1 //===- CodeGenPrepare.cpp - Prepare a function for code generation --------===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // This pass munges the code in the input function to better prepare it for
11 // SelectionDAG-based code generation. This works around limitations in it's
12 // basic-block-at-a-time approach. It should eventually be removed.
14 //===----------------------------------------------------------------------===//
16 #include "llvm/CodeGen/Passes.h"
17 #include "llvm/ADT/DenseMap.h"
18 #include "llvm/ADT/SmallSet.h"
19 #include "llvm/ADT/Statistic.h"
20 #include "llvm/Analysis/InstructionSimplify.h"
21 #include "llvm/Analysis/TargetLibraryInfo.h"
22 #include "llvm/Analysis/TargetTransformInfo.h"
23 #include "llvm/Analysis/ValueTracking.h"
24 #include "llvm/IR/CallSite.h"
25 #include "llvm/IR/Constants.h"
26 #include "llvm/IR/DataLayout.h"
27 #include "llvm/IR/DerivedTypes.h"
28 #include "llvm/IR/Dominators.h"
29 #include "llvm/IR/Function.h"
30 #include "llvm/IR/GetElementPtrTypeIterator.h"
31 #include "llvm/IR/IRBuilder.h"
32 #include "llvm/IR/InlineAsm.h"
33 #include "llvm/IR/Instructions.h"
34 #include "llvm/IR/IntrinsicInst.h"
35 #include "llvm/IR/MDBuilder.h"
36 #include "llvm/IR/PatternMatch.h"
37 #include "llvm/IR/Statepoint.h"
38 #include "llvm/IR/ValueHandle.h"
39 #include "llvm/IR/ValueMap.h"
40 #include "llvm/Pass.h"
41 #include "llvm/Support/CommandLine.h"
42 #include "llvm/Support/Debug.h"
43 #include "llvm/Support/raw_ostream.h"
44 #include "llvm/Target/TargetLowering.h"
45 #include "llvm/Target/TargetSubtargetInfo.h"
46 #include "llvm/Transforms/Utils/BasicBlockUtils.h"
47 #include "llvm/Transforms/Utils/BuildLibCalls.h"
48 #include "llvm/Transforms/Utils/BypassSlowDivision.h"
49 #include "llvm/Transforms/Utils/Local.h"
50 #include "llvm/Transforms/Utils/SimplifyLibCalls.h"
52 using namespace llvm::PatternMatch;
54 #define DEBUG_TYPE "codegenprepare"
56 STATISTIC(NumBlocksElim, "Number of blocks eliminated");
57 STATISTIC(NumPHIsElim, "Number of trivial PHIs eliminated");
58 STATISTIC(NumGEPsElim, "Number of GEPs converted to casts");
59 STATISTIC(NumCmpUses, "Number of uses of Cmp expressions replaced with uses of "
61 STATISTIC(NumCastUses, "Number of uses of Cast expressions replaced with uses "
63 STATISTIC(NumMemoryInsts, "Number of memory instructions whose address "
64 "computations were sunk");
65 STATISTIC(NumExtsMoved, "Number of [s|z]ext instructions combined with loads");
66 STATISTIC(NumExtUses, "Number of uses of [s|z]ext instructions optimized");
67 STATISTIC(NumAndsAdded,
68 "Number of and mask instructions added to form ext loads");
69 STATISTIC(NumAndUses, "Number of uses of and mask instructions optimized");
70 STATISTIC(NumRetsDup, "Number of return instructions duplicated");
71 STATISTIC(NumDbgValueMoved, "Number of debug value instructions moved");
72 STATISTIC(NumSelectsExpanded, "Number of selects turned into branches");
73 STATISTIC(NumAndCmpsMoved, "Number of and/cmp's pushed into branches");
74 STATISTIC(NumStoreExtractExposed, "Number of store(extractelement) exposed");
76 static cl::opt<bool> DisableBranchOpts(
77 "disable-cgp-branch-opts", cl::Hidden, cl::init(false),
78 cl::desc("Disable branch optimizations in CodeGenPrepare"));
81 DisableGCOpts("disable-cgp-gc-opts", cl::Hidden, cl::init(false),
82 cl::desc("Disable GC optimizations in CodeGenPrepare"));
84 static cl::opt<bool> DisableSelectToBranch(
85 "disable-cgp-select2branch", cl::Hidden, cl::init(false),
86 cl::desc("Disable select to branch conversion."));
88 static cl::opt<bool> AddrSinkUsingGEPs(
89 "addr-sink-using-gep", cl::Hidden, cl::init(false),
90 cl::desc("Address sinking in CGP using GEPs."));
92 static cl::opt<bool> EnableAndCmpSinking(
93 "enable-andcmp-sinking", cl::Hidden, cl::init(true),
94 cl::desc("Enable sinkinig and/cmp into branches."));
96 static cl::opt<bool> DisableStoreExtract(
97 "disable-cgp-store-extract", cl::Hidden, cl::init(false),
98 cl::desc("Disable store(extract) optimizations in CodeGenPrepare"));
100 static cl::opt<bool> StressStoreExtract(
101 "stress-cgp-store-extract", cl::Hidden, cl::init(false),
102 cl::desc("Stress test store(extract) optimizations in CodeGenPrepare"));
104 static cl::opt<bool> DisableExtLdPromotion(
105 "disable-cgp-ext-ld-promotion", cl::Hidden, cl::init(false),
106 cl::desc("Disable ext(promotable(ld)) -> promoted(ext(ld)) optimization in "
109 static cl::opt<bool> StressExtLdPromotion(
110 "stress-cgp-ext-ld-promotion", cl::Hidden, cl::init(false),
111 cl::desc("Stress test ext(promotable(ld)) -> promoted(ext(ld)) "
112 "optimization in CodeGenPrepare"));
115 typedef SmallPtrSet<Instruction *, 16> SetOfInstrs;
116 typedef PointerIntPair<Type *, 1, bool> TypeIsSExt;
117 typedef DenseMap<Instruction *, TypeIsSExt> InstrToOrigTy;
118 class TypePromotionTransaction;
120 class CodeGenPrepare : public FunctionPass {
121 const TargetMachine *TM;
122 const TargetLowering *TLI;
123 const TargetTransformInfo *TTI;
124 const TargetLibraryInfo *TLInfo;
126 /// As we scan instructions optimizing them, this is the next instruction
127 /// to optimize. Transforms that can invalidate this should update it.
128 BasicBlock::iterator CurInstIterator;
130 /// Keeps track of non-local addresses that have been sunk into a block.
131 /// This allows us to avoid inserting duplicate code for blocks with
132 /// multiple load/stores of the same address.
133 ValueMap<Value*, Value*> SunkAddrs;
135 /// Keeps track of all instructions inserted for the current function.
136 SetOfInstrs InsertedInsts;
137 /// Keeps track of the type of the related instruction before their
138 /// promotion for the current function.
139 InstrToOrigTy PromotedInsts;
141 /// True if CFG is modified in any way.
144 /// True if optimizing for size.
147 /// DataLayout for the Function being processed.
148 const DataLayout *DL;
151 static char ID; // Pass identification, replacement for typeid
152 explicit CodeGenPrepare(const TargetMachine *TM = nullptr)
153 : FunctionPass(ID), TM(TM), TLI(nullptr), TTI(nullptr), DL(nullptr) {
154 initializeCodeGenPreparePass(*PassRegistry::getPassRegistry());
156 bool runOnFunction(Function &F) override;
158 const char *getPassName() const override { return "CodeGen Prepare"; }
160 void getAnalysisUsage(AnalysisUsage &AU) const override {
161 AU.addPreserved<DominatorTreeWrapperPass>();
162 AU.addRequired<TargetLibraryInfoWrapperPass>();
163 AU.addRequired<TargetTransformInfoWrapperPass>();
167 bool eliminateFallThrough(Function &F);
168 bool eliminateMostlyEmptyBlocks(Function &F);
169 bool canMergeBlocks(const BasicBlock *BB, const BasicBlock *DestBB) const;
170 void eliminateMostlyEmptyBlock(BasicBlock *BB);
171 bool optimizeBlock(BasicBlock &BB, bool& ModifiedDT);
172 bool optimizeInst(Instruction *I, bool& ModifiedDT);
173 bool optimizeMemoryInst(Instruction *I, Value *Addr,
174 Type *AccessTy, unsigned AS);
175 bool optimizeInlineAsmInst(CallInst *CS);
176 bool optimizeCallInst(CallInst *CI, bool& ModifiedDT);
177 bool moveExtToFormExtLoad(Instruction *&I);
178 bool optimizeExtUses(Instruction *I);
179 bool optimizeLoadExt(LoadInst *I);
180 bool optimizeSelectInst(SelectInst *SI);
181 bool optimizeShuffleVectorInst(ShuffleVectorInst *SI);
182 bool optimizeSwitchInst(SwitchInst *CI);
183 bool optimizeExtractElementInst(Instruction *Inst);
184 bool dupRetToEnableTailCallOpts(BasicBlock *BB);
185 bool placeDbgValues(Function &F);
186 bool sinkAndCmp(Function &F);
187 bool extLdPromotion(TypePromotionTransaction &TPT, LoadInst *&LI,
189 const SmallVectorImpl<Instruction *> &Exts,
190 unsigned CreatedInstCost);
191 bool splitBranchCondition(Function &F);
192 bool simplifyOffsetableRelocate(Instruction &I);
193 void stripInvariantGroupMetadata(Instruction &I);
197 char CodeGenPrepare::ID = 0;
198 INITIALIZE_TM_PASS(CodeGenPrepare, "codegenprepare",
199 "Optimize for code generation", false, false)
201 FunctionPass *llvm::createCodeGenPreparePass(const TargetMachine *TM) {
202 return new CodeGenPrepare(TM);
205 bool CodeGenPrepare::runOnFunction(Function &F) {
206 if (skipOptnoneFunction(F))
209 DL = &F.getParent()->getDataLayout();
211 bool EverMadeChange = false;
212 // Clear per function information.
213 InsertedInsts.clear();
214 PromotedInsts.clear();
218 TLI = TM->getSubtargetImpl(F)->getTargetLowering();
219 TLInfo = &getAnalysis<TargetLibraryInfoWrapperPass>().getTLI();
220 TTI = &getAnalysis<TargetTransformInfoWrapperPass>().getTTI(F);
221 OptSize = F.optForSize();
223 /// This optimization identifies DIV instructions that can be
224 /// profitably bypassed and carried out with a shorter, faster divide.
225 if (!OptSize && TLI && TLI->isSlowDivBypassed()) {
226 const DenseMap<unsigned int, unsigned int> &BypassWidths =
227 TLI->getBypassSlowDivWidths();
228 for (Function::iterator I = F.begin(); I != F.end(); I++)
229 EverMadeChange |= bypassSlowDivision(F, I, BypassWidths);
232 // Eliminate blocks that contain only PHI nodes and an
233 // unconditional branch.
234 EverMadeChange |= eliminateMostlyEmptyBlocks(F);
236 // llvm.dbg.value is far away from the value then iSel may not be able
237 // handle it properly. iSel will drop llvm.dbg.value if it can not
238 // find a node corresponding to the value.
239 EverMadeChange |= placeDbgValues(F);
241 // If there is a mask, compare against zero, and branch that can be combined
242 // into a single target instruction, push the mask and compare into branch
243 // users. Do this before OptimizeBlock -> OptimizeInst ->
244 // OptimizeCmpExpression, which perturbs the pattern being searched for.
245 if (!DisableBranchOpts) {
246 EverMadeChange |= sinkAndCmp(F);
247 EverMadeChange |= splitBranchCondition(F);
250 bool MadeChange = true;
253 for (Function::iterator I = F.begin(); I != F.end(); ) {
254 BasicBlock *BB = &*I++;
255 bool ModifiedDTOnIteration = false;
256 MadeChange |= optimizeBlock(*BB, ModifiedDTOnIteration);
258 // Restart BB iteration if the dominator tree of the Function was changed
259 if (ModifiedDTOnIteration)
262 EverMadeChange |= MadeChange;
267 if (!DisableBranchOpts) {
269 SmallPtrSet<BasicBlock*, 8> WorkList;
270 for (BasicBlock &BB : F) {
271 SmallVector<BasicBlock *, 2> Successors(succ_begin(&BB), succ_end(&BB));
272 MadeChange |= ConstantFoldTerminator(&BB, true);
273 if (!MadeChange) continue;
275 for (SmallVectorImpl<BasicBlock*>::iterator
276 II = Successors.begin(), IE = Successors.end(); II != IE; ++II)
277 if (pred_begin(*II) == pred_end(*II))
278 WorkList.insert(*II);
281 // Delete the dead blocks and any of their dead successors.
282 MadeChange |= !WorkList.empty();
283 while (!WorkList.empty()) {
284 BasicBlock *BB = *WorkList.begin();
286 SmallVector<BasicBlock*, 2> Successors(succ_begin(BB), succ_end(BB));
290 for (SmallVectorImpl<BasicBlock*>::iterator
291 II = Successors.begin(), IE = Successors.end(); II != IE; ++II)
292 if (pred_begin(*II) == pred_end(*II))
293 WorkList.insert(*II);
296 // Merge pairs of basic blocks with unconditional branches, connected by
298 if (EverMadeChange || MadeChange)
299 MadeChange |= eliminateFallThrough(F);
301 EverMadeChange |= MadeChange;
304 if (!DisableGCOpts) {
305 SmallVector<Instruction *, 2> Statepoints;
306 for (BasicBlock &BB : F)
307 for (Instruction &I : BB)
309 Statepoints.push_back(&I);
310 for (auto &I : Statepoints)
311 EverMadeChange |= simplifyOffsetableRelocate(*I);
314 return EverMadeChange;
317 /// Merge basic blocks which are connected by a single edge, where one of the
318 /// basic blocks has a single successor pointing to the other basic block,
319 /// which has a single predecessor.
320 bool CodeGenPrepare::eliminateFallThrough(Function &F) {
321 bool Changed = false;
322 // Scan all of the blocks in the function, except for the entry block.
323 for (Function::iterator I = std::next(F.begin()), E = F.end(); I != E;) {
324 BasicBlock *BB = &*I++;
325 // If the destination block has a single pred, then this is a trivial
326 // edge, just collapse it.
327 BasicBlock *SinglePred = BB->getSinglePredecessor();
329 // Don't merge if BB's address is taken.
330 if (!SinglePred || SinglePred == BB || BB->hasAddressTaken()) continue;
332 BranchInst *Term = dyn_cast<BranchInst>(SinglePred->getTerminator());
333 if (Term && !Term->isConditional()) {
335 DEBUG(dbgs() << "To merge:\n"<< *SinglePred << "\n\n\n");
336 // Remember if SinglePred was the entry block of the function.
337 // If so, we will need to move BB back to the entry position.
338 bool isEntry = SinglePred == &SinglePred->getParent()->getEntryBlock();
339 MergeBasicBlockIntoOnlyPred(BB, nullptr);
341 if (isEntry && BB != &BB->getParent()->getEntryBlock())
342 BB->moveBefore(&BB->getParent()->getEntryBlock());
344 // We have erased a block. Update the iterator.
345 I = BB->getIterator();
351 /// Eliminate blocks that contain only PHI nodes, debug info directives, and an
352 /// unconditional branch. Passes before isel (e.g. LSR/loopsimplify) often split
353 /// edges in ways that are non-optimal for isel. Start by eliminating these
354 /// blocks so we can split them the way we want them.
355 bool CodeGenPrepare::eliminateMostlyEmptyBlocks(Function &F) {
356 bool MadeChange = false;
357 // Note that this intentionally skips the entry block.
358 for (Function::iterator I = std::next(F.begin()), E = F.end(); I != E;) {
359 BasicBlock *BB = &*I++;
361 // If this block doesn't end with an uncond branch, ignore it.
362 BranchInst *BI = dyn_cast<BranchInst>(BB->getTerminator());
363 if (!BI || !BI->isUnconditional())
366 // If the instruction before the branch (skipping debug info) isn't a phi
367 // node, then other stuff is happening here.
368 BasicBlock::iterator BBI = BI->getIterator();
369 if (BBI != BB->begin()) {
371 while (isa<DbgInfoIntrinsic>(BBI)) {
372 if (BBI == BB->begin())
376 if (!isa<DbgInfoIntrinsic>(BBI) && !isa<PHINode>(BBI))
380 // Do not break infinite loops.
381 BasicBlock *DestBB = BI->getSuccessor(0);
385 if (!canMergeBlocks(BB, DestBB))
388 eliminateMostlyEmptyBlock(BB);
394 /// Return true if we can merge BB into DestBB if there is a single
395 /// unconditional branch between them, and BB contains no other non-phi
397 bool CodeGenPrepare::canMergeBlocks(const BasicBlock *BB,
398 const BasicBlock *DestBB) const {
399 // We only want to eliminate blocks whose phi nodes are used by phi nodes in
400 // the successor. If there are more complex condition (e.g. preheaders),
401 // don't mess around with them.
402 BasicBlock::const_iterator BBI = BB->begin();
403 while (const PHINode *PN = dyn_cast<PHINode>(BBI++)) {
404 for (const User *U : PN->users()) {
405 const Instruction *UI = cast<Instruction>(U);
406 if (UI->getParent() != DestBB || !isa<PHINode>(UI))
408 // If User is inside DestBB block and it is a PHINode then check
409 // incoming value. If incoming value is not from BB then this is
410 // a complex condition (e.g. preheaders) we want to avoid here.
411 if (UI->getParent() == DestBB) {
412 if (const PHINode *UPN = dyn_cast<PHINode>(UI))
413 for (unsigned I = 0, E = UPN->getNumIncomingValues(); I != E; ++I) {
414 Instruction *Insn = dyn_cast<Instruction>(UPN->getIncomingValue(I));
415 if (Insn && Insn->getParent() == BB &&
416 Insn->getParent() != UPN->getIncomingBlock(I))
423 // If BB and DestBB contain any common predecessors, then the phi nodes in BB
424 // and DestBB may have conflicting incoming values for the block. If so, we
425 // can't merge the block.
426 const PHINode *DestBBPN = dyn_cast<PHINode>(DestBB->begin());
427 if (!DestBBPN) return true; // no conflict.
429 // Collect the preds of BB.
430 SmallPtrSet<const BasicBlock*, 16> BBPreds;
431 if (const PHINode *BBPN = dyn_cast<PHINode>(BB->begin())) {
432 // It is faster to get preds from a PHI than with pred_iterator.
433 for (unsigned i = 0, e = BBPN->getNumIncomingValues(); i != e; ++i)
434 BBPreds.insert(BBPN->getIncomingBlock(i));
436 BBPreds.insert(pred_begin(BB), pred_end(BB));
439 // Walk the preds of DestBB.
440 for (unsigned i = 0, e = DestBBPN->getNumIncomingValues(); i != e; ++i) {
441 BasicBlock *Pred = DestBBPN->getIncomingBlock(i);
442 if (BBPreds.count(Pred)) { // Common predecessor?
443 BBI = DestBB->begin();
444 while (const PHINode *PN = dyn_cast<PHINode>(BBI++)) {
445 const Value *V1 = PN->getIncomingValueForBlock(Pred);
446 const Value *V2 = PN->getIncomingValueForBlock(BB);
448 // If V2 is a phi node in BB, look up what the mapped value will be.
449 if (const PHINode *V2PN = dyn_cast<PHINode>(V2))
450 if (V2PN->getParent() == BB)
451 V2 = V2PN->getIncomingValueForBlock(Pred);
453 // If there is a conflict, bail out.
454 if (V1 != V2) return false;
463 /// Eliminate a basic block that has only phi's and an unconditional branch in
465 void CodeGenPrepare::eliminateMostlyEmptyBlock(BasicBlock *BB) {
466 BranchInst *BI = cast<BranchInst>(BB->getTerminator());
467 BasicBlock *DestBB = BI->getSuccessor(0);
469 DEBUG(dbgs() << "MERGING MOSTLY EMPTY BLOCKS - BEFORE:\n" << *BB << *DestBB);
471 // If the destination block has a single pred, then this is a trivial edge,
473 if (BasicBlock *SinglePred = DestBB->getSinglePredecessor()) {
474 if (SinglePred != DestBB) {
475 // Remember if SinglePred was the entry block of the function. If so, we
476 // will need to move BB back to the entry position.
477 bool isEntry = SinglePred == &SinglePred->getParent()->getEntryBlock();
478 MergeBasicBlockIntoOnlyPred(DestBB, nullptr);
480 if (isEntry && BB != &BB->getParent()->getEntryBlock())
481 BB->moveBefore(&BB->getParent()->getEntryBlock());
483 DEBUG(dbgs() << "AFTER:\n" << *DestBB << "\n\n\n");
488 // Otherwise, we have multiple predecessors of BB. Update the PHIs in DestBB
489 // to handle the new incoming edges it is about to have.
491 for (BasicBlock::iterator BBI = DestBB->begin();
492 (PN = dyn_cast<PHINode>(BBI)); ++BBI) {
493 // Remove the incoming value for BB, and remember it.
494 Value *InVal = PN->removeIncomingValue(BB, false);
496 // Two options: either the InVal is a phi node defined in BB or it is some
497 // value that dominates BB.
498 PHINode *InValPhi = dyn_cast<PHINode>(InVal);
499 if (InValPhi && InValPhi->getParent() == BB) {
500 // Add all of the input values of the input PHI as inputs of this phi.
501 for (unsigned i = 0, e = InValPhi->getNumIncomingValues(); i != e; ++i)
502 PN->addIncoming(InValPhi->getIncomingValue(i),
503 InValPhi->getIncomingBlock(i));
505 // Otherwise, add one instance of the dominating value for each edge that
506 // we will be adding.
507 if (PHINode *BBPN = dyn_cast<PHINode>(BB->begin())) {
508 for (unsigned i = 0, e = BBPN->getNumIncomingValues(); i != e; ++i)
509 PN->addIncoming(InVal, BBPN->getIncomingBlock(i));
511 for (pred_iterator PI = pred_begin(BB), E = pred_end(BB); PI != E; ++PI)
512 PN->addIncoming(InVal, *PI);
517 // The PHIs are now updated, change everything that refers to BB to use
518 // DestBB and remove BB.
519 BB->replaceAllUsesWith(DestBB);
520 BB->eraseFromParent();
523 DEBUG(dbgs() << "AFTER:\n" << *DestBB << "\n\n\n");
526 // Computes a map of base pointer relocation instructions to corresponding
527 // derived pointer relocation instructions given a vector of all relocate calls
528 static void computeBaseDerivedRelocateMap(
529 const SmallVectorImpl<User *> &AllRelocateCalls,
530 DenseMap<IntrinsicInst *, SmallVector<IntrinsicInst *, 2>> &
532 // Collect information in two maps: one primarily for locating the base object
533 // while filling the second map; the second map is the final structure holding
534 // a mapping between Base and corresponding Derived relocate calls
535 DenseMap<std::pair<unsigned, unsigned>, IntrinsicInst *> RelocateIdxMap;
536 for (auto &U : AllRelocateCalls) {
537 GCRelocateOperands ThisRelocate(U);
538 IntrinsicInst *I = cast<IntrinsicInst>(U);
539 auto K = std::make_pair(ThisRelocate.getBasePtrIndex(),
540 ThisRelocate.getDerivedPtrIndex());
541 RelocateIdxMap.insert(std::make_pair(K, I));
543 for (auto &Item : RelocateIdxMap) {
544 std::pair<unsigned, unsigned> Key = Item.first;
545 if (Key.first == Key.second)
546 // Base relocation: nothing to insert
549 IntrinsicInst *I = Item.second;
550 auto BaseKey = std::make_pair(Key.first, Key.first);
552 // We're iterating over RelocateIdxMap so we cannot modify it.
553 auto MaybeBase = RelocateIdxMap.find(BaseKey);
554 if (MaybeBase == RelocateIdxMap.end())
555 // TODO: We might want to insert a new base object relocate and gep off
556 // that, if there are enough derived object relocates.
559 RelocateInstMap[MaybeBase->second].push_back(I);
563 // Accepts a GEP and extracts the operands into a vector provided they're all
564 // small integer constants
565 static bool getGEPSmallConstantIntOffsetV(GetElementPtrInst *GEP,
566 SmallVectorImpl<Value *> &OffsetV) {
567 for (unsigned i = 1; i < GEP->getNumOperands(); i++) {
568 // Only accept small constant integer operands
569 auto Op = dyn_cast<ConstantInt>(GEP->getOperand(i));
570 if (!Op || Op->getZExtValue() > 20)
574 for (unsigned i = 1; i < GEP->getNumOperands(); i++)
575 OffsetV.push_back(GEP->getOperand(i));
579 // Takes a RelocatedBase (base pointer relocation instruction) and Targets to
580 // replace, computes a replacement, and affects it.
582 simplifyRelocatesOffABase(IntrinsicInst *RelocatedBase,
583 const SmallVectorImpl<IntrinsicInst *> &Targets) {
584 bool MadeChange = false;
585 for (auto &ToReplace : Targets) {
586 GCRelocateOperands MasterRelocate(RelocatedBase);
587 GCRelocateOperands ThisRelocate(ToReplace);
589 assert(ThisRelocate.getBasePtrIndex() == MasterRelocate.getBasePtrIndex() &&
590 "Not relocating a derived object of the original base object");
591 if (ThisRelocate.getBasePtrIndex() == ThisRelocate.getDerivedPtrIndex()) {
592 // A duplicate relocate call. TODO: coalesce duplicates.
596 if (RelocatedBase->getParent() != ToReplace->getParent()) {
597 // Base and derived relocates are in different basic blocks.
598 // In this case transform is only valid when base dominates derived
599 // relocate. However it would be too expensive to check dominance
600 // for each such relocate, so we skip the whole transformation.
604 Value *Base = ThisRelocate.getBasePtr();
605 auto Derived = dyn_cast<GetElementPtrInst>(ThisRelocate.getDerivedPtr());
606 if (!Derived || Derived->getPointerOperand() != Base)
609 SmallVector<Value *, 2> OffsetV;
610 if (!getGEPSmallConstantIntOffsetV(Derived, OffsetV))
613 // Create a Builder and replace the target callsite with a gep
614 assert(RelocatedBase->getNextNode() && "Should always have one since it's not a terminator");
616 // Insert after RelocatedBase
617 IRBuilder<> Builder(RelocatedBase->getNextNode());
618 Builder.SetCurrentDebugLocation(ToReplace->getDebugLoc());
620 // If gc_relocate does not match the actual type, cast it to the right type.
621 // In theory, there must be a bitcast after gc_relocate if the type does not
622 // match, and we should reuse it to get the derived pointer. But it could be
626 // %g1 = call coldcc i8 addrspace(1)* @llvm.experimental.gc.relocate.p1i8(...)
631 // %g2 = call coldcc i8 addrspace(1)* @llvm.experimental.gc.relocate.p1i8(...)
635 // %p1 = phi i8 addrspace(1)* [ %g1, %bb1 ], [ %g2, %bb2 ]
636 // %cast = bitcast i8 addrspace(1)* %p1 in to i32 addrspace(1)*
638 // In this case, we can not find the bitcast any more. So we insert a new bitcast
639 // no matter there is already one or not. In this way, we can handle all cases, and
640 // the extra bitcast should be optimized away in later passes.
641 Instruction *ActualRelocatedBase = RelocatedBase;
642 if (RelocatedBase->getType() != Base->getType()) {
643 ActualRelocatedBase =
644 cast<Instruction>(Builder.CreateBitCast(RelocatedBase, Base->getType()));
646 Value *Replacement = Builder.CreateGEP(
647 Derived->getSourceElementType(), ActualRelocatedBase, makeArrayRef(OffsetV));
648 Instruction *ReplacementInst = cast<Instruction>(Replacement);
649 Replacement->takeName(ToReplace);
650 // If the newly generated derived pointer's type does not match the original derived
651 // pointer's type, cast the new derived pointer to match it. Same reasoning as above.
652 Instruction *ActualReplacement = ReplacementInst;
653 if (ReplacementInst->getType() != ToReplace->getType()) {
655 cast<Instruction>(Builder.CreateBitCast(ReplacementInst, ToReplace->getType()));
657 ToReplace->replaceAllUsesWith(ActualReplacement);
658 ToReplace->eraseFromParent();
668 // %ptr = gep %base + 15
669 // %tok = statepoint (%fun, i32 0, i32 0, i32 0, %base, %ptr)
670 // %base' = relocate(%tok, i32 4, i32 4)
671 // %ptr' = relocate(%tok, i32 4, i32 5)
677 // %ptr = gep %base + 15
678 // %tok = statepoint (%fun, i32 0, i32 0, i32 0, %base, %ptr)
679 // %base' = gc.relocate(%tok, i32 4, i32 4)
680 // %ptr' = gep %base' + 15
682 bool CodeGenPrepare::simplifyOffsetableRelocate(Instruction &I) {
683 bool MadeChange = false;
684 SmallVector<User *, 2> AllRelocateCalls;
686 for (auto *U : I.users())
687 if (isGCRelocate(dyn_cast<Instruction>(U)))
688 // Collect all the relocate calls associated with a statepoint
689 AllRelocateCalls.push_back(U);
691 // We need atleast one base pointer relocation + one derived pointer
692 // relocation to mangle
693 if (AllRelocateCalls.size() < 2)
696 // RelocateInstMap is a mapping from the base relocate instruction to the
697 // corresponding derived relocate instructions
698 DenseMap<IntrinsicInst *, SmallVector<IntrinsicInst *, 2>> RelocateInstMap;
699 computeBaseDerivedRelocateMap(AllRelocateCalls, RelocateInstMap);
700 if (RelocateInstMap.empty())
703 for (auto &Item : RelocateInstMap)
704 // Item.first is the RelocatedBase to offset against
705 // Item.second is the vector of Targets to replace
706 MadeChange = simplifyRelocatesOffABase(Item.first, Item.second);
710 /// SinkCast - Sink the specified cast instruction into its user blocks
711 static bool SinkCast(CastInst *CI) {
712 BasicBlock *DefBB = CI->getParent();
714 /// InsertedCasts - Only insert a cast in each block once.
715 DenseMap<BasicBlock*, CastInst*> InsertedCasts;
717 bool MadeChange = false;
718 for (Value::user_iterator UI = CI->user_begin(), E = CI->user_end();
720 Use &TheUse = UI.getUse();
721 Instruction *User = cast<Instruction>(*UI);
723 // Figure out which BB this cast is used in. For PHI's this is the
724 // appropriate predecessor block.
725 BasicBlock *UserBB = User->getParent();
726 if (PHINode *PN = dyn_cast<PHINode>(User)) {
727 UserBB = PN->getIncomingBlock(TheUse);
730 // Preincrement use iterator so we don't invalidate it.
733 // If this user is in the same block as the cast, don't change the cast.
734 if (UserBB == DefBB) continue;
736 // If we have already inserted a cast into this block, use it.
737 CastInst *&InsertedCast = InsertedCasts[UserBB];
740 BasicBlock::iterator InsertPt = UserBB->getFirstInsertionPt();
741 assert(InsertPt != UserBB->end());
742 InsertedCast = CastInst::Create(CI->getOpcode(), CI->getOperand(0),
743 CI->getType(), "", &*InsertPt);
746 // Replace a use of the cast with a use of the new cast.
747 TheUse = InsertedCast;
752 // If we removed all uses, nuke the cast.
753 if (CI->use_empty()) {
754 CI->eraseFromParent();
761 /// If the specified cast instruction is a noop copy (e.g. it's casting from
762 /// one pointer type to another, i32->i8 on PPC), sink it into user blocks to
763 /// reduce the number of virtual registers that must be created and coalesced.
765 /// Return true if any changes are made.
767 static bool OptimizeNoopCopyExpression(CastInst *CI, const TargetLowering &TLI,
768 const DataLayout &DL) {
769 // If this is a noop copy,
770 EVT SrcVT = TLI.getValueType(DL, CI->getOperand(0)->getType());
771 EVT DstVT = TLI.getValueType(DL, CI->getType());
773 // This is an fp<->int conversion?
774 if (SrcVT.isInteger() != DstVT.isInteger())
777 // If this is an extension, it will be a zero or sign extension, which
779 if (SrcVT.bitsLT(DstVT)) return false;
781 // If these values will be promoted, find out what they will be promoted
782 // to. This helps us consider truncates on PPC as noop copies when they
784 if (TLI.getTypeAction(CI->getContext(), SrcVT) ==
785 TargetLowering::TypePromoteInteger)
786 SrcVT = TLI.getTypeToTransformTo(CI->getContext(), SrcVT);
787 if (TLI.getTypeAction(CI->getContext(), DstVT) ==
788 TargetLowering::TypePromoteInteger)
789 DstVT = TLI.getTypeToTransformTo(CI->getContext(), DstVT);
791 // If, after promotion, these are the same types, this is a noop copy.
798 /// Try to combine CI into a call to the llvm.uadd.with.overflow intrinsic if
801 /// Return true if any changes were made.
802 static bool CombineUAddWithOverflow(CmpInst *CI) {
806 m_UAddWithOverflow(m_Value(A), m_Value(B), m_Instruction(AddI))))
809 Type *Ty = AddI->getType();
810 if (!isa<IntegerType>(Ty))
813 // We don't want to move around uses of condition values this late, so we we
814 // check if it is legal to create the call to the intrinsic in the basic
815 // block containing the icmp:
817 if (AddI->getParent() != CI->getParent() && !AddI->hasOneUse())
821 // Someday m_UAddWithOverflow may get smarter, but this is a safe assumption
823 if (AddI->hasOneUse())
824 assert(*AddI->user_begin() == CI && "expected!");
827 Module *M = CI->getParent()->getParent()->getParent();
828 Value *F = Intrinsic::getDeclaration(M, Intrinsic::uadd_with_overflow, Ty);
830 auto *InsertPt = AddI->hasOneUse() ? CI : AddI;
832 auto *UAddWithOverflow =
833 CallInst::Create(F, {A, B}, "uadd.overflow", InsertPt);
834 auto *UAdd = ExtractValueInst::Create(UAddWithOverflow, 0, "uadd", InsertPt);
836 ExtractValueInst::Create(UAddWithOverflow, 1, "overflow", InsertPt);
838 CI->replaceAllUsesWith(Overflow);
839 AddI->replaceAllUsesWith(UAdd);
840 CI->eraseFromParent();
841 AddI->eraseFromParent();
845 /// Sink the given CmpInst into user blocks to reduce the number of virtual
846 /// registers that must be created and coalesced. This is a clear win except on
847 /// targets with multiple condition code registers (PowerPC), where it might
848 /// lose; some adjustment may be wanted there.
850 /// Return true if any changes are made.
851 static bool SinkCmpExpression(CmpInst *CI) {
852 BasicBlock *DefBB = CI->getParent();
854 /// Only insert a cmp in each block once.
855 DenseMap<BasicBlock*, CmpInst*> InsertedCmps;
857 bool MadeChange = false;
858 for (Value::user_iterator UI = CI->user_begin(), E = CI->user_end();
860 Use &TheUse = UI.getUse();
861 Instruction *User = cast<Instruction>(*UI);
863 // Preincrement use iterator so we don't invalidate it.
866 // Don't bother for PHI nodes.
867 if (isa<PHINode>(User))
870 // Figure out which BB this cmp is used in.
871 BasicBlock *UserBB = User->getParent();
873 // If this user is in the same block as the cmp, don't change the cmp.
874 if (UserBB == DefBB) continue;
876 // If we have already inserted a cmp into this block, use it.
877 CmpInst *&InsertedCmp = InsertedCmps[UserBB];
880 BasicBlock::iterator InsertPt = UserBB->getFirstInsertionPt();
881 assert(InsertPt != UserBB->end());
883 CmpInst::Create(CI->getOpcode(), CI->getPredicate(),
884 CI->getOperand(0), CI->getOperand(1), "", &*InsertPt);
887 // Replace a use of the cmp with a use of the new cmp.
888 TheUse = InsertedCmp;
893 // If we removed all uses, nuke the cmp.
894 if (CI->use_empty()) {
895 CI->eraseFromParent();
902 static bool OptimizeCmpExpression(CmpInst *CI) {
903 if (SinkCmpExpression(CI))
906 if (CombineUAddWithOverflow(CI))
912 /// Check if the candidates could be combined with a shift instruction, which
914 /// 1. Truncate instruction
915 /// 2. And instruction and the imm is a mask of the low bits:
916 /// imm & (imm+1) == 0
917 static bool isExtractBitsCandidateUse(Instruction *User) {
918 if (!isa<TruncInst>(User)) {
919 if (User->getOpcode() != Instruction::And ||
920 !isa<ConstantInt>(User->getOperand(1)))
923 const APInt &Cimm = cast<ConstantInt>(User->getOperand(1))->getValue();
925 if ((Cimm & (Cimm + 1)).getBoolValue())
931 /// Sink both shift and truncate instruction to the use of truncate's BB.
933 SinkShiftAndTruncate(BinaryOperator *ShiftI, Instruction *User, ConstantInt *CI,
934 DenseMap<BasicBlock *, BinaryOperator *> &InsertedShifts,
935 const TargetLowering &TLI, const DataLayout &DL) {
936 BasicBlock *UserBB = User->getParent();
937 DenseMap<BasicBlock *, CastInst *> InsertedTruncs;
938 TruncInst *TruncI = dyn_cast<TruncInst>(User);
939 bool MadeChange = false;
941 for (Value::user_iterator TruncUI = TruncI->user_begin(),
942 TruncE = TruncI->user_end();
943 TruncUI != TruncE;) {
945 Use &TruncTheUse = TruncUI.getUse();
946 Instruction *TruncUser = cast<Instruction>(*TruncUI);
947 // Preincrement use iterator so we don't invalidate it.
951 int ISDOpcode = TLI.InstructionOpcodeToISD(TruncUser->getOpcode());
955 // If the use is actually a legal node, there will not be an
956 // implicit truncate.
957 // FIXME: always querying the result type is just an
958 // approximation; some nodes' legality is determined by the
959 // operand or other means. There's no good way to find out though.
960 if (TLI.isOperationLegalOrCustom(
961 ISDOpcode, TLI.getValueType(DL, TruncUser->getType(), true)))
964 // Don't bother for PHI nodes.
965 if (isa<PHINode>(TruncUser))
968 BasicBlock *TruncUserBB = TruncUser->getParent();
970 if (UserBB == TruncUserBB)
973 BinaryOperator *&InsertedShift = InsertedShifts[TruncUserBB];
974 CastInst *&InsertedTrunc = InsertedTruncs[TruncUserBB];
976 if (!InsertedShift && !InsertedTrunc) {
977 BasicBlock::iterator InsertPt = TruncUserBB->getFirstInsertionPt();
978 assert(InsertPt != TruncUserBB->end());
980 if (ShiftI->getOpcode() == Instruction::AShr)
981 InsertedShift = BinaryOperator::CreateAShr(ShiftI->getOperand(0), CI,
984 InsertedShift = BinaryOperator::CreateLShr(ShiftI->getOperand(0), CI,
988 BasicBlock::iterator TruncInsertPt = TruncUserBB->getFirstInsertionPt();
990 assert(TruncInsertPt != TruncUserBB->end());
992 InsertedTrunc = CastInst::Create(TruncI->getOpcode(), InsertedShift,
993 TruncI->getType(), "", &*TruncInsertPt);
997 TruncTheUse = InsertedTrunc;
1003 /// Sink the shift *right* instruction into user blocks if the uses could
1004 /// potentially be combined with this shift instruction and generate BitExtract
1005 /// instruction. It will only be applied if the architecture supports BitExtract
1006 /// instruction. Here is an example:
1008 /// %x.extract.shift = lshr i64 %arg1, 32
1010 /// %x.extract.trunc = trunc i64 %x.extract.shift to i16
1014 /// %x.extract.shift.1 = lshr i64 %arg1, 32
1015 /// %x.extract.trunc = trunc i64 %x.extract.shift.1 to i16
1017 /// CodeGen will recoginze the pattern in BB2 and generate BitExtract
1019 /// Return true if any changes are made.
1020 static bool OptimizeExtractBits(BinaryOperator *ShiftI, ConstantInt *CI,
1021 const TargetLowering &TLI,
1022 const DataLayout &DL) {
1023 BasicBlock *DefBB = ShiftI->getParent();
1025 /// Only insert instructions in each block once.
1026 DenseMap<BasicBlock *, BinaryOperator *> InsertedShifts;
1028 bool shiftIsLegal = TLI.isTypeLegal(TLI.getValueType(DL, ShiftI->getType()));
1030 bool MadeChange = false;
1031 for (Value::user_iterator UI = ShiftI->user_begin(), E = ShiftI->user_end();
1033 Use &TheUse = UI.getUse();
1034 Instruction *User = cast<Instruction>(*UI);
1035 // Preincrement use iterator so we don't invalidate it.
1038 // Don't bother for PHI nodes.
1039 if (isa<PHINode>(User))
1042 if (!isExtractBitsCandidateUse(User))
1045 BasicBlock *UserBB = User->getParent();
1047 if (UserBB == DefBB) {
1048 // If the shift and truncate instruction are in the same BB. The use of
1049 // the truncate(TruncUse) may still introduce another truncate if not
1050 // legal. In this case, we would like to sink both shift and truncate
1051 // instruction to the BB of TruncUse.
1054 // i64 shift.result = lshr i64 opnd, imm
1055 // trunc.result = trunc shift.result to i16
1058 // ----> We will have an implicit truncate here if the architecture does
1059 // not have i16 compare.
1060 // cmp i16 trunc.result, opnd2
1062 if (isa<TruncInst>(User) && shiftIsLegal
1063 // If the type of the truncate is legal, no trucate will be
1064 // introduced in other basic blocks.
1066 (!TLI.isTypeLegal(TLI.getValueType(DL, User->getType()))))
1068 SinkShiftAndTruncate(ShiftI, User, CI, InsertedShifts, TLI, DL);
1072 // If we have already inserted a shift into this block, use it.
1073 BinaryOperator *&InsertedShift = InsertedShifts[UserBB];
1075 if (!InsertedShift) {
1076 BasicBlock::iterator InsertPt = UserBB->getFirstInsertionPt();
1077 assert(InsertPt != UserBB->end());
1079 if (ShiftI->getOpcode() == Instruction::AShr)
1080 InsertedShift = BinaryOperator::CreateAShr(ShiftI->getOperand(0), CI,
1083 InsertedShift = BinaryOperator::CreateLShr(ShiftI->getOperand(0), CI,
1089 // Replace a use of the shift with a use of the new shift.
1090 TheUse = InsertedShift;
1093 // If we removed all uses, nuke the shift.
1094 if (ShiftI->use_empty())
1095 ShiftI->eraseFromParent();
1100 // Translate a masked load intrinsic like
1101 // <16 x i32 > @llvm.masked.load( <16 x i32>* %addr, i32 align,
1102 // <16 x i1> %mask, <16 x i32> %passthru)
1103 // to a chain of basic blocks, with loading element one-by-one if
1104 // the appropriate mask bit is set
1106 // %1 = bitcast i8* %addr to i32*
1107 // %2 = extractelement <16 x i1> %mask, i32 0
1108 // %3 = icmp eq i1 %2, true
1109 // br i1 %3, label %cond.load, label %else
1111 //cond.load: ; preds = %0
1112 // %4 = getelementptr i32* %1, i32 0
1113 // %5 = load i32* %4
1114 // %6 = insertelement <16 x i32> undef, i32 %5, i32 0
1117 //else: ; preds = %0, %cond.load
1118 // %res.phi.else = phi <16 x i32> [ %6, %cond.load ], [ undef, %0 ]
1119 // %7 = extractelement <16 x i1> %mask, i32 1
1120 // %8 = icmp eq i1 %7, true
1121 // br i1 %8, label %cond.load1, label %else2
1123 //cond.load1: ; preds = %else
1124 // %9 = getelementptr i32* %1, i32 1
1125 // %10 = load i32* %9
1126 // %11 = insertelement <16 x i32> %res.phi.else, i32 %10, i32 1
1129 //else2: ; preds = %else, %cond.load1
1130 // %res.phi.else3 = phi <16 x i32> [ %11, %cond.load1 ], [ %res.phi.else, %else ]
1131 // %12 = extractelement <16 x i1> %mask, i32 2
1132 // %13 = icmp eq i1 %12, true
1133 // br i1 %13, label %cond.load4, label %else5
1135 static void ScalarizeMaskedLoad(CallInst *CI) {
1136 Value *Ptr = CI->getArgOperand(0);
1137 Value *Alignment = CI->getArgOperand(1);
1138 Value *Mask = CI->getArgOperand(2);
1139 Value *Src0 = CI->getArgOperand(3);
1141 unsigned AlignVal = cast<ConstantInt>(Alignment)->getZExtValue();
1142 VectorType *VecType = dyn_cast<VectorType>(CI->getType());
1143 assert(VecType && "Unexpected return type of masked load intrinsic");
1145 Type *EltTy = CI->getType()->getVectorElementType();
1147 IRBuilder<> Builder(CI->getContext());
1148 Instruction *InsertPt = CI;
1149 BasicBlock *IfBlock = CI->getParent();
1150 BasicBlock *CondBlock = nullptr;
1151 BasicBlock *PrevIfBlock = CI->getParent();
1153 Builder.SetInsertPoint(InsertPt);
1154 Builder.SetCurrentDebugLocation(CI->getDebugLoc());
1156 // Short-cut if the mask is all-true.
1157 bool IsAllOnesMask = isa<Constant>(Mask) &&
1158 cast<Constant>(Mask)->isAllOnesValue();
1160 if (IsAllOnesMask) {
1161 Value *NewI = Builder.CreateAlignedLoad(Ptr, AlignVal);
1162 CI->replaceAllUsesWith(NewI);
1163 CI->eraseFromParent();
1167 // Adjust alignment for the scalar instruction.
1168 AlignVal = std::min(AlignVal, VecType->getScalarSizeInBits()/8);
1169 // Bitcast %addr fron i8* to EltTy*
1171 EltTy->getPointerTo(cast<PointerType>(Ptr->getType())->getAddressSpace());
1172 Value *FirstEltPtr = Builder.CreateBitCast(Ptr, NewPtrType);
1173 unsigned VectorWidth = VecType->getNumElements();
1175 Value *UndefVal = UndefValue::get(VecType);
1177 // The result vector
1178 Value *VResult = UndefVal;
1180 if (isa<ConstantVector>(Mask)) {
1181 for (unsigned Idx = 0; Idx < VectorWidth; ++Idx) {
1182 if (cast<ConstantVector>(Mask)->getOperand(Idx)->isNullValue())
1185 Builder.CreateInBoundsGEP(EltTy, FirstEltPtr, Builder.getInt32(Idx));
1186 LoadInst* Load = Builder.CreateAlignedLoad(Gep, AlignVal);
1187 VResult = Builder.CreateInsertElement(VResult, Load,
1188 Builder.getInt32(Idx));
1190 Value *NewI = Builder.CreateSelect(Mask, VResult, Src0);
1191 CI->replaceAllUsesWith(NewI);
1192 CI->eraseFromParent();
1196 PHINode *Phi = nullptr;
1197 Value *PrevPhi = UndefVal;
1199 for (unsigned Idx = 0; Idx < VectorWidth; ++Idx) {
1201 // Fill the "else" block, created in the previous iteration
1203 // %res.phi.else3 = phi <16 x i32> [ %11, %cond.load1 ], [ %res.phi.else, %else ]
1204 // %mask_1 = extractelement <16 x i1> %mask, i32 Idx
1205 // %to_load = icmp eq i1 %mask_1, true
1206 // br i1 %to_load, label %cond.load, label %else
1209 Phi = Builder.CreatePHI(VecType, 2, "res.phi.else");
1210 Phi->addIncoming(VResult, CondBlock);
1211 Phi->addIncoming(PrevPhi, PrevIfBlock);
1216 Value *Predicate = Builder.CreateExtractElement(Mask, Builder.getInt32(Idx));
1217 Value *Cmp = Builder.CreateICmp(ICmpInst::ICMP_EQ, Predicate,
1218 ConstantInt::get(Predicate->getType(), 1));
1220 // Create "cond" block
1222 // %EltAddr = getelementptr i32* %1, i32 0
1223 // %Elt = load i32* %EltAddr
1224 // VResult = insertelement <16 x i32> VResult, i32 %Elt, i32 Idx
1226 CondBlock = IfBlock->splitBasicBlock(InsertPt->getIterator(), "cond.load");
1227 Builder.SetInsertPoint(InsertPt);
1230 Builder.CreateInBoundsGEP(EltTy, FirstEltPtr, Builder.getInt32(Idx));
1231 LoadInst *Load = Builder.CreateAlignedLoad(Gep, AlignVal);
1232 VResult = Builder.CreateInsertElement(VResult, Load, Builder.getInt32(Idx));
1234 // Create "else" block, fill it in the next iteration
1235 BasicBlock *NewIfBlock =
1236 CondBlock->splitBasicBlock(InsertPt->getIterator(), "else");
1237 Builder.SetInsertPoint(InsertPt);
1238 Instruction *OldBr = IfBlock->getTerminator();
1239 BranchInst::Create(CondBlock, NewIfBlock, Cmp, OldBr);
1240 OldBr->eraseFromParent();
1241 PrevIfBlock = IfBlock;
1242 IfBlock = NewIfBlock;
1245 Phi = Builder.CreatePHI(VecType, 2, "res.phi.select");
1246 Phi->addIncoming(VResult, CondBlock);
1247 Phi->addIncoming(PrevPhi, PrevIfBlock);
1248 Value *NewI = Builder.CreateSelect(Mask, Phi, Src0);
1249 CI->replaceAllUsesWith(NewI);
1250 CI->eraseFromParent();
1253 // Translate a masked store intrinsic, like
1254 // void @llvm.masked.store(<16 x i32> %src, <16 x i32>* %addr, i32 align,
1256 // to a chain of basic blocks, that stores element one-by-one if
1257 // the appropriate mask bit is set
1259 // %1 = bitcast i8* %addr to i32*
1260 // %2 = extractelement <16 x i1> %mask, i32 0
1261 // %3 = icmp eq i1 %2, true
1262 // br i1 %3, label %cond.store, label %else
1264 // cond.store: ; preds = %0
1265 // %4 = extractelement <16 x i32> %val, i32 0
1266 // %5 = getelementptr i32* %1, i32 0
1267 // store i32 %4, i32* %5
1270 // else: ; preds = %0, %cond.store
1271 // %6 = extractelement <16 x i1> %mask, i32 1
1272 // %7 = icmp eq i1 %6, true
1273 // br i1 %7, label %cond.store1, label %else2
1275 // cond.store1: ; preds = %else
1276 // %8 = extractelement <16 x i32> %val, i32 1
1277 // %9 = getelementptr i32* %1, i32 1
1278 // store i32 %8, i32* %9
1281 static void ScalarizeMaskedStore(CallInst *CI) {
1282 Value *Src = CI->getArgOperand(0);
1283 Value *Ptr = CI->getArgOperand(1);
1284 Value *Alignment = CI->getArgOperand(2);
1285 Value *Mask = CI->getArgOperand(3);
1287 unsigned AlignVal = cast<ConstantInt>(Alignment)->getZExtValue();
1288 VectorType *VecType = dyn_cast<VectorType>(Src->getType());
1289 assert(VecType && "Unexpected data type in masked store intrinsic");
1291 Type *EltTy = VecType->getElementType();
1293 IRBuilder<> Builder(CI->getContext());
1294 Instruction *InsertPt = CI;
1295 BasicBlock *IfBlock = CI->getParent();
1296 Builder.SetInsertPoint(InsertPt);
1297 Builder.SetCurrentDebugLocation(CI->getDebugLoc());
1299 // Short-cut if the mask is all-true.
1300 bool IsAllOnesMask = isa<Constant>(Mask) &&
1301 cast<Constant>(Mask)->isAllOnesValue();
1303 if (IsAllOnesMask) {
1304 Builder.CreateAlignedStore(Src, Ptr, AlignVal);
1305 CI->eraseFromParent();
1309 // Adjust alignment for the scalar instruction.
1310 AlignVal = std::max(AlignVal, VecType->getScalarSizeInBits()/8);
1311 // Bitcast %addr fron i8* to EltTy*
1313 EltTy->getPointerTo(cast<PointerType>(Ptr->getType())->getAddressSpace());
1314 Value *FirstEltPtr = Builder.CreateBitCast(Ptr, NewPtrType);
1315 unsigned VectorWidth = VecType->getNumElements();
1317 if (isa<ConstantVector>(Mask)) {
1318 for (unsigned Idx = 0; Idx < VectorWidth; ++Idx) {
1319 if (cast<ConstantVector>(Mask)->getOperand(Idx)->isNullValue())
1321 Value *OneElt = Builder.CreateExtractElement(Src, Builder.getInt32(Idx));
1323 Builder.CreateInBoundsGEP(EltTy, FirstEltPtr, Builder.getInt32(Idx));
1324 Builder.CreateAlignedStore(OneElt, Gep, AlignVal);
1326 CI->eraseFromParent();
1330 for (unsigned Idx = 0; Idx < VectorWidth; ++Idx) {
1332 // Fill the "else" block, created in the previous iteration
1334 // %mask_1 = extractelement <16 x i1> %mask, i32 Idx
1335 // %to_store = icmp eq i1 %mask_1, true
1336 // br i1 %to_store, label %cond.store, label %else
1338 Value *Predicate = Builder.CreateExtractElement(Mask, Builder.getInt32(Idx));
1339 Value *Cmp = Builder.CreateICmp(ICmpInst::ICMP_EQ, Predicate,
1340 ConstantInt::get(Predicate->getType(), 1));
1342 // Create "cond" block
1344 // %OneElt = extractelement <16 x i32> %Src, i32 Idx
1345 // %EltAddr = getelementptr i32* %1, i32 0
1346 // %store i32 %OneElt, i32* %EltAddr
1348 BasicBlock *CondBlock =
1349 IfBlock->splitBasicBlock(InsertPt->getIterator(), "cond.store");
1350 Builder.SetInsertPoint(InsertPt);
1352 Value *OneElt = Builder.CreateExtractElement(Src, Builder.getInt32(Idx));
1354 Builder.CreateInBoundsGEP(EltTy, FirstEltPtr, Builder.getInt32(Idx));
1355 Builder.CreateAlignedStore(OneElt, Gep, AlignVal);
1357 // Create "else" block, fill it in the next iteration
1358 BasicBlock *NewIfBlock =
1359 CondBlock->splitBasicBlock(InsertPt->getIterator(), "else");
1360 Builder.SetInsertPoint(InsertPt);
1361 Instruction *OldBr = IfBlock->getTerminator();
1362 BranchInst::Create(CondBlock, NewIfBlock, Cmp, OldBr);
1363 OldBr->eraseFromParent();
1364 IfBlock = NewIfBlock;
1366 CI->eraseFromParent();
1369 // Translate a masked gather intrinsic like
1370 // <16 x i32 > @llvm.masked.gather.v16i32( <16 x i32*> %Ptrs, i32 4,
1371 // <16 x i1> %Mask, <16 x i32> %Src)
1372 // to a chain of basic blocks, with loading element one-by-one if
1373 // the appropriate mask bit is set
1375 // % Ptrs = getelementptr i32, i32* %base, <16 x i64> %ind
1376 // % Mask0 = extractelement <16 x i1> %Mask, i32 0
1377 // % ToLoad0 = icmp eq i1 % Mask0, true
1378 // br i1 % ToLoad0, label %cond.load, label %else
1381 // % Ptr0 = extractelement <16 x i32*> %Ptrs, i32 0
1382 // % Load0 = load i32, i32* % Ptr0, align 4
1383 // % Res0 = insertelement <16 x i32> undef, i32 % Load0, i32 0
1387 // %res.phi.else = phi <16 x i32>[% Res0, %cond.load], [undef, % 0]
1388 // % Mask1 = extractelement <16 x i1> %Mask, i32 1
1389 // % ToLoad1 = icmp eq i1 % Mask1, true
1390 // br i1 % ToLoad1, label %cond.load1, label %else2
1393 // % Ptr1 = extractelement <16 x i32*> %Ptrs, i32 1
1394 // % Load1 = load i32, i32* % Ptr1, align 4
1395 // % Res1 = insertelement <16 x i32> %res.phi.else, i32 % Load1, i32 1
1398 // % Result = select <16 x i1> %Mask, <16 x i32> %res.phi.select, <16 x i32> %Src
1399 // ret <16 x i32> %Result
1400 static void ScalarizeMaskedGather(CallInst *CI) {
1401 Value *Ptrs = CI->getArgOperand(0);
1402 Value *Alignment = CI->getArgOperand(1);
1403 Value *Mask = CI->getArgOperand(2);
1404 Value *Src0 = CI->getArgOperand(3);
1406 VectorType *VecType = dyn_cast<VectorType>(CI->getType());
1408 assert(VecType && "Unexpected return type of masked load intrinsic");
1410 IRBuilder<> Builder(CI->getContext());
1411 Instruction *InsertPt = CI;
1412 BasicBlock *IfBlock = CI->getParent();
1413 BasicBlock *CondBlock = nullptr;
1414 BasicBlock *PrevIfBlock = CI->getParent();
1415 Builder.SetInsertPoint(InsertPt);
1416 unsigned AlignVal = cast<ConstantInt>(Alignment)->getZExtValue();
1418 Builder.SetCurrentDebugLocation(CI->getDebugLoc());
1420 Value *UndefVal = UndefValue::get(VecType);
1422 // The result vector
1423 Value *VResult = UndefVal;
1424 unsigned VectorWidth = VecType->getNumElements();
1426 // Shorten the way if the mask is a vector of constants.
1427 bool IsConstMask = isa<ConstantVector>(Mask);
1430 for (unsigned Idx = 0; Idx < VectorWidth; ++Idx) {
1431 if (cast<ConstantVector>(Mask)->getOperand(Idx)->isNullValue())
1433 Value *Ptr = Builder.CreateExtractElement(Ptrs, Builder.getInt32(Idx),
1434 "Ptr" + Twine(Idx));
1435 LoadInst *Load = Builder.CreateAlignedLoad(Ptr, AlignVal,
1436 "Load" + Twine(Idx));
1437 VResult = Builder.CreateInsertElement(VResult, Load,
1438 Builder.getInt32(Idx),
1439 "Res" + Twine(Idx));
1441 Value *NewI = Builder.CreateSelect(Mask, VResult, Src0);
1442 CI->replaceAllUsesWith(NewI);
1443 CI->eraseFromParent();
1447 PHINode *Phi = nullptr;
1448 Value *PrevPhi = UndefVal;
1450 for (unsigned Idx = 0; Idx < VectorWidth; ++Idx) {
1452 // Fill the "else" block, created in the previous iteration
1454 // %Mask1 = extractelement <16 x i1> %Mask, i32 1
1455 // %ToLoad1 = icmp eq i1 %Mask1, true
1456 // br i1 %ToLoad1, label %cond.load, label %else
1459 Phi = Builder.CreatePHI(VecType, 2, "res.phi.else");
1460 Phi->addIncoming(VResult, CondBlock);
1461 Phi->addIncoming(PrevPhi, PrevIfBlock);
1466 Value *Predicate = Builder.CreateExtractElement(Mask,
1467 Builder.getInt32(Idx),
1468 "Mask" + Twine(Idx));
1469 Value *Cmp = Builder.CreateICmp(ICmpInst::ICMP_EQ, Predicate,
1470 ConstantInt::get(Predicate->getType(), 1),
1471 "ToLoad" + Twine(Idx));
1473 // Create "cond" block
1475 // %EltAddr = getelementptr i32* %1, i32 0
1476 // %Elt = load i32* %EltAddr
1477 // VResult = insertelement <16 x i32> VResult, i32 %Elt, i32 Idx
1479 CondBlock = IfBlock->splitBasicBlock(InsertPt, "cond.load");
1480 Builder.SetInsertPoint(InsertPt);
1482 Value *Ptr = Builder.CreateExtractElement(Ptrs, Builder.getInt32(Idx),
1483 "Ptr" + Twine(Idx));
1484 LoadInst *Load = Builder.CreateAlignedLoad(Ptr, AlignVal,
1485 "Load" + Twine(Idx));
1486 VResult = Builder.CreateInsertElement(VResult, Load, Builder.getInt32(Idx),
1487 "Res" + Twine(Idx));
1489 // Create "else" block, fill it in the next iteration
1490 BasicBlock *NewIfBlock = CondBlock->splitBasicBlock(InsertPt, "else");
1491 Builder.SetInsertPoint(InsertPt);
1492 Instruction *OldBr = IfBlock->getTerminator();
1493 BranchInst::Create(CondBlock, NewIfBlock, Cmp, OldBr);
1494 OldBr->eraseFromParent();
1495 PrevIfBlock = IfBlock;
1496 IfBlock = NewIfBlock;
1499 Phi = Builder.CreatePHI(VecType, 2, "res.phi.select");
1500 Phi->addIncoming(VResult, CondBlock);
1501 Phi->addIncoming(PrevPhi, PrevIfBlock);
1502 Value *NewI = Builder.CreateSelect(Mask, Phi, Src0);
1503 CI->replaceAllUsesWith(NewI);
1504 CI->eraseFromParent();
1507 // Translate a masked scatter intrinsic, like
1508 // void @llvm.masked.scatter.v16i32(<16 x i32> %Src, <16 x i32*>* %Ptrs, i32 4,
1510 // to a chain of basic blocks, that stores element one-by-one if
1511 // the appropriate mask bit is set.
1513 // % Ptrs = getelementptr i32, i32* %ptr, <16 x i64> %ind
1514 // % Mask0 = extractelement <16 x i1> % Mask, i32 0
1515 // % ToStore0 = icmp eq i1 % Mask0, true
1516 // br i1 %ToStore0, label %cond.store, label %else
1519 // % Elt0 = extractelement <16 x i32> %Src, i32 0
1520 // % Ptr0 = extractelement <16 x i32*> %Ptrs, i32 0
1521 // store i32 %Elt0, i32* % Ptr0, align 4
1525 // % Mask1 = extractelement <16 x i1> % Mask, i32 1
1526 // % ToStore1 = icmp eq i1 % Mask1, true
1527 // br i1 % ToStore1, label %cond.store1, label %else2
1530 // % Elt1 = extractelement <16 x i32> %Src, i32 1
1531 // % Ptr1 = extractelement <16 x i32*> %Ptrs, i32 1
1532 // store i32 % Elt1, i32* % Ptr1, align 4
1535 static void ScalarizeMaskedScatter(CallInst *CI) {
1536 Value *Src = CI->getArgOperand(0);
1537 Value *Ptrs = CI->getArgOperand(1);
1538 Value *Alignment = CI->getArgOperand(2);
1539 Value *Mask = CI->getArgOperand(3);
1541 assert(isa<VectorType>(Src->getType()) &&
1542 "Unexpected data type in masked scatter intrinsic");
1543 assert(isa<VectorType>(Ptrs->getType()) &&
1544 isa<PointerType>(Ptrs->getType()->getVectorElementType()) &&
1545 "Vector of pointers is expected in masked scatter intrinsic");
1547 IRBuilder<> Builder(CI->getContext());
1548 Instruction *InsertPt = CI;
1549 BasicBlock *IfBlock = CI->getParent();
1550 Builder.SetInsertPoint(InsertPt);
1551 Builder.SetCurrentDebugLocation(CI->getDebugLoc());
1553 unsigned AlignVal = cast<ConstantInt>(Alignment)->getZExtValue();
1554 unsigned VectorWidth = Src->getType()->getVectorNumElements();
1556 // Shorten the way if the mask is a vector of constants.
1557 bool IsConstMask = isa<ConstantVector>(Mask);
1560 for (unsigned Idx = 0; Idx < VectorWidth; ++Idx) {
1561 if (cast<ConstantVector>(Mask)->getOperand(Idx)->isNullValue())
1563 Value *OneElt = Builder.CreateExtractElement(Src, Builder.getInt32(Idx),
1564 "Elt" + Twine(Idx));
1565 Value *Ptr = Builder.CreateExtractElement(Ptrs, Builder.getInt32(Idx),
1566 "Ptr" + Twine(Idx));
1567 Builder.CreateAlignedStore(OneElt, Ptr, AlignVal);
1569 CI->eraseFromParent();
1572 for (unsigned Idx = 0; Idx < VectorWidth; ++Idx) {
1573 // Fill the "else" block, created in the previous iteration
1575 // % Mask1 = extractelement <16 x i1> % Mask, i32 Idx
1576 // % ToStore = icmp eq i1 % Mask1, true
1577 // br i1 % ToStore, label %cond.store, label %else
1579 Value *Predicate = Builder.CreateExtractElement(Mask,
1580 Builder.getInt32(Idx),
1581 "Mask" + Twine(Idx));
1583 Builder.CreateICmp(ICmpInst::ICMP_EQ, Predicate,
1584 ConstantInt::get(Predicate->getType(), 1),
1585 "ToStore" + Twine(Idx));
1587 // Create "cond" block
1589 // % Elt1 = extractelement <16 x i32> %Src, i32 1
1590 // % Ptr1 = extractelement <16 x i32*> %Ptrs, i32 1
1591 // %store i32 % Elt1, i32* % Ptr1
1593 BasicBlock *CondBlock = IfBlock->splitBasicBlock(InsertPt, "cond.store");
1594 Builder.SetInsertPoint(InsertPt);
1596 Value *OneElt = Builder.CreateExtractElement(Src, Builder.getInt32(Idx),
1597 "Elt" + Twine(Idx));
1598 Value *Ptr = Builder.CreateExtractElement(Ptrs, Builder.getInt32(Idx),
1599 "Ptr" + Twine(Idx));
1600 Builder.CreateAlignedStore(OneElt, Ptr, AlignVal);
1602 // Create "else" block, fill it in the next iteration
1603 BasicBlock *NewIfBlock = CondBlock->splitBasicBlock(InsertPt, "else");
1604 Builder.SetInsertPoint(InsertPt);
1605 Instruction *OldBr = IfBlock->getTerminator();
1606 BranchInst::Create(CondBlock, NewIfBlock, Cmp, OldBr);
1607 OldBr->eraseFromParent();
1608 IfBlock = NewIfBlock;
1610 CI->eraseFromParent();
1613 /// If counting leading or trailing zeros is an expensive operation and a zero
1614 /// input is defined, add a check for zero to avoid calling the intrinsic.
1616 /// We want to transform:
1617 /// %z = call i64 @llvm.cttz.i64(i64 %A, i1 false)
1621 /// %cmpz = icmp eq i64 %A, 0
1622 /// br i1 %cmpz, label %cond.end, label %cond.false
1624 /// %z = call i64 @llvm.cttz.i64(i64 %A, i1 true)
1625 /// br label %cond.end
1627 /// %ctz = phi i64 [ 64, %entry ], [ %z, %cond.false ]
1629 /// If the transform is performed, return true and set ModifiedDT to true.
1630 static bool despeculateCountZeros(IntrinsicInst *CountZeros,
1631 const TargetLowering *TLI,
1632 const DataLayout *DL,
1637 // If a zero input is undefined, it doesn't make sense to despeculate that.
1638 if (match(CountZeros->getOperand(1), m_One()))
1641 // If it's cheap to speculate, there's nothing to do.
1642 auto IntrinsicID = CountZeros->getIntrinsicID();
1643 if ((IntrinsicID == Intrinsic::cttz && TLI->isCheapToSpeculateCttz()) ||
1644 (IntrinsicID == Intrinsic::ctlz && TLI->isCheapToSpeculateCtlz()))
1647 // Only handle legal scalar cases. Anything else requires too much work.
1648 Type *Ty = CountZeros->getType();
1649 unsigned SizeInBits = Ty->getPrimitiveSizeInBits();
1650 if (Ty->isVectorTy() || SizeInBits > DL->getLargestLegalIntTypeSize())
1653 // The intrinsic will be sunk behind a compare against zero and branch.
1654 BasicBlock *StartBlock = CountZeros->getParent();
1655 BasicBlock *CallBlock = StartBlock->splitBasicBlock(CountZeros, "cond.false");
1657 // Create another block after the count zero intrinsic. A PHI will be added
1658 // in this block to select the result of the intrinsic or the bit-width
1659 // constant if the input to the intrinsic is zero.
1660 BasicBlock::iterator SplitPt = ++(BasicBlock::iterator(CountZeros));
1661 BasicBlock *EndBlock = CallBlock->splitBasicBlock(SplitPt, "cond.end");
1663 // Set up a builder to create a compare, conditional branch, and PHI.
1664 IRBuilder<> Builder(CountZeros->getContext());
1665 Builder.SetInsertPoint(StartBlock->getTerminator());
1666 Builder.SetCurrentDebugLocation(CountZeros->getDebugLoc());
1668 // Replace the unconditional branch that was created by the first split with
1669 // a compare against zero and a conditional branch.
1670 Value *Zero = Constant::getNullValue(Ty);
1671 Value *Cmp = Builder.CreateICmpEQ(CountZeros->getOperand(0), Zero, "cmpz");
1672 Builder.CreateCondBr(Cmp, EndBlock, CallBlock);
1673 StartBlock->getTerminator()->eraseFromParent();
1675 // Create a PHI in the end block to select either the output of the intrinsic
1676 // or the bit width of the operand.
1677 Builder.SetInsertPoint(&EndBlock->front());
1678 PHINode *PN = Builder.CreatePHI(Ty, 2, "ctz");
1679 CountZeros->replaceAllUsesWith(PN);
1680 Value *BitWidth = Builder.getInt(APInt(SizeInBits, SizeInBits));
1681 PN->addIncoming(BitWidth, StartBlock);
1682 PN->addIncoming(CountZeros, CallBlock);
1684 // We are explicitly handling the zero case, so we can set the intrinsic's
1685 // undefined zero argument to 'true'. This will also prevent reprocessing the
1686 // intrinsic; we only despeculate when a zero input is defined.
1687 CountZeros->setArgOperand(1, Builder.getTrue());
1692 bool CodeGenPrepare::optimizeCallInst(CallInst *CI, bool& ModifiedDT) {
1693 BasicBlock *BB = CI->getParent();
1695 // Lower inline assembly if we can.
1696 // If we found an inline asm expession, and if the target knows how to
1697 // lower it to normal LLVM code, do so now.
1698 if (TLI && isa<InlineAsm>(CI->getCalledValue())) {
1699 if (TLI->ExpandInlineAsm(CI)) {
1700 // Avoid invalidating the iterator.
1701 CurInstIterator = BB->begin();
1702 // Avoid processing instructions out of order, which could cause
1703 // reuse before a value is defined.
1707 // Sink address computing for memory operands into the block.
1708 if (optimizeInlineAsmInst(CI))
1712 // Align the pointer arguments to this call if the target thinks it's a good
1714 unsigned MinSize, PrefAlign;
1715 if (TLI && TLI->shouldAlignPointerArgs(CI, MinSize, PrefAlign)) {
1716 for (auto &Arg : CI->arg_operands()) {
1717 // We want to align both objects whose address is used directly and
1718 // objects whose address is used in casts and GEPs, though it only makes
1719 // sense for GEPs if the offset is a multiple of the desired alignment and
1720 // if size - offset meets the size threshold.
1721 if (!Arg->getType()->isPointerTy())
1723 APInt Offset(DL->getPointerSizeInBits(
1724 cast<PointerType>(Arg->getType())->getAddressSpace()),
1726 Value *Val = Arg->stripAndAccumulateInBoundsConstantOffsets(*DL, Offset);
1727 uint64_t Offset2 = Offset.getLimitedValue();
1728 if ((Offset2 & (PrefAlign-1)) != 0)
1731 if ((AI = dyn_cast<AllocaInst>(Val)) && AI->getAlignment() < PrefAlign &&
1732 DL->getTypeAllocSize(AI->getAllocatedType()) >= MinSize + Offset2)
1733 AI->setAlignment(PrefAlign);
1734 // Global variables can only be aligned if they are defined in this
1735 // object (i.e. they are uniquely initialized in this object), and
1736 // over-aligning global variables that have an explicit section is
1739 if ((GV = dyn_cast<GlobalVariable>(Val)) && GV->hasUniqueInitializer() &&
1740 !GV->hasSection() && GV->getAlignment() < PrefAlign &&
1741 DL->getTypeAllocSize(GV->getType()->getElementType()) >=
1743 GV->setAlignment(PrefAlign);
1745 // If this is a memcpy (or similar) then we may be able to improve the
1747 if (MemIntrinsic *MI = dyn_cast<MemIntrinsic>(CI)) {
1748 unsigned Align = getKnownAlignment(MI->getDest(), *DL);
1749 if (MemTransferInst *MTI = dyn_cast<MemTransferInst>(MI))
1750 Align = std::min(Align, getKnownAlignment(MTI->getSource(), *DL));
1751 if (Align > MI->getAlignment())
1752 MI->setAlignment(ConstantInt::get(MI->getAlignmentType(), Align));
1756 IntrinsicInst *II = dyn_cast<IntrinsicInst>(CI);
1758 switch (II->getIntrinsicID()) {
1760 case Intrinsic::objectsize: {
1761 // Lower all uses of llvm.objectsize.*
1762 bool Min = (cast<ConstantInt>(II->getArgOperand(1))->getZExtValue() == 1);
1763 Type *ReturnTy = CI->getType();
1764 Constant *RetVal = ConstantInt::get(ReturnTy, Min ? 0 : -1ULL);
1766 // Substituting this can cause recursive simplifications, which can
1767 // invalidate our iterator. Use a WeakVH to hold onto it in case this
1769 WeakVH IterHandle(&*CurInstIterator);
1771 replaceAndRecursivelySimplify(CI, RetVal,
1774 // If the iterator instruction was recursively deleted, start over at the
1775 // start of the block.
1776 if (IterHandle != CurInstIterator.getNodePtrUnchecked()) {
1777 CurInstIterator = BB->begin();
1782 case Intrinsic::masked_load: {
1783 // Scalarize unsupported vector masked load
1784 if (!TTI->isLegalMaskedLoad(CI->getType())) {
1785 ScalarizeMaskedLoad(CI);
1791 case Intrinsic::masked_store: {
1792 if (!TTI->isLegalMaskedStore(CI->getArgOperand(0)->getType())) {
1793 ScalarizeMaskedStore(CI);
1799 case Intrinsic::masked_gather: {
1800 if (!TTI->isLegalMaskedGather(CI->getType())) {
1801 ScalarizeMaskedGather(CI);
1807 case Intrinsic::masked_scatter: {
1808 if (!TTI->isLegalMaskedScatter(CI->getArgOperand(0)->getType())) {
1809 ScalarizeMaskedScatter(CI);
1815 case Intrinsic::aarch64_stlxr:
1816 case Intrinsic::aarch64_stxr: {
1817 ZExtInst *ExtVal = dyn_cast<ZExtInst>(CI->getArgOperand(0));
1818 if (!ExtVal || !ExtVal->hasOneUse() ||
1819 ExtVal->getParent() == CI->getParent())
1821 // Sink a zext feeding stlxr/stxr before it, so it can be folded into it.
1822 ExtVal->moveBefore(CI);
1823 // Mark this instruction as "inserted by CGP", so that other
1824 // optimizations don't touch it.
1825 InsertedInsts.insert(ExtVal);
1828 case Intrinsic::invariant_group_barrier:
1829 II->replaceAllUsesWith(II->getArgOperand(0));
1830 II->eraseFromParent();
1833 case Intrinsic::cttz:
1834 case Intrinsic::ctlz:
1835 // If counting zeros is expensive, try to avoid it.
1836 return despeculateCountZeros(II, TLI, DL, ModifiedDT);
1840 // Unknown address space.
1841 // TODO: Target hook to pick which address space the intrinsic cares
1843 unsigned AddrSpace = ~0u;
1844 SmallVector<Value*, 2> PtrOps;
1846 if (TLI->GetAddrModeArguments(II, PtrOps, AccessTy, AddrSpace))
1847 while (!PtrOps.empty())
1848 if (optimizeMemoryInst(II, PtrOps.pop_back_val(), AccessTy, AddrSpace))
1853 // From here on out we're working with named functions.
1854 if (!CI->getCalledFunction()) return false;
1856 // Lower all default uses of _chk calls. This is very similar
1857 // to what InstCombineCalls does, but here we are only lowering calls
1858 // to fortified library functions (e.g. __memcpy_chk) that have the default
1859 // "don't know" as the objectsize. Anything else should be left alone.
1860 FortifiedLibCallSimplifier Simplifier(TLInfo, true);
1861 if (Value *V = Simplifier.optimizeCall(CI)) {
1862 CI->replaceAllUsesWith(V);
1863 CI->eraseFromParent();
1869 /// Look for opportunities to duplicate return instructions to the predecessor
1870 /// to enable tail call optimizations. The case it is currently looking for is:
1873 /// %tmp0 = tail call i32 @f0()
1874 /// br label %return
1876 /// %tmp1 = tail call i32 @f1()
1877 /// br label %return
1879 /// %tmp2 = tail call i32 @f2()
1880 /// br label %return
1882 /// %retval = phi i32 [ %tmp0, %bb0 ], [ %tmp1, %bb1 ], [ %tmp2, %bb2 ]
1890 /// %tmp0 = tail call i32 @f0()
1893 /// %tmp1 = tail call i32 @f1()
1896 /// %tmp2 = tail call i32 @f2()
1899 bool CodeGenPrepare::dupRetToEnableTailCallOpts(BasicBlock *BB) {
1903 ReturnInst *RI = dyn_cast<ReturnInst>(BB->getTerminator());
1907 PHINode *PN = nullptr;
1908 BitCastInst *BCI = nullptr;
1909 Value *V = RI->getReturnValue();
1911 BCI = dyn_cast<BitCastInst>(V);
1913 V = BCI->getOperand(0);
1915 PN = dyn_cast<PHINode>(V);
1920 if (PN && PN->getParent() != BB)
1923 // It's not safe to eliminate the sign / zero extension of the return value.
1924 // See llvm::isInTailCallPosition().
1925 const Function *F = BB->getParent();
1926 AttributeSet CallerAttrs = F->getAttributes();
1927 if (CallerAttrs.hasAttribute(AttributeSet::ReturnIndex, Attribute::ZExt) ||
1928 CallerAttrs.hasAttribute(AttributeSet::ReturnIndex, Attribute::SExt))
1931 // Make sure there are no instructions between the PHI and return, or that the
1932 // return is the first instruction in the block.
1934 BasicBlock::iterator BI = BB->begin();
1935 do { ++BI; } while (isa<DbgInfoIntrinsic>(BI));
1937 // Also skip over the bitcast.
1942 BasicBlock::iterator BI = BB->begin();
1943 while (isa<DbgInfoIntrinsic>(BI)) ++BI;
1948 /// Only dup the ReturnInst if the CallInst is likely to be emitted as a tail
1950 SmallVector<CallInst*, 4> TailCalls;
1952 for (unsigned I = 0, E = PN->getNumIncomingValues(); I != E; ++I) {
1953 CallInst *CI = dyn_cast<CallInst>(PN->getIncomingValue(I));
1954 // Make sure the phi value is indeed produced by the tail call.
1955 if (CI && CI->hasOneUse() && CI->getParent() == PN->getIncomingBlock(I) &&
1956 TLI->mayBeEmittedAsTailCall(CI))
1957 TailCalls.push_back(CI);
1960 SmallPtrSet<BasicBlock*, 4> VisitedBBs;
1961 for (pred_iterator PI = pred_begin(BB), PE = pred_end(BB); PI != PE; ++PI) {
1962 if (!VisitedBBs.insert(*PI).second)
1965 BasicBlock::InstListType &InstList = (*PI)->getInstList();
1966 BasicBlock::InstListType::reverse_iterator RI = InstList.rbegin();
1967 BasicBlock::InstListType::reverse_iterator RE = InstList.rend();
1968 do { ++RI; } while (RI != RE && isa<DbgInfoIntrinsic>(&*RI));
1972 CallInst *CI = dyn_cast<CallInst>(&*RI);
1973 if (CI && CI->use_empty() && TLI->mayBeEmittedAsTailCall(CI))
1974 TailCalls.push_back(CI);
1978 bool Changed = false;
1979 for (unsigned i = 0, e = TailCalls.size(); i != e; ++i) {
1980 CallInst *CI = TailCalls[i];
1983 // Conservatively require the attributes of the call to match those of the
1984 // return. Ignore noalias because it doesn't affect the call sequence.
1985 AttributeSet CalleeAttrs = CS.getAttributes();
1986 if (AttrBuilder(CalleeAttrs, AttributeSet::ReturnIndex).
1987 removeAttribute(Attribute::NoAlias) !=
1988 AttrBuilder(CalleeAttrs, AttributeSet::ReturnIndex).
1989 removeAttribute(Attribute::NoAlias))
1992 // Make sure the call instruction is followed by an unconditional branch to
1993 // the return block.
1994 BasicBlock *CallBB = CI->getParent();
1995 BranchInst *BI = dyn_cast<BranchInst>(CallBB->getTerminator());
1996 if (!BI || !BI->isUnconditional() || BI->getSuccessor(0) != BB)
1999 // Duplicate the return into CallBB.
2000 (void)FoldReturnIntoUncondBranch(RI, BB, CallBB);
2001 ModifiedDT = Changed = true;
2005 // If we eliminated all predecessors of the block, delete the block now.
2006 if (Changed && !BB->hasAddressTaken() && pred_begin(BB) == pred_end(BB))
2007 BB->eraseFromParent();
2012 //===----------------------------------------------------------------------===//
2013 // Memory Optimization
2014 //===----------------------------------------------------------------------===//
2018 /// This is an extended version of TargetLowering::AddrMode
2019 /// which holds actual Value*'s for register values.
2020 struct ExtAddrMode : public TargetLowering::AddrMode {
2023 ExtAddrMode() : BaseReg(nullptr), ScaledReg(nullptr) {}
2024 void print(raw_ostream &OS) const;
2027 bool operator==(const ExtAddrMode& O) const {
2028 return (BaseReg == O.BaseReg) && (ScaledReg == O.ScaledReg) &&
2029 (BaseGV == O.BaseGV) && (BaseOffs == O.BaseOffs) &&
2030 (HasBaseReg == O.HasBaseReg) && (Scale == O.Scale);
2035 static inline raw_ostream &operator<<(raw_ostream &OS, const ExtAddrMode &AM) {
2041 void ExtAddrMode::print(raw_ostream &OS) const {
2042 bool NeedPlus = false;
2045 OS << (NeedPlus ? " + " : "")
2047 BaseGV->printAsOperand(OS, /*PrintType=*/false);
2052 OS << (NeedPlus ? " + " : "")
2058 OS << (NeedPlus ? " + " : "")
2060 BaseReg->printAsOperand(OS, /*PrintType=*/false);
2064 OS << (NeedPlus ? " + " : "")
2066 ScaledReg->printAsOperand(OS, /*PrintType=*/false);
2072 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
2073 void ExtAddrMode::dump() const {
2079 /// \brief This class provides transaction based operation on the IR.
2080 /// Every change made through this class is recorded in the internal state and
2081 /// can be undone (rollback) until commit is called.
2082 class TypePromotionTransaction {
2084 /// \brief This represents the common interface of the individual transaction.
2085 /// Each class implements the logic for doing one specific modification on
2086 /// the IR via the TypePromotionTransaction.
2087 class TypePromotionAction {
2089 /// The Instruction modified.
2093 /// \brief Constructor of the action.
2094 /// The constructor performs the related action on the IR.
2095 TypePromotionAction(Instruction *Inst) : Inst(Inst) {}
2097 virtual ~TypePromotionAction() {}
2099 /// \brief Undo the modification done by this action.
2100 /// When this method is called, the IR must be in the same state as it was
2101 /// before this action was applied.
2102 /// \pre Undoing the action works if and only if the IR is in the exact same
2103 /// state as it was directly after this action was applied.
2104 virtual void undo() = 0;
2106 /// \brief Advocate every change made by this action.
2107 /// When the results on the IR of the action are to be kept, it is important
2108 /// to call this function, otherwise hidden information may be kept forever.
2109 virtual void commit() {
2110 // Nothing to be done, this action is not doing anything.
2114 /// \brief Utility to remember the position of an instruction.
2115 class InsertionHandler {
2116 /// Position of an instruction.
2117 /// Either an instruction:
2118 /// - Is the first in a basic block: BB is used.
2119 /// - Has a previous instructon: PrevInst is used.
2121 Instruction *PrevInst;
2124 /// Remember whether or not the instruction had a previous instruction.
2125 bool HasPrevInstruction;
2128 /// \brief Record the position of \p Inst.
2129 InsertionHandler(Instruction *Inst) {
2130 BasicBlock::iterator It = Inst->getIterator();
2131 HasPrevInstruction = (It != (Inst->getParent()->begin()));
2132 if (HasPrevInstruction)
2133 Point.PrevInst = &*--It;
2135 Point.BB = Inst->getParent();
2138 /// \brief Insert \p Inst at the recorded position.
2139 void insert(Instruction *Inst) {
2140 if (HasPrevInstruction) {
2141 if (Inst->getParent())
2142 Inst->removeFromParent();
2143 Inst->insertAfter(Point.PrevInst);
2145 Instruction *Position = &*Point.BB->getFirstInsertionPt();
2146 if (Inst->getParent())
2147 Inst->moveBefore(Position);
2149 Inst->insertBefore(Position);
2154 /// \brief Move an instruction before another.
2155 class InstructionMoveBefore : public TypePromotionAction {
2156 /// Original position of the instruction.
2157 InsertionHandler Position;
2160 /// \brief Move \p Inst before \p Before.
2161 InstructionMoveBefore(Instruction *Inst, Instruction *Before)
2162 : TypePromotionAction(Inst), Position(Inst) {
2163 DEBUG(dbgs() << "Do: move: " << *Inst << "\nbefore: " << *Before << "\n");
2164 Inst->moveBefore(Before);
2167 /// \brief Move the instruction back to its original position.
2168 void undo() override {
2169 DEBUG(dbgs() << "Undo: moveBefore: " << *Inst << "\n");
2170 Position.insert(Inst);
2174 /// \brief Set the operand of an instruction with a new value.
2175 class OperandSetter : public TypePromotionAction {
2176 /// Original operand of the instruction.
2178 /// Index of the modified instruction.
2182 /// \brief Set \p Idx operand of \p Inst with \p NewVal.
2183 OperandSetter(Instruction *Inst, unsigned Idx, Value *NewVal)
2184 : TypePromotionAction(Inst), Idx(Idx) {
2185 DEBUG(dbgs() << "Do: setOperand: " << Idx << "\n"
2186 << "for:" << *Inst << "\n"
2187 << "with:" << *NewVal << "\n");
2188 Origin = Inst->getOperand(Idx);
2189 Inst->setOperand(Idx, NewVal);
2192 /// \brief Restore the original value of the instruction.
2193 void undo() override {
2194 DEBUG(dbgs() << "Undo: setOperand:" << Idx << "\n"
2195 << "for: " << *Inst << "\n"
2196 << "with: " << *Origin << "\n");
2197 Inst->setOperand(Idx, Origin);
2201 /// \brief Hide the operands of an instruction.
2202 /// Do as if this instruction was not using any of its operands.
2203 class OperandsHider : public TypePromotionAction {
2204 /// The list of original operands.
2205 SmallVector<Value *, 4> OriginalValues;
2208 /// \brief Remove \p Inst from the uses of the operands of \p Inst.
2209 OperandsHider(Instruction *Inst) : TypePromotionAction(Inst) {
2210 DEBUG(dbgs() << "Do: OperandsHider: " << *Inst << "\n");
2211 unsigned NumOpnds = Inst->getNumOperands();
2212 OriginalValues.reserve(NumOpnds);
2213 for (unsigned It = 0; It < NumOpnds; ++It) {
2214 // Save the current operand.
2215 Value *Val = Inst->getOperand(It);
2216 OriginalValues.push_back(Val);
2218 // We could use OperandSetter here, but that would imply an overhead
2219 // that we are not willing to pay.
2220 Inst->setOperand(It, UndefValue::get(Val->getType()));
2224 /// \brief Restore the original list of uses.
2225 void undo() override {
2226 DEBUG(dbgs() << "Undo: OperandsHider: " << *Inst << "\n");
2227 for (unsigned It = 0, EndIt = OriginalValues.size(); It != EndIt; ++It)
2228 Inst->setOperand(It, OriginalValues[It]);
2232 /// \brief Build a truncate instruction.
2233 class TruncBuilder : public TypePromotionAction {
2236 /// \brief Build a truncate instruction of \p Opnd producing a \p Ty
2238 /// trunc Opnd to Ty.
2239 TruncBuilder(Instruction *Opnd, Type *Ty) : TypePromotionAction(Opnd) {
2240 IRBuilder<> Builder(Opnd);
2241 Val = Builder.CreateTrunc(Opnd, Ty, "promoted");
2242 DEBUG(dbgs() << "Do: TruncBuilder: " << *Val << "\n");
2245 /// \brief Get the built value.
2246 Value *getBuiltValue() { return Val; }
2248 /// \brief Remove the built instruction.
2249 void undo() override {
2250 DEBUG(dbgs() << "Undo: TruncBuilder: " << *Val << "\n");
2251 if (Instruction *IVal = dyn_cast<Instruction>(Val))
2252 IVal->eraseFromParent();
2256 /// \brief Build a sign extension instruction.
2257 class SExtBuilder : public TypePromotionAction {
2260 /// \brief Build a sign extension instruction of \p Opnd producing a \p Ty
2262 /// sext Opnd to Ty.
2263 SExtBuilder(Instruction *InsertPt, Value *Opnd, Type *Ty)
2264 : TypePromotionAction(InsertPt) {
2265 IRBuilder<> Builder(InsertPt);
2266 Val = Builder.CreateSExt(Opnd, Ty, "promoted");
2267 DEBUG(dbgs() << "Do: SExtBuilder: " << *Val << "\n");
2270 /// \brief Get the built value.
2271 Value *getBuiltValue() { return Val; }
2273 /// \brief Remove the built instruction.
2274 void undo() override {
2275 DEBUG(dbgs() << "Undo: SExtBuilder: " << *Val << "\n");
2276 if (Instruction *IVal = dyn_cast<Instruction>(Val))
2277 IVal->eraseFromParent();
2281 /// \brief Build a zero extension instruction.
2282 class ZExtBuilder : public TypePromotionAction {
2285 /// \brief Build a zero extension instruction of \p Opnd producing a \p Ty
2287 /// zext Opnd to Ty.
2288 ZExtBuilder(Instruction *InsertPt, Value *Opnd, Type *Ty)
2289 : TypePromotionAction(InsertPt) {
2290 IRBuilder<> Builder(InsertPt);
2291 Val = Builder.CreateZExt(Opnd, Ty, "promoted");
2292 DEBUG(dbgs() << "Do: ZExtBuilder: " << *Val << "\n");
2295 /// \brief Get the built value.
2296 Value *getBuiltValue() { return Val; }
2298 /// \brief Remove the built instruction.
2299 void undo() override {
2300 DEBUG(dbgs() << "Undo: ZExtBuilder: " << *Val << "\n");
2301 if (Instruction *IVal = dyn_cast<Instruction>(Val))
2302 IVal->eraseFromParent();
2306 /// \brief Mutate an instruction to another type.
2307 class TypeMutator : public TypePromotionAction {
2308 /// Record the original type.
2312 /// \brief Mutate the type of \p Inst into \p NewTy.
2313 TypeMutator(Instruction *Inst, Type *NewTy)
2314 : TypePromotionAction(Inst), OrigTy(Inst->getType()) {
2315 DEBUG(dbgs() << "Do: MutateType: " << *Inst << " with " << *NewTy
2317 Inst->mutateType(NewTy);
2320 /// \brief Mutate the instruction back to its original type.
2321 void undo() override {
2322 DEBUG(dbgs() << "Undo: MutateType: " << *Inst << " with " << *OrigTy
2324 Inst->mutateType(OrigTy);
2328 /// \brief Replace the uses of an instruction by another instruction.
2329 class UsesReplacer : public TypePromotionAction {
2330 /// Helper structure to keep track of the replaced uses.
2331 struct InstructionAndIdx {
2332 /// The instruction using the instruction.
2334 /// The index where this instruction is used for Inst.
2336 InstructionAndIdx(Instruction *Inst, unsigned Idx)
2337 : Inst(Inst), Idx(Idx) {}
2340 /// Keep track of the original uses (pair Instruction, Index).
2341 SmallVector<InstructionAndIdx, 4> OriginalUses;
2342 typedef SmallVectorImpl<InstructionAndIdx>::iterator use_iterator;
2345 /// \brief Replace all the use of \p Inst by \p New.
2346 UsesReplacer(Instruction *Inst, Value *New) : TypePromotionAction(Inst) {
2347 DEBUG(dbgs() << "Do: UsersReplacer: " << *Inst << " with " << *New
2349 // Record the original uses.
2350 for (Use &U : Inst->uses()) {
2351 Instruction *UserI = cast<Instruction>(U.getUser());
2352 OriginalUses.push_back(InstructionAndIdx(UserI, U.getOperandNo()));
2354 // Now, we can replace the uses.
2355 Inst->replaceAllUsesWith(New);
2358 /// \brief Reassign the original uses of Inst to Inst.
2359 void undo() override {
2360 DEBUG(dbgs() << "Undo: UsersReplacer: " << *Inst << "\n");
2361 for (use_iterator UseIt = OriginalUses.begin(),
2362 EndIt = OriginalUses.end();
2363 UseIt != EndIt; ++UseIt) {
2364 UseIt->Inst->setOperand(UseIt->Idx, Inst);
2369 /// \brief Remove an instruction from the IR.
2370 class InstructionRemover : public TypePromotionAction {
2371 /// Original position of the instruction.
2372 InsertionHandler Inserter;
2373 /// Helper structure to hide all the link to the instruction. In other
2374 /// words, this helps to do as if the instruction was removed.
2375 OperandsHider Hider;
2376 /// Keep track of the uses replaced, if any.
2377 UsesReplacer *Replacer;
2380 /// \brief Remove all reference of \p Inst and optinally replace all its
2382 /// \pre If !Inst->use_empty(), then New != nullptr
2383 InstructionRemover(Instruction *Inst, Value *New = nullptr)
2384 : TypePromotionAction(Inst), Inserter(Inst), Hider(Inst),
2387 Replacer = new UsesReplacer(Inst, New);
2388 DEBUG(dbgs() << "Do: InstructionRemover: " << *Inst << "\n");
2389 Inst->removeFromParent();
2392 ~InstructionRemover() override { delete Replacer; }
2394 /// \brief Really remove the instruction.
2395 void commit() override { delete Inst; }
2397 /// \brief Resurrect the instruction and reassign it to the proper uses if
2398 /// new value was provided when build this action.
2399 void undo() override {
2400 DEBUG(dbgs() << "Undo: InstructionRemover: " << *Inst << "\n");
2401 Inserter.insert(Inst);
2409 /// Restoration point.
2410 /// The restoration point is a pointer to an action instead of an iterator
2411 /// because the iterator may be invalidated but not the pointer.
2412 typedef const TypePromotionAction *ConstRestorationPt;
2413 /// Advocate every changes made in that transaction.
2415 /// Undo all the changes made after the given point.
2416 void rollback(ConstRestorationPt Point);
2417 /// Get the current restoration point.
2418 ConstRestorationPt getRestorationPoint() const;
2420 /// \name API for IR modification with state keeping to support rollback.
2422 /// Same as Instruction::setOperand.
2423 void setOperand(Instruction *Inst, unsigned Idx, Value *NewVal);
2424 /// Same as Instruction::eraseFromParent.
2425 void eraseInstruction(Instruction *Inst, Value *NewVal = nullptr);
2426 /// Same as Value::replaceAllUsesWith.
2427 void replaceAllUsesWith(Instruction *Inst, Value *New);
2428 /// Same as Value::mutateType.
2429 void mutateType(Instruction *Inst, Type *NewTy);
2430 /// Same as IRBuilder::createTrunc.
2431 Value *createTrunc(Instruction *Opnd, Type *Ty);
2432 /// Same as IRBuilder::createSExt.
2433 Value *createSExt(Instruction *Inst, Value *Opnd, Type *Ty);
2434 /// Same as IRBuilder::createZExt.
2435 Value *createZExt(Instruction *Inst, Value *Opnd, Type *Ty);
2436 /// Same as Instruction::moveBefore.
2437 void moveBefore(Instruction *Inst, Instruction *Before);
2441 /// The ordered list of actions made so far.
2442 SmallVector<std::unique_ptr<TypePromotionAction>, 16> Actions;
2443 typedef SmallVectorImpl<std::unique_ptr<TypePromotionAction>>::iterator CommitPt;
2446 void TypePromotionTransaction::setOperand(Instruction *Inst, unsigned Idx,
2449 make_unique<TypePromotionTransaction::OperandSetter>(Inst, Idx, NewVal));
2452 void TypePromotionTransaction::eraseInstruction(Instruction *Inst,
2455 make_unique<TypePromotionTransaction::InstructionRemover>(Inst, NewVal));
2458 void TypePromotionTransaction::replaceAllUsesWith(Instruction *Inst,
2460 Actions.push_back(make_unique<TypePromotionTransaction::UsesReplacer>(Inst, New));
2463 void TypePromotionTransaction::mutateType(Instruction *Inst, Type *NewTy) {
2464 Actions.push_back(make_unique<TypePromotionTransaction::TypeMutator>(Inst, NewTy));
2467 Value *TypePromotionTransaction::createTrunc(Instruction *Opnd,
2469 std::unique_ptr<TruncBuilder> Ptr(new TruncBuilder(Opnd, Ty));
2470 Value *Val = Ptr->getBuiltValue();
2471 Actions.push_back(std::move(Ptr));
2475 Value *TypePromotionTransaction::createSExt(Instruction *Inst,
2476 Value *Opnd, Type *Ty) {
2477 std::unique_ptr<SExtBuilder> Ptr(new SExtBuilder(Inst, Opnd, Ty));
2478 Value *Val = Ptr->getBuiltValue();
2479 Actions.push_back(std::move(Ptr));
2483 Value *TypePromotionTransaction::createZExt(Instruction *Inst,
2484 Value *Opnd, Type *Ty) {
2485 std::unique_ptr<ZExtBuilder> Ptr(new ZExtBuilder(Inst, Opnd, Ty));
2486 Value *Val = Ptr->getBuiltValue();
2487 Actions.push_back(std::move(Ptr));
2491 void TypePromotionTransaction::moveBefore(Instruction *Inst,
2492 Instruction *Before) {
2494 make_unique<TypePromotionTransaction::InstructionMoveBefore>(Inst, Before));
2497 TypePromotionTransaction::ConstRestorationPt
2498 TypePromotionTransaction::getRestorationPoint() const {
2499 return !Actions.empty() ? Actions.back().get() : nullptr;
2502 void TypePromotionTransaction::commit() {
2503 for (CommitPt It = Actions.begin(), EndIt = Actions.end(); It != EndIt;
2509 void TypePromotionTransaction::rollback(
2510 TypePromotionTransaction::ConstRestorationPt Point) {
2511 while (!Actions.empty() && Point != Actions.back().get()) {
2512 std::unique_ptr<TypePromotionAction> Curr = Actions.pop_back_val();
2517 /// \brief A helper class for matching addressing modes.
2519 /// This encapsulates the logic for matching the target-legal addressing modes.
2520 class AddressingModeMatcher {
2521 SmallVectorImpl<Instruction*> &AddrModeInsts;
2522 const TargetMachine &TM;
2523 const TargetLowering &TLI;
2524 const DataLayout &DL;
2526 /// AccessTy/MemoryInst - This is the type for the access (e.g. double) and
2527 /// the memory instruction that we're computing this address for.
2530 Instruction *MemoryInst;
2532 /// This is the addressing mode that we're building up. This is
2533 /// part of the return value of this addressing mode matching stuff.
2534 ExtAddrMode &AddrMode;
2536 /// The instructions inserted by other CodeGenPrepare optimizations.
2537 const SetOfInstrs &InsertedInsts;
2538 /// A map from the instructions to their type before promotion.
2539 InstrToOrigTy &PromotedInsts;
2540 /// The ongoing transaction where every action should be registered.
2541 TypePromotionTransaction &TPT;
2543 /// This is set to true when we should not do profitability checks.
2544 /// When true, IsProfitableToFoldIntoAddressingMode always returns true.
2545 bool IgnoreProfitability;
2547 AddressingModeMatcher(SmallVectorImpl<Instruction *> &AMI,
2548 const TargetMachine &TM, Type *AT, unsigned AS,
2549 Instruction *MI, ExtAddrMode &AM,
2550 const SetOfInstrs &InsertedInsts,
2551 InstrToOrigTy &PromotedInsts,
2552 TypePromotionTransaction &TPT)
2553 : AddrModeInsts(AMI), TM(TM),
2554 TLI(*TM.getSubtargetImpl(*MI->getParent()->getParent())
2555 ->getTargetLowering()),
2556 DL(MI->getModule()->getDataLayout()), AccessTy(AT), AddrSpace(AS),
2557 MemoryInst(MI), AddrMode(AM), InsertedInsts(InsertedInsts),
2558 PromotedInsts(PromotedInsts), TPT(TPT) {
2559 IgnoreProfitability = false;
2563 /// Find the maximal addressing mode that a load/store of V can fold,
2564 /// give an access type of AccessTy. This returns a list of involved
2565 /// instructions in AddrModeInsts.
2566 /// \p InsertedInsts The instructions inserted by other CodeGenPrepare
2568 /// \p PromotedInsts maps the instructions to their type before promotion.
2569 /// \p The ongoing transaction where every action should be registered.
2570 static ExtAddrMode Match(Value *V, Type *AccessTy, unsigned AS,
2571 Instruction *MemoryInst,
2572 SmallVectorImpl<Instruction*> &AddrModeInsts,
2573 const TargetMachine &TM,
2574 const SetOfInstrs &InsertedInsts,
2575 InstrToOrigTy &PromotedInsts,
2576 TypePromotionTransaction &TPT) {
2579 bool Success = AddressingModeMatcher(AddrModeInsts, TM, AccessTy, AS,
2580 MemoryInst, Result, InsertedInsts,
2581 PromotedInsts, TPT).matchAddr(V, 0);
2582 (void)Success; assert(Success && "Couldn't select *anything*?");
2586 bool matchScaledValue(Value *ScaleReg, int64_t Scale, unsigned Depth);
2587 bool matchAddr(Value *V, unsigned Depth);
2588 bool matchOperationAddr(User *Operation, unsigned Opcode, unsigned Depth,
2589 bool *MovedAway = nullptr);
2590 bool isProfitableToFoldIntoAddressingMode(Instruction *I,
2591 ExtAddrMode &AMBefore,
2592 ExtAddrMode &AMAfter);
2593 bool valueAlreadyLiveAtInst(Value *Val, Value *KnownLive1, Value *KnownLive2);
2594 bool isPromotionProfitable(unsigned NewCost, unsigned OldCost,
2595 Value *PromotedOperand) const;
2598 /// Try adding ScaleReg*Scale to the current addressing mode.
2599 /// Return true and update AddrMode if this addr mode is legal for the target,
2601 bool AddressingModeMatcher::matchScaledValue(Value *ScaleReg, int64_t Scale,
2603 // If Scale is 1, then this is the same as adding ScaleReg to the addressing
2604 // mode. Just process that directly.
2606 return matchAddr(ScaleReg, Depth);
2608 // If the scale is 0, it takes nothing to add this.
2612 // If we already have a scale of this value, we can add to it, otherwise, we
2613 // need an available scale field.
2614 if (AddrMode.Scale != 0 && AddrMode.ScaledReg != ScaleReg)
2617 ExtAddrMode TestAddrMode = AddrMode;
2619 // Add scale to turn X*4+X*3 -> X*7. This could also do things like
2620 // [A+B + A*7] -> [B+A*8].
2621 TestAddrMode.Scale += Scale;
2622 TestAddrMode.ScaledReg = ScaleReg;
2624 // If the new address isn't legal, bail out.
2625 if (!TLI.isLegalAddressingMode(DL, TestAddrMode, AccessTy, AddrSpace))
2628 // It was legal, so commit it.
2629 AddrMode = TestAddrMode;
2631 // Okay, we decided that we can add ScaleReg+Scale to AddrMode. Check now
2632 // to see if ScaleReg is actually X+C. If so, we can turn this into adding
2633 // X*Scale + C*Scale to addr mode.
2634 ConstantInt *CI = nullptr; Value *AddLHS = nullptr;
2635 if (isa<Instruction>(ScaleReg) && // not a constant expr.
2636 match(ScaleReg, m_Add(m_Value(AddLHS), m_ConstantInt(CI)))) {
2637 TestAddrMode.ScaledReg = AddLHS;
2638 TestAddrMode.BaseOffs += CI->getSExtValue()*TestAddrMode.Scale;
2640 // If this addressing mode is legal, commit it and remember that we folded
2641 // this instruction.
2642 if (TLI.isLegalAddressingMode(DL, TestAddrMode, AccessTy, AddrSpace)) {
2643 AddrModeInsts.push_back(cast<Instruction>(ScaleReg));
2644 AddrMode = TestAddrMode;
2649 // Otherwise, not (x+c)*scale, just return what we have.
2653 /// This is a little filter, which returns true if an addressing computation
2654 /// involving I might be folded into a load/store accessing it.
2655 /// This doesn't need to be perfect, but needs to accept at least
2656 /// the set of instructions that MatchOperationAddr can.
2657 static bool MightBeFoldableInst(Instruction *I) {
2658 switch (I->getOpcode()) {
2659 case Instruction::BitCast:
2660 case Instruction::AddrSpaceCast:
2661 // Don't touch identity bitcasts.
2662 if (I->getType() == I->getOperand(0)->getType())
2664 return I->getType()->isPointerTy() || I->getType()->isIntegerTy();
2665 case Instruction::PtrToInt:
2666 // PtrToInt is always a noop, as we know that the int type is pointer sized.
2668 case Instruction::IntToPtr:
2669 // We know the input is intptr_t, so this is foldable.
2671 case Instruction::Add:
2673 case Instruction::Mul:
2674 case Instruction::Shl:
2675 // Can only handle X*C and X << C.
2676 return isa<ConstantInt>(I->getOperand(1));
2677 case Instruction::GetElementPtr:
2684 /// \brief Check whether or not \p Val is a legal instruction for \p TLI.
2685 /// \note \p Val is assumed to be the product of some type promotion.
2686 /// Therefore if \p Val has an undefined state in \p TLI, this is assumed
2687 /// to be legal, as the non-promoted value would have had the same state.
2688 static bool isPromotedInstructionLegal(const TargetLowering &TLI,
2689 const DataLayout &DL, Value *Val) {
2690 Instruction *PromotedInst = dyn_cast<Instruction>(Val);
2693 int ISDOpcode = TLI.InstructionOpcodeToISD(PromotedInst->getOpcode());
2694 // If the ISDOpcode is undefined, it was undefined before the promotion.
2697 // Otherwise, check if the promoted instruction is legal or not.
2698 return TLI.isOperationLegalOrCustom(
2699 ISDOpcode, TLI.getValueType(DL, PromotedInst->getType()));
2702 /// \brief Hepler class to perform type promotion.
2703 class TypePromotionHelper {
2704 /// \brief Utility function to check whether or not a sign or zero extension
2705 /// of \p Inst with \p ConsideredExtType can be moved through \p Inst by
2706 /// either using the operands of \p Inst or promoting \p Inst.
2707 /// The type of the extension is defined by \p IsSExt.
2708 /// In other words, check if:
2709 /// ext (Ty Inst opnd1 opnd2 ... opndN) to ConsideredExtType.
2710 /// #1 Promotion applies:
2711 /// ConsideredExtType Inst (ext opnd1 to ConsideredExtType, ...).
2712 /// #2 Operand reuses:
2713 /// ext opnd1 to ConsideredExtType.
2714 /// \p PromotedInsts maps the instructions to their type before promotion.
2715 static bool canGetThrough(const Instruction *Inst, Type *ConsideredExtType,
2716 const InstrToOrigTy &PromotedInsts, bool IsSExt);
2718 /// \brief Utility function to determine if \p OpIdx should be promoted when
2719 /// promoting \p Inst.
2720 static bool shouldExtOperand(const Instruction *Inst, int OpIdx) {
2721 return !(isa<SelectInst>(Inst) && OpIdx == 0);
2724 /// \brief Utility function to promote the operand of \p Ext when this
2725 /// operand is a promotable trunc or sext or zext.
2726 /// \p PromotedInsts maps the instructions to their type before promotion.
2727 /// \p CreatedInstsCost[out] contains the cost of all instructions
2728 /// created to promote the operand of Ext.
2729 /// Newly added extensions are inserted in \p Exts.
2730 /// Newly added truncates are inserted in \p Truncs.
2731 /// Should never be called directly.
2732 /// \return The promoted value which is used instead of Ext.
2733 static Value *promoteOperandForTruncAndAnyExt(
2734 Instruction *Ext, TypePromotionTransaction &TPT,
2735 InstrToOrigTy &PromotedInsts, unsigned &CreatedInstsCost,
2736 SmallVectorImpl<Instruction *> *Exts,
2737 SmallVectorImpl<Instruction *> *Truncs, const TargetLowering &TLI);
2739 /// \brief Utility function to promote the operand of \p Ext when this
2740 /// operand is promotable and is not a supported trunc or sext.
2741 /// \p PromotedInsts maps the instructions to their type before promotion.
2742 /// \p CreatedInstsCost[out] contains the cost of all the instructions
2743 /// created to promote the operand of Ext.
2744 /// Newly added extensions are inserted in \p Exts.
2745 /// Newly added truncates are inserted in \p Truncs.
2746 /// Should never be called directly.
2747 /// \return The promoted value which is used instead of Ext.
2748 static Value *promoteOperandForOther(Instruction *Ext,
2749 TypePromotionTransaction &TPT,
2750 InstrToOrigTy &PromotedInsts,
2751 unsigned &CreatedInstsCost,
2752 SmallVectorImpl<Instruction *> *Exts,
2753 SmallVectorImpl<Instruction *> *Truncs,
2754 const TargetLowering &TLI, bool IsSExt);
2756 /// \see promoteOperandForOther.
2757 static Value *signExtendOperandForOther(
2758 Instruction *Ext, TypePromotionTransaction &TPT,
2759 InstrToOrigTy &PromotedInsts, unsigned &CreatedInstsCost,
2760 SmallVectorImpl<Instruction *> *Exts,
2761 SmallVectorImpl<Instruction *> *Truncs, const TargetLowering &TLI) {
2762 return promoteOperandForOther(Ext, TPT, PromotedInsts, CreatedInstsCost,
2763 Exts, Truncs, TLI, true);
2766 /// \see promoteOperandForOther.
2767 static Value *zeroExtendOperandForOther(
2768 Instruction *Ext, TypePromotionTransaction &TPT,
2769 InstrToOrigTy &PromotedInsts, unsigned &CreatedInstsCost,
2770 SmallVectorImpl<Instruction *> *Exts,
2771 SmallVectorImpl<Instruction *> *Truncs, const TargetLowering &TLI) {
2772 return promoteOperandForOther(Ext, TPT, PromotedInsts, CreatedInstsCost,
2773 Exts, Truncs, TLI, false);
2777 /// Type for the utility function that promotes the operand of Ext.
2778 typedef Value *(*Action)(Instruction *Ext, TypePromotionTransaction &TPT,
2779 InstrToOrigTy &PromotedInsts,
2780 unsigned &CreatedInstsCost,
2781 SmallVectorImpl<Instruction *> *Exts,
2782 SmallVectorImpl<Instruction *> *Truncs,
2783 const TargetLowering &TLI);
2784 /// \brief Given a sign/zero extend instruction \p Ext, return the approriate
2785 /// action to promote the operand of \p Ext instead of using Ext.
2786 /// \return NULL if no promotable action is possible with the current
2788 /// \p InsertedInsts keeps track of all the instructions inserted by the
2789 /// other CodeGenPrepare optimizations. This information is important
2790 /// because we do not want to promote these instructions as CodeGenPrepare
2791 /// will reinsert them later. Thus creating an infinite loop: create/remove.
2792 /// \p PromotedInsts maps the instructions to their type before promotion.
2793 static Action getAction(Instruction *Ext, const SetOfInstrs &InsertedInsts,
2794 const TargetLowering &TLI,
2795 const InstrToOrigTy &PromotedInsts);
2798 bool TypePromotionHelper::canGetThrough(const Instruction *Inst,
2799 Type *ConsideredExtType,
2800 const InstrToOrigTy &PromotedInsts,
2802 // The promotion helper does not know how to deal with vector types yet.
2803 // To be able to fix that, we would need to fix the places where we
2804 // statically extend, e.g., constants and such.
2805 if (Inst->getType()->isVectorTy())
2808 // We can always get through zext.
2809 if (isa<ZExtInst>(Inst))
2812 // sext(sext) is ok too.
2813 if (IsSExt && isa<SExtInst>(Inst))
2816 // We can get through binary operator, if it is legal. In other words, the
2817 // binary operator must have a nuw or nsw flag.
2818 const BinaryOperator *BinOp = dyn_cast<BinaryOperator>(Inst);
2819 if (BinOp && isa<OverflowingBinaryOperator>(BinOp) &&
2820 ((!IsSExt && BinOp->hasNoUnsignedWrap()) ||
2821 (IsSExt && BinOp->hasNoSignedWrap())))
2824 // Check if we can do the following simplification.
2825 // ext(trunc(opnd)) --> ext(opnd)
2826 if (!isa<TruncInst>(Inst))
2829 Value *OpndVal = Inst->getOperand(0);
2830 // Check if we can use this operand in the extension.
2831 // If the type is larger than the result type of the extension, we cannot.
2832 if (!OpndVal->getType()->isIntegerTy() ||
2833 OpndVal->getType()->getIntegerBitWidth() >
2834 ConsideredExtType->getIntegerBitWidth())
2837 // If the operand of the truncate is not an instruction, we will not have
2838 // any information on the dropped bits.
2839 // (Actually we could for constant but it is not worth the extra logic).
2840 Instruction *Opnd = dyn_cast<Instruction>(OpndVal);
2844 // Check if the source of the type is narrow enough.
2845 // I.e., check that trunc just drops extended bits of the same kind of
2847 // #1 get the type of the operand and check the kind of the extended bits.
2848 const Type *OpndType;
2849 InstrToOrigTy::const_iterator It = PromotedInsts.find(Opnd);
2850 if (It != PromotedInsts.end() && It->second.getInt() == IsSExt)
2851 OpndType = It->second.getPointer();
2852 else if ((IsSExt && isa<SExtInst>(Opnd)) || (!IsSExt && isa<ZExtInst>(Opnd)))
2853 OpndType = Opnd->getOperand(0)->getType();
2857 // #2 check that the truncate just drops extended bits.
2858 return Inst->getType()->getIntegerBitWidth() >=
2859 OpndType->getIntegerBitWidth();
2862 TypePromotionHelper::Action TypePromotionHelper::getAction(
2863 Instruction *Ext, const SetOfInstrs &InsertedInsts,
2864 const TargetLowering &TLI, const InstrToOrigTy &PromotedInsts) {
2865 assert((isa<SExtInst>(Ext) || isa<ZExtInst>(Ext)) &&
2866 "Unexpected instruction type");
2867 Instruction *ExtOpnd = dyn_cast<Instruction>(Ext->getOperand(0));
2868 Type *ExtTy = Ext->getType();
2869 bool IsSExt = isa<SExtInst>(Ext);
2870 // If the operand of the extension is not an instruction, we cannot
2872 // If it, check we can get through.
2873 if (!ExtOpnd || !canGetThrough(ExtOpnd, ExtTy, PromotedInsts, IsSExt))
2876 // Do not promote if the operand has been added by codegenprepare.
2877 // Otherwise, it means we are undoing an optimization that is likely to be
2878 // redone, thus causing potential infinite loop.
2879 if (isa<TruncInst>(ExtOpnd) && InsertedInsts.count(ExtOpnd))
2882 // SExt or Trunc instructions.
2883 // Return the related handler.
2884 if (isa<SExtInst>(ExtOpnd) || isa<TruncInst>(ExtOpnd) ||
2885 isa<ZExtInst>(ExtOpnd))
2886 return promoteOperandForTruncAndAnyExt;
2888 // Regular instruction.
2889 // Abort early if we will have to insert non-free instructions.
2890 if (!ExtOpnd->hasOneUse() && !TLI.isTruncateFree(ExtTy, ExtOpnd->getType()))
2892 return IsSExt ? signExtendOperandForOther : zeroExtendOperandForOther;
2895 Value *TypePromotionHelper::promoteOperandForTruncAndAnyExt(
2896 llvm::Instruction *SExt, TypePromotionTransaction &TPT,
2897 InstrToOrigTy &PromotedInsts, unsigned &CreatedInstsCost,
2898 SmallVectorImpl<Instruction *> *Exts,
2899 SmallVectorImpl<Instruction *> *Truncs, const TargetLowering &TLI) {
2900 // By construction, the operand of SExt is an instruction. Otherwise we cannot
2901 // get through it and this method should not be called.
2902 Instruction *SExtOpnd = cast<Instruction>(SExt->getOperand(0));
2903 Value *ExtVal = SExt;
2904 bool HasMergedNonFreeExt = false;
2905 if (isa<ZExtInst>(SExtOpnd)) {
2906 // Replace s|zext(zext(opnd))
2908 HasMergedNonFreeExt = !TLI.isExtFree(SExtOpnd);
2910 TPT.createZExt(SExt, SExtOpnd->getOperand(0), SExt->getType());
2911 TPT.replaceAllUsesWith(SExt, ZExt);
2912 TPT.eraseInstruction(SExt);
2915 // Replace z|sext(trunc(opnd)) or sext(sext(opnd))
2917 TPT.setOperand(SExt, 0, SExtOpnd->getOperand(0));
2919 CreatedInstsCost = 0;
2921 // Remove dead code.
2922 if (SExtOpnd->use_empty())
2923 TPT.eraseInstruction(SExtOpnd);
2925 // Check if the extension is still needed.
2926 Instruction *ExtInst = dyn_cast<Instruction>(ExtVal);
2927 if (!ExtInst || ExtInst->getType() != ExtInst->getOperand(0)->getType()) {
2930 Exts->push_back(ExtInst);
2931 CreatedInstsCost = !TLI.isExtFree(ExtInst) && !HasMergedNonFreeExt;
2936 // At this point we have: ext ty opnd to ty.
2937 // Reassign the uses of ExtInst to the opnd and remove ExtInst.
2938 Value *NextVal = ExtInst->getOperand(0);
2939 TPT.eraseInstruction(ExtInst, NextVal);
2943 Value *TypePromotionHelper::promoteOperandForOther(
2944 Instruction *Ext, TypePromotionTransaction &TPT,
2945 InstrToOrigTy &PromotedInsts, unsigned &CreatedInstsCost,
2946 SmallVectorImpl<Instruction *> *Exts,
2947 SmallVectorImpl<Instruction *> *Truncs, const TargetLowering &TLI,
2949 // By construction, the operand of Ext is an instruction. Otherwise we cannot
2950 // get through it and this method should not be called.
2951 Instruction *ExtOpnd = cast<Instruction>(Ext->getOperand(0));
2952 CreatedInstsCost = 0;
2953 if (!ExtOpnd->hasOneUse()) {
2954 // ExtOpnd will be promoted.
2955 // All its uses, but Ext, will need to use a truncated value of the
2956 // promoted version.
2957 // Create the truncate now.
2958 Value *Trunc = TPT.createTrunc(Ext, ExtOpnd->getType());
2959 if (Instruction *ITrunc = dyn_cast<Instruction>(Trunc)) {
2960 ITrunc->removeFromParent();
2961 // Insert it just after the definition.
2962 ITrunc->insertAfter(ExtOpnd);
2964 Truncs->push_back(ITrunc);
2967 TPT.replaceAllUsesWith(ExtOpnd, Trunc);
2968 // Restore the operand of Ext (which has been replaced by the previous call
2969 // to replaceAllUsesWith) to avoid creating a cycle trunc <-> sext.
2970 TPT.setOperand(Ext, 0, ExtOpnd);
2973 // Get through the Instruction:
2974 // 1. Update its type.
2975 // 2. Replace the uses of Ext by Inst.
2976 // 3. Extend each operand that needs to be extended.
2978 // Remember the original type of the instruction before promotion.
2979 // This is useful to know that the high bits are sign extended bits.
2980 PromotedInsts.insert(std::pair<Instruction *, TypeIsSExt>(
2981 ExtOpnd, TypeIsSExt(ExtOpnd->getType(), IsSExt)));
2983 TPT.mutateType(ExtOpnd, Ext->getType());
2985 TPT.replaceAllUsesWith(Ext, ExtOpnd);
2987 Instruction *ExtForOpnd = Ext;
2989 DEBUG(dbgs() << "Propagate Ext to operands\n");
2990 for (int OpIdx = 0, EndOpIdx = ExtOpnd->getNumOperands(); OpIdx != EndOpIdx;
2992 DEBUG(dbgs() << "Operand:\n" << *(ExtOpnd->getOperand(OpIdx)) << '\n');
2993 if (ExtOpnd->getOperand(OpIdx)->getType() == Ext->getType() ||
2994 !shouldExtOperand(ExtOpnd, OpIdx)) {
2995 DEBUG(dbgs() << "No need to propagate\n");
2998 // Check if we can statically extend the operand.
2999 Value *Opnd = ExtOpnd->getOperand(OpIdx);
3000 if (const ConstantInt *Cst = dyn_cast<ConstantInt>(Opnd)) {
3001 DEBUG(dbgs() << "Statically extend\n");
3002 unsigned BitWidth = Ext->getType()->getIntegerBitWidth();
3003 APInt CstVal = IsSExt ? Cst->getValue().sext(BitWidth)
3004 : Cst->getValue().zext(BitWidth);
3005 TPT.setOperand(ExtOpnd, OpIdx, ConstantInt::get(Ext->getType(), CstVal));
3008 // UndefValue are typed, so we have to statically sign extend them.
3009 if (isa<UndefValue>(Opnd)) {
3010 DEBUG(dbgs() << "Statically extend\n");
3011 TPT.setOperand(ExtOpnd, OpIdx, UndefValue::get(Ext->getType()));
3015 // Otherwise we have to explicity sign extend the operand.
3016 // Check if Ext was reused to extend an operand.
3018 // If yes, create a new one.
3019 DEBUG(dbgs() << "More operands to ext\n");
3020 Value *ValForExtOpnd = IsSExt ? TPT.createSExt(Ext, Opnd, Ext->getType())
3021 : TPT.createZExt(Ext, Opnd, Ext->getType());
3022 if (!isa<Instruction>(ValForExtOpnd)) {
3023 TPT.setOperand(ExtOpnd, OpIdx, ValForExtOpnd);
3026 ExtForOpnd = cast<Instruction>(ValForExtOpnd);
3029 Exts->push_back(ExtForOpnd);
3030 TPT.setOperand(ExtForOpnd, 0, Opnd);
3032 // Move the sign extension before the insertion point.
3033 TPT.moveBefore(ExtForOpnd, ExtOpnd);
3034 TPT.setOperand(ExtOpnd, OpIdx, ExtForOpnd);
3035 CreatedInstsCost += !TLI.isExtFree(ExtForOpnd);
3036 // If more sext are required, new instructions will have to be created.
3037 ExtForOpnd = nullptr;
3039 if (ExtForOpnd == Ext) {
3040 DEBUG(dbgs() << "Extension is useless now\n");
3041 TPT.eraseInstruction(Ext);
3046 /// Check whether or not promoting an instruction to a wider type is profitable.
3047 /// \p NewCost gives the cost of extension instructions created by the
3049 /// \p OldCost gives the cost of extension instructions before the promotion
3050 /// plus the number of instructions that have been
3051 /// matched in the addressing mode the promotion.
3052 /// \p PromotedOperand is the value that has been promoted.
3053 /// \return True if the promotion is profitable, false otherwise.
3054 bool AddressingModeMatcher::isPromotionProfitable(
3055 unsigned NewCost, unsigned OldCost, Value *PromotedOperand) const {
3056 DEBUG(dbgs() << "OldCost: " << OldCost << "\tNewCost: " << NewCost << '\n');
3057 // The cost of the new extensions is greater than the cost of the
3058 // old extension plus what we folded.
3059 // This is not profitable.
3060 if (NewCost > OldCost)
3062 if (NewCost < OldCost)
3064 // The promotion is neutral but it may help folding the sign extension in
3065 // loads for instance.
3066 // Check that we did not create an illegal instruction.
3067 return isPromotedInstructionLegal(TLI, DL, PromotedOperand);
3070 /// Given an instruction or constant expr, see if we can fold the operation
3071 /// into the addressing mode. If so, update the addressing mode and return
3072 /// true, otherwise return false without modifying AddrMode.
3073 /// If \p MovedAway is not NULL, it contains the information of whether or
3074 /// not AddrInst has to be folded into the addressing mode on success.
3075 /// If \p MovedAway == true, \p AddrInst will not be part of the addressing
3076 /// because it has been moved away.
3077 /// Thus AddrInst must not be added in the matched instructions.
3078 /// This state can happen when AddrInst is a sext, since it may be moved away.
3079 /// Therefore, AddrInst may not be valid when MovedAway is true and it must
3080 /// not be referenced anymore.
3081 bool AddressingModeMatcher::matchOperationAddr(User *AddrInst, unsigned Opcode,
3084 // Avoid exponential behavior on extremely deep expression trees.
3085 if (Depth >= 5) return false;
3087 // By default, all matched instructions stay in place.
3092 case Instruction::PtrToInt:
3093 // PtrToInt is always a noop, as we know that the int type is pointer sized.
3094 return matchAddr(AddrInst->getOperand(0), Depth);
3095 case Instruction::IntToPtr: {
3096 auto AS = AddrInst->getType()->getPointerAddressSpace();
3097 auto PtrTy = MVT::getIntegerVT(DL.getPointerSizeInBits(AS));
3098 // This inttoptr is a no-op if the integer type is pointer sized.
3099 if (TLI.getValueType(DL, AddrInst->getOperand(0)->getType()) == PtrTy)
3100 return matchAddr(AddrInst->getOperand(0), Depth);
3103 case Instruction::BitCast:
3104 // BitCast is always a noop, and we can handle it as long as it is
3105 // int->int or pointer->pointer (we don't want int<->fp or something).
3106 if ((AddrInst->getOperand(0)->getType()->isPointerTy() ||
3107 AddrInst->getOperand(0)->getType()->isIntegerTy()) &&
3108 // Don't touch identity bitcasts. These were probably put here by LSR,
3109 // and we don't want to mess around with them. Assume it knows what it
3111 AddrInst->getOperand(0)->getType() != AddrInst->getType())
3112 return matchAddr(AddrInst->getOperand(0), Depth);
3114 case Instruction::AddrSpaceCast: {
3116 = AddrInst->getOperand(0)->getType()->getPointerAddressSpace();
3117 unsigned DestAS = AddrInst->getType()->getPointerAddressSpace();
3118 if (TLI.isNoopAddrSpaceCast(SrcAS, DestAS))
3119 return matchAddr(AddrInst->getOperand(0), Depth);
3122 case Instruction::Add: {
3123 // Check to see if we can merge in the RHS then the LHS. If so, we win.
3124 ExtAddrMode BackupAddrMode = AddrMode;
3125 unsigned OldSize = AddrModeInsts.size();
3126 // Start a transaction at this point.
3127 // The LHS may match but not the RHS.
3128 // Therefore, we need a higher level restoration point to undo partially
3129 // matched operation.
3130 TypePromotionTransaction::ConstRestorationPt LastKnownGood =
3131 TPT.getRestorationPoint();
3133 if (matchAddr(AddrInst->getOperand(1), Depth+1) &&
3134 matchAddr(AddrInst->getOperand(0), Depth+1))
3137 // Restore the old addr mode info.
3138 AddrMode = BackupAddrMode;
3139 AddrModeInsts.resize(OldSize);
3140 TPT.rollback(LastKnownGood);
3142 // Otherwise this was over-aggressive. Try merging in the LHS then the RHS.
3143 if (matchAddr(AddrInst->getOperand(0), Depth+1) &&
3144 matchAddr(AddrInst->getOperand(1), Depth+1))
3147 // Otherwise we definitely can't merge the ADD in.
3148 AddrMode = BackupAddrMode;
3149 AddrModeInsts.resize(OldSize);
3150 TPT.rollback(LastKnownGood);
3153 //case Instruction::Or:
3154 // TODO: We can handle "Or Val, Imm" iff this OR is equivalent to an ADD.
3156 case Instruction::Mul:
3157 case Instruction::Shl: {
3158 // Can only handle X*C and X << C.
3159 ConstantInt *RHS = dyn_cast<ConstantInt>(AddrInst->getOperand(1));
3162 int64_t Scale = RHS->getSExtValue();
3163 if (Opcode == Instruction::Shl)
3164 Scale = 1LL << Scale;
3166 return matchScaledValue(AddrInst->getOperand(0), Scale, Depth);
3168 case Instruction::GetElementPtr: {
3169 // Scan the GEP. We check it if it contains constant offsets and at most
3170 // one variable offset.
3171 int VariableOperand = -1;
3172 unsigned VariableScale = 0;
3174 int64_t ConstantOffset = 0;
3175 gep_type_iterator GTI = gep_type_begin(AddrInst);
3176 for (unsigned i = 1, e = AddrInst->getNumOperands(); i != e; ++i, ++GTI) {
3177 if (StructType *STy = dyn_cast<StructType>(*GTI)) {
3178 const StructLayout *SL = DL.getStructLayout(STy);
3180 cast<ConstantInt>(AddrInst->getOperand(i))->getZExtValue();
3181 ConstantOffset += SL->getElementOffset(Idx);
3183 uint64_t TypeSize = DL.getTypeAllocSize(GTI.getIndexedType());
3184 if (ConstantInt *CI = dyn_cast<ConstantInt>(AddrInst->getOperand(i))) {
3185 ConstantOffset += CI->getSExtValue()*TypeSize;
3186 } else if (TypeSize) { // Scales of zero don't do anything.
3187 // We only allow one variable index at the moment.
3188 if (VariableOperand != -1)
3191 // Remember the variable index.
3192 VariableOperand = i;
3193 VariableScale = TypeSize;
3198 // A common case is for the GEP to only do a constant offset. In this case,
3199 // just add it to the disp field and check validity.
3200 if (VariableOperand == -1) {
3201 AddrMode.BaseOffs += ConstantOffset;
3202 if (ConstantOffset == 0 ||
3203 TLI.isLegalAddressingMode(DL, AddrMode, AccessTy, AddrSpace)) {
3204 // Check to see if we can fold the base pointer in too.
3205 if (matchAddr(AddrInst->getOperand(0), Depth+1))
3208 AddrMode.BaseOffs -= ConstantOffset;
3212 // Save the valid addressing mode in case we can't match.
3213 ExtAddrMode BackupAddrMode = AddrMode;
3214 unsigned OldSize = AddrModeInsts.size();
3216 // See if the scale and offset amount is valid for this target.
3217 AddrMode.BaseOffs += ConstantOffset;
3219 // Match the base operand of the GEP.
3220 if (!matchAddr(AddrInst->getOperand(0), Depth+1)) {
3221 // If it couldn't be matched, just stuff the value in a register.
3222 if (AddrMode.HasBaseReg) {
3223 AddrMode = BackupAddrMode;
3224 AddrModeInsts.resize(OldSize);
3227 AddrMode.HasBaseReg = true;
3228 AddrMode.BaseReg = AddrInst->getOperand(0);
3231 // Match the remaining variable portion of the GEP.
3232 if (!matchScaledValue(AddrInst->getOperand(VariableOperand), VariableScale,
3234 // If it couldn't be matched, try stuffing the base into a register
3235 // instead of matching it, and retrying the match of the scale.
3236 AddrMode = BackupAddrMode;
3237 AddrModeInsts.resize(OldSize);
3238 if (AddrMode.HasBaseReg)
3240 AddrMode.HasBaseReg = true;
3241 AddrMode.BaseReg = AddrInst->getOperand(0);
3242 AddrMode.BaseOffs += ConstantOffset;
3243 if (!matchScaledValue(AddrInst->getOperand(VariableOperand),
3244 VariableScale, Depth)) {
3245 // If even that didn't work, bail.
3246 AddrMode = BackupAddrMode;
3247 AddrModeInsts.resize(OldSize);
3254 case Instruction::SExt:
3255 case Instruction::ZExt: {
3256 Instruction *Ext = dyn_cast<Instruction>(AddrInst);
3260 // Try to move this ext out of the way of the addressing mode.
3261 // Ask for a method for doing so.
3262 TypePromotionHelper::Action TPH =
3263 TypePromotionHelper::getAction(Ext, InsertedInsts, TLI, PromotedInsts);
3267 TypePromotionTransaction::ConstRestorationPt LastKnownGood =
3268 TPT.getRestorationPoint();
3269 unsigned CreatedInstsCost = 0;
3270 unsigned ExtCost = !TLI.isExtFree(Ext);
3271 Value *PromotedOperand =
3272 TPH(Ext, TPT, PromotedInsts, CreatedInstsCost, nullptr, nullptr, TLI);
3273 // SExt has been moved away.
3274 // Thus either it will be rematched later in the recursive calls or it is
3275 // gone. Anyway, we must not fold it into the addressing mode at this point.
3279 // addr = gep base, idx
3281 // promotedOpnd = ext opnd <- no match here
3282 // op = promoted_add promotedOpnd, 1 <- match (later in recursive calls)
3283 // addr = gep base, op <- match
3287 assert(PromotedOperand &&
3288 "TypePromotionHelper should have filtered out those cases");
3290 ExtAddrMode BackupAddrMode = AddrMode;
3291 unsigned OldSize = AddrModeInsts.size();
3293 if (!matchAddr(PromotedOperand, Depth) ||
3294 // The total of the new cost is equal to the cost of the created
3296 // The total of the old cost is equal to the cost of the extension plus
3297 // what we have saved in the addressing mode.
3298 !isPromotionProfitable(CreatedInstsCost,
3299 ExtCost + (AddrModeInsts.size() - OldSize),
3301 AddrMode = BackupAddrMode;
3302 AddrModeInsts.resize(OldSize);
3303 DEBUG(dbgs() << "Sign extension does not pay off: rollback\n");
3304 TPT.rollback(LastKnownGood);
3313 /// If we can, try to add the value of 'Addr' into the current addressing mode.
3314 /// If Addr can't be added to AddrMode this returns false and leaves AddrMode
3315 /// unmodified. This assumes that Addr is either a pointer type or intptr_t
3318 bool AddressingModeMatcher::matchAddr(Value *Addr, unsigned Depth) {
3319 // Start a transaction at this point that we will rollback if the matching
3321 TypePromotionTransaction::ConstRestorationPt LastKnownGood =
3322 TPT.getRestorationPoint();
3323 if (ConstantInt *CI = dyn_cast<ConstantInt>(Addr)) {
3324 // Fold in immediates if legal for the target.
3325 AddrMode.BaseOffs += CI->getSExtValue();
3326 if (TLI.isLegalAddressingMode(DL, AddrMode, AccessTy, AddrSpace))
3328 AddrMode.BaseOffs -= CI->getSExtValue();
3329 } else if (GlobalValue *GV = dyn_cast<GlobalValue>(Addr)) {
3330 // If this is a global variable, try to fold it into the addressing mode.
3331 if (!AddrMode.BaseGV) {
3332 AddrMode.BaseGV = GV;
3333 if (TLI.isLegalAddressingMode(DL, AddrMode, AccessTy, AddrSpace))
3335 AddrMode.BaseGV = nullptr;
3337 } else if (Instruction *I = dyn_cast<Instruction>(Addr)) {
3338 ExtAddrMode BackupAddrMode = AddrMode;
3339 unsigned OldSize = AddrModeInsts.size();
3341 // Check to see if it is possible to fold this operation.
3342 bool MovedAway = false;
3343 if (matchOperationAddr(I, I->getOpcode(), Depth, &MovedAway)) {
3344 // This instruction may have been moved away. If so, there is nothing
3348 // Okay, it's possible to fold this. Check to see if it is actually
3349 // *profitable* to do so. We use a simple cost model to avoid increasing
3350 // register pressure too much.
3351 if (I->hasOneUse() ||
3352 isProfitableToFoldIntoAddressingMode(I, BackupAddrMode, AddrMode)) {
3353 AddrModeInsts.push_back(I);
3357 // It isn't profitable to do this, roll back.
3358 //cerr << "NOT FOLDING: " << *I;
3359 AddrMode = BackupAddrMode;
3360 AddrModeInsts.resize(OldSize);
3361 TPT.rollback(LastKnownGood);
3363 } else if (ConstantExpr *CE = dyn_cast<ConstantExpr>(Addr)) {
3364 if (matchOperationAddr(CE, CE->getOpcode(), Depth))
3366 TPT.rollback(LastKnownGood);
3367 } else if (isa<ConstantPointerNull>(Addr)) {
3368 // Null pointer gets folded without affecting the addressing mode.
3372 // Worse case, the target should support [reg] addressing modes. :)
3373 if (!AddrMode.HasBaseReg) {
3374 AddrMode.HasBaseReg = true;
3375 AddrMode.BaseReg = Addr;
3376 // Still check for legality in case the target supports [imm] but not [i+r].
3377 if (TLI.isLegalAddressingMode(DL, AddrMode, AccessTy, AddrSpace))
3379 AddrMode.HasBaseReg = false;
3380 AddrMode.BaseReg = nullptr;
3383 // If the base register is already taken, see if we can do [r+r].
3384 if (AddrMode.Scale == 0) {
3386 AddrMode.ScaledReg = Addr;
3387 if (TLI.isLegalAddressingMode(DL, AddrMode, AccessTy, AddrSpace))
3390 AddrMode.ScaledReg = nullptr;
3393 TPT.rollback(LastKnownGood);
3397 /// Check to see if all uses of OpVal by the specified inline asm call are due
3398 /// to memory operands. If so, return true, otherwise return false.
3399 static bool IsOperandAMemoryOperand(CallInst *CI, InlineAsm *IA, Value *OpVal,
3400 const TargetMachine &TM) {
3401 const Function *F = CI->getParent()->getParent();
3402 const TargetLowering *TLI = TM.getSubtargetImpl(*F)->getTargetLowering();
3403 const TargetRegisterInfo *TRI = TM.getSubtargetImpl(*F)->getRegisterInfo();
3404 TargetLowering::AsmOperandInfoVector TargetConstraints =
3405 TLI->ParseConstraints(F->getParent()->getDataLayout(), TRI,
3406 ImmutableCallSite(CI));
3407 for (unsigned i = 0, e = TargetConstraints.size(); i != e; ++i) {
3408 TargetLowering::AsmOperandInfo &OpInfo = TargetConstraints[i];
3410 // Compute the constraint code and ConstraintType to use.
3411 TLI->ComputeConstraintToUse(OpInfo, SDValue());
3413 // If this asm operand is our Value*, and if it isn't an indirect memory
3414 // operand, we can't fold it!
3415 if (OpInfo.CallOperandVal == OpVal &&
3416 (OpInfo.ConstraintType != TargetLowering::C_Memory ||
3417 !OpInfo.isIndirect))
3424 /// Recursively walk all the uses of I until we find a memory use.
3425 /// If we find an obviously non-foldable instruction, return true.
3426 /// Add the ultimately found memory instructions to MemoryUses.
3427 static bool FindAllMemoryUses(
3429 SmallVectorImpl<std::pair<Instruction *, unsigned>> &MemoryUses,
3430 SmallPtrSetImpl<Instruction *> &ConsideredInsts, const TargetMachine &TM) {
3431 // If we already considered this instruction, we're done.
3432 if (!ConsideredInsts.insert(I).second)
3435 // If this is an obviously unfoldable instruction, bail out.
3436 if (!MightBeFoldableInst(I))
3439 // Loop over all the uses, recursively processing them.
3440 for (Use &U : I->uses()) {
3441 Instruction *UserI = cast<Instruction>(U.getUser());
3443 if (LoadInst *LI = dyn_cast<LoadInst>(UserI)) {
3444 MemoryUses.push_back(std::make_pair(LI, U.getOperandNo()));
3448 if (StoreInst *SI = dyn_cast<StoreInst>(UserI)) {
3449 unsigned opNo = U.getOperandNo();
3450 if (opNo == 0) return true; // Storing addr, not into addr.
3451 MemoryUses.push_back(std::make_pair(SI, opNo));
3455 if (CallInst *CI = dyn_cast<CallInst>(UserI)) {
3456 InlineAsm *IA = dyn_cast<InlineAsm>(CI->getCalledValue());
3457 if (!IA) return true;
3459 // If this is a memory operand, we're cool, otherwise bail out.
3460 if (!IsOperandAMemoryOperand(CI, IA, I, TM))
3465 if (FindAllMemoryUses(UserI, MemoryUses, ConsideredInsts, TM))
3472 /// Return true if Val is already known to be live at the use site that we're
3473 /// folding it into. If so, there is no cost to include it in the addressing
3474 /// mode. KnownLive1 and KnownLive2 are two values that we know are live at the
3475 /// instruction already.
3476 bool AddressingModeMatcher::valueAlreadyLiveAtInst(Value *Val,Value *KnownLive1,
3477 Value *KnownLive2) {
3478 // If Val is either of the known-live values, we know it is live!
3479 if (Val == nullptr || Val == KnownLive1 || Val == KnownLive2)
3482 // All values other than instructions and arguments (e.g. constants) are live.
3483 if (!isa<Instruction>(Val) && !isa<Argument>(Val)) return true;
3485 // If Val is a constant sized alloca in the entry block, it is live, this is
3486 // true because it is just a reference to the stack/frame pointer, which is
3487 // live for the whole function.
3488 if (AllocaInst *AI = dyn_cast<AllocaInst>(Val))
3489 if (AI->isStaticAlloca())
3492 // Check to see if this value is already used in the memory instruction's
3493 // block. If so, it's already live into the block at the very least, so we
3494 // can reasonably fold it.
3495 return Val->isUsedInBasicBlock(MemoryInst->getParent());
3498 /// It is possible for the addressing mode of the machine to fold the specified
3499 /// instruction into a load or store that ultimately uses it.
3500 /// However, the specified instruction has multiple uses.
3501 /// Given this, it may actually increase register pressure to fold it
3502 /// into the load. For example, consider this code:
3506 /// use(Y) -> nonload/store
3510 /// In this case, Y has multiple uses, and can be folded into the load of Z
3511 /// (yielding load [X+2]). However, doing this will cause both "X" and "X+1" to
3512 /// be live at the use(Y) line. If we don't fold Y into load Z, we use one
3513 /// fewer register. Since Y can't be folded into "use(Y)" we don't increase the
3514 /// number of computations either.
3516 /// Note that this (like most of CodeGenPrepare) is just a rough heuristic. If
3517 /// X was live across 'load Z' for other reasons, we actually *would* want to
3518 /// fold the addressing mode in the Z case. This would make Y die earlier.
3519 bool AddressingModeMatcher::
3520 isProfitableToFoldIntoAddressingMode(Instruction *I, ExtAddrMode &AMBefore,
3521 ExtAddrMode &AMAfter) {
3522 if (IgnoreProfitability) return true;
3524 // AMBefore is the addressing mode before this instruction was folded into it,
3525 // and AMAfter is the addressing mode after the instruction was folded. Get
3526 // the set of registers referenced by AMAfter and subtract out those
3527 // referenced by AMBefore: this is the set of values which folding in this
3528 // address extends the lifetime of.
3530 // Note that there are only two potential values being referenced here,
3531 // BaseReg and ScaleReg (global addresses are always available, as are any
3532 // folded immediates).
3533 Value *BaseReg = AMAfter.BaseReg, *ScaledReg = AMAfter.ScaledReg;
3535 // If the BaseReg or ScaledReg was referenced by the previous addrmode, their
3536 // lifetime wasn't extended by adding this instruction.
3537 if (valueAlreadyLiveAtInst(BaseReg, AMBefore.BaseReg, AMBefore.ScaledReg))
3539 if (valueAlreadyLiveAtInst(ScaledReg, AMBefore.BaseReg, AMBefore.ScaledReg))
3540 ScaledReg = nullptr;
3542 // If folding this instruction (and it's subexprs) didn't extend any live
3543 // ranges, we're ok with it.
3544 if (!BaseReg && !ScaledReg)
3547 // If all uses of this instruction are ultimately load/store/inlineasm's,
3548 // check to see if their addressing modes will include this instruction. If
3549 // so, we can fold it into all uses, so it doesn't matter if it has multiple
3551 SmallVector<std::pair<Instruction*,unsigned>, 16> MemoryUses;
3552 SmallPtrSet<Instruction*, 16> ConsideredInsts;
3553 if (FindAllMemoryUses(I, MemoryUses, ConsideredInsts, TM))
3554 return false; // Has a non-memory, non-foldable use!
3556 // Now that we know that all uses of this instruction are part of a chain of
3557 // computation involving only operations that could theoretically be folded
3558 // into a memory use, loop over each of these uses and see if they could
3559 // *actually* fold the instruction.
3560 SmallVector<Instruction*, 32> MatchedAddrModeInsts;
3561 for (unsigned i = 0, e = MemoryUses.size(); i != e; ++i) {
3562 Instruction *User = MemoryUses[i].first;
3563 unsigned OpNo = MemoryUses[i].second;
3565 // Get the access type of this use. If the use isn't a pointer, we don't
3566 // know what it accesses.
3567 Value *Address = User->getOperand(OpNo);
3568 PointerType *AddrTy = dyn_cast<PointerType>(Address->getType());
3571 Type *AddressAccessTy = AddrTy->getElementType();
3572 unsigned AS = AddrTy->getAddressSpace();
3574 // Do a match against the root of this address, ignoring profitability. This
3575 // will tell us if the addressing mode for the memory operation will
3576 // *actually* cover the shared instruction.
3578 TypePromotionTransaction::ConstRestorationPt LastKnownGood =
3579 TPT.getRestorationPoint();
3580 AddressingModeMatcher Matcher(MatchedAddrModeInsts, TM, AddressAccessTy, AS,
3581 MemoryInst, Result, InsertedInsts,
3582 PromotedInsts, TPT);
3583 Matcher.IgnoreProfitability = true;
3584 bool Success = Matcher.matchAddr(Address, 0);
3585 (void)Success; assert(Success && "Couldn't select *anything*?");
3587 // The match was to check the profitability, the changes made are not
3588 // part of the original matcher. Therefore, they should be dropped
3589 // otherwise the original matcher will not present the right state.
3590 TPT.rollback(LastKnownGood);
3592 // If the match didn't cover I, then it won't be shared by it.
3593 if (std::find(MatchedAddrModeInsts.begin(), MatchedAddrModeInsts.end(),
3594 I) == MatchedAddrModeInsts.end())
3597 MatchedAddrModeInsts.clear();
3603 } // end anonymous namespace
3605 /// Return true if the specified values are defined in a
3606 /// different basic block than BB.
3607 static bool IsNonLocalValue(Value *V, BasicBlock *BB) {
3608 if (Instruction *I = dyn_cast<Instruction>(V))
3609 return I->getParent() != BB;
3613 /// Load and Store Instructions often have addressing modes that can do
3614 /// significant amounts of computation. As such, instruction selection will try
3615 /// to get the load or store to do as much computation as possible for the
3616 /// program. The problem is that isel can only see within a single block. As
3617 /// such, we sink as much legal addressing mode work into the block as possible.
3619 /// This method is used to optimize both load/store and inline asms with memory
3621 bool CodeGenPrepare::optimizeMemoryInst(Instruction *MemoryInst, Value *Addr,
3622 Type *AccessTy, unsigned AddrSpace) {
3625 // Try to collapse single-value PHI nodes. This is necessary to undo
3626 // unprofitable PRE transformations.
3627 SmallVector<Value*, 8> worklist;
3628 SmallPtrSet<Value*, 16> Visited;
3629 worklist.push_back(Addr);
3631 // Use a worklist to iteratively look through PHI nodes, and ensure that
3632 // the addressing mode obtained from the non-PHI roots of the graph
3634 Value *Consensus = nullptr;
3635 unsigned NumUsesConsensus = 0;
3636 bool IsNumUsesConsensusValid = false;
3637 SmallVector<Instruction*, 16> AddrModeInsts;
3638 ExtAddrMode AddrMode;
3639 TypePromotionTransaction TPT;
3640 TypePromotionTransaction::ConstRestorationPt LastKnownGood =
3641 TPT.getRestorationPoint();
3642 while (!worklist.empty()) {
3643 Value *V = worklist.back();
3644 worklist.pop_back();
3646 // Break use-def graph loops.
3647 if (!Visited.insert(V).second) {
3648 Consensus = nullptr;
3652 // For a PHI node, push all of its incoming values.
3653 if (PHINode *P = dyn_cast<PHINode>(V)) {
3654 for (Value *IncValue : P->incoming_values())
3655 worklist.push_back(IncValue);
3659 // For non-PHIs, determine the addressing mode being computed.
3660 SmallVector<Instruction*, 16> NewAddrModeInsts;
3661 ExtAddrMode NewAddrMode = AddressingModeMatcher::Match(
3662 V, AccessTy, AddrSpace, MemoryInst, NewAddrModeInsts, *TM,
3663 InsertedInsts, PromotedInsts, TPT);
3665 // This check is broken into two cases with very similar code to avoid using
3666 // getNumUses() as much as possible. Some values have a lot of uses, so
3667 // calling getNumUses() unconditionally caused a significant compile-time
3671 AddrMode = NewAddrMode;
3672 AddrModeInsts = NewAddrModeInsts;
3674 } else if (NewAddrMode == AddrMode) {
3675 if (!IsNumUsesConsensusValid) {
3676 NumUsesConsensus = Consensus->getNumUses();
3677 IsNumUsesConsensusValid = true;
3680 // Ensure that the obtained addressing mode is equivalent to that obtained
3681 // for all other roots of the PHI traversal. Also, when choosing one
3682 // such root as representative, select the one with the most uses in order
3683 // to keep the cost modeling heuristics in AddressingModeMatcher
3685 unsigned NumUses = V->getNumUses();
3686 if (NumUses > NumUsesConsensus) {
3688 NumUsesConsensus = NumUses;
3689 AddrModeInsts = NewAddrModeInsts;
3694 Consensus = nullptr;
3698 // If the addressing mode couldn't be determined, or if multiple different
3699 // ones were determined, bail out now.
3701 TPT.rollback(LastKnownGood);
3706 // Check to see if any of the instructions supersumed by this addr mode are
3707 // non-local to I's BB.
3708 bool AnyNonLocal = false;
3709 for (unsigned i = 0, e = AddrModeInsts.size(); i != e; ++i) {
3710 if (IsNonLocalValue(AddrModeInsts[i], MemoryInst->getParent())) {
3716 // If all the instructions matched are already in this BB, don't do anything.
3718 DEBUG(dbgs() << "CGP: Found local addrmode: " << AddrMode << "\n");
3722 // Insert this computation right after this user. Since our caller is
3723 // scanning from the top of the BB to the bottom, reuse of the expr are
3724 // guaranteed to happen later.
3725 IRBuilder<> Builder(MemoryInst);
3727 // Now that we determined the addressing expression we want to use and know
3728 // that we have to sink it into this block. Check to see if we have already
3729 // done this for some other load/store instr in this block. If so, reuse the
3731 Value *&SunkAddr = SunkAddrs[Addr];
3733 DEBUG(dbgs() << "CGP: Reusing nonlocal addrmode: " << AddrMode << " for "
3734 << *MemoryInst << "\n");
3735 if (SunkAddr->getType() != Addr->getType())
3736 SunkAddr = Builder.CreateBitCast(SunkAddr, Addr->getType());
3737 } else if (AddrSinkUsingGEPs ||
3738 (!AddrSinkUsingGEPs.getNumOccurrences() && TM &&
3739 TM->getSubtargetImpl(*MemoryInst->getParent()->getParent())
3741 // By default, we use the GEP-based method when AA is used later. This
3742 // prevents new inttoptr/ptrtoint pairs from degrading AA capabilities.
3743 DEBUG(dbgs() << "CGP: SINKING nonlocal addrmode: " << AddrMode << " for "
3744 << *MemoryInst << "\n");
3745 Type *IntPtrTy = DL->getIntPtrType(Addr->getType());
3746 Value *ResultPtr = nullptr, *ResultIndex = nullptr;
3748 // First, find the pointer.
3749 if (AddrMode.BaseReg && AddrMode.BaseReg->getType()->isPointerTy()) {
3750 ResultPtr = AddrMode.BaseReg;
3751 AddrMode.BaseReg = nullptr;
3754 if (AddrMode.Scale && AddrMode.ScaledReg->getType()->isPointerTy()) {
3755 // We can't add more than one pointer together, nor can we scale a
3756 // pointer (both of which seem meaningless).
3757 if (ResultPtr || AddrMode.Scale != 1)
3760 ResultPtr = AddrMode.ScaledReg;
3764 if (AddrMode.BaseGV) {
3768 ResultPtr = AddrMode.BaseGV;
3771 // If the real base value actually came from an inttoptr, then the matcher
3772 // will look through it and provide only the integer value. In that case,
3774 if (!ResultPtr && AddrMode.BaseReg) {
3776 Builder.CreateIntToPtr(AddrMode.BaseReg, Addr->getType(), "sunkaddr");
3777 AddrMode.BaseReg = nullptr;
3778 } else if (!ResultPtr && AddrMode.Scale == 1) {
3780 Builder.CreateIntToPtr(AddrMode.ScaledReg, Addr->getType(), "sunkaddr");
3785 !AddrMode.BaseReg && !AddrMode.Scale && !AddrMode.BaseOffs) {
3786 SunkAddr = Constant::getNullValue(Addr->getType());
3787 } else if (!ResultPtr) {
3791 Builder.getInt8PtrTy(Addr->getType()->getPointerAddressSpace());
3792 Type *I8Ty = Builder.getInt8Ty();
3794 // Start with the base register. Do this first so that subsequent address
3795 // matching finds it last, which will prevent it from trying to match it
3796 // as the scaled value in case it happens to be a mul. That would be
3797 // problematic if we've sunk a different mul for the scale, because then
3798 // we'd end up sinking both muls.
3799 if (AddrMode.BaseReg) {
3800 Value *V = AddrMode.BaseReg;
3801 if (V->getType() != IntPtrTy)
3802 V = Builder.CreateIntCast(V, IntPtrTy, /*isSigned=*/true, "sunkaddr");
3807 // Add the scale value.
3808 if (AddrMode.Scale) {
3809 Value *V = AddrMode.ScaledReg;
3810 if (V->getType() == IntPtrTy) {
3812 } else if (cast<IntegerType>(IntPtrTy)->getBitWidth() <
3813 cast<IntegerType>(V->getType())->getBitWidth()) {
3814 V = Builder.CreateTrunc(V, IntPtrTy, "sunkaddr");
3816 // It is only safe to sign extend the BaseReg if we know that the math
3817 // required to create it did not overflow before we extend it. Since
3818 // the original IR value was tossed in favor of a constant back when
3819 // the AddrMode was created we need to bail out gracefully if widths
3820 // do not match instead of extending it.
3821 Instruction *I = dyn_cast_or_null<Instruction>(ResultIndex);
3822 if (I && (ResultIndex != AddrMode.BaseReg))
3823 I->eraseFromParent();
3827 if (AddrMode.Scale != 1)
3828 V = Builder.CreateMul(V, ConstantInt::get(IntPtrTy, AddrMode.Scale),
3831 ResultIndex = Builder.CreateAdd(ResultIndex, V, "sunkaddr");
3836 // Add in the Base Offset if present.
3837 if (AddrMode.BaseOffs) {
3838 Value *V = ConstantInt::get(IntPtrTy, AddrMode.BaseOffs);
3840 // We need to add this separately from the scale above to help with
3841 // SDAG consecutive load/store merging.
3842 if (ResultPtr->getType() != I8PtrTy)
3843 ResultPtr = Builder.CreateBitCast(ResultPtr, I8PtrTy);
3844 ResultPtr = Builder.CreateGEP(I8Ty, ResultPtr, ResultIndex, "sunkaddr");
3851 SunkAddr = ResultPtr;
3853 if (ResultPtr->getType() != I8PtrTy)
3854 ResultPtr = Builder.CreateBitCast(ResultPtr, I8PtrTy);
3855 SunkAddr = Builder.CreateGEP(I8Ty, ResultPtr, ResultIndex, "sunkaddr");
3858 if (SunkAddr->getType() != Addr->getType())
3859 SunkAddr = Builder.CreateBitCast(SunkAddr, Addr->getType());
3862 DEBUG(dbgs() << "CGP: SINKING nonlocal addrmode: " << AddrMode << " for "
3863 << *MemoryInst << "\n");
3864 Type *IntPtrTy = DL->getIntPtrType(Addr->getType());
3865 Value *Result = nullptr;
3867 // Start with the base register. Do this first so that subsequent address
3868 // matching finds it last, which will prevent it from trying to match it
3869 // as the scaled value in case it happens to be a mul. That would be
3870 // problematic if we've sunk a different mul for the scale, because then
3871 // we'd end up sinking both muls.
3872 if (AddrMode.BaseReg) {
3873 Value *V = AddrMode.BaseReg;
3874 if (V->getType()->isPointerTy())
3875 V = Builder.CreatePtrToInt(V, IntPtrTy, "sunkaddr");
3876 if (V->getType() != IntPtrTy)
3877 V = Builder.CreateIntCast(V, IntPtrTy, /*isSigned=*/true, "sunkaddr");
3881 // Add the scale value.
3882 if (AddrMode.Scale) {
3883 Value *V = AddrMode.ScaledReg;
3884 if (V->getType() == IntPtrTy) {
3886 } else if (V->getType()->isPointerTy()) {
3887 V = Builder.CreatePtrToInt(V, IntPtrTy, "sunkaddr");
3888 } else if (cast<IntegerType>(IntPtrTy)->getBitWidth() <
3889 cast<IntegerType>(V->getType())->getBitWidth()) {
3890 V = Builder.CreateTrunc(V, IntPtrTy, "sunkaddr");
3892 // It is only safe to sign extend the BaseReg if we know that the math
3893 // required to create it did not overflow before we extend it. Since
3894 // the original IR value was tossed in favor of a constant back when
3895 // the AddrMode was created we need to bail out gracefully if widths
3896 // do not match instead of extending it.
3897 Instruction *I = dyn_cast_or_null<Instruction>(Result);
3898 if (I && (Result != AddrMode.BaseReg))
3899 I->eraseFromParent();
3902 if (AddrMode.Scale != 1)
3903 V = Builder.CreateMul(V, ConstantInt::get(IntPtrTy, AddrMode.Scale),
3906 Result = Builder.CreateAdd(Result, V, "sunkaddr");
3911 // Add in the BaseGV if present.
3912 if (AddrMode.BaseGV) {
3913 Value *V = Builder.CreatePtrToInt(AddrMode.BaseGV, IntPtrTy, "sunkaddr");
3915 Result = Builder.CreateAdd(Result, V, "sunkaddr");
3920 // Add in the Base Offset if present.
3921 if (AddrMode.BaseOffs) {
3922 Value *V = ConstantInt::get(IntPtrTy, AddrMode.BaseOffs);
3924 Result = Builder.CreateAdd(Result, V, "sunkaddr");
3930 SunkAddr = Constant::getNullValue(Addr->getType());
3932 SunkAddr = Builder.CreateIntToPtr(Result, Addr->getType(), "sunkaddr");
3935 MemoryInst->replaceUsesOfWith(Repl, SunkAddr);
3937 // If we have no uses, recursively delete the value and all dead instructions
3939 if (Repl->use_empty()) {
3940 // This can cause recursive deletion, which can invalidate our iterator.
3941 // Use a WeakVH to hold onto it in case this happens.
3942 WeakVH IterHandle(&*CurInstIterator);
3943 BasicBlock *BB = CurInstIterator->getParent();
3945 RecursivelyDeleteTriviallyDeadInstructions(Repl, TLInfo);
3947 if (IterHandle != CurInstIterator.getNodePtrUnchecked()) {
3948 // If the iterator instruction was recursively deleted, start over at the
3949 // start of the block.
3950 CurInstIterator = BB->begin();
3958 /// If there are any memory operands, use OptimizeMemoryInst to sink their
3959 /// address computing into the block when possible / profitable.
3960 bool CodeGenPrepare::optimizeInlineAsmInst(CallInst *CS) {
3961 bool MadeChange = false;
3963 const TargetRegisterInfo *TRI =
3964 TM->getSubtargetImpl(*CS->getParent()->getParent())->getRegisterInfo();
3965 TargetLowering::AsmOperandInfoVector TargetConstraints =
3966 TLI->ParseConstraints(*DL, TRI, CS);
3968 for (unsigned i = 0, e = TargetConstraints.size(); i != e; ++i) {
3969 TargetLowering::AsmOperandInfo &OpInfo = TargetConstraints[i];
3971 // Compute the constraint code and ConstraintType to use.
3972 TLI->ComputeConstraintToUse(OpInfo, SDValue());
3974 if (OpInfo.ConstraintType == TargetLowering::C_Memory &&
3975 OpInfo.isIndirect) {
3976 Value *OpVal = CS->getArgOperand(ArgNo++);
3977 MadeChange |= optimizeMemoryInst(CS, OpVal, OpVal->getType(), ~0u);
3978 } else if (OpInfo.Type == InlineAsm::isInput)
3985 /// \brief Check if all the uses of \p Inst are equivalent (or free) zero or
3986 /// sign extensions.
3987 static bool hasSameExtUse(Instruction *Inst, const TargetLowering &TLI) {
3988 assert(!Inst->use_empty() && "Input must have at least one use");
3989 const Instruction *FirstUser = cast<Instruction>(*Inst->user_begin());
3990 bool IsSExt = isa<SExtInst>(FirstUser);
3991 Type *ExtTy = FirstUser->getType();
3992 for (const User *U : Inst->users()) {
3993 const Instruction *UI = cast<Instruction>(U);
3994 if ((IsSExt && !isa<SExtInst>(UI)) || (!IsSExt && !isa<ZExtInst>(UI)))
3996 Type *CurTy = UI->getType();
3997 // Same input and output types: Same instruction after CSE.
4001 // If IsSExt is true, we are in this situation:
4003 // b = sext ty1 a to ty2
4004 // c = sext ty1 a to ty3
4005 // Assuming ty2 is shorter than ty3, this could be turned into:
4007 // b = sext ty1 a to ty2
4008 // c = sext ty2 b to ty3
4009 // However, the last sext is not free.
4013 // This is a ZExt, maybe this is free to extend from one type to another.
4014 // In that case, we would not account for a different use.
4017 if (ExtTy->getScalarType()->getIntegerBitWidth() >
4018 CurTy->getScalarType()->getIntegerBitWidth()) {
4026 if (!TLI.isZExtFree(NarrowTy, LargeTy))
4029 // All uses are the same or can be derived from one another for free.
4033 /// \brief Try to form ExtLd by promoting \p Exts until they reach a
4034 /// load instruction.
4035 /// If an ext(load) can be formed, it is returned via \p LI for the load
4036 /// and \p Inst for the extension.
4037 /// Otherwise LI == nullptr and Inst == nullptr.
4038 /// When some promotion happened, \p TPT contains the proper state to
4041 /// \return true when promoting was necessary to expose the ext(load)
4042 /// opportunity, false otherwise.
4046 /// %ld = load i32* %addr
4047 /// %add = add nuw i32 %ld, 4
4048 /// %zext = zext i32 %add to i64
4052 /// %ld = load i32* %addr
4053 /// %zext = zext i32 %ld to i64
4054 /// %add = add nuw i64 %zext, 4
4056 /// Thanks to the promotion, we can match zext(load i32*) to i64.
4057 bool CodeGenPrepare::extLdPromotion(TypePromotionTransaction &TPT,
4058 LoadInst *&LI, Instruction *&Inst,
4059 const SmallVectorImpl<Instruction *> &Exts,
4060 unsigned CreatedInstsCost = 0) {
4061 // Iterate over all the extensions to see if one form an ext(load).
4062 for (auto I : Exts) {
4063 // Check if we directly have ext(load).
4064 if ((LI = dyn_cast<LoadInst>(I->getOperand(0)))) {
4066 // No promotion happened here.
4069 // Check whether or not we want to do any promotion.
4070 if (!TLI || !TLI->enableExtLdPromotion() || DisableExtLdPromotion)
4072 // Get the action to perform the promotion.
4073 TypePromotionHelper::Action TPH = TypePromotionHelper::getAction(
4074 I, InsertedInsts, *TLI, PromotedInsts);
4075 // Check if we can promote.
4078 // Save the current state.
4079 TypePromotionTransaction::ConstRestorationPt LastKnownGood =
4080 TPT.getRestorationPoint();
4081 SmallVector<Instruction *, 4> NewExts;
4082 unsigned NewCreatedInstsCost = 0;
4083 unsigned ExtCost = !TLI->isExtFree(I);
4085 Value *PromotedVal = TPH(I, TPT, PromotedInsts, NewCreatedInstsCost,
4086 &NewExts, nullptr, *TLI);
4087 assert(PromotedVal &&
4088 "TypePromotionHelper should have filtered out those cases");
4090 // We would be able to merge only one extension in a load.
4091 // Therefore, if we have more than 1 new extension we heuristically
4092 // cut this search path, because it means we degrade the code quality.
4093 // With exactly 2, the transformation is neutral, because we will merge
4094 // one extension but leave one. However, we optimistically keep going,
4095 // because the new extension may be removed too.
4096 long long TotalCreatedInstsCost = CreatedInstsCost + NewCreatedInstsCost;
4097 TotalCreatedInstsCost -= ExtCost;
4098 if (!StressExtLdPromotion &&
4099 (TotalCreatedInstsCost > 1 ||
4100 !isPromotedInstructionLegal(*TLI, *DL, PromotedVal))) {
4101 // The promotion is not profitable, rollback to the previous state.
4102 TPT.rollback(LastKnownGood);
4105 // The promotion is profitable.
4106 // Check if it exposes an ext(load).
4107 (void)extLdPromotion(TPT, LI, Inst, NewExts, TotalCreatedInstsCost);
4108 if (LI && (StressExtLdPromotion || NewCreatedInstsCost <= ExtCost ||
4109 // If we have created a new extension, i.e., now we have two
4110 // extensions. We must make sure one of them is merged with
4111 // the load, otherwise we may degrade the code quality.
4112 (LI->hasOneUse() || hasSameExtUse(LI, *TLI))))
4113 // Promotion happened.
4115 // If this does not help to expose an ext(load) then, rollback.
4116 TPT.rollback(LastKnownGood);
4118 // None of the extension can form an ext(load).
4124 /// Move a zext or sext fed by a load into the same basic block as the load,
4125 /// unless conditions are unfavorable. This allows SelectionDAG to fold the
4126 /// extend into the load.
4127 /// \p I[in/out] the extension may be modified during the process if some
4128 /// promotions apply.
4130 bool CodeGenPrepare::moveExtToFormExtLoad(Instruction *&I) {
4131 // Try to promote a chain of computation if it allows to form
4132 // an extended load.
4133 TypePromotionTransaction TPT;
4134 TypePromotionTransaction::ConstRestorationPt LastKnownGood =
4135 TPT.getRestorationPoint();
4136 SmallVector<Instruction *, 1> Exts;
4138 // Look for a load being extended.
4139 LoadInst *LI = nullptr;
4140 Instruction *OldExt = I;
4141 bool HasPromoted = extLdPromotion(TPT, LI, I, Exts);
4143 assert(!HasPromoted && !LI && "If we did not match any load instruction "
4144 "the code must remain the same");
4149 // If they're already in the same block, there's nothing to do.
4150 // Make the cheap checks first if we did not promote.
4151 // If we promoted, we need to check if it is indeed profitable.
4152 if (!HasPromoted && LI->getParent() == I->getParent())
4155 EVT VT = TLI->getValueType(*DL, I->getType());
4156 EVT LoadVT = TLI->getValueType(*DL, LI->getType());
4158 // If the load has other users and the truncate is not free, this probably
4159 // isn't worthwhile.
4160 if (!LI->hasOneUse() && TLI &&
4161 (TLI->isTypeLegal(LoadVT) || !TLI->isTypeLegal(VT)) &&
4162 !TLI->isTruncateFree(I->getType(), LI->getType())) {
4164 TPT.rollback(LastKnownGood);
4168 // Check whether the target supports casts folded into loads.
4170 if (isa<ZExtInst>(I))
4171 LType = ISD::ZEXTLOAD;
4173 assert(isa<SExtInst>(I) && "Unexpected ext type!");
4174 LType = ISD::SEXTLOAD;
4176 if (TLI && !TLI->isLoadExtLegal(LType, VT, LoadVT)) {
4178 TPT.rollback(LastKnownGood);
4182 // Move the extend into the same block as the load, so that SelectionDAG
4185 I->removeFromParent();
4191 bool CodeGenPrepare::optimizeExtUses(Instruction *I) {
4192 BasicBlock *DefBB = I->getParent();
4194 // If the result of a {s|z}ext and its source are both live out, rewrite all
4195 // other uses of the source with result of extension.
4196 Value *Src = I->getOperand(0);
4197 if (Src->hasOneUse())
4200 // Only do this xform if truncating is free.
4201 if (TLI && !TLI->isTruncateFree(I->getType(), Src->getType()))
4204 // Only safe to perform the optimization if the source is also defined in
4206 if (!isa<Instruction>(Src) || DefBB != cast<Instruction>(Src)->getParent())
4209 bool DefIsLiveOut = false;
4210 for (User *U : I->users()) {
4211 Instruction *UI = cast<Instruction>(U);
4213 // Figure out which BB this ext is used in.
4214 BasicBlock *UserBB = UI->getParent();
4215 if (UserBB == DefBB) continue;
4216 DefIsLiveOut = true;
4222 // Make sure none of the uses are PHI nodes.
4223 for (User *U : Src->users()) {
4224 Instruction *UI = cast<Instruction>(U);
4225 BasicBlock *UserBB = UI->getParent();
4226 if (UserBB == DefBB) continue;
4227 // Be conservative. We don't want this xform to end up introducing
4228 // reloads just before load / store instructions.
4229 if (isa<PHINode>(UI) || isa<LoadInst>(UI) || isa<StoreInst>(UI))
4233 // InsertedTruncs - Only insert one trunc in each block once.
4234 DenseMap<BasicBlock*, Instruction*> InsertedTruncs;
4236 bool MadeChange = false;
4237 for (Use &U : Src->uses()) {
4238 Instruction *User = cast<Instruction>(U.getUser());
4240 // Figure out which BB this ext is used in.
4241 BasicBlock *UserBB = User->getParent();
4242 if (UserBB == DefBB) continue;
4244 // Both src and def are live in this block. Rewrite the use.
4245 Instruction *&InsertedTrunc = InsertedTruncs[UserBB];
4247 if (!InsertedTrunc) {
4248 BasicBlock::iterator InsertPt = UserBB->getFirstInsertionPt();
4249 assert(InsertPt != UserBB->end());
4250 InsertedTrunc = new TruncInst(I, Src->getType(), "", &*InsertPt);
4251 InsertedInsts.insert(InsertedTrunc);
4254 // Replace a use of the {s|z}ext source with a use of the result.
4263 // Find loads whose uses only use some of the loaded value's bits. Add an "and"
4264 // just after the load if the target can fold this into one extload instruction,
4265 // with the hope of eliminating some of the other later "and" instructions using
4266 // the loaded value. "and"s that are made trivially redundant by the insertion
4267 // of the new "and" are removed by this function, while others (e.g. those whose
4268 // path from the load goes through a phi) are left for isel to potentially
4301 // becomes (after a call to optimizeLoadExt for each load):
4305 // x1' = and x1, 0xff
4309 // x2' = and x2, 0xff
4316 bool CodeGenPrepare::optimizeLoadExt(LoadInst *Load) {
4318 if (!Load->isSimple() ||
4319 !(Load->getType()->isIntegerTy() || Load->getType()->isPointerTy()))
4322 // Skip loads we've already transformed or have no reason to transform.
4323 if (Load->hasOneUse()) {
4324 User *LoadUser = *Load->user_begin();
4325 if (cast<Instruction>(LoadUser)->getParent() == Load->getParent() &&
4326 !dyn_cast<PHINode>(LoadUser))
4330 // Look at all uses of Load, looking through phis, to determine how many bits
4331 // of the loaded value are needed.
4332 SmallVector<Instruction *, 8> WorkList;
4333 SmallPtrSet<Instruction *, 16> Visited;
4334 SmallVector<Instruction *, 8> AndsToMaybeRemove;
4335 for (auto *U : Load->users())
4336 WorkList.push_back(cast<Instruction>(U));
4338 EVT LoadResultVT = TLI->getValueType(*DL, Load->getType());
4339 unsigned BitWidth = LoadResultVT.getSizeInBits();
4340 APInt DemandBits(BitWidth, 0);
4341 APInt WidestAndBits(BitWidth, 0);
4343 while (!WorkList.empty()) {
4344 Instruction *I = WorkList.back();
4345 WorkList.pop_back();
4347 // Break use-def graph loops.
4348 if (!Visited.insert(I).second)
4351 // For a PHI node, push all of its users.
4352 if (auto *Phi = dyn_cast<PHINode>(I)) {
4353 for (auto *U : Phi->users())
4354 WorkList.push_back(cast<Instruction>(U));
4358 switch (I->getOpcode()) {
4359 case llvm::Instruction::And: {
4360 auto *AndC = dyn_cast<ConstantInt>(I->getOperand(1));
4363 APInt AndBits = AndC->getValue();
4364 DemandBits |= AndBits;
4365 // Keep track of the widest and mask we see.
4366 if (AndBits.ugt(WidestAndBits))
4367 WidestAndBits = AndBits;
4368 if (AndBits == WidestAndBits && I->getOperand(0) == Load)
4369 AndsToMaybeRemove.push_back(I);
4373 case llvm::Instruction::Shl: {
4374 auto *ShlC = dyn_cast<ConstantInt>(I->getOperand(1));
4377 uint64_t ShiftAmt = ShlC->getLimitedValue(BitWidth - 1);
4378 auto ShlDemandBits = APInt::getAllOnesValue(BitWidth).lshr(ShiftAmt);
4379 DemandBits |= ShlDemandBits;
4383 case llvm::Instruction::Trunc: {
4384 EVT TruncVT = TLI->getValueType(*DL, I->getType());
4385 unsigned TruncBitWidth = TruncVT.getSizeInBits();
4386 auto TruncBits = APInt::getAllOnesValue(TruncBitWidth).zext(BitWidth);
4387 DemandBits |= TruncBits;
4396 uint32_t ActiveBits = DemandBits.getActiveBits();
4397 // Avoid hoisting (and (load x) 1) since it is unlikely to be folded by the
4398 // target even if isLoadExtLegal says an i1 EXTLOAD is valid. For example,
4399 // for the AArch64 target isLoadExtLegal(ZEXTLOAD, i32, i1) returns true, but
4400 // (and (load x) 1) is not matched as a single instruction, rather as a LDR
4401 // followed by an AND.
4402 // TODO: Look into removing this restriction by fixing backends to either
4403 // return false for isLoadExtLegal for i1 or have them select this pattern to
4404 // a single instruction.
4406 // Also avoid hoisting if we didn't see any ands with the exact DemandBits
4407 // mask, since these are the only ands that will be removed by isel.
4408 if (ActiveBits <= 1 || !APIntOps::isMask(ActiveBits, DemandBits) ||
4409 WidestAndBits != DemandBits)
4412 LLVMContext &Ctx = Load->getType()->getContext();
4413 Type *TruncTy = Type::getIntNTy(Ctx, ActiveBits);
4414 EVT TruncVT = TLI->getValueType(*DL, TruncTy);
4416 // Reject cases that won't be matched as extloads.
4417 if (!LoadResultVT.bitsGT(TruncVT) || !TruncVT.isRound() ||
4418 !TLI->isLoadExtLegal(ISD::ZEXTLOAD, LoadResultVT, TruncVT))
4421 IRBuilder<> Builder(Load->getNextNode());
4422 auto *NewAnd = dyn_cast<Instruction>(
4423 Builder.CreateAnd(Load, ConstantInt::get(Ctx, DemandBits)));
4425 // Replace all uses of load with new and (except for the use of load in the
4427 Load->replaceAllUsesWith(NewAnd);
4428 NewAnd->setOperand(0, Load);
4430 // Remove any and instructions that are now redundant.
4431 for (auto *And : AndsToMaybeRemove)
4432 // Check that the and mask is the same as the one we decided to put on the
4434 if (cast<ConstantInt>(And->getOperand(1))->getValue() == DemandBits) {
4435 And->replaceAllUsesWith(NewAnd);
4436 if (&*CurInstIterator == And)
4437 CurInstIterator = std::next(And->getIterator());
4438 And->eraseFromParent();
4446 /// Check if V (an operand of a select instruction) is an expensive instruction
4447 /// that is only used once.
4448 static bool sinkSelectOperand(const TargetTransformInfo *TTI, Value *V) {
4449 auto *I = dyn_cast<Instruction>(V);
4450 // If it's safe to speculatively execute, then it should not have side
4451 // effects; therefore, it's safe to sink and possibly *not* execute.
4452 return I && I->hasOneUse() && isSafeToSpeculativelyExecute(I) &&
4453 TTI->getUserCost(I) >= TargetTransformInfo::TCC_Expensive;
4456 /// Returns true if a SelectInst should be turned into an explicit branch.
4457 static bool isFormingBranchFromSelectProfitable(const TargetTransformInfo *TTI,
4459 // FIXME: This should use the same heuristics as IfConversion to determine
4460 // whether a select is better represented as a branch. This requires that
4461 // branch probability metadata is preserved for the select, which is not the
4464 CmpInst *Cmp = dyn_cast<CmpInst>(SI->getCondition());
4466 // If a branch is predictable, an out-of-order CPU can avoid blocking on its
4467 // comparison condition. If the compare has more than one use, there's
4468 // probably another cmov or setcc around, so it's not worth emitting a branch.
4469 if (!Cmp || !Cmp->hasOneUse())
4472 Value *CmpOp0 = Cmp->getOperand(0);
4473 Value *CmpOp1 = Cmp->getOperand(1);
4475 // Emit "cmov on compare with a memory operand" as a branch to avoid stalls
4476 // on a load from memory. But if the load is used more than once, do not
4477 // change the select to a branch because the load is probably needed
4478 // regardless of whether the branch is taken or not.
4479 if ((isa<LoadInst>(CmpOp0) && CmpOp0->hasOneUse()) ||
4480 (isa<LoadInst>(CmpOp1) && CmpOp1->hasOneUse()))
4483 // If either operand of the select is expensive and only needed on one side
4484 // of the select, we should form a branch.
4485 if (sinkSelectOperand(TTI, SI->getTrueValue()) ||
4486 sinkSelectOperand(TTI, SI->getFalseValue()))
4493 /// If we have a SelectInst that will likely profit from branch prediction,
4494 /// turn it into a branch.
4495 bool CodeGenPrepare::optimizeSelectInst(SelectInst *SI) {
4496 bool VectorCond = !SI->getCondition()->getType()->isIntegerTy(1);
4498 // Can we convert the 'select' to CF ?
4499 if (DisableSelectToBranch || OptSize || !TLI || VectorCond)
4502 TargetLowering::SelectSupportKind SelectKind;
4504 SelectKind = TargetLowering::VectorMaskSelect;
4505 else if (SI->getType()->isVectorTy())
4506 SelectKind = TargetLowering::ScalarCondVectorVal;
4508 SelectKind = TargetLowering::ScalarValSelect;
4510 // Do we have efficient codegen support for this kind of 'selects' ?
4511 if (TLI->isSelectSupported(SelectKind)) {
4512 // We have efficient codegen support for the select instruction.
4513 // Check if it is profitable to keep this 'select'.
4514 if (!TLI->isPredictableSelectExpensive() ||
4515 !isFormingBranchFromSelectProfitable(TTI, SI))
4521 // Transform a sequence like this:
4523 // %cmp = cmp uge i32 %a, %b
4524 // %sel = select i1 %cmp, i32 %c, i32 %d
4528 // %cmp = cmp uge i32 %a, %b
4529 // br i1 %cmp, label %select.true, label %select.false
4531 // br label %select.end
4533 // br label %select.end
4535 // %sel = phi i32 [ %c, %select.true ], [ %d, %select.false ]
4537 // In addition, we may sink instructions that produce %c or %d from
4538 // the entry block into the destination(s) of the new branch.
4539 // If the true or false blocks do not contain a sunken instruction, that
4540 // block and its branch may be optimized away. In that case, one side of the
4541 // first branch will point directly to select.end, and the corresponding PHI
4542 // predecessor block will be the start block.
4544 // First, we split the block containing the select into 2 blocks.
4545 BasicBlock *StartBlock = SI->getParent();
4546 BasicBlock::iterator SplitPt = ++(BasicBlock::iterator(SI));
4547 BasicBlock *EndBlock = StartBlock->splitBasicBlock(SplitPt, "select.end");
4549 // Delete the unconditional branch that was just created by the split.
4550 StartBlock->getTerminator()->eraseFromParent();
4552 // These are the new basic blocks for the conditional branch.
4553 // At least one will become an actual new basic block.
4554 BasicBlock *TrueBlock = nullptr;
4555 BasicBlock *FalseBlock = nullptr;
4557 // Sink expensive instructions into the conditional blocks to avoid executing
4558 // them speculatively.
4559 if (sinkSelectOperand(TTI, SI->getTrueValue())) {
4560 TrueBlock = BasicBlock::Create(SI->getContext(), "select.true.sink",
4561 EndBlock->getParent(), EndBlock);
4562 auto *TrueBranch = BranchInst::Create(EndBlock, TrueBlock);
4563 auto *TrueInst = cast<Instruction>(SI->getTrueValue());
4564 TrueInst->moveBefore(TrueBranch);
4566 if (sinkSelectOperand(TTI, SI->getFalseValue())) {
4567 FalseBlock = BasicBlock::Create(SI->getContext(), "select.false.sink",
4568 EndBlock->getParent(), EndBlock);
4569 auto *FalseBranch = BranchInst::Create(EndBlock, FalseBlock);
4570 auto *FalseInst = cast<Instruction>(SI->getFalseValue());
4571 FalseInst->moveBefore(FalseBranch);
4574 // If there was nothing to sink, then arbitrarily choose the 'false' side
4575 // for a new input value to the PHI.
4576 if (TrueBlock == FalseBlock) {
4577 assert(TrueBlock == nullptr &&
4578 "Unexpected basic block transform while optimizing select");
4580 FalseBlock = BasicBlock::Create(SI->getContext(), "select.false",
4581 EndBlock->getParent(), EndBlock);
4582 BranchInst::Create(EndBlock, FalseBlock);
4585 // Insert the real conditional branch based on the original condition.
4586 // If we did not create a new block for one of the 'true' or 'false' paths
4587 // of the condition, it means that side of the branch goes to the end block
4588 // directly and the path originates from the start block from the point of
4589 // view of the new PHI.
4590 if (TrueBlock == nullptr) {
4591 BranchInst::Create(EndBlock, FalseBlock, SI->getCondition(), SI);
4592 TrueBlock = StartBlock;
4593 } else if (FalseBlock == nullptr) {
4594 BranchInst::Create(TrueBlock, EndBlock, SI->getCondition(), SI);
4595 FalseBlock = StartBlock;
4597 BranchInst::Create(TrueBlock, FalseBlock, SI->getCondition(), SI);
4600 // The select itself is replaced with a PHI Node.
4601 PHINode *PN = PHINode::Create(SI->getType(), 2, "", &EndBlock->front());
4603 PN->addIncoming(SI->getTrueValue(), TrueBlock);
4604 PN->addIncoming(SI->getFalseValue(), FalseBlock);
4606 SI->replaceAllUsesWith(PN);
4607 SI->eraseFromParent();
4609 // Instruct OptimizeBlock to skip to the next block.
4610 CurInstIterator = StartBlock->end();
4611 ++NumSelectsExpanded;
4615 static bool isBroadcastShuffle(ShuffleVectorInst *SVI) {
4616 SmallVector<int, 16> Mask(SVI->getShuffleMask());
4618 for (unsigned i = 0; i < Mask.size(); ++i) {
4619 if (SplatElem != -1 && Mask[i] != -1 && Mask[i] != SplatElem)
4621 SplatElem = Mask[i];
4627 /// Some targets have expensive vector shifts if the lanes aren't all the same
4628 /// (e.g. x86 only introduced "vpsllvd" and friends with AVX2). In these cases
4629 /// it's often worth sinking a shufflevector splat down to its use so that
4630 /// codegen can spot all lanes are identical.
4631 bool CodeGenPrepare::optimizeShuffleVectorInst(ShuffleVectorInst *SVI) {
4632 BasicBlock *DefBB = SVI->getParent();
4634 // Only do this xform if variable vector shifts are particularly expensive.
4635 if (!TLI || !TLI->isVectorShiftByScalarCheap(SVI->getType()))
4638 // We only expect better codegen by sinking a shuffle if we can recognise a
4640 if (!isBroadcastShuffle(SVI))
4643 // InsertedShuffles - Only insert a shuffle in each block once.
4644 DenseMap<BasicBlock*, Instruction*> InsertedShuffles;
4646 bool MadeChange = false;
4647 for (User *U : SVI->users()) {
4648 Instruction *UI = cast<Instruction>(U);
4650 // Figure out which BB this ext is used in.
4651 BasicBlock *UserBB = UI->getParent();
4652 if (UserBB == DefBB) continue;
4654 // For now only apply this when the splat is used by a shift instruction.
4655 if (!UI->isShift()) continue;
4657 // Everything checks out, sink the shuffle if the user's block doesn't
4658 // already have a copy.
4659 Instruction *&InsertedShuffle = InsertedShuffles[UserBB];
4661 if (!InsertedShuffle) {
4662 BasicBlock::iterator InsertPt = UserBB->getFirstInsertionPt();
4663 assert(InsertPt != UserBB->end());
4665 new ShuffleVectorInst(SVI->getOperand(0), SVI->getOperand(1),
4666 SVI->getOperand(2), "", &*InsertPt);
4669 UI->replaceUsesOfWith(SVI, InsertedShuffle);
4673 // If we removed all uses, nuke the shuffle.
4674 if (SVI->use_empty()) {
4675 SVI->eraseFromParent();
4682 bool CodeGenPrepare::optimizeSwitchInst(SwitchInst *SI) {
4686 Value *Cond = SI->getCondition();
4687 Type *OldType = Cond->getType();
4688 LLVMContext &Context = Cond->getContext();
4689 MVT RegType = TLI->getRegisterType(Context, TLI->getValueType(*DL, OldType));
4690 unsigned RegWidth = RegType.getSizeInBits();
4692 if (RegWidth <= cast<IntegerType>(OldType)->getBitWidth())
4695 // If the register width is greater than the type width, expand the condition
4696 // of the switch instruction and each case constant to the width of the
4697 // register. By widening the type of the switch condition, subsequent
4698 // comparisons (for case comparisons) will not need to be extended to the
4699 // preferred register width, so we will potentially eliminate N-1 extends,
4700 // where N is the number of cases in the switch.
4701 auto *NewType = Type::getIntNTy(Context, RegWidth);
4703 // Zero-extend the switch condition and case constants unless the switch
4704 // condition is a function argument that is already being sign-extended.
4705 // In that case, we can avoid an unnecessary mask/extension by sign-extending
4706 // everything instead.
4707 Instruction::CastOps ExtType = Instruction::ZExt;
4708 if (auto *Arg = dyn_cast<Argument>(Cond))
4709 if (Arg->hasSExtAttr())
4710 ExtType = Instruction::SExt;
4712 auto *ExtInst = CastInst::Create(ExtType, Cond, NewType);
4713 ExtInst->insertBefore(SI);
4714 SI->setCondition(ExtInst);
4715 for (SwitchInst::CaseIt Case : SI->cases()) {
4716 APInt NarrowConst = Case.getCaseValue()->getValue();
4717 APInt WideConst = (ExtType == Instruction::ZExt) ?
4718 NarrowConst.zext(RegWidth) : NarrowConst.sext(RegWidth);
4719 Case.setValue(ConstantInt::get(Context, WideConst));
4726 /// \brief Helper class to promote a scalar operation to a vector one.
4727 /// This class is used to move downward extractelement transition.
4729 /// a = vector_op <2 x i32>
4730 /// b = extractelement <2 x i32> a, i32 0
4735 /// a = vector_op <2 x i32>
4736 /// c = vector_op a (equivalent to scalar_op on the related lane)
4737 /// * d = extractelement <2 x i32> c, i32 0
4739 /// Assuming both extractelement and store can be combine, we get rid of the
4741 class VectorPromoteHelper {
4742 /// DataLayout associated with the current module.
4743 const DataLayout &DL;
4745 /// Used to perform some checks on the legality of vector operations.
4746 const TargetLowering &TLI;
4748 /// Used to estimated the cost of the promoted chain.
4749 const TargetTransformInfo &TTI;
4751 /// The transition being moved downwards.
4752 Instruction *Transition;
4753 /// The sequence of instructions to be promoted.
4754 SmallVector<Instruction *, 4> InstsToBePromoted;
4755 /// Cost of combining a store and an extract.
4756 unsigned StoreExtractCombineCost;
4757 /// Instruction that will be combined with the transition.
4758 Instruction *CombineInst;
4760 /// \brief The instruction that represents the current end of the transition.
4761 /// Since we are faking the promotion until we reach the end of the chain
4762 /// of computation, we need a way to get the current end of the transition.
4763 Instruction *getEndOfTransition() const {
4764 if (InstsToBePromoted.empty())
4766 return InstsToBePromoted.back();
4769 /// \brief Return the index of the original value in the transition.
4770 /// E.g., for "extractelement <2 x i32> c, i32 1" the original value,
4771 /// c, is at index 0.
4772 unsigned getTransitionOriginalValueIdx() const {
4773 assert(isa<ExtractElementInst>(Transition) &&
4774 "Other kind of transitions are not supported yet");
4778 /// \brief Return the index of the index in the transition.
4779 /// E.g., for "extractelement <2 x i32> c, i32 0" the index
4781 unsigned getTransitionIdx() const {
4782 assert(isa<ExtractElementInst>(Transition) &&
4783 "Other kind of transitions are not supported yet");
4787 /// \brief Get the type of the transition.
4788 /// This is the type of the original value.
4789 /// E.g., for "extractelement <2 x i32> c, i32 1" the type of the
4790 /// transition is <2 x i32>.
4791 Type *getTransitionType() const {
4792 return Transition->getOperand(getTransitionOriginalValueIdx())->getType();
4795 /// \brief Promote \p ToBePromoted by moving \p Def downward through.
4796 /// I.e., we have the following sequence:
4797 /// Def = Transition <ty1> a to <ty2>
4798 /// b = ToBePromoted <ty2> Def, ...
4800 /// b = ToBePromoted <ty1> a, ...
4801 /// Def = Transition <ty1> ToBePromoted to <ty2>
4802 void promoteImpl(Instruction *ToBePromoted);
4804 /// \brief Check whether or not it is profitable to promote all the
4805 /// instructions enqueued to be promoted.
4806 bool isProfitableToPromote() {
4807 Value *ValIdx = Transition->getOperand(getTransitionOriginalValueIdx());
4808 unsigned Index = isa<ConstantInt>(ValIdx)
4809 ? cast<ConstantInt>(ValIdx)->getZExtValue()
4811 Type *PromotedType = getTransitionType();
4813 StoreInst *ST = cast<StoreInst>(CombineInst);
4814 unsigned AS = ST->getPointerAddressSpace();
4815 unsigned Align = ST->getAlignment();
4816 // Check if this store is supported.
4817 if (!TLI.allowsMisalignedMemoryAccesses(
4818 TLI.getValueType(DL, ST->getValueOperand()->getType()), AS,
4820 // If this is not supported, there is no way we can combine
4821 // the extract with the store.
4825 // The scalar chain of computation has to pay for the transition
4826 // scalar to vector.
4827 // The vector chain has to account for the combining cost.
4828 uint64_t ScalarCost =
4829 TTI.getVectorInstrCost(Transition->getOpcode(), PromotedType, Index);
4830 uint64_t VectorCost = StoreExtractCombineCost;
4831 for (const auto &Inst : InstsToBePromoted) {
4832 // Compute the cost.
4833 // By construction, all instructions being promoted are arithmetic ones.
4834 // Moreover, one argument is a constant that can be viewed as a splat
4836 Value *Arg0 = Inst->getOperand(0);
4837 bool IsArg0Constant = isa<UndefValue>(Arg0) || isa<ConstantInt>(Arg0) ||
4838 isa<ConstantFP>(Arg0);
4839 TargetTransformInfo::OperandValueKind Arg0OVK =
4840 IsArg0Constant ? TargetTransformInfo::OK_UniformConstantValue
4841 : TargetTransformInfo::OK_AnyValue;
4842 TargetTransformInfo::OperandValueKind Arg1OVK =
4843 !IsArg0Constant ? TargetTransformInfo::OK_UniformConstantValue
4844 : TargetTransformInfo::OK_AnyValue;
4845 ScalarCost += TTI.getArithmeticInstrCost(
4846 Inst->getOpcode(), Inst->getType(), Arg0OVK, Arg1OVK);
4847 VectorCost += TTI.getArithmeticInstrCost(Inst->getOpcode(), PromotedType,
4850 DEBUG(dbgs() << "Estimated cost of computation to be promoted:\nScalar: "
4851 << ScalarCost << "\nVector: " << VectorCost << '\n');
4852 return ScalarCost > VectorCost;
4855 /// \brief Generate a constant vector with \p Val with the same
4856 /// number of elements as the transition.
4857 /// \p UseSplat defines whether or not \p Val should be replicated
4858 /// across the whole vector.
4859 /// In other words, if UseSplat == true, we generate <Val, Val, ..., Val>,
4860 /// otherwise we generate a vector with as many undef as possible:
4861 /// <undef, ..., undef, Val, undef, ..., undef> where \p Val is only
4862 /// used at the index of the extract.
4863 Value *getConstantVector(Constant *Val, bool UseSplat) const {
4864 unsigned ExtractIdx = UINT_MAX;
4866 // If we cannot determine where the constant must be, we have to
4867 // use a splat constant.
4868 Value *ValExtractIdx = Transition->getOperand(getTransitionIdx());
4869 if (ConstantInt *CstVal = dyn_cast<ConstantInt>(ValExtractIdx))
4870 ExtractIdx = CstVal->getSExtValue();
4875 unsigned End = getTransitionType()->getVectorNumElements();
4877 return ConstantVector::getSplat(End, Val);
4879 SmallVector<Constant *, 4> ConstVec;
4880 UndefValue *UndefVal = UndefValue::get(Val->getType());
4881 for (unsigned Idx = 0; Idx != End; ++Idx) {
4882 if (Idx == ExtractIdx)
4883 ConstVec.push_back(Val);
4885 ConstVec.push_back(UndefVal);
4887 return ConstantVector::get(ConstVec);
4890 /// \brief Check if promoting to a vector type an operand at \p OperandIdx
4891 /// in \p Use can trigger undefined behavior.
4892 static bool canCauseUndefinedBehavior(const Instruction *Use,
4893 unsigned OperandIdx) {
4894 // This is not safe to introduce undef when the operand is on
4895 // the right hand side of a division-like instruction.
4896 if (OperandIdx != 1)
4898 switch (Use->getOpcode()) {
4901 case Instruction::SDiv:
4902 case Instruction::UDiv:
4903 case Instruction::SRem:
4904 case Instruction::URem:
4906 case Instruction::FDiv:
4907 case Instruction::FRem:
4908 return !Use->hasNoNaNs();
4910 llvm_unreachable(nullptr);
4914 VectorPromoteHelper(const DataLayout &DL, const TargetLowering &TLI,
4915 const TargetTransformInfo &TTI, Instruction *Transition,
4916 unsigned CombineCost)
4917 : DL(DL), TLI(TLI), TTI(TTI), Transition(Transition),
4918 StoreExtractCombineCost(CombineCost), CombineInst(nullptr) {
4919 assert(Transition && "Do not know how to promote null");
4922 /// \brief Check if we can promote \p ToBePromoted to \p Type.
4923 bool canPromote(const Instruction *ToBePromoted) const {
4924 // We could support CastInst too.
4925 return isa<BinaryOperator>(ToBePromoted);
4928 /// \brief Check if it is profitable to promote \p ToBePromoted
4929 /// by moving downward the transition through.
4930 bool shouldPromote(const Instruction *ToBePromoted) const {
4931 // Promote only if all the operands can be statically expanded.
4932 // Indeed, we do not want to introduce any new kind of transitions.
4933 for (const Use &U : ToBePromoted->operands()) {
4934 const Value *Val = U.get();
4935 if (Val == getEndOfTransition()) {
4936 // If the use is a division and the transition is on the rhs,
4937 // we cannot promote the operation, otherwise we may create a
4938 // division by zero.
4939 if (canCauseUndefinedBehavior(ToBePromoted, U.getOperandNo()))
4943 if (!isa<ConstantInt>(Val) && !isa<UndefValue>(Val) &&
4944 !isa<ConstantFP>(Val))
4947 // Check that the resulting operation is legal.
4948 int ISDOpcode = TLI.InstructionOpcodeToISD(ToBePromoted->getOpcode());
4951 return StressStoreExtract ||
4952 TLI.isOperationLegalOrCustom(
4953 ISDOpcode, TLI.getValueType(DL, getTransitionType(), true));
4956 /// \brief Check whether or not \p Use can be combined
4957 /// with the transition.
4958 /// I.e., is it possible to do Use(Transition) => AnotherUse?
4959 bool canCombine(const Instruction *Use) { return isa<StoreInst>(Use); }
4961 /// \brief Record \p ToBePromoted as part of the chain to be promoted.
4962 void enqueueForPromotion(Instruction *ToBePromoted) {
4963 InstsToBePromoted.push_back(ToBePromoted);
4966 /// \brief Set the instruction that will be combined with the transition.
4967 void recordCombineInstruction(Instruction *ToBeCombined) {
4968 assert(canCombine(ToBeCombined) && "Unsupported instruction to combine");
4969 CombineInst = ToBeCombined;
4972 /// \brief Promote all the instructions enqueued for promotion if it is
4974 /// \return True if the promotion happened, false otherwise.
4976 // Check if there is something to promote.
4977 // Right now, if we do not have anything to combine with,
4978 // we assume the promotion is not profitable.
4979 if (InstsToBePromoted.empty() || !CombineInst)
4983 if (!StressStoreExtract && !isProfitableToPromote())
4987 for (auto &ToBePromoted : InstsToBePromoted)
4988 promoteImpl(ToBePromoted);
4989 InstsToBePromoted.clear();
4993 } // End of anonymous namespace.
4995 void VectorPromoteHelper::promoteImpl(Instruction *ToBePromoted) {
4996 // At this point, we know that all the operands of ToBePromoted but Def
4997 // can be statically promoted.
4998 // For Def, we need to use its parameter in ToBePromoted:
4999 // b = ToBePromoted ty1 a
5000 // Def = Transition ty1 b to ty2
5001 // Move the transition down.
5002 // 1. Replace all uses of the promoted operation by the transition.
5003 // = ... b => = ... Def.
5004 assert(ToBePromoted->getType() == Transition->getType() &&
5005 "The type of the result of the transition does not match "
5007 ToBePromoted->replaceAllUsesWith(Transition);
5008 // 2. Update the type of the uses.
5009 // b = ToBePromoted ty2 Def => b = ToBePromoted ty1 Def.
5010 Type *TransitionTy = getTransitionType();
5011 ToBePromoted->mutateType(TransitionTy);
5012 // 3. Update all the operands of the promoted operation with promoted
5014 // b = ToBePromoted ty1 Def => b = ToBePromoted ty1 a.
5015 for (Use &U : ToBePromoted->operands()) {
5016 Value *Val = U.get();
5017 Value *NewVal = nullptr;
5018 if (Val == Transition)
5019 NewVal = Transition->getOperand(getTransitionOriginalValueIdx());
5020 else if (isa<UndefValue>(Val) || isa<ConstantInt>(Val) ||
5021 isa<ConstantFP>(Val)) {
5022 // Use a splat constant if it is not safe to use undef.
5023 NewVal = getConstantVector(
5024 cast<Constant>(Val),
5025 isa<UndefValue>(Val) ||
5026 canCauseUndefinedBehavior(ToBePromoted, U.getOperandNo()));
5028 llvm_unreachable("Did you modified shouldPromote and forgot to update "
5030 ToBePromoted->setOperand(U.getOperandNo(), NewVal);
5032 Transition->removeFromParent();
5033 Transition->insertAfter(ToBePromoted);
5034 Transition->setOperand(getTransitionOriginalValueIdx(), ToBePromoted);
5037 /// Some targets can do store(extractelement) with one instruction.
5038 /// Try to push the extractelement towards the stores when the target
5039 /// has this feature and this is profitable.
5040 bool CodeGenPrepare::optimizeExtractElementInst(Instruction *Inst) {
5041 unsigned CombineCost = UINT_MAX;
5042 if (DisableStoreExtract || !TLI ||
5043 (!StressStoreExtract &&
5044 !TLI->canCombineStoreAndExtract(Inst->getOperand(0)->getType(),
5045 Inst->getOperand(1), CombineCost)))
5048 // At this point we know that Inst is a vector to scalar transition.
5049 // Try to move it down the def-use chain, until:
5050 // - We can combine the transition with its single use
5051 // => we got rid of the transition.
5052 // - We escape the current basic block
5053 // => we would need to check that we are moving it at a cheaper place and
5054 // we do not do that for now.
5055 BasicBlock *Parent = Inst->getParent();
5056 DEBUG(dbgs() << "Found an interesting transition: " << *Inst << '\n');
5057 VectorPromoteHelper VPH(*DL, *TLI, *TTI, Inst, CombineCost);
5058 // If the transition has more than one use, assume this is not going to be
5060 while (Inst->hasOneUse()) {
5061 Instruction *ToBePromoted = cast<Instruction>(*Inst->user_begin());
5062 DEBUG(dbgs() << "Use: " << *ToBePromoted << '\n');
5064 if (ToBePromoted->getParent() != Parent) {
5065 DEBUG(dbgs() << "Instruction to promote is in a different block ("
5066 << ToBePromoted->getParent()->getName()
5067 << ") than the transition (" << Parent->getName() << ").\n");
5071 if (VPH.canCombine(ToBePromoted)) {
5072 DEBUG(dbgs() << "Assume " << *Inst << '\n'
5073 << "will be combined with: " << *ToBePromoted << '\n');
5074 VPH.recordCombineInstruction(ToBePromoted);
5075 bool Changed = VPH.promote();
5076 NumStoreExtractExposed += Changed;
5080 DEBUG(dbgs() << "Try promoting.\n");
5081 if (!VPH.canPromote(ToBePromoted) || !VPH.shouldPromote(ToBePromoted))
5084 DEBUG(dbgs() << "Promoting is possible... Enqueue for promotion!\n");
5086 VPH.enqueueForPromotion(ToBePromoted);
5087 Inst = ToBePromoted;
5092 bool CodeGenPrepare::optimizeInst(Instruction *I, bool& ModifiedDT) {
5093 // Bail out if we inserted the instruction to prevent optimizations from
5094 // stepping on each other's toes.
5095 if (InsertedInsts.count(I))
5098 if (PHINode *P = dyn_cast<PHINode>(I)) {
5099 // It is possible for very late stage optimizations (such as SimplifyCFG)
5100 // to introduce PHI nodes too late to be cleaned up. If we detect such a
5101 // trivial PHI, go ahead and zap it here.
5102 if (Value *V = SimplifyInstruction(P, *DL, TLInfo, nullptr)) {
5103 P->replaceAllUsesWith(V);
5104 P->eraseFromParent();
5111 if (CastInst *CI = dyn_cast<CastInst>(I)) {
5112 // If the source of the cast is a constant, then this should have
5113 // already been constant folded. The only reason NOT to constant fold
5114 // it is if something (e.g. LSR) was careful to place the constant
5115 // evaluation in a block other than then one that uses it (e.g. to hoist
5116 // the address of globals out of a loop). If this is the case, we don't
5117 // want to forward-subst the cast.
5118 if (isa<Constant>(CI->getOperand(0)))
5121 if (TLI && OptimizeNoopCopyExpression(CI, *TLI, *DL))
5124 if (isa<ZExtInst>(I) || isa<SExtInst>(I)) {
5125 /// Sink a zext or sext into its user blocks if the target type doesn't
5126 /// fit in one register
5128 TLI->getTypeAction(CI->getContext(),
5129 TLI->getValueType(*DL, CI->getType())) ==
5130 TargetLowering::TypeExpandInteger) {
5131 return SinkCast(CI);
5133 bool MadeChange = moveExtToFormExtLoad(I);
5134 return MadeChange | optimizeExtUses(I);
5140 if (CmpInst *CI = dyn_cast<CmpInst>(I))
5141 if (!TLI || !TLI->hasMultipleConditionRegisters())
5142 return OptimizeCmpExpression(CI);
5144 if (LoadInst *LI = dyn_cast<LoadInst>(I)) {
5145 stripInvariantGroupMetadata(*LI);
5147 bool Modified = optimizeLoadExt(LI);
5148 unsigned AS = LI->getPointerAddressSpace();
5149 Modified |= optimizeMemoryInst(I, I->getOperand(0), LI->getType(), AS);
5155 if (StoreInst *SI = dyn_cast<StoreInst>(I)) {
5156 stripInvariantGroupMetadata(*SI);
5158 unsigned AS = SI->getPointerAddressSpace();
5159 return optimizeMemoryInst(I, SI->getOperand(1),
5160 SI->getOperand(0)->getType(), AS);
5165 BinaryOperator *BinOp = dyn_cast<BinaryOperator>(I);
5167 if (BinOp && (BinOp->getOpcode() == Instruction::AShr ||
5168 BinOp->getOpcode() == Instruction::LShr)) {
5169 ConstantInt *CI = dyn_cast<ConstantInt>(BinOp->getOperand(1));
5170 if (TLI && CI && TLI->hasExtractBitsInsn())
5171 return OptimizeExtractBits(BinOp, CI, *TLI, *DL);
5176 if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(I)) {
5177 if (GEPI->hasAllZeroIndices()) {
5178 /// The GEP operand must be a pointer, so must its result -> BitCast
5179 Instruction *NC = new BitCastInst(GEPI->getOperand(0), GEPI->getType(),
5180 GEPI->getName(), GEPI);
5181 GEPI->replaceAllUsesWith(NC);
5182 GEPI->eraseFromParent();
5184 optimizeInst(NC, ModifiedDT);
5190 if (CallInst *CI = dyn_cast<CallInst>(I))
5191 return optimizeCallInst(CI, ModifiedDT);
5193 if (SelectInst *SI = dyn_cast<SelectInst>(I))
5194 return optimizeSelectInst(SI);
5196 if (ShuffleVectorInst *SVI = dyn_cast<ShuffleVectorInst>(I))
5197 return optimizeShuffleVectorInst(SVI);
5199 if (auto *Switch = dyn_cast<SwitchInst>(I))
5200 return optimizeSwitchInst(Switch);
5202 if (isa<ExtractElementInst>(I))
5203 return optimizeExtractElementInst(I);
5208 // In this pass we look for GEP and cast instructions that are used
5209 // across basic blocks and rewrite them to improve basic-block-at-a-time
5211 bool CodeGenPrepare::optimizeBlock(BasicBlock &BB, bool& ModifiedDT) {
5213 bool MadeChange = false;
5215 CurInstIterator = BB.begin();
5216 while (CurInstIterator != BB.end()) {
5217 MadeChange |= optimizeInst(&*CurInstIterator++, ModifiedDT);
5221 MadeChange |= dupRetToEnableTailCallOpts(&BB);
5226 // llvm.dbg.value is far away from the value then iSel may not be able
5227 // handle it properly. iSel will drop llvm.dbg.value if it can not
5228 // find a node corresponding to the value.
5229 bool CodeGenPrepare::placeDbgValues(Function &F) {
5230 bool MadeChange = false;
5231 for (BasicBlock &BB : F) {
5232 Instruction *PrevNonDbgInst = nullptr;
5233 for (BasicBlock::iterator BI = BB.begin(), BE = BB.end(); BI != BE;) {
5234 Instruction *Insn = &*BI++;
5235 DbgValueInst *DVI = dyn_cast<DbgValueInst>(Insn);
5236 // Leave dbg.values that refer to an alloca alone. These
5237 // instrinsics describe the address of a variable (= the alloca)
5238 // being taken. They should not be moved next to the alloca
5239 // (and to the beginning of the scope), but rather stay close to
5240 // where said address is used.
5241 if (!DVI || (DVI->getValue() && isa<AllocaInst>(DVI->getValue()))) {
5242 PrevNonDbgInst = Insn;
5246 Instruction *VI = dyn_cast_or_null<Instruction>(DVI->getValue());
5247 if (VI && VI != PrevNonDbgInst && !VI->isTerminator()) {
5248 DEBUG(dbgs() << "Moving Debug Value before :\n" << *DVI << ' ' << *VI);
5249 DVI->removeFromParent();
5250 if (isa<PHINode>(VI))
5251 DVI->insertBefore(&*VI->getParent()->getFirstInsertionPt());
5253 DVI->insertAfter(VI);
5262 // If there is a sequence that branches based on comparing a single bit
5263 // against zero that can be combined into a single instruction, and the
5264 // target supports folding these into a single instruction, sink the
5265 // mask and compare into the branch uses. Do this before OptimizeBlock ->
5266 // OptimizeInst -> OptimizeCmpExpression, which perturbs the pattern being
5268 bool CodeGenPrepare::sinkAndCmp(Function &F) {
5269 if (!EnableAndCmpSinking)
5271 if (!TLI || !TLI->isMaskAndBranchFoldingLegal())
5273 bool MadeChange = false;
5274 for (Function::iterator I = F.begin(), E = F.end(); I != E; ) {
5275 BasicBlock *BB = &*I++;
5277 // Does this BB end with the following?
5278 // %andVal = and %val, #single-bit-set
5279 // %icmpVal = icmp %andResult, 0
5280 // br i1 %cmpVal label %dest1, label %dest2"
5281 BranchInst *Brcc = dyn_cast<BranchInst>(BB->getTerminator());
5282 if (!Brcc || !Brcc->isConditional())
5284 ICmpInst *Cmp = dyn_cast<ICmpInst>(Brcc->getOperand(0));
5285 if (!Cmp || Cmp->getParent() != BB)
5287 ConstantInt *Zero = dyn_cast<ConstantInt>(Cmp->getOperand(1));
5288 if (!Zero || !Zero->isZero())
5290 Instruction *And = dyn_cast<Instruction>(Cmp->getOperand(0));
5291 if (!And || And->getOpcode() != Instruction::And || And->getParent() != BB)
5293 ConstantInt* Mask = dyn_cast<ConstantInt>(And->getOperand(1));
5294 if (!Mask || !Mask->getUniqueInteger().isPowerOf2())
5296 DEBUG(dbgs() << "found and; icmp ?,0; brcc\n"); DEBUG(BB->dump());
5298 // Push the "and; icmp" for any users that are conditional branches.
5299 // Since there can only be one branch use per BB, we don't need to keep
5300 // track of which BBs we insert into.
5301 for (Value::use_iterator UI = Cmp->use_begin(), E = Cmp->use_end();
5305 BranchInst *BrccUser = dyn_cast<BranchInst>(*UI);
5307 if (!BrccUser || !BrccUser->isConditional())
5309 BasicBlock *UserBB = BrccUser->getParent();
5310 if (UserBB == BB) continue;
5311 DEBUG(dbgs() << "found Brcc use\n");
5313 // Sink the "and; icmp" to use.
5315 BinaryOperator *NewAnd =
5316 BinaryOperator::CreateAnd(And->getOperand(0), And->getOperand(1), "",
5319 CmpInst::Create(Cmp->getOpcode(), Cmp->getPredicate(), NewAnd, Zero,
5323 DEBUG(BrccUser->getParent()->dump());
5329 /// \brief Retrieve the probabilities of a conditional branch. Returns true on
5330 /// success, or returns false if no or invalid metadata was found.
5331 static bool extractBranchMetadata(BranchInst *BI,
5332 uint64_t &ProbTrue, uint64_t &ProbFalse) {
5333 assert(BI->isConditional() &&
5334 "Looking for probabilities on unconditional branch?");
5335 auto *ProfileData = BI->getMetadata(LLVMContext::MD_prof);
5336 if (!ProfileData || ProfileData->getNumOperands() != 3)
5339 const auto *CITrue =
5340 mdconst::dyn_extract<ConstantInt>(ProfileData->getOperand(1));
5341 const auto *CIFalse =
5342 mdconst::dyn_extract<ConstantInt>(ProfileData->getOperand(2));
5343 if (!CITrue || !CIFalse)
5346 ProbTrue = CITrue->getValue().getZExtValue();
5347 ProbFalse = CIFalse->getValue().getZExtValue();
5352 /// \brief Scale down both weights to fit into uint32_t.
5353 static void scaleWeights(uint64_t &NewTrue, uint64_t &NewFalse) {
5354 uint64_t NewMax = (NewTrue > NewFalse) ? NewTrue : NewFalse;
5355 uint32_t Scale = (NewMax / UINT32_MAX) + 1;
5356 NewTrue = NewTrue / Scale;
5357 NewFalse = NewFalse / Scale;
5360 /// \brief Some targets prefer to split a conditional branch like:
5362 /// %0 = icmp ne i32 %a, 0
5363 /// %1 = icmp ne i32 %b, 0
5364 /// %or.cond = or i1 %0, %1
5365 /// br i1 %or.cond, label %TrueBB, label %FalseBB
5367 /// into multiple branch instructions like:
5370 /// %0 = icmp ne i32 %a, 0
5371 /// br i1 %0, label %TrueBB, label %bb2
5373 /// %1 = icmp ne i32 %b, 0
5374 /// br i1 %1, label %TrueBB, label %FalseBB
5376 /// This usually allows instruction selection to do even further optimizations
5377 /// and combine the compare with the branch instruction. Currently this is
5378 /// applied for targets which have "cheap" jump instructions.
5380 /// FIXME: Remove the (equivalent?) implementation in SelectionDAG.
5382 bool CodeGenPrepare::splitBranchCondition(Function &F) {
5383 if (!TM || !TM->Options.EnableFastISel || !TLI || TLI->isJumpExpensive())
5386 bool MadeChange = false;
5387 for (auto &BB : F) {
5388 // Does this BB end with the following?
5389 // %cond1 = icmp|fcmp|binary instruction ...
5390 // %cond2 = icmp|fcmp|binary instruction ...
5391 // %cond.or = or|and i1 %cond1, cond2
5392 // br i1 %cond.or label %dest1, label %dest2"
5393 BinaryOperator *LogicOp;
5394 BasicBlock *TBB, *FBB;
5395 if (!match(BB.getTerminator(), m_Br(m_OneUse(m_BinOp(LogicOp)), TBB, FBB)))
5398 auto *Br1 = cast<BranchInst>(BB.getTerminator());
5399 if (Br1->getMetadata(LLVMContext::MD_unpredictable))
5403 Value *Cond1, *Cond2;
5404 if (match(LogicOp, m_And(m_OneUse(m_Value(Cond1)),
5405 m_OneUse(m_Value(Cond2)))))
5406 Opc = Instruction::And;
5407 else if (match(LogicOp, m_Or(m_OneUse(m_Value(Cond1)),
5408 m_OneUse(m_Value(Cond2)))))
5409 Opc = Instruction::Or;
5413 if (!match(Cond1, m_CombineOr(m_Cmp(), m_BinOp())) ||
5414 !match(Cond2, m_CombineOr(m_Cmp(), m_BinOp())) )
5417 DEBUG(dbgs() << "Before branch condition splitting\n"; BB.dump());
5420 auto *InsertBefore = std::next(Function::iterator(BB))
5421 .getNodePtrUnchecked();
5422 auto TmpBB = BasicBlock::Create(BB.getContext(),
5423 BB.getName() + ".cond.split",
5424 BB.getParent(), InsertBefore);
5426 // Update original basic block by using the first condition directly by the
5427 // branch instruction and removing the no longer needed and/or instruction.
5428 Br1->setCondition(Cond1);
5429 LogicOp->eraseFromParent();
5431 // Depending on the conditon we have to either replace the true or the false
5432 // successor of the original branch instruction.
5433 if (Opc == Instruction::And)
5434 Br1->setSuccessor(0, TmpBB);
5436 Br1->setSuccessor(1, TmpBB);
5438 // Fill in the new basic block.
5439 auto *Br2 = IRBuilder<>(TmpBB).CreateCondBr(Cond2, TBB, FBB);
5440 if (auto *I = dyn_cast<Instruction>(Cond2)) {
5441 I->removeFromParent();
5442 I->insertBefore(Br2);
5445 // Update PHI nodes in both successors. The original BB needs to be
5446 // replaced in one succesor's PHI nodes, because the branch comes now from
5447 // the newly generated BB (NewBB). In the other successor we need to add one
5448 // incoming edge to the PHI nodes, because both branch instructions target
5449 // now the same successor. Depending on the original branch condition
5450 // (and/or) we have to swap the successors (TrueDest, FalseDest), so that
5451 // we perfrom the correct update for the PHI nodes.
5452 // This doesn't change the successor order of the just created branch
5453 // instruction (or any other instruction).
5454 if (Opc == Instruction::Or)
5455 std::swap(TBB, FBB);
5457 // Replace the old BB with the new BB.
5458 for (auto &I : *TBB) {
5459 PHINode *PN = dyn_cast<PHINode>(&I);
5463 while ((i = PN->getBasicBlockIndex(&BB)) >= 0)
5464 PN->setIncomingBlock(i, TmpBB);
5467 // Add another incoming edge form the new BB.
5468 for (auto &I : *FBB) {
5469 PHINode *PN = dyn_cast<PHINode>(&I);
5472 auto *Val = PN->getIncomingValueForBlock(&BB);
5473 PN->addIncoming(Val, TmpBB);
5476 // Update the branch weights (from SelectionDAGBuilder::
5477 // FindMergedConditions).
5478 if (Opc == Instruction::Or) {
5479 // Codegen X | Y as:
5488 // We have flexibility in setting Prob for BB1 and Prob for NewBB.
5489 // The requirement is that
5490 // TrueProb for BB1 + (FalseProb for BB1 * TrueProb for TmpBB)
5491 // = TrueProb for orignal BB.
5492 // Assuming the orignal weights are A and B, one choice is to set BB1's
5493 // weights to A and A+2B, and set TmpBB's weights to A and 2B. This choice
5495 // TrueProb for BB1 == FalseProb for BB1 * TrueProb for TmpBB.
5496 // Another choice is to assume TrueProb for BB1 equals to TrueProb for
5497 // TmpBB, but the math is more complicated.
5498 uint64_t TrueWeight, FalseWeight;
5499 if (extractBranchMetadata(Br1, TrueWeight, FalseWeight)) {
5500 uint64_t NewTrueWeight = TrueWeight;
5501 uint64_t NewFalseWeight = TrueWeight + 2 * FalseWeight;
5502 scaleWeights(NewTrueWeight, NewFalseWeight);
5503 Br1->setMetadata(LLVMContext::MD_prof, MDBuilder(Br1->getContext())
5504 .createBranchWeights(TrueWeight, FalseWeight));
5506 NewTrueWeight = TrueWeight;
5507 NewFalseWeight = 2 * FalseWeight;
5508 scaleWeights(NewTrueWeight, NewFalseWeight);
5509 Br2->setMetadata(LLVMContext::MD_prof, MDBuilder(Br2->getContext())
5510 .createBranchWeights(TrueWeight, FalseWeight));
5513 // Codegen X & Y as:
5521 // This requires creation of TmpBB after CurBB.
5523 // We have flexibility in setting Prob for BB1 and Prob for TmpBB.
5524 // The requirement is that
5525 // FalseProb for BB1 + (TrueProb for BB1 * FalseProb for TmpBB)
5526 // = FalseProb for orignal BB.
5527 // Assuming the orignal weights are A and B, one choice is to set BB1's
5528 // weights to 2A+B and B, and set TmpBB's weights to 2A and B. This choice
5530 // FalseProb for BB1 == TrueProb for BB1 * FalseProb for TmpBB.
5531 uint64_t TrueWeight, FalseWeight;
5532 if (extractBranchMetadata(Br1, TrueWeight, FalseWeight)) {
5533 uint64_t NewTrueWeight = 2 * TrueWeight + FalseWeight;
5534 uint64_t NewFalseWeight = FalseWeight;
5535 scaleWeights(NewTrueWeight, NewFalseWeight);
5536 Br1->setMetadata(LLVMContext::MD_prof, MDBuilder(Br1->getContext())
5537 .createBranchWeights(TrueWeight, FalseWeight));
5539 NewTrueWeight = 2 * TrueWeight;
5540 NewFalseWeight = FalseWeight;
5541 scaleWeights(NewTrueWeight, NewFalseWeight);
5542 Br2->setMetadata(LLVMContext::MD_prof, MDBuilder(Br2->getContext())
5543 .createBranchWeights(TrueWeight, FalseWeight));
5547 // Note: No point in getting fancy here, since the DT info is never
5548 // available to CodeGenPrepare.
5553 DEBUG(dbgs() << "After branch condition splitting\n"; BB.dump();
5559 void CodeGenPrepare::stripInvariantGroupMetadata(Instruction &I) {
5560 if (auto *InvariantMD = I.getMetadata(LLVMContext::MD_invariant_group))
5561 I.dropUnknownNonDebugMetadata(InvariantMD->getMetadataID());