1 //===- CodeGenPrepare.cpp - Prepare a function for code generation --------===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // This pass munges the code in the input function to better prepare it for
11 // SelectionDAG-based code generation. This works around limitations in it's
12 // basic-block-at-a-time approach. It should eventually be removed.
14 //===----------------------------------------------------------------------===//
16 #include "llvm/CodeGen/Passes.h"
17 #include "llvm/ADT/DenseMap.h"
18 #include "llvm/ADT/SmallSet.h"
19 #include "llvm/ADT/Statistic.h"
20 #include "llvm/Analysis/InstructionSimplify.h"
21 #include "llvm/Analysis/TargetLibraryInfo.h"
22 #include "llvm/Analysis/TargetTransformInfo.h"
23 #include "llvm/Analysis/ValueTracking.h"
24 #include "llvm/IR/CallSite.h"
25 #include "llvm/IR/Constants.h"
26 #include "llvm/IR/DataLayout.h"
27 #include "llvm/IR/DerivedTypes.h"
28 #include "llvm/IR/Dominators.h"
29 #include "llvm/IR/Function.h"
30 #include "llvm/IR/GetElementPtrTypeIterator.h"
31 #include "llvm/IR/IRBuilder.h"
32 #include "llvm/IR/InlineAsm.h"
33 #include "llvm/IR/Instructions.h"
34 #include "llvm/IR/IntrinsicInst.h"
35 #include "llvm/IR/MDBuilder.h"
36 #include "llvm/IR/PatternMatch.h"
37 #include "llvm/IR/Statepoint.h"
38 #include "llvm/IR/ValueHandle.h"
39 #include "llvm/IR/ValueMap.h"
40 #include "llvm/Pass.h"
41 #include "llvm/Support/CommandLine.h"
42 #include "llvm/Support/Debug.h"
43 #include "llvm/Support/raw_ostream.h"
44 #include "llvm/Target/TargetLowering.h"
45 #include "llvm/Target/TargetSubtargetInfo.h"
46 #include "llvm/Transforms/Utils/BasicBlockUtils.h"
47 #include "llvm/Transforms/Utils/BuildLibCalls.h"
48 #include "llvm/Transforms/Utils/BypassSlowDivision.h"
49 #include "llvm/Transforms/Utils/Local.h"
50 #include "llvm/Transforms/Utils/SimplifyLibCalls.h"
52 using namespace llvm::PatternMatch;
54 #define DEBUG_TYPE "codegenprepare"
56 STATISTIC(NumBlocksElim, "Number of blocks eliminated");
57 STATISTIC(NumPHIsElim, "Number of trivial PHIs eliminated");
58 STATISTIC(NumGEPsElim, "Number of GEPs converted to casts");
59 STATISTIC(NumCmpUses, "Number of uses of Cmp expressions replaced with uses of "
61 STATISTIC(NumCastUses, "Number of uses of Cast expressions replaced with uses "
63 STATISTIC(NumMemoryInsts, "Number of memory instructions whose address "
64 "computations were sunk");
65 STATISTIC(NumExtsMoved, "Number of [s|z]ext instructions combined with loads");
66 STATISTIC(NumExtUses, "Number of uses of [s|z]ext instructions optimized");
67 STATISTIC(NumAndsAdded,
68 "Number of and mask instructions added to form ext loads");
69 STATISTIC(NumAndUses, "Number of uses of and mask instructions optimized");
70 STATISTIC(NumRetsDup, "Number of return instructions duplicated");
71 STATISTIC(NumDbgValueMoved, "Number of debug value instructions moved");
72 STATISTIC(NumSelectsExpanded, "Number of selects turned into branches");
73 STATISTIC(NumAndCmpsMoved, "Number of and/cmp's pushed into branches");
74 STATISTIC(NumStoreExtractExposed, "Number of store(extractelement) exposed");
76 static cl::opt<bool> DisableBranchOpts(
77 "disable-cgp-branch-opts", cl::Hidden, cl::init(false),
78 cl::desc("Disable branch optimizations in CodeGenPrepare"));
81 DisableGCOpts("disable-cgp-gc-opts", cl::Hidden, cl::init(false),
82 cl::desc("Disable GC optimizations in CodeGenPrepare"));
84 static cl::opt<bool> DisableSelectToBranch(
85 "disable-cgp-select2branch", cl::Hidden, cl::init(false),
86 cl::desc("Disable select to branch conversion."));
88 static cl::opt<bool> AddrSinkUsingGEPs(
89 "addr-sink-using-gep", cl::Hidden, cl::init(false),
90 cl::desc("Address sinking in CGP using GEPs."));
92 static cl::opt<bool> EnableAndCmpSinking(
93 "enable-andcmp-sinking", cl::Hidden, cl::init(true),
94 cl::desc("Enable sinkinig and/cmp into branches."));
96 static cl::opt<bool> DisableStoreExtract(
97 "disable-cgp-store-extract", cl::Hidden, cl::init(false),
98 cl::desc("Disable store(extract) optimizations in CodeGenPrepare"));
100 static cl::opt<bool> StressStoreExtract(
101 "stress-cgp-store-extract", cl::Hidden, cl::init(false),
102 cl::desc("Stress test store(extract) optimizations in CodeGenPrepare"));
104 static cl::opt<bool> DisableExtLdPromotion(
105 "disable-cgp-ext-ld-promotion", cl::Hidden, cl::init(false),
106 cl::desc("Disable ext(promotable(ld)) -> promoted(ext(ld)) optimization in "
109 static cl::opt<bool> StressExtLdPromotion(
110 "stress-cgp-ext-ld-promotion", cl::Hidden, cl::init(false),
111 cl::desc("Stress test ext(promotable(ld)) -> promoted(ext(ld)) "
112 "optimization in CodeGenPrepare"));
115 typedef SmallPtrSet<Instruction *, 16> SetOfInstrs;
116 typedef PointerIntPair<Type *, 1, bool> TypeIsSExt;
117 typedef DenseMap<Instruction *, TypeIsSExt> InstrToOrigTy;
118 class TypePromotionTransaction;
120 class CodeGenPrepare : public FunctionPass {
121 const TargetMachine *TM;
122 const TargetLowering *TLI;
123 const TargetTransformInfo *TTI;
124 const TargetLibraryInfo *TLInfo;
126 /// As we scan instructions optimizing them, this is the next instruction
127 /// to optimize. Transforms that can invalidate this should update it.
128 BasicBlock::iterator CurInstIterator;
130 /// Keeps track of non-local addresses that have been sunk into a block.
131 /// This allows us to avoid inserting duplicate code for blocks with
132 /// multiple load/stores of the same address.
133 ValueMap<Value*, Value*> SunkAddrs;
135 /// Keeps track of all instructions inserted for the current function.
136 SetOfInstrs InsertedInsts;
137 /// Keeps track of the type of the related instruction before their
138 /// promotion for the current function.
139 InstrToOrigTy PromotedInsts;
141 /// True if CFG is modified in any way.
144 /// True if optimizing for size.
147 /// DataLayout for the Function being processed.
148 const DataLayout *DL;
151 static char ID; // Pass identification, replacement for typeid
152 explicit CodeGenPrepare(const TargetMachine *TM = nullptr)
153 : FunctionPass(ID), TM(TM), TLI(nullptr), TTI(nullptr), DL(nullptr) {
154 initializeCodeGenPreparePass(*PassRegistry::getPassRegistry());
156 bool runOnFunction(Function &F) override;
158 const char *getPassName() const override { return "CodeGen Prepare"; }
160 void getAnalysisUsage(AnalysisUsage &AU) const override {
161 AU.addPreserved<DominatorTreeWrapperPass>();
162 AU.addRequired<TargetLibraryInfoWrapperPass>();
163 AU.addRequired<TargetTransformInfoWrapperPass>();
167 bool eliminateFallThrough(Function &F);
168 bool eliminateMostlyEmptyBlocks(Function &F);
169 bool canMergeBlocks(const BasicBlock *BB, const BasicBlock *DestBB) const;
170 void eliminateMostlyEmptyBlock(BasicBlock *BB);
171 bool optimizeBlock(BasicBlock &BB, bool& ModifiedDT);
172 bool optimizeInst(Instruction *I, bool& ModifiedDT);
173 bool optimizeMemoryInst(Instruction *I, Value *Addr,
174 Type *AccessTy, unsigned AS);
175 bool optimizeInlineAsmInst(CallInst *CS);
176 bool optimizeCallInst(CallInst *CI, bool& ModifiedDT);
177 bool moveExtToFormExtLoad(Instruction *&I);
178 bool optimizeExtUses(Instruction *I);
179 bool optimizeLoadExt(LoadInst *I);
180 bool optimizeSelectInst(SelectInst *SI);
181 bool optimizeShuffleVectorInst(ShuffleVectorInst *SI);
182 bool optimizeSwitchInst(SwitchInst *CI);
183 bool optimizeExtractElementInst(Instruction *Inst);
184 bool dupRetToEnableTailCallOpts(BasicBlock *BB);
185 bool placeDbgValues(Function &F);
186 bool sinkAndCmp(Function &F);
187 bool extLdPromotion(TypePromotionTransaction &TPT, LoadInst *&LI,
189 const SmallVectorImpl<Instruction *> &Exts,
190 unsigned CreatedInstCost);
191 bool splitBranchCondition(Function &F);
192 bool simplifyOffsetableRelocate(Instruction &I);
193 void stripInvariantGroupMetadata(Instruction &I);
197 char CodeGenPrepare::ID = 0;
198 INITIALIZE_TM_PASS(CodeGenPrepare, "codegenprepare",
199 "Optimize for code generation", false, false)
201 FunctionPass *llvm::createCodeGenPreparePass(const TargetMachine *TM) {
202 return new CodeGenPrepare(TM);
205 bool CodeGenPrepare::runOnFunction(Function &F) {
206 if (skipOptnoneFunction(F))
209 DL = &F.getParent()->getDataLayout();
211 bool EverMadeChange = false;
212 // Clear per function information.
213 InsertedInsts.clear();
214 PromotedInsts.clear();
218 TLI = TM->getSubtargetImpl(F)->getTargetLowering();
219 TLInfo = &getAnalysis<TargetLibraryInfoWrapperPass>().getTLI();
220 TTI = &getAnalysis<TargetTransformInfoWrapperPass>().getTTI(F);
221 OptSize = F.optForSize();
223 /// This optimization identifies DIV instructions that can be
224 /// profitably bypassed and carried out with a shorter, faster divide.
225 if (!OptSize && TLI && TLI->isSlowDivBypassed()) {
226 const DenseMap<unsigned int, unsigned int> &BypassWidths =
227 TLI->getBypassSlowDivWidths();
228 BasicBlock* BB = &*F.begin();
229 while (BB != nullptr) {
230 // bypassSlowDivision may create new BBs, but we don't want to reapply the
231 // optimization to those blocks.
232 BasicBlock* Next = BB->getNextNode();
233 EverMadeChange |= bypassSlowDivision(BB, BypassWidths);
238 // Eliminate blocks that contain only PHI nodes and an
239 // unconditional branch.
240 EverMadeChange |= eliminateMostlyEmptyBlocks(F);
242 // llvm.dbg.value is far away from the value then iSel may not be able
243 // handle it properly. iSel will drop llvm.dbg.value if it can not
244 // find a node corresponding to the value.
245 EverMadeChange |= placeDbgValues(F);
247 // If there is a mask, compare against zero, and branch that can be combined
248 // into a single target instruction, push the mask and compare into branch
249 // users. Do this before OptimizeBlock -> OptimizeInst ->
250 // OptimizeCmpExpression, which perturbs the pattern being searched for.
251 if (!DisableBranchOpts) {
252 EverMadeChange |= sinkAndCmp(F);
253 EverMadeChange |= splitBranchCondition(F);
256 bool MadeChange = true;
259 for (Function::iterator I = F.begin(); I != F.end(); ) {
260 BasicBlock *BB = &*I++;
261 bool ModifiedDTOnIteration = false;
262 MadeChange |= optimizeBlock(*BB, ModifiedDTOnIteration);
264 // Restart BB iteration if the dominator tree of the Function was changed
265 if (ModifiedDTOnIteration)
268 EverMadeChange |= MadeChange;
273 if (!DisableBranchOpts) {
275 SmallPtrSet<BasicBlock*, 8> WorkList;
276 for (BasicBlock &BB : F) {
277 SmallVector<BasicBlock *, 2> Successors(succ_begin(&BB), succ_end(&BB));
278 MadeChange |= ConstantFoldTerminator(&BB, true);
279 if (!MadeChange) continue;
281 for (SmallVectorImpl<BasicBlock*>::iterator
282 II = Successors.begin(), IE = Successors.end(); II != IE; ++II)
283 if (pred_begin(*II) == pred_end(*II))
284 WorkList.insert(*II);
287 // Delete the dead blocks and any of their dead successors.
288 MadeChange |= !WorkList.empty();
289 while (!WorkList.empty()) {
290 BasicBlock *BB = *WorkList.begin();
292 SmallVector<BasicBlock*, 2> Successors(succ_begin(BB), succ_end(BB));
296 for (SmallVectorImpl<BasicBlock*>::iterator
297 II = Successors.begin(), IE = Successors.end(); II != IE; ++II)
298 if (pred_begin(*II) == pred_end(*II))
299 WorkList.insert(*II);
302 // Merge pairs of basic blocks with unconditional branches, connected by
304 if (EverMadeChange || MadeChange)
305 MadeChange |= eliminateFallThrough(F);
307 EverMadeChange |= MadeChange;
310 if (!DisableGCOpts) {
311 SmallVector<Instruction *, 2> Statepoints;
312 for (BasicBlock &BB : F)
313 for (Instruction &I : BB)
315 Statepoints.push_back(&I);
316 for (auto &I : Statepoints)
317 EverMadeChange |= simplifyOffsetableRelocate(*I);
320 return EverMadeChange;
323 /// Merge basic blocks which are connected by a single edge, where one of the
324 /// basic blocks has a single successor pointing to the other basic block,
325 /// which has a single predecessor.
326 bool CodeGenPrepare::eliminateFallThrough(Function &F) {
327 bool Changed = false;
328 // Scan all of the blocks in the function, except for the entry block.
329 for (Function::iterator I = std::next(F.begin()), E = F.end(); I != E;) {
330 BasicBlock *BB = &*I++;
331 // If the destination block has a single pred, then this is a trivial
332 // edge, just collapse it.
333 BasicBlock *SinglePred = BB->getSinglePredecessor();
335 // Don't merge if BB's address is taken.
336 if (!SinglePred || SinglePred == BB || BB->hasAddressTaken()) continue;
338 BranchInst *Term = dyn_cast<BranchInst>(SinglePred->getTerminator());
339 if (Term && !Term->isConditional()) {
341 DEBUG(dbgs() << "To merge:\n"<< *SinglePred << "\n\n\n");
342 // Remember if SinglePred was the entry block of the function.
343 // If so, we will need to move BB back to the entry position.
344 bool isEntry = SinglePred == &SinglePred->getParent()->getEntryBlock();
345 MergeBasicBlockIntoOnlyPred(BB, nullptr);
347 if (isEntry && BB != &BB->getParent()->getEntryBlock())
348 BB->moveBefore(&BB->getParent()->getEntryBlock());
350 // We have erased a block. Update the iterator.
351 I = BB->getIterator();
357 /// Eliminate blocks that contain only PHI nodes, debug info directives, and an
358 /// unconditional branch. Passes before isel (e.g. LSR/loopsimplify) often split
359 /// edges in ways that are non-optimal for isel. Start by eliminating these
360 /// blocks so we can split them the way we want them.
361 bool CodeGenPrepare::eliminateMostlyEmptyBlocks(Function &F) {
362 bool MadeChange = false;
363 // Note that this intentionally skips the entry block.
364 for (Function::iterator I = std::next(F.begin()), E = F.end(); I != E;) {
365 BasicBlock *BB = &*I++;
367 // If this block doesn't end with an uncond branch, ignore it.
368 BranchInst *BI = dyn_cast<BranchInst>(BB->getTerminator());
369 if (!BI || !BI->isUnconditional())
372 // If the instruction before the branch (skipping debug info) isn't a phi
373 // node, then other stuff is happening here.
374 BasicBlock::iterator BBI = BI->getIterator();
375 if (BBI != BB->begin()) {
377 while (isa<DbgInfoIntrinsic>(BBI)) {
378 if (BBI == BB->begin())
382 if (!isa<DbgInfoIntrinsic>(BBI) && !isa<PHINode>(BBI))
386 // Do not break infinite loops.
387 BasicBlock *DestBB = BI->getSuccessor(0);
391 if (!canMergeBlocks(BB, DestBB))
394 eliminateMostlyEmptyBlock(BB);
400 /// Return true if we can merge BB into DestBB if there is a single
401 /// unconditional branch between them, and BB contains no other non-phi
403 bool CodeGenPrepare::canMergeBlocks(const BasicBlock *BB,
404 const BasicBlock *DestBB) const {
405 // We only want to eliminate blocks whose phi nodes are used by phi nodes in
406 // the successor. If there are more complex condition (e.g. preheaders),
407 // don't mess around with them.
408 BasicBlock::const_iterator BBI = BB->begin();
409 while (const PHINode *PN = dyn_cast<PHINode>(BBI++)) {
410 for (const User *U : PN->users()) {
411 const Instruction *UI = cast<Instruction>(U);
412 if (UI->getParent() != DestBB || !isa<PHINode>(UI))
414 // If User is inside DestBB block and it is a PHINode then check
415 // incoming value. If incoming value is not from BB then this is
416 // a complex condition (e.g. preheaders) we want to avoid here.
417 if (UI->getParent() == DestBB) {
418 if (const PHINode *UPN = dyn_cast<PHINode>(UI))
419 for (unsigned I = 0, E = UPN->getNumIncomingValues(); I != E; ++I) {
420 Instruction *Insn = dyn_cast<Instruction>(UPN->getIncomingValue(I));
421 if (Insn && Insn->getParent() == BB &&
422 Insn->getParent() != UPN->getIncomingBlock(I))
429 // If BB and DestBB contain any common predecessors, then the phi nodes in BB
430 // and DestBB may have conflicting incoming values for the block. If so, we
431 // can't merge the block.
432 const PHINode *DestBBPN = dyn_cast<PHINode>(DestBB->begin());
433 if (!DestBBPN) return true; // no conflict.
435 // Collect the preds of BB.
436 SmallPtrSet<const BasicBlock*, 16> BBPreds;
437 if (const PHINode *BBPN = dyn_cast<PHINode>(BB->begin())) {
438 // It is faster to get preds from a PHI than with pred_iterator.
439 for (unsigned i = 0, e = BBPN->getNumIncomingValues(); i != e; ++i)
440 BBPreds.insert(BBPN->getIncomingBlock(i));
442 BBPreds.insert(pred_begin(BB), pred_end(BB));
445 // Walk the preds of DestBB.
446 for (unsigned i = 0, e = DestBBPN->getNumIncomingValues(); i != e; ++i) {
447 BasicBlock *Pred = DestBBPN->getIncomingBlock(i);
448 if (BBPreds.count(Pred)) { // Common predecessor?
449 BBI = DestBB->begin();
450 while (const PHINode *PN = dyn_cast<PHINode>(BBI++)) {
451 const Value *V1 = PN->getIncomingValueForBlock(Pred);
452 const Value *V2 = PN->getIncomingValueForBlock(BB);
454 // If V2 is a phi node in BB, look up what the mapped value will be.
455 if (const PHINode *V2PN = dyn_cast<PHINode>(V2))
456 if (V2PN->getParent() == BB)
457 V2 = V2PN->getIncomingValueForBlock(Pred);
459 // If there is a conflict, bail out.
460 if (V1 != V2) return false;
469 /// Eliminate a basic block that has only phi's and an unconditional branch in
471 void CodeGenPrepare::eliminateMostlyEmptyBlock(BasicBlock *BB) {
472 BranchInst *BI = cast<BranchInst>(BB->getTerminator());
473 BasicBlock *DestBB = BI->getSuccessor(0);
475 DEBUG(dbgs() << "MERGING MOSTLY EMPTY BLOCKS - BEFORE:\n" << *BB << *DestBB);
477 // If the destination block has a single pred, then this is a trivial edge,
479 if (BasicBlock *SinglePred = DestBB->getSinglePredecessor()) {
480 if (SinglePred != DestBB) {
481 // Remember if SinglePred was the entry block of the function. If so, we
482 // will need to move BB back to the entry position.
483 bool isEntry = SinglePred == &SinglePred->getParent()->getEntryBlock();
484 MergeBasicBlockIntoOnlyPred(DestBB, nullptr);
486 if (isEntry && BB != &BB->getParent()->getEntryBlock())
487 BB->moveBefore(&BB->getParent()->getEntryBlock());
489 DEBUG(dbgs() << "AFTER:\n" << *DestBB << "\n\n\n");
494 // Otherwise, we have multiple predecessors of BB. Update the PHIs in DestBB
495 // to handle the new incoming edges it is about to have.
497 for (BasicBlock::iterator BBI = DestBB->begin();
498 (PN = dyn_cast<PHINode>(BBI)); ++BBI) {
499 // Remove the incoming value for BB, and remember it.
500 Value *InVal = PN->removeIncomingValue(BB, false);
502 // Two options: either the InVal is a phi node defined in BB or it is some
503 // value that dominates BB.
504 PHINode *InValPhi = dyn_cast<PHINode>(InVal);
505 if (InValPhi && InValPhi->getParent() == BB) {
506 // Add all of the input values of the input PHI as inputs of this phi.
507 for (unsigned i = 0, e = InValPhi->getNumIncomingValues(); i != e; ++i)
508 PN->addIncoming(InValPhi->getIncomingValue(i),
509 InValPhi->getIncomingBlock(i));
511 // Otherwise, add one instance of the dominating value for each edge that
512 // we will be adding.
513 if (PHINode *BBPN = dyn_cast<PHINode>(BB->begin())) {
514 for (unsigned i = 0, e = BBPN->getNumIncomingValues(); i != e; ++i)
515 PN->addIncoming(InVal, BBPN->getIncomingBlock(i));
517 for (pred_iterator PI = pred_begin(BB), E = pred_end(BB); PI != E; ++PI)
518 PN->addIncoming(InVal, *PI);
523 // The PHIs are now updated, change everything that refers to BB to use
524 // DestBB and remove BB.
525 BB->replaceAllUsesWith(DestBB);
526 BB->eraseFromParent();
529 DEBUG(dbgs() << "AFTER:\n" << *DestBB << "\n\n\n");
532 // Computes a map of base pointer relocation instructions to corresponding
533 // derived pointer relocation instructions given a vector of all relocate calls
534 static void computeBaseDerivedRelocateMap(
535 const SmallVectorImpl<User *> &AllRelocateCalls,
536 DenseMap<IntrinsicInst *, SmallVector<IntrinsicInst *, 2>> &
538 // Collect information in two maps: one primarily for locating the base object
539 // while filling the second map; the second map is the final structure holding
540 // a mapping between Base and corresponding Derived relocate calls
541 DenseMap<std::pair<unsigned, unsigned>, IntrinsicInst *> RelocateIdxMap;
542 for (auto &U : AllRelocateCalls) {
543 GCRelocateOperands ThisRelocate(U);
544 IntrinsicInst *I = cast<IntrinsicInst>(U);
545 auto K = std::make_pair(ThisRelocate.getBasePtrIndex(),
546 ThisRelocate.getDerivedPtrIndex());
547 RelocateIdxMap.insert(std::make_pair(K, I));
549 for (auto &Item : RelocateIdxMap) {
550 std::pair<unsigned, unsigned> Key = Item.first;
551 if (Key.first == Key.second)
552 // Base relocation: nothing to insert
555 IntrinsicInst *I = Item.second;
556 auto BaseKey = std::make_pair(Key.first, Key.first);
558 // We're iterating over RelocateIdxMap so we cannot modify it.
559 auto MaybeBase = RelocateIdxMap.find(BaseKey);
560 if (MaybeBase == RelocateIdxMap.end())
561 // TODO: We might want to insert a new base object relocate and gep off
562 // that, if there are enough derived object relocates.
565 RelocateInstMap[MaybeBase->second].push_back(I);
569 // Accepts a GEP and extracts the operands into a vector provided they're all
570 // small integer constants
571 static bool getGEPSmallConstantIntOffsetV(GetElementPtrInst *GEP,
572 SmallVectorImpl<Value *> &OffsetV) {
573 for (unsigned i = 1; i < GEP->getNumOperands(); i++) {
574 // Only accept small constant integer operands
575 auto Op = dyn_cast<ConstantInt>(GEP->getOperand(i));
576 if (!Op || Op->getZExtValue() > 20)
580 for (unsigned i = 1; i < GEP->getNumOperands(); i++)
581 OffsetV.push_back(GEP->getOperand(i));
585 // Takes a RelocatedBase (base pointer relocation instruction) and Targets to
586 // replace, computes a replacement, and affects it.
588 simplifyRelocatesOffABase(IntrinsicInst *RelocatedBase,
589 const SmallVectorImpl<IntrinsicInst *> &Targets) {
590 bool MadeChange = false;
591 for (auto &ToReplace : Targets) {
592 GCRelocateOperands MasterRelocate(RelocatedBase);
593 GCRelocateOperands ThisRelocate(ToReplace);
595 assert(ThisRelocate.getBasePtrIndex() == MasterRelocate.getBasePtrIndex() &&
596 "Not relocating a derived object of the original base object");
597 if (ThisRelocate.getBasePtrIndex() == ThisRelocate.getDerivedPtrIndex()) {
598 // A duplicate relocate call. TODO: coalesce duplicates.
602 if (RelocatedBase->getParent() != ToReplace->getParent()) {
603 // Base and derived relocates are in different basic blocks.
604 // In this case transform is only valid when base dominates derived
605 // relocate. However it would be too expensive to check dominance
606 // for each such relocate, so we skip the whole transformation.
610 Value *Base = ThisRelocate.getBasePtr();
611 auto Derived = dyn_cast<GetElementPtrInst>(ThisRelocate.getDerivedPtr());
612 if (!Derived || Derived->getPointerOperand() != Base)
615 SmallVector<Value *, 2> OffsetV;
616 if (!getGEPSmallConstantIntOffsetV(Derived, OffsetV))
619 // Create a Builder and replace the target callsite with a gep
620 assert(RelocatedBase->getNextNode() && "Should always have one since it's not a terminator");
622 // Insert after RelocatedBase
623 IRBuilder<> Builder(RelocatedBase->getNextNode());
624 Builder.SetCurrentDebugLocation(ToReplace->getDebugLoc());
626 // If gc_relocate does not match the actual type, cast it to the right type.
627 // In theory, there must be a bitcast after gc_relocate if the type does not
628 // match, and we should reuse it to get the derived pointer. But it could be
632 // %g1 = call coldcc i8 addrspace(1)* @llvm.experimental.gc.relocate.p1i8(...)
637 // %g2 = call coldcc i8 addrspace(1)* @llvm.experimental.gc.relocate.p1i8(...)
641 // %p1 = phi i8 addrspace(1)* [ %g1, %bb1 ], [ %g2, %bb2 ]
642 // %cast = bitcast i8 addrspace(1)* %p1 in to i32 addrspace(1)*
644 // In this case, we can not find the bitcast any more. So we insert a new bitcast
645 // no matter there is already one or not. In this way, we can handle all cases, and
646 // the extra bitcast should be optimized away in later passes.
647 Value *ActualRelocatedBase = RelocatedBase;
648 if (RelocatedBase->getType() != Base->getType()) {
649 ActualRelocatedBase =
650 Builder.CreateBitCast(RelocatedBase, Base->getType());
652 Value *Replacement = Builder.CreateGEP(
653 Derived->getSourceElementType(), ActualRelocatedBase, makeArrayRef(OffsetV));
654 Replacement->takeName(ToReplace);
655 // If the newly generated derived pointer's type does not match the original derived
656 // pointer's type, cast the new derived pointer to match it. Same reasoning as above.
657 Value *ActualReplacement = Replacement;
658 if (Replacement->getType() != ToReplace->getType()) {
660 Builder.CreateBitCast(Replacement, ToReplace->getType());
662 ToReplace->replaceAllUsesWith(ActualReplacement);
663 ToReplace->eraseFromParent();
673 // %ptr = gep %base + 15
674 // %tok = statepoint (%fun, i32 0, i32 0, i32 0, %base, %ptr)
675 // %base' = relocate(%tok, i32 4, i32 4)
676 // %ptr' = relocate(%tok, i32 4, i32 5)
682 // %ptr = gep %base + 15
683 // %tok = statepoint (%fun, i32 0, i32 0, i32 0, %base, %ptr)
684 // %base' = gc.relocate(%tok, i32 4, i32 4)
685 // %ptr' = gep %base' + 15
687 bool CodeGenPrepare::simplifyOffsetableRelocate(Instruction &I) {
688 bool MadeChange = false;
689 SmallVector<User *, 2> AllRelocateCalls;
691 for (auto *U : I.users())
692 if (isGCRelocate(dyn_cast<Instruction>(U)))
693 // Collect all the relocate calls associated with a statepoint
694 AllRelocateCalls.push_back(U);
696 // We need atleast one base pointer relocation + one derived pointer
697 // relocation to mangle
698 if (AllRelocateCalls.size() < 2)
701 // RelocateInstMap is a mapping from the base relocate instruction to the
702 // corresponding derived relocate instructions
703 DenseMap<IntrinsicInst *, SmallVector<IntrinsicInst *, 2>> RelocateInstMap;
704 computeBaseDerivedRelocateMap(AllRelocateCalls, RelocateInstMap);
705 if (RelocateInstMap.empty())
708 for (auto &Item : RelocateInstMap)
709 // Item.first is the RelocatedBase to offset against
710 // Item.second is the vector of Targets to replace
711 MadeChange = simplifyRelocatesOffABase(Item.first, Item.second);
715 /// SinkCast - Sink the specified cast instruction into its user blocks
716 static bool SinkCast(CastInst *CI) {
717 BasicBlock *DefBB = CI->getParent();
719 /// InsertedCasts - Only insert a cast in each block once.
720 DenseMap<BasicBlock*, CastInst*> InsertedCasts;
722 bool MadeChange = false;
723 for (Value::user_iterator UI = CI->user_begin(), E = CI->user_end();
725 Use &TheUse = UI.getUse();
726 Instruction *User = cast<Instruction>(*UI);
728 // Figure out which BB this cast is used in. For PHI's this is the
729 // appropriate predecessor block.
730 BasicBlock *UserBB = User->getParent();
731 if (PHINode *PN = dyn_cast<PHINode>(User)) {
732 UserBB = PN->getIncomingBlock(TheUse);
735 // Preincrement use iterator so we don't invalidate it.
738 // If the block selected to receive the cast is an EH pad that does not
739 // allow non-PHI instructions before the terminator, we can't sink the
741 if (UserBB->getTerminator()->isEHPad())
744 // If this user is in the same block as the cast, don't change the cast.
745 if (UserBB == DefBB) continue;
747 // If we have already inserted a cast into this block, use it.
748 CastInst *&InsertedCast = InsertedCasts[UserBB];
751 BasicBlock::iterator InsertPt = UserBB->getFirstInsertionPt();
752 assert(InsertPt != UserBB->end());
753 InsertedCast = CastInst::Create(CI->getOpcode(), CI->getOperand(0),
754 CI->getType(), "", &*InsertPt);
757 // Replace a use of the cast with a use of the new cast.
758 TheUse = InsertedCast;
763 // If we removed all uses, nuke the cast.
764 if (CI->use_empty()) {
765 CI->eraseFromParent();
772 /// If the specified cast instruction is a noop copy (e.g. it's casting from
773 /// one pointer type to another, i32->i8 on PPC), sink it into user blocks to
774 /// reduce the number of virtual registers that must be created and coalesced.
776 /// Return true if any changes are made.
778 static bool OptimizeNoopCopyExpression(CastInst *CI, const TargetLowering &TLI,
779 const DataLayout &DL) {
780 // If this is a noop copy,
781 EVT SrcVT = TLI.getValueType(DL, CI->getOperand(0)->getType());
782 EVT DstVT = TLI.getValueType(DL, CI->getType());
784 // This is an fp<->int conversion?
785 if (SrcVT.isInteger() != DstVT.isInteger())
788 // If this is an extension, it will be a zero or sign extension, which
790 if (SrcVT.bitsLT(DstVT)) return false;
792 // If these values will be promoted, find out what they will be promoted
793 // to. This helps us consider truncates on PPC as noop copies when they
795 if (TLI.getTypeAction(CI->getContext(), SrcVT) ==
796 TargetLowering::TypePromoteInteger)
797 SrcVT = TLI.getTypeToTransformTo(CI->getContext(), SrcVT);
798 if (TLI.getTypeAction(CI->getContext(), DstVT) ==
799 TargetLowering::TypePromoteInteger)
800 DstVT = TLI.getTypeToTransformTo(CI->getContext(), DstVT);
802 // If, after promotion, these are the same types, this is a noop copy.
809 /// Try to combine CI into a call to the llvm.uadd.with.overflow intrinsic if
812 /// Return true if any changes were made.
813 static bool CombineUAddWithOverflow(CmpInst *CI) {
817 m_UAddWithOverflow(m_Value(A), m_Value(B), m_Instruction(AddI))))
820 Type *Ty = AddI->getType();
821 if (!isa<IntegerType>(Ty))
824 // We don't want to move around uses of condition values this late, so we we
825 // check if it is legal to create the call to the intrinsic in the basic
826 // block containing the icmp:
828 if (AddI->getParent() != CI->getParent() && !AddI->hasOneUse())
832 // Someday m_UAddWithOverflow may get smarter, but this is a safe assumption
834 if (AddI->hasOneUse())
835 assert(*AddI->user_begin() == CI && "expected!");
838 Module *M = CI->getModule();
839 Value *F = Intrinsic::getDeclaration(M, Intrinsic::uadd_with_overflow, Ty);
841 auto *InsertPt = AddI->hasOneUse() ? CI : AddI;
843 auto *UAddWithOverflow =
844 CallInst::Create(F, {A, B}, "uadd.overflow", InsertPt);
845 auto *UAdd = ExtractValueInst::Create(UAddWithOverflow, 0, "uadd", InsertPt);
847 ExtractValueInst::Create(UAddWithOverflow, 1, "overflow", InsertPt);
849 CI->replaceAllUsesWith(Overflow);
850 AddI->replaceAllUsesWith(UAdd);
851 CI->eraseFromParent();
852 AddI->eraseFromParent();
856 /// Sink the given CmpInst into user blocks to reduce the number of virtual
857 /// registers that must be created and coalesced. This is a clear win except on
858 /// targets with multiple condition code registers (PowerPC), where it might
859 /// lose; some adjustment may be wanted there.
861 /// Return true if any changes are made.
862 static bool SinkCmpExpression(CmpInst *CI) {
863 BasicBlock *DefBB = CI->getParent();
865 /// Only insert a cmp in each block once.
866 DenseMap<BasicBlock*, CmpInst*> InsertedCmps;
868 bool MadeChange = false;
869 for (Value::user_iterator UI = CI->user_begin(), E = CI->user_end();
871 Use &TheUse = UI.getUse();
872 Instruction *User = cast<Instruction>(*UI);
874 // Preincrement use iterator so we don't invalidate it.
877 // Don't bother for PHI nodes.
878 if (isa<PHINode>(User))
881 // Figure out which BB this cmp is used in.
882 BasicBlock *UserBB = User->getParent();
884 // If this user is in the same block as the cmp, don't change the cmp.
885 if (UserBB == DefBB) continue;
887 // If we have already inserted a cmp into this block, use it.
888 CmpInst *&InsertedCmp = InsertedCmps[UserBB];
891 BasicBlock::iterator InsertPt = UserBB->getFirstInsertionPt();
892 assert(InsertPt != UserBB->end());
894 CmpInst::Create(CI->getOpcode(), CI->getPredicate(),
895 CI->getOperand(0), CI->getOperand(1), "", &*InsertPt);
898 // Replace a use of the cmp with a use of the new cmp.
899 TheUse = InsertedCmp;
904 // If we removed all uses, nuke the cmp.
905 if (CI->use_empty()) {
906 CI->eraseFromParent();
913 static bool OptimizeCmpExpression(CmpInst *CI) {
914 if (SinkCmpExpression(CI))
917 if (CombineUAddWithOverflow(CI))
923 /// Check if the candidates could be combined with a shift instruction, which
925 /// 1. Truncate instruction
926 /// 2. And instruction and the imm is a mask of the low bits:
927 /// imm & (imm+1) == 0
928 static bool isExtractBitsCandidateUse(Instruction *User) {
929 if (!isa<TruncInst>(User)) {
930 if (User->getOpcode() != Instruction::And ||
931 !isa<ConstantInt>(User->getOperand(1)))
934 const APInt &Cimm = cast<ConstantInt>(User->getOperand(1))->getValue();
936 if ((Cimm & (Cimm + 1)).getBoolValue())
942 /// Sink both shift and truncate instruction to the use of truncate's BB.
944 SinkShiftAndTruncate(BinaryOperator *ShiftI, Instruction *User, ConstantInt *CI,
945 DenseMap<BasicBlock *, BinaryOperator *> &InsertedShifts,
946 const TargetLowering &TLI, const DataLayout &DL) {
947 BasicBlock *UserBB = User->getParent();
948 DenseMap<BasicBlock *, CastInst *> InsertedTruncs;
949 TruncInst *TruncI = dyn_cast<TruncInst>(User);
950 bool MadeChange = false;
952 for (Value::user_iterator TruncUI = TruncI->user_begin(),
953 TruncE = TruncI->user_end();
954 TruncUI != TruncE;) {
956 Use &TruncTheUse = TruncUI.getUse();
957 Instruction *TruncUser = cast<Instruction>(*TruncUI);
958 // Preincrement use iterator so we don't invalidate it.
962 int ISDOpcode = TLI.InstructionOpcodeToISD(TruncUser->getOpcode());
966 // If the use is actually a legal node, there will not be an
967 // implicit truncate.
968 // FIXME: always querying the result type is just an
969 // approximation; some nodes' legality is determined by the
970 // operand or other means. There's no good way to find out though.
971 if (TLI.isOperationLegalOrCustom(
972 ISDOpcode, TLI.getValueType(DL, TruncUser->getType(), true)))
975 // Don't bother for PHI nodes.
976 if (isa<PHINode>(TruncUser))
979 BasicBlock *TruncUserBB = TruncUser->getParent();
981 if (UserBB == TruncUserBB)
984 BinaryOperator *&InsertedShift = InsertedShifts[TruncUserBB];
985 CastInst *&InsertedTrunc = InsertedTruncs[TruncUserBB];
987 if (!InsertedShift && !InsertedTrunc) {
988 BasicBlock::iterator InsertPt = TruncUserBB->getFirstInsertionPt();
989 assert(InsertPt != TruncUserBB->end());
991 if (ShiftI->getOpcode() == Instruction::AShr)
992 InsertedShift = BinaryOperator::CreateAShr(ShiftI->getOperand(0), CI,
995 InsertedShift = BinaryOperator::CreateLShr(ShiftI->getOperand(0), CI,
999 BasicBlock::iterator TruncInsertPt = TruncUserBB->getFirstInsertionPt();
1001 assert(TruncInsertPt != TruncUserBB->end());
1003 InsertedTrunc = CastInst::Create(TruncI->getOpcode(), InsertedShift,
1004 TruncI->getType(), "", &*TruncInsertPt);
1008 TruncTheUse = InsertedTrunc;
1014 /// Sink the shift *right* instruction into user blocks if the uses could
1015 /// potentially be combined with this shift instruction and generate BitExtract
1016 /// instruction. It will only be applied if the architecture supports BitExtract
1017 /// instruction. Here is an example:
1019 /// %x.extract.shift = lshr i64 %arg1, 32
1021 /// %x.extract.trunc = trunc i64 %x.extract.shift to i16
1025 /// %x.extract.shift.1 = lshr i64 %arg1, 32
1026 /// %x.extract.trunc = trunc i64 %x.extract.shift.1 to i16
1028 /// CodeGen will recoginze the pattern in BB2 and generate BitExtract
1030 /// Return true if any changes are made.
1031 static bool OptimizeExtractBits(BinaryOperator *ShiftI, ConstantInt *CI,
1032 const TargetLowering &TLI,
1033 const DataLayout &DL) {
1034 BasicBlock *DefBB = ShiftI->getParent();
1036 /// Only insert instructions in each block once.
1037 DenseMap<BasicBlock *, BinaryOperator *> InsertedShifts;
1039 bool shiftIsLegal = TLI.isTypeLegal(TLI.getValueType(DL, ShiftI->getType()));
1041 bool MadeChange = false;
1042 for (Value::user_iterator UI = ShiftI->user_begin(), E = ShiftI->user_end();
1044 Use &TheUse = UI.getUse();
1045 Instruction *User = cast<Instruction>(*UI);
1046 // Preincrement use iterator so we don't invalidate it.
1049 // Don't bother for PHI nodes.
1050 if (isa<PHINode>(User))
1053 if (!isExtractBitsCandidateUse(User))
1056 BasicBlock *UserBB = User->getParent();
1058 if (UserBB == DefBB) {
1059 // If the shift and truncate instruction are in the same BB. The use of
1060 // the truncate(TruncUse) may still introduce another truncate if not
1061 // legal. In this case, we would like to sink both shift and truncate
1062 // instruction to the BB of TruncUse.
1065 // i64 shift.result = lshr i64 opnd, imm
1066 // trunc.result = trunc shift.result to i16
1069 // ----> We will have an implicit truncate here if the architecture does
1070 // not have i16 compare.
1071 // cmp i16 trunc.result, opnd2
1073 if (isa<TruncInst>(User) && shiftIsLegal
1074 // If the type of the truncate is legal, no trucate will be
1075 // introduced in other basic blocks.
1077 (!TLI.isTypeLegal(TLI.getValueType(DL, User->getType()))))
1079 SinkShiftAndTruncate(ShiftI, User, CI, InsertedShifts, TLI, DL);
1083 // If we have already inserted a shift into this block, use it.
1084 BinaryOperator *&InsertedShift = InsertedShifts[UserBB];
1086 if (!InsertedShift) {
1087 BasicBlock::iterator InsertPt = UserBB->getFirstInsertionPt();
1088 assert(InsertPt != UserBB->end());
1090 if (ShiftI->getOpcode() == Instruction::AShr)
1091 InsertedShift = BinaryOperator::CreateAShr(ShiftI->getOperand(0), CI,
1094 InsertedShift = BinaryOperator::CreateLShr(ShiftI->getOperand(0), CI,
1100 // Replace a use of the shift with a use of the new shift.
1101 TheUse = InsertedShift;
1104 // If we removed all uses, nuke the shift.
1105 if (ShiftI->use_empty())
1106 ShiftI->eraseFromParent();
1111 // Translate a masked load intrinsic like
1112 // <16 x i32 > @llvm.masked.load( <16 x i32>* %addr, i32 align,
1113 // <16 x i1> %mask, <16 x i32> %passthru)
1114 // to a chain of basic blocks, with loading element one-by-one if
1115 // the appropriate mask bit is set
1117 // %1 = bitcast i8* %addr to i32*
1118 // %2 = extractelement <16 x i1> %mask, i32 0
1119 // %3 = icmp eq i1 %2, true
1120 // br i1 %3, label %cond.load, label %else
1122 //cond.load: ; preds = %0
1123 // %4 = getelementptr i32* %1, i32 0
1124 // %5 = load i32* %4
1125 // %6 = insertelement <16 x i32> undef, i32 %5, i32 0
1128 //else: ; preds = %0, %cond.load
1129 // %res.phi.else = phi <16 x i32> [ %6, %cond.load ], [ undef, %0 ]
1130 // %7 = extractelement <16 x i1> %mask, i32 1
1131 // %8 = icmp eq i1 %7, true
1132 // br i1 %8, label %cond.load1, label %else2
1134 //cond.load1: ; preds = %else
1135 // %9 = getelementptr i32* %1, i32 1
1136 // %10 = load i32* %9
1137 // %11 = insertelement <16 x i32> %res.phi.else, i32 %10, i32 1
1140 //else2: ; preds = %else, %cond.load1
1141 // %res.phi.else3 = phi <16 x i32> [ %11, %cond.load1 ], [ %res.phi.else, %else ]
1142 // %12 = extractelement <16 x i1> %mask, i32 2
1143 // %13 = icmp eq i1 %12, true
1144 // br i1 %13, label %cond.load4, label %else5
1146 static void ScalarizeMaskedLoad(CallInst *CI) {
1147 Value *Ptr = CI->getArgOperand(0);
1148 Value *Alignment = CI->getArgOperand(1);
1149 Value *Mask = CI->getArgOperand(2);
1150 Value *Src0 = CI->getArgOperand(3);
1152 unsigned AlignVal = cast<ConstantInt>(Alignment)->getZExtValue();
1153 VectorType *VecType = dyn_cast<VectorType>(CI->getType());
1154 assert(VecType && "Unexpected return type of masked load intrinsic");
1156 Type *EltTy = CI->getType()->getVectorElementType();
1158 IRBuilder<> Builder(CI->getContext());
1159 Instruction *InsertPt = CI;
1160 BasicBlock *IfBlock = CI->getParent();
1161 BasicBlock *CondBlock = nullptr;
1162 BasicBlock *PrevIfBlock = CI->getParent();
1164 Builder.SetInsertPoint(InsertPt);
1165 Builder.SetCurrentDebugLocation(CI->getDebugLoc());
1167 // Short-cut if the mask is all-true.
1168 bool IsAllOnesMask = isa<Constant>(Mask) &&
1169 cast<Constant>(Mask)->isAllOnesValue();
1171 if (IsAllOnesMask) {
1172 Value *NewI = Builder.CreateAlignedLoad(Ptr, AlignVal);
1173 CI->replaceAllUsesWith(NewI);
1174 CI->eraseFromParent();
1178 // Adjust alignment for the scalar instruction.
1179 AlignVal = std::min(AlignVal, VecType->getScalarSizeInBits()/8);
1180 // Bitcast %addr fron i8* to EltTy*
1182 EltTy->getPointerTo(cast<PointerType>(Ptr->getType())->getAddressSpace());
1183 Value *FirstEltPtr = Builder.CreateBitCast(Ptr, NewPtrType);
1184 unsigned VectorWidth = VecType->getNumElements();
1186 Value *UndefVal = UndefValue::get(VecType);
1188 // The result vector
1189 Value *VResult = UndefVal;
1191 if (isa<ConstantVector>(Mask)) {
1192 for (unsigned Idx = 0; Idx < VectorWidth; ++Idx) {
1193 if (cast<ConstantVector>(Mask)->getOperand(Idx)->isNullValue())
1196 Builder.CreateInBoundsGEP(EltTy, FirstEltPtr, Builder.getInt32(Idx));
1197 LoadInst* Load = Builder.CreateAlignedLoad(Gep, AlignVal);
1198 VResult = Builder.CreateInsertElement(VResult, Load,
1199 Builder.getInt32(Idx));
1201 Value *NewI = Builder.CreateSelect(Mask, VResult, Src0);
1202 CI->replaceAllUsesWith(NewI);
1203 CI->eraseFromParent();
1207 PHINode *Phi = nullptr;
1208 Value *PrevPhi = UndefVal;
1210 for (unsigned Idx = 0; Idx < VectorWidth; ++Idx) {
1212 // Fill the "else" block, created in the previous iteration
1214 // %res.phi.else3 = phi <16 x i32> [ %11, %cond.load1 ], [ %res.phi.else, %else ]
1215 // %mask_1 = extractelement <16 x i1> %mask, i32 Idx
1216 // %to_load = icmp eq i1 %mask_1, true
1217 // br i1 %to_load, label %cond.load, label %else
1220 Phi = Builder.CreatePHI(VecType, 2, "res.phi.else");
1221 Phi->addIncoming(VResult, CondBlock);
1222 Phi->addIncoming(PrevPhi, PrevIfBlock);
1227 Value *Predicate = Builder.CreateExtractElement(Mask, Builder.getInt32(Idx));
1228 Value *Cmp = Builder.CreateICmp(ICmpInst::ICMP_EQ, Predicate,
1229 ConstantInt::get(Predicate->getType(), 1));
1231 // Create "cond" block
1233 // %EltAddr = getelementptr i32* %1, i32 0
1234 // %Elt = load i32* %EltAddr
1235 // VResult = insertelement <16 x i32> VResult, i32 %Elt, i32 Idx
1237 CondBlock = IfBlock->splitBasicBlock(InsertPt->getIterator(), "cond.load");
1238 Builder.SetInsertPoint(InsertPt);
1241 Builder.CreateInBoundsGEP(EltTy, FirstEltPtr, Builder.getInt32(Idx));
1242 LoadInst *Load = Builder.CreateAlignedLoad(Gep, AlignVal);
1243 VResult = Builder.CreateInsertElement(VResult, Load, Builder.getInt32(Idx));
1245 // Create "else" block, fill it in the next iteration
1246 BasicBlock *NewIfBlock =
1247 CondBlock->splitBasicBlock(InsertPt->getIterator(), "else");
1248 Builder.SetInsertPoint(InsertPt);
1249 Instruction *OldBr = IfBlock->getTerminator();
1250 BranchInst::Create(CondBlock, NewIfBlock, Cmp, OldBr);
1251 OldBr->eraseFromParent();
1252 PrevIfBlock = IfBlock;
1253 IfBlock = NewIfBlock;
1256 Phi = Builder.CreatePHI(VecType, 2, "res.phi.select");
1257 Phi->addIncoming(VResult, CondBlock);
1258 Phi->addIncoming(PrevPhi, PrevIfBlock);
1259 Value *NewI = Builder.CreateSelect(Mask, Phi, Src0);
1260 CI->replaceAllUsesWith(NewI);
1261 CI->eraseFromParent();
1264 // Translate a masked store intrinsic, like
1265 // void @llvm.masked.store(<16 x i32> %src, <16 x i32>* %addr, i32 align,
1267 // to a chain of basic blocks, that stores element one-by-one if
1268 // the appropriate mask bit is set
1270 // %1 = bitcast i8* %addr to i32*
1271 // %2 = extractelement <16 x i1> %mask, i32 0
1272 // %3 = icmp eq i1 %2, true
1273 // br i1 %3, label %cond.store, label %else
1275 // cond.store: ; preds = %0
1276 // %4 = extractelement <16 x i32> %val, i32 0
1277 // %5 = getelementptr i32* %1, i32 0
1278 // store i32 %4, i32* %5
1281 // else: ; preds = %0, %cond.store
1282 // %6 = extractelement <16 x i1> %mask, i32 1
1283 // %7 = icmp eq i1 %6, true
1284 // br i1 %7, label %cond.store1, label %else2
1286 // cond.store1: ; preds = %else
1287 // %8 = extractelement <16 x i32> %val, i32 1
1288 // %9 = getelementptr i32* %1, i32 1
1289 // store i32 %8, i32* %9
1292 static void ScalarizeMaskedStore(CallInst *CI) {
1293 Value *Src = CI->getArgOperand(0);
1294 Value *Ptr = CI->getArgOperand(1);
1295 Value *Alignment = CI->getArgOperand(2);
1296 Value *Mask = CI->getArgOperand(3);
1298 unsigned AlignVal = cast<ConstantInt>(Alignment)->getZExtValue();
1299 VectorType *VecType = dyn_cast<VectorType>(Src->getType());
1300 assert(VecType && "Unexpected data type in masked store intrinsic");
1302 Type *EltTy = VecType->getElementType();
1304 IRBuilder<> Builder(CI->getContext());
1305 Instruction *InsertPt = CI;
1306 BasicBlock *IfBlock = CI->getParent();
1307 Builder.SetInsertPoint(InsertPt);
1308 Builder.SetCurrentDebugLocation(CI->getDebugLoc());
1310 // Short-cut if the mask is all-true.
1311 bool IsAllOnesMask = isa<Constant>(Mask) &&
1312 cast<Constant>(Mask)->isAllOnesValue();
1314 if (IsAllOnesMask) {
1315 Builder.CreateAlignedStore(Src, Ptr, AlignVal);
1316 CI->eraseFromParent();
1320 // Adjust alignment for the scalar instruction.
1321 AlignVal = std::max(AlignVal, VecType->getScalarSizeInBits()/8);
1322 // Bitcast %addr fron i8* to EltTy*
1324 EltTy->getPointerTo(cast<PointerType>(Ptr->getType())->getAddressSpace());
1325 Value *FirstEltPtr = Builder.CreateBitCast(Ptr, NewPtrType);
1326 unsigned VectorWidth = VecType->getNumElements();
1328 if (isa<ConstantVector>(Mask)) {
1329 for (unsigned Idx = 0; Idx < VectorWidth; ++Idx) {
1330 if (cast<ConstantVector>(Mask)->getOperand(Idx)->isNullValue())
1332 Value *OneElt = Builder.CreateExtractElement(Src, Builder.getInt32(Idx));
1334 Builder.CreateInBoundsGEP(EltTy, FirstEltPtr, Builder.getInt32(Idx));
1335 Builder.CreateAlignedStore(OneElt, Gep, AlignVal);
1337 CI->eraseFromParent();
1341 for (unsigned Idx = 0; Idx < VectorWidth; ++Idx) {
1343 // Fill the "else" block, created in the previous iteration
1345 // %mask_1 = extractelement <16 x i1> %mask, i32 Idx
1346 // %to_store = icmp eq i1 %mask_1, true
1347 // br i1 %to_store, label %cond.store, label %else
1349 Value *Predicate = Builder.CreateExtractElement(Mask, Builder.getInt32(Idx));
1350 Value *Cmp = Builder.CreateICmp(ICmpInst::ICMP_EQ, Predicate,
1351 ConstantInt::get(Predicate->getType(), 1));
1353 // Create "cond" block
1355 // %OneElt = extractelement <16 x i32> %Src, i32 Idx
1356 // %EltAddr = getelementptr i32* %1, i32 0
1357 // %store i32 %OneElt, i32* %EltAddr
1359 BasicBlock *CondBlock =
1360 IfBlock->splitBasicBlock(InsertPt->getIterator(), "cond.store");
1361 Builder.SetInsertPoint(InsertPt);
1363 Value *OneElt = Builder.CreateExtractElement(Src, Builder.getInt32(Idx));
1365 Builder.CreateInBoundsGEP(EltTy, FirstEltPtr, Builder.getInt32(Idx));
1366 Builder.CreateAlignedStore(OneElt, Gep, AlignVal);
1368 // Create "else" block, fill it in the next iteration
1369 BasicBlock *NewIfBlock =
1370 CondBlock->splitBasicBlock(InsertPt->getIterator(), "else");
1371 Builder.SetInsertPoint(InsertPt);
1372 Instruction *OldBr = IfBlock->getTerminator();
1373 BranchInst::Create(CondBlock, NewIfBlock, Cmp, OldBr);
1374 OldBr->eraseFromParent();
1375 IfBlock = NewIfBlock;
1377 CI->eraseFromParent();
1380 // Translate a masked gather intrinsic like
1381 // <16 x i32 > @llvm.masked.gather.v16i32( <16 x i32*> %Ptrs, i32 4,
1382 // <16 x i1> %Mask, <16 x i32> %Src)
1383 // to a chain of basic blocks, with loading element one-by-one if
1384 // the appropriate mask bit is set
1386 // % Ptrs = getelementptr i32, i32* %base, <16 x i64> %ind
1387 // % Mask0 = extractelement <16 x i1> %Mask, i32 0
1388 // % ToLoad0 = icmp eq i1 % Mask0, true
1389 // br i1 % ToLoad0, label %cond.load, label %else
1392 // % Ptr0 = extractelement <16 x i32*> %Ptrs, i32 0
1393 // % Load0 = load i32, i32* % Ptr0, align 4
1394 // % Res0 = insertelement <16 x i32> undef, i32 % Load0, i32 0
1398 // %res.phi.else = phi <16 x i32>[% Res0, %cond.load], [undef, % 0]
1399 // % Mask1 = extractelement <16 x i1> %Mask, i32 1
1400 // % ToLoad1 = icmp eq i1 % Mask1, true
1401 // br i1 % ToLoad1, label %cond.load1, label %else2
1404 // % Ptr1 = extractelement <16 x i32*> %Ptrs, i32 1
1405 // % Load1 = load i32, i32* % Ptr1, align 4
1406 // % Res1 = insertelement <16 x i32> %res.phi.else, i32 % Load1, i32 1
1409 // % Result = select <16 x i1> %Mask, <16 x i32> %res.phi.select, <16 x i32> %Src
1410 // ret <16 x i32> %Result
1411 static void ScalarizeMaskedGather(CallInst *CI) {
1412 Value *Ptrs = CI->getArgOperand(0);
1413 Value *Alignment = CI->getArgOperand(1);
1414 Value *Mask = CI->getArgOperand(2);
1415 Value *Src0 = CI->getArgOperand(3);
1417 VectorType *VecType = dyn_cast<VectorType>(CI->getType());
1419 assert(VecType && "Unexpected return type of masked load intrinsic");
1421 IRBuilder<> Builder(CI->getContext());
1422 Instruction *InsertPt = CI;
1423 BasicBlock *IfBlock = CI->getParent();
1424 BasicBlock *CondBlock = nullptr;
1425 BasicBlock *PrevIfBlock = CI->getParent();
1426 Builder.SetInsertPoint(InsertPt);
1427 unsigned AlignVal = cast<ConstantInt>(Alignment)->getZExtValue();
1429 Builder.SetCurrentDebugLocation(CI->getDebugLoc());
1431 Value *UndefVal = UndefValue::get(VecType);
1433 // The result vector
1434 Value *VResult = UndefVal;
1435 unsigned VectorWidth = VecType->getNumElements();
1437 // Shorten the way if the mask is a vector of constants.
1438 bool IsConstMask = isa<ConstantVector>(Mask);
1441 for (unsigned Idx = 0; Idx < VectorWidth; ++Idx) {
1442 if (cast<ConstantVector>(Mask)->getOperand(Idx)->isNullValue())
1444 Value *Ptr = Builder.CreateExtractElement(Ptrs, Builder.getInt32(Idx),
1445 "Ptr" + Twine(Idx));
1446 LoadInst *Load = Builder.CreateAlignedLoad(Ptr, AlignVal,
1447 "Load" + Twine(Idx));
1448 VResult = Builder.CreateInsertElement(VResult, Load,
1449 Builder.getInt32(Idx),
1450 "Res" + Twine(Idx));
1452 Value *NewI = Builder.CreateSelect(Mask, VResult, Src0);
1453 CI->replaceAllUsesWith(NewI);
1454 CI->eraseFromParent();
1458 PHINode *Phi = nullptr;
1459 Value *PrevPhi = UndefVal;
1461 for (unsigned Idx = 0; Idx < VectorWidth; ++Idx) {
1463 // Fill the "else" block, created in the previous iteration
1465 // %Mask1 = extractelement <16 x i1> %Mask, i32 1
1466 // %ToLoad1 = icmp eq i1 %Mask1, true
1467 // br i1 %ToLoad1, label %cond.load, label %else
1470 Phi = Builder.CreatePHI(VecType, 2, "res.phi.else");
1471 Phi->addIncoming(VResult, CondBlock);
1472 Phi->addIncoming(PrevPhi, PrevIfBlock);
1477 Value *Predicate = Builder.CreateExtractElement(Mask,
1478 Builder.getInt32(Idx),
1479 "Mask" + Twine(Idx));
1480 Value *Cmp = Builder.CreateICmp(ICmpInst::ICMP_EQ, Predicate,
1481 ConstantInt::get(Predicate->getType(), 1),
1482 "ToLoad" + Twine(Idx));
1484 // Create "cond" block
1486 // %EltAddr = getelementptr i32* %1, i32 0
1487 // %Elt = load i32* %EltAddr
1488 // VResult = insertelement <16 x i32> VResult, i32 %Elt, i32 Idx
1490 CondBlock = IfBlock->splitBasicBlock(InsertPt, "cond.load");
1491 Builder.SetInsertPoint(InsertPt);
1493 Value *Ptr = Builder.CreateExtractElement(Ptrs, Builder.getInt32(Idx),
1494 "Ptr" + Twine(Idx));
1495 LoadInst *Load = Builder.CreateAlignedLoad(Ptr, AlignVal,
1496 "Load" + Twine(Idx));
1497 VResult = Builder.CreateInsertElement(VResult, Load, Builder.getInt32(Idx),
1498 "Res" + Twine(Idx));
1500 // Create "else" block, fill it in the next iteration
1501 BasicBlock *NewIfBlock = CondBlock->splitBasicBlock(InsertPt, "else");
1502 Builder.SetInsertPoint(InsertPt);
1503 Instruction *OldBr = IfBlock->getTerminator();
1504 BranchInst::Create(CondBlock, NewIfBlock, Cmp, OldBr);
1505 OldBr->eraseFromParent();
1506 PrevIfBlock = IfBlock;
1507 IfBlock = NewIfBlock;
1510 Phi = Builder.CreatePHI(VecType, 2, "res.phi.select");
1511 Phi->addIncoming(VResult, CondBlock);
1512 Phi->addIncoming(PrevPhi, PrevIfBlock);
1513 Value *NewI = Builder.CreateSelect(Mask, Phi, Src0);
1514 CI->replaceAllUsesWith(NewI);
1515 CI->eraseFromParent();
1518 // Translate a masked scatter intrinsic, like
1519 // void @llvm.masked.scatter.v16i32(<16 x i32> %Src, <16 x i32*>* %Ptrs, i32 4,
1521 // to a chain of basic blocks, that stores element one-by-one if
1522 // the appropriate mask bit is set.
1524 // % Ptrs = getelementptr i32, i32* %ptr, <16 x i64> %ind
1525 // % Mask0 = extractelement <16 x i1> % Mask, i32 0
1526 // % ToStore0 = icmp eq i1 % Mask0, true
1527 // br i1 %ToStore0, label %cond.store, label %else
1530 // % Elt0 = extractelement <16 x i32> %Src, i32 0
1531 // % Ptr0 = extractelement <16 x i32*> %Ptrs, i32 0
1532 // store i32 %Elt0, i32* % Ptr0, align 4
1536 // % Mask1 = extractelement <16 x i1> % Mask, i32 1
1537 // % ToStore1 = icmp eq i1 % Mask1, true
1538 // br i1 % ToStore1, label %cond.store1, label %else2
1541 // % Elt1 = extractelement <16 x i32> %Src, i32 1
1542 // % Ptr1 = extractelement <16 x i32*> %Ptrs, i32 1
1543 // store i32 % Elt1, i32* % Ptr1, align 4
1546 static void ScalarizeMaskedScatter(CallInst *CI) {
1547 Value *Src = CI->getArgOperand(0);
1548 Value *Ptrs = CI->getArgOperand(1);
1549 Value *Alignment = CI->getArgOperand(2);
1550 Value *Mask = CI->getArgOperand(3);
1552 assert(isa<VectorType>(Src->getType()) &&
1553 "Unexpected data type in masked scatter intrinsic");
1554 assert(isa<VectorType>(Ptrs->getType()) &&
1555 isa<PointerType>(Ptrs->getType()->getVectorElementType()) &&
1556 "Vector of pointers is expected in masked scatter intrinsic");
1558 IRBuilder<> Builder(CI->getContext());
1559 Instruction *InsertPt = CI;
1560 BasicBlock *IfBlock = CI->getParent();
1561 Builder.SetInsertPoint(InsertPt);
1562 Builder.SetCurrentDebugLocation(CI->getDebugLoc());
1564 unsigned AlignVal = cast<ConstantInt>(Alignment)->getZExtValue();
1565 unsigned VectorWidth = Src->getType()->getVectorNumElements();
1567 // Shorten the way if the mask is a vector of constants.
1568 bool IsConstMask = isa<ConstantVector>(Mask);
1571 for (unsigned Idx = 0; Idx < VectorWidth; ++Idx) {
1572 if (cast<ConstantVector>(Mask)->getOperand(Idx)->isNullValue())
1574 Value *OneElt = Builder.CreateExtractElement(Src, Builder.getInt32(Idx),
1575 "Elt" + Twine(Idx));
1576 Value *Ptr = Builder.CreateExtractElement(Ptrs, Builder.getInt32(Idx),
1577 "Ptr" + Twine(Idx));
1578 Builder.CreateAlignedStore(OneElt, Ptr, AlignVal);
1580 CI->eraseFromParent();
1583 for (unsigned Idx = 0; Idx < VectorWidth; ++Idx) {
1584 // Fill the "else" block, created in the previous iteration
1586 // % Mask1 = extractelement <16 x i1> % Mask, i32 Idx
1587 // % ToStore = icmp eq i1 % Mask1, true
1588 // br i1 % ToStore, label %cond.store, label %else
1590 Value *Predicate = Builder.CreateExtractElement(Mask,
1591 Builder.getInt32(Idx),
1592 "Mask" + Twine(Idx));
1594 Builder.CreateICmp(ICmpInst::ICMP_EQ, Predicate,
1595 ConstantInt::get(Predicate->getType(), 1),
1596 "ToStore" + Twine(Idx));
1598 // Create "cond" block
1600 // % Elt1 = extractelement <16 x i32> %Src, i32 1
1601 // % Ptr1 = extractelement <16 x i32*> %Ptrs, i32 1
1602 // %store i32 % Elt1, i32* % Ptr1
1604 BasicBlock *CondBlock = IfBlock->splitBasicBlock(InsertPt, "cond.store");
1605 Builder.SetInsertPoint(InsertPt);
1607 Value *OneElt = Builder.CreateExtractElement(Src, Builder.getInt32(Idx),
1608 "Elt" + Twine(Idx));
1609 Value *Ptr = Builder.CreateExtractElement(Ptrs, Builder.getInt32(Idx),
1610 "Ptr" + Twine(Idx));
1611 Builder.CreateAlignedStore(OneElt, Ptr, AlignVal);
1613 // Create "else" block, fill it in the next iteration
1614 BasicBlock *NewIfBlock = CondBlock->splitBasicBlock(InsertPt, "else");
1615 Builder.SetInsertPoint(InsertPt);
1616 Instruction *OldBr = IfBlock->getTerminator();
1617 BranchInst::Create(CondBlock, NewIfBlock, Cmp, OldBr);
1618 OldBr->eraseFromParent();
1619 IfBlock = NewIfBlock;
1621 CI->eraseFromParent();
1624 /// If counting leading or trailing zeros is an expensive operation and a zero
1625 /// input is defined, add a check for zero to avoid calling the intrinsic.
1627 /// We want to transform:
1628 /// %z = call i64 @llvm.cttz.i64(i64 %A, i1 false)
1632 /// %cmpz = icmp eq i64 %A, 0
1633 /// br i1 %cmpz, label %cond.end, label %cond.false
1635 /// %z = call i64 @llvm.cttz.i64(i64 %A, i1 true)
1636 /// br label %cond.end
1638 /// %ctz = phi i64 [ 64, %entry ], [ %z, %cond.false ]
1640 /// If the transform is performed, return true and set ModifiedDT to true.
1641 static bool despeculateCountZeros(IntrinsicInst *CountZeros,
1642 const TargetLowering *TLI,
1643 const DataLayout *DL,
1648 // If a zero input is undefined, it doesn't make sense to despeculate that.
1649 if (match(CountZeros->getOperand(1), m_One()))
1652 // If it's cheap to speculate, there's nothing to do.
1653 auto IntrinsicID = CountZeros->getIntrinsicID();
1654 if ((IntrinsicID == Intrinsic::cttz && TLI->isCheapToSpeculateCttz()) ||
1655 (IntrinsicID == Intrinsic::ctlz && TLI->isCheapToSpeculateCtlz()))
1658 // Only handle legal scalar cases. Anything else requires too much work.
1659 Type *Ty = CountZeros->getType();
1660 unsigned SizeInBits = Ty->getPrimitiveSizeInBits();
1661 if (Ty->isVectorTy() || SizeInBits > DL->getLargestLegalIntTypeSize())
1664 // The intrinsic will be sunk behind a compare against zero and branch.
1665 BasicBlock *StartBlock = CountZeros->getParent();
1666 BasicBlock *CallBlock = StartBlock->splitBasicBlock(CountZeros, "cond.false");
1668 // Create another block after the count zero intrinsic. A PHI will be added
1669 // in this block to select the result of the intrinsic or the bit-width
1670 // constant if the input to the intrinsic is zero.
1671 BasicBlock::iterator SplitPt = ++(BasicBlock::iterator(CountZeros));
1672 BasicBlock *EndBlock = CallBlock->splitBasicBlock(SplitPt, "cond.end");
1674 // Set up a builder to create a compare, conditional branch, and PHI.
1675 IRBuilder<> Builder(CountZeros->getContext());
1676 Builder.SetInsertPoint(StartBlock->getTerminator());
1677 Builder.SetCurrentDebugLocation(CountZeros->getDebugLoc());
1679 // Replace the unconditional branch that was created by the first split with
1680 // a compare against zero and a conditional branch.
1681 Value *Zero = Constant::getNullValue(Ty);
1682 Value *Cmp = Builder.CreateICmpEQ(CountZeros->getOperand(0), Zero, "cmpz");
1683 Builder.CreateCondBr(Cmp, EndBlock, CallBlock);
1684 StartBlock->getTerminator()->eraseFromParent();
1686 // Create a PHI in the end block to select either the output of the intrinsic
1687 // or the bit width of the operand.
1688 Builder.SetInsertPoint(&EndBlock->front());
1689 PHINode *PN = Builder.CreatePHI(Ty, 2, "ctz");
1690 CountZeros->replaceAllUsesWith(PN);
1691 Value *BitWidth = Builder.getInt(APInt(SizeInBits, SizeInBits));
1692 PN->addIncoming(BitWidth, StartBlock);
1693 PN->addIncoming(CountZeros, CallBlock);
1695 // We are explicitly handling the zero case, so we can set the intrinsic's
1696 // undefined zero argument to 'true'. This will also prevent reprocessing the
1697 // intrinsic; we only despeculate when a zero input is defined.
1698 CountZeros->setArgOperand(1, Builder.getTrue());
1703 bool CodeGenPrepare::optimizeCallInst(CallInst *CI, bool& ModifiedDT) {
1704 BasicBlock *BB = CI->getParent();
1706 // Lower inline assembly if we can.
1707 // If we found an inline asm expession, and if the target knows how to
1708 // lower it to normal LLVM code, do so now.
1709 if (TLI && isa<InlineAsm>(CI->getCalledValue())) {
1710 if (TLI->ExpandInlineAsm(CI)) {
1711 // Avoid invalidating the iterator.
1712 CurInstIterator = BB->begin();
1713 // Avoid processing instructions out of order, which could cause
1714 // reuse before a value is defined.
1718 // Sink address computing for memory operands into the block.
1719 if (optimizeInlineAsmInst(CI))
1723 // Align the pointer arguments to this call if the target thinks it's a good
1725 unsigned MinSize, PrefAlign;
1726 if (TLI && TLI->shouldAlignPointerArgs(CI, MinSize, PrefAlign)) {
1727 for (auto &Arg : CI->arg_operands()) {
1728 // We want to align both objects whose address is used directly and
1729 // objects whose address is used in casts and GEPs, though it only makes
1730 // sense for GEPs if the offset is a multiple of the desired alignment and
1731 // if size - offset meets the size threshold.
1732 if (!Arg->getType()->isPointerTy())
1734 APInt Offset(DL->getPointerSizeInBits(
1735 cast<PointerType>(Arg->getType())->getAddressSpace()),
1737 Value *Val = Arg->stripAndAccumulateInBoundsConstantOffsets(*DL, Offset);
1738 uint64_t Offset2 = Offset.getLimitedValue();
1739 if ((Offset2 & (PrefAlign-1)) != 0)
1742 if ((AI = dyn_cast<AllocaInst>(Val)) && AI->getAlignment() < PrefAlign &&
1743 DL->getTypeAllocSize(AI->getAllocatedType()) >= MinSize + Offset2)
1744 AI->setAlignment(PrefAlign);
1745 // Global variables can only be aligned if they are defined in this
1746 // object (i.e. they are uniquely initialized in this object), and
1747 // over-aligning global variables that have an explicit section is
1750 if ((GV = dyn_cast<GlobalVariable>(Val)) && GV->hasUniqueInitializer() &&
1751 !GV->hasSection() && GV->getAlignment() < PrefAlign &&
1752 DL->getTypeAllocSize(GV->getType()->getElementType()) >=
1754 GV->setAlignment(PrefAlign);
1756 // If this is a memcpy (or similar) then we may be able to improve the
1758 if (MemIntrinsic *MI = dyn_cast<MemIntrinsic>(CI)) {
1759 unsigned Align = getKnownAlignment(MI->getDest(), *DL);
1760 if (MemTransferInst *MTI = dyn_cast<MemTransferInst>(MI))
1761 Align = std::min(Align, getKnownAlignment(MTI->getSource(), *DL));
1762 if (Align > MI->getAlignment())
1763 MI->setAlignment(ConstantInt::get(MI->getAlignmentType(), Align));
1767 IntrinsicInst *II = dyn_cast<IntrinsicInst>(CI);
1769 switch (II->getIntrinsicID()) {
1771 case Intrinsic::objectsize: {
1772 // Lower all uses of llvm.objectsize.*
1773 bool Min = (cast<ConstantInt>(II->getArgOperand(1))->getZExtValue() == 1);
1774 Type *ReturnTy = CI->getType();
1775 Constant *RetVal = ConstantInt::get(ReturnTy, Min ? 0 : -1ULL);
1777 // Substituting this can cause recursive simplifications, which can
1778 // invalidate our iterator. Use a WeakVH to hold onto it in case this
1780 WeakVH IterHandle(&*CurInstIterator);
1782 replaceAndRecursivelySimplify(CI, RetVal,
1785 // If the iterator instruction was recursively deleted, start over at the
1786 // start of the block.
1787 if (IterHandle != CurInstIterator.getNodePtrUnchecked()) {
1788 CurInstIterator = BB->begin();
1793 case Intrinsic::masked_load: {
1794 // Scalarize unsupported vector masked load
1795 if (!TTI->isLegalMaskedLoad(CI->getType())) {
1796 ScalarizeMaskedLoad(CI);
1802 case Intrinsic::masked_store: {
1803 if (!TTI->isLegalMaskedStore(CI->getArgOperand(0)->getType())) {
1804 ScalarizeMaskedStore(CI);
1810 case Intrinsic::masked_gather: {
1811 if (!TTI->isLegalMaskedGather(CI->getType())) {
1812 ScalarizeMaskedGather(CI);
1818 case Intrinsic::masked_scatter: {
1819 if (!TTI->isLegalMaskedScatter(CI->getArgOperand(0)->getType())) {
1820 ScalarizeMaskedScatter(CI);
1826 case Intrinsic::aarch64_stlxr:
1827 case Intrinsic::aarch64_stxr: {
1828 ZExtInst *ExtVal = dyn_cast<ZExtInst>(CI->getArgOperand(0));
1829 if (!ExtVal || !ExtVal->hasOneUse() ||
1830 ExtVal->getParent() == CI->getParent())
1832 // Sink a zext feeding stlxr/stxr before it, so it can be folded into it.
1833 ExtVal->moveBefore(CI);
1834 // Mark this instruction as "inserted by CGP", so that other
1835 // optimizations don't touch it.
1836 InsertedInsts.insert(ExtVal);
1839 case Intrinsic::invariant_group_barrier:
1840 II->replaceAllUsesWith(II->getArgOperand(0));
1841 II->eraseFromParent();
1844 case Intrinsic::cttz:
1845 case Intrinsic::ctlz:
1846 // If counting zeros is expensive, try to avoid it.
1847 return despeculateCountZeros(II, TLI, DL, ModifiedDT);
1851 // Unknown address space.
1852 // TODO: Target hook to pick which address space the intrinsic cares
1854 unsigned AddrSpace = ~0u;
1855 SmallVector<Value*, 2> PtrOps;
1857 if (TLI->GetAddrModeArguments(II, PtrOps, AccessTy, AddrSpace))
1858 while (!PtrOps.empty())
1859 if (optimizeMemoryInst(II, PtrOps.pop_back_val(), AccessTy, AddrSpace))
1864 // From here on out we're working with named functions.
1865 if (!CI->getCalledFunction()) return false;
1867 // Lower all default uses of _chk calls. This is very similar
1868 // to what InstCombineCalls does, but here we are only lowering calls
1869 // to fortified library functions (e.g. __memcpy_chk) that have the default
1870 // "don't know" as the objectsize. Anything else should be left alone.
1871 FortifiedLibCallSimplifier Simplifier(TLInfo, true);
1872 if (Value *V = Simplifier.optimizeCall(CI)) {
1873 CI->replaceAllUsesWith(V);
1874 CI->eraseFromParent();
1880 /// Look for opportunities to duplicate return instructions to the predecessor
1881 /// to enable tail call optimizations. The case it is currently looking for is:
1884 /// %tmp0 = tail call i32 @f0()
1885 /// br label %return
1887 /// %tmp1 = tail call i32 @f1()
1888 /// br label %return
1890 /// %tmp2 = tail call i32 @f2()
1891 /// br label %return
1893 /// %retval = phi i32 [ %tmp0, %bb0 ], [ %tmp1, %bb1 ], [ %tmp2, %bb2 ]
1901 /// %tmp0 = tail call i32 @f0()
1904 /// %tmp1 = tail call i32 @f1()
1907 /// %tmp2 = tail call i32 @f2()
1910 bool CodeGenPrepare::dupRetToEnableTailCallOpts(BasicBlock *BB) {
1914 ReturnInst *RI = dyn_cast<ReturnInst>(BB->getTerminator());
1918 PHINode *PN = nullptr;
1919 BitCastInst *BCI = nullptr;
1920 Value *V = RI->getReturnValue();
1922 BCI = dyn_cast<BitCastInst>(V);
1924 V = BCI->getOperand(0);
1926 PN = dyn_cast<PHINode>(V);
1931 if (PN && PN->getParent() != BB)
1934 // It's not safe to eliminate the sign / zero extension of the return value.
1935 // See llvm::isInTailCallPosition().
1936 const Function *F = BB->getParent();
1937 AttributeSet CallerAttrs = F->getAttributes();
1938 if (CallerAttrs.hasAttribute(AttributeSet::ReturnIndex, Attribute::ZExt) ||
1939 CallerAttrs.hasAttribute(AttributeSet::ReturnIndex, Attribute::SExt))
1942 // Make sure there are no instructions between the PHI and return, or that the
1943 // return is the first instruction in the block.
1945 BasicBlock::iterator BI = BB->begin();
1946 do { ++BI; } while (isa<DbgInfoIntrinsic>(BI));
1948 // Also skip over the bitcast.
1953 BasicBlock::iterator BI = BB->begin();
1954 while (isa<DbgInfoIntrinsic>(BI)) ++BI;
1959 /// Only dup the ReturnInst if the CallInst is likely to be emitted as a tail
1961 SmallVector<CallInst*, 4> TailCalls;
1963 for (unsigned I = 0, E = PN->getNumIncomingValues(); I != E; ++I) {
1964 CallInst *CI = dyn_cast<CallInst>(PN->getIncomingValue(I));
1965 // Make sure the phi value is indeed produced by the tail call.
1966 if (CI && CI->hasOneUse() && CI->getParent() == PN->getIncomingBlock(I) &&
1967 TLI->mayBeEmittedAsTailCall(CI))
1968 TailCalls.push_back(CI);
1971 SmallPtrSet<BasicBlock*, 4> VisitedBBs;
1972 for (pred_iterator PI = pred_begin(BB), PE = pred_end(BB); PI != PE; ++PI) {
1973 if (!VisitedBBs.insert(*PI).second)
1976 BasicBlock::InstListType &InstList = (*PI)->getInstList();
1977 BasicBlock::InstListType::reverse_iterator RI = InstList.rbegin();
1978 BasicBlock::InstListType::reverse_iterator RE = InstList.rend();
1979 do { ++RI; } while (RI != RE && isa<DbgInfoIntrinsic>(&*RI));
1983 CallInst *CI = dyn_cast<CallInst>(&*RI);
1984 if (CI && CI->use_empty() && TLI->mayBeEmittedAsTailCall(CI))
1985 TailCalls.push_back(CI);
1989 bool Changed = false;
1990 for (unsigned i = 0, e = TailCalls.size(); i != e; ++i) {
1991 CallInst *CI = TailCalls[i];
1994 // Conservatively require the attributes of the call to match those of the
1995 // return. Ignore noalias because it doesn't affect the call sequence.
1996 AttributeSet CalleeAttrs = CS.getAttributes();
1997 if (AttrBuilder(CalleeAttrs, AttributeSet::ReturnIndex).
1998 removeAttribute(Attribute::NoAlias) !=
1999 AttrBuilder(CalleeAttrs, AttributeSet::ReturnIndex).
2000 removeAttribute(Attribute::NoAlias))
2003 // Make sure the call instruction is followed by an unconditional branch to
2004 // the return block.
2005 BasicBlock *CallBB = CI->getParent();
2006 BranchInst *BI = dyn_cast<BranchInst>(CallBB->getTerminator());
2007 if (!BI || !BI->isUnconditional() || BI->getSuccessor(0) != BB)
2010 // Duplicate the return into CallBB.
2011 (void)FoldReturnIntoUncondBranch(RI, BB, CallBB);
2012 ModifiedDT = Changed = true;
2016 // If we eliminated all predecessors of the block, delete the block now.
2017 if (Changed && !BB->hasAddressTaken() && pred_begin(BB) == pred_end(BB))
2018 BB->eraseFromParent();
2023 //===----------------------------------------------------------------------===//
2024 // Memory Optimization
2025 //===----------------------------------------------------------------------===//
2029 /// This is an extended version of TargetLowering::AddrMode
2030 /// which holds actual Value*'s for register values.
2031 struct ExtAddrMode : public TargetLowering::AddrMode {
2034 ExtAddrMode() : BaseReg(nullptr), ScaledReg(nullptr) {}
2035 void print(raw_ostream &OS) const;
2038 bool operator==(const ExtAddrMode& O) const {
2039 return (BaseReg == O.BaseReg) && (ScaledReg == O.ScaledReg) &&
2040 (BaseGV == O.BaseGV) && (BaseOffs == O.BaseOffs) &&
2041 (HasBaseReg == O.HasBaseReg) && (Scale == O.Scale);
2046 static inline raw_ostream &operator<<(raw_ostream &OS, const ExtAddrMode &AM) {
2052 void ExtAddrMode::print(raw_ostream &OS) const {
2053 bool NeedPlus = false;
2056 OS << (NeedPlus ? " + " : "")
2058 BaseGV->printAsOperand(OS, /*PrintType=*/false);
2063 OS << (NeedPlus ? " + " : "")
2069 OS << (NeedPlus ? " + " : "")
2071 BaseReg->printAsOperand(OS, /*PrintType=*/false);
2075 OS << (NeedPlus ? " + " : "")
2077 ScaledReg->printAsOperand(OS, /*PrintType=*/false);
2083 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
2084 void ExtAddrMode::dump() const {
2090 /// \brief This class provides transaction based operation on the IR.
2091 /// Every change made through this class is recorded in the internal state and
2092 /// can be undone (rollback) until commit is called.
2093 class TypePromotionTransaction {
2095 /// \brief This represents the common interface of the individual transaction.
2096 /// Each class implements the logic for doing one specific modification on
2097 /// the IR via the TypePromotionTransaction.
2098 class TypePromotionAction {
2100 /// The Instruction modified.
2104 /// \brief Constructor of the action.
2105 /// The constructor performs the related action on the IR.
2106 TypePromotionAction(Instruction *Inst) : Inst(Inst) {}
2108 virtual ~TypePromotionAction() {}
2110 /// \brief Undo the modification done by this action.
2111 /// When this method is called, the IR must be in the same state as it was
2112 /// before this action was applied.
2113 /// \pre Undoing the action works if and only if the IR is in the exact same
2114 /// state as it was directly after this action was applied.
2115 virtual void undo() = 0;
2117 /// \brief Advocate every change made by this action.
2118 /// When the results on the IR of the action are to be kept, it is important
2119 /// to call this function, otherwise hidden information may be kept forever.
2120 virtual void commit() {
2121 // Nothing to be done, this action is not doing anything.
2125 /// \brief Utility to remember the position of an instruction.
2126 class InsertionHandler {
2127 /// Position of an instruction.
2128 /// Either an instruction:
2129 /// - Is the first in a basic block: BB is used.
2130 /// - Has a previous instructon: PrevInst is used.
2132 Instruction *PrevInst;
2135 /// Remember whether or not the instruction had a previous instruction.
2136 bool HasPrevInstruction;
2139 /// \brief Record the position of \p Inst.
2140 InsertionHandler(Instruction *Inst) {
2141 BasicBlock::iterator It = Inst->getIterator();
2142 HasPrevInstruction = (It != (Inst->getParent()->begin()));
2143 if (HasPrevInstruction)
2144 Point.PrevInst = &*--It;
2146 Point.BB = Inst->getParent();
2149 /// \brief Insert \p Inst at the recorded position.
2150 void insert(Instruction *Inst) {
2151 if (HasPrevInstruction) {
2152 if (Inst->getParent())
2153 Inst->removeFromParent();
2154 Inst->insertAfter(Point.PrevInst);
2156 Instruction *Position = &*Point.BB->getFirstInsertionPt();
2157 if (Inst->getParent())
2158 Inst->moveBefore(Position);
2160 Inst->insertBefore(Position);
2165 /// \brief Move an instruction before another.
2166 class InstructionMoveBefore : public TypePromotionAction {
2167 /// Original position of the instruction.
2168 InsertionHandler Position;
2171 /// \brief Move \p Inst before \p Before.
2172 InstructionMoveBefore(Instruction *Inst, Instruction *Before)
2173 : TypePromotionAction(Inst), Position(Inst) {
2174 DEBUG(dbgs() << "Do: move: " << *Inst << "\nbefore: " << *Before << "\n");
2175 Inst->moveBefore(Before);
2178 /// \brief Move the instruction back to its original position.
2179 void undo() override {
2180 DEBUG(dbgs() << "Undo: moveBefore: " << *Inst << "\n");
2181 Position.insert(Inst);
2185 /// \brief Set the operand of an instruction with a new value.
2186 class OperandSetter : public TypePromotionAction {
2187 /// Original operand of the instruction.
2189 /// Index of the modified instruction.
2193 /// \brief Set \p Idx operand of \p Inst with \p NewVal.
2194 OperandSetter(Instruction *Inst, unsigned Idx, Value *NewVal)
2195 : TypePromotionAction(Inst), Idx(Idx) {
2196 DEBUG(dbgs() << "Do: setOperand: " << Idx << "\n"
2197 << "for:" << *Inst << "\n"
2198 << "with:" << *NewVal << "\n");
2199 Origin = Inst->getOperand(Idx);
2200 Inst->setOperand(Idx, NewVal);
2203 /// \brief Restore the original value of the instruction.
2204 void undo() override {
2205 DEBUG(dbgs() << "Undo: setOperand:" << Idx << "\n"
2206 << "for: " << *Inst << "\n"
2207 << "with: " << *Origin << "\n");
2208 Inst->setOperand(Idx, Origin);
2212 /// \brief Hide the operands of an instruction.
2213 /// Do as if this instruction was not using any of its operands.
2214 class OperandsHider : public TypePromotionAction {
2215 /// The list of original operands.
2216 SmallVector<Value *, 4> OriginalValues;
2219 /// \brief Remove \p Inst from the uses of the operands of \p Inst.
2220 OperandsHider(Instruction *Inst) : TypePromotionAction(Inst) {
2221 DEBUG(dbgs() << "Do: OperandsHider: " << *Inst << "\n");
2222 unsigned NumOpnds = Inst->getNumOperands();
2223 OriginalValues.reserve(NumOpnds);
2224 for (unsigned It = 0; It < NumOpnds; ++It) {
2225 // Save the current operand.
2226 Value *Val = Inst->getOperand(It);
2227 OriginalValues.push_back(Val);
2229 // We could use OperandSetter here, but that would imply an overhead
2230 // that we are not willing to pay.
2231 Inst->setOperand(It, UndefValue::get(Val->getType()));
2235 /// \brief Restore the original list of uses.
2236 void undo() override {
2237 DEBUG(dbgs() << "Undo: OperandsHider: " << *Inst << "\n");
2238 for (unsigned It = 0, EndIt = OriginalValues.size(); It != EndIt; ++It)
2239 Inst->setOperand(It, OriginalValues[It]);
2243 /// \brief Build a truncate instruction.
2244 class TruncBuilder : public TypePromotionAction {
2247 /// \brief Build a truncate instruction of \p Opnd producing a \p Ty
2249 /// trunc Opnd to Ty.
2250 TruncBuilder(Instruction *Opnd, Type *Ty) : TypePromotionAction(Opnd) {
2251 IRBuilder<> Builder(Opnd);
2252 Val = Builder.CreateTrunc(Opnd, Ty, "promoted");
2253 DEBUG(dbgs() << "Do: TruncBuilder: " << *Val << "\n");
2256 /// \brief Get the built value.
2257 Value *getBuiltValue() { return Val; }
2259 /// \brief Remove the built instruction.
2260 void undo() override {
2261 DEBUG(dbgs() << "Undo: TruncBuilder: " << *Val << "\n");
2262 if (Instruction *IVal = dyn_cast<Instruction>(Val))
2263 IVal->eraseFromParent();
2267 /// \brief Build a sign extension instruction.
2268 class SExtBuilder : public TypePromotionAction {
2271 /// \brief Build a sign extension instruction of \p Opnd producing a \p Ty
2273 /// sext Opnd to Ty.
2274 SExtBuilder(Instruction *InsertPt, Value *Opnd, Type *Ty)
2275 : TypePromotionAction(InsertPt) {
2276 IRBuilder<> Builder(InsertPt);
2277 Val = Builder.CreateSExt(Opnd, Ty, "promoted");
2278 DEBUG(dbgs() << "Do: SExtBuilder: " << *Val << "\n");
2281 /// \brief Get the built value.
2282 Value *getBuiltValue() { return Val; }
2284 /// \brief Remove the built instruction.
2285 void undo() override {
2286 DEBUG(dbgs() << "Undo: SExtBuilder: " << *Val << "\n");
2287 if (Instruction *IVal = dyn_cast<Instruction>(Val))
2288 IVal->eraseFromParent();
2292 /// \brief Build a zero extension instruction.
2293 class ZExtBuilder : public TypePromotionAction {
2296 /// \brief Build a zero extension instruction of \p Opnd producing a \p Ty
2298 /// zext Opnd to Ty.
2299 ZExtBuilder(Instruction *InsertPt, Value *Opnd, Type *Ty)
2300 : TypePromotionAction(InsertPt) {
2301 IRBuilder<> Builder(InsertPt);
2302 Val = Builder.CreateZExt(Opnd, Ty, "promoted");
2303 DEBUG(dbgs() << "Do: ZExtBuilder: " << *Val << "\n");
2306 /// \brief Get the built value.
2307 Value *getBuiltValue() { return Val; }
2309 /// \brief Remove the built instruction.
2310 void undo() override {
2311 DEBUG(dbgs() << "Undo: ZExtBuilder: " << *Val << "\n");
2312 if (Instruction *IVal = dyn_cast<Instruction>(Val))
2313 IVal->eraseFromParent();
2317 /// \brief Mutate an instruction to another type.
2318 class TypeMutator : public TypePromotionAction {
2319 /// Record the original type.
2323 /// \brief Mutate the type of \p Inst into \p NewTy.
2324 TypeMutator(Instruction *Inst, Type *NewTy)
2325 : TypePromotionAction(Inst), OrigTy(Inst->getType()) {
2326 DEBUG(dbgs() << "Do: MutateType: " << *Inst << " with " << *NewTy
2328 Inst->mutateType(NewTy);
2331 /// \brief Mutate the instruction back to its original type.
2332 void undo() override {
2333 DEBUG(dbgs() << "Undo: MutateType: " << *Inst << " with " << *OrigTy
2335 Inst->mutateType(OrigTy);
2339 /// \brief Replace the uses of an instruction by another instruction.
2340 class UsesReplacer : public TypePromotionAction {
2341 /// Helper structure to keep track of the replaced uses.
2342 struct InstructionAndIdx {
2343 /// The instruction using the instruction.
2345 /// The index where this instruction is used for Inst.
2347 InstructionAndIdx(Instruction *Inst, unsigned Idx)
2348 : Inst(Inst), Idx(Idx) {}
2351 /// Keep track of the original uses (pair Instruction, Index).
2352 SmallVector<InstructionAndIdx, 4> OriginalUses;
2353 typedef SmallVectorImpl<InstructionAndIdx>::iterator use_iterator;
2356 /// \brief Replace all the use of \p Inst by \p New.
2357 UsesReplacer(Instruction *Inst, Value *New) : TypePromotionAction(Inst) {
2358 DEBUG(dbgs() << "Do: UsersReplacer: " << *Inst << " with " << *New
2360 // Record the original uses.
2361 for (Use &U : Inst->uses()) {
2362 Instruction *UserI = cast<Instruction>(U.getUser());
2363 OriginalUses.push_back(InstructionAndIdx(UserI, U.getOperandNo()));
2365 // Now, we can replace the uses.
2366 Inst->replaceAllUsesWith(New);
2369 /// \brief Reassign the original uses of Inst to Inst.
2370 void undo() override {
2371 DEBUG(dbgs() << "Undo: UsersReplacer: " << *Inst << "\n");
2372 for (use_iterator UseIt = OriginalUses.begin(),
2373 EndIt = OriginalUses.end();
2374 UseIt != EndIt; ++UseIt) {
2375 UseIt->Inst->setOperand(UseIt->Idx, Inst);
2380 /// \brief Remove an instruction from the IR.
2381 class InstructionRemover : public TypePromotionAction {
2382 /// Original position of the instruction.
2383 InsertionHandler Inserter;
2384 /// Helper structure to hide all the link to the instruction. In other
2385 /// words, this helps to do as if the instruction was removed.
2386 OperandsHider Hider;
2387 /// Keep track of the uses replaced, if any.
2388 UsesReplacer *Replacer;
2391 /// \brief Remove all reference of \p Inst and optinally replace all its
2393 /// \pre If !Inst->use_empty(), then New != nullptr
2394 InstructionRemover(Instruction *Inst, Value *New = nullptr)
2395 : TypePromotionAction(Inst), Inserter(Inst), Hider(Inst),
2398 Replacer = new UsesReplacer(Inst, New);
2399 DEBUG(dbgs() << "Do: InstructionRemover: " << *Inst << "\n");
2400 Inst->removeFromParent();
2403 ~InstructionRemover() override { delete Replacer; }
2405 /// \brief Really remove the instruction.
2406 void commit() override { delete Inst; }
2408 /// \brief Resurrect the instruction and reassign it to the proper uses if
2409 /// new value was provided when build this action.
2410 void undo() override {
2411 DEBUG(dbgs() << "Undo: InstructionRemover: " << *Inst << "\n");
2412 Inserter.insert(Inst);
2420 /// Restoration point.
2421 /// The restoration point is a pointer to an action instead of an iterator
2422 /// because the iterator may be invalidated but not the pointer.
2423 typedef const TypePromotionAction *ConstRestorationPt;
2424 /// Advocate every changes made in that transaction.
2426 /// Undo all the changes made after the given point.
2427 void rollback(ConstRestorationPt Point);
2428 /// Get the current restoration point.
2429 ConstRestorationPt getRestorationPoint() const;
2431 /// \name API for IR modification with state keeping to support rollback.
2433 /// Same as Instruction::setOperand.
2434 void setOperand(Instruction *Inst, unsigned Idx, Value *NewVal);
2435 /// Same as Instruction::eraseFromParent.
2436 void eraseInstruction(Instruction *Inst, Value *NewVal = nullptr);
2437 /// Same as Value::replaceAllUsesWith.
2438 void replaceAllUsesWith(Instruction *Inst, Value *New);
2439 /// Same as Value::mutateType.
2440 void mutateType(Instruction *Inst, Type *NewTy);
2441 /// Same as IRBuilder::createTrunc.
2442 Value *createTrunc(Instruction *Opnd, Type *Ty);
2443 /// Same as IRBuilder::createSExt.
2444 Value *createSExt(Instruction *Inst, Value *Opnd, Type *Ty);
2445 /// Same as IRBuilder::createZExt.
2446 Value *createZExt(Instruction *Inst, Value *Opnd, Type *Ty);
2447 /// Same as Instruction::moveBefore.
2448 void moveBefore(Instruction *Inst, Instruction *Before);
2452 /// The ordered list of actions made so far.
2453 SmallVector<std::unique_ptr<TypePromotionAction>, 16> Actions;
2454 typedef SmallVectorImpl<std::unique_ptr<TypePromotionAction>>::iterator CommitPt;
2457 void TypePromotionTransaction::setOperand(Instruction *Inst, unsigned Idx,
2460 make_unique<TypePromotionTransaction::OperandSetter>(Inst, Idx, NewVal));
2463 void TypePromotionTransaction::eraseInstruction(Instruction *Inst,
2466 make_unique<TypePromotionTransaction::InstructionRemover>(Inst, NewVal));
2469 void TypePromotionTransaction::replaceAllUsesWith(Instruction *Inst,
2471 Actions.push_back(make_unique<TypePromotionTransaction::UsesReplacer>(Inst, New));
2474 void TypePromotionTransaction::mutateType(Instruction *Inst, Type *NewTy) {
2475 Actions.push_back(make_unique<TypePromotionTransaction::TypeMutator>(Inst, NewTy));
2478 Value *TypePromotionTransaction::createTrunc(Instruction *Opnd,
2480 std::unique_ptr<TruncBuilder> Ptr(new TruncBuilder(Opnd, Ty));
2481 Value *Val = Ptr->getBuiltValue();
2482 Actions.push_back(std::move(Ptr));
2486 Value *TypePromotionTransaction::createSExt(Instruction *Inst,
2487 Value *Opnd, Type *Ty) {
2488 std::unique_ptr<SExtBuilder> Ptr(new SExtBuilder(Inst, Opnd, Ty));
2489 Value *Val = Ptr->getBuiltValue();
2490 Actions.push_back(std::move(Ptr));
2494 Value *TypePromotionTransaction::createZExt(Instruction *Inst,
2495 Value *Opnd, Type *Ty) {
2496 std::unique_ptr<ZExtBuilder> Ptr(new ZExtBuilder(Inst, Opnd, Ty));
2497 Value *Val = Ptr->getBuiltValue();
2498 Actions.push_back(std::move(Ptr));
2502 void TypePromotionTransaction::moveBefore(Instruction *Inst,
2503 Instruction *Before) {
2505 make_unique<TypePromotionTransaction::InstructionMoveBefore>(Inst, Before));
2508 TypePromotionTransaction::ConstRestorationPt
2509 TypePromotionTransaction::getRestorationPoint() const {
2510 return !Actions.empty() ? Actions.back().get() : nullptr;
2513 void TypePromotionTransaction::commit() {
2514 for (CommitPt It = Actions.begin(), EndIt = Actions.end(); It != EndIt;
2520 void TypePromotionTransaction::rollback(
2521 TypePromotionTransaction::ConstRestorationPt Point) {
2522 while (!Actions.empty() && Point != Actions.back().get()) {
2523 std::unique_ptr<TypePromotionAction> Curr = Actions.pop_back_val();
2528 /// \brief A helper class for matching addressing modes.
2530 /// This encapsulates the logic for matching the target-legal addressing modes.
2531 class AddressingModeMatcher {
2532 SmallVectorImpl<Instruction*> &AddrModeInsts;
2533 const TargetMachine &TM;
2534 const TargetLowering &TLI;
2535 const DataLayout &DL;
2537 /// AccessTy/MemoryInst - This is the type for the access (e.g. double) and
2538 /// the memory instruction that we're computing this address for.
2541 Instruction *MemoryInst;
2543 /// This is the addressing mode that we're building up. This is
2544 /// part of the return value of this addressing mode matching stuff.
2545 ExtAddrMode &AddrMode;
2547 /// The instructions inserted by other CodeGenPrepare optimizations.
2548 const SetOfInstrs &InsertedInsts;
2549 /// A map from the instructions to their type before promotion.
2550 InstrToOrigTy &PromotedInsts;
2551 /// The ongoing transaction where every action should be registered.
2552 TypePromotionTransaction &TPT;
2554 /// This is set to true when we should not do profitability checks.
2555 /// When true, IsProfitableToFoldIntoAddressingMode always returns true.
2556 bool IgnoreProfitability;
2558 AddressingModeMatcher(SmallVectorImpl<Instruction *> &AMI,
2559 const TargetMachine &TM, Type *AT, unsigned AS,
2560 Instruction *MI, ExtAddrMode &AM,
2561 const SetOfInstrs &InsertedInsts,
2562 InstrToOrigTy &PromotedInsts,
2563 TypePromotionTransaction &TPT)
2564 : AddrModeInsts(AMI), TM(TM),
2565 TLI(*TM.getSubtargetImpl(*MI->getParent()->getParent())
2566 ->getTargetLowering()),
2567 DL(MI->getModule()->getDataLayout()), AccessTy(AT), AddrSpace(AS),
2568 MemoryInst(MI), AddrMode(AM), InsertedInsts(InsertedInsts),
2569 PromotedInsts(PromotedInsts), TPT(TPT) {
2570 IgnoreProfitability = false;
2574 /// Find the maximal addressing mode that a load/store of V can fold,
2575 /// give an access type of AccessTy. This returns a list of involved
2576 /// instructions in AddrModeInsts.
2577 /// \p InsertedInsts The instructions inserted by other CodeGenPrepare
2579 /// \p PromotedInsts maps the instructions to their type before promotion.
2580 /// \p The ongoing transaction where every action should be registered.
2581 static ExtAddrMode Match(Value *V, Type *AccessTy, unsigned AS,
2582 Instruction *MemoryInst,
2583 SmallVectorImpl<Instruction*> &AddrModeInsts,
2584 const TargetMachine &TM,
2585 const SetOfInstrs &InsertedInsts,
2586 InstrToOrigTy &PromotedInsts,
2587 TypePromotionTransaction &TPT) {
2590 bool Success = AddressingModeMatcher(AddrModeInsts, TM, AccessTy, AS,
2591 MemoryInst, Result, InsertedInsts,
2592 PromotedInsts, TPT).matchAddr(V, 0);
2593 (void)Success; assert(Success && "Couldn't select *anything*?");
2597 bool matchScaledValue(Value *ScaleReg, int64_t Scale, unsigned Depth);
2598 bool matchAddr(Value *V, unsigned Depth);
2599 bool matchOperationAddr(User *Operation, unsigned Opcode, unsigned Depth,
2600 bool *MovedAway = nullptr);
2601 bool isProfitableToFoldIntoAddressingMode(Instruction *I,
2602 ExtAddrMode &AMBefore,
2603 ExtAddrMode &AMAfter);
2604 bool valueAlreadyLiveAtInst(Value *Val, Value *KnownLive1, Value *KnownLive2);
2605 bool isPromotionProfitable(unsigned NewCost, unsigned OldCost,
2606 Value *PromotedOperand) const;
2609 /// Try adding ScaleReg*Scale to the current addressing mode.
2610 /// Return true and update AddrMode if this addr mode is legal for the target,
2612 bool AddressingModeMatcher::matchScaledValue(Value *ScaleReg, int64_t Scale,
2614 // If Scale is 1, then this is the same as adding ScaleReg to the addressing
2615 // mode. Just process that directly.
2617 return matchAddr(ScaleReg, Depth);
2619 // If the scale is 0, it takes nothing to add this.
2623 // If we already have a scale of this value, we can add to it, otherwise, we
2624 // need an available scale field.
2625 if (AddrMode.Scale != 0 && AddrMode.ScaledReg != ScaleReg)
2628 ExtAddrMode TestAddrMode = AddrMode;
2630 // Add scale to turn X*4+X*3 -> X*7. This could also do things like
2631 // [A+B + A*7] -> [B+A*8].
2632 TestAddrMode.Scale += Scale;
2633 TestAddrMode.ScaledReg = ScaleReg;
2635 // If the new address isn't legal, bail out.
2636 if (!TLI.isLegalAddressingMode(DL, TestAddrMode, AccessTy, AddrSpace))
2639 // It was legal, so commit it.
2640 AddrMode = TestAddrMode;
2642 // Okay, we decided that we can add ScaleReg+Scale to AddrMode. Check now
2643 // to see if ScaleReg is actually X+C. If so, we can turn this into adding
2644 // X*Scale + C*Scale to addr mode.
2645 ConstantInt *CI = nullptr; Value *AddLHS = nullptr;
2646 if (isa<Instruction>(ScaleReg) && // not a constant expr.
2647 match(ScaleReg, m_Add(m_Value(AddLHS), m_ConstantInt(CI)))) {
2648 TestAddrMode.ScaledReg = AddLHS;
2649 TestAddrMode.BaseOffs += CI->getSExtValue()*TestAddrMode.Scale;
2651 // If this addressing mode is legal, commit it and remember that we folded
2652 // this instruction.
2653 if (TLI.isLegalAddressingMode(DL, TestAddrMode, AccessTy, AddrSpace)) {
2654 AddrModeInsts.push_back(cast<Instruction>(ScaleReg));
2655 AddrMode = TestAddrMode;
2660 // Otherwise, not (x+c)*scale, just return what we have.
2664 /// This is a little filter, which returns true if an addressing computation
2665 /// involving I might be folded into a load/store accessing it.
2666 /// This doesn't need to be perfect, but needs to accept at least
2667 /// the set of instructions that MatchOperationAddr can.
2668 static bool MightBeFoldableInst(Instruction *I) {
2669 switch (I->getOpcode()) {
2670 case Instruction::BitCast:
2671 case Instruction::AddrSpaceCast:
2672 // Don't touch identity bitcasts.
2673 if (I->getType() == I->getOperand(0)->getType())
2675 return I->getType()->isPointerTy() || I->getType()->isIntegerTy();
2676 case Instruction::PtrToInt:
2677 // PtrToInt is always a noop, as we know that the int type is pointer sized.
2679 case Instruction::IntToPtr:
2680 // We know the input is intptr_t, so this is foldable.
2682 case Instruction::Add:
2684 case Instruction::Mul:
2685 case Instruction::Shl:
2686 // Can only handle X*C and X << C.
2687 return isa<ConstantInt>(I->getOperand(1));
2688 case Instruction::GetElementPtr:
2695 /// \brief Check whether or not \p Val is a legal instruction for \p TLI.
2696 /// \note \p Val is assumed to be the product of some type promotion.
2697 /// Therefore if \p Val has an undefined state in \p TLI, this is assumed
2698 /// to be legal, as the non-promoted value would have had the same state.
2699 static bool isPromotedInstructionLegal(const TargetLowering &TLI,
2700 const DataLayout &DL, Value *Val) {
2701 Instruction *PromotedInst = dyn_cast<Instruction>(Val);
2704 int ISDOpcode = TLI.InstructionOpcodeToISD(PromotedInst->getOpcode());
2705 // If the ISDOpcode is undefined, it was undefined before the promotion.
2708 // Otherwise, check if the promoted instruction is legal or not.
2709 return TLI.isOperationLegalOrCustom(
2710 ISDOpcode, TLI.getValueType(DL, PromotedInst->getType()));
2713 /// \brief Hepler class to perform type promotion.
2714 class TypePromotionHelper {
2715 /// \brief Utility function to check whether or not a sign or zero extension
2716 /// of \p Inst with \p ConsideredExtType can be moved through \p Inst by
2717 /// either using the operands of \p Inst or promoting \p Inst.
2718 /// The type of the extension is defined by \p IsSExt.
2719 /// In other words, check if:
2720 /// ext (Ty Inst opnd1 opnd2 ... opndN) to ConsideredExtType.
2721 /// #1 Promotion applies:
2722 /// ConsideredExtType Inst (ext opnd1 to ConsideredExtType, ...).
2723 /// #2 Operand reuses:
2724 /// ext opnd1 to ConsideredExtType.
2725 /// \p PromotedInsts maps the instructions to their type before promotion.
2726 static bool canGetThrough(const Instruction *Inst, Type *ConsideredExtType,
2727 const InstrToOrigTy &PromotedInsts, bool IsSExt);
2729 /// \brief Utility function to determine if \p OpIdx should be promoted when
2730 /// promoting \p Inst.
2731 static bool shouldExtOperand(const Instruction *Inst, int OpIdx) {
2732 return !(isa<SelectInst>(Inst) && OpIdx == 0);
2735 /// \brief Utility function to promote the operand of \p Ext when this
2736 /// operand is a promotable trunc or sext or zext.
2737 /// \p PromotedInsts maps the instructions to their type before promotion.
2738 /// \p CreatedInstsCost[out] contains the cost of all instructions
2739 /// created to promote the operand of Ext.
2740 /// Newly added extensions are inserted in \p Exts.
2741 /// Newly added truncates are inserted in \p Truncs.
2742 /// Should never be called directly.
2743 /// \return The promoted value which is used instead of Ext.
2744 static Value *promoteOperandForTruncAndAnyExt(
2745 Instruction *Ext, TypePromotionTransaction &TPT,
2746 InstrToOrigTy &PromotedInsts, unsigned &CreatedInstsCost,
2747 SmallVectorImpl<Instruction *> *Exts,
2748 SmallVectorImpl<Instruction *> *Truncs, const TargetLowering &TLI);
2750 /// \brief Utility function to promote the operand of \p Ext when this
2751 /// operand is promotable and is not a supported trunc or sext.
2752 /// \p PromotedInsts maps the instructions to their type before promotion.
2753 /// \p CreatedInstsCost[out] contains the cost of all the instructions
2754 /// created to promote the operand of Ext.
2755 /// Newly added extensions are inserted in \p Exts.
2756 /// Newly added truncates are inserted in \p Truncs.
2757 /// Should never be called directly.
2758 /// \return The promoted value which is used instead of Ext.
2759 static Value *promoteOperandForOther(Instruction *Ext,
2760 TypePromotionTransaction &TPT,
2761 InstrToOrigTy &PromotedInsts,
2762 unsigned &CreatedInstsCost,
2763 SmallVectorImpl<Instruction *> *Exts,
2764 SmallVectorImpl<Instruction *> *Truncs,
2765 const TargetLowering &TLI, bool IsSExt);
2767 /// \see promoteOperandForOther.
2768 static Value *signExtendOperandForOther(
2769 Instruction *Ext, TypePromotionTransaction &TPT,
2770 InstrToOrigTy &PromotedInsts, unsigned &CreatedInstsCost,
2771 SmallVectorImpl<Instruction *> *Exts,
2772 SmallVectorImpl<Instruction *> *Truncs, const TargetLowering &TLI) {
2773 return promoteOperandForOther(Ext, TPT, PromotedInsts, CreatedInstsCost,
2774 Exts, Truncs, TLI, true);
2777 /// \see promoteOperandForOther.
2778 static Value *zeroExtendOperandForOther(
2779 Instruction *Ext, TypePromotionTransaction &TPT,
2780 InstrToOrigTy &PromotedInsts, unsigned &CreatedInstsCost,
2781 SmallVectorImpl<Instruction *> *Exts,
2782 SmallVectorImpl<Instruction *> *Truncs, const TargetLowering &TLI) {
2783 return promoteOperandForOther(Ext, TPT, PromotedInsts, CreatedInstsCost,
2784 Exts, Truncs, TLI, false);
2788 /// Type for the utility function that promotes the operand of Ext.
2789 typedef Value *(*Action)(Instruction *Ext, TypePromotionTransaction &TPT,
2790 InstrToOrigTy &PromotedInsts,
2791 unsigned &CreatedInstsCost,
2792 SmallVectorImpl<Instruction *> *Exts,
2793 SmallVectorImpl<Instruction *> *Truncs,
2794 const TargetLowering &TLI);
2795 /// \brief Given a sign/zero extend instruction \p Ext, return the approriate
2796 /// action to promote the operand of \p Ext instead of using Ext.
2797 /// \return NULL if no promotable action is possible with the current
2799 /// \p InsertedInsts keeps track of all the instructions inserted by the
2800 /// other CodeGenPrepare optimizations. This information is important
2801 /// because we do not want to promote these instructions as CodeGenPrepare
2802 /// will reinsert them later. Thus creating an infinite loop: create/remove.
2803 /// \p PromotedInsts maps the instructions to their type before promotion.
2804 static Action getAction(Instruction *Ext, const SetOfInstrs &InsertedInsts,
2805 const TargetLowering &TLI,
2806 const InstrToOrigTy &PromotedInsts);
2809 bool TypePromotionHelper::canGetThrough(const Instruction *Inst,
2810 Type *ConsideredExtType,
2811 const InstrToOrigTy &PromotedInsts,
2813 // The promotion helper does not know how to deal with vector types yet.
2814 // To be able to fix that, we would need to fix the places where we
2815 // statically extend, e.g., constants and such.
2816 if (Inst->getType()->isVectorTy())
2819 // We can always get through zext.
2820 if (isa<ZExtInst>(Inst))
2823 // sext(sext) is ok too.
2824 if (IsSExt && isa<SExtInst>(Inst))
2827 // We can get through binary operator, if it is legal. In other words, the
2828 // binary operator must have a nuw or nsw flag.
2829 const BinaryOperator *BinOp = dyn_cast<BinaryOperator>(Inst);
2830 if (BinOp && isa<OverflowingBinaryOperator>(BinOp) &&
2831 ((!IsSExt && BinOp->hasNoUnsignedWrap()) ||
2832 (IsSExt && BinOp->hasNoSignedWrap())))
2835 // Check if we can do the following simplification.
2836 // ext(trunc(opnd)) --> ext(opnd)
2837 if (!isa<TruncInst>(Inst))
2840 Value *OpndVal = Inst->getOperand(0);
2841 // Check if we can use this operand in the extension.
2842 // If the type is larger than the result type of the extension, we cannot.
2843 if (!OpndVal->getType()->isIntegerTy() ||
2844 OpndVal->getType()->getIntegerBitWidth() >
2845 ConsideredExtType->getIntegerBitWidth())
2848 // If the operand of the truncate is not an instruction, we will not have
2849 // any information on the dropped bits.
2850 // (Actually we could for constant but it is not worth the extra logic).
2851 Instruction *Opnd = dyn_cast<Instruction>(OpndVal);
2855 // Check if the source of the type is narrow enough.
2856 // I.e., check that trunc just drops extended bits of the same kind of
2858 // #1 get the type of the operand and check the kind of the extended bits.
2859 const Type *OpndType;
2860 InstrToOrigTy::const_iterator It = PromotedInsts.find(Opnd);
2861 if (It != PromotedInsts.end() && It->second.getInt() == IsSExt)
2862 OpndType = It->second.getPointer();
2863 else if ((IsSExt && isa<SExtInst>(Opnd)) || (!IsSExt && isa<ZExtInst>(Opnd)))
2864 OpndType = Opnd->getOperand(0)->getType();
2868 // #2 check that the truncate just drops extended bits.
2869 return Inst->getType()->getIntegerBitWidth() >=
2870 OpndType->getIntegerBitWidth();
2873 TypePromotionHelper::Action TypePromotionHelper::getAction(
2874 Instruction *Ext, const SetOfInstrs &InsertedInsts,
2875 const TargetLowering &TLI, const InstrToOrigTy &PromotedInsts) {
2876 assert((isa<SExtInst>(Ext) || isa<ZExtInst>(Ext)) &&
2877 "Unexpected instruction type");
2878 Instruction *ExtOpnd = dyn_cast<Instruction>(Ext->getOperand(0));
2879 Type *ExtTy = Ext->getType();
2880 bool IsSExt = isa<SExtInst>(Ext);
2881 // If the operand of the extension is not an instruction, we cannot
2883 // If it, check we can get through.
2884 if (!ExtOpnd || !canGetThrough(ExtOpnd, ExtTy, PromotedInsts, IsSExt))
2887 // Do not promote if the operand has been added by codegenprepare.
2888 // Otherwise, it means we are undoing an optimization that is likely to be
2889 // redone, thus causing potential infinite loop.
2890 if (isa<TruncInst>(ExtOpnd) && InsertedInsts.count(ExtOpnd))
2893 // SExt or Trunc instructions.
2894 // Return the related handler.
2895 if (isa<SExtInst>(ExtOpnd) || isa<TruncInst>(ExtOpnd) ||
2896 isa<ZExtInst>(ExtOpnd))
2897 return promoteOperandForTruncAndAnyExt;
2899 // Regular instruction.
2900 // Abort early if we will have to insert non-free instructions.
2901 if (!ExtOpnd->hasOneUse() && !TLI.isTruncateFree(ExtTy, ExtOpnd->getType()))
2903 return IsSExt ? signExtendOperandForOther : zeroExtendOperandForOther;
2906 Value *TypePromotionHelper::promoteOperandForTruncAndAnyExt(
2907 llvm::Instruction *SExt, TypePromotionTransaction &TPT,
2908 InstrToOrigTy &PromotedInsts, unsigned &CreatedInstsCost,
2909 SmallVectorImpl<Instruction *> *Exts,
2910 SmallVectorImpl<Instruction *> *Truncs, const TargetLowering &TLI) {
2911 // By construction, the operand of SExt is an instruction. Otherwise we cannot
2912 // get through it and this method should not be called.
2913 Instruction *SExtOpnd = cast<Instruction>(SExt->getOperand(0));
2914 Value *ExtVal = SExt;
2915 bool HasMergedNonFreeExt = false;
2916 if (isa<ZExtInst>(SExtOpnd)) {
2917 // Replace s|zext(zext(opnd))
2919 HasMergedNonFreeExt = !TLI.isExtFree(SExtOpnd);
2921 TPT.createZExt(SExt, SExtOpnd->getOperand(0), SExt->getType());
2922 TPT.replaceAllUsesWith(SExt, ZExt);
2923 TPT.eraseInstruction(SExt);
2926 // Replace z|sext(trunc(opnd)) or sext(sext(opnd))
2928 TPT.setOperand(SExt, 0, SExtOpnd->getOperand(0));
2930 CreatedInstsCost = 0;
2932 // Remove dead code.
2933 if (SExtOpnd->use_empty())
2934 TPT.eraseInstruction(SExtOpnd);
2936 // Check if the extension is still needed.
2937 Instruction *ExtInst = dyn_cast<Instruction>(ExtVal);
2938 if (!ExtInst || ExtInst->getType() != ExtInst->getOperand(0)->getType()) {
2941 Exts->push_back(ExtInst);
2942 CreatedInstsCost = !TLI.isExtFree(ExtInst) && !HasMergedNonFreeExt;
2947 // At this point we have: ext ty opnd to ty.
2948 // Reassign the uses of ExtInst to the opnd and remove ExtInst.
2949 Value *NextVal = ExtInst->getOperand(0);
2950 TPT.eraseInstruction(ExtInst, NextVal);
2954 Value *TypePromotionHelper::promoteOperandForOther(
2955 Instruction *Ext, TypePromotionTransaction &TPT,
2956 InstrToOrigTy &PromotedInsts, unsigned &CreatedInstsCost,
2957 SmallVectorImpl<Instruction *> *Exts,
2958 SmallVectorImpl<Instruction *> *Truncs, const TargetLowering &TLI,
2960 // By construction, the operand of Ext is an instruction. Otherwise we cannot
2961 // get through it and this method should not be called.
2962 Instruction *ExtOpnd = cast<Instruction>(Ext->getOperand(0));
2963 CreatedInstsCost = 0;
2964 if (!ExtOpnd->hasOneUse()) {
2965 // ExtOpnd will be promoted.
2966 // All its uses, but Ext, will need to use a truncated value of the
2967 // promoted version.
2968 // Create the truncate now.
2969 Value *Trunc = TPT.createTrunc(Ext, ExtOpnd->getType());
2970 if (Instruction *ITrunc = dyn_cast<Instruction>(Trunc)) {
2971 ITrunc->removeFromParent();
2972 // Insert it just after the definition.
2973 ITrunc->insertAfter(ExtOpnd);
2975 Truncs->push_back(ITrunc);
2978 TPT.replaceAllUsesWith(ExtOpnd, Trunc);
2979 // Restore the operand of Ext (which has been replaced by the previous call
2980 // to replaceAllUsesWith) to avoid creating a cycle trunc <-> sext.
2981 TPT.setOperand(Ext, 0, ExtOpnd);
2984 // Get through the Instruction:
2985 // 1. Update its type.
2986 // 2. Replace the uses of Ext by Inst.
2987 // 3. Extend each operand that needs to be extended.
2989 // Remember the original type of the instruction before promotion.
2990 // This is useful to know that the high bits are sign extended bits.
2991 PromotedInsts.insert(std::pair<Instruction *, TypeIsSExt>(
2992 ExtOpnd, TypeIsSExt(ExtOpnd->getType(), IsSExt)));
2994 TPT.mutateType(ExtOpnd, Ext->getType());
2996 TPT.replaceAllUsesWith(Ext, ExtOpnd);
2998 Instruction *ExtForOpnd = Ext;
3000 DEBUG(dbgs() << "Propagate Ext to operands\n");
3001 for (int OpIdx = 0, EndOpIdx = ExtOpnd->getNumOperands(); OpIdx != EndOpIdx;
3003 DEBUG(dbgs() << "Operand:\n" << *(ExtOpnd->getOperand(OpIdx)) << '\n');
3004 if (ExtOpnd->getOperand(OpIdx)->getType() == Ext->getType() ||
3005 !shouldExtOperand(ExtOpnd, OpIdx)) {
3006 DEBUG(dbgs() << "No need to propagate\n");
3009 // Check if we can statically extend the operand.
3010 Value *Opnd = ExtOpnd->getOperand(OpIdx);
3011 if (const ConstantInt *Cst = dyn_cast<ConstantInt>(Opnd)) {
3012 DEBUG(dbgs() << "Statically extend\n");
3013 unsigned BitWidth = Ext->getType()->getIntegerBitWidth();
3014 APInt CstVal = IsSExt ? Cst->getValue().sext(BitWidth)
3015 : Cst->getValue().zext(BitWidth);
3016 TPT.setOperand(ExtOpnd, OpIdx, ConstantInt::get(Ext->getType(), CstVal));
3019 // UndefValue are typed, so we have to statically sign extend them.
3020 if (isa<UndefValue>(Opnd)) {
3021 DEBUG(dbgs() << "Statically extend\n");
3022 TPT.setOperand(ExtOpnd, OpIdx, UndefValue::get(Ext->getType()));
3026 // Otherwise we have to explicity sign extend the operand.
3027 // Check if Ext was reused to extend an operand.
3029 // If yes, create a new one.
3030 DEBUG(dbgs() << "More operands to ext\n");
3031 Value *ValForExtOpnd = IsSExt ? TPT.createSExt(Ext, Opnd, Ext->getType())
3032 : TPT.createZExt(Ext, Opnd, Ext->getType());
3033 if (!isa<Instruction>(ValForExtOpnd)) {
3034 TPT.setOperand(ExtOpnd, OpIdx, ValForExtOpnd);
3037 ExtForOpnd = cast<Instruction>(ValForExtOpnd);
3040 Exts->push_back(ExtForOpnd);
3041 TPT.setOperand(ExtForOpnd, 0, Opnd);
3043 // Move the sign extension before the insertion point.
3044 TPT.moveBefore(ExtForOpnd, ExtOpnd);
3045 TPT.setOperand(ExtOpnd, OpIdx, ExtForOpnd);
3046 CreatedInstsCost += !TLI.isExtFree(ExtForOpnd);
3047 // If more sext are required, new instructions will have to be created.
3048 ExtForOpnd = nullptr;
3050 if (ExtForOpnd == Ext) {
3051 DEBUG(dbgs() << "Extension is useless now\n");
3052 TPT.eraseInstruction(Ext);
3057 /// Check whether or not promoting an instruction to a wider type is profitable.
3058 /// \p NewCost gives the cost of extension instructions created by the
3060 /// \p OldCost gives the cost of extension instructions before the promotion
3061 /// plus the number of instructions that have been
3062 /// matched in the addressing mode the promotion.
3063 /// \p PromotedOperand is the value that has been promoted.
3064 /// \return True if the promotion is profitable, false otherwise.
3065 bool AddressingModeMatcher::isPromotionProfitable(
3066 unsigned NewCost, unsigned OldCost, Value *PromotedOperand) const {
3067 DEBUG(dbgs() << "OldCost: " << OldCost << "\tNewCost: " << NewCost << '\n');
3068 // The cost of the new extensions is greater than the cost of the
3069 // old extension plus what we folded.
3070 // This is not profitable.
3071 if (NewCost > OldCost)
3073 if (NewCost < OldCost)
3075 // The promotion is neutral but it may help folding the sign extension in
3076 // loads for instance.
3077 // Check that we did not create an illegal instruction.
3078 return isPromotedInstructionLegal(TLI, DL, PromotedOperand);
3081 /// Given an instruction or constant expr, see if we can fold the operation
3082 /// into the addressing mode. If so, update the addressing mode and return
3083 /// true, otherwise return false without modifying AddrMode.
3084 /// If \p MovedAway is not NULL, it contains the information of whether or
3085 /// not AddrInst has to be folded into the addressing mode on success.
3086 /// If \p MovedAway == true, \p AddrInst will not be part of the addressing
3087 /// because it has been moved away.
3088 /// Thus AddrInst must not be added in the matched instructions.
3089 /// This state can happen when AddrInst is a sext, since it may be moved away.
3090 /// Therefore, AddrInst may not be valid when MovedAway is true and it must
3091 /// not be referenced anymore.
3092 bool AddressingModeMatcher::matchOperationAddr(User *AddrInst, unsigned Opcode,
3095 // Avoid exponential behavior on extremely deep expression trees.
3096 if (Depth >= 5) return false;
3098 // By default, all matched instructions stay in place.
3103 case Instruction::PtrToInt:
3104 // PtrToInt is always a noop, as we know that the int type is pointer sized.
3105 return matchAddr(AddrInst->getOperand(0), Depth);
3106 case Instruction::IntToPtr: {
3107 auto AS = AddrInst->getType()->getPointerAddressSpace();
3108 auto PtrTy = MVT::getIntegerVT(DL.getPointerSizeInBits(AS));
3109 // This inttoptr is a no-op if the integer type is pointer sized.
3110 if (TLI.getValueType(DL, AddrInst->getOperand(0)->getType()) == PtrTy)
3111 return matchAddr(AddrInst->getOperand(0), Depth);
3114 case Instruction::BitCast:
3115 // BitCast is always a noop, and we can handle it as long as it is
3116 // int->int or pointer->pointer (we don't want int<->fp or something).
3117 if ((AddrInst->getOperand(0)->getType()->isPointerTy() ||
3118 AddrInst->getOperand(0)->getType()->isIntegerTy()) &&
3119 // Don't touch identity bitcasts. These were probably put here by LSR,
3120 // and we don't want to mess around with them. Assume it knows what it
3122 AddrInst->getOperand(0)->getType() != AddrInst->getType())
3123 return matchAddr(AddrInst->getOperand(0), Depth);
3125 case Instruction::AddrSpaceCast: {
3127 = AddrInst->getOperand(0)->getType()->getPointerAddressSpace();
3128 unsigned DestAS = AddrInst->getType()->getPointerAddressSpace();
3129 if (TLI.isNoopAddrSpaceCast(SrcAS, DestAS))
3130 return matchAddr(AddrInst->getOperand(0), Depth);
3133 case Instruction::Add: {
3134 // Check to see if we can merge in the RHS then the LHS. If so, we win.
3135 ExtAddrMode BackupAddrMode = AddrMode;
3136 unsigned OldSize = AddrModeInsts.size();
3137 // Start a transaction at this point.
3138 // The LHS may match but not the RHS.
3139 // Therefore, we need a higher level restoration point to undo partially
3140 // matched operation.
3141 TypePromotionTransaction::ConstRestorationPt LastKnownGood =
3142 TPT.getRestorationPoint();
3144 if (matchAddr(AddrInst->getOperand(1), Depth+1) &&
3145 matchAddr(AddrInst->getOperand(0), Depth+1))
3148 // Restore the old addr mode info.
3149 AddrMode = BackupAddrMode;
3150 AddrModeInsts.resize(OldSize);
3151 TPT.rollback(LastKnownGood);
3153 // Otherwise this was over-aggressive. Try merging in the LHS then the RHS.
3154 if (matchAddr(AddrInst->getOperand(0), Depth+1) &&
3155 matchAddr(AddrInst->getOperand(1), Depth+1))
3158 // Otherwise we definitely can't merge the ADD in.
3159 AddrMode = BackupAddrMode;
3160 AddrModeInsts.resize(OldSize);
3161 TPT.rollback(LastKnownGood);
3164 //case Instruction::Or:
3165 // TODO: We can handle "Or Val, Imm" iff this OR is equivalent to an ADD.
3167 case Instruction::Mul:
3168 case Instruction::Shl: {
3169 // Can only handle X*C and X << C.
3170 ConstantInt *RHS = dyn_cast<ConstantInt>(AddrInst->getOperand(1));
3173 int64_t Scale = RHS->getSExtValue();
3174 if (Opcode == Instruction::Shl)
3175 Scale = 1LL << Scale;
3177 return matchScaledValue(AddrInst->getOperand(0), Scale, Depth);
3179 case Instruction::GetElementPtr: {
3180 // Scan the GEP. We check it if it contains constant offsets and at most
3181 // one variable offset.
3182 int VariableOperand = -1;
3183 unsigned VariableScale = 0;
3185 int64_t ConstantOffset = 0;
3186 gep_type_iterator GTI = gep_type_begin(AddrInst);
3187 for (unsigned i = 1, e = AddrInst->getNumOperands(); i != e; ++i, ++GTI) {
3188 if (StructType *STy = dyn_cast<StructType>(*GTI)) {
3189 const StructLayout *SL = DL.getStructLayout(STy);
3191 cast<ConstantInt>(AddrInst->getOperand(i))->getZExtValue();
3192 ConstantOffset += SL->getElementOffset(Idx);
3194 uint64_t TypeSize = DL.getTypeAllocSize(GTI.getIndexedType());
3195 if (ConstantInt *CI = dyn_cast<ConstantInt>(AddrInst->getOperand(i))) {
3196 ConstantOffset += CI->getSExtValue()*TypeSize;
3197 } else if (TypeSize) { // Scales of zero don't do anything.
3198 // We only allow one variable index at the moment.
3199 if (VariableOperand != -1)
3202 // Remember the variable index.
3203 VariableOperand = i;
3204 VariableScale = TypeSize;
3209 // A common case is for the GEP to only do a constant offset. In this case,
3210 // just add it to the disp field and check validity.
3211 if (VariableOperand == -1) {
3212 AddrMode.BaseOffs += ConstantOffset;
3213 if (ConstantOffset == 0 ||
3214 TLI.isLegalAddressingMode(DL, AddrMode, AccessTy, AddrSpace)) {
3215 // Check to see if we can fold the base pointer in too.
3216 if (matchAddr(AddrInst->getOperand(0), Depth+1))
3219 AddrMode.BaseOffs -= ConstantOffset;
3223 // Save the valid addressing mode in case we can't match.
3224 ExtAddrMode BackupAddrMode = AddrMode;
3225 unsigned OldSize = AddrModeInsts.size();
3227 // See if the scale and offset amount is valid for this target.
3228 AddrMode.BaseOffs += ConstantOffset;
3230 // Match the base operand of the GEP.
3231 if (!matchAddr(AddrInst->getOperand(0), Depth+1)) {
3232 // If it couldn't be matched, just stuff the value in a register.
3233 if (AddrMode.HasBaseReg) {
3234 AddrMode = BackupAddrMode;
3235 AddrModeInsts.resize(OldSize);
3238 AddrMode.HasBaseReg = true;
3239 AddrMode.BaseReg = AddrInst->getOperand(0);
3242 // Match the remaining variable portion of the GEP.
3243 if (!matchScaledValue(AddrInst->getOperand(VariableOperand), VariableScale,
3245 // If it couldn't be matched, try stuffing the base into a register
3246 // instead of matching it, and retrying the match of the scale.
3247 AddrMode = BackupAddrMode;
3248 AddrModeInsts.resize(OldSize);
3249 if (AddrMode.HasBaseReg)
3251 AddrMode.HasBaseReg = true;
3252 AddrMode.BaseReg = AddrInst->getOperand(0);
3253 AddrMode.BaseOffs += ConstantOffset;
3254 if (!matchScaledValue(AddrInst->getOperand(VariableOperand),
3255 VariableScale, Depth)) {
3256 // If even that didn't work, bail.
3257 AddrMode = BackupAddrMode;
3258 AddrModeInsts.resize(OldSize);
3265 case Instruction::SExt:
3266 case Instruction::ZExt: {
3267 Instruction *Ext = dyn_cast<Instruction>(AddrInst);
3271 // Try to move this ext out of the way of the addressing mode.
3272 // Ask for a method for doing so.
3273 TypePromotionHelper::Action TPH =
3274 TypePromotionHelper::getAction(Ext, InsertedInsts, TLI, PromotedInsts);
3278 TypePromotionTransaction::ConstRestorationPt LastKnownGood =
3279 TPT.getRestorationPoint();
3280 unsigned CreatedInstsCost = 0;
3281 unsigned ExtCost = !TLI.isExtFree(Ext);
3282 Value *PromotedOperand =
3283 TPH(Ext, TPT, PromotedInsts, CreatedInstsCost, nullptr, nullptr, TLI);
3284 // SExt has been moved away.
3285 // Thus either it will be rematched later in the recursive calls or it is
3286 // gone. Anyway, we must not fold it into the addressing mode at this point.
3290 // addr = gep base, idx
3292 // promotedOpnd = ext opnd <- no match here
3293 // op = promoted_add promotedOpnd, 1 <- match (later in recursive calls)
3294 // addr = gep base, op <- match
3298 assert(PromotedOperand &&
3299 "TypePromotionHelper should have filtered out those cases");
3301 ExtAddrMode BackupAddrMode = AddrMode;
3302 unsigned OldSize = AddrModeInsts.size();
3304 if (!matchAddr(PromotedOperand, Depth) ||
3305 // The total of the new cost is equal to the cost of the created
3307 // The total of the old cost is equal to the cost of the extension plus
3308 // what we have saved in the addressing mode.
3309 !isPromotionProfitable(CreatedInstsCost,
3310 ExtCost + (AddrModeInsts.size() - OldSize),
3312 AddrMode = BackupAddrMode;
3313 AddrModeInsts.resize(OldSize);
3314 DEBUG(dbgs() << "Sign extension does not pay off: rollback\n");
3315 TPT.rollback(LastKnownGood);
3324 /// If we can, try to add the value of 'Addr' into the current addressing mode.
3325 /// If Addr can't be added to AddrMode this returns false and leaves AddrMode
3326 /// unmodified. This assumes that Addr is either a pointer type or intptr_t
3329 bool AddressingModeMatcher::matchAddr(Value *Addr, unsigned Depth) {
3330 // Start a transaction at this point that we will rollback if the matching
3332 TypePromotionTransaction::ConstRestorationPt LastKnownGood =
3333 TPT.getRestorationPoint();
3334 if (ConstantInt *CI = dyn_cast<ConstantInt>(Addr)) {
3335 // Fold in immediates if legal for the target.
3336 AddrMode.BaseOffs += CI->getSExtValue();
3337 if (TLI.isLegalAddressingMode(DL, AddrMode, AccessTy, AddrSpace))
3339 AddrMode.BaseOffs -= CI->getSExtValue();
3340 } else if (GlobalValue *GV = dyn_cast<GlobalValue>(Addr)) {
3341 // If this is a global variable, try to fold it into the addressing mode.
3342 if (!AddrMode.BaseGV) {
3343 AddrMode.BaseGV = GV;
3344 if (TLI.isLegalAddressingMode(DL, AddrMode, AccessTy, AddrSpace))
3346 AddrMode.BaseGV = nullptr;
3348 } else if (Instruction *I = dyn_cast<Instruction>(Addr)) {
3349 ExtAddrMode BackupAddrMode = AddrMode;
3350 unsigned OldSize = AddrModeInsts.size();
3352 // Check to see if it is possible to fold this operation.
3353 bool MovedAway = false;
3354 if (matchOperationAddr(I, I->getOpcode(), Depth, &MovedAway)) {
3355 // This instruction may have been moved away. If so, there is nothing
3359 // Okay, it's possible to fold this. Check to see if it is actually
3360 // *profitable* to do so. We use a simple cost model to avoid increasing
3361 // register pressure too much.
3362 if (I->hasOneUse() ||
3363 isProfitableToFoldIntoAddressingMode(I, BackupAddrMode, AddrMode)) {
3364 AddrModeInsts.push_back(I);
3368 // It isn't profitable to do this, roll back.
3369 //cerr << "NOT FOLDING: " << *I;
3370 AddrMode = BackupAddrMode;
3371 AddrModeInsts.resize(OldSize);
3372 TPT.rollback(LastKnownGood);
3374 } else if (ConstantExpr *CE = dyn_cast<ConstantExpr>(Addr)) {
3375 if (matchOperationAddr(CE, CE->getOpcode(), Depth))
3377 TPT.rollback(LastKnownGood);
3378 } else if (isa<ConstantPointerNull>(Addr)) {
3379 // Null pointer gets folded without affecting the addressing mode.
3383 // Worse case, the target should support [reg] addressing modes. :)
3384 if (!AddrMode.HasBaseReg) {
3385 AddrMode.HasBaseReg = true;
3386 AddrMode.BaseReg = Addr;
3387 // Still check for legality in case the target supports [imm] but not [i+r].
3388 if (TLI.isLegalAddressingMode(DL, AddrMode, AccessTy, AddrSpace))
3390 AddrMode.HasBaseReg = false;
3391 AddrMode.BaseReg = nullptr;
3394 // If the base register is already taken, see if we can do [r+r].
3395 if (AddrMode.Scale == 0) {
3397 AddrMode.ScaledReg = Addr;
3398 if (TLI.isLegalAddressingMode(DL, AddrMode, AccessTy, AddrSpace))
3401 AddrMode.ScaledReg = nullptr;
3404 TPT.rollback(LastKnownGood);
3408 /// Check to see if all uses of OpVal by the specified inline asm call are due
3409 /// to memory operands. If so, return true, otherwise return false.
3410 static bool IsOperandAMemoryOperand(CallInst *CI, InlineAsm *IA, Value *OpVal,
3411 const TargetMachine &TM) {
3412 const Function *F = CI->getParent()->getParent();
3413 const TargetLowering *TLI = TM.getSubtargetImpl(*F)->getTargetLowering();
3414 const TargetRegisterInfo *TRI = TM.getSubtargetImpl(*F)->getRegisterInfo();
3415 TargetLowering::AsmOperandInfoVector TargetConstraints =
3416 TLI->ParseConstraints(F->getParent()->getDataLayout(), TRI,
3417 ImmutableCallSite(CI));
3418 for (unsigned i = 0, e = TargetConstraints.size(); i != e; ++i) {
3419 TargetLowering::AsmOperandInfo &OpInfo = TargetConstraints[i];
3421 // Compute the constraint code and ConstraintType to use.
3422 TLI->ComputeConstraintToUse(OpInfo, SDValue());
3424 // If this asm operand is our Value*, and if it isn't an indirect memory
3425 // operand, we can't fold it!
3426 if (OpInfo.CallOperandVal == OpVal &&
3427 (OpInfo.ConstraintType != TargetLowering::C_Memory ||
3428 !OpInfo.isIndirect))
3435 /// Recursively walk all the uses of I until we find a memory use.
3436 /// If we find an obviously non-foldable instruction, return true.
3437 /// Add the ultimately found memory instructions to MemoryUses.
3438 static bool FindAllMemoryUses(
3440 SmallVectorImpl<std::pair<Instruction *, unsigned>> &MemoryUses,
3441 SmallPtrSetImpl<Instruction *> &ConsideredInsts, const TargetMachine &TM) {
3442 // If we already considered this instruction, we're done.
3443 if (!ConsideredInsts.insert(I).second)
3446 // If this is an obviously unfoldable instruction, bail out.
3447 if (!MightBeFoldableInst(I))
3450 // Loop over all the uses, recursively processing them.
3451 for (Use &U : I->uses()) {
3452 Instruction *UserI = cast<Instruction>(U.getUser());
3454 if (LoadInst *LI = dyn_cast<LoadInst>(UserI)) {
3455 MemoryUses.push_back(std::make_pair(LI, U.getOperandNo()));
3459 if (StoreInst *SI = dyn_cast<StoreInst>(UserI)) {
3460 unsigned opNo = U.getOperandNo();
3461 if (opNo == 0) return true; // Storing addr, not into addr.
3462 MemoryUses.push_back(std::make_pair(SI, opNo));
3466 if (CallInst *CI = dyn_cast<CallInst>(UserI)) {
3467 InlineAsm *IA = dyn_cast<InlineAsm>(CI->getCalledValue());
3468 if (!IA) return true;
3470 // If this is a memory operand, we're cool, otherwise bail out.
3471 if (!IsOperandAMemoryOperand(CI, IA, I, TM))
3476 if (FindAllMemoryUses(UserI, MemoryUses, ConsideredInsts, TM))
3483 /// Return true if Val is already known to be live at the use site that we're
3484 /// folding it into. If so, there is no cost to include it in the addressing
3485 /// mode. KnownLive1 and KnownLive2 are two values that we know are live at the
3486 /// instruction already.
3487 bool AddressingModeMatcher::valueAlreadyLiveAtInst(Value *Val,Value *KnownLive1,
3488 Value *KnownLive2) {
3489 // If Val is either of the known-live values, we know it is live!
3490 if (Val == nullptr || Val == KnownLive1 || Val == KnownLive2)
3493 // All values other than instructions and arguments (e.g. constants) are live.
3494 if (!isa<Instruction>(Val) && !isa<Argument>(Val)) return true;
3496 // If Val is a constant sized alloca in the entry block, it is live, this is
3497 // true because it is just a reference to the stack/frame pointer, which is
3498 // live for the whole function.
3499 if (AllocaInst *AI = dyn_cast<AllocaInst>(Val))
3500 if (AI->isStaticAlloca())
3503 // Check to see if this value is already used in the memory instruction's
3504 // block. If so, it's already live into the block at the very least, so we
3505 // can reasonably fold it.
3506 return Val->isUsedInBasicBlock(MemoryInst->getParent());
3509 /// It is possible for the addressing mode of the machine to fold the specified
3510 /// instruction into a load or store that ultimately uses it.
3511 /// However, the specified instruction has multiple uses.
3512 /// Given this, it may actually increase register pressure to fold it
3513 /// into the load. For example, consider this code:
3517 /// use(Y) -> nonload/store
3521 /// In this case, Y has multiple uses, and can be folded into the load of Z
3522 /// (yielding load [X+2]). However, doing this will cause both "X" and "X+1" to
3523 /// be live at the use(Y) line. If we don't fold Y into load Z, we use one
3524 /// fewer register. Since Y can't be folded into "use(Y)" we don't increase the
3525 /// number of computations either.
3527 /// Note that this (like most of CodeGenPrepare) is just a rough heuristic. If
3528 /// X was live across 'load Z' for other reasons, we actually *would* want to
3529 /// fold the addressing mode in the Z case. This would make Y die earlier.
3530 bool AddressingModeMatcher::
3531 isProfitableToFoldIntoAddressingMode(Instruction *I, ExtAddrMode &AMBefore,
3532 ExtAddrMode &AMAfter) {
3533 if (IgnoreProfitability) return true;
3535 // AMBefore is the addressing mode before this instruction was folded into it,
3536 // and AMAfter is the addressing mode after the instruction was folded. Get
3537 // the set of registers referenced by AMAfter and subtract out those
3538 // referenced by AMBefore: this is the set of values which folding in this
3539 // address extends the lifetime of.
3541 // Note that there are only two potential values being referenced here,
3542 // BaseReg and ScaleReg (global addresses are always available, as are any
3543 // folded immediates).
3544 Value *BaseReg = AMAfter.BaseReg, *ScaledReg = AMAfter.ScaledReg;
3546 // If the BaseReg or ScaledReg was referenced by the previous addrmode, their
3547 // lifetime wasn't extended by adding this instruction.
3548 if (valueAlreadyLiveAtInst(BaseReg, AMBefore.BaseReg, AMBefore.ScaledReg))
3550 if (valueAlreadyLiveAtInst(ScaledReg, AMBefore.BaseReg, AMBefore.ScaledReg))
3551 ScaledReg = nullptr;
3553 // If folding this instruction (and it's subexprs) didn't extend any live
3554 // ranges, we're ok with it.
3555 if (!BaseReg && !ScaledReg)
3558 // If all uses of this instruction are ultimately load/store/inlineasm's,
3559 // check to see if their addressing modes will include this instruction. If
3560 // so, we can fold it into all uses, so it doesn't matter if it has multiple
3562 SmallVector<std::pair<Instruction*,unsigned>, 16> MemoryUses;
3563 SmallPtrSet<Instruction*, 16> ConsideredInsts;
3564 if (FindAllMemoryUses(I, MemoryUses, ConsideredInsts, TM))
3565 return false; // Has a non-memory, non-foldable use!
3567 // Now that we know that all uses of this instruction are part of a chain of
3568 // computation involving only operations that could theoretically be folded
3569 // into a memory use, loop over each of these uses and see if they could
3570 // *actually* fold the instruction.
3571 SmallVector<Instruction*, 32> MatchedAddrModeInsts;
3572 for (unsigned i = 0, e = MemoryUses.size(); i != e; ++i) {
3573 Instruction *User = MemoryUses[i].first;
3574 unsigned OpNo = MemoryUses[i].second;
3576 // Get the access type of this use. If the use isn't a pointer, we don't
3577 // know what it accesses.
3578 Value *Address = User->getOperand(OpNo);
3579 PointerType *AddrTy = dyn_cast<PointerType>(Address->getType());
3582 Type *AddressAccessTy = AddrTy->getElementType();
3583 unsigned AS = AddrTy->getAddressSpace();
3585 // Do a match against the root of this address, ignoring profitability. This
3586 // will tell us if the addressing mode for the memory operation will
3587 // *actually* cover the shared instruction.
3589 TypePromotionTransaction::ConstRestorationPt LastKnownGood =
3590 TPT.getRestorationPoint();
3591 AddressingModeMatcher Matcher(MatchedAddrModeInsts, TM, AddressAccessTy, AS,
3592 MemoryInst, Result, InsertedInsts,
3593 PromotedInsts, TPT);
3594 Matcher.IgnoreProfitability = true;
3595 bool Success = Matcher.matchAddr(Address, 0);
3596 (void)Success; assert(Success && "Couldn't select *anything*?");
3598 // The match was to check the profitability, the changes made are not
3599 // part of the original matcher. Therefore, they should be dropped
3600 // otherwise the original matcher will not present the right state.
3601 TPT.rollback(LastKnownGood);
3603 // If the match didn't cover I, then it won't be shared by it.
3604 if (std::find(MatchedAddrModeInsts.begin(), MatchedAddrModeInsts.end(),
3605 I) == MatchedAddrModeInsts.end())
3608 MatchedAddrModeInsts.clear();
3614 } // end anonymous namespace
3616 /// Return true if the specified values are defined in a
3617 /// different basic block than BB.
3618 static bool IsNonLocalValue(Value *V, BasicBlock *BB) {
3619 if (Instruction *I = dyn_cast<Instruction>(V))
3620 return I->getParent() != BB;
3624 /// Load and Store Instructions often have addressing modes that can do
3625 /// significant amounts of computation. As such, instruction selection will try
3626 /// to get the load or store to do as much computation as possible for the
3627 /// program. The problem is that isel can only see within a single block. As
3628 /// such, we sink as much legal addressing mode work into the block as possible.
3630 /// This method is used to optimize both load/store and inline asms with memory
3632 bool CodeGenPrepare::optimizeMemoryInst(Instruction *MemoryInst, Value *Addr,
3633 Type *AccessTy, unsigned AddrSpace) {
3636 // Try to collapse single-value PHI nodes. This is necessary to undo
3637 // unprofitable PRE transformations.
3638 SmallVector<Value*, 8> worklist;
3639 SmallPtrSet<Value*, 16> Visited;
3640 worklist.push_back(Addr);
3642 // Use a worklist to iteratively look through PHI nodes, and ensure that
3643 // the addressing mode obtained from the non-PHI roots of the graph
3645 Value *Consensus = nullptr;
3646 unsigned NumUsesConsensus = 0;
3647 bool IsNumUsesConsensusValid = false;
3648 SmallVector<Instruction*, 16> AddrModeInsts;
3649 ExtAddrMode AddrMode;
3650 TypePromotionTransaction TPT;
3651 TypePromotionTransaction::ConstRestorationPt LastKnownGood =
3652 TPT.getRestorationPoint();
3653 while (!worklist.empty()) {
3654 Value *V = worklist.back();
3655 worklist.pop_back();
3657 // Break use-def graph loops.
3658 if (!Visited.insert(V).second) {
3659 Consensus = nullptr;
3663 // For a PHI node, push all of its incoming values.
3664 if (PHINode *P = dyn_cast<PHINode>(V)) {
3665 for (Value *IncValue : P->incoming_values())
3666 worklist.push_back(IncValue);
3670 // For non-PHIs, determine the addressing mode being computed.
3671 SmallVector<Instruction*, 16> NewAddrModeInsts;
3672 ExtAddrMode NewAddrMode = AddressingModeMatcher::Match(
3673 V, AccessTy, AddrSpace, MemoryInst, NewAddrModeInsts, *TM,
3674 InsertedInsts, PromotedInsts, TPT);
3676 // This check is broken into two cases with very similar code to avoid using
3677 // getNumUses() as much as possible. Some values have a lot of uses, so
3678 // calling getNumUses() unconditionally caused a significant compile-time
3682 AddrMode = NewAddrMode;
3683 AddrModeInsts = NewAddrModeInsts;
3685 } else if (NewAddrMode == AddrMode) {
3686 if (!IsNumUsesConsensusValid) {
3687 NumUsesConsensus = Consensus->getNumUses();
3688 IsNumUsesConsensusValid = true;
3691 // Ensure that the obtained addressing mode is equivalent to that obtained
3692 // for all other roots of the PHI traversal. Also, when choosing one
3693 // such root as representative, select the one with the most uses in order
3694 // to keep the cost modeling heuristics in AddressingModeMatcher
3696 unsigned NumUses = V->getNumUses();
3697 if (NumUses > NumUsesConsensus) {
3699 NumUsesConsensus = NumUses;
3700 AddrModeInsts = NewAddrModeInsts;
3705 Consensus = nullptr;
3709 // If the addressing mode couldn't be determined, or if multiple different
3710 // ones were determined, bail out now.
3712 TPT.rollback(LastKnownGood);
3717 // Check to see if any of the instructions supersumed by this addr mode are
3718 // non-local to I's BB.
3719 bool AnyNonLocal = false;
3720 for (unsigned i = 0, e = AddrModeInsts.size(); i != e; ++i) {
3721 if (IsNonLocalValue(AddrModeInsts[i], MemoryInst->getParent())) {
3727 // If all the instructions matched are already in this BB, don't do anything.
3729 DEBUG(dbgs() << "CGP: Found local addrmode: " << AddrMode << "\n");
3733 // Insert this computation right after this user. Since our caller is
3734 // scanning from the top of the BB to the bottom, reuse of the expr are
3735 // guaranteed to happen later.
3736 IRBuilder<> Builder(MemoryInst);
3738 // Now that we determined the addressing expression we want to use and know
3739 // that we have to sink it into this block. Check to see if we have already
3740 // done this for some other load/store instr in this block. If so, reuse the
3742 Value *&SunkAddr = SunkAddrs[Addr];
3744 DEBUG(dbgs() << "CGP: Reusing nonlocal addrmode: " << AddrMode << " for "
3745 << *MemoryInst << "\n");
3746 if (SunkAddr->getType() != Addr->getType())
3747 SunkAddr = Builder.CreateBitCast(SunkAddr, Addr->getType());
3748 } else if (AddrSinkUsingGEPs ||
3749 (!AddrSinkUsingGEPs.getNumOccurrences() && TM &&
3750 TM->getSubtargetImpl(*MemoryInst->getParent()->getParent())
3752 // By default, we use the GEP-based method when AA is used later. This
3753 // prevents new inttoptr/ptrtoint pairs from degrading AA capabilities.
3754 DEBUG(dbgs() << "CGP: SINKING nonlocal addrmode: " << AddrMode << " for "
3755 << *MemoryInst << "\n");
3756 Type *IntPtrTy = DL->getIntPtrType(Addr->getType());
3757 Value *ResultPtr = nullptr, *ResultIndex = nullptr;
3759 // First, find the pointer.
3760 if (AddrMode.BaseReg && AddrMode.BaseReg->getType()->isPointerTy()) {
3761 ResultPtr = AddrMode.BaseReg;
3762 AddrMode.BaseReg = nullptr;
3765 if (AddrMode.Scale && AddrMode.ScaledReg->getType()->isPointerTy()) {
3766 // We can't add more than one pointer together, nor can we scale a
3767 // pointer (both of which seem meaningless).
3768 if (ResultPtr || AddrMode.Scale != 1)
3771 ResultPtr = AddrMode.ScaledReg;
3775 if (AddrMode.BaseGV) {
3779 ResultPtr = AddrMode.BaseGV;
3782 // If the real base value actually came from an inttoptr, then the matcher
3783 // will look through it and provide only the integer value. In that case,
3785 if (!ResultPtr && AddrMode.BaseReg) {
3787 Builder.CreateIntToPtr(AddrMode.BaseReg, Addr->getType(), "sunkaddr");
3788 AddrMode.BaseReg = nullptr;
3789 } else if (!ResultPtr && AddrMode.Scale == 1) {
3791 Builder.CreateIntToPtr(AddrMode.ScaledReg, Addr->getType(), "sunkaddr");
3796 !AddrMode.BaseReg && !AddrMode.Scale && !AddrMode.BaseOffs) {
3797 SunkAddr = Constant::getNullValue(Addr->getType());
3798 } else if (!ResultPtr) {
3802 Builder.getInt8PtrTy(Addr->getType()->getPointerAddressSpace());
3803 Type *I8Ty = Builder.getInt8Ty();
3805 // Start with the base register. Do this first so that subsequent address
3806 // matching finds it last, which will prevent it from trying to match it
3807 // as the scaled value in case it happens to be a mul. That would be
3808 // problematic if we've sunk a different mul for the scale, because then
3809 // we'd end up sinking both muls.
3810 if (AddrMode.BaseReg) {
3811 Value *V = AddrMode.BaseReg;
3812 if (V->getType() != IntPtrTy)
3813 V = Builder.CreateIntCast(V, IntPtrTy, /*isSigned=*/true, "sunkaddr");
3818 // Add the scale value.
3819 if (AddrMode.Scale) {
3820 Value *V = AddrMode.ScaledReg;
3821 if (V->getType() == IntPtrTy) {
3823 } else if (cast<IntegerType>(IntPtrTy)->getBitWidth() <
3824 cast<IntegerType>(V->getType())->getBitWidth()) {
3825 V = Builder.CreateTrunc(V, IntPtrTy, "sunkaddr");
3827 // It is only safe to sign extend the BaseReg if we know that the math
3828 // required to create it did not overflow before we extend it. Since
3829 // the original IR value was tossed in favor of a constant back when
3830 // the AddrMode was created we need to bail out gracefully if widths
3831 // do not match instead of extending it.
3832 Instruction *I = dyn_cast_or_null<Instruction>(ResultIndex);
3833 if (I && (ResultIndex != AddrMode.BaseReg))
3834 I->eraseFromParent();
3838 if (AddrMode.Scale != 1)
3839 V = Builder.CreateMul(V, ConstantInt::get(IntPtrTy, AddrMode.Scale),
3842 ResultIndex = Builder.CreateAdd(ResultIndex, V, "sunkaddr");
3847 // Add in the Base Offset if present.
3848 if (AddrMode.BaseOffs) {
3849 Value *V = ConstantInt::get(IntPtrTy, AddrMode.BaseOffs);
3851 // We need to add this separately from the scale above to help with
3852 // SDAG consecutive load/store merging.
3853 if (ResultPtr->getType() != I8PtrTy)
3854 ResultPtr = Builder.CreateBitCast(ResultPtr, I8PtrTy);
3855 ResultPtr = Builder.CreateGEP(I8Ty, ResultPtr, ResultIndex, "sunkaddr");
3862 SunkAddr = ResultPtr;
3864 if (ResultPtr->getType() != I8PtrTy)
3865 ResultPtr = Builder.CreateBitCast(ResultPtr, I8PtrTy);
3866 SunkAddr = Builder.CreateGEP(I8Ty, ResultPtr, ResultIndex, "sunkaddr");
3869 if (SunkAddr->getType() != Addr->getType())
3870 SunkAddr = Builder.CreateBitCast(SunkAddr, Addr->getType());
3873 DEBUG(dbgs() << "CGP: SINKING nonlocal addrmode: " << AddrMode << " for "
3874 << *MemoryInst << "\n");
3875 Type *IntPtrTy = DL->getIntPtrType(Addr->getType());
3876 Value *Result = nullptr;
3878 // Start with the base register. Do this first so that subsequent address
3879 // matching finds it last, which will prevent it from trying to match it
3880 // as the scaled value in case it happens to be a mul. That would be
3881 // problematic if we've sunk a different mul for the scale, because then
3882 // we'd end up sinking both muls.
3883 if (AddrMode.BaseReg) {
3884 Value *V = AddrMode.BaseReg;
3885 if (V->getType()->isPointerTy())
3886 V = Builder.CreatePtrToInt(V, IntPtrTy, "sunkaddr");
3887 if (V->getType() != IntPtrTy)
3888 V = Builder.CreateIntCast(V, IntPtrTy, /*isSigned=*/true, "sunkaddr");
3892 // Add the scale value.
3893 if (AddrMode.Scale) {
3894 Value *V = AddrMode.ScaledReg;
3895 if (V->getType() == IntPtrTy) {
3897 } else if (V->getType()->isPointerTy()) {
3898 V = Builder.CreatePtrToInt(V, IntPtrTy, "sunkaddr");
3899 } else if (cast<IntegerType>(IntPtrTy)->getBitWidth() <
3900 cast<IntegerType>(V->getType())->getBitWidth()) {
3901 V = Builder.CreateTrunc(V, IntPtrTy, "sunkaddr");
3903 // It is only safe to sign extend the BaseReg if we know that the math
3904 // required to create it did not overflow before we extend it. Since
3905 // the original IR value was tossed in favor of a constant back when
3906 // the AddrMode was created we need to bail out gracefully if widths
3907 // do not match instead of extending it.
3908 Instruction *I = dyn_cast_or_null<Instruction>(Result);
3909 if (I && (Result != AddrMode.BaseReg))
3910 I->eraseFromParent();
3913 if (AddrMode.Scale != 1)
3914 V = Builder.CreateMul(V, ConstantInt::get(IntPtrTy, AddrMode.Scale),
3917 Result = Builder.CreateAdd(Result, V, "sunkaddr");
3922 // Add in the BaseGV if present.
3923 if (AddrMode.BaseGV) {
3924 Value *V = Builder.CreatePtrToInt(AddrMode.BaseGV, IntPtrTy, "sunkaddr");
3926 Result = Builder.CreateAdd(Result, V, "sunkaddr");
3931 // Add in the Base Offset if present.
3932 if (AddrMode.BaseOffs) {
3933 Value *V = ConstantInt::get(IntPtrTy, AddrMode.BaseOffs);
3935 Result = Builder.CreateAdd(Result, V, "sunkaddr");
3941 SunkAddr = Constant::getNullValue(Addr->getType());
3943 SunkAddr = Builder.CreateIntToPtr(Result, Addr->getType(), "sunkaddr");
3946 MemoryInst->replaceUsesOfWith(Repl, SunkAddr);
3948 // If we have no uses, recursively delete the value and all dead instructions
3950 if (Repl->use_empty()) {
3951 // This can cause recursive deletion, which can invalidate our iterator.
3952 // Use a WeakVH to hold onto it in case this happens.
3953 WeakVH IterHandle(&*CurInstIterator);
3954 BasicBlock *BB = CurInstIterator->getParent();
3956 RecursivelyDeleteTriviallyDeadInstructions(Repl, TLInfo);
3958 if (IterHandle != CurInstIterator.getNodePtrUnchecked()) {
3959 // If the iterator instruction was recursively deleted, start over at the
3960 // start of the block.
3961 CurInstIterator = BB->begin();
3969 /// If there are any memory operands, use OptimizeMemoryInst to sink their
3970 /// address computing into the block when possible / profitable.
3971 bool CodeGenPrepare::optimizeInlineAsmInst(CallInst *CS) {
3972 bool MadeChange = false;
3974 const TargetRegisterInfo *TRI =
3975 TM->getSubtargetImpl(*CS->getParent()->getParent())->getRegisterInfo();
3976 TargetLowering::AsmOperandInfoVector TargetConstraints =
3977 TLI->ParseConstraints(*DL, TRI, CS);
3979 for (unsigned i = 0, e = TargetConstraints.size(); i != e; ++i) {
3980 TargetLowering::AsmOperandInfo &OpInfo = TargetConstraints[i];
3982 // Compute the constraint code and ConstraintType to use.
3983 TLI->ComputeConstraintToUse(OpInfo, SDValue());
3985 if (OpInfo.ConstraintType == TargetLowering::C_Memory &&
3986 OpInfo.isIndirect) {
3987 Value *OpVal = CS->getArgOperand(ArgNo++);
3988 MadeChange |= optimizeMemoryInst(CS, OpVal, OpVal->getType(), ~0u);
3989 } else if (OpInfo.Type == InlineAsm::isInput)
3996 /// \brief Check if all the uses of \p Inst are equivalent (or free) zero or
3997 /// sign extensions.
3998 static bool hasSameExtUse(Instruction *Inst, const TargetLowering &TLI) {
3999 assert(!Inst->use_empty() && "Input must have at least one use");
4000 const Instruction *FirstUser = cast<Instruction>(*Inst->user_begin());
4001 bool IsSExt = isa<SExtInst>(FirstUser);
4002 Type *ExtTy = FirstUser->getType();
4003 for (const User *U : Inst->users()) {
4004 const Instruction *UI = cast<Instruction>(U);
4005 if ((IsSExt && !isa<SExtInst>(UI)) || (!IsSExt && !isa<ZExtInst>(UI)))
4007 Type *CurTy = UI->getType();
4008 // Same input and output types: Same instruction after CSE.
4012 // If IsSExt is true, we are in this situation:
4014 // b = sext ty1 a to ty2
4015 // c = sext ty1 a to ty3
4016 // Assuming ty2 is shorter than ty3, this could be turned into:
4018 // b = sext ty1 a to ty2
4019 // c = sext ty2 b to ty3
4020 // However, the last sext is not free.
4024 // This is a ZExt, maybe this is free to extend from one type to another.
4025 // In that case, we would not account for a different use.
4028 if (ExtTy->getScalarType()->getIntegerBitWidth() >
4029 CurTy->getScalarType()->getIntegerBitWidth()) {
4037 if (!TLI.isZExtFree(NarrowTy, LargeTy))
4040 // All uses are the same or can be derived from one another for free.
4044 /// \brief Try to form ExtLd by promoting \p Exts until they reach a
4045 /// load instruction.
4046 /// If an ext(load) can be formed, it is returned via \p LI for the load
4047 /// and \p Inst for the extension.
4048 /// Otherwise LI == nullptr and Inst == nullptr.
4049 /// When some promotion happened, \p TPT contains the proper state to
4052 /// \return true when promoting was necessary to expose the ext(load)
4053 /// opportunity, false otherwise.
4057 /// %ld = load i32* %addr
4058 /// %add = add nuw i32 %ld, 4
4059 /// %zext = zext i32 %add to i64
4063 /// %ld = load i32* %addr
4064 /// %zext = zext i32 %ld to i64
4065 /// %add = add nuw i64 %zext, 4
4067 /// Thanks to the promotion, we can match zext(load i32*) to i64.
4068 bool CodeGenPrepare::extLdPromotion(TypePromotionTransaction &TPT,
4069 LoadInst *&LI, Instruction *&Inst,
4070 const SmallVectorImpl<Instruction *> &Exts,
4071 unsigned CreatedInstsCost = 0) {
4072 // Iterate over all the extensions to see if one form an ext(load).
4073 for (auto I : Exts) {
4074 // Check if we directly have ext(load).
4075 if ((LI = dyn_cast<LoadInst>(I->getOperand(0)))) {
4077 // No promotion happened here.
4080 // Check whether or not we want to do any promotion.
4081 if (!TLI || !TLI->enableExtLdPromotion() || DisableExtLdPromotion)
4083 // Get the action to perform the promotion.
4084 TypePromotionHelper::Action TPH = TypePromotionHelper::getAction(
4085 I, InsertedInsts, *TLI, PromotedInsts);
4086 // Check if we can promote.
4089 // Save the current state.
4090 TypePromotionTransaction::ConstRestorationPt LastKnownGood =
4091 TPT.getRestorationPoint();
4092 SmallVector<Instruction *, 4> NewExts;
4093 unsigned NewCreatedInstsCost = 0;
4094 unsigned ExtCost = !TLI->isExtFree(I);
4096 Value *PromotedVal = TPH(I, TPT, PromotedInsts, NewCreatedInstsCost,
4097 &NewExts, nullptr, *TLI);
4098 assert(PromotedVal &&
4099 "TypePromotionHelper should have filtered out those cases");
4101 // We would be able to merge only one extension in a load.
4102 // Therefore, if we have more than 1 new extension we heuristically
4103 // cut this search path, because it means we degrade the code quality.
4104 // With exactly 2, the transformation is neutral, because we will merge
4105 // one extension but leave one. However, we optimistically keep going,
4106 // because the new extension may be removed too.
4107 long long TotalCreatedInstsCost = CreatedInstsCost + NewCreatedInstsCost;
4108 TotalCreatedInstsCost -= ExtCost;
4109 if (!StressExtLdPromotion &&
4110 (TotalCreatedInstsCost > 1 ||
4111 !isPromotedInstructionLegal(*TLI, *DL, PromotedVal))) {
4112 // The promotion is not profitable, rollback to the previous state.
4113 TPT.rollback(LastKnownGood);
4116 // The promotion is profitable.
4117 // Check if it exposes an ext(load).
4118 (void)extLdPromotion(TPT, LI, Inst, NewExts, TotalCreatedInstsCost);
4119 if (LI && (StressExtLdPromotion || NewCreatedInstsCost <= ExtCost ||
4120 // If we have created a new extension, i.e., now we have two
4121 // extensions. We must make sure one of them is merged with
4122 // the load, otherwise we may degrade the code quality.
4123 (LI->hasOneUse() || hasSameExtUse(LI, *TLI))))
4124 // Promotion happened.
4126 // If this does not help to expose an ext(load) then, rollback.
4127 TPT.rollback(LastKnownGood);
4129 // None of the extension can form an ext(load).
4135 /// Move a zext or sext fed by a load into the same basic block as the load,
4136 /// unless conditions are unfavorable. This allows SelectionDAG to fold the
4137 /// extend into the load.
4138 /// \p I[in/out] the extension may be modified during the process if some
4139 /// promotions apply.
4141 bool CodeGenPrepare::moveExtToFormExtLoad(Instruction *&I) {
4142 // Try to promote a chain of computation if it allows to form
4143 // an extended load.
4144 TypePromotionTransaction TPT;
4145 TypePromotionTransaction::ConstRestorationPt LastKnownGood =
4146 TPT.getRestorationPoint();
4147 SmallVector<Instruction *, 1> Exts;
4149 // Look for a load being extended.
4150 LoadInst *LI = nullptr;
4151 Instruction *OldExt = I;
4152 bool HasPromoted = extLdPromotion(TPT, LI, I, Exts);
4154 assert(!HasPromoted && !LI && "If we did not match any load instruction "
4155 "the code must remain the same");
4160 // If they're already in the same block, there's nothing to do.
4161 // Make the cheap checks first if we did not promote.
4162 // If we promoted, we need to check if it is indeed profitable.
4163 if (!HasPromoted && LI->getParent() == I->getParent())
4166 EVT VT = TLI->getValueType(*DL, I->getType());
4167 EVT LoadVT = TLI->getValueType(*DL, LI->getType());
4169 // If the load has other users and the truncate is not free, this probably
4170 // isn't worthwhile.
4171 if (!LI->hasOneUse() && TLI &&
4172 (TLI->isTypeLegal(LoadVT) || !TLI->isTypeLegal(VT)) &&
4173 !TLI->isTruncateFree(I->getType(), LI->getType())) {
4175 TPT.rollback(LastKnownGood);
4179 // Check whether the target supports casts folded into loads.
4181 if (isa<ZExtInst>(I))
4182 LType = ISD::ZEXTLOAD;
4184 assert(isa<SExtInst>(I) && "Unexpected ext type!");
4185 LType = ISD::SEXTLOAD;
4187 if (TLI && !TLI->isLoadExtLegal(LType, VT, LoadVT)) {
4189 TPT.rollback(LastKnownGood);
4193 // Move the extend into the same block as the load, so that SelectionDAG
4196 I->removeFromParent();
4202 bool CodeGenPrepare::optimizeExtUses(Instruction *I) {
4203 BasicBlock *DefBB = I->getParent();
4205 // If the result of a {s|z}ext and its source are both live out, rewrite all
4206 // other uses of the source with result of extension.
4207 Value *Src = I->getOperand(0);
4208 if (Src->hasOneUse())
4211 // Only do this xform if truncating is free.
4212 if (TLI && !TLI->isTruncateFree(I->getType(), Src->getType()))
4215 // Only safe to perform the optimization if the source is also defined in
4217 if (!isa<Instruction>(Src) || DefBB != cast<Instruction>(Src)->getParent())
4220 bool DefIsLiveOut = false;
4221 for (User *U : I->users()) {
4222 Instruction *UI = cast<Instruction>(U);
4224 // Figure out which BB this ext is used in.
4225 BasicBlock *UserBB = UI->getParent();
4226 if (UserBB == DefBB) continue;
4227 DefIsLiveOut = true;
4233 // Make sure none of the uses are PHI nodes.
4234 for (User *U : Src->users()) {
4235 Instruction *UI = cast<Instruction>(U);
4236 BasicBlock *UserBB = UI->getParent();
4237 if (UserBB == DefBB) continue;
4238 // Be conservative. We don't want this xform to end up introducing
4239 // reloads just before load / store instructions.
4240 if (isa<PHINode>(UI) || isa<LoadInst>(UI) || isa<StoreInst>(UI))
4244 // InsertedTruncs - Only insert one trunc in each block once.
4245 DenseMap<BasicBlock*, Instruction*> InsertedTruncs;
4247 bool MadeChange = false;
4248 for (Use &U : Src->uses()) {
4249 Instruction *User = cast<Instruction>(U.getUser());
4251 // Figure out which BB this ext is used in.
4252 BasicBlock *UserBB = User->getParent();
4253 if (UserBB == DefBB) continue;
4255 // Both src and def are live in this block. Rewrite the use.
4256 Instruction *&InsertedTrunc = InsertedTruncs[UserBB];
4258 if (!InsertedTrunc) {
4259 BasicBlock::iterator InsertPt = UserBB->getFirstInsertionPt();
4260 assert(InsertPt != UserBB->end());
4261 InsertedTrunc = new TruncInst(I, Src->getType(), "", &*InsertPt);
4262 InsertedInsts.insert(InsertedTrunc);
4265 // Replace a use of the {s|z}ext source with a use of the result.
4274 // Find loads whose uses only use some of the loaded value's bits. Add an "and"
4275 // just after the load if the target can fold this into one extload instruction,
4276 // with the hope of eliminating some of the other later "and" instructions using
4277 // the loaded value. "and"s that are made trivially redundant by the insertion
4278 // of the new "and" are removed by this function, while others (e.g. those whose
4279 // path from the load goes through a phi) are left for isel to potentially
4312 // becomes (after a call to optimizeLoadExt for each load):
4316 // x1' = and x1, 0xff
4320 // x2' = and x2, 0xff
4327 bool CodeGenPrepare::optimizeLoadExt(LoadInst *Load) {
4329 if (!Load->isSimple() ||
4330 !(Load->getType()->isIntegerTy() || Load->getType()->isPointerTy()))
4333 // Skip loads we've already transformed or have no reason to transform.
4334 if (Load->hasOneUse()) {
4335 User *LoadUser = *Load->user_begin();
4336 if (cast<Instruction>(LoadUser)->getParent() == Load->getParent() &&
4337 !dyn_cast<PHINode>(LoadUser))
4341 // Look at all uses of Load, looking through phis, to determine how many bits
4342 // of the loaded value are needed.
4343 SmallVector<Instruction *, 8> WorkList;
4344 SmallPtrSet<Instruction *, 16> Visited;
4345 SmallVector<Instruction *, 8> AndsToMaybeRemove;
4346 for (auto *U : Load->users())
4347 WorkList.push_back(cast<Instruction>(U));
4349 EVT LoadResultVT = TLI->getValueType(*DL, Load->getType());
4350 unsigned BitWidth = LoadResultVT.getSizeInBits();
4351 APInt DemandBits(BitWidth, 0);
4352 APInt WidestAndBits(BitWidth, 0);
4354 while (!WorkList.empty()) {
4355 Instruction *I = WorkList.back();
4356 WorkList.pop_back();
4358 // Break use-def graph loops.
4359 if (!Visited.insert(I).second)
4362 // For a PHI node, push all of its users.
4363 if (auto *Phi = dyn_cast<PHINode>(I)) {
4364 for (auto *U : Phi->users())
4365 WorkList.push_back(cast<Instruction>(U));
4369 switch (I->getOpcode()) {
4370 case llvm::Instruction::And: {
4371 auto *AndC = dyn_cast<ConstantInt>(I->getOperand(1));
4374 APInt AndBits = AndC->getValue();
4375 DemandBits |= AndBits;
4376 // Keep track of the widest and mask we see.
4377 if (AndBits.ugt(WidestAndBits))
4378 WidestAndBits = AndBits;
4379 if (AndBits == WidestAndBits && I->getOperand(0) == Load)
4380 AndsToMaybeRemove.push_back(I);
4384 case llvm::Instruction::Shl: {
4385 auto *ShlC = dyn_cast<ConstantInt>(I->getOperand(1));
4388 uint64_t ShiftAmt = ShlC->getLimitedValue(BitWidth - 1);
4389 auto ShlDemandBits = APInt::getAllOnesValue(BitWidth).lshr(ShiftAmt);
4390 DemandBits |= ShlDemandBits;
4394 case llvm::Instruction::Trunc: {
4395 EVT TruncVT = TLI->getValueType(*DL, I->getType());
4396 unsigned TruncBitWidth = TruncVT.getSizeInBits();
4397 auto TruncBits = APInt::getAllOnesValue(TruncBitWidth).zext(BitWidth);
4398 DemandBits |= TruncBits;
4407 uint32_t ActiveBits = DemandBits.getActiveBits();
4408 // Avoid hoisting (and (load x) 1) since it is unlikely to be folded by the
4409 // target even if isLoadExtLegal says an i1 EXTLOAD is valid. For example,
4410 // for the AArch64 target isLoadExtLegal(ZEXTLOAD, i32, i1) returns true, but
4411 // (and (load x) 1) is not matched as a single instruction, rather as a LDR
4412 // followed by an AND.
4413 // TODO: Look into removing this restriction by fixing backends to either
4414 // return false for isLoadExtLegal for i1 or have them select this pattern to
4415 // a single instruction.
4417 // Also avoid hoisting if we didn't see any ands with the exact DemandBits
4418 // mask, since these are the only ands that will be removed by isel.
4419 if (ActiveBits <= 1 || !APIntOps::isMask(ActiveBits, DemandBits) ||
4420 WidestAndBits != DemandBits)
4423 LLVMContext &Ctx = Load->getType()->getContext();
4424 Type *TruncTy = Type::getIntNTy(Ctx, ActiveBits);
4425 EVT TruncVT = TLI->getValueType(*DL, TruncTy);
4427 // Reject cases that won't be matched as extloads.
4428 if (!LoadResultVT.bitsGT(TruncVT) || !TruncVT.isRound() ||
4429 !TLI->isLoadExtLegal(ISD::ZEXTLOAD, LoadResultVT, TruncVT))
4432 IRBuilder<> Builder(Load->getNextNode());
4433 auto *NewAnd = dyn_cast<Instruction>(
4434 Builder.CreateAnd(Load, ConstantInt::get(Ctx, DemandBits)));
4436 // Replace all uses of load with new and (except for the use of load in the
4438 Load->replaceAllUsesWith(NewAnd);
4439 NewAnd->setOperand(0, Load);
4441 // Remove any and instructions that are now redundant.
4442 for (auto *And : AndsToMaybeRemove)
4443 // Check that the and mask is the same as the one we decided to put on the
4445 if (cast<ConstantInt>(And->getOperand(1))->getValue() == DemandBits) {
4446 And->replaceAllUsesWith(NewAnd);
4447 if (&*CurInstIterator == And)
4448 CurInstIterator = std::next(And->getIterator());
4449 And->eraseFromParent();
4457 /// Check if V (an operand of a select instruction) is an expensive instruction
4458 /// that is only used once.
4459 static bool sinkSelectOperand(const TargetTransformInfo *TTI, Value *V) {
4460 auto *I = dyn_cast<Instruction>(V);
4461 // If it's safe to speculatively execute, then it should not have side
4462 // effects; therefore, it's safe to sink and possibly *not* execute.
4463 return I && I->hasOneUse() && isSafeToSpeculativelyExecute(I) &&
4464 TTI->getUserCost(I) >= TargetTransformInfo::TCC_Expensive;
4467 /// Returns true if a SelectInst should be turned into an explicit branch.
4468 static bool isFormingBranchFromSelectProfitable(const TargetTransformInfo *TTI,
4470 // FIXME: This should use the same heuristics as IfConversion to determine
4471 // whether a select is better represented as a branch. This requires that
4472 // branch probability metadata is preserved for the select, which is not the
4475 CmpInst *Cmp = dyn_cast<CmpInst>(SI->getCondition());
4477 // If a branch is predictable, an out-of-order CPU can avoid blocking on its
4478 // comparison condition. If the compare has more than one use, there's
4479 // probably another cmov or setcc around, so it's not worth emitting a branch.
4480 if (!Cmp || !Cmp->hasOneUse())
4483 Value *CmpOp0 = Cmp->getOperand(0);
4484 Value *CmpOp1 = Cmp->getOperand(1);
4486 // Emit "cmov on compare with a memory operand" as a branch to avoid stalls
4487 // on a load from memory. But if the load is used more than once, do not
4488 // change the select to a branch because the load is probably needed
4489 // regardless of whether the branch is taken or not.
4490 if ((isa<LoadInst>(CmpOp0) && CmpOp0->hasOneUse()) ||
4491 (isa<LoadInst>(CmpOp1) && CmpOp1->hasOneUse()))
4494 // If either operand of the select is expensive and only needed on one side
4495 // of the select, we should form a branch.
4496 if (sinkSelectOperand(TTI, SI->getTrueValue()) ||
4497 sinkSelectOperand(TTI, SI->getFalseValue()))
4504 /// If we have a SelectInst that will likely profit from branch prediction,
4505 /// turn it into a branch.
4506 bool CodeGenPrepare::optimizeSelectInst(SelectInst *SI) {
4507 bool VectorCond = !SI->getCondition()->getType()->isIntegerTy(1);
4509 // Can we convert the 'select' to CF ?
4510 if (DisableSelectToBranch || OptSize || !TLI || VectorCond)
4513 TargetLowering::SelectSupportKind SelectKind;
4515 SelectKind = TargetLowering::VectorMaskSelect;
4516 else if (SI->getType()->isVectorTy())
4517 SelectKind = TargetLowering::ScalarCondVectorVal;
4519 SelectKind = TargetLowering::ScalarValSelect;
4521 // Do we have efficient codegen support for this kind of 'selects' ?
4522 if (TLI->isSelectSupported(SelectKind)) {
4523 // We have efficient codegen support for the select instruction.
4524 // Check if it is profitable to keep this 'select'.
4525 if (!TLI->isPredictableSelectExpensive() ||
4526 !isFormingBranchFromSelectProfitable(TTI, SI))
4532 // Transform a sequence like this:
4534 // %cmp = cmp uge i32 %a, %b
4535 // %sel = select i1 %cmp, i32 %c, i32 %d
4539 // %cmp = cmp uge i32 %a, %b
4540 // br i1 %cmp, label %select.true, label %select.false
4542 // br label %select.end
4544 // br label %select.end
4546 // %sel = phi i32 [ %c, %select.true ], [ %d, %select.false ]
4548 // In addition, we may sink instructions that produce %c or %d from
4549 // the entry block into the destination(s) of the new branch.
4550 // If the true or false blocks do not contain a sunken instruction, that
4551 // block and its branch may be optimized away. In that case, one side of the
4552 // first branch will point directly to select.end, and the corresponding PHI
4553 // predecessor block will be the start block.
4555 // First, we split the block containing the select into 2 blocks.
4556 BasicBlock *StartBlock = SI->getParent();
4557 BasicBlock::iterator SplitPt = ++(BasicBlock::iterator(SI));
4558 BasicBlock *EndBlock = StartBlock->splitBasicBlock(SplitPt, "select.end");
4560 // Delete the unconditional branch that was just created by the split.
4561 StartBlock->getTerminator()->eraseFromParent();
4563 // These are the new basic blocks for the conditional branch.
4564 // At least one will become an actual new basic block.
4565 BasicBlock *TrueBlock = nullptr;
4566 BasicBlock *FalseBlock = nullptr;
4568 // Sink expensive instructions into the conditional blocks to avoid executing
4569 // them speculatively.
4570 if (sinkSelectOperand(TTI, SI->getTrueValue())) {
4571 TrueBlock = BasicBlock::Create(SI->getContext(), "select.true.sink",
4572 EndBlock->getParent(), EndBlock);
4573 auto *TrueBranch = BranchInst::Create(EndBlock, TrueBlock);
4574 auto *TrueInst = cast<Instruction>(SI->getTrueValue());
4575 TrueInst->moveBefore(TrueBranch);
4577 if (sinkSelectOperand(TTI, SI->getFalseValue())) {
4578 FalseBlock = BasicBlock::Create(SI->getContext(), "select.false.sink",
4579 EndBlock->getParent(), EndBlock);
4580 auto *FalseBranch = BranchInst::Create(EndBlock, FalseBlock);
4581 auto *FalseInst = cast<Instruction>(SI->getFalseValue());
4582 FalseInst->moveBefore(FalseBranch);
4585 // If there was nothing to sink, then arbitrarily choose the 'false' side
4586 // for a new input value to the PHI.
4587 if (TrueBlock == FalseBlock) {
4588 assert(TrueBlock == nullptr &&
4589 "Unexpected basic block transform while optimizing select");
4591 FalseBlock = BasicBlock::Create(SI->getContext(), "select.false",
4592 EndBlock->getParent(), EndBlock);
4593 BranchInst::Create(EndBlock, FalseBlock);
4596 // Insert the real conditional branch based on the original condition.
4597 // If we did not create a new block for one of the 'true' or 'false' paths
4598 // of the condition, it means that side of the branch goes to the end block
4599 // directly and the path originates from the start block from the point of
4600 // view of the new PHI.
4601 if (TrueBlock == nullptr) {
4602 BranchInst::Create(EndBlock, FalseBlock, SI->getCondition(), SI);
4603 TrueBlock = StartBlock;
4604 } else if (FalseBlock == nullptr) {
4605 BranchInst::Create(TrueBlock, EndBlock, SI->getCondition(), SI);
4606 FalseBlock = StartBlock;
4608 BranchInst::Create(TrueBlock, FalseBlock, SI->getCondition(), SI);
4611 // The select itself is replaced with a PHI Node.
4612 PHINode *PN = PHINode::Create(SI->getType(), 2, "", &EndBlock->front());
4614 PN->addIncoming(SI->getTrueValue(), TrueBlock);
4615 PN->addIncoming(SI->getFalseValue(), FalseBlock);
4617 SI->replaceAllUsesWith(PN);
4618 SI->eraseFromParent();
4620 // Instruct OptimizeBlock to skip to the next block.
4621 CurInstIterator = StartBlock->end();
4622 ++NumSelectsExpanded;
4626 static bool isBroadcastShuffle(ShuffleVectorInst *SVI) {
4627 SmallVector<int, 16> Mask(SVI->getShuffleMask());
4629 for (unsigned i = 0; i < Mask.size(); ++i) {
4630 if (SplatElem != -1 && Mask[i] != -1 && Mask[i] != SplatElem)
4632 SplatElem = Mask[i];
4638 /// Some targets have expensive vector shifts if the lanes aren't all the same
4639 /// (e.g. x86 only introduced "vpsllvd" and friends with AVX2). In these cases
4640 /// it's often worth sinking a shufflevector splat down to its use so that
4641 /// codegen can spot all lanes are identical.
4642 bool CodeGenPrepare::optimizeShuffleVectorInst(ShuffleVectorInst *SVI) {
4643 BasicBlock *DefBB = SVI->getParent();
4645 // Only do this xform if variable vector shifts are particularly expensive.
4646 if (!TLI || !TLI->isVectorShiftByScalarCheap(SVI->getType()))
4649 // We only expect better codegen by sinking a shuffle if we can recognise a
4651 if (!isBroadcastShuffle(SVI))
4654 // InsertedShuffles - Only insert a shuffle in each block once.
4655 DenseMap<BasicBlock*, Instruction*> InsertedShuffles;
4657 bool MadeChange = false;
4658 for (User *U : SVI->users()) {
4659 Instruction *UI = cast<Instruction>(U);
4661 // Figure out which BB this ext is used in.
4662 BasicBlock *UserBB = UI->getParent();
4663 if (UserBB == DefBB) continue;
4665 // For now only apply this when the splat is used by a shift instruction.
4666 if (!UI->isShift()) continue;
4668 // Everything checks out, sink the shuffle if the user's block doesn't
4669 // already have a copy.
4670 Instruction *&InsertedShuffle = InsertedShuffles[UserBB];
4672 if (!InsertedShuffle) {
4673 BasicBlock::iterator InsertPt = UserBB->getFirstInsertionPt();
4674 assert(InsertPt != UserBB->end());
4676 new ShuffleVectorInst(SVI->getOperand(0), SVI->getOperand(1),
4677 SVI->getOperand(2), "", &*InsertPt);
4680 UI->replaceUsesOfWith(SVI, InsertedShuffle);
4684 // If we removed all uses, nuke the shuffle.
4685 if (SVI->use_empty()) {
4686 SVI->eraseFromParent();
4693 bool CodeGenPrepare::optimizeSwitchInst(SwitchInst *SI) {
4697 Value *Cond = SI->getCondition();
4698 Type *OldType = Cond->getType();
4699 LLVMContext &Context = Cond->getContext();
4700 MVT RegType = TLI->getRegisterType(Context, TLI->getValueType(*DL, OldType));
4701 unsigned RegWidth = RegType.getSizeInBits();
4703 if (RegWidth <= cast<IntegerType>(OldType)->getBitWidth())
4706 // If the register width is greater than the type width, expand the condition
4707 // of the switch instruction and each case constant to the width of the
4708 // register. By widening the type of the switch condition, subsequent
4709 // comparisons (for case comparisons) will not need to be extended to the
4710 // preferred register width, so we will potentially eliminate N-1 extends,
4711 // where N is the number of cases in the switch.
4712 auto *NewType = Type::getIntNTy(Context, RegWidth);
4714 // Zero-extend the switch condition and case constants unless the switch
4715 // condition is a function argument that is already being sign-extended.
4716 // In that case, we can avoid an unnecessary mask/extension by sign-extending
4717 // everything instead.
4718 Instruction::CastOps ExtType = Instruction::ZExt;
4719 if (auto *Arg = dyn_cast<Argument>(Cond))
4720 if (Arg->hasSExtAttr())
4721 ExtType = Instruction::SExt;
4723 auto *ExtInst = CastInst::Create(ExtType, Cond, NewType);
4724 ExtInst->insertBefore(SI);
4725 SI->setCondition(ExtInst);
4726 for (SwitchInst::CaseIt Case : SI->cases()) {
4727 APInt NarrowConst = Case.getCaseValue()->getValue();
4728 APInt WideConst = (ExtType == Instruction::ZExt) ?
4729 NarrowConst.zext(RegWidth) : NarrowConst.sext(RegWidth);
4730 Case.setValue(ConstantInt::get(Context, WideConst));
4737 /// \brief Helper class to promote a scalar operation to a vector one.
4738 /// This class is used to move downward extractelement transition.
4740 /// a = vector_op <2 x i32>
4741 /// b = extractelement <2 x i32> a, i32 0
4746 /// a = vector_op <2 x i32>
4747 /// c = vector_op a (equivalent to scalar_op on the related lane)
4748 /// * d = extractelement <2 x i32> c, i32 0
4750 /// Assuming both extractelement and store can be combine, we get rid of the
4752 class VectorPromoteHelper {
4753 /// DataLayout associated with the current module.
4754 const DataLayout &DL;
4756 /// Used to perform some checks on the legality of vector operations.
4757 const TargetLowering &TLI;
4759 /// Used to estimated the cost of the promoted chain.
4760 const TargetTransformInfo &TTI;
4762 /// The transition being moved downwards.
4763 Instruction *Transition;
4764 /// The sequence of instructions to be promoted.
4765 SmallVector<Instruction *, 4> InstsToBePromoted;
4766 /// Cost of combining a store and an extract.
4767 unsigned StoreExtractCombineCost;
4768 /// Instruction that will be combined with the transition.
4769 Instruction *CombineInst;
4771 /// \brief The instruction that represents the current end of the transition.
4772 /// Since we are faking the promotion until we reach the end of the chain
4773 /// of computation, we need a way to get the current end of the transition.
4774 Instruction *getEndOfTransition() const {
4775 if (InstsToBePromoted.empty())
4777 return InstsToBePromoted.back();
4780 /// \brief Return the index of the original value in the transition.
4781 /// E.g., for "extractelement <2 x i32> c, i32 1" the original value,
4782 /// c, is at index 0.
4783 unsigned getTransitionOriginalValueIdx() const {
4784 assert(isa<ExtractElementInst>(Transition) &&
4785 "Other kind of transitions are not supported yet");
4789 /// \brief Return the index of the index in the transition.
4790 /// E.g., for "extractelement <2 x i32> c, i32 0" the index
4792 unsigned getTransitionIdx() const {
4793 assert(isa<ExtractElementInst>(Transition) &&
4794 "Other kind of transitions are not supported yet");
4798 /// \brief Get the type of the transition.
4799 /// This is the type of the original value.
4800 /// E.g., for "extractelement <2 x i32> c, i32 1" the type of the
4801 /// transition is <2 x i32>.
4802 Type *getTransitionType() const {
4803 return Transition->getOperand(getTransitionOriginalValueIdx())->getType();
4806 /// \brief Promote \p ToBePromoted by moving \p Def downward through.
4807 /// I.e., we have the following sequence:
4808 /// Def = Transition <ty1> a to <ty2>
4809 /// b = ToBePromoted <ty2> Def, ...
4811 /// b = ToBePromoted <ty1> a, ...
4812 /// Def = Transition <ty1> ToBePromoted to <ty2>
4813 void promoteImpl(Instruction *ToBePromoted);
4815 /// \brief Check whether or not it is profitable to promote all the
4816 /// instructions enqueued to be promoted.
4817 bool isProfitableToPromote() {
4818 Value *ValIdx = Transition->getOperand(getTransitionOriginalValueIdx());
4819 unsigned Index = isa<ConstantInt>(ValIdx)
4820 ? cast<ConstantInt>(ValIdx)->getZExtValue()
4822 Type *PromotedType = getTransitionType();
4824 StoreInst *ST = cast<StoreInst>(CombineInst);
4825 unsigned AS = ST->getPointerAddressSpace();
4826 unsigned Align = ST->getAlignment();
4827 // Check if this store is supported.
4828 if (!TLI.allowsMisalignedMemoryAccesses(
4829 TLI.getValueType(DL, ST->getValueOperand()->getType()), AS,
4831 // If this is not supported, there is no way we can combine
4832 // the extract with the store.
4836 // The scalar chain of computation has to pay for the transition
4837 // scalar to vector.
4838 // The vector chain has to account for the combining cost.
4839 uint64_t ScalarCost =
4840 TTI.getVectorInstrCost(Transition->getOpcode(), PromotedType, Index);
4841 uint64_t VectorCost = StoreExtractCombineCost;
4842 for (const auto &Inst : InstsToBePromoted) {
4843 // Compute the cost.
4844 // By construction, all instructions being promoted are arithmetic ones.
4845 // Moreover, one argument is a constant that can be viewed as a splat
4847 Value *Arg0 = Inst->getOperand(0);
4848 bool IsArg0Constant = isa<UndefValue>(Arg0) || isa<ConstantInt>(Arg0) ||
4849 isa<ConstantFP>(Arg0);
4850 TargetTransformInfo::OperandValueKind Arg0OVK =
4851 IsArg0Constant ? TargetTransformInfo::OK_UniformConstantValue
4852 : TargetTransformInfo::OK_AnyValue;
4853 TargetTransformInfo::OperandValueKind Arg1OVK =
4854 !IsArg0Constant ? TargetTransformInfo::OK_UniformConstantValue
4855 : TargetTransformInfo::OK_AnyValue;
4856 ScalarCost += TTI.getArithmeticInstrCost(
4857 Inst->getOpcode(), Inst->getType(), Arg0OVK, Arg1OVK);
4858 VectorCost += TTI.getArithmeticInstrCost(Inst->getOpcode(), PromotedType,
4861 DEBUG(dbgs() << "Estimated cost of computation to be promoted:\nScalar: "
4862 << ScalarCost << "\nVector: " << VectorCost << '\n');
4863 return ScalarCost > VectorCost;
4866 /// \brief Generate a constant vector with \p Val with the same
4867 /// number of elements as the transition.
4868 /// \p UseSplat defines whether or not \p Val should be replicated
4869 /// across the whole vector.
4870 /// In other words, if UseSplat == true, we generate <Val, Val, ..., Val>,
4871 /// otherwise we generate a vector with as many undef as possible:
4872 /// <undef, ..., undef, Val, undef, ..., undef> where \p Val is only
4873 /// used at the index of the extract.
4874 Value *getConstantVector(Constant *Val, bool UseSplat) const {
4875 unsigned ExtractIdx = UINT_MAX;
4877 // If we cannot determine where the constant must be, we have to
4878 // use a splat constant.
4879 Value *ValExtractIdx = Transition->getOperand(getTransitionIdx());
4880 if (ConstantInt *CstVal = dyn_cast<ConstantInt>(ValExtractIdx))
4881 ExtractIdx = CstVal->getSExtValue();
4886 unsigned End = getTransitionType()->getVectorNumElements();
4888 return ConstantVector::getSplat(End, Val);
4890 SmallVector<Constant *, 4> ConstVec;
4891 UndefValue *UndefVal = UndefValue::get(Val->getType());
4892 for (unsigned Idx = 0; Idx != End; ++Idx) {
4893 if (Idx == ExtractIdx)
4894 ConstVec.push_back(Val);
4896 ConstVec.push_back(UndefVal);
4898 return ConstantVector::get(ConstVec);
4901 /// \brief Check if promoting to a vector type an operand at \p OperandIdx
4902 /// in \p Use can trigger undefined behavior.
4903 static bool canCauseUndefinedBehavior(const Instruction *Use,
4904 unsigned OperandIdx) {
4905 // This is not safe to introduce undef when the operand is on
4906 // the right hand side of a division-like instruction.
4907 if (OperandIdx != 1)
4909 switch (Use->getOpcode()) {
4912 case Instruction::SDiv:
4913 case Instruction::UDiv:
4914 case Instruction::SRem:
4915 case Instruction::URem:
4917 case Instruction::FDiv:
4918 case Instruction::FRem:
4919 return !Use->hasNoNaNs();
4921 llvm_unreachable(nullptr);
4925 VectorPromoteHelper(const DataLayout &DL, const TargetLowering &TLI,
4926 const TargetTransformInfo &TTI, Instruction *Transition,
4927 unsigned CombineCost)
4928 : DL(DL), TLI(TLI), TTI(TTI), Transition(Transition),
4929 StoreExtractCombineCost(CombineCost), CombineInst(nullptr) {
4930 assert(Transition && "Do not know how to promote null");
4933 /// \brief Check if we can promote \p ToBePromoted to \p Type.
4934 bool canPromote(const Instruction *ToBePromoted) const {
4935 // We could support CastInst too.
4936 return isa<BinaryOperator>(ToBePromoted);
4939 /// \brief Check if it is profitable to promote \p ToBePromoted
4940 /// by moving downward the transition through.
4941 bool shouldPromote(const Instruction *ToBePromoted) const {
4942 // Promote only if all the operands can be statically expanded.
4943 // Indeed, we do not want to introduce any new kind of transitions.
4944 for (const Use &U : ToBePromoted->operands()) {
4945 const Value *Val = U.get();
4946 if (Val == getEndOfTransition()) {
4947 // If the use is a division and the transition is on the rhs,
4948 // we cannot promote the operation, otherwise we may create a
4949 // division by zero.
4950 if (canCauseUndefinedBehavior(ToBePromoted, U.getOperandNo()))
4954 if (!isa<ConstantInt>(Val) && !isa<UndefValue>(Val) &&
4955 !isa<ConstantFP>(Val))
4958 // Check that the resulting operation is legal.
4959 int ISDOpcode = TLI.InstructionOpcodeToISD(ToBePromoted->getOpcode());
4962 return StressStoreExtract ||
4963 TLI.isOperationLegalOrCustom(
4964 ISDOpcode, TLI.getValueType(DL, getTransitionType(), true));
4967 /// \brief Check whether or not \p Use can be combined
4968 /// with the transition.
4969 /// I.e., is it possible to do Use(Transition) => AnotherUse?
4970 bool canCombine(const Instruction *Use) { return isa<StoreInst>(Use); }
4972 /// \brief Record \p ToBePromoted as part of the chain to be promoted.
4973 void enqueueForPromotion(Instruction *ToBePromoted) {
4974 InstsToBePromoted.push_back(ToBePromoted);
4977 /// \brief Set the instruction that will be combined with the transition.
4978 void recordCombineInstruction(Instruction *ToBeCombined) {
4979 assert(canCombine(ToBeCombined) && "Unsupported instruction to combine");
4980 CombineInst = ToBeCombined;
4983 /// \brief Promote all the instructions enqueued for promotion if it is
4985 /// \return True if the promotion happened, false otherwise.
4987 // Check if there is something to promote.
4988 // Right now, if we do not have anything to combine with,
4989 // we assume the promotion is not profitable.
4990 if (InstsToBePromoted.empty() || !CombineInst)
4994 if (!StressStoreExtract && !isProfitableToPromote())
4998 for (auto &ToBePromoted : InstsToBePromoted)
4999 promoteImpl(ToBePromoted);
5000 InstsToBePromoted.clear();
5004 } // End of anonymous namespace.
5006 void VectorPromoteHelper::promoteImpl(Instruction *ToBePromoted) {
5007 // At this point, we know that all the operands of ToBePromoted but Def
5008 // can be statically promoted.
5009 // For Def, we need to use its parameter in ToBePromoted:
5010 // b = ToBePromoted ty1 a
5011 // Def = Transition ty1 b to ty2
5012 // Move the transition down.
5013 // 1. Replace all uses of the promoted operation by the transition.
5014 // = ... b => = ... Def.
5015 assert(ToBePromoted->getType() == Transition->getType() &&
5016 "The type of the result of the transition does not match "
5018 ToBePromoted->replaceAllUsesWith(Transition);
5019 // 2. Update the type of the uses.
5020 // b = ToBePromoted ty2 Def => b = ToBePromoted ty1 Def.
5021 Type *TransitionTy = getTransitionType();
5022 ToBePromoted->mutateType(TransitionTy);
5023 // 3. Update all the operands of the promoted operation with promoted
5025 // b = ToBePromoted ty1 Def => b = ToBePromoted ty1 a.
5026 for (Use &U : ToBePromoted->operands()) {
5027 Value *Val = U.get();
5028 Value *NewVal = nullptr;
5029 if (Val == Transition)
5030 NewVal = Transition->getOperand(getTransitionOriginalValueIdx());
5031 else if (isa<UndefValue>(Val) || isa<ConstantInt>(Val) ||
5032 isa<ConstantFP>(Val)) {
5033 // Use a splat constant if it is not safe to use undef.
5034 NewVal = getConstantVector(
5035 cast<Constant>(Val),
5036 isa<UndefValue>(Val) ||
5037 canCauseUndefinedBehavior(ToBePromoted, U.getOperandNo()));
5039 llvm_unreachable("Did you modified shouldPromote and forgot to update "
5041 ToBePromoted->setOperand(U.getOperandNo(), NewVal);
5043 Transition->removeFromParent();
5044 Transition->insertAfter(ToBePromoted);
5045 Transition->setOperand(getTransitionOriginalValueIdx(), ToBePromoted);
5048 /// Some targets can do store(extractelement) with one instruction.
5049 /// Try to push the extractelement towards the stores when the target
5050 /// has this feature and this is profitable.
5051 bool CodeGenPrepare::optimizeExtractElementInst(Instruction *Inst) {
5052 unsigned CombineCost = UINT_MAX;
5053 if (DisableStoreExtract || !TLI ||
5054 (!StressStoreExtract &&
5055 !TLI->canCombineStoreAndExtract(Inst->getOperand(0)->getType(),
5056 Inst->getOperand(1), CombineCost)))
5059 // At this point we know that Inst is a vector to scalar transition.
5060 // Try to move it down the def-use chain, until:
5061 // - We can combine the transition with its single use
5062 // => we got rid of the transition.
5063 // - We escape the current basic block
5064 // => we would need to check that we are moving it at a cheaper place and
5065 // we do not do that for now.
5066 BasicBlock *Parent = Inst->getParent();
5067 DEBUG(dbgs() << "Found an interesting transition: " << *Inst << '\n');
5068 VectorPromoteHelper VPH(*DL, *TLI, *TTI, Inst, CombineCost);
5069 // If the transition has more than one use, assume this is not going to be
5071 while (Inst->hasOneUse()) {
5072 Instruction *ToBePromoted = cast<Instruction>(*Inst->user_begin());
5073 DEBUG(dbgs() << "Use: " << *ToBePromoted << '\n');
5075 if (ToBePromoted->getParent() != Parent) {
5076 DEBUG(dbgs() << "Instruction to promote is in a different block ("
5077 << ToBePromoted->getParent()->getName()
5078 << ") than the transition (" << Parent->getName() << ").\n");
5082 if (VPH.canCombine(ToBePromoted)) {
5083 DEBUG(dbgs() << "Assume " << *Inst << '\n'
5084 << "will be combined with: " << *ToBePromoted << '\n');
5085 VPH.recordCombineInstruction(ToBePromoted);
5086 bool Changed = VPH.promote();
5087 NumStoreExtractExposed += Changed;
5091 DEBUG(dbgs() << "Try promoting.\n");
5092 if (!VPH.canPromote(ToBePromoted) || !VPH.shouldPromote(ToBePromoted))
5095 DEBUG(dbgs() << "Promoting is possible... Enqueue for promotion!\n");
5097 VPH.enqueueForPromotion(ToBePromoted);
5098 Inst = ToBePromoted;
5103 bool CodeGenPrepare::optimizeInst(Instruction *I, bool& ModifiedDT) {
5104 // Bail out if we inserted the instruction to prevent optimizations from
5105 // stepping on each other's toes.
5106 if (InsertedInsts.count(I))
5109 if (PHINode *P = dyn_cast<PHINode>(I)) {
5110 // It is possible for very late stage optimizations (such as SimplifyCFG)
5111 // to introduce PHI nodes too late to be cleaned up. If we detect such a
5112 // trivial PHI, go ahead and zap it here.
5113 if (Value *V = SimplifyInstruction(P, *DL, TLInfo, nullptr)) {
5114 P->replaceAllUsesWith(V);
5115 P->eraseFromParent();
5122 if (CastInst *CI = dyn_cast<CastInst>(I)) {
5123 // If the source of the cast is a constant, then this should have
5124 // already been constant folded. The only reason NOT to constant fold
5125 // it is if something (e.g. LSR) was careful to place the constant
5126 // evaluation in a block other than then one that uses it (e.g. to hoist
5127 // the address of globals out of a loop). If this is the case, we don't
5128 // want to forward-subst the cast.
5129 if (isa<Constant>(CI->getOperand(0)))
5132 if (TLI && OptimizeNoopCopyExpression(CI, *TLI, *DL))
5135 if (isa<ZExtInst>(I) || isa<SExtInst>(I)) {
5136 /// Sink a zext or sext into its user blocks if the target type doesn't
5137 /// fit in one register
5139 TLI->getTypeAction(CI->getContext(),
5140 TLI->getValueType(*DL, CI->getType())) ==
5141 TargetLowering::TypeExpandInteger) {
5142 return SinkCast(CI);
5144 bool MadeChange = moveExtToFormExtLoad(I);
5145 return MadeChange | optimizeExtUses(I);
5151 if (CmpInst *CI = dyn_cast<CmpInst>(I))
5152 if (!TLI || !TLI->hasMultipleConditionRegisters())
5153 return OptimizeCmpExpression(CI);
5155 if (LoadInst *LI = dyn_cast<LoadInst>(I)) {
5156 stripInvariantGroupMetadata(*LI);
5158 bool Modified = optimizeLoadExt(LI);
5159 unsigned AS = LI->getPointerAddressSpace();
5160 Modified |= optimizeMemoryInst(I, I->getOperand(0), LI->getType(), AS);
5166 if (StoreInst *SI = dyn_cast<StoreInst>(I)) {
5167 stripInvariantGroupMetadata(*SI);
5169 unsigned AS = SI->getPointerAddressSpace();
5170 return optimizeMemoryInst(I, SI->getOperand(1),
5171 SI->getOperand(0)->getType(), AS);
5176 BinaryOperator *BinOp = dyn_cast<BinaryOperator>(I);
5178 if (BinOp && (BinOp->getOpcode() == Instruction::AShr ||
5179 BinOp->getOpcode() == Instruction::LShr)) {
5180 ConstantInt *CI = dyn_cast<ConstantInt>(BinOp->getOperand(1));
5181 if (TLI && CI && TLI->hasExtractBitsInsn())
5182 return OptimizeExtractBits(BinOp, CI, *TLI, *DL);
5187 if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(I)) {
5188 if (GEPI->hasAllZeroIndices()) {
5189 /// The GEP operand must be a pointer, so must its result -> BitCast
5190 Instruction *NC = new BitCastInst(GEPI->getOperand(0), GEPI->getType(),
5191 GEPI->getName(), GEPI);
5192 GEPI->replaceAllUsesWith(NC);
5193 GEPI->eraseFromParent();
5195 optimizeInst(NC, ModifiedDT);
5201 if (CallInst *CI = dyn_cast<CallInst>(I))
5202 return optimizeCallInst(CI, ModifiedDT);
5204 if (SelectInst *SI = dyn_cast<SelectInst>(I))
5205 return optimizeSelectInst(SI);
5207 if (ShuffleVectorInst *SVI = dyn_cast<ShuffleVectorInst>(I))
5208 return optimizeShuffleVectorInst(SVI);
5210 if (auto *Switch = dyn_cast<SwitchInst>(I))
5211 return optimizeSwitchInst(Switch);
5213 if (isa<ExtractElementInst>(I))
5214 return optimizeExtractElementInst(I);
5219 // In this pass we look for GEP and cast instructions that are used
5220 // across basic blocks and rewrite them to improve basic-block-at-a-time
5222 bool CodeGenPrepare::optimizeBlock(BasicBlock &BB, bool& ModifiedDT) {
5224 bool MadeChange = false;
5226 CurInstIterator = BB.begin();
5227 while (CurInstIterator != BB.end()) {
5228 MadeChange |= optimizeInst(&*CurInstIterator++, ModifiedDT);
5232 MadeChange |= dupRetToEnableTailCallOpts(&BB);
5237 // llvm.dbg.value is far away from the value then iSel may not be able
5238 // handle it properly. iSel will drop llvm.dbg.value if it can not
5239 // find a node corresponding to the value.
5240 bool CodeGenPrepare::placeDbgValues(Function &F) {
5241 bool MadeChange = false;
5242 for (BasicBlock &BB : F) {
5243 Instruction *PrevNonDbgInst = nullptr;
5244 for (BasicBlock::iterator BI = BB.begin(), BE = BB.end(); BI != BE;) {
5245 Instruction *Insn = &*BI++;
5246 DbgValueInst *DVI = dyn_cast<DbgValueInst>(Insn);
5247 // Leave dbg.values that refer to an alloca alone. These
5248 // instrinsics describe the address of a variable (= the alloca)
5249 // being taken. They should not be moved next to the alloca
5250 // (and to the beginning of the scope), but rather stay close to
5251 // where said address is used.
5252 if (!DVI || (DVI->getValue() && isa<AllocaInst>(DVI->getValue()))) {
5253 PrevNonDbgInst = Insn;
5257 Instruction *VI = dyn_cast_or_null<Instruction>(DVI->getValue());
5258 if (VI && VI != PrevNonDbgInst && !VI->isTerminator()) {
5259 // If VI is a phi in a block with an EHPad terminator, we can't insert
5261 if (isa<PHINode>(VI) && VI->getParent()->getTerminator()->isEHPad())
5263 DEBUG(dbgs() << "Moving Debug Value before :\n" << *DVI << ' ' << *VI);
5264 DVI->removeFromParent();
5265 if (isa<PHINode>(VI))
5266 DVI->insertBefore(&*VI->getParent()->getFirstInsertionPt());
5268 DVI->insertAfter(VI);
5277 // If there is a sequence that branches based on comparing a single bit
5278 // against zero that can be combined into a single instruction, and the
5279 // target supports folding these into a single instruction, sink the
5280 // mask and compare into the branch uses. Do this before OptimizeBlock ->
5281 // OptimizeInst -> OptimizeCmpExpression, which perturbs the pattern being
5283 bool CodeGenPrepare::sinkAndCmp(Function &F) {
5284 if (!EnableAndCmpSinking)
5286 if (!TLI || !TLI->isMaskAndBranchFoldingLegal())
5288 bool MadeChange = false;
5289 for (Function::iterator I = F.begin(), E = F.end(); I != E; ) {
5290 BasicBlock *BB = &*I++;
5292 // Does this BB end with the following?
5293 // %andVal = and %val, #single-bit-set
5294 // %icmpVal = icmp %andResult, 0
5295 // br i1 %cmpVal label %dest1, label %dest2"
5296 BranchInst *Brcc = dyn_cast<BranchInst>(BB->getTerminator());
5297 if (!Brcc || !Brcc->isConditional())
5299 ICmpInst *Cmp = dyn_cast<ICmpInst>(Brcc->getOperand(0));
5300 if (!Cmp || Cmp->getParent() != BB)
5302 ConstantInt *Zero = dyn_cast<ConstantInt>(Cmp->getOperand(1));
5303 if (!Zero || !Zero->isZero())
5305 Instruction *And = dyn_cast<Instruction>(Cmp->getOperand(0));
5306 if (!And || And->getOpcode() != Instruction::And || And->getParent() != BB)
5308 ConstantInt* Mask = dyn_cast<ConstantInt>(And->getOperand(1));
5309 if (!Mask || !Mask->getUniqueInteger().isPowerOf2())
5311 DEBUG(dbgs() << "found and; icmp ?,0; brcc\n"); DEBUG(BB->dump());
5313 // Push the "and; icmp" for any users that are conditional branches.
5314 // Since there can only be one branch use per BB, we don't need to keep
5315 // track of which BBs we insert into.
5316 for (Value::use_iterator UI = Cmp->use_begin(), E = Cmp->use_end();
5320 BranchInst *BrccUser = dyn_cast<BranchInst>(*UI);
5322 if (!BrccUser || !BrccUser->isConditional())
5324 BasicBlock *UserBB = BrccUser->getParent();
5325 if (UserBB == BB) continue;
5326 DEBUG(dbgs() << "found Brcc use\n");
5328 // Sink the "and; icmp" to use.
5330 BinaryOperator *NewAnd =
5331 BinaryOperator::CreateAnd(And->getOperand(0), And->getOperand(1), "",
5334 CmpInst::Create(Cmp->getOpcode(), Cmp->getPredicate(), NewAnd, Zero,
5338 DEBUG(BrccUser->getParent()->dump());
5344 /// \brief Retrieve the probabilities of a conditional branch. Returns true on
5345 /// success, or returns false if no or invalid metadata was found.
5346 static bool extractBranchMetadata(BranchInst *BI,
5347 uint64_t &ProbTrue, uint64_t &ProbFalse) {
5348 assert(BI->isConditional() &&
5349 "Looking for probabilities on unconditional branch?");
5350 auto *ProfileData = BI->getMetadata(LLVMContext::MD_prof);
5351 if (!ProfileData || ProfileData->getNumOperands() != 3)
5354 const auto *CITrue =
5355 mdconst::dyn_extract<ConstantInt>(ProfileData->getOperand(1));
5356 const auto *CIFalse =
5357 mdconst::dyn_extract<ConstantInt>(ProfileData->getOperand(2));
5358 if (!CITrue || !CIFalse)
5361 ProbTrue = CITrue->getValue().getZExtValue();
5362 ProbFalse = CIFalse->getValue().getZExtValue();
5367 /// \brief Scale down both weights to fit into uint32_t.
5368 static void scaleWeights(uint64_t &NewTrue, uint64_t &NewFalse) {
5369 uint64_t NewMax = (NewTrue > NewFalse) ? NewTrue : NewFalse;
5370 uint32_t Scale = (NewMax / UINT32_MAX) + 1;
5371 NewTrue = NewTrue / Scale;
5372 NewFalse = NewFalse / Scale;
5375 /// \brief Some targets prefer to split a conditional branch like:
5377 /// %0 = icmp ne i32 %a, 0
5378 /// %1 = icmp ne i32 %b, 0
5379 /// %or.cond = or i1 %0, %1
5380 /// br i1 %or.cond, label %TrueBB, label %FalseBB
5382 /// into multiple branch instructions like:
5385 /// %0 = icmp ne i32 %a, 0
5386 /// br i1 %0, label %TrueBB, label %bb2
5388 /// %1 = icmp ne i32 %b, 0
5389 /// br i1 %1, label %TrueBB, label %FalseBB
5391 /// This usually allows instruction selection to do even further optimizations
5392 /// and combine the compare with the branch instruction. Currently this is
5393 /// applied for targets which have "cheap" jump instructions.
5395 /// FIXME: Remove the (equivalent?) implementation in SelectionDAG.
5397 bool CodeGenPrepare::splitBranchCondition(Function &F) {
5398 if (!TM || !TM->Options.EnableFastISel || !TLI || TLI->isJumpExpensive())
5401 bool MadeChange = false;
5402 for (auto &BB : F) {
5403 // Does this BB end with the following?
5404 // %cond1 = icmp|fcmp|binary instruction ...
5405 // %cond2 = icmp|fcmp|binary instruction ...
5406 // %cond.or = or|and i1 %cond1, cond2
5407 // br i1 %cond.or label %dest1, label %dest2"
5408 BinaryOperator *LogicOp;
5409 BasicBlock *TBB, *FBB;
5410 if (!match(BB.getTerminator(), m_Br(m_OneUse(m_BinOp(LogicOp)), TBB, FBB)))
5413 auto *Br1 = cast<BranchInst>(BB.getTerminator());
5414 if (Br1->getMetadata(LLVMContext::MD_unpredictable))
5418 Value *Cond1, *Cond2;
5419 if (match(LogicOp, m_And(m_OneUse(m_Value(Cond1)),
5420 m_OneUse(m_Value(Cond2)))))
5421 Opc = Instruction::And;
5422 else if (match(LogicOp, m_Or(m_OneUse(m_Value(Cond1)),
5423 m_OneUse(m_Value(Cond2)))))
5424 Opc = Instruction::Or;
5428 if (!match(Cond1, m_CombineOr(m_Cmp(), m_BinOp())) ||
5429 !match(Cond2, m_CombineOr(m_Cmp(), m_BinOp())) )
5432 DEBUG(dbgs() << "Before branch condition splitting\n"; BB.dump());
5435 auto *InsertBefore = std::next(Function::iterator(BB))
5436 .getNodePtrUnchecked();
5437 auto TmpBB = BasicBlock::Create(BB.getContext(),
5438 BB.getName() + ".cond.split",
5439 BB.getParent(), InsertBefore);
5441 // Update original basic block by using the first condition directly by the
5442 // branch instruction and removing the no longer needed and/or instruction.
5443 Br1->setCondition(Cond1);
5444 LogicOp->eraseFromParent();
5446 // Depending on the conditon we have to either replace the true or the false
5447 // successor of the original branch instruction.
5448 if (Opc == Instruction::And)
5449 Br1->setSuccessor(0, TmpBB);
5451 Br1->setSuccessor(1, TmpBB);
5453 // Fill in the new basic block.
5454 auto *Br2 = IRBuilder<>(TmpBB).CreateCondBr(Cond2, TBB, FBB);
5455 if (auto *I = dyn_cast<Instruction>(Cond2)) {
5456 I->removeFromParent();
5457 I->insertBefore(Br2);
5460 // Update PHI nodes in both successors. The original BB needs to be
5461 // replaced in one succesor's PHI nodes, because the branch comes now from
5462 // the newly generated BB (NewBB). In the other successor we need to add one
5463 // incoming edge to the PHI nodes, because both branch instructions target
5464 // now the same successor. Depending on the original branch condition
5465 // (and/or) we have to swap the successors (TrueDest, FalseDest), so that
5466 // we perfrom the correct update for the PHI nodes.
5467 // This doesn't change the successor order of the just created branch
5468 // instruction (or any other instruction).
5469 if (Opc == Instruction::Or)
5470 std::swap(TBB, FBB);
5472 // Replace the old BB with the new BB.
5473 for (auto &I : *TBB) {
5474 PHINode *PN = dyn_cast<PHINode>(&I);
5478 while ((i = PN->getBasicBlockIndex(&BB)) >= 0)
5479 PN->setIncomingBlock(i, TmpBB);
5482 // Add another incoming edge form the new BB.
5483 for (auto &I : *FBB) {
5484 PHINode *PN = dyn_cast<PHINode>(&I);
5487 auto *Val = PN->getIncomingValueForBlock(&BB);
5488 PN->addIncoming(Val, TmpBB);
5491 // Update the branch weights (from SelectionDAGBuilder::
5492 // FindMergedConditions).
5493 if (Opc == Instruction::Or) {
5494 // Codegen X | Y as:
5503 // We have flexibility in setting Prob for BB1 and Prob for NewBB.
5504 // The requirement is that
5505 // TrueProb for BB1 + (FalseProb for BB1 * TrueProb for TmpBB)
5506 // = TrueProb for orignal BB.
5507 // Assuming the orignal weights are A and B, one choice is to set BB1's
5508 // weights to A and A+2B, and set TmpBB's weights to A and 2B. This choice
5510 // TrueProb for BB1 == FalseProb for BB1 * TrueProb for TmpBB.
5511 // Another choice is to assume TrueProb for BB1 equals to TrueProb for
5512 // TmpBB, but the math is more complicated.
5513 uint64_t TrueWeight, FalseWeight;
5514 if (extractBranchMetadata(Br1, TrueWeight, FalseWeight)) {
5515 uint64_t NewTrueWeight = TrueWeight;
5516 uint64_t NewFalseWeight = TrueWeight + 2 * FalseWeight;
5517 scaleWeights(NewTrueWeight, NewFalseWeight);
5518 Br1->setMetadata(LLVMContext::MD_prof, MDBuilder(Br1->getContext())
5519 .createBranchWeights(TrueWeight, FalseWeight));
5521 NewTrueWeight = TrueWeight;
5522 NewFalseWeight = 2 * FalseWeight;
5523 scaleWeights(NewTrueWeight, NewFalseWeight);
5524 Br2->setMetadata(LLVMContext::MD_prof, MDBuilder(Br2->getContext())
5525 .createBranchWeights(TrueWeight, FalseWeight));
5528 // Codegen X & Y as:
5536 // This requires creation of TmpBB after CurBB.
5538 // We have flexibility in setting Prob for BB1 and Prob for TmpBB.
5539 // The requirement is that
5540 // FalseProb for BB1 + (TrueProb for BB1 * FalseProb for TmpBB)
5541 // = FalseProb for orignal BB.
5542 // Assuming the orignal weights are A and B, one choice is to set BB1's
5543 // weights to 2A+B and B, and set TmpBB's weights to 2A and B. This choice
5545 // FalseProb for BB1 == TrueProb for BB1 * FalseProb for TmpBB.
5546 uint64_t TrueWeight, FalseWeight;
5547 if (extractBranchMetadata(Br1, TrueWeight, FalseWeight)) {
5548 uint64_t NewTrueWeight = 2 * TrueWeight + FalseWeight;
5549 uint64_t NewFalseWeight = FalseWeight;
5550 scaleWeights(NewTrueWeight, NewFalseWeight);
5551 Br1->setMetadata(LLVMContext::MD_prof, MDBuilder(Br1->getContext())
5552 .createBranchWeights(TrueWeight, FalseWeight));
5554 NewTrueWeight = 2 * TrueWeight;
5555 NewFalseWeight = FalseWeight;
5556 scaleWeights(NewTrueWeight, NewFalseWeight);
5557 Br2->setMetadata(LLVMContext::MD_prof, MDBuilder(Br2->getContext())
5558 .createBranchWeights(TrueWeight, FalseWeight));
5562 // Note: No point in getting fancy here, since the DT info is never
5563 // available to CodeGenPrepare.
5568 DEBUG(dbgs() << "After branch condition splitting\n"; BB.dump();
5574 void CodeGenPrepare::stripInvariantGroupMetadata(Instruction &I) {
5575 if (auto *InvariantMD = I.getMetadata(LLVMContext::MD_invariant_group))
5576 I.dropUnknownNonDebugMetadata(InvariantMD->getMetadataID());