1//===-- SPIRVLegalizePointerCast.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// The LLVM IR has multiple legal patterns we cannot lower to Logical SPIR-V.
10// This pass modifies such loads to have an IR we can directly lower to valid
11// logical SPIR-V.
12// OpenCL can avoid this because they rely on ptrcast, which is not supported
13// by logical SPIR-V.
14//
15// This pass relies on the assign_ptr_type intrinsic to deduce the type of the
16// pointed values, must replace all occurences of `ptrcast`. This is why
17// unhandled cases are reported as unreachable: we MUST cover all cases.
18//
19// 1. Loading the first element of an array
20//
21// %array = [10 x i32]
22// %value = load i32, ptr %array
23//
24// LLVM can skip the GEP instruction, and only request loading the first 4
25// bytes. In logical SPIR-V, we need an OpAccessChain to access the first
26// element. This pass will add a getelementptr instruction before the load.
27//
28//
29// 2. Implicit downcast from load
30//
31// %1 = getelementptr <4 x i32>, ptr %vec4, i64 0
32// %2 = load <3 x i32>, ptr %1
33//
34// The pointer in the GEP instruction is only used for offset computations,
35// but it doesn't NEED to match the pointed type. OpAccessChain however
36// requires this. Also, LLVM loads define the bitwidth of the load, not the
37// pointer. In this example, we can guess %vec4 is a vec4 thanks to the GEP
38// instruction basetype, but we only want to load the first 3 elements, hence
39// do a partial load. In logical SPIR-V, this is not legal. What we must do
40// is load the full vector (basetype), extract 3 elements, and recombine them
41// to form a 3-element vector.
42//
43//===----------------------------------------------------------------------===//
44
45#include "SPIRV.h"
46#include "SPIRVSubtarget.h"
47#include "SPIRVTargetMachine.h"
48#include "SPIRVUtils.h"
49#include "llvm/IR/IRBuilder.h"
50#include "llvm/IR/IntrinsicInst.h"
51#include "llvm/IR/Intrinsics.h"
52#include "llvm/IR/IntrinsicsSPIRV.h"
53#include "llvm/Transforms/Utils/Cloning.h"
54#include "llvm/Transforms/Utils/LowerMemIntrinsics.h"
55
56using namespace llvm;
57
58namespace {
59class SPIRVLegalizePointerCastImpl {
60
61 // Builds the `spv_assign_type` assigning |Ty| to |Value| at the current
62 // builder position.
63 void buildAssignType(IRBuilder<> &B, Type *Ty, Value *Arg) {
64 Value *OfType = PoisonValue::get(T: Ty);
65 CallInst *AssignCI = buildIntrWithMD(IntrID: Intrinsic::spv_assign_type,
66 Types: {Arg->getType()}, Arg: OfType, Arg2: Arg, Imms: {}, B);
67 GR->addAssignPtrTypeInstr(Val: Arg, AssignPtrTyCI: AssignCI);
68 }
69
70 static FixedVectorType *makeVectorFromTotalBits(Type *ElemTy,
71 TypeSize TotalBits) {
72 unsigned ElemBits = ElemTy->getScalarSizeInBits();
73 assert(ElemBits && TotalBits % ElemBits == 0 &&
74 "TotalBits must be divisible by element bit size");
75 return FixedVectorType::get(ElementType: ElemTy, NumElts: TotalBits / ElemBits);
76 }
77
78 Value *resizeVectorBitsWithShuffle(IRBuilder<> &B, Value *V,
79 FixedVectorType *DstTy) {
80 auto *SrcTy = cast<FixedVectorType>(Val: V->getType());
81 assert(SrcTy->getElementType() == DstTy->getElementType() &&
82 "shuffle resize expects identical element types");
83
84 const unsigned NumNeeded = DstTy->getNumElements();
85 const unsigned NumSource = SrcTy->getNumElements();
86
87 SmallVector<int> Mask(NumNeeded);
88 for (unsigned I = 0; I < NumNeeded; ++I)
89 Mask[I] = (I < NumSource) ? static_cast<int>(I) : -1;
90
91 Value *Resized = B.CreateShuffleVector(V1: V, V2: V, Mask);
92 buildAssignType(B, Ty: DstTy, Arg: Resized);
93 return Resized;
94 }
95
96 // Loads parts of the vector of type |SourceType| from the pointer |Source|
97 // and create a new vector of type |TargetType|. |TargetType| must be a vector
98 // type.
99 // Returns the loaded value.
100 Value *loadVectorFromVector(IRBuilder<> &B, FixedVectorType *SourceType,
101 FixedVectorType *TargetType, Value *Source,
102 Align OriginalAlign) {
103 LoadInst *NewLoad = B.CreateLoad(Ty: SourceType, Ptr: Source);
104 NewLoad->setAlignment(OriginalAlign);
105 buildAssignType(B, Ty: SourceType, Arg: NewLoad);
106 Value *AssignValue = NewLoad;
107 if (TargetType->getElementType() != SourceType->getElementType()) {
108 const DataLayout &DL = B.getDataLayout();
109 TypeSize TargetTypeSize = DL.getTypeSizeInBits(Ty: TargetType);
110 TypeSize SourceTypeSize = DL.getTypeSizeInBits(Ty: SourceType);
111
112 Value *BitcastSrcVal = NewLoad;
113 FixedVectorType *BitcastSrcTy =
114 cast<FixedVectorType>(Val: BitcastSrcVal->getType());
115 FixedVectorType *BitcastDstTy = TargetType;
116
117 if (TargetTypeSize != SourceTypeSize) {
118 unsigned TargetElemBits =
119 TargetType->getElementType()->getScalarSizeInBits();
120 if (SourceTypeSize % TargetElemBits == 0) {
121 // No Resize needed. Same total bits as source, but use target element
122 // type.
123 BitcastDstTy = makeVectorFromTotalBits(ElemTy: TargetType->getElementType(),
124 TotalBits: SourceTypeSize);
125 } else {
126 // Resize source to target total bitwidth using source element type.
127 BitcastSrcTy = makeVectorFromTotalBits(ElemTy: SourceType->getElementType(),
128 TotalBits: TargetTypeSize);
129 BitcastSrcVal = resizeVectorBitsWithShuffle(B, V: NewLoad, DstTy: BitcastSrcTy);
130 }
131 }
132 AssignValue =
133 B.CreateIntrinsic(ID: Intrinsic::spv_bitcast,
134 OverloadTypes: {BitcastDstTy, BitcastSrcTy}, Args: {BitcastSrcVal});
135 buildAssignType(B, Ty: BitcastDstTy, Arg: AssignValue);
136 if (BitcastDstTy == TargetType)
137 return AssignValue;
138 }
139
140 auto *AssignVecTy = cast<FixedVectorType>(Val: AssignValue->getType());
141 const unsigned NumTarget = TargetType->getNumElements();
142 const unsigned NumSource = AssignVecTy->getNumElements();
143
144 // Optimizations may widen a narrow load to cover padding (e.g., loading a
145 // <1 x float> column as <4 x float>). Since extra lanes read trailing
146 // padding, insert only the valid lanes into a poison vector to avoid poison
147 // scalars.
148 if (NumTarget > NumSource) {
149 Value *Result = PoisonValue::get(T: TargetType);
150 buildAssignType(B, Ty: TargetType, Arg: Result);
151 for (unsigned I = 0; I < NumSource; ++I) {
152 Value *Scalar = extractScalarFromVector(B, Vector: AssignValue, Index: I);
153 Result = makeInsertElement(B, Vector: Result, Element: Scalar, Index: I);
154 }
155 return Result;
156 }
157
158 assert(NumTarget < NumSource);
159 SmallVector<int> Mask(/* Size= */ NumTarget);
160 for (unsigned I = 0; I < NumTarget; ++I)
161 Mask[I] = I;
162 Value *Output = B.CreateShuffleVector(V1: AssignValue, V2: AssignValue, Mask);
163 buildAssignType(B, Ty: TargetType, Arg: Output);
164 return Output;
165 }
166
167 // Returns true if |FromTy| has a memory layout compatible with loading or
168 // storing |ToTy|.
169 bool isCompatibleMemoryLayout(Type *ToTy, Type *FromTy) {
170 if (ToTy == FromTy)
171 return true;
172 auto *SVT = dyn_cast<FixedVectorType>(Val: FromTy);
173 auto *DVT = dyn_cast<FixedVectorType>(Val: ToTy);
174 if (SVT && DVT)
175 return true;
176 auto *SAT = dyn_cast<ArrayType>(Val: FromTy);
177 if (SAT && DVT) {
178 if (SAT->getElementType() == DVT->getElementType())
179 return true;
180 if (auto *MAT = dyn_cast<FixedVectorType>(Val: SAT->getElementType()))
181 if (MAT->getElementType() == DVT->getElementType())
182 return true;
183 }
184 return false;
185 }
186
187 // Traverses the aggregate type to find the first sub-type that matches
188 // the TargetElemType's memory layout, optionally emitting a GEP intrinsic.
189 std::optional<std::pair<Value *, Type *>>
190 getPointerToFirstCompatibleType(IRBuilder<> &B, Value *BasePtr,
191 Type *PointerType, Type *TargetElemType,
192 bool IsInBounds) {
193 Type *CurrentTy = GR->findDeducedElementType(Val: BasePtr);
194 assert(CurrentTy && "Could not deduce aggregate type");
195 SmallVector<Value *, 8> Args{/* isInBounds= */ B.getInt1(V: IsInBounds),
196 BasePtr};
197 Args.push_back(Elt: B.getInt32(C: 0)); // Pointer offset
198
199 while (!isCompatibleMemoryLayout(ToTy: TargetElemType, FromTy: CurrentTy)) {
200 if (auto *ST = dyn_cast<StructType>(Val: CurrentTy)) {
201 if (ST->getNumElements() == 0)
202 return std::nullopt;
203 CurrentTy = ST->getTypeAtIndex(N: 0u);
204 } else if (auto *AT = dyn_cast<ArrayType>(Val: CurrentTy)) {
205 CurrentTy = AT->getElementType();
206 } else if (auto *VT = dyn_cast<FixedVectorType>(Val: CurrentTy)) {
207 CurrentTy = VT->getElementType();
208 } else {
209 return std::nullopt;
210 }
211 Args.push_back(Elt: B.getInt32(C: 0));
212 }
213
214 Value *GEP = BasePtr;
215 if (Args.size() > 3) {
216 std::array<Type *, 2> Types = {PointerType, BasePtr->getType()};
217 GEP = B.CreateIntrinsic(ID: Intrinsic::spv_gep, OverloadTypes: {Types}, Args: {Args});
218 GR->buildAssignPtr(B, ElemTy: CurrentTy, Arg: GEP);
219 }
220
221 return std::make_pair(x&: GEP, y&: CurrentTy);
222 }
223
224 static IntrinsicInst *getResourceGetPointer(Value *Ptr) {
225 if (auto *II = dyn_cast<IntrinsicInst>(Val: Ptr))
226 if (II->getIntrinsicID() == Intrinsic::spv_resource_getpointer)
227 return II;
228 return nullptr;
229 }
230
231 Value *gepByteOffset(IRBuilder<> &B, Value *BasePtr, unsigned ByteOffset) {
232 if (ByteOffset == 0)
233 return BasePtr;
234
235 if (IntrinsicInst *ResourcePtr = getResourceGetPointer(Ptr: BasePtr)) {
236 Value *Handle = ResourcePtr->getOperand(i_nocapture: 0);
237 Value *BaseOffset = ResourcePtr->getOperand(i_nocapture: 1);
238 Value *NewOffset;
239 if (auto *CI = dyn_cast<ConstantInt>(Val: BaseOffset)) {
240 NewOffset =
241 ConstantInt::get(Ty: CI->getType(), V: CI->getZExtValue() + ByteOffset);
242 } else {
243 NewOffset = B.CreateAdd(
244 LHS: BaseOffset, RHS: ConstantInt::get(Ty: BaseOffset->getType(), V: ByteOffset));
245 buildAssignType(B, Ty: NewOffset->getType(), Arg: NewOffset);
246 }
247 SmallVector<OperandBundleDef> OpBundles;
248 ResourcePtr->getOperandBundlesAsDefs(Defs&: OpBundles);
249 CallInst *ResourcePtrAtOffset = B.CreateCall(
250 FTy: ResourcePtr->getFunctionType(), Callee: ResourcePtr->getCalledOperand(),
251 Args: {Handle, NewOffset}, OpBundles);
252 ResourcePtrAtOffset->setAttributes(ResourcePtr->getAttributes());
253 ResourcePtrAtOffset->setCallingConv(ResourcePtr->getCallingConv());
254 Type *I8Ty = Type::getInt8Ty(C&: B.getContext());
255 GR->buildAssignPtr(B, ElemTy: I8Ty, Arg: ResourcePtrAtOffset);
256 return ResourcePtrAtOffset;
257 }
258 llvm_unreachable(
259 "byte layout pointer must come from spv.resource.getpointer");
260 }
261
262 Value *scalarToStoreInt(IRBuilder<> &B, Value *Scalar) {
263 Type *Ty = Scalar->getType();
264 const DataLayout &DL = B.getDataLayout();
265 Type *IntTy =
266 IntegerType::get(C&: B.getContext(), NumBits: DL.getTypeStoreSizeInBits(Ty));
267 if (Ty == IntTy)
268 return Scalar;
269 if (Ty->isIntOrIntVectorTy())
270 return B.CreateIntCast(V: Scalar, DestTy: IntTy, /*isSigned=*/false);
271 return B.CreateIntrinsic(ID: Intrinsic::spv_bitcast, OverloadTypes: {IntTy, Ty}, Args: {Scalar});
272 }
273
274 Value *storeIntToScalar(IRBuilder<> &B, Value *IntVal, Type *ScalarTy) {
275 if (IntVal->getType() == ScalarTy)
276 return IntVal;
277 if (ScalarTy->isIntOrIntVectorTy())
278 return B.CreateIntCast(V: IntVal, DestTy: ScalarTy, /*isSigned=*/false);
279 return B.CreateIntrinsic(ID: Intrinsic::spv_bitcast,
280 OverloadTypes: {ScalarTy, IntVal->getType()}, Args: {IntVal});
281 }
282
283 void storeScalarToByteLayout(IRBuilder<> &B, Value *Src, Value *Dst,
284 Align Alignment) {
285 LLVMContext &Ctx = B.getContext();
286 Type *I8Ty = Type::getInt8Ty(C&: Ctx);
287 const DataLayout &DL = B.getDataLayout();
288 Value *IntVal = scalarToStoreInt(B, Scalar: Src);
289 if (IntVal != Src)
290 buildAssignType(B, Ty: IntVal->getType(), Arg: IntVal);
291 unsigned NumBytes = DL.getTypeStoreSize(Ty: Src->getType());
292
293 auto StoreByte = [&](unsigned I, Value *Shifted) {
294 Value *Byte = B.CreateTrunc(V: Shifted, DestTy: I8Ty);
295 buildAssignType(B, Ty: I8Ty, Arg: Byte);
296 Value *Ptr = gepByteOffset(B, BasePtr: Dst, ByteOffset: I);
297 StoreInst *SI = B.CreateStore(Val: Byte, Ptr);
298 SI->setAlignment(commonAlignment(A: Alignment, Offset: I));
299 };
300
301 if (NumBytes > 0)
302 StoreByte(0, IntVal);
303
304 for (unsigned I = 1; I < NumBytes; ++I) {
305 Value *Shifted =
306 B.CreateLShr(LHS: IntVal, RHS: ConstantInt::get(Ty: IntVal->getType(), V: 8 * I));
307 StoreByte(I, Shifted);
308 }
309 }
310
311 Value *loadScalarFromByteLayout(IRBuilder<> &B, Type *AccessTy, Value *Src,
312 Align Alignment) {
313 LLVMContext &Ctx = B.getContext();
314 Type *I8Ty = Type::getInt8Ty(C&: Ctx);
315 const DataLayout &DL = B.getDataLayout();
316 unsigned NumBytes = DL.getTypeStoreSize(Ty: AccessTy);
317 Type *IntTy = IntegerType::get(C&: Ctx, NumBits: DL.getTypeStoreSizeInBits(Ty: AccessTy));
318 Value *IntVal = ConstantInt::get(Ty: IntTy, V: 0);
319
320 for (unsigned I = 0; I < NumBytes; ++I) {
321 Value *Ptr = gepByteOffset(B, BasePtr: Src, ByteOffset: I);
322 LoadInst *LI = B.CreateLoad(Ty: I8Ty, Ptr);
323 LI->setAlignment(commonAlignment(A: Alignment, Offset: I));
324 buildAssignType(B, Ty: I8Ty, Arg: LI);
325 Value *Extended = B.CreateZExt(V: LI, DestTy: IntTy);
326 buildAssignType(B, Ty: IntTy, Arg: Extended);
327
328 if (I == 0) {
329 IntVal = Extended;
330 } else {
331 Value *Shifted = B.CreateShl(LHS: Extended, RHS: ConstantInt::get(Ty: IntTy, V: 8 * I));
332 buildAssignType(B, Ty: IntTy, Arg: Shifted);
333 IntVal = B.CreateOr(LHS: IntVal, RHS: Shifted);
334 }
335 buildAssignType(B, Ty: IntTy, Arg: IntVal);
336 }
337
338 Value *Result = storeIntToScalar(B, IntVal, ScalarTy: AccessTy);
339 if (Result != IntVal)
340 buildAssignType(B, Ty: AccessTy, Arg: Result);
341 return Result;
342 }
343
344 // Classifies a ptrcast reinterpretation: casted pointee matches the access
345 // type but differs from the original storage layout (e.g. i8 byte buffer as
346 // i32). ByteWise means multi-byte access must use per-byte i8 load/store.
347 bool shouldReinterpretByteWise(IRBuilder<> &B, Type *AccessTy,
348 Value *OriginalPtr) {
349 Type *OriginalElemTy = GR->findDeducedElementType(Val: OriginalPtr);
350 const DataLayout &DL = B.getDataLayout();
351 if (OriginalElemTy && OriginalElemTy == Type::getInt8Ty(C&: B.getContext()) &&
352 AccessTy->isSingleValueType() && DL.getTypeStoreSize(Ty: AccessTy) > 1)
353 return true;
354
355 return false;
356 }
357
358 bool tryReinterpretLoad(IRBuilder<> &B, Type *AccessTy, Value *OriginalPtr,
359 Value *CastedPtr, LoadInst *IllegalLoad) {
360 Type *CastedElemTy = GR->findDeducedElementType(Val: CastedPtr);
361 if (!CastedElemTy || CastedElemTy != AccessTy)
362 return false;
363
364 Align Alignment = IllegalLoad->getAlign();
365 if (shouldReinterpretByteWise(B, AccessTy, OriginalPtr)) {
366 const DataLayout &DL = B.getDataLayout();
367 Value *Loaded;
368 if (auto *VT = dyn_cast<FixedVectorType>(Val: AccessTy)) {
369 unsigned ElemSize = DL.getTypeStoreSize(Ty: VT->getElementType());
370 SmallVector<Value *, 4> LoadedElements;
371 for (unsigned I = 0; I < VT->getNumElements(); ++I) {
372 Value *ElemPtr = gepByteOffset(B, BasePtr: OriginalPtr, ByteOffset: I * ElemSize);
373 LoadedElements.push_back(Elt: loadScalarFromByteLayout(
374 B, AccessTy: VT->getElementType(), Src: ElemPtr,
375 Alignment: commonAlignment(A: Alignment, Offset: I * ElemSize)));
376 }
377 Loaded = buildVectorFromLoadedElements(B, TargetType: VT, LoadedElements);
378 } else {
379 Loaded = loadScalarFromByteLayout(B, AccessTy, Src: OriginalPtr, Alignment);
380 buildAssignType(B, Ty: AccessTy, Arg: Loaded);
381 }
382 GR->replaceAllUsesWith(Old: IllegalLoad, New: Loaded, /* DeleteOld= */ true);
383 DeadInstructions.push_back(x: IllegalLoad);
384 return true;
385 }
386
387 GR->buildAssignPtr(B, ElemTy: AccessTy, Arg: OriginalPtr);
388 LoadInst *LI = B.CreateLoad(Ty: AccessTy, Ptr: OriginalPtr);
389 LI->setAlignment(Alignment);
390 buildAssignType(B, Ty: AccessTy, Arg: LI);
391 GR->replaceAllUsesWith(Old: IllegalLoad, New: LI, /* DeleteOld= */ true);
392 DeadInstructions.push_back(x: IllegalLoad);
393 return true;
394 }
395
396 bool tryReinterpretStore(IRBuilder<> &B, Type *AccessTy, Value *OriginalPtr,
397 Value *CastedPtr, Value *StoreSrc, Align Alignment) {
398 Type *CastedElemTy = GR->findDeducedElementType(Val: CastedPtr);
399 if (!CastedElemTy || CastedElemTy != AccessTy)
400 return false;
401
402 if (shouldReinterpretByteWise(B, AccessTy, OriginalPtr)) {
403 const DataLayout &DL = B.getDataLayout();
404 if (auto *VT = dyn_cast<FixedVectorType>(Val: StoreSrc->getType())) {
405 unsigned ElemSize = DL.getTypeStoreSize(Ty: VT->getElementType());
406 for (unsigned I = 0; I < VT->getNumElements(); ++I) {
407 Value *Elem = extractScalarFromVector(B, Vector: StoreSrc, Index: I);
408 Value *ElemPtr = gepByteOffset(B, BasePtr: OriginalPtr, ByteOffset: I * ElemSize);
409 storeScalarToByteLayout(B, Src: Elem, Dst: ElemPtr,
410 Alignment: commonAlignment(A: Alignment, Offset: I * ElemSize));
411 }
412 } else {
413 storeScalarToByteLayout(B, Src: StoreSrc, Dst: OriginalPtr, Alignment);
414 }
415 return true;
416 }
417
418 GR->buildAssignPtr(B, ElemTy: AccessTy, Arg: OriginalPtr);
419 StoreInst *SI = B.CreateStore(Val: StoreSrc, Ptr: OriginalPtr);
420 SI->setAlignment(Alignment);
421 return true;
422 }
423
424 // Builds a legalized load from a pointer, drilling down through
425 // memory layouts to find a compatible type. Load flags will be
426 // copied from |IllegalLoad|, which should be the load being legalized.
427 Value *buildLegalizedLoad(IRBuilder<> &B, Type *ElementType, Value *Source,
428 LoadInst *IllegalLoad, Value *CastedPtr) {
429 auto ResultOpt = getPointerToFirstCompatibleType(
430 B, BasePtr: Source, PointerType: IllegalLoad->getPointerOperandType(), TargetElemType: ElementType, IsInBounds: false);
431 if (!ResultOpt) {
432 if (tryReinterpretLoad(B, AccessTy: ElementType, OriginalPtr: Source, CastedPtr, IllegalLoad))
433 return nullptr;
434 llvm_unreachable("Failed to load from aggregate: "
435 "Could not find compatible memory layout.");
436 }
437 auto [GEP, CurrentTy] = *ResultOpt;
438
439 auto *SAT = dyn_cast<ArrayType>(Val: CurrentTy);
440 auto *SVT = dyn_cast<FixedVectorType>(Val: CurrentTy);
441 auto *DVT = dyn_cast<FixedVectorType>(Val: ElementType);
442 auto *MAT =
443 SAT ? dyn_cast<FixedVectorType>(Val: SAT->getElementType()) : nullptr;
444
445 if (ElementType == CurrentTy) {
446 LoadInst *LI = B.CreateLoad(Ty: ElementType, Ptr: GEP);
447 LI->setAlignment(IllegalLoad->getAlign());
448 buildAssignType(B, Ty: ElementType, Arg: LI);
449 return LI;
450 }
451 if (SVT && DVT)
452 return loadVectorFromVector(B, SourceType: SVT, TargetType: DVT, Source: GEP, OriginalAlign: IllegalLoad->getAlign());
453 if (SAT && DVT && SAT->getElementType() == DVT->getElementType())
454 return loadVectorFromArray(B, TargetType: DVT, Source: GEP, OriginalAlign: IllegalLoad->getAlign());
455 if (MAT && DVT && MAT->getElementType() == DVT->getElementType())
456 return loadVectorFromMatrixArray(B, TargetType: DVT, Source: GEP, ArrElemVecTy: MAT,
457 OriginalAlign: IllegalLoad->getAlign());
458
459 llvm_unreachable("Failed to load from aggregate.");
460 }
461
462 Value *
463 buildVectorFromLoadedElements(IRBuilder<> &B, FixedVectorType *TargetType,
464 SmallVector<Value *, 4> &LoadedElements) {
465 // <1 x T> shares the SPIR-V type with T, so emitting OpCompositeInsert on
466 // a scalar would be invalid. Bridge with spv_bitcast instead unless
467 // SPV_EXT_long_vector is available.
468 bool CanUseAnyVectorRank = TM.getSubtargetImpl()->canUseExtension(
469 E: SPIRV::Extension::SPV_EXT_long_vector);
470 if (TargetType->getNumElements() == 1 && !CanUseAnyVectorRank) {
471 Value *Scalar = LoadedElements[0];
472 Value *NewVector = B.CreateIntrinsic(
473 ID: Intrinsic::spv_bitcast, OverloadTypes: {TargetType, Scalar->getType()}, Args: {Scalar});
474 buildAssignType(B, Ty: TargetType, Arg: NewVector);
475 return NewVector;
476 }
477
478 // Build the vector from the loaded elements.
479 Value *NewVector = PoisonValue::get(T: TargetType);
480 buildAssignType(B, Ty: TargetType, Arg: NewVector);
481
482 for (unsigned I = 0, E = TargetType->getNumElements(); I < E; ++I) {
483 Value *Index = B.getInt32(C: I);
484 SmallVector<Type *, 4> Types = {TargetType, TargetType,
485 TargetType->getElementType(),
486 Index->getType()};
487 SmallVector<Value *> Args = {NewVector, LoadedElements[I], Index};
488 NewVector = B.CreateIntrinsic(ID: Intrinsic::spv_insertelt, OverloadTypes: {Types}, Args: {Args});
489 buildAssignType(B, Ty: TargetType, Arg: NewVector);
490 }
491 return NewVector;
492 }
493
494 // Loads elements from a matrix with an array of vector memory layout and
495 // constructs a vector.
496 Value *loadVectorFromMatrixArray(IRBuilder<> &B, FixedVectorType *TargetType,
497 Value *Source, FixedVectorType *ArrElemVecTy,
498 Align OriginalAlign) {
499 Type *TargetElemTy = TargetType->getElementType();
500 unsigned ScalarsPerArrayElement = ArrElemVecTy->getNumElements();
501 const DataLayout &DL = B.getDataLayout();
502 uint64_t ArrElemVecSize = DL.getTypeAllocSize(Ty: ArrElemVecTy);
503 // Load each element of the array.
504 SmallVector<Value *, 4> LoadedElements;
505 std::array<Type *, 2> Types = {Source->getType(), Source->getType()};
506 for (unsigned I = 0, E = TargetType->getNumElements(); I < E; ++I) {
507 unsigned ArrayIndex = I / ScalarsPerArrayElement;
508 unsigned ElementIndexInArrayElem = I % ScalarsPerArrayElement;
509 // Create a GEP to access the i-th element of the array.
510 std::array<Value *, 4> Args = {
511 B.getInt1(/*Inbounds=*/V: false), Source, B.getInt32(C: 0),
512 ConstantInt::get(Ty: B.getInt32Ty(), V: ArrayIndex)};
513 auto *ElementPtr = B.CreateIntrinsic(ID: Intrinsic::spv_gep, OverloadTypes: {Types}, Args: {Args});
514 GR->buildAssignPtr(B, ElemTy: ArrElemVecTy, Arg: ElementPtr);
515 LoadInst *LoadVec = B.CreateLoad(Ty: ArrElemVecTy, Ptr: ElementPtr);
516 LoadVec->setAlignment(
517 commonAlignment(A: OriginalAlign, Offset: ArrayIndex * ArrElemVecSize));
518 buildAssignType(B, Ty: ArrElemVecTy, Arg: LoadVec);
519 LoadedElements.push_back(Elt: makeExtractElement(B, ElementType: TargetElemTy, Vector: LoadVec,
520 Index: ElementIndexInArrayElem));
521 }
522 return buildVectorFromLoadedElements(B, TargetType, LoadedElements);
523 }
524
525 // Loads elements from an array and constructs a vector.
526 Value *loadVectorFromArray(IRBuilder<> &B, FixedVectorType *TargetType,
527 Value *Source, Align OriginalAlign) {
528 // Load each element of the array.
529 SmallVector<Value *, 4> LoadedElements;
530 std::array<Type *, 2> Types = {Source->getType(), Source->getType()};
531 const DataLayout &DL = B.getDataLayout();
532 uint64_t ElemSize = DL.getTypeAllocSize(Ty: TargetType->getElementType());
533 for (unsigned I = 0, E = TargetType->getNumElements(); I < E; ++I) {
534 // Create a GEP to access the i-th element of the array.
535 std::array<Value *, 4> Args = {B.getInt1(/*Inbounds=*/V: false), Source,
536 B.getInt32(C: 0),
537 ConstantInt::get(Ty: B.getInt32Ty(), V: I)};
538 auto *ElementPtr = B.CreateIntrinsic(ID: Intrinsic::spv_gep, OverloadTypes: {Types}, Args: {Args});
539 GR->buildAssignPtr(B, ElemTy: TargetType->getElementType(), Arg: ElementPtr);
540
541 // Load the value from the element pointer.
542 LoadInst *Load = B.CreateLoad(Ty: TargetType->getElementType(), Ptr: ElementPtr);
543 Load->setAlignment(commonAlignment(A: OriginalAlign, Offset: I * ElemSize));
544 buildAssignType(B, Ty: TargetType->getElementType(), Arg: Load);
545 LoadedElements.push_back(Elt: Load);
546 }
547 return buildVectorFromLoadedElements(B, TargetType, LoadedElements);
548 }
549
550 // Stores elements from a vector into a matrix (an array of vectors).
551 void storeMatrixArrayFromVector(IRBuilder<> &B, Value *SrcVector,
552 Value *DstArrayPtr, ArrayType *ArrTy,
553 Align Alignment) {
554 auto *SrcVecTy = cast<FixedVectorType>(Val: SrcVector->getType());
555 auto *ArrElemVecTy = cast<FixedVectorType>(Val: ArrTy->getElementType());
556 Type *ElemTy = ArrElemVecTy->getElementType();
557 unsigned ScalarsPerArrayElement = ArrElemVecTy->getNumElements();
558 unsigned SrcNumElements = SrcVecTy->getNumElements();
559 assert(
560 SrcNumElements % ScalarsPerArrayElement == 0 &&
561 "Source vector size must be a multiple of array element vector size");
562
563 std::array<Type *, 2> Types = {DstArrayPtr->getType(),
564 DstArrayPtr->getType()};
565 const DataLayout &DL = B.getDataLayout();
566 uint64_t ArrElemVecSize = DL.getTypeAllocSize(Ty: ArrElemVecTy);
567
568 for (unsigned I = 0; I < SrcNumElements; I += ScalarsPerArrayElement) {
569 unsigned ArrayIndex = I / ScalarsPerArrayElement;
570 // Create a GEP to access the array element.
571 std::array<Value *, 4> Args = {
572 B.getInt1(/*Inbounds=*/V: false), DstArrayPtr, B.getInt32(C: 0),
573 ConstantInt::get(Ty: B.getInt32Ty(), V: ArrayIndex)};
574 auto *ElementPtr = B.CreateIntrinsic(ID: Intrinsic::spv_gep, OverloadTypes: {Types}, Args: {Args});
575 GR->buildAssignPtr(B, ElemTy: ArrElemVecTy, Arg: ElementPtr);
576
577 // Extract scalar elements from the source vector for this array slot.
578 SmallVector<Value *, 4> Elements;
579 for (unsigned J = 0; J < ScalarsPerArrayElement; ++J)
580 Elements.push_back(Elt: makeExtractElement(B, ElementType: ElemTy, Vector: SrcVector, Index: I + J));
581
582 // Build a vector from the extracted elements and store it.
583 Value *Vec = buildVectorFromLoadedElements(B, TargetType: ArrElemVecTy, LoadedElements&: Elements);
584 StoreInst *SI = B.CreateStore(Val: Vec, Ptr: ElementPtr);
585 SI->setAlignment(commonAlignment(A: Alignment, Offset: ArrayIndex * ArrElemVecSize));
586 }
587 }
588
589 // Stores elements from a vector into an array.
590 void storeArrayFromVector(IRBuilder<> &B, Value *SrcVector,
591 Value *DstArrayPtr, ArrayType *ArrTy,
592 Align Alignment) {
593 auto *VecTy = cast<FixedVectorType>(Val: SrcVector->getType());
594 Type *ElemTy = ArrTy->getElementType();
595
596 // Ensure the element types of the array and vector are the same.
597 assert(VecTy->getElementType() == ElemTy &&
598 "Element types of array and vector must be the same.");
599 std::array<Type *, 2> Types = {DstArrayPtr->getType(),
600 DstArrayPtr->getType()};
601 const DataLayout &DL = B.getDataLayout();
602 uint64_t ElemSize = DL.getTypeAllocSize(Ty: ElemTy);
603
604 for (unsigned I = 0, E = VecTy->getNumElements(); I < E; ++I) {
605 // Create a GEP to access the i-th element of the array.
606 std::array<Value *, 4> Args = {B.getInt1(/*Inbounds=*/V: false), DstArrayPtr,
607 B.getInt32(C: 0),
608 ConstantInt::get(Ty: B.getInt32Ty(), V: I)};
609 auto *ElementPtr = B.CreateIntrinsic(ID: Intrinsic::spv_gep, OverloadTypes: {Types}, Args: {Args});
610 GR->buildAssignPtr(B, ElemTy, Arg: ElementPtr);
611
612 // Extract the element from the vector and store it.
613 bool CanUseAnyVectorRank = TM.getSubtargetImpl()->canUseExtension(
614 E: SPIRV::Extension::SPV_EXT_long_vector);
615 Value *Element = (E == 1 && !CanUseAnyVectorRank)
616 ? SrcVector
617 : makeExtractElement(B, ElementType: ElemTy, Vector: SrcVector, Index: I);
618 StoreInst *SI = B.CreateStore(Val: Element, Ptr: ElementPtr);
619 SI->setAlignment(commonAlignment(A: Alignment, Offset: I * ElemSize));
620 }
621 }
622
623 // Replaces the load instruction to get rid of the ptrcast used as source
624 // operand.
625 void transformLoad(IRBuilder<> &B, LoadInst *LI, Value *CastedOperand,
626 Value *OriginalOperand) {
627 Type *ToTy = GR->findDeducedElementType(Val: CastedOperand);
628 B.SetInsertPoint(LI);
629
630 Value *Output =
631 buildLegalizedLoad(B, ElementType: ToTy, Source: OriginalOperand, IllegalLoad: LI, CastedPtr: CastedOperand);
632 if (!Output)
633 return;
634
635 GR->replaceAllUsesWith(Old: LI, New: Output, /* DeleteOld= */ true);
636 DeadInstructions.push_back(x: LI);
637 }
638
639 // Creates an spv_insertelt instruction (equivalent to llvm's insertelement).
640 Value *makeInsertElement(IRBuilder<> &B, Value *Vector, Value *Element,
641 unsigned Index) {
642 Type *Int32Ty = Type::getInt32Ty(C&: B.getContext());
643 SmallVector<Type *, 4> Types = {Vector->getType(), Vector->getType(),
644 Element->getType(), Int32Ty};
645 SmallVector<Value *> Args = {Vector, Element, B.getInt32(C: Index)};
646 Value *NewI = B.CreateIntrinsic(ID: Intrinsic::spv_insertelt, OverloadTypes: {Types}, Args: {Args});
647 buildAssignType(B, Ty: Vector->getType(), Arg: NewI);
648 return NewI;
649 }
650
651 // Creates an spv_extractelt instruction (equivalent to llvm's
652 // extractelement).
653 Value *makeExtractElement(IRBuilder<> &B, Type *ElementType, Value *Vector,
654 unsigned Index) {
655 Type *Int32Ty = Type::getInt32Ty(C&: B.getContext());
656 SmallVector<Type *, 3> Types = {ElementType, Vector->getType(), Int32Ty};
657 SmallVector<Value *> Args = {Vector, B.getInt32(C: Index)};
658 Value *NewI = B.CreateIntrinsic(ID: Intrinsic::spv_extractelt, OverloadTypes: {Types}, Args: {Args});
659 buildAssignType(B, Ty: ElementType, Arg: NewI);
660 return NewI;
661 }
662
663 // Extracts scalar element |Index| from |Vector|. A <1 x T> vector shares its
664 // SPIR-V type with the scalar T, so a plain extractelement would be invalid;
665 // bridge it with spv_bitcast instead.
666 Value *extractScalarFromVector(IRBuilder<> &B, Value *Vector,
667 unsigned Index) {
668 auto *VecTy = cast<FixedVectorType>(Val: Vector->getType());
669 Type *ElemTy = VecTy->getElementType();
670 if (VecTy->getNumElements() == 1) {
671 Value *Scalar =
672 B.CreateIntrinsic(ID: Intrinsic::spv_bitcast, OverloadTypes: {ElemTy, VecTy}, Args: {Vector});
673 buildAssignType(B, Ty: ElemTy, Arg: Scalar);
674 return Scalar;
675 }
676 return makeExtractElement(B, ElementType: ElemTy, Vector, Index);
677 }
678
679 // Stores the given Src vector operand into the Dst vector, adjusting the size
680 // if required.
681 Value *storeVectorFromVector(IRBuilder<> &B, Value *Src, Value *Dst,
682 Align Alignment) {
683 FixedVectorType *SrcType = cast<FixedVectorType>(Val: Src->getType());
684 FixedVectorType *DstType =
685 cast<FixedVectorType>(Val: GR->findDeducedElementType(Val: Dst));
686 auto dstNumElements = DstType->getNumElements();
687 auto srcNumElements = SrcType->getNumElements();
688
689 // if the element type differs, it is a bitcast.
690 if (DstType->getElementType() != SrcType->getElementType()) {
691 // Support bitcast between vectors of different sizes only if
692 // the total bitwidth is the same.
693 [[maybe_unused]] auto dstBitWidth =
694 DstType->getElementType()->getScalarSizeInBits() * dstNumElements;
695 [[maybe_unused]] auto srcBitWidth =
696 SrcType->getElementType()->getScalarSizeInBits() * srcNumElements;
697 assert(dstBitWidth == srcBitWidth &&
698 "Unsupported bitcast between vectors of different sizes.");
699
700 Src =
701 B.CreateIntrinsic(ID: Intrinsic::spv_bitcast, OverloadTypes: {DstType, SrcType}, Args: {Src});
702 buildAssignType(B, Ty: DstType, Arg: Src);
703 SrcType = DstType;
704
705 StoreInst *SI = B.CreateStore(Val: Src, Ptr: Dst);
706 SI->setAlignment(Alignment);
707 return SI;
708 }
709
710 assert(DstType->getNumElements() >= SrcType->getNumElements());
711 LoadInst *LI = B.CreateLoad(Ty: DstType, Ptr: Dst);
712 LI->setAlignment(Alignment);
713 Value *OldValues = LI;
714 buildAssignType(B, Ty: OldValues->getType(), Arg: OldValues);
715
716 for (unsigned I = 0; I < SrcType->getNumElements(); ++I) {
717 Value *Element = extractScalarFromVector(B, Vector: Src, Index: I);
718 OldValues = makeInsertElement(B, Vector: OldValues, Element, Index: I);
719 }
720
721 StoreInst *SI = B.CreateStore(Val: OldValues, Ptr: Dst);
722 SI->setAlignment(Alignment);
723 return SI;
724 }
725
726 // Builds a legalized store to a pointer, drilling down through
727 // memory layouts to find a compatible type.
728 void buildLegalizedStore(IRBuilder<> &B, Value *Src, Value *Dst,
729 Align Alignment, Value *CastedPtr,
730 Instruction *IllegalStore) {
731 auto ResultOpt = getPointerToFirstCompatibleType(B, BasePtr: Dst, PointerType: Dst->getType(),
732 TargetElemType: Src->getType(), IsInBounds: true);
733 if (!ResultOpt) {
734 if (tryReinterpretStore(B, AccessTy: Src->getType(), OriginalPtr: Dst, CastedPtr, StoreSrc: Src,
735 Alignment))
736 return;
737 llvm_unreachable("Failed to store to aggregate: "
738 "Could not find compatible memory layout.");
739 }
740 auto [GEP, CurrentTy] = *ResultOpt;
741
742 auto *DAT = dyn_cast<ArrayType>(Val: CurrentTy);
743 auto *DVT = dyn_cast<FixedVectorType>(Val: CurrentTy);
744 auto *SVT = dyn_cast<FixedVectorType>(Val: Src->getType());
745 auto *DMAT =
746 DAT ? dyn_cast<FixedVectorType>(Val: DAT->getElementType()) : nullptr;
747
748 if (Src->getType() == CurrentTy) {
749 StoreInst *SI = B.CreateStore(Val: Src, Ptr: GEP);
750 SI->setAlignment(Alignment);
751 return;
752 }
753 if (DVT && SVT) {
754 storeVectorFromVector(B, Src, Dst: GEP, Alignment);
755 return;
756 }
757 if (DAT && SVT && SVT->getElementType() == DAT->getElementType()) {
758 storeArrayFromVector(B, SrcVector: Src, DstArrayPtr: GEP, ArrTy: DAT, Alignment);
759 return;
760 }
761 if (DMAT && SVT && DMAT->getElementType() == SVT->getElementType()) {
762 storeMatrixArrayFromVector(B, SrcVector: Src, DstArrayPtr: GEP, ArrTy: DAT, Alignment);
763 return;
764 }
765
766 llvm_unreachable("Failed to store to aggregate.");
767 }
768
769 // Transforms a store instruction (or SPV intrinsic) using a ptrcast as
770 // operand into a valid logical SPIR-V store with no ptrcast.
771 void transformStore(IRBuilder<> &B, Instruction *IllegalStore, Value *Src,
772 Value *Dst, Value *CastedOperand, Align Alignment) {
773 B.SetInsertPoint(IllegalStore);
774 buildLegalizedStore(B, Src, Dst, Alignment, CastedPtr: CastedOperand, IllegalStore);
775 DeadInstructions.push_back(x: IllegalStore);
776 }
777
778 void legalizePointerCast(IntrinsicInst *II) {
779 Value *CastedOperand = II;
780 Value *OriginalOperand = II->getOperand(i_nocapture: 0);
781
782 IRBuilder<> B(*II->getModule());
783 std::vector<Value *> Users;
784 for (Use &U : II->uses())
785 Users.push_back(x: U.getUser());
786
787 for (Value *User : Users) {
788 if (LoadInst *LI = dyn_cast<LoadInst>(Val: User)) {
789 transformLoad(B, LI, CastedOperand, OriginalOperand);
790 continue;
791 }
792
793 if (StoreInst *SI = dyn_cast<StoreInst>(Val: User)) {
794 transformStore(B, IllegalStore: SI, Src: SI->getValueOperand(), Dst: OriginalOperand,
795 CastedOperand, Alignment: SI->getAlign());
796 continue;
797 }
798
799 if (IntrinsicInst *Intrin = dyn_cast<IntrinsicInst>(Val: User)) {
800 if (Intrin->getIntrinsicID() == Intrinsic::spv_assign_ptr_type) {
801 DeadInstructions.push_back(x: Intrin);
802 continue;
803 }
804
805 if (Intrin->getIntrinsicID() == Intrinsic::spv_gep) {
806 GR->replaceAllUsesWith(Old: CastedOperand, New: OriginalOperand,
807 /* DeleteOld= */ false);
808 continue;
809 }
810
811 if (Intrin->getIntrinsicID() == Intrinsic::spv_store) {
812 Align Alignment;
813 if (ConstantInt *C = dyn_cast<ConstantInt>(Val: Intrin->getOperand(i_nocapture: 3)))
814 Alignment = Align(C->getZExtValue());
815 transformStore(B, IllegalStore: Intrin, Src: Intrin->getArgOperand(i: 0), Dst: OriginalOperand,
816 CastedOperand, Alignment);
817 continue;
818 }
819 }
820
821 llvm_unreachable("Unsupported ptrcast user. Please fix.");
822 }
823
824 DeadInstructions.push_back(x: II);
825 }
826
827public:
828 SPIRVLegalizePointerCastImpl(const SPIRVTargetMachine &TM) : TM(TM) {}
829
830 bool run(Function &F) {
831 const SPIRVSubtarget &ST = TM.getSubtarget<SPIRVSubtarget>(F);
832 GR = ST.getSPIRVGlobalRegistry();
833 DeadInstructions.clear();
834
835 std::vector<IntrinsicInst *> WorkList;
836 for (auto &BB : F) {
837 for (auto &I : BB) {
838 auto *II = dyn_cast<IntrinsicInst>(Val: &I);
839 if (II && II->getIntrinsicID() == Intrinsic::spv_ptrcast)
840 WorkList.push_back(x: II);
841 }
842 }
843
844 for (IntrinsicInst *II : WorkList)
845 legalizePointerCast(II);
846
847 for (Instruction *I : DeadInstructions)
848 I->eraseFromParent();
849
850 return DeadInstructions.size() != 0;
851 }
852
853private:
854 const SPIRVTargetMachine &TM;
855 SPIRVGlobalRegistry *GR = nullptr;
856 std::vector<Instruction *> DeadInstructions;
857};
858
859class SPIRVLegalizePointerCastLegacy : public FunctionPass {
860public:
861 static char ID;
862 SPIRVLegalizePointerCastLegacy(const SPIRVTargetMachine &TM)
863 : FunctionPass(ID), TM(TM) {}
864
865 bool runOnFunction(Function &F) override {
866 return SPIRVLegalizePointerCastImpl(TM).run(F);
867 }
868
869private:
870 const SPIRVTargetMachine &TM;
871};
872} // namespace
873
874PreservedAnalyses
875SPIRVLegalizePointerCastPass::run(Function &F, FunctionAnalysisManager &AM) {
876 return SPIRVLegalizePointerCastImpl(TM).run(F) ? PreservedAnalyses::none()
877 : PreservedAnalyses::all();
878}
879
880char SPIRVLegalizePointerCastLegacy::ID = 0;
881INITIALIZE_PASS(SPIRVLegalizePointerCastLegacy, "spirv-legalize-pointer-cast",
882 "SPIRV legalize pointer cast pass", false, false)
883
884FunctionPass *llvm::createSPIRVLegalizePointerCastPass(SPIRVTargetMachine *TM) {
885 return new SPIRVLegalizePointerCastLegacy(*TM);
886}
887