1 //===- Scalarizer.cpp - Scalarize vector operations -----------------------===//
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 pass converts vector operations into scalar operations, in order
10 // to expose optimization opportunities on the individual scalar operations.
11 // It is mainly intended for targets that do not have vector units, but it
12 // may also be useful for revectorizing code to different vector widths.
13 //
14 //===----------------------------------------------------------------------===//
15
16 #include "llvm/Transforms/Scalar/Scalarizer.h"
17 #include "llvm/ADT/PostOrderIterator.h"
18 #include "llvm/ADT/SmallVector.h"
19 #include "llvm/ADT/Twine.h"
20 #include "llvm/Analysis/VectorUtils.h"
21 #include "llvm/IR/Argument.h"
22 #include "llvm/IR/BasicBlock.h"
23 #include "llvm/IR/Constants.h"
24 #include "llvm/IR/DataLayout.h"
25 #include "llvm/IR/DerivedTypes.h"
26 #include "llvm/IR/Dominators.h"
27 #include "llvm/IR/Function.h"
28 #include "llvm/IR/IRBuilder.h"
29 #include "llvm/IR/InstVisitor.h"
30 #include "llvm/IR/InstrTypes.h"
31 #include "llvm/IR/Instruction.h"
32 #include "llvm/IR/Instructions.h"
33 #include "llvm/IR/Intrinsics.h"
34 #include "llvm/IR/LLVMContext.h"
35 #include "llvm/IR/Module.h"
36 #include "llvm/IR/Type.h"
37 #include "llvm/IR/Value.h"
38 #include "llvm/InitializePasses.h"
39 #include "llvm/Pass.h"
40 #include "llvm/Support/Casting.h"
41 #include "llvm/Support/CommandLine.h"
42 #include "llvm/Support/MathExtras.h"
43 #include "llvm/Transforms/Scalar.h"
44 #include "llvm/Transforms/Utils/Local.h"
45 #include <cassert>
46 #include <cstdint>
47 #include <iterator>
48 #include <map>
49 #include <utility>
50
51 using namespace llvm;
52
53 #define DEBUG_TYPE "scalarizer"
54
55 static cl::opt<bool> ScalarizeVariableInsertExtract(
56 "scalarize-variable-insert-extract", cl::init(true), cl::Hidden,
57 cl::desc("Allow the scalarizer pass to scalarize "
58 "insertelement/extractelement with variable index"));
59
60 // This is disabled by default because having separate loads and stores
61 // makes it more likely that the -combiner-alias-analysis limits will be
62 // reached.
63 static cl::opt<bool>
64 ScalarizeLoadStore("scalarize-load-store", cl::init(false), cl::Hidden,
65 cl::desc("Allow the scalarizer pass to scalarize loads and store"));
66
67 namespace {
68
69 // Used to store the scattered form of a vector.
70 using ValueVector = SmallVector<Value *, 8>;
71
72 // Used to map a vector Value to its scattered form. We use std::map
73 // because we want iterators to persist across insertion and because the
74 // values are relatively large.
75 using ScatterMap = std::map<Value *, ValueVector>;
76
77 // Lists Instructions that have been replaced with scalar implementations,
78 // along with a pointer to their scattered forms.
79 using GatherList = SmallVector<std::pair<Instruction *, ValueVector *>, 16>;
80
81 // Provides a very limited vector-like interface for lazily accessing one
82 // component of a scattered vector or vector pointer.
83 class Scatterer {
84 public:
85 Scatterer() = default;
86
87 // Scatter V into Size components. If new instructions are needed,
88 // insert them before BBI in BB. If Cache is nonnull, use it to cache
89 // the results.
90 Scatterer(BasicBlock *bb, BasicBlock::iterator bbi, Value *v,
91 ValueVector *cachePtr = nullptr);
92
93 // Return component I, creating a new Value for it if necessary.
94 Value *operator[](unsigned I);
95
96 // Return the number of components.
size() const97 unsigned size() const { return Size; }
98
99 private:
100 BasicBlock *BB;
101 BasicBlock::iterator BBI;
102 Value *V;
103 ValueVector *CachePtr;
104 PointerType *PtrTy;
105 ValueVector Tmp;
106 unsigned Size;
107 };
108
109 // FCmpSpliiter(FCI)(Builder, X, Y, Name) uses Builder to create an FCmp
110 // called Name that compares X and Y in the same way as FCI.
111 struct FCmpSplitter {
FCmpSplitter__anon0932baa40111::FCmpSplitter112 FCmpSplitter(FCmpInst &fci) : FCI(fci) {}
113
operator ()__anon0932baa40111::FCmpSplitter114 Value *operator()(IRBuilder<> &Builder, Value *Op0, Value *Op1,
115 const Twine &Name) const {
116 return Builder.CreateFCmp(FCI.getPredicate(), Op0, Op1, Name);
117 }
118
119 FCmpInst &FCI;
120 };
121
122 // ICmpSpliiter(ICI)(Builder, X, Y, Name) uses Builder to create an ICmp
123 // called Name that compares X and Y in the same way as ICI.
124 struct ICmpSplitter {
ICmpSplitter__anon0932baa40111::ICmpSplitter125 ICmpSplitter(ICmpInst &ici) : ICI(ici) {}
126
operator ()__anon0932baa40111::ICmpSplitter127 Value *operator()(IRBuilder<> &Builder, Value *Op0, Value *Op1,
128 const Twine &Name) const {
129 return Builder.CreateICmp(ICI.getPredicate(), Op0, Op1, Name);
130 }
131
132 ICmpInst &ICI;
133 };
134
135 // UnarySpliiter(UO)(Builder, X, Name) uses Builder to create
136 // a unary operator like UO called Name with operand X.
137 struct UnarySplitter {
UnarySplitter__anon0932baa40111::UnarySplitter138 UnarySplitter(UnaryOperator &uo) : UO(uo) {}
139
operator ()__anon0932baa40111::UnarySplitter140 Value *operator()(IRBuilder<> &Builder, Value *Op, const Twine &Name) const {
141 return Builder.CreateUnOp(UO.getOpcode(), Op, Name);
142 }
143
144 UnaryOperator &UO;
145 };
146
147 // BinarySpliiter(BO)(Builder, X, Y, Name) uses Builder to create
148 // a binary operator like BO called Name with operands X and Y.
149 struct BinarySplitter {
BinarySplitter__anon0932baa40111::BinarySplitter150 BinarySplitter(BinaryOperator &bo) : BO(bo) {}
151
operator ()__anon0932baa40111::BinarySplitter152 Value *operator()(IRBuilder<> &Builder, Value *Op0, Value *Op1,
153 const Twine &Name) const {
154 return Builder.CreateBinOp(BO.getOpcode(), Op0, Op1, Name);
155 }
156
157 BinaryOperator &BO;
158 };
159
160 // Information about a load or store that we're scalarizing.
161 struct VectorLayout {
162 VectorLayout() = default;
163
164 // Return the alignment of element I.
getElemAlign__anon0932baa40111::VectorLayout165 Align getElemAlign(unsigned I) {
166 return commonAlignment(VecAlign, I * ElemSize);
167 }
168
169 // The type of the vector.
170 VectorType *VecTy = nullptr;
171
172 // The type of each element.
173 Type *ElemTy = nullptr;
174
175 // The alignment of the vector.
176 Align VecAlign;
177
178 // The size of each element.
179 uint64_t ElemSize = 0;
180 };
181
182 class ScalarizerVisitor : public InstVisitor<ScalarizerVisitor, bool> {
183 public:
ScalarizerVisitor(unsigned ParallelLoopAccessMDKind,DominatorTree * DT)184 ScalarizerVisitor(unsigned ParallelLoopAccessMDKind, DominatorTree *DT)
185 : ParallelLoopAccessMDKind(ParallelLoopAccessMDKind), DT(DT) {
186 }
187
188 bool visit(Function &F);
189
190 // InstVisitor methods. They return true if the instruction was scalarized,
191 // false if nothing changed.
visitInstruction(Instruction & I)192 bool visitInstruction(Instruction &I) { return false; }
193 bool visitSelectInst(SelectInst &SI);
194 bool visitICmpInst(ICmpInst &ICI);
195 bool visitFCmpInst(FCmpInst &FCI);
196 bool visitUnaryOperator(UnaryOperator &UO);
197 bool visitBinaryOperator(BinaryOperator &BO);
198 bool visitGetElementPtrInst(GetElementPtrInst &GEPI);
199 bool visitCastInst(CastInst &CI);
200 bool visitBitCastInst(BitCastInst &BCI);
201 bool visitInsertElementInst(InsertElementInst &IEI);
202 bool visitExtractElementInst(ExtractElementInst &EEI);
203 bool visitShuffleVectorInst(ShuffleVectorInst &SVI);
204 bool visitPHINode(PHINode &PHI);
205 bool visitLoadInst(LoadInst &LI);
206 bool visitStoreInst(StoreInst &SI);
207 bool visitCallInst(CallInst &ICI);
208
209 private:
210 Scatterer scatter(Instruction *Point, Value *V);
211 void gather(Instruction *Op, const ValueVector &CV);
212 bool canTransferMetadata(unsigned Kind);
213 void transferMetadataAndIRFlags(Instruction *Op, const ValueVector &CV);
214 Optional<VectorLayout> getVectorLayout(Type *Ty, Align Alignment,
215 const DataLayout &DL);
216 bool finish();
217
218 template<typename T> bool splitUnary(Instruction &, const T &);
219 template<typename T> bool splitBinary(Instruction &, const T &);
220
221 bool splitCall(CallInst &CI);
222
223 ScatterMap Scattered;
224 GatherList Gathered;
225
226 SmallVector<WeakTrackingVH, 32> PotentiallyDeadInstrs;
227
228 unsigned ParallelLoopAccessMDKind;
229
230 DominatorTree *DT;
231 };
232
233 class ScalarizerLegacyPass : public FunctionPass {
234 public:
235 static char ID;
236
ScalarizerLegacyPass()237 ScalarizerLegacyPass() : FunctionPass(ID) {
238 initializeScalarizerLegacyPassPass(*PassRegistry::getPassRegistry());
239 }
240
241 bool runOnFunction(Function &F) override;
242
getAnalysisUsage(AnalysisUsage & AU) const243 void getAnalysisUsage(AnalysisUsage& AU) const override {
244 AU.addRequired<DominatorTreeWrapperPass>();
245 AU.addPreserved<DominatorTreeWrapperPass>();
246 }
247 };
248
249 } // end anonymous namespace
250
251 char ScalarizerLegacyPass::ID = 0;
252 INITIALIZE_PASS_BEGIN(ScalarizerLegacyPass, "scalarizer",
253 "Scalarize vector operations", false, false)
INITIALIZE_PASS_DEPENDENCY(DominatorTreeWrapperPass)254 INITIALIZE_PASS_DEPENDENCY(DominatorTreeWrapperPass)
255 INITIALIZE_PASS_END(ScalarizerLegacyPass, "scalarizer",
256 "Scalarize vector operations", false, false)
257
258 Scatterer::Scatterer(BasicBlock *bb, BasicBlock::iterator bbi, Value *v,
259 ValueVector *cachePtr)
260 : BB(bb), BBI(bbi), V(v), CachePtr(cachePtr) {
261 Type *Ty = V->getType();
262 PtrTy = dyn_cast<PointerType>(Ty);
263 if (PtrTy)
264 Ty = PtrTy->getElementType();
265 Size = cast<FixedVectorType>(Ty)->getNumElements();
266 if (!CachePtr)
267 Tmp.resize(Size, nullptr);
268 else if (CachePtr->empty())
269 CachePtr->resize(Size, nullptr);
270 else
271 assert(Size == CachePtr->size() && "Inconsistent vector sizes");
272 }
273
274 // Return component I, creating a new Value for it if necessary.
operator [](unsigned I)275 Value *Scatterer::operator[](unsigned I) {
276 ValueVector &CV = (CachePtr ? *CachePtr : Tmp);
277 // Try to reuse a previous value.
278 if (CV[I])
279 return CV[I];
280 IRBuilder<> Builder(BB, BBI);
281 if (PtrTy) {
282 Type *ElTy = cast<VectorType>(PtrTy->getElementType())->getElementType();
283 if (!CV[0]) {
284 Type *NewPtrTy = PointerType::get(ElTy, PtrTy->getAddressSpace());
285 CV[0] = Builder.CreateBitCast(V, NewPtrTy, V->getName() + ".i0");
286 }
287 if (I != 0)
288 CV[I] = Builder.CreateConstGEP1_32(ElTy, CV[0], I,
289 V->getName() + ".i" + Twine(I));
290 } else {
291 // Search through a chain of InsertElementInsts looking for element I.
292 // Record other elements in the cache. The new V is still suitable
293 // for all uncached indices.
294 while (true) {
295 InsertElementInst *Insert = dyn_cast<InsertElementInst>(V);
296 if (!Insert)
297 break;
298 ConstantInt *Idx = dyn_cast<ConstantInt>(Insert->getOperand(2));
299 if (!Idx)
300 break;
301 unsigned J = Idx->getZExtValue();
302 V = Insert->getOperand(0);
303 if (I == J) {
304 CV[J] = Insert->getOperand(1);
305 return CV[J];
306 } else if (!CV[J]) {
307 // Only cache the first entry we find for each index we're not actively
308 // searching for. This prevents us from going too far up the chain and
309 // caching incorrect entries.
310 CV[J] = Insert->getOperand(1);
311 }
312 }
313 CV[I] = Builder.CreateExtractElement(V, Builder.getInt32(I),
314 V->getName() + ".i" + Twine(I));
315 }
316 return CV[I];
317 }
318
runOnFunction(Function & F)319 bool ScalarizerLegacyPass::runOnFunction(Function &F) {
320 if (skipFunction(F))
321 return false;
322
323 Module &M = *F.getParent();
324 unsigned ParallelLoopAccessMDKind =
325 M.getContext().getMDKindID("llvm.mem.parallel_loop_access");
326 DominatorTree *DT = &getAnalysis<DominatorTreeWrapperPass>().getDomTree();
327 ScalarizerVisitor Impl(ParallelLoopAccessMDKind, DT);
328 return Impl.visit(F);
329 }
330
createScalarizerPass()331 FunctionPass *llvm::createScalarizerPass() {
332 return new ScalarizerLegacyPass();
333 }
334
visit(Function & F)335 bool ScalarizerVisitor::visit(Function &F) {
336 assert(Gathered.empty() && Scattered.empty());
337
338 // To ensure we replace gathered components correctly we need to do an ordered
339 // traversal of the basic blocks in the function.
340 ReversePostOrderTraversal<BasicBlock *> RPOT(&F.getEntryBlock());
341 for (BasicBlock *BB : RPOT) {
342 for (BasicBlock::iterator II = BB->begin(), IE = BB->end(); II != IE;) {
343 Instruction *I = &*II;
344 bool Done = InstVisitor::visit(I);
345 ++II;
346 if (Done && I->getType()->isVoidTy())
347 I->eraseFromParent();
348 }
349 }
350 return finish();
351 }
352
353 // Return a scattered form of V that can be accessed by Point. V must be a
354 // vector or a pointer to a vector.
scatter(Instruction * Point,Value * V)355 Scatterer ScalarizerVisitor::scatter(Instruction *Point, Value *V) {
356 if (Argument *VArg = dyn_cast<Argument>(V)) {
357 // Put the scattered form of arguments in the entry block,
358 // so that it can be used everywhere.
359 Function *F = VArg->getParent();
360 BasicBlock *BB = &F->getEntryBlock();
361 return Scatterer(BB, BB->begin(), V, &Scattered[V]);
362 }
363 if (Instruction *VOp = dyn_cast<Instruction>(V)) {
364 // When scalarizing PHI nodes we might try to examine/rewrite InsertElement
365 // nodes in predecessors. If those predecessors are unreachable from entry,
366 // then the IR in those blocks could have unexpected properties resulting in
367 // infinite loops in Scatterer::operator[]. By simply treating values
368 // originating from instructions in unreachable blocks as undef we do not
369 // need to analyse them further.
370 if (!DT->isReachableFromEntry(VOp->getParent()))
371 return Scatterer(Point->getParent(), Point->getIterator(),
372 UndefValue::get(V->getType()));
373 // Put the scattered form of an instruction directly after the
374 // instruction.
375 BasicBlock *BB = VOp->getParent();
376 return Scatterer(BB, std::next(BasicBlock::iterator(VOp)),
377 V, &Scattered[V]);
378 }
379 // In the fallback case, just put the scattered before Point and
380 // keep the result local to Point.
381 return Scatterer(Point->getParent(), Point->getIterator(), V);
382 }
383
384 // Replace Op with the gathered form of the components in CV. Defer the
385 // deletion of Op and creation of the gathered form to the end of the pass,
386 // so that we can avoid creating the gathered form if all uses of Op are
387 // replaced with uses of CV.
gather(Instruction * Op,const ValueVector & CV)388 void ScalarizerVisitor::gather(Instruction *Op, const ValueVector &CV) {
389 transferMetadataAndIRFlags(Op, CV);
390
391 // If we already have a scattered form of Op (created from ExtractElements
392 // of Op itself), replace them with the new form.
393 ValueVector &SV = Scattered[Op];
394 if (!SV.empty()) {
395 for (unsigned I = 0, E = SV.size(); I != E; ++I) {
396 Value *V = SV[I];
397 if (V == nullptr || SV[I] == CV[I])
398 continue;
399
400 Instruction *Old = cast<Instruction>(V);
401 if (isa<Instruction>(CV[I]))
402 CV[I]->takeName(Old);
403 Old->replaceAllUsesWith(CV[I]);
404 PotentiallyDeadInstrs.emplace_back(Old);
405 }
406 }
407 SV = CV;
408 Gathered.push_back(GatherList::value_type(Op, &SV));
409 }
410
411 // Return true if it is safe to transfer the given metadata tag from
412 // vector to scalar instructions.
canTransferMetadata(unsigned Tag)413 bool ScalarizerVisitor::canTransferMetadata(unsigned Tag) {
414 return (Tag == LLVMContext::MD_tbaa
415 || Tag == LLVMContext::MD_fpmath
416 || Tag == LLVMContext::MD_tbaa_struct
417 || Tag == LLVMContext::MD_invariant_load
418 || Tag == LLVMContext::MD_alias_scope
419 || Tag == LLVMContext::MD_noalias
420 || Tag == ParallelLoopAccessMDKind
421 || Tag == LLVMContext::MD_access_group);
422 }
423
424 // Transfer metadata from Op to the instructions in CV if it is known
425 // to be safe to do so.
transferMetadataAndIRFlags(Instruction * Op,const ValueVector & CV)426 void ScalarizerVisitor::transferMetadataAndIRFlags(Instruction *Op,
427 const ValueVector &CV) {
428 SmallVector<std::pair<unsigned, MDNode *>, 4> MDs;
429 Op->getAllMetadataOtherThanDebugLoc(MDs);
430 for (unsigned I = 0, E = CV.size(); I != E; ++I) {
431 if (Instruction *New = dyn_cast<Instruction>(CV[I])) {
432 for (const auto &MD : MDs)
433 if (canTransferMetadata(MD.first))
434 New->setMetadata(MD.first, MD.second);
435 New->copyIRFlags(Op);
436 if (Op->getDebugLoc() && !New->getDebugLoc())
437 New->setDebugLoc(Op->getDebugLoc());
438 }
439 }
440 }
441
442 // Try to fill in Layout from Ty, returning true on success. Alignment is
443 // the alignment of the vector, or None if the ABI default should be used.
444 Optional<VectorLayout>
getVectorLayout(Type * Ty,Align Alignment,const DataLayout & DL)445 ScalarizerVisitor::getVectorLayout(Type *Ty, Align Alignment,
446 const DataLayout &DL) {
447 VectorLayout Layout;
448 // Make sure we're dealing with a vector.
449 Layout.VecTy = dyn_cast<VectorType>(Ty);
450 if (!Layout.VecTy)
451 return None;
452 // Check that we're dealing with full-byte elements.
453 Layout.ElemTy = Layout.VecTy->getElementType();
454 if (!DL.typeSizeEqualsStoreSize(Layout.ElemTy))
455 return None;
456 Layout.VecAlign = Alignment;
457 Layout.ElemSize = DL.getTypeStoreSize(Layout.ElemTy);
458 return Layout;
459 }
460
461 // Scalarize one-operand instruction I, using Split(Builder, X, Name)
462 // to create an instruction like I with operand X and name Name.
463 template<typename Splitter>
splitUnary(Instruction & I,const Splitter & Split)464 bool ScalarizerVisitor::splitUnary(Instruction &I, const Splitter &Split) {
465 VectorType *VT = dyn_cast<VectorType>(I.getType());
466 if (!VT)
467 return false;
468
469 unsigned NumElems = cast<FixedVectorType>(VT)->getNumElements();
470 IRBuilder<> Builder(&I);
471 Scatterer Op = scatter(&I, I.getOperand(0));
472 assert(Op.size() == NumElems && "Mismatched unary operation");
473 ValueVector Res;
474 Res.resize(NumElems);
475 for (unsigned Elem = 0; Elem < NumElems; ++Elem)
476 Res[Elem] = Split(Builder, Op[Elem], I.getName() + ".i" + Twine(Elem));
477 gather(&I, Res);
478 return true;
479 }
480
481 // Scalarize two-operand instruction I, using Split(Builder, X, Y, Name)
482 // to create an instruction like I with operands X and Y and name Name.
483 template<typename Splitter>
splitBinary(Instruction & I,const Splitter & Split)484 bool ScalarizerVisitor::splitBinary(Instruction &I, const Splitter &Split) {
485 VectorType *VT = dyn_cast<VectorType>(I.getType());
486 if (!VT)
487 return false;
488
489 unsigned NumElems = cast<FixedVectorType>(VT)->getNumElements();
490 IRBuilder<> Builder(&I);
491 Scatterer VOp0 = scatter(&I, I.getOperand(0));
492 Scatterer VOp1 = scatter(&I, I.getOperand(1));
493 assert(VOp0.size() == NumElems && "Mismatched binary operation");
494 assert(VOp1.size() == NumElems && "Mismatched binary operation");
495 ValueVector Res;
496 Res.resize(NumElems);
497 for (unsigned Elem = 0; Elem < NumElems; ++Elem) {
498 Value *Op0 = VOp0[Elem];
499 Value *Op1 = VOp1[Elem];
500 Res[Elem] = Split(Builder, Op0, Op1, I.getName() + ".i" + Twine(Elem));
501 }
502 gather(&I, Res);
503 return true;
504 }
505
isTriviallyScalariable(Intrinsic::ID ID)506 static bool isTriviallyScalariable(Intrinsic::ID ID) {
507 return isTriviallyVectorizable(ID);
508 }
509
510 // All of the current scalarizable intrinsics only have one mangled type.
getScalarIntrinsicDeclaration(Module * M,Intrinsic::ID ID,VectorType * Ty)511 static Function *getScalarIntrinsicDeclaration(Module *M,
512 Intrinsic::ID ID,
513 VectorType *Ty) {
514 return Intrinsic::getDeclaration(M, ID, { Ty->getScalarType() });
515 }
516
517 /// If a call to a vector typed intrinsic function, split into a scalar call per
518 /// element if possible for the intrinsic.
splitCall(CallInst & CI)519 bool ScalarizerVisitor::splitCall(CallInst &CI) {
520 VectorType *VT = dyn_cast<VectorType>(CI.getType());
521 if (!VT)
522 return false;
523
524 Function *F = CI.getCalledFunction();
525 if (!F)
526 return false;
527
528 Intrinsic::ID ID = F->getIntrinsicID();
529 if (ID == Intrinsic::not_intrinsic || !isTriviallyScalariable(ID))
530 return false;
531
532 unsigned NumElems = cast<FixedVectorType>(VT)->getNumElements();
533 unsigned NumArgs = CI.getNumArgOperands();
534
535 ValueVector ScalarOperands(NumArgs);
536 SmallVector<Scatterer, 8> Scattered(NumArgs);
537
538 Scattered.resize(NumArgs);
539
540 // Assumes that any vector type has the same number of elements as the return
541 // vector type, which is true for all current intrinsics.
542 for (unsigned I = 0; I != NumArgs; ++I) {
543 Value *OpI = CI.getOperand(I);
544 if (OpI->getType()->isVectorTy()) {
545 Scattered[I] = scatter(&CI, OpI);
546 assert(Scattered[I].size() == NumElems && "mismatched call operands");
547 } else {
548 ScalarOperands[I] = OpI;
549 }
550 }
551
552 ValueVector Res(NumElems);
553 ValueVector ScalarCallOps(NumArgs);
554
555 Function *NewIntrin = getScalarIntrinsicDeclaration(F->getParent(), ID, VT);
556 IRBuilder<> Builder(&CI);
557
558 // Perform actual scalarization, taking care to preserve any scalar operands.
559 for (unsigned Elem = 0; Elem < NumElems; ++Elem) {
560 ScalarCallOps.clear();
561
562 for (unsigned J = 0; J != NumArgs; ++J) {
563 if (hasVectorInstrinsicScalarOpd(ID, J))
564 ScalarCallOps.push_back(ScalarOperands[J]);
565 else
566 ScalarCallOps.push_back(Scattered[J][Elem]);
567 }
568
569 Res[Elem] = Builder.CreateCall(NewIntrin, ScalarCallOps,
570 CI.getName() + ".i" + Twine(Elem));
571 }
572
573 gather(&CI, Res);
574 return true;
575 }
576
visitSelectInst(SelectInst & SI)577 bool ScalarizerVisitor::visitSelectInst(SelectInst &SI) {
578 VectorType *VT = dyn_cast<VectorType>(SI.getType());
579 if (!VT)
580 return false;
581
582 unsigned NumElems = cast<FixedVectorType>(VT)->getNumElements();
583 IRBuilder<> Builder(&SI);
584 Scatterer VOp1 = scatter(&SI, SI.getOperand(1));
585 Scatterer VOp2 = scatter(&SI, SI.getOperand(2));
586 assert(VOp1.size() == NumElems && "Mismatched select");
587 assert(VOp2.size() == NumElems && "Mismatched select");
588 ValueVector Res;
589 Res.resize(NumElems);
590
591 if (SI.getOperand(0)->getType()->isVectorTy()) {
592 Scatterer VOp0 = scatter(&SI, SI.getOperand(0));
593 assert(VOp0.size() == NumElems && "Mismatched select");
594 for (unsigned I = 0; I < NumElems; ++I) {
595 Value *Op0 = VOp0[I];
596 Value *Op1 = VOp1[I];
597 Value *Op2 = VOp2[I];
598 Res[I] = Builder.CreateSelect(Op0, Op1, Op2,
599 SI.getName() + ".i" + Twine(I));
600 }
601 } else {
602 Value *Op0 = SI.getOperand(0);
603 for (unsigned I = 0; I < NumElems; ++I) {
604 Value *Op1 = VOp1[I];
605 Value *Op2 = VOp2[I];
606 Res[I] = Builder.CreateSelect(Op0, Op1, Op2,
607 SI.getName() + ".i" + Twine(I));
608 }
609 }
610 gather(&SI, Res);
611 return true;
612 }
613
visitICmpInst(ICmpInst & ICI)614 bool ScalarizerVisitor::visitICmpInst(ICmpInst &ICI) {
615 return splitBinary(ICI, ICmpSplitter(ICI));
616 }
617
visitFCmpInst(FCmpInst & FCI)618 bool ScalarizerVisitor::visitFCmpInst(FCmpInst &FCI) {
619 return splitBinary(FCI, FCmpSplitter(FCI));
620 }
621
visitUnaryOperator(UnaryOperator & UO)622 bool ScalarizerVisitor::visitUnaryOperator(UnaryOperator &UO) {
623 return splitUnary(UO, UnarySplitter(UO));
624 }
625
visitBinaryOperator(BinaryOperator & BO)626 bool ScalarizerVisitor::visitBinaryOperator(BinaryOperator &BO) {
627 return splitBinary(BO, BinarySplitter(BO));
628 }
629
visitGetElementPtrInst(GetElementPtrInst & GEPI)630 bool ScalarizerVisitor::visitGetElementPtrInst(GetElementPtrInst &GEPI) {
631 VectorType *VT = dyn_cast<VectorType>(GEPI.getType());
632 if (!VT)
633 return false;
634
635 IRBuilder<> Builder(&GEPI);
636 unsigned NumElems = cast<FixedVectorType>(VT)->getNumElements();
637 unsigned NumIndices = GEPI.getNumIndices();
638
639 // The base pointer might be scalar even if it's a vector GEP. In those cases,
640 // splat the pointer into a vector value, and scatter that vector.
641 Value *Op0 = GEPI.getOperand(0);
642 if (!Op0->getType()->isVectorTy())
643 Op0 = Builder.CreateVectorSplat(NumElems, Op0);
644 Scatterer Base = scatter(&GEPI, Op0);
645
646 SmallVector<Scatterer, 8> Ops;
647 Ops.resize(NumIndices);
648 for (unsigned I = 0; I < NumIndices; ++I) {
649 Value *Op = GEPI.getOperand(I + 1);
650
651 // The indices might be scalars even if it's a vector GEP. In those cases,
652 // splat the scalar into a vector value, and scatter that vector.
653 if (!Op->getType()->isVectorTy())
654 Op = Builder.CreateVectorSplat(NumElems, Op);
655
656 Ops[I] = scatter(&GEPI, Op);
657 }
658
659 ValueVector Res;
660 Res.resize(NumElems);
661 for (unsigned I = 0; I < NumElems; ++I) {
662 SmallVector<Value *, 8> Indices;
663 Indices.resize(NumIndices);
664 for (unsigned J = 0; J < NumIndices; ++J)
665 Indices[J] = Ops[J][I];
666 Res[I] = Builder.CreateGEP(GEPI.getSourceElementType(), Base[I], Indices,
667 GEPI.getName() + ".i" + Twine(I));
668 if (GEPI.isInBounds())
669 if (GetElementPtrInst *NewGEPI = dyn_cast<GetElementPtrInst>(Res[I]))
670 NewGEPI->setIsInBounds();
671 }
672 gather(&GEPI, Res);
673 return true;
674 }
675
visitCastInst(CastInst & CI)676 bool ScalarizerVisitor::visitCastInst(CastInst &CI) {
677 VectorType *VT = dyn_cast<VectorType>(CI.getDestTy());
678 if (!VT)
679 return false;
680
681 unsigned NumElems = cast<FixedVectorType>(VT)->getNumElements();
682 IRBuilder<> Builder(&CI);
683 Scatterer Op0 = scatter(&CI, CI.getOperand(0));
684 assert(Op0.size() == NumElems && "Mismatched cast");
685 ValueVector Res;
686 Res.resize(NumElems);
687 for (unsigned I = 0; I < NumElems; ++I)
688 Res[I] = Builder.CreateCast(CI.getOpcode(), Op0[I], VT->getElementType(),
689 CI.getName() + ".i" + Twine(I));
690 gather(&CI, Res);
691 return true;
692 }
693
visitBitCastInst(BitCastInst & BCI)694 bool ScalarizerVisitor::visitBitCastInst(BitCastInst &BCI) {
695 VectorType *DstVT = dyn_cast<VectorType>(BCI.getDestTy());
696 VectorType *SrcVT = dyn_cast<VectorType>(BCI.getSrcTy());
697 if (!DstVT || !SrcVT)
698 return false;
699
700 unsigned DstNumElems = cast<FixedVectorType>(DstVT)->getNumElements();
701 unsigned SrcNumElems = cast<FixedVectorType>(SrcVT)->getNumElements();
702 IRBuilder<> Builder(&BCI);
703 Scatterer Op0 = scatter(&BCI, BCI.getOperand(0));
704 ValueVector Res;
705 Res.resize(DstNumElems);
706
707 if (DstNumElems == SrcNumElems) {
708 for (unsigned I = 0; I < DstNumElems; ++I)
709 Res[I] = Builder.CreateBitCast(Op0[I], DstVT->getElementType(),
710 BCI.getName() + ".i" + Twine(I));
711 } else if (DstNumElems > SrcNumElems) {
712 // <M x t1> -> <N*M x t2>. Convert each t1 to <N x t2> and copy the
713 // individual elements to the destination.
714 unsigned FanOut = DstNumElems / SrcNumElems;
715 auto *MidTy = FixedVectorType::get(DstVT->getElementType(), FanOut);
716 unsigned ResI = 0;
717 for (unsigned Op0I = 0; Op0I < SrcNumElems; ++Op0I) {
718 Value *V = Op0[Op0I];
719 Instruction *VI;
720 // Look through any existing bitcasts before converting to <N x t2>.
721 // In the best case, the resulting conversion might be a no-op.
722 while ((VI = dyn_cast<Instruction>(V)) &&
723 VI->getOpcode() == Instruction::BitCast)
724 V = VI->getOperand(0);
725 V = Builder.CreateBitCast(V, MidTy, V->getName() + ".cast");
726 Scatterer Mid = scatter(&BCI, V);
727 for (unsigned MidI = 0; MidI < FanOut; ++MidI)
728 Res[ResI++] = Mid[MidI];
729 }
730 } else {
731 // <N*M x t1> -> <M x t2>. Convert each group of <N x t1> into a t2.
732 unsigned FanIn = SrcNumElems / DstNumElems;
733 auto *MidTy = FixedVectorType::get(SrcVT->getElementType(), FanIn);
734 unsigned Op0I = 0;
735 for (unsigned ResI = 0; ResI < DstNumElems; ++ResI) {
736 Value *V = PoisonValue::get(MidTy);
737 for (unsigned MidI = 0; MidI < FanIn; ++MidI)
738 V = Builder.CreateInsertElement(V, Op0[Op0I++], Builder.getInt32(MidI),
739 BCI.getName() + ".i" + Twine(ResI)
740 + ".upto" + Twine(MidI));
741 Res[ResI] = Builder.CreateBitCast(V, DstVT->getElementType(),
742 BCI.getName() + ".i" + Twine(ResI));
743 }
744 }
745 gather(&BCI, Res);
746 return true;
747 }
748
visitInsertElementInst(InsertElementInst & IEI)749 bool ScalarizerVisitor::visitInsertElementInst(InsertElementInst &IEI) {
750 VectorType *VT = dyn_cast<VectorType>(IEI.getType());
751 if (!VT)
752 return false;
753
754 unsigned NumElems = cast<FixedVectorType>(VT)->getNumElements();
755 IRBuilder<> Builder(&IEI);
756 Scatterer Op0 = scatter(&IEI, IEI.getOperand(0));
757 Value *NewElt = IEI.getOperand(1);
758 Value *InsIdx = IEI.getOperand(2);
759
760 ValueVector Res;
761 Res.resize(NumElems);
762
763 if (auto *CI = dyn_cast<ConstantInt>(InsIdx)) {
764 for (unsigned I = 0; I < NumElems; ++I)
765 Res[I] = CI->getValue().getZExtValue() == I ? NewElt : Op0[I];
766 } else {
767 if (!ScalarizeVariableInsertExtract)
768 return false;
769
770 for (unsigned I = 0; I < NumElems; ++I) {
771 Value *ShouldReplace =
772 Builder.CreateICmpEQ(InsIdx, ConstantInt::get(InsIdx->getType(), I),
773 InsIdx->getName() + ".is." + Twine(I));
774 Value *OldElt = Op0[I];
775 Res[I] = Builder.CreateSelect(ShouldReplace, NewElt, OldElt,
776 IEI.getName() + ".i" + Twine(I));
777 }
778 }
779
780 gather(&IEI, Res);
781 return true;
782 }
783
visitExtractElementInst(ExtractElementInst & EEI)784 bool ScalarizerVisitor::visitExtractElementInst(ExtractElementInst &EEI) {
785 VectorType *VT = dyn_cast<VectorType>(EEI.getOperand(0)->getType());
786 if (!VT)
787 return false;
788
789 unsigned NumSrcElems = cast<FixedVectorType>(VT)->getNumElements();
790 IRBuilder<> Builder(&EEI);
791 Scatterer Op0 = scatter(&EEI, EEI.getOperand(0));
792 Value *ExtIdx = EEI.getOperand(1);
793
794 if (auto *CI = dyn_cast<ConstantInt>(ExtIdx)) {
795 Value *Res = Op0[CI->getValue().getZExtValue()];
796 gather(&EEI, {Res});
797 return true;
798 }
799
800 if (!ScalarizeVariableInsertExtract)
801 return false;
802
803 Value *Res = UndefValue::get(VT->getElementType());
804 for (unsigned I = 0; I < NumSrcElems; ++I) {
805 Value *ShouldExtract =
806 Builder.CreateICmpEQ(ExtIdx, ConstantInt::get(ExtIdx->getType(), I),
807 ExtIdx->getName() + ".is." + Twine(I));
808 Value *Elt = Op0[I];
809 Res = Builder.CreateSelect(ShouldExtract, Elt, Res,
810 EEI.getName() + ".upto" + Twine(I));
811 }
812 gather(&EEI, {Res});
813 return true;
814 }
815
visitShuffleVectorInst(ShuffleVectorInst & SVI)816 bool ScalarizerVisitor::visitShuffleVectorInst(ShuffleVectorInst &SVI) {
817 VectorType *VT = dyn_cast<VectorType>(SVI.getType());
818 if (!VT)
819 return false;
820
821 unsigned NumElems = cast<FixedVectorType>(VT)->getNumElements();
822 Scatterer Op0 = scatter(&SVI, SVI.getOperand(0));
823 Scatterer Op1 = scatter(&SVI, SVI.getOperand(1));
824 ValueVector Res;
825 Res.resize(NumElems);
826
827 for (unsigned I = 0; I < NumElems; ++I) {
828 int Selector = SVI.getMaskValue(I);
829 if (Selector < 0)
830 Res[I] = UndefValue::get(VT->getElementType());
831 else if (unsigned(Selector) < Op0.size())
832 Res[I] = Op0[Selector];
833 else
834 Res[I] = Op1[Selector - Op0.size()];
835 }
836 gather(&SVI, Res);
837 return true;
838 }
839
visitPHINode(PHINode & PHI)840 bool ScalarizerVisitor::visitPHINode(PHINode &PHI) {
841 VectorType *VT = dyn_cast<VectorType>(PHI.getType());
842 if (!VT)
843 return false;
844
845 unsigned NumElems = cast<FixedVectorType>(VT)->getNumElements();
846 IRBuilder<> Builder(&PHI);
847 ValueVector Res;
848 Res.resize(NumElems);
849
850 unsigned NumOps = PHI.getNumOperands();
851 for (unsigned I = 0; I < NumElems; ++I)
852 Res[I] = Builder.CreatePHI(VT->getElementType(), NumOps,
853 PHI.getName() + ".i" + Twine(I));
854
855 for (unsigned I = 0; I < NumOps; ++I) {
856 Scatterer Op = scatter(&PHI, PHI.getIncomingValue(I));
857 BasicBlock *IncomingBlock = PHI.getIncomingBlock(I);
858 for (unsigned J = 0; J < NumElems; ++J)
859 cast<PHINode>(Res[J])->addIncoming(Op[J], IncomingBlock);
860 }
861 gather(&PHI, Res);
862 return true;
863 }
864
visitLoadInst(LoadInst & LI)865 bool ScalarizerVisitor::visitLoadInst(LoadInst &LI) {
866 if (!ScalarizeLoadStore)
867 return false;
868 if (!LI.isSimple())
869 return false;
870
871 Optional<VectorLayout> Layout = getVectorLayout(
872 LI.getType(), LI.getAlign(), LI.getModule()->getDataLayout());
873 if (!Layout)
874 return false;
875
876 unsigned NumElems = cast<FixedVectorType>(Layout->VecTy)->getNumElements();
877 IRBuilder<> Builder(&LI);
878 Scatterer Ptr = scatter(&LI, LI.getPointerOperand());
879 ValueVector Res;
880 Res.resize(NumElems);
881
882 for (unsigned I = 0; I < NumElems; ++I)
883 Res[I] = Builder.CreateAlignedLoad(Layout->VecTy->getElementType(), Ptr[I],
884 Align(Layout->getElemAlign(I)),
885 LI.getName() + ".i" + Twine(I));
886 gather(&LI, Res);
887 return true;
888 }
889
visitStoreInst(StoreInst & SI)890 bool ScalarizerVisitor::visitStoreInst(StoreInst &SI) {
891 if (!ScalarizeLoadStore)
892 return false;
893 if (!SI.isSimple())
894 return false;
895
896 Value *FullValue = SI.getValueOperand();
897 Optional<VectorLayout> Layout = getVectorLayout(
898 FullValue->getType(), SI.getAlign(), SI.getModule()->getDataLayout());
899 if (!Layout)
900 return false;
901
902 unsigned NumElems = cast<FixedVectorType>(Layout->VecTy)->getNumElements();
903 IRBuilder<> Builder(&SI);
904 Scatterer VPtr = scatter(&SI, SI.getPointerOperand());
905 Scatterer VVal = scatter(&SI, FullValue);
906
907 ValueVector Stores;
908 Stores.resize(NumElems);
909 for (unsigned I = 0; I < NumElems; ++I) {
910 Value *Val = VVal[I];
911 Value *Ptr = VPtr[I];
912 Stores[I] = Builder.CreateAlignedStore(Val, Ptr, Layout->getElemAlign(I));
913 }
914 transferMetadataAndIRFlags(&SI, Stores);
915 return true;
916 }
917
visitCallInst(CallInst & CI)918 bool ScalarizerVisitor::visitCallInst(CallInst &CI) {
919 return splitCall(CI);
920 }
921
922 // Delete the instructions that we scalarized. If a full vector result
923 // is still needed, recreate it using InsertElements.
finish()924 bool ScalarizerVisitor::finish() {
925 // The presence of data in Gathered or Scattered indicates changes
926 // made to the Function.
927 if (Gathered.empty() && Scattered.empty())
928 return false;
929 for (const auto &GMI : Gathered) {
930 Instruction *Op = GMI.first;
931 ValueVector &CV = *GMI.second;
932 if (!Op->use_empty()) {
933 // The value is still needed, so recreate it using a series of
934 // InsertElements.
935 Value *Res = PoisonValue::get(Op->getType());
936 if (auto *Ty = dyn_cast<VectorType>(Op->getType())) {
937 BasicBlock *BB = Op->getParent();
938 unsigned Count = cast<FixedVectorType>(Ty)->getNumElements();
939 IRBuilder<> Builder(Op);
940 if (isa<PHINode>(Op))
941 Builder.SetInsertPoint(BB, BB->getFirstInsertionPt());
942 for (unsigned I = 0; I < Count; ++I)
943 Res = Builder.CreateInsertElement(Res, CV[I], Builder.getInt32(I),
944 Op->getName() + ".upto" + Twine(I));
945 Res->takeName(Op);
946 } else {
947 assert(CV.size() == 1 && Op->getType() == CV[0]->getType());
948 Res = CV[0];
949 if (Op == Res)
950 continue;
951 }
952 Op->replaceAllUsesWith(Res);
953 }
954 PotentiallyDeadInstrs.emplace_back(Op);
955 }
956 Gathered.clear();
957 Scattered.clear();
958
959 RecursivelyDeleteTriviallyDeadInstructionsPermissive(PotentiallyDeadInstrs);
960
961 return true;
962 }
963
run(Function & F,FunctionAnalysisManager & AM)964 PreservedAnalyses ScalarizerPass::run(Function &F, FunctionAnalysisManager &AM) {
965 Module &M = *F.getParent();
966 unsigned ParallelLoopAccessMDKind =
967 M.getContext().getMDKindID("llvm.mem.parallel_loop_access");
968 DominatorTree *DT = &AM.getResult<DominatorTreeAnalysis>(F);
969 ScalarizerVisitor Impl(ParallelLoopAccessMDKind, DT);
970 bool Changed = Impl.visit(F);
971 PreservedAnalyses PA;
972 PA.preserve<DominatorTreeAnalysis>();
973 return Changed ? PA : PreservedAnalyses::all();
974 }
975