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(NumRetsDup, "Number of return instructions duplicated");
68 STATISTIC(NumDbgValueMoved, "Number of debug value instructions moved");
69 STATISTIC(NumSelectsExpanded, "Number of selects turned into branches");
70 STATISTIC(NumAndCmpsMoved, "Number of and/cmp's pushed into branches");
71 STATISTIC(NumStoreExtractExposed, "Number of store(extractelement) exposed");
73 static cl::opt<bool> DisableBranchOpts(
74 "disable-cgp-branch-opts", cl::Hidden, cl::init(false),
75 cl::desc("Disable branch optimizations in CodeGenPrepare"));
78 DisableGCOpts("disable-cgp-gc-opts", cl::Hidden, cl::init(false),
79 cl::desc("Disable GC optimizations in CodeGenPrepare"));
81 static cl::opt<bool> DisableSelectToBranch(
82 "disable-cgp-select2branch", cl::Hidden, cl::init(false),
83 cl::desc("Disable select to branch conversion."));
85 static cl::opt<bool> AddrSinkUsingGEPs(
86 "addr-sink-using-gep", cl::Hidden, cl::init(false),
87 cl::desc("Address sinking in CGP using GEPs."));
89 static cl::opt<bool> EnableAndCmpSinking(
90 "enable-andcmp-sinking", cl::Hidden, cl::init(true),
91 cl::desc("Enable sinkinig and/cmp into branches."));
93 static cl::opt<bool> DisableStoreExtract(
94 "disable-cgp-store-extract", cl::Hidden, cl::init(false),
95 cl::desc("Disable store(extract) optimizations in CodeGenPrepare"));
97 static cl::opt<bool> StressStoreExtract(
98 "stress-cgp-store-extract", cl::Hidden, cl::init(false),
99 cl::desc("Stress test store(extract) optimizations in CodeGenPrepare"));
101 static cl::opt<bool> DisableExtLdPromotion(
102 "disable-cgp-ext-ld-promotion", cl::Hidden, cl::init(false),
103 cl::desc("Disable ext(promotable(ld)) -> promoted(ext(ld)) optimization in "
106 static cl::opt<bool> StressExtLdPromotion(
107 "stress-cgp-ext-ld-promotion", cl::Hidden, cl::init(false),
108 cl::desc("Stress test ext(promotable(ld)) -> promoted(ext(ld)) "
109 "optimization in CodeGenPrepare"));
112 typedef SmallPtrSet<Instruction *, 16> SetOfInstrs;
113 typedef PointerIntPair<Type *, 1, bool> TypeIsSExt;
114 typedef DenseMap<Instruction *, TypeIsSExt> InstrToOrigTy;
115 class TypePromotionTransaction;
117 class CodeGenPrepare : public FunctionPass {
118 const TargetMachine *TM;
119 const TargetLowering *TLI;
120 const TargetTransformInfo *TTI;
121 const TargetLibraryInfo *TLInfo;
123 /// As we scan instructions optimizing them, this is the next instruction
124 /// to optimize. Transforms that can invalidate this should update it.
125 BasicBlock::iterator CurInstIterator;
127 /// Keeps track of non-local addresses that have been sunk into a block.
128 /// This allows us to avoid inserting duplicate code for blocks with
129 /// multiple load/stores of the same address.
130 ValueMap<Value*, Value*> SunkAddrs;
132 /// Keeps track of all instructions inserted for the current function.
133 SetOfInstrs InsertedInsts;
134 /// Keeps track of the type of the related instruction before their
135 /// promotion for the current function.
136 InstrToOrigTy PromotedInsts;
138 /// True if CFG is modified in any way.
141 /// True if optimizing for size.
144 /// DataLayout for the Function being processed.
145 const DataLayout *DL;
148 static char ID; // Pass identification, replacement for typeid
149 explicit CodeGenPrepare(const TargetMachine *TM = nullptr)
150 : FunctionPass(ID), TM(TM), TLI(nullptr), TTI(nullptr), DL(nullptr) {
151 initializeCodeGenPreparePass(*PassRegistry::getPassRegistry());
153 bool runOnFunction(Function &F) override;
155 const char *getPassName() const override { return "CodeGen Prepare"; }
157 void getAnalysisUsage(AnalysisUsage &AU) const override {
158 AU.addPreserved<DominatorTreeWrapperPass>();
159 AU.addRequired<TargetLibraryInfoWrapperPass>();
160 AU.addRequired<TargetTransformInfoWrapperPass>();
164 bool eliminateFallThrough(Function &F);
165 bool eliminateMostlyEmptyBlocks(Function &F);
166 bool canMergeBlocks(const BasicBlock *BB, const BasicBlock *DestBB) const;
167 void eliminateMostlyEmptyBlock(BasicBlock *BB);
168 bool optimizeBlock(BasicBlock &BB, bool& ModifiedDT);
169 bool optimizeInst(Instruction *I, bool& ModifiedDT);
170 bool optimizeMemoryInst(Instruction *I, Value *Addr,
171 Type *AccessTy, unsigned AS);
172 bool optimizeInlineAsmInst(CallInst *CS);
173 bool optimizeCallInst(CallInst *CI, bool& ModifiedDT);
174 bool moveExtToFormExtLoad(Instruction *&I);
175 bool optimizeExtUses(Instruction *I);
176 bool optimizeSelectInst(SelectInst *SI);
177 bool optimizeShuffleVectorInst(ShuffleVectorInst *SI);
178 bool optimizeSwitchInst(SwitchInst *CI);
179 bool optimizeExtractElementInst(Instruction *Inst);
180 bool dupRetToEnableTailCallOpts(BasicBlock *BB);
181 bool placeDbgValues(Function &F);
182 bool sinkAndCmp(Function &F);
183 bool extLdPromotion(TypePromotionTransaction &TPT, LoadInst *&LI,
185 const SmallVectorImpl<Instruction *> &Exts,
186 unsigned CreatedInstCost);
187 bool splitBranchCondition(Function &F);
188 bool simplifyOffsetableRelocate(Instruction &I);
189 void stripInvariantGroupMetadata(Instruction &I);
193 char CodeGenPrepare::ID = 0;
194 INITIALIZE_TM_PASS(CodeGenPrepare, "codegenprepare",
195 "Optimize for code generation", false, false)
197 FunctionPass *llvm::createCodeGenPreparePass(const TargetMachine *TM) {
198 return new CodeGenPrepare(TM);
201 bool CodeGenPrepare::runOnFunction(Function &F) {
202 if (skipOptnoneFunction(F))
205 DL = &F.getParent()->getDataLayout();
207 bool EverMadeChange = false;
208 // Clear per function information.
209 InsertedInsts.clear();
210 PromotedInsts.clear();
214 TLI = TM->getSubtargetImpl(F)->getTargetLowering();
215 TLInfo = &getAnalysis<TargetLibraryInfoWrapperPass>().getTLI();
216 TTI = &getAnalysis<TargetTransformInfoWrapperPass>().getTTI(F);
217 OptSize = F.optForSize();
219 /// This optimization identifies DIV instructions that can be
220 /// profitably bypassed and carried out with a shorter, faster divide.
221 if (!OptSize && TLI && TLI->isSlowDivBypassed()) {
222 const DenseMap<unsigned int, unsigned int> &BypassWidths =
223 TLI->getBypassSlowDivWidths();
224 for (Function::iterator I = F.begin(); I != F.end(); I++)
225 EverMadeChange |= bypassSlowDivision(F, I, BypassWidths);
228 // Eliminate blocks that contain only PHI nodes and an
229 // unconditional branch.
230 EverMadeChange |= eliminateMostlyEmptyBlocks(F);
232 // llvm.dbg.value is far away from the value then iSel may not be able
233 // handle it properly. iSel will drop llvm.dbg.value if it can not
234 // find a node corresponding to the value.
235 EverMadeChange |= placeDbgValues(F);
237 // If there is a mask, compare against zero, and branch that can be combined
238 // into a single target instruction, push the mask and compare into branch
239 // users. Do this before OptimizeBlock -> OptimizeInst ->
240 // OptimizeCmpExpression, which perturbs the pattern being searched for.
241 if (!DisableBranchOpts) {
242 EverMadeChange |= sinkAndCmp(F);
243 EverMadeChange |= splitBranchCondition(F);
246 bool MadeChange = true;
249 for (Function::iterator I = F.begin(); I != F.end(); ) {
250 BasicBlock *BB = &*I++;
251 bool ModifiedDTOnIteration = false;
252 MadeChange |= optimizeBlock(*BB, ModifiedDTOnIteration);
254 // Restart BB iteration if the dominator tree of the Function was changed
255 if (ModifiedDTOnIteration)
258 EverMadeChange |= MadeChange;
263 if (!DisableBranchOpts) {
265 SmallPtrSet<BasicBlock*, 8> WorkList;
266 for (BasicBlock &BB : F) {
267 SmallVector<BasicBlock *, 2> Successors(succ_begin(&BB), succ_end(&BB));
268 MadeChange |= ConstantFoldTerminator(&BB, true);
269 if (!MadeChange) continue;
271 for (SmallVectorImpl<BasicBlock*>::iterator
272 II = Successors.begin(), IE = Successors.end(); II != IE; ++II)
273 if (pred_begin(*II) == pred_end(*II))
274 WorkList.insert(*II);
277 // Delete the dead blocks and any of their dead successors.
278 MadeChange |= !WorkList.empty();
279 while (!WorkList.empty()) {
280 BasicBlock *BB = *WorkList.begin();
282 SmallVector<BasicBlock*, 2> Successors(succ_begin(BB), succ_end(BB));
286 for (SmallVectorImpl<BasicBlock*>::iterator
287 II = Successors.begin(), IE = Successors.end(); II != IE; ++II)
288 if (pred_begin(*II) == pred_end(*II))
289 WorkList.insert(*II);
292 // Merge pairs of basic blocks with unconditional branches, connected by
294 if (EverMadeChange || MadeChange)
295 MadeChange |= eliminateFallThrough(F);
297 EverMadeChange |= MadeChange;
300 if (!DisableGCOpts) {
301 SmallVector<Instruction *, 2> Statepoints;
302 for (BasicBlock &BB : F)
303 for (Instruction &I : BB)
305 Statepoints.push_back(&I);
306 for (auto &I : Statepoints)
307 EverMadeChange |= simplifyOffsetableRelocate(*I);
310 return EverMadeChange;
313 /// Merge basic blocks which are connected by a single edge, where one of the
314 /// basic blocks has a single successor pointing to the other basic block,
315 /// which has a single predecessor.
316 bool CodeGenPrepare::eliminateFallThrough(Function &F) {
317 bool Changed = false;
318 // Scan all of the blocks in the function, except for the entry block.
319 for (Function::iterator I = std::next(F.begin()), E = F.end(); I != E;) {
320 BasicBlock *BB = &*I++;
321 // If the destination block has a single pred, then this is a trivial
322 // edge, just collapse it.
323 BasicBlock *SinglePred = BB->getSinglePredecessor();
325 // Don't merge if BB's address is taken.
326 if (!SinglePred || SinglePred == BB || BB->hasAddressTaken()) continue;
328 BranchInst *Term = dyn_cast<BranchInst>(SinglePred->getTerminator());
329 if (Term && !Term->isConditional()) {
331 DEBUG(dbgs() << "To merge:\n"<< *SinglePred << "\n\n\n");
332 // Remember if SinglePred was the entry block of the function.
333 // If so, we will need to move BB back to the entry position.
334 bool isEntry = SinglePred == &SinglePred->getParent()->getEntryBlock();
335 MergeBasicBlockIntoOnlyPred(BB, nullptr);
337 if (isEntry && BB != &BB->getParent()->getEntryBlock())
338 BB->moveBefore(&BB->getParent()->getEntryBlock());
340 // We have erased a block. Update the iterator.
341 I = BB->getIterator();
347 /// Eliminate blocks that contain only PHI nodes, debug info directives, and an
348 /// unconditional branch. Passes before isel (e.g. LSR/loopsimplify) often split
349 /// edges in ways that are non-optimal for isel. Start by eliminating these
350 /// blocks so we can split them the way we want them.
351 bool CodeGenPrepare::eliminateMostlyEmptyBlocks(Function &F) {
352 bool MadeChange = false;
353 // Note that this intentionally skips the entry block.
354 for (Function::iterator I = std::next(F.begin()), E = F.end(); I != E;) {
355 BasicBlock *BB = &*I++;
357 // If this block doesn't end with an uncond branch, ignore it.
358 BranchInst *BI = dyn_cast<BranchInst>(BB->getTerminator());
359 if (!BI || !BI->isUnconditional())
362 // If the instruction before the branch (skipping debug info) isn't a phi
363 // node, then other stuff is happening here.
364 BasicBlock::iterator BBI = BI->getIterator();
365 if (BBI != BB->begin()) {
367 while (isa<DbgInfoIntrinsic>(BBI)) {
368 if (BBI == BB->begin())
372 if (!isa<DbgInfoIntrinsic>(BBI) && !isa<PHINode>(BBI))
376 // Do not break infinite loops.
377 BasicBlock *DestBB = BI->getSuccessor(0);
381 if (!canMergeBlocks(BB, DestBB))
384 eliminateMostlyEmptyBlock(BB);
390 /// Return true if we can merge BB into DestBB if there is a single
391 /// unconditional branch between them, and BB contains no other non-phi
393 bool CodeGenPrepare::canMergeBlocks(const BasicBlock *BB,
394 const BasicBlock *DestBB) const {
395 // We only want to eliminate blocks whose phi nodes are used by phi nodes in
396 // the successor. If there are more complex condition (e.g. preheaders),
397 // don't mess around with them.
398 BasicBlock::const_iterator BBI = BB->begin();
399 while (const PHINode *PN = dyn_cast<PHINode>(BBI++)) {
400 for (const User *U : PN->users()) {
401 const Instruction *UI = cast<Instruction>(U);
402 if (UI->getParent() != DestBB || !isa<PHINode>(UI))
404 // If User is inside DestBB block and it is a PHINode then check
405 // incoming value. If incoming value is not from BB then this is
406 // a complex condition (e.g. preheaders) we want to avoid here.
407 if (UI->getParent() == DestBB) {
408 if (const PHINode *UPN = dyn_cast<PHINode>(UI))
409 for (unsigned I = 0, E = UPN->getNumIncomingValues(); I != E; ++I) {
410 Instruction *Insn = dyn_cast<Instruction>(UPN->getIncomingValue(I));
411 if (Insn && Insn->getParent() == BB &&
412 Insn->getParent() != UPN->getIncomingBlock(I))
419 // If BB and DestBB contain any common predecessors, then the phi nodes in BB
420 // and DestBB may have conflicting incoming values for the block. If so, we
421 // can't merge the block.
422 const PHINode *DestBBPN = dyn_cast<PHINode>(DestBB->begin());
423 if (!DestBBPN) return true; // no conflict.
425 // Collect the preds of BB.
426 SmallPtrSet<const BasicBlock*, 16> BBPreds;
427 if (const PHINode *BBPN = dyn_cast<PHINode>(BB->begin())) {
428 // It is faster to get preds from a PHI than with pred_iterator.
429 for (unsigned i = 0, e = BBPN->getNumIncomingValues(); i != e; ++i)
430 BBPreds.insert(BBPN->getIncomingBlock(i));
432 BBPreds.insert(pred_begin(BB), pred_end(BB));
435 // Walk the preds of DestBB.
436 for (unsigned i = 0, e = DestBBPN->getNumIncomingValues(); i != e; ++i) {
437 BasicBlock *Pred = DestBBPN->getIncomingBlock(i);
438 if (BBPreds.count(Pred)) { // Common predecessor?
439 BBI = DestBB->begin();
440 while (const PHINode *PN = dyn_cast<PHINode>(BBI++)) {
441 const Value *V1 = PN->getIncomingValueForBlock(Pred);
442 const Value *V2 = PN->getIncomingValueForBlock(BB);
444 // If V2 is a phi node in BB, look up what the mapped value will be.
445 if (const PHINode *V2PN = dyn_cast<PHINode>(V2))
446 if (V2PN->getParent() == BB)
447 V2 = V2PN->getIncomingValueForBlock(Pred);
449 // If there is a conflict, bail out.
450 if (V1 != V2) return false;
459 /// Eliminate a basic block that has only phi's and an unconditional branch in
461 void CodeGenPrepare::eliminateMostlyEmptyBlock(BasicBlock *BB) {
462 BranchInst *BI = cast<BranchInst>(BB->getTerminator());
463 BasicBlock *DestBB = BI->getSuccessor(0);
465 DEBUG(dbgs() << "MERGING MOSTLY EMPTY BLOCKS - BEFORE:\n" << *BB << *DestBB);
467 // If the destination block has a single pred, then this is a trivial edge,
469 if (BasicBlock *SinglePred = DestBB->getSinglePredecessor()) {
470 if (SinglePred != DestBB) {
471 // Remember if SinglePred was the entry block of the function. If so, we
472 // will need to move BB back to the entry position.
473 bool isEntry = SinglePred == &SinglePred->getParent()->getEntryBlock();
474 MergeBasicBlockIntoOnlyPred(DestBB, nullptr);
476 if (isEntry && BB != &BB->getParent()->getEntryBlock())
477 BB->moveBefore(&BB->getParent()->getEntryBlock());
479 DEBUG(dbgs() << "AFTER:\n" << *DestBB << "\n\n\n");
484 // Otherwise, we have multiple predecessors of BB. Update the PHIs in DestBB
485 // to handle the new incoming edges it is about to have.
487 for (BasicBlock::iterator BBI = DestBB->begin();
488 (PN = dyn_cast<PHINode>(BBI)); ++BBI) {
489 // Remove the incoming value for BB, and remember it.
490 Value *InVal = PN->removeIncomingValue(BB, false);
492 // Two options: either the InVal is a phi node defined in BB or it is some
493 // value that dominates BB.
494 PHINode *InValPhi = dyn_cast<PHINode>(InVal);
495 if (InValPhi && InValPhi->getParent() == BB) {
496 // Add all of the input values of the input PHI as inputs of this phi.
497 for (unsigned i = 0, e = InValPhi->getNumIncomingValues(); i != e; ++i)
498 PN->addIncoming(InValPhi->getIncomingValue(i),
499 InValPhi->getIncomingBlock(i));
501 // Otherwise, add one instance of the dominating value for each edge that
502 // we will be adding.
503 if (PHINode *BBPN = dyn_cast<PHINode>(BB->begin())) {
504 for (unsigned i = 0, e = BBPN->getNumIncomingValues(); i != e; ++i)
505 PN->addIncoming(InVal, BBPN->getIncomingBlock(i));
507 for (pred_iterator PI = pred_begin(BB), E = pred_end(BB); PI != E; ++PI)
508 PN->addIncoming(InVal, *PI);
513 // The PHIs are now updated, change everything that refers to BB to use
514 // DestBB and remove BB.
515 BB->replaceAllUsesWith(DestBB);
516 BB->eraseFromParent();
519 DEBUG(dbgs() << "AFTER:\n" << *DestBB << "\n\n\n");
522 // Computes a map of base pointer relocation instructions to corresponding
523 // derived pointer relocation instructions given a vector of all relocate calls
524 static void computeBaseDerivedRelocateMap(
525 const SmallVectorImpl<User *> &AllRelocateCalls,
526 DenseMap<IntrinsicInst *, SmallVector<IntrinsicInst *, 2>> &
528 // Collect information in two maps: one primarily for locating the base object
529 // while filling the second map; the second map is the final structure holding
530 // a mapping between Base and corresponding Derived relocate calls
531 DenseMap<std::pair<unsigned, unsigned>, IntrinsicInst *> RelocateIdxMap;
532 for (auto &U : AllRelocateCalls) {
533 GCRelocateOperands ThisRelocate(U);
534 IntrinsicInst *I = cast<IntrinsicInst>(U);
535 auto K = std::make_pair(ThisRelocate.getBasePtrIndex(),
536 ThisRelocate.getDerivedPtrIndex());
537 RelocateIdxMap.insert(std::make_pair(K, I));
539 for (auto &Item : RelocateIdxMap) {
540 std::pair<unsigned, unsigned> Key = Item.first;
541 if (Key.first == Key.second)
542 // Base relocation: nothing to insert
545 IntrinsicInst *I = Item.second;
546 auto BaseKey = std::make_pair(Key.first, Key.first);
548 // We're iterating over RelocateIdxMap so we cannot modify it.
549 auto MaybeBase = RelocateIdxMap.find(BaseKey);
550 if (MaybeBase == RelocateIdxMap.end())
551 // TODO: We might want to insert a new base object relocate and gep off
552 // that, if there are enough derived object relocates.
555 RelocateInstMap[MaybeBase->second].push_back(I);
559 // Accepts a GEP and extracts the operands into a vector provided they're all
560 // small integer constants
561 static bool getGEPSmallConstantIntOffsetV(GetElementPtrInst *GEP,
562 SmallVectorImpl<Value *> &OffsetV) {
563 for (unsigned i = 1; i < GEP->getNumOperands(); i++) {
564 // Only accept small constant integer operands
565 auto Op = dyn_cast<ConstantInt>(GEP->getOperand(i));
566 if (!Op || Op->getZExtValue() > 20)
570 for (unsigned i = 1; i < GEP->getNumOperands(); i++)
571 OffsetV.push_back(GEP->getOperand(i));
575 // Takes a RelocatedBase (base pointer relocation instruction) and Targets to
576 // replace, computes a replacement, and affects it.
578 simplifyRelocatesOffABase(IntrinsicInst *RelocatedBase,
579 const SmallVectorImpl<IntrinsicInst *> &Targets) {
580 bool MadeChange = false;
581 for (auto &ToReplace : Targets) {
582 GCRelocateOperands MasterRelocate(RelocatedBase);
583 GCRelocateOperands ThisRelocate(ToReplace);
585 assert(ThisRelocate.getBasePtrIndex() == MasterRelocate.getBasePtrIndex() &&
586 "Not relocating a derived object of the original base object");
587 if (ThisRelocate.getBasePtrIndex() == ThisRelocate.getDerivedPtrIndex()) {
588 // A duplicate relocate call. TODO: coalesce duplicates.
592 if (RelocatedBase->getParent() != ToReplace->getParent()) {
593 // Base and derived relocates are in different basic blocks.
594 // In this case transform is only valid when base dominates derived
595 // relocate. However it would be too expensive to check dominance
596 // for each such relocate, so we skip the whole transformation.
600 Value *Base = ThisRelocate.getBasePtr();
601 auto Derived = dyn_cast<GetElementPtrInst>(ThisRelocate.getDerivedPtr());
602 if (!Derived || Derived->getPointerOperand() != Base)
605 SmallVector<Value *, 2> OffsetV;
606 if (!getGEPSmallConstantIntOffsetV(Derived, OffsetV))
609 // Create a Builder and replace the target callsite with a gep
610 assert(RelocatedBase->getNextNode() && "Should always have one since it's not a terminator");
612 // Insert after RelocatedBase
613 IRBuilder<> Builder(RelocatedBase->getNextNode());
614 Builder.SetCurrentDebugLocation(ToReplace->getDebugLoc());
616 // If gc_relocate does not match the actual type, cast it to the right type.
617 // In theory, there must be a bitcast after gc_relocate if the type does not
618 // match, and we should reuse it to get the derived pointer. But it could be
622 // %g1 = call coldcc i8 addrspace(1)* @llvm.experimental.gc.relocate.p1i8(...)
627 // %g2 = call coldcc i8 addrspace(1)* @llvm.experimental.gc.relocate.p1i8(...)
631 // %p1 = phi i8 addrspace(1)* [ %g1, %bb1 ], [ %g2, %bb2 ]
632 // %cast = bitcast i8 addrspace(1)* %p1 in to i32 addrspace(1)*
634 // In this case, we can not find the bitcast any more. So we insert a new bitcast
635 // no matter there is already one or not. In this way, we can handle all cases, and
636 // the extra bitcast should be optimized away in later passes.
637 Instruction *ActualRelocatedBase = RelocatedBase;
638 if (RelocatedBase->getType() != Base->getType()) {
639 ActualRelocatedBase =
640 cast<Instruction>(Builder.CreateBitCast(RelocatedBase, Base->getType()));
642 Value *Replacement = Builder.CreateGEP(
643 Derived->getSourceElementType(), ActualRelocatedBase, makeArrayRef(OffsetV));
644 Instruction *ReplacementInst = cast<Instruction>(Replacement);
645 Replacement->takeName(ToReplace);
646 // If the newly generated derived pointer's type does not match the original derived
647 // pointer's type, cast the new derived pointer to match it. Same reasoning as above.
648 Instruction *ActualReplacement = ReplacementInst;
649 if (ReplacementInst->getType() != ToReplace->getType()) {
651 cast<Instruction>(Builder.CreateBitCast(ReplacementInst, ToReplace->getType()));
653 ToReplace->replaceAllUsesWith(ActualReplacement);
654 ToReplace->eraseFromParent();
664 // %ptr = gep %base + 15
665 // %tok = statepoint (%fun, i32 0, i32 0, i32 0, %base, %ptr)
666 // %base' = relocate(%tok, i32 4, i32 4)
667 // %ptr' = relocate(%tok, i32 4, i32 5)
673 // %ptr = gep %base + 15
674 // %tok = statepoint (%fun, i32 0, i32 0, i32 0, %base, %ptr)
675 // %base' = gc.relocate(%tok, i32 4, i32 4)
676 // %ptr' = gep %base' + 15
678 bool CodeGenPrepare::simplifyOffsetableRelocate(Instruction &I) {
679 bool MadeChange = false;
680 SmallVector<User *, 2> AllRelocateCalls;
682 for (auto *U : I.users())
683 if (isGCRelocate(dyn_cast<Instruction>(U)))
684 // Collect all the relocate calls associated with a statepoint
685 AllRelocateCalls.push_back(U);
687 // We need atleast one base pointer relocation + one derived pointer
688 // relocation to mangle
689 if (AllRelocateCalls.size() < 2)
692 // RelocateInstMap is a mapping from the base relocate instruction to the
693 // corresponding derived relocate instructions
694 DenseMap<IntrinsicInst *, SmallVector<IntrinsicInst *, 2>> RelocateInstMap;
695 computeBaseDerivedRelocateMap(AllRelocateCalls, RelocateInstMap);
696 if (RelocateInstMap.empty())
699 for (auto &Item : RelocateInstMap)
700 // Item.first is the RelocatedBase to offset against
701 // Item.second is the vector of Targets to replace
702 MadeChange = simplifyRelocatesOffABase(Item.first, Item.second);
706 /// SinkCast - Sink the specified cast instruction into its user blocks
707 static bool SinkCast(CastInst *CI) {
708 BasicBlock *DefBB = CI->getParent();
710 /// InsertedCasts - Only insert a cast in each block once.
711 DenseMap<BasicBlock*, CastInst*> InsertedCasts;
713 bool MadeChange = false;
714 for (Value::user_iterator UI = CI->user_begin(), E = CI->user_end();
716 Use &TheUse = UI.getUse();
717 Instruction *User = cast<Instruction>(*UI);
719 // Figure out which BB this cast is used in. For PHI's this is the
720 // appropriate predecessor block.
721 BasicBlock *UserBB = User->getParent();
722 if (PHINode *PN = dyn_cast<PHINode>(User)) {
723 UserBB = PN->getIncomingBlock(TheUse);
726 // Preincrement use iterator so we don't invalidate it.
729 // If this user is in the same block as the cast, don't change the cast.
730 if (UserBB == DefBB) continue;
732 // If we have already inserted a cast into this block, use it.
733 CastInst *&InsertedCast = InsertedCasts[UserBB];
736 BasicBlock::iterator InsertPt = UserBB->getFirstInsertionPt();
737 assert(InsertPt != UserBB->end());
738 InsertedCast = CastInst::Create(CI->getOpcode(), CI->getOperand(0),
739 CI->getType(), "", &*InsertPt);
742 // Replace a use of the cast with a use of the new cast.
743 TheUse = InsertedCast;
748 // If we removed all uses, nuke the cast.
749 if (CI->use_empty()) {
750 CI->eraseFromParent();
757 /// If the specified cast instruction is a noop copy (e.g. it's casting from
758 /// one pointer type to another, i32->i8 on PPC), sink it into user blocks to
759 /// reduce the number of virtual registers that must be created and coalesced.
761 /// Return true if any changes are made.
763 static bool OptimizeNoopCopyExpression(CastInst *CI, const TargetLowering &TLI,
764 const DataLayout &DL) {
765 // If this is a noop copy,
766 EVT SrcVT = TLI.getValueType(DL, CI->getOperand(0)->getType());
767 EVT DstVT = TLI.getValueType(DL, CI->getType());
769 // This is an fp<->int conversion?
770 if (SrcVT.isInteger() != DstVT.isInteger())
773 // If this is an extension, it will be a zero or sign extension, which
775 if (SrcVT.bitsLT(DstVT)) return false;
777 // If these values will be promoted, find out what they will be promoted
778 // to. This helps us consider truncates on PPC as noop copies when they
780 if (TLI.getTypeAction(CI->getContext(), SrcVT) ==
781 TargetLowering::TypePromoteInteger)
782 SrcVT = TLI.getTypeToTransformTo(CI->getContext(), SrcVT);
783 if (TLI.getTypeAction(CI->getContext(), DstVT) ==
784 TargetLowering::TypePromoteInteger)
785 DstVT = TLI.getTypeToTransformTo(CI->getContext(), DstVT);
787 // If, after promotion, these are the same types, this is a noop copy.
794 /// Try to combine CI into a call to the llvm.uadd.with.overflow intrinsic if
797 /// Return true if any changes were made.
798 static bool CombineUAddWithOverflow(CmpInst *CI) {
802 m_UAddWithOverflow(m_Value(A), m_Value(B), m_Instruction(AddI))))
805 Type *Ty = AddI->getType();
806 if (!isa<IntegerType>(Ty))
809 // We don't want to move around uses of condition values this late, so we we
810 // check if it is legal to create the call to the intrinsic in the basic
811 // block containing the icmp:
813 if (AddI->getParent() != CI->getParent() && !AddI->hasOneUse())
817 // Someday m_UAddWithOverflow may get smarter, but this is a safe assumption
819 if (AddI->hasOneUse())
820 assert(*AddI->user_begin() == CI && "expected!");
823 Module *M = CI->getParent()->getParent()->getParent();
824 Value *F = Intrinsic::getDeclaration(M, Intrinsic::uadd_with_overflow, Ty);
826 auto *InsertPt = AddI->hasOneUse() ? CI : AddI;
828 auto *UAddWithOverflow =
829 CallInst::Create(F, {A, B}, "uadd.overflow", InsertPt);
830 auto *UAdd = ExtractValueInst::Create(UAddWithOverflow, 0, "uadd", InsertPt);
832 ExtractValueInst::Create(UAddWithOverflow, 1, "overflow", InsertPt);
834 CI->replaceAllUsesWith(Overflow);
835 AddI->replaceAllUsesWith(UAdd);
836 CI->eraseFromParent();
837 AddI->eraseFromParent();
841 /// Sink the given CmpInst into user blocks to reduce the number of virtual
842 /// registers that must be created and coalesced. This is a clear win except on
843 /// targets with multiple condition code registers (PowerPC), where it might
844 /// lose; some adjustment may be wanted there.
846 /// Return true if any changes are made.
847 static bool SinkCmpExpression(CmpInst *CI) {
848 BasicBlock *DefBB = CI->getParent();
850 /// Only insert a cmp in each block once.
851 DenseMap<BasicBlock*, CmpInst*> InsertedCmps;
853 bool MadeChange = false;
854 for (Value::user_iterator UI = CI->user_begin(), E = CI->user_end();
856 Use &TheUse = UI.getUse();
857 Instruction *User = cast<Instruction>(*UI);
859 // Preincrement use iterator so we don't invalidate it.
862 // Don't bother for PHI nodes.
863 if (isa<PHINode>(User))
866 // Figure out which BB this cmp is used in.
867 BasicBlock *UserBB = User->getParent();
869 // If this user is in the same block as the cmp, don't change the cmp.
870 if (UserBB == DefBB) continue;
872 // If we have already inserted a cmp into this block, use it.
873 CmpInst *&InsertedCmp = InsertedCmps[UserBB];
876 BasicBlock::iterator InsertPt = UserBB->getFirstInsertionPt();
877 assert(InsertPt != UserBB->end());
879 CmpInst::Create(CI->getOpcode(), CI->getPredicate(),
880 CI->getOperand(0), CI->getOperand(1), "", &*InsertPt);
883 // Replace a use of the cmp with a use of the new cmp.
884 TheUse = InsertedCmp;
889 // If we removed all uses, nuke the cmp.
890 if (CI->use_empty()) {
891 CI->eraseFromParent();
898 static bool OptimizeCmpExpression(CmpInst *CI) {
899 if (SinkCmpExpression(CI))
902 if (CombineUAddWithOverflow(CI))
908 /// Check if the candidates could be combined with a shift instruction, which
910 /// 1. Truncate instruction
911 /// 2. And instruction and the imm is a mask of the low bits:
912 /// imm & (imm+1) == 0
913 static bool isExtractBitsCandidateUse(Instruction *User) {
914 if (!isa<TruncInst>(User)) {
915 if (User->getOpcode() != Instruction::And ||
916 !isa<ConstantInt>(User->getOperand(1)))
919 const APInt &Cimm = cast<ConstantInt>(User->getOperand(1))->getValue();
921 if ((Cimm & (Cimm + 1)).getBoolValue())
927 /// Sink both shift and truncate instruction to the use of truncate's BB.
929 SinkShiftAndTruncate(BinaryOperator *ShiftI, Instruction *User, ConstantInt *CI,
930 DenseMap<BasicBlock *, BinaryOperator *> &InsertedShifts,
931 const TargetLowering &TLI, const DataLayout &DL) {
932 BasicBlock *UserBB = User->getParent();
933 DenseMap<BasicBlock *, CastInst *> InsertedTruncs;
934 TruncInst *TruncI = dyn_cast<TruncInst>(User);
935 bool MadeChange = false;
937 for (Value::user_iterator TruncUI = TruncI->user_begin(),
938 TruncE = TruncI->user_end();
939 TruncUI != TruncE;) {
941 Use &TruncTheUse = TruncUI.getUse();
942 Instruction *TruncUser = cast<Instruction>(*TruncUI);
943 // Preincrement use iterator so we don't invalidate it.
947 int ISDOpcode = TLI.InstructionOpcodeToISD(TruncUser->getOpcode());
951 // If the use is actually a legal node, there will not be an
952 // implicit truncate.
953 // FIXME: always querying the result type is just an
954 // approximation; some nodes' legality is determined by the
955 // operand or other means. There's no good way to find out though.
956 if (TLI.isOperationLegalOrCustom(
957 ISDOpcode, TLI.getValueType(DL, TruncUser->getType(), true)))
960 // Don't bother for PHI nodes.
961 if (isa<PHINode>(TruncUser))
964 BasicBlock *TruncUserBB = TruncUser->getParent();
966 if (UserBB == TruncUserBB)
969 BinaryOperator *&InsertedShift = InsertedShifts[TruncUserBB];
970 CastInst *&InsertedTrunc = InsertedTruncs[TruncUserBB];
972 if (!InsertedShift && !InsertedTrunc) {
973 BasicBlock::iterator InsertPt = TruncUserBB->getFirstInsertionPt();
974 assert(InsertPt != TruncUserBB->end());
976 if (ShiftI->getOpcode() == Instruction::AShr)
977 InsertedShift = BinaryOperator::CreateAShr(ShiftI->getOperand(0), CI,
980 InsertedShift = BinaryOperator::CreateLShr(ShiftI->getOperand(0), CI,
984 BasicBlock::iterator TruncInsertPt = TruncUserBB->getFirstInsertionPt();
986 assert(TruncInsertPt != TruncUserBB->end());
988 InsertedTrunc = CastInst::Create(TruncI->getOpcode(), InsertedShift,
989 TruncI->getType(), "", &*TruncInsertPt);
993 TruncTheUse = InsertedTrunc;
999 /// Sink the shift *right* instruction into user blocks if the uses could
1000 /// potentially be combined with this shift instruction and generate BitExtract
1001 /// instruction. It will only be applied if the architecture supports BitExtract
1002 /// instruction. Here is an example:
1004 /// %x.extract.shift = lshr i64 %arg1, 32
1006 /// %x.extract.trunc = trunc i64 %x.extract.shift to i16
1010 /// %x.extract.shift.1 = lshr i64 %arg1, 32
1011 /// %x.extract.trunc = trunc i64 %x.extract.shift.1 to i16
1013 /// CodeGen will recoginze the pattern in BB2 and generate BitExtract
1015 /// Return true if any changes are made.
1016 static bool OptimizeExtractBits(BinaryOperator *ShiftI, ConstantInt *CI,
1017 const TargetLowering &TLI,
1018 const DataLayout &DL) {
1019 BasicBlock *DefBB = ShiftI->getParent();
1021 /// Only insert instructions in each block once.
1022 DenseMap<BasicBlock *, BinaryOperator *> InsertedShifts;
1024 bool shiftIsLegal = TLI.isTypeLegal(TLI.getValueType(DL, ShiftI->getType()));
1026 bool MadeChange = false;
1027 for (Value::user_iterator UI = ShiftI->user_begin(), E = ShiftI->user_end();
1029 Use &TheUse = UI.getUse();
1030 Instruction *User = cast<Instruction>(*UI);
1031 // Preincrement use iterator so we don't invalidate it.
1034 // Don't bother for PHI nodes.
1035 if (isa<PHINode>(User))
1038 if (!isExtractBitsCandidateUse(User))
1041 BasicBlock *UserBB = User->getParent();
1043 if (UserBB == DefBB) {
1044 // If the shift and truncate instruction are in the same BB. The use of
1045 // the truncate(TruncUse) may still introduce another truncate if not
1046 // legal. In this case, we would like to sink both shift and truncate
1047 // instruction to the BB of TruncUse.
1050 // i64 shift.result = lshr i64 opnd, imm
1051 // trunc.result = trunc shift.result to i16
1054 // ----> We will have an implicit truncate here if the architecture does
1055 // not have i16 compare.
1056 // cmp i16 trunc.result, opnd2
1058 if (isa<TruncInst>(User) && shiftIsLegal
1059 // If the type of the truncate is legal, no trucate will be
1060 // introduced in other basic blocks.
1062 (!TLI.isTypeLegal(TLI.getValueType(DL, User->getType()))))
1064 SinkShiftAndTruncate(ShiftI, User, CI, InsertedShifts, TLI, DL);
1068 // If we have already inserted a shift into this block, use it.
1069 BinaryOperator *&InsertedShift = InsertedShifts[UserBB];
1071 if (!InsertedShift) {
1072 BasicBlock::iterator InsertPt = UserBB->getFirstInsertionPt();
1073 assert(InsertPt != UserBB->end());
1075 if (ShiftI->getOpcode() == Instruction::AShr)
1076 InsertedShift = BinaryOperator::CreateAShr(ShiftI->getOperand(0), CI,
1079 InsertedShift = BinaryOperator::CreateLShr(ShiftI->getOperand(0), CI,
1085 // Replace a use of the shift with a use of the new shift.
1086 TheUse = InsertedShift;
1089 // If we removed all uses, nuke the shift.
1090 if (ShiftI->use_empty())
1091 ShiftI->eraseFromParent();
1096 // Translate a masked load intrinsic like
1097 // <16 x i32 > @llvm.masked.load( <16 x i32>* %addr, i32 align,
1098 // <16 x i1> %mask, <16 x i32> %passthru)
1099 // to a chain of basic blocks, with loading element one-by-one if
1100 // the appropriate mask bit is set
1102 // %1 = bitcast i8* %addr to i32*
1103 // %2 = extractelement <16 x i1> %mask, i32 0
1104 // %3 = icmp eq i1 %2, true
1105 // br i1 %3, label %cond.load, label %else
1107 //cond.load: ; preds = %0
1108 // %4 = getelementptr i32* %1, i32 0
1109 // %5 = load i32* %4
1110 // %6 = insertelement <16 x i32> undef, i32 %5, i32 0
1113 //else: ; preds = %0, %cond.load
1114 // %res.phi.else = phi <16 x i32> [ %6, %cond.load ], [ undef, %0 ]
1115 // %7 = extractelement <16 x i1> %mask, i32 1
1116 // %8 = icmp eq i1 %7, true
1117 // br i1 %8, label %cond.load1, label %else2
1119 //cond.load1: ; preds = %else
1120 // %9 = getelementptr i32* %1, i32 1
1121 // %10 = load i32* %9
1122 // %11 = insertelement <16 x i32> %res.phi.else, i32 %10, i32 1
1125 //else2: ; preds = %else, %cond.load1
1126 // %res.phi.else3 = phi <16 x i32> [ %11, %cond.load1 ], [ %res.phi.else, %else ]
1127 // %12 = extractelement <16 x i1> %mask, i32 2
1128 // %13 = icmp eq i1 %12, true
1129 // br i1 %13, label %cond.load4, label %else5
1131 static void ScalarizeMaskedLoad(CallInst *CI) {
1132 Value *Ptr = CI->getArgOperand(0);
1133 Value *Alignment = CI->getArgOperand(1);
1134 Value *Mask = CI->getArgOperand(2);
1135 Value *Src0 = CI->getArgOperand(3);
1137 unsigned AlignVal = cast<ConstantInt>(Alignment)->getZExtValue();
1138 VectorType *VecType = dyn_cast<VectorType>(CI->getType());
1139 assert(VecType && "Unexpected return type of masked load intrinsic");
1141 Type *EltTy = CI->getType()->getVectorElementType();
1143 IRBuilder<> Builder(CI->getContext());
1144 Instruction *InsertPt = CI;
1145 BasicBlock *IfBlock = CI->getParent();
1146 BasicBlock *CondBlock = nullptr;
1147 BasicBlock *PrevIfBlock = CI->getParent();
1149 Builder.SetInsertPoint(InsertPt);
1150 Builder.SetCurrentDebugLocation(CI->getDebugLoc());
1152 // Short-cut if the mask is all-true.
1153 bool IsAllOnesMask = isa<Constant>(Mask) &&
1154 cast<Constant>(Mask)->isAllOnesValue();
1156 if (IsAllOnesMask) {
1157 Value *NewI = Builder.CreateAlignedLoad(Ptr, AlignVal);
1158 CI->replaceAllUsesWith(NewI);
1159 CI->eraseFromParent();
1163 // Adjust alignment for the scalar instruction.
1164 AlignVal = std::min(AlignVal, VecType->getScalarSizeInBits()/8);
1165 // Bitcast %addr fron i8* to EltTy*
1167 EltTy->getPointerTo(cast<PointerType>(Ptr->getType())->getAddressSpace());
1168 Value *FirstEltPtr = Builder.CreateBitCast(Ptr, NewPtrType);
1169 unsigned VectorWidth = VecType->getNumElements();
1171 Value *UndefVal = UndefValue::get(VecType);
1173 // The result vector
1174 Value *VResult = UndefVal;
1176 if (isa<ConstantVector>(Mask)) {
1177 for (unsigned Idx = 0; Idx < VectorWidth; ++Idx) {
1178 if (cast<ConstantVector>(Mask)->getOperand(Idx)->isNullValue())
1181 Builder.CreateInBoundsGEP(EltTy, FirstEltPtr, Builder.getInt32(Idx));
1182 LoadInst* Load = Builder.CreateAlignedLoad(Gep, AlignVal);
1183 VResult = Builder.CreateInsertElement(VResult, Load,
1184 Builder.getInt32(Idx));
1186 Value *NewI = Builder.CreateSelect(Mask, VResult, Src0);
1187 CI->replaceAllUsesWith(NewI);
1188 CI->eraseFromParent();
1192 PHINode *Phi = nullptr;
1193 Value *PrevPhi = UndefVal;
1195 for (unsigned Idx = 0; Idx < VectorWidth; ++Idx) {
1197 // Fill the "else" block, created in the previous iteration
1199 // %res.phi.else3 = phi <16 x i32> [ %11, %cond.load1 ], [ %res.phi.else, %else ]
1200 // %mask_1 = extractelement <16 x i1> %mask, i32 Idx
1201 // %to_load = icmp eq i1 %mask_1, true
1202 // br i1 %to_load, label %cond.load, label %else
1205 Phi = Builder.CreatePHI(VecType, 2, "res.phi.else");
1206 Phi->addIncoming(VResult, CondBlock);
1207 Phi->addIncoming(PrevPhi, PrevIfBlock);
1212 Value *Predicate = Builder.CreateExtractElement(Mask, Builder.getInt32(Idx));
1213 Value *Cmp = Builder.CreateICmp(ICmpInst::ICMP_EQ, Predicate,
1214 ConstantInt::get(Predicate->getType(), 1));
1216 // Create "cond" block
1218 // %EltAddr = getelementptr i32* %1, i32 0
1219 // %Elt = load i32* %EltAddr
1220 // VResult = insertelement <16 x i32> VResult, i32 %Elt, i32 Idx
1222 CondBlock = IfBlock->splitBasicBlock(InsertPt->getIterator(), "cond.load");
1223 Builder.SetInsertPoint(InsertPt);
1226 Builder.CreateInBoundsGEP(EltTy, FirstEltPtr, Builder.getInt32(Idx));
1227 LoadInst *Load = Builder.CreateAlignedLoad(Gep, AlignVal);
1228 VResult = Builder.CreateInsertElement(VResult, Load, Builder.getInt32(Idx));
1230 // Create "else" block, fill it in the next iteration
1231 BasicBlock *NewIfBlock =
1232 CondBlock->splitBasicBlock(InsertPt->getIterator(), "else");
1233 Builder.SetInsertPoint(InsertPt);
1234 Instruction *OldBr = IfBlock->getTerminator();
1235 BranchInst::Create(CondBlock, NewIfBlock, Cmp, OldBr);
1236 OldBr->eraseFromParent();
1237 PrevIfBlock = IfBlock;
1238 IfBlock = NewIfBlock;
1241 Phi = Builder.CreatePHI(VecType, 2, "res.phi.select");
1242 Phi->addIncoming(VResult, CondBlock);
1243 Phi->addIncoming(PrevPhi, PrevIfBlock);
1244 Value *NewI = Builder.CreateSelect(Mask, Phi, Src0);
1245 CI->replaceAllUsesWith(NewI);
1246 CI->eraseFromParent();
1249 // Translate a masked store intrinsic, like
1250 // void @llvm.masked.store(<16 x i32> %src, <16 x i32>* %addr, i32 align,
1252 // to a chain of basic blocks, that stores element one-by-one if
1253 // the appropriate mask bit is set
1255 // %1 = bitcast i8* %addr to i32*
1256 // %2 = extractelement <16 x i1> %mask, i32 0
1257 // %3 = icmp eq i1 %2, true
1258 // br i1 %3, label %cond.store, label %else
1260 // cond.store: ; preds = %0
1261 // %4 = extractelement <16 x i32> %val, i32 0
1262 // %5 = getelementptr i32* %1, i32 0
1263 // store i32 %4, i32* %5
1266 // else: ; preds = %0, %cond.store
1267 // %6 = extractelement <16 x i1> %mask, i32 1
1268 // %7 = icmp eq i1 %6, true
1269 // br i1 %7, label %cond.store1, label %else2
1271 // cond.store1: ; preds = %else
1272 // %8 = extractelement <16 x i32> %val, i32 1
1273 // %9 = getelementptr i32* %1, i32 1
1274 // store i32 %8, i32* %9
1277 static void ScalarizeMaskedStore(CallInst *CI) {
1278 Value *Src = CI->getArgOperand(0);
1279 Value *Ptr = CI->getArgOperand(1);
1280 Value *Alignment = CI->getArgOperand(2);
1281 Value *Mask = CI->getArgOperand(3);
1283 unsigned AlignVal = cast<ConstantInt>(Alignment)->getZExtValue();
1284 VectorType *VecType = dyn_cast<VectorType>(Src->getType());
1285 assert(VecType && "Unexpected data type in masked store intrinsic");
1287 Type *EltTy = VecType->getElementType();
1289 IRBuilder<> Builder(CI->getContext());
1290 Instruction *InsertPt = CI;
1291 BasicBlock *IfBlock = CI->getParent();
1292 Builder.SetInsertPoint(InsertPt);
1293 Builder.SetCurrentDebugLocation(CI->getDebugLoc());
1295 // Short-cut if the mask is all-true.
1296 bool IsAllOnesMask = isa<Constant>(Mask) &&
1297 cast<Constant>(Mask)->isAllOnesValue();
1299 if (IsAllOnesMask) {
1300 Builder.CreateAlignedStore(Src, Ptr, AlignVal);
1301 CI->eraseFromParent();
1305 // Adjust alignment for the scalar instruction.
1306 AlignVal = std::max(AlignVal, VecType->getScalarSizeInBits()/8);
1307 // Bitcast %addr fron i8* to EltTy*
1309 EltTy->getPointerTo(cast<PointerType>(Ptr->getType())->getAddressSpace());
1310 Value *FirstEltPtr = Builder.CreateBitCast(Ptr, NewPtrType);
1311 unsigned VectorWidth = VecType->getNumElements();
1313 if (isa<ConstantVector>(Mask)) {
1314 for (unsigned Idx = 0; Idx < VectorWidth; ++Idx) {
1315 if (cast<ConstantVector>(Mask)->getOperand(Idx)->isNullValue())
1317 Value *OneElt = Builder.CreateExtractElement(Src, Builder.getInt32(Idx));
1319 Builder.CreateInBoundsGEP(EltTy, FirstEltPtr, Builder.getInt32(Idx));
1320 Builder.CreateAlignedStore(OneElt, Gep, AlignVal);
1322 CI->eraseFromParent();
1326 for (unsigned Idx = 0; Idx < VectorWidth; ++Idx) {
1328 // Fill the "else" block, created in the previous iteration
1330 // %mask_1 = extractelement <16 x i1> %mask, i32 Idx
1331 // %to_store = icmp eq i1 %mask_1, true
1332 // br i1 %to_store, label %cond.store, label %else
1334 Value *Predicate = Builder.CreateExtractElement(Mask, Builder.getInt32(Idx));
1335 Value *Cmp = Builder.CreateICmp(ICmpInst::ICMP_EQ, Predicate,
1336 ConstantInt::get(Predicate->getType(), 1));
1338 // Create "cond" block
1340 // %OneElt = extractelement <16 x i32> %Src, i32 Idx
1341 // %EltAddr = getelementptr i32* %1, i32 0
1342 // %store i32 %OneElt, i32* %EltAddr
1344 BasicBlock *CondBlock =
1345 IfBlock->splitBasicBlock(InsertPt->getIterator(), "cond.store");
1346 Builder.SetInsertPoint(InsertPt);
1348 Value *OneElt = Builder.CreateExtractElement(Src, Builder.getInt32(Idx));
1350 Builder.CreateInBoundsGEP(EltTy, FirstEltPtr, Builder.getInt32(Idx));
1351 Builder.CreateAlignedStore(OneElt, Gep, AlignVal);
1353 // Create "else" block, fill it in the next iteration
1354 BasicBlock *NewIfBlock =
1355 CondBlock->splitBasicBlock(InsertPt->getIterator(), "else");
1356 Builder.SetInsertPoint(InsertPt);
1357 Instruction *OldBr = IfBlock->getTerminator();
1358 BranchInst::Create(CondBlock, NewIfBlock, Cmp, OldBr);
1359 OldBr->eraseFromParent();
1360 IfBlock = NewIfBlock;
1362 CI->eraseFromParent();
1365 // Translate a masked gather intrinsic like
1366 // <16 x i32 > @llvm.masked.gather.v16i32( <16 x i32*> %Ptrs, i32 4,
1367 // <16 x i1> %Mask, <16 x i32> %Src)
1368 // to a chain of basic blocks, with loading element one-by-one if
1369 // the appropriate mask bit is set
1371 // % Ptrs = getelementptr i32, i32* %base, <16 x i64> %ind
1372 // % Mask0 = extractelement <16 x i1> %Mask, i32 0
1373 // % ToLoad0 = icmp eq i1 % Mask0, true
1374 // br i1 % ToLoad0, label %cond.load, label %else
1377 // % Ptr0 = extractelement <16 x i32*> %Ptrs, i32 0
1378 // % Load0 = load i32, i32* % Ptr0, align 4
1379 // % Res0 = insertelement <16 x i32> undef, i32 % Load0, i32 0
1383 // %res.phi.else = phi <16 x i32>[% Res0, %cond.load], [undef, % 0]
1384 // % Mask1 = extractelement <16 x i1> %Mask, i32 1
1385 // % ToLoad1 = icmp eq i1 % Mask1, true
1386 // br i1 % ToLoad1, label %cond.load1, label %else2
1389 // % Ptr1 = extractelement <16 x i32*> %Ptrs, i32 1
1390 // % Load1 = load i32, i32* % Ptr1, align 4
1391 // % Res1 = insertelement <16 x i32> %res.phi.else, i32 % Load1, i32 1
1394 // % Result = select <16 x i1> %Mask, <16 x i32> %res.phi.select, <16 x i32> %Src
1395 // ret <16 x i32> %Result
1396 static void ScalarizeMaskedGather(CallInst *CI) {
1397 Value *Ptrs = CI->getArgOperand(0);
1398 Value *Alignment = CI->getArgOperand(1);
1399 Value *Mask = CI->getArgOperand(2);
1400 Value *Src0 = CI->getArgOperand(3);
1402 VectorType *VecType = dyn_cast<VectorType>(CI->getType());
1404 assert(VecType && "Unexpected return type of masked load intrinsic");
1406 IRBuilder<> Builder(CI->getContext());
1407 Instruction *InsertPt = CI;
1408 BasicBlock *IfBlock = CI->getParent();
1409 BasicBlock *CondBlock = nullptr;
1410 BasicBlock *PrevIfBlock = CI->getParent();
1411 Builder.SetInsertPoint(InsertPt);
1412 unsigned AlignVal = cast<ConstantInt>(Alignment)->getZExtValue();
1414 Builder.SetCurrentDebugLocation(CI->getDebugLoc());
1416 Value *UndefVal = UndefValue::get(VecType);
1418 // The result vector
1419 Value *VResult = UndefVal;
1420 unsigned VectorWidth = VecType->getNumElements();
1422 // Shorten the way if the mask is a vector of constants.
1423 bool IsConstMask = isa<ConstantVector>(Mask);
1426 for (unsigned Idx = 0; Idx < VectorWidth; ++Idx) {
1427 if (cast<ConstantVector>(Mask)->getOperand(Idx)->isNullValue())
1429 Value *Ptr = Builder.CreateExtractElement(Ptrs, Builder.getInt32(Idx),
1430 "Ptr" + Twine(Idx));
1431 LoadInst *Load = Builder.CreateAlignedLoad(Ptr, AlignVal,
1432 "Load" + Twine(Idx));
1433 VResult = Builder.CreateInsertElement(VResult, Load,
1434 Builder.getInt32(Idx),
1435 "Res" + Twine(Idx));
1437 Value *NewI = Builder.CreateSelect(Mask, VResult, Src0);
1438 CI->replaceAllUsesWith(NewI);
1439 CI->eraseFromParent();
1443 PHINode *Phi = nullptr;
1444 Value *PrevPhi = UndefVal;
1446 for (unsigned Idx = 0; Idx < VectorWidth; ++Idx) {
1448 // Fill the "else" block, created in the previous iteration
1450 // %Mask1 = extractelement <16 x i1> %Mask, i32 1
1451 // %ToLoad1 = icmp eq i1 %Mask1, true
1452 // br i1 %ToLoad1, label %cond.load, label %else
1455 Phi = Builder.CreatePHI(VecType, 2, "res.phi.else");
1456 Phi->addIncoming(VResult, CondBlock);
1457 Phi->addIncoming(PrevPhi, PrevIfBlock);
1462 Value *Predicate = Builder.CreateExtractElement(Mask,
1463 Builder.getInt32(Idx),
1464 "Mask" + Twine(Idx));
1465 Value *Cmp = Builder.CreateICmp(ICmpInst::ICMP_EQ, Predicate,
1466 ConstantInt::get(Predicate->getType(), 1),
1467 "ToLoad" + Twine(Idx));
1469 // Create "cond" block
1471 // %EltAddr = getelementptr i32* %1, i32 0
1472 // %Elt = load i32* %EltAddr
1473 // VResult = insertelement <16 x i32> VResult, i32 %Elt, i32 Idx
1475 CondBlock = IfBlock->splitBasicBlock(InsertPt, "cond.load");
1476 Builder.SetInsertPoint(InsertPt);
1478 Value *Ptr = Builder.CreateExtractElement(Ptrs, Builder.getInt32(Idx),
1479 "Ptr" + Twine(Idx));
1480 LoadInst *Load = Builder.CreateAlignedLoad(Ptr, AlignVal,
1481 "Load" + Twine(Idx));
1482 VResult = Builder.CreateInsertElement(VResult, Load, Builder.getInt32(Idx),
1483 "Res" + Twine(Idx));
1485 // Create "else" block, fill it in the next iteration
1486 BasicBlock *NewIfBlock = CondBlock->splitBasicBlock(InsertPt, "else");
1487 Builder.SetInsertPoint(InsertPt);
1488 Instruction *OldBr = IfBlock->getTerminator();
1489 BranchInst::Create(CondBlock, NewIfBlock, Cmp, OldBr);
1490 OldBr->eraseFromParent();
1491 PrevIfBlock = IfBlock;
1492 IfBlock = NewIfBlock;
1495 Phi = Builder.CreatePHI(VecType, 2, "res.phi.select");
1496 Phi->addIncoming(VResult, CondBlock);
1497 Phi->addIncoming(PrevPhi, PrevIfBlock);
1498 Value *NewI = Builder.CreateSelect(Mask, Phi, Src0);
1499 CI->replaceAllUsesWith(NewI);
1500 CI->eraseFromParent();
1503 // Translate a masked scatter intrinsic, like
1504 // void @llvm.masked.scatter.v16i32(<16 x i32> %Src, <16 x i32*>* %Ptrs, i32 4,
1506 // to a chain of basic blocks, that stores element one-by-one if
1507 // the appropriate mask bit is set.
1509 // % Ptrs = getelementptr i32, i32* %ptr, <16 x i64> %ind
1510 // % Mask0 = extractelement <16 x i1> % Mask, i32 0
1511 // % ToStore0 = icmp eq i1 % Mask0, true
1512 // br i1 %ToStore0, label %cond.store, label %else
1515 // % Elt0 = extractelement <16 x i32> %Src, i32 0
1516 // % Ptr0 = extractelement <16 x i32*> %Ptrs, i32 0
1517 // store i32 %Elt0, i32* % Ptr0, align 4
1521 // % Mask1 = extractelement <16 x i1> % Mask, i32 1
1522 // % ToStore1 = icmp eq i1 % Mask1, true
1523 // br i1 % ToStore1, label %cond.store1, label %else2
1526 // % Elt1 = extractelement <16 x i32> %Src, i32 1
1527 // % Ptr1 = extractelement <16 x i32*> %Ptrs, i32 1
1528 // store i32 % Elt1, i32* % Ptr1, align 4
1531 static void ScalarizeMaskedScatter(CallInst *CI) {
1532 Value *Src = CI->getArgOperand(0);
1533 Value *Ptrs = CI->getArgOperand(1);
1534 Value *Alignment = CI->getArgOperand(2);
1535 Value *Mask = CI->getArgOperand(3);
1537 assert(isa<VectorType>(Src->getType()) &&
1538 "Unexpected data type in masked scatter intrinsic");
1539 assert(isa<VectorType>(Ptrs->getType()) &&
1540 isa<PointerType>(Ptrs->getType()->getVectorElementType()) &&
1541 "Vector of pointers is expected in masked scatter intrinsic");
1543 IRBuilder<> Builder(CI->getContext());
1544 Instruction *InsertPt = CI;
1545 BasicBlock *IfBlock = CI->getParent();
1546 Builder.SetInsertPoint(InsertPt);
1547 Builder.SetCurrentDebugLocation(CI->getDebugLoc());
1549 unsigned AlignVal = cast<ConstantInt>(Alignment)->getZExtValue();
1550 unsigned VectorWidth = Src->getType()->getVectorNumElements();
1552 // Shorten the way if the mask is a vector of constants.
1553 bool IsConstMask = isa<ConstantVector>(Mask);
1556 for (unsigned Idx = 0; Idx < VectorWidth; ++Idx) {
1557 if (cast<ConstantVector>(Mask)->getOperand(Idx)->isNullValue())
1559 Value *OneElt = Builder.CreateExtractElement(Src, Builder.getInt32(Idx),
1560 "Elt" + Twine(Idx));
1561 Value *Ptr = Builder.CreateExtractElement(Ptrs, Builder.getInt32(Idx),
1562 "Ptr" + Twine(Idx));
1563 Builder.CreateAlignedStore(OneElt, Ptr, AlignVal);
1565 CI->eraseFromParent();
1568 for (unsigned Idx = 0; Idx < VectorWidth; ++Idx) {
1569 // Fill the "else" block, created in the previous iteration
1571 // % Mask1 = extractelement <16 x i1> % Mask, i32 Idx
1572 // % ToStore = icmp eq i1 % Mask1, true
1573 // br i1 % ToStore, label %cond.store, label %else
1575 Value *Predicate = Builder.CreateExtractElement(Mask,
1576 Builder.getInt32(Idx),
1577 "Mask" + Twine(Idx));
1579 Builder.CreateICmp(ICmpInst::ICMP_EQ, Predicate,
1580 ConstantInt::get(Predicate->getType(), 1),
1581 "ToStore" + Twine(Idx));
1583 // Create "cond" block
1585 // % Elt1 = extractelement <16 x i32> %Src, i32 1
1586 // % Ptr1 = extractelement <16 x i32*> %Ptrs, i32 1
1587 // %store i32 % Elt1, i32* % Ptr1
1589 BasicBlock *CondBlock = IfBlock->splitBasicBlock(InsertPt, "cond.store");
1590 Builder.SetInsertPoint(InsertPt);
1592 Value *OneElt = Builder.CreateExtractElement(Src, Builder.getInt32(Idx),
1593 "Elt" + Twine(Idx));
1594 Value *Ptr = Builder.CreateExtractElement(Ptrs, Builder.getInt32(Idx),
1595 "Ptr" + Twine(Idx));
1596 Builder.CreateAlignedStore(OneElt, Ptr, AlignVal);
1598 // Create "else" block, fill it in the next iteration
1599 BasicBlock *NewIfBlock = CondBlock->splitBasicBlock(InsertPt, "else");
1600 Builder.SetInsertPoint(InsertPt);
1601 Instruction *OldBr = IfBlock->getTerminator();
1602 BranchInst::Create(CondBlock, NewIfBlock, Cmp, OldBr);
1603 OldBr->eraseFromParent();
1604 IfBlock = NewIfBlock;
1606 CI->eraseFromParent();
1609 /// If counting leading or trailing zeros is an expensive operation and a zero
1610 /// input is defined, add a check for zero to avoid calling the intrinsic.
1612 /// We want to transform:
1613 /// %z = call i64 @llvm.cttz.i64(i64 %A, i1 false)
1617 /// %cmpz = icmp eq i64 %A, 0
1618 /// br i1 %cmpz, label %cond.end, label %cond.false
1620 /// %z = call i64 @llvm.cttz.i64(i64 %A, i1 true)
1621 /// br label %cond.end
1623 /// %ctz = phi i64 [ 64, %entry ], [ %z, %cond.false ]
1625 /// If the transform is performed, return true and set ModifiedDT to true.
1626 static bool despeculateCountZeros(IntrinsicInst *CountZeros,
1627 const TargetLowering *TLI,
1628 const DataLayout *DL,
1633 // If a zero input is undefined, it doesn't make sense to despeculate that.
1634 if (match(CountZeros->getOperand(1), m_One()))
1637 // If it's cheap to speculate, there's nothing to do.
1638 auto IntrinsicID = CountZeros->getIntrinsicID();
1639 if ((IntrinsicID == Intrinsic::cttz && TLI->isCheapToSpeculateCttz()) ||
1640 (IntrinsicID == Intrinsic::ctlz && TLI->isCheapToSpeculateCtlz()))
1643 // Only handle legal scalar cases. Anything else requires too much work.
1644 Type *Ty = CountZeros->getType();
1645 unsigned SizeInBits = Ty->getPrimitiveSizeInBits();
1646 if (Ty->isVectorTy() || SizeInBits > DL->getLargestLegalIntTypeSize())
1649 // The intrinsic will be sunk behind a compare against zero and branch.
1650 BasicBlock *StartBlock = CountZeros->getParent();
1651 BasicBlock *CallBlock = StartBlock->splitBasicBlock(CountZeros, "cond.false");
1653 // Create another block after the count zero intrinsic. A PHI will be added
1654 // in this block to select the result of the intrinsic or the bit-width
1655 // constant if the input to the intrinsic is zero.
1656 BasicBlock::iterator SplitPt = ++(BasicBlock::iterator(CountZeros));
1657 BasicBlock *EndBlock = CallBlock->splitBasicBlock(SplitPt, "cond.end");
1659 // Set up a builder to create a compare, conditional branch, and PHI.
1660 IRBuilder<> Builder(CountZeros->getContext());
1661 Builder.SetInsertPoint(StartBlock->getTerminator());
1662 Builder.SetCurrentDebugLocation(CountZeros->getDebugLoc());
1664 // Replace the unconditional branch that was created by the first split with
1665 // a compare against zero and a conditional branch.
1666 Value *Zero = Constant::getNullValue(Ty);
1667 Value *Cmp = Builder.CreateICmpEQ(CountZeros->getOperand(0), Zero, "cmpz");
1668 Builder.CreateCondBr(Cmp, EndBlock, CallBlock);
1669 StartBlock->getTerminator()->eraseFromParent();
1671 // Create a PHI in the end block to select either the output of the intrinsic
1672 // or the bit width of the operand.
1673 Builder.SetInsertPoint(&EndBlock->front());
1674 PHINode *PN = Builder.CreatePHI(Ty, 2, "ctz");
1675 CountZeros->replaceAllUsesWith(PN);
1676 Value *BitWidth = Builder.getInt(APInt(SizeInBits, SizeInBits));
1677 PN->addIncoming(BitWidth, StartBlock);
1678 PN->addIncoming(CountZeros, CallBlock);
1680 // We are explicitly handling the zero case, so we can set the intrinsic's
1681 // undefined zero argument to 'true'. This will also prevent reprocessing the
1682 // intrinsic; we only despeculate when a zero input is defined.
1683 CountZeros->setArgOperand(1, Builder.getTrue());
1688 bool CodeGenPrepare::optimizeCallInst(CallInst *CI, bool& ModifiedDT) {
1689 BasicBlock *BB = CI->getParent();
1691 // Lower inline assembly if we can.
1692 // If we found an inline asm expession, and if the target knows how to
1693 // lower it to normal LLVM code, do so now.
1694 if (TLI && isa<InlineAsm>(CI->getCalledValue())) {
1695 if (TLI->ExpandInlineAsm(CI)) {
1696 // Avoid invalidating the iterator.
1697 CurInstIterator = BB->begin();
1698 // Avoid processing instructions out of order, which could cause
1699 // reuse before a value is defined.
1703 // Sink address computing for memory operands into the block.
1704 if (optimizeInlineAsmInst(CI))
1708 // Align the pointer arguments to this call if the target thinks it's a good
1710 unsigned MinSize, PrefAlign;
1711 if (TLI && TLI->shouldAlignPointerArgs(CI, MinSize, PrefAlign)) {
1712 for (auto &Arg : CI->arg_operands()) {
1713 // We want to align both objects whose address is used directly and
1714 // objects whose address is used in casts and GEPs, though it only makes
1715 // sense for GEPs if the offset is a multiple of the desired alignment and
1716 // if size - offset meets the size threshold.
1717 if (!Arg->getType()->isPointerTy())
1719 APInt Offset(DL->getPointerSizeInBits(
1720 cast<PointerType>(Arg->getType())->getAddressSpace()),
1722 Value *Val = Arg->stripAndAccumulateInBoundsConstantOffsets(*DL, Offset);
1723 uint64_t Offset2 = Offset.getLimitedValue();
1724 if ((Offset2 & (PrefAlign-1)) != 0)
1727 if ((AI = dyn_cast<AllocaInst>(Val)) && AI->getAlignment() < PrefAlign &&
1728 DL->getTypeAllocSize(AI->getAllocatedType()) >= MinSize + Offset2)
1729 AI->setAlignment(PrefAlign);
1730 // Global variables can only be aligned if they are defined in this
1731 // object (i.e. they are uniquely initialized in this object), and
1732 // over-aligning global variables that have an explicit section is
1735 if ((GV = dyn_cast<GlobalVariable>(Val)) && GV->hasUniqueInitializer() &&
1736 !GV->hasSection() && GV->getAlignment() < PrefAlign &&
1737 DL->getTypeAllocSize(GV->getType()->getElementType()) >=
1739 GV->setAlignment(PrefAlign);
1741 // If this is a memcpy (or similar) then we may be able to improve the
1743 if (MemIntrinsic *MI = dyn_cast<MemIntrinsic>(CI)) {
1744 unsigned Align = getKnownAlignment(MI->getDest(), *DL);
1745 if (MemTransferInst *MTI = dyn_cast<MemTransferInst>(MI))
1746 Align = std::min(Align, getKnownAlignment(MTI->getSource(), *DL));
1747 if (Align > MI->getAlignment())
1748 MI->setAlignment(ConstantInt::get(MI->getAlignmentType(), Align));
1752 IntrinsicInst *II = dyn_cast<IntrinsicInst>(CI);
1754 switch (II->getIntrinsicID()) {
1756 case Intrinsic::objectsize: {
1757 // Lower all uses of llvm.objectsize.*
1758 bool Min = (cast<ConstantInt>(II->getArgOperand(1))->getZExtValue() == 1);
1759 Type *ReturnTy = CI->getType();
1760 Constant *RetVal = ConstantInt::get(ReturnTy, Min ? 0 : -1ULL);
1762 // Substituting this can cause recursive simplifications, which can
1763 // invalidate our iterator. Use a WeakVH to hold onto it in case this
1765 WeakVH IterHandle(&*CurInstIterator);
1767 replaceAndRecursivelySimplify(CI, RetVal,
1770 // If the iterator instruction was recursively deleted, start over at the
1771 // start of the block.
1772 if (IterHandle != CurInstIterator.getNodePtrUnchecked()) {
1773 CurInstIterator = BB->begin();
1778 case Intrinsic::masked_load: {
1779 // Scalarize unsupported vector masked load
1780 if (!TTI->isLegalMaskedLoad(CI->getType())) {
1781 ScalarizeMaskedLoad(CI);
1787 case Intrinsic::masked_store: {
1788 if (!TTI->isLegalMaskedStore(CI->getArgOperand(0)->getType())) {
1789 ScalarizeMaskedStore(CI);
1795 case Intrinsic::masked_gather: {
1796 if (!TTI->isLegalMaskedGather(CI->getType())) {
1797 ScalarizeMaskedGather(CI);
1803 case Intrinsic::masked_scatter: {
1804 if (!TTI->isLegalMaskedScatter(CI->getArgOperand(0)->getType())) {
1805 ScalarizeMaskedScatter(CI);
1811 case Intrinsic::aarch64_stlxr:
1812 case Intrinsic::aarch64_stxr: {
1813 ZExtInst *ExtVal = dyn_cast<ZExtInst>(CI->getArgOperand(0));
1814 if (!ExtVal || !ExtVal->hasOneUse() ||
1815 ExtVal->getParent() == CI->getParent())
1817 // Sink a zext feeding stlxr/stxr before it, so it can be folded into it.
1818 ExtVal->moveBefore(CI);
1819 // Mark this instruction as "inserted by CGP", so that other
1820 // optimizations don't touch it.
1821 InsertedInsts.insert(ExtVal);
1824 case Intrinsic::invariant_group_barrier:
1825 II->replaceAllUsesWith(II->getArgOperand(0));
1826 II->eraseFromParent();
1829 case Intrinsic::cttz:
1830 case Intrinsic::ctlz:
1831 // If counting zeros is expensive, try to avoid it.
1832 return despeculateCountZeros(II, TLI, DL, ModifiedDT);
1836 // Unknown address space.
1837 // TODO: Target hook to pick which address space the intrinsic cares
1839 unsigned AddrSpace = ~0u;
1840 SmallVector<Value*, 2> PtrOps;
1842 if (TLI->GetAddrModeArguments(II, PtrOps, AccessTy, AddrSpace))
1843 while (!PtrOps.empty())
1844 if (optimizeMemoryInst(II, PtrOps.pop_back_val(), AccessTy, AddrSpace))
1849 // From here on out we're working with named functions.
1850 if (!CI->getCalledFunction()) return false;
1852 // Lower all default uses of _chk calls. This is very similar
1853 // to what InstCombineCalls does, but here we are only lowering calls
1854 // to fortified library functions (e.g. __memcpy_chk) that have the default
1855 // "don't know" as the objectsize. Anything else should be left alone.
1856 FortifiedLibCallSimplifier Simplifier(TLInfo, true);
1857 if (Value *V = Simplifier.optimizeCall(CI)) {
1858 CI->replaceAllUsesWith(V);
1859 CI->eraseFromParent();
1865 /// Look for opportunities to duplicate return instructions to the predecessor
1866 /// to enable tail call optimizations. The case it is currently looking for is:
1869 /// %tmp0 = tail call i32 @f0()
1870 /// br label %return
1872 /// %tmp1 = tail call i32 @f1()
1873 /// br label %return
1875 /// %tmp2 = tail call i32 @f2()
1876 /// br label %return
1878 /// %retval = phi i32 [ %tmp0, %bb0 ], [ %tmp1, %bb1 ], [ %tmp2, %bb2 ]
1886 /// %tmp0 = tail call i32 @f0()
1889 /// %tmp1 = tail call i32 @f1()
1892 /// %tmp2 = tail call i32 @f2()
1895 bool CodeGenPrepare::dupRetToEnableTailCallOpts(BasicBlock *BB) {
1899 ReturnInst *RI = dyn_cast<ReturnInst>(BB->getTerminator());
1903 PHINode *PN = nullptr;
1904 BitCastInst *BCI = nullptr;
1905 Value *V = RI->getReturnValue();
1907 BCI = dyn_cast<BitCastInst>(V);
1909 V = BCI->getOperand(0);
1911 PN = dyn_cast<PHINode>(V);
1916 if (PN && PN->getParent() != BB)
1919 // It's not safe to eliminate the sign / zero extension of the return value.
1920 // See llvm::isInTailCallPosition().
1921 const Function *F = BB->getParent();
1922 AttributeSet CallerAttrs = F->getAttributes();
1923 if (CallerAttrs.hasAttribute(AttributeSet::ReturnIndex, Attribute::ZExt) ||
1924 CallerAttrs.hasAttribute(AttributeSet::ReturnIndex, Attribute::SExt))
1927 // Make sure there are no instructions between the PHI and return, or that the
1928 // return is the first instruction in the block.
1930 BasicBlock::iterator BI = BB->begin();
1931 do { ++BI; } while (isa<DbgInfoIntrinsic>(BI));
1933 // Also skip over the bitcast.
1938 BasicBlock::iterator BI = BB->begin();
1939 while (isa<DbgInfoIntrinsic>(BI)) ++BI;
1944 /// Only dup the ReturnInst if the CallInst is likely to be emitted as a tail
1946 SmallVector<CallInst*, 4> TailCalls;
1948 for (unsigned I = 0, E = PN->getNumIncomingValues(); I != E; ++I) {
1949 CallInst *CI = dyn_cast<CallInst>(PN->getIncomingValue(I));
1950 // Make sure the phi value is indeed produced by the tail call.
1951 if (CI && CI->hasOneUse() && CI->getParent() == PN->getIncomingBlock(I) &&
1952 TLI->mayBeEmittedAsTailCall(CI))
1953 TailCalls.push_back(CI);
1956 SmallPtrSet<BasicBlock*, 4> VisitedBBs;
1957 for (pred_iterator PI = pred_begin(BB), PE = pred_end(BB); PI != PE; ++PI) {
1958 if (!VisitedBBs.insert(*PI).second)
1961 BasicBlock::InstListType &InstList = (*PI)->getInstList();
1962 BasicBlock::InstListType::reverse_iterator RI = InstList.rbegin();
1963 BasicBlock::InstListType::reverse_iterator RE = InstList.rend();
1964 do { ++RI; } while (RI != RE && isa<DbgInfoIntrinsic>(&*RI));
1968 CallInst *CI = dyn_cast<CallInst>(&*RI);
1969 if (CI && CI->use_empty() && TLI->mayBeEmittedAsTailCall(CI))
1970 TailCalls.push_back(CI);
1974 bool Changed = false;
1975 for (unsigned i = 0, e = TailCalls.size(); i != e; ++i) {
1976 CallInst *CI = TailCalls[i];
1979 // Conservatively require the attributes of the call to match those of the
1980 // return. Ignore noalias because it doesn't affect the call sequence.
1981 AttributeSet CalleeAttrs = CS.getAttributes();
1982 if (AttrBuilder(CalleeAttrs, AttributeSet::ReturnIndex).
1983 removeAttribute(Attribute::NoAlias) !=
1984 AttrBuilder(CalleeAttrs, AttributeSet::ReturnIndex).
1985 removeAttribute(Attribute::NoAlias))
1988 // Make sure the call instruction is followed by an unconditional branch to
1989 // the return block.
1990 BasicBlock *CallBB = CI->getParent();
1991 BranchInst *BI = dyn_cast<BranchInst>(CallBB->getTerminator());
1992 if (!BI || !BI->isUnconditional() || BI->getSuccessor(0) != BB)
1995 // Duplicate the return into CallBB.
1996 (void)FoldReturnIntoUncondBranch(RI, BB, CallBB);
1997 ModifiedDT = Changed = true;
2001 // If we eliminated all predecessors of the block, delete the block now.
2002 if (Changed && !BB->hasAddressTaken() && pred_begin(BB) == pred_end(BB))
2003 BB->eraseFromParent();
2008 //===----------------------------------------------------------------------===//
2009 // Memory Optimization
2010 //===----------------------------------------------------------------------===//
2014 /// This is an extended version of TargetLowering::AddrMode
2015 /// which holds actual Value*'s for register values.
2016 struct ExtAddrMode : public TargetLowering::AddrMode {
2019 ExtAddrMode() : BaseReg(nullptr), ScaledReg(nullptr) {}
2020 void print(raw_ostream &OS) const;
2023 bool operator==(const ExtAddrMode& O) const {
2024 return (BaseReg == O.BaseReg) && (ScaledReg == O.ScaledReg) &&
2025 (BaseGV == O.BaseGV) && (BaseOffs == O.BaseOffs) &&
2026 (HasBaseReg == O.HasBaseReg) && (Scale == O.Scale);
2031 static inline raw_ostream &operator<<(raw_ostream &OS, const ExtAddrMode &AM) {
2037 void ExtAddrMode::print(raw_ostream &OS) const {
2038 bool NeedPlus = false;
2041 OS << (NeedPlus ? " + " : "")
2043 BaseGV->printAsOperand(OS, /*PrintType=*/false);
2048 OS << (NeedPlus ? " + " : "")
2054 OS << (NeedPlus ? " + " : "")
2056 BaseReg->printAsOperand(OS, /*PrintType=*/false);
2060 OS << (NeedPlus ? " + " : "")
2062 ScaledReg->printAsOperand(OS, /*PrintType=*/false);
2068 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
2069 void ExtAddrMode::dump() const {
2075 /// \brief This class provides transaction based operation on the IR.
2076 /// Every change made through this class is recorded in the internal state and
2077 /// can be undone (rollback) until commit is called.
2078 class TypePromotionTransaction {
2080 /// \brief This represents the common interface of the individual transaction.
2081 /// Each class implements the logic for doing one specific modification on
2082 /// the IR via the TypePromotionTransaction.
2083 class TypePromotionAction {
2085 /// The Instruction modified.
2089 /// \brief Constructor of the action.
2090 /// The constructor performs the related action on the IR.
2091 TypePromotionAction(Instruction *Inst) : Inst(Inst) {}
2093 virtual ~TypePromotionAction() {}
2095 /// \brief Undo the modification done by this action.
2096 /// When this method is called, the IR must be in the same state as it was
2097 /// before this action was applied.
2098 /// \pre Undoing the action works if and only if the IR is in the exact same
2099 /// state as it was directly after this action was applied.
2100 virtual void undo() = 0;
2102 /// \brief Advocate every change made by this action.
2103 /// When the results on the IR of the action are to be kept, it is important
2104 /// to call this function, otherwise hidden information may be kept forever.
2105 virtual void commit() {
2106 // Nothing to be done, this action is not doing anything.
2110 /// \brief Utility to remember the position of an instruction.
2111 class InsertionHandler {
2112 /// Position of an instruction.
2113 /// Either an instruction:
2114 /// - Is the first in a basic block: BB is used.
2115 /// - Has a previous instructon: PrevInst is used.
2117 Instruction *PrevInst;
2120 /// Remember whether or not the instruction had a previous instruction.
2121 bool HasPrevInstruction;
2124 /// \brief Record the position of \p Inst.
2125 InsertionHandler(Instruction *Inst) {
2126 BasicBlock::iterator It = Inst->getIterator();
2127 HasPrevInstruction = (It != (Inst->getParent()->begin()));
2128 if (HasPrevInstruction)
2129 Point.PrevInst = &*--It;
2131 Point.BB = Inst->getParent();
2134 /// \brief Insert \p Inst at the recorded position.
2135 void insert(Instruction *Inst) {
2136 if (HasPrevInstruction) {
2137 if (Inst->getParent())
2138 Inst->removeFromParent();
2139 Inst->insertAfter(Point.PrevInst);
2141 Instruction *Position = &*Point.BB->getFirstInsertionPt();
2142 if (Inst->getParent())
2143 Inst->moveBefore(Position);
2145 Inst->insertBefore(Position);
2150 /// \brief Move an instruction before another.
2151 class InstructionMoveBefore : public TypePromotionAction {
2152 /// Original position of the instruction.
2153 InsertionHandler Position;
2156 /// \brief Move \p Inst before \p Before.
2157 InstructionMoveBefore(Instruction *Inst, Instruction *Before)
2158 : TypePromotionAction(Inst), Position(Inst) {
2159 DEBUG(dbgs() << "Do: move: " << *Inst << "\nbefore: " << *Before << "\n");
2160 Inst->moveBefore(Before);
2163 /// \brief Move the instruction back to its original position.
2164 void undo() override {
2165 DEBUG(dbgs() << "Undo: moveBefore: " << *Inst << "\n");
2166 Position.insert(Inst);
2170 /// \brief Set the operand of an instruction with a new value.
2171 class OperandSetter : public TypePromotionAction {
2172 /// Original operand of the instruction.
2174 /// Index of the modified instruction.
2178 /// \brief Set \p Idx operand of \p Inst with \p NewVal.
2179 OperandSetter(Instruction *Inst, unsigned Idx, Value *NewVal)
2180 : TypePromotionAction(Inst), Idx(Idx) {
2181 DEBUG(dbgs() << "Do: setOperand: " << Idx << "\n"
2182 << "for:" << *Inst << "\n"
2183 << "with:" << *NewVal << "\n");
2184 Origin = Inst->getOperand(Idx);
2185 Inst->setOperand(Idx, NewVal);
2188 /// \brief Restore the original value of the instruction.
2189 void undo() override {
2190 DEBUG(dbgs() << "Undo: setOperand:" << Idx << "\n"
2191 << "for: " << *Inst << "\n"
2192 << "with: " << *Origin << "\n");
2193 Inst->setOperand(Idx, Origin);
2197 /// \brief Hide the operands of an instruction.
2198 /// Do as if this instruction was not using any of its operands.
2199 class OperandsHider : public TypePromotionAction {
2200 /// The list of original operands.
2201 SmallVector<Value *, 4> OriginalValues;
2204 /// \brief Remove \p Inst from the uses of the operands of \p Inst.
2205 OperandsHider(Instruction *Inst) : TypePromotionAction(Inst) {
2206 DEBUG(dbgs() << "Do: OperandsHider: " << *Inst << "\n");
2207 unsigned NumOpnds = Inst->getNumOperands();
2208 OriginalValues.reserve(NumOpnds);
2209 for (unsigned It = 0; It < NumOpnds; ++It) {
2210 // Save the current operand.
2211 Value *Val = Inst->getOperand(It);
2212 OriginalValues.push_back(Val);
2214 // We could use OperandSetter here, but that would imply an overhead
2215 // that we are not willing to pay.
2216 Inst->setOperand(It, UndefValue::get(Val->getType()));
2220 /// \brief Restore the original list of uses.
2221 void undo() override {
2222 DEBUG(dbgs() << "Undo: OperandsHider: " << *Inst << "\n");
2223 for (unsigned It = 0, EndIt = OriginalValues.size(); It != EndIt; ++It)
2224 Inst->setOperand(It, OriginalValues[It]);
2228 /// \brief Build a truncate instruction.
2229 class TruncBuilder : public TypePromotionAction {
2232 /// \brief Build a truncate instruction of \p Opnd producing a \p Ty
2234 /// trunc Opnd to Ty.
2235 TruncBuilder(Instruction *Opnd, Type *Ty) : TypePromotionAction(Opnd) {
2236 IRBuilder<> Builder(Opnd);
2237 Val = Builder.CreateTrunc(Opnd, Ty, "promoted");
2238 DEBUG(dbgs() << "Do: TruncBuilder: " << *Val << "\n");
2241 /// \brief Get the built value.
2242 Value *getBuiltValue() { return Val; }
2244 /// \brief Remove the built instruction.
2245 void undo() override {
2246 DEBUG(dbgs() << "Undo: TruncBuilder: " << *Val << "\n");
2247 if (Instruction *IVal = dyn_cast<Instruction>(Val))
2248 IVal->eraseFromParent();
2252 /// \brief Build a sign extension instruction.
2253 class SExtBuilder : public TypePromotionAction {
2256 /// \brief Build a sign extension instruction of \p Opnd producing a \p Ty
2258 /// sext Opnd to Ty.
2259 SExtBuilder(Instruction *InsertPt, Value *Opnd, Type *Ty)
2260 : TypePromotionAction(InsertPt) {
2261 IRBuilder<> Builder(InsertPt);
2262 Val = Builder.CreateSExt(Opnd, Ty, "promoted");
2263 DEBUG(dbgs() << "Do: SExtBuilder: " << *Val << "\n");
2266 /// \brief Get the built value.
2267 Value *getBuiltValue() { return Val; }
2269 /// \brief Remove the built instruction.
2270 void undo() override {
2271 DEBUG(dbgs() << "Undo: SExtBuilder: " << *Val << "\n");
2272 if (Instruction *IVal = dyn_cast<Instruction>(Val))
2273 IVal->eraseFromParent();
2277 /// \brief Build a zero extension instruction.
2278 class ZExtBuilder : public TypePromotionAction {
2281 /// \brief Build a zero extension instruction of \p Opnd producing a \p Ty
2283 /// zext Opnd to Ty.
2284 ZExtBuilder(Instruction *InsertPt, Value *Opnd, Type *Ty)
2285 : TypePromotionAction(InsertPt) {
2286 IRBuilder<> Builder(InsertPt);
2287 Val = Builder.CreateZExt(Opnd, Ty, "promoted");
2288 DEBUG(dbgs() << "Do: ZExtBuilder: " << *Val << "\n");
2291 /// \brief Get the built value.
2292 Value *getBuiltValue() { return Val; }
2294 /// \brief Remove the built instruction.
2295 void undo() override {
2296 DEBUG(dbgs() << "Undo: ZExtBuilder: " << *Val << "\n");
2297 if (Instruction *IVal = dyn_cast<Instruction>(Val))
2298 IVal->eraseFromParent();
2302 /// \brief Mutate an instruction to another type.
2303 class TypeMutator : public TypePromotionAction {
2304 /// Record the original type.
2308 /// \brief Mutate the type of \p Inst into \p NewTy.
2309 TypeMutator(Instruction *Inst, Type *NewTy)
2310 : TypePromotionAction(Inst), OrigTy(Inst->getType()) {
2311 DEBUG(dbgs() << "Do: MutateType: " << *Inst << " with " << *NewTy
2313 Inst->mutateType(NewTy);
2316 /// \brief Mutate the instruction back to its original type.
2317 void undo() override {
2318 DEBUG(dbgs() << "Undo: MutateType: " << *Inst << " with " << *OrigTy
2320 Inst->mutateType(OrigTy);
2324 /// \brief Replace the uses of an instruction by another instruction.
2325 class UsesReplacer : public TypePromotionAction {
2326 /// Helper structure to keep track of the replaced uses.
2327 struct InstructionAndIdx {
2328 /// The instruction using the instruction.
2330 /// The index where this instruction is used for Inst.
2332 InstructionAndIdx(Instruction *Inst, unsigned Idx)
2333 : Inst(Inst), Idx(Idx) {}
2336 /// Keep track of the original uses (pair Instruction, Index).
2337 SmallVector<InstructionAndIdx, 4> OriginalUses;
2338 typedef SmallVectorImpl<InstructionAndIdx>::iterator use_iterator;
2341 /// \brief Replace all the use of \p Inst by \p New.
2342 UsesReplacer(Instruction *Inst, Value *New) : TypePromotionAction(Inst) {
2343 DEBUG(dbgs() << "Do: UsersReplacer: " << *Inst << " with " << *New
2345 // Record the original uses.
2346 for (Use &U : Inst->uses()) {
2347 Instruction *UserI = cast<Instruction>(U.getUser());
2348 OriginalUses.push_back(InstructionAndIdx(UserI, U.getOperandNo()));
2350 // Now, we can replace the uses.
2351 Inst->replaceAllUsesWith(New);
2354 /// \brief Reassign the original uses of Inst to Inst.
2355 void undo() override {
2356 DEBUG(dbgs() << "Undo: UsersReplacer: " << *Inst << "\n");
2357 for (use_iterator UseIt = OriginalUses.begin(),
2358 EndIt = OriginalUses.end();
2359 UseIt != EndIt; ++UseIt) {
2360 UseIt->Inst->setOperand(UseIt->Idx, Inst);
2365 /// \brief Remove an instruction from the IR.
2366 class InstructionRemover : public TypePromotionAction {
2367 /// Original position of the instruction.
2368 InsertionHandler Inserter;
2369 /// Helper structure to hide all the link to the instruction. In other
2370 /// words, this helps to do as if the instruction was removed.
2371 OperandsHider Hider;
2372 /// Keep track of the uses replaced, if any.
2373 UsesReplacer *Replacer;
2376 /// \brief Remove all reference of \p Inst and optinally replace all its
2378 /// \pre If !Inst->use_empty(), then New != nullptr
2379 InstructionRemover(Instruction *Inst, Value *New = nullptr)
2380 : TypePromotionAction(Inst), Inserter(Inst), Hider(Inst),
2383 Replacer = new UsesReplacer(Inst, New);
2384 DEBUG(dbgs() << "Do: InstructionRemover: " << *Inst << "\n");
2385 Inst->removeFromParent();
2388 ~InstructionRemover() override { delete Replacer; }
2390 /// \brief Really remove the instruction.
2391 void commit() override { delete Inst; }
2393 /// \brief Resurrect the instruction and reassign it to the proper uses if
2394 /// new value was provided when build this action.
2395 void undo() override {
2396 DEBUG(dbgs() << "Undo: InstructionRemover: " << *Inst << "\n");
2397 Inserter.insert(Inst);
2405 /// Restoration point.
2406 /// The restoration point is a pointer to an action instead of an iterator
2407 /// because the iterator may be invalidated but not the pointer.
2408 typedef const TypePromotionAction *ConstRestorationPt;
2409 /// Advocate every changes made in that transaction.
2411 /// Undo all the changes made after the given point.
2412 void rollback(ConstRestorationPt Point);
2413 /// Get the current restoration point.
2414 ConstRestorationPt getRestorationPoint() const;
2416 /// \name API for IR modification with state keeping to support rollback.
2418 /// Same as Instruction::setOperand.
2419 void setOperand(Instruction *Inst, unsigned Idx, Value *NewVal);
2420 /// Same as Instruction::eraseFromParent.
2421 void eraseInstruction(Instruction *Inst, Value *NewVal = nullptr);
2422 /// Same as Value::replaceAllUsesWith.
2423 void replaceAllUsesWith(Instruction *Inst, Value *New);
2424 /// Same as Value::mutateType.
2425 void mutateType(Instruction *Inst, Type *NewTy);
2426 /// Same as IRBuilder::createTrunc.
2427 Value *createTrunc(Instruction *Opnd, Type *Ty);
2428 /// Same as IRBuilder::createSExt.
2429 Value *createSExt(Instruction *Inst, Value *Opnd, Type *Ty);
2430 /// Same as IRBuilder::createZExt.
2431 Value *createZExt(Instruction *Inst, Value *Opnd, Type *Ty);
2432 /// Same as Instruction::moveBefore.
2433 void moveBefore(Instruction *Inst, Instruction *Before);
2437 /// The ordered list of actions made so far.
2438 SmallVector<std::unique_ptr<TypePromotionAction>, 16> Actions;
2439 typedef SmallVectorImpl<std::unique_ptr<TypePromotionAction>>::iterator CommitPt;
2442 void TypePromotionTransaction::setOperand(Instruction *Inst, unsigned Idx,
2445 make_unique<TypePromotionTransaction::OperandSetter>(Inst, Idx, NewVal));
2448 void TypePromotionTransaction::eraseInstruction(Instruction *Inst,
2451 make_unique<TypePromotionTransaction::InstructionRemover>(Inst, NewVal));
2454 void TypePromotionTransaction::replaceAllUsesWith(Instruction *Inst,
2456 Actions.push_back(make_unique<TypePromotionTransaction::UsesReplacer>(Inst, New));
2459 void TypePromotionTransaction::mutateType(Instruction *Inst, Type *NewTy) {
2460 Actions.push_back(make_unique<TypePromotionTransaction::TypeMutator>(Inst, NewTy));
2463 Value *TypePromotionTransaction::createTrunc(Instruction *Opnd,
2465 std::unique_ptr<TruncBuilder> Ptr(new TruncBuilder(Opnd, Ty));
2466 Value *Val = Ptr->getBuiltValue();
2467 Actions.push_back(std::move(Ptr));
2471 Value *TypePromotionTransaction::createSExt(Instruction *Inst,
2472 Value *Opnd, Type *Ty) {
2473 std::unique_ptr<SExtBuilder> Ptr(new SExtBuilder(Inst, Opnd, Ty));
2474 Value *Val = Ptr->getBuiltValue();
2475 Actions.push_back(std::move(Ptr));
2479 Value *TypePromotionTransaction::createZExt(Instruction *Inst,
2480 Value *Opnd, Type *Ty) {
2481 std::unique_ptr<ZExtBuilder> Ptr(new ZExtBuilder(Inst, Opnd, Ty));
2482 Value *Val = Ptr->getBuiltValue();
2483 Actions.push_back(std::move(Ptr));
2487 void TypePromotionTransaction::moveBefore(Instruction *Inst,
2488 Instruction *Before) {
2490 make_unique<TypePromotionTransaction::InstructionMoveBefore>(Inst, Before));
2493 TypePromotionTransaction::ConstRestorationPt
2494 TypePromotionTransaction::getRestorationPoint() const {
2495 return !Actions.empty() ? Actions.back().get() : nullptr;
2498 void TypePromotionTransaction::commit() {
2499 for (CommitPt It = Actions.begin(), EndIt = Actions.end(); It != EndIt;
2505 void TypePromotionTransaction::rollback(
2506 TypePromotionTransaction::ConstRestorationPt Point) {
2507 while (!Actions.empty() && Point != Actions.back().get()) {
2508 std::unique_ptr<TypePromotionAction> Curr = Actions.pop_back_val();
2513 /// \brief A helper class for matching addressing modes.
2515 /// This encapsulates the logic for matching the target-legal addressing modes.
2516 class AddressingModeMatcher {
2517 SmallVectorImpl<Instruction*> &AddrModeInsts;
2518 const TargetMachine &TM;
2519 const TargetLowering &TLI;
2520 const DataLayout &DL;
2522 /// AccessTy/MemoryInst - This is the type for the access (e.g. double) and
2523 /// the memory instruction that we're computing this address for.
2526 Instruction *MemoryInst;
2528 /// This is the addressing mode that we're building up. This is
2529 /// part of the return value of this addressing mode matching stuff.
2530 ExtAddrMode &AddrMode;
2532 /// The instructions inserted by other CodeGenPrepare optimizations.
2533 const SetOfInstrs &InsertedInsts;
2534 /// A map from the instructions to their type before promotion.
2535 InstrToOrigTy &PromotedInsts;
2536 /// The ongoing transaction where every action should be registered.
2537 TypePromotionTransaction &TPT;
2539 /// This is set to true when we should not do profitability checks.
2540 /// When true, IsProfitableToFoldIntoAddressingMode always returns true.
2541 bool IgnoreProfitability;
2543 AddressingModeMatcher(SmallVectorImpl<Instruction *> &AMI,
2544 const TargetMachine &TM, Type *AT, unsigned AS,
2545 Instruction *MI, ExtAddrMode &AM,
2546 const SetOfInstrs &InsertedInsts,
2547 InstrToOrigTy &PromotedInsts,
2548 TypePromotionTransaction &TPT)
2549 : AddrModeInsts(AMI), TM(TM),
2550 TLI(*TM.getSubtargetImpl(*MI->getParent()->getParent())
2551 ->getTargetLowering()),
2552 DL(MI->getModule()->getDataLayout()), AccessTy(AT), AddrSpace(AS),
2553 MemoryInst(MI), AddrMode(AM), InsertedInsts(InsertedInsts),
2554 PromotedInsts(PromotedInsts), TPT(TPT) {
2555 IgnoreProfitability = false;
2559 /// Find the maximal addressing mode that a load/store of V can fold,
2560 /// give an access type of AccessTy. This returns a list of involved
2561 /// instructions in AddrModeInsts.
2562 /// \p InsertedInsts The instructions inserted by other CodeGenPrepare
2564 /// \p PromotedInsts maps the instructions to their type before promotion.
2565 /// \p The ongoing transaction where every action should be registered.
2566 static ExtAddrMode Match(Value *V, Type *AccessTy, unsigned AS,
2567 Instruction *MemoryInst,
2568 SmallVectorImpl<Instruction*> &AddrModeInsts,
2569 const TargetMachine &TM,
2570 const SetOfInstrs &InsertedInsts,
2571 InstrToOrigTy &PromotedInsts,
2572 TypePromotionTransaction &TPT) {
2575 bool Success = AddressingModeMatcher(AddrModeInsts, TM, AccessTy, AS,
2576 MemoryInst, Result, InsertedInsts,
2577 PromotedInsts, TPT).matchAddr(V, 0);
2578 (void)Success; assert(Success && "Couldn't select *anything*?");
2582 bool matchScaledValue(Value *ScaleReg, int64_t Scale, unsigned Depth);
2583 bool matchAddr(Value *V, unsigned Depth);
2584 bool matchOperationAddr(User *Operation, unsigned Opcode, unsigned Depth,
2585 bool *MovedAway = nullptr);
2586 bool isProfitableToFoldIntoAddressingMode(Instruction *I,
2587 ExtAddrMode &AMBefore,
2588 ExtAddrMode &AMAfter);
2589 bool valueAlreadyLiveAtInst(Value *Val, Value *KnownLive1, Value *KnownLive2);
2590 bool isPromotionProfitable(unsigned NewCost, unsigned OldCost,
2591 Value *PromotedOperand) const;
2594 /// Try adding ScaleReg*Scale to the current addressing mode.
2595 /// Return true and update AddrMode if this addr mode is legal for the target,
2597 bool AddressingModeMatcher::matchScaledValue(Value *ScaleReg, int64_t Scale,
2599 // If Scale is 1, then this is the same as adding ScaleReg to the addressing
2600 // mode. Just process that directly.
2602 return matchAddr(ScaleReg, Depth);
2604 // If the scale is 0, it takes nothing to add this.
2608 // If we already have a scale of this value, we can add to it, otherwise, we
2609 // need an available scale field.
2610 if (AddrMode.Scale != 0 && AddrMode.ScaledReg != ScaleReg)
2613 ExtAddrMode TestAddrMode = AddrMode;
2615 // Add scale to turn X*4+X*3 -> X*7. This could also do things like
2616 // [A+B + A*7] -> [B+A*8].
2617 TestAddrMode.Scale += Scale;
2618 TestAddrMode.ScaledReg = ScaleReg;
2620 // If the new address isn't legal, bail out.
2621 if (!TLI.isLegalAddressingMode(DL, TestAddrMode, AccessTy, AddrSpace))
2624 // It was legal, so commit it.
2625 AddrMode = TestAddrMode;
2627 // Okay, we decided that we can add ScaleReg+Scale to AddrMode. Check now
2628 // to see if ScaleReg is actually X+C. If so, we can turn this into adding
2629 // X*Scale + C*Scale to addr mode.
2630 ConstantInt *CI = nullptr; Value *AddLHS = nullptr;
2631 if (isa<Instruction>(ScaleReg) && // not a constant expr.
2632 match(ScaleReg, m_Add(m_Value(AddLHS), m_ConstantInt(CI)))) {
2633 TestAddrMode.ScaledReg = AddLHS;
2634 TestAddrMode.BaseOffs += CI->getSExtValue()*TestAddrMode.Scale;
2636 // If this addressing mode is legal, commit it and remember that we folded
2637 // this instruction.
2638 if (TLI.isLegalAddressingMode(DL, TestAddrMode, AccessTy, AddrSpace)) {
2639 AddrModeInsts.push_back(cast<Instruction>(ScaleReg));
2640 AddrMode = TestAddrMode;
2645 // Otherwise, not (x+c)*scale, just return what we have.
2649 /// This is a little filter, which returns true if an addressing computation
2650 /// involving I might be folded into a load/store accessing it.
2651 /// This doesn't need to be perfect, but needs to accept at least
2652 /// the set of instructions that MatchOperationAddr can.
2653 static bool MightBeFoldableInst(Instruction *I) {
2654 switch (I->getOpcode()) {
2655 case Instruction::BitCast:
2656 case Instruction::AddrSpaceCast:
2657 // Don't touch identity bitcasts.
2658 if (I->getType() == I->getOperand(0)->getType())
2660 return I->getType()->isPointerTy() || I->getType()->isIntegerTy();
2661 case Instruction::PtrToInt:
2662 // PtrToInt is always a noop, as we know that the int type is pointer sized.
2664 case Instruction::IntToPtr:
2665 // We know the input is intptr_t, so this is foldable.
2667 case Instruction::Add:
2669 case Instruction::Mul:
2670 case Instruction::Shl:
2671 // Can only handle X*C and X << C.
2672 return isa<ConstantInt>(I->getOperand(1));
2673 case Instruction::GetElementPtr:
2680 /// \brief Check whether or not \p Val is a legal instruction for \p TLI.
2681 /// \note \p Val is assumed to be the product of some type promotion.
2682 /// Therefore if \p Val has an undefined state in \p TLI, this is assumed
2683 /// to be legal, as the non-promoted value would have had the same state.
2684 static bool isPromotedInstructionLegal(const TargetLowering &TLI,
2685 const DataLayout &DL, Value *Val) {
2686 Instruction *PromotedInst = dyn_cast<Instruction>(Val);
2689 int ISDOpcode = TLI.InstructionOpcodeToISD(PromotedInst->getOpcode());
2690 // If the ISDOpcode is undefined, it was undefined before the promotion.
2693 // Otherwise, check if the promoted instruction is legal or not.
2694 return TLI.isOperationLegalOrCustom(
2695 ISDOpcode, TLI.getValueType(DL, PromotedInst->getType()));
2698 /// \brief Hepler class to perform type promotion.
2699 class TypePromotionHelper {
2700 /// \brief Utility function to check whether or not a sign or zero extension
2701 /// of \p Inst with \p ConsideredExtType can be moved through \p Inst by
2702 /// either using the operands of \p Inst or promoting \p Inst.
2703 /// The type of the extension is defined by \p IsSExt.
2704 /// In other words, check if:
2705 /// ext (Ty Inst opnd1 opnd2 ... opndN) to ConsideredExtType.
2706 /// #1 Promotion applies:
2707 /// ConsideredExtType Inst (ext opnd1 to ConsideredExtType, ...).
2708 /// #2 Operand reuses:
2709 /// ext opnd1 to ConsideredExtType.
2710 /// \p PromotedInsts maps the instructions to their type before promotion.
2711 static bool canGetThrough(const Instruction *Inst, Type *ConsideredExtType,
2712 const InstrToOrigTy &PromotedInsts, bool IsSExt);
2714 /// \brief Utility function to determine if \p OpIdx should be promoted when
2715 /// promoting \p Inst.
2716 static bool shouldExtOperand(const Instruction *Inst, int OpIdx) {
2717 return !(isa<SelectInst>(Inst) && OpIdx == 0);
2720 /// \brief Utility function to promote the operand of \p Ext when this
2721 /// operand is a promotable trunc or sext or zext.
2722 /// \p PromotedInsts maps the instructions to their type before promotion.
2723 /// \p CreatedInstsCost[out] contains the cost of all instructions
2724 /// created to promote the operand of Ext.
2725 /// Newly added extensions are inserted in \p Exts.
2726 /// Newly added truncates are inserted in \p Truncs.
2727 /// Should never be called directly.
2728 /// \return The promoted value which is used instead of Ext.
2729 static Value *promoteOperandForTruncAndAnyExt(
2730 Instruction *Ext, TypePromotionTransaction &TPT,
2731 InstrToOrigTy &PromotedInsts, unsigned &CreatedInstsCost,
2732 SmallVectorImpl<Instruction *> *Exts,
2733 SmallVectorImpl<Instruction *> *Truncs, const TargetLowering &TLI);
2735 /// \brief Utility function to promote the operand of \p Ext when this
2736 /// operand is promotable and is not a supported trunc or sext.
2737 /// \p PromotedInsts maps the instructions to their type before promotion.
2738 /// \p CreatedInstsCost[out] contains the cost of all the 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 *promoteOperandForOther(Instruction *Ext,
2745 TypePromotionTransaction &TPT,
2746 InstrToOrigTy &PromotedInsts,
2747 unsigned &CreatedInstsCost,
2748 SmallVectorImpl<Instruction *> *Exts,
2749 SmallVectorImpl<Instruction *> *Truncs,
2750 const TargetLowering &TLI, bool IsSExt);
2752 /// \see promoteOperandForOther.
2753 static Value *signExtendOperandForOther(
2754 Instruction *Ext, TypePromotionTransaction &TPT,
2755 InstrToOrigTy &PromotedInsts, unsigned &CreatedInstsCost,
2756 SmallVectorImpl<Instruction *> *Exts,
2757 SmallVectorImpl<Instruction *> *Truncs, const TargetLowering &TLI) {
2758 return promoteOperandForOther(Ext, TPT, PromotedInsts, CreatedInstsCost,
2759 Exts, Truncs, TLI, true);
2762 /// \see promoteOperandForOther.
2763 static Value *zeroExtendOperandForOther(
2764 Instruction *Ext, TypePromotionTransaction &TPT,
2765 InstrToOrigTy &PromotedInsts, unsigned &CreatedInstsCost,
2766 SmallVectorImpl<Instruction *> *Exts,
2767 SmallVectorImpl<Instruction *> *Truncs, const TargetLowering &TLI) {
2768 return promoteOperandForOther(Ext, TPT, PromotedInsts, CreatedInstsCost,
2769 Exts, Truncs, TLI, false);
2773 /// Type for the utility function that promotes the operand of Ext.
2774 typedef Value *(*Action)(Instruction *Ext, TypePromotionTransaction &TPT,
2775 InstrToOrigTy &PromotedInsts,
2776 unsigned &CreatedInstsCost,
2777 SmallVectorImpl<Instruction *> *Exts,
2778 SmallVectorImpl<Instruction *> *Truncs,
2779 const TargetLowering &TLI);
2780 /// \brief Given a sign/zero extend instruction \p Ext, return the approriate
2781 /// action to promote the operand of \p Ext instead of using Ext.
2782 /// \return NULL if no promotable action is possible with the current
2784 /// \p InsertedInsts keeps track of all the instructions inserted by the
2785 /// other CodeGenPrepare optimizations. This information is important
2786 /// because we do not want to promote these instructions as CodeGenPrepare
2787 /// will reinsert them later. Thus creating an infinite loop: create/remove.
2788 /// \p PromotedInsts maps the instructions to their type before promotion.
2789 static Action getAction(Instruction *Ext, const SetOfInstrs &InsertedInsts,
2790 const TargetLowering &TLI,
2791 const InstrToOrigTy &PromotedInsts);
2794 bool TypePromotionHelper::canGetThrough(const Instruction *Inst,
2795 Type *ConsideredExtType,
2796 const InstrToOrigTy &PromotedInsts,
2798 // The promotion helper does not know how to deal with vector types yet.
2799 // To be able to fix that, we would need to fix the places where we
2800 // statically extend, e.g., constants and such.
2801 if (Inst->getType()->isVectorTy())
2804 // We can always get through zext.
2805 if (isa<ZExtInst>(Inst))
2808 // sext(sext) is ok too.
2809 if (IsSExt && isa<SExtInst>(Inst))
2812 // We can get through binary operator, if it is legal. In other words, the
2813 // binary operator must have a nuw or nsw flag.
2814 const BinaryOperator *BinOp = dyn_cast<BinaryOperator>(Inst);
2815 if (BinOp && isa<OverflowingBinaryOperator>(BinOp) &&
2816 ((!IsSExt && BinOp->hasNoUnsignedWrap()) ||
2817 (IsSExt && BinOp->hasNoSignedWrap())))
2820 // Check if we can do the following simplification.
2821 // ext(trunc(opnd)) --> ext(opnd)
2822 if (!isa<TruncInst>(Inst))
2825 Value *OpndVal = Inst->getOperand(0);
2826 // Check if we can use this operand in the extension.
2827 // If the type is larger than the result type of the extension, we cannot.
2828 if (!OpndVal->getType()->isIntegerTy() ||
2829 OpndVal->getType()->getIntegerBitWidth() >
2830 ConsideredExtType->getIntegerBitWidth())
2833 // If the operand of the truncate is not an instruction, we will not have
2834 // any information on the dropped bits.
2835 // (Actually we could for constant but it is not worth the extra logic).
2836 Instruction *Opnd = dyn_cast<Instruction>(OpndVal);
2840 // Check if the source of the type is narrow enough.
2841 // I.e., check that trunc just drops extended bits of the same kind of
2843 // #1 get the type of the operand and check the kind of the extended bits.
2844 const Type *OpndType;
2845 InstrToOrigTy::const_iterator It = PromotedInsts.find(Opnd);
2846 if (It != PromotedInsts.end() && It->second.getInt() == IsSExt)
2847 OpndType = It->second.getPointer();
2848 else if ((IsSExt && isa<SExtInst>(Opnd)) || (!IsSExt && isa<ZExtInst>(Opnd)))
2849 OpndType = Opnd->getOperand(0)->getType();
2853 // #2 check that the truncate just drops extended bits.
2854 return Inst->getType()->getIntegerBitWidth() >=
2855 OpndType->getIntegerBitWidth();
2858 TypePromotionHelper::Action TypePromotionHelper::getAction(
2859 Instruction *Ext, const SetOfInstrs &InsertedInsts,
2860 const TargetLowering &TLI, const InstrToOrigTy &PromotedInsts) {
2861 assert((isa<SExtInst>(Ext) || isa<ZExtInst>(Ext)) &&
2862 "Unexpected instruction type");
2863 Instruction *ExtOpnd = dyn_cast<Instruction>(Ext->getOperand(0));
2864 Type *ExtTy = Ext->getType();
2865 bool IsSExt = isa<SExtInst>(Ext);
2866 // If the operand of the extension is not an instruction, we cannot
2868 // If it, check we can get through.
2869 if (!ExtOpnd || !canGetThrough(ExtOpnd, ExtTy, PromotedInsts, IsSExt))
2872 // Do not promote if the operand has been added by codegenprepare.
2873 // Otherwise, it means we are undoing an optimization that is likely to be
2874 // redone, thus causing potential infinite loop.
2875 if (isa<TruncInst>(ExtOpnd) && InsertedInsts.count(ExtOpnd))
2878 // SExt or Trunc instructions.
2879 // Return the related handler.
2880 if (isa<SExtInst>(ExtOpnd) || isa<TruncInst>(ExtOpnd) ||
2881 isa<ZExtInst>(ExtOpnd))
2882 return promoteOperandForTruncAndAnyExt;
2884 // Regular instruction.
2885 // Abort early if we will have to insert non-free instructions.
2886 if (!ExtOpnd->hasOneUse() && !TLI.isTruncateFree(ExtTy, ExtOpnd->getType()))
2888 return IsSExt ? signExtendOperandForOther : zeroExtendOperandForOther;
2891 Value *TypePromotionHelper::promoteOperandForTruncAndAnyExt(
2892 llvm::Instruction *SExt, TypePromotionTransaction &TPT,
2893 InstrToOrigTy &PromotedInsts, unsigned &CreatedInstsCost,
2894 SmallVectorImpl<Instruction *> *Exts,
2895 SmallVectorImpl<Instruction *> *Truncs, const TargetLowering &TLI) {
2896 // By construction, the operand of SExt is an instruction. Otherwise we cannot
2897 // get through it and this method should not be called.
2898 Instruction *SExtOpnd = cast<Instruction>(SExt->getOperand(0));
2899 Value *ExtVal = SExt;
2900 bool HasMergedNonFreeExt = false;
2901 if (isa<ZExtInst>(SExtOpnd)) {
2902 // Replace s|zext(zext(opnd))
2904 HasMergedNonFreeExt = !TLI.isExtFree(SExtOpnd);
2906 TPT.createZExt(SExt, SExtOpnd->getOperand(0), SExt->getType());
2907 TPT.replaceAllUsesWith(SExt, ZExt);
2908 TPT.eraseInstruction(SExt);
2911 // Replace z|sext(trunc(opnd)) or sext(sext(opnd))
2913 TPT.setOperand(SExt, 0, SExtOpnd->getOperand(0));
2915 CreatedInstsCost = 0;
2917 // Remove dead code.
2918 if (SExtOpnd->use_empty())
2919 TPT.eraseInstruction(SExtOpnd);
2921 // Check if the extension is still needed.
2922 Instruction *ExtInst = dyn_cast<Instruction>(ExtVal);
2923 if (!ExtInst || ExtInst->getType() != ExtInst->getOperand(0)->getType()) {
2926 Exts->push_back(ExtInst);
2927 CreatedInstsCost = !TLI.isExtFree(ExtInst) && !HasMergedNonFreeExt;
2932 // At this point we have: ext ty opnd to ty.
2933 // Reassign the uses of ExtInst to the opnd and remove ExtInst.
2934 Value *NextVal = ExtInst->getOperand(0);
2935 TPT.eraseInstruction(ExtInst, NextVal);
2939 Value *TypePromotionHelper::promoteOperandForOther(
2940 Instruction *Ext, TypePromotionTransaction &TPT,
2941 InstrToOrigTy &PromotedInsts, unsigned &CreatedInstsCost,
2942 SmallVectorImpl<Instruction *> *Exts,
2943 SmallVectorImpl<Instruction *> *Truncs, const TargetLowering &TLI,
2945 // By construction, the operand of Ext is an instruction. Otherwise we cannot
2946 // get through it and this method should not be called.
2947 Instruction *ExtOpnd = cast<Instruction>(Ext->getOperand(0));
2948 CreatedInstsCost = 0;
2949 if (!ExtOpnd->hasOneUse()) {
2950 // ExtOpnd will be promoted.
2951 // All its uses, but Ext, will need to use a truncated value of the
2952 // promoted version.
2953 // Create the truncate now.
2954 Value *Trunc = TPT.createTrunc(Ext, ExtOpnd->getType());
2955 if (Instruction *ITrunc = dyn_cast<Instruction>(Trunc)) {
2956 ITrunc->removeFromParent();
2957 // Insert it just after the definition.
2958 ITrunc->insertAfter(ExtOpnd);
2960 Truncs->push_back(ITrunc);
2963 TPT.replaceAllUsesWith(ExtOpnd, Trunc);
2964 // Restore the operand of Ext (which has been replaced by the previous call
2965 // to replaceAllUsesWith) to avoid creating a cycle trunc <-> sext.
2966 TPT.setOperand(Ext, 0, ExtOpnd);
2969 // Get through the Instruction:
2970 // 1. Update its type.
2971 // 2. Replace the uses of Ext by Inst.
2972 // 3. Extend each operand that needs to be extended.
2974 // Remember the original type of the instruction before promotion.
2975 // This is useful to know that the high bits are sign extended bits.
2976 PromotedInsts.insert(std::pair<Instruction *, TypeIsSExt>(
2977 ExtOpnd, TypeIsSExt(ExtOpnd->getType(), IsSExt)));
2979 TPT.mutateType(ExtOpnd, Ext->getType());
2981 TPT.replaceAllUsesWith(Ext, ExtOpnd);
2983 Instruction *ExtForOpnd = Ext;
2985 DEBUG(dbgs() << "Propagate Ext to operands\n");
2986 for (int OpIdx = 0, EndOpIdx = ExtOpnd->getNumOperands(); OpIdx != EndOpIdx;
2988 DEBUG(dbgs() << "Operand:\n" << *(ExtOpnd->getOperand(OpIdx)) << '\n');
2989 if (ExtOpnd->getOperand(OpIdx)->getType() == Ext->getType() ||
2990 !shouldExtOperand(ExtOpnd, OpIdx)) {
2991 DEBUG(dbgs() << "No need to propagate\n");
2994 // Check if we can statically extend the operand.
2995 Value *Opnd = ExtOpnd->getOperand(OpIdx);
2996 if (const ConstantInt *Cst = dyn_cast<ConstantInt>(Opnd)) {
2997 DEBUG(dbgs() << "Statically extend\n");
2998 unsigned BitWidth = Ext->getType()->getIntegerBitWidth();
2999 APInt CstVal = IsSExt ? Cst->getValue().sext(BitWidth)
3000 : Cst->getValue().zext(BitWidth);
3001 TPT.setOperand(ExtOpnd, OpIdx, ConstantInt::get(Ext->getType(), CstVal));
3004 // UndefValue are typed, so we have to statically sign extend them.
3005 if (isa<UndefValue>(Opnd)) {
3006 DEBUG(dbgs() << "Statically extend\n");
3007 TPT.setOperand(ExtOpnd, OpIdx, UndefValue::get(Ext->getType()));
3011 // Otherwise we have to explicity sign extend the operand.
3012 // Check if Ext was reused to extend an operand.
3014 // If yes, create a new one.
3015 DEBUG(dbgs() << "More operands to ext\n");
3016 Value *ValForExtOpnd = IsSExt ? TPT.createSExt(Ext, Opnd, Ext->getType())
3017 : TPT.createZExt(Ext, Opnd, Ext->getType());
3018 if (!isa<Instruction>(ValForExtOpnd)) {
3019 TPT.setOperand(ExtOpnd, OpIdx, ValForExtOpnd);
3022 ExtForOpnd = cast<Instruction>(ValForExtOpnd);
3025 Exts->push_back(ExtForOpnd);
3026 TPT.setOperand(ExtForOpnd, 0, Opnd);
3028 // Move the sign extension before the insertion point.
3029 TPT.moveBefore(ExtForOpnd, ExtOpnd);
3030 TPT.setOperand(ExtOpnd, OpIdx, ExtForOpnd);
3031 CreatedInstsCost += !TLI.isExtFree(ExtForOpnd);
3032 // If more sext are required, new instructions will have to be created.
3033 ExtForOpnd = nullptr;
3035 if (ExtForOpnd == Ext) {
3036 DEBUG(dbgs() << "Extension is useless now\n");
3037 TPT.eraseInstruction(Ext);
3042 /// Check whether or not promoting an instruction to a wider type is profitable.
3043 /// \p NewCost gives the cost of extension instructions created by the
3045 /// \p OldCost gives the cost of extension instructions before the promotion
3046 /// plus the number of instructions that have been
3047 /// matched in the addressing mode the promotion.
3048 /// \p PromotedOperand is the value that has been promoted.
3049 /// \return True if the promotion is profitable, false otherwise.
3050 bool AddressingModeMatcher::isPromotionProfitable(
3051 unsigned NewCost, unsigned OldCost, Value *PromotedOperand) const {
3052 DEBUG(dbgs() << "OldCost: " << OldCost << "\tNewCost: " << NewCost << '\n');
3053 // The cost of the new extensions is greater than the cost of the
3054 // old extension plus what we folded.
3055 // This is not profitable.
3056 if (NewCost > OldCost)
3058 if (NewCost < OldCost)
3060 // The promotion is neutral but it may help folding the sign extension in
3061 // loads for instance.
3062 // Check that we did not create an illegal instruction.
3063 return isPromotedInstructionLegal(TLI, DL, PromotedOperand);
3066 /// Given an instruction or constant expr, see if we can fold the operation
3067 /// into the addressing mode. If so, update the addressing mode and return
3068 /// true, otherwise return false without modifying AddrMode.
3069 /// If \p MovedAway is not NULL, it contains the information of whether or
3070 /// not AddrInst has to be folded into the addressing mode on success.
3071 /// If \p MovedAway == true, \p AddrInst will not be part of the addressing
3072 /// because it has been moved away.
3073 /// Thus AddrInst must not be added in the matched instructions.
3074 /// This state can happen when AddrInst is a sext, since it may be moved away.
3075 /// Therefore, AddrInst may not be valid when MovedAway is true and it must
3076 /// not be referenced anymore.
3077 bool AddressingModeMatcher::matchOperationAddr(User *AddrInst, unsigned Opcode,
3080 // Avoid exponential behavior on extremely deep expression trees.
3081 if (Depth >= 5) return false;
3083 // By default, all matched instructions stay in place.
3088 case Instruction::PtrToInt:
3089 // PtrToInt is always a noop, as we know that the int type is pointer sized.
3090 return matchAddr(AddrInst->getOperand(0), Depth);
3091 case Instruction::IntToPtr: {
3092 auto AS = AddrInst->getType()->getPointerAddressSpace();
3093 auto PtrTy = MVT::getIntegerVT(DL.getPointerSizeInBits(AS));
3094 // This inttoptr is a no-op if the integer type is pointer sized.
3095 if (TLI.getValueType(DL, AddrInst->getOperand(0)->getType()) == PtrTy)
3096 return matchAddr(AddrInst->getOperand(0), Depth);
3099 case Instruction::BitCast:
3100 // BitCast is always a noop, and we can handle it as long as it is
3101 // int->int or pointer->pointer (we don't want int<->fp or something).
3102 if ((AddrInst->getOperand(0)->getType()->isPointerTy() ||
3103 AddrInst->getOperand(0)->getType()->isIntegerTy()) &&
3104 // Don't touch identity bitcasts. These were probably put here by LSR,
3105 // and we don't want to mess around with them. Assume it knows what it
3107 AddrInst->getOperand(0)->getType() != AddrInst->getType())
3108 return matchAddr(AddrInst->getOperand(0), Depth);
3110 case Instruction::AddrSpaceCast: {
3112 = AddrInst->getOperand(0)->getType()->getPointerAddressSpace();
3113 unsigned DestAS = AddrInst->getType()->getPointerAddressSpace();
3114 if (TLI.isNoopAddrSpaceCast(SrcAS, DestAS))
3115 return matchAddr(AddrInst->getOperand(0), Depth);
3118 case Instruction::Add: {
3119 // Check to see if we can merge in the RHS then the LHS. If so, we win.
3120 ExtAddrMode BackupAddrMode = AddrMode;
3121 unsigned OldSize = AddrModeInsts.size();
3122 // Start a transaction at this point.
3123 // The LHS may match but not the RHS.
3124 // Therefore, we need a higher level restoration point to undo partially
3125 // matched operation.
3126 TypePromotionTransaction::ConstRestorationPt LastKnownGood =
3127 TPT.getRestorationPoint();
3129 if (matchAddr(AddrInst->getOperand(1), Depth+1) &&
3130 matchAddr(AddrInst->getOperand(0), Depth+1))
3133 // Restore the old addr mode info.
3134 AddrMode = BackupAddrMode;
3135 AddrModeInsts.resize(OldSize);
3136 TPT.rollback(LastKnownGood);
3138 // Otherwise this was over-aggressive. Try merging in the LHS then the RHS.
3139 if (matchAddr(AddrInst->getOperand(0), Depth+1) &&
3140 matchAddr(AddrInst->getOperand(1), Depth+1))
3143 // Otherwise we definitely can't merge the ADD in.
3144 AddrMode = BackupAddrMode;
3145 AddrModeInsts.resize(OldSize);
3146 TPT.rollback(LastKnownGood);
3149 //case Instruction::Or:
3150 // TODO: We can handle "Or Val, Imm" iff this OR is equivalent to an ADD.
3152 case Instruction::Mul:
3153 case Instruction::Shl: {
3154 // Can only handle X*C and X << C.
3155 ConstantInt *RHS = dyn_cast<ConstantInt>(AddrInst->getOperand(1));
3158 int64_t Scale = RHS->getSExtValue();
3159 if (Opcode == Instruction::Shl)
3160 Scale = 1LL << Scale;
3162 return matchScaledValue(AddrInst->getOperand(0), Scale, Depth);
3164 case Instruction::GetElementPtr: {
3165 // Scan the GEP. We check it if it contains constant offsets and at most
3166 // one variable offset.
3167 int VariableOperand = -1;
3168 unsigned VariableScale = 0;
3170 int64_t ConstantOffset = 0;
3171 gep_type_iterator GTI = gep_type_begin(AddrInst);
3172 for (unsigned i = 1, e = AddrInst->getNumOperands(); i != e; ++i, ++GTI) {
3173 if (StructType *STy = dyn_cast<StructType>(*GTI)) {
3174 const StructLayout *SL = DL.getStructLayout(STy);
3176 cast<ConstantInt>(AddrInst->getOperand(i))->getZExtValue();
3177 ConstantOffset += SL->getElementOffset(Idx);
3179 uint64_t TypeSize = DL.getTypeAllocSize(GTI.getIndexedType());
3180 if (ConstantInt *CI = dyn_cast<ConstantInt>(AddrInst->getOperand(i))) {
3181 ConstantOffset += CI->getSExtValue()*TypeSize;
3182 } else if (TypeSize) { // Scales of zero don't do anything.
3183 // We only allow one variable index at the moment.
3184 if (VariableOperand != -1)
3187 // Remember the variable index.
3188 VariableOperand = i;
3189 VariableScale = TypeSize;
3194 // A common case is for the GEP to only do a constant offset. In this case,
3195 // just add it to the disp field and check validity.
3196 if (VariableOperand == -1) {
3197 AddrMode.BaseOffs += ConstantOffset;
3198 if (ConstantOffset == 0 ||
3199 TLI.isLegalAddressingMode(DL, AddrMode, AccessTy, AddrSpace)) {
3200 // Check to see if we can fold the base pointer in too.
3201 if (matchAddr(AddrInst->getOperand(0), Depth+1))
3204 AddrMode.BaseOffs -= ConstantOffset;
3208 // Save the valid addressing mode in case we can't match.
3209 ExtAddrMode BackupAddrMode = AddrMode;
3210 unsigned OldSize = AddrModeInsts.size();
3212 // See if the scale and offset amount is valid for this target.
3213 AddrMode.BaseOffs += ConstantOffset;
3215 // Match the base operand of the GEP.
3216 if (!matchAddr(AddrInst->getOperand(0), Depth+1)) {
3217 // If it couldn't be matched, just stuff the value in a register.
3218 if (AddrMode.HasBaseReg) {
3219 AddrMode = BackupAddrMode;
3220 AddrModeInsts.resize(OldSize);
3223 AddrMode.HasBaseReg = true;
3224 AddrMode.BaseReg = AddrInst->getOperand(0);
3227 // Match the remaining variable portion of the GEP.
3228 if (!matchScaledValue(AddrInst->getOperand(VariableOperand), VariableScale,
3230 // If it couldn't be matched, try stuffing the base into a register
3231 // instead of matching it, and retrying the match of the scale.
3232 AddrMode = BackupAddrMode;
3233 AddrModeInsts.resize(OldSize);
3234 if (AddrMode.HasBaseReg)
3236 AddrMode.HasBaseReg = true;
3237 AddrMode.BaseReg = AddrInst->getOperand(0);
3238 AddrMode.BaseOffs += ConstantOffset;
3239 if (!matchScaledValue(AddrInst->getOperand(VariableOperand),
3240 VariableScale, Depth)) {
3241 // If even that didn't work, bail.
3242 AddrMode = BackupAddrMode;
3243 AddrModeInsts.resize(OldSize);
3250 case Instruction::SExt:
3251 case Instruction::ZExt: {
3252 Instruction *Ext = dyn_cast<Instruction>(AddrInst);
3256 // Try to move this ext out of the way of the addressing mode.
3257 // Ask for a method for doing so.
3258 TypePromotionHelper::Action TPH =
3259 TypePromotionHelper::getAction(Ext, InsertedInsts, TLI, PromotedInsts);
3263 TypePromotionTransaction::ConstRestorationPt LastKnownGood =
3264 TPT.getRestorationPoint();
3265 unsigned CreatedInstsCost = 0;
3266 unsigned ExtCost = !TLI.isExtFree(Ext);
3267 Value *PromotedOperand =
3268 TPH(Ext, TPT, PromotedInsts, CreatedInstsCost, nullptr, nullptr, TLI);
3269 // SExt has been moved away.
3270 // Thus either it will be rematched later in the recursive calls or it is
3271 // gone. Anyway, we must not fold it into the addressing mode at this point.
3275 // addr = gep base, idx
3277 // promotedOpnd = ext opnd <- no match here
3278 // op = promoted_add promotedOpnd, 1 <- match (later in recursive calls)
3279 // addr = gep base, op <- match
3283 assert(PromotedOperand &&
3284 "TypePromotionHelper should have filtered out those cases");
3286 ExtAddrMode BackupAddrMode = AddrMode;
3287 unsigned OldSize = AddrModeInsts.size();
3289 if (!matchAddr(PromotedOperand, Depth) ||
3290 // The total of the new cost is equal to the cost of the created
3292 // The total of the old cost is equal to the cost of the extension plus
3293 // what we have saved in the addressing mode.
3294 !isPromotionProfitable(CreatedInstsCost,
3295 ExtCost + (AddrModeInsts.size() - OldSize),
3297 AddrMode = BackupAddrMode;
3298 AddrModeInsts.resize(OldSize);
3299 DEBUG(dbgs() << "Sign extension does not pay off: rollback\n");
3300 TPT.rollback(LastKnownGood);
3309 /// If we can, try to add the value of 'Addr' into the current addressing mode.
3310 /// If Addr can't be added to AddrMode this returns false and leaves AddrMode
3311 /// unmodified. This assumes that Addr is either a pointer type or intptr_t
3314 bool AddressingModeMatcher::matchAddr(Value *Addr, unsigned Depth) {
3315 // Start a transaction at this point that we will rollback if the matching
3317 TypePromotionTransaction::ConstRestorationPt LastKnownGood =
3318 TPT.getRestorationPoint();
3319 if (ConstantInt *CI = dyn_cast<ConstantInt>(Addr)) {
3320 // Fold in immediates if legal for the target.
3321 AddrMode.BaseOffs += CI->getSExtValue();
3322 if (TLI.isLegalAddressingMode(DL, AddrMode, AccessTy, AddrSpace))
3324 AddrMode.BaseOffs -= CI->getSExtValue();
3325 } else if (GlobalValue *GV = dyn_cast<GlobalValue>(Addr)) {
3326 // If this is a global variable, try to fold it into the addressing mode.
3327 if (!AddrMode.BaseGV) {
3328 AddrMode.BaseGV = GV;
3329 if (TLI.isLegalAddressingMode(DL, AddrMode, AccessTy, AddrSpace))
3331 AddrMode.BaseGV = nullptr;
3333 } else if (Instruction *I = dyn_cast<Instruction>(Addr)) {
3334 ExtAddrMode BackupAddrMode = AddrMode;
3335 unsigned OldSize = AddrModeInsts.size();
3337 // Check to see if it is possible to fold this operation.
3338 bool MovedAway = false;
3339 if (matchOperationAddr(I, I->getOpcode(), Depth, &MovedAway)) {
3340 // This instruction may have been moved away. If so, there is nothing
3344 // Okay, it's possible to fold this. Check to see if it is actually
3345 // *profitable* to do so. We use a simple cost model to avoid increasing
3346 // register pressure too much.
3347 if (I->hasOneUse() ||
3348 isProfitableToFoldIntoAddressingMode(I, BackupAddrMode, AddrMode)) {
3349 AddrModeInsts.push_back(I);
3353 // It isn't profitable to do this, roll back.
3354 //cerr << "NOT FOLDING: " << *I;
3355 AddrMode = BackupAddrMode;
3356 AddrModeInsts.resize(OldSize);
3357 TPT.rollback(LastKnownGood);
3359 } else if (ConstantExpr *CE = dyn_cast<ConstantExpr>(Addr)) {
3360 if (matchOperationAddr(CE, CE->getOpcode(), Depth))
3362 TPT.rollback(LastKnownGood);
3363 } else if (isa<ConstantPointerNull>(Addr)) {
3364 // Null pointer gets folded without affecting the addressing mode.
3368 // Worse case, the target should support [reg] addressing modes. :)
3369 if (!AddrMode.HasBaseReg) {
3370 AddrMode.HasBaseReg = true;
3371 AddrMode.BaseReg = Addr;
3372 // Still check for legality in case the target supports [imm] but not [i+r].
3373 if (TLI.isLegalAddressingMode(DL, AddrMode, AccessTy, AddrSpace))
3375 AddrMode.HasBaseReg = false;
3376 AddrMode.BaseReg = nullptr;
3379 // If the base register is already taken, see if we can do [r+r].
3380 if (AddrMode.Scale == 0) {
3382 AddrMode.ScaledReg = Addr;
3383 if (TLI.isLegalAddressingMode(DL, AddrMode, AccessTy, AddrSpace))
3386 AddrMode.ScaledReg = nullptr;
3389 TPT.rollback(LastKnownGood);
3393 /// Check to see if all uses of OpVal by the specified inline asm call are due
3394 /// to memory operands. If so, return true, otherwise return false.
3395 static bool IsOperandAMemoryOperand(CallInst *CI, InlineAsm *IA, Value *OpVal,
3396 const TargetMachine &TM) {
3397 const Function *F = CI->getParent()->getParent();
3398 const TargetLowering *TLI = TM.getSubtargetImpl(*F)->getTargetLowering();
3399 const TargetRegisterInfo *TRI = TM.getSubtargetImpl(*F)->getRegisterInfo();
3400 TargetLowering::AsmOperandInfoVector TargetConstraints =
3401 TLI->ParseConstraints(F->getParent()->getDataLayout(), TRI,
3402 ImmutableCallSite(CI));
3403 for (unsigned i = 0, e = TargetConstraints.size(); i != e; ++i) {
3404 TargetLowering::AsmOperandInfo &OpInfo = TargetConstraints[i];
3406 // Compute the constraint code and ConstraintType to use.
3407 TLI->ComputeConstraintToUse(OpInfo, SDValue());
3409 // If this asm operand is our Value*, and if it isn't an indirect memory
3410 // operand, we can't fold it!
3411 if (OpInfo.CallOperandVal == OpVal &&
3412 (OpInfo.ConstraintType != TargetLowering::C_Memory ||
3413 !OpInfo.isIndirect))
3420 /// Recursively walk all the uses of I until we find a memory use.
3421 /// If we find an obviously non-foldable instruction, return true.
3422 /// Add the ultimately found memory instructions to MemoryUses.
3423 static bool FindAllMemoryUses(
3425 SmallVectorImpl<std::pair<Instruction *, unsigned>> &MemoryUses,
3426 SmallPtrSetImpl<Instruction *> &ConsideredInsts, const TargetMachine &TM) {
3427 // If we already considered this instruction, we're done.
3428 if (!ConsideredInsts.insert(I).second)
3431 // If this is an obviously unfoldable instruction, bail out.
3432 if (!MightBeFoldableInst(I))
3435 // Loop over all the uses, recursively processing them.
3436 for (Use &U : I->uses()) {
3437 Instruction *UserI = cast<Instruction>(U.getUser());
3439 if (LoadInst *LI = dyn_cast<LoadInst>(UserI)) {
3440 MemoryUses.push_back(std::make_pair(LI, U.getOperandNo()));
3444 if (StoreInst *SI = dyn_cast<StoreInst>(UserI)) {
3445 unsigned opNo = U.getOperandNo();
3446 if (opNo == 0) return true; // Storing addr, not into addr.
3447 MemoryUses.push_back(std::make_pair(SI, opNo));
3451 if (CallInst *CI = dyn_cast<CallInst>(UserI)) {
3452 InlineAsm *IA = dyn_cast<InlineAsm>(CI->getCalledValue());
3453 if (!IA) return true;
3455 // If this is a memory operand, we're cool, otherwise bail out.
3456 if (!IsOperandAMemoryOperand(CI, IA, I, TM))
3461 if (FindAllMemoryUses(UserI, MemoryUses, ConsideredInsts, TM))
3468 /// Return true if Val is already known to be live at the use site that we're
3469 /// folding it into. If so, there is no cost to include it in the addressing
3470 /// mode. KnownLive1 and KnownLive2 are two values that we know are live at the
3471 /// instruction already.
3472 bool AddressingModeMatcher::valueAlreadyLiveAtInst(Value *Val,Value *KnownLive1,
3473 Value *KnownLive2) {
3474 // If Val is either of the known-live values, we know it is live!
3475 if (Val == nullptr || Val == KnownLive1 || Val == KnownLive2)
3478 // All values other than instructions and arguments (e.g. constants) are live.
3479 if (!isa<Instruction>(Val) && !isa<Argument>(Val)) return true;
3481 // If Val is a constant sized alloca in the entry block, it is live, this is
3482 // true because it is just a reference to the stack/frame pointer, which is
3483 // live for the whole function.
3484 if (AllocaInst *AI = dyn_cast<AllocaInst>(Val))
3485 if (AI->isStaticAlloca())
3488 // Check to see if this value is already used in the memory instruction's
3489 // block. If so, it's already live into the block at the very least, so we
3490 // can reasonably fold it.
3491 return Val->isUsedInBasicBlock(MemoryInst->getParent());
3494 /// It is possible for the addressing mode of the machine to fold the specified
3495 /// instruction into a load or store that ultimately uses it.
3496 /// However, the specified instruction has multiple uses.
3497 /// Given this, it may actually increase register pressure to fold it
3498 /// into the load. For example, consider this code:
3502 /// use(Y) -> nonload/store
3506 /// In this case, Y has multiple uses, and can be folded into the load of Z
3507 /// (yielding load [X+2]). However, doing this will cause both "X" and "X+1" to
3508 /// be live at the use(Y) line. If we don't fold Y into load Z, we use one
3509 /// fewer register. Since Y can't be folded into "use(Y)" we don't increase the
3510 /// number of computations either.
3512 /// Note that this (like most of CodeGenPrepare) is just a rough heuristic. If
3513 /// X was live across 'load Z' for other reasons, we actually *would* want to
3514 /// fold the addressing mode in the Z case. This would make Y die earlier.
3515 bool AddressingModeMatcher::
3516 isProfitableToFoldIntoAddressingMode(Instruction *I, ExtAddrMode &AMBefore,
3517 ExtAddrMode &AMAfter) {
3518 if (IgnoreProfitability) return true;
3520 // AMBefore is the addressing mode before this instruction was folded into it,
3521 // and AMAfter is the addressing mode after the instruction was folded. Get
3522 // the set of registers referenced by AMAfter and subtract out those
3523 // referenced by AMBefore: this is the set of values which folding in this
3524 // address extends the lifetime of.
3526 // Note that there are only two potential values being referenced here,
3527 // BaseReg and ScaleReg (global addresses are always available, as are any
3528 // folded immediates).
3529 Value *BaseReg = AMAfter.BaseReg, *ScaledReg = AMAfter.ScaledReg;
3531 // If the BaseReg or ScaledReg was referenced by the previous addrmode, their
3532 // lifetime wasn't extended by adding this instruction.
3533 if (valueAlreadyLiveAtInst(BaseReg, AMBefore.BaseReg, AMBefore.ScaledReg))
3535 if (valueAlreadyLiveAtInst(ScaledReg, AMBefore.BaseReg, AMBefore.ScaledReg))
3536 ScaledReg = nullptr;
3538 // If folding this instruction (and it's subexprs) didn't extend any live
3539 // ranges, we're ok with it.
3540 if (!BaseReg && !ScaledReg)
3543 // If all uses of this instruction are ultimately load/store/inlineasm's,
3544 // check to see if their addressing modes will include this instruction. If
3545 // so, we can fold it into all uses, so it doesn't matter if it has multiple
3547 SmallVector<std::pair<Instruction*,unsigned>, 16> MemoryUses;
3548 SmallPtrSet<Instruction*, 16> ConsideredInsts;
3549 if (FindAllMemoryUses(I, MemoryUses, ConsideredInsts, TM))
3550 return false; // Has a non-memory, non-foldable use!
3552 // Now that we know that all uses of this instruction are part of a chain of
3553 // computation involving only operations that could theoretically be folded
3554 // into a memory use, loop over each of these uses and see if they could
3555 // *actually* fold the instruction.
3556 SmallVector<Instruction*, 32> MatchedAddrModeInsts;
3557 for (unsigned i = 0, e = MemoryUses.size(); i != e; ++i) {
3558 Instruction *User = MemoryUses[i].first;
3559 unsigned OpNo = MemoryUses[i].second;
3561 // Get the access type of this use. If the use isn't a pointer, we don't
3562 // know what it accesses.
3563 Value *Address = User->getOperand(OpNo);
3564 PointerType *AddrTy = dyn_cast<PointerType>(Address->getType());
3567 Type *AddressAccessTy = AddrTy->getElementType();
3568 unsigned AS = AddrTy->getAddressSpace();
3570 // Do a match against the root of this address, ignoring profitability. This
3571 // will tell us if the addressing mode for the memory operation will
3572 // *actually* cover the shared instruction.
3574 TypePromotionTransaction::ConstRestorationPt LastKnownGood =
3575 TPT.getRestorationPoint();
3576 AddressingModeMatcher Matcher(MatchedAddrModeInsts, TM, AddressAccessTy, AS,
3577 MemoryInst, Result, InsertedInsts,
3578 PromotedInsts, TPT);
3579 Matcher.IgnoreProfitability = true;
3580 bool Success = Matcher.matchAddr(Address, 0);
3581 (void)Success; assert(Success && "Couldn't select *anything*?");
3583 // The match was to check the profitability, the changes made are not
3584 // part of the original matcher. Therefore, they should be dropped
3585 // otherwise the original matcher will not present the right state.
3586 TPT.rollback(LastKnownGood);
3588 // If the match didn't cover I, then it won't be shared by it.
3589 if (std::find(MatchedAddrModeInsts.begin(), MatchedAddrModeInsts.end(),
3590 I) == MatchedAddrModeInsts.end())
3593 MatchedAddrModeInsts.clear();
3599 } // end anonymous namespace
3601 /// Return true if the specified values are defined in a
3602 /// different basic block than BB.
3603 static bool IsNonLocalValue(Value *V, BasicBlock *BB) {
3604 if (Instruction *I = dyn_cast<Instruction>(V))
3605 return I->getParent() != BB;
3609 /// Load and Store Instructions often have addressing modes that can do
3610 /// significant amounts of computation. As such, instruction selection will try
3611 /// to get the load or store to do as much computation as possible for the
3612 /// program. The problem is that isel can only see within a single block. As
3613 /// such, we sink as much legal addressing mode work into the block as possible.
3615 /// This method is used to optimize both load/store and inline asms with memory
3617 bool CodeGenPrepare::optimizeMemoryInst(Instruction *MemoryInst, Value *Addr,
3618 Type *AccessTy, unsigned AddrSpace) {
3621 // Try to collapse single-value PHI nodes. This is necessary to undo
3622 // unprofitable PRE transformations.
3623 SmallVector<Value*, 8> worklist;
3624 SmallPtrSet<Value*, 16> Visited;
3625 worklist.push_back(Addr);
3627 // Use a worklist to iteratively look through PHI nodes, and ensure that
3628 // the addressing mode obtained from the non-PHI roots of the graph
3630 Value *Consensus = nullptr;
3631 unsigned NumUsesConsensus = 0;
3632 bool IsNumUsesConsensusValid = false;
3633 SmallVector<Instruction*, 16> AddrModeInsts;
3634 ExtAddrMode AddrMode;
3635 TypePromotionTransaction TPT;
3636 TypePromotionTransaction::ConstRestorationPt LastKnownGood =
3637 TPT.getRestorationPoint();
3638 while (!worklist.empty()) {
3639 Value *V = worklist.back();
3640 worklist.pop_back();
3642 // Break use-def graph loops.
3643 if (!Visited.insert(V).second) {
3644 Consensus = nullptr;
3648 // For a PHI node, push all of its incoming values.
3649 if (PHINode *P = dyn_cast<PHINode>(V)) {
3650 for (Value *IncValue : P->incoming_values())
3651 worklist.push_back(IncValue);
3655 // For non-PHIs, determine the addressing mode being computed.
3656 SmallVector<Instruction*, 16> NewAddrModeInsts;
3657 ExtAddrMode NewAddrMode = AddressingModeMatcher::Match(
3658 V, AccessTy, AddrSpace, MemoryInst, NewAddrModeInsts, *TM,
3659 InsertedInsts, PromotedInsts, TPT);
3661 // This check is broken into two cases with very similar code to avoid using
3662 // getNumUses() as much as possible. Some values have a lot of uses, so
3663 // calling getNumUses() unconditionally caused a significant compile-time
3667 AddrMode = NewAddrMode;
3668 AddrModeInsts = NewAddrModeInsts;
3670 } else if (NewAddrMode == AddrMode) {
3671 if (!IsNumUsesConsensusValid) {
3672 NumUsesConsensus = Consensus->getNumUses();
3673 IsNumUsesConsensusValid = true;
3676 // Ensure that the obtained addressing mode is equivalent to that obtained
3677 // for all other roots of the PHI traversal. Also, when choosing one
3678 // such root as representative, select the one with the most uses in order
3679 // to keep the cost modeling heuristics in AddressingModeMatcher
3681 unsigned NumUses = V->getNumUses();
3682 if (NumUses > NumUsesConsensus) {
3684 NumUsesConsensus = NumUses;
3685 AddrModeInsts = NewAddrModeInsts;
3690 Consensus = nullptr;
3694 // If the addressing mode couldn't be determined, or if multiple different
3695 // ones were determined, bail out now.
3697 TPT.rollback(LastKnownGood);
3702 // Check to see if any of the instructions supersumed by this addr mode are
3703 // non-local to I's BB.
3704 bool AnyNonLocal = false;
3705 for (unsigned i = 0, e = AddrModeInsts.size(); i != e; ++i) {
3706 if (IsNonLocalValue(AddrModeInsts[i], MemoryInst->getParent())) {
3712 // If all the instructions matched are already in this BB, don't do anything.
3714 DEBUG(dbgs() << "CGP: Found local addrmode: " << AddrMode << "\n");
3718 // Insert this computation right after this user. Since our caller is
3719 // scanning from the top of the BB to the bottom, reuse of the expr are
3720 // guaranteed to happen later.
3721 IRBuilder<> Builder(MemoryInst);
3723 // Now that we determined the addressing expression we want to use and know
3724 // that we have to sink it into this block. Check to see if we have already
3725 // done this for some other load/store instr in this block. If so, reuse the
3727 Value *&SunkAddr = SunkAddrs[Addr];
3729 DEBUG(dbgs() << "CGP: Reusing nonlocal addrmode: " << AddrMode << " for "
3730 << *MemoryInst << "\n");
3731 if (SunkAddr->getType() != Addr->getType())
3732 SunkAddr = Builder.CreateBitCast(SunkAddr, Addr->getType());
3733 } else if (AddrSinkUsingGEPs ||
3734 (!AddrSinkUsingGEPs.getNumOccurrences() && TM &&
3735 TM->getSubtargetImpl(*MemoryInst->getParent()->getParent())
3737 // By default, we use the GEP-based method when AA is used later. This
3738 // prevents new inttoptr/ptrtoint pairs from degrading AA capabilities.
3739 DEBUG(dbgs() << "CGP: SINKING nonlocal addrmode: " << AddrMode << " for "
3740 << *MemoryInst << "\n");
3741 Type *IntPtrTy = DL->getIntPtrType(Addr->getType());
3742 Value *ResultPtr = nullptr, *ResultIndex = nullptr;
3744 // First, find the pointer.
3745 if (AddrMode.BaseReg && AddrMode.BaseReg->getType()->isPointerTy()) {
3746 ResultPtr = AddrMode.BaseReg;
3747 AddrMode.BaseReg = nullptr;
3750 if (AddrMode.Scale && AddrMode.ScaledReg->getType()->isPointerTy()) {
3751 // We can't add more than one pointer together, nor can we scale a
3752 // pointer (both of which seem meaningless).
3753 if (ResultPtr || AddrMode.Scale != 1)
3756 ResultPtr = AddrMode.ScaledReg;
3760 if (AddrMode.BaseGV) {
3764 ResultPtr = AddrMode.BaseGV;
3767 // If the real base value actually came from an inttoptr, then the matcher
3768 // will look through it and provide only the integer value. In that case,
3770 if (!ResultPtr && AddrMode.BaseReg) {
3772 Builder.CreateIntToPtr(AddrMode.BaseReg, Addr->getType(), "sunkaddr");
3773 AddrMode.BaseReg = nullptr;
3774 } else if (!ResultPtr && AddrMode.Scale == 1) {
3776 Builder.CreateIntToPtr(AddrMode.ScaledReg, Addr->getType(), "sunkaddr");
3781 !AddrMode.BaseReg && !AddrMode.Scale && !AddrMode.BaseOffs) {
3782 SunkAddr = Constant::getNullValue(Addr->getType());
3783 } else if (!ResultPtr) {
3787 Builder.getInt8PtrTy(Addr->getType()->getPointerAddressSpace());
3788 Type *I8Ty = Builder.getInt8Ty();
3790 // Start with the base register. Do this first so that subsequent address
3791 // matching finds it last, which will prevent it from trying to match it
3792 // as the scaled value in case it happens to be a mul. That would be
3793 // problematic if we've sunk a different mul for the scale, because then
3794 // we'd end up sinking both muls.
3795 if (AddrMode.BaseReg) {
3796 Value *V = AddrMode.BaseReg;
3797 if (V->getType() != IntPtrTy)
3798 V = Builder.CreateIntCast(V, IntPtrTy, /*isSigned=*/true, "sunkaddr");
3803 // Add the scale value.
3804 if (AddrMode.Scale) {
3805 Value *V = AddrMode.ScaledReg;
3806 if (V->getType() == IntPtrTy) {
3808 } else if (cast<IntegerType>(IntPtrTy)->getBitWidth() <
3809 cast<IntegerType>(V->getType())->getBitWidth()) {
3810 V = Builder.CreateTrunc(V, IntPtrTy, "sunkaddr");
3812 // It is only safe to sign extend the BaseReg if we know that the math
3813 // required to create it did not overflow before we extend it. Since
3814 // the original IR value was tossed in favor of a constant back when
3815 // the AddrMode was created we need to bail out gracefully if widths
3816 // do not match instead of extending it.
3817 Instruction *I = dyn_cast_or_null<Instruction>(ResultIndex);
3818 if (I && (ResultIndex != AddrMode.BaseReg))
3819 I->eraseFromParent();
3823 if (AddrMode.Scale != 1)
3824 V = Builder.CreateMul(V, ConstantInt::get(IntPtrTy, AddrMode.Scale),
3827 ResultIndex = Builder.CreateAdd(ResultIndex, V, "sunkaddr");
3832 // Add in the Base Offset if present.
3833 if (AddrMode.BaseOffs) {
3834 Value *V = ConstantInt::get(IntPtrTy, AddrMode.BaseOffs);
3836 // We need to add this separately from the scale above to help with
3837 // SDAG consecutive load/store merging.
3838 if (ResultPtr->getType() != I8PtrTy)
3839 ResultPtr = Builder.CreateBitCast(ResultPtr, I8PtrTy);
3840 ResultPtr = Builder.CreateGEP(I8Ty, ResultPtr, ResultIndex, "sunkaddr");
3847 SunkAddr = ResultPtr;
3849 if (ResultPtr->getType() != I8PtrTy)
3850 ResultPtr = Builder.CreateBitCast(ResultPtr, I8PtrTy);
3851 SunkAddr = Builder.CreateGEP(I8Ty, ResultPtr, ResultIndex, "sunkaddr");
3854 if (SunkAddr->getType() != Addr->getType())
3855 SunkAddr = Builder.CreateBitCast(SunkAddr, Addr->getType());
3858 DEBUG(dbgs() << "CGP: SINKING nonlocal addrmode: " << AddrMode << " for "
3859 << *MemoryInst << "\n");
3860 Type *IntPtrTy = DL->getIntPtrType(Addr->getType());
3861 Value *Result = nullptr;
3863 // Start with the base register. Do this first so that subsequent address
3864 // matching finds it last, which will prevent it from trying to match it
3865 // as the scaled value in case it happens to be a mul. That would be
3866 // problematic if we've sunk a different mul for the scale, because then
3867 // we'd end up sinking both muls.
3868 if (AddrMode.BaseReg) {
3869 Value *V = AddrMode.BaseReg;
3870 if (V->getType()->isPointerTy())
3871 V = Builder.CreatePtrToInt(V, IntPtrTy, "sunkaddr");
3872 if (V->getType() != IntPtrTy)
3873 V = Builder.CreateIntCast(V, IntPtrTy, /*isSigned=*/true, "sunkaddr");
3877 // Add the scale value.
3878 if (AddrMode.Scale) {
3879 Value *V = AddrMode.ScaledReg;
3880 if (V->getType() == IntPtrTy) {
3882 } else if (V->getType()->isPointerTy()) {
3883 V = Builder.CreatePtrToInt(V, IntPtrTy, "sunkaddr");
3884 } else if (cast<IntegerType>(IntPtrTy)->getBitWidth() <
3885 cast<IntegerType>(V->getType())->getBitWidth()) {
3886 V = Builder.CreateTrunc(V, IntPtrTy, "sunkaddr");
3888 // It is only safe to sign extend the BaseReg if we know that the math
3889 // required to create it did not overflow before we extend it. Since
3890 // the original IR value was tossed in favor of a constant back when
3891 // the AddrMode was created we need to bail out gracefully if widths
3892 // do not match instead of extending it.
3893 Instruction *I = dyn_cast_or_null<Instruction>(Result);
3894 if (I && (Result != AddrMode.BaseReg))
3895 I->eraseFromParent();
3898 if (AddrMode.Scale != 1)
3899 V = Builder.CreateMul(V, ConstantInt::get(IntPtrTy, AddrMode.Scale),
3902 Result = Builder.CreateAdd(Result, V, "sunkaddr");
3907 // Add in the BaseGV if present.
3908 if (AddrMode.BaseGV) {
3909 Value *V = Builder.CreatePtrToInt(AddrMode.BaseGV, IntPtrTy, "sunkaddr");
3911 Result = Builder.CreateAdd(Result, V, "sunkaddr");
3916 // Add in the Base Offset if present.
3917 if (AddrMode.BaseOffs) {
3918 Value *V = ConstantInt::get(IntPtrTy, AddrMode.BaseOffs);
3920 Result = Builder.CreateAdd(Result, V, "sunkaddr");
3926 SunkAddr = Constant::getNullValue(Addr->getType());
3928 SunkAddr = Builder.CreateIntToPtr(Result, Addr->getType(), "sunkaddr");
3931 MemoryInst->replaceUsesOfWith(Repl, SunkAddr);
3933 // If we have no uses, recursively delete the value and all dead instructions
3935 if (Repl->use_empty()) {
3936 // This can cause recursive deletion, which can invalidate our iterator.
3937 // Use a WeakVH to hold onto it in case this happens.
3938 WeakVH IterHandle(&*CurInstIterator);
3939 BasicBlock *BB = CurInstIterator->getParent();
3941 RecursivelyDeleteTriviallyDeadInstructions(Repl, TLInfo);
3943 if (IterHandle != CurInstIterator.getNodePtrUnchecked()) {
3944 // If the iterator instruction was recursively deleted, start over at the
3945 // start of the block.
3946 CurInstIterator = BB->begin();
3954 /// If there are any memory operands, use OptimizeMemoryInst to sink their
3955 /// address computing into the block when possible / profitable.
3956 bool CodeGenPrepare::optimizeInlineAsmInst(CallInst *CS) {
3957 bool MadeChange = false;
3959 const TargetRegisterInfo *TRI =
3960 TM->getSubtargetImpl(*CS->getParent()->getParent())->getRegisterInfo();
3961 TargetLowering::AsmOperandInfoVector TargetConstraints =
3962 TLI->ParseConstraints(*DL, TRI, CS);
3964 for (unsigned i = 0, e = TargetConstraints.size(); i != e; ++i) {
3965 TargetLowering::AsmOperandInfo &OpInfo = TargetConstraints[i];
3967 // Compute the constraint code and ConstraintType to use.
3968 TLI->ComputeConstraintToUse(OpInfo, SDValue());
3970 if (OpInfo.ConstraintType == TargetLowering::C_Memory &&
3971 OpInfo.isIndirect) {
3972 Value *OpVal = CS->getArgOperand(ArgNo++);
3973 MadeChange |= optimizeMemoryInst(CS, OpVal, OpVal->getType(), ~0u);
3974 } else if (OpInfo.Type == InlineAsm::isInput)
3981 /// \brief Check if all the uses of \p Inst are equivalent (or free) zero or
3982 /// sign extensions.
3983 static bool hasSameExtUse(Instruction *Inst, const TargetLowering &TLI) {
3984 assert(!Inst->use_empty() && "Input must have at least one use");
3985 const Instruction *FirstUser = cast<Instruction>(*Inst->user_begin());
3986 bool IsSExt = isa<SExtInst>(FirstUser);
3987 Type *ExtTy = FirstUser->getType();
3988 for (const User *U : Inst->users()) {
3989 const Instruction *UI = cast<Instruction>(U);
3990 if ((IsSExt && !isa<SExtInst>(UI)) || (!IsSExt && !isa<ZExtInst>(UI)))
3992 Type *CurTy = UI->getType();
3993 // Same input and output types: Same instruction after CSE.
3997 // If IsSExt is true, we are in this situation:
3999 // b = sext ty1 a to ty2
4000 // c = sext ty1 a to ty3
4001 // Assuming ty2 is shorter than ty3, this could be turned into:
4003 // b = sext ty1 a to ty2
4004 // c = sext ty2 b to ty3
4005 // However, the last sext is not free.
4009 // This is a ZExt, maybe this is free to extend from one type to another.
4010 // In that case, we would not account for a different use.
4013 if (ExtTy->getScalarType()->getIntegerBitWidth() >
4014 CurTy->getScalarType()->getIntegerBitWidth()) {
4022 if (!TLI.isZExtFree(NarrowTy, LargeTy))
4025 // All uses are the same or can be derived from one another for free.
4029 /// \brief Try to form ExtLd by promoting \p Exts until they reach a
4030 /// load instruction.
4031 /// If an ext(load) can be formed, it is returned via \p LI for the load
4032 /// and \p Inst for the extension.
4033 /// Otherwise LI == nullptr and Inst == nullptr.
4034 /// When some promotion happened, \p TPT contains the proper state to
4037 /// \return true when promoting was necessary to expose the ext(load)
4038 /// opportunity, false otherwise.
4042 /// %ld = load i32* %addr
4043 /// %add = add nuw i32 %ld, 4
4044 /// %zext = zext i32 %add to i64
4048 /// %ld = load i32* %addr
4049 /// %zext = zext i32 %ld to i64
4050 /// %add = add nuw i64 %zext, 4
4052 /// Thanks to the promotion, we can match zext(load i32*) to i64.
4053 bool CodeGenPrepare::extLdPromotion(TypePromotionTransaction &TPT,
4054 LoadInst *&LI, Instruction *&Inst,
4055 const SmallVectorImpl<Instruction *> &Exts,
4056 unsigned CreatedInstsCost = 0) {
4057 // Iterate over all the extensions to see if one form an ext(load).
4058 for (auto I : Exts) {
4059 // Check if we directly have ext(load).
4060 if ((LI = dyn_cast<LoadInst>(I->getOperand(0)))) {
4062 // No promotion happened here.
4065 // Check whether or not we want to do any promotion.
4066 if (!TLI || !TLI->enableExtLdPromotion() || DisableExtLdPromotion)
4068 // Get the action to perform the promotion.
4069 TypePromotionHelper::Action TPH = TypePromotionHelper::getAction(
4070 I, InsertedInsts, *TLI, PromotedInsts);
4071 // Check if we can promote.
4074 // Save the current state.
4075 TypePromotionTransaction::ConstRestorationPt LastKnownGood =
4076 TPT.getRestorationPoint();
4077 SmallVector<Instruction *, 4> NewExts;
4078 unsigned NewCreatedInstsCost = 0;
4079 unsigned ExtCost = !TLI->isExtFree(I);
4081 Value *PromotedVal = TPH(I, TPT, PromotedInsts, NewCreatedInstsCost,
4082 &NewExts, nullptr, *TLI);
4083 assert(PromotedVal &&
4084 "TypePromotionHelper should have filtered out those cases");
4086 // We would be able to merge only one extension in a load.
4087 // Therefore, if we have more than 1 new extension we heuristically
4088 // cut this search path, because it means we degrade the code quality.
4089 // With exactly 2, the transformation is neutral, because we will merge
4090 // one extension but leave one. However, we optimistically keep going,
4091 // because the new extension may be removed too.
4092 long long TotalCreatedInstsCost = CreatedInstsCost + NewCreatedInstsCost;
4093 TotalCreatedInstsCost -= ExtCost;
4094 if (!StressExtLdPromotion &&
4095 (TotalCreatedInstsCost > 1 ||
4096 !isPromotedInstructionLegal(*TLI, *DL, PromotedVal))) {
4097 // The promotion is not profitable, rollback to the previous state.
4098 TPT.rollback(LastKnownGood);
4101 // The promotion is profitable.
4102 // Check if it exposes an ext(load).
4103 (void)extLdPromotion(TPT, LI, Inst, NewExts, TotalCreatedInstsCost);
4104 if (LI && (StressExtLdPromotion || NewCreatedInstsCost <= ExtCost ||
4105 // If we have created a new extension, i.e., now we have two
4106 // extensions. We must make sure one of them is merged with
4107 // the load, otherwise we may degrade the code quality.
4108 (LI->hasOneUse() || hasSameExtUse(LI, *TLI))))
4109 // Promotion happened.
4111 // If this does not help to expose an ext(load) then, rollback.
4112 TPT.rollback(LastKnownGood);
4114 // None of the extension can form an ext(load).
4120 /// Move a zext or sext fed by a load into the same basic block as the load,
4121 /// unless conditions are unfavorable. This allows SelectionDAG to fold the
4122 /// extend into the load.
4123 /// \p I[in/out] the extension may be modified during the process if some
4124 /// promotions apply.
4126 bool CodeGenPrepare::moveExtToFormExtLoad(Instruction *&I) {
4127 // Try to promote a chain of computation if it allows to form
4128 // an extended load.
4129 TypePromotionTransaction TPT;
4130 TypePromotionTransaction::ConstRestorationPt LastKnownGood =
4131 TPT.getRestorationPoint();
4132 SmallVector<Instruction *, 1> Exts;
4134 // Look for a load being extended.
4135 LoadInst *LI = nullptr;
4136 Instruction *OldExt = I;
4137 bool HasPromoted = extLdPromotion(TPT, LI, I, Exts);
4139 assert(!HasPromoted && !LI && "If we did not match any load instruction "
4140 "the code must remain the same");
4145 // If they're already in the same block, there's nothing to do.
4146 // Make the cheap checks first if we did not promote.
4147 // If we promoted, we need to check if it is indeed profitable.
4148 if (!HasPromoted && LI->getParent() == I->getParent())
4151 EVT VT = TLI->getValueType(*DL, I->getType());
4152 EVT LoadVT = TLI->getValueType(*DL, LI->getType());
4154 // If the load has other users and the truncate is not free, this probably
4155 // isn't worthwhile.
4156 if (!LI->hasOneUse() && TLI &&
4157 (TLI->isTypeLegal(LoadVT) || !TLI->isTypeLegal(VT)) &&
4158 !TLI->isTruncateFree(I->getType(), LI->getType())) {
4160 TPT.rollback(LastKnownGood);
4164 // Check whether the target supports casts folded into loads.
4166 if (isa<ZExtInst>(I))
4167 LType = ISD::ZEXTLOAD;
4169 assert(isa<SExtInst>(I) && "Unexpected ext type!");
4170 LType = ISD::SEXTLOAD;
4172 if (TLI && !TLI->isLoadExtLegal(LType, VT, LoadVT)) {
4174 TPT.rollback(LastKnownGood);
4178 // Move the extend into the same block as the load, so that SelectionDAG
4181 I->removeFromParent();
4187 bool CodeGenPrepare::optimizeExtUses(Instruction *I) {
4188 BasicBlock *DefBB = I->getParent();
4190 // If the result of a {s|z}ext and its source are both live out, rewrite all
4191 // other uses of the source with result of extension.
4192 Value *Src = I->getOperand(0);
4193 if (Src->hasOneUse())
4196 // Only do this xform if truncating is free.
4197 if (TLI && !TLI->isTruncateFree(I->getType(), Src->getType()))
4200 // Only safe to perform the optimization if the source is also defined in
4202 if (!isa<Instruction>(Src) || DefBB != cast<Instruction>(Src)->getParent())
4205 bool DefIsLiveOut = false;
4206 for (User *U : I->users()) {
4207 Instruction *UI = cast<Instruction>(U);
4209 // Figure out which BB this ext is used in.
4210 BasicBlock *UserBB = UI->getParent();
4211 if (UserBB == DefBB) continue;
4212 DefIsLiveOut = true;
4218 // Make sure none of the uses are PHI nodes.
4219 for (User *U : Src->users()) {
4220 Instruction *UI = cast<Instruction>(U);
4221 BasicBlock *UserBB = UI->getParent();
4222 if (UserBB == DefBB) continue;
4223 // Be conservative. We don't want this xform to end up introducing
4224 // reloads just before load / store instructions.
4225 if (isa<PHINode>(UI) || isa<LoadInst>(UI) || isa<StoreInst>(UI))
4229 // InsertedTruncs - Only insert one trunc in each block once.
4230 DenseMap<BasicBlock*, Instruction*> InsertedTruncs;
4232 bool MadeChange = false;
4233 for (Use &U : Src->uses()) {
4234 Instruction *User = cast<Instruction>(U.getUser());
4236 // Figure out which BB this ext is used in.
4237 BasicBlock *UserBB = User->getParent();
4238 if (UserBB == DefBB) continue;
4240 // Both src and def are live in this block. Rewrite the use.
4241 Instruction *&InsertedTrunc = InsertedTruncs[UserBB];
4243 if (!InsertedTrunc) {
4244 BasicBlock::iterator InsertPt = UserBB->getFirstInsertionPt();
4245 assert(InsertPt != UserBB->end());
4246 InsertedTrunc = new TruncInst(I, Src->getType(), "", &*InsertPt);
4247 InsertedInsts.insert(InsertedTrunc);
4250 // Replace a use of the {s|z}ext source with a use of the result.
4259 /// Check if V (an operand of a select instruction) is an expensive instruction
4260 /// that is only used once.
4261 static bool sinkSelectOperand(const TargetTransformInfo *TTI, Value *V) {
4262 auto *I = dyn_cast<Instruction>(V);
4263 // If it's safe to speculatively execute, then it should not have side
4264 // effects; therefore, it's safe to sink and possibly *not* execute.
4265 return I && I->hasOneUse() && isSafeToSpeculativelyExecute(I) &&
4266 TTI->getUserCost(I) >= TargetTransformInfo::TCC_Expensive;
4269 /// Returns true if a SelectInst should be turned into an explicit branch.
4270 static bool isFormingBranchFromSelectProfitable(const TargetTransformInfo *TTI,
4272 // FIXME: This should use the same heuristics as IfConversion to determine
4273 // whether a select is better represented as a branch. This requires that
4274 // branch probability metadata is preserved for the select, which is not the
4277 CmpInst *Cmp = dyn_cast<CmpInst>(SI->getCondition());
4279 // If a branch is predictable, an out-of-order CPU can avoid blocking on its
4280 // comparison condition. If the compare has more than one use, there's
4281 // probably another cmov or setcc around, so it's not worth emitting a branch.
4282 if (!Cmp || !Cmp->hasOneUse())
4285 Value *CmpOp0 = Cmp->getOperand(0);
4286 Value *CmpOp1 = Cmp->getOperand(1);
4288 // Emit "cmov on compare with a memory operand" as a branch to avoid stalls
4289 // on a load from memory. But if the load is used more than once, do not
4290 // change the select to a branch because the load is probably needed
4291 // regardless of whether the branch is taken or not.
4292 if ((isa<LoadInst>(CmpOp0) && CmpOp0->hasOneUse()) ||
4293 (isa<LoadInst>(CmpOp1) && CmpOp1->hasOneUse()))
4296 // If either operand of the select is expensive and only needed on one side
4297 // of the select, we should form a branch.
4298 if (sinkSelectOperand(TTI, SI->getTrueValue()) ||
4299 sinkSelectOperand(TTI, SI->getFalseValue()))
4306 /// If we have a SelectInst that will likely profit from branch prediction,
4307 /// turn it into a branch.
4308 bool CodeGenPrepare::optimizeSelectInst(SelectInst *SI) {
4309 bool VectorCond = !SI->getCondition()->getType()->isIntegerTy(1);
4311 // Can we convert the 'select' to CF ?
4312 if (DisableSelectToBranch || OptSize || !TLI || VectorCond)
4315 TargetLowering::SelectSupportKind SelectKind;
4317 SelectKind = TargetLowering::VectorMaskSelect;
4318 else if (SI->getType()->isVectorTy())
4319 SelectKind = TargetLowering::ScalarCondVectorVal;
4321 SelectKind = TargetLowering::ScalarValSelect;
4323 // Do we have efficient codegen support for this kind of 'selects' ?
4324 if (TLI->isSelectSupported(SelectKind)) {
4325 // We have efficient codegen support for the select instruction.
4326 // Check if it is profitable to keep this 'select'.
4327 if (!TLI->isPredictableSelectExpensive() ||
4328 !isFormingBranchFromSelectProfitable(TTI, SI))
4334 // Transform a sequence like this:
4336 // %cmp = cmp uge i32 %a, %b
4337 // %sel = select i1 %cmp, i32 %c, i32 %d
4341 // %cmp = cmp uge i32 %a, %b
4342 // br i1 %cmp, label %select.true, label %select.false
4344 // br label %select.end
4346 // br label %select.end
4348 // %sel = phi i32 [ %c, %select.true ], [ %d, %select.false ]
4350 // In addition, we may sink instructions that produce %c or %d from
4351 // the entry block into the destination(s) of the new branch.
4352 // If the true or false blocks do not contain a sunken instruction, that
4353 // block and its branch may be optimized away. In that case, one side of the
4354 // first branch will point directly to select.end, and the corresponding PHI
4355 // predecessor block will be the start block.
4357 // First, we split the block containing the select into 2 blocks.
4358 BasicBlock *StartBlock = SI->getParent();
4359 BasicBlock::iterator SplitPt = ++(BasicBlock::iterator(SI));
4360 BasicBlock *EndBlock = StartBlock->splitBasicBlock(SplitPt, "select.end");
4362 // Delete the unconditional branch that was just created by the split.
4363 StartBlock->getTerminator()->eraseFromParent();
4365 // These are the new basic blocks for the conditional branch.
4366 // At least one will become an actual new basic block.
4367 BasicBlock *TrueBlock = nullptr;
4368 BasicBlock *FalseBlock = nullptr;
4370 // Sink expensive instructions into the conditional blocks to avoid executing
4371 // them speculatively.
4372 if (sinkSelectOperand(TTI, SI->getTrueValue())) {
4373 TrueBlock = BasicBlock::Create(SI->getContext(), "select.true.sink",
4374 EndBlock->getParent(), EndBlock);
4375 auto *TrueBranch = BranchInst::Create(EndBlock, TrueBlock);
4376 auto *TrueInst = cast<Instruction>(SI->getTrueValue());
4377 TrueInst->moveBefore(TrueBranch);
4379 if (sinkSelectOperand(TTI, SI->getFalseValue())) {
4380 FalseBlock = BasicBlock::Create(SI->getContext(), "select.false.sink",
4381 EndBlock->getParent(), EndBlock);
4382 auto *FalseBranch = BranchInst::Create(EndBlock, FalseBlock);
4383 auto *FalseInst = cast<Instruction>(SI->getFalseValue());
4384 FalseInst->moveBefore(FalseBranch);
4387 // If there was nothing to sink, then arbitrarily choose the 'false' side
4388 // for a new input value to the PHI.
4389 if (TrueBlock == FalseBlock) {
4390 assert(TrueBlock == nullptr &&
4391 "Unexpected basic block transform while optimizing select");
4393 FalseBlock = BasicBlock::Create(SI->getContext(), "select.false",
4394 EndBlock->getParent(), EndBlock);
4395 BranchInst::Create(EndBlock, FalseBlock);
4398 // Insert the real conditional branch based on the original condition.
4399 // If we did not create a new block for one of the 'true' or 'false' paths
4400 // of the condition, it means that side of the branch goes to the end block
4401 // directly and the path originates from the start block from the point of
4402 // view of the new PHI.
4403 if (TrueBlock == nullptr) {
4404 BranchInst::Create(EndBlock, FalseBlock, SI->getCondition(), SI);
4405 TrueBlock = StartBlock;
4406 } else if (FalseBlock == nullptr) {
4407 BranchInst::Create(TrueBlock, EndBlock, SI->getCondition(), SI);
4408 FalseBlock = StartBlock;
4410 BranchInst::Create(TrueBlock, FalseBlock, SI->getCondition(), SI);
4413 // The select itself is replaced with a PHI Node.
4414 PHINode *PN = PHINode::Create(SI->getType(), 2, "", &EndBlock->front());
4416 PN->addIncoming(SI->getTrueValue(), TrueBlock);
4417 PN->addIncoming(SI->getFalseValue(), FalseBlock);
4419 SI->replaceAllUsesWith(PN);
4420 SI->eraseFromParent();
4422 // Instruct OptimizeBlock to skip to the next block.
4423 CurInstIterator = StartBlock->end();
4424 ++NumSelectsExpanded;
4428 static bool isBroadcastShuffle(ShuffleVectorInst *SVI) {
4429 SmallVector<int, 16> Mask(SVI->getShuffleMask());
4431 for (unsigned i = 0; i < Mask.size(); ++i) {
4432 if (SplatElem != -1 && Mask[i] != -1 && Mask[i] != SplatElem)
4434 SplatElem = Mask[i];
4440 /// Some targets have expensive vector shifts if the lanes aren't all the same
4441 /// (e.g. x86 only introduced "vpsllvd" and friends with AVX2). In these cases
4442 /// it's often worth sinking a shufflevector splat down to its use so that
4443 /// codegen can spot all lanes are identical.
4444 bool CodeGenPrepare::optimizeShuffleVectorInst(ShuffleVectorInst *SVI) {
4445 BasicBlock *DefBB = SVI->getParent();
4447 // Only do this xform if variable vector shifts are particularly expensive.
4448 if (!TLI || !TLI->isVectorShiftByScalarCheap(SVI->getType()))
4451 // We only expect better codegen by sinking a shuffle if we can recognise a
4453 if (!isBroadcastShuffle(SVI))
4456 // InsertedShuffles - Only insert a shuffle in each block once.
4457 DenseMap<BasicBlock*, Instruction*> InsertedShuffles;
4459 bool MadeChange = false;
4460 for (User *U : SVI->users()) {
4461 Instruction *UI = cast<Instruction>(U);
4463 // Figure out which BB this ext is used in.
4464 BasicBlock *UserBB = UI->getParent();
4465 if (UserBB == DefBB) continue;
4467 // For now only apply this when the splat is used by a shift instruction.
4468 if (!UI->isShift()) continue;
4470 // Everything checks out, sink the shuffle if the user's block doesn't
4471 // already have a copy.
4472 Instruction *&InsertedShuffle = InsertedShuffles[UserBB];
4474 if (!InsertedShuffle) {
4475 BasicBlock::iterator InsertPt = UserBB->getFirstInsertionPt();
4476 assert(InsertPt != UserBB->end());
4478 new ShuffleVectorInst(SVI->getOperand(0), SVI->getOperand(1),
4479 SVI->getOperand(2), "", &*InsertPt);
4482 UI->replaceUsesOfWith(SVI, InsertedShuffle);
4486 // If we removed all uses, nuke the shuffle.
4487 if (SVI->use_empty()) {
4488 SVI->eraseFromParent();
4495 bool CodeGenPrepare::optimizeSwitchInst(SwitchInst *SI) {
4499 Value *Cond = SI->getCondition();
4500 Type *OldType = Cond->getType();
4501 LLVMContext &Context = Cond->getContext();
4502 MVT RegType = TLI->getRegisterType(Context, TLI->getValueType(*DL, OldType));
4503 unsigned RegWidth = RegType.getSizeInBits();
4505 if (RegWidth <= cast<IntegerType>(OldType)->getBitWidth())
4508 // If the register width is greater than the type width, expand the condition
4509 // of the switch instruction and each case constant to the width of the
4510 // register. By widening the type of the switch condition, subsequent
4511 // comparisons (for case comparisons) will not need to be extended to the
4512 // preferred register width, so we will potentially eliminate N-1 extends,
4513 // where N is the number of cases in the switch.
4514 auto *NewType = Type::getIntNTy(Context, RegWidth);
4516 // Zero-extend the switch condition and case constants unless the switch
4517 // condition is a function argument that is already being sign-extended.
4518 // In that case, we can avoid an unnecessary mask/extension by sign-extending
4519 // everything instead.
4520 Instruction::CastOps ExtType = Instruction::ZExt;
4521 if (auto *Arg = dyn_cast<Argument>(Cond))
4522 if (Arg->hasSExtAttr())
4523 ExtType = Instruction::SExt;
4525 auto *ExtInst = CastInst::Create(ExtType, Cond, NewType);
4526 ExtInst->insertBefore(SI);
4527 SI->setCondition(ExtInst);
4528 for (SwitchInst::CaseIt Case : SI->cases()) {
4529 APInt NarrowConst = Case.getCaseValue()->getValue();
4530 APInt WideConst = (ExtType == Instruction::ZExt) ?
4531 NarrowConst.zext(RegWidth) : NarrowConst.sext(RegWidth);
4532 Case.setValue(ConstantInt::get(Context, WideConst));
4539 /// \brief Helper class to promote a scalar operation to a vector one.
4540 /// This class is used to move downward extractelement transition.
4542 /// a = vector_op <2 x i32>
4543 /// b = extractelement <2 x i32> a, i32 0
4548 /// a = vector_op <2 x i32>
4549 /// c = vector_op a (equivalent to scalar_op on the related lane)
4550 /// * d = extractelement <2 x i32> c, i32 0
4552 /// Assuming both extractelement and store can be combine, we get rid of the
4554 class VectorPromoteHelper {
4555 /// DataLayout associated with the current module.
4556 const DataLayout &DL;
4558 /// Used to perform some checks on the legality of vector operations.
4559 const TargetLowering &TLI;
4561 /// Used to estimated the cost of the promoted chain.
4562 const TargetTransformInfo &TTI;
4564 /// The transition being moved downwards.
4565 Instruction *Transition;
4566 /// The sequence of instructions to be promoted.
4567 SmallVector<Instruction *, 4> InstsToBePromoted;
4568 /// Cost of combining a store and an extract.
4569 unsigned StoreExtractCombineCost;
4570 /// Instruction that will be combined with the transition.
4571 Instruction *CombineInst;
4573 /// \brief The instruction that represents the current end of the transition.
4574 /// Since we are faking the promotion until we reach the end of the chain
4575 /// of computation, we need a way to get the current end of the transition.
4576 Instruction *getEndOfTransition() const {
4577 if (InstsToBePromoted.empty())
4579 return InstsToBePromoted.back();
4582 /// \brief Return the index of the original value in the transition.
4583 /// E.g., for "extractelement <2 x i32> c, i32 1" the original value,
4584 /// c, is at index 0.
4585 unsigned getTransitionOriginalValueIdx() const {
4586 assert(isa<ExtractElementInst>(Transition) &&
4587 "Other kind of transitions are not supported yet");
4591 /// \brief Return the index of the index in the transition.
4592 /// E.g., for "extractelement <2 x i32> c, i32 0" the index
4594 unsigned getTransitionIdx() const {
4595 assert(isa<ExtractElementInst>(Transition) &&
4596 "Other kind of transitions are not supported yet");
4600 /// \brief Get the type of the transition.
4601 /// This is the type of the original value.
4602 /// E.g., for "extractelement <2 x i32> c, i32 1" the type of the
4603 /// transition is <2 x i32>.
4604 Type *getTransitionType() const {
4605 return Transition->getOperand(getTransitionOriginalValueIdx())->getType();
4608 /// \brief Promote \p ToBePromoted by moving \p Def downward through.
4609 /// I.e., we have the following sequence:
4610 /// Def = Transition <ty1> a to <ty2>
4611 /// b = ToBePromoted <ty2> Def, ...
4613 /// b = ToBePromoted <ty1> a, ...
4614 /// Def = Transition <ty1> ToBePromoted to <ty2>
4615 void promoteImpl(Instruction *ToBePromoted);
4617 /// \brief Check whether or not it is profitable to promote all the
4618 /// instructions enqueued to be promoted.
4619 bool isProfitableToPromote() {
4620 Value *ValIdx = Transition->getOperand(getTransitionOriginalValueIdx());
4621 unsigned Index = isa<ConstantInt>(ValIdx)
4622 ? cast<ConstantInt>(ValIdx)->getZExtValue()
4624 Type *PromotedType = getTransitionType();
4626 StoreInst *ST = cast<StoreInst>(CombineInst);
4627 unsigned AS = ST->getPointerAddressSpace();
4628 unsigned Align = ST->getAlignment();
4629 // Check if this store is supported.
4630 if (!TLI.allowsMisalignedMemoryAccesses(
4631 TLI.getValueType(DL, ST->getValueOperand()->getType()), AS,
4633 // If this is not supported, there is no way we can combine
4634 // the extract with the store.
4638 // The scalar chain of computation has to pay for the transition
4639 // scalar to vector.
4640 // The vector chain has to account for the combining cost.
4641 uint64_t ScalarCost =
4642 TTI.getVectorInstrCost(Transition->getOpcode(), PromotedType, Index);
4643 uint64_t VectorCost = StoreExtractCombineCost;
4644 for (const auto &Inst : InstsToBePromoted) {
4645 // Compute the cost.
4646 // By construction, all instructions being promoted are arithmetic ones.
4647 // Moreover, one argument is a constant that can be viewed as a splat
4649 Value *Arg0 = Inst->getOperand(0);
4650 bool IsArg0Constant = isa<UndefValue>(Arg0) || isa<ConstantInt>(Arg0) ||
4651 isa<ConstantFP>(Arg0);
4652 TargetTransformInfo::OperandValueKind Arg0OVK =
4653 IsArg0Constant ? TargetTransformInfo::OK_UniformConstantValue
4654 : TargetTransformInfo::OK_AnyValue;
4655 TargetTransformInfo::OperandValueKind Arg1OVK =
4656 !IsArg0Constant ? TargetTransformInfo::OK_UniformConstantValue
4657 : TargetTransformInfo::OK_AnyValue;
4658 ScalarCost += TTI.getArithmeticInstrCost(
4659 Inst->getOpcode(), Inst->getType(), Arg0OVK, Arg1OVK);
4660 VectorCost += TTI.getArithmeticInstrCost(Inst->getOpcode(), PromotedType,
4663 DEBUG(dbgs() << "Estimated cost of computation to be promoted:\nScalar: "
4664 << ScalarCost << "\nVector: " << VectorCost << '\n');
4665 return ScalarCost > VectorCost;
4668 /// \brief Generate a constant vector with \p Val with the same
4669 /// number of elements as the transition.
4670 /// \p UseSplat defines whether or not \p Val should be replicated
4671 /// across the whole vector.
4672 /// In other words, if UseSplat == true, we generate <Val, Val, ..., Val>,
4673 /// otherwise we generate a vector with as many undef as possible:
4674 /// <undef, ..., undef, Val, undef, ..., undef> where \p Val is only
4675 /// used at the index of the extract.
4676 Value *getConstantVector(Constant *Val, bool UseSplat) const {
4677 unsigned ExtractIdx = UINT_MAX;
4679 // If we cannot determine where the constant must be, we have to
4680 // use a splat constant.
4681 Value *ValExtractIdx = Transition->getOperand(getTransitionIdx());
4682 if (ConstantInt *CstVal = dyn_cast<ConstantInt>(ValExtractIdx))
4683 ExtractIdx = CstVal->getSExtValue();
4688 unsigned End = getTransitionType()->getVectorNumElements();
4690 return ConstantVector::getSplat(End, Val);
4692 SmallVector<Constant *, 4> ConstVec;
4693 UndefValue *UndefVal = UndefValue::get(Val->getType());
4694 for (unsigned Idx = 0; Idx != End; ++Idx) {
4695 if (Idx == ExtractIdx)
4696 ConstVec.push_back(Val);
4698 ConstVec.push_back(UndefVal);
4700 return ConstantVector::get(ConstVec);
4703 /// \brief Check if promoting to a vector type an operand at \p OperandIdx
4704 /// in \p Use can trigger undefined behavior.
4705 static bool canCauseUndefinedBehavior(const Instruction *Use,
4706 unsigned OperandIdx) {
4707 // This is not safe to introduce undef when the operand is on
4708 // the right hand side of a division-like instruction.
4709 if (OperandIdx != 1)
4711 switch (Use->getOpcode()) {
4714 case Instruction::SDiv:
4715 case Instruction::UDiv:
4716 case Instruction::SRem:
4717 case Instruction::URem:
4719 case Instruction::FDiv:
4720 case Instruction::FRem:
4721 return !Use->hasNoNaNs();
4723 llvm_unreachable(nullptr);
4727 VectorPromoteHelper(const DataLayout &DL, const TargetLowering &TLI,
4728 const TargetTransformInfo &TTI, Instruction *Transition,
4729 unsigned CombineCost)
4730 : DL(DL), TLI(TLI), TTI(TTI), Transition(Transition),
4731 StoreExtractCombineCost(CombineCost), CombineInst(nullptr) {
4732 assert(Transition && "Do not know how to promote null");
4735 /// \brief Check if we can promote \p ToBePromoted to \p Type.
4736 bool canPromote(const Instruction *ToBePromoted) const {
4737 // We could support CastInst too.
4738 return isa<BinaryOperator>(ToBePromoted);
4741 /// \brief Check if it is profitable to promote \p ToBePromoted
4742 /// by moving downward the transition through.
4743 bool shouldPromote(const Instruction *ToBePromoted) const {
4744 // Promote only if all the operands can be statically expanded.
4745 // Indeed, we do not want to introduce any new kind of transitions.
4746 for (const Use &U : ToBePromoted->operands()) {
4747 const Value *Val = U.get();
4748 if (Val == getEndOfTransition()) {
4749 // If the use is a division and the transition is on the rhs,
4750 // we cannot promote the operation, otherwise we may create a
4751 // division by zero.
4752 if (canCauseUndefinedBehavior(ToBePromoted, U.getOperandNo()))
4756 if (!isa<ConstantInt>(Val) && !isa<UndefValue>(Val) &&
4757 !isa<ConstantFP>(Val))
4760 // Check that the resulting operation is legal.
4761 int ISDOpcode = TLI.InstructionOpcodeToISD(ToBePromoted->getOpcode());
4764 return StressStoreExtract ||
4765 TLI.isOperationLegalOrCustom(
4766 ISDOpcode, TLI.getValueType(DL, getTransitionType(), true));
4769 /// \brief Check whether or not \p Use can be combined
4770 /// with the transition.
4771 /// I.e., is it possible to do Use(Transition) => AnotherUse?
4772 bool canCombine(const Instruction *Use) { return isa<StoreInst>(Use); }
4774 /// \brief Record \p ToBePromoted as part of the chain to be promoted.
4775 void enqueueForPromotion(Instruction *ToBePromoted) {
4776 InstsToBePromoted.push_back(ToBePromoted);
4779 /// \brief Set the instruction that will be combined with the transition.
4780 void recordCombineInstruction(Instruction *ToBeCombined) {
4781 assert(canCombine(ToBeCombined) && "Unsupported instruction to combine");
4782 CombineInst = ToBeCombined;
4785 /// \brief Promote all the instructions enqueued for promotion if it is
4787 /// \return True if the promotion happened, false otherwise.
4789 // Check if there is something to promote.
4790 // Right now, if we do not have anything to combine with,
4791 // we assume the promotion is not profitable.
4792 if (InstsToBePromoted.empty() || !CombineInst)
4796 if (!StressStoreExtract && !isProfitableToPromote())
4800 for (auto &ToBePromoted : InstsToBePromoted)
4801 promoteImpl(ToBePromoted);
4802 InstsToBePromoted.clear();
4806 } // End of anonymous namespace.
4808 void VectorPromoteHelper::promoteImpl(Instruction *ToBePromoted) {
4809 // At this point, we know that all the operands of ToBePromoted but Def
4810 // can be statically promoted.
4811 // For Def, we need to use its parameter in ToBePromoted:
4812 // b = ToBePromoted ty1 a
4813 // Def = Transition ty1 b to ty2
4814 // Move the transition down.
4815 // 1. Replace all uses of the promoted operation by the transition.
4816 // = ... b => = ... Def.
4817 assert(ToBePromoted->getType() == Transition->getType() &&
4818 "The type of the result of the transition does not match "
4820 ToBePromoted->replaceAllUsesWith(Transition);
4821 // 2. Update the type of the uses.
4822 // b = ToBePromoted ty2 Def => b = ToBePromoted ty1 Def.
4823 Type *TransitionTy = getTransitionType();
4824 ToBePromoted->mutateType(TransitionTy);
4825 // 3. Update all the operands of the promoted operation with promoted
4827 // b = ToBePromoted ty1 Def => b = ToBePromoted ty1 a.
4828 for (Use &U : ToBePromoted->operands()) {
4829 Value *Val = U.get();
4830 Value *NewVal = nullptr;
4831 if (Val == Transition)
4832 NewVal = Transition->getOperand(getTransitionOriginalValueIdx());
4833 else if (isa<UndefValue>(Val) || isa<ConstantInt>(Val) ||
4834 isa<ConstantFP>(Val)) {
4835 // Use a splat constant if it is not safe to use undef.
4836 NewVal = getConstantVector(
4837 cast<Constant>(Val),
4838 isa<UndefValue>(Val) ||
4839 canCauseUndefinedBehavior(ToBePromoted, U.getOperandNo()));
4841 llvm_unreachable("Did you modified shouldPromote and forgot to update "
4843 ToBePromoted->setOperand(U.getOperandNo(), NewVal);
4845 Transition->removeFromParent();
4846 Transition->insertAfter(ToBePromoted);
4847 Transition->setOperand(getTransitionOriginalValueIdx(), ToBePromoted);
4850 /// Some targets can do store(extractelement) with one instruction.
4851 /// Try to push the extractelement towards the stores when the target
4852 /// has this feature and this is profitable.
4853 bool CodeGenPrepare::optimizeExtractElementInst(Instruction *Inst) {
4854 unsigned CombineCost = UINT_MAX;
4855 if (DisableStoreExtract || !TLI ||
4856 (!StressStoreExtract &&
4857 !TLI->canCombineStoreAndExtract(Inst->getOperand(0)->getType(),
4858 Inst->getOperand(1), CombineCost)))
4861 // At this point we know that Inst is a vector to scalar transition.
4862 // Try to move it down the def-use chain, until:
4863 // - We can combine the transition with its single use
4864 // => we got rid of the transition.
4865 // - We escape the current basic block
4866 // => we would need to check that we are moving it at a cheaper place and
4867 // we do not do that for now.
4868 BasicBlock *Parent = Inst->getParent();
4869 DEBUG(dbgs() << "Found an interesting transition: " << *Inst << '\n');
4870 VectorPromoteHelper VPH(*DL, *TLI, *TTI, Inst, CombineCost);
4871 // If the transition has more than one use, assume this is not going to be
4873 while (Inst->hasOneUse()) {
4874 Instruction *ToBePromoted = cast<Instruction>(*Inst->user_begin());
4875 DEBUG(dbgs() << "Use: " << *ToBePromoted << '\n');
4877 if (ToBePromoted->getParent() != Parent) {
4878 DEBUG(dbgs() << "Instruction to promote is in a different block ("
4879 << ToBePromoted->getParent()->getName()
4880 << ") than the transition (" << Parent->getName() << ").\n");
4884 if (VPH.canCombine(ToBePromoted)) {
4885 DEBUG(dbgs() << "Assume " << *Inst << '\n'
4886 << "will be combined with: " << *ToBePromoted << '\n');
4887 VPH.recordCombineInstruction(ToBePromoted);
4888 bool Changed = VPH.promote();
4889 NumStoreExtractExposed += Changed;
4893 DEBUG(dbgs() << "Try promoting.\n");
4894 if (!VPH.canPromote(ToBePromoted) || !VPH.shouldPromote(ToBePromoted))
4897 DEBUG(dbgs() << "Promoting is possible... Enqueue for promotion!\n");
4899 VPH.enqueueForPromotion(ToBePromoted);
4900 Inst = ToBePromoted;
4905 bool CodeGenPrepare::optimizeInst(Instruction *I, bool& ModifiedDT) {
4906 // Bail out if we inserted the instruction to prevent optimizations from
4907 // stepping on each other's toes.
4908 if (InsertedInsts.count(I))
4911 if (PHINode *P = dyn_cast<PHINode>(I)) {
4912 // It is possible for very late stage optimizations (such as SimplifyCFG)
4913 // to introduce PHI nodes too late to be cleaned up. If we detect such a
4914 // trivial PHI, go ahead and zap it here.
4915 if (Value *V = SimplifyInstruction(P, *DL, TLInfo, nullptr)) {
4916 P->replaceAllUsesWith(V);
4917 P->eraseFromParent();
4924 if (CastInst *CI = dyn_cast<CastInst>(I)) {
4925 // If the source of the cast is a constant, then this should have
4926 // already been constant folded. The only reason NOT to constant fold
4927 // it is if something (e.g. LSR) was careful to place the constant
4928 // evaluation in a block other than then one that uses it (e.g. to hoist
4929 // the address of globals out of a loop). If this is the case, we don't
4930 // want to forward-subst the cast.
4931 if (isa<Constant>(CI->getOperand(0)))
4934 if (TLI && OptimizeNoopCopyExpression(CI, *TLI, *DL))
4937 if (isa<ZExtInst>(I) || isa<SExtInst>(I)) {
4938 /// Sink a zext or sext into its user blocks if the target type doesn't
4939 /// fit in one register
4941 TLI->getTypeAction(CI->getContext(),
4942 TLI->getValueType(*DL, CI->getType())) ==
4943 TargetLowering::TypeExpandInteger) {
4944 return SinkCast(CI);
4946 bool MadeChange = moveExtToFormExtLoad(I);
4947 return MadeChange | optimizeExtUses(I);
4953 if (CmpInst *CI = dyn_cast<CmpInst>(I))
4954 if (!TLI || !TLI->hasMultipleConditionRegisters())
4955 return OptimizeCmpExpression(CI);
4957 if (LoadInst *LI = dyn_cast<LoadInst>(I)) {
4958 stripInvariantGroupMetadata(*LI);
4960 unsigned AS = LI->getPointerAddressSpace();
4961 return optimizeMemoryInst(I, I->getOperand(0), LI->getType(), AS);
4966 if (StoreInst *SI = dyn_cast<StoreInst>(I)) {
4967 stripInvariantGroupMetadata(*SI);
4969 unsigned AS = SI->getPointerAddressSpace();
4970 return optimizeMemoryInst(I, SI->getOperand(1),
4971 SI->getOperand(0)->getType(), AS);
4976 BinaryOperator *BinOp = dyn_cast<BinaryOperator>(I);
4978 if (BinOp && (BinOp->getOpcode() == Instruction::AShr ||
4979 BinOp->getOpcode() == Instruction::LShr)) {
4980 ConstantInt *CI = dyn_cast<ConstantInt>(BinOp->getOperand(1));
4981 if (TLI && CI && TLI->hasExtractBitsInsn())
4982 return OptimizeExtractBits(BinOp, CI, *TLI, *DL);
4987 if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(I)) {
4988 if (GEPI->hasAllZeroIndices()) {
4989 /// The GEP operand must be a pointer, so must its result -> BitCast
4990 Instruction *NC = new BitCastInst(GEPI->getOperand(0), GEPI->getType(),
4991 GEPI->getName(), GEPI);
4992 GEPI->replaceAllUsesWith(NC);
4993 GEPI->eraseFromParent();
4995 optimizeInst(NC, ModifiedDT);
5001 if (CallInst *CI = dyn_cast<CallInst>(I))
5002 return optimizeCallInst(CI, ModifiedDT);
5004 if (SelectInst *SI = dyn_cast<SelectInst>(I))
5005 return optimizeSelectInst(SI);
5007 if (ShuffleVectorInst *SVI = dyn_cast<ShuffleVectorInst>(I))
5008 return optimizeShuffleVectorInst(SVI);
5010 if (auto *Switch = dyn_cast<SwitchInst>(I))
5011 return optimizeSwitchInst(Switch);
5013 if (isa<ExtractElementInst>(I))
5014 return optimizeExtractElementInst(I);
5019 // In this pass we look for GEP and cast instructions that are used
5020 // across basic blocks and rewrite them to improve basic-block-at-a-time
5022 bool CodeGenPrepare::optimizeBlock(BasicBlock &BB, bool& ModifiedDT) {
5024 bool MadeChange = false;
5026 CurInstIterator = BB.begin();
5027 while (CurInstIterator != BB.end()) {
5028 MadeChange |= optimizeInst(&*CurInstIterator++, ModifiedDT);
5032 MadeChange |= dupRetToEnableTailCallOpts(&BB);
5037 // llvm.dbg.value is far away from the value then iSel may not be able
5038 // handle it properly. iSel will drop llvm.dbg.value if it can not
5039 // find a node corresponding to the value.
5040 bool CodeGenPrepare::placeDbgValues(Function &F) {
5041 bool MadeChange = false;
5042 for (BasicBlock &BB : F) {
5043 Instruction *PrevNonDbgInst = nullptr;
5044 for (BasicBlock::iterator BI = BB.begin(), BE = BB.end(); BI != BE;) {
5045 Instruction *Insn = &*BI++;
5046 DbgValueInst *DVI = dyn_cast<DbgValueInst>(Insn);
5047 // Leave dbg.values that refer to an alloca alone. These
5048 // instrinsics describe the address of a variable (= the alloca)
5049 // being taken. They should not be moved next to the alloca
5050 // (and to the beginning of the scope), but rather stay close to
5051 // where said address is used.
5052 if (!DVI || (DVI->getValue() && isa<AllocaInst>(DVI->getValue()))) {
5053 PrevNonDbgInst = Insn;
5057 Instruction *VI = dyn_cast_or_null<Instruction>(DVI->getValue());
5058 if (VI && VI != PrevNonDbgInst && !VI->isTerminator()) {
5059 DEBUG(dbgs() << "Moving Debug Value before :\n" << *DVI << ' ' << *VI);
5060 DVI->removeFromParent();
5061 if (isa<PHINode>(VI))
5062 DVI->insertBefore(&*VI->getParent()->getFirstInsertionPt());
5064 DVI->insertAfter(VI);
5073 // If there is a sequence that branches based on comparing a single bit
5074 // against zero that can be combined into a single instruction, and the
5075 // target supports folding these into a single instruction, sink the
5076 // mask and compare into the branch uses. Do this before OptimizeBlock ->
5077 // OptimizeInst -> OptimizeCmpExpression, which perturbs the pattern being
5079 bool CodeGenPrepare::sinkAndCmp(Function &F) {
5080 if (!EnableAndCmpSinking)
5082 if (!TLI || !TLI->isMaskAndBranchFoldingLegal())
5084 bool MadeChange = false;
5085 for (Function::iterator I = F.begin(), E = F.end(); I != E; ) {
5086 BasicBlock *BB = &*I++;
5088 // Does this BB end with the following?
5089 // %andVal = and %val, #single-bit-set
5090 // %icmpVal = icmp %andResult, 0
5091 // br i1 %cmpVal label %dest1, label %dest2"
5092 BranchInst *Brcc = dyn_cast<BranchInst>(BB->getTerminator());
5093 if (!Brcc || !Brcc->isConditional())
5095 ICmpInst *Cmp = dyn_cast<ICmpInst>(Brcc->getOperand(0));
5096 if (!Cmp || Cmp->getParent() != BB)
5098 ConstantInt *Zero = dyn_cast<ConstantInt>(Cmp->getOperand(1));
5099 if (!Zero || !Zero->isZero())
5101 Instruction *And = dyn_cast<Instruction>(Cmp->getOperand(0));
5102 if (!And || And->getOpcode() != Instruction::And || And->getParent() != BB)
5104 ConstantInt* Mask = dyn_cast<ConstantInt>(And->getOperand(1));
5105 if (!Mask || !Mask->getUniqueInteger().isPowerOf2())
5107 DEBUG(dbgs() << "found and; icmp ?,0; brcc\n"); DEBUG(BB->dump());
5109 // Push the "and; icmp" for any users that are conditional branches.
5110 // Since there can only be one branch use per BB, we don't need to keep
5111 // track of which BBs we insert into.
5112 for (Value::use_iterator UI = Cmp->use_begin(), E = Cmp->use_end();
5116 BranchInst *BrccUser = dyn_cast<BranchInst>(*UI);
5118 if (!BrccUser || !BrccUser->isConditional())
5120 BasicBlock *UserBB = BrccUser->getParent();
5121 if (UserBB == BB) continue;
5122 DEBUG(dbgs() << "found Brcc use\n");
5124 // Sink the "and; icmp" to use.
5126 BinaryOperator *NewAnd =
5127 BinaryOperator::CreateAnd(And->getOperand(0), And->getOperand(1), "",
5130 CmpInst::Create(Cmp->getOpcode(), Cmp->getPredicate(), NewAnd, Zero,
5134 DEBUG(BrccUser->getParent()->dump());
5140 /// \brief Retrieve the probabilities of a conditional branch. Returns true on
5141 /// success, or returns false if no or invalid metadata was found.
5142 static bool extractBranchMetadata(BranchInst *BI,
5143 uint64_t &ProbTrue, uint64_t &ProbFalse) {
5144 assert(BI->isConditional() &&
5145 "Looking for probabilities on unconditional branch?");
5146 auto *ProfileData = BI->getMetadata(LLVMContext::MD_prof);
5147 if (!ProfileData || ProfileData->getNumOperands() != 3)
5150 const auto *CITrue =
5151 mdconst::dyn_extract<ConstantInt>(ProfileData->getOperand(1));
5152 const auto *CIFalse =
5153 mdconst::dyn_extract<ConstantInt>(ProfileData->getOperand(2));
5154 if (!CITrue || !CIFalse)
5157 ProbTrue = CITrue->getValue().getZExtValue();
5158 ProbFalse = CIFalse->getValue().getZExtValue();
5163 /// \brief Scale down both weights to fit into uint32_t.
5164 static void scaleWeights(uint64_t &NewTrue, uint64_t &NewFalse) {
5165 uint64_t NewMax = (NewTrue > NewFalse) ? NewTrue : NewFalse;
5166 uint32_t Scale = (NewMax / UINT32_MAX) + 1;
5167 NewTrue = NewTrue / Scale;
5168 NewFalse = NewFalse / Scale;
5171 /// \brief Some targets prefer to split a conditional branch like:
5173 /// %0 = icmp ne i32 %a, 0
5174 /// %1 = icmp ne i32 %b, 0
5175 /// %or.cond = or i1 %0, %1
5176 /// br i1 %or.cond, label %TrueBB, label %FalseBB
5178 /// into multiple branch instructions like:
5181 /// %0 = icmp ne i32 %a, 0
5182 /// br i1 %0, label %TrueBB, label %bb2
5184 /// %1 = icmp ne i32 %b, 0
5185 /// br i1 %1, label %TrueBB, label %FalseBB
5187 /// This usually allows instruction selection to do even further optimizations
5188 /// and combine the compare with the branch instruction. Currently this is
5189 /// applied for targets which have "cheap" jump instructions.
5191 /// FIXME: Remove the (equivalent?) implementation in SelectionDAG.
5193 bool CodeGenPrepare::splitBranchCondition(Function &F) {
5194 if (!TM || !TM->Options.EnableFastISel || !TLI || TLI->isJumpExpensive())
5197 bool MadeChange = false;
5198 for (auto &BB : F) {
5199 // Does this BB end with the following?
5200 // %cond1 = icmp|fcmp|binary instruction ...
5201 // %cond2 = icmp|fcmp|binary instruction ...
5202 // %cond.or = or|and i1 %cond1, cond2
5203 // br i1 %cond.or label %dest1, label %dest2"
5204 BinaryOperator *LogicOp;
5205 BasicBlock *TBB, *FBB;
5206 if (!match(BB.getTerminator(), m_Br(m_OneUse(m_BinOp(LogicOp)), TBB, FBB)))
5209 auto *Br1 = cast<BranchInst>(BB.getTerminator());
5210 if (Br1->getMetadata(LLVMContext::MD_unpredictable))
5214 Value *Cond1, *Cond2;
5215 if (match(LogicOp, m_And(m_OneUse(m_Value(Cond1)),
5216 m_OneUse(m_Value(Cond2)))))
5217 Opc = Instruction::And;
5218 else if (match(LogicOp, m_Or(m_OneUse(m_Value(Cond1)),
5219 m_OneUse(m_Value(Cond2)))))
5220 Opc = Instruction::Or;
5224 if (!match(Cond1, m_CombineOr(m_Cmp(), m_BinOp())) ||
5225 !match(Cond2, m_CombineOr(m_Cmp(), m_BinOp())) )
5228 DEBUG(dbgs() << "Before branch condition splitting\n"; BB.dump());
5231 auto *InsertBefore = std::next(Function::iterator(BB))
5232 .getNodePtrUnchecked();
5233 auto TmpBB = BasicBlock::Create(BB.getContext(),
5234 BB.getName() + ".cond.split",
5235 BB.getParent(), InsertBefore);
5237 // Update original basic block by using the first condition directly by the
5238 // branch instruction and removing the no longer needed and/or instruction.
5239 Br1->setCondition(Cond1);
5240 LogicOp->eraseFromParent();
5242 // Depending on the conditon we have to either replace the true or the false
5243 // successor of the original branch instruction.
5244 if (Opc == Instruction::And)
5245 Br1->setSuccessor(0, TmpBB);
5247 Br1->setSuccessor(1, TmpBB);
5249 // Fill in the new basic block.
5250 auto *Br2 = IRBuilder<>(TmpBB).CreateCondBr(Cond2, TBB, FBB);
5251 if (auto *I = dyn_cast<Instruction>(Cond2)) {
5252 I->removeFromParent();
5253 I->insertBefore(Br2);
5256 // Update PHI nodes in both successors. The original BB needs to be
5257 // replaced in one succesor's PHI nodes, because the branch comes now from
5258 // the newly generated BB (NewBB). In the other successor we need to add one
5259 // incoming edge to the PHI nodes, because both branch instructions target
5260 // now the same successor. Depending on the original branch condition
5261 // (and/or) we have to swap the successors (TrueDest, FalseDest), so that
5262 // we perfrom the correct update for the PHI nodes.
5263 // This doesn't change the successor order of the just created branch
5264 // instruction (or any other instruction).
5265 if (Opc == Instruction::Or)
5266 std::swap(TBB, FBB);
5268 // Replace the old BB with the new BB.
5269 for (auto &I : *TBB) {
5270 PHINode *PN = dyn_cast<PHINode>(&I);
5274 while ((i = PN->getBasicBlockIndex(&BB)) >= 0)
5275 PN->setIncomingBlock(i, TmpBB);
5278 // Add another incoming edge form the new BB.
5279 for (auto &I : *FBB) {
5280 PHINode *PN = dyn_cast<PHINode>(&I);
5283 auto *Val = PN->getIncomingValueForBlock(&BB);
5284 PN->addIncoming(Val, TmpBB);
5287 // Update the branch weights (from SelectionDAGBuilder::
5288 // FindMergedConditions).
5289 if (Opc == Instruction::Or) {
5290 // Codegen X | Y as:
5299 // We have flexibility in setting Prob for BB1 and Prob for NewBB.
5300 // The requirement is that
5301 // TrueProb for BB1 + (FalseProb for BB1 * TrueProb for TmpBB)
5302 // = TrueProb for orignal BB.
5303 // Assuming the orignal weights are A and B, one choice is to set BB1's
5304 // weights to A and A+2B, and set TmpBB's weights to A and 2B. This choice
5306 // TrueProb for BB1 == FalseProb for BB1 * TrueProb for TmpBB.
5307 // Another choice is to assume TrueProb for BB1 equals to TrueProb for
5308 // TmpBB, but the math is more complicated.
5309 uint64_t TrueWeight, FalseWeight;
5310 if (extractBranchMetadata(Br1, TrueWeight, FalseWeight)) {
5311 uint64_t NewTrueWeight = TrueWeight;
5312 uint64_t NewFalseWeight = TrueWeight + 2 * FalseWeight;
5313 scaleWeights(NewTrueWeight, NewFalseWeight);
5314 Br1->setMetadata(LLVMContext::MD_prof, MDBuilder(Br1->getContext())
5315 .createBranchWeights(TrueWeight, FalseWeight));
5317 NewTrueWeight = TrueWeight;
5318 NewFalseWeight = 2 * FalseWeight;
5319 scaleWeights(NewTrueWeight, NewFalseWeight);
5320 Br2->setMetadata(LLVMContext::MD_prof, MDBuilder(Br2->getContext())
5321 .createBranchWeights(TrueWeight, FalseWeight));
5324 // Codegen X & Y as:
5332 // This requires creation of TmpBB after CurBB.
5334 // We have flexibility in setting Prob for BB1 and Prob for TmpBB.
5335 // The requirement is that
5336 // FalseProb for BB1 + (TrueProb for BB1 * FalseProb for TmpBB)
5337 // = FalseProb for orignal BB.
5338 // Assuming the orignal weights are A and B, one choice is to set BB1's
5339 // weights to 2A+B and B, and set TmpBB's weights to 2A and B. This choice
5341 // FalseProb for BB1 == TrueProb for BB1 * FalseProb for TmpBB.
5342 uint64_t TrueWeight, FalseWeight;
5343 if (extractBranchMetadata(Br1, TrueWeight, FalseWeight)) {
5344 uint64_t NewTrueWeight = 2 * TrueWeight + FalseWeight;
5345 uint64_t NewFalseWeight = FalseWeight;
5346 scaleWeights(NewTrueWeight, NewFalseWeight);
5347 Br1->setMetadata(LLVMContext::MD_prof, MDBuilder(Br1->getContext())
5348 .createBranchWeights(TrueWeight, FalseWeight));
5350 NewTrueWeight = 2 * TrueWeight;
5351 NewFalseWeight = FalseWeight;
5352 scaleWeights(NewTrueWeight, NewFalseWeight);
5353 Br2->setMetadata(LLVMContext::MD_prof, MDBuilder(Br2->getContext())
5354 .createBranchWeights(TrueWeight, FalseWeight));
5358 // Note: No point in getting fancy here, since the DT info is never
5359 // available to CodeGenPrepare.
5364 DEBUG(dbgs() << "After branch condition splitting\n"; BB.dump();
5370 void CodeGenPrepare::stripInvariantGroupMetadata(Instruction &I) {
5371 if (auto *InvariantMD = I.getMetadata(LLVMContext::MD_invariant_group))
5372 I.dropUnknownNonDebugMetadata(InvariantMD->getMetadataID());