVInit) {
1273 SPIRVTypeInst GVType = GR.getSPIRVTypeForVReg(GV);
1274 SPIRVTypeInst GVPointeeType = GR.getPointeeType(GVType);
1275 SPIRVTypeInst ResPointeeType = GR.getPointeeType(ResType);
1276 if (GVPointeeType && ResPointeeType && GVPointeeType != ResPointeeType) {
1277 // Build a new virtual register that is associated with the required
1278 // data type.
1279 Register NewVReg = MRI->createGenericVirtualRegister(MRI->getType(GV));
1280 MRI->setRegClass(NewVReg, MRI->getRegClass(GV));
1281 // Having a correctly typed base we are ready to build the actually
1282 // required GEP. It may not be a constant though, because all Operands
1283 // of OpSpecConstantOp is to originate from other const instructions,
1284 // and only the AccessChain named opcodes accept a global OpVariable
1285 // instruction. We can't use an AccessChain opcode because of the type
1286 // mismatch between result and base types.
1287 if (!GR.isBitcastCompatible(ResType, GVType))
1288 report_fatal_error(
1289 "incompatible result and operand types in a bitcast");
1290 Register ResTypeReg = GR.getSPIRVTypeID(ResType);
1291 MachineInstrBuilder MIB =
1292 BuildMI(BB, I, I.getDebugLoc(), TII.get(SPIRV::OpBitcast))
1293 .addDef(NewVReg)
1294 .addUse(ResTypeReg)
1295 .addUse(GV);
1296 MIB.constrainAllUses(TII, TRI, RBI);
1297 BuildMI(BB, I, I.getDebugLoc(),
1298 TII.get(STI.isLogicalSPIRV() ? SPIRV::OpInBoundsAccessChain
1299 : SPIRV::OpInBoundsPtrAccessChain))
1300 .addDef(ResVReg)
1301 .addUse(ResTypeReg)
1302 .addUse(NewVReg)
1303 .addUse(I.getOperand(2).getReg())
1304 .constrainAllUses(TII, TRI, RBI);
1305 } else {
1306 BuildMI(BB, I, I.getDebugLoc(), TII.get(SPIRV::OpSpecConstantOp))
1307 .addDef(ResVReg)
1308 .addUse(GR.getSPIRVTypeID(ResType))
1309 .addImm(
1310 static_cast<uint32_t>(SPIRV::Opcode::InBoundsPtrAccessChain))
1311 .addUse(GV)
1312 .addUse(I.getOperand(2).getReg())
1313 .constrainAllUses(TII, TRI, RBI);
1314 }
1315 return true;
1316 }
1317 // It's possible to translate G_PTR_ADD to OpSpecConstantOp: either to
1318 // initialize a global variable with a constant expression (e.g., the test
1319 // case opencl/basic/progvar_prog_scope_init.ll), or for another use case
1320 Register Idx = buildZerosVal(GR.getOrCreateSPIRVIntegerType(32, I, TII), I);
1321 auto MIB = BuildMI(BB, I, I.getDebugLoc(), TII.get(SPIRV::OpSpecConstantOp))
1322 .addDef(ResVReg)
1323 .addUse(GR.getSPIRVTypeID(ResType))
1324 .addImm(static_cast<uint32_t>(
1325 SPIRV::Opcode::InBoundsPtrAccessChain))
1326 .addUse(GV)
1327 .addUse(Idx)
1328 .addUse(I.getOperand(2).getReg());
1329 MIB.constrainAllUses(TII, TRI, RBI);
1330 return true;
1331 }
1332
1333 case TargetOpcode::G_ATOMICRMW_OR:
1334 return selectAtomicRMW(ResVReg, ResType, I, SPIRV::OpAtomicOr);
1335 case TargetOpcode::G_ATOMICRMW_ADD:
1336 return selectAtomicRMW(ResVReg, ResType, I, SPIRV::OpAtomicIAdd);
1337 case TargetOpcode::G_ATOMICRMW_AND:
1338 return selectAtomicRMW(ResVReg, ResType, I, SPIRV::OpAtomicAnd);
1339 case TargetOpcode::G_ATOMICRMW_MAX:
1340 return selectAtomicRMW(ResVReg, ResType, I, SPIRV::OpAtomicSMax);
1341 case TargetOpcode::G_ATOMICRMW_MIN:
1342 return selectAtomicRMW(ResVReg, ResType, I, SPIRV::OpAtomicSMin);
1343 case TargetOpcode::G_ATOMICRMW_SUB:
1344 return selectAtomicRMW(ResVReg, ResType, I, SPIRV::OpAtomicISub);
1345 case TargetOpcode::G_ATOMICRMW_XOR:
1346 return selectAtomicRMW(ResVReg, ResType, I, SPIRV::OpAtomicXor);
1347 case TargetOpcode::G_ATOMICRMW_UMAX:
1348 return selectAtomicRMW(ResVReg, ResType, I, SPIRV::OpAtomicUMax);
1349 case TargetOpcode::G_ATOMICRMW_UMIN:
1350 return selectAtomicRMW(ResVReg, ResType, I, SPIRV::OpAtomicUMin);
1351 case TargetOpcode::G_ATOMICRMW_XCHG:
1352 return selectAtomicRMW(ResVReg, ResType, I, SPIRV::OpAtomicExchange);
1353
1354 case TargetOpcode::G_ATOMICRMW_FADD:
1355 return selectAtomicRMW(ResVReg, ResType, I, SPIRV::OpAtomicFAddEXT);
1356 case TargetOpcode::G_ATOMICRMW_FSUB:
1357 // Translate G_ATOMICRMW_FSUB to OpAtomicFAddEXT with negative value operand
1358 return selectAtomicRMW(ResVReg, ResType, I, SPIRV::OpAtomicFAddEXT,
1359 ResType->getOpcode() == SPIRV::OpTypeVector
1360 ? SPIRV::OpFNegateV
1361 : SPIRV::OpFNegate);
1362 case TargetOpcode::G_ATOMICRMW_FMIN:
1363 return selectAtomicRMW(ResVReg, ResType, I, SPIRV::OpAtomicFMinEXT);
1364 case TargetOpcode::G_ATOMICRMW_FMAX:
1365 return selectAtomicRMW(ResVReg, ResType, I, SPIRV::OpAtomicFMaxEXT);
1366
1367 case TargetOpcode::G_FENCE:
1368 return selectFence(I);
1369
1370 case TargetOpcode::G_STACKSAVE:
1371 return selectStackSave(ResVReg, ResType, I);
1372 case TargetOpcode::G_STACKRESTORE:
1373 return selectStackRestore(I);
1374
1375 case TargetOpcode::G_UNMERGE_VALUES:
1376 return selectUnmergeValues(I);
1377
1378 // Discard gen opcodes for intrinsics which we do not expect to actually
1379 // represent code after lowering or intrinsics which are not implemented but
1380 // should not crash when found in a customer's LLVM IR input.
1381 case TargetOpcode::G_TRAP:
1382 case TargetOpcode::G_UBSANTRAP:
1383 case TargetOpcode::DBG_LABEL:
1384 return true;
1385 case TargetOpcode::G_DEBUGTRAP:
1386 return selectDebugTrap(ResVReg, ResType, I);
1387
1388 default:
1389 return false;
1390 }
1391}
1392
1393bool SPIRVInstructionSelector::selectDebugTrap(Register ResVReg,
1394 SPIRVTypeInst ResType,
1395 MachineInstr &I) const {
1396 unsigned Opcode = SPIRV::OpNop;
1397 MachineBasicBlock &BB = *I.getParent();
1398 BuildMI(BB, I, I.getDebugLoc(), TII.get(Opcode))
1399 .constrainAllUses(TII, TRI, RBI);
1400 return true;
1401}
1402
1403bool SPIRVInstructionSelector::selectExtInst(Register ResVReg,
1404 SPIRVTypeInst ResType,
1405 MachineInstr &I,
1406 GL::GLSLExtInst GLInst,
1407 bool setMIFlags, bool useMISrc,
1408 ArrayRef<Register> SrcRegs) const {
1409 if (!STI.canUseExtInstSet(
1410 SPIRV::InstructionSet::InstructionSet::GLSL_std_450)) {
1411 std::string DiagMsg;
1412 raw_string_ostream OS(DiagMsg);
1413 I.print(OS, true, false, false, false);
1414 DiagMsg += " is only supported with the GLSL extended instruction set.\n";
1415 report_fatal_error(DiagMsg.c_str(), false);
1416 }
1417 return selectExtInst(ResVReg, ResType, I,
1418 {{SPIRV::InstructionSet::GLSL_std_450, GLInst}},
1419 setMIFlags, useMISrc, SrcRegs);
1420}
1421
1422bool SPIRVInstructionSelector::selectExtInst(Register ResVReg,
1423 SPIRVTypeInst ResType,
1424 MachineInstr &I,
1425 CL::OpenCLExtInst CLInst,
1426 bool setMIFlags, bool useMISrc,
1427 ArrayRef<Register> SrcRegs) const {
1428 return selectExtInst(ResVReg, ResType, I,
1429 {{SPIRV::InstructionSet::OpenCL_std, CLInst}},
1430 setMIFlags, useMISrc, SrcRegs);
1431}
1432
1433bool SPIRVInstructionSelector::selectExtInst(
1434 Register ResVReg, SPIRVTypeInst ResType, MachineInstr &I,
1435 CL::OpenCLExtInst CLInst, GL::GLSLExtInst GLInst, bool setMIFlags,
1436 bool useMISrc, ArrayRef<Register> SrcRegs) const {
1437 ExtInstList ExtInsts = {{SPIRV::InstructionSet::OpenCL_std, CLInst},
1438 {SPIRV::InstructionSet::GLSL_std_450, GLInst}};
1439 return selectExtInst(ResVReg, ResType, I, ExtInsts, setMIFlags, useMISrc,
1440 SrcRegs);
1441}
1442
1443bool SPIRVInstructionSelector::selectExtInst(Register ResVReg,
1444 SPIRVTypeInst ResType,
1445 MachineInstr &I,
1446 const ExtInstList &Insts,
1447 bool setMIFlags, bool useMISrc,
1448 ArrayRef<Register> SrcRegs) const {
1449
1450 for (const auto &[InstructionSet, Opcode] : Insts) {
1451 if (!STI.canUseExtInstSet(InstructionSet))
1452 continue;
1453 MachineBasicBlock &BB = *I.getParent();
1454 auto MIB = BuildMI(BB, I, I.getDebugLoc(), TII.get(SPIRV::OpExtInst))
1455 .addDef(ResVReg)
1456 .addUse(GR.getSPIRVTypeID(ResType))
1457 .addImm(static_cast<uint32_t>(InstructionSet))
1458 .addImm(Opcode);
1459 if (setMIFlags)
1460 MIB.setMIFlags(I.getFlags());
1461 if (useMISrc) {
1462 const unsigned NumOps = I.getNumOperands();
1463 unsigned Index = 1;
1464 if (Index < NumOps &&
1465 I.getOperand(Index).getType() ==
1466 MachineOperand::MachineOperandType::MO_IntrinsicID)
1467 Index = 2;
1468 for (; Index < NumOps; ++Index)
1469 MIB.add(I.getOperand(Index));
1470 } else {
1471 for (Register SReg : SrcRegs) {
1472 MIB.addUse(SReg);
1473 }
1474 }
1475 MIB.constrainAllUses(TII, TRI, RBI);
1476 return true;
1477 }
1478 return false;
1479}
1480
1481bool SPIRVInstructionSelector::selectFrexp(Register ResVReg,
1482 SPIRVTypeInst ResType,
1483 MachineInstr &I) const {
1484 ExtInstList ExtInsts = {{SPIRV::InstructionSet::OpenCL_std, CL::frexp},
1485 {SPIRV::InstructionSet::GLSL_std_450, GL::Frexp}};
1486 for (const auto &Ex : ExtInsts) {
1487 SPIRV::InstructionSet::InstructionSet Set = Ex.first;
1488 uint32_t Opcode = Ex.second;
1489 if (!STI.canUseExtInstSet(Set))
1490 continue;
1491
1492 MachineIRBuilder MIRBuilder(I);
1493 SPIRVTypeInst PointeeTy = GR.getSPIRVTypeForVReg(I.getOperand(1).getReg());
1494 const SPIRVTypeInst PointerType = GR.getOrCreateSPIRVPointerType(
1495 PointeeTy, MIRBuilder, SPIRV::StorageClass::Function);
1496 Register PointerVReg =
1497 createVirtualRegister(PointerType, &GR, MRI, MRI->getMF());
1498
1499 auto It = getOpVariableMBBIt(*I.getMF());
1500 BuildMI(*It->getParent(), It, It->getDebugLoc(), TII.get(SPIRV::OpVariable))
1501 .addDef(PointerVReg)
1502 .addUse(GR.getSPIRVTypeID(PointerType))
1503 .addImm(static_cast<uint32_t>(SPIRV::StorageClass::Function))
1504 .constrainAllUses(TII, TRI, RBI);
1505
1506 BuildMI(*I.getParent(), I, I.getDebugLoc(), TII.get(SPIRV::OpExtInst))
1507 .addDef(ResVReg)
1508 .addUse(GR.getSPIRVTypeID(ResType))
1509 .addImm(static_cast<uint32_t>(Ex.first))
1510 .addImm(Opcode)
1511 .add(I.getOperand(2))
1512 .addUse(PointerVReg)
1513 .constrainAllUses(TII, TRI, RBI);
1514
1515 BuildMI(*I.getParent(), I, I.getDebugLoc(), TII.get(SPIRV::OpLoad))
1516 .addDef(I.getOperand(1).getReg())
1517 .addUse(GR.getSPIRVTypeID(PointeeTy))
1518 .addUse(PointerVReg)
1519 .constrainAllUses(TII, TRI, RBI);
1520 return true;
1521 }
1522 return false;
1523}
1524
1525bool SPIRVInstructionSelector::selectSincos(Register ResVReg,
1526 SPIRVTypeInst ResType,
1527 MachineInstr &I) const {
1528 Register CosResVReg = I.getOperand(1).getReg();
1529 unsigned SrcIdx = I.getNumExplicitDefs();
1530 Register ResTypeReg = GR.getSPIRVTypeID(ResType);
1531
1532 if (STI.canUseExtInstSet(SPIRV::InstructionSet::OpenCL_std)) {
1533 // OpenCL.std sincos(x, cosval*) -> returns sin(x), writes cos(x) to ptr.
1534 MachineIRBuilder MIRBuilder(I);
1535 const SPIRVTypeInst PointerType = GR.getOrCreateSPIRVPointerType(
1536 ResType, MIRBuilder, SPIRV::StorageClass::Function);
1537 Register PointerVReg =
1538 createVirtualRegister(PointerType, &GR, MRI, MRI->getMF());
1539
1540 auto It = getOpVariableMBBIt(*I.getMF());
1541 BuildMI(*It->getParent(), It, It->getDebugLoc(), TII.get(SPIRV::OpVariable))
1542 .addDef(PointerVReg)
1543 .addUse(GR.getSPIRVTypeID(PointerType))
1544 .addImm(static_cast<uint32_t>(SPIRV::StorageClass::Function))
1545 .constrainAllUses(TII, TRI, RBI);
1546 BuildMI(*I.getParent(), I, I.getDebugLoc(), TII.get(SPIRV::OpExtInst))
1547 .addDef(ResVReg)
1548 .addUse(ResTypeReg)
1549 .addImm(static_cast<uint32_t>(SPIRV::InstructionSet::OpenCL_std))
1550 .addImm(CL::sincos)
1551 .add(I.getOperand(SrcIdx))
1552 .addUse(PointerVReg)
1553 .constrainAllUses(TII, TRI, RBI);
1554 BuildMI(*I.getParent(), I, I.getDebugLoc(), TII.get(SPIRV::OpLoad))
1555 .addDef(CosResVReg)
1556 .addUse(ResTypeReg)
1557 .addUse(PointerVReg)
1558 .constrainAllUses(TII, TRI, RBI);
1559 return true;
1560 } else if (STI.canUseExtInstSet(SPIRV::InstructionSet::GLSL_std_450)) {
1561 // GLSL.std.450 has no combined sincos; emit separate Sin and Cos.
1562 BuildMI(*I.getParent(), I, I.getDebugLoc(), TII.get(SPIRV::OpExtInst))
1563 .addDef(ResVReg)
1564 .addUse(ResTypeReg)
1565 .addImm(static_cast<uint32_t>(SPIRV::InstructionSet::GLSL_std_450))
1566 .addImm(GL::Sin)
1567 .add(I.getOperand(SrcIdx))
1568 .constrainAllUses(TII, TRI, RBI);
1569 BuildMI(*I.getParent(), I, I.getDebugLoc(), TII.get(SPIRV::OpExtInst))
1570 .addDef(CosResVReg)
1571 .addUse(ResTypeReg)
1572 .addImm(static_cast<uint32_t>(SPIRV::InstructionSet::GLSL_std_450))
1573 .addImm(GL::Cos)
1574 .add(I.getOperand(SrcIdx))
1575 .constrainAllUses(TII, TRI, RBI);
1576 return true;
1577 }
1578 return false;
1579}
1580
1581bool SPIRVInstructionSelector::selectOpWithSrcs(Register ResVReg,
1582 SPIRVTypeInst ResType,
1583 MachineInstr &I,
1584 std::vector<Register> Srcs,
1585 unsigned Opcode) const {
1586 auto MIB = BuildMI(*I.getParent(), I, I.getDebugLoc(), TII.get(Opcode))
1587 .addDef(ResVReg)
1588 .addUse(GR.getSPIRVTypeID(ResType));
1589 for (Register SReg : Srcs) {
1590 MIB.addUse(SReg);
1591 }
1592 MIB.constrainAllUses(TII, TRI, RBI);
1593 return true;
1594}
1595
1596std::optional<SplitParts> SPIRVInstructionSelector::splitEvenOddLanes(
1597 Register PopCountReg, unsigned ComponentCount, MachineInstr &I,
1598 SPIRVTypeInst I32Type) const {
1599 SplitParts Parts;
1600
1601 if (ComponentCount == 1) {
1602 // ---- Scalar path: extract element 1 (high word) and element 0 (low word)
1603 // ----
1604 Parts.IsScalar = true;
1605 Parts.Type = I32Type;
1606 Parts.High = MRI->createVirtualRegister(GR.getRegClass(I32Type));
1607 Parts.Low = MRI->createVirtualRegister(GR.getRegClass(I32Type));
1608
1609 bool ZeroAsNull = !STI.isShader();
1610 Register IdxZero = GR.getOrCreateConstInt(0, I, I32Type, TII, ZeroAsNull);
1611 Register IdxOne = GR.getOrCreateConstInt(1, I, I32Type, TII, ZeroAsNull);
1612
1613 if (!selectOpWithSrcs(Parts.High, I32Type, I, {PopCountReg, IdxOne},
1614 SPIRV::OpVectorExtractDynamic))
1615 return std::nullopt;
1616
1617 if (!selectOpWithSrcs(Parts.Low, I32Type, I, {PopCountReg, IdxZero},
1618 SPIRV::OpVectorExtractDynamic))
1619 return std::nullopt;
1620
1621 } else {
1622 // ---- Vector path: shuffle odd lanes → High, even lanes → Low ----
1623 MachineIRBuilder MIRBuilder(I);
1624 Parts.IsScalar = false;
1625 Parts.Type = GR.getOrCreateSPIRVVectorType(I32Type, ComponentCount,
1626 MIRBuilder, /*IsSigned=*/false);
1627 Parts.High = MRI->createVirtualRegister(GR.getRegClass(Parts.Type));
1628 Parts.Low = MRI->createVirtualRegister(GR.getRegClass(Parts.Type));
1629
1630 // High = odd-indexed elements (1, 3, 5, …) — the upper 32-bit halves.
1631 auto MIB = BuildMI(*I.getParent(), I, I.getDebugLoc(),
1632 TII.get(SPIRV::OpVectorShuffle))
1633 .addDef(Parts.High)
1634 .addUse(GR.getSPIRVTypeID(Parts.Type))
1635 .addUse(PopCountReg)
1636 .addUse(PopCountReg);
1637 for (unsigned J = 1; J < ComponentCount * 2; J += 2)
1638 MIB.addImm(J);
1639 MIB.constrainAllUses(TII, TRI, RBI);
1640
1641 // Low = even-indexed elements (0, 2, 4, …) — the lower 32-bit halves.
1642 MIB = BuildMI(*I.getParent(), I, I.getDebugLoc(),
1643 TII.get(SPIRV::OpVectorShuffle))
1644 .addDef(Parts.Low)
1645 .addUse(GR.getSPIRVTypeID(Parts.Type))
1646 .addUse(PopCountReg)
1647 .addUse(PopCountReg);
1648 for (unsigned J = 0; J < ComponentCount * 2; J += 2)
1649 MIB.addImm(J);
1650 MIB.constrainAllUses(TII, TRI, RBI);
1651 }
1652
1653 return Parts;
1654}
1655
1656bool SPIRVInstructionSelector::selectPopCount16(Register ResVReg,
1657 SPIRVTypeInst ResType,
1658 MachineInstr &I,
1659 unsigned ExtOpcode,
1660 unsigned Opcode) const {
1661 Register OpReg = I.getOperand(1).getReg();
1662 unsigned NumElems = GR.getScalarOrVectorComponentCount(OpReg);
1663
1664 MachineIRBuilder MIRBuilder(I);
1665 SPIRVTypeInst I32Type = GR.getOrCreateSPIRVIntegerType(32, MIRBuilder);
1666 SPIRVTypeInst I32VectorType =
1667 GR.getOrCreateSPIRVVectorType(I32Type, NumElems, MIRBuilder, false);
1668
1669 bool IsVector = NumElems > 1;
1670 SPIRVTypeInst ExtType = IsVector ? I32VectorType : I32Type;
1671 Register ExtReg = MRI->createVirtualRegister(GR.getRegClass(ExtType));
1672 // Always use OpUConvert to always use a 0 extend
1673 if (!selectOpWithSrcs(ExtReg, ExtType, I, {OpReg}, SPIRV::OpUConvert))
1674 return false;
1675
1676 Register PopCountReg = MRI->createVirtualRegister(GR.getRegClass(ExtType));
1677 if (!selectPopCount32(PopCountReg, ExtType, I, ExtReg, Opcode))
1678 return false;
1679
1680 return selectOpWithSrcs(ResVReg, ResType, I, {PopCountReg}, ExtOpcode);
1681}
1682
1683bool SPIRVInstructionSelector::selectPopCount32(Register ResVReg,
1684 SPIRVTypeInst ResType,
1685 MachineInstr &I,
1686 Register SrcReg,
1687 unsigned Opcode) const {
1688 return selectOpWithSrcs(ResVReg, ResType, I, {SrcReg}, Opcode);
1689}
1690
1691bool SPIRVInstructionSelector::selectPopCount64(Register ResVReg,
1692 SPIRVTypeInst ResType,
1693 MachineInstr &I,
1694 Register SrcReg,
1695 unsigned Opcode) const {
1696 unsigned ComponentCount = GR.getScalarOrVectorComponentCount(ResType);
1697 if (ComponentCount > 2)
1698 return handle64BitOverflow(
1699 ResVReg, ResType, I, SrcReg, Opcode,
1700 [this](Register R, SPIRVTypeInst T, MachineInstr &I, Register S,
1701 unsigned O) { return this->selectPopCount64(R, T, I, S, O); });
1702
1703 MachineIRBuilder MIRBuilder(I);
1704
1705 // ---- Types ----
1706 SPIRVTypeInst I32Type = GR.getOrCreateSPIRVIntegerType(32, MIRBuilder);
1707 SPIRVTypeInst VecI32Type = GR.getOrCreateSPIRVVectorType(
1708 I32Type, 2 * ComponentCount, MIRBuilder, /*IsSigned=*/false);
1709
1710 // Converts 64 bit into and array of 32 bit, containing 2 elements.
1711 Register Vec32 = MRI->createVirtualRegister(GR.getRegClass(VecI32Type));
1712 if (!selectOpWithSrcs(Vec32, VecI32Type, I, {SrcReg}, SPIRV::OpBitcast))
1713 return false;
1714
1715 // Apply popcount on each 32 bit lane
1716 Register Pop32 = MRI->createVirtualRegister(GR.getRegClass(VecI32Type));
1717 if (!selectPopCount32(Pop32, VecI32Type, I, Vec32, Opcode))
1718 return false;
1719
1720 // Splits result into highbit lane and lowbit lane
1721 auto MaybeParts = splitEvenOddLanes(Pop32, ComponentCount, I, I32Type);
1722 if (!MaybeParts)
1723 return false;
1724 SplitParts &Parts = *MaybeParts;
1725
1726 // Sum high part and low part
1727 unsigned OpAdd = Parts.IsScalar ? SPIRV::OpIAddS : SPIRV::OpIAddV;
1728 Register Sum = MRI->createVirtualRegister(GR.getRegClass(Parts.Type));
1729 if (!selectOpWithSrcs(Sum, Parts.Type, I, {Parts.High, Parts.Low}, OpAdd))
1730 return false;
1731
1732 // Convert 32 bit sum into 64 bit scalar
1733 bool IsSigned = GR.isScalarOrVectorSigned(ResType);
1734 unsigned ConvOp = IsSigned ? SPIRV::OpSConvert : SPIRV::OpUConvert;
1735 return selectOpWithSrcs(ResVReg, ResType, I, {Sum}, ConvOp);
1736}
1737
1738bool SPIRVInstructionSelector::selectPopCount(Register ResVReg,
1739 SPIRVTypeInst ResType,
1740 MachineInstr &I,
1741 unsigned Opcode) const {
1742 // Vulkan restricts OpBitCount to 32-bit integers or vectors of 32-bit
1743 // integers unless VK_KHR_maintenance9 is enabled. Until VK_KHR_maintaince9
1744 // is core we will not generate OpBitCount with any other types when
1745 // targeting Vulkan.
1746 if (!STI.getTargetTriple().isVulkanOS())
1747 return selectUnOp(ResVReg, ResType, I, Opcode);
1748
1749 Register OpReg = I.getOperand(1).getReg();
1750 SPIRVTypeInst OpType = GR.getSPIRVTypeForVReg(OpReg);
1751 unsigned ExtOpcode = GR.isScalarOrVectorSigned(ResType) ? SPIRV::OpSConvert
1752 : SPIRV::OpUConvert;
1753 switch (GR.getScalarOrVectorBitWidth(OpType)) {
1754 case 8:
1755 case 16:
1756 return selectPopCount16(ResVReg, ResType, I, ExtOpcode, Opcode);
1757 case 32:
1758 return selectPopCount32(ResVReg, ResType, I, OpReg, Opcode);
1759 case 64:
1760 return selectPopCount64(ResVReg, ResType, I, OpReg, Opcode);
1761 default:
1762 report_fatal_error("unsupported operand bit width for popcount");
1763 }
1764}
1765
1766bool SPIRVInstructionSelector::selectUnOp(Register ResVReg,
1767 SPIRVTypeInst ResType,
1768 MachineInstr &I,
1769 unsigned Opcode) const {
1770 if (STI.isPhysicalSPIRV() && I.getOperand(1).isReg()) {
1771 Register SrcReg = I.getOperand(1).getReg();
1772 bool IsGV = false;
1773 for (MachineRegisterInfo::def_instr_iterator DefIt =
1774 MRI->def_instr_begin(SrcReg);
1775 DefIt != MRI->def_instr_end(); DefIt = std::next(DefIt)) {
1776 unsigned DefOpCode = DefIt->getOpcode();
1777 if (DefOpCode == SPIRV::ASSIGN_TYPE || DefOpCode == TargetOpcode::COPY) {
1778 // We need special handling to look through the type assignment or the
1779 // COPY pseudo-op and see if this is a constant or a global.
1780 if (auto *VRD = getVRegDef(*MRI, DefIt->getOperand(1).getReg()))
1781 DefOpCode = VRD->getOpcode();
1782 }
1783 if (DefOpCode == TargetOpcode::G_GLOBAL_VALUE ||
1784 DefOpCode == TargetOpcode::G_CONSTANT ||
1785 DefOpCode == SPIRV::OpVariable || DefOpCode == SPIRV::OpConstantI) {
1786 IsGV = true;
1787 break;
1788 }
1789 }
1790 if (IsGV) {
1791 uint32_t SpecOpcode = 0;
1792 switch (Opcode) {
1793 case SPIRV::OpConvertPtrToU:
1794 SpecOpcode = static_cast<uint32_t>(SPIRV::Opcode::ConvertPtrToU);
1795 break;
1796 case SPIRV::OpConvertUToPtr:
1797 SpecOpcode = static_cast<uint32_t>(SPIRV::Opcode::ConvertUToPtr);
1798 break;
1799 }
1800 if (SpecOpcode) {
1801 BuildMI(*I.getParent(), I, I.getDebugLoc(),
1802 TII.get(SPIRV::OpSpecConstantOp))
1803 .addDef(ResVReg)
1804 .addUse(GR.getSPIRVTypeID(ResType))
1805 .addImm(SpecOpcode)
1806 .addUse(SrcReg)
1807 .constrainAllUses(TII, TRI, RBI);
1808 return true;
1809 }
1810 }
1811 }
1812 return selectOpWithSrcs(ResVReg, ResType, I, {I.getOperand(1).getReg()},
1813 Opcode);
1814}
1815
1816bool SPIRVInstructionSelector::selectBitcast(Register ResVReg,
1817 SPIRVTypeInst ResType,
1818 MachineInstr &I) const {
1819 Register OpReg = I.getOperand(1).getReg();
1820 SPIRVTypeInst OpType =
1821 OpReg.isValid() ? GR.getSPIRVTypeForVReg(OpReg) : nullptr;
1822 if (!GR.isBitcastCompatible(ResType, OpType))
1823 report_fatal_error("incompatible result and operand types in a bitcast");
1824 return selectUnOp(ResVReg, ResType, I, SPIRV::OpBitcast);
1825}
1826
1827static void addMemoryOperands(MachineMemOperand *MemOp,
1828 MachineInstrBuilder &MIB,
1829 MachineIRBuilder &MIRBuilder,
1830 SPIRVGlobalRegistry &GR) {
1831 const SPIRVSubtarget *ST =
1832 static_cast<const SPIRVSubtarget *>(&MIRBuilder.getMF().getSubtarget());
1833 uint32_t SpvMemOp = static_cast<uint32_t>(SPIRV::MemoryOperand::None);
1834 if (MemOp->isVolatile())
1835 SpvMemOp |= static_cast<uint32_t>(SPIRV::MemoryOperand::Volatile);
1836 if (MemOp->isNonTemporal())
1837 SpvMemOp |= static_cast<uint32_t>(SPIRV::MemoryOperand::Nontemporal);
1838 // Aligned memory operand requires the Kernel capability.
1839 if (!ST->isShader() && MemOp->getAlign().value())
1840 SpvMemOp |= static_cast<uint32_t>(SPIRV::MemoryOperand::Aligned);
1841
1842 [[maybe_unused]] MachineInstr *AliasList = nullptr;
1843 [[maybe_unused]] MachineInstr *NoAliasList = nullptr;
1844 if (ST->canUseExtension(SPIRV::Extension::SPV_INTEL_memory_access_aliasing)) {
1845 if (auto *MD = MemOp->getAAInfo().Scope) {
1846 AliasList = GR.getOrAddMemAliasingINTELInst(MIRBuilder, MD);
1847 if (AliasList)
1848 SpvMemOp |=
1849 static_cast<uint32_t>(SPIRV::MemoryOperand::AliasScopeINTELMask);
1850 }
1851 if (auto *MD = MemOp->getAAInfo().NoAlias) {
1852 NoAliasList = GR.getOrAddMemAliasingINTELInst(MIRBuilder, MD);
1853 if (NoAliasList)
1854 SpvMemOp |=
1855 static_cast<uint32_t>(SPIRV::MemoryOperand::NoAliasINTELMask);
1856 }
1857 }
1858
1859 if (SpvMemOp != static_cast<uint32_t>(SPIRV::MemoryOperand::None)) {
1860 MIB.addImm(SpvMemOp);
1861 if (SpvMemOp & static_cast<uint32_t>(SPIRV::MemoryOperand::Aligned))
1862 MIB.addImm(MemOp->getAlign().value());
1863 if (AliasList)
1864 MIB.addUse(AliasList->getOperand(0).getReg());
1865 if (NoAliasList)
1866 MIB.addUse(NoAliasList->getOperand(0).getReg());
1867 }
1868}
1869
1870static void addMemoryOperands(uint64_t Flags, MachineInstrBuilder &MIB) {
1871 uint32_t SpvMemOp = static_cast<uint32_t>(SPIRV::MemoryOperand::None);
1872 if (Flags & MachineMemOperand::Flags::MOVolatile)
1873 SpvMemOp |= static_cast<uint32_t>(SPIRV::MemoryOperand::Volatile);
1874 if (Flags & MachineMemOperand::Flags::MONonTemporal)
1875 SpvMemOp |= static_cast<uint32_t>(SPIRV::MemoryOperand::Nontemporal);
1876
1877 if (SpvMemOp != static_cast<uint32_t>(SPIRV::MemoryOperand::None))
1878 MIB.addImm(SpvMemOp);
1879}
1880
1881bool SPIRVInstructionSelector::selectLoad(Register ResVReg,
1882 SPIRVTypeInst ResType,
1883 MachineInstr &I) const {
1884 unsigned OpOffset = isa<GIntrinsic>(I) ? 1 : 0;
1885 Register Ptr = I.getOperand(1 + OpOffset).getReg();
1886
1887 auto *PtrDef = getVRegDef(*MRI, Ptr);
1888 auto *IntPtrDef = dyn_cast<GIntrinsic>(PtrDef);
1889 if (IntPtrDef &&
1890 IntPtrDef->getIntrinsicID() == Intrinsic::spv_resource_getpointer) {
1891 Register HandleReg = IntPtrDef->getOperand(2).getReg();
1892 SPIRVTypeInst HandleType = GR.getSPIRVTypeForVReg(HandleReg);
1893 if (HandleType->getOpcode() == SPIRV::OpTypeImage) {
1894 Register NewHandleReg =
1895 MRI->createVirtualRegister(MRI->getRegClass(HandleReg));
1896 auto *HandleDef = cast<GIntrinsic>(getVRegDef(*MRI, HandleReg));
1897 if (!loadHandleBeforePosition(NewHandleReg, HandleType, *HandleDef, I)) {
1898 return false;
1899 }
1900
1901 Register IdxReg = IntPtrDef->getOperand(3).getReg();
1902 return generateImageReadOrFetch(ResVReg, ResType, NewHandleReg, IdxReg,
1903 I.getDebugLoc(), I);
1904 }
1905 }
1906
1907 MachineIRBuilder MIRBuilder(I);
1908
1909 if (I.getNumMemOperands()) {
1910 const MachineMemOperand *MemOp = *I.memoperands_begin();
1911 if (MemOp->isAtomic())
1912 return selectAtomicLoad(ResVReg, ResType, I);
1913 }
1914
1915 auto MIB = MIRBuilder.buildInstr(SPIRV::OpLoad)
1916 .addDef(ResVReg)
1917 .addUse(GR.getSPIRVTypeID(ResType))
1918 .addUse(Ptr);
1919 if (!I.getNumMemOperands()) {
1920 assert(I.getOpcode() == TargetOpcode::G_INTRINSIC_W_SIDE_EFFECTS ||
1921 I.getOpcode() ==
1922 TargetOpcode::G_INTRINSIC_CONVERGENT_W_SIDE_EFFECTS);
1923 addMemoryOperands(I.getOperand(2 + OpOffset).getImm(), MIB);
1924 } else {
1925 addMemoryOperands(*I.memoperands_begin(), MIB, MIRBuilder, GR);
1926 }
1927 MIB.constrainAllUses(TII, TRI, RBI);
1928 return true;
1929}
1930
1931bool SPIRVInstructionSelector::selectAtomicLoad(Register ResVReg,
1932 SPIRVTypeInst ResType,
1933 MachineInstr &I) const {
1934 LLVMContext &Context = I.getMF()->getFunction().getContext();
1935
1936 unsigned OpOffset = isa<GIntrinsic>(I) ? 1 : 0;
1937 Register Ptr = I.getOperand(1 + OpOffset).getReg();
1938
1939 if (!ResType.isTypeIntOrFloat())
1940 return diagnoseUnsupported(I,
1941 "Lowering to SPIR-V of atomic load is only "
1942 "allowed for integer or floating point types");
1943
1944 assert(I.getNumMemOperands());
1945 const MachineMemOperand &MemOp = **I.memoperands_begin();
1946 assert(MemOp.isAtomic());
1947 // TODO: This must be relaxed since the volatile attribute on atomic load is
1948 // supported with the VulkanMemoryModelKHR capability.
1949 if (MemOp.isVolatile())
1950 return diagnoseUnsupported(I, "Lowering to SPIR-V of atomic load of "
1951 "volatile memory is not supported");
1952
1953 uint32_t Scope =
1954 static_cast<uint32_t>(getMemScope(Context, MemOp.getSyncScopeID()));
1955 Register ScopeReg = buildI32Constant(Scope, I);
1956
1957 AtomicOrdering AO = MemOp.getSuccessOrdering();
1958 uint32_t StorageClass = static_cast<uint32_t>(getMemSemanticsForStorageClass(
1959 addressSpaceToStorageClass(MemOp.getAddrSpace(), STI)));
1960 uint32_t MemSem = static_cast<uint32_t>(getMemSemantics(AO));
1961 Register MemSemReg = buildI32Constant(MemSem | StorageClass, I);
1962
1963 MachineIRBuilder MIRBuilder(I);
1964 auto AtomicLoad = MIRBuilder.buildInstr(SPIRV::OpAtomicLoad)
1965 .addDef(ResVReg)
1966 .addUse(GR.getSPIRVTypeID(ResType))
1967 .addUse(Ptr)
1968 .addUse(ScopeReg)
1969 .addUse(MemSemReg);
1970 AtomicLoad.constrainAllUses(TII, TRI, RBI);
1971 return true;
1972}
1973
1974bool SPIRVInstructionSelector::selectStore(MachineInstr &I) const {
1975 unsigned OpOffset = isa<GIntrinsic>(I) ? 1 : 0;
1976 Register StoreVal = I.getOperand(0 + OpOffset).getReg();
1977 Register Ptr = I.getOperand(1 + OpOffset).getReg();
1978
1979 auto *PtrDef = getVRegDef(*MRI, Ptr);
1980 auto *IntPtrDef = dyn_cast<GIntrinsic>(PtrDef);
1981 if (IntPtrDef &&
1982 IntPtrDef->getIntrinsicID() == Intrinsic::spv_resource_getpointer) {
1983 Register HandleReg = IntPtrDef->getOperand(2).getReg();
1984 Register NewHandleReg =
1985 MRI->createVirtualRegister(MRI->getRegClass(HandleReg));
1986 auto *HandleDef = cast<GIntrinsic>(getVRegDef(*MRI, HandleReg));
1987 SPIRVTypeInst HandleType = GR.getSPIRVTypeForVReg(HandleReg);
1988 if (!loadHandleBeforePosition(NewHandleReg, HandleType, *HandleDef, I)) {
1989 return false;
1990 }
1991
1992 Register IdxReg = IntPtrDef->getOperand(3).getReg();
1993 if (HandleType->getOpcode() == SPIRV::OpTypeImage) {
1994 auto BMI = BuildMI(*I.getParent(), I, I.getDebugLoc(),
1995 TII.get(SPIRV::OpImageWrite))
1996 .addUse(NewHandleReg)
1997 .addUse(IdxReg)
1998 .addUse(StoreVal);
1999
2000 const llvm::Type *LLVMHandleType = GR.getTypeForSPIRVType(HandleType);
2001 if (sampledTypeIsSignedInteger(LLVMHandleType))
2002 BMI.addImm(0x1000); // SignExtend
2003
2004 BMI.constrainAllUses(TII, TRI, RBI);
2005 return true;
2006 }
2007 }
2008
2009 if (I.getNumMemOperands()) {
2010 const MachineMemOperand *MemOp = *I.memoperands_begin();
2011 if (MemOp->isAtomic())
2012 return selectAtomicStore(I);
2013 }
2014
2015 MachineIRBuilder MIRBuilder(I);
2016 auto MIB = MIRBuilder.buildInstr(SPIRV::OpStore).addUse(Ptr).addUse(StoreVal);
2017 if (!I.getNumMemOperands()) {
2018 assert(I.getOpcode() == TargetOpcode::G_INTRINSIC_W_SIDE_EFFECTS ||
2019 I.getOpcode() ==
2020 TargetOpcode::G_INTRINSIC_CONVERGENT_W_SIDE_EFFECTS);
2021 addMemoryOperands(I.getOperand(2 + OpOffset).getImm(), MIB);
2022 } else {
2023 addMemoryOperands(*I.memoperands_begin(), MIB, MIRBuilder, GR);
2024 }
2025 MIB.constrainAllUses(TII, TRI, RBI);
2026 return true;
2027}
2028
2029bool SPIRVInstructionSelector::selectAtomicStore(MachineInstr &I) const {
2030 LLVMContext &Context = I.getMF()->getFunction().getContext();
2031
2032 unsigned OpOffset = isa<GIntrinsic>(I) ? 1 : 0;
2033 Register StoreVal = I.getOperand(0 + OpOffset).getReg();
2034 Register Ptr = I.getOperand(1 + OpOffset).getReg();
2035
2036 SPIRVTypeInst PtrType = GR.getSPIRVTypeForVReg(Ptr);
2037 SPIRVTypeInst PointeeType = GR.getPointeeType(PtrType);
2038 if (!PointeeType.isTypeIntOrFloat())
2039 return diagnoseUnsupported(I,
2040 "Lowering to SPIR-V of atomic store is only "
2041 "allowed for integer or floating point types");
2042
2043 assert(I.getNumMemOperands());
2044 const MachineMemOperand &MemOp = **I.memoperands_begin();
2045 assert(MemOp.isAtomic());
2046
2047 // TODO: This must be relaxed since the volatile attribute on atomic store is
2048 // supported with the VulkanMemoryModelKHR capability.
2049 if (MemOp.isVolatile())
2050 return diagnoseUnsupported(I, "Lowering to SPIR-V of atomic store of "
2051 "volatile memory is not supported");
2052
2053 uint32_t Scope =
2054 static_cast<uint32_t>(getMemScope(Context, MemOp.getSyncScopeID()));
2055 Register ScopeReg = buildI32Constant(Scope, I);
2056
2057 AtomicOrdering AO = MemOp.getSuccessOrdering();
2058 uint32_t StorageClass = static_cast<uint32_t>(getMemSemanticsForStorageClass(
2059 addressSpaceToStorageClass(MemOp.getAddrSpace(), STI)));
2060 uint32_t MemSem = static_cast<uint32_t>(getMemSemantics(AO));
2061 Register MemSemReg = buildI32Constant(MemSem | StorageClass, I);
2062
2063 MachineIRBuilder MIRBuilder(I);
2064 auto AtomicStore = MIRBuilder.buildInstr(SPIRV::OpAtomicStore)
2065 .addUse(Ptr)
2066 .addUse(ScopeReg)
2067 .addUse(MemSemReg)
2068 .addUse(StoreVal);
2069 AtomicStore.constrainAllUses(TII, TRI, RBI);
2070 return true;
2071}
2072
2073bool SPIRVInstructionSelector::selectMaskedGather(Register ResVReg,
2074 SPIRVTypeInst ResType,
2075 MachineInstr &I) const {
2076 assert(I.getNumExplicitDefs() == 1 && "Expected single def for gather");
2077 // Operand indices:
2078 // 0: result (def)
2079 // 1: intrinsic ID
2080 // 2: vector of pointers
2081 // 3: alignment (i32 immediate)
2082 // 4: mask (vector of i1)
2083 // 5: passthru/fill value
2084 const Register PtrsReg = I.getOperand(2).getReg();
2085 const uint32_t Alignment = I.getOperand(3).getImm();
2086 const Register MaskReg = I.getOperand(4).getReg();
2087 const Register PassthruReg = I.getOperand(5).getReg();
2088 const Register AlignmentReg = buildI32Constant(Alignment, I);
2089
2090 MachineBasicBlock &BB = *I.getParent();
2091 auto MIB =
2092 BuildMI(BB, I, I.getDebugLoc(), TII.get(SPIRV::OpMaskedGatherINTEL))
2093 .addDef(ResVReg)
2094 .addUse(GR.getSPIRVTypeID(ResType))
2095 .addUse(PtrsReg)
2096 .addUse(AlignmentReg)
2097 .addUse(MaskReg)
2098 .addUse(PassthruReg);
2099 MIB.constrainAllUses(TII, TRI, RBI);
2100 return true;
2101}
2102
2103bool SPIRVInstructionSelector::selectMaskedScatter(MachineInstr &I) const {
2104 assert(I.getNumExplicitDefs() == 0 && "Expected no defs for scatter");
2105 // Operand indices (no explicit defs):
2106 // 0: intrinsic ID
2107 // 1: value vector
2108 // 2: vector of pointers
2109 // 3: alignment (i32 immediate)
2110 // 4: mask (vector of i1)
2111 const Register ValuesReg = I.getOperand(1).getReg();
2112 const Register PtrsReg = I.getOperand(2).getReg();
2113 const uint32_t Alignment = I.getOperand(3).getImm();
2114 const Register MaskReg = I.getOperand(4).getReg();
2115 const Register AlignmentReg = buildI32Constant(Alignment, I);
2116 MachineBasicBlock &BB = *I.getParent();
2117
2118 auto MIB =
2119 BuildMI(BB, I, I.getDebugLoc(), TII.get(SPIRV::OpMaskedScatterINTEL))
2120 .addUse(PtrsReg)
2121 .addUse(AlignmentReg)
2122 .addUse(MaskReg)
2123 .addUse(ValuesReg);
2124 MIB.constrainAllUses(TII, TRI, RBI);
2125 return true;
2126}
2127
2128bool SPIRVInstructionSelector::diagnoseUnsupported(const MachineInstr &I,
2129 const Twine &Msg) const {
2130 const Function &F = I.getMF()->getFunction();
2131 F.getContext().diagnose(
2132 DiagnosticInfoUnsupported(F, Msg, I.getDebugLoc(), DS_Error));
2133 return false;
2134}
2135
2136bool SPIRVInstructionSelector::selectStackSave(Register ResVReg,
2137 SPIRVTypeInst ResType,
2138 MachineInstr &I) const {
2139 if (!STI.canUseExtension(SPIRV::Extension::SPV_INTEL_variable_length_array))
2140 report_fatal_error(
2141 "llvm.stacksave intrinsic: this instruction requires the following "
2142 "SPIR-V extension: SPV_INTEL_variable_length_array",
2143 false);
2144 MachineBasicBlock &BB = *I.getParent();
2145 BuildMI(BB, I, I.getDebugLoc(), TII.get(SPIRV::OpSaveMemoryINTEL))
2146 .addDef(ResVReg)
2147 .addUse(GR.getSPIRVTypeID(ResType))
2148 .constrainAllUses(TII, TRI, RBI);
2149 return true;
2150}
2151
2152bool SPIRVInstructionSelector::selectStackRestore(MachineInstr &I) const {
2153 if (!STI.canUseExtension(SPIRV::Extension::SPV_INTEL_variable_length_array))
2154 report_fatal_error(
2155 "llvm.stackrestore intrinsic: this instruction requires the following "
2156 "SPIR-V extension: SPV_INTEL_variable_length_array",
2157 false);
2158 if (!I.getOperand(0).isReg())
2159 return false;
2160 MachineBasicBlock &BB = *I.getParent();
2161 BuildMI(BB, I, I.getDebugLoc(), TII.get(SPIRV::OpRestoreMemoryINTEL))
2162 .addUse(I.getOperand(0).getReg())
2163 .constrainAllUses(TII, TRI, RBI);
2164 return true;
2165}
2166
2167Register
2168SPIRVInstructionSelector::getOrCreateMemSetGlobal(MachineInstr &I) const {
2169 MachineIRBuilder MIRBuilder(I);
2170 assert(I.getOperand(1).isReg() && I.getOperand(2).isReg());
2171
2172 // TODO: check if we have such GV, add init, use buildGlobalVariable.
2173 unsigned Num = getIConstVal(I.getOperand(2).getReg(), MRI);
2174 Function &CurFunction = GR.CurMF->getFunction();
2175 Type *LLVMArrTy =
2176 ArrayType::get(IntegerType::get(CurFunction.getContext(), 8), Num);
2177 GlobalVariable *GV = new GlobalVariable(*CurFunction.getParent(), LLVMArrTy,
2178 true, GlobalValue::InternalLinkage,
2179 Constant::getNullValue(LLVMArrTy));
2180
2181 Type *ValTy = Type::getInt8Ty(I.getMF()->getFunction().getContext());
2182 Type *ArrTy = ArrayType::get(ValTy, Num);
2183 SPIRVTypeInst VarTy = GR.getOrCreateSPIRVPointerType(
2184 ArrTy, MIRBuilder, SPIRV::StorageClass::UniformConstant);
2185
2186 SPIRVTypeInst SpvArrTy = GR.getOrCreateSPIRVType(
2187 ArrTy, MIRBuilder, SPIRV::AccessQualifier::None, false);
2188
2189 unsigned Val = getIConstVal(I.getOperand(1).getReg(), MRI);
2190 Register Const = GR.getOrCreateConstIntArray(Val, Num, I, SpvArrTy, TII);
2191
2192 Register VarReg = MRI->createGenericVirtualRegister(LLT::scalar(64));
2193 auto MIBVar =
2194 BuildMI(*I.getParent(), I, I.getDebugLoc(), TII.get(SPIRV::OpVariable))
2195 .addDef(VarReg)
2196 .addUse(GR.getSPIRVTypeID(VarTy))
2197 .addImm(SPIRV::StorageClass::UniformConstant)
2198 .addUse(Const);
2199 MIBVar.constrainAllUses(TII, TRI, RBI);
2200
2201 GR.add(GV, MIBVar);
2202 GR.addGlobalObject(GV, GR.CurMF, VarReg);
2203
2204 buildOpDecorate(VarReg, I, TII, SPIRV::Decoration::Constant, {});
2205 return VarReg;
2206}
2207
2208bool SPIRVInstructionSelector::selectCopyMemory(MachineInstr &I,
2209 Register SrcReg) const {
2210 MachineBasicBlock &BB = *I.getParent();
2211 Register DstReg = I.getOperand(0).getReg();
2212 SPIRVTypeInst DstTy = GR.getSPIRVTypeForVReg(DstReg);
2213 SPIRVTypeInst SrcTy = GR.getSPIRVTypeForVReg(SrcReg);
2214 if (GR.getPointeeType(DstTy) != GR.getPointeeType(SrcTy))
2215 report_fatal_error("OpCopyMemory requires operands to have the same type");
2216 uint64_t CopySize = getIConstVal(I.getOperand(2).getReg(), MRI);
2217 SPIRVTypeInst PointeeTy = GR.getPointeeType(DstTy);
2218 const Type *LLVMPointeeTy = GR.getTypeForSPIRVType(PointeeTy);
2219 if (!LLVMPointeeTy)
2220 report_fatal_error(
2221 "Unable to determine pointee type size for OpCopyMemory");
2222 const DataLayout &DL = I.getMF()->getFunction().getDataLayout();
2223 if (CopySize != DL.getTypeStoreSize(const_cast<Type *>(LLVMPointeeTy)))
2224 report_fatal_error(
2225 "OpCopyMemory requires the size to match the pointee type size");
2226 auto MIB = BuildMI(BB, I, I.getDebugLoc(), TII.get(SPIRV::OpCopyMemory))
2227 .addUse(DstReg)
2228 .addUse(SrcReg);
2229 if (I.getNumMemOperands()) {
2230 MachineIRBuilder MIRBuilder(I);
2231 addMemoryOperands(*I.memoperands_begin(), MIB, MIRBuilder, GR);
2232 }
2233 MIB.constrainAllUses(TII, TRI, RBI);
2234 return true;
2235}
2236
2237bool SPIRVInstructionSelector::selectCopyMemorySized(MachineInstr &I,
2238 Register SrcReg) const {
2239 MachineBasicBlock &BB = *I.getParent();
2240 auto MIB = BuildMI(BB, I, I.getDebugLoc(), TII.get(SPIRV::OpCopyMemorySized))
2241 .addUse(I.getOperand(0).getReg())
2242 .addUse(SrcReg)
2243 .addUse(I.getOperand(2).getReg());
2244 if (I.getNumMemOperands()) {
2245 MachineIRBuilder MIRBuilder(I);
2246 addMemoryOperands(*I.memoperands_begin(), MIB, MIRBuilder, GR);
2247 }
2248 MIB.constrainAllUses(TII, TRI, RBI);
2249 return true;
2250}
2251
2252bool SPIRVInstructionSelector::selectMemOperation(Register ResVReg,
2253 MachineInstr &I) const {
2254 Register SrcReg = I.getOperand(1).getReg();
2255 if (I.getOpcode() == TargetOpcode::G_MEMSET) {
2256 Register VarReg = getOrCreateMemSetGlobal(I);
2257 if (!VarReg.isValid())
2258 return false;
2259 Type *ValTy = Type::getInt8Ty(I.getMF()->getFunction().getContext());
2260 SPIRVTypeInst SourceTy = GR.getOrCreateSPIRVPointerType(
2261 ValTy, I, SPIRV::StorageClass::UniformConstant);
2262 SrcReg = MRI->createGenericVirtualRegister(LLT::scalar(64));
2263 if (!selectOpWithSrcs(SrcReg, SourceTy, I, {VarReg}, SPIRV::OpBitcast))
2264 return false;
2265 }
2266 if (STI.isLogicalSPIRV()) {
2267 if (!selectCopyMemory(I, SrcReg))
2268 return false;
2269 } else {
2270 if (!selectCopyMemorySized(I, SrcReg))
2271 return false;
2272 }
2273 if (ResVReg.isValid() && ResVReg != I.getOperand(0).getReg())
2274 if (!BuildCOPY(ResVReg, I.getOperand(0).getReg(), I))
2275 return false;
2276 return true;
2277}
2278
2279bool SPIRVInstructionSelector::selectAtomicRMW(Register ResVReg,
2280 SPIRVTypeInst ResType,
2281 MachineInstr &I,
2282 unsigned NewOpcode,
2283 unsigned NegateOpcode) const {
2284 assert(I.hasOneMemOperand());
2285 const MachineMemOperand *MemOp = *I.memoperands_begin();
2286 uint32_t Scope = static_cast<uint32_t>(getMemScope(
2287 GR.CurMF->getFunction().getContext(), MemOp->getSyncScopeID()));
2288 Register ScopeReg = buildI32Constant(Scope, I);
2289
2290 Register Ptr = I.getOperand(1).getReg();
2291 uint32_t ScSem = static_cast<uint32_t>(
2292 getMemSemanticsForStorageClass(GR.getPointerStorageClass(Ptr)));
2293 AtomicOrdering AO = MemOp->getSuccessOrdering();
2294 uint32_t MemSem = static_cast<uint32_t>(getMemSemantics(AO)) | ScSem;
2295 Register MemSemReg = buildI32Constant(MemSem, I);
2296
2297 Register ValueReg = I.getOperand(2).getReg();
2298 if (NegateOpcode != 0) {
2299 // Translation with negative value operand is requested
2300 Register TmpReg = createVirtualRegister(ResType, &GR, MRI, MRI->getMF());
2301 if (!selectOpWithSrcs(TmpReg, ResType, I, {ValueReg}, NegateOpcode))
2302 return false;
2303 ValueReg = TmpReg;
2304 }
2305
2306 BuildMI(*I.getParent(), I, I.getDebugLoc(), TII.get(NewOpcode))
2307 .addDef(ResVReg)
2308 .addUse(GR.getSPIRVTypeID(ResType))
2309 .addUse(Ptr)
2310 .addUse(ScopeReg)
2311 .addUse(MemSemReg)
2312 .addUse(ValueReg)
2313 .constrainAllUses(TII, TRI, RBI);
2314 return true;
2315}
2316
2317bool SPIRVInstructionSelector::selectUnmergeValues(MachineInstr &I) const {
2318 unsigned ArgI = I.getNumOperands() - 1;
2319 Register SrcReg =
2320 I.getOperand(ArgI).isReg() ? I.getOperand(ArgI).getReg() : Register(0);
2321 SPIRVTypeInst SrcType =
2322 SrcReg.isValid() ? GR.getSPIRVTypeForVReg(SrcReg) : nullptr;
2323 if (!SrcType || SrcType->getOpcode() != SPIRV::OpTypeVector)
2324 report_fatal_error(
2325 "cannot select G_UNMERGE_VALUES with a non-vector argument");
2326
2327 SPIRVTypeInst ScalarType = GR.getScalarOrVectorComponentType(SrcType);
2328 MachineBasicBlock &BB = *I.getParent();
2329 unsigned CurrentIndex = 0;
2330 for (unsigned i = 0; i < I.getNumDefs(); ++i) {
2331 Register ResVReg = I.getOperand(i).getReg();
2332 SPIRVTypeInst ResType = GR.getSPIRVTypeForVReg(ResVReg);
2333 if (!ResType) {
2334 LLT ResLLT = MRI->getType(ResVReg);
2335 assert(ResLLT.isValid());
2336 if (ResLLT.isVector()) {
2337 ResType = GR.getOrCreateSPIRVVectorType(
2338 ScalarType, ResLLT.getNumElements(), I, TII);
2339 } else {
2340 ResType = ScalarType;
2341 }
2342 MRI->setRegClass(ResVReg, GR.getRegClass(ResType));
2343 GR.assignSPIRVTypeToVReg(ResType, ResVReg, *GR.CurMF);
2344 }
2345
2346 if (ResType->getOpcode() == SPIRV::OpTypeVector) {
2347 Register UndefReg = GR.getOrCreateUndef(I, SrcType, TII);
2348 auto MIB =
2349 BuildMI(BB, I, I.getDebugLoc(), TII.get(SPIRV::OpVectorShuffle))
2350 .addDef(ResVReg)
2351 .addUse(GR.getSPIRVTypeID(ResType))
2352 .addUse(SrcReg)
2353 .addUse(UndefReg);
2354 unsigned NumElements = GR.getScalarOrVectorComponentCount(ResType);
2355 for (unsigned j = 0; j < NumElements; ++j) {
2356 MIB.addImm(CurrentIndex + j);
2357 }
2358 CurrentIndex += NumElements;
2359 MIB.constrainAllUses(TII, TRI, RBI);
2360 } else {
2361 auto MIB =
2362 BuildMI(BB, I, I.getDebugLoc(), TII.get(SPIRV::OpCompositeExtract))
2363 .addDef(ResVReg)
2364 .addUse(GR.getSPIRVTypeID(ResType))
2365 .addUse(SrcReg)
2366 .addImm(CurrentIndex);
2367 CurrentIndex++;
2368 MIB.constrainAllUses(TII, TRI, RBI);
2369 }
2370 }
2371 return true;
2372}
2373
2374bool SPIRVInstructionSelector::selectFence(MachineInstr &I) const {
2375 AtomicOrdering AO = AtomicOrdering(I.getOperand(0).getImm());
2376 uint32_t MemSem = static_cast<uint32_t>(getMemSemantics(AO));
2377 Register MemSemReg = buildI32Constant(MemSem, I);
2378 SyncScope::ID Ord = SyncScope::ID(I.getOperand(1).getImm());
2379 uint32_t Scope = static_cast<uint32_t>(
2380 getMemScope(GR.CurMF->getFunction().getContext(), Ord));
2381 Register ScopeReg = buildI32Constant(Scope, I);
2382 MachineBasicBlock &BB = *I.getParent();
2383 BuildMI(BB, I, I.getDebugLoc(), TII.get(SPIRV::OpMemoryBarrier))
2384 .addUse(ScopeReg)
2385 .addUse(MemSemReg)
2386 .constrainAllUses(TII, TRI, RBI);
2387 return true;
2388}
2389
2390bool SPIRVInstructionSelector::selectOverflowArith(Register ResVReg,
2391 SPIRVTypeInst ResType,
2392 MachineInstr &I,
2393 unsigned Opcode) const {
2394 Type *ResTy = nullptr;
2395 StringRef ResName;
2396 if (!GR.findValueAttrs(&I, ResTy, ResName))
2397 report_fatal_error(
2398 "Not enough info to select the arithmetic with overflow instruction");
2399 if (!ResTy || !ResTy->isStructTy())
2400 report_fatal_error("Expect struct type result for the arithmetic "
2401 "with overflow instruction");
2402 // "Result Type must be from OpTypeStruct. The struct must have two members,
2403 // and the two members must be the same type."
2404 Type *ResElemTy = cast<StructType>(ResTy)->getElementType(0);
2405 ResTy = StructType::get(ResElemTy, ResElemTy);
2406 // Build SPIR-V types and constant(s) if needed.
2407 MachineIRBuilder MIRBuilder(I);
2408 SPIRVTypeInst StructType = GR.getOrCreateSPIRVType(
2409 ResTy, MIRBuilder, SPIRV::AccessQualifier::ReadWrite, false);
2410 assert(I.getNumDefs() > 1 && "Not enought operands");
2411 SPIRVTypeInst BoolType = GR.getOrCreateSPIRVBoolType(I, TII);
2412 unsigned N = GR.getScalarOrVectorComponentCount(ResType);
2413 if (N > 1)
2414 BoolType = GR.getOrCreateSPIRVVectorType(BoolType, N, I, TII);
2415 Register BoolTypeReg = GR.getSPIRVTypeID(BoolType);
2416 Register ZeroReg = buildZerosVal(ResType, I);
2417 // A new virtual register to store the result struct.
2418 Register StructVReg = MRI->createGenericVirtualRegister(LLT::scalar(64));
2419 MRI->setRegClass(StructVReg, &SPIRV::IDRegClass);
2420 // Build the result name if needed.
2421 if (ResName.size() > 0)
2422 buildOpName(StructVReg, ResName, MIRBuilder);
2423 // Build the arithmetic with overflow instruction.
2424 MachineBasicBlock &BB = *I.getParent();
2425 auto MIB =
2426 BuildMI(BB, MIRBuilder.getInsertPt(), I.getDebugLoc(), TII.get(Opcode))
2427 .addDef(StructVReg)
2428 .addUse(GR.getSPIRVTypeID(StructType));
2429 for (unsigned i = I.getNumDefs(); i < I.getNumOperands(); ++i)
2430 MIB.addUse(I.getOperand(i).getReg());
2431 MIB.constrainAllUses(TII, TRI, RBI);
2432 // Build instructions to extract fields of the instruction's result.
2433 // A new virtual register to store the higher part of the result struct.
2434 Register HigherVReg = MRI->createGenericVirtualRegister(LLT::scalar(64));
2435 MRI->setRegClass(HigherVReg, &SPIRV::iIDRegClass);
2436 for (unsigned i = 0; i < I.getNumDefs(); ++i) {
2437 auto MIB =
2438 BuildMI(BB, I, I.getDebugLoc(), TII.get(SPIRV::OpCompositeExtract))
2439 .addDef(i == 1 ? HigherVReg : I.getOperand(i).getReg())
2440 .addUse(GR.getSPIRVTypeID(ResType))
2441 .addUse(StructVReg)
2442 .addImm(i);
2443 MIB.constrainAllUses(TII, TRI, RBI);
2444 }
2445 // Build boolean value from the higher part.
2446 BuildMI(BB, I, I.getDebugLoc(), TII.get(SPIRV::OpINotEqual))
2447 .addDef(I.getOperand(1).getReg())
2448 .addUse(BoolTypeReg)
2449 .addUse(HigherVReg)
2450 .addUse(ZeroReg)
2451 .constrainAllUses(TII, TRI, RBI);
2452 return true;
2453}
2454
2455bool SPIRVInstructionSelector::selectAtomicCmpXchg(Register ResVReg,
2456 SPIRVTypeInst ResType,
2457 MachineInstr &I) const {
2458 assert(isa<GIntrinsic>(I) &&
2459 "selectAtomicCmpXchg only handles the spv_cmpxchg intrinsic");
2460 Register Ptr = I.getOperand(2).getReg();
2461 Register ScopeReg = I.getOperand(5).getReg();
2462 Register MemSemEqReg = I.getOperand(6).getReg();
2463 Register MemSemNeqReg = I.getOperand(7).getReg();
2464 Register Cmp = I.getOperand(3).getReg();
2465 Register Val = I.getOperand(4).getReg();
2466 SPIRVTypeInst SpvValTy = GR.getSPIRVTypeForVReg(Val);
2467 Register ACmpRes = createVirtualRegister(SpvValTy, &GR, MRI, *I.getMF());
2468 const DebugLoc &DL = I.getDebugLoc();
2469 BuildMI(*I.getParent(), I, DL, TII.get(SPIRV::OpAtomicCompareExchange))
2470 .addDef(ACmpRes)
2471 .addUse(GR.getSPIRVTypeID(SpvValTy))
2472 .addUse(Ptr)
2473 .addUse(ScopeReg)
2474 .addUse(MemSemEqReg)
2475 .addUse(MemSemNeqReg)
2476 .addUse(Val)
2477 .addUse(Cmp)
2478 .constrainAllUses(TII, TRI, RBI);
2479 SPIRVTypeInst BoolTy = GR.getOrCreateSPIRVBoolType(I, TII);
2480 Register CmpSuccReg = createVirtualRegister(BoolTy, &GR, MRI, *I.getMF());
2481 BuildMI(*I.getParent(), I, DL, TII.get(SPIRV::OpIEqual))
2482 .addDef(CmpSuccReg)
2483 .addUse(GR.getSPIRVTypeID(BoolTy))
2484 .addUse(ACmpRes)
2485 .addUse(Cmp)
2486 .constrainAllUses(TII, TRI, RBI);
2487 Register TmpReg = createVirtualRegister(ResType, &GR, MRI, *I.getMF());
2488 BuildMI(*I.getParent(), I, DL, TII.get(SPIRV::OpCompositeInsert))
2489 .addDef(TmpReg)
2490 .addUse(GR.getSPIRVTypeID(ResType))
2491 .addUse(ACmpRes)
2492 .addUse(GR.getOrCreateUndef(I, ResType, TII))
2493 .addImm(0)
2494 .constrainAllUses(TII, TRI, RBI);
2495 BuildMI(*I.getParent(), I, DL, TII.get(SPIRV::OpCompositeInsert))
2496 .addDef(ResVReg)
2497 .addUse(GR.getSPIRVTypeID(ResType))
2498 .addUse(CmpSuccReg)
2499 .addUse(TmpReg)
2500 .addImm(1)
2501 .constrainAllUses(TII, TRI, RBI);
2502 return true;
2503}
2504
2505static bool isUSMStorageClass(SPIRV::StorageClass::StorageClass SC) {
2506 switch (SC) {
2507 case SPIRV::StorageClass::DeviceOnlyINTEL:
2508 case SPIRV::StorageClass::HostOnlyINTEL:
2509 return true;
2510 default:
2511 return false;
2512 }
2513}
2514
2515// Returns true ResVReg is referred only from global vars and OpName's.
2516static bool isASCastInGVar(MachineRegisterInfo *MRI, Register ResVReg) {
2517 bool IsGRef = false;
2518 bool IsAllowedRefs =
2519 llvm::all_of(MRI->use_instructions(ResVReg), [&IsGRef](auto const &It) {
2520 unsigned Opcode = It.getOpcode();
2521 if (Opcode == SPIRV::OpConstantComposite ||
2522 Opcode == SPIRV::OpSpecConstantComposite ||
2523 Opcode == SPIRV::OpVariable ||
2524 isSpvIntrinsic(It, Intrinsic::spv_init_global))
2525 return IsGRef = true;
2526 return Opcode == SPIRV::OpName;
2527 });
2528 return IsAllowedRefs && IsGRef;
2529}
2530
2531Register SPIRVInstructionSelector::getUcharPtrTypeReg(
2532 MachineInstr &I, SPIRV::StorageClass::StorageClass SC) const {
2533 return GR.getSPIRVTypeID(GR.getOrCreateSPIRVPointerType(
2534 Type::getInt8Ty(I.getMF()->getFunction().getContext()), I, SC));
2535}
2536
2537MachineInstrBuilder
2538SPIRVInstructionSelector::buildSpecConstantOp(MachineInstr &I, Register Dest,
2539 Register Src, Register DestType,
2540 uint32_t Opcode) const {
2541 return BuildMI(*I.getParent(), I, I.getDebugLoc(),
2542 TII.get(SPIRV::OpSpecConstantOp))
2543 .addDef(Dest)
2544 .addUse(DestType)
2545 .addImm(Opcode)
2546 .addUse(Src);
2547}
2548
2549MachineInstrBuilder
2550SPIRVInstructionSelector::buildConstGenericPtr(MachineInstr &I, Register SrcPtr,
2551 SPIRVTypeInst SrcPtrTy) const {
2552 SPIRVTypeInst GenericPtrTy =
2553 GR.changePointerStorageClass(SrcPtrTy, SPIRV::StorageClass::Generic, I);
2554 Register Tmp = MRI->createVirtualRegister(&SPIRV::pIDRegClass);
2555 MRI->setType(Tmp, LLT::pointer(storageClassToAddressSpace(
2556 SPIRV::StorageClass::Generic),
2557 GR.getPointerSize()));
2558 MachineFunction *MF = I.getParent()->getParent();
2559 GR.assignSPIRVTypeToVReg(GenericPtrTy, Tmp, *MF);
2560 MachineInstrBuilder MIB = buildSpecConstantOp(
2561 I, Tmp, SrcPtr, GR.getSPIRVTypeID(GenericPtrTy),
2562 static_cast<uint32_t>(SPIRV::Opcode::PtrCastToGeneric));
2563 GR.add(MIB.getInstr(), MIB);
2564 return MIB;
2565}
2566
2567// In SPIR-V address space casting can only happen to and from the Generic
2568// storage class. We can also only cast Workgroup, CrossWorkgroup, or Function
2569// pointers to and from Generic pointers. As such, we can convert e.g. from
2570// Workgroup to Function by going via a Generic pointer as an intermediary. All
2571// other combinations can only be done by a bitcast, and are probably not safe.
2572bool SPIRVInstructionSelector::selectAddrSpaceCast(Register ResVReg,
2573 SPIRVTypeInst ResType,
2574 MachineInstr &I) const {
2575 MachineBasicBlock &BB = *I.getParent();
2576 const DebugLoc &DL = I.getDebugLoc();
2577
2578 Register SrcPtr = I.getOperand(1).getReg();
2579 SPIRVTypeInst SrcPtrTy = GR.getSPIRVTypeForVReg(SrcPtr);
2580
2581 // don't generate a cast for a null that may be represented by OpTypeInt
2582 if (SrcPtrTy->getOpcode() != SPIRV::OpTypePointer ||
2583 ResType->getOpcode() != SPIRV::OpTypePointer)
2584 return BuildCOPY(ResVReg, SrcPtr, I);
2585
2586 SPIRV::StorageClass::StorageClass SrcSC = GR.getPointerStorageClass(SrcPtrTy);
2587 SPIRV::StorageClass::StorageClass DstSC = GR.getPointerStorageClass(ResType);
2588
2589 if (isASCastInGVar(MRI, ResVReg)) {
2590 // AddrSpaceCast uses within OpVariable and OpConstantComposite instructions
2591 // are expressed by OpSpecConstantOp with an Opcode.
2592 // TODO: maybe insert a check whether the Kernel capability was declared and
2593 // so PtrCastToGeneric/GenericCastToPtr are available.
2594 unsigned SpecOpcode =
2595 DstSC == SPIRV::StorageClass::Generic && isGenericCastablePtr(SrcSC)
2596 ? static_cast<uint32_t>(SPIRV::Opcode::PtrCastToGeneric)
2597 : (SrcSC == SPIRV::StorageClass::Generic &&
2598 isGenericCastablePtr(DstSC)
2599 ? static_cast<uint32_t>(SPIRV::Opcode::GenericCastToPtr)
2600 : 0);
2601 // TODO: OpConstantComposite expects i8*, so we are forced to forget a
2602 // correct value of ResType and use general i8* instead. Maybe this should
2603 // be addressed in the emit-intrinsic step to infer a correct
2604 // OpConstantComposite type.
2605 if (SpecOpcode) {
2606 buildSpecConstantOp(I, ResVReg, SrcPtr, getUcharPtrTypeReg(I, DstSC),
2607 SpecOpcode)
2608 .constrainAllUses(TII, TRI, RBI);
2609 } else if (isGenericCastablePtr(SrcSC) && isGenericCastablePtr(DstSC)) {
2610 MachineInstrBuilder MIB = buildConstGenericPtr(I, SrcPtr, SrcPtrTy);
2611 MIB.constrainAllUses(TII, TRI, RBI);
2612 buildSpecConstantOp(
2613 I, ResVReg, MIB->getOperand(0).getReg(), getUcharPtrTypeReg(I, DstSC),
2614 static_cast<uint32_t>(SPIRV::Opcode::GenericCastToPtr))
2615 .constrainAllUses(TII, TRI, RBI);
2616 }
2617 return true;
2618 }
2619
2620 // don't generate a cast between identical storage classes
2621 if (SrcSC == DstSC)
2622 return BuildCOPY(ResVReg, SrcPtr, I);
2623
2624 if ((SrcSC == SPIRV::StorageClass::Function &&
2625 DstSC == SPIRV::StorageClass::Private) ||
2626 (DstSC == SPIRV::StorageClass::Function &&
2627 SrcSC == SPIRV::StorageClass::Private))
2628 return BuildCOPY(ResVReg, SrcPtr, I);
2629
2630 // Casting from an eligible pointer to Generic.
2631 if (DstSC == SPIRV::StorageClass::Generic && isGenericCastablePtr(SrcSC))
2632 return selectUnOp(ResVReg, ResType, I, SPIRV::OpPtrCastToGeneric);
2633 // Casting from Generic to an eligible pointer.
2634 if (SrcSC == SPIRV::StorageClass::Generic && isGenericCastablePtr(DstSC))
2635 return selectUnOp(ResVReg, ResType, I, SPIRV::OpGenericCastToPtr);
2636 // Casting between 2 eligible pointers using Generic as an intermediary.
2637 if (isGenericCastablePtr(SrcSC) && isGenericCastablePtr(DstSC)) {
2638 SPIRVTypeInst GenericPtrTy =
2639 GR.changePointerStorageClass(SrcPtrTy, SPIRV::StorageClass::Generic, I);
2640 Register Tmp = createVirtualRegister(GenericPtrTy, &GR, MRI, MRI->getMF());
2641 BuildMI(BB, I, DL, TII.get(SPIRV::OpPtrCastToGeneric))
2642 .addDef(Tmp)
2643 .addUse(GR.getSPIRVTypeID(GenericPtrTy))
2644 .addUse(SrcPtr)
2645 .constrainAllUses(TII, TRI, RBI);
2646 BuildMI(BB, I, DL, TII.get(SPIRV::OpGenericCastToPtr))
2647 .addDef(ResVReg)
2648 .addUse(GR.getSPIRVTypeID(ResType))
2649 .addUse(Tmp)
2650 .constrainAllUses(TII, TRI, RBI);
2651 return true;
2652 }
2653
2654 // Check if instructions from the SPV_INTEL_usm_storage_classes extension may
2655 // be applied
2656 if (isUSMStorageClass(SrcSC) && DstSC == SPIRV::StorageClass::CrossWorkgroup)
2657 return selectUnOp(ResVReg, ResType, I,
2658 SPIRV::OpPtrCastToCrossWorkgroupINTEL);
2659 if (SrcSC == SPIRV::StorageClass::CrossWorkgroup && isUSMStorageClass(DstSC))
2660 return selectUnOp(ResVReg, ResType, I,
2661 SPIRV::OpCrossWorkgroupCastToPtrINTEL);
2662 if (isUSMStorageClass(SrcSC) && DstSC == SPIRV::StorageClass::Generic)
2663 return selectUnOp(ResVReg, ResType, I, SPIRV::OpPtrCastToGeneric);
2664 if (SrcSC == SPIRV::StorageClass::Generic && isUSMStorageClass(DstSC))
2665 return selectUnOp(ResVReg, ResType, I, SPIRV::OpGenericCastToPtr);
2666
2667 // Bitcast for pointers requires that the address spaces must match
2668 return false;
2669}
2670
2671static unsigned getFCmpOpcode(unsigned PredNum) {
2672 auto Pred = static_cast<CmpInst::Predicate>(PredNum);
2673 switch (Pred) {
2674 case CmpInst::FCMP_OEQ:
2675 return SPIRV::OpFOrdEqual;
2676 case CmpInst::FCMP_OGE:
2677 return SPIRV::OpFOrdGreaterThanEqual;
2678 case CmpInst::FCMP_OGT:
2679 return SPIRV::OpFOrdGreaterThan;
2680 case CmpInst::FCMP_OLE:
2681 return SPIRV::OpFOrdLessThanEqual;
2682 case CmpInst::FCMP_OLT:
2683 return SPIRV::OpFOrdLessThan;
2684 case CmpInst::FCMP_ONE:
2685 return SPIRV::OpFOrdNotEqual;
2686 case CmpInst::FCMP_ORD:
2687 return SPIRV::OpOrdered;
2688 case CmpInst::FCMP_UEQ:
2689 return SPIRV::OpFUnordEqual;
2690 case CmpInst::FCMP_UGE:
2691 return SPIRV::OpFUnordGreaterThanEqual;
2692 case CmpInst::FCMP_UGT:
2693 return SPIRV::OpFUnordGreaterThan;
2694 case CmpInst::FCMP_ULE:
2695 return SPIRV::OpFUnordLessThanEqual;
2696 case CmpInst::FCMP_ULT:
2697 return SPIRV::OpFUnordLessThan;
2698 case CmpInst::FCMP_UNE:
2699 return SPIRV::OpFUnordNotEqual;
2700 case CmpInst::FCMP_UNO:
2701 return SPIRV::OpUnordered;
2702 default:
2703 llvm_unreachable("Unknown predicate type for FCmp");
2704 }
2705}
2706
2707static unsigned getICmpOpcode(unsigned PredNum) {
2708 auto Pred = static_cast<CmpInst::Predicate>(PredNum);
2709 switch (Pred) {
2710 case CmpInst::ICMP_EQ:
2711 return SPIRV::OpIEqual;
2712 case CmpInst::ICMP_NE:
2713 return SPIRV::OpINotEqual;
2714 case CmpInst::ICMP_SGE:
2715 return SPIRV::OpSGreaterThanEqual;
2716 case CmpInst::ICMP_SGT:
2717 return SPIRV::OpSGreaterThan;
2718 case CmpInst::ICMP_SLE:
2719 return SPIRV::OpSLessThanEqual;
2720 case CmpInst::ICMP_SLT:
2721 return SPIRV::OpSLessThan;
2722 case CmpInst::ICMP_UGE:
2723 return SPIRV::OpUGreaterThanEqual;
2724 case CmpInst::ICMP_UGT:
2725 return SPIRV::OpUGreaterThan;
2726 case CmpInst::ICMP_ULE:
2727 return SPIRV::OpULessThanEqual;
2728 case CmpInst::ICMP_ULT:
2729 return SPIRV::OpULessThan;
2730 default:
2731 llvm_unreachable("Unknown predicate type for ICmp");
2732 }
2733}
2734
2735static unsigned getPtrCmpOpcode(unsigned Pred) {
2736 switch (static_cast<CmpInst::Predicate>(Pred)) {
2737 case CmpInst::ICMP_EQ:
2738 return SPIRV::OpPtrEqual;
2739 case CmpInst::ICMP_NE:
2740 return SPIRV::OpPtrNotEqual;
2741 default:
2742 llvm_unreachable("Unknown predicate type for pointer comparison");
2743 }
2744}
2745
2746// Return the logical operation, or abort if none exists.
2747static unsigned getBoolCmpOpcode(unsigned PredNum) {
2748 auto Pred = static_cast<CmpInst::Predicate>(PredNum);
2749 switch (Pred) {
2750 case CmpInst::ICMP_EQ:
2751 return SPIRV::OpLogicalEqual;
2752 case CmpInst::ICMP_NE:
2753 return SPIRV::OpLogicalNotEqual;
2754 default:
2755 llvm_unreachable("Unknown predicate type for Bool comparison");
2756 }
2757}
2758
2759static APFloat getZeroFP(const Type *LLVMFloatTy) {
2760 if (!LLVMFloatTy)
2761 return APFloat::getZero(APFloat::IEEEsingle());
2762 switch (LLVMFloatTy->getScalarType()->getTypeID()) {
2763 case Type::HalfTyID:
2764 return APFloat::getZero(APFloat::IEEEhalf());
2765 default:
2766 case Type::FloatTyID:
2767 return APFloat::getZero(APFloat::IEEEsingle());
2768 case Type::DoubleTyID:
2769 return APFloat::getZero(APFloat::IEEEdouble());
2770 }
2771}
2772
2773static APFloat getOneFP(const Type *LLVMFloatTy) {
2774 if (!LLVMFloatTy)
2775 return APFloat::getOne(APFloat::IEEEsingle());
2776 switch (LLVMFloatTy->getScalarType()->getTypeID()) {
2777 case Type::HalfTyID:
2778 return APFloat::getOne(APFloat::IEEEhalf());
2779 default:
2780 case Type::FloatTyID:
2781 return APFloat::getOne(APFloat::IEEEsingle());
2782 case Type::DoubleTyID:
2783 return APFloat::getOne(APFloat::IEEEdouble());
2784 }
2785}
2786
2787bool SPIRVInstructionSelector::selectAnyOrAll(Register ResVReg,
2788 SPIRVTypeInst ResType,
2789 MachineInstr &I,
2790 unsigned OpAnyOrAll) const {
2791 assert(I.getNumOperands() == 3);
2792 assert(I.getOperand(2).isReg());
2793 MachineBasicBlock &BB = *I.getParent();
2794 Register InputRegister = I.getOperand(2).getReg();
2795 SPIRVTypeInst InputType = GR.getSPIRVTypeForVReg(InputRegister);
2796
2797 assert(InputType && "VReg has no type assigned");
2798
2799 bool IsBoolTy = GR.isScalarOrVectorOfType(InputRegister, SPIRV::OpTypeBool);
2800 bool IsVectorTy = InputType->getOpcode() == SPIRV::OpTypeVector;
2801 if (IsBoolTy && !IsVectorTy) {
2802 assert(ResVReg == I.getOperand(0).getReg());
2803 return BuildCOPY(ResVReg, InputRegister, I);
2804 }
2805
2806 bool IsFloatTy = GR.isScalarOrVectorOfType(InputRegister, SPIRV::OpTypeFloat);
2807 unsigned SpirvNotEqualId =
2808 IsFloatTy ? SPIRV::OpFOrdNotEqual : SPIRV::OpINotEqual;
2809 SPIRVTypeInst SpvBoolScalarTy = GR.getOrCreateSPIRVBoolType(I, TII);
2810 SPIRVTypeInst SpvBoolTy = SpvBoolScalarTy;
2811 Register NotEqualReg = ResVReg;
2812
2813 if (IsVectorTy) {
2814 NotEqualReg =
2815 IsBoolTy ? InputRegister
2816 : createVirtualRegister(SpvBoolTy, &GR, MRI, MRI->getMF());
2817 const unsigned NumElts = GR.getScalarOrVectorComponentCount(InputType);
2818 SpvBoolTy = GR.getOrCreateSPIRVVectorType(SpvBoolTy, NumElts, I, TII);
2819 }
2820
2821 if (!IsBoolTy) {
2822 Register ConstZeroReg =
2823 IsFloatTy ? buildZerosValF(InputType, I) : buildZerosVal(InputType, I);
2824
2825 BuildMI(BB, I, I.getDebugLoc(), TII.get(SpirvNotEqualId))
2826 .addDef(NotEqualReg)
2827 .addUse(GR.getSPIRVTypeID(SpvBoolTy))
2828 .addUse(InputRegister)
2829 .addUse(ConstZeroReg)
2830 .constrainAllUses(TII, TRI, RBI);
2831 }
2832
2833 if (IsVectorTy)
2834 BuildMI(BB, I, I.getDebugLoc(), TII.get(OpAnyOrAll))
2835 .addDef(ResVReg)
2836 .addUse(GR.getSPIRVTypeID(SpvBoolScalarTy))
2837 .addUse(NotEqualReg)
2838 .constrainAllUses(TII, TRI, RBI);
2839 return true;
2840}
2841
2842bool SPIRVInstructionSelector::selectAll(Register ResVReg,
2843 SPIRVTypeInst ResType,
2844 MachineInstr &I) const {
2845 return selectAnyOrAll(ResVReg, ResType, I, SPIRV::OpAll);
2846}
2847
2848bool SPIRVInstructionSelector::selectAny(Register ResVReg,
2849 SPIRVTypeInst ResType,
2850 MachineInstr &I) const {
2851 return selectAnyOrAll(ResVReg, ResType, I, SPIRV::OpAny);
2852}
2853
2854// Select the OpDot instruction for the given float dot
2855bool SPIRVInstructionSelector::selectFloatDot(Register ResVReg,
2856 SPIRVTypeInst ResType,
2857 MachineInstr &I) const {
2858 assert(I.getNumOperands() == 4);
2859 assert(I.getOperand(2).isReg());
2860 assert(I.getOperand(3).isReg());
2861
2862 [[maybe_unused]] SPIRVTypeInst VecType =
2863 GR.getSPIRVTypeForVReg(I.getOperand(2).getReg());
2864
2865 assert(VecType->getOpcode() == SPIRV::OpTypeVector &&
2866 GR.getScalarOrVectorComponentCount(VecType) > 1 &&
2867 "dot product requires a vector of at least 2 components");
2868
2869 [[maybe_unused]] SPIRVTypeInst EltType =
2870 GR.getScalarOrVectorComponentType(VecType);
2871
2872 assert(EltType->getOpcode() == SPIRV::OpTypeFloat);
2873
2874 MachineBasicBlock &BB = *I.getParent();
2875 BuildMI(BB, I, I.getDebugLoc(), TII.get(SPIRV::OpDot))
2876 .addDef(ResVReg)
2877 .addUse(GR.getSPIRVTypeID(ResType))
2878 .addUse(I.getOperand(2).getReg())
2879 .addUse(I.getOperand(3).getReg())
2880 .constrainAllUses(TII, TRI, RBI);
2881 return true;
2882}
2883
2884bool SPIRVInstructionSelector::selectIntegerDot(Register ResVReg,
2885 SPIRVTypeInst ResType,
2886 MachineInstr &I,
2887 bool Signed) const {
2888 assert(I.getNumOperands() == 4);
2889 assert(I.getOperand(2).isReg());
2890 assert(I.getOperand(3).isReg());
2891 MachineBasicBlock &BB = *I.getParent();
2892
2893 auto DotOp = Signed ? SPIRV::OpSDot : SPIRV::OpUDot;
2894 BuildMI(BB, I, I.getDebugLoc(), TII.get(DotOp))
2895 .addDef(ResVReg)
2896 .addUse(GR.getSPIRVTypeID(ResType))
2897 .addUse(I.getOperand(2).getReg())
2898 .addUse(I.getOperand(3).getReg())
2899 .constrainAllUses(TII, TRI, RBI);
2900 return true;
2901}
2902
2903// Since pre-1.6 SPIRV has no integer dot implementation,
2904// expand by piecewise multiplying and adding the results
2905bool SPIRVInstructionSelector::selectIntegerDotExpansion(
2906 Register ResVReg, SPIRVTypeInst ResType, MachineInstr &I) const {
2907 assert(I.getNumOperands() == 4);
2908 assert(I.getOperand(2).isReg());
2909 assert(I.getOperand(3).isReg());
2910 MachineBasicBlock &BB = *I.getParent();
2911
2912 // Multiply the vectors, then sum the results
2913 Register Vec0 = I.getOperand(2).getReg();
2914 Register Vec1 = I.getOperand(3).getReg();
2915 Register TmpVec = MRI->createVirtualRegister(GR.getRegClass(ResType));
2916 SPIRVTypeInst VecType = GR.getSPIRVTypeForVReg(Vec0);
2917
2918 BuildMI(BB, I, I.getDebugLoc(), TII.get(SPIRV::OpIMulV))
2919 .addDef(TmpVec)
2920 .addUse(GR.getSPIRVTypeID(VecType))
2921 .addUse(Vec0)
2922 .addUse(Vec1)
2923 .constrainAllUses(TII, TRI, RBI);
2924
2925 assert(VecType->getOpcode() == SPIRV::OpTypeVector &&
2926 GR.getScalarOrVectorComponentCount(VecType) > 1 &&
2927 "dot product requires a vector of at least 2 components");
2928
2929 Register Res = MRI->createVirtualRegister(GR.getRegClass(ResType));
2930 BuildMI(BB, I, I.getDebugLoc(), TII.get(SPIRV::OpCompositeExtract))
2931 .addDef(Res)
2932 .addUse(GR.getSPIRVTypeID(ResType))
2933 .addUse(TmpVec)
2934 .addImm(0)
2935 .constrainAllUses(TII, TRI, RBI);
2936
2937 for (unsigned i = 1; i < GR.getScalarOrVectorComponentCount(VecType); i++) {
2938 Register Elt = MRI->createVirtualRegister(GR.getRegClass(ResType));
2939
2940 BuildMI(BB, I, I.getDebugLoc(), TII.get(SPIRV::OpCompositeExtract))
2941 .addDef(Elt)
2942 .addUse(GR.getSPIRVTypeID(ResType))
2943 .addUse(TmpVec)
2944 .addImm(i)
2945 .constrainAllUses(TII, TRI, RBI);
2946
2947 Register Sum = i < GR.getScalarOrVectorComponentCount(VecType) - 1
2948 ? MRI->createVirtualRegister(GR.getRegClass(ResType))
2949 : ResVReg;
2950
2951 BuildMI(BB, I, I.getDebugLoc(), TII.get(SPIRV::OpIAddS))
2952 .addDef(Sum)
2953 .addUse(GR.getSPIRVTypeID(ResType))
2954 .addUse(Res)
2955 .addUse(Elt)
2956 .constrainAllUses(TII, TRI, RBI);
2957 Res = Sum;
2958 }
2959
2960 return true;
2961}
2962
2963bool SPIRVInstructionSelector::selectOpIsInf(Register ResVReg,
2964 SPIRVTypeInst ResType,
2965 MachineInstr &I) const {
2966 MachineBasicBlock &BB = *I.getParent();
2967 BuildMI(BB, I, I.getDebugLoc(), TII.get(SPIRV::OpIsInf))
2968 .addDef(ResVReg)
2969 .addUse(GR.getSPIRVTypeID(ResType))
2970 .addUse(I.getOperand(2).getReg())
2971 .constrainAllUses(TII, TRI, RBI);
2972 return true;
2973}
2974
2975bool SPIRVInstructionSelector::selectOpIsNan(Register ResVReg,
2976 SPIRVTypeInst ResType,
2977 MachineInstr &I) const {
2978 MachineBasicBlock &BB = *I.getParent();
2979 BuildMI(BB, I, I.getDebugLoc(), TII.get(SPIRV::OpIsNan))
2980 .addDef(ResVReg)
2981 .addUse(GR.getSPIRVTypeID(ResType))
2982 .addUse(I.getOperand(2).getReg())
2983 .constrainAllUses(TII, TRI, RBI);
2984 return true;
2985}
2986
2987template <bool Signed>
2988bool SPIRVInstructionSelector::selectDot4AddPacked(Register ResVReg,
2989 SPIRVTypeInst ResType,
2990 MachineInstr &I) const {
2991 assert(I.getNumOperands() == 5);
2992 assert(I.getOperand(2).isReg());
2993 assert(I.getOperand(3).isReg());
2994 assert(I.getOperand(4).isReg());
2995 MachineBasicBlock &BB = *I.getParent();
2996
2997 Register Acc = I.getOperand(2).getReg();
2998 Register X = I.getOperand(3).getReg();
2999 Register Y = I.getOperand(4).getReg();
3000
3001 auto DotOp = Signed ? SPIRV::OpSDot : SPIRV::OpUDot;
3002 Register Dot = MRI->createVirtualRegister(GR.getRegClass(ResType));
3003 auto MIB = BuildMI(BB, I, I.getDebugLoc(), TII.get(DotOp))
3004 .addDef(Dot)
3005 .addUse(GR.getSPIRVTypeID(ResType))
3006 .addUse(X)
3007 .addUse(Y);
3008 MIB.addImm(SPIRV::BuiltIn::PackedVectorFormat4x8Bit);
3009 MIB.constrainAllUses(TII, TRI, RBI);
3010
3011 BuildMI(BB, I, I.getDebugLoc(), TII.get(SPIRV::OpIAddS))
3012 .addDef(ResVReg)
3013 .addUse(GR.getSPIRVTypeID(ResType))
3014 .addUse(Dot)
3015 .addUse(Acc)
3016 .constrainAllUses(TII, TRI, RBI);
3017 return true;
3018}
3019
3020// Since pre-1.6 SPIRV has no DotProductInput4x8BitPacked implementation,
3021// extract the elements of the packed inputs, multiply them and add the result
3022// to the accumulator.
3023template <bool Signed>
3024bool SPIRVInstructionSelector::selectDot4AddPackedExpansion(
3025 Register ResVReg, SPIRVTypeInst ResType, MachineInstr &I) const {
3026 assert(I.getNumOperands() == 5);
3027 assert(I.getOperand(2).isReg());
3028 assert(I.getOperand(3).isReg());
3029 assert(I.getOperand(4).isReg());
3030 MachineBasicBlock &BB = *I.getParent();
3031
3032 Register Acc = I.getOperand(2).getReg();
3033 Register X = I.getOperand(3).getReg();
3034 Register Y = I.getOperand(4).getReg();
3035
3036 SPIRVTypeInst EltType = GR.getOrCreateSPIRVIntegerType(8, I, TII);
3037 auto ExtractOp =
3038 Signed ? SPIRV::OpBitFieldSExtract : SPIRV::OpBitFieldUExtract;
3039
3040 bool ZeroAsNull = !STI.isShader();
3041 // Extract the i8 element, multiply and add it to the accumulator
3042 for (unsigned i = 0; i < 4; i++) {
3043 // A[i]
3044 Register AElt = MRI->createVirtualRegister(&SPIRV::IDRegClass);
3045 BuildMI(BB, I, I.getDebugLoc(), TII.get(ExtractOp))
3046 .addDef(AElt)
3047 .addUse(GR.getSPIRVTypeID(ResType))
3048 .addUse(X)
3049 .addUse(GR.getOrCreateConstInt(i * 8, I, EltType, TII, ZeroAsNull))
3050 .addUse(GR.getOrCreateConstInt(8, I, EltType, TII, ZeroAsNull))
3051 .constrainAllUses(TII, TRI, RBI);
3052
3053 // B[i]
3054 Register BElt = MRI->createVirtualRegister(&SPIRV::IDRegClass);
3055 BuildMI(BB, I, I.getDebugLoc(), TII.get(ExtractOp))
3056 .addDef(BElt)
3057 .addUse(GR.getSPIRVTypeID(ResType))
3058 .addUse(Y)
3059 .addUse(GR.getOrCreateConstInt(i * 8, I, EltType, TII, ZeroAsNull))
3060 .addUse(GR.getOrCreateConstInt(8, I, EltType, TII, ZeroAsNull))
3061 .constrainAllUses(TII, TRI, RBI);
3062
3063 // A[i] * B[i]
3064 Register Mul = MRI->createVirtualRegister(&SPIRV::IDRegClass);
3065 BuildMI(BB, I, I.getDebugLoc(), TII.get(SPIRV::OpIMulS))
3066 .addDef(Mul)
3067 .addUse(GR.getSPIRVTypeID(ResType))
3068 .addUse(AElt)
3069 .addUse(BElt)
3070 .constrainAllUses(TII, TRI, RBI);
3071
3072 // Discard 24 highest-bits so that stored i32 register is i8 equivalent
3073 Register MaskMul = MRI->createVirtualRegister(&SPIRV::IDRegClass);
3074 BuildMI(BB, I, I.getDebugLoc(), TII.get(ExtractOp))
3075 .addDef(MaskMul)
3076 .addUse(GR.getSPIRVTypeID(ResType))
3077 .addUse(Mul)
3078 .addUse(GR.getOrCreateConstInt(0, I, EltType, TII, ZeroAsNull))
3079 .addUse(GR.getOrCreateConstInt(8, I, EltType, TII, ZeroAsNull))
3080 .constrainAllUses(TII, TRI, RBI);
3081
3082 // Acc = Acc + A[i] * B[i]
3083 Register Sum =
3084 i < 3 ? MRI->createVirtualRegister(&SPIRV::IDRegClass) : ResVReg;
3085 BuildMI(BB, I, I.getDebugLoc(), TII.get(SPIRV::OpIAddS))
3086 .addDef(Sum)
3087 .addUse(GR.getSPIRVTypeID(ResType))
3088 .addUse(Acc)
3089 .addUse(MaskMul)
3090 .constrainAllUses(TII, TRI, RBI);
3091
3092 Acc = Sum;
3093 }
3094
3095 return true;
3096}
3097
3098/// Transform saturate(x) to clamp(x, 0.0f, 1.0f) as SPIRV
3099/// does not have a saturate builtin.
3100bool SPIRVInstructionSelector::selectSaturate(Register ResVReg,
3101 SPIRVTypeInst ResType,
3102 MachineInstr &I) const {
3103 assert(I.getNumOperands() == 3);
3104 assert(I.getOperand(2).isReg());
3105 MachineBasicBlock &BB = *I.getParent();
3106 Register VZero = buildZerosValF(ResType, I);
3107 Register VOne = buildOnesValF(ResType, I);
3108
3109 BuildMI(BB, I, I.getDebugLoc(), TII.get(SPIRV::OpExtInst))
3110 .addDef(ResVReg)
3111 .addUse(GR.getSPIRVTypeID(ResType))
3112 .addImm(static_cast<uint32_t>(SPIRV::InstructionSet::GLSL_std_450))
3113 .addImm(GL::FClamp)
3114 .addUse(I.getOperand(2).getReg())
3115 .addUse(VZero)
3116 .addUse(VOne)
3117 .constrainAllUses(TII, TRI, RBI);
3118 return true;
3119}
3120
3121bool SPIRVInstructionSelector::selectSign(Register ResVReg,
3122 SPIRVTypeInst ResType,
3123 MachineInstr &I) const {
3124 assert(I.getNumOperands() == 3);
3125 assert(I.getOperand(2).isReg());
3126 MachineBasicBlock &BB = *I.getParent();
3127 Register InputRegister = I.getOperand(2).getReg();
3128 SPIRVTypeInst InputType = GR.getSPIRVTypeForVReg(InputRegister);
3129 auto &DL = I.getDebugLoc();
3130
3131 if (!InputType)
3132 report_fatal_error("Input Type could not be determined.");
3133
3134 bool IsFloatTy = GR.isScalarOrVectorOfType(InputRegister, SPIRV::OpTypeFloat);
3135
3136 unsigned SignBitWidth = GR.getScalarOrVectorBitWidth(InputType);
3137 unsigned ResBitWidth = GR.getScalarOrVectorBitWidth(ResType);
3138
3139 bool NeedsConversion = IsFloatTy || SignBitWidth != ResBitWidth;
3140
3141 auto SignOpcode = IsFloatTy ? GL::FSign : GL::SSign;
3142 Register SignReg = NeedsConversion
3143 ? MRI->createVirtualRegister(&SPIRV::IDRegClass)
3144 : ResVReg;
3145
3146 BuildMI(BB, I, DL, TII.get(SPIRV::OpExtInst))
3147 .addDef(SignReg)
3148 .addUse(GR.getSPIRVTypeID(InputType))
3149 .addImm(static_cast<uint32_t>(SPIRV::InstructionSet::GLSL_std_450))
3150 .addImm(SignOpcode)
3151 .addUse(InputRegister)
3152 .constrainAllUses(TII, TRI, RBI);
3153
3154 if (NeedsConversion) {
3155 auto ConvertOpcode = IsFloatTy ? SPIRV::OpConvertFToS : SPIRV::OpSConvert;
3156 BuildMI(*I.getParent(), I, DL, TII.get(ConvertOpcode))
3157 .addDef(ResVReg)
3158 .addUse(GR.getSPIRVTypeID(ResType))
3159 .addUse(SignReg)
3160 .constrainAllUses(TII, TRI, RBI);
3161 }
3162
3163 return true;
3164}
3165
3166bool SPIRVInstructionSelector::selectWaveOpInst(Register ResVReg,
3167 SPIRVTypeInst ResType,
3168 MachineInstr &I,
3169 unsigned Opcode) const {
3170 MachineBasicBlock &BB = *I.getParent();
3171 SPIRVTypeInst IntTy = GR.getOrCreateSPIRVIntegerType(32, I, TII);
3172
3173 auto BMI = BuildMI(BB, I, I.getDebugLoc(), TII.get(Opcode))
3174 .addDef(ResVReg)
3175 .addUse(GR.getSPIRVTypeID(ResType))
3176 .addUse(GR.getOrCreateConstInt(SPIRV::Scope::Subgroup, I,
3177 IntTy, TII, !STI.isShader()));
3178
3179 for (unsigned J = 2; J < I.getNumOperands(); J++) {
3180 BMI.addUse(I.getOperand(J).getReg());
3181 }
3182
3183 BMI.constrainAllUses(TII, TRI, RBI);
3184 return true;
3185}
3186
3187bool SPIRVInstructionSelector::selectBarrierInst(MachineInstr &I,
3188 unsigned Scope,
3189 unsigned MemSem,
3190 bool WithGroupSync) const {
3191 auto BarrierType =
3192 WithGroupSync ? SPIRV::OpControlBarrier : SPIRV::OpMemoryBarrier;
3193
3194 MemSem |= SPIRV::MemorySemantics::AcquireRelease;
3195
3196 assert(((Scope != SPIRV::Scope::Workgroup) ||
3197 ((MemSem & SPIRV::MemorySemantics::WorkgroupMemory) > 0)) &&
3198 "Workgroup Scope must set WorkGroupMemory semantic "
3199 "in Barrier instruction");
3200
3201 assert(((Scope != SPIRV::Scope::Device) ||
3202 ((MemSem & SPIRV::MemorySemantics::UniformMemory) > 0 &&
3203 (MemSem & SPIRV::MemorySemantics::ImageMemory) > 0)) &&
3204 "Device Scope must set UniformMemory and ImageMemory semantic "
3205 "in Barrier instruction");
3206
3207 MachineBasicBlock &BB = *I.getParent();
3208 auto MI = BuildMI(BB, I, I.getDebugLoc(), TII.get(BarrierType));
3209
3210 // OpControlBarrier needs to also set Execution Scope
3211 if (WithGroupSync) {
3212 Register ExecReg = buildI32Constant(SPIRV::Scope::Workgroup, I);
3213 MI.addUse(ExecReg);
3214 }
3215
3216 Register ScopeReg = buildI32Constant(Scope, I);
3217 Register MemSemReg = buildI32Constant(MemSem, I);
3218
3219 MI.addUse(ScopeReg).addUse(MemSemReg).constrainAllUses(TII, TRI, RBI);
3220 return true;
3221}
3222
3223bool SPIRVInstructionSelector::selectWaveActiveCountBits(
3224 Register ResVReg, SPIRVTypeInst ResType, MachineInstr &I) const {
3225
3226 SPIRVTypeInst IntTy = GR.getOrCreateSPIRVIntegerType(32, I, TII);
3227 SPIRVTypeInst BallotType = GR.getOrCreateSPIRVVectorType(IntTy, 4, I, TII);
3228 Register BallotReg = MRI->createVirtualRegister(GR.getRegClass(BallotType));
3229 if (!selectWaveOpInst(BallotReg, BallotType, I,
3230 SPIRV::OpGroupNonUniformBallot))
3231 return false;
3232
3233 MachineBasicBlock &BB = *I.getParent();
3234 BuildMI(BB, I, I.getDebugLoc(),
3235 TII.get(SPIRV::OpGroupNonUniformBallotBitCount))
3236 .addDef(ResVReg)
3237 .addUse(GR.getSPIRVTypeID(ResType))
3238 .addUse(GR.getOrCreateConstInt(SPIRV::Scope::Subgroup, I, IntTy, TII,
3239 !STI.isShader()))
3240 .addImm(SPIRV::GroupOperation::Reduce)
3241 .addUse(BallotReg)
3242 .constrainAllUses(TII, TRI, RBI);
3243
3244 return true;
3245}
3246
3247bool SPIRVInstructionSelector::selectWaveActiveAllEqual(Register ResVReg,
3248 SPIRVTypeInst ResType,
3249 MachineInstr &I) const {
3250 MachineBasicBlock &BB = *I.getParent();
3251 const DebugLoc &DL = I.getDebugLoc();
3252
3253 // Input to the intrinsic
3254 Register InputReg = I.getOperand(2).getReg();
3255 SPIRVTypeInst InputType = GR.getSPIRVTypeForVReg(InputReg);
3256
3257 // Determine if input is vector
3258 unsigned NumElems = GR.getScalarOrVectorComponentCount(InputType);
3259 bool IsVector = NumElems > 1;
3260
3261 // Determine element types
3262 SPIRVTypeInst ElemInputType = GR.getScalarOrVectorComponentType(InputType);
3263 SPIRVTypeInst ElemBoolType = GR.getScalarOrVectorComponentType(ResType);
3264
3265 // Subgroup scope constant
3266 SPIRVTypeInst IntTy = GR.getOrCreateSPIRVIntegerType(32, I, TII);
3267 Register ScopeConst = GR.getOrCreateConstInt(SPIRV::Scope::Subgroup, I, IntTy,
3268 TII, !STI.isShader());
3269
3270 // Scalar case
3271 if (!IsVector) {
3272 return selectWaveOpInst(ResVReg, ElemBoolType, I,
3273 SPIRV::OpGroupNonUniformAllEqual);
3274 }
3275
3276 // Vector case
3277 SmallVector<Register, 4> ElementResults;
3278 ElementResults.reserve(NumElems);
3279
3280 for (unsigned Idx = 0; Idx < NumElems; ++Idx) {
3281 // Extract element
3282 Register ElemInput = InputReg;
3283 Register Extracted =
3284 MRI->createVirtualRegister(GR.getRegClass(ElemInputType));
3285
3286 BuildMI(BB, I, DL, TII.get(SPIRV::OpCompositeExtract))
3287 .addDef(Extracted)
3288 .addUse(GR.getSPIRVTypeID(ElemInputType))
3289 .addUse(InputReg)
3290 .addImm(Idx)
3291 .constrainAllUses(TII, TRI, RBI);
3292
3293 ElemInput = Extracted;
3294
3295 // Emit per-element AllEqual
3296 Register ElemResult =
3297 MRI->createVirtualRegister(GR.getRegClass(ElemBoolType));
3298
3299 BuildMI(BB, I, DL, TII.get(SPIRV::OpGroupNonUniformAllEqual))
3300 .addDef(ElemResult)
3301 .addUse(GR.getSPIRVTypeID(ElemBoolType))
3302 .addUse(ScopeConst)
3303 .addUse(ElemInput)
3304 .constrainAllUses(TII, TRI, RBI);
3305
3306 ElementResults.push_back(ElemResult);
3307 }
3308
3309 // Reconstruct vector<bool>
3310 auto MIB = BuildMI(BB, I, DL, TII.get(SPIRV::OpCompositeConstruct))
3311 .addDef(ResVReg)
3312 .addUse(GR.getSPIRVTypeID(ResType));
3313 for (Register R : ElementResults)
3314 MIB.addUse(R);
3315
3316 MIB.constrainAllUses(TII, TRI, RBI);
3317
3318 return true;
3319}
3320
3321bool SPIRVInstructionSelector::selectWavePrefixBitCount(Register ResVReg,
3322 SPIRVTypeInst ResType,
3323 MachineInstr &I) const {
3324
3325 assert(I.getNumOperands() == 3);
3326
3327 auto Op = I.getOperand(2);
3328 assert(Op.isReg());
3329
3330 MachineBasicBlock &BB = *I.getParent();
3331 DebugLoc DL = I.getDebugLoc();
3332
3333 Register InputRegister = Op.getReg();
3334 SPIRVTypeInst InputType = GR.getSPIRVTypeForVReg(InputRegister);
3335
3336 if (!InputType)
3337 report_fatal_error("Input Type could not be determined.");
3338
3339 if (InputType->getOpcode() != SPIRV::OpTypeBool)
3340 report_fatal_error("WavePrefixBitCount requires boolean input");
3341
3342 // Types
3343 SPIRVTypeInst Int32Ty = GR.getOrCreateSPIRVIntegerType(32, I, TII);
3344
3345 // Ballot result type: vector<uint32>
3346 // Match DXC: %v4uint for Subgroup size
3347 SPIRVTypeInst BallotTy = GR.getOrCreateSPIRVVectorType(Int32Ty, 4, I, TII);
3348
3349 // Create a vreg for the ballot result
3350 Register BallotVReg = MRI->createVirtualRegister(&SPIRV::IDRegClass);
3351
3352 // 1. OpGroupNonUniformBallot
3353 BuildMI(BB, I, DL, TII.get(SPIRV::OpGroupNonUniformBallot))
3354 .addDef(BallotVReg)
3355 .addUse(GR.getSPIRVTypeID(BallotTy))
3356 .addUse(GR.getOrCreateConstInt(SPIRV::Scope::Subgroup, I, Int32Ty, TII))
3357 .addUse(InputRegister)
3358 .constrainAllUses(TII, TRI, RBI);
3359
3360 // 2. OpGroupNonUniformBallotBitCount
3361 BuildMI(BB, I, DL, TII.get(SPIRV::OpGroupNonUniformBallotBitCount))
3362 .addDef(ResVReg)
3363 .addUse(GR.getSPIRVTypeID(ResType))
3364 .addUse(GR.getOrCreateConstInt(SPIRV::Scope::Subgroup, I, Int32Ty, TII))
3365 .addImm(SPIRV::GroupOperation::ExclusiveScan)
3366 .addUse(BallotVReg)
3367 .constrainAllUses(TII, TRI, RBI);
3368
3369 return true;
3370}
3371
3372bool SPIRVInstructionSelector::selectWaveReduceMax(Register ResVReg,
3373 SPIRVTypeInst ResType,
3374 MachineInstr &I,
3375 bool IsUnsigned) const {
3376 return selectWaveReduce(
3377 ResVReg, ResType, I, IsUnsigned,
3378 [&](Register InputRegister, bool IsUnsigned) {
3379 const bool IsFloatTy =
3380 GR.isScalarOrVectorOfType(InputRegister, SPIRV::OpTypeFloat);
3381 const auto IntOp = IsUnsigned ? SPIRV::OpGroupNonUniformUMax
3382 : SPIRV::OpGroupNonUniformSMax;
3383 return IsFloatTy ? SPIRV::OpGroupNonUniformFMax : IntOp;
3384 });
3385}
3386
3387bool SPIRVInstructionSelector::selectWaveReduceMin(Register ResVReg,
3388 SPIRVTypeInst ResType,
3389 MachineInstr &I,
3390 bool IsUnsigned) const {
3391 return selectWaveReduce(
3392 ResVReg, ResType, I, IsUnsigned,
3393 [&](Register InputRegister, bool IsUnsigned) {
3394 const bool IsFloatTy =
3395 GR.isScalarOrVectorOfType(InputRegister, SPIRV::OpTypeFloat);
3396 const auto IntOp = IsUnsigned ? SPIRV::OpGroupNonUniformUMin
3397 : SPIRV::OpGroupNonUniformSMin;
3398 return IsFloatTy ? SPIRV::OpGroupNonUniformFMin : IntOp;
3399 });
3400}
3401
3402bool SPIRVInstructionSelector::selectWaveReduceSum(Register ResVReg,
3403 SPIRVTypeInst ResType,
3404 MachineInstr &I) const {
3405 return selectWaveReduce(ResVReg, ResType, I, /*IsUnsigned*/ false,
3406 [&](Register InputRegister, bool IsUnsigned) {
3407 bool IsFloatTy = GR.isScalarOrVectorOfType(
3408 InputRegister, SPIRV::OpTypeFloat);
3409 return IsFloatTy ? SPIRV::OpGroupNonUniformFAdd
3410 : SPIRV::OpGroupNonUniformIAdd;
3411 });
3412}
3413
3414bool SPIRVInstructionSelector::selectWaveReduceProduct(Register ResVReg,
3415 SPIRVTypeInst ResType,
3416 MachineInstr &I) const {
3417 return selectWaveReduce(ResVReg, ResType, I, /*IsUnsigned*/ false,
3418 [&](Register InputRegister, bool IsUnsigned) {
3419 bool IsFloatTy = GR.isScalarOrVectorOfType(
3420 InputRegister, SPIRV::OpTypeFloat);
3421 return IsFloatTy ? SPIRV::OpGroupNonUniformFMul
3422 : SPIRV::OpGroupNonUniformIMul;
3423 });
3424}
3425
3426template <typename PickOpcodeFn>
3427bool SPIRVInstructionSelector::selectWaveReduce(
3428 Register ResVReg, SPIRVTypeInst ResType, MachineInstr &I, bool IsUnsigned,
3429 PickOpcodeFn &&PickOpcode) const {
3430 assert(I.getNumOperands() == 3);
3431 assert(I.getOperand(2).isReg());
3432 MachineBasicBlock &BB = *I.getParent();
3433 Register InputRegister = I.getOperand(2).getReg();
3434 SPIRVTypeInst InputType = GR.getSPIRVTypeForVReg(InputRegister);
3435
3436 if (!InputType)
3437 report_fatal_error("Input Type could not be determined.");
3438
3439 SPIRVTypeInst IntTy = GR.getOrCreateSPIRVIntegerType(32, I, TII);
3440 const unsigned Opcode = PickOpcode(InputRegister, IsUnsigned);
3441 BuildMI(BB, I, I.getDebugLoc(), TII.get(Opcode))
3442 .addDef(ResVReg)
3443 .addUse(GR.getSPIRVTypeID(ResType))
3444 .addUse(GR.getOrCreateConstInt(SPIRV::Scope::Subgroup, I, IntTy, TII,
3445 !STI.isShader()))
3446 .addImm(SPIRV::GroupOperation::Reduce)
3447 .addUse(I.getOperand(2).getReg())
3448 .constrainAllUses(TII, TRI, RBI);
3449 return true;
3450}
3451
3452bool SPIRVInstructionSelector::selectWaveReduceOp(Register ResVReg,
3453 SPIRVTypeInst ResType,
3454 MachineInstr &I,
3455 unsigned Opcode) const {
3456 return selectWaveReduce(
3457 ResVReg, ResType, I, false,
3458 [&](Register InputRegister, bool IsUnsigned) { return Opcode; });
3459}
3460
3461bool SPIRVInstructionSelector::selectWaveExclusiveScanSum(
3462 Register ResVReg, SPIRVTypeInst ResType, MachineInstr &I) const {
3463 return selectWaveExclusiveScan(ResVReg, ResType, I, /*IsUnsigned*/ false,
3464 [&](Register InputRegister, bool IsUnsigned) {
3465 bool IsFloatTy = GR.isScalarOrVectorOfType(
3466 InputRegister, SPIRV::OpTypeFloat);
3467 return IsFloatTy
3468 ? SPIRV::OpGroupNonUniformFAdd
3469 : SPIRV::OpGroupNonUniformIAdd;
3470 });
3471}
3472
3473bool SPIRVInstructionSelector::selectWaveExclusiveScanProduct(
3474 Register ResVReg, SPIRVTypeInst ResType, MachineInstr &I) const {
3475 return selectWaveExclusiveScan(ResVReg, ResType, I, /*IsUnsigned*/ false,
3476 [&](Register InputRegister, bool IsUnsigned) {
3477 bool IsFloatTy = GR.isScalarOrVectorOfType(
3478 InputRegister, SPIRV::OpTypeFloat);
3479 return IsFloatTy
3480 ? SPIRV::OpGroupNonUniformFMul
3481 : SPIRV::OpGroupNonUniformIMul;
3482 });
3483}
3484
3485template <typename PickOpcodeFn>
3486bool SPIRVInstructionSelector::selectWaveExclusiveScan(
3487 Register ResVReg, SPIRVTypeInst ResType, MachineInstr &I, bool IsUnsigned,
3488 PickOpcodeFn &&PickOpcode) const {
3489 assert(I.getNumOperands() == 3);
3490 assert(I.getOperand(2).isReg());
3491 MachineBasicBlock &BB = *I.getParent();
3492 Register InputRegister = I.getOperand(2).getReg();
3493 SPIRVTypeInst InputType = GR.getSPIRVTypeForVReg(InputRegister);
3494
3495 if (!InputType)
3496 report_fatal_error("Input Type could not be determined.");
3497
3498 SPIRVTypeInst IntTy = GR.getOrCreateSPIRVIntegerType(32, I, TII);
3499 const unsigned Opcode = PickOpcode(InputRegister, IsUnsigned);
3500 BuildMI(BB, I, I.getDebugLoc(), TII.get(Opcode))
3501 .addDef(ResVReg)
3502 .addUse(GR.getSPIRVTypeID(ResType))
3503 .addUse(GR.getOrCreateConstInt(SPIRV::Scope::Subgroup, I, IntTy, TII,
3504 !STI.isShader()))
3505 .addImm(SPIRV::GroupOperation::ExclusiveScan)
3506 .addUse(I.getOperand(2).getReg())
3507 .constrainAllUses(TII, TRI, RBI);
3508 return true;
3509}
3510
3511bool SPIRVInstructionSelector::selectQuadSwap(Register ResVReg,
3512 SPIRVTypeInst ResType,
3513 MachineInstr &I,
3514 unsigned Direction) const {
3515 assert(I.getNumOperands() == 3);
3516 assert(I.getOperand(2).isReg());
3517 MachineBasicBlock &BB = *I.getParent();
3518 Register InputRegister = I.getOperand(2).getReg();
3519
3520 SPIRVTypeInst IntTy = GR.getOrCreateSPIRVIntegerType(32, I, TII);
3521 bool ZeroAsNull = !STI.isShader();
3522 Register DirectionReg =
3523 GR.getOrCreateConstInt(Direction, I, IntTy, TII, ZeroAsNull);
3524 BuildMI(BB, I, I.getDebugLoc(), TII.get(SPIRV::OpGroupNonUniformQuadSwap))
3525 .addDef(ResVReg)
3526 .addUse(GR.getSPIRVTypeID(ResType))
3527 .addUse(GR.getOrCreateConstInt(SPIRV::Scope::Subgroup, I, IntTy, TII,
3528 ZeroAsNull))
3529 .addUse(InputRegister)
3530 .addUse(DirectionReg)
3531 .constrainAllUses(TII, TRI, RBI);
3532 return true;
3533}
3534
3535bool SPIRVInstructionSelector::selectBitreverse16(Register ResVReg,
3536 SPIRVTypeInst ResType,
3537 MachineInstr &I,
3538 Register Op) const {
3539 SPIRVTypeInst Int32Type = GR.getOrCreateSPIRVIntegerType(32, I, TII);
3540 Register ShiftConst = GR.getOrCreateConstInt(16, I, Int32Type, TII);
3541 unsigned ShiftOp = SPIRV::OpShiftRightLogicalS;
3542
3543 const unsigned N = GR.getScalarOrVectorComponentCount(ResType);
3544 const unsigned ExtendOpcode = GR.isScalarOrVectorSigned(ResType)
3545 ? SPIRV::OpSConvert
3546 : SPIRV::OpUConvert;
3547
3548 if (N > 1) {
3549 Int32Type = GR.getOrCreateSPIRVVectorType(Int32Type, N, I, TII);
3550 ShiftOp = SPIRV::OpShiftRightLogicalV;
3551
3552 // Vector shifts require a composite constant
3553 const Register CompositeReg =
3554 MRI->createVirtualRegister(GR.getRegClass(Int32Type));
3555 auto MIB = BuildMI(*I.getParent(), I, I.getDebugLoc(),
3556 TII.get(SPIRV::OpConstantComposite))
3557 .addDef(CompositeReg)
3558 .addUse(GR.getSPIRVTypeID(Int32Type));
3559 for (unsigned It = 0; It < N; ++It)
3560 MIB.addUse(ShiftConst);
3561 MIB.constrainAllUses(TII, TRI, RBI);
3562
3563 ShiftConst = CompositeReg;
3564 }
3565
3566 // Converts the i16 input to i32 (or vector of i32)
3567 Register ExtReg = MRI->createVirtualRegister(GR.getRegClass(Int32Type));
3568 if (!selectOpWithSrcs(ExtReg, Int32Type, I, {Op}, ExtendOpcode))
3569 return false;
3570
3571 // Perform bitreverse on the i32 value
3572 Register BitrevReg = MRI->createVirtualRegister(GR.getRegClass(Int32Type));
3573 if (!selectBitreverseNative(BitrevReg, Int32Type, I, ExtReg))
3574 return false;
3575
3576 // Shift the bit-reversed value to get the final result.
3577 Register ShiftReg = MRI->createVirtualRegister(GR.getRegClass(Int32Type));
3578 if (!selectOpWithSrcs(ShiftReg, Int32Type, I, {BitrevReg, ShiftConst},
3579 ShiftOp))
3580 return false;
3581
3582 // Finally, convert the result back to i16 (or vector of i16).
3583 return selectOpWithSrcs(ResVReg, ResType, I, {ShiftReg}, ExtendOpcode);
3584}
3585
3586bool SPIRVInstructionSelector::handle64BitOverflow(
3587 Register ResVReg, SPIRVTypeInst ResType, MachineInstr &I, Register SrcReg,
3588 unsigned int Opcode,
3589 std::function<bool(Register, SPIRVTypeInst, MachineInstr &, Register,
3590 unsigned)>
3591 CallbackFunction) const {
3592
3593 SPIRVTypeInst BaseType = GR.retrieveScalarOrVectorIntType(ResType);
3594 assert(BaseType->getOpcode() == SPIRV::OpTypeInt &&
3595 "handle64BitOverflow should only be used for integer types");
3596 unsigned ComponentCount = GR.getScalarOrVectorComponentCount(ResType);
3597 assert(ComponentCount < 5 && "Vec 5+ will generate invalid SPIR-V ops");
3598
3599 MachineIRBuilder MIRBuilder(I);
3600 SPIRVTypeInst I64Type = GR.getOrCreateSPIRVIntegerType(64, MIRBuilder);
3601 SPIRVTypeInst I64x2Type =
3602 GR.getOrCreateSPIRVVectorType(I64Type, 2, MIRBuilder, false);
3603 SPIRVTypeInst Vec2ResType =
3604 GR.getOrCreateSPIRVVectorType(BaseType, 2, MIRBuilder, false);
3605
3606 std::vector<Register> PartialRegs;
3607
3608 unsigned CurrentComponent = 0;
3609 for (; CurrentComponent + 1 < ComponentCount; CurrentComponent += 2) {
3610 Register PopCountResult =
3611 MRI->createVirtualRegister(GR.getRegClass(I64x2Type));
3612
3613 auto MIB = BuildMI(*I.getParent(), I, I.getDebugLoc(),
3614 TII.get(SPIRV::OpVectorShuffle))
3615 .addDef(PopCountResult)
3616 .addUse(GR.getSPIRVTypeID(I64x2Type))
3617 .addUse(SrcReg)
3618 .addUse(SrcReg)
3619 .addImm(CurrentComponent)
3620 .addImm(CurrentComponent + 1);
3621
3622 MIB.constrainAllUses(TII, TRI, RBI);
3623
3624 Register SubVecReg =
3625 MRI->createVirtualRegister(GR.getRegClass(Vec2ResType));
3626
3627 if (!CallbackFunction(SubVecReg, Vec2ResType, I, PopCountResult, Opcode))
3628 return false;
3629
3630 PartialRegs.push_back(SubVecReg);
3631 }
3632 // On odd component counts we need to handle one more component
3633 if (CurrentComponent != ComponentCount) {
3634 bool ZeroAsNull = !STI.isShader();
3635 Register FinalElemReg = MRI->createVirtualRegister(GR.getRegClass(I64Type));
3636 Register ConstIntLastIdx = GR.getOrCreateConstInt(
3637 ComponentCount - 1, I, BaseType, TII, ZeroAsNull);
3638
3639 if (!selectOpWithSrcs(FinalElemReg, I64Type, I, {SrcReg, ConstIntLastIdx},
3640 SPIRV::OpVectorExtractDynamic))
3641 return false;
3642
3643 Register FinalElemResReg =
3644 MRI->createVirtualRegister(GR.getRegClass(BaseType));
3645
3646 if (!CallbackFunction(FinalElemResReg, BaseType, I, FinalElemReg, Opcode))
3647 return false;
3648
3649 PartialRegs.push_back(FinalElemResReg);
3650 }
3651 // Join all the resulting registers back into the return type in order
3652 // (ie i32x2, i32x2, i32x1 -> i32x5)
3653 return selectOpWithSrcs(ResVReg, ResType, I, std::move(PartialRegs),
3654 SPIRV::OpCompositeConstruct);
3655}
3656
3657bool SPIRVInstructionSelector::selectBitreverse64(Register ResVReg,
3658 SPIRVTypeInst ResType,
3659 MachineInstr &I,
3660 Register SrcReg) const {
3661 unsigned ComponentCount = GR.getScalarOrVectorComponentCount(ResType);
3662 if (ComponentCount > 2)
3663 return handle64BitOverflow(
3664 ResVReg, ResType, I, SrcReg, SPIRV::OpBitReverse,
3665 [this](Register R, SPIRVTypeInst T, MachineInstr &I, Register S,
3666 unsigned O) { return this->selectBitreverse64(R, T, I, S); });
3667
3668 MachineIRBuilder MIRBuilder(I);
3669
3670 SPIRVTypeInst I32Type = GR.getOrCreateSPIRVIntegerType(32, MIRBuilder);
3671 SPIRVTypeInst VecI32Type = GR.getOrCreateSPIRVVectorType(
3672 I32Type, 2 * ComponentCount, MIRBuilder, /*IsSigned=*/false);
3673
3674 // Converts 64 bit into and array of 32 bit, containing 2 elements.
3675 Register Vec32 = MRI->createVirtualRegister(GR.getRegClass(VecI32Type));
3676 if (!selectOpWithSrcs(Vec32, VecI32Type, I, {SrcReg}, SPIRV::OpBitcast))
3677 return false;
3678
3679 // Apply bitreverse on each 32 bit lane
3680 Register Reverse32 = MRI->createVirtualRegister(GR.getRegClass(VecI32Type));
3681 if (!selectBitreverseNative(Reverse32, VecI32Type, I, Vec32))
3682 return false;
3683
3684 // Splits result into highbit lane and lowbit lane
3685 auto MaybeParts = splitEvenOddLanes(Reverse32, ComponentCount, I, I32Type);
3686 if (!MaybeParts)
3687 return false;
3688 SplitParts &Parts = *MaybeParts;
3689
3690 // Reversing a 64-bit value = reverse each 32-bit half AND swap them,
3691 // so the old High word becomes lane 0 (low) and old Low becomes lane 1
3692 // (high).
3693 Register SwappedVec = MRI->createVirtualRegister(GR.getRegClass(VecI32Type));
3694 if (!selectOpWithSrcs(SwappedVec, VecI32Type, I, {Parts.High, Parts.Low},
3695 SPIRV::OpCompositeConstruct))
3696 return false;
3697
3698 // Groups 32 bit vector back to 64 bit scalar.
3699 return selectOpWithSrcs(ResVReg, ResType, I, {SwappedVec}, SPIRV::OpBitcast);
3700}
3701
3702bool SPIRVInstructionSelector::selectBitreverseNative(Register ResVReg,
3703 SPIRVTypeInst ResType,
3704 MachineInstr &I,
3705 Register Op) const {
3706 MachineBasicBlock &BB = *I.getParent();
3707 BuildMI(BB, I, I.getDebugLoc(), TII.get(SPIRV::OpBitReverse))
3708 .addDef(ResVReg)
3709 .addUse(GR.getSPIRVTypeID(ResType))
3710 .addUse(Op)
3711 .constrainAllUses(TII, TRI, RBI);
3712 return true;
3713}
3714
3715bool SPIRVInstructionSelector::selectBitreverse(Register ResVReg,
3716 SPIRVTypeInst ResType,
3717 MachineInstr &I) const {
3718 Register OpReg = I.getOperand(1).getReg();
3719
3720 // TODO: Fix shader behavior in case of VK_KHR_maintenance9 extension is set
3721 if (STI.isShader()) {
3722 SPIRVTypeInst OpType = GR.getSPIRVTypeForVReg(OpReg);
3723 switch (GR.getScalarOrVectorBitWidth(OpType)) {
3724 case 8:
3725 case 16:
3726 return selectBitreverse16(ResVReg, ResType, I, OpReg);
3727 case 32:
3728 return selectBitreverseNative(ResVReg, ResType, I, OpReg);
3729 case 64:
3730 return selectBitreverse64(ResVReg, ResType, I, OpReg);
3731 }
3732 return SPIRVInstructionSelector::diagnoseUnsupported(
3733 I, "G_BITREVERSE only support 16,32,64 bits.");
3734 }
3735
3736 if (STI.canUseExtension(SPIRV::Extension::SPV_KHR_bit_instructions))
3737 return selectBitreverseNative(ResVReg, ResType, I, OpReg);
3738
3739 // Expansion bitreverse using bit manipulation operations
3740 // Algo: https://graphics.stanford.edu/~seander/bithacks.html#ReverseParallel
3741 const unsigned BitWidth = GR.getScalarOrVectorBitWidth(ResType);
3742 // TODO: add support for any bit width and bitwidth more than 64.
3743 if (BitWidth > 64 || !isPowerOf2_32(BitWidth))
3744 return false;
3745
3746 const unsigned N = GR.getScalarOrVectorComponentCount(ResType);
3747
3748 unsigned AndOp = SPIRV::OpBitwiseAndS;
3749 unsigned OrOp = SPIRV::OpBitwiseOrS;
3750 unsigned ShlOp = SPIRV::OpShiftLeftLogicalS;
3751 unsigned ShrOp = SPIRV::OpShiftRightLogicalS;
3752 if (N > 1) {
3753 AndOp = SPIRV::OpBitwiseAndV;
3754 OrOp = SPIRV::OpBitwiseOrV;
3755 ShlOp = SPIRV::OpShiftLeftLogicalV;
3756 ShrOp = SPIRV::OpShiftRightLogicalV;
3757 }
3758
3759 // Helper, one swap per step: ((input>>shift)&mask)|((input&mask)<<shift),
3760 // RPN: input shift >> mask & input mask & shift << |
3761 auto SwapBits = [&](const Register Input, const uint64_t Mask,
3762 const unsigned Shift) -> Register {
3763 auto CreateConst = [&](const uint64_t Value) -> Register {
3764 if (N == 1)
3765 return GR.getOrCreateConstInt(
3766 Value, I, GR.retrieveScalarOrVectorIntType(ResType), TII);
3767 return GR.getOrCreateConstVector(Value, I, ResType, TII);
3768 };
3769
3770 Register MaskReg = CreateConst(Mask);
3771 Register ShiftReg = CreateConst(Shift);
3772 Register T1 = MRI->createVirtualRegister(GR.getRegClass(ResType));
3773 Register T2 = MRI->createVirtualRegister(GR.getRegClass(ResType));
3774 Register T3 = MRI->createVirtualRegister(GR.getRegClass(ResType));
3775 Register T4 = MRI->createVirtualRegister(GR.getRegClass(ResType));
3776 Register Result = MRI->createVirtualRegister(GR.getRegClass(ResType));
3777
3778 if (!selectOpWithSrcs(T1, ResType, I, {Input, ShiftReg}, ShrOp) ||
3779 !selectOpWithSrcs(T2, ResType, I, {T1, MaskReg}, AndOp) ||
3780 !selectOpWithSrcs(T3, ResType, I, {Input, MaskReg}, AndOp) ||
3781 !selectOpWithSrcs(T4, ResType, I, {T3, ShiftReg}, ShlOp) ||
3782 !selectOpWithSrcs(Result, ResType, I, {T2, T4}, OrOp))
3783 return Register();
3784
3785 return Result;
3786 };
3787
3788 unsigned Shift = BitWidth;
3789 Register Result = OpReg;
3790 uint64_t Mask = ~0ull;
3791 while ((Shift >>= 1) > 0) {
3792 Mask ^= (Mask << Shift);
3793 Result = SwapBits(Result, Mask, Shift);
3794 if (!Result.isValid())
3795 return false;
3796 }
3797
3798 return BuildCOPY(ResVReg, Result, I);
3799}
3800
3801bool SPIRVInstructionSelector::selectFreeze(Register ResVReg,
3802 SPIRVTypeInst ResType,
3803 MachineInstr &I) const {
3804 // There is no way to implement `freeze` correctly without support on SPIR-V
3805 // standard side, but we may at least address a simple (static) case when
3806 // undef/poison value presence is obvious. The main benefit of even
3807 // incomplete `freeze` support is preventing of translation from crashing due
3808 // to lack of support on legalization and instruction selection steps.
3809 if (!I.getOperand(0).isReg() || !I.getOperand(1).isReg())
3810 return false;
3811 Register OpReg = I.getOperand(1).getReg();
3812 if (MachineInstr *Def = MRI->getVRegDef(OpReg)) {
3813 if (Def->getOpcode() == TargetOpcode::COPY)
3814 Def = MRI->getVRegDef(Def->getOperand(1).getReg());
3815 Register Reg;
3816 switch (Def->getOpcode()) {
3817 case SPIRV::ASSIGN_TYPE:
3818 if (MachineInstr *AssignToDef =
3819 MRI->getVRegDef(Def->getOperand(1).getReg())) {
3820 if (AssignToDef->getOpcode() == TargetOpcode::G_IMPLICIT_DEF)
3821 Reg = Def->getOperand(2).getReg();
3822 }
3823 break;
3824 case SPIRV::OpUndef:
3825 Reg = Def->getOperand(1).getReg();
3826 break;
3827 }
3828 unsigned DestOpCode;
3829 if (Reg.isValid()) {
3830 DestOpCode = SPIRV::OpConstantNull;
3831 } else {
3832 DestOpCode = TargetOpcode::COPY;
3833 Reg = OpReg;
3834 }
3835 BuildMI(*I.getParent(), I, I.getDebugLoc(), TII.get(DestOpCode))
3836 .addDef(I.getOperand(0).getReg())
3837 .addUse(Reg)
3838 .constrainAllUses(TII, TRI, RBI);
3839 return true;
3840 }
3841 return false;
3842}
3843
3844bool SPIRVInstructionSelector::selectBuildVector(Register ResVReg,
3845 SPIRVTypeInst ResType,
3846 MachineInstr &I) const {
3847 unsigned N = 0;
3848 if (ResType->getOpcode() == SPIRV::OpTypeVector)
3849 N = GR.getScalarOrVectorComponentCount(ResType);
3850 else if (ResType->getOpcode() == SPIRV::OpTypeArray)
3851 N = getArrayComponentCount(MRI, ResType);
3852 else
3853 report_fatal_error("Cannot select G_BUILD_VECTOR with a non-vector result");
3854 if (I.getNumExplicitOperands() - I.getNumExplicitDefs() != N)
3855 report_fatal_error("G_BUILD_VECTOR and the result type are inconsistent");
3856
3857 // check if we may construct a constant vector
3858 bool IsConst = true;
3859 for (unsigned i = I.getNumExplicitDefs();
3860 i < I.getNumExplicitOperands() && IsConst; ++i)
3861 if (!isConstReg(MRI, I.getOperand(i).getReg()))
3862 IsConst = false;
3863
3864 if (!IsConst && N < 2)
3865 report_fatal_error(
3866 "There must be at least two constituent operands in a vector");
3867
3868 MRI->setRegClass(ResVReg, GR.getRegClass(ResType));
3869 auto MIB = BuildMI(*I.getParent(), I, I.getDebugLoc(),
3870 TII.get(IsConst ? SPIRV::OpConstantComposite
3871 : SPIRV::OpCompositeConstruct))
3872 .addDef(ResVReg)
3873 .addUse(GR.getSPIRVTypeID(ResType));
3874 for (unsigned i = I.getNumExplicitDefs(); i < I.getNumExplicitOperands(); ++i)
3875 MIB.addUse(I.getOperand(i).getReg());
3876 MIB.constrainAllUses(TII, TRI, RBI);
3877 return true;
3878}
3879
3880bool SPIRVInstructionSelector::selectSplatVector(Register ResVReg,
3881 SPIRVTypeInst ResType,
3882 MachineInstr &I) const {
3883 unsigned N = 0;
3884 if (ResType->getOpcode() == SPIRV::OpTypeVector)
3885 N = GR.getScalarOrVectorComponentCount(ResType);
3886 else if (ResType->getOpcode() == SPIRV::OpTypeArray)
3887 N = getArrayComponentCount(MRI, ResType);
3888 else
3889 report_fatal_error("Cannot select G_SPLAT_VECTOR with a non-vector result");
3890
3891 unsigned OpIdx = I.getNumExplicitDefs();
3892 if (!I.getOperand(OpIdx).isReg())
3893 report_fatal_error("Unexpected argument in G_SPLAT_VECTOR");
3894
3895 // check if we may construct a constant vector
3896 Register OpReg = I.getOperand(OpIdx).getReg();
3897 bool IsConst = isConstReg(MRI, OpReg);
3898
3899 if (!IsConst && N < 2)
3900 report_fatal_error(
3901 "There must be at least two constituent operands in a vector");
3902
3903 MRI->setRegClass(ResVReg, GR.getRegClass(ResType));
3904 auto MIB = BuildMI(*I.getParent(), I, I.getDebugLoc(),
3905 TII.get(IsConst ? SPIRV::OpConstantComposite
3906 : SPIRV::OpCompositeConstruct))
3907 .addDef(ResVReg)
3908 .addUse(GR.getSPIRVTypeID(ResType));
3909 for (unsigned i = 0; i < N; ++i)
3910 MIB.addUse(OpReg);
3911 MIB.constrainAllUses(TII, TRI, RBI);
3912 return true;
3913}
3914
3915bool SPIRVInstructionSelector::selectDiscard(Register ResVReg,
3916 SPIRVTypeInst ResType,
3917 MachineInstr &I) const {
3918
3919 unsigned Opcode;
3920
3921 if (STI.canUseExtension(
3922 SPIRV::Extension::SPV_EXT_demote_to_helper_invocation) ||
3923 STI.isAtLeastSPIRVVer(llvm::VersionTuple(1, 6))) {
3924 Opcode = SPIRV::OpDemoteToHelperInvocation;
3925 } else {
3926 Opcode = SPIRV::OpKill;
3927 // OpKill must be the last operation of any basic block.
3928 if (MachineInstr *NextI = I.getNextNode()) {
3929 GR.invalidateMachineInstr(NextI);
3930 NextI->eraseFromParent();
3931 }
3932 }
3933
3934 MachineBasicBlock &BB = *I.getParent();
3935 BuildMI(BB, I, I.getDebugLoc(), TII.get(Opcode))
3936 .constrainAllUses(TII, TRI, RBI);
3937 return true;
3938}
3939
3940bool SPIRVInstructionSelector::selectCmp(Register ResVReg,
3941 SPIRVTypeInst ResType, unsigned CmpOpc,
3942 MachineInstr &I) const {
3943 Register Cmp0 = I.getOperand(2).getReg();
3944 Register Cmp1 = I.getOperand(3).getReg();
3945 assert(GR.getSPIRVTypeForVReg(Cmp0)->getOpcode() ==
3946 GR.getSPIRVTypeForVReg(Cmp1)->getOpcode() &&
3947 "CMP operands should have the same type");
3948 BuildMI(*I.getParent(), I, I.getDebugLoc(), TII.get(CmpOpc))
3949 .addDef(ResVReg)
3950 .addUse(GR.getSPIRVTypeID(ResType))
3951 .addUse(Cmp0)
3952 .addUse(Cmp1)
3953 .setMIFlags(I.getFlags())
3954 .constrainAllUses(TII, TRI, RBI);
3955 return true;
3956}
3957
3958bool SPIRVInstructionSelector::selectICmp(Register ResVReg,
3959 SPIRVTypeInst ResType,
3960 MachineInstr &I) const {
3961 auto Pred = I.getOperand(1).getPredicate();
3962 unsigned CmpOpc;
3963
3964 Register CmpOperand = I.getOperand(2).getReg();
3965 if (GR.isScalarOfType(CmpOperand, SPIRV::OpTypePointer))
3966 CmpOpc = getPtrCmpOpcode(Pred);
3967 else if (GR.isScalarOrVectorOfType(CmpOperand, SPIRV::OpTypeBool))
3968 CmpOpc = getBoolCmpOpcode(Pred);
3969 else
3970 CmpOpc = getICmpOpcode(Pred);
3971 return selectCmp(ResVReg, ResType, CmpOpc, I);
3972}
3973
3974Register
3975SPIRVInstructionSelector::buildI32Constant(uint32_t Val, MachineInstr &I,
3976 SPIRVTypeInst ResType) const {
3977 Type *LLVMTy = IntegerType::get(GR.CurMF->getFunction().getContext(), 32);
3978 SPIRVTypeInst SpvI32Ty =
3979 ResType ? ResType : GR.getOrCreateSPIRVIntegerType(32, I, TII);
3980 // Find a constant in DT or build a new one.
3981 auto ConstInt = ConstantInt::get(LLVMTy, Val);
3982 Register NewReg = GR.find(ConstInt, GR.CurMF);
3983 if (!NewReg.isValid()) {
3984 NewReg = MRI->createGenericVirtualRegister(LLT::scalar(64));
3985 MachineBasicBlock &BB = *I.getParent();
3986 MachineInstr *MI =
3987 Val == 0
3988 ? BuildMI(BB, I, I.getDebugLoc(), TII.get(SPIRV::OpConstantNull))
3989 .addDef(NewReg)
3990 .addUse(GR.getSPIRVTypeID(SpvI32Ty))
3991 : BuildMI(BB, I, I.getDebugLoc(), TII.get(SPIRV::OpConstantI))
3992 .addDef(NewReg)
3993 .addUse(GR.getSPIRVTypeID(SpvI32Ty))
3994 .addImm(APInt(32, Val).getZExtValue());
3995 constrainSelectedInstRegOperands(*MI, TII, TRI, RBI);
3996 GR.add(ConstInt, MI);
3997 }
3998 return NewReg;
3999}
4000
4001bool SPIRVInstructionSelector::selectFCmp(Register ResVReg,
4002 SPIRVTypeInst ResType,
4003 MachineInstr &I) const {
4004 unsigned CmpOp = getFCmpOpcode(I.getOperand(1).getPredicate());
4005 return selectCmp(ResVReg, ResType, CmpOp, I);
4006}
4007
4008bool SPIRVInstructionSelector::selectExp10(Register ResVReg,
4009 SPIRVTypeInst ResType,
4010 MachineInstr &I) const {
4011 if (STI.canUseExtInstSet(SPIRV::InstructionSet::OpenCL_std)) {
4012 return selectExtInst(ResVReg, ResType, I, CL::exp10);
4013 }
4014
4015 if (STI.canUseExtInstSet(SPIRV::InstructionSet::GLSL_std_450)) {
4016 /// There is no exp10 in GLSL. Use exp10(x) = exp2(x * log2(10)) instead
4017 /// log2(10) ~= 3.3219280948874l
4018
4019 if (ResType->getOpcode() != SPIRV::OpTypeVector &&
4020 ResType->getOpcode() != SPIRV::OpTypeFloat)
4021 return false;
4022
4023 MachineIRBuilder MIRBuilder(I);
4024
4025 SPIRVTypeInst SpirvScalarType = GR.getScalarOrVectorComponentType(ResType);
4026
4027 // Match the literal precision to the scalar type so the OpConstant
4028 // literal does not contain non-zero high-order bits that would fail
4029 // SPIR-V validation when the type is narrower than 32 bits (e.g. half).
4030 APFloat ConstVal(3.3219280948873623);
4031 bool LosesInfo;
4032 ConstVal.convert(
4033 getZeroFP(GR.getTypeForSPIRVType(SpirvScalarType)).getSemantics(),
4034 APFloat::rmNearestTiesToEven, &LosesInfo);
4035 Register ConstReg =
4036 GR.buildConstantFP(ConstVal, MIRBuilder, SpirvScalarType);
4037 Register ArgReg = MRI->createVirtualRegister(GR.getRegClass(ResType));
4038 auto Opcode = ResType->getOpcode() == SPIRV::OpTypeVector
4039 ? SPIRV::OpVectorTimesScalar
4040 : SPIRV::OpFMulS;
4041
4042 if (!selectOpWithSrcs(ArgReg, ResType, I,
4043 {I.getOperand(1).getReg(), ConstReg}, Opcode))
4044 return false;
4045 if (!selectExtInst(ResVReg, ResType, I,
4046 {{SPIRV::InstructionSet::GLSL_std_450, GL::Exp2}}, false,
4047 false, {ArgReg}))
4048 return false;
4049
4050 return true;
4051 }
4052
4053 return false;
4054}
4055
4056Register SPIRVInstructionSelector::buildZerosVal(SPIRVTypeInst ResType,
4057 MachineInstr &I) const {
4058 // OpenCL uses nulls for Zero. In HLSL we don't use null constants.
4059 bool ZeroAsNull = !STI.isShader();
4060 if (ResType->getOpcode() == SPIRV::OpTypeVector)
4061 return GR.getOrCreateConstVector(0UL, I, ResType, TII, ZeroAsNull);
4062 return GR.getOrCreateConstInt(0, I, ResType, TII, ZeroAsNull);
4063}
4064
4065bool SPIRVInstructionSelector::isScalarOrVectorIntConstantZero(
4066 Register Reg) const {
4067 SPIRVTypeInst Type = GR.getSPIRVTypeForVReg(Reg);
4068 if (!Type)
4069 return false;
4070 SPIRVTypeInst CompType = GR.getScalarOrVectorComponentType(Type);
4071 if (!CompType || CompType->getOpcode() != SPIRV::OpTypeInt)
4072 return false;
4073
4074 auto IsZero = [this](Register Reg) {
4075 MachineInstr *Def = getDefInstrMaybeConstant(Reg, MRI);
4076 if (!Def)
4077 return false;
4078
4079 if (Def->getOpcode() == SPIRV::OpConstantNull)
4080 return true;
4081
4082 if (Def->getOpcode() == TargetOpcode::G_CONSTANT ||
4083 Def->getOpcode() == SPIRV::OpConstantI)
4084 return getIConstVal(Reg, MRI) == 0;
4085
4086 return false;
4087 };
4088
4089 if (IsZero(Reg))
4090 return true;
4091
4092 MachineInstr *Def = MRI->getVRegDef(Reg);
4093 if (!Def)
4094 return false;
4095
4096 if (Def->getOpcode() == TargetOpcode::G_BUILD_VECTOR ||
4097 (Def->getOpcode() == TargetOpcode::G_INTRINSIC_W_SIDE_EFFECTS &&
4098 cast<GIntrinsic>(Def)->getIntrinsicID() ==
4099 Intrinsic::spv_const_composite)) {
4100 unsigned StartOp = Def->getOpcode() == TargetOpcode::G_BUILD_VECTOR ? 1 : 2;
4101 for (unsigned i = StartOp; i < Def->getNumOperands(); ++i) {
4102 if (!IsZero(Def->getOperand(i).getReg()))
4103 return false;
4104 }
4105 return true;
4106 }
4107
4108 return false;
4109}
4110
4111Register SPIRVInstructionSelector::buildZerosValF(SPIRVTypeInst ResType,
4112 MachineInstr &I) const {
4113 // OpenCL uses nulls for Zero. In HLSL we don't use null constants.
4114 bool ZeroAsNull = !STI.isShader();
4115 APFloat VZero = getZeroFP(GR.getTypeForSPIRVType(ResType));
4116 if (ResType->getOpcode() == SPIRV::OpTypeVector)
4117 return GR.getOrCreateConstVector(VZero, I, ResType, TII, ZeroAsNull);
4118 return GR.getOrCreateConstFP(VZero, I, ResType, TII, ZeroAsNull);
4119}
4120
4121Register SPIRVInstructionSelector::buildOnesValF(SPIRVTypeInst ResType,
4122 MachineInstr &I) const {
4123 // OpenCL uses nulls for Zero. In HLSL we don't use null constants.
4124 bool ZeroAsNull = !STI.isShader();
4125 APFloat VOne = getOneFP(GR.getTypeForSPIRVType(ResType));
4126 if (ResType->getOpcode() == SPIRV::OpTypeVector)
4127 return GR.getOrCreateConstVector(VOne, I, ResType, TII, ZeroAsNull);
4128 return GR.getOrCreateConstFP(VOne, I, ResType, TII, ZeroAsNull);
4129}
4130
4131Register SPIRVInstructionSelector::buildOnesVal(bool AllOnes,
4132 SPIRVTypeInst ResType,
4133 MachineInstr &I) const {
4134 unsigned BitWidth = GR.getScalarOrVectorBitWidth(ResType);
4135 APInt One =
4136 AllOnes ? APInt::getAllOnes(BitWidth) : APInt::getOneBitSet(BitWidth, 0);
4137 if (ResType->getOpcode() == SPIRV::OpTypeVector)
4138 return GR.getOrCreateConstVector(One, I, ResType, TII);
4139 return GR.getOrCreateConstInt(One, I, ResType, TII);
4140}
4141
4142bool SPIRVInstructionSelector::selectSelect(Register ResVReg,
4143 SPIRVTypeInst ResType,
4144 MachineInstr &I) const {
4145 Register SelectFirstArg = I.getOperand(2).getReg();
4146 Register SelectSecondArg = I.getOperand(3).getReg();
4147 assert(ResType == GR.getSPIRVTypeForVReg(SelectFirstArg) &&
4148 ResType == GR.getSPIRVTypeForVReg(SelectSecondArg));
4149
4150 bool IsFloatTy =
4151 GR.isScalarOrVectorOfType(SelectFirstArg, SPIRV::OpTypeFloat);
4152 bool IsPtrTy =
4153 GR.isScalarOrVectorOfType(SelectFirstArg, SPIRV::OpTypePointer);
4154 bool IsVectorTy = GR.getSPIRVTypeForVReg(SelectFirstArg)->getOpcode() ==
4155 SPIRV::OpTypeVector;
4156
4157 bool IsScalarBool =
4158 GR.isScalarOfType(I.getOperand(1).getReg(), SPIRV::OpTypeBool);
4159 unsigned Opcode;
4160 if (IsVectorTy) {
4161 if (IsFloatTy) {
4162 Opcode = IsScalarBool ? SPIRV::OpSelectVFSCond : SPIRV::OpSelectVFVCond;
4163 } else if (IsPtrTy) {
4164 Opcode = IsScalarBool ? SPIRV::OpSelectVPSCond : SPIRV::OpSelectVPVCond;
4165 } else {
4166 Opcode = IsScalarBool ? SPIRV::OpSelectVISCond : SPIRV::OpSelectVIVCond;
4167 }
4168 } else {
4169 if (IsFloatTy) {
4170 Opcode = IsScalarBool ? SPIRV::OpSelectSFSCond : SPIRV::OpSelectVFVCond;
4171 } else if (IsPtrTy) {
4172 Opcode = IsScalarBool ? SPIRV::OpSelectSPSCond : SPIRV::OpSelectVPVCond;
4173 } else {
4174 Opcode = IsScalarBool ? SPIRV::OpSelectSISCond : SPIRV::OpSelectVIVCond;
4175 }
4176 }
4177 BuildMI(*I.getParent(), I, I.getDebugLoc(), TII.get(Opcode))
4178 .addDef(ResVReg)
4179 .addUse(GR.getSPIRVTypeID(ResType))
4180 .addUse(I.getOperand(1).getReg())
4181 .addUse(SelectFirstArg)
4182 .addUse(SelectSecondArg)
4183 .constrainAllUses(TII, TRI, RBI);
4184 return true;
4185}
4186
4187// This function is used to extend a bool or a vector of bools into an integer
4188// or vector of integers.
4189bool SPIRVInstructionSelector::selectBoolToInt(Register ResVReg,
4190 SPIRVTypeInst ResType,
4191 Register BooleanVReg,
4192 MachineInstr &InsertAt,
4193 bool IsSigned) const {
4194 // To extend a bool, we need to use OpSelect between constants.
4195 Register ZeroReg = buildZerosVal(ResType, InsertAt);
4196 Register OneReg = buildOnesVal(IsSigned, ResType, InsertAt);
4197 bool IsScalarBool = GR.isScalarOfType(BooleanVReg, SPIRV::OpTypeBool);
4198 unsigned Opcode =
4199 IsScalarBool ? SPIRV::OpSelectSISCond : SPIRV::OpSelectVIVCond;
4200 BuildMI(*InsertAt.getParent(), InsertAt, InsertAt.getDebugLoc(),
4201 TII.get(Opcode))
4202 .addDef(ResVReg)
4203 .addUse(GR.getSPIRVTypeID(ResType))
4204 .addUse(BooleanVReg)
4205 .addUse(OneReg)
4206 .addUse(ZeroReg)
4207 .constrainAllUses(TII, TRI, RBI);
4208 return true;
4209}
4210
4211bool SPIRVInstructionSelector::selectIToF(Register ResVReg,
4212 SPIRVTypeInst ResType,
4213 MachineInstr &I, bool IsSigned,
4214 unsigned Opcode) const {
4215 Register SrcReg = I.getOperand(1).getReg();
4216 // We can convert bool value directly to float type without OpConvert*ToF,
4217 // however the translator generates OpSelect+OpConvert*ToF, so we do the same.
4218 if (GR.isScalarOrVectorOfType(I.getOperand(1).getReg(), SPIRV::OpTypeBool)) {
4219 unsigned BitWidth = GR.getScalarOrVectorBitWidth(ResType);
4220 SPIRVTypeInst TmpType = GR.getOrCreateSPIRVIntegerType(BitWidth, I, TII);
4221 if (ResType->getOpcode() == SPIRV::OpTypeVector) {
4222 const unsigned NumElts = GR.getScalarOrVectorComponentCount(ResType);
4223 TmpType = GR.getOrCreateSPIRVVectorType(TmpType, NumElts, I, TII);
4224 }
4225 SrcReg = createVirtualRegister(TmpType, &GR, MRI, MRI->getMF());
4226 selectBoolToInt(SrcReg, TmpType, I.getOperand(1).getReg(), I, false);
4227 }
4228 return selectOpWithSrcs(ResVReg, ResType, I, {SrcReg}, Opcode);
4229}
4230
4231bool SPIRVInstructionSelector::selectExt(Register ResVReg,
4232 SPIRVTypeInst ResType, MachineInstr &I,
4233 bool IsSigned) const {
4234 Register SrcReg = I.getOperand(1).getReg();
4235 if (GR.isScalarOrVectorOfType(SrcReg, SPIRV::OpTypeBool))
4236 return selectBoolToInt(ResVReg, ResType, I.getOperand(1).getReg(), I,
4237 IsSigned);
4238
4239 SPIRVTypeInst SrcType = GR.getSPIRVTypeForVReg(SrcReg);
4240 if (ResType == SrcType)
4241 return BuildCOPY(ResVReg, SrcReg, I);
4242
4243 unsigned Opcode = IsSigned ? SPIRV::OpSConvert : SPIRV::OpUConvert;
4244 return selectUnOp(ResVReg, ResType, I, Opcode);
4245}
4246
4247bool SPIRVInstructionSelector::selectSUCmp(Register ResVReg,
4248 SPIRVTypeInst ResType,
4249 MachineInstr &I,
4250 bool IsSigned) const {
4251 MachineIRBuilder MIRBuilder(I);
4252 MachineRegisterInfo *MRI = MIRBuilder.getMRI();
4253 MachineBasicBlock &BB = *I.getParent();
4254 // Ensure we have bool.
4255 SPIRVTypeInst BoolType = GR.getOrCreateSPIRVBoolType(I, TII);
4256 unsigned N = GR.getScalarOrVectorComponentCount(ResType);
4257 if (N > 1)
4258 BoolType = GR.getOrCreateSPIRVVectorType(BoolType, N, I, TII);
4259 Register BoolTypeReg = GR.getSPIRVTypeID(BoolType);
4260 // Build less-than-equal and less-than.
4261 // TODO: replace with one-liner createVirtualRegister() from
4262 // llvm/lib/Target/SPIRV/SPIRVUtils.cpp when PR #116609 is merged.
4263 Register IsLessEqReg = MRI->createVirtualRegister(GR.getRegClass(ResType));
4264 MRI->setType(IsLessEqReg, LLT::scalar(64));
4265 GR.assignSPIRVTypeToVReg(ResType, IsLessEqReg, MIRBuilder.getMF());
4266 BuildMI(BB, I, I.getDebugLoc(),
4267 TII.get(IsSigned ? SPIRV::OpSLessThanEqual : SPIRV::OpULessThanEqual))
4268 .addDef(IsLessEqReg)
4269 .addUse(BoolTypeReg)
4270 .addUse(I.getOperand(1).getReg())
4271 .addUse(I.getOperand(2).getReg())
4272 .constrainAllUses(TII, TRI, RBI);
4273 Register IsLessReg = MRI->createVirtualRegister(GR.getRegClass(ResType));
4274 MRI->setType(IsLessReg, LLT::scalar(64));
4275 GR.assignSPIRVTypeToVReg(ResType, IsLessReg, MIRBuilder.getMF());
4276 BuildMI(BB, I, I.getDebugLoc(),
4277 TII.get(IsSigned ? SPIRV::OpSLessThan : SPIRV::OpULessThan))
4278 .addDef(IsLessReg)
4279 .addUse(BoolTypeReg)
4280 .addUse(I.getOperand(1).getReg())
4281 .addUse(I.getOperand(2).getReg())
4282 .constrainAllUses(TII, TRI, RBI);
4283 // Build selects.
4284 Register ResTypeReg = GR.getSPIRVTypeID(ResType);
4285 Register NegOneOrZeroReg =
4286 MRI->createVirtualRegister(GR.getRegClass(ResType));
4287 MRI->setType(NegOneOrZeroReg, LLT::scalar(64));
4288 GR.assignSPIRVTypeToVReg(ResType, NegOneOrZeroReg, MIRBuilder.getMF());
4289 unsigned SelectOpcode =
4290 N > 1 ? SPIRV::OpSelectVIVCond : SPIRV::OpSelectSISCond;
4291 BuildMI(BB, I, I.getDebugLoc(), TII.get(SelectOpcode))
4292 .addDef(NegOneOrZeroReg)
4293 .addUse(ResTypeReg)
4294 .addUse(IsLessReg)
4295 .addUse(buildOnesVal(true, ResType, I)) // -1
4296 .addUse(buildZerosVal(ResType, I))
4297 .constrainAllUses(TII, TRI, RBI);
4298 BuildMI(BB, I, I.getDebugLoc(), TII.get(SelectOpcode))
4299 .addDef(ResVReg)
4300 .addUse(ResTypeReg)
4301 .addUse(IsLessEqReg)
4302 .addUse(NegOneOrZeroReg) // -1 or 0
4303 .addUse(buildOnesVal(false, ResType, I))
4304 .constrainAllUses(TII, TRI, RBI);
4305 return true;
4306}
4307
4308bool SPIRVInstructionSelector::selectIntToBool(Register IntReg,
4309 Register ResVReg,
4310 MachineInstr &I,
4311 SPIRVTypeInst IntTy,
4312 SPIRVTypeInst BoolTy) const {
4313 // To truncate to a bool, we use OpBitwiseAnd 1 and OpINotEqual to zero.
4314 Register BitIntReg = createVirtualRegister(IntTy, &GR, MRI, MRI->getMF());
4315 bool IsVectorTy = IntTy->getOpcode() == SPIRV::OpTypeVector;
4316 unsigned Opcode = IsVectorTy ? SPIRV::OpBitwiseAndV : SPIRV::OpBitwiseAndS;
4317 Register Zero = buildZerosVal(IntTy, I);
4318 Register One = buildOnesVal(false, IntTy, I);
4319 MachineBasicBlock &BB = *I.getParent();
4320 BuildMI(BB, I, I.getDebugLoc(), TII.get(Opcode))
4321 .addDef(BitIntReg)
4322 .addUse(GR.getSPIRVTypeID(IntTy))
4323 .addUse(IntReg)
4324 .addUse(One)
4325 .constrainAllUses(TII, TRI, RBI);
4326 BuildMI(BB, I, I.getDebugLoc(), TII.get(SPIRV::OpINotEqual))
4327 .addDef(ResVReg)
4328 .addUse(GR.getSPIRVTypeID(BoolTy))
4329 .addUse(BitIntReg)
4330 .addUse(Zero)
4331 .constrainAllUses(TII, TRI, RBI);
4332 return true;
4333}
4334
4335bool SPIRVInstructionSelector::selectTrunc(Register ResVReg,
4336 SPIRVTypeInst ResType,
4337 MachineInstr &I) const {
4338 Register IntReg = I.getOperand(1).getReg();
4339 const SPIRVTypeInst ArgType = GR.getSPIRVTypeForVReg(IntReg);
4340 if (GR.isScalarOrVectorOfType(ResVReg, SPIRV::OpTypeBool))
4341 return selectIntToBool(IntReg, ResVReg, I, ArgType, ResType);
4342 if (ArgType == ResType)
4343 return BuildCOPY(ResVReg, IntReg, I);
4344 bool IsSigned = GR.isScalarOrVectorSigned(ResType);
4345 unsigned Opcode = IsSigned ? SPIRV::OpSConvert : SPIRV::OpUConvert;
4346 return selectUnOp(ResVReg, ResType, I, Opcode);
4347}
4348
4349bool SPIRVInstructionSelector::selectConst(Register ResVReg,
4350 SPIRVTypeInst ResType,
4351 MachineInstr &I) const {
4352 unsigned Opcode = I.getOpcode();
4353 unsigned TpOpcode = ResType->getOpcode();
4354 Register Reg;
4355 if (TpOpcode == SPIRV::OpTypePointer || TpOpcode == SPIRV::OpTypeEvent) {
4356 assert(Opcode == TargetOpcode::G_CONSTANT &&
4357 I.getOperand(1).getCImm()->isZero());
4358 MachineBasicBlock &DepMBB = I.getMF()->front();
4359 MachineIRBuilder MIRBuilder(DepMBB, DepMBB.getFirstNonPHI());
4360 Reg = GR.getOrCreateConstNullPtr(MIRBuilder, ResType);
4361 } else if (Opcode == TargetOpcode::G_FCONSTANT) {
4362 Reg = GR.getOrCreateConstFP(I.getOperand(1).getFPImm()->getValue(), I,
4363 ResType, TII, !STI.isShader());
4364 } else {
4365 Reg = GR.getOrCreateConstInt(I.getOperand(1).getCImm()->getValue(), I,
4366 ResType, TII, !STI.isShader());
4367 }
4368 return Reg == ResVReg ? true : BuildCOPY(ResVReg, Reg, I);
4369}
4370
4371bool SPIRVInstructionSelector::selectOpUndef(Register ResVReg,
4372 SPIRVTypeInst ResType,
4373 MachineInstr &I) const {
4374 BuildMI(*I.getParent(), I, I.getDebugLoc(), TII.get(SPIRV::OpUndef))
4375 .addDef(ResVReg)
4376 .addUse(GR.getSPIRVTypeID(ResType))
4377 .constrainAllUses(TII, TRI, RBI);
4378 return true;
4379}
4380
4381bool SPIRVInstructionSelector::selectInsertVal(Register ResVReg,
4382 SPIRVTypeInst ResType,
4383 MachineInstr &I) const {
4384 MachineBasicBlock &BB = *I.getParent();
4385 auto MIB = BuildMI(BB, I, I.getDebugLoc(), TII.get(SPIRV::OpCompositeInsert))
4386 .addDef(ResVReg)
4387 .addUse(GR.getSPIRVTypeID(ResType))
4388 // object to insert
4389 .addUse(I.getOperand(3).getReg())
4390 // composite to insert into
4391 .addUse(I.getOperand(2).getReg());
4392 for (unsigned i = 4; i < I.getNumOperands(); i++)
4393 MIB.addImm(foldImm(I.getOperand(i), MRI));
4394 MIB.constrainAllUses(TII, TRI, RBI);
4395 return true;
4396}
4397
4398bool SPIRVInstructionSelector::selectExtractVal(Register ResVReg,
4399 SPIRVTypeInst ResType,
4400 MachineInstr &I) const {
4401 Type *MaybeResTy = nullptr;
4402 StringRef ResName;
4403 if (GR.findValueAttrs(&I, MaybeResTy, ResName) &&
4404 MaybeResTy != GR.getTypeForSPIRVType(ResType)) {
4405 assert((!MaybeResTy || MaybeResTy->isAggregateType()) &&
4406 "Expected aggregate type for extractv instruction");
4407 ResType = GR.getOrCreateSPIRVType(MaybeResTy, I,
4408 SPIRV::AccessQualifier::ReadWrite, false);
4409 GR.assignSPIRVTypeToVReg(ResType, ResVReg, *I.getMF());
4410 }
4411 MachineBasicBlock &BB = *I.getParent();
4412 auto MIB = BuildMI(BB, I, I.getDebugLoc(), TII.get(SPIRV::OpCompositeExtract))
4413 .addDef(ResVReg)
4414 .addUse(GR.getSPIRVTypeID(ResType))
4415 .addUse(I.getOperand(2).getReg());
4416 for (unsigned i = 3; i < I.getNumOperands(); i++)
4417 MIB.addImm(foldImm(I.getOperand(i), MRI));
4418 MIB.constrainAllUses(TII, TRI, RBI);
4419 return true;
4420}
4421
4422bool SPIRVInstructionSelector::selectInsertElt(Register ResVReg,
4423 SPIRVTypeInst ResType,
4424 MachineInstr &I) const {
4425 if (getImm(I.getOperand(4), MRI))
4426 return selectInsertVal(ResVReg, ResType, I);
4427 MachineBasicBlock &BB = *I.getParent();
4428 BuildMI(BB, I, I.getDebugLoc(), TII.get(SPIRV::OpVectorInsertDynamic))
4429 .addDef(ResVReg)
4430 .addUse(GR.getSPIRVTypeID(ResType))
4431 .addUse(I.getOperand(2).getReg())
4432 .addUse(I.getOperand(3).getReg())
4433 .addUse(I.getOperand(4).getReg())
4434 .constrainAllUses(TII, TRI, RBI);
4435 return true;
4436}
4437
4438bool SPIRVInstructionSelector::selectExtractElt(Register ResVReg,
4439 SPIRVTypeInst ResType,
4440 MachineInstr &I) const {
4441 if (getImm(I.getOperand(3), MRI))
4442 return selectExtractVal(ResVReg, ResType, I);
4443 MachineBasicBlock &BB = *I.getParent();
4444 BuildMI(BB, I, I.getDebugLoc(), TII.get(SPIRV::OpVectorExtractDynamic))
4445 .addDef(ResVReg)
4446 .addUse(GR.getSPIRVTypeID(ResType))
4447 .addUse(I.getOperand(2).getReg())
4448 .addUse(I.getOperand(3).getReg())
4449 .constrainAllUses(TII, TRI, RBI);
4450 return true;
4451}
4452
4453bool SPIRVInstructionSelector::selectGEP(Register ResVReg,
4454 SPIRVTypeInst ResType,
4455 MachineInstr &I) const {
4456 const bool IsGEPInBounds = I.getOperand(2).getImm();
4457
4458 // OpAccessChain could be used for OpenCL, but the SPIRV-LLVM Translator only
4459 // relies on PtrAccessChain, so we'll try not to deviate. For Vulkan however,
4460 // we have to use Op[InBounds]AccessChain.
4461 const unsigned Opcode = STI.isLogicalSPIRV()
4462 ? (IsGEPInBounds ? SPIRV::OpInBoundsAccessChain
4463 : SPIRV::OpAccessChain)
4464 : (IsGEPInBounds ? SPIRV::OpInBoundsPtrAccessChain
4465 : SPIRV::OpPtrAccessChain);
4466
4467 auto Res = BuildMI(*I.getParent(), I, I.getDebugLoc(), TII.get(Opcode))
4468 .addDef(ResVReg)
4469 .addUse(GR.getSPIRVTypeID(ResType))
4470 // Object to get a pointer to.
4471 .addUse(I.getOperand(3).getReg());
4472 assert(
4473 (Opcode == SPIRV::OpPtrAccessChain ||
4474 Opcode == SPIRV::OpInBoundsPtrAccessChain ||
4475 (getImm(I.getOperand(4), MRI) && foldImm(I.getOperand(4), MRI) == 0)) &&
4476 "Cannot translate GEP to OpAccessChain. First index must be 0.");
4477
4478 // Adding indices.
4479 const unsigned StartingIndex =
4480 (Opcode == SPIRV::OpAccessChain || Opcode == SPIRV::OpInBoundsAccessChain)
4481 ? 5
4482 : 4;
4483 for (unsigned i = StartingIndex; i < I.getNumExplicitOperands(); ++i)
4484 Res.addUse(I.getOperand(i).getReg());
4485 Res.constrainAllUses(TII, TRI, RBI);
4486 return true;
4487}
4488
4489// Maybe wrap a value into OpSpecConstantOp
4490bool SPIRVInstructionSelector::wrapIntoSpecConstantOp(
4491 MachineInstr &I, SmallVector<Register> &CompositeArgs) const {
4492 unsigned Lim = I.getNumExplicitOperands();
4493 for (unsigned i = I.getNumExplicitDefs() + 1; i < Lim; ++i) {
4494 Register OpReg = I.getOperand(i).getReg();
4495 MachineInstr *OpDefine = MRI->getVRegDef(OpReg);
4496 SPIRVTypeInst OpType = GR.getSPIRVTypeForVReg(OpReg);
4497 if (!OpDefine || !OpType || isConstReg(MRI, OpDefine) ||
4498 OpDefine->getOpcode() == TargetOpcode::G_ADDRSPACE_CAST ||
4499 OpDefine->getOpcode() == TargetOpcode::G_INTTOPTR ||
4500 GR.isAggregateType(OpType)) {
4501 // The case of G_ADDRSPACE_CAST inside spv_const_composite() is processed
4502 // by selectAddrSpaceCast(), and G_INTTOPTR is processed by selectUnOp()
4503 CompositeArgs.push_back(OpReg);
4504 continue;
4505 }
4506 MachineFunction *MF = I.getMF();
4507 Register WrapReg = GR.find(OpDefine, MF);
4508 if (WrapReg.isValid()) {
4509 CompositeArgs.push_back(WrapReg);
4510 continue;
4511 }
4512 // Create a new register for the wrapper
4513 WrapReg = MRI->createVirtualRegister(GR.getRegClass(OpType));
4514 CompositeArgs.push_back(WrapReg);
4515 // Decorate the wrapper register and generate a new instruction
4516 MRI->setType(WrapReg, LLT::pointer(0, 64));
4517 GR.assignSPIRVTypeToVReg(OpType, WrapReg, *MF);
4518 auto MIB = BuildMI(*I.getParent(), I, I.getDebugLoc(),
4519 TII.get(SPIRV::OpSpecConstantOp))
4520 .addDef(WrapReg)
4521 .addUse(GR.getSPIRVTypeID(OpType))
4522 .addImm(static_cast<uint32_t>(SPIRV::Opcode::Bitcast))
4523 .addUse(OpReg);
4524 GR.add(OpDefine, MIB);
4525 MIB.constrainAllUses(TII, TRI, RBI);
4526 }
4527 return true;
4528}
4529
4530bool SPIRVInstructionSelector::selectDerivativeInst(
4531 Register ResVReg, SPIRVTypeInst ResType, MachineInstr &I,
4532 const unsigned DPdOpCode) const {
4533 // TODO: This should check specifically for Fragment Execution Model, but STI
4534 // doesn't provide that information yet. See #167562
4535 errorIfInstrOutsideShader(I);
4536
4537 // If the arg/result types are half then we need to wrap the instr in
4538 // conversions to float
4539 // This case occurs because a half arg/result is legal in HLSL but not spirv.
4540 Register SrcReg = I.getOperand(2).getReg();
4541 SPIRVTypeInst SrcType = GR.getSPIRVTypeForVReg(SrcReg);
4542 unsigned BitWidth = std::min(GR.getScalarOrVectorBitWidth(SrcType),
4543 GR.getScalarOrVectorBitWidth(ResType));
4544 if (BitWidth == 32)
4545 return BuildMI(*I.getParent(), I, I.getDebugLoc(), TII.get(DPdOpCode))
4546 .addDef(ResVReg)
4547 .addUse(GR.getSPIRVTypeID(ResType))
4548 .addUse(I.getOperand(2).getReg());
4549
4550 MachineIRBuilder MIRBuilder(I);
4551 unsigned componentCount = GR.getScalarOrVectorComponentCount(SrcType);
4552 SPIRVTypeInst F32ConvertTy = GR.getOrCreateSPIRVFloatType(32, I, TII);
4553 if (componentCount != 1)
4554 F32ConvertTy = GR.getOrCreateSPIRVVectorType(F32ConvertTy, componentCount,
4555 MIRBuilder, false);
4556
4557 const TargetRegisterClass *RegClass = GR.getRegClass(SrcType);
4558 Register ConvertToVReg = MRI->createVirtualRegister(RegClass);
4559 Register DpdOpVReg = MRI->createVirtualRegister(RegClass);
4560
4561 BuildMI(*I.getParent(), I, I.getDebugLoc(), TII.get(SPIRV::OpFConvert))
4562 .addDef(ConvertToVReg)
4563 .addUse(GR.getSPIRVTypeID(F32ConvertTy))
4564 .addUse(SrcReg)
4565 .constrainAllUses(TII, TRI, RBI);
4566 BuildMI(*I.getParent(), I, I.getDebugLoc(), TII.get(DPdOpCode))
4567 .addDef(DpdOpVReg)
4568 .addUse(GR.getSPIRVTypeID(F32ConvertTy))
4569 .addUse(ConvertToVReg)
4570 .constrainAllUses(TII, TRI, RBI);
4571 BuildMI(*I.getParent(), I, I.getDebugLoc(), TII.get(SPIRV::OpFConvert))
4572 .addDef(ResVReg)
4573 .addUse(GR.getSPIRVTypeID(ResType))
4574 .addUse(DpdOpVReg)
4575 .constrainAllUses(TII, TRI, RBI);
4576 return true;
4577}
4578
4579bool SPIRVInstructionSelector::selectIntrinsic(Register ResVReg,
4580 SPIRVTypeInst ResType,
4581 MachineInstr &I) const {
4582 MachineBasicBlock &BB = *I.getParent();
4583 Intrinsic::ID IID = cast<GIntrinsic>(I).getIntrinsicID();
4584 switch (IID) {
4585 case Intrinsic::spv_load:
4586 return selectLoad(ResVReg, ResType, I);
4587 case Intrinsic::spv_atomic_load:
4588 return selectAtomicLoad(ResVReg, ResType, I);
4589 case Intrinsic::spv_store:
4590 return selectStore(I);
4591 case Intrinsic::spv_atomic_store:
4592 return selectAtomicStore(I);
4593 case Intrinsic::spv_extractv:
4594 return selectExtractVal(ResVReg, ResType, I);
4595 case Intrinsic::spv_insertv:
4596 return selectInsertVal(ResVReg, ResType, I);
4597 case Intrinsic::spv_extractelt:
4598 return selectExtractElt(ResVReg, ResType, I);
4599 case Intrinsic::spv_insertelt:
4600 return selectInsertElt(ResVReg, ResType, I);
4601 case Intrinsic::spv_gep:
4602 return selectGEP(ResVReg, ResType, I);
4603 case Intrinsic::spv_bitcast: {
4604 Register OpReg = I.getOperand(2).getReg();
4605 SPIRVTypeInst OpType =
4606 OpReg.isValid() ? GR.getSPIRVTypeForVReg(OpReg) : nullptr;
4607 if (!GR.isBitcastCompatible(ResType, OpType))
4608 report_fatal_error("incompatible result and operand types in a bitcast");
4609 return selectOpWithSrcs(ResVReg, ResType, I, {OpReg}, SPIRV::OpBitcast);
4610 }
4611 case Intrinsic::spv_unref_global:
4612 case Intrinsic::spv_init_global: {
4613 MachineInstr *MI = MRI->getVRegDef(I.getOperand(1).getReg());
4614 MachineInstr *Init = I.getNumExplicitOperands() > 2
4615 ? MRI->getVRegDef(I.getOperand(2).getReg())
4616 : nullptr;
4617 assert(MI);
4618 Register GVarVReg = MI->getOperand(0).getReg();
4619 if (!selectGlobalValue(GVarVReg, *MI, Init))
4620 return false;
4621 // We violate SSA form by inserting OpVariable and still having a gMIR
4622 // instruction %vreg = G_GLOBAL_VALUE @gvar. We need to fix this by erasing
4623 // the duplicated definition.
4624 if (MI->getOpcode() == TargetOpcode::G_GLOBAL_VALUE) {
4625 GR.invalidateMachineInstr(MI);
4626 MI->eraseFromParent();
4627 }
4628 return true;
4629 }
4630 case Intrinsic::spv_undef: {
4631 auto MIB = BuildMI(BB, I, I.getDebugLoc(), TII.get(SPIRV::OpUndef))
4632 .addDef(ResVReg)
4633 .addUse(GR.getSPIRVTypeID(ResType));
4634 MIB.constrainAllUses(TII, TRI, RBI);
4635 return true;
4636 }
4637 case Intrinsic::spv_named_boolean_spec_constant: {
4638 auto Opcode = I.getOperand(3).getImm() ? SPIRV::OpSpecConstantTrue
4639 : SPIRV::OpSpecConstantFalse;
4640
4641 auto MIB = BuildMI(BB, I, I.getDebugLoc(), TII.get(Opcode))
4642 .addDef(I.getOperand(0).getReg())
4643 .addUse(GR.getSPIRVTypeID(ResType));
4644 MIB.constrainAllUses(TII, TRI, RBI);
4645 unsigned SpecId = I.getOperand(2).getImm();
4646 buildOpDecorate(I.getOperand(0).getReg(), *++MIB->getIterator(), TII,
4647 SPIRV::Decoration::SpecId, {SpecId});
4648
4649 return true;
4650 }
4651 case Intrinsic::spv_const_composite: {
4652 // If no values are attached, the composite is null constant.
4653 bool IsNull = I.getNumExplicitDefs() + 1 == I.getNumExplicitOperands();
4654 SmallVector<Register> CompositeArgs;
4655 MRI->setRegClass(ResVReg, GR.getRegClass(ResType));
4656
4657 // skip type MD node we already used when generated assign.type for this
4658 if (!IsNull) {
4659 if (!wrapIntoSpecConstantOp(I, CompositeArgs))
4660 return false;
4661 std::function<bool(Register)> HasSpecConstOperand =
4662 [&](Register Reg) -> bool {
4663 MachineInstr *Def = MRI->getVRegDef(Reg);
4664 if (!Def)
4665 return false;
4666 if (!isConstReg(MRI, Def))
4667 return true;
4668 // Recurse into not-yet-selected spv_const_composite intrinsics
4669 // to detect transitive spec constant operands.
4670 if (isSpvIntrinsic(*Def, Intrinsic::spv_const_composite)) {
4671 for (unsigned J = Def->getNumExplicitDefs() + 1;
4672 J < Def->getNumExplicitOperands(); ++J) {
4673 if (Def->getOperand(J).isReg() &&
4674 HasSpecConstOperand(Def->getOperand(J).getReg()))
4675 return true;
4676 }
4677 }
4678 return false;
4679 };
4680 bool HasSpecConst = llvm::any_of(CompositeArgs, HasSpecConstOperand);
4681 unsigned CompositeOpc = HasSpecConst ? SPIRV::OpSpecConstantComposite
4682 : SPIRV::OpConstantComposite;
4683 unsigned ContinuedOpc = HasSpecConst
4684 ? SPIRV::OpSpecConstantCompositeContinuedINTEL
4685 : SPIRV::OpConstantCompositeContinuedINTEL;
4686 MachineIRBuilder MIR(I);
4687 SmallVector<MachineInstr *, 4> Instructions = createContinuedInstructions(
4688 MIR, CompositeOpc, 3, ContinuedOpc, CompositeArgs, ResVReg,
4689 GR.getSPIRVTypeID(ResType));
4690 for (auto *Instr : Instructions) {
4691 Instr->setDebugLoc(I.getDebugLoc());
4692 constrainSelectedInstRegOperands(*Instr, TII, TRI, RBI);
4693 }
4694 return true;
4695 } else {
4696 auto MIB = BuildMI(BB, I, I.getDebugLoc(), TII.get(SPIRV::OpConstantNull))
4697 .addDef(ResVReg)
4698 .addUse(GR.getSPIRVTypeID(ResType));
4699 MIB.constrainAllUses(TII, TRI, RBI);
4700 return true;
4701 }
4702 }
4703 case Intrinsic::spv_assign_name: {
4704 auto MIB = BuildMI(BB, I, I.getDebugLoc(), TII.get(SPIRV::OpName));
4705 MIB.addUse(I.getOperand(I.getNumExplicitDefs() + 1).getReg());
4706 for (unsigned i = I.getNumExplicitDefs() + 2;
4707 i < I.getNumExplicitOperands(); ++i) {
4708 MIB.addImm(I.getOperand(i).getImm());
4709 }
4710 MIB.constrainAllUses(TII, TRI, RBI);
4711 return true;
4712 }
4713 case Intrinsic::spv_switch: {
4714 auto MIB = BuildMI(BB, I, I.getDebugLoc(), TII.get(SPIRV::OpSwitch));
4715 for (unsigned i = 1; i < I.getNumExplicitOperands(); ++i) {
4716 if (I.getOperand(i).isReg())
4717 MIB.addReg(I.getOperand(i).getReg());
4718 else if (I.getOperand(i).isCImm())
4719 addNumImm(I.getOperand(i).getCImm()->getValue(), MIB);
4720 else if (I.getOperand(i).isMBB())
4721 MIB.addMBB(I.getOperand(i).getMBB());
4722 else
4723 llvm_unreachable("Unexpected OpSwitch operand");
4724 }
4725 MIB.constrainAllUses(TII, TRI, RBI);
4726 return true;
4727 }
4728 case Intrinsic::spv_loop_merge: {
4729 auto MIB = BuildMI(BB, I, I.getDebugLoc(), TII.get(SPIRV::OpLoopMerge));
4730 for (unsigned i = 1; i < I.getNumExplicitOperands(); ++i) {
4731 if (I.getOperand(i).isMBB())
4732 MIB.addMBB(I.getOperand(i).getMBB());
4733 else
4734 MIB.addImm(foldImm(I.getOperand(i), MRI));
4735 }
4736 MIB.constrainAllUses(TII, TRI, RBI);
4737 return true;
4738 }
4739 case Intrinsic::spv_loop_control_intel: {
4740 auto MIB =
4741 BuildMI(BB, I, I.getDebugLoc(), TII.get(SPIRV::OpLoopControlINTEL));
4742 for (unsigned J = 1; J < I.getNumExplicitOperands(); ++J)
4743 MIB.addImm(foldImm(I.getOperand(J), MRI));
4744 MIB.constrainAllUses(TII, TRI, RBI);
4745 return true;
4746 }
4747 case Intrinsic::spv_selection_merge: {
4748 auto MIB =
4749 BuildMI(BB, I, I.getDebugLoc(), TII.get(SPIRV::OpSelectionMerge));
4750 assert(I.getOperand(1).isMBB() &&
4751 "operand 1 to spv_selection_merge must be a basic block");
4752 MIB.addMBB(I.getOperand(1).getMBB());
4753 MIB.addImm(getSelectionOperandForImm(I.getOperand(2).getImm()));
4754 MIB.constrainAllUses(TII, TRI, RBI);
4755 return true;
4756 }
4757 case Intrinsic::spv_cmpxchg:
4758 return selectAtomicCmpXchg(ResVReg, ResType, I);
4759 case Intrinsic::spv_unreachable:
4760 BuildMI(BB, I, I.getDebugLoc(), TII.get(SPIRV::OpUnreachable))
4761 .constrainAllUses(TII, TRI, RBI);
4762 return true;
4763 case Intrinsic::spv_alloca:
4764 return selectFrameIndex(ResVReg, ResType, I);
4765 case Intrinsic::spv_alloca_array:
4766 return selectAllocaArray(ResVReg, ResType, I);
4767 case Intrinsic::spv_assume:
4768 if (STI.canUseExtension(SPIRV::Extension::SPV_KHR_expect_assume)) {
4769 BuildMI(BB, I, I.getDebugLoc(), TII.get(SPIRV::OpAssumeTrueKHR))
4770 .addUse(I.getOperand(1).getReg())
4771 .constrainAllUses(TII, TRI, RBI);
4772 return true;
4773 }
4774 break;
4775 case Intrinsic::spv_expect:
4776 if (STI.canUseExtension(SPIRV::Extension::SPV_KHR_expect_assume)) {
4777 BuildMI(BB, I, I.getDebugLoc(), TII.get(SPIRV::OpExpectKHR))
4778 .addDef(ResVReg)
4779 .addUse(GR.getSPIRVTypeID(ResType))
4780 .addUse(I.getOperand(2).getReg())
4781 .addUse(I.getOperand(3).getReg())
4782 .constrainAllUses(TII, TRI, RBI);
4783 return true;
4784 }
4785 break;
4786 case Intrinsic::arithmetic_fence:
4787 if (STI.canUseExtension(SPIRV::Extension::SPV_EXT_arithmetic_fence)) {
4788 BuildMI(BB, I, I.getDebugLoc(), TII.get(SPIRV::OpArithmeticFenceEXT))
4789 .addDef(ResVReg)
4790 .addUse(GR.getSPIRVTypeID(ResType))
4791 .addUse(I.getOperand(2).getReg())
4792 .constrainAllUses(TII, TRI, RBI);
4793 return true;
4794 } else
4795 return BuildCOPY(ResVReg, I.getOperand(2).getReg(), I);
4796 break;
4797 case Intrinsic::spv_thread_id:
4798 // The HLSL SV_DispatchThreadID semantic is lowered to llvm.spv.thread.id
4799 // intrinsic in LLVM IR for SPIR-V backend.
4800 //
4801 // In SPIR-V backend, llvm.spv.thread.id is now correctly translated to a
4802 // `GlobalInvocationId` builtin variable
4803 return loadVec3BuiltinInputID(SPIRV::BuiltIn::GlobalInvocationId, ResVReg,
4804 ResType, I);
4805 case Intrinsic::spv_thread_id_in_group:
4806 // The HLSL SV_GroupThreadId semantic is lowered to
4807 // llvm.spv.thread.id.in.group intrinsic in LLVM IR for SPIR-V backend.
4808 //
4809 // In SPIR-V backend, llvm.spv.thread.id.in.group is now correctly
4810 // translated to a `LocalInvocationId` builtin variable
4811 return loadVec3BuiltinInputID(SPIRV::BuiltIn::LocalInvocationId, ResVReg,
4812 ResType, I);
4813 case Intrinsic::spv_group_id:
4814 // The HLSL SV_GroupId semantic is lowered to
4815 // llvm.spv.group.id intrinsic in LLVM IR for SPIR-V backend.
4816 //
4817 // In SPIR-V backend, llvm.spv.group.id is now translated to a `WorkgroupId`
4818 // builtin variable
4819 return loadVec3BuiltinInputID(SPIRV::BuiltIn::WorkgroupId, ResVReg, ResType,
4820 I);
4821 case Intrinsic::spv_flattened_thread_id_in_group:
4822 // The HLSL SV_GroupIndex semantic is lowered to
4823 // llvm.spv.flattened.thread.id.in.group() intrinsic in LLVM IR for SPIR-V
4824 // backend.
4825 //
4826 // In SPIR-V backend, llvm.spv.flattened.thread.id.in.group is translated to
4827 // a `LocalInvocationIndex` builtin variable
4828 return loadBuiltinInputID(SPIRV::BuiltIn::LocalInvocationIndex, ResVReg,
4829 ResType, I);
4830 case Intrinsic::spv_workgroup_size:
4831 return loadVec3BuiltinInputID(SPIRV::BuiltIn::WorkgroupSize, ResVReg,
4832 ResType, I);
4833 case Intrinsic::spv_global_size:
4834 return loadVec3BuiltinInputID(SPIRV::BuiltIn::GlobalSize, ResVReg, ResType,
4835 I);
4836 case Intrinsic::spv_global_offset:
4837 return loadVec3BuiltinInputID(SPIRV::BuiltIn::GlobalOffset, ResVReg,
4838 ResType, I);
4839 case Intrinsic::spv_num_workgroups:
4840 return loadVec3BuiltinInputID(SPIRV::BuiltIn::NumWorkgroups, ResVReg,
4841 ResType, I);
4842 case Intrinsic::spv_subgroup_size:
4843 return loadBuiltinInputID(SPIRV::BuiltIn::SubgroupSize, ResVReg, ResType,
4844 I);
4845 case Intrinsic::spv_num_subgroups:
4846 return loadBuiltinInputID(SPIRV::BuiltIn::NumSubgroups, ResVReg, ResType,
4847 I);
4848 case Intrinsic::spv_subgroup_id:
4849 return loadBuiltinInputID(SPIRV::BuiltIn::SubgroupId, ResVReg, ResType, I);
4850 case Intrinsic::spv_subgroup_local_invocation_id:
4851 return loadBuiltinInputID(SPIRV::BuiltIn::SubgroupLocalInvocationId,
4852 ResVReg, ResType, I);
4853 case Intrinsic::spv_subgroup_max_size:
4854 return loadBuiltinInputID(SPIRV::BuiltIn::SubgroupMaxSize, ResVReg, ResType,
4855 I);
4856 case Intrinsic::spv_fdot:
4857 return selectFloatDot(ResVReg, ResType, I);
4858 case Intrinsic::spv_udot:
4859 case Intrinsic::spv_sdot:
4860 if (STI.canUseExtension(SPIRV::Extension::SPV_KHR_integer_dot_product) ||
4861 STI.isAtLeastSPIRVVer(VersionTuple(1, 6)))
4862 return selectIntegerDot(ResVReg, ResType, I,
4863 /*Signed=*/IID == Intrinsic::spv_sdot);
4864 return selectIntegerDotExpansion(ResVReg, ResType, I);
4865 case Intrinsic::spv_dot4add_i8packed:
4866 if (STI.canUseExtension(SPIRV::Extension::SPV_KHR_integer_dot_product) ||
4867 STI.isAtLeastSPIRVVer(VersionTuple(1, 6)))
4868 return selectDot4AddPacked<true>(ResVReg, ResType, I);
4869 return selectDot4AddPackedExpansion<true>(ResVReg, ResType, I);
4870 case Intrinsic::spv_dot4add_u8packed:
4871 if (STI.canUseExtension(SPIRV::Extension::SPV_KHR_integer_dot_product) ||
4872 STI.isAtLeastSPIRVVer(VersionTuple(1, 6)))
4873 return selectDot4AddPacked<false>(ResVReg, ResType, I);
4874 return selectDot4AddPackedExpansion<false>(ResVReg, ResType, I);
4875 case Intrinsic::spv_all:
4876 return selectAll(ResVReg, ResType, I);
4877 case Intrinsic::spv_any:
4878 return selectAny(ResVReg, ResType, I);
4879 case Intrinsic::spv_cross:
4880 return selectExtInst(ResVReg, ResType, I, CL::cross, GL::Cross);
4881 case Intrinsic::spv_distance:
4882 return selectExtInst(ResVReg, ResType, I, CL::distance, GL::Distance);
4883 case Intrinsic::spv_lerp:
4884 return selectExtInst(ResVReg, ResType, I, CL::mix, GL::FMix);
4885 case Intrinsic::spv_length:
4886 return selectExtInst(ResVReg, ResType, I, CL::length, GL::Length);
4887 case Intrinsic::spv_degrees:
4888 return selectExtInst(ResVReg, ResType, I, CL::degrees, GL::Degrees);
4889 case Intrinsic::spv_faceforward:
4890 return selectExtInst(ResVReg, ResType, I, GL::FaceForward);
4891 case Intrinsic::spv_frac:
4892 return selectExtInst(ResVReg, ResType, I, CL::fract, GL::Fract);
4893 case Intrinsic::spv_isinf:
4894 return selectOpIsInf(ResVReg, ResType, I);
4895 case Intrinsic::spv_isnan:
4896 return selectOpIsNan(ResVReg, ResType, I);
4897 case Intrinsic::spv_normalize:
4898 return selectExtInst(ResVReg, ResType, I, CL::normalize, GL::Normalize);
4899 case Intrinsic::spv_refract:
4900 return selectExtInst(ResVReg, ResType, I, GL::Refract);
4901 case Intrinsic::spv_reflect:
4902 return selectExtInst(ResVReg, ResType, I, GL::Reflect);
4903 case Intrinsic::spv_rsqrt:
4904 return selectExtInst(ResVReg, ResType, I, CL::rsqrt, GL::InverseSqrt);
4905 case Intrinsic::spv_sign:
4906 return selectSign(ResVReg, ResType, I);
4907 case Intrinsic::spv_smoothstep:
4908 return selectExtInst(ResVReg, ResType, I, CL::smoothstep, GL::SmoothStep);
4909 case Intrinsic::spv_firstbituhigh: // There is no CL equivalent of FindUMsb
4910 return selectFirstBitHigh(ResVReg, ResType, I, /*IsSigned=*/false);
4911 case Intrinsic::spv_firstbitshigh: // There is no CL equivalent of FindSMsb
4912 return selectFirstBitHigh(ResVReg, ResType, I, /*IsSigned=*/true);
4913 case Intrinsic::spv_firstbitlow: // There is no CL equivlent of FindILsb
4914 return selectFirstBitLow(ResVReg, ResType, I);
4915 case Intrinsic::spv_all_memory_barrier:
4916 return selectBarrierInst(I, SPIRV::Scope::Device,
4917 SPIRV::MemorySemantics::UniformMemory |
4918 SPIRV::MemorySemantics::ImageMemory |
4919 SPIRV::MemorySemantics::WorkgroupMemory,
4920 /*WithGroupSync*/ false);
4921 case Intrinsic::spv_all_memory_barrier_with_group_sync:
4922 return selectBarrierInst(I, SPIRV::Scope::Device,
4923 SPIRV::MemorySemantics::UniformMemory |
4924 SPIRV::MemorySemantics::ImageMemory |
4925 SPIRV::MemorySemantics::WorkgroupMemory,
4926 /*WithGroupSync*/ true);
4927 case Intrinsic::spv_device_memory_barrier:
4928 return selectBarrierInst(I, SPIRV::Scope::Device,
4929 SPIRV::MemorySemantics::UniformMemory |
4930 SPIRV::MemorySemantics::ImageMemory,
4931 /*WithGroupSync*/ false);
4932 case Intrinsic::spv_device_memory_barrier_with_group_sync:
4933 return selectBarrierInst(I, SPIRV::Scope::Device,
4934 SPIRV::MemorySemantics::UniformMemory |
4935 SPIRV::MemorySemantics::ImageMemory,
4936 /*WithGroupSync*/ true);
4937 case Intrinsic::spv_group_memory_barrier:
4938 return selectBarrierInst(I, SPIRV::Scope::Workgroup,
4939 SPIRV::MemorySemantics::WorkgroupMemory,
4940 /*WithGroupSync*/ false);
4941 case Intrinsic::spv_group_memory_barrier_with_group_sync:
4942 return selectBarrierInst(I, SPIRV::Scope::Workgroup,
4943 SPIRV::MemorySemantics::WorkgroupMemory,
4944 /*WithGroupSync*/ true);
4945 case Intrinsic::spv_generic_cast_to_ptr_explicit: {
4946 Register PtrReg = I.getOperand(I.getNumExplicitDefs() + 1).getReg();
4947 SPIRV::StorageClass::StorageClass ResSC =
4948 GR.getPointerStorageClass(ResType);
4949 if (!isGenericCastablePtr(ResSC))
4950 report_fatal_error("The target storage class is not castable from the "
4951 "Generic storage class");
4952 BuildMI(BB, I, I.getDebugLoc(), TII.get(SPIRV::OpGenericCastToPtrExplicit))
4953 .addDef(ResVReg)
4954 .addUse(GR.getSPIRVTypeID(ResType))
4955 .addUse(PtrReg)
4956 .addImm(ResSC)
4957 .constrainAllUses(TII, TRI, RBI);
4958 return true;
4959 }
4960 case Intrinsic::spv_lifetime_start:
4961 case Intrinsic::spv_lifetime_end: {
4962 unsigned Op = IID == Intrinsic::spv_lifetime_start ? SPIRV::OpLifetimeStart
4963 : SPIRV::OpLifetimeStop;
4964 int64_t Size = I.getOperand(I.getNumExplicitDefs() + 1).getImm();
4965 Register PtrReg = I.getOperand(I.getNumExplicitDefs() + 2).getReg();
4966 if (Size == -1)
4967 Size = 0;
4968 BuildMI(BB, I, I.getDebugLoc(), TII.get(Op))
4969 .addUse(PtrReg)
4970 .addImm(Size)
4971 .constrainAllUses(TII, TRI, RBI);
4972 return true;
4973 }
4974 case Intrinsic::spv_saturate:
4975 return selectSaturate(ResVReg, ResType, I);
4976 case Intrinsic::spv_nclamp:
4977 return selectExtInst(ResVReg, ResType, I, CL::fclamp, GL::NClamp);
4978 case Intrinsic::spv_uclamp:
4979 return selectExtInst(ResVReg, ResType, I, CL::u_clamp, GL::UClamp);
4980 case Intrinsic::spv_sclamp:
4981 return selectExtInst(ResVReg, ResType, I, CL::s_clamp, GL::SClamp);
4982 case Intrinsic::spv_subgroup_prefix_bit_count:
4983 return selectWavePrefixBitCount(ResVReg, ResType, I);
4984 case Intrinsic::spv_wave_active_countbits:
4985 return selectWaveActiveCountBits(ResVReg, ResType, I);
4986 case Intrinsic::spv_wave_all_equal:
4987 return selectWaveActiveAllEqual(ResVReg, ResType, I);
4988 case Intrinsic::spv_wave_all:
4989 return selectWaveOpInst(ResVReg, ResType, I, SPIRV::OpGroupNonUniformAll);
4990 case Intrinsic::spv_wave_any:
4991 return selectWaveOpInst(ResVReg, ResType, I, SPIRV::OpGroupNonUniformAny);
4992 case Intrinsic::spv_subgroup_ballot:
4993 return selectWaveOpInst(ResVReg, ResType, I,
4994 SPIRV::OpGroupNonUniformBallot);
4995 case Intrinsic::spv_wave_is_first_lane:
4996 return selectWaveOpInst(ResVReg, ResType, I, SPIRV::OpGroupNonUniformElect);
4997 case Intrinsic::spv_wave_reduce_or:
4998 return selectWaveReduceOp(ResVReg, ResType, I,
4999 SPIRV::OpGroupNonUniformBitwiseOr);
5000 case Intrinsic::spv_wave_reduce_xor:
5001 return selectWaveReduceOp(ResVReg, ResType, I,
5002 SPIRV::OpGroupNonUniformBitwiseXor);
5003 case Intrinsic::spv_wave_reduce_and:
5004 return selectWaveReduceOp(ResVReg, ResType, I,
5005 SPIRV::OpGroupNonUniformBitwiseAnd);
5006 case Intrinsic::spv_wave_reduce_umax:
5007 return selectWaveReduceMax(ResVReg, ResType, I, /*IsUnsigned*/ true);
5008 case Intrinsic::spv_wave_reduce_max:
5009 return selectWaveReduceMax(ResVReg, ResType, I, /*IsUnsigned*/ false);
5010 case Intrinsic::spv_wave_reduce_umin:
5011 return selectWaveReduceMin(ResVReg, ResType, I, /*IsUnsigned*/ true);
5012 case Intrinsic::spv_wave_reduce_min:
5013 return selectWaveReduceMin(ResVReg, ResType, I, /*IsUnsigned*/ false);
5014 case Intrinsic::spv_wave_reduce_sum:
5015 return selectWaveReduceSum(ResVReg, ResType, I);
5016 case Intrinsic::spv_wave_product:
5017 return selectWaveReduceProduct(ResVReg, ResType, I);
5018 case Intrinsic::spv_wave_readlane:
5019 return selectWaveOpInst(ResVReg, ResType, I,
5020 SPIRV::OpGroupNonUniformShuffle);
5021 case Intrinsic::spv_wave_prefix_sum:
5022 return selectWaveExclusiveScanSum(ResVReg, ResType, I);
5023 case Intrinsic::spv_wave_prefix_product:
5024 return selectWaveExclusiveScanProduct(ResVReg, ResType, I);
5025 case Intrinsic::spv_quad_read_across_x: {
5026 return selectQuadSwap(ResVReg, ResType, I, /*Direction*/ 0);
5027 }
5028 case Intrinsic::spv_quad_read_across_y: {
5029 return selectQuadSwap(ResVReg, ResType, I, /*Direction*/ 1);
5030 }
5031 case Intrinsic::spv_step:
5032 return selectExtInst(ResVReg, ResType, I, CL::step, GL::Step);
5033 case Intrinsic::spv_radians:
5034 return selectExtInst(ResVReg, ResType, I, CL::radians, GL::Radians);
5035 // Discard intrinsics which we do not expect to actually represent code after
5036 // lowering or intrinsics which are not implemented but should not crash when
5037 // found in a customer's LLVM IR input.
5038 case Intrinsic::instrprof_increment:
5039 case Intrinsic::instrprof_increment_step:
5040 case Intrinsic::instrprof_value_profile:
5041 break;
5042 // Discard internal intrinsics.
5043 case Intrinsic::spv_value_md:
5044 break;
5045 case Intrinsic::spv_resource_handlefrombinding: {
5046 return selectHandleFromBinding(ResVReg, ResType, I);
5047 }
5048 case Intrinsic::spv_resource_counterhandlefrombinding:
5049 return selectCounterHandleFromBinding(ResVReg, ResType, I);
5050 case Intrinsic::spv_resource_updatecounter:
5051 return selectUpdateCounter(ResVReg, ResType, I);
5052 case Intrinsic::spv_resource_store_typedbuffer: {
5053 return selectImageWriteIntrinsic(I);
5054 }
5055 case Intrinsic::spv_resource_load_typedbuffer: {
5056 return selectReadImageIntrinsic(ResVReg, ResType, I);
5057 }
5058 case Intrinsic::spv_resource_load_level: {
5059 return selectLoadLevelIntrinsic(ResVReg, ResType, I);
5060 }
5061 case Intrinsic::spv_resource_getdimensions_x:
5062 case Intrinsic::spv_resource_getdimensions_xy:
5063 case Intrinsic::spv_resource_getdimensions_xyz: {
5064 return selectGetDimensionsIntrinsic(ResVReg, ResType, I);
5065 }
5066 case Intrinsic::spv_resource_getdimensions_levels_x:
5067 case Intrinsic::spv_resource_getdimensions_levels_xy:
5068 case Intrinsic::spv_resource_getdimensions_levels_xyz: {
5069 return selectGetDimensionsLevelsIntrinsic(ResVReg, ResType, I);
5070 }
5071 case Intrinsic::spv_resource_getdimensions_ms_xy:
5072 case Intrinsic::spv_resource_getdimensions_ms_xyz: {
5073 return selectGetDimensionsMSIntrinsic(ResVReg, ResType, I);
5074 }
5075 case Intrinsic::spv_resource_calculate_lod:
5076 case Intrinsic::spv_resource_calculate_lod_unclamped:
5077 return selectCalculateLodIntrinsic(ResVReg, ResType, I);
5078 case Intrinsic::spv_resource_sample:
5079 case Intrinsic::spv_resource_sample_clamp:
5080 return selectSampleBasicIntrinsic(ResVReg, ResType, I);
5081 case Intrinsic::spv_resource_samplebias:
5082 case Intrinsic::spv_resource_samplebias_clamp:
5083 return selectSampleBiasIntrinsic(ResVReg, ResType, I);
5084 case Intrinsic::spv_resource_samplegrad:
5085 case Intrinsic::spv_resource_samplegrad_clamp:
5086 return selectSampleGradIntrinsic(ResVReg, ResType, I);
5087 case Intrinsic::spv_resource_samplelevel:
5088 return selectSampleLevelIntrinsic(ResVReg, ResType, I);
5089 case Intrinsic::spv_resource_samplecmp:
5090 case Intrinsic::spv_resource_samplecmp_clamp:
5091 return selectSampleCmpIntrinsic(ResVReg, ResType, I);
5092 case Intrinsic::spv_resource_samplecmplevelzero:
5093 return selectSampleCmpLevelZeroIntrinsic(ResVReg, ResType, I);
5094 case Intrinsic::spv_resource_gather:
5095 case Intrinsic::spv_resource_gather_cmp:
5096 return selectGatherIntrinsic(ResVReg, ResType, I);
5097 case Intrinsic::spv_resource_getpointer: {
5098 return selectResourceGetPointer(ResVReg, ResType, I);
5099 }
5100 case Intrinsic::spv_pushconstant_getpointer: {
5101 return selectPushConstantGetPointer(ResVReg, ResType, I);
5102 }
5103 case Intrinsic::spv_discard: {
5104 return selectDiscard(ResVReg, ResType, I);
5105 }
5106 case Intrinsic::spv_resource_nonuniformindex: {
5107 return selectResourceNonUniformIndex(ResVReg, ResType, I);
5108 }
5109 case Intrinsic::spv_unpackhalf2x16: {
5110 return selectExtInst(ResVReg, ResType, I, GL::UnpackHalf2x16);
5111 }
5112 case Intrinsic::spv_packhalf2x16: {
5113 return selectExtInst(ResVReg, ResType, I, GL::PackHalf2x16);
5114 }
5115 case Intrinsic::spv_ddx:
5116 return selectDerivativeInst(ResVReg, ResType, I, SPIRV::OpDPdx);
5117 case Intrinsic::spv_ddy:
5118 return selectDerivativeInst(ResVReg, ResType, I, SPIRV::OpDPdy);
5119 case Intrinsic::spv_ddx_coarse:
5120 return selectDerivativeInst(ResVReg, ResType, I, SPIRV::OpDPdxCoarse);
5121 case Intrinsic::spv_ddy_coarse:
5122 return selectDerivativeInst(ResVReg, ResType, I, SPIRV::OpDPdyCoarse);
5123 case Intrinsic::spv_ddx_fine:
5124 return selectDerivativeInst(ResVReg, ResType, I, SPIRV::OpDPdxFine);
5125 case Intrinsic::spv_ddy_fine:
5126 return selectDerivativeInst(ResVReg, ResType, I, SPIRV::OpDPdyFine);
5127 case Intrinsic::spv_fwidth:
5128 return selectDerivativeInst(ResVReg, ResType, I, SPIRV::OpFwidth);
5129 case Intrinsic::spv_masked_gather:
5130 if (STI.canUseExtension(SPIRV::Extension::SPV_INTEL_masked_gather_scatter))
5131 return selectMaskedGather(ResVReg, ResType, I);
5132 return diagnoseUnsupported(
5133 I, "llvm.masked.gather requires SPV_INTEL_masked_gather_scatter");
5134 case Intrinsic::spv_masked_scatter:
5135 if (STI.canUseExtension(SPIRV::Extension::SPV_INTEL_masked_gather_scatter))
5136 return selectMaskedScatter(I);
5137 return diagnoseUnsupported(
5138 I, "llvm.masked.scatter requires SPV_INTEL_masked_gather_scatter");
5139 default: {
5140 std::string DiagMsg;
5141 raw_string_ostream OS(DiagMsg);
5142 I.print(OS);
5143 DiagMsg = "Intrinsic selection not implemented: " + DiagMsg;
5144 report_fatal_error(DiagMsg.c_str(), false);
5145 }
5146 }
5147 return true;
5148}
5149
5150bool SPIRVInstructionSelector::selectHandleFromBinding(Register &ResVReg,
5151 SPIRVTypeInst ResType,
5152 MachineInstr &I) const {
5153 // The images need to be loaded in the same basic block as their use. We defer
5154 // loading the image to the intrinsic that uses it.
5155 if (ResType->getOpcode() == SPIRV::OpTypeImage)
5156 return true;
5157
5158 return loadHandleBeforePosition(ResVReg, GR.getSPIRVTypeForVReg(ResVReg),
5159 *cast<GIntrinsic>(&I), I);
5160}
5161
5162bool SPIRVInstructionSelector::selectCounterHandleFromBinding(
5163 Register &ResVReg, SPIRVTypeInst ResType, MachineInstr &I) const {
5164 auto &Intr = cast<GIntrinsic>(I);
5165 assert(Intr.getIntrinsicID() ==
5166 Intrinsic::spv_resource_counterhandlefrombinding);
5167
5168 // Extract information from the intrinsic call.
5169 Register MainHandleReg = Intr.getOperand(2).getReg();
5170 auto *MainHandleDef = cast<GIntrinsic>(getVRegDef(*MRI, MainHandleReg));
5171 assert(MainHandleDef->getIntrinsicID() ==
5172 Intrinsic::spv_resource_handlefrombinding);
5173
5174 uint32_t Set = getIConstVal(Intr.getOperand(4).getReg(), MRI);
5175 uint32_t Binding = getIConstVal(Intr.getOperand(3).getReg(), MRI);
5176 uint32_t ArraySize = getIConstVal(MainHandleDef->getOperand(4).getReg(), MRI);
5177 Register IndexReg = MainHandleDef->getOperand(5).getReg();
5178 std::string CounterName =
5179 getStringValueFromReg(MainHandleDef->getOperand(6).getReg(), *MRI) +
5180 ".counter";
5181
5182 // Create the counter variable.
5183 MachineIRBuilder MIRBuilder(I);
5184 Register CounterVarReg =
5185 buildPointerToResource(SPIRVTypeInst(GR.getPointeeType(ResType)),
5186 GR.getPointerStorageClass(ResType), Set, Binding,
5187 ArraySize, IndexReg, CounterName, MIRBuilder);
5188
5189 return BuildCOPY(ResVReg, CounterVarReg, I);
5190}
5191
5192bool SPIRVInstructionSelector::selectUpdateCounter(Register &ResVReg,
5193 SPIRVTypeInst ResType,
5194 MachineInstr &I) const {
5195 auto &Intr = cast<GIntrinsic>(I);
5196 assert(Intr.getIntrinsicID() == Intrinsic::spv_resource_updatecounter);
5197
5198 Register CounterHandleReg = Intr.getOperand(2).getReg();
5199 Register IncrReg = Intr.getOperand(3).getReg();
5200
5201 // The counter handle is a pointer to the counter variable (which is a struct
5202 // containing an i32). We need to get a pointer to that i32 member to do the
5203 // atomic operation.
5204#ifndef NDEBUG
5205 SPIRVTypeInst CounterVarType = GR.getSPIRVTypeForVReg(CounterHandleReg);
5206 SPIRVTypeInst CounterVarPointeeType = GR.getPointeeType(CounterVarType);
5207 assert(CounterVarPointeeType &&
5208 CounterVarPointeeType->getOpcode() == SPIRV::OpTypeStruct &&
5209 "Counter variable must be a struct");
5210 assert(GR.getPointerStorageClass(CounterVarType) ==
5211 SPIRV::StorageClass::StorageBuffer &&
5212 "Counter variable must be in the storage buffer storage class");
5213 assert(CounterVarPointeeType->getNumOperands() == 2 &&
5214 "Counter variable must have exactly 1 member in the struct");
5215 const SPIRVTypeInst MemberType =
5216 GR.getSPIRVTypeForVReg(CounterVarPointeeType->getOperand(1).getReg());
5217 assert(MemberType->getOpcode() == SPIRV::OpTypeInt &&
5218 "Counter variable struct must have a single i32 member");
5219#endif
5220
5221 // The struct has a single i32 member.
5222 MachineIRBuilder MIRBuilder(I);
5223 const Type *LLVMIntType =
5224 Type::getInt32Ty(I.getMF()->getFunction().getContext());
5225
5226 SPIRVTypeInst IntPtrType = GR.getOrCreateSPIRVPointerType(
5227 LLVMIntType, MIRBuilder, SPIRV::StorageClass::StorageBuffer);
5228
5229 Register Zero = buildI32Constant(0, I);
5230
5231 Register PtrToCounter =
5232 MRI->createVirtualRegister(GR.getRegClass(IntPtrType));
5233 BuildMI(*I.getParent(), I, I.getDebugLoc(), TII.get(SPIRV::OpAccessChain))
5234 .addDef(PtrToCounter)
5235 .addUse(GR.getSPIRVTypeID(IntPtrType))
5236 .addUse(CounterHandleReg)
5237 .addUse(Zero)
5238 .constrainAllUses(TII, TRI, RBI);
5239
5240 // For UAV/SSBO counters, the scope is Device. The counter variable is not
5241 // used as a flag. So the memory semantics can be None.
5242 Register Scope = buildI32Constant(SPIRV::Scope::Device, I);
5243 Register Semantics = buildI32Constant(SPIRV::MemorySemantics::None, I);
5244
5245 int64_t IncrVal = getIConstValSext(IncrReg, MRI);
5246 Register Incr = buildI32Constant(static_cast<uint32_t>(IncrVal), I);
5247
5248 Register AtomicRes = MRI->createVirtualRegister(GR.getRegClass(ResType));
5249 BuildMI(*I.getParent(), I, I.getDebugLoc(), TII.get(SPIRV::OpAtomicIAdd))
5250 .addDef(AtomicRes)
5251 .addUse(GR.getSPIRVTypeID(ResType))
5252 .addUse(PtrToCounter)
5253 .addUse(Scope)
5254 .addUse(Semantics)
5255 .addUse(Incr)
5256 .constrainAllUses(TII, TRI, RBI);
5257 if (IncrVal >= 0) {
5258 return BuildCOPY(ResVReg, AtomicRes, I);
5259 }
5260
5261 // In HLSL, IncrementCounter returns the value *before* the increment, while
5262 // DecrementCounter returns the value *after* the decrement. Both are lowered
5263 // to the same atomic intrinsic which returns the value *before* the
5264 // operation. So for decrements (negative IncrVal), we must subtract the
5265 // increment value from the result to get the post-decrement value.
5266 BuildMI(*I.getParent(), I, I.getDebugLoc(), TII.get(SPIRV::OpIAddS))
5267 .addDef(ResVReg)
5268 .addUse(GR.getSPIRVTypeID(ResType))
5269 .addUse(AtomicRes)
5270 .addUse(Incr)
5271 .constrainAllUses(TII, TRI, RBI);
5272 return true;
5273}
5274bool SPIRVInstructionSelector::selectReadImageIntrinsic(Register &ResVReg,
5275 SPIRVTypeInst ResType,
5276 MachineInstr &I) const {
5277
5278 // If the load of the image is in a different basic block, then
5279 // this will generate invalid code. A proper solution is to move
5280 // the OpLoad from selectHandleFromBinding here. However, to do
5281 // that we will need to change the return type of the intrinsic.
5282 // We will do that when we can, but for now trying to move forward with other
5283 // issues.
5284 Register ImageReg = I.getOperand(2).getReg();
5285 auto *ImageDef = cast<GIntrinsic>(getVRegDef(*MRI, ImageReg));
5286 Register NewImageReg = MRI->createVirtualRegister(MRI->getRegClass(ImageReg));
5287 if (!loadHandleBeforePosition(NewImageReg, GR.getSPIRVTypeForVReg(ImageReg),
5288 *ImageDef, I)) {
5289 return false;
5290 }
5291
5292 Register IdxReg = I.getOperand(3).getReg();
5293 DebugLoc Loc = I.getDebugLoc();
5294 MachineInstr &Pos = I;
5295
5296 return generateImageReadOrFetch(ResVReg, ResType, NewImageReg, IdxReg, Loc,
5297 Pos);
5298}
5299
5300bool SPIRVInstructionSelector::generateSampleImage(
5301 Register ResVReg, SPIRVTypeInst ResType, Register ImageReg,
5302 Register SamplerReg, Register CoordinateReg, const ImageOperands &ImOps,
5303 DebugLoc Loc, MachineInstr &Pos) const {
5304 auto *ImageDef = cast<GIntrinsic>(getVRegDef(*MRI, ImageReg));
5305 Register NewImageReg = MRI->createVirtualRegister(MRI->getRegClass(ImageReg));
5306 if (!loadHandleBeforePosition(NewImageReg, GR.getSPIRVTypeForVReg(ImageReg),
5307 *ImageDef, Pos)) {
5308 return false;
5309 }
5310
5311 auto *SamplerDef = cast<GIntrinsic>(getVRegDef(*MRI, SamplerReg));
5312 Register NewSamplerReg =
5313 MRI->createVirtualRegister(MRI->getRegClass(SamplerReg));
5314 if (!loadHandleBeforePosition(NewSamplerReg,
5315 GR.getSPIRVTypeForVReg(SamplerReg), *SamplerDef,
5316 Pos)) {
5317 return false;
5318 }
5319
5320 MachineIRBuilder MIRBuilder(Pos);
5321 SPIRVTypeInst SampledImageType = GR.getOrCreateOpTypeSampledImage(
5322 GR.getSPIRVTypeForVReg(ImageReg), MIRBuilder);
5323 Register SampledImageReg =
5324 MRI->createVirtualRegister(GR.getRegClass(SampledImageType));
5325
5326 BuildMI(*Pos.getParent(), Pos, Loc, TII.get(SPIRV::OpSampledImage))
5327 .addDef(SampledImageReg)
5328 .addUse(GR.getSPIRVTypeID(SampledImageType))
5329 .addUse(NewImageReg)
5330 .addUse(NewSamplerReg)
5331 .constrainAllUses(TII, TRI, RBI);
5332
5333 bool IsExplicitLod = ImOps.GradX.has_value() || ImOps.GradY.has_value() ||
5334 ImOps.Lod.has_value();
5335 unsigned Opcode = IsExplicitLod ? SPIRV::OpImageSampleExplicitLod
5336 : SPIRV::OpImageSampleImplicitLod;
5337 if (ImOps.Compare)
5338 Opcode = IsExplicitLod ? SPIRV::OpImageSampleDrefExplicitLod
5339 : SPIRV::OpImageSampleDrefImplicitLod;
5340
5341 auto MIB = BuildMI(*Pos.getParent(), Pos, Loc, TII.get(Opcode))
5342 .addDef(ResVReg)
5343 .addUse(GR.getSPIRVTypeID(ResType))
5344 .addUse(SampledImageReg)
5345 .addUse(CoordinateReg);
5346
5347 if (ImOps.Compare)
5348 MIB.addUse(*ImOps.Compare);
5349
5350 uint32_t ImageOperands = 0;
5351 if (ImOps.Bias)
5352 ImageOperands |= SPIRV::ImageOperand::Bias;
5353 if (ImOps.Lod)
5354 ImageOperands |= SPIRV::ImageOperand::Lod;
5355 if (ImOps.GradX && ImOps.GradY)
5356 ImageOperands |= SPIRV::ImageOperand::Grad;
5357 if (ImOps.Offset && !isScalarOrVectorIntConstantZero(*ImOps.Offset)) {
5358 if (isConstReg(MRI, *ImOps.Offset))
5359 ImageOperands |= SPIRV::ImageOperand::ConstOffset;
5360 else {
5361 Pos.emitGenericError(
5362 "Non-constant offsets are not supported in sample instructions.");
5363 }
5364 }
5365 if (ImOps.MinLod)
5366 ImageOperands |= SPIRV::ImageOperand::MinLod;
5367
5368 if (ImageOperands != 0) {
5369 MIB.addImm(ImageOperands);
5370 if (ImageOperands & SPIRV::ImageOperand::Bias)
5371 MIB.addUse(*ImOps.Bias);
5372 if (ImageOperands & SPIRV::ImageOperand::Lod)
5373 MIB.addUse(*ImOps.Lod);
5374 if (ImageOperands & SPIRV::ImageOperand::Grad) {
5375 MIB.addUse(*ImOps.GradX);
5376 MIB.addUse(*ImOps.GradY);
5377 }
5378 if (ImageOperands &
5379 (SPIRV::ImageOperand::ConstOffset | SPIRV::ImageOperand::Offset))
5380 MIB.addUse(*ImOps.Offset);
5381 if (ImageOperands & SPIRV::ImageOperand::MinLod)
5382 MIB.addUse(*ImOps.MinLod);
5383 }
5384
5385 MIB.constrainAllUses(TII, TRI, RBI);
5386 return true;
5387}
5388
5389bool SPIRVInstructionSelector::selectImageQuerySize(
5390 Register ImageReg, Register &ResVReg, MachineInstr &I,
5391 std::optional<Register> LodReg) const {
5392 unsigned Opcode =
5393 LodReg ? SPIRV::OpImageQuerySizeLod : SPIRV::OpImageQuerySize;
5394 SPIRVTypeInst ImageType = GR.getSPIRVTypeForVReg(ImageReg);
5395 assert(ImageType && ImageType->getOpcode() == SPIRV::OpTypeImage &&
5396 "ImageReg is not an image type.");
5397
5398 auto Dim = static_cast<SPIRV::Dim::Dim>(ImageType->getOperand(2).getImm());
5399 bool IsArray = ImageType->getOperand(4).getImm() != 0;
5400 unsigned NumComponents = 0;
5401 switch (Dim) {
5402 case SPIRV::Dim::DIM_1D:
5403 case SPIRV::Dim::DIM_Buffer:
5404 NumComponents = IsArray ? 2 : 1;
5405 break;
5406 case SPIRV::Dim::DIM_2D:
5407 case SPIRV::Dim::DIM_Cube:
5408 case SPIRV::Dim::DIM_Rect:
5409 NumComponents = IsArray ? 3 : 2;
5410 break;
5411 case SPIRV::Dim::DIM_3D:
5412 NumComponents = 3;
5413 break;
5414 default:
5415 I.emitGenericError("Unsupported image dimension for OpImageQuerySize.");
5416 return false;
5417 }
5418
5419 SPIRVTypeInst I32Ty = GR.getOrCreateSPIRVIntegerType(32, I, TII);
5420 SPIRVTypeInst ResType =
5421 NumComponents == 1
5422 ? I32Ty
5423 : GR.getOrCreateSPIRVVectorType(I32Ty, NumComponents, I, TII);
5424
5425 auto MIB = BuildMI(*I.getParent(), I, I.getDebugLoc(), TII.get(Opcode))
5426 .addDef(ResVReg)
5427 .addUse(GR.getSPIRVTypeID(ResType))
5428 .addUse(ImageReg);
5429 if (LodReg)
5430 MIB.addUse(*LodReg);
5431 MIB.constrainAllUses(TII, TRI, RBI);
5432 return true;
5433}
5434
5435bool SPIRVInstructionSelector::selectGetDimensionsIntrinsic(
5436 Register &ResVReg, SPIRVTypeInst ResType, MachineInstr &I) const {
5437 Register ImageReg = I.getOperand(2).getReg();
5438 auto *ImageDef = cast<GIntrinsic>(getVRegDef(*MRI, ImageReg));
5439 Register NewImageReg = MRI->createVirtualRegister(MRI->getRegClass(ImageReg));
5440 if (!loadHandleBeforePosition(NewImageReg, GR.getSPIRVTypeForVReg(ImageReg),
5441 *ImageDef, I)) {
5442 return false;
5443 }
5444 return selectImageQuerySize(NewImageReg, ResVReg, I);
5445}
5446
5447bool SPIRVInstructionSelector::selectGetDimensionsLevelsIntrinsic(
5448 Register &ResVReg, SPIRVTypeInst ResType, MachineInstr &I) const {
5449 Register ImageReg = I.getOperand(2).getReg();
5450 auto *ImageDef = cast<GIntrinsic>(getVRegDef(*MRI, ImageReg));
5451 Register NewImageReg = MRI->createVirtualRegister(MRI->getRegClass(ImageReg));
5452 if (!loadHandleBeforePosition(NewImageReg, GR.getSPIRVTypeForVReg(ImageReg),
5453 *ImageDef, I)) {
5454 return false;
5455 }
5456
5457 Register SizeReg = MRI->createVirtualRegister(&SPIRV::iIDRegClass);
5458 Register LodReg = I.getOperand(3).getReg();
5459
5460 assert(GR.getSPIRVTypeForVReg(NewImageReg)->getOperand(6).getImm() == 1 &&
5461 "OpImageQuerySizeLod and OpImageQueryLevels require a sampled image");
5462
5463 if (!selectImageQuerySize(NewImageReg, SizeReg, I, LodReg)) {
5464 return false;
5465 }
5466
5467 SPIRVTypeInst I32Ty = GR.getOrCreateSPIRVIntegerType(32, I, TII);
5468 Register LevelsReg = MRI->createVirtualRegister(&SPIRV::iIDRegClass);
5469 BuildMI(*I.getParent(), I, I.getDebugLoc(),
5470 TII.get(SPIRV::OpImageQueryLevels))
5471 .addDef(LevelsReg)
5472 .addUse(GR.getSPIRVTypeID(I32Ty))
5473 .addUse(NewImageReg)
5474 .constrainAllUses(TII, TRI, RBI);
5475
5476 BuildMI(*I.getParent(), I, I.getDebugLoc(),
5477 TII.get(SPIRV::OpCompositeConstruct))
5478 .addDef(ResVReg)
5479 .addUse(GR.getSPIRVTypeID(ResType))
5480 .addUse(SizeReg)
5481 .addUse(LevelsReg)
5482 .constrainAllUses(TII, TRI, RBI);
5483
5484 return true;
5485}
5486
5487bool SPIRVInstructionSelector::selectGetDimensionsMSIntrinsic(
5488 Register &ResVReg, SPIRVTypeInst ResType, MachineInstr &I) const {
5489 Register ImageReg = I.getOperand(2).getReg();
5490 auto *ImageDef = cast<GIntrinsic>(getVRegDef(*MRI, ImageReg));
5491 Register NewImageReg = MRI->createVirtualRegister(MRI->getRegClass(ImageReg));
5492 if (!loadHandleBeforePosition(NewImageReg, GR.getSPIRVTypeForVReg(ImageReg),
5493 *ImageDef, I)) {
5494 return false;
5495 }
5496
5497 Register SizeReg = MRI->createVirtualRegister(&SPIRV::iIDRegClass);
5498
5499 assert(GR.getSPIRVTypeForVReg(NewImageReg)->getOperand(5).getImm() == 1 &&
5500 "OpImageQuerySamples requires a multisampled image");
5501
5502 if (!selectImageQuerySize(NewImageReg, SizeReg, I)) {
5503 return false;
5504 }
5505
5506 Register SamplesReg = MRI->createVirtualRegister(&SPIRV::iIDRegClass);
5507
5508 SPIRVTypeInst I32Ty = GR.getOrCreateSPIRVIntegerType(32, I, TII);
5509 BuildMI(*I.getParent(), I, I.getDebugLoc(),
5510 TII.get(SPIRV::OpImageQuerySamples))
5511 .addDef(SamplesReg)
5512 .addUse(GR.getSPIRVTypeID(I32Ty))
5513 .addUse(NewImageReg)
5514 .constrainAllUses(TII, TRI, RBI);
5515
5516 BuildMI(*I.getParent(), I, I.getDebugLoc(),
5517 TII.get(SPIRV::OpCompositeConstruct))
5518 .addDef(ResVReg)
5519 .addUse(GR.getSPIRVTypeID(ResType))
5520 .addUse(SizeReg)
5521 .addUse(SamplesReg)
5522 .constrainAllUses(TII, TRI, RBI);
5523
5524 return true;
5525}
5526
5527bool SPIRVInstructionSelector::selectCalculateLodIntrinsic(
5528 Register &ResVReg, SPIRVTypeInst ResType, MachineInstr &I) const {
5529 Register ImageReg = I.getOperand(2).getReg();
5530 Register SamplerReg = I.getOperand(3).getReg();
5531 Register CoordinateReg = I.getOperand(4).getReg();
5532
5533 auto *ImageDef = dyn_cast<GIntrinsic>(getVRegDef(*MRI, ImageReg));
5534 if (!ImageDef)
5535 return false;
5536 Register NewImageReg = MRI->createVirtualRegister(MRI->getRegClass(ImageReg));
5537 if (!loadHandleBeforePosition(NewImageReg, GR.getSPIRVTypeForVReg(ImageReg),
5538 *ImageDef, I)) {
5539 return false;
5540 }
5541
5542 auto *SamplerDef = dyn_cast<GIntrinsic>(getVRegDef(*MRI, SamplerReg));
5543 if (!SamplerDef)
5544 return false;
5545 Register NewSamplerReg =
5546 MRI->createVirtualRegister(MRI->getRegClass(SamplerReg));
5547 if (!loadHandleBeforePosition(
5548 NewSamplerReg, GR.getSPIRVTypeForVReg(SamplerReg), *SamplerDef, I)) {
5549 return false;
5550 }
5551
5552 MachineIRBuilder MIRBuilder(I);
5553 SPIRVTypeInst SampledImageType = GR.getOrCreateOpTypeSampledImage(
5554 GR.getSPIRVTypeForVReg(ImageReg), MIRBuilder);
5555 Register SampledImageReg =
5556 MRI->createVirtualRegister(GR.getRegClass(SampledImageType));
5557
5558 BuildMI(*I.getParent(), I, I.getDebugLoc(), TII.get(SPIRV::OpSampledImage))
5559 .addDef(SampledImageReg)
5560 .addUse(GR.getSPIRVTypeID(SampledImageType))
5561 .addUse(NewImageReg)
5562 .addUse(NewSamplerReg)
5563 .constrainAllUses(TII, TRI, RBI);
5564
5565 SPIRVTypeInst Vec2Ty = GR.getOrCreateSPIRVVectorType(ResType, 2, I, TII);
5566 Register QueryResultReg = MRI->createVirtualRegister(GR.getRegClass(Vec2Ty));
5567
5568 BuildMI(*I.getParent(), I, I.getDebugLoc(), TII.get(SPIRV::OpImageQueryLod))
5569 .addDef(QueryResultReg)
5570 .addUse(GR.getSPIRVTypeID(Vec2Ty))
5571 .addUse(SampledImageReg)
5572 .addUse(CoordinateReg)
5573 .constrainAllUses(TII, TRI, RBI);
5574
5575 unsigned ExtractedIndex =
5576 cast<GIntrinsic>(I).getIntrinsicID() ==
5577 Intrinsic::spv_resource_calculate_lod_unclamped
5578 ? 1
5579 : 0;
5580
5581 MachineInstrBuilder MIB = BuildMI(*I.getParent(), I, I.getDebugLoc(),
5582 TII.get(SPIRV::OpCompositeExtract))
5583 .addDef(ResVReg)
5584 .addUse(GR.getSPIRVTypeID(ResType))
5585 .addUse(QueryResultReg)
5586 .addImm(ExtractedIndex);
5587
5588 MIB.constrainAllUses(TII, TRI, RBI);
5589 return true;
5590}
5591
5592bool SPIRVInstructionSelector::selectSampleBasicIntrinsic(
5593 Register &ResVReg, SPIRVTypeInst ResType, MachineInstr &I) const {
5594 Register ImageReg = I.getOperand(2).getReg();
5595 Register SamplerReg = I.getOperand(3).getReg();
5596 Register CoordinateReg = I.getOperand(4).getReg();
5597 ImageOperands ImOps;
5598 if (I.getNumOperands() > 5)
5599 ImOps.Offset = I.getOperand(5).getReg();
5600 if (I.getNumOperands() > 6)
5601 ImOps.MinLod = I.getOperand(6).getReg();
5602 return generateSampleImage(ResVReg, ResType, ImageReg, SamplerReg,
5603 CoordinateReg, ImOps, I.getDebugLoc(), I);
5604}
5605
5606bool SPIRVInstructionSelector::selectSampleBiasIntrinsic(
5607 Register &ResVReg, SPIRVTypeInst ResType, MachineInstr &I) const {
5608 Register ImageReg = I.getOperand(2).getReg();
5609 Register SamplerReg = I.getOperand(3).getReg();
5610 Register CoordinateReg = I.getOperand(4).getReg();
5611 ImageOperands ImOps;
5612 ImOps.Bias = I.getOperand(5).getReg();
5613 if (I.getNumOperands() > 6)
5614 ImOps.Offset = I.getOperand(6).getReg();
5615 if (I.getNumOperands() > 7)
5616 ImOps.MinLod = I.getOperand(7).getReg();
5617 return generateSampleImage(ResVReg, ResType, ImageReg, SamplerReg,
5618 CoordinateReg, ImOps, I.getDebugLoc(), I);
5619}
5620
5621bool SPIRVInstructionSelector::selectSampleGradIntrinsic(
5622 Register &ResVReg, SPIRVTypeInst ResType, MachineInstr &I) const {
5623 Register ImageReg = I.getOperand(2).getReg();
5624 Register SamplerReg = I.getOperand(3).getReg();
5625 Register CoordinateReg = I.getOperand(4).getReg();
5626 ImageOperands ImOps;
5627 ImOps.GradX = I.getOperand(5).getReg();
5628 ImOps.GradY = I.getOperand(6).getReg();
5629 if (I.getNumOperands() > 7)
5630 ImOps.Offset = I.getOperand(7).getReg();
5631 if (I.getNumOperands() > 8)
5632 ImOps.MinLod = I.getOperand(8).getReg();
5633 return generateSampleImage(ResVReg, ResType, ImageReg, SamplerReg,
5634 CoordinateReg, ImOps, I.getDebugLoc(), I);
5635}
5636
5637bool SPIRVInstructionSelector::selectSampleLevelIntrinsic(
5638 Register &ResVReg, SPIRVTypeInst ResType, MachineInstr &I) const {
5639 Register ImageReg = I.getOperand(2).getReg();
5640 Register SamplerReg = I.getOperand(3).getReg();
5641 Register CoordinateReg = I.getOperand(4).getReg();
5642 ImageOperands ImOps;
5643 ImOps.Lod = I.getOperand(5).getReg();
5644 if (I.getNumOperands() > 6)
5645 ImOps.Offset = I.getOperand(6).getReg();
5646 return generateSampleImage(ResVReg, ResType, ImageReg, SamplerReg,
5647 CoordinateReg, ImOps, I.getDebugLoc(), I);
5648}
5649
5650bool SPIRVInstructionSelector::selectSampleCmpIntrinsic(Register &ResVReg,
5651 SPIRVTypeInst ResType,
5652 MachineInstr &I) const {
5653 Register ImageReg = I.getOperand(2).getReg();
5654 Register SamplerReg = I.getOperand(3).getReg();
5655 Register CoordinateReg = I.getOperand(4).getReg();
5656 ImageOperands ImOps;
5657 ImOps.Compare = I.getOperand(5).getReg();
5658 if (I.getNumOperands() > 6)
5659 ImOps.Offset = I.getOperand(6).getReg();
5660 if (I.getNumOperands() > 7)
5661 ImOps.MinLod = I.getOperand(7).getReg();
5662 return generateSampleImage(ResVReg, ResType, ImageReg, SamplerReg,
5663 CoordinateReg, ImOps, I.getDebugLoc(), I);
5664}
5665
5666bool SPIRVInstructionSelector::selectLoadLevelIntrinsic(Register &ResVReg,
5667 SPIRVTypeInst ResType,
5668 MachineInstr &I) const {
5669 Register ImageReg = I.getOperand(2).getReg();
5670 Register CoordinateReg = I.getOperand(3).getReg();
5671 Register LodReg = I.getOperand(4).getReg();
5672
5673 ImageOperands ImOps;
5674 ImOps.Lod = LodReg;
5675 if (I.getNumOperands() > 5)
5676 ImOps.Offset = I.getOperand(5).getReg();
5677
5678 auto *ImageDef = dyn_cast<GIntrinsic>(getVRegDef(*MRI, ImageReg));
5679 if (!ImageDef)
5680 return false;
5681
5682 Register NewImageReg = MRI->createVirtualRegister(MRI->getRegClass(ImageReg));
5683 if (!loadHandleBeforePosition(NewImageReg, GR.getSPIRVTypeForVReg(ImageReg),
5684 *ImageDef, I)) {
5685 return false;
5686 }
5687
5688 return generateImageReadOrFetch(ResVReg, ResType, NewImageReg, CoordinateReg,
5689 I.getDebugLoc(), I, &ImOps);
5690}
5691
5692bool SPIRVInstructionSelector::selectSampleCmpLevelZeroIntrinsic(
5693 Register &ResVReg, SPIRVTypeInst ResType, MachineInstr &I) const {
5694 Register ImageReg = I.getOperand(2).getReg();
5695 Register SamplerReg = I.getOperand(3).getReg();
5696 Register CoordinateReg = I.getOperand(4).getReg();
5697 ImageOperands ImOps;
5698 ImOps.Compare = I.getOperand(5).getReg();
5699 if (I.getNumOperands() > 6)
5700 ImOps.Offset = I.getOperand(6).getReg();
5701 SPIRVTypeInst FloatTy = GR.getOrCreateSPIRVFloatType(32, I, TII);
5702 ImOps.Lod = GR.getOrCreateConstFP(APFloat(0.0f), I, FloatTy, TII);
5703 return generateSampleImage(ResVReg, ResType, ImageReg, SamplerReg,
5704 CoordinateReg, ImOps, I.getDebugLoc(), I);
5705}
5706
5707bool SPIRVInstructionSelector::selectGatherIntrinsic(Register &ResVReg,
5708 SPIRVTypeInst ResType,
5709 MachineInstr &I) const {
5710 Register ImageReg = I.getOperand(2).getReg();
5711 Register SamplerReg = I.getOperand(3).getReg();
5712 Register CoordinateReg = I.getOperand(4).getReg();
5713 SPIRVTypeInst ImageType = GR.getSPIRVTypeForVReg(ImageReg);
5714 assert(ImageType && ImageType->getOpcode() == SPIRV::OpTypeImage &&
5715 "ImageReg is not an image type.");
5716
5717 Register ComponentOrCompareReg;
5718 Register OffsetReg;
5719
5720 ComponentOrCompareReg = I.getOperand(5).getReg();
5721 OffsetReg = I.getOperand(6).getReg();
5722 auto *ImageDef = cast<GIntrinsic>(getVRegDef(*MRI, ImageReg));
5723 Register NewImageReg = MRI->createVirtualRegister(MRI->getRegClass(ImageReg));
5724 if (!loadHandleBeforePosition(NewImageReg, ImageType, *ImageDef, I)) {
5725 return false;
5726 }
5727
5728 auto Dim = static_cast<SPIRV::Dim::Dim>(ImageType->getOperand(2).getImm());
5729 if (Dim != SPIRV::Dim::DIM_2D && Dim != SPIRV::Dim::DIM_Cube &&
5730 Dim != SPIRV::Dim::DIM_Rect) {
5731 I.emitGenericError(
5732 "Gather operations are only supported for 2D, Cube, and Rect images.");
5733 return false;
5734 }
5735
5736 auto *SamplerDef = cast<GIntrinsic>(getVRegDef(*MRI, SamplerReg));
5737 Register NewSamplerReg =
5738 MRI->createVirtualRegister(MRI->getRegClass(SamplerReg));
5739 if (!loadHandleBeforePosition(
5740 NewSamplerReg, GR.getSPIRVTypeForVReg(SamplerReg), *SamplerDef, I)) {
5741 return false;
5742 }
5743
5744 MachineIRBuilder MIRBuilder(I);
5745 SPIRVTypeInst SampledImageType =
5746 GR.getOrCreateOpTypeSampledImage(ImageType, MIRBuilder);
5747 Register SampledImageReg =
5748 MRI->createVirtualRegister(GR.getRegClass(SampledImageType));
5749
5750 BuildMI(*I.getParent(), I, I.getDebugLoc(), TII.get(SPIRV::OpSampledImage))
5751 .addDef(SampledImageReg)
5752 .addUse(GR.getSPIRVTypeID(SampledImageType))
5753 .addUse(NewImageReg)
5754 .addUse(NewSamplerReg)
5755 .constrainAllUses(TII, TRI, RBI);
5756
5757 auto IntrId = cast<GIntrinsic>(I).getIntrinsicID();
5758 bool IsGatherCmp = IntrId == Intrinsic::spv_resource_gather_cmp;
5759 unsigned Opcode =
5760 IsGatherCmp ? SPIRV::OpImageDrefGather : SPIRV::OpImageGather;
5761
5762 auto MIB = BuildMI(*I.getParent(), I, I.getDebugLoc(), TII.get(Opcode))
5763 .addDef(ResVReg)
5764 .addUse(GR.getSPIRVTypeID(ResType))
5765 .addUse(SampledImageReg)
5766 .addUse(CoordinateReg)
5767 .addUse(ComponentOrCompareReg);
5768
5769 uint32_t ImageOperands = 0;
5770 if (OffsetReg && !isScalarOrVectorIntConstantZero(OffsetReg)) {
5771 if (Dim == SPIRV::Dim::DIM_Cube) {
5772 I.emitGenericError(
5773 "Gather operations with offset are not supported for Cube images.");
5774 return false;
5775 }
5776 if (isConstReg(MRI, OffsetReg))
5777 ImageOperands |= SPIRV::ImageOperand::ConstOffset;
5778 else {
5779 ImageOperands |= SPIRV::ImageOperand::Offset;
5780 }
5781 }
5782
5783 if (ImageOperands != 0) {
5784 MIB.addImm(ImageOperands);
5785 if (ImageOperands &
5786 (SPIRV::ImageOperand::ConstOffset | SPIRV::ImageOperand::Offset))
5787 MIB.addUse(OffsetReg);
5788 }
5789
5790 MIB.constrainAllUses(TII, TRI, RBI);
5791 return true;
5792}
5793
5794bool SPIRVInstructionSelector::generateImageReadOrFetch(
5795 Register &ResVReg, SPIRVTypeInst ResType, Register ImageReg,
5796 Register IdxReg, DebugLoc Loc, MachineInstr &Pos,
5797 const ImageOperands *ImOps) const {
5798 SPIRVTypeInst ImageType = GR.getSPIRVTypeForVReg(ImageReg);
5799 assert(ImageType && ImageType->getOpcode() == SPIRV::OpTypeImage &&
5800 "ImageReg is not an image type.");
5801
5802 bool IsSignedInteger =
5803 sampledTypeIsSignedInteger(GR.getTypeForSPIRVType(ImageType));
5804 // Check if the "sampled" operand of the image type is 1.
5805 // https://registry.khronos.org/SPIR-V/specs/unified1/SPIRV.html#OpImageFetch
5806 auto SampledOp = ImageType->getOperand(6);
5807 bool IsFetch = (SampledOp.getImm() == 1);
5808
5809 auto AddOperands = [&](MachineInstrBuilder &MIB) {
5810 uint32_t ImageOperandsMask = 0;
5811 if (IsSignedInteger)
5812 ImageOperandsMask |= 0x1000; // SignExtend
5813
5814 if (IsFetch && ImOps) {
5815 if (ImOps->Lod)
5816 ImageOperandsMask |= SPIRV::ImageOperand::Lod;
5817 if (ImOps->Offset && !isScalarOrVectorIntConstantZero(*ImOps->Offset)) {
5818 if (isConstReg(MRI, *ImOps->Offset))
5819 ImageOperandsMask |= SPIRV::ImageOperand::ConstOffset;
5820 else
5821 ImageOperandsMask |= SPIRV::ImageOperand::Offset;
5822 }
5823 }
5824
5825 if (ImageOperandsMask != 0) {
5826 MIB.addImm(ImageOperandsMask);
5827 if (IsFetch && ImOps) {
5828 if (ImOps->Lod)
5829 MIB.addUse(*ImOps->Lod);
5830 if (ImOps->Offset &&
5831 (ImageOperandsMask &
5832 (SPIRV::ImageOperand::Offset | SPIRV::ImageOperand::ConstOffset)))
5833 MIB.addUse(*ImOps->Offset);
5834 }
5835 }
5836 };
5837
5838 uint64_t ResultSize = GR.getScalarOrVectorComponentCount(ResType);
5839 if (ResultSize == 4) {
5840 auto BMI =
5841 BuildMI(*Pos.getParent(), Pos, Loc,
5842 TII.get(IsFetch ? SPIRV::OpImageFetch : SPIRV::OpImageRead))
5843 .addDef(ResVReg)
5844 .addUse(GR.getSPIRVTypeID(ResType))
5845 .addUse(ImageReg)
5846 .addUse(IdxReg);
5847
5848 AddOperands(BMI);
5849 BMI.constrainAllUses(TII, TRI, RBI);
5850 return true;
5851 }
5852
5853 SPIRVTypeInst ReadType = widenTypeToVec4(ResType, Pos);
5854 Register ReadReg = MRI->createVirtualRegister(GR.getRegClass(ReadType));
5855 auto BMI =
5856 BuildMI(*Pos.getParent(), Pos, Loc,
5857 TII.get(IsFetch ? SPIRV::OpImageFetch : SPIRV::OpImageRead))
5858 .addDef(ReadReg)
5859 .addUse(GR.getSPIRVTypeID(ReadType))
5860 .addUse(ImageReg)
5861 .addUse(IdxReg);
5862 AddOperands(BMI);
5863 BMI.constrainAllUses(TII, TRI, RBI);
5864
5865 if (ResultSize == 1) {
5866 BuildMI(*Pos.getParent(), Pos, Loc, TII.get(SPIRV::OpCompositeExtract))
5867 .addDef(ResVReg)
5868 .addUse(GR.getSPIRVTypeID(ResType))
5869 .addUse(ReadReg)
5870 .addImm(0)
5871 .constrainAllUses(TII, TRI, RBI);
5872 return true;
5873 }
5874 return extractSubvector(ResVReg, ResType, ReadReg, Pos);
5875}
5876
5877bool SPIRVInstructionSelector::selectResourceGetPointer(Register &ResVReg,
5878 SPIRVTypeInst ResType,
5879 MachineInstr &I) const {
5880 Register ResourcePtr = I.getOperand(2).getReg();
5881 SPIRVTypeInst RegType = GR.getSPIRVTypeForVReg(ResourcePtr, I.getMF());
5882 if (RegType->getOpcode() == SPIRV::OpTypeImage) {
5883 // For texel buffers, the index into the image is part of the OpImageRead or
5884 // OpImageWrite instructions. So we will do nothing in this case. This
5885 // intrinsic will be combined with the load or store when selecting the load
5886 // or store.
5887 return true;
5888 }
5889
5890 assert(ResType->getOpcode() == SPIRV::OpTypePointer);
5891 MachineIRBuilder MIRBuilder(I);
5892
5893 Register IndexReg = I.getOperand(3).getReg();
5894 Register ZeroReg =
5895 buildZerosVal(GR.getOrCreateSPIRVIntegerType(32, I, TII), I);
5896 BuildMI(*I.getParent(), I, I.getDebugLoc(), TII.get(SPIRV::OpAccessChain))
5897 .addDef(ResVReg)
5898 .addUse(GR.getSPIRVTypeID(ResType))
5899 .addUse(ResourcePtr)
5900 .addUse(ZeroReg)
5901 .addUse(IndexReg)
5902 .constrainAllUses(TII, TRI, RBI);
5903 return true;
5904}
5905
5906bool SPIRVInstructionSelector::selectPushConstantGetPointer(
5907 Register &ResVReg, SPIRVTypeInst ResType, MachineInstr &I) const {
5908 MRI->replaceRegWith(ResVReg, I.getOperand(2).getReg());
5909 return true;
5910}
5911
5912bool SPIRVInstructionSelector::selectResourceNonUniformIndex(
5913 Register &ResVReg, SPIRVTypeInst ResType, MachineInstr &I) const {
5914 Register ObjReg = I.getOperand(2).getReg();
5915 if (!BuildCOPY(ResVReg, ObjReg, I))
5916 return false;
5917
5918 buildOpDecorate(ResVReg, I, TII, SPIRV::Decoration::NonUniformEXT, {});
5919 // Check for the registers that use the index marked as non-uniform
5920 // and recursively mark them as non-uniform.
5921 // Per the spec, it's necessary that the final argument used for
5922 // load/store/sample/atomic must be decorated, so we need to propagate the
5923 // decoration through access chains and copies.
5924 // https://docs.vulkan.org/samples/latest/samples/extensions/descriptor_indexing/README.html#_when_to_use_non_uniform_indexing_qualifier
5925 decorateUsesAsNonUniform(ResVReg);
5926 return true;
5927}
5928
5929void SPIRVInstructionSelector::decorateUsesAsNonUniform(
5930 Register &NonUniformReg) const {
5931 llvm::SmallVector<Register> WorkList = {NonUniformReg};
5932 while (WorkList.size() > 0) {
5933 Register CurrentReg = WorkList.back();
5934 WorkList.pop_back();
5935
5936 bool IsDecorated = false;
5937 for (MachineInstr &Use : MRI->use_instructions(CurrentReg)) {
5938 if (Use.getOpcode() == SPIRV::OpDecorate &&
5939 Use.getOperand(1).getImm() == SPIRV::Decoration::NonUniformEXT) {
5940 IsDecorated = true;
5941 continue;
5942 }
5943 // Check if the instruction has the result register and add it to the
5944 // worklist.
5945 if (Use.getOperand(0).isReg() && Use.getOperand(0).isDef()) {
5946 Register ResultReg = Use.getOperand(0).getReg();
5947 if (ResultReg == CurrentReg)
5948 continue;
5949 WorkList.push_back(ResultReg);
5950 }
5951 }
5952
5953 if (!IsDecorated) {
5954 buildOpDecorate(CurrentReg, *MRI->getVRegDef(CurrentReg), TII,
5955 SPIRV::Decoration::NonUniformEXT, {});
5956 }
5957 }
5958}
5959
5960bool SPIRVInstructionSelector::extractSubvector(
5961 Register &ResVReg, SPIRVTypeInst ResType, Register &ReadReg,
5962 MachineInstr &InsertionPoint) const {
5963 SPIRVTypeInst InputType = GR.getResultType(ReadReg);
5964 [[maybe_unused]] uint64_t InputSize =
5965 GR.getScalarOrVectorComponentCount(InputType);
5966 uint64_t ResultSize = GR.getScalarOrVectorComponentCount(ResType);
5967 assert(InputSize > 1 && "The input must be a vector.");
5968 assert(ResultSize > 1 && "The result must be a vector.");
5969 assert(ResultSize < InputSize &&
5970 "Cannot extract more element than there are in the input.");
5971 SmallVector<Register> ComponentRegisters;
5972 SPIRVTypeInst ScalarType = GR.getScalarOrVectorComponentType(ResType);
5973 const TargetRegisterClass *ScalarRegClass = GR.getRegClass(ScalarType);
5974 for (uint64_t I = 0; I < ResultSize; I++) {
5975 Register ComponentReg = MRI->createVirtualRegister(ScalarRegClass);
5976 BuildMI(*InsertionPoint.getParent(), InsertionPoint,
5977 InsertionPoint.getDebugLoc(), TII.get(SPIRV::OpCompositeExtract))
5978 .addDef(ComponentReg)
5979 .addUse(ScalarType->getOperand(0).getReg())
5980 .addUse(ReadReg)
5981 .addImm(I)
5982 .constrainAllUses(TII, TRI, RBI);
5983 ComponentRegisters.emplace_back(ComponentReg);
5984 }
5985
5986 MachineInstrBuilder MIB = BuildMI(*InsertionPoint.getParent(), InsertionPoint,
5987 InsertionPoint.getDebugLoc(),
5988 TII.get(SPIRV::OpCompositeConstruct))
5989 .addDef(ResVReg)
5990 .addUse(GR.getSPIRVTypeID(ResType));
5991
5992 for (Register ComponentReg : ComponentRegisters)
5993 MIB.addUse(ComponentReg);
5994 MIB.constrainAllUses(TII, TRI, RBI);
5995 return true;
5996}
5997
5998bool SPIRVInstructionSelector::selectImageWriteIntrinsic(
5999 MachineInstr &I) const {
6000 // If the load of the image is in a different basic block, then
6001 // this will generate invalid code. A proper solution is to move
6002 // the OpLoad from selectHandleFromBinding here. However, to do
6003 // that we will need to change the return type of the intrinsic.
6004 // We will do that when we can, but for now trying to move forward with other
6005 // issues.
6006 Register ImageReg = I.getOperand(1).getReg();
6007 auto *ImageDef = cast<GIntrinsic>(getVRegDef(*MRI, ImageReg));
6008 Register NewImageReg = MRI->createVirtualRegister(MRI->getRegClass(ImageReg));
6009 if (!loadHandleBeforePosition(NewImageReg, GR.getSPIRVTypeForVReg(ImageReg),
6010 *ImageDef, I)) {
6011 return false;
6012 }
6013
6014 Register CoordinateReg = I.getOperand(2).getReg();
6015 Register DataReg = I.getOperand(3).getReg();
6016 assert(GR.getResultType(DataReg)->getOpcode() == SPIRV::OpTypeVector);
6017 assert(GR.getScalarOrVectorComponentCount(GR.getResultType(DataReg)) == 4);
6018 BuildMI(*I.getParent(), I, I.getDebugLoc(), TII.get(SPIRV::OpImageWrite))
6019 .addUse(NewImageReg)
6020 .addUse(CoordinateReg)
6021 .addUse(DataReg)
6022 .constrainAllUses(TII, TRI, RBI);
6023 return true;
6024}
6025
6026Register SPIRVInstructionSelector::buildPointerToResource(
6027 SPIRVTypeInst SpirvResType, SPIRV::StorageClass::StorageClass SC,
6028 uint32_t Set, uint32_t Binding, uint32_t ArraySize, Register IndexReg,
6029 StringRef Name, MachineIRBuilder MIRBuilder) const {
6030 const Type *ResType = GR.getTypeForSPIRVType(SpirvResType);
6031 if (ArraySize == 1) {
6032 SPIRVTypeInst PtrType =
6033 GR.getOrCreateSPIRVPointerType(ResType, MIRBuilder, SC);
6034 assert(GR.getPointeeType(PtrType) == SpirvResType &&
6035 "SpirvResType did not have an explicit layout.");
6036 return GR.getOrCreateGlobalVariableWithBinding(PtrType, Set, Binding, Name,
6037 MIRBuilder);
6038 }
6039
6040 const Type *VarType = ArrayType::get(const_cast<Type *>(ResType), ArraySize);
6041 SPIRVTypeInst VarPointerType =
6042 GR.getOrCreateSPIRVPointerType(VarType, MIRBuilder, SC);
6043 Register VarReg = GR.getOrCreateGlobalVariableWithBinding(
6044 VarPointerType, Set, Binding, Name, MIRBuilder);
6045
6046 SPIRVTypeInst ResPointerType =
6047 GR.getOrCreateSPIRVPointerType(ResType, MIRBuilder, SC);
6048 Register AcReg = MRI->createVirtualRegister(GR.getRegClass(ResPointerType));
6049
6050 MIRBuilder.buildInstr(SPIRV::OpAccessChain)
6051 .addDef(AcReg)
6052 .addUse(GR.getSPIRVTypeID(ResPointerType))
6053 .addUse(VarReg)
6054 .addUse(IndexReg);
6055
6056 return AcReg;
6057}
6058
6059bool SPIRVInstructionSelector::selectFirstBitSet16(
6060 Register ResVReg, SPIRVTypeInst ResType, MachineInstr &I,
6061 unsigned ExtendOpcode, unsigned BitSetOpcode) const {
6062 Register ExtReg = MRI->createVirtualRegister(GR.getRegClass(ResType));
6063 if (!selectOpWithSrcs(ExtReg, ResType, I, {I.getOperand(2).getReg()},
6064 ExtendOpcode))
6065 return false;
6066
6067 return selectFirstBitSet32(ResVReg, ResType, I, ExtReg, BitSetOpcode);
6068}
6069
6070bool SPIRVInstructionSelector::selectFirstBitSet32(
6071 Register ResVReg, SPIRVTypeInst ResType, MachineInstr &I, Register SrcReg,
6072 unsigned BitSetOpcode) const {
6073 BuildMI(*I.getParent(), I, I.getDebugLoc(), TII.get(SPIRV::OpExtInst))
6074 .addDef(ResVReg)
6075 .addUse(GR.getSPIRVTypeID(ResType))
6076 .addImm(static_cast<uint32_t>(SPIRV::InstructionSet::GLSL_std_450))
6077 .addImm(BitSetOpcode)
6078 .addUse(SrcReg)
6079 .constrainAllUses(TII, TRI, RBI);
6080 return true;
6081}
6082
6083bool SPIRVInstructionSelector::selectFirstBitSet64(
6084 Register ResVReg, SPIRVTypeInst ResType, MachineInstr &I, Register SrcReg,
6085 unsigned BitSetOpcode, bool SwapPrimarySide) const {
6086 unsigned ComponentCount = GR.getScalarOrVectorComponentCount(ResType);
6087 SPIRVTypeInst BaseType = GR.retrieveScalarOrVectorIntType(ResType);
6088 bool ZeroAsNull = !STI.isShader();
6089 Register ConstIntZero =
6090 GR.getOrCreateConstInt(0, I, BaseType, TII, ZeroAsNull);
6091 Register ConstIntOne =
6092 GR.getOrCreateConstInt(1, I, BaseType, TII, ZeroAsNull);
6093
6094 // SPIRV doesn't support vectors with more than 4 components. Since the
6095 // algoritm below converts i64 -> i32x2 and i64x4 -> i32x8 it can only
6096 // operate on vectors with 2 or less components. When largers vectors are
6097 // seen. Split them, recurse, then recombine them.
6098 if (ComponentCount > 2) {
6099 auto Func = [this, SwapPrimarySide](Register ResVReg, SPIRVTypeInst ResType,
6100 MachineInstr &I, Register SrcReg,
6101 unsigned Opcode) -> bool {
6102 return this->selectFirstBitSet64(ResVReg, ResType, I, SrcReg, Opcode,
6103 SwapPrimarySide);
6104 };
6105
6106 return handle64BitOverflow(ResVReg, ResType, I, SrcReg, BitSetOpcode, Func);
6107 }
6108
6109 // 1. Split int64 into 2 pieces using a bitcast
6110 MachineIRBuilder MIRBuilder(I);
6111 SPIRVTypeInst PostCastType = GR.getOrCreateSPIRVVectorType(
6112 BaseType, 2 * ComponentCount, MIRBuilder, false);
6113 Register BitcastReg =
6114 MRI->createVirtualRegister(GR.getRegClass(PostCastType));
6115
6116 if (!selectOpWithSrcs(BitcastReg, PostCastType, I, {SrcReg},
6117 SPIRV::OpBitcast))
6118 return false;
6119
6120 // 2. Find the first set bit from the primary side for all the pieces in #1
6121 Register FBSReg = MRI->createVirtualRegister(GR.getRegClass(PostCastType));
6122 if (!selectFirstBitSet32(FBSReg, PostCastType, I, BitcastReg, BitSetOpcode))
6123 return false;
6124
6125 // 3. Split result vector into high bits and low bits
6126 Register HighReg = MRI->createVirtualRegister(GR.getRegClass(ResType));
6127 Register LowReg = MRI->createVirtualRegister(GR.getRegClass(ResType));
6128
6129 bool IsScalarRes = ResType->getOpcode() != SPIRV::OpTypeVector;
6130 if (IsScalarRes) {
6131 // if scalar do a vector extract
6132 if (!selectOpWithSrcs(HighReg, ResType, I, {FBSReg, ConstIntOne},
6133 SPIRV::OpVectorExtractDynamic))
6134 return false;
6135 if (!selectOpWithSrcs(LowReg, ResType, I, {FBSReg, ConstIntZero},
6136 SPIRV::OpVectorExtractDynamic))
6137 return false;
6138 } else {
6139 // if vector do a shufflevector
6140 auto MIB = BuildMI(*I.getParent(), I, I.getDebugLoc(),
6141 TII.get(SPIRV::OpVectorShuffle))
6142 .addDef(HighReg)
6143 .addUse(GR.getSPIRVTypeID(ResType))
6144 .addUse(FBSReg)
6145 // Per the spec, repeat the vector if only one vec is needed
6146 .addUse(FBSReg);
6147
6148 // high bits are stored in even natural indexes. Extract them from FBSReg
6149 for (unsigned J = 1; J < ComponentCount * 2; J += 2) {
6150 MIB.addImm(J);
6151 }
6152
6153 MIB.constrainAllUses(TII, TRI, RBI);
6154
6155 MIB = BuildMI(*I.getParent(), I, I.getDebugLoc(),
6156 TII.get(SPIRV::OpVectorShuffle))
6157 .addDef(LowReg)
6158 .addUse(GR.getSPIRVTypeID(ResType))
6159 .addUse(FBSReg)
6160 // Per the spec, repeat the vector if only one vec is needed
6161 .addUse(FBSReg);
6162
6163 // low bits are stored in odd natural indices. Extract them from FBSReg
6164 for (unsigned J = 0; J < ComponentCount * 2; J += 2) {
6165 MIB.addImm(J);
6166 }
6167 MIB.constrainAllUses(TII, TRI, RBI);
6168 }
6169
6170 // 4. Check the result. When primary bits == -1 use secondary, otherwise use
6171 // primary
6172 SPIRVTypeInst BoolType = GR.getOrCreateSPIRVBoolType(I, TII);
6173 Register NegOneReg;
6174 Register Reg0;
6175 Register Reg32;
6176 unsigned SelectOp;
6177 unsigned AddOp;
6178
6179 if (IsScalarRes) {
6180 NegOneReg =
6181 GR.getOrCreateConstInt((unsigned)-1, I, ResType, TII, ZeroAsNull);
6182 Reg0 = GR.getOrCreateConstInt(0, I, ResType, TII, ZeroAsNull);
6183 Reg32 = GR.getOrCreateConstInt(32, I, ResType, TII, ZeroAsNull);
6184 SelectOp = SPIRV::OpSelectSISCond;
6185 AddOp = SPIRV::OpIAddS;
6186 } else {
6187 BoolType = GR.getOrCreateSPIRVVectorType(BoolType, ComponentCount,
6188 MIRBuilder, false);
6189 NegOneReg =
6190 GR.getOrCreateConstVector((unsigned)-1, I, ResType, TII, ZeroAsNull);
6191 Reg0 = GR.getOrCreateConstVector(0, I, ResType, TII, ZeroAsNull);
6192 Reg32 = GR.getOrCreateConstVector(32, I, ResType, TII, ZeroAsNull);
6193 SelectOp = SPIRV::OpSelectVIVCond;
6194 AddOp = SPIRV::OpIAddV;
6195 }
6196
6197 Register PrimaryReg = HighReg;
6198 Register SecondaryReg = LowReg;
6199 Register RegPrimaryOffset = Reg32;
6200 Register RegSecondaryOffset = Reg0;
6201
6202 // By default the emitted opcodes check for the set bit from the MSB side.
6203 // Setting SwapPrimarySide checks the set bit from the LSB side
6204 if (SwapPrimarySide) {
6205 PrimaryReg = LowReg;
6206 SecondaryReg = HighReg;
6207 RegPrimaryOffset = Reg0;
6208 RegSecondaryOffset = Reg32;
6209 }
6210
6211 Register RegSecondaryHasVal =
6212 MRI->createVirtualRegister(GR.getRegClass(BoolType));
6213 if (!selectOpWithSrcs(RegSecondaryHasVal, BoolType, I,
6214 {SecondaryReg, NegOneReg}, SPIRV::OpINotEqual))
6215 return false;
6216
6217 Register RegPrimaryHasVal =
6218 MRI->createVirtualRegister(GR.getRegClass(BoolType));
6219 if (!selectOpWithSrcs(RegPrimaryHasVal, BoolType, I, {PrimaryReg, NegOneReg},
6220 SPIRV::OpINotEqual))
6221 return false;
6222
6223 // Pass 1: seed with secondary (lower-priority fallback)
6224 // ReturnBits = secondaryHasVal ? SecondaryBits : -1
6225 // Add = secondaryHasVal ? SecondaryOffset : 0
6226 Register RegReturnBits = MRI->createVirtualRegister(GR.getRegClass(ResType));
6227 if (!selectOpWithSrcs(RegReturnBits, ResType, I,
6228 {RegSecondaryHasVal, SecondaryReg, NegOneReg},
6229 SelectOp))
6230 return false;
6231
6232 Register RegAdd;
6233 if (SwapPrimarySide) {
6234 RegAdd = MRI->createVirtualRegister(GR.getRegClass(ResType));
6235 if (!selectOpWithSrcs(RegAdd, ResType, I,
6236 {RegSecondaryHasVal, RegSecondaryOffset, Reg0},
6237 SelectOp))
6238 return false;
6239 } else {
6240 RegAdd = Reg0;
6241 }
6242
6243 // Pass 2: override with primary (higher priority) if it has a valid result
6244 // ReturnBits2 = primaryHasVal ? PrimaryBits : ReturnBits
6245 // Add2 = primaryHasVal ? PrimaryOffset : Add
6246 Register RegReturnBits2 = MRI->createVirtualRegister(GR.getRegClass(ResType));
6247 if (!selectOpWithSrcs(RegReturnBits2, ResType, I,
6248 {RegPrimaryHasVal, PrimaryReg, RegReturnBits},
6249 SelectOp))
6250 return false;
6251
6252 Register RegAdd2 = MRI->createVirtualRegister(GR.getRegClass(ResType));
6253 if (!selectOpWithSrcs(RegAdd2, ResType, I,
6254 {RegPrimaryHasVal, RegPrimaryOffset, RegAdd}, SelectOp))
6255 return false;
6256
6257 return selectOpWithSrcs(ResVReg, ResType, I, {RegReturnBits2, RegAdd2},
6258 AddOp);
6259}
6260
6261bool SPIRVInstructionSelector::selectFirstBitHigh(Register ResVReg,
6262 SPIRVTypeInst ResType,
6263 MachineInstr &I,
6264 bool IsSigned) const {
6265 // FindUMsb and FindSMsb intrinsics only support 32 bit integers
6266 Register OpReg = I.getOperand(2).getReg();
6267 SPIRVTypeInst OpType = GR.getSPIRVTypeForVReg(OpReg);
6268 // zero or sign extend
6269 unsigned ExtendOpcode = IsSigned ? SPIRV::OpSConvert : SPIRV::OpUConvert;
6270 unsigned BitSetOpcode = IsSigned ? GL::FindSMsb : GL::FindUMsb;
6271
6272 switch (GR.getScalarOrVectorBitWidth(OpType)) {
6273 case 16:
6274 return selectFirstBitSet16(ResVReg, ResType, I, ExtendOpcode, BitSetOpcode);
6275 case 32:
6276 return selectFirstBitSet32(ResVReg, ResType, I, OpReg, BitSetOpcode);
6277 case 64:
6278 return selectFirstBitSet64(ResVReg, ResType, I, OpReg, BitSetOpcode,
6279 /*SwapPrimarySide=*/false);
6280 default:
6281 report_fatal_error(
6282 "spv_firstbituhigh and spv_firstbitshigh only support 16,32,64 bits.");
6283 }
6284}
6285
6286bool SPIRVInstructionSelector::selectFirstBitLow(Register ResVReg,
6287 SPIRVTypeInst ResType,
6288 MachineInstr &I) const {
6289 // FindILsb intrinsic only supports 32 bit integers
6290 Register OpReg = I.getOperand(2).getReg();
6291 SPIRVTypeInst OpType = GR.getSPIRVTypeForVReg(OpReg);
6292 // OpUConvert treats the operand bits as an unsigned i16 and zero extends it
6293 // to an unsigned i32. As this leaves all the least significant bits unchanged
6294 // so the first set bit from the LSB side doesn't change.
6295 unsigned ExtendOpcode = SPIRV::OpUConvert;
6296 unsigned BitSetOpcode = GL::FindILsb;
6297
6298 switch (GR.getScalarOrVectorBitWidth(OpType)) {
6299 case 16:
6300 return selectFirstBitSet16(ResVReg, ResType, I, ExtendOpcode, BitSetOpcode);
6301 case 32:
6302 return selectFirstBitSet32(ResVReg, ResType, I, OpReg, BitSetOpcode);
6303 case 64:
6304 return selectFirstBitSet64(ResVReg, ResType, I, OpReg, BitSetOpcode,
6305 /*SwapPrimarySide=*/true);
6306 default:
6307 report_fatal_error("spv_firstbitlow only supports 16,32,64 bits.");
6308 }
6309}
6310
6311bool SPIRVInstructionSelector::selectAllocaArray(Register ResVReg,
6312 SPIRVTypeInst ResType,
6313 MachineInstr &I) const {
6314 // there was an allocation size parameter to the allocation instruction
6315 // that is not 1
6316 MachineBasicBlock &BB = *I.getParent();
6317 BuildMI(BB, I, I.getDebugLoc(), TII.get(SPIRV::OpVariableLengthArrayINTEL))
6318 .addDef(ResVReg)
6319 .addUse(GR.getSPIRVTypeID(ResType))
6320 .addUse(I.getOperand(2).getReg())
6321 .constrainAllUses(TII, TRI, RBI);
6322 if (!STI.isShader()) {
6323 unsigned Alignment = I.getOperand(3).getImm();
6324 buildOpDecorate(ResVReg, I, TII, SPIRV::Decoration::Alignment, {Alignment});
6325 }
6326 return true;
6327}
6328
6329bool SPIRVInstructionSelector::selectFrameIndex(Register ResVReg,
6330 SPIRVTypeInst ResType,
6331 MachineInstr &I) const {
6332 // Change order of instructions if needed: all OpVariable instructions in a
6333 // function must be the first instructions in the first block
6334 auto It = getOpVariableMBBIt(*I.getMF());
6335 BuildMI(*It->getParent(), It, It->getDebugLoc(), TII.get(SPIRV::OpVariable))
6336 .addDef(ResVReg)
6337 .addUse(GR.getSPIRVTypeID(ResType))
6338 .addImm(static_cast<uint32_t>(SPIRV::StorageClass::Function))
6339 .constrainAllUses(TII, TRI, RBI);
6340 if (!STI.isShader()) {
6341 unsigned Alignment = I.getOperand(2).getImm();
6342 buildOpDecorate(ResVReg, *It, TII, SPIRV::Decoration::Alignment,
6343 {Alignment});
6344 }
6345 return true;
6346}
6347
6348bool SPIRVInstructionSelector::selectBranch(MachineInstr &I) const {
6349 // InstructionSelector walks backwards through the instructions. We can use
6350 // both a G_BR and a G_BRCOND to create an OpBranchConditional. We hit G_BR
6351 // first, so can generate an OpBranchConditional here. If there is no
6352 // G_BRCOND, we just use OpBranch for a regular unconditional branch.
6353 const MachineInstr *PrevI = I.getPrevNode();
6354 MachineBasicBlock &MBB = *I.getParent();
6355 if (PrevI != nullptr && PrevI->getOpcode() == TargetOpcode::G_BRCOND) {
6356 BuildMI(MBB, I, I.getDebugLoc(), TII.get(SPIRV::OpBranchConditional))
6357 .addUse(PrevI->getOperand(0).getReg())
6358 .addMBB(PrevI->getOperand(1).getMBB())
6359 .addMBB(I.getOperand(0).getMBB())
6360 .constrainAllUses(TII, TRI, RBI);
6361 return true;
6362 }
6363 BuildMI(MBB, I, I.getDebugLoc(), TII.get(SPIRV::OpBranch))
6364 .addMBB(I.getOperand(0).getMBB())
6365 .constrainAllUses(TII, TRI, RBI);
6366 return true;
6367}
6368
6369bool SPIRVInstructionSelector::selectBranchCond(MachineInstr &I) const {
6370 // InstructionSelector walks backwards through the instructions. For an
6371 // explicit conditional branch with no fallthrough, we use both a G_BR and a
6372 // G_BRCOND to create an OpBranchConditional. We should hit G_BR first, and
6373 // generate the OpBranchConditional in selectBranch above.
6374 //
6375 // If an OpBranchConditional has been generated, we simply return, as the work
6376 // is alread done. If there is no OpBranchConditional, LLVM must be relying on
6377 // implicit fallthrough to the next basic block, so we need to create an
6378 // OpBranchConditional with an explicit "false" argument pointing to the next
6379 // basic block that LLVM would fall through to.
6380 const MachineInstr *NextI = I.getNextNode();
6381 // Check if this has already been successfully selected.
6382 if (NextI != nullptr && NextI->getOpcode() == SPIRV::OpBranchConditional)
6383 return true;
6384 // Must be relying on implicit block fallthrough, so generate an
6385 // OpBranchConditional with the "next" basic block as the "false" target.
6386 MachineBasicBlock &MBB = *I.getParent();
6387 unsigned NextMBBNum = MBB.getNextNode()->getNumber();
6388 MachineBasicBlock *NextMBB = I.getMF()->getBlockNumbered(NextMBBNum);
6389 BuildMI(MBB, I, I.getDebugLoc(), TII.get(SPIRV::OpBranchConditional))
6390 .addUse(I.getOperand(0).getReg())
6391 .addMBB(I.getOperand(1).getMBB())
6392 .addMBB(NextMBB)
6393 .constrainAllUses(TII, TRI, RBI);
6394 return true;
6395}
6396
6397bool SPIRVInstructionSelector::selectPhi(Register ResVReg,
6398 MachineInstr &I) const {
6399 auto MIB =
6400 BuildMI(*I.getParent(), I, I.getDebugLoc(), TII.get(TargetOpcode::PHI))
6401 .addDef(ResVReg);
6402 const unsigned NumOps = I.getNumOperands();
6403 for (unsigned i = 1; i < NumOps; i += 2) {
6404 MIB.addUse(I.getOperand(i + 0).getReg());
6405 MIB.addMBB(I.getOperand(i + 1).getMBB());
6406 }
6407 MIB.constrainAllUses(TII, TRI, RBI);
6408 return true;
6409}
6410
6411bool SPIRVInstructionSelector::selectGlobalValue(
6412 Register ResVReg, MachineInstr &I, const MachineInstr *Init) const {
6413 // FIXME: don't use MachineIRBuilder here, replace it with BuildMI.
6414 MachineIRBuilder MIRBuilder(I);
6415 const GlobalValue *GV = I.getOperand(1).getGlobal();
6416 Type *GVType = toTypedPointer(GR.getDeducedGlobalValueType(GV));
6417
6418 std::string GlobalIdent;
6419 if (!GV->hasName()) {
6420 unsigned &ID = UnnamedGlobalIDs[GV];
6421 if (ID == 0)
6422 ID = UnnamedGlobalIDs.size();
6423 GlobalIdent = "__unnamed_" + Twine(ID).str();
6424 } else {
6425 GlobalIdent = GV->getName();
6426 }
6427
6428 // Behaviour of functions as operands depends on availability of the
6429 // corresponding extension (SPV_INTEL_function_pointers):
6430 // - If there is an extension to operate with functions as operands:
6431 // We create a proper constant operand and evaluate a correct type for a
6432 // function pointer.
6433 // - Without the required extension:
6434 // We have functions as operands in tests with blocks of instruction e.g. in
6435 // transcoding/global_block.ll. These operands are not used and should be
6436 // substituted by zero constants. Their type is expected to be always
6437 // OpTypePointer Function %uchar.
6438 if (isa<Function>(GV)) {
6439 const Constant *ConstVal = GV;
6440 MachineBasicBlock &BB = *I.getParent();
6441 Register NewReg = GR.find(ConstVal, GR.CurMF);
6442 if (!NewReg.isValid()) {
6443 const Function *GVFun =
6444 STI.canUseExtension(SPIRV::Extension::SPV_INTEL_function_pointers)
6445 ? dyn_cast<Function>(GV)
6446 : nullptr;
6447 SPIRVTypeInst ResType = GR.getOrCreateSPIRVPointerType(
6448 GVType, I,
6449 GVFun ? SPIRV::StorageClass::CodeSectionINTEL
6450 : addressSpaceToStorageClass(GV->getAddressSpace(), STI));
6451 if (GVFun) {
6452 // References to a function via function pointers generate virtual
6453 // registers without a definition. We will resolve it later, during
6454 // module analysis stage.
6455 Register ResTypeReg = GR.getSPIRVTypeID(ResType);
6456 MachineRegisterInfo *MRI = MIRBuilder.getMRI();
6457 Register FuncVReg =
6458 MRI->createGenericVirtualRegister(GR.getRegType(ResType));
6459 MRI->setRegClass(FuncVReg, &SPIRV::pIDRegClass);
6460 GR.assignSPIRVTypeToVReg(ResType, FuncVReg, *GR.CurMF);
6461 MachineInstrBuilder MIB1 =
6462 BuildMI(BB, I, I.getDebugLoc(), TII.get(SPIRV::OpUndef))
6463 .addDef(FuncVReg)
6464 .addUse(ResTypeReg);
6465 MachineInstrBuilder MIB2 =
6466 BuildMI(BB, I, I.getDebugLoc(),
6467 TII.get(SPIRV::OpConstantFunctionPointerINTEL))
6468 .addDef(ResVReg)
6469 .addUse(ResTypeReg)
6470 .addUse(FuncVReg);
6471 GR.add(ConstVal, MIB2);
6472 // mapping the function pointer to the used Function
6473 GR.recordFunctionPointer(&MIB2.getInstr()->getOperand(2), GVFun);
6474 GR.assignSPIRVTypeToVReg(ResType, ResVReg, *GR.CurMF);
6475 MIB1.constrainAllUses(TII, TRI, RBI);
6476 MIB2.constrainAllUses(TII, TRI, RBI);
6477 return true;
6478 }
6479 MachineInstrBuilder MIB3 =
6480 BuildMI(BB, I, I.getDebugLoc(), TII.get(SPIRV::OpConstantNull))
6481 .addDef(ResVReg)
6482 .addUse(GR.getSPIRVTypeID(ResType));
6483 GR.add(ConstVal, MIB3);
6484 MIB3.constrainAllUses(TII, TRI, RBI);
6485 return true;
6486 }
6487 assert(NewReg != ResVReg);
6488 return BuildCOPY(ResVReg, NewReg, I);
6489 }
6490 auto GlobalVar = cast<GlobalVariable>(GV);
6491 assert(GlobalVar->getName() != "llvm.global.annotations");
6492
6493 // Skip empty declaration for GVs with initializers till we get the decl with
6494 // passed initializer.
6495 if (hasInitializer(GlobalVar) && !Init)
6496 return true;
6497
6498 const std::optional<SPIRV::LinkageType::LinkageType> LnkType =
6499 getSpirvLinkageTypeFor(STI, *GV);
6500
6501 if (LnkType && *LnkType == SPIRV::LinkageType::Import)
6502 Init = nullptr;
6503
6504 const unsigned AddrSpace = GV->getAddressSpace();
6505 SPIRV::StorageClass::StorageClass StorageClass =
6506 addressSpaceToStorageClass(AddrSpace, STI);
6507 SPIRVTypeInst ResType =
6508 GR.getOrCreateSPIRVPointerType(GVType, I, StorageClass);
6509 Register Reg = GR.buildGlobalVariable(
6510 ResVReg, ResType, GlobalIdent, GV, StorageClass, Init,
6511 GlobalVar->isConstant(), LnkType, MIRBuilder, true);
6512 // TODO: For AMDGCN, we pipe externally_initialized through via
6513 // HostAccessINTEL, with ReadWrite (3) access, which is we then handle during
6514 // reverse translation. We should remove this once SPIR-V gains the ability to
6515 // express the concept.
6516 if (GlobalVar->isExternallyInitialized() &&
6517 STI.getTargetTriple().getVendor() == Triple::AMD) {
6518 constexpr unsigned ReadWriteINTEL = 3u;
6519 buildOpDecorate(Reg, MIRBuilder, SPIRV::Decoration::HostAccessINTEL,
6520 {ReadWriteINTEL});
6521 MachineInstrBuilder MIB(*MF, --MIRBuilder.getInsertPt());
6522 addStringImm(GV->getName(), MIB);
6523 }
6524 return Reg.isValid();
6525}
6526
6527bool SPIRVInstructionSelector::selectLog10(Register ResVReg,
6528 SPIRVTypeInst ResType,
6529 MachineInstr &I) const {
6530 if (STI.canUseExtInstSet(SPIRV::InstructionSet::OpenCL_std)) {
6531 return selectExtInst(ResVReg, ResType, I, CL::log10);
6532 }
6533
6534 // There is no log10 instruction in the GLSL Extended Instruction set, so it
6535 // is implemented as:
6536 // log10(x) = log2(x) * (1 / log2(10))
6537 // = log2(x) * 0.30103
6538
6539 MachineIRBuilder MIRBuilder(I);
6540 MachineBasicBlock &BB = *I.getParent();
6541
6542 // Build log2(x).
6543 Register VarReg = MRI->createVirtualRegister(GR.getRegClass(ResType));
6544 BuildMI(BB, I, I.getDebugLoc(), TII.get(SPIRV::OpExtInst))
6545 .addDef(VarReg)
6546 .addUse(GR.getSPIRVTypeID(ResType))
6547 .addImm(static_cast<uint32_t>(SPIRV::InstructionSet::GLSL_std_450))
6548 .addImm(GL::Log2)
6549 .add(I.getOperand(1))
6550 .constrainAllUses(TII, TRI, RBI);
6551
6552 // Build 0.30103.
6553 assert(ResType->getOpcode() == SPIRV::OpTypeVector ||
6554 ResType->getOpcode() == SPIRV::OpTypeFloat);
6555 // TODO: Add matrix implementation once supported by the HLSL frontend.
6556 SPIRVTypeInst SpirvScalarType = GR.getScalarOrVectorComponentType(ResType);
6557 // The literal must match the precision of the scalar type, otherwise the
6558 // OpConstant will contain non-zero high-order bits and fail SPIR-V
6559 // validation when the type is narrower than 32 bits (e.g. half).
6560 APFloat ScaleVal(0.30103);
6561 bool LosesInfo;
6562 ScaleVal.convert(
6563 getZeroFP(GR.getTypeForSPIRVType(SpirvScalarType)).getSemantics(),
6564 APFloat::rmNearestTiesToEven, &LosesInfo);
6565 Register ScaleReg = GR.buildConstantFP(ScaleVal, MIRBuilder, SpirvScalarType);
6566
6567 // Multiply log2(x) by 0.30103 to get log10(x) result.
6568 auto Opcode = ResType->getOpcode() == SPIRV::OpTypeVector
6569 ? SPIRV::OpVectorTimesScalar
6570 : SPIRV::OpFMulS;
6571 BuildMI(BB, I, I.getDebugLoc(), TII.get(Opcode))
6572 .addDef(ResVReg)
6573 .addUse(GR.getSPIRVTypeID(ResType))
6574 .addUse(VarReg)
6575 .addUse(ScaleReg)
6576 .constrainAllUses(TII, TRI, RBI);
6577 return true;
6578}
6579
6580bool SPIRVInstructionSelector::selectFpowi(Register ResVReg,
6581 SPIRVTypeInst ResType,
6582 MachineInstr &I) const {
6583 // On OpenCL targets, pown(gentype x, intn n) maps directly.
6584 if (STI.canUseExtInstSet(SPIRV::InstructionSet::OpenCL_std))
6585 return selectExtInst(ResVReg, ResType, I, CL::pown);
6586
6587 // On GLSL (Vulkan) targets, there is no integer-exponent power instruction.
6588 // Lower as: Pow(base, OpConvertSToF(exp)).
6589 if (STI.canUseExtInstSet(SPIRV::InstructionSet::GLSL_std_450)) {
6590 Register BaseReg = I.getOperand(1).getReg();
6591 Register ExpReg = I.getOperand(2).getReg();
6592 Register FloatExpReg = MRI->createVirtualRegister(GR.getRegClass(ResType));
6593 if (!selectOpWithSrcs(FloatExpReg, ResType, I, {ExpReg},
6594 SPIRV::OpConvertSToF))
6595 return false;
6596 return selectExtInst(ResVReg, ResType, I, GL::Pow,
6597 /*setMIFlags=*/true, /*useMISrc=*/false,
6598 {BaseReg, FloatExpReg});
6599 }
6600 return false;
6601}
6602
6603bool SPIRVInstructionSelector::selectModf(Register ResVReg,
6604 SPIRVTypeInst ResType,
6605 MachineInstr &I) const {
6606 // llvm.modf has a single arg --the number to be decomposed-- and returns a
6607 // struct { restype, restype }, while OpenCLLIB::modf has two args --the
6608 // number to be decomposed and a pointer--, returns the fractional part and
6609 // the integral part is stored in the pointer argument. Therefore, we can't
6610 // use directly the OpenCLLIB::modf intrinsic. However, we can do some
6611 // scaffolding to make it work. The idea is to create an alloca instruction
6612 // to get a ptr, pass this ptr to OpenCL::modf, and then load the value
6613 // from this ptr to place it in the struct. llvm.modf returns the fractional
6614 // part as the first element of the result, and the integral part as the
6615 // second element of the result.
6616
6617 // At this point, the return type is not a struct anymore, but rather two
6618 // independent elements of SPIRVResType. We can get each independent element
6619 // from I.getDefs() or I.getOperands().
6620 if (STI.canUseExtInstSet(SPIRV::InstructionSet::OpenCL_std)) {
6621 MachineIRBuilder MIRBuilder(I);
6622 // Get pointer type for alloca variable.
6623 const SPIRVTypeInst PtrType = GR.getOrCreateSPIRVPointerType(
6624 ResType, MIRBuilder, SPIRV::StorageClass::Function);
6625 // Create new register for the pointer type of alloca variable.
6626 Register PtrTyReg =
6627 MIRBuilder.getMRI()->createVirtualRegister(&SPIRV::iIDRegClass);
6628 MIRBuilder.getMRI()->setType(
6629 PtrTyReg,
6630 LLT::pointer(storageClassToAddressSpace(SPIRV::StorageClass::Function),
6631 GR.getPointerSize()));
6632
6633 // Assign SPIR-V type of the pointer type of the alloca variable to the
6634 // new register.
6635 GR.assignSPIRVTypeToVReg(PtrType, PtrTyReg, MIRBuilder.getMF());
6636 MachineBasicBlock::iterator VarPos = getOpVariableMBBIt(*I.getMF());
6637 MachineBasicBlock &EntryBB = I.getMF()->front();
6638 auto AllocaMIB =
6639 BuildMI(EntryBB, VarPos, I.getDebugLoc(), TII.get(SPIRV::OpVariable))
6640 .addDef(PtrTyReg)
6641 .addUse(GR.getSPIRVTypeID(PtrType))
6642 .addImm(static_cast<uint32_t>(SPIRV::StorageClass::Function));
6643 Register Variable = AllocaMIB->getOperand(0).getReg();
6644
6645 MachineBasicBlock &BB = *I.getParent();
6646 // Create the OpenCLLIB::modf instruction.
6647 auto MIB =
6648 BuildMI(BB, I, I.getDebugLoc(), TII.get(SPIRV::OpExtInst))
6649 .addDef(ResVReg)
6650 .addUse(GR.getSPIRVTypeID(ResType))
6651 .addImm(static_cast<uint32_t>(SPIRV::InstructionSet::OpenCL_std))
6652 .addImm(CL::modf)
6653 .setMIFlags(I.getFlags())
6654 .add(I.getOperand(I.getNumExplicitDefs())) // Floating point value.
6655 .addUse(Variable); // Pointer to integral part.
6656 // Assign the integral part stored in the ptr to the second element of the
6657 // result.
6658 Register IntegralPartReg = I.getOperand(1).getReg();
6659 if (IntegralPartReg.isValid()) {
6660 // Load the value from the pointer to integral part.
6661 auto LoadMIB = BuildMI(BB, I, I.getDebugLoc(), TII.get(SPIRV::OpLoad))
6662 .addDef(IntegralPartReg)
6663 .addUse(GR.getSPIRVTypeID(ResType))
6664 .addUse(Variable);
6665 LoadMIB.constrainAllUses(TII, TRI, RBI);
6666 return true;
6667 }
6668
6669 MIB.constrainAllUses(TII, TRI, RBI);
6670 return true;
6671 } else if (STI.canUseExtInstSet(SPIRV::InstructionSet::GLSL_std_450)) {
6672 assert(false && "GLSL::Modf is deprecated.");
6673 // FIXME: GL::Modf is deprecated, use Modfstruct instead.
6674 return false;
6675 }
6676 return false;
6677}
6678
6679// Generate the instructions to load 3-element vector builtin input
6680// IDs/Indices.
6681// Like: GlobalInvocationId, LocalInvocationId, etc....
6682
6683bool SPIRVInstructionSelector::loadVec3BuiltinInputID(
6684 SPIRV::BuiltIn::BuiltIn BuiltInValue, Register ResVReg,
6685 SPIRVTypeInst ResType, MachineInstr &I) const {
6686 MachineIRBuilder MIRBuilder(I);
6687 const SPIRVTypeInst Vec3Ty =
6688 GR.getOrCreateSPIRVVectorType(ResType, 3, MIRBuilder, false);
6689 const SPIRVTypeInst PtrType = GR.getOrCreateSPIRVPointerType(
6690 Vec3Ty, MIRBuilder, SPIRV::StorageClass::Input);
6691
6692 // Create new register for the input ID builtin variable.
6693 Register NewRegister =
6694 MIRBuilder.getMRI()->createVirtualRegister(&SPIRV::iIDRegClass);
6695 MIRBuilder.getMRI()->setType(NewRegister, LLT::pointer(0, 64));
6696 GR.assignSPIRVTypeToVReg(PtrType, NewRegister, MIRBuilder.getMF());
6697
6698 // Build global variable with the necessary decorations for the input ID
6699 // builtin variable.
6700 Register Variable = GR.buildGlobalVariable(
6701 NewRegister, PtrType, getLinkStringForBuiltIn(BuiltInValue), nullptr,
6702 SPIRV::StorageClass::Input, nullptr, true, std::nullopt, MIRBuilder,
6703 false);
6704
6705 // Create new register for loading value.
6706 MachineRegisterInfo *MRI = MIRBuilder.getMRI();
6707 Register LoadedRegister = MRI->createVirtualRegister(&SPIRV::iIDRegClass);
6708 MIRBuilder.getMRI()->setType(LoadedRegister, LLT::pointer(0, 64));
6709 GR.assignSPIRVTypeToVReg(Vec3Ty, LoadedRegister, MIRBuilder.getMF());
6710
6711 // Load v3uint value from the global variable.
6712 BuildMI(*I.getParent(), I, I.getDebugLoc(), TII.get(SPIRV::OpLoad))
6713 .addDef(LoadedRegister)
6714 .addUse(GR.getSPIRVTypeID(Vec3Ty))
6715 .addUse(Variable);
6716
6717 // Get the input ID index. Expecting operand is a constant immediate value,
6718 // wrapped in a type assignment.
6719 assert(I.getOperand(2).isReg());
6720 const uint32_t ThreadId = foldImm(I.getOperand(2), MRI);
6721
6722 // Extract the input ID from the loaded vector value.
6723 MachineBasicBlock &BB = *I.getParent();
6724 auto MIB = BuildMI(BB, I, I.getDebugLoc(), TII.get(SPIRV::OpCompositeExtract))
6725 .addDef(ResVReg)
6726 .addUse(GR.getSPIRVTypeID(ResType))
6727 .addUse(LoadedRegister)
6728 .addImm(ThreadId);
6729 MIB.constrainAllUses(TII, TRI, RBI);
6730 return true;
6731}
6732
6733// Generate the instructions to load 32-bit integer builtin input IDs/Indices.
6734// Like LocalInvocationIndex
6735bool SPIRVInstructionSelector::loadBuiltinInputID(
6736 SPIRV::BuiltIn::BuiltIn BuiltInValue, Register ResVReg,
6737 SPIRVTypeInst ResType, MachineInstr &I) const {
6738 MachineIRBuilder MIRBuilder(I);
6739 const SPIRVTypeInst PtrType = GR.getOrCreateSPIRVPointerType(
6740 ResType, MIRBuilder, SPIRV::StorageClass::Input);
6741
6742 // Create new register for the input ID builtin variable.
6743 Register NewRegister =
6744 MIRBuilder.getMRI()->createVirtualRegister(GR.getRegClass(PtrType));
6745 MIRBuilder.getMRI()->setType(
6746 NewRegister,
6747 LLT::pointer(storageClassToAddressSpace(SPIRV::StorageClass::Input),
6748 GR.getPointerSize()));
6749 GR.assignSPIRVTypeToVReg(PtrType, NewRegister, MIRBuilder.getMF());
6750
6751 // Build global variable with the necessary decorations for the input ID
6752 // builtin variable.
6753 Register Variable = GR.buildGlobalVariable(
6754 NewRegister, PtrType, getLinkStringForBuiltIn(BuiltInValue), nullptr,
6755 SPIRV::StorageClass::Input, nullptr, true, std::nullopt, MIRBuilder,
6756 false);
6757
6758 // Load uint value from the global variable.
6759 auto MIB = BuildMI(*I.getParent(), I, I.getDebugLoc(), TII.get(SPIRV::OpLoad))
6760 .addDef(ResVReg)
6761 .addUse(GR.getSPIRVTypeID(ResType))
6762 .addUse(Variable);
6763
6764 MIB.constrainAllUses(TII, TRI, RBI);
6765 return true;
6766}
6767
6768SPIRVTypeInst SPIRVInstructionSelector::widenTypeToVec4(SPIRVTypeInst Type,
6769 MachineInstr &I) const {
6770 MachineIRBuilder MIRBuilder(I);
6771 if (Type->getOpcode() != SPIRV::OpTypeVector)
6772 return GR.getOrCreateSPIRVVectorType(Type, 4, MIRBuilder, false);
6773
6774 if (GR.getScalarOrVectorComponentCount(Type) == 4)
6775 return Type;
6776
6777 SPIRVTypeInst ScalarType = GR.getScalarOrVectorComponentType(Type);
6778 return GR.getOrCreateSPIRVVectorType(ScalarType, 4, MIRBuilder, false);
6779}
6780
6781bool SPIRVInstructionSelector::loadHandleBeforePosition(
6782 Register &HandleReg, SPIRVTypeInst ResType, GIntrinsic &HandleDef,
6783 MachineInstr &Pos) const {
6784
6785 assert(HandleDef.getIntrinsicID() ==
6786 Intrinsic::spv_resource_handlefrombinding);
6787 uint32_t Set = foldImm(HandleDef.getOperand(2), MRI);
6788 uint32_t Binding = foldImm(HandleDef.getOperand(3), MRI);
6789 uint32_t ArraySize = foldImm(HandleDef.getOperand(4), MRI);
6790 Register IndexReg = HandleDef.getOperand(5).getReg();
6791 std::string Name =
6792 getStringValueFromReg(HandleDef.getOperand(6).getReg(), *MRI);
6793
6794 bool IsStructuredBuffer = ResType->getOpcode() == SPIRV::OpTypePointer;
6795 MachineIRBuilder MIRBuilder(HandleDef);
6796 SPIRVTypeInst VarType = ResType;
6797 SPIRV::StorageClass::StorageClass SC = SPIRV::StorageClass::UniformConstant;
6798
6799 if (IsStructuredBuffer) {
6800 VarType = GR.getPointeeType(ResType);
6801 SC = GR.getPointerStorageClass(ResType);
6802 }
6803
6804 if (ResType->getOpcode() == SPIRV::OpTypeImage && ArraySize == 0)
6805 MIRBuilder.buildInstr(SPIRV::OpCapability)
6806 .addImm(SPIRV::Capability::RuntimeDescriptorArrayEXT);
6807
6808 Register VarReg =
6809 buildPointerToResource(SPIRVTypeInst(VarType), SC, Set, Binding,
6810 ArraySize, IndexReg, Name, MIRBuilder);
6811
6812 // The handle for the buffer is the pointer to the resource. For an image, the
6813 // handle is the image object. So images get an extra load.
6814 uint32_t LoadOpcode =
6815 IsStructuredBuffer ? SPIRV::OpCopyObject : SPIRV::OpLoad;
6816 GR.assignSPIRVTypeToVReg(ResType, HandleReg, *Pos.getMF());
6817 BuildMI(*Pos.getParent(), Pos, HandleDef.getDebugLoc(), TII.get(LoadOpcode))
6818 .addDef(HandleReg)
6819 .addUse(GR.getSPIRVTypeID(ResType))
6820 .addUse(VarReg)
6821 .constrainAllUses(TII, TRI, RBI);
6822 return true;
6823}
6824
6825void SPIRVInstructionSelector::errorIfInstrOutsideShader(
6826 MachineInstr &I) const {
6827 if (!STI.isShader()) {
6828 std::string DiagMsg;
6829 raw_string_ostream OS(DiagMsg);
6830 I.print(OS, true, false, false, false);
6831 DiagMsg += " is only supported in shaders.\n";
6832 report_fatal_error(DiagMsg.c_str(), false);
6833 }
6834}
6835
6836namespace llvm {
6837InstructionSelector *
6838createSPIRVInstructionSelector(const SPIRVTargetMachine &TM,
6839 const SPIRVSubtarget &Subtarget,
6840 const RegisterBankInfo &RBI) {
6841 return new SPIRVInstructionSelector(TM, Subtarget, RBI);
6842}
6843} // namespace llvm
UseMI
MachineInstrBuilder & UseMI
Definition AArch64ExpandPseudoInsts.cpp:127
GET_GLOBALISEL_PREDICATES_INIT
#define GET_GLOBALISEL_PREDICATES_INIT
GET_GLOBALISEL_TEMPORARIES_INIT
#define GET_GLOBALISEL_TEMPORARIES_INIT
Generic
@ Generic
Definition AArch64MCAsmInfo.cpp:24
assert
assert(UImm &&(UImm !=~static_cast< T >(0)) &&"Invalid immediate!")
APFloat.h
This file declares a class to represent arbitrary precision floating point values and provide a varie...
selectUnmergeValues
static bool selectUnmergeValues(MachineInstrBuilder &MIB, const ARMBaseInstrInfo &TII, MachineRegisterInfo &MRI, const TargetRegisterInfo &TRI, const RegisterBankInfo &RBI)
Definition ARMInstructionSelector.cpp:264
MBB
MachineBasicBlock & MBB
Definition ARMSLSHardening.cpp:71
DL
MachineBasicBlock MachineBasicBlock::iterator DebugLoc DL
Definition ARMSLSHardening.cpp:73
SwapBits
static uint8_t SwapBits(uint8_t Val)
Definition BPFMCCodeEmitter.cpp:122
true
basic Basic Alias true
Definition BasicAliasAnalysis.cpp:2057
X
#define X(NUM, ENUM, NAME)
Definition ELF.h:851
E
static GCRegistry::Add< CoreCLRGC > E("coreclr", "CoreCLR-compatible GC")
Binding
DXIL Resource Implicit Binding
Definition DXILResourceImplicitBinding.cpp:185
DEBUG_TYPE
#define DEBUG_TYPE
Definition GenericCycleImpl.h:31
GenericMachineInstrs.h
Declares convenience wrapper classes for interpreting MachineInstr instances as specific generic oper...
TII
const HexagonInstrInfo * TII
Definition HexagonCopyToCombine.cpp:118
MI
IRTranslator LLVM IR MI
Definition IRTranslator.cpp:110
LLVMIntType
LLVMTypeRef LLVMIntType(unsigned NumBits)
Definition Core.cpp:729
NumOps
const size_t AbstractManglingParser< Derived, Alloc >::NumOps
Definition ItaniumDemangle.h:3452
Direction
Loop::LoopBounds::Direction Direction
Definition LoopInfo.cpp:253
F
#define F(x, y, z)
Definition MD5.cpp:54
I
#define I(x, y, z)
Definition MD5.cpp:57
Reg
Register Reg
Definition MachineSink.cpp:2126
TRI
Register const TargetRegisterInfo * TRI
Definition MachineSink.cpp:2127
Register
Promote Memory to Register
Definition Mem2Reg.cpp:110
T
#define T
Definition Mips16ISelLowering.cpp:282
T1
#define T1
Definition Mips16ISelLowering.cpp:281
OpIdx
MachineInstr unsigned OpIdx
Definition NVPTXPrologEpilogPass.cpp:56
High
uint64_t High
Definition NVVMIntrRange.cpp:46
II
uint64_t IntrinsicInst * II
Definition NVVMIntrRange.cpp:46
getName
static StringRef getName(Value *V)
Definition ProvenanceAnalysisEvaluator.cpp:20
getFCmpOpcode
static unsigned getFCmpOpcode(CmpInst::Predicate Pred, unsigned Size)
Definition RISCVInstructionSelector.cpp:1851
getOneFP
static APFloat getOneFP(const Type *LLVMFloatTy)
Definition SPIRVInstructionSelector.cpp:2773
isUSMStorageClass
static bool isUSMStorageClass(SPIRV::StorageClass::StorageClass SC)
Definition SPIRVInstructionSelector.cpp:2505
isASCastInGVar
static bool isASCastInGVar(MachineRegisterInfo *MRI, Register ResVReg)
Definition SPIRVInstructionSelector.cpp:2516
mayApplyGenericSelection
static bool mayApplyGenericSelection(unsigned Opcode)
Definition SPIRVInstructionSelector.cpp:953
getZeroFP
static APFloat getZeroFP(const Type *LLVMFloatTy)
Definition SPIRVInstructionSelector.cpp:2759
ExtInstList
std::vector< std::pair< SPIRV::InstructionSet::InstructionSet, uint32_t > > ExtInstList
Definition SPIRVInstructionSelector.cpp:46
intrinsicHasSideEffects
static bool intrinsicHasSideEffects(Intrinsic::ID ID)
Definition SPIRVInstructionSelector.cpp:667
getBoolCmpOpcode
static unsigned getBoolCmpOpcode(unsigned PredNum)
Definition SPIRVInstructionSelector.cpp:2747
getICmpOpcode
static unsigned getICmpOpcode(unsigned PredNum)
Definition SPIRVInstructionSelector.cpp:2707
isOpcodeWithNoSideEffects
static bool isOpcodeWithNoSideEffects(unsigned Opcode)
Definition SPIRVInstructionSelector.cpp:733
addMemoryOperands
static void addMemoryOperands(MachineMemOperand *MemOp, MachineInstrBuilder &MIB, MachineIRBuilder &MIRBuilder, SPIRVGlobalRegistry &GR)
Definition SPIRVInstructionSelector.cpp:1827
isConstReg
static bool isConstReg(MachineRegisterInfo *MRI, MachineInstr *OpDef)
Definition SPIRVInstructionSelector.cpp:599
getPtrCmpOpcode
static unsigned getPtrCmpOpcode(unsigned Pred)
Definition SPIRVInstructionSelector.cpp:2735
isDead
bool isDead(const MachineInstr &MI, const MachineRegisterInfo &MRI)
Definition SPIRVInstructionSelector.cpp:767
SPIRV
spirv structurize SPIRV
Definition SPIRVStructurizer.cpp:1218
BaseType
BaseType
A given derived pointer can have multiple base pointers through phi/selects.
Definition SafepointIRVerifier.cpp:311
StringExtras.h
This file contains some functions that are useful when dealing with strings.
LLVM_DEBUG
#define LLVM_DEBUG(...)
Definition Debug.h:114
Y
static TableGen::Emitter::Opt Y("gen-skeleton-entry", EmitSkeleton, "Generate example skeleton entry")
CallbackFunction
static ManagedStatic< cl::opt< FnT >, OptCreatorT > CallbackFunction
Definition TableGenBackend.cpp:53
Mul
BinaryOperator * Mul
Definition X86PartialReduction.cpp:75
llvm::APFloatBase::IEEEsingle
static const fltSemantics & IEEEsingle()
Definition APFloat.h:296
llvm::APFloatBase::IEEEdouble
static const fltSemantics & IEEEdouble()
Definition APFloat.h:297
llvm::APFloatBase::IEEEhalf
static const fltSemantics & IEEEhalf()
Definition APFloat.h:294
llvm::APFloat::getSemantics
const fltSemantics & getSemantics() const
Definition APFloat.h:1542
llvm::APFloat::getOne
static APFloat getOne(const fltSemantics &Sem, bool Negative=false)
Factory for Positive and Negative One.
Definition APFloat.h:1143
llvm::APFloat::getZero
static APFloat getZero(const fltSemantics &Sem, bool Negative=false)
Factory for Positive and Negative Zero.
Definition APFloat.h:1134
llvm::APInt::getAllOnes
static APInt getAllOnes(unsigned numBits)
Return an APInt of a specified width with all bits set.
Definition APInt.h:235
llvm::ArrayRef
ArrayRef - Represent a constant reference to an array (0 or more elements consecutively in memory),...
Definition ArrayRef.h:40
llvm::BlockFrequencyInfo
BlockFrequencyInfo pass uses BlockFrequencyInfoImpl implementation to estimate IR basic block frequen...
Definition BlockFrequencyInfo.h:38
llvm::CmpInst::Predicate
Predicate
This enumeration lists the possible predicates for CmpInst subclasses.
Definition InstrTypes.h:676
llvm::CmpInst::FCMP_OEQ
@ FCMP_OEQ
0 0 0 1 True if ordered and equal
Definition InstrTypes.h:679
llvm::CmpInst::ICMP_SLT
@ ICMP_SLT
signed less than
Definition InstrTypes.h:705
llvm::CmpInst::ICMP_SLE
@ ICMP_SLE
signed less or equal
Definition InstrTypes.h:706
llvm::CmpInst::FCMP_OLT
@ FCMP_OLT
0 1 0 0 True if ordered and less than
Definition InstrTypes.h:682
llvm::CmpInst::FCMP_ULE
@ FCMP_ULE
1 1 0 1 True if unordered, less than, or equal
Definition InstrTypes.h:691
llvm::CmpInst::FCMP_OGT
@ FCMP_OGT
0 0 1 0 True if ordered and greater than
Definition InstrTypes.h:680
llvm::CmpInst::FCMP_OGE
@ FCMP_OGE
0 0 1 1 True if ordered and greater than or equal
Definition InstrTypes.h:681
llvm::CmpInst::ICMP_UGE
@ ICMP_UGE
unsigned greater or equal
Definition InstrTypes.h:700
llvm::CmpInst::ICMP_UGT
@ ICMP_UGT
unsigned greater than
Definition InstrTypes.h:699
llvm::CmpInst::ICMP_SGT
@ ICMP_SGT
signed greater than
Definition InstrTypes.h:703
llvm::CmpInst::FCMP_ULT
@ FCMP_ULT
1 1 0 0 True if unordered or less than
Definition InstrTypes.h:690
llvm::CmpInst::FCMP_ONE
@ FCMP_ONE
0 1 1 0 True if ordered and operands are unequal
Definition InstrTypes.h:684
llvm::CmpInst::FCMP_UEQ
@ FCMP_UEQ
1 0 0 1 True if unordered or equal
Definition InstrTypes.h:687
llvm::CmpInst::ICMP_ULT
@ ICMP_ULT
unsigned less than
Definition InstrTypes.h:701
llvm::CmpInst::FCMP_UGT
@ FCMP_UGT
1 0 1 0 True if unordered or greater than
Definition InstrTypes.h:688
llvm::CmpInst::FCMP_OLE
@ FCMP_OLE
0 1 0 1 True if ordered and less than or equal
Definition InstrTypes.h:683
llvm::CmpInst::FCMP_ORD
@ FCMP_ORD
0 1 1 1 True if ordered (no nans)
Definition InstrTypes.h:685
llvm::CmpInst::ICMP_NE
@ ICMP_NE
not equal
Definition InstrTypes.h:698
llvm::CmpInst::ICMP_SGE
@ ICMP_SGE
signed greater or equal
Definition InstrTypes.h:704
llvm::CmpInst::FCMP_UNE
@ FCMP_UNE
1 1 1 0 True if unordered or not equal
Definition InstrTypes.h:692
llvm::CmpInst::ICMP_ULE
@ ICMP_ULE
unsigned less or equal
Definition InstrTypes.h:702
llvm::CmpInst::FCMP_UGE
@ FCMP_UGE
1 0 1 1 True if unordered, greater than, or equal
Definition InstrTypes.h:689
llvm::CmpInst::FCMP_UNO
@ FCMP_UNO
1 0 0 0 True if unordered: isnan(X) | isnan(Y)
Definition InstrTypes.h:686
llvm::Constant::getNullValue
static LLVM_ABI Constant * getNullValue(Type *Ty)
Constructor to create a '0' constant of arbitrary type.
Definition Constants.cpp:390
llvm::DenseMapBase::size
unsigned size() const
Definition DenseMap.h:110
llvm::Function::getContext
LLVMContext & getContext() const
getContext - Return a reference to the LLVMContext associated with this function.
Definition Function.cpp:358
llvm::GIntrinsic
Represents a call to an intrinsic.
Definition GenericMachineInstrs.h:529
llvm::GIntrinsic::getIntrinsicID
Intrinsic::ID getIntrinsicID() const
Definition GenericMachineInstrs.h:531
llvm::GlobalValue::getAddressSpace
unsigned getAddressSpace() const
Definition GlobalValue.h:207
llvm::GlobalValue::getParent
Module * getParent()
Get the module that this global value is contained inside of...
Definition GlobalValue.h:663
llvm::GlobalValue::InternalLinkage
@ InternalLinkage
Rename collisions when linking (static functions).
Definition GlobalValue.h:60
llvm::IntegerType::get
static LLVM_ABI IntegerType * get(LLVMContext &C, unsigned NumBits)
This static method is the primary way of constructing an IntegerType.
Definition Type.cpp:354
llvm::LLT::scalar
static constexpr LLT scalar(unsigned SizeInBits)
Get a low-level scalar or aggregate "bag of bits".
Definition LowLevelType.h:88
llvm::LLT::isValid
constexpr bool isValid() const
Definition LowLevelType.h:262
llvm::LLT::getNumElements
constexpr uint16_t getNumElements() const
Returns the number of elements in a vector LLT.
Definition LowLevelType.h:347
llvm::LLT::isVector
constexpr bool isVector() const
Definition LowLevelType.h:286
llvm::LLT::pointer
static constexpr LLT pointer(unsigned AddressSpace, unsigned SizeInBits)
Get a low-level pointer in the given address space.
Definition LowLevelType.h:115
llvm::LLT::fixed_vector
static constexpr LLT fixed_vector(unsigned NumElements, unsigned ScalarSizeInBits)
Get a low-level fixed-width vector of some number of elements and element width.
Definition LowLevelType.h:203
llvm::MachineBasicBlock::getNumber
int getNumber() const
MachineBasicBlocks are uniquely numbered at the function level, unless they're not in a MachineFuncti...
Definition MachineBasicBlock.h:1274
llvm::MachineBasicBlock::getFirstNonPHI
LLVM_ABI iterator getFirstNonPHI()
Returns a pointer to the first instruction in this block that is not a PHINode instruction.
Definition MachineBasicBlock.cpp:202
llvm::MachineBasicBlock::getParent
const MachineFunction * getParent() const
Return the MachineFunction containing this basic block.
Definition MachineBasicBlock.h:330
llvm::MachineBasicBlock::iterator
MachineInstrBundleIterator< MachineInstr > iterator
Definition MachineBasicBlock.h:348
llvm::MachineFunction::getRegInfo
MachineRegisterInfo & getRegInfo()
getRegInfo - Return information about the registers currently in use.
Definition MachineFunction.h:798
llvm::MachineFunction::getFunction
Function & getFunction()
Return the LLVM function that this machine code represents.
Definition MachineFunction.h:749
llvm::MachineIRBuilder
Helper class to build MachineInstr.
Definition MachineIRBuilder.h:237
llvm::MachineIRBuilder::getInsertPt
MachineBasicBlock::iterator getInsertPt()
Current insertion point for new instructions.
Definition MachineIRBuilder.h:335
llvm::MachineIRBuilder::buildInstr
MachineInstrBuilder buildInstr(unsigned Opcode)
Build and insert <empty> = Opcode <empty>.
Definition MachineIRBuilder.h:423
llvm::MachineIRBuilder::getMF
MachineFunction & getMF()
Getter for the function we currently build.
Definition MachineIRBuilder.h:289
llvm::MachineIRBuilder::getMRI
MachineRegisterInfo * getMRI()
Getter for MRI.
Definition MachineIRBuilder.h:311
llvm::MachineInstrBuilder::constrainAllUses
void constrainAllUses(const TargetInstrInfo &TII, const TargetRegisterInfo &TRI, const RegisterBankInfo &RBI) const
Definition MachineInstrBuilder.h:441
llvm::MachineInstrBuilder::addUse
const MachineInstrBuilder & addUse(Register RegNo, RegState Flags={}, unsigned SubReg=0) const
Add a virtual register use operand.
Definition MachineInstrBuilder.h:225
llvm::MachineInstrBuilder::addReg
const MachineInstrBuilder & addReg(Register RegNo, RegState Flags={}, unsigned SubReg=0) const
Add a new virtual register operand.
Definition MachineInstrBuilder.h:199
llvm::MachineInstrBuilder::addImm
const MachineInstrBuilder & addImm(int64_t Val) const
Add a new immediate operand.
Definition MachineInstrBuilder.h:233
llvm::MachineInstrBuilder::add
const MachineInstrBuilder & add(const MachineOperand &MO) const
Definition MachineInstrBuilder.h:326
llvm::MachineInstrBuilder::addMBB
const MachineInstrBuilder & addMBB(MachineBasicBlock *MBB, unsigned TargetFlags=0) const
Definition MachineInstrBuilder.h:248
llvm::MachineInstrBuilder::addDef
const MachineInstrBuilder & addDef(Register RegNo, RegState Flags={}, unsigned SubReg=0) const
Add a virtual register definition operand.
Definition MachineInstrBuilder.h:218
llvm::MachineInstrBuilder::setMIFlags
const MachineInstrBuilder & setMIFlags(unsigned Flags) const
Definition MachineInstrBuilder.h:379
llvm::MachineInstrBuilder::getInstr
MachineInstr * getInstr() const
If conversion operators fail, use this method to get the MachineInstr explicitly.
Definition MachineInstrBuilder.h:191
llvm::MachineInstr
Representation of each machine instruction.
Definition MachineInstr.h:73
llvm::MachineInstr::getOpcode
unsigned getOpcode() const
Returns the opcode of this MachineInstr.
Definition MachineInstr.h:601
llvm::MachineInstr::getParent
const MachineBasicBlock * getParent() const
Definition MachineInstr.h:373
llvm::MachineInstr::getNumOperands
unsigned getNumOperands() const
Retuns the total number of operands.
Definition MachineInstr.h:604
llvm::MachineInstr::getNumExplicitOperands
LLVM_ABI unsigned getNumExplicitOperands() const
Returns the number of non-implicit operands.
Definition MachineInstr.cpp:851
llvm::MachineInstr::getNumExplicitDefs
LLVM_ABI unsigned getNumExplicitDefs() const
Returns the number of non-implicit definitions.
Definition MachineInstr.cpp:869
llvm::MachineInstr::emitGenericError
LLVM_ABI void emitGenericError(const Twine &ErrMsg) const
Definition MachineInstr.cpp:2373
llvm::MachineInstr::getMF
LLVM_ABI const MachineFunction * getMF() const
Return the function that contains the basic block that this instruction belongs to.
Definition MachineInstr.cpp:786
llvm::MachineInstr::getDebugLoc
const DebugLoc & getDebugLoc() const
Returns the debug location id of this MachineInstr.
Definition MachineInstr.h:525
llvm::MachineInstr::getOperand
const MachineOperand & getOperand(unsigned i) const
Definition MachineInstr.h:609
llvm::MachineMemOperand
A description of a memory reference used in the backend.
Definition MachineMemOperand.h:130
llvm::MachineMemOperand::MOVolatile
@ MOVolatile
The memory access is volatile.
Definition MachineMemOperand.h:141
llvm::MachineMemOperand::MONonTemporal
@ MONonTemporal
The memory access is non-temporal.
Definition MachineMemOperand.h:143
llvm::MachineOperand::getImm
int64_t getImm() const
Definition MachineOperand.h:560
llvm::MachineOperand::isReg
bool isReg() const
isReg - Tests if this is a MO_Register operand.
Definition MachineOperand.h:331
llvm::MachineOperand::getMBB
MachineBasicBlock * getMBB() const
Definition MachineOperand.h:575
llvm::MachineOperand::getReg
Register getReg() const
getReg - Returns the register number.
Definition MachineOperand.h:372
llvm::MachineRegisterInfo
MachineRegisterInfo - Keep track of information for virtual and physical registers,...
Definition MachineRegisterInfo.h:53
llvm::MachineRegisterInfo::use_instr_iterator
defusechain_instr_iterator< true, false, false, true > use_instr_iterator
use_instr_iterator/use_instr_begin/use_instr_end - Walk all uses of the specified register,...
Definition MachineRegisterInfo.h:478
llvm::MachineRegisterInfo::getRegClass
const TargetRegisterClass * getRegClass(Register Reg) const
Return the register class of the specified virtual register.
Definition MachineRegisterInfo.h:648
llvm::MachineRegisterInfo::getVRegDef
LLVM_ABI MachineInstr * getVRegDef(Register Reg) const
getVRegDef - Return the machine instr that defines the specified virtual register or null if none is ...
Definition MachineRegisterInfo.cpp:404
llvm::MachineRegisterInfo::use_instr_begin
use_instr_iterator use_instr_begin(Register RegNo) const
Definition MachineRegisterInfo.h:480
llvm::MachineRegisterInfo::def_instr_end
static def_instr_iterator def_instr_end()
Definition MachineRegisterInfo.h:404
llvm::MachineRegisterInfo::def_instr_iterator
defusechain_instr_iterator< false, true, false, true > def_instr_iterator
def_instr_iterator/def_instr_begin/def_instr_end - Walk all defs of the specified register,...
Definition MachineRegisterInfo.h:399
llvm::MachineRegisterInfo::createVirtualRegister
LLVM_ABI Register createVirtualRegister(const TargetRegisterClass *RegClass, StringRef Name="")
createVirtualRegister - Create and return a new virtual register in the function with the specified r...
Definition MachineRegisterInfo.cpp:154
llvm::MachineRegisterInfo::def_instr_begin
def_instr_iterator def_instr_begin(Register RegNo) const
Definition MachineRegisterInfo.h:401
llvm::MachineRegisterInfo::getType
LLT getType(Register Reg) const
Get the low-level type of Reg or LLT{} if Reg is not a generic (target independent) virtual register.
Definition MachineRegisterInfo.h:771
llvm::MachineRegisterInfo::use_instr_end
static use_instr_iterator use_instr_end()
Definition MachineRegisterInfo.h:483
llvm::MachineRegisterInfo::use_nodbg_instructions
iterator_range< use_instr_nodbg_iterator > use_nodbg_instructions(Register Reg) const
Definition MachineRegisterInfo.h:546
llvm::MachineRegisterInfo::setType
LLVM_ABI void setType(Register VReg, LLT Ty)
Set the low-level type of VReg to Ty.
Definition MachineRegisterInfo.cpp:185
llvm::MachineRegisterInfo::getMF
const MachineFunction & getMF() const
Definition MachineRegisterInfo.h:189
llvm::MachineRegisterInfo::setRegClass
LLVM_ABI void setRegClass(Register Reg, const TargetRegisterClass *RC)
setRegClass - Set the register class of the specified virtual register.
Definition MachineRegisterInfo.cpp:58
llvm::MachineRegisterInfo::createGenericVirtualRegister
LLVM_ABI Register createGenericVirtualRegister(LLT Ty, StringRef Name="")
Create and return a new generic virtual register with low-level type Ty.
Definition MachineRegisterInfo.cpp:191
llvm::MachineRegisterInfo::getRegClassOrNull
const TargetRegisterClass * getRegClassOrNull(Register Reg) const
Return the register class of Reg, or null if Reg has not been assigned a register class yet.
Definition MachineRegisterInfo.h:665
llvm::MachineRegisterInfo::use_instructions
iterator_range< use_instr_iterator > use_instructions(Register Reg) const
Definition MachineRegisterInfo.h:488
llvm::MachineRegisterInfo::getNumVirtRegs
unsigned getNumVirtRegs() const
getNumVirtRegs - Return the number of virtual registers created.
Definition MachineRegisterInfo.h:796
llvm::MachineRegisterInfo::replaceRegWith
LLVM_ABI void replaceRegWith(Register FromReg, Register ToReg)
replaceRegWith - Replace all instances of FromReg with ToReg in the machine function.
Definition MachineRegisterInfo.cpp:386
llvm::ProfileSummaryInfo
Analysis providing profile information.
Definition ProfileSummaryInfo.h:42
llvm::RegisterBankInfo
Holds all the information related to register banks.
Definition RegisterBankInfo.h:40
llvm::Register
Wrapper class representing virtual and physical registers.
Definition Register.h:20
llvm::Register::isValid
constexpr bool isValid() const
Definition Register.h:112
llvm::Register::isPhysical
constexpr bool isPhysical() const
Return true if the specified register number is in the physical register namespace.
Definition Register.h:83
llvm::SPIRVGlobalRegistry::isScalarOrVectorSigned
bool isScalarOrVectorSigned(SPIRVTypeInst Type) const
Definition SPIRVGlobalRegistry.cpp:1478
llvm::SPIRVGlobalRegistry::getOrCreateOpTypeSampledImage
SPIRVTypeInst getOrCreateOpTypeSampledImage(SPIRVTypeInst ImageType, MachineIRBuilder &MIRBuilder)
Definition SPIRVGlobalRegistry.cpp:1726
llvm::SPIRVGlobalRegistry::assignSPIRVTypeToVReg
void assignSPIRVTypeToVReg(SPIRVTypeInst Type, Register VReg, const MachineFunction &MF)
Definition SPIRVGlobalRegistry.cpp:130
llvm::SPIRVGlobalRegistry::getRegClass
const TargetRegisterClass * getRegClass(SPIRVTypeInst SpvType) const
Definition SPIRVGlobalRegistry.cpp:2088
llvm::SPIRVGlobalRegistry::getOrAddMemAliasingINTELInst
MachineInstr * getOrAddMemAliasingINTELInst(MachineIRBuilder &MIRBuilder, const MDNode *AliasingListMD)
Definition SPIRVGlobalRegistry.cpp:2150
llvm::SPIRVGlobalRegistry::isAggregateType
bool isAggregateType(SPIRVTypeInst Type) const
Definition SPIRVGlobalRegistry.h:371
llvm::SPIRVGlobalRegistry::getScalarOrVectorBitWidth
unsigned getScalarOrVectorBitWidth(SPIRVTypeInst Type) const
Definition SPIRVGlobalRegistry.cpp:1449
llvm::SPIRVGlobalRegistry::getOrCreateSPIRVIntegerType
SPIRVTypeInst getOrCreateSPIRVIntegerType(unsigned BitWidth, MachineIRBuilder &MIRBuilder)
Definition SPIRVGlobalRegistry.cpp:1856
llvm::SPIRVGlobalRegistry::getOrCreateSPIRVVectorType
SPIRVTypeInst getOrCreateSPIRVVectorType(SPIRVTypeInst BaseType, unsigned NumElements, MachineIRBuilder &MIRBuilder, bool EmitIR)
Definition SPIRVGlobalRegistry.cpp:1955
llvm::SPIRVGlobalRegistry::buildGlobalVariable
Register buildGlobalVariable(Register Reg, SPIRVTypeInst BaseType, StringRef Name, const GlobalValue *GV, SPIRV::StorageClass::StorageClass Storage, const MachineInstr *Init, bool IsConst, const std::optional< SPIRV::LinkageType::LinkageType > &LinkageType, MachineIRBuilder &MIRBuilder, bool IsInstSelector)
Definition SPIRVGlobalRegistry.cpp:805
llvm::SPIRVGlobalRegistry::getResultType
SPIRVTypeInst getResultType(Register VReg, MachineFunction *MF=nullptr)
Definition SPIRVGlobalRegistry.cpp:1332
llvm::SPIRVGlobalRegistry::getScalarOrVectorComponentCount
unsigned getScalarOrVectorComponentCount(Register VReg) const
Definition SPIRVGlobalRegistry.cpp:1422
llvm::SPIRVGlobalRegistry::getTypeForSPIRVType
const Type * getTypeForSPIRVType(SPIRVTypeInst Ty) const
Definition SPIRVGlobalRegistry.h:324
llvm::SPIRVGlobalRegistry::isBitcastCompatible
bool isBitcastCompatible(SPIRVTypeInst Type1, SPIRVTypeInst Type2) const
Definition SPIRVGlobalRegistry.cpp:1494
llvm::SPIRVGlobalRegistry::getOrCreateConstFP
Register getOrCreateConstFP(APFloat Val, MachineInstr &I, SPIRVTypeInst SpvType, const SPIRVInstrInfo &TII, bool ZeroAsNull=true)
Definition SPIRVGlobalRegistry.cpp:359
llvm::SPIRVGlobalRegistry::getRegType
LLT getRegType(SPIRVTypeInst SpvType) const
Definition SPIRVGlobalRegistry.cpp:2114
llvm::SPIRVGlobalRegistry::invalidateMachineInstr
void invalidateMachineInstr(MachineInstr *MI)
Definition SPIRVGlobalRegistry.cpp:233
llvm::SPIRVGlobalRegistry::getOrCreateSPIRVBoolType
SPIRVTypeInst getOrCreateSPIRVBoolType(MachineIRBuilder &MIRBuilder, bool EmitIR)
Definition SPIRVGlobalRegistry.cpp:1930
llvm::SPIRVGlobalRegistry::getOrCreateSPIRVPointerType
SPIRVTypeInst getOrCreateSPIRVPointerType(const Type *BaseType, MachineIRBuilder &MIRBuilder, SPIRV::StorageClass::StorageClass SC)
Definition SPIRVGlobalRegistry.cpp:1996
llvm::SPIRVGlobalRegistry::isScalarOfType
bool isScalarOfType(Register VReg, unsigned TypeOpcode) const
Definition SPIRVGlobalRegistry.cpp:1387
llvm::SPIRVGlobalRegistry::getSPIRVTypeID
Register getSPIRVTypeID(SPIRVTypeInst SpirvType) const
Definition SPIRVGlobalRegistry.cpp:1159
llvm::SPIRVGlobalRegistry::getOrCreateConstInt
Register getOrCreateConstInt(uint64_t Val, MachineInstr &I, SPIRVTypeInst SpvType, const SPIRVInstrInfo &TII, bool ZeroAsNull=true)
Definition SPIRVGlobalRegistry.cpp:412
llvm::SPIRVGlobalRegistry::getOrCreateConstIntArray
Register getOrCreateConstIntArray(uint64_t Val, size_t Num, MachineInstr &I, SPIRVTypeInst SpvType, const SPIRVInstrInfo &TII)
Definition SPIRVGlobalRegistry.cpp:671
llvm::SPIRVGlobalRegistry::findValueAttrs
bool findValueAttrs(const MachineInstr *Key, Type *&Ty, StringRef &Name)
Definition SPIRVGlobalRegistry.h:176
llvm::SPIRVGlobalRegistry::retrieveScalarOrVectorIntType
SPIRVTypeInst retrieveScalarOrVectorIntType(SPIRVTypeInst Type) const
Definition SPIRVGlobalRegistry.cpp:1472
llvm::SPIRVGlobalRegistry::getOrCreateGlobalVariableWithBinding
Register getOrCreateGlobalVariableWithBinding(SPIRVTypeInst VarType, uint32_t Set, uint32_t Binding, StringRef Name, MachineIRBuilder &MIRBuilder)
Definition SPIRVGlobalRegistry.cpp:917
llvm::SPIRVGlobalRegistry::changePointerStorageClass
SPIRVTypeInst changePointerStorageClass(SPIRVTypeInst PtrType, SPIRV::StorageClass::StorageClass SC, MachineInstr &I)
Definition SPIRVGlobalRegistry.cpp:2007
llvm::SPIRVGlobalRegistry::getOrCreateConstVector
Register getOrCreateConstVector(uint64_t Val, MachineInstr &I, SPIRVTypeInst SpvType, const SPIRVInstrInfo &TII, bool ZeroAsNull=true)
Definition SPIRVGlobalRegistry.cpp:622
llvm::SPIRVGlobalRegistry::buildConstantFP
Register buildConstantFP(APFloat Val, MachineIRBuilder &MIRBuilder, SPIRVTypeInst SpvType=nullptr)
Definition SPIRVGlobalRegistry.cpp:526
llvm::SPIRVGlobalRegistry::addGlobalObject
void addGlobalObject(const Value *V, const MachineFunction *MF, Register R)
Definition SPIRVGlobalRegistry.h:122
llvm::SPIRVGlobalRegistry::getScalarOrVectorComponentType
SPIRVTypeInst getScalarOrVectorComponentType(SPIRVTypeInst Type) const
Definition SPIRVGlobalRegistry.cpp:1436
llvm::SPIRVGlobalRegistry::recordFunctionPointer
void recordFunctionPointer(const MachineOperand *MO, const Function *F)
Definition SPIRVGlobalRegistry.h:259
llvm::SPIRVGlobalRegistry::getOrCreateSPIRVFloatType
SPIRVTypeInst getOrCreateSPIRVFloatType(unsigned BitWidth, MachineInstr &I, const SPIRVInstrInfo &TII)
Definition SPIRVGlobalRegistry.cpp:1909
llvm::SPIRVGlobalRegistry::getPointeeType
SPIRVTypeInst getPointeeType(SPIRVTypeInst PtrType)
Definition SPIRVGlobalRegistry.cpp:1483
llvm::SPIRVGlobalRegistry::getOrCreateSPIRVType
SPIRVTypeInst getOrCreateSPIRVType(const Type *Type, MachineInstr &I, SPIRV::AccessQualifier::AccessQualifier AQ, bool EmitIR)
Definition SPIRVGlobalRegistry.h:310
llvm::SPIRVGlobalRegistry::isScalarOrVectorOfType
bool isScalarOrVectorOfType(Register VReg, unsigned TypeOpcode) const
Definition SPIRVGlobalRegistry.cpp:1394
llvm::SPIRVGlobalRegistry::setCurrentFunc
MachineFunction * setCurrentFunc(MachineFunction &MF)
Definition SPIRVGlobalRegistry.h:364
llvm::SPIRVGlobalRegistry::getOrCreateConstNullPtr
Register getOrCreateConstNullPtr(MachineIRBuilder &MIRBuilder, SPIRVTypeInst SpvType)
Definition SPIRVGlobalRegistry.cpp:754
llvm::SPIRVGlobalRegistry::getSPIRVTypeForVReg
SPIRVTypeInst getSPIRVTypeForVReg(Register VReg, const MachineFunction *MF=nullptr) const
Definition SPIRVGlobalRegistry.cpp:1321
llvm::SPIRVGlobalRegistry::getDeducedGlobalValueType
Type * getDeducedGlobalValueType(const GlobalValue *Global)
Definition SPIRVGlobalRegistry.h:212
llvm::SPIRVGlobalRegistry::getOrCreateUndef
Register getOrCreateUndef(MachineInstr &I, SPIRVTypeInst SpvType, const SPIRVInstrInfo &TII)
Definition SPIRVGlobalRegistry.cpp:2054
llvm::SPIRVGlobalRegistry::getPointerStorageClass
SPIRV::StorageClass::StorageClass getPointerStorageClass(Register VReg) const
Definition SPIRVGlobalRegistry.cpp:1514
llvm::SPIRVIRMapping::erase
bool erase(const MachineInstr *MI)
Definition SPIRVIRMapping.h:211
llvm::SPIRVIRMapping::add
bool add(SPIRV::IRHandle Handle, const MachineInstr *MI)
Definition SPIRVIRMapping.h:189
llvm::SPIRVIRMapping::find
Register find(SPIRV::IRHandle Handle, const MachineFunction *MF)
Definition SPIRVIRMapping.h:235
llvm::SPIRVSubtarget::isPhysicalSPIRV
bool isPhysicalSPIRV() const
Definition SPIRVSubtarget.h:92
llvm::SPIRVSubtarget::isAtLeastSPIRVVer
bool isAtLeastSPIRVVer(VersionTuple VerToCompareTo) const
Definition SPIRVSubtarget.cpp:143
llvm::SPIRVSubtarget::canUseExtInstSet
bool canUseExtInstSet(SPIRV::InstructionSet::InstructionSet E) const
Definition SPIRVSubtarget.cpp:130
llvm::SPIRVSubtarget::isLogicalSPIRV
bool isLogicalSPIRV() const
Definition SPIRVSubtarget.h:89
llvm::SPIRVSubtarget::canUseExtension
bool canUseExtension(SPIRV::Extension::Extension E) const
Definition SPIRVSubtarget.cpp:126
llvm::SPIRVTypeInst::isTypeIntOrFloat
bool isTypeIntOrFloat() const
Definition SPIRVTypeInst.h:57
llvm::SmallPtrSetImpl::erase
bool erase(PtrType Ptr)
Remove pointer from the set.
Definition SmallPtrSet.h:404
llvm::SmallPtrSetImpl::insert
std::pair< iterator, bool > insert(PtrType Ptr)
Inserts Ptr if and only if there is no element in the container equal to Ptr.
Definition SmallPtrSet.h:389
llvm::SmallPtrSetImpl::contains
bool contains(ConstPtrType Ptr) const
Definition SmallPtrSet.h:461
llvm::SmallPtrSet
SmallPtrSet - This class implements a set which is optimized for holding SmallSize or less elements.
Definition SmallPtrSet.h:527
llvm::SmallVectorImpl::emplace_back
reference emplace_back(ArgTypes &&... Args)
Definition SmallVector.h:964
llvm::SmallVectorImpl::reserve
void reserve(size_type N)
Definition SmallVector.h:669
llvm::SmallVectorTemplateBase::push_back
void push_back(const T &Elt)
Definition SmallVector.h:422
llvm::SmallVector
This is a 'vector' (really, a variable-sized array), optimized for the case when the array is small.
Definition SmallVector.h:1223
llvm::StringRef::size
constexpr size_t size() const
size - Get the string size.
Definition StringRef.h:143
llvm::StructType::get
static LLVM_ABI StructType * get(LLVMContext &Context, ArrayRef< Type * > Elements, bool isPacked=false)
This static method is the primary way to create a literal StructType.
Definition Type.cpp:483
llvm::Twine
Twine - A lightweight data structure for efficiently representing the concatenation of temporary valu...
Definition Twine.h:82
llvm::Type
The instances of the Type class are immutable: once they are created, they are never changed.
Definition Type.h:46
llvm::Type::HalfTyID
@ HalfTyID
16-bit floating point type
Definition Type.h:57
llvm::Type::FloatTyID
@ FloatTyID
32-bit floating point type
Definition Type.h:59
llvm::Type::DoubleTyID
@ DoubleTyID
64-bit floating point type
Definition Type.h:60
llvm::Type::getScalarType
Type * getScalarType() const
If this is a vector type, return the element type, otherwise return 'this'.
Definition Type.h:370
llvm::Type::isStructTy
bool isStructTy() const
True if this is an instance of StructType.
Definition Type.h:278
llvm::Type::isAggregateType
bool isAggregateType() const
Return true if the type is an aggregate type.
Definition Type.h:321
llvm::Type::getTypeID
TypeID getTypeID() const
Return the type id for the type.
Definition Type.h:138
llvm::User::getOperand
Value * getOperand(unsigned i) const
Definition User.h:207
llvm::Value::hasName
bool hasName() const
Definition Value.h:261
llvm::Value::getName
LLVM_ABI StringRef getName() const
Return a constant reference to the value's name.
Definition Value.cpp:318
llvm::ilist_node_impl::getIterator
self_iterator getIterator()
Definition ilist_node.h:123
llvm::ilist_node_with_parent::getNextNode
NodeTy * getNextNode()
Get the next node, or nullptr for the list tail.
Definition ilist_node.h:348
uint32_t
uint64_t
llvm_unreachable
#define llvm_unreachable(msg)
Marks that the current location is not supposed to be reachable.
Definition ErrorHandling.h:164
llvm::AMDGPU::HSAMD::Kernel::Arg::Key::IsConst
constexpr char IsConst[]
Key for Kernel::Arg::Metadata::mIsConst.
Definition AMDGPUMetadata.h:197
llvm::BitmaskEnumDetail::Mask
constexpr std::underlying_type_t< E > Mask()
Get a bitmask with 1s in all places up to the high-order bit of E's largest value.
Definition BitmaskEnum.h:126
llvm::CallingConv::ID
unsigned ID
LLVM IR allows to use arbitrary numbers as calling convention identifiers.
Definition CallingConv.h:24
llvm::dwarf_linker::DebugSectionKind::DebugLoc
@ DebugLoc
Definition DWARFLinkerBase.h:34
llvm::ms_demangle::QualifierMangleMode::Result
@ Result
Definition MicrosoftDemangle.h:132
llvm::rdf::Def
NodeAddr< DefNode * > Def
Definition RDFGraph.h:384
llvm::rdf::Instr
NodeAddr< InstrNode * > Instr
Definition RDFGraph.h:389
llvm::rdf::Use
NodeAddr< UseNode * > Use
Definition RDFGraph.h:385
llvm::rdf::Func
NodeAddr< FuncNode * > Func
Definition RDFGraph.h:393
llvm::sframe::BaseReg
BaseReg
Stack frame base register. Bit 0 of FREInfo.Info.
Definition SFrame.h:77
llvm
This is an optimization pass for GlobalISel generic memory operations.
Definition FunctionInfo.h:25
llvm::buildOpName
void buildOpName(Register Target, const StringRef &Name, MachineIRBuilder &MIRBuilder)
Definition SPIRVUtils.cpp:220
llvm::ThreadPriority::Low
@ Low
Lower the current thread's priority such that it does not affect foreground tasks significantly.
Definition Threading.h:280
llvm::Offset
@ Offset
Definition DWP.cpp:557
llvm::Value
FunctionAddr VTableAddr Value
Definition InstrProf.h:137
llvm::all_of
bool all_of(R &&range, UnaryPredicate P)
Provide wrappers to std::all_of which take ranges instead of having to pass begin/end explicitly.
Definition STLExtras.h:1738
llvm::getOpVariableMBBIt
MachineBasicBlock::iterator getOpVariableMBBIt(MachineFunction &MF)
Definition SPIRVUtils.cpp:332
llvm::getIConstValSext
int64_t getIConstValSext(Register ConstReg, const MachineRegisterInfo *MRI)
Definition SPIRVUtils.cpp:502
llvm::BuildMI
MachineInstrBuilder BuildMI(MachineFunction &MF, const MIMetadata &MIMD, const MCInstrDesc &MCID)
Builder interface. Specify how to create the initial instruction itself.
Definition MachineInstrBuilder.h:449
llvm::isTypeFoldingSupported
bool isTypeFoldingSupported(unsigned Opcode)
Definition SPIRVUtils.cpp:1104
llvm::dyn_cast
decltype(auto) dyn_cast(const From &Val)
dyn_cast<X> - Return the argument parameter cast to the specified type.
Definition Casting.h:643
llvm::Int32Ty
FunctionAddr VTableAddr uintptr_t uintptr_t Int32Ty
Definition InstrProf.h:328
llvm::addNumImm
void addNumImm(const APInt &Imm, MachineInstrBuilder &MIB)
Definition SPIRVUtils.cpp:189
llvm::salvageDebugInfo
LLVM_ABI void salvageDebugInfo(const MachineRegisterInfo &MRI, MachineInstr &MI)
Assuming the instruction MI is going to be deleted, attempt to salvage debug users of MI by writing t...
Definition Utils.cpp:1682
llvm::constrainSelectedInstRegOperands
LLVM_ABI void constrainSelectedInstRegOperands(MachineInstr &I, const TargetInstrInfo &TII, const TargetRegisterInfo &TRI, const RegisterBankInfo &RBI)
Mutate the newly-selected instruction I to constrain its (possibly generic) virtual register operands...
Definition Utils.cpp:155
llvm::isPreISelGenericOpcode
bool isPreISelGenericOpcode(unsigned Opcode)
Check whether the given Opcode is a generic opcode that is not supposed to appear after ISel.
Definition TargetOpcodes.h:30
llvm::createVirtualRegister
Register createVirtualRegister(SPIRVTypeInst SpvType, SPIRVGlobalRegistry *GR, MachineRegisterInfo *MRI, const MachineFunction &MF)
Definition SPIRVUtils.cpp:930
llvm::getArrayComponentCount
unsigned getArrayComponentCount(const MachineRegisterInfo *MRI, const MachineInstr *ResType)
Definition SPIRVUtils.cpp:1141
llvm::getIConstVal
uint64_t getIConstVal(Register ConstReg, const MachineRegisterInfo *MRI)
Definition SPIRVUtils.cpp:496
llvm::createContinuedInstructions
SmallVector< MachineInstr *, 4 > createContinuedInstructions(MachineIRBuilder &MIRBuilder, unsigned Opcode, unsigned MinWC, unsigned ContinuedOpcode, ArrayRef< Register > Args, Register ReturnRegister, Register TypeID)
Definition SPIRVUtils.cpp:995
llvm::getMemSemanticsForStorageClass
SPIRV::MemorySemantics::MemorySemantics getMemSemanticsForStorageClass(SPIRV::StorageClass::StorageClass SC)
Definition SPIRVUtils.cpp:411
llvm::storageClassToAddressSpace
constexpr unsigned storageClassToAddressSpace(SPIRV::StorageClass::StorageClass SC)
Definition SPIRVUtils.h:251
llvm::any_of
bool any_of(R &&range, UnaryPredicate P)
Provide wrappers to std::any_of which take ranges instead of having to pass begin/end explicitly.
Definition STLExtras.h:1745
llvm::buildOpDecorate
void buildOpDecorate(Register Reg, MachineIRBuilder &MIRBuilder, SPIRV::Decoration::Decoration Dec, const std::vector< uint32_t > &DecArgs, StringRef StrImm)
Definition SPIRVUtils.cpp:247
llvm::getImm
MachineInstr * getImm(const MachineOperand &MO, const MachineRegisterInfo *MRI)
Definition SPIRVUtils.cpp:1122
llvm::isPowerOf2_32
constexpr bool isPowerOf2_32(uint32_t Value)
Return true if the argument is a power of two > 0.
Definition MathExtras.h:279
llvm::toTypedPointer
Type * toTypedPointer(Type *Ty)
Definition SPIRVUtils.h:469
llvm::dbgs
LLVM_ABI raw_ostream & dbgs()
dbgs() - This returns a reference to a raw_ostream for debugging messages.
Definition Debug.cpp:207
llvm::report_fatal_error
LLVM_ABI void report_fatal_error(Error Err, bool gen_crash_diag=true)
Definition Error.cpp:163
llvm::isGenericCastablePtr
constexpr bool isGenericCastablePtr(SPIRV::StorageClass::StorageClass SC)
Definition SPIRVUtils.h:236
llvm::SmallVector
class LLVM_GSL_OWNER SmallVector
Forward declaration of SmallVector so that calculateSmallVectorDefaultInlinedElements can reference s...
Definition SmallVector.h:1149
llvm::passCopy
MachineInstr * passCopy(MachineInstr *Def, const MachineRegisterInfo *MRI)
Definition SPIRVUtils.cpp:1109
llvm::isa
bool isa(const From &Val)
isa<X> - Return true if the parameter to the template is an instance of one of the template type argu...
Definition Casting.h:547
llvm::getSpirvLinkageTypeFor
std::optional< SPIRV::LinkageType::LinkageType > getSpirvLinkageTypeFor(const SPIRVSubtarget &ST, const GlobalValue &GV)
Definition SPIRVUtils.cpp:1216
llvm::errs
LLVM_ABI raw_fd_ostream & errs()
This returns a reference to a raw_ostream for standard error.
Definition raw_ostream.cpp:904
llvm::addressSpaceToStorageClass
SPIRV::StorageClass::StorageClass addressSpaceToStorageClass(unsigned AddrSpace, const SPIRVSubtarget &STI)
Definition SPIRVUtils.cpp:371
llvm::AtomicOrdering
AtomicOrdering
Atomic ordering for LLVM's memory model.
Definition AtomicOrdering.h:56
llvm::getMemScope
SPIRV::Scope::Scope getMemScope(LLVMContext &Ctx, SyncScope::ID Id)
Definition SPIRVUtils.cpp:447
llvm::createSPIRVInstructionSelector
InstructionSelector * createSPIRVInstructionSelector(const SPIRVTargetMachine &TM, const SPIRVSubtarget &Subtarget, const RegisterBankInfo &RBI)
Definition SPIRVInstructionSelector.cpp:6838
llvm::getStringValueFromReg
std::string getStringValueFromReg(Register Reg, MachineRegisterInfo &MRI)
Definition SPIRVUtils.cpp:179
llvm::foldImm
int64_t foldImm(const MachineOperand &MO, const MachineRegisterInfo *MRI)
Definition SPIRVUtils.cpp:1131
llvm::Op
DWARFExpression::Operation Op
Definition DWARFExpressionPrinter.cpp:23
llvm::ArrayRef
ArrayRef(const T &OneElt) -> ArrayRef< T >
llvm::getDefInstrMaybeConstant
MachineInstr * getDefInstrMaybeConstant(Register &ConstReg, const MachineRegisterInfo *MRI)
Definition SPIRVUtils.cpp:473
llvm::BitWidth
constexpr unsigned BitWidth
Definition BitmaskEnum.h:219
llvm::cast
decltype(auto) cast(const From &Val)
cast<X> - Return the argument parameter cast to the specified type.
Definition Casting.h:559
llvm::hasInitializer
bool hasInitializer(const GlobalVariable *GV)
Definition SPIRVUtils.h:353
llvm::isSpvIntrinsic
bool isSpvIntrinsic(const MachineInstr &MI, Intrinsic::ID IntrinsicID)
Definition SPIRVUtils.cpp:508
llvm::addStringImm
void addStringImm(const StringRef &Str, MCInst &Inst)
Definition SPIRVUtils.cpp:150
llvm::getVRegDef
MachineInstr * getVRegDef(MachineRegisterInfo &MRI, Register Reg)
Definition SPIRVUtils.cpp:881
llvm::getMemSemantics
SPIRV::MemorySemantics::MemorySemantics getMemSemantics(AtomicOrdering Ord)
Definition SPIRVUtils.cpp:429
llvm::getLinkStringForBuiltIn
std::string getLinkStringForBuiltIn(SPIRV::BuiltIn::BuiltIn BuiltInValue)
Definition SPIRVBaseInfo.cpp:202
llvm::isTriviallyDead
LLVM_ABI bool isTriviallyDead(const MachineInstr &MI, const MachineRegisterInfo &MRI)
Check whether an instruction MI is dead: it only defines dead virtual registers, and doesn't have oth...
Definition Utils.cpp:220
N
#define N
Generated on for LLVM by doxygen 1.14.0