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/IR/CallSite.h"
24 #include "llvm/IR/Constants.h"
25 #include "llvm/IR/DataLayout.h"
26 #include "llvm/IR/DerivedTypes.h"
27 #include "llvm/IR/Dominators.h"
28 #include "llvm/IR/Function.h"
29 #include "llvm/IR/GetElementPtrTypeIterator.h"
30 #include "llvm/IR/IRBuilder.h"
31 #include "llvm/IR/InlineAsm.h"
32 #include "llvm/IR/Instructions.h"
33 #include "llvm/IR/IntrinsicInst.h"
34 #include "llvm/IR/MDBuilder.h"
35 #include "llvm/IR/PatternMatch.h"
36 #include "llvm/IR/Statepoint.h"
37 #include "llvm/IR/ValueHandle.h"
38 #include "llvm/IR/ValueMap.h"
39 #include "llvm/Pass.h"
40 #include "llvm/Support/CommandLine.h"
41 #include "llvm/Support/Debug.h"
42 #include "llvm/Support/raw_ostream.h"
43 #include "llvm/Target/TargetLowering.h"
44 #include "llvm/Target/TargetSubtargetInfo.h"
45 #include "llvm/Transforms/Utils/BasicBlockUtils.h"
46 #include "llvm/Transforms/Utils/BuildLibCalls.h"
47 #include "llvm/Transforms/Utils/BypassSlowDivision.h"
48 #include "llvm/Transforms/Utils/Local.h"
49 #include "llvm/Transforms/Utils/SimplifyLibCalls.h"
51 using namespace llvm::PatternMatch;
53 #define DEBUG_TYPE "codegenprepare"
55 STATISTIC(NumBlocksElim, "Number of blocks eliminated");
56 STATISTIC(NumPHIsElim, "Number of trivial PHIs eliminated");
57 STATISTIC(NumGEPsElim, "Number of GEPs converted to casts");
58 STATISTIC(NumCmpUses, "Number of uses of Cmp expressions replaced with uses of "
60 STATISTIC(NumCastUses, "Number of uses of Cast expressions replaced with uses "
62 STATISTIC(NumMemoryInsts, "Number of memory instructions whose address "
63 "computations were sunk");
64 STATISTIC(NumExtsMoved, "Number of [s|z]ext instructions combined with loads");
65 STATISTIC(NumExtUses, "Number of uses of [s|z]ext instructions optimized");
66 STATISTIC(NumRetsDup, "Number of return instructions duplicated");
67 STATISTIC(NumDbgValueMoved, "Number of debug value instructions moved");
68 STATISTIC(NumSelectsExpanded, "Number of selects turned into branches");
69 STATISTIC(NumAndCmpsMoved, "Number of and/cmp's pushed into branches");
70 STATISTIC(NumStoreExtractExposed, "Number of store(extractelement) exposed");
72 static cl::opt<bool> DisableBranchOpts(
73 "disable-cgp-branch-opts", cl::Hidden, cl::init(false),
74 cl::desc("Disable branch optimizations in CodeGenPrepare"));
77 DisableGCOpts("disable-cgp-gc-opts", cl::Hidden, cl::init(false),
78 cl::desc("Disable GC optimizations in CodeGenPrepare"));
80 static cl::opt<bool> DisableSelectToBranch(
81 "disable-cgp-select2branch", cl::Hidden, cl::init(false),
82 cl::desc("Disable select to branch conversion."));
84 static cl::opt<bool> AddrSinkUsingGEPs(
85 "addr-sink-using-gep", cl::Hidden, cl::init(false),
86 cl::desc("Address sinking in CGP using GEPs."));
88 static cl::opt<bool> EnableAndCmpSinking(
89 "enable-andcmp-sinking", cl::Hidden, cl::init(true),
90 cl::desc("Enable sinkinig and/cmp into branches."));
92 static cl::opt<bool> DisableStoreExtract(
93 "disable-cgp-store-extract", cl::Hidden, cl::init(false),
94 cl::desc("Disable store(extract) optimizations in CodeGenPrepare"));
96 static cl::opt<bool> StressStoreExtract(
97 "stress-cgp-store-extract", cl::Hidden, cl::init(false),
98 cl::desc("Stress test store(extract) optimizations in CodeGenPrepare"));
100 static cl::opt<bool> DisableExtLdPromotion(
101 "disable-cgp-ext-ld-promotion", cl::Hidden, cl::init(false),
102 cl::desc("Disable ext(promotable(ld)) -> promoted(ext(ld)) optimization in "
105 static cl::opt<bool> StressExtLdPromotion(
106 "stress-cgp-ext-ld-promotion", cl::Hidden, cl::init(false),
107 cl::desc("Stress test ext(promotable(ld)) -> promoted(ext(ld)) "
108 "optimization in CodeGenPrepare"));
111 typedef SmallPtrSet<Instruction *, 16> SetOfInstrs;
115 TypeIsSExt(Type *Ty, bool IsSExt) : Ty(Ty), IsSExt(IsSExt) {}
117 typedef DenseMap<Instruction *, TypeIsSExt> InstrToOrigTy;
118 class TypePromotionTransaction;
120 class CodeGenPrepare : public FunctionPass {
121 /// TLI - Keep a pointer of a TargetLowering to consult for determining
122 /// transformation profitability.
123 const TargetMachine *TM;
124 const TargetLowering *TLI;
125 const TargetTransformInfo *TTI;
126 const TargetLibraryInfo *TLInfo;
129 /// CurInstIterator - As we scan instructions optimizing them, this is the
130 /// next instruction to optimize. Xforms that can invalidate this should
132 BasicBlock::iterator CurInstIterator;
134 /// Keeps track of non-local addresses that have been sunk into a block.
135 /// This allows us to avoid inserting duplicate code for blocks with
136 /// multiple load/stores of the same address.
137 ValueMap<Value*, Value*> SunkAddrs;
139 /// Keeps track of all truncates inserted for the current function.
140 SetOfInstrs InsertedTruncsSet;
141 /// Keeps track of the type of the related instruction before their
142 /// promotion for the current function.
143 InstrToOrigTy PromotedInsts;
145 /// ModifiedDT - If CFG is modified in anyway, dominator tree may need to
149 /// OptSize - True if optimizing for size.
153 static char ID; // Pass identification, replacement for typeid
154 explicit CodeGenPrepare(const TargetMachine *TM = nullptr)
155 : FunctionPass(ID), TM(TM), TLI(nullptr), TTI(nullptr) {
156 initializeCodeGenPreparePass(*PassRegistry::getPassRegistry());
158 bool runOnFunction(Function &F) override;
160 const char *getPassName() const override { return "CodeGen Prepare"; }
162 void getAnalysisUsage(AnalysisUsage &AU) const override {
163 AU.addPreserved<DominatorTreeWrapperPass>();
164 AU.addRequired<TargetLibraryInfoWrapperPass>();
165 AU.addRequired<TargetTransformInfo>();
169 bool EliminateFallThrough(Function &F);
170 bool EliminateMostlyEmptyBlocks(Function &F);
171 bool CanMergeBlocks(const BasicBlock *BB, const BasicBlock *DestBB) const;
172 void EliminateMostlyEmptyBlock(BasicBlock *BB);
173 bool OptimizeBlock(BasicBlock &BB, bool& ModifiedDT);
174 bool OptimizeInst(Instruction *I, bool& ModifiedDT);
175 bool OptimizeMemoryInst(Instruction *I, Value *Addr, Type *AccessTy);
176 bool OptimizeInlineAsmInst(CallInst *CS);
177 bool OptimizeCallInst(CallInst *CI, bool& ModifiedDT);
178 bool MoveExtToFormExtLoad(Instruction *&I);
179 bool OptimizeExtUses(Instruction *I);
180 bool OptimizeSelectInst(SelectInst *SI);
181 bool OptimizeShuffleVectorInst(ShuffleVectorInst *SI);
182 bool OptimizeExtractElementInst(Instruction *Inst);
183 bool DupRetToEnableTailCallOpts(BasicBlock *BB);
184 bool PlaceDbgValues(Function &F);
185 bool sinkAndCmp(Function &F);
186 bool ExtLdPromotion(TypePromotionTransaction &TPT, LoadInst *&LI,
188 const SmallVectorImpl<Instruction *> &Exts,
189 unsigned CreatedInst);
190 bool splitBranchCondition(Function &F);
191 bool simplifyOffsetableRelocate(Instruction &I);
195 char CodeGenPrepare::ID = 0;
196 INITIALIZE_TM_PASS(CodeGenPrepare, "codegenprepare",
197 "Optimize for code generation", false, false)
199 FunctionPass *llvm::createCodeGenPreparePass(const TargetMachine *TM) {
200 return new CodeGenPrepare(TM);
203 bool CodeGenPrepare::runOnFunction(Function &F) {
204 if (skipOptnoneFunction(F))
207 bool EverMadeChange = false;
208 // Clear per function information.
209 InsertedTruncsSet.clear();
210 PromotedInsts.clear();
214 TLI = TM->getSubtargetImpl()->getTargetLowering();
215 TLInfo = &getAnalysis<TargetLibraryInfoWrapperPass>().getTLI();
216 TTI = &getAnalysis<TargetTransformInfo>();
217 DominatorTreeWrapperPass *DTWP =
218 getAnalysisIfAvailable<DominatorTreeWrapperPass>();
219 DT = DTWP ? &DTWP->getDomTree() : nullptr;
220 OptSize = F.getAttributes().hasAttribute(AttributeSet::FunctionIndex,
221 Attribute::OptimizeForSize);
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 ModifiedDT |= ModifiedDTOnIteration;
260 if (ModifiedDTOnIteration)
263 EverMadeChange |= MadeChange;
268 if (!DisableBranchOpts) {
270 SmallPtrSet<BasicBlock*, 8> WorkList;
271 for (BasicBlock &BB : F) {
272 SmallVector<BasicBlock *, 2> Successors(succ_begin(&BB), succ_end(&BB));
273 MadeChange |= ConstantFoldTerminator(&BB, true);
274 if (!MadeChange) continue;
276 for (SmallVectorImpl<BasicBlock*>::iterator
277 II = Successors.begin(), IE = Successors.end(); II != IE; ++II)
278 if (pred_begin(*II) == pred_end(*II))
279 WorkList.insert(*II);
282 // Delete the dead blocks and any of their dead successors.
283 MadeChange |= !WorkList.empty();
284 while (!WorkList.empty()) {
285 BasicBlock *BB = *WorkList.begin();
287 SmallVector<BasicBlock*, 2> Successors(succ_begin(BB), succ_end(BB));
291 for (SmallVectorImpl<BasicBlock*>::iterator
292 II = Successors.begin(), IE = Successors.end(); II != IE; ++II)
293 if (pred_begin(*II) == pred_end(*II))
294 WorkList.insert(*II);
297 // Merge pairs of basic blocks with unconditional branches, connected by
299 if (EverMadeChange || MadeChange)
300 MadeChange |= EliminateFallThrough(F);
304 EverMadeChange |= MadeChange;
307 if (!DisableGCOpts) {
308 SmallVector<Instruction *, 2> Statepoints;
309 for (BasicBlock &BB : F)
310 for (Instruction &I : BB)
312 Statepoints.push_back(&I);
313 for (auto &I : Statepoints)
314 EverMadeChange |= simplifyOffsetableRelocate(*I);
317 if (ModifiedDT && DT)
320 return EverMadeChange;
323 /// EliminateFallThrough - Merge basic blocks which are connected
324 /// by a single edge, where one of the basic blocks has a single successor
325 /// pointing to the other basic block, which has a single predecessor.
326 bool CodeGenPrepare::EliminateFallThrough(Function &F) {
327 bool Changed = false;
328 // Scan all of the blocks in the function, except for the entry block.
329 for (Function::iterator I = std::next(F.begin()), E = F.end(); I != E;) {
330 BasicBlock *BB = I++;
331 // If the destination block has a single pred, then this is a trivial
332 // edge, just collapse it.
333 BasicBlock *SinglePred = BB->getSinglePredecessor();
335 // Don't merge if BB's address is taken.
336 if (!SinglePred || SinglePred == BB || BB->hasAddressTaken()) continue;
338 BranchInst *Term = dyn_cast<BranchInst>(SinglePred->getTerminator());
339 if (Term && !Term->isConditional()) {
341 DEBUG(dbgs() << "To merge:\n"<< *SinglePred << "\n\n\n");
342 // Remember if SinglePred was the entry block of the function.
343 // If so, we will need to move BB back to the entry position.
344 bool isEntry = SinglePred == &SinglePred->getParent()->getEntryBlock();
345 MergeBasicBlockIntoOnlyPred(BB, DT);
347 if (isEntry && BB != &BB->getParent()->getEntryBlock())
348 BB->moveBefore(&BB->getParent()->getEntryBlock());
350 // We have erased a block. Update the iterator.
357 /// EliminateMostlyEmptyBlocks - eliminate blocks that contain only PHI nodes,
358 /// debug info directives, and an unconditional branch. Passes before isel
359 /// (e.g. LSR/loopsimplify) often split edges in ways that are non-optimal for
360 /// isel. Start by eliminating these blocks so we can split them the way we
362 bool CodeGenPrepare::EliminateMostlyEmptyBlocks(Function &F) {
363 bool MadeChange = false;
364 // Note that this intentionally skips the entry block.
365 for (Function::iterator I = std::next(F.begin()), E = F.end(); I != E;) {
366 BasicBlock *BB = I++;
368 // If this block doesn't end with an uncond branch, ignore it.
369 BranchInst *BI = dyn_cast<BranchInst>(BB->getTerminator());
370 if (!BI || !BI->isUnconditional())
373 // If the instruction before the branch (skipping debug info) isn't a phi
374 // node, then other stuff is happening here.
375 BasicBlock::iterator BBI = BI;
376 if (BBI != BB->begin()) {
378 while (isa<DbgInfoIntrinsic>(BBI)) {
379 if (BBI == BB->begin())
383 if (!isa<DbgInfoIntrinsic>(BBI) && !isa<PHINode>(BBI))
387 // Do not break infinite loops.
388 BasicBlock *DestBB = BI->getSuccessor(0);
392 if (!CanMergeBlocks(BB, DestBB))
395 EliminateMostlyEmptyBlock(BB);
401 /// CanMergeBlocks - Return true if we can merge BB into DestBB if there is a
402 /// single uncond branch between them, and BB contains no other non-phi
404 bool CodeGenPrepare::CanMergeBlocks(const BasicBlock *BB,
405 const BasicBlock *DestBB) const {
406 // We only want to eliminate blocks whose phi nodes are used by phi nodes in
407 // the successor. If there are more complex condition (e.g. preheaders),
408 // don't mess around with them.
409 BasicBlock::const_iterator BBI = BB->begin();
410 while (const PHINode *PN = dyn_cast<PHINode>(BBI++)) {
411 for (const User *U : PN->users()) {
412 const Instruction *UI = cast<Instruction>(U);
413 if (UI->getParent() != DestBB || !isa<PHINode>(UI))
415 // If User is inside DestBB block and it is a PHINode then check
416 // incoming value. If incoming value is not from BB then this is
417 // a complex condition (e.g. preheaders) we want to avoid here.
418 if (UI->getParent() == DestBB) {
419 if (const PHINode *UPN = dyn_cast<PHINode>(UI))
420 for (unsigned I = 0, E = UPN->getNumIncomingValues(); I != E; ++I) {
421 Instruction *Insn = dyn_cast<Instruction>(UPN->getIncomingValue(I));
422 if (Insn && Insn->getParent() == BB &&
423 Insn->getParent() != UPN->getIncomingBlock(I))
430 // If BB and DestBB contain any common predecessors, then the phi nodes in BB
431 // and DestBB may have conflicting incoming values for the block. If so, we
432 // can't merge the block.
433 const PHINode *DestBBPN = dyn_cast<PHINode>(DestBB->begin());
434 if (!DestBBPN) return true; // no conflict.
436 // Collect the preds of BB.
437 SmallPtrSet<const BasicBlock*, 16> BBPreds;
438 if (const PHINode *BBPN = dyn_cast<PHINode>(BB->begin())) {
439 // It is faster to get preds from a PHI than with pred_iterator.
440 for (unsigned i = 0, e = BBPN->getNumIncomingValues(); i != e; ++i)
441 BBPreds.insert(BBPN->getIncomingBlock(i));
443 BBPreds.insert(pred_begin(BB), pred_end(BB));
446 // Walk the preds of DestBB.
447 for (unsigned i = 0, e = DestBBPN->getNumIncomingValues(); i != e; ++i) {
448 BasicBlock *Pred = DestBBPN->getIncomingBlock(i);
449 if (BBPreds.count(Pred)) { // Common predecessor?
450 BBI = DestBB->begin();
451 while (const PHINode *PN = dyn_cast<PHINode>(BBI++)) {
452 const Value *V1 = PN->getIncomingValueForBlock(Pred);
453 const Value *V2 = PN->getIncomingValueForBlock(BB);
455 // If V2 is a phi node in BB, look up what the mapped value will be.
456 if (const PHINode *V2PN = dyn_cast<PHINode>(V2))
457 if (V2PN->getParent() == BB)
458 V2 = V2PN->getIncomingValueForBlock(Pred);
460 // If there is a conflict, bail out.
461 if (V1 != V2) return false;
470 /// EliminateMostlyEmptyBlock - Eliminate a basic block that have only phi's and
471 /// an unconditional branch in it.
472 void CodeGenPrepare::EliminateMostlyEmptyBlock(BasicBlock *BB) {
473 BranchInst *BI = cast<BranchInst>(BB->getTerminator());
474 BasicBlock *DestBB = BI->getSuccessor(0);
476 DEBUG(dbgs() << "MERGING MOSTLY EMPTY BLOCKS - BEFORE:\n" << *BB << *DestBB);
478 // If the destination block has a single pred, then this is a trivial edge,
480 if (BasicBlock *SinglePred = DestBB->getSinglePredecessor()) {
481 if (SinglePred != DestBB) {
482 // Remember if SinglePred was the entry block of the function. If so, we
483 // will need to move BB back to the entry position.
484 bool isEntry = SinglePred == &SinglePred->getParent()->getEntryBlock();
485 MergeBasicBlockIntoOnlyPred(DestBB, DT);
487 if (isEntry && BB != &BB->getParent()->getEntryBlock())
488 BB->moveBefore(&BB->getParent()->getEntryBlock());
490 DEBUG(dbgs() << "AFTER:\n" << *DestBB << "\n\n\n");
495 // Otherwise, we have multiple predecessors of BB. Update the PHIs in DestBB
496 // to handle the new incoming edges it is about to have.
498 for (BasicBlock::iterator BBI = DestBB->begin();
499 (PN = dyn_cast<PHINode>(BBI)); ++BBI) {
500 // Remove the incoming value for BB, and remember it.
501 Value *InVal = PN->removeIncomingValue(BB, false);
503 // Two options: either the InVal is a phi node defined in BB or it is some
504 // value that dominates BB.
505 PHINode *InValPhi = dyn_cast<PHINode>(InVal);
506 if (InValPhi && InValPhi->getParent() == BB) {
507 // Add all of the input values of the input PHI as inputs of this phi.
508 for (unsigned i = 0, e = InValPhi->getNumIncomingValues(); i != e; ++i)
509 PN->addIncoming(InValPhi->getIncomingValue(i),
510 InValPhi->getIncomingBlock(i));
512 // Otherwise, add one instance of the dominating value for each edge that
513 // we will be adding.
514 if (PHINode *BBPN = dyn_cast<PHINode>(BB->begin())) {
515 for (unsigned i = 0, e = BBPN->getNumIncomingValues(); i != e; ++i)
516 PN->addIncoming(InVal, BBPN->getIncomingBlock(i));
518 for (pred_iterator PI = pred_begin(BB), E = pred_end(BB); PI != E; ++PI)
519 PN->addIncoming(InVal, *PI);
524 // The PHIs are now updated, change everything that refers to BB to use
525 // DestBB and remove BB.
526 BB->replaceAllUsesWith(DestBB);
527 if (DT && !ModifiedDT) {
528 BasicBlock *BBIDom = DT->getNode(BB)->getIDom()->getBlock();
529 BasicBlock *DestBBIDom = DT->getNode(DestBB)->getIDom()->getBlock();
530 BasicBlock *NewIDom = DT->findNearestCommonDominator(BBIDom, DestBBIDom);
531 DT->changeImmediateDominator(DestBB, NewIDom);
534 BB->eraseFromParent();
537 DEBUG(dbgs() << "AFTER:\n" << *DestBB << "\n\n\n");
540 // Computes a map of base pointer relocation instructions to corresponding
541 // derived pointer relocation instructions given a vector of all relocate calls
542 static void computeBaseDerivedRelocateMap(
543 const SmallVectorImpl<User *> &AllRelocateCalls,
544 DenseMap<IntrinsicInst *, SmallVector<IntrinsicInst *, 2>> &
546 // Collect information in two maps: one primarily for locating the base object
547 // while filling the second map; the second map is the final structure holding
548 // a mapping between Base and corresponding Derived relocate calls
549 DenseMap<std::pair<unsigned, unsigned>, IntrinsicInst *> RelocateIdxMap;
550 for (auto &U : AllRelocateCalls) {
551 GCRelocateOperands ThisRelocate(U);
552 IntrinsicInst *I = cast<IntrinsicInst>(U);
553 auto K = std::make_pair(ThisRelocate.basePtrIndex(),
554 ThisRelocate.derivedPtrIndex());
555 RelocateIdxMap.insert(std::make_pair(K, I));
557 for (auto &Item : RelocateIdxMap) {
558 std::pair<unsigned, unsigned> Key = Item.first;
559 if (Key.first == Key.second)
560 // Base relocation: nothing to insert
563 IntrinsicInst *I = Item.second;
564 auto BaseKey = std::make_pair(Key.first, Key.first);
565 IntrinsicInst *Base = RelocateIdxMap[BaseKey];
567 // TODO: We might want to insert a new base object relocate and gep off
568 // that, if there are enough derived object relocates.
570 RelocateInstMap[Base].push_back(I);
574 // Accepts a GEP and extracts the operands into a vector provided they're all
575 // small integer constants
576 static bool getGEPSmallConstantIntOffsetV(GetElementPtrInst *GEP,
577 SmallVectorImpl<Value *> &OffsetV) {
578 for (unsigned i = 1; i < GEP->getNumOperands(); i++) {
579 // Only accept small constant integer operands
580 auto Op = dyn_cast<ConstantInt>(GEP->getOperand(i));
581 if (!Op || Op->getZExtValue() > 20)
585 for (unsigned i = 1; i < GEP->getNumOperands(); i++)
586 OffsetV.push_back(GEP->getOperand(i));
590 // Takes a RelocatedBase (base pointer relocation instruction) and Targets to
591 // replace, computes a replacement, and affects it.
593 simplifyRelocatesOffABase(IntrinsicInst *RelocatedBase,
594 const SmallVectorImpl<IntrinsicInst *> &Targets) {
595 bool MadeChange = false;
596 for (auto &ToReplace : Targets) {
597 GCRelocateOperands MasterRelocate(RelocatedBase);
598 GCRelocateOperands ThisRelocate(ToReplace);
600 assert(ThisRelocate.basePtrIndex() == MasterRelocate.basePtrIndex() &&
601 "Not relocating a derived object of the original base object");
602 if (ThisRelocate.basePtrIndex() == ThisRelocate.derivedPtrIndex()) {
603 // A duplicate relocate call. TODO: coalesce duplicates.
607 Value *Base = ThisRelocate.basePtr();
608 auto Derived = dyn_cast<GetElementPtrInst>(ThisRelocate.derivedPtr());
609 if (!Derived || Derived->getPointerOperand() != Base)
612 SmallVector<Value *, 2> OffsetV;
613 if (!getGEPSmallConstantIntOffsetV(Derived, OffsetV))
616 // Create a Builder and replace the target callsite with a gep
617 IRBuilder<> Builder(ToReplace);
618 Builder.SetCurrentDebugLocation(ToReplace->getDebugLoc());
620 Builder.CreateGEP(RelocatedBase, makeArrayRef(OffsetV));
621 Instruction *ReplacementInst = cast<Instruction>(Replacement);
622 ReplacementInst->removeFromParent();
623 ReplacementInst->insertAfter(RelocatedBase);
624 Replacement->takeName(ToReplace);
625 ToReplace->replaceAllUsesWith(Replacement);
626 ToReplace->eraseFromParent();
636 // %ptr = gep %base + 15
637 // %tok = statepoint (%fun, i32 0, i32 0, i32 0, %base, %ptr)
638 // %base' = relocate(%tok, i32 4, i32 4)
639 // %ptr' = relocate(%tok, i32 4, i32 5)
645 // %ptr = gep %base + 15
646 // %tok = statepoint (%fun, i32 0, i32 0, i32 0, %base, %ptr)
647 // %base' = gc.relocate(%tok, i32 4, i32 4)
648 // %ptr' = gep %base' + 15
650 bool CodeGenPrepare::simplifyOffsetableRelocate(Instruction &I) {
651 bool MadeChange = false;
652 SmallVector<User *, 2> AllRelocateCalls;
654 for (auto *U : I.users())
655 if (isGCRelocate(dyn_cast<Instruction>(U)))
656 // Collect all the relocate calls associated with a statepoint
657 AllRelocateCalls.push_back(U);
659 // We need atleast one base pointer relocation + one derived pointer
660 // relocation to mangle
661 if (AllRelocateCalls.size() < 2)
664 // RelocateInstMap is a mapping from the base relocate instruction to the
665 // corresponding derived relocate instructions
666 DenseMap<IntrinsicInst *, SmallVector<IntrinsicInst *, 2>> RelocateInstMap;
667 computeBaseDerivedRelocateMap(AllRelocateCalls, RelocateInstMap);
668 if (RelocateInstMap.empty())
671 for (auto &Item : RelocateInstMap)
672 // Item.first is the RelocatedBase to offset against
673 // Item.second is the vector of Targets to replace
674 MadeChange = simplifyRelocatesOffABase(Item.first, Item.second);
678 /// SinkCast - Sink the specified cast instruction into its user blocks
679 static bool SinkCast(CastInst *CI) {
680 BasicBlock *DefBB = CI->getParent();
682 /// InsertedCasts - Only insert a cast in each block once.
683 DenseMap<BasicBlock*, CastInst*> InsertedCasts;
685 bool MadeChange = false;
686 for (Value::user_iterator UI = CI->user_begin(), E = CI->user_end();
688 Use &TheUse = UI.getUse();
689 Instruction *User = cast<Instruction>(*UI);
691 // Figure out which BB this cast is used in. For PHI's this is the
692 // appropriate predecessor block.
693 BasicBlock *UserBB = User->getParent();
694 if (PHINode *PN = dyn_cast<PHINode>(User)) {
695 UserBB = PN->getIncomingBlock(TheUse);
698 // Preincrement use iterator so we don't invalidate it.
701 // If this user is in the same block as the cast, don't change the cast.
702 if (UserBB == DefBB) continue;
704 // If we have already inserted a cast into this block, use it.
705 CastInst *&InsertedCast = InsertedCasts[UserBB];
708 BasicBlock::iterator InsertPt = UserBB->getFirstInsertionPt();
710 CastInst::Create(CI->getOpcode(), CI->getOperand(0), CI->getType(), "",
715 // Replace a use of the cast with a use of the new cast.
716 TheUse = InsertedCast;
720 // If we removed all uses, nuke the cast.
721 if (CI->use_empty()) {
722 CI->eraseFromParent();
729 /// OptimizeNoopCopyExpression - If the specified cast instruction is a noop
730 /// copy (e.g. it's casting from one pointer type to another, i32->i8 on PPC),
731 /// sink it into user blocks to reduce the number of virtual
732 /// registers that must be created and coalesced.
734 /// Return true if any changes are made.
736 static bool OptimizeNoopCopyExpression(CastInst *CI, const TargetLowering &TLI){
737 // If this is a noop copy,
738 EVT SrcVT = TLI.getValueType(CI->getOperand(0)->getType());
739 EVT DstVT = TLI.getValueType(CI->getType());
741 // This is an fp<->int conversion?
742 if (SrcVT.isInteger() != DstVT.isInteger())
745 // If this is an extension, it will be a zero or sign extension, which
747 if (SrcVT.bitsLT(DstVT)) return false;
749 // If these values will be promoted, find out what they will be promoted
750 // to. This helps us consider truncates on PPC as noop copies when they
752 if (TLI.getTypeAction(CI->getContext(), SrcVT) ==
753 TargetLowering::TypePromoteInteger)
754 SrcVT = TLI.getTypeToTransformTo(CI->getContext(), SrcVT);
755 if (TLI.getTypeAction(CI->getContext(), DstVT) ==
756 TargetLowering::TypePromoteInteger)
757 DstVT = TLI.getTypeToTransformTo(CI->getContext(), DstVT);
759 // If, after promotion, these are the same types, this is a noop copy.
766 /// OptimizeCmpExpression - sink the given CmpInst into user blocks to reduce
767 /// the number of virtual registers that must be created and coalesced. This is
768 /// a clear win except on targets with multiple condition code registers
769 /// (PowerPC), where it might lose; some adjustment may be wanted there.
771 /// Return true if any changes are made.
772 static bool OptimizeCmpExpression(CmpInst *CI) {
773 BasicBlock *DefBB = CI->getParent();
775 /// InsertedCmp - Only insert a cmp in each block once.
776 DenseMap<BasicBlock*, CmpInst*> InsertedCmps;
778 bool MadeChange = false;
779 for (Value::user_iterator UI = CI->user_begin(), E = CI->user_end();
781 Use &TheUse = UI.getUse();
782 Instruction *User = cast<Instruction>(*UI);
784 // Preincrement use iterator so we don't invalidate it.
787 // Don't bother for PHI nodes.
788 if (isa<PHINode>(User))
791 // Figure out which BB this cmp is used in.
792 BasicBlock *UserBB = User->getParent();
794 // If this user is in the same block as the cmp, don't change the cmp.
795 if (UserBB == DefBB) continue;
797 // If we have already inserted a cmp into this block, use it.
798 CmpInst *&InsertedCmp = InsertedCmps[UserBB];
801 BasicBlock::iterator InsertPt = UserBB->getFirstInsertionPt();
803 CmpInst::Create(CI->getOpcode(),
804 CI->getPredicate(), CI->getOperand(0),
805 CI->getOperand(1), "", InsertPt);
809 // Replace a use of the cmp with a use of the new cmp.
810 TheUse = InsertedCmp;
814 // If we removed all uses, nuke the cmp.
816 CI->eraseFromParent();
821 /// isExtractBitsCandidateUse - Check if the candidates could
822 /// be combined with shift instruction, which includes:
823 /// 1. Truncate instruction
824 /// 2. And instruction and the imm is a mask of the low bits:
825 /// imm & (imm+1) == 0
826 static bool isExtractBitsCandidateUse(Instruction *User) {
827 if (!isa<TruncInst>(User)) {
828 if (User->getOpcode() != Instruction::And ||
829 !isa<ConstantInt>(User->getOperand(1)))
832 const APInt &Cimm = cast<ConstantInt>(User->getOperand(1))->getValue();
834 if ((Cimm & (Cimm + 1)).getBoolValue())
840 /// SinkShiftAndTruncate - sink both shift and truncate instruction
841 /// to the use of truncate's BB.
843 SinkShiftAndTruncate(BinaryOperator *ShiftI, Instruction *User, ConstantInt *CI,
844 DenseMap<BasicBlock *, BinaryOperator *> &InsertedShifts,
845 const TargetLowering &TLI) {
846 BasicBlock *UserBB = User->getParent();
847 DenseMap<BasicBlock *, CastInst *> InsertedTruncs;
848 TruncInst *TruncI = dyn_cast<TruncInst>(User);
849 bool MadeChange = false;
851 for (Value::user_iterator TruncUI = TruncI->user_begin(),
852 TruncE = TruncI->user_end();
853 TruncUI != TruncE;) {
855 Use &TruncTheUse = TruncUI.getUse();
856 Instruction *TruncUser = cast<Instruction>(*TruncUI);
857 // Preincrement use iterator so we don't invalidate it.
861 int ISDOpcode = TLI.InstructionOpcodeToISD(TruncUser->getOpcode());
865 // If the use is actually a legal node, there will not be an
866 // implicit truncate.
867 // FIXME: always querying the result type is just an
868 // approximation; some nodes' legality is determined by the
869 // operand or other means. There's no good way to find out though.
870 if (TLI.isOperationLegalOrCustom(
871 ISDOpcode, TLI.getValueType(TruncUser->getType(), true)))
874 // Don't bother for PHI nodes.
875 if (isa<PHINode>(TruncUser))
878 BasicBlock *TruncUserBB = TruncUser->getParent();
880 if (UserBB == TruncUserBB)
883 BinaryOperator *&InsertedShift = InsertedShifts[TruncUserBB];
884 CastInst *&InsertedTrunc = InsertedTruncs[TruncUserBB];
886 if (!InsertedShift && !InsertedTrunc) {
887 BasicBlock::iterator InsertPt = TruncUserBB->getFirstInsertionPt();
889 if (ShiftI->getOpcode() == Instruction::AShr)
891 BinaryOperator::CreateAShr(ShiftI->getOperand(0), CI, "", InsertPt);
894 BinaryOperator::CreateLShr(ShiftI->getOperand(0), CI, "", InsertPt);
897 BasicBlock::iterator TruncInsertPt = TruncUserBB->getFirstInsertionPt();
900 InsertedTrunc = CastInst::Create(TruncI->getOpcode(), InsertedShift,
901 TruncI->getType(), "", TruncInsertPt);
905 TruncTheUse = InsertedTrunc;
911 /// OptimizeExtractBits - sink the shift *right* instruction into user blocks if
912 /// the uses could potentially be combined with this shift instruction and
913 /// generate BitExtract instruction. It will only be applied if the architecture
914 /// supports BitExtract instruction. Here is an example:
916 /// %x.extract.shift = lshr i64 %arg1, 32
918 /// %x.extract.trunc = trunc i64 %x.extract.shift to i16
922 /// %x.extract.shift.1 = lshr i64 %arg1, 32
923 /// %x.extract.trunc = trunc i64 %x.extract.shift.1 to i16
925 /// CodeGen will recoginze the pattern in BB2 and generate BitExtract
927 /// Return true if any changes are made.
928 static bool OptimizeExtractBits(BinaryOperator *ShiftI, ConstantInt *CI,
929 const TargetLowering &TLI) {
930 BasicBlock *DefBB = ShiftI->getParent();
932 /// Only insert instructions in each block once.
933 DenseMap<BasicBlock *, BinaryOperator *> InsertedShifts;
935 bool shiftIsLegal = TLI.isTypeLegal(TLI.getValueType(ShiftI->getType()));
937 bool MadeChange = false;
938 for (Value::user_iterator UI = ShiftI->user_begin(), E = ShiftI->user_end();
940 Use &TheUse = UI.getUse();
941 Instruction *User = cast<Instruction>(*UI);
942 // Preincrement use iterator so we don't invalidate it.
945 // Don't bother for PHI nodes.
946 if (isa<PHINode>(User))
949 if (!isExtractBitsCandidateUse(User))
952 BasicBlock *UserBB = User->getParent();
954 if (UserBB == DefBB) {
955 // If the shift and truncate instruction are in the same BB. The use of
956 // the truncate(TruncUse) may still introduce another truncate if not
957 // legal. In this case, we would like to sink both shift and truncate
958 // instruction to the BB of TruncUse.
961 // i64 shift.result = lshr i64 opnd, imm
962 // trunc.result = trunc shift.result to i16
965 // ----> We will have an implicit truncate here if the architecture does
966 // not have i16 compare.
967 // cmp i16 trunc.result, opnd2
969 if (isa<TruncInst>(User) && shiftIsLegal
970 // If the type of the truncate is legal, no trucate will be
971 // introduced in other basic blocks.
972 && (!TLI.isTypeLegal(TLI.getValueType(User->getType()))))
974 SinkShiftAndTruncate(ShiftI, User, CI, InsertedShifts, TLI);
978 // If we have already inserted a shift into this block, use it.
979 BinaryOperator *&InsertedShift = InsertedShifts[UserBB];
981 if (!InsertedShift) {
982 BasicBlock::iterator InsertPt = UserBB->getFirstInsertionPt();
984 if (ShiftI->getOpcode() == Instruction::AShr)
986 BinaryOperator::CreateAShr(ShiftI->getOperand(0), CI, "", InsertPt);
989 BinaryOperator::CreateLShr(ShiftI->getOperand(0), CI, "", InsertPt);
994 // Replace a use of the shift with a use of the new shift.
995 TheUse = InsertedShift;
998 // If we removed all uses, nuke the shift.
999 if (ShiftI->use_empty())
1000 ShiftI->eraseFromParent();
1005 // ScalarizeMaskedLoad() translates masked load intrinsic, like
1006 // <16 x i32 > @llvm.masked.load( <16 x i32>* %addr, i32 align,
1007 // <16 x i1> %mask, <16 x i32> %passthru)
1008 // to a chain of basic blocks, whith loading element one-by-one if
1009 // the appropriate mask bit is set
1011 // %1 = bitcast i8* %addr to i32*
1012 // %2 = extractelement <16 x i1> %mask, i32 0
1013 // %3 = icmp eq i1 %2, true
1014 // br i1 %3, label %cond.load, label %else
1016 //cond.load: ; preds = %0
1017 // %4 = getelementptr i32* %1, i32 0
1018 // %5 = load i32* %4
1019 // %6 = insertelement <16 x i32> undef, i32 %5, i32 0
1022 //else: ; preds = %0, %cond.load
1023 // %res.phi.else = phi <16 x i32> [ %6, %cond.load ], [ undef, %0 ]
1024 // %7 = extractelement <16 x i1> %mask, i32 1
1025 // %8 = icmp eq i1 %7, true
1026 // br i1 %8, label %cond.load1, label %else2
1028 //cond.load1: ; preds = %else
1029 // %9 = getelementptr i32* %1, i32 1
1030 // %10 = load i32* %9
1031 // %11 = insertelement <16 x i32> %res.phi.else, i32 %10, i32 1
1034 //else2: ; preds = %else, %cond.load1
1035 // %res.phi.else3 = phi <16 x i32> [ %11, %cond.load1 ], [ %res.phi.else, %else ]
1036 // %12 = extractelement <16 x i1> %mask, i32 2
1037 // %13 = icmp eq i1 %12, true
1038 // br i1 %13, label %cond.load4, label %else5
1040 static void ScalarizeMaskedLoad(CallInst *CI) {
1041 Value *Ptr = CI->getArgOperand(0);
1042 Value *Src0 = CI->getArgOperand(3);
1043 Value *Mask = CI->getArgOperand(2);
1044 VectorType *VecType = dyn_cast<VectorType>(CI->getType());
1045 Type *EltTy = VecType->getElementType();
1047 assert(VecType && "Unexpected return type of masked load intrinsic");
1049 IRBuilder<> Builder(CI->getContext());
1050 Instruction *InsertPt = CI;
1051 BasicBlock *IfBlock = CI->getParent();
1052 BasicBlock *CondBlock = nullptr;
1053 BasicBlock *PrevIfBlock = CI->getParent();
1054 Builder.SetInsertPoint(InsertPt);
1056 Builder.SetCurrentDebugLocation(CI->getDebugLoc());
1058 // Bitcast %addr fron i8* to EltTy*
1060 EltTy->getPointerTo(cast<PointerType>(Ptr->getType())->getAddressSpace());
1061 Value *FirstEltPtr = Builder.CreateBitCast(Ptr, NewPtrType);
1062 Value *UndefVal = UndefValue::get(VecType);
1064 // The result vector
1065 Value *VResult = UndefVal;
1067 PHINode *Phi = nullptr;
1068 Value *PrevPhi = UndefVal;
1070 unsigned VectorWidth = VecType->getNumElements();
1071 for (unsigned Idx = 0; Idx < VectorWidth; ++Idx) {
1073 // Fill the "else" block, created in the previous iteration
1075 // %res.phi.else3 = phi <16 x i32> [ %11, %cond.load1 ], [ %res.phi.else, %else ]
1076 // %mask_1 = extractelement <16 x i1> %mask, i32 Idx
1077 // %to_load = icmp eq i1 %mask_1, true
1078 // br i1 %to_load, label %cond.load, label %else
1081 Phi = Builder.CreatePHI(VecType, 2, "res.phi.else");
1082 Phi->addIncoming(VResult, CondBlock);
1083 Phi->addIncoming(PrevPhi, PrevIfBlock);
1088 Value *Predicate = Builder.CreateExtractElement(Mask, Builder.getInt32(Idx));
1089 Value *Cmp = Builder.CreateICmp(ICmpInst::ICMP_EQ, Predicate,
1090 ConstantInt::get(Predicate->getType(), 1));
1092 // Create "cond" block
1094 // %EltAddr = getelementptr i32* %1, i32 0
1095 // %Elt = load i32* %EltAddr
1096 // VResult = insertelement <16 x i32> VResult, i32 %Elt, i32 Idx
1098 CondBlock = IfBlock->splitBasicBlock(InsertPt, "cond.load");
1099 Builder.SetInsertPoint(InsertPt);
1101 Value* Gep = Builder.CreateInBoundsGEP(FirstEltPtr, Builder.getInt32(Idx));
1102 LoadInst* Load = Builder.CreateLoad(Gep, false);
1103 VResult = Builder.CreateInsertElement(VResult, Load, Builder.getInt32(Idx));
1105 // Create "else" block, fill it in the next iteration
1106 BasicBlock *NewIfBlock = CondBlock->splitBasicBlock(InsertPt, "else");
1107 Builder.SetInsertPoint(InsertPt);
1108 Instruction *OldBr = IfBlock->getTerminator();
1109 BranchInst::Create(CondBlock, NewIfBlock, Cmp, OldBr);
1110 OldBr->eraseFromParent();
1111 PrevIfBlock = IfBlock;
1112 IfBlock = NewIfBlock;
1115 Phi = Builder.CreatePHI(VecType, 2, "res.phi.select");
1116 Phi->addIncoming(VResult, CondBlock);
1117 Phi->addIncoming(PrevPhi, PrevIfBlock);
1118 Value *NewI = Builder.CreateSelect(Mask, Phi, Src0);
1119 CI->replaceAllUsesWith(NewI);
1120 CI->eraseFromParent();
1123 // ScalarizeMaskedStore() translates masked store intrinsic, like
1124 // void @llvm.masked.store(<16 x i32> %src, <16 x i32>* %addr, i32 align,
1126 // to a chain of basic blocks, that stores element one-by-one if
1127 // the appropriate mask bit is set
1129 // %1 = bitcast i8* %addr to i32*
1130 // %2 = extractelement <16 x i1> %mask, i32 0
1131 // %3 = icmp eq i1 %2, true
1132 // br i1 %3, label %cond.store, label %else
1134 // cond.store: ; preds = %0
1135 // %4 = extractelement <16 x i32> %val, i32 0
1136 // %5 = getelementptr i32* %1, i32 0
1137 // store i32 %4, i32* %5
1140 // else: ; preds = %0, %cond.store
1141 // %6 = extractelement <16 x i1> %mask, i32 1
1142 // %7 = icmp eq i1 %6, true
1143 // br i1 %7, label %cond.store1, label %else2
1145 // cond.store1: ; preds = %else
1146 // %8 = extractelement <16 x i32> %val, i32 1
1147 // %9 = getelementptr i32* %1, i32 1
1148 // store i32 %8, i32* %9
1151 static void ScalarizeMaskedStore(CallInst *CI) {
1152 Value *Ptr = CI->getArgOperand(1);
1153 Value *Src = CI->getArgOperand(0);
1154 Value *Mask = CI->getArgOperand(3);
1156 VectorType *VecType = dyn_cast<VectorType>(Src->getType());
1157 Type *EltTy = VecType->getElementType();
1159 assert(VecType && "Unexpected data type in masked store intrinsic");
1161 IRBuilder<> Builder(CI->getContext());
1162 Instruction *InsertPt = CI;
1163 BasicBlock *IfBlock = CI->getParent();
1164 Builder.SetInsertPoint(InsertPt);
1165 Builder.SetCurrentDebugLocation(CI->getDebugLoc());
1167 // Bitcast %addr fron i8* to EltTy*
1169 EltTy->getPointerTo(cast<PointerType>(Ptr->getType())->getAddressSpace());
1170 Value *FirstEltPtr = Builder.CreateBitCast(Ptr, NewPtrType);
1172 unsigned VectorWidth = VecType->getNumElements();
1173 for (unsigned Idx = 0; Idx < VectorWidth; ++Idx) {
1175 // Fill the "else" block, created in the previous iteration
1177 // %mask_1 = extractelement <16 x i1> %mask, i32 Idx
1178 // %to_store = icmp eq i1 %mask_1, true
1179 // br i1 %to_load, label %cond.store, label %else
1181 Value *Predicate = Builder.CreateExtractElement(Mask, Builder.getInt32(Idx));
1182 Value *Cmp = Builder.CreateICmp(ICmpInst::ICMP_EQ, Predicate,
1183 ConstantInt::get(Predicate->getType(), 1));
1185 // Create "cond" block
1187 // %OneElt = extractelement <16 x i32> %Src, i32 Idx
1188 // %EltAddr = getelementptr i32* %1, i32 0
1189 // %store i32 %OneElt, i32* %EltAddr
1191 BasicBlock *CondBlock = IfBlock->splitBasicBlock(InsertPt, "cond.store");
1192 Builder.SetInsertPoint(InsertPt);
1194 Value *OneElt = Builder.CreateExtractElement(Src, Builder.getInt32(Idx));
1195 Value* Gep = Builder.CreateInBoundsGEP(FirstEltPtr, Builder.getInt32(Idx));
1196 Builder.CreateStore(OneElt, Gep);
1198 // Create "else" block, fill it in the next iteration
1199 BasicBlock *NewIfBlock = CondBlock->splitBasicBlock(InsertPt, "else");
1200 Builder.SetInsertPoint(InsertPt);
1201 Instruction *OldBr = IfBlock->getTerminator();
1202 BranchInst::Create(CondBlock, NewIfBlock, Cmp, OldBr);
1203 OldBr->eraseFromParent();
1204 IfBlock = NewIfBlock;
1206 CI->eraseFromParent();
1209 bool CodeGenPrepare::OptimizeCallInst(CallInst *CI, bool& ModifiedDT) {
1210 BasicBlock *BB = CI->getParent();
1212 // Lower inline assembly if we can.
1213 // If we found an inline asm expession, and if the target knows how to
1214 // lower it to normal LLVM code, do so now.
1215 if (TLI && isa<InlineAsm>(CI->getCalledValue())) {
1216 if (TLI->ExpandInlineAsm(CI)) {
1217 // Avoid invalidating the iterator.
1218 CurInstIterator = BB->begin();
1219 // Avoid processing instructions out of order, which could cause
1220 // reuse before a value is defined.
1224 // Sink address computing for memory operands into the block.
1225 if (OptimizeInlineAsmInst(CI))
1229 IntrinsicInst *II = dyn_cast<IntrinsicInst>(CI);
1231 switch (II->getIntrinsicID()) {
1233 case Intrinsic::objectsize: {
1234 // Lower all uses of llvm.objectsize.*
1235 bool Min = (cast<ConstantInt>(II->getArgOperand(1))->getZExtValue() == 1);
1236 Type *ReturnTy = CI->getType();
1237 Constant *RetVal = ConstantInt::get(ReturnTy, Min ? 0 : -1ULL);
1239 // Substituting this can cause recursive simplifications, which can
1240 // invalidate our iterator. Use a WeakVH to hold onto it in case this
1242 WeakVH IterHandle(CurInstIterator);
1244 replaceAndRecursivelySimplify(CI, RetVal,
1245 TLI ? TLI->getDataLayout() : nullptr,
1246 TLInfo, ModifiedDT ? nullptr : DT);
1248 // If the iterator instruction was recursively deleted, start over at the
1249 // start of the block.
1250 if (IterHandle != CurInstIterator) {
1251 CurInstIterator = BB->begin();
1256 case Intrinsic::masked_load: {
1257 // Scalarize unsupported vector masked load
1258 if (!TTI->isLegalMaskedLoad(CI->getType(), 1)) {
1259 ScalarizeMaskedLoad(CI);
1265 case Intrinsic::masked_store: {
1266 if (!TTI->isLegalMaskedStore(CI->getArgOperand(0)->getType(), 1)) {
1267 ScalarizeMaskedStore(CI);
1276 SmallVector<Value*, 2> PtrOps;
1278 if (TLI->GetAddrModeArguments(II, PtrOps, AccessTy))
1279 while (!PtrOps.empty())
1280 if (OptimizeMemoryInst(II, PtrOps.pop_back_val(), AccessTy))
1285 // From here on out we're working with named functions.
1286 if (!CI->getCalledFunction()) return false;
1288 // We'll need DataLayout from here on out.
1289 const DataLayout *TD = TLI ? TLI->getDataLayout() : nullptr;
1290 if (!TD) return false;
1292 // Lower all default uses of _chk calls. This is very similar
1293 // to what InstCombineCalls does, but here we are only lowering calls
1294 // to fortified library functions (e.g. __memcpy_chk) that have the default
1295 // "don't know" as the objectsize. Anything else should be left alone.
1296 FortifiedLibCallSimplifier Simplifier(TD, TLInfo, true);
1297 if (Value *V = Simplifier.optimizeCall(CI)) {
1298 CI->replaceAllUsesWith(V);
1299 CI->eraseFromParent();
1305 /// DupRetToEnableTailCallOpts - Look for opportunities to duplicate return
1306 /// instructions to the predecessor to enable tail call optimizations. The
1307 /// case it is currently looking for is:
1310 /// %tmp0 = tail call i32 @f0()
1311 /// br label %return
1313 /// %tmp1 = tail call i32 @f1()
1314 /// br label %return
1316 /// %tmp2 = tail call i32 @f2()
1317 /// br label %return
1319 /// %retval = phi i32 [ %tmp0, %bb0 ], [ %tmp1, %bb1 ], [ %tmp2, %bb2 ]
1327 /// %tmp0 = tail call i32 @f0()
1330 /// %tmp1 = tail call i32 @f1()
1333 /// %tmp2 = tail call i32 @f2()
1336 bool CodeGenPrepare::DupRetToEnableTailCallOpts(BasicBlock *BB) {
1340 ReturnInst *RI = dyn_cast<ReturnInst>(BB->getTerminator());
1344 PHINode *PN = nullptr;
1345 BitCastInst *BCI = nullptr;
1346 Value *V = RI->getReturnValue();
1348 BCI = dyn_cast<BitCastInst>(V);
1350 V = BCI->getOperand(0);
1352 PN = dyn_cast<PHINode>(V);
1357 if (PN && PN->getParent() != BB)
1360 // It's not safe to eliminate the sign / zero extension of the return value.
1361 // See llvm::isInTailCallPosition().
1362 const Function *F = BB->getParent();
1363 AttributeSet CallerAttrs = F->getAttributes();
1364 if (CallerAttrs.hasAttribute(AttributeSet::ReturnIndex, Attribute::ZExt) ||
1365 CallerAttrs.hasAttribute(AttributeSet::ReturnIndex, Attribute::SExt))
1368 // Make sure there are no instructions between the PHI and return, or that the
1369 // return is the first instruction in the block.
1371 BasicBlock::iterator BI = BB->begin();
1372 do { ++BI; } while (isa<DbgInfoIntrinsic>(BI));
1374 // Also skip over the bitcast.
1379 BasicBlock::iterator BI = BB->begin();
1380 while (isa<DbgInfoIntrinsic>(BI)) ++BI;
1385 /// Only dup the ReturnInst if the CallInst is likely to be emitted as a tail
1387 SmallVector<CallInst*, 4> TailCalls;
1389 for (unsigned I = 0, E = PN->getNumIncomingValues(); I != E; ++I) {
1390 CallInst *CI = dyn_cast<CallInst>(PN->getIncomingValue(I));
1391 // Make sure the phi value is indeed produced by the tail call.
1392 if (CI && CI->hasOneUse() && CI->getParent() == PN->getIncomingBlock(I) &&
1393 TLI->mayBeEmittedAsTailCall(CI))
1394 TailCalls.push_back(CI);
1397 SmallPtrSet<BasicBlock*, 4> VisitedBBs;
1398 for (pred_iterator PI = pred_begin(BB), PE = pred_end(BB); PI != PE; ++PI) {
1399 if (!VisitedBBs.insert(*PI).second)
1402 BasicBlock::InstListType &InstList = (*PI)->getInstList();
1403 BasicBlock::InstListType::reverse_iterator RI = InstList.rbegin();
1404 BasicBlock::InstListType::reverse_iterator RE = InstList.rend();
1405 do { ++RI; } while (RI != RE && isa<DbgInfoIntrinsic>(&*RI));
1409 CallInst *CI = dyn_cast<CallInst>(&*RI);
1410 if (CI && CI->use_empty() && TLI->mayBeEmittedAsTailCall(CI))
1411 TailCalls.push_back(CI);
1415 bool Changed = false;
1416 for (unsigned i = 0, e = TailCalls.size(); i != e; ++i) {
1417 CallInst *CI = TailCalls[i];
1420 // Conservatively require the attributes of the call to match those of the
1421 // return. Ignore noalias because it doesn't affect the call sequence.
1422 AttributeSet CalleeAttrs = CS.getAttributes();
1423 if (AttrBuilder(CalleeAttrs, AttributeSet::ReturnIndex).
1424 removeAttribute(Attribute::NoAlias) !=
1425 AttrBuilder(CalleeAttrs, AttributeSet::ReturnIndex).
1426 removeAttribute(Attribute::NoAlias))
1429 // Make sure the call instruction is followed by an unconditional branch to
1430 // the return block.
1431 BasicBlock *CallBB = CI->getParent();
1432 BranchInst *BI = dyn_cast<BranchInst>(CallBB->getTerminator());
1433 if (!BI || !BI->isUnconditional() || BI->getSuccessor(0) != BB)
1436 // Duplicate the return into CallBB.
1437 (void)FoldReturnIntoUncondBranch(RI, BB, CallBB);
1438 ModifiedDT = Changed = true;
1442 // If we eliminated all predecessors of the block, delete the block now.
1443 if (Changed && !BB->hasAddressTaken() && pred_begin(BB) == pred_end(BB))
1444 BB->eraseFromParent();
1449 //===----------------------------------------------------------------------===//
1450 // Memory Optimization
1451 //===----------------------------------------------------------------------===//
1455 /// ExtAddrMode - This is an extended version of TargetLowering::AddrMode
1456 /// which holds actual Value*'s for register values.
1457 struct ExtAddrMode : public TargetLowering::AddrMode {
1460 ExtAddrMode() : BaseReg(nullptr), ScaledReg(nullptr) {}
1461 void print(raw_ostream &OS) const;
1464 bool operator==(const ExtAddrMode& O) const {
1465 return (BaseReg == O.BaseReg) && (ScaledReg == O.ScaledReg) &&
1466 (BaseGV == O.BaseGV) && (BaseOffs == O.BaseOffs) &&
1467 (HasBaseReg == O.HasBaseReg) && (Scale == O.Scale);
1472 static inline raw_ostream &operator<<(raw_ostream &OS, const ExtAddrMode &AM) {
1478 void ExtAddrMode::print(raw_ostream &OS) const {
1479 bool NeedPlus = false;
1482 OS << (NeedPlus ? " + " : "")
1484 BaseGV->printAsOperand(OS, /*PrintType=*/false);
1489 OS << (NeedPlus ? " + " : "")
1495 OS << (NeedPlus ? " + " : "")
1497 BaseReg->printAsOperand(OS, /*PrintType=*/false);
1501 OS << (NeedPlus ? " + " : "")
1503 ScaledReg->printAsOperand(OS, /*PrintType=*/false);
1509 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
1510 void ExtAddrMode::dump() const {
1516 /// \brief This class provides transaction based operation on the IR.
1517 /// Every change made through this class is recorded in the internal state and
1518 /// can be undone (rollback) until commit is called.
1519 class TypePromotionTransaction {
1521 /// \brief This represents the common interface of the individual transaction.
1522 /// Each class implements the logic for doing one specific modification on
1523 /// the IR via the TypePromotionTransaction.
1524 class TypePromotionAction {
1526 /// The Instruction modified.
1530 /// \brief Constructor of the action.
1531 /// The constructor performs the related action on the IR.
1532 TypePromotionAction(Instruction *Inst) : Inst(Inst) {}
1534 virtual ~TypePromotionAction() {}
1536 /// \brief Undo the modification done by this action.
1537 /// When this method is called, the IR must be in the same state as it was
1538 /// before this action was applied.
1539 /// \pre Undoing the action works if and only if the IR is in the exact same
1540 /// state as it was directly after this action was applied.
1541 virtual void undo() = 0;
1543 /// \brief Advocate every change made by this action.
1544 /// When the results on the IR of the action are to be kept, it is important
1545 /// to call this function, otherwise hidden information may be kept forever.
1546 virtual void commit() {
1547 // Nothing to be done, this action is not doing anything.
1551 /// \brief Utility to remember the position of an instruction.
1552 class InsertionHandler {
1553 /// Position of an instruction.
1554 /// Either an instruction:
1555 /// - Is the first in a basic block: BB is used.
1556 /// - Has a previous instructon: PrevInst is used.
1558 Instruction *PrevInst;
1561 /// Remember whether or not the instruction had a previous instruction.
1562 bool HasPrevInstruction;
1565 /// \brief Record the position of \p Inst.
1566 InsertionHandler(Instruction *Inst) {
1567 BasicBlock::iterator It = Inst;
1568 HasPrevInstruction = (It != (Inst->getParent()->begin()));
1569 if (HasPrevInstruction)
1570 Point.PrevInst = --It;
1572 Point.BB = Inst->getParent();
1575 /// \brief Insert \p Inst at the recorded position.
1576 void insert(Instruction *Inst) {
1577 if (HasPrevInstruction) {
1578 if (Inst->getParent())
1579 Inst->removeFromParent();
1580 Inst->insertAfter(Point.PrevInst);
1582 Instruction *Position = Point.BB->getFirstInsertionPt();
1583 if (Inst->getParent())
1584 Inst->moveBefore(Position);
1586 Inst->insertBefore(Position);
1591 /// \brief Move an instruction before another.
1592 class InstructionMoveBefore : public TypePromotionAction {
1593 /// Original position of the instruction.
1594 InsertionHandler Position;
1597 /// \brief Move \p Inst before \p Before.
1598 InstructionMoveBefore(Instruction *Inst, Instruction *Before)
1599 : TypePromotionAction(Inst), Position(Inst) {
1600 DEBUG(dbgs() << "Do: move: " << *Inst << "\nbefore: " << *Before << "\n");
1601 Inst->moveBefore(Before);
1604 /// \brief Move the instruction back to its original position.
1605 void undo() override {
1606 DEBUG(dbgs() << "Undo: moveBefore: " << *Inst << "\n");
1607 Position.insert(Inst);
1611 /// \brief Set the operand of an instruction with a new value.
1612 class OperandSetter : public TypePromotionAction {
1613 /// Original operand of the instruction.
1615 /// Index of the modified instruction.
1619 /// \brief Set \p Idx operand of \p Inst with \p NewVal.
1620 OperandSetter(Instruction *Inst, unsigned Idx, Value *NewVal)
1621 : TypePromotionAction(Inst), Idx(Idx) {
1622 DEBUG(dbgs() << "Do: setOperand: " << Idx << "\n"
1623 << "for:" << *Inst << "\n"
1624 << "with:" << *NewVal << "\n");
1625 Origin = Inst->getOperand(Idx);
1626 Inst->setOperand(Idx, NewVal);
1629 /// \brief Restore the original value of the instruction.
1630 void undo() override {
1631 DEBUG(dbgs() << "Undo: setOperand:" << Idx << "\n"
1632 << "for: " << *Inst << "\n"
1633 << "with: " << *Origin << "\n");
1634 Inst->setOperand(Idx, Origin);
1638 /// \brief Hide the operands of an instruction.
1639 /// Do as if this instruction was not using any of its operands.
1640 class OperandsHider : public TypePromotionAction {
1641 /// The list of original operands.
1642 SmallVector<Value *, 4> OriginalValues;
1645 /// \brief Remove \p Inst from the uses of the operands of \p Inst.
1646 OperandsHider(Instruction *Inst) : TypePromotionAction(Inst) {
1647 DEBUG(dbgs() << "Do: OperandsHider: " << *Inst << "\n");
1648 unsigned NumOpnds = Inst->getNumOperands();
1649 OriginalValues.reserve(NumOpnds);
1650 for (unsigned It = 0; It < NumOpnds; ++It) {
1651 // Save the current operand.
1652 Value *Val = Inst->getOperand(It);
1653 OriginalValues.push_back(Val);
1655 // We could use OperandSetter here, but that would implied an overhead
1656 // that we are not willing to pay.
1657 Inst->setOperand(It, UndefValue::get(Val->getType()));
1661 /// \brief Restore the original list of uses.
1662 void undo() override {
1663 DEBUG(dbgs() << "Undo: OperandsHider: " << *Inst << "\n");
1664 for (unsigned It = 0, EndIt = OriginalValues.size(); It != EndIt; ++It)
1665 Inst->setOperand(It, OriginalValues[It]);
1669 /// \brief Build a truncate instruction.
1670 class TruncBuilder : public TypePromotionAction {
1673 /// \brief Build a truncate instruction of \p Opnd producing a \p Ty
1675 /// trunc Opnd to Ty.
1676 TruncBuilder(Instruction *Opnd, Type *Ty) : TypePromotionAction(Opnd) {
1677 IRBuilder<> Builder(Opnd);
1678 Val = Builder.CreateTrunc(Opnd, Ty, "promoted");
1679 DEBUG(dbgs() << "Do: TruncBuilder: " << *Val << "\n");
1682 /// \brief Get the built value.
1683 Value *getBuiltValue() { return Val; }
1685 /// \brief Remove the built instruction.
1686 void undo() override {
1687 DEBUG(dbgs() << "Undo: TruncBuilder: " << *Val << "\n");
1688 if (Instruction *IVal = dyn_cast<Instruction>(Val))
1689 IVal->eraseFromParent();
1693 /// \brief Build a sign extension instruction.
1694 class SExtBuilder : public TypePromotionAction {
1697 /// \brief Build a sign extension instruction of \p Opnd producing a \p Ty
1699 /// sext Opnd to Ty.
1700 SExtBuilder(Instruction *InsertPt, Value *Opnd, Type *Ty)
1701 : TypePromotionAction(InsertPt) {
1702 IRBuilder<> Builder(InsertPt);
1703 Val = Builder.CreateSExt(Opnd, Ty, "promoted");
1704 DEBUG(dbgs() << "Do: SExtBuilder: " << *Val << "\n");
1707 /// \brief Get the built value.
1708 Value *getBuiltValue() { return Val; }
1710 /// \brief Remove the built instruction.
1711 void undo() override {
1712 DEBUG(dbgs() << "Undo: SExtBuilder: " << *Val << "\n");
1713 if (Instruction *IVal = dyn_cast<Instruction>(Val))
1714 IVal->eraseFromParent();
1718 /// \brief Build a zero extension instruction.
1719 class ZExtBuilder : public TypePromotionAction {
1722 /// \brief Build a zero extension instruction of \p Opnd producing a \p Ty
1724 /// zext Opnd to Ty.
1725 ZExtBuilder(Instruction *InsertPt, Value *Opnd, Type *Ty)
1726 : TypePromotionAction(InsertPt) {
1727 IRBuilder<> Builder(InsertPt);
1728 Val = Builder.CreateZExt(Opnd, Ty, "promoted");
1729 DEBUG(dbgs() << "Do: ZExtBuilder: " << *Val << "\n");
1732 /// \brief Get the built value.
1733 Value *getBuiltValue() { return Val; }
1735 /// \brief Remove the built instruction.
1736 void undo() override {
1737 DEBUG(dbgs() << "Undo: ZExtBuilder: " << *Val << "\n");
1738 if (Instruction *IVal = dyn_cast<Instruction>(Val))
1739 IVal->eraseFromParent();
1743 /// \brief Mutate an instruction to another type.
1744 class TypeMutator : public TypePromotionAction {
1745 /// Record the original type.
1749 /// \brief Mutate the type of \p Inst into \p NewTy.
1750 TypeMutator(Instruction *Inst, Type *NewTy)
1751 : TypePromotionAction(Inst), OrigTy(Inst->getType()) {
1752 DEBUG(dbgs() << "Do: MutateType: " << *Inst << " with " << *NewTy
1754 Inst->mutateType(NewTy);
1757 /// \brief Mutate the instruction back to its original type.
1758 void undo() override {
1759 DEBUG(dbgs() << "Undo: MutateType: " << *Inst << " with " << *OrigTy
1761 Inst->mutateType(OrigTy);
1765 /// \brief Replace the uses of an instruction by another instruction.
1766 class UsesReplacer : public TypePromotionAction {
1767 /// Helper structure to keep track of the replaced uses.
1768 struct InstructionAndIdx {
1769 /// The instruction using the instruction.
1771 /// The index where this instruction is used for Inst.
1773 InstructionAndIdx(Instruction *Inst, unsigned Idx)
1774 : Inst(Inst), Idx(Idx) {}
1777 /// Keep track of the original uses (pair Instruction, Index).
1778 SmallVector<InstructionAndIdx, 4> OriginalUses;
1779 typedef SmallVectorImpl<InstructionAndIdx>::iterator use_iterator;
1782 /// \brief Replace all the use of \p Inst by \p New.
1783 UsesReplacer(Instruction *Inst, Value *New) : TypePromotionAction(Inst) {
1784 DEBUG(dbgs() << "Do: UsersReplacer: " << *Inst << " with " << *New
1786 // Record the original uses.
1787 for (Use &U : Inst->uses()) {
1788 Instruction *UserI = cast<Instruction>(U.getUser());
1789 OriginalUses.push_back(InstructionAndIdx(UserI, U.getOperandNo()));
1791 // Now, we can replace the uses.
1792 Inst->replaceAllUsesWith(New);
1795 /// \brief Reassign the original uses of Inst to Inst.
1796 void undo() override {
1797 DEBUG(dbgs() << "Undo: UsersReplacer: " << *Inst << "\n");
1798 for (use_iterator UseIt = OriginalUses.begin(),
1799 EndIt = OriginalUses.end();
1800 UseIt != EndIt; ++UseIt) {
1801 UseIt->Inst->setOperand(UseIt->Idx, Inst);
1806 /// \brief Remove an instruction from the IR.
1807 class InstructionRemover : public TypePromotionAction {
1808 /// Original position of the instruction.
1809 InsertionHandler Inserter;
1810 /// Helper structure to hide all the link to the instruction. In other
1811 /// words, this helps to do as if the instruction was removed.
1812 OperandsHider Hider;
1813 /// Keep track of the uses replaced, if any.
1814 UsesReplacer *Replacer;
1817 /// \brief Remove all reference of \p Inst and optinally replace all its
1819 /// \pre If !Inst->use_empty(), then New != nullptr
1820 InstructionRemover(Instruction *Inst, Value *New = nullptr)
1821 : TypePromotionAction(Inst), Inserter(Inst), Hider(Inst),
1824 Replacer = new UsesReplacer(Inst, New);
1825 DEBUG(dbgs() << "Do: InstructionRemover: " << *Inst << "\n");
1826 Inst->removeFromParent();
1829 ~InstructionRemover() { delete Replacer; }
1831 /// \brief Really remove the instruction.
1832 void commit() override { delete Inst; }
1834 /// \brief Resurrect the instruction and reassign it to the proper uses if
1835 /// new value was provided when build this action.
1836 void undo() override {
1837 DEBUG(dbgs() << "Undo: InstructionRemover: " << *Inst << "\n");
1838 Inserter.insert(Inst);
1846 /// Restoration point.
1847 /// The restoration point is a pointer to an action instead of an iterator
1848 /// because the iterator may be invalidated but not the pointer.
1849 typedef const TypePromotionAction *ConstRestorationPt;
1850 /// Advocate every changes made in that transaction.
1852 /// Undo all the changes made after the given point.
1853 void rollback(ConstRestorationPt Point);
1854 /// Get the current restoration point.
1855 ConstRestorationPt getRestorationPoint() const;
1857 /// \name API for IR modification with state keeping to support rollback.
1859 /// Same as Instruction::setOperand.
1860 void setOperand(Instruction *Inst, unsigned Idx, Value *NewVal);
1861 /// Same as Instruction::eraseFromParent.
1862 void eraseInstruction(Instruction *Inst, Value *NewVal = nullptr);
1863 /// Same as Value::replaceAllUsesWith.
1864 void replaceAllUsesWith(Instruction *Inst, Value *New);
1865 /// Same as Value::mutateType.
1866 void mutateType(Instruction *Inst, Type *NewTy);
1867 /// Same as IRBuilder::createTrunc.
1868 Value *createTrunc(Instruction *Opnd, Type *Ty);
1869 /// Same as IRBuilder::createSExt.
1870 Value *createSExt(Instruction *Inst, Value *Opnd, Type *Ty);
1871 /// Same as IRBuilder::createZExt.
1872 Value *createZExt(Instruction *Inst, Value *Opnd, Type *Ty);
1873 /// Same as Instruction::moveBefore.
1874 void moveBefore(Instruction *Inst, Instruction *Before);
1878 /// The ordered list of actions made so far.
1879 SmallVector<std::unique_ptr<TypePromotionAction>, 16> Actions;
1880 typedef SmallVectorImpl<std::unique_ptr<TypePromotionAction>>::iterator CommitPt;
1883 void TypePromotionTransaction::setOperand(Instruction *Inst, unsigned Idx,
1886 make_unique<TypePromotionTransaction::OperandSetter>(Inst, Idx, NewVal));
1889 void TypePromotionTransaction::eraseInstruction(Instruction *Inst,
1892 make_unique<TypePromotionTransaction::InstructionRemover>(Inst, NewVal));
1895 void TypePromotionTransaction::replaceAllUsesWith(Instruction *Inst,
1897 Actions.push_back(make_unique<TypePromotionTransaction::UsesReplacer>(Inst, New));
1900 void TypePromotionTransaction::mutateType(Instruction *Inst, Type *NewTy) {
1901 Actions.push_back(make_unique<TypePromotionTransaction::TypeMutator>(Inst, NewTy));
1904 Value *TypePromotionTransaction::createTrunc(Instruction *Opnd,
1906 std::unique_ptr<TruncBuilder> Ptr(new TruncBuilder(Opnd, Ty));
1907 Value *Val = Ptr->getBuiltValue();
1908 Actions.push_back(std::move(Ptr));
1912 Value *TypePromotionTransaction::createSExt(Instruction *Inst,
1913 Value *Opnd, Type *Ty) {
1914 std::unique_ptr<SExtBuilder> Ptr(new SExtBuilder(Inst, Opnd, Ty));
1915 Value *Val = Ptr->getBuiltValue();
1916 Actions.push_back(std::move(Ptr));
1920 Value *TypePromotionTransaction::createZExt(Instruction *Inst,
1921 Value *Opnd, Type *Ty) {
1922 std::unique_ptr<ZExtBuilder> Ptr(new ZExtBuilder(Inst, Opnd, Ty));
1923 Value *Val = Ptr->getBuiltValue();
1924 Actions.push_back(std::move(Ptr));
1928 void TypePromotionTransaction::moveBefore(Instruction *Inst,
1929 Instruction *Before) {
1931 make_unique<TypePromotionTransaction::InstructionMoveBefore>(Inst, Before));
1934 TypePromotionTransaction::ConstRestorationPt
1935 TypePromotionTransaction::getRestorationPoint() const {
1936 return !Actions.empty() ? Actions.back().get() : nullptr;
1939 void TypePromotionTransaction::commit() {
1940 for (CommitPt It = Actions.begin(), EndIt = Actions.end(); It != EndIt;
1946 void TypePromotionTransaction::rollback(
1947 TypePromotionTransaction::ConstRestorationPt Point) {
1948 while (!Actions.empty() && Point != Actions.back().get()) {
1949 std::unique_ptr<TypePromotionAction> Curr = Actions.pop_back_val();
1954 /// \brief A helper class for matching addressing modes.
1956 /// This encapsulates the logic for matching the target-legal addressing modes.
1957 class AddressingModeMatcher {
1958 SmallVectorImpl<Instruction*> &AddrModeInsts;
1959 const TargetLowering &TLI;
1961 /// AccessTy/MemoryInst - This is the type for the access (e.g. double) and
1962 /// the memory instruction that we're computing this address for.
1964 Instruction *MemoryInst;
1966 /// AddrMode - This is the addressing mode that we're building up. This is
1967 /// part of the return value of this addressing mode matching stuff.
1968 ExtAddrMode &AddrMode;
1970 /// The truncate instruction inserted by other CodeGenPrepare optimizations.
1971 const SetOfInstrs &InsertedTruncs;
1972 /// A map from the instructions to their type before promotion.
1973 InstrToOrigTy &PromotedInsts;
1974 /// The ongoing transaction where every action should be registered.
1975 TypePromotionTransaction &TPT;
1977 /// IgnoreProfitability - This is set to true when we should not do
1978 /// profitability checks. When true, IsProfitableToFoldIntoAddressingMode
1979 /// always returns true.
1980 bool IgnoreProfitability;
1982 AddressingModeMatcher(SmallVectorImpl<Instruction*> &AMI,
1983 const TargetLowering &T, Type *AT,
1984 Instruction *MI, ExtAddrMode &AM,
1985 const SetOfInstrs &InsertedTruncs,
1986 InstrToOrigTy &PromotedInsts,
1987 TypePromotionTransaction &TPT)
1988 : AddrModeInsts(AMI), TLI(T), AccessTy(AT), MemoryInst(MI), AddrMode(AM),
1989 InsertedTruncs(InsertedTruncs), PromotedInsts(PromotedInsts), TPT(TPT) {
1990 IgnoreProfitability = false;
1994 /// Match - Find the maximal addressing mode that a load/store of V can fold,
1995 /// give an access type of AccessTy. This returns a list of involved
1996 /// instructions in AddrModeInsts.
1997 /// \p InsertedTruncs The truncate instruction inserted by other
2000 /// \p PromotedInsts maps the instructions to their type before promotion.
2001 /// \p The ongoing transaction where every action should be registered.
2002 static ExtAddrMode Match(Value *V, Type *AccessTy,
2003 Instruction *MemoryInst,
2004 SmallVectorImpl<Instruction*> &AddrModeInsts,
2005 const TargetLowering &TLI,
2006 const SetOfInstrs &InsertedTruncs,
2007 InstrToOrigTy &PromotedInsts,
2008 TypePromotionTransaction &TPT) {
2011 bool Success = AddressingModeMatcher(AddrModeInsts, TLI, AccessTy,
2012 MemoryInst, Result, InsertedTruncs,
2013 PromotedInsts, TPT).MatchAddr(V, 0);
2014 (void)Success; assert(Success && "Couldn't select *anything*?");
2018 bool MatchScaledValue(Value *ScaleReg, int64_t Scale, unsigned Depth);
2019 bool MatchAddr(Value *V, unsigned Depth);
2020 bool MatchOperationAddr(User *Operation, unsigned Opcode, unsigned Depth,
2021 bool *MovedAway = nullptr);
2022 bool IsProfitableToFoldIntoAddressingMode(Instruction *I,
2023 ExtAddrMode &AMBefore,
2024 ExtAddrMode &AMAfter);
2025 bool ValueAlreadyLiveAtInst(Value *Val, Value *KnownLive1, Value *KnownLive2);
2026 bool IsPromotionProfitable(unsigned MatchedSize, unsigned SizeWithPromotion,
2027 Value *PromotedOperand) const;
2030 /// MatchScaledValue - Try adding ScaleReg*Scale to the current addressing mode.
2031 /// Return true and update AddrMode if this addr mode is legal for the target,
2033 bool AddressingModeMatcher::MatchScaledValue(Value *ScaleReg, int64_t Scale,
2035 // If Scale is 1, then this is the same as adding ScaleReg to the addressing
2036 // mode. Just process that directly.
2038 return MatchAddr(ScaleReg, Depth);
2040 // If the scale is 0, it takes nothing to add this.
2044 // If we already have a scale of this value, we can add to it, otherwise, we
2045 // need an available scale field.
2046 if (AddrMode.Scale != 0 && AddrMode.ScaledReg != ScaleReg)
2049 ExtAddrMode TestAddrMode = AddrMode;
2051 // Add scale to turn X*4+X*3 -> X*7. This could also do things like
2052 // [A+B + A*7] -> [B+A*8].
2053 TestAddrMode.Scale += Scale;
2054 TestAddrMode.ScaledReg = ScaleReg;
2056 // If the new address isn't legal, bail out.
2057 if (!TLI.isLegalAddressingMode(TestAddrMode, AccessTy))
2060 // It was legal, so commit it.
2061 AddrMode = TestAddrMode;
2063 // Okay, we decided that we can add ScaleReg+Scale to AddrMode. Check now
2064 // to see if ScaleReg is actually X+C. If so, we can turn this into adding
2065 // X*Scale + C*Scale to addr mode.
2066 ConstantInt *CI = nullptr; Value *AddLHS = nullptr;
2067 if (isa<Instruction>(ScaleReg) && // not a constant expr.
2068 match(ScaleReg, m_Add(m_Value(AddLHS), m_ConstantInt(CI)))) {
2069 TestAddrMode.ScaledReg = AddLHS;
2070 TestAddrMode.BaseOffs += CI->getSExtValue()*TestAddrMode.Scale;
2072 // If this addressing mode is legal, commit it and remember that we folded
2073 // this instruction.
2074 if (TLI.isLegalAddressingMode(TestAddrMode, AccessTy)) {
2075 AddrModeInsts.push_back(cast<Instruction>(ScaleReg));
2076 AddrMode = TestAddrMode;
2081 // Otherwise, not (x+c)*scale, just return what we have.
2085 /// MightBeFoldableInst - This is a little filter, which returns true if an
2086 /// addressing computation involving I might be folded into a load/store
2087 /// accessing it. This doesn't need to be perfect, but needs to accept at least
2088 /// the set of instructions that MatchOperationAddr can.
2089 static bool MightBeFoldableInst(Instruction *I) {
2090 switch (I->getOpcode()) {
2091 case Instruction::BitCast:
2092 case Instruction::AddrSpaceCast:
2093 // Don't touch identity bitcasts.
2094 if (I->getType() == I->getOperand(0)->getType())
2096 return I->getType()->isPointerTy() || I->getType()->isIntegerTy();
2097 case Instruction::PtrToInt:
2098 // PtrToInt is always a noop, as we know that the int type is pointer sized.
2100 case Instruction::IntToPtr:
2101 // We know the input is intptr_t, so this is foldable.
2103 case Instruction::Add:
2105 case Instruction::Mul:
2106 case Instruction::Shl:
2107 // Can only handle X*C and X << C.
2108 return isa<ConstantInt>(I->getOperand(1));
2109 case Instruction::GetElementPtr:
2116 /// \brief Check whether or not \p Val is a legal instruction for \p TLI.
2117 /// \note \p Val is assumed to be the product of some type promotion.
2118 /// Therefore if \p Val has an undefined state in \p TLI, this is assumed
2119 /// to be legal, as the non-promoted value would have had the same state.
2120 static bool isPromotedInstructionLegal(const TargetLowering &TLI, Value *Val) {
2121 Instruction *PromotedInst = dyn_cast<Instruction>(Val);
2124 int ISDOpcode = TLI.InstructionOpcodeToISD(PromotedInst->getOpcode());
2125 // If the ISDOpcode is undefined, it was undefined before the promotion.
2128 // Otherwise, check if the promoted instruction is legal or not.
2129 return TLI.isOperationLegalOrCustom(
2130 ISDOpcode, TLI.getValueType(PromotedInst->getType()));
2133 /// \brief Hepler class to perform type promotion.
2134 class TypePromotionHelper {
2135 /// \brief Utility function to check whether or not a sign or zero extension
2136 /// of \p Inst with \p ConsideredExtType can be moved through \p Inst by
2137 /// either using the operands of \p Inst or promoting \p Inst.
2138 /// The type of the extension is defined by \p IsSExt.
2139 /// In other words, check if:
2140 /// ext (Ty Inst opnd1 opnd2 ... opndN) to ConsideredExtType.
2141 /// #1 Promotion applies:
2142 /// ConsideredExtType Inst (ext opnd1 to ConsideredExtType, ...).
2143 /// #2 Operand reuses:
2144 /// ext opnd1 to ConsideredExtType.
2145 /// \p PromotedInsts maps the instructions to their type before promotion.
2146 static bool canGetThrough(const Instruction *Inst, Type *ConsideredExtType,
2147 const InstrToOrigTy &PromotedInsts, bool IsSExt);
2149 /// \brief Utility function to determine if \p OpIdx should be promoted when
2150 /// promoting \p Inst.
2151 static bool shouldExtOperand(const Instruction *Inst, int OpIdx) {
2152 if (isa<SelectInst>(Inst) && OpIdx == 0)
2157 /// \brief Utility function to promote the operand of \p Ext when this
2158 /// operand is a promotable trunc or sext or zext.
2159 /// \p PromotedInsts maps the instructions to their type before promotion.
2160 /// \p CreatedInsts[out] contains how many non-free instructions have been
2161 /// created to promote the operand of Ext.
2162 /// Newly added extensions are inserted in \p Exts.
2163 /// Newly added truncates are inserted in \p Truncs.
2164 /// Should never be called directly.
2165 /// \return The promoted value which is used instead of Ext.
2166 static Value *promoteOperandForTruncAndAnyExt(
2167 Instruction *Ext, TypePromotionTransaction &TPT,
2168 InstrToOrigTy &PromotedInsts, unsigned &CreatedInsts,
2169 SmallVectorImpl<Instruction *> *Exts,
2170 SmallVectorImpl<Instruction *> *Truncs);
2172 /// \brief Utility function to promote the operand of \p Ext when this
2173 /// operand is promotable and is not a supported trunc or sext.
2174 /// \p PromotedInsts maps the instructions to their type before promotion.
2175 /// \p CreatedInsts[out] contains how many non-free instructions have been
2176 /// created to promote the operand of Ext.
2177 /// Newly added extensions are inserted in \p Exts.
2178 /// Newly added truncates are inserted in \p Truncs.
2179 /// Should never be called directly.
2180 /// \return The promoted value which is used instead of Ext.
2182 promoteOperandForOther(Instruction *Ext, TypePromotionTransaction &TPT,
2183 InstrToOrigTy &PromotedInsts, unsigned &CreatedInsts,
2184 SmallVectorImpl<Instruction *> *Exts,
2185 SmallVectorImpl<Instruction *> *Truncs, bool IsSExt);
2187 /// \see promoteOperandForOther.
2189 signExtendOperandForOther(Instruction *Ext, TypePromotionTransaction &TPT,
2190 InstrToOrigTy &PromotedInsts,
2191 unsigned &CreatedInsts,
2192 SmallVectorImpl<Instruction *> *Exts,
2193 SmallVectorImpl<Instruction *> *Truncs) {
2194 return promoteOperandForOther(Ext, TPT, PromotedInsts, CreatedInsts, Exts,
2198 /// \see promoteOperandForOther.
2200 zeroExtendOperandForOther(Instruction *Ext, TypePromotionTransaction &TPT,
2201 InstrToOrigTy &PromotedInsts,
2202 unsigned &CreatedInsts,
2203 SmallVectorImpl<Instruction *> *Exts,
2204 SmallVectorImpl<Instruction *> *Truncs) {
2205 return promoteOperandForOther(Ext, TPT, PromotedInsts, CreatedInsts, Exts,
2210 /// Type for the utility function that promotes the operand of Ext.
2211 typedef Value *(*Action)(Instruction *Ext, TypePromotionTransaction &TPT,
2212 InstrToOrigTy &PromotedInsts, unsigned &CreatedInsts,
2213 SmallVectorImpl<Instruction *> *Exts,
2214 SmallVectorImpl<Instruction *> *Truncs);
2215 /// \brief Given a sign/zero extend instruction \p Ext, return the approriate
2216 /// action to promote the operand of \p Ext instead of using Ext.
2217 /// \return NULL if no promotable action is possible with the current
2219 /// \p InsertedTruncs keeps track of all the truncate instructions inserted by
2220 /// the others CodeGenPrepare optimizations. This information is important
2221 /// because we do not want to promote these instructions as CodeGenPrepare
2222 /// will reinsert them later. Thus creating an infinite loop: create/remove.
2223 /// \p PromotedInsts maps the instructions to their type before promotion.
2224 static Action getAction(Instruction *Ext, const SetOfInstrs &InsertedTruncs,
2225 const TargetLowering &TLI,
2226 const InstrToOrigTy &PromotedInsts);
2229 bool TypePromotionHelper::canGetThrough(const Instruction *Inst,
2230 Type *ConsideredExtType,
2231 const InstrToOrigTy &PromotedInsts,
2233 // The promotion helper does not know how to deal with vector types yet.
2234 // To be able to fix that, we would need to fix the places where we
2235 // statically extend, e.g., constants and such.
2236 if (Inst->getType()->isVectorTy())
2239 // We can always get through zext.
2240 if (isa<ZExtInst>(Inst))
2243 // sext(sext) is ok too.
2244 if (IsSExt && isa<SExtInst>(Inst))
2247 // We can get through binary operator, if it is legal. In other words, the
2248 // binary operator must have a nuw or nsw flag.
2249 const BinaryOperator *BinOp = dyn_cast<BinaryOperator>(Inst);
2250 if (BinOp && isa<OverflowingBinaryOperator>(BinOp) &&
2251 ((!IsSExt && BinOp->hasNoUnsignedWrap()) ||
2252 (IsSExt && BinOp->hasNoSignedWrap())))
2255 // Check if we can do the following simplification.
2256 // ext(trunc(opnd)) --> ext(opnd)
2257 if (!isa<TruncInst>(Inst))
2260 Value *OpndVal = Inst->getOperand(0);
2261 // Check if we can use this operand in the extension.
2262 // If the type is larger than the result type of the extension,
2264 if (!OpndVal->getType()->isIntegerTy() ||
2265 OpndVal->getType()->getIntegerBitWidth() >
2266 ConsideredExtType->getIntegerBitWidth())
2269 // If the operand of the truncate is not an instruction, we will not have
2270 // any information on the dropped bits.
2271 // (Actually we could for constant but it is not worth the extra logic).
2272 Instruction *Opnd = dyn_cast<Instruction>(OpndVal);
2276 // Check if the source of the type is narrow enough.
2277 // I.e., check that trunc just drops extended bits of the same kind of
2279 // #1 get the type of the operand and check the kind of the extended bits.
2280 const Type *OpndType;
2281 InstrToOrigTy::const_iterator It = PromotedInsts.find(Opnd);
2282 if (It != PromotedInsts.end() && It->second.IsSExt == IsSExt)
2283 OpndType = It->second.Ty;
2284 else if ((IsSExt && isa<SExtInst>(Opnd)) || (!IsSExt && isa<ZExtInst>(Opnd)))
2285 OpndType = Opnd->getOperand(0)->getType();
2289 // #2 check that the truncate just drop extended bits.
2290 if (Inst->getType()->getIntegerBitWidth() >= OpndType->getIntegerBitWidth())
2296 TypePromotionHelper::Action TypePromotionHelper::getAction(
2297 Instruction *Ext, const SetOfInstrs &InsertedTruncs,
2298 const TargetLowering &TLI, const InstrToOrigTy &PromotedInsts) {
2299 assert((isa<SExtInst>(Ext) || isa<ZExtInst>(Ext)) &&
2300 "Unexpected instruction type");
2301 Instruction *ExtOpnd = dyn_cast<Instruction>(Ext->getOperand(0));
2302 Type *ExtTy = Ext->getType();
2303 bool IsSExt = isa<SExtInst>(Ext);
2304 // If the operand of the extension is not an instruction, we cannot
2306 // If it, check we can get through.
2307 if (!ExtOpnd || !canGetThrough(ExtOpnd, ExtTy, PromotedInsts, IsSExt))
2310 // Do not promote if the operand has been added by codegenprepare.
2311 // Otherwise, it means we are undoing an optimization that is likely to be
2312 // redone, thus causing potential infinite loop.
2313 if (isa<TruncInst>(ExtOpnd) && InsertedTruncs.count(ExtOpnd))
2316 // SExt or Trunc instructions.
2317 // Return the related handler.
2318 if (isa<SExtInst>(ExtOpnd) || isa<TruncInst>(ExtOpnd) ||
2319 isa<ZExtInst>(ExtOpnd))
2320 return promoteOperandForTruncAndAnyExt;
2322 // Regular instruction.
2323 // Abort early if we will have to insert non-free instructions.
2324 if (!ExtOpnd->hasOneUse() && !TLI.isTruncateFree(ExtTy, ExtOpnd->getType()))
2326 return IsSExt ? signExtendOperandForOther : zeroExtendOperandForOther;
2329 Value *TypePromotionHelper::promoteOperandForTruncAndAnyExt(
2330 llvm::Instruction *SExt, TypePromotionTransaction &TPT,
2331 InstrToOrigTy &PromotedInsts, unsigned &CreatedInsts,
2332 SmallVectorImpl<Instruction *> *Exts,
2333 SmallVectorImpl<Instruction *> *Truncs) {
2334 // By construction, the operand of SExt is an instruction. Otherwise we cannot
2335 // get through it and this method should not be called.
2336 Instruction *SExtOpnd = cast<Instruction>(SExt->getOperand(0));
2337 Value *ExtVal = SExt;
2338 if (isa<ZExtInst>(SExtOpnd)) {
2339 // Replace s|zext(zext(opnd))
2342 TPT.createZExt(SExt, SExtOpnd->getOperand(0), SExt->getType());
2343 TPT.replaceAllUsesWith(SExt, ZExt);
2344 TPT.eraseInstruction(SExt);
2347 // Replace z|sext(trunc(opnd)) or sext(sext(opnd))
2349 TPT.setOperand(SExt, 0, SExtOpnd->getOperand(0));
2353 // Remove dead code.
2354 if (SExtOpnd->use_empty())
2355 TPT.eraseInstruction(SExtOpnd);
2357 // Check if the extension is still needed.
2358 Instruction *ExtInst = dyn_cast<Instruction>(ExtVal);
2359 if (!ExtInst || ExtInst->getType() != ExtInst->getOperand(0)->getType()) {
2360 if (ExtInst && Exts)
2361 Exts->push_back(ExtInst);
2365 // At this point we have: ext ty opnd to ty.
2366 // Reassign the uses of ExtInst to the opnd and remove ExtInst.
2367 Value *NextVal = ExtInst->getOperand(0);
2368 TPT.eraseInstruction(ExtInst, NextVal);
2372 Value *TypePromotionHelper::promoteOperandForOther(
2373 Instruction *Ext, TypePromotionTransaction &TPT,
2374 InstrToOrigTy &PromotedInsts, unsigned &CreatedInsts,
2375 SmallVectorImpl<Instruction *> *Exts,
2376 SmallVectorImpl<Instruction *> *Truncs, bool IsSExt) {
2377 // By construction, the operand of Ext is an instruction. Otherwise we cannot
2378 // get through it and this method should not be called.
2379 Instruction *ExtOpnd = cast<Instruction>(Ext->getOperand(0));
2381 if (!ExtOpnd->hasOneUse()) {
2382 // ExtOpnd will be promoted.
2383 // All its uses, but Ext, will need to use a truncated value of the
2384 // promoted version.
2385 // Create the truncate now.
2386 Value *Trunc = TPT.createTrunc(Ext, ExtOpnd->getType());
2387 if (Instruction *ITrunc = dyn_cast<Instruction>(Trunc)) {
2388 ITrunc->removeFromParent();
2389 // Insert it just after the definition.
2390 ITrunc->insertAfter(ExtOpnd);
2392 Truncs->push_back(ITrunc);
2395 TPT.replaceAllUsesWith(ExtOpnd, Trunc);
2396 // Restore the operand of Ext (which has been replace by the previous call
2397 // to replaceAllUsesWith) to avoid creating a cycle trunc <-> sext.
2398 TPT.setOperand(Ext, 0, ExtOpnd);
2401 // Get through the Instruction:
2402 // 1. Update its type.
2403 // 2. Replace the uses of Ext by Inst.
2404 // 3. Extend each operand that needs to be extended.
2406 // Remember the original type of the instruction before promotion.
2407 // This is useful to know that the high bits are sign extended bits.
2408 PromotedInsts.insert(std::pair<Instruction *, TypeIsSExt>(
2409 ExtOpnd, TypeIsSExt(ExtOpnd->getType(), IsSExt)));
2411 TPT.mutateType(ExtOpnd, Ext->getType());
2413 TPT.replaceAllUsesWith(Ext, ExtOpnd);
2415 Instruction *ExtForOpnd = Ext;
2417 DEBUG(dbgs() << "Propagate Ext to operands\n");
2418 for (int OpIdx = 0, EndOpIdx = ExtOpnd->getNumOperands(); OpIdx != EndOpIdx;
2420 DEBUG(dbgs() << "Operand:\n" << *(ExtOpnd->getOperand(OpIdx)) << '\n');
2421 if (ExtOpnd->getOperand(OpIdx)->getType() == Ext->getType() ||
2422 !shouldExtOperand(ExtOpnd, OpIdx)) {
2423 DEBUG(dbgs() << "No need to propagate\n");
2426 // Check if we can statically extend the operand.
2427 Value *Opnd = ExtOpnd->getOperand(OpIdx);
2428 if (const ConstantInt *Cst = dyn_cast<ConstantInt>(Opnd)) {
2429 DEBUG(dbgs() << "Statically extend\n");
2430 unsigned BitWidth = Ext->getType()->getIntegerBitWidth();
2431 APInt CstVal = IsSExt ? Cst->getValue().sext(BitWidth)
2432 : Cst->getValue().zext(BitWidth);
2433 TPT.setOperand(ExtOpnd, OpIdx, ConstantInt::get(Ext->getType(), CstVal));
2436 // UndefValue are typed, so we have to statically sign extend them.
2437 if (isa<UndefValue>(Opnd)) {
2438 DEBUG(dbgs() << "Statically extend\n");
2439 TPT.setOperand(ExtOpnd, OpIdx, UndefValue::get(Ext->getType()));
2443 // Otherwise we have to explicity sign extend the operand.
2444 // Check if Ext was reused to extend an operand.
2446 // If yes, create a new one.
2447 DEBUG(dbgs() << "More operands to ext\n");
2448 Value *ValForExtOpnd = IsSExt ? TPT.createSExt(Ext, Opnd, Ext->getType())
2449 : TPT.createZExt(Ext, Opnd, Ext->getType());
2450 if (!isa<Instruction>(ValForExtOpnd)) {
2451 TPT.setOperand(ExtOpnd, OpIdx, ValForExtOpnd);
2454 ExtForOpnd = cast<Instruction>(ValForExtOpnd);
2458 Exts->push_back(ExtForOpnd);
2459 TPT.setOperand(ExtForOpnd, 0, Opnd);
2461 // Move the sign extension before the insertion point.
2462 TPT.moveBefore(ExtForOpnd, ExtOpnd);
2463 TPT.setOperand(ExtOpnd, OpIdx, ExtForOpnd);
2464 // If more sext are required, new instructions will have to be created.
2465 ExtForOpnd = nullptr;
2467 if (ExtForOpnd == Ext) {
2468 DEBUG(dbgs() << "Extension is useless now\n");
2469 TPT.eraseInstruction(Ext);
2474 /// IsPromotionProfitable - Check whether or not promoting an instruction
2475 /// to a wider type was profitable.
2476 /// \p MatchedSize gives the number of instructions that have been matched
2477 /// in the addressing mode after the promotion was applied.
2478 /// \p SizeWithPromotion gives the number of created instructions for
2479 /// the promotion plus the number of instructions that have been
2480 /// matched in the addressing mode before the promotion.
2481 /// \p PromotedOperand is the value that has been promoted.
2482 /// \return True if the promotion is profitable, false otherwise.
2484 AddressingModeMatcher::IsPromotionProfitable(unsigned MatchedSize,
2485 unsigned SizeWithPromotion,
2486 Value *PromotedOperand) const {
2487 // We folded less instructions than what we created to promote the operand.
2488 // This is not profitable.
2489 if (MatchedSize < SizeWithPromotion)
2491 if (MatchedSize > SizeWithPromotion)
2493 // The promotion is neutral but it may help folding the sign extension in
2494 // loads for instance.
2495 // Check that we did not create an illegal instruction.
2496 return isPromotedInstructionLegal(TLI, PromotedOperand);
2499 /// MatchOperationAddr - Given an instruction or constant expr, see if we can
2500 /// fold the operation into the addressing mode. If so, update the addressing
2501 /// mode and return true, otherwise return false without modifying AddrMode.
2502 /// If \p MovedAway is not NULL, it contains the information of whether or
2503 /// not AddrInst has to be folded into the addressing mode on success.
2504 /// If \p MovedAway == true, \p AddrInst will not be part of the addressing
2505 /// because it has been moved away.
2506 /// Thus AddrInst must not be added in the matched instructions.
2507 /// This state can happen when AddrInst is a sext, since it may be moved away.
2508 /// Therefore, AddrInst may not be valid when MovedAway is true and it must
2509 /// not be referenced anymore.
2510 bool AddressingModeMatcher::MatchOperationAddr(User *AddrInst, unsigned Opcode,
2513 // Avoid exponential behavior on extremely deep expression trees.
2514 if (Depth >= 5) return false;
2516 // By default, all matched instructions stay in place.
2521 case Instruction::PtrToInt:
2522 // PtrToInt is always a noop, as we know that the int type is pointer sized.
2523 return MatchAddr(AddrInst->getOperand(0), Depth);
2524 case Instruction::IntToPtr:
2525 // This inttoptr is a no-op if the integer type is pointer sized.
2526 if (TLI.getValueType(AddrInst->getOperand(0)->getType()) ==
2527 TLI.getPointerTy(AddrInst->getType()->getPointerAddressSpace()))
2528 return MatchAddr(AddrInst->getOperand(0), Depth);
2530 case Instruction::BitCast:
2531 case Instruction::AddrSpaceCast:
2532 // BitCast is always a noop, and we can handle it as long as it is
2533 // int->int or pointer->pointer (we don't want int<->fp or something).
2534 if ((AddrInst->getOperand(0)->getType()->isPointerTy() ||
2535 AddrInst->getOperand(0)->getType()->isIntegerTy()) &&
2536 // Don't touch identity bitcasts. These were probably put here by LSR,
2537 // and we don't want to mess around with them. Assume it knows what it
2539 AddrInst->getOperand(0)->getType() != AddrInst->getType())
2540 return MatchAddr(AddrInst->getOperand(0), Depth);
2542 case Instruction::Add: {
2543 // Check to see if we can merge in the RHS then the LHS. If so, we win.
2544 ExtAddrMode BackupAddrMode = AddrMode;
2545 unsigned OldSize = AddrModeInsts.size();
2546 // Start a transaction at this point.
2547 // The LHS may match but not the RHS.
2548 // Therefore, we need a higher level restoration point to undo partially
2549 // matched operation.
2550 TypePromotionTransaction::ConstRestorationPt LastKnownGood =
2551 TPT.getRestorationPoint();
2553 if (MatchAddr(AddrInst->getOperand(1), Depth+1) &&
2554 MatchAddr(AddrInst->getOperand(0), Depth+1))
2557 // Restore the old addr mode info.
2558 AddrMode = BackupAddrMode;
2559 AddrModeInsts.resize(OldSize);
2560 TPT.rollback(LastKnownGood);
2562 // Otherwise this was over-aggressive. Try merging in the LHS then the RHS.
2563 if (MatchAddr(AddrInst->getOperand(0), Depth+1) &&
2564 MatchAddr(AddrInst->getOperand(1), Depth+1))
2567 // Otherwise we definitely can't merge the ADD in.
2568 AddrMode = BackupAddrMode;
2569 AddrModeInsts.resize(OldSize);
2570 TPT.rollback(LastKnownGood);
2573 //case Instruction::Or:
2574 // TODO: We can handle "Or Val, Imm" iff this OR is equivalent to an ADD.
2576 case Instruction::Mul:
2577 case Instruction::Shl: {
2578 // Can only handle X*C and X << C.
2579 ConstantInt *RHS = dyn_cast<ConstantInt>(AddrInst->getOperand(1));
2582 int64_t Scale = RHS->getSExtValue();
2583 if (Opcode == Instruction::Shl)
2584 Scale = 1LL << Scale;
2586 return MatchScaledValue(AddrInst->getOperand(0), Scale, Depth);
2588 case Instruction::GetElementPtr: {
2589 // Scan the GEP. We check it if it contains constant offsets and at most
2590 // one variable offset.
2591 int VariableOperand = -1;
2592 unsigned VariableScale = 0;
2594 int64_t ConstantOffset = 0;
2595 const DataLayout *TD = TLI.getDataLayout();
2596 gep_type_iterator GTI = gep_type_begin(AddrInst);
2597 for (unsigned i = 1, e = AddrInst->getNumOperands(); i != e; ++i, ++GTI) {
2598 if (StructType *STy = dyn_cast<StructType>(*GTI)) {
2599 const StructLayout *SL = TD->getStructLayout(STy);
2601 cast<ConstantInt>(AddrInst->getOperand(i))->getZExtValue();
2602 ConstantOffset += SL->getElementOffset(Idx);
2604 uint64_t TypeSize = TD->getTypeAllocSize(GTI.getIndexedType());
2605 if (ConstantInt *CI = dyn_cast<ConstantInt>(AddrInst->getOperand(i))) {
2606 ConstantOffset += CI->getSExtValue()*TypeSize;
2607 } else if (TypeSize) { // Scales of zero don't do anything.
2608 // We only allow one variable index at the moment.
2609 if (VariableOperand != -1)
2612 // Remember the variable index.
2613 VariableOperand = i;
2614 VariableScale = TypeSize;
2619 // A common case is for the GEP to only do a constant offset. In this case,
2620 // just add it to the disp field and check validity.
2621 if (VariableOperand == -1) {
2622 AddrMode.BaseOffs += ConstantOffset;
2623 if (ConstantOffset == 0 || TLI.isLegalAddressingMode(AddrMode, AccessTy)){
2624 // Check to see if we can fold the base pointer in too.
2625 if (MatchAddr(AddrInst->getOperand(0), Depth+1))
2628 AddrMode.BaseOffs -= ConstantOffset;
2632 // Save the valid addressing mode in case we can't match.
2633 ExtAddrMode BackupAddrMode = AddrMode;
2634 unsigned OldSize = AddrModeInsts.size();
2636 // See if the scale and offset amount is valid for this target.
2637 AddrMode.BaseOffs += ConstantOffset;
2639 // Match the base operand of the GEP.
2640 if (!MatchAddr(AddrInst->getOperand(0), Depth+1)) {
2641 // If it couldn't be matched, just stuff the value in a register.
2642 if (AddrMode.HasBaseReg) {
2643 AddrMode = BackupAddrMode;
2644 AddrModeInsts.resize(OldSize);
2647 AddrMode.HasBaseReg = true;
2648 AddrMode.BaseReg = AddrInst->getOperand(0);
2651 // Match the remaining variable portion of the GEP.
2652 if (!MatchScaledValue(AddrInst->getOperand(VariableOperand), VariableScale,
2654 // If it couldn't be matched, try stuffing the base into a register
2655 // instead of matching it, and retrying the match of the scale.
2656 AddrMode = BackupAddrMode;
2657 AddrModeInsts.resize(OldSize);
2658 if (AddrMode.HasBaseReg)
2660 AddrMode.HasBaseReg = true;
2661 AddrMode.BaseReg = AddrInst->getOperand(0);
2662 AddrMode.BaseOffs += ConstantOffset;
2663 if (!MatchScaledValue(AddrInst->getOperand(VariableOperand),
2664 VariableScale, Depth)) {
2665 // If even that didn't work, bail.
2666 AddrMode = BackupAddrMode;
2667 AddrModeInsts.resize(OldSize);
2674 case Instruction::SExt:
2675 case Instruction::ZExt: {
2676 Instruction *Ext = dyn_cast<Instruction>(AddrInst);
2680 // Try to move this ext out of the way of the addressing mode.
2681 // Ask for a method for doing so.
2682 TypePromotionHelper::Action TPH =
2683 TypePromotionHelper::getAction(Ext, InsertedTruncs, TLI, PromotedInsts);
2687 TypePromotionTransaction::ConstRestorationPt LastKnownGood =
2688 TPT.getRestorationPoint();
2689 unsigned CreatedInsts = 0;
2690 Value *PromotedOperand =
2691 TPH(Ext, TPT, PromotedInsts, CreatedInsts, nullptr, nullptr);
2692 // SExt has been moved away.
2693 // Thus either it will be rematched later in the recursive calls or it is
2694 // gone. Anyway, we must not fold it into the addressing mode at this point.
2698 // addr = gep base, idx
2700 // promotedOpnd = ext opnd <- no match here
2701 // op = promoted_add promotedOpnd, 1 <- match (later in recursive calls)
2702 // addr = gep base, op <- match
2706 assert(PromotedOperand &&
2707 "TypePromotionHelper should have filtered out those cases");
2709 ExtAddrMode BackupAddrMode = AddrMode;
2710 unsigned OldSize = AddrModeInsts.size();
2712 if (!MatchAddr(PromotedOperand, Depth) ||
2713 !IsPromotionProfitable(AddrModeInsts.size(), OldSize + CreatedInsts,
2715 AddrMode = BackupAddrMode;
2716 AddrModeInsts.resize(OldSize);
2717 DEBUG(dbgs() << "Sign extension does not pay off: rollback\n");
2718 TPT.rollback(LastKnownGood);
2727 /// MatchAddr - If we can, try to add the value of 'Addr' into the current
2728 /// addressing mode. If Addr can't be added to AddrMode this returns false and
2729 /// leaves AddrMode unmodified. This assumes that Addr is either a pointer type
2730 /// or intptr_t for the target.
2732 bool AddressingModeMatcher::MatchAddr(Value *Addr, unsigned Depth) {
2733 // Start a transaction at this point that we will rollback if the matching
2735 TypePromotionTransaction::ConstRestorationPt LastKnownGood =
2736 TPT.getRestorationPoint();
2737 if (ConstantInt *CI = dyn_cast<ConstantInt>(Addr)) {
2738 // Fold in immediates if legal for the target.
2739 AddrMode.BaseOffs += CI->getSExtValue();
2740 if (TLI.isLegalAddressingMode(AddrMode, AccessTy))
2742 AddrMode.BaseOffs -= CI->getSExtValue();
2743 } else if (GlobalValue *GV = dyn_cast<GlobalValue>(Addr)) {
2744 // If this is a global variable, try to fold it into the addressing mode.
2745 if (!AddrMode.BaseGV) {
2746 AddrMode.BaseGV = GV;
2747 if (TLI.isLegalAddressingMode(AddrMode, AccessTy))
2749 AddrMode.BaseGV = nullptr;
2751 } else if (Instruction *I = dyn_cast<Instruction>(Addr)) {
2752 ExtAddrMode BackupAddrMode = AddrMode;
2753 unsigned OldSize = AddrModeInsts.size();
2755 // Check to see if it is possible to fold this operation.
2756 bool MovedAway = false;
2757 if (MatchOperationAddr(I, I->getOpcode(), Depth, &MovedAway)) {
2758 // This instruction may have been move away. If so, there is nothing
2762 // Okay, it's possible to fold this. Check to see if it is actually
2763 // *profitable* to do so. We use a simple cost model to avoid increasing
2764 // register pressure too much.
2765 if (I->hasOneUse() ||
2766 IsProfitableToFoldIntoAddressingMode(I, BackupAddrMode, AddrMode)) {
2767 AddrModeInsts.push_back(I);
2771 // It isn't profitable to do this, roll back.
2772 //cerr << "NOT FOLDING: " << *I;
2773 AddrMode = BackupAddrMode;
2774 AddrModeInsts.resize(OldSize);
2775 TPT.rollback(LastKnownGood);
2777 } else if (ConstantExpr *CE = dyn_cast<ConstantExpr>(Addr)) {
2778 if (MatchOperationAddr(CE, CE->getOpcode(), Depth))
2780 TPT.rollback(LastKnownGood);
2781 } else if (isa<ConstantPointerNull>(Addr)) {
2782 // Null pointer gets folded without affecting the addressing mode.
2786 // Worse case, the target should support [reg] addressing modes. :)
2787 if (!AddrMode.HasBaseReg) {
2788 AddrMode.HasBaseReg = true;
2789 AddrMode.BaseReg = Addr;
2790 // Still check for legality in case the target supports [imm] but not [i+r].
2791 if (TLI.isLegalAddressingMode(AddrMode, AccessTy))
2793 AddrMode.HasBaseReg = false;
2794 AddrMode.BaseReg = nullptr;
2797 // If the base register is already taken, see if we can do [r+r].
2798 if (AddrMode.Scale == 0) {
2800 AddrMode.ScaledReg = Addr;
2801 if (TLI.isLegalAddressingMode(AddrMode, AccessTy))
2804 AddrMode.ScaledReg = nullptr;
2807 TPT.rollback(LastKnownGood);
2811 /// IsOperandAMemoryOperand - Check to see if all uses of OpVal by the specified
2812 /// inline asm call are due to memory operands. If so, return true, otherwise
2814 static bool IsOperandAMemoryOperand(CallInst *CI, InlineAsm *IA, Value *OpVal,
2815 const TargetLowering &TLI) {
2816 TargetLowering::AsmOperandInfoVector TargetConstraints = TLI.ParseConstraints(ImmutableCallSite(CI));
2817 for (unsigned i = 0, e = TargetConstraints.size(); i != e; ++i) {
2818 TargetLowering::AsmOperandInfo &OpInfo = TargetConstraints[i];
2820 // Compute the constraint code and ConstraintType to use.
2821 TLI.ComputeConstraintToUse(OpInfo, SDValue());
2823 // If this asm operand is our Value*, and if it isn't an indirect memory
2824 // operand, we can't fold it!
2825 if (OpInfo.CallOperandVal == OpVal &&
2826 (OpInfo.ConstraintType != TargetLowering::C_Memory ||
2827 !OpInfo.isIndirect))
2834 /// FindAllMemoryUses - Recursively walk all the uses of I until we find a
2835 /// memory use. If we find an obviously non-foldable instruction, return true.
2836 /// Add the ultimately found memory instructions to MemoryUses.
2837 static bool FindAllMemoryUses(Instruction *I,
2838 SmallVectorImpl<std::pair<Instruction*,unsigned> > &MemoryUses,
2839 SmallPtrSetImpl<Instruction*> &ConsideredInsts,
2840 const TargetLowering &TLI) {
2841 // If we already considered this instruction, we're done.
2842 if (!ConsideredInsts.insert(I).second)
2845 // If this is an obviously unfoldable instruction, bail out.
2846 if (!MightBeFoldableInst(I))
2849 // Loop over all the uses, recursively processing them.
2850 for (Use &U : I->uses()) {
2851 Instruction *UserI = cast<Instruction>(U.getUser());
2853 if (LoadInst *LI = dyn_cast<LoadInst>(UserI)) {
2854 MemoryUses.push_back(std::make_pair(LI, U.getOperandNo()));
2858 if (StoreInst *SI = dyn_cast<StoreInst>(UserI)) {
2859 unsigned opNo = U.getOperandNo();
2860 if (opNo == 0) return true; // Storing addr, not into addr.
2861 MemoryUses.push_back(std::make_pair(SI, opNo));
2865 if (CallInst *CI = dyn_cast<CallInst>(UserI)) {
2866 InlineAsm *IA = dyn_cast<InlineAsm>(CI->getCalledValue());
2867 if (!IA) return true;
2869 // If this is a memory operand, we're cool, otherwise bail out.
2870 if (!IsOperandAMemoryOperand(CI, IA, I, TLI))
2875 if (FindAllMemoryUses(UserI, MemoryUses, ConsideredInsts, TLI))
2882 /// ValueAlreadyLiveAtInst - Retrn true if Val is already known to be live at
2883 /// the use site that we're folding it into. If so, there is no cost to
2884 /// include it in the addressing mode. KnownLive1 and KnownLive2 are two values
2885 /// that we know are live at the instruction already.
2886 bool AddressingModeMatcher::ValueAlreadyLiveAtInst(Value *Val,Value *KnownLive1,
2887 Value *KnownLive2) {
2888 // If Val is either of the known-live values, we know it is live!
2889 if (Val == nullptr || Val == KnownLive1 || Val == KnownLive2)
2892 // All values other than instructions and arguments (e.g. constants) are live.
2893 if (!isa<Instruction>(Val) && !isa<Argument>(Val)) return true;
2895 // If Val is a constant sized alloca in the entry block, it is live, this is
2896 // true because it is just a reference to the stack/frame pointer, which is
2897 // live for the whole function.
2898 if (AllocaInst *AI = dyn_cast<AllocaInst>(Val))
2899 if (AI->isStaticAlloca())
2902 // Check to see if this value is already used in the memory instruction's
2903 // block. If so, it's already live into the block at the very least, so we
2904 // can reasonably fold it.
2905 return Val->isUsedInBasicBlock(MemoryInst->getParent());
2908 /// IsProfitableToFoldIntoAddressingMode - It is possible for the addressing
2909 /// mode of the machine to fold the specified instruction into a load or store
2910 /// that ultimately uses it. However, the specified instruction has multiple
2911 /// uses. Given this, it may actually increase register pressure to fold it
2912 /// into the load. For example, consider this code:
2916 /// use(Y) -> nonload/store
2920 /// In this case, Y has multiple uses, and can be folded into the load of Z
2921 /// (yielding load [X+2]). However, doing this will cause both "X" and "X+1" to
2922 /// be live at the use(Y) line. If we don't fold Y into load Z, we use one
2923 /// fewer register. Since Y can't be folded into "use(Y)" we don't increase the
2924 /// number of computations either.
2926 /// Note that this (like most of CodeGenPrepare) is just a rough heuristic. If
2927 /// X was live across 'load Z' for other reasons, we actually *would* want to
2928 /// fold the addressing mode in the Z case. This would make Y die earlier.
2929 bool AddressingModeMatcher::
2930 IsProfitableToFoldIntoAddressingMode(Instruction *I, ExtAddrMode &AMBefore,
2931 ExtAddrMode &AMAfter) {
2932 if (IgnoreProfitability) return true;
2934 // AMBefore is the addressing mode before this instruction was folded into it,
2935 // and AMAfter is the addressing mode after the instruction was folded. Get
2936 // the set of registers referenced by AMAfter and subtract out those
2937 // referenced by AMBefore: this is the set of values which folding in this
2938 // address extends the lifetime of.
2940 // Note that there are only two potential values being referenced here,
2941 // BaseReg and ScaleReg (global addresses are always available, as are any
2942 // folded immediates).
2943 Value *BaseReg = AMAfter.BaseReg, *ScaledReg = AMAfter.ScaledReg;
2945 // If the BaseReg or ScaledReg was referenced by the previous addrmode, their
2946 // lifetime wasn't extended by adding this instruction.
2947 if (ValueAlreadyLiveAtInst(BaseReg, AMBefore.BaseReg, AMBefore.ScaledReg))
2949 if (ValueAlreadyLiveAtInst(ScaledReg, AMBefore.BaseReg, AMBefore.ScaledReg))
2950 ScaledReg = nullptr;
2952 // If folding this instruction (and it's subexprs) didn't extend any live
2953 // ranges, we're ok with it.
2954 if (!BaseReg && !ScaledReg)
2957 // If all uses of this instruction are ultimately load/store/inlineasm's,
2958 // check to see if their addressing modes will include this instruction. If
2959 // so, we can fold it into all uses, so it doesn't matter if it has multiple
2961 SmallVector<std::pair<Instruction*,unsigned>, 16> MemoryUses;
2962 SmallPtrSet<Instruction*, 16> ConsideredInsts;
2963 if (FindAllMemoryUses(I, MemoryUses, ConsideredInsts, TLI))
2964 return false; // Has a non-memory, non-foldable use!
2966 // Now that we know that all uses of this instruction are part of a chain of
2967 // computation involving only operations that could theoretically be folded
2968 // into a memory use, loop over each of these uses and see if they could
2969 // *actually* fold the instruction.
2970 SmallVector<Instruction*, 32> MatchedAddrModeInsts;
2971 for (unsigned i = 0, e = MemoryUses.size(); i != e; ++i) {
2972 Instruction *User = MemoryUses[i].first;
2973 unsigned OpNo = MemoryUses[i].second;
2975 // Get the access type of this use. If the use isn't a pointer, we don't
2976 // know what it accesses.
2977 Value *Address = User->getOperand(OpNo);
2978 if (!Address->getType()->isPointerTy())
2980 Type *AddressAccessTy = Address->getType()->getPointerElementType();
2982 // Do a match against the root of this address, ignoring profitability. This
2983 // will tell us if the addressing mode for the memory operation will
2984 // *actually* cover the shared instruction.
2986 TypePromotionTransaction::ConstRestorationPt LastKnownGood =
2987 TPT.getRestorationPoint();
2988 AddressingModeMatcher Matcher(MatchedAddrModeInsts, TLI, AddressAccessTy,
2989 MemoryInst, Result, InsertedTruncs,
2990 PromotedInsts, TPT);
2991 Matcher.IgnoreProfitability = true;
2992 bool Success = Matcher.MatchAddr(Address, 0);
2993 (void)Success; assert(Success && "Couldn't select *anything*?");
2995 // The match was to check the profitability, the changes made are not
2996 // part of the original matcher. Therefore, they should be dropped
2997 // otherwise the original matcher will not present the right state.
2998 TPT.rollback(LastKnownGood);
3000 // If the match didn't cover I, then it won't be shared by it.
3001 if (std::find(MatchedAddrModeInsts.begin(), MatchedAddrModeInsts.end(),
3002 I) == MatchedAddrModeInsts.end())
3005 MatchedAddrModeInsts.clear();
3011 } // end anonymous namespace
3013 /// IsNonLocalValue - Return true if the specified values are defined in a
3014 /// different basic block than BB.
3015 static bool IsNonLocalValue(Value *V, BasicBlock *BB) {
3016 if (Instruction *I = dyn_cast<Instruction>(V))
3017 return I->getParent() != BB;
3021 /// OptimizeMemoryInst - Load and Store Instructions often have
3022 /// addressing modes that can do significant amounts of computation. As such,
3023 /// instruction selection will try to get the load or store to do as much
3024 /// computation as possible for the program. The problem is that isel can only
3025 /// see within a single block. As such, we sink as much legal addressing mode
3026 /// stuff into the block as possible.
3028 /// This method is used to optimize both load/store and inline asms with memory
3030 bool CodeGenPrepare::OptimizeMemoryInst(Instruction *MemoryInst, Value *Addr,
3034 // Try to collapse single-value PHI nodes. This is necessary to undo
3035 // unprofitable PRE transformations.
3036 SmallVector<Value*, 8> worklist;
3037 SmallPtrSet<Value*, 16> Visited;
3038 worklist.push_back(Addr);
3040 // Use a worklist to iteratively look through PHI nodes, and ensure that
3041 // the addressing mode obtained from the non-PHI roots of the graph
3043 Value *Consensus = nullptr;
3044 unsigned NumUsesConsensus = 0;
3045 bool IsNumUsesConsensusValid = false;
3046 SmallVector<Instruction*, 16> AddrModeInsts;
3047 ExtAddrMode AddrMode;
3048 TypePromotionTransaction TPT;
3049 TypePromotionTransaction::ConstRestorationPt LastKnownGood =
3050 TPT.getRestorationPoint();
3051 while (!worklist.empty()) {
3052 Value *V = worklist.back();
3053 worklist.pop_back();
3055 // Break use-def graph loops.
3056 if (!Visited.insert(V).second) {
3057 Consensus = nullptr;
3061 // For a PHI node, push all of its incoming values.
3062 if (PHINode *P = dyn_cast<PHINode>(V)) {
3063 for (unsigned i = 0, e = P->getNumIncomingValues(); i != e; ++i)
3064 worklist.push_back(P->getIncomingValue(i));
3068 // For non-PHIs, determine the addressing mode being computed.
3069 SmallVector<Instruction*, 16> NewAddrModeInsts;
3070 ExtAddrMode NewAddrMode = AddressingModeMatcher::Match(
3071 V, AccessTy, MemoryInst, NewAddrModeInsts, *TLI, InsertedTruncsSet,
3072 PromotedInsts, TPT);
3074 // This check is broken into two cases with very similar code to avoid using
3075 // getNumUses() as much as possible. Some values have a lot of uses, so
3076 // calling getNumUses() unconditionally caused a significant compile-time
3080 AddrMode = NewAddrMode;
3081 AddrModeInsts = NewAddrModeInsts;
3083 } else if (NewAddrMode == AddrMode) {
3084 if (!IsNumUsesConsensusValid) {
3085 NumUsesConsensus = Consensus->getNumUses();
3086 IsNumUsesConsensusValid = true;
3089 // Ensure that the obtained addressing mode is equivalent to that obtained
3090 // for all other roots of the PHI traversal. Also, when choosing one
3091 // such root as representative, select the one with the most uses in order
3092 // to keep the cost modeling heuristics in AddressingModeMatcher
3094 unsigned NumUses = V->getNumUses();
3095 if (NumUses > NumUsesConsensus) {
3097 NumUsesConsensus = NumUses;
3098 AddrModeInsts = NewAddrModeInsts;
3103 Consensus = nullptr;
3107 // If the addressing mode couldn't be determined, or if multiple different
3108 // ones were determined, bail out now.
3110 TPT.rollback(LastKnownGood);
3115 // Check to see if any of the instructions supersumed by this addr mode are
3116 // non-local to I's BB.
3117 bool AnyNonLocal = false;
3118 for (unsigned i = 0, e = AddrModeInsts.size(); i != e; ++i) {
3119 if (IsNonLocalValue(AddrModeInsts[i], MemoryInst->getParent())) {
3125 // If all the instructions matched are already in this BB, don't do anything.
3127 DEBUG(dbgs() << "CGP: Found local addrmode: " << AddrMode << "\n");
3131 // Insert this computation right after this user. Since our caller is
3132 // scanning from the top of the BB to the bottom, reuse of the expr are
3133 // guaranteed to happen later.
3134 IRBuilder<> Builder(MemoryInst);
3136 // Now that we determined the addressing expression we want to use and know
3137 // that we have to sink it into this block. Check to see if we have already
3138 // done this for some other load/store instr in this block. If so, reuse the
3140 Value *&SunkAddr = SunkAddrs[Addr];
3142 DEBUG(dbgs() << "CGP: Reusing nonlocal addrmode: " << AddrMode << " for "
3143 << *MemoryInst << "\n");
3144 if (SunkAddr->getType() != Addr->getType())
3145 SunkAddr = Builder.CreateBitCast(SunkAddr, Addr->getType());
3146 } else if (AddrSinkUsingGEPs || (!AddrSinkUsingGEPs.getNumOccurrences() &&
3147 TM && TM->getSubtarget<TargetSubtargetInfo>().useAA())) {
3148 // By default, we use the GEP-based method when AA is used later. This
3149 // prevents new inttoptr/ptrtoint pairs from degrading AA capabilities.
3150 DEBUG(dbgs() << "CGP: SINKING nonlocal addrmode: " << AddrMode << " for "
3151 << *MemoryInst << "\n");
3152 Type *IntPtrTy = TLI->getDataLayout()->getIntPtrType(Addr->getType());
3153 Value *ResultPtr = nullptr, *ResultIndex = nullptr;
3155 // First, find the pointer.
3156 if (AddrMode.BaseReg && AddrMode.BaseReg->getType()->isPointerTy()) {
3157 ResultPtr = AddrMode.BaseReg;
3158 AddrMode.BaseReg = nullptr;
3161 if (AddrMode.Scale && AddrMode.ScaledReg->getType()->isPointerTy()) {
3162 // We can't add more than one pointer together, nor can we scale a
3163 // pointer (both of which seem meaningless).
3164 if (ResultPtr || AddrMode.Scale != 1)
3167 ResultPtr = AddrMode.ScaledReg;
3171 if (AddrMode.BaseGV) {
3175 ResultPtr = AddrMode.BaseGV;
3178 // If the real base value actually came from an inttoptr, then the matcher
3179 // will look through it and provide only the integer value. In that case,
3181 if (!ResultPtr && AddrMode.BaseReg) {
3183 Builder.CreateIntToPtr(AddrMode.BaseReg, Addr->getType(), "sunkaddr");
3184 AddrMode.BaseReg = nullptr;
3185 } else if (!ResultPtr && AddrMode.Scale == 1) {
3187 Builder.CreateIntToPtr(AddrMode.ScaledReg, Addr->getType(), "sunkaddr");
3192 !AddrMode.BaseReg && !AddrMode.Scale && !AddrMode.BaseOffs) {
3193 SunkAddr = Constant::getNullValue(Addr->getType());
3194 } else if (!ResultPtr) {
3198 Builder.getInt8PtrTy(Addr->getType()->getPointerAddressSpace());
3200 // Start with the base register. Do this first so that subsequent address
3201 // matching finds it last, which will prevent it from trying to match it
3202 // as the scaled value in case it happens to be a mul. That would be
3203 // problematic if we've sunk a different mul for the scale, because then
3204 // we'd end up sinking both muls.
3205 if (AddrMode.BaseReg) {
3206 Value *V = AddrMode.BaseReg;
3207 if (V->getType() != IntPtrTy)
3208 V = Builder.CreateIntCast(V, IntPtrTy, /*isSigned=*/true, "sunkaddr");
3213 // Add the scale value.
3214 if (AddrMode.Scale) {
3215 Value *V = AddrMode.ScaledReg;
3216 if (V->getType() == IntPtrTy) {
3218 } else if (cast<IntegerType>(IntPtrTy)->getBitWidth() <
3219 cast<IntegerType>(V->getType())->getBitWidth()) {
3220 V = Builder.CreateTrunc(V, IntPtrTy, "sunkaddr");
3222 // It is only safe to sign extend the BaseReg if we know that the math
3223 // required to create it did not overflow before we extend it. Since
3224 // the original IR value was tossed in favor of a constant back when
3225 // the AddrMode was created we need to bail out gracefully if widths
3226 // do not match instead of extending it.
3227 Instruction *I = dyn_cast_or_null<Instruction>(ResultIndex);
3228 if (I && (ResultIndex != AddrMode.BaseReg))
3229 I->eraseFromParent();
3233 if (AddrMode.Scale != 1)
3234 V = Builder.CreateMul(V, ConstantInt::get(IntPtrTy, AddrMode.Scale),
3237 ResultIndex = Builder.CreateAdd(ResultIndex, V, "sunkaddr");
3242 // Add in the Base Offset if present.
3243 if (AddrMode.BaseOffs) {
3244 Value *V = ConstantInt::get(IntPtrTy, AddrMode.BaseOffs);
3246 // We need to add this separately from the scale above to help with
3247 // SDAG consecutive load/store merging.
3248 if (ResultPtr->getType() != I8PtrTy)
3249 ResultPtr = Builder.CreateBitCast(ResultPtr, I8PtrTy);
3250 ResultPtr = Builder.CreateGEP(ResultPtr, ResultIndex, "sunkaddr");
3257 SunkAddr = ResultPtr;
3259 if (ResultPtr->getType() != I8PtrTy)
3260 ResultPtr = Builder.CreateBitCast(ResultPtr, I8PtrTy);
3261 SunkAddr = Builder.CreateGEP(ResultPtr, ResultIndex, "sunkaddr");
3264 if (SunkAddr->getType() != Addr->getType())
3265 SunkAddr = Builder.CreateBitCast(SunkAddr, Addr->getType());
3268 DEBUG(dbgs() << "CGP: SINKING nonlocal addrmode: " << AddrMode << " for "
3269 << *MemoryInst << "\n");
3270 Type *IntPtrTy = TLI->getDataLayout()->getIntPtrType(Addr->getType());
3271 Value *Result = nullptr;
3273 // Start with the base register. Do this first so that subsequent address
3274 // matching finds it last, which will prevent it from trying to match it
3275 // as the scaled value in case it happens to be a mul. That would be
3276 // problematic if we've sunk a different mul for the scale, because then
3277 // we'd end up sinking both muls.
3278 if (AddrMode.BaseReg) {
3279 Value *V = AddrMode.BaseReg;
3280 if (V->getType()->isPointerTy())
3281 V = Builder.CreatePtrToInt(V, IntPtrTy, "sunkaddr");
3282 if (V->getType() != IntPtrTy)
3283 V = Builder.CreateIntCast(V, IntPtrTy, /*isSigned=*/true, "sunkaddr");
3287 // Add the scale value.
3288 if (AddrMode.Scale) {
3289 Value *V = AddrMode.ScaledReg;
3290 if (V->getType() == IntPtrTy) {
3292 } else if (V->getType()->isPointerTy()) {
3293 V = Builder.CreatePtrToInt(V, IntPtrTy, "sunkaddr");
3294 } else if (cast<IntegerType>(IntPtrTy)->getBitWidth() <
3295 cast<IntegerType>(V->getType())->getBitWidth()) {
3296 V = Builder.CreateTrunc(V, IntPtrTy, "sunkaddr");
3298 // It is only safe to sign extend the BaseReg if we know that the math
3299 // required to create it did not overflow before we extend it. Since
3300 // the original IR value was tossed in favor of a constant back when
3301 // the AddrMode was created we need to bail out gracefully if widths
3302 // do not match instead of extending it.
3303 Instruction *I = dyn_cast_or_null<Instruction>(Result);
3304 if (I && (Result != AddrMode.BaseReg))
3305 I->eraseFromParent();
3308 if (AddrMode.Scale != 1)
3309 V = Builder.CreateMul(V, ConstantInt::get(IntPtrTy, AddrMode.Scale),
3312 Result = Builder.CreateAdd(Result, V, "sunkaddr");
3317 // Add in the BaseGV if present.
3318 if (AddrMode.BaseGV) {
3319 Value *V = Builder.CreatePtrToInt(AddrMode.BaseGV, IntPtrTy, "sunkaddr");
3321 Result = Builder.CreateAdd(Result, V, "sunkaddr");
3326 // Add in the Base Offset if present.
3327 if (AddrMode.BaseOffs) {
3328 Value *V = ConstantInt::get(IntPtrTy, AddrMode.BaseOffs);
3330 Result = Builder.CreateAdd(Result, V, "sunkaddr");
3336 SunkAddr = Constant::getNullValue(Addr->getType());
3338 SunkAddr = Builder.CreateIntToPtr(Result, Addr->getType(), "sunkaddr");
3341 MemoryInst->replaceUsesOfWith(Repl, SunkAddr);
3343 // If we have no uses, recursively delete the value and all dead instructions
3345 if (Repl->use_empty()) {
3346 // This can cause recursive deletion, which can invalidate our iterator.
3347 // Use a WeakVH to hold onto it in case this happens.
3348 WeakVH IterHandle(CurInstIterator);
3349 BasicBlock *BB = CurInstIterator->getParent();
3351 RecursivelyDeleteTriviallyDeadInstructions(Repl, TLInfo);
3353 if (IterHandle != CurInstIterator) {
3354 // If the iterator instruction was recursively deleted, start over at the
3355 // start of the block.
3356 CurInstIterator = BB->begin();
3364 /// OptimizeInlineAsmInst - If there are any memory operands, use
3365 /// OptimizeMemoryInst to sink their address computing into the block when
3366 /// possible / profitable.
3367 bool CodeGenPrepare::OptimizeInlineAsmInst(CallInst *CS) {
3368 bool MadeChange = false;
3370 TargetLowering::AsmOperandInfoVector
3371 TargetConstraints = TLI->ParseConstraints(CS);
3373 for (unsigned i = 0, e = TargetConstraints.size(); i != e; ++i) {
3374 TargetLowering::AsmOperandInfo &OpInfo = TargetConstraints[i];
3376 // Compute the constraint code and ConstraintType to use.
3377 TLI->ComputeConstraintToUse(OpInfo, SDValue());
3379 if (OpInfo.ConstraintType == TargetLowering::C_Memory &&
3380 OpInfo.isIndirect) {
3381 Value *OpVal = CS->getArgOperand(ArgNo++);
3382 MadeChange |= OptimizeMemoryInst(CS, OpVal, OpVal->getType());
3383 } else if (OpInfo.Type == InlineAsm::isInput)
3390 /// \brief Check if all the uses of \p Inst are equivalent (or free) zero or
3391 /// sign extensions.
3392 static bool hasSameExtUse(Instruction *Inst, const TargetLowering &TLI) {
3393 assert(!Inst->use_empty() && "Input must have at least one use");
3394 const Instruction *FirstUser = cast<Instruction>(*Inst->user_begin());
3395 bool IsSExt = isa<SExtInst>(FirstUser);
3396 Type *ExtTy = FirstUser->getType();
3397 for (const User *U : Inst->users()) {
3398 const Instruction *UI = cast<Instruction>(U);
3399 if ((IsSExt && !isa<SExtInst>(UI)) || (!IsSExt && !isa<ZExtInst>(UI)))
3401 Type *CurTy = UI->getType();
3402 // Same input and output types: Same instruction after CSE.
3406 // If IsSExt is true, we are in this situation:
3408 // b = sext ty1 a to ty2
3409 // c = sext ty1 a to ty3
3410 // Assuming ty2 is shorter than ty3, this could be turned into:
3412 // b = sext ty1 a to ty2
3413 // c = sext ty2 b to ty3
3414 // However, the last sext is not free.
3418 // This is a ZExt, maybe this is free to extend from one type to another.
3419 // In that case, we would not account for a different use.
3422 if (ExtTy->getScalarType()->getIntegerBitWidth() >
3423 CurTy->getScalarType()->getIntegerBitWidth()) {
3431 if (!TLI.isZExtFree(NarrowTy, LargeTy))
3434 // All uses are the same or can be derived from one another for free.
3438 /// \brief Try to form ExtLd by promoting \p Exts until they reach a
3439 /// load instruction.
3440 /// If an ext(load) can be formed, it is returned via \p LI for the load
3441 /// and \p Inst for the extension.
3442 /// Otherwise LI == nullptr and Inst == nullptr.
3443 /// When some promotion happened, \p TPT contains the proper state to
3446 /// \return true when promoting was necessary to expose the ext(load)
3447 /// opportunity, false otherwise.
3451 /// %ld = load i32* %addr
3452 /// %add = add nuw i32 %ld, 4
3453 /// %zext = zext i32 %add to i64
3457 /// %ld = load i32* %addr
3458 /// %zext = zext i32 %ld to i64
3459 /// %add = add nuw i64 %zext, 4
3461 /// Thanks to the promotion, we can match zext(load i32*) to i64.
3462 bool CodeGenPrepare::ExtLdPromotion(TypePromotionTransaction &TPT,
3463 LoadInst *&LI, Instruction *&Inst,
3464 const SmallVectorImpl<Instruction *> &Exts,
3465 unsigned CreatedInsts = 0) {
3466 // Iterate over all the extensions to see if one form an ext(load).
3467 for (auto I : Exts) {
3468 // Check if we directly have ext(load).
3469 if ((LI = dyn_cast<LoadInst>(I->getOperand(0)))) {
3471 // No promotion happened here.
3474 // Check whether or not we want to do any promotion.
3475 if (!TLI || !TLI->enableExtLdPromotion() || DisableExtLdPromotion)
3477 // Get the action to perform the promotion.
3478 TypePromotionHelper::Action TPH = TypePromotionHelper::getAction(
3479 I, InsertedTruncsSet, *TLI, PromotedInsts);
3480 // Check if we can promote.
3483 // Save the current state.
3484 TypePromotionTransaction::ConstRestorationPt LastKnownGood =
3485 TPT.getRestorationPoint();
3486 SmallVector<Instruction *, 4> NewExts;
3487 unsigned NewCreatedInsts = 0;
3489 Value *PromotedVal =
3490 TPH(I, TPT, PromotedInsts, NewCreatedInsts, &NewExts, nullptr);
3491 assert(PromotedVal &&
3492 "TypePromotionHelper should have filtered out those cases");
3494 // We would be able to merge only one extension in a load.
3495 // Therefore, if we have more than 1 new extension we heuristically
3496 // cut this search path, because it means we degrade the code quality.
3497 // With exactly 2, the transformation is neutral, because we will merge
3498 // one extension but leave one. However, we optimistically keep going,
3499 // because the new extension may be removed too.
3500 unsigned TotalCreatedInsts = CreatedInsts + NewCreatedInsts;
3501 if (!StressExtLdPromotion &&
3502 (TotalCreatedInsts > 1 ||
3503 !isPromotedInstructionLegal(*TLI, PromotedVal))) {
3504 // The promotion is not profitable, rollback to the previous state.
3505 TPT.rollback(LastKnownGood);
3508 // The promotion is profitable.
3509 // Check if it exposes an ext(load).
3510 (void)ExtLdPromotion(TPT, LI, Inst, NewExts, TotalCreatedInsts);
3511 if (LI && (StressExtLdPromotion || NewCreatedInsts == 0 ||
3512 // If we have created a new extension, i.e., now we have two
3513 // extensions. We must make sure one of them is merged with
3514 // the load, otherwise we may degrade the code quality.
3515 (LI->hasOneUse() || hasSameExtUse(LI, *TLI))))
3516 // Promotion happened.
3518 // If this does not help to expose an ext(load) then, rollback.
3519 TPT.rollback(LastKnownGood);
3521 // None of the extension can form an ext(load).
3527 /// MoveExtToFormExtLoad - Move a zext or sext fed by a load into the same
3528 /// basic block as the load, unless conditions are unfavorable. This allows
3529 /// SelectionDAG to fold the extend into the load.
3530 /// \p I[in/out] the extension may be modified during the process if some
3531 /// promotions apply.
3533 bool CodeGenPrepare::MoveExtToFormExtLoad(Instruction *&I) {
3534 // Try to promote a chain of computation if it allows to form
3535 // an extended load.
3536 TypePromotionTransaction TPT;
3537 TypePromotionTransaction::ConstRestorationPt LastKnownGood =
3538 TPT.getRestorationPoint();
3539 SmallVector<Instruction *, 1> Exts;
3541 // Look for a load being extended.
3542 LoadInst *LI = nullptr;
3543 Instruction *OldExt = I;
3544 bool HasPromoted = ExtLdPromotion(TPT, LI, I, Exts);
3546 assert(!HasPromoted && !LI && "If we did not match any load instruction "
3547 "the code must remain the same");
3552 // If they're already in the same block, there's nothing to do.
3553 // Make the cheap checks first if we did not promote.
3554 // If we promoted, we need to check if it is indeed profitable.
3555 if (!HasPromoted && LI->getParent() == I->getParent())
3558 EVT VT = TLI->getValueType(I->getType());
3559 EVT LoadVT = TLI->getValueType(LI->getType());
3561 // If the load has other users and the truncate is not free, this probably
3562 // isn't worthwhile.
3563 if (!LI->hasOneUse() && TLI &&
3564 (TLI->isTypeLegal(LoadVT) || !TLI->isTypeLegal(VT)) &&
3565 !TLI->isTruncateFree(I->getType(), LI->getType())) {
3567 TPT.rollback(LastKnownGood);
3571 // Check whether the target supports casts folded into loads.
3573 if (isa<ZExtInst>(I))
3574 LType = ISD::ZEXTLOAD;
3576 assert(isa<SExtInst>(I) && "Unexpected ext type!");
3577 LType = ISD::SEXTLOAD;
3579 if (TLI && !TLI->isLoadExtLegal(LType, VT, LoadVT)) {
3581 TPT.rollback(LastKnownGood);
3585 // Move the extend into the same block as the load, so that SelectionDAG
3588 I->removeFromParent();
3594 bool CodeGenPrepare::OptimizeExtUses(Instruction *I) {
3595 BasicBlock *DefBB = I->getParent();
3597 // If the result of a {s|z}ext and its source are both live out, rewrite all
3598 // other uses of the source with result of extension.
3599 Value *Src = I->getOperand(0);
3600 if (Src->hasOneUse())
3603 // Only do this xform if truncating is free.
3604 if (TLI && !TLI->isTruncateFree(I->getType(), Src->getType()))
3607 // Only safe to perform the optimization if the source is also defined in
3609 if (!isa<Instruction>(Src) || DefBB != cast<Instruction>(Src)->getParent())
3612 bool DefIsLiveOut = false;
3613 for (User *U : I->users()) {
3614 Instruction *UI = cast<Instruction>(U);
3616 // Figure out which BB this ext is used in.
3617 BasicBlock *UserBB = UI->getParent();
3618 if (UserBB == DefBB) continue;
3619 DefIsLiveOut = true;
3625 // Make sure none of the uses are PHI nodes.
3626 for (User *U : Src->users()) {
3627 Instruction *UI = cast<Instruction>(U);
3628 BasicBlock *UserBB = UI->getParent();
3629 if (UserBB == DefBB) continue;
3630 // Be conservative. We don't want this xform to end up introducing
3631 // reloads just before load / store instructions.
3632 if (isa<PHINode>(UI) || isa<LoadInst>(UI) || isa<StoreInst>(UI))
3636 // InsertedTruncs - Only insert one trunc in each block once.
3637 DenseMap<BasicBlock*, Instruction*> InsertedTruncs;
3639 bool MadeChange = false;
3640 for (Use &U : Src->uses()) {
3641 Instruction *User = cast<Instruction>(U.getUser());
3643 // Figure out which BB this ext is used in.
3644 BasicBlock *UserBB = User->getParent();
3645 if (UserBB == DefBB) continue;
3647 // Both src and def are live in this block. Rewrite the use.
3648 Instruction *&InsertedTrunc = InsertedTruncs[UserBB];
3650 if (!InsertedTrunc) {
3651 BasicBlock::iterator InsertPt = UserBB->getFirstInsertionPt();
3652 InsertedTrunc = new TruncInst(I, Src->getType(), "", InsertPt);
3653 InsertedTruncsSet.insert(InsertedTrunc);
3656 // Replace a use of the {s|z}ext source with a use of the result.
3665 /// isFormingBranchFromSelectProfitable - Returns true if a SelectInst should be
3666 /// turned into an explicit branch.
3667 static bool isFormingBranchFromSelectProfitable(SelectInst *SI) {
3668 // FIXME: This should use the same heuristics as IfConversion to determine
3669 // whether a select is better represented as a branch. This requires that
3670 // branch probability metadata is preserved for the select, which is not the
3673 CmpInst *Cmp = dyn_cast<CmpInst>(SI->getCondition());
3675 // If the branch is predicted right, an out of order CPU can avoid blocking on
3676 // the compare. Emit cmovs on compares with a memory operand as branches to
3677 // avoid stalls on the load from memory. If the compare has more than one use
3678 // there's probably another cmov or setcc around so it's not worth emitting a
3683 Value *CmpOp0 = Cmp->getOperand(0);
3684 Value *CmpOp1 = Cmp->getOperand(1);
3686 // We check that the memory operand has one use to avoid uses of the loaded
3687 // value directly after the compare, making branches unprofitable.
3688 return Cmp->hasOneUse() &&
3689 ((isa<LoadInst>(CmpOp0) && CmpOp0->hasOneUse()) ||
3690 (isa<LoadInst>(CmpOp1) && CmpOp1->hasOneUse()));
3694 /// If we have a SelectInst that will likely profit from branch prediction,
3695 /// turn it into a branch.
3696 bool CodeGenPrepare::OptimizeSelectInst(SelectInst *SI) {
3697 bool VectorCond = !SI->getCondition()->getType()->isIntegerTy(1);
3699 // Can we convert the 'select' to CF ?
3700 if (DisableSelectToBranch || OptSize || !TLI || VectorCond)
3703 TargetLowering::SelectSupportKind SelectKind;
3705 SelectKind = TargetLowering::VectorMaskSelect;
3706 else if (SI->getType()->isVectorTy())
3707 SelectKind = TargetLowering::ScalarCondVectorVal;
3709 SelectKind = TargetLowering::ScalarValSelect;
3711 // Do we have efficient codegen support for this kind of 'selects' ?
3712 if (TLI->isSelectSupported(SelectKind)) {
3713 // We have efficient codegen support for the select instruction.
3714 // Check if it is profitable to keep this 'select'.
3715 if (!TLI->isPredictableSelectExpensive() ||
3716 !isFormingBranchFromSelectProfitable(SI))
3722 // First, we split the block containing the select into 2 blocks.
3723 BasicBlock *StartBlock = SI->getParent();
3724 BasicBlock::iterator SplitPt = ++(BasicBlock::iterator(SI));
3725 BasicBlock *NextBlock = StartBlock->splitBasicBlock(SplitPt, "select.end");
3727 // Create a new block serving as the landing pad for the branch.
3728 BasicBlock *SmallBlock = BasicBlock::Create(SI->getContext(), "select.mid",
3729 NextBlock->getParent(), NextBlock);
3731 // Move the unconditional branch from the block with the select in it into our
3732 // landing pad block.
3733 StartBlock->getTerminator()->eraseFromParent();
3734 BranchInst::Create(NextBlock, SmallBlock);
3736 // Insert the real conditional branch based on the original condition.
3737 BranchInst::Create(NextBlock, SmallBlock, SI->getCondition(), SI);
3739 // The select itself is replaced with a PHI Node.
3740 PHINode *PN = PHINode::Create(SI->getType(), 2, "", NextBlock->begin());
3742 PN->addIncoming(SI->getTrueValue(), StartBlock);
3743 PN->addIncoming(SI->getFalseValue(), SmallBlock);
3744 SI->replaceAllUsesWith(PN);
3745 SI->eraseFromParent();
3747 // Instruct OptimizeBlock to skip to the next block.
3748 CurInstIterator = StartBlock->end();
3749 ++NumSelectsExpanded;
3753 static bool isBroadcastShuffle(ShuffleVectorInst *SVI) {
3754 SmallVector<int, 16> Mask(SVI->getShuffleMask());
3756 for (unsigned i = 0; i < Mask.size(); ++i) {
3757 if (SplatElem != -1 && Mask[i] != -1 && Mask[i] != SplatElem)
3759 SplatElem = Mask[i];
3765 /// Some targets have expensive vector shifts if the lanes aren't all the same
3766 /// (e.g. x86 only introduced "vpsllvd" and friends with AVX2). In these cases
3767 /// it's often worth sinking a shufflevector splat down to its use so that
3768 /// codegen can spot all lanes are identical.
3769 bool CodeGenPrepare::OptimizeShuffleVectorInst(ShuffleVectorInst *SVI) {
3770 BasicBlock *DefBB = SVI->getParent();
3772 // Only do this xform if variable vector shifts are particularly expensive.
3773 if (!TLI || !TLI->isVectorShiftByScalarCheap(SVI->getType()))
3776 // We only expect better codegen by sinking a shuffle if we can recognise a
3778 if (!isBroadcastShuffle(SVI))
3781 // InsertedShuffles - Only insert a shuffle in each block once.
3782 DenseMap<BasicBlock*, Instruction*> InsertedShuffles;
3784 bool MadeChange = false;
3785 for (User *U : SVI->users()) {
3786 Instruction *UI = cast<Instruction>(U);
3788 // Figure out which BB this ext is used in.
3789 BasicBlock *UserBB = UI->getParent();
3790 if (UserBB == DefBB) continue;
3792 // For now only apply this when the splat is used by a shift instruction.
3793 if (!UI->isShift()) continue;
3795 // Everything checks out, sink the shuffle if the user's block doesn't
3796 // already have a copy.
3797 Instruction *&InsertedShuffle = InsertedShuffles[UserBB];
3799 if (!InsertedShuffle) {
3800 BasicBlock::iterator InsertPt = UserBB->getFirstInsertionPt();
3801 InsertedShuffle = new ShuffleVectorInst(SVI->getOperand(0),
3803 SVI->getOperand(2), "", InsertPt);
3806 UI->replaceUsesOfWith(SVI, InsertedShuffle);
3810 // If we removed all uses, nuke the shuffle.
3811 if (SVI->use_empty()) {
3812 SVI->eraseFromParent();
3820 /// \brief Helper class to promote a scalar operation to a vector one.
3821 /// This class is used to move downward extractelement transition.
3823 /// a = vector_op <2 x i32>
3824 /// b = extractelement <2 x i32> a, i32 0
3829 /// a = vector_op <2 x i32>
3830 /// c = vector_op a (equivalent to scalar_op on the related lane)
3831 /// * d = extractelement <2 x i32> c, i32 0
3833 /// Assuming both extractelement and store can be combine, we get rid of the
3835 class VectorPromoteHelper {
3836 /// Used to perform some checks on the legality of vector operations.
3837 const TargetLowering &TLI;
3839 /// Used to estimated the cost of the promoted chain.
3840 const TargetTransformInfo &TTI;
3842 /// The transition being moved downwards.
3843 Instruction *Transition;
3844 /// The sequence of instructions to be promoted.
3845 SmallVector<Instruction *, 4> InstsToBePromoted;
3846 /// Cost of combining a store and an extract.
3847 unsigned StoreExtractCombineCost;
3848 /// Instruction that will be combined with the transition.
3849 Instruction *CombineInst;
3851 /// \brief The instruction that represents the current end of the transition.
3852 /// Since we are faking the promotion until we reach the end of the chain
3853 /// of computation, we need a way to get the current end of the transition.
3854 Instruction *getEndOfTransition() const {
3855 if (InstsToBePromoted.empty())
3857 return InstsToBePromoted.back();
3860 /// \brief Return the index of the original value in the transition.
3861 /// E.g., for "extractelement <2 x i32> c, i32 1" the original value,
3862 /// c, is at index 0.
3863 unsigned getTransitionOriginalValueIdx() const {
3864 assert(isa<ExtractElementInst>(Transition) &&
3865 "Other kind of transitions are not supported yet");
3869 /// \brief Return the index of the index in the transition.
3870 /// E.g., for "extractelement <2 x i32> c, i32 0" the index
3872 unsigned getTransitionIdx() const {
3873 assert(isa<ExtractElementInst>(Transition) &&
3874 "Other kind of transitions are not supported yet");
3878 /// \brief Get the type of the transition.
3879 /// This is the type of the original value.
3880 /// E.g., for "extractelement <2 x i32> c, i32 1" the type of the
3881 /// transition is <2 x i32>.
3882 Type *getTransitionType() const {
3883 return Transition->getOperand(getTransitionOriginalValueIdx())->getType();
3886 /// \brief Promote \p ToBePromoted by moving \p Def downward through.
3887 /// I.e., we have the following sequence:
3888 /// Def = Transition <ty1> a to <ty2>
3889 /// b = ToBePromoted <ty2> Def, ...
3891 /// b = ToBePromoted <ty1> a, ...
3892 /// Def = Transition <ty1> ToBePromoted to <ty2>
3893 void promoteImpl(Instruction *ToBePromoted);
3895 /// \brief Check whether or not it is profitable to promote all the
3896 /// instructions enqueued to be promoted.
3897 bool isProfitableToPromote() {
3898 Value *ValIdx = Transition->getOperand(getTransitionOriginalValueIdx());
3899 unsigned Index = isa<ConstantInt>(ValIdx)
3900 ? cast<ConstantInt>(ValIdx)->getZExtValue()
3902 Type *PromotedType = getTransitionType();
3904 StoreInst *ST = cast<StoreInst>(CombineInst);
3905 unsigned AS = ST->getPointerAddressSpace();
3906 unsigned Align = ST->getAlignment();
3907 // Check if this store is supported.
3908 if (!TLI.allowsMisalignedMemoryAccesses(
3909 TLI.getValueType(ST->getValueOperand()->getType()), AS, Align)) {
3910 // If this is not supported, there is no way we can combine
3911 // the extract with the store.
3915 // The scalar chain of computation has to pay for the transition
3916 // scalar to vector.
3917 // The vector chain has to account for the combining cost.
3918 uint64_t ScalarCost =
3919 TTI.getVectorInstrCost(Transition->getOpcode(), PromotedType, Index);
3920 uint64_t VectorCost = StoreExtractCombineCost;
3921 for (const auto &Inst : InstsToBePromoted) {
3922 // Compute the cost.
3923 // By construction, all instructions being promoted are arithmetic ones.
3924 // Moreover, one argument is a constant that can be viewed as a splat
3926 Value *Arg0 = Inst->getOperand(0);
3927 bool IsArg0Constant = isa<UndefValue>(Arg0) || isa<ConstantInt>(Arg0) ||
3928 isa<ConstantFP>(Arg0);
3929 TargetTransformInfo::OperandValueKind Arg0OVK =
3930 IsArg0Constant ? TargetTransformInfo::OK_UniformConstantValue
3931 : TargetTransformInfo::OK_AnyValue;
3932 TargetTransformInfo::OperandValueKind Arg1OVK =
3933 !IsArg0Constant ? TargetTransformInfo::OK_UniformConstantValue
3934 : TargetTransformInfo::OK_AnyValue;
3935 ScalarCost += TTI.getArithmeticInstrCost(
3936 Inst->getOpcode(), Inst->getType(), Arg0OVK, Arg1OVK);
3937 VectorCost += TTI.getArithmeticInstrCost(Inst->getOpcode(), PromotedType,
3940 DEBUG(dbgs() << "Estimated cost of computation to be promoted:\nScalar: "
3941 << ScalarCost << "\nVector: " << VectorCost << '\n');
3942 return ScalarCost > VectorCost;
3945 /// \brief Generate a constant vector with \p Val with the same
3946 /// number of elements as the transition.
3947 /// \p UseSplat defines whether or not \p Val should be replicated
3948 /// accross the whole vector.
3949 /// In other words, if UseSplat == true, we generate <Val, Val, ..., Val>,
3950 /// otherwise we generate a vector with as many undef as possible:
3951 /// <undef, ..., undef, Val, undef, ..., undef> where \p Val is only
3952 /// used at the index of the extract.
3953 Value *getConstantVector(Constant *Val, bool UseSplat) const {
3954 unsigned ExtractIdx = UINT_MAX;
3956 // If we cannot determine where the constant must be, we have to
3957 // use a splat constant.
3958 Value *ValExtractIdx = Transition->getOperand(getTransitionIdx());
3959 if (ConstantInt *CstVal = dyn_cast<ConstantInt>(ValExtractIdx))
3960 ExtractIdx = CstVal->getSExtValue();
3965 unsigned End = getTransitionType()->getVectorNumElements();
3967 return ConstantVector::getSplat(End, Val);
3969 SmallVector<Constant *, 4> ConstVec;
3970 UndefValue *UndefVal = UndefValue::get(Val->getType());
3971 for (unsigned Idx = 0; Idx != End; ++Idx) {
3972 if (Idx == ExtractIdx)
3973 ConstVec.push_back(Val);
3975 ConstVec.push_back(UndefVal);
3977 return ConstantVector::get(ConstVec);
3980 /// \brief Check if promoting to a vector type an operand at \p OperandIdx
3981 /// in \p Use can trigger undefined behavior.
3982 static bool canCauseUndefinedBehavior(const Instruction *Use,
3983 unsigned OperandIdx) {
3984 // This is not safe to introduce undef when the operand is on
3985 // the right hand side of a division-like instruction.
3986 if (OperandIdx != 1)
3988 switch (Use->getOpcode()) {
3991 case Instruction::SDiv:
3992 case Instruction::UDiv:
3993 case Instruction::SRem:
3994 case Instruction::URem:
3996 case Instruction::FDiv:
3997 case Instruction::FRem:
3998 return !Use->hasNoNaNs();
4000 llvm_unreachable(nullptr);
4004 VectorPromoteHelper(const TargetLowering &TLI, const TargetTransformInfo &TTI,
4005 Instruction *Transition, unsigned CombineCost)
4006 : TLI(TLI), TTI(TTI), Transition(Transition),
4007 StoreExtractCombineCost(CombineCost), CombineInst(nullptr) {
4008 assert(Transition && "Do not know how to promote null");
4011 /// \brief Check if we can promote \p ToBePromoted to \p Type.
4012 bool canPromote(const Instruction *ToBePromoted) const {
4013 // We could support CastInst too.
4014 return isa<BinaryOperator>(ToBePromoted);
4017 /// \brief Check if it is profitable to promote \p ToBePromoted
4018 /// by moving downward the transition through.
4019 bool shouldPromote(const Instruction *ToBePromoted) const {
4020 // Promote only if all the operands can be statically expanded.
4021 // Indeed, we do not want to introduce any new kind of transitions.
4022 for (const Use &U : ToBePromoted->operands()) {
4023 const Value *Val = U.get();
4024 if (Val == getEndOfTransition()) {
4025 // If the use is a division and the transition is on the rhs,
4026 // we cannot promote the operation, otherwise we may create a
4027 // division by zero.
4028 if (canCauseUndefinedBehavior(ToBePromoted, U.getOperandNo()))
4032 if (!isa<ConstantInt>(Val) && !isa<UndefValue>(Val) &&
4033 !isa<ConstantFP>(Val))
4036 // Check that the resulting operation is legal.
4037 int ISDOpcode = TLI.InstructionOpcodeToISD(ToBePromoted->getOpcode());
4040 return StressStoreExtract ||
4041 TLI.isOperationLegalOrCustom(
4042 ISDOpcode, TLI.getValueType(getTransitionType(), true));
4045 /// \brief Check whether or not \p Use can be combined
4046 /// with the transition.
4047 /// I.e., is it possible to do Use(Transition) => AnotherUse?
4048 bool canCombine(const Instruction *Use) { return isa<StoreInst>(Use); }
4050 /// \brief Record \p ToBePromoted as part of the chain to be promoted.
4051 void enqueueForPromotion(Instruction *ToBePromoted) {
4052 InstsToBePromoted.push_back(ToBePromoted);
4055 /// \brief Set the instruction that will be combined with the transition.
4056 void recordCombineInstruction(Instruction *ToBeCombined) {
4057 assert(canCombine(ToBeCombined) && "Unsupported instruction to combine");
4058 CombineInst = ToBeCombined;
4061 /// \brief Promote all the instructions enqueued for promotion if it is
4063 /// \return True if the promotion happened, false otherwise.
4065 // Check if there is something to promote.
4066 // Right now, if we do not have anything to combine with,
4067 // we assume the promotion is not profitable.
4068 if (InstsToBePromoted.empty() || !CombineInst)
4072 if (!StressStoreExtract && !isProfitableToPromote())
4076 for (auto &ToBePromoted : InstsToBePromoted)
4077 promoteImpl(ToBePromoted);
4078 InstsToBePromoted.clear();
4082 } // End of anonymous namespace.
4084 void VectorPromoteHelper::promoteImpl(Instruction *ToBePromoted) {
4085 // At this point, we know that all the operands of ToBePromoted but Def
4086 // can be statically promoted.
4087 // For Def, we need to use its parameter in ToBePromoted:
4088 // b = ToBePromoted ty1 a
4089 // Def = Transition ty1 b to ty2
4090 // Move the transition down.
4091 // 1. Replace all uses of the promoted operation by the transition.
4092 // = ... b => = ... Def.
4093 assert(ToBePromoted->getType() == Transition->getType() &&
4094 "The type of the result of the transition does not match "
4096 ToBePromoted->replaceAllUsesWith(Transition);
4097 // 2. Update the type of the uses.
4098 // b = ToBePromoted ty2 Def => b = ToBePromoted ty1 Def.
4099 Type *TransitionTy = getTransitionType();
4100 ToBePromoted->mutateType(TransitionTy);
4101 // 3. Update all the operands of the promoted operation with promoted
4103 // b = ToBePromoted ty1 Def => b = ToBePromoted ty1 a.
4104 for (Use &U : ToBePromoted->operands()) {
4105 Value *Val = U.get();
4106 Value *NewVal = nullptr;
4107 if (Val == Transition)
4108 NewVal = Transition->getOperand(getTransitionOriginalValueIdx());
4109 else if (isa<UndefValue>(Val) || isa<ConstantInt>(Val) ||
4110 isa<ConstantFP>(Val)) {
4111 // Use a splat constant if it is not safe to use undef.
4112 NewVal = getConstantVector(
4113 cast<Constant>(Val),
4114 isa<UndefValue>(Val) ||
4115 canCauseUndefinedBehavior(ToBePromoted, U.getOperandNo()));
4117 llvm_unreachable("Did you modified shouldPromote and forgot to update "
4119 ToBePromoted->setOperand(U.getOperandNo(), NewVal);
4121 Transition->removeFromParent();
4122 Transition->insertAfter(ToBePromoted);
4123 Transition->setOperand(getTransitionOriginalValueIdx(), ToBePromoted);
4126 // See if we can speculate calls to intrinsic cttz/ctlz.
4131 // %cmp = icmp eq i64 %val, 0
4132 // br i1 %cmp, label %end.bb, label %then.bb
4135 // %c = tail call i64 @llvm.cttz.i64(i64 %val, i1 true)
4139 // %cond = phi i64 [ %c, %then.bb ], [ 64, %entry ]
4145 // %c = tail call i64 @llvm.cttz.i64(i64 %val, i1 false)
4147 static bool OptimizeBranchInst(BranchInst *BrInst, const TargetLowering &TLI) {
4148 assert(BrInst->isConditional() && "Expected a conditional branch!");
4149 BasicBlock *ThenBB = BrInst->getSuccessor(1);
4150 BasicBlock *EndBB = BrInst->getSuccessor(0);
4152 // See if ThenBB contains only one instruction (excluding the
4153 // terminator and DbgInfoIntrinsic calls).
4154 IntrinsicInst *II = nullptr;
4155 CastInst *CI = nullptr;
4156 for (BasicBlock::iterator I = ThenBB->begin(),
4157 E = std::prev(ThenBB->end()); I != E; ++I) {
4159 if (isa<DbgInfoIntrinsic>(I))
4162 // Check if this is a zero extension or a truncate of a previously
4163 // matched call to intrinsic cttz/ctlz.
4165 // Early exit if we already found a "free" zero extend/truncate.
4169 Type *SrcTy = II->getType();
4170 Type *DestTy = I->getType();
4173 if (match(cast<Instruction>(I), m_ZExt(m_Value(V))) && V == II) {
4174 // Speculate this zero extend only if it is "free" for the target.
4175 if (TLI.isZExtFree(SrcTy, DestTy)) {
4176 CI = cast<CastInst>(I);
4179 } else if (match(cast<Instruction>(I), m_Trunc(m_Value(V))) && V == II) {
4180 // Speculate this truncate only if it is "free" for the target.
4181 if (TLI.isTruncateFree(SrcTy, DestTy)) {
4182 CI = cast<CastInst>(I);
4186 // Avoid speculating more than one instruction.
4191 // See if this is a call to intrinsic cttz/ctlz.
4192 if (match(cast<Instruction>(I), m_Intrinsic<Intrinsic::cttz>())) {
4193 // Avoid speculating expensive intrinsic calls.
4194 if (!TLI.isCheapToSpeculateCttz())
4197 else if (match(cast<Instruction>(I), m_Intrinsic<Intrinsic::ctlz>())) {
4198 // Avoid speculating expensive intrinsic calls.
4199 if (!TLI.isCheapToSpeculateCtlz())
4204 II = cast<IntrinsicInst>(I);
4207 // Look for PHI nodes with 'II' as the incoming value from 'ThenBB'.
4208 BasicBlock *EntryBB = BrInst->getParent();
4209 for (BasicBlock::iterator I = EndBB->begin();
4210 PHINode *PN = dyn_cast<PHINode>(I); ++I) {
4211 Value *ThenV = PN->getIncomingValueForBlock(ThenBB);
4212 Value *OrigV = PN->getIncomingValueForBlock(EntryBB);
4217 if (ThenV != II && (!CI || ThenV != CI))
4220 if (ConstantInt *CInt = dyn_cast<ConstantInt>(OrigV)) {
4221 unsigned BitWidth = II->getType()->getIntegerBitWidth();
4223 // Don't try to simplify this phi node if 'ThenV' is a cttz/ctlz
4224 // intrinsic call, but 'OrigV' is not equal to the 'size-of' in bits
4225 // of the value in input to the cttz/ctlz.
4226 if (CInt->getValue() != BitWidth)
4229 // Hoist the call to cttz/ctlz from ThenBB into EntryBB.
4230 EntryBB->getInstList().splice(BrInst, ThenBB->getInstList(),
4231 ThenBB->begin(), std::prev(ThenBB->end()));
4233 // Update PN setting ThenV as the incoming value from both 'EntryBB'
4234 // and 'ThenBB'. Eventually, method 'OptimizeInst' will fold this
4235 // phi node if all the incoming values are the same.
4236 PN->setIncomingValue(PN->getBasicBlockIndex(EntryBB), ThenV);
4237 PN->setIncomingValue(PN->getBasicBlockIndex(ThenBB), ThenV);
4239 // Clear the 'undef on zero' flag of the cttz/ctlz intrinsic call.
4240 if (cast<ConstantInt>(II->getArgOperand(1))->isOne()) {
4241 Type *Ty = II->getArgOperand(0)->getType();
4242 Value *Args[] = { II->getArgOperand(0),
4243 ConstantInt::getFalse(II->getContext()) };
4244 Module *M = EntryBB->getParent()->getParent();
4245 Value *IF = Intrinsic::getDeclaration(M, II->getIntrinsicID(), Ty);
4246 IRBuilder<> Builder(II);
4247 Instruction *NewI = Builder.CreateCall(IF, Args);
4249 // Replace the old call to cttz/ctlz.
4250 II->replaceAllUsesWith(NewI);
4251 II->eraseFromParent();
4254 // Update BrInst condition so that the branch to EndBB is always taken.
4255 // Later on, method 'ConstantFoldTerminator' will simplify this branch
4256 // replacing it with a direct branch to 'EndBB'.
4257 // As a side effect, CodeGenPrepare will attempt to simplify the control
4258 // flow graph by deleting basic block 'ThenBB' and merging 'EntryBB' into
4259 // 'EndBB' (calling method 'EliminateFallThrough').
4260 BrInst->setCondition(ConstantInt::getTrue(BrInst->getContext()));
4268 /// Some targets can do store(extractelement) with one instruction.
4269 /// Try to push the extractelement towards the stores when the target
4270 /// has this feature and this is profitable.
4271 bool CodeGenPrepare::OptimizeExtractElementInst(Instruction *Inst) {
4272 unsigned CombineCost = UINT_MAX;
4273 if (DisableStoreExtract || !TLI ||
4274 (!StressStoreExtract &&
4275 !TLI->canCombineStoreAndExtract(Inst->getOperand(0)->getType(),
4276 Inst->getOperand(1), CombineCost)))
4279 // At this point we know that Inst is a vector to scalar transition.
4280 // Try to move it down the def-use chain, until:
4281 // - We can combine the transition with its single use
4282 // => we got rid of the transition.
4283 // - We escape the current basic block
4284 // => we would need to check that we are moving it at a cheaper place and
4285 // we do not do that for now.
4286 BasicBlock *Parent = Inst->getParent();
4287 DEBUG(dbgs() << "Found an interesting transition: " << *Inst << '\n');
4288 VectorPromoteHelper VPH(*TLI, *TTI, Inst, CombineCost);
4289 // If the transition has more than one use, assume this is not going to be
4291 while (Inst->hasOneUse()) {
4292 Instruction *ToBePromoted = cast<Instruction>(*Inst->user_begin());
4293 DEBUG(dbgs() << "Use: " << *ToBePromoted << '\n');
4295 if (ToBePromoted->getParent() != Parent) {
4296 DEBUG(dbgs() << "Instruction to promote is in a different block ("
4297 << ToBePromoted->getParent()->getName()
4298 << ") than the transition (" << Parent->getName() << ").\n");
4302 if (VPH.canCombine(ToBePromoted)) {
4303 DEBUG(dbgs() << "Assume " << *Inst << '\n'
4304 << "will be combined with: " << *ToBePromoted << '\n');
4305 VPH.recordCombineInstruction(ToBePromoted);
4306 bool Changed = VPH.promote();
4307 NumStoreExtractExposed += Changed;
4311 DEBUG(dbgs() << "Try promoting.\n");
4312 if (!VPH.canPromote(ToBePromoted) || !VPH.shouldPromote(ToBePromoted))
4315 DEBUG(dbgs() << "Promoting is possible... Enqueue for promotion!\n");
4317 VPH.enqueueForPromotion(ToBePromoted);
4318 Inst = ToBePromoted;
4323 bool CodeGenPrepare::OptimizeInst(Instruction *I, bool& ModifiedDT) {
4324 if (PHINode *P = dyn_cast<PHINode>(I)) {
4325 // It is possible for very late stage optimizations (such as SimplifyCFG)
4326 // to introduce PHI nodes too late to be cleaned up. If we detect such a
4327 // trivial PHI, go ahead and zap it here.
4328 if (Value *V = SimplifyInstruction(P, TLI ? TLI->getDataLayout() : nullptr,
4330 P->replaceAllUsesWith(V);
4331 P->eraseFromParent();
4338 if (CastInst *CI = dyn_cast<CastInst>(I)) {
4339 // If the source of the cast is a constant, then this should have
4340 // already been constant folded. The only reason NOT to constant fold
4341 // it is if something (e.g. LSR) was careful to place the constant
4342 // evaluation in a block other than then one that uses it (e.g. to hoist
4343 // the address of globals out of a loop). If this is the case, we don't
4344 // want to forward-subst the cast.
4345 if (isa<Constant>(CI->getOperand(0)))
4348 if (TLI && OptimizeNoopCopyExpression(CI, *TLI))
4351 if (isa<ZExtInst>(I) || isa<SExtInst>(I)) {
4352 /// Sink a zext or sext into its user blocks if the target type doesn't
4353 /// fit in one register
4354 if (TLI && TLI->getTypeAction(CI->getContext(),
4355 TLI->getValueType(CI->getType())) ==
4356 TargetLowering::TypeExpandInteger) {
4357 return SinkCast(CI);
4359 bool MadeChange = MoveExtToFormExtLoad(I);
4360 return MadeChange | OptimizeExtUses(I);
4366 if (CmpInst *CI = dyn_cast<CmpInst>(I))
4367 if (!TLI || !TLI->hasMultipleConditionRegisters())
4368 return OptimizeCmpExpression(CI);
4370 if (LoadInst *LI = dyn_cast<LoadInst>(I)) {
4372 return OptimizeMemoryInst(I, I->getOperand(0), LI->getType());
4376 if (StoreInst *SI = dyn_cast<StoreInst>(I)) {
4378 return OptimizeMemoryInst(I, SI->getOperand(1),
4379 SI->getOperand(0)->getType());
4383 BinaryOperator *BinOp = dyn_cast<BinaryOperator>(I);
4385 if (BinOp && (BinOp->getOpcode() == Instruction::AShr ||
4386 BinOp->getOpcode() == Instruction::LShr)) {
4387 ConstantInt *CI = dyn_cast<ConstantInt>(BinOp->getOperand(1));
4388 if (TLI && CI && TLI->hasExtractBitsInsn())
4389 return OptimizeExtractBits(BinOp, CI, *TLI);
4394 if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(I)) {
4395 if (GEPI->hasAllZeroIndices()) {
4396 /// The GEP operand must be a pointer, so must its result -> BitCast
4397 Instruction *NC = new BitCastInst(GEPI->getOperand(0), GEPI->getType(),
4398 GEPI->getName(), GEPI);
4399 GEPI->replaceAllUsesWith(NC);
4400 GEPI->eraseFromParent();
4402 OptimizeInst(NC, ModifiedDT);
4408 if (CallInst *CI = dyn_cast<CallInst>(I))
4409 return OptimizeCallInst(CI, ModifiedDT);
4411 if (SelectInst *SI = dyn_cast<SelectInst>(I))
4412 return OptimizeSelectInst(SI);
4414 if (ShuffleVectorInst *SVI = dyn_cast<ShuffleVectorInst>(I))
4415 return OptimizeShuffleVectorInst(SVI);
4417 if (isa<ExtractElementInst>(I))
4418 return OptimizeExtractElementInst(I);
4420 if (BranchInst *BI = dyn_cast<BranchInst>(I)) {
4421 if (TLI && BI->isConditional() && BI->getCondition()->hasOneUse()) {
4422 // Check if the branch condition compares a value agaist zero.
4423 if (ICmpInst *ICI = dyn_cast<ICmpInst>(BI->getCondition())) {
4424 if (ICI->getPredicate() == ICmpInst::ICMP_EQ &&
4425 match(ICI->getOperand(1), m_Zero())) {
4426 BasicBlock *ThenBB = BI->getSuccessor(1);
4427 BasicBlock *EndBB = BI->getSuccessor(0);
4429 // Check if ThenBB is only reachable from this basic block; also,
4430 // check if EndBB has more than one predecessor.
4431 if (ThenBB->getSinglePredecessor() &&
4432 !EndBB->getSinglePredecessor()) {
4433 TerminatorInst *TI = ThenBB->getTerminator();
4435 if (TI->getNumSuccessors() == 1 && TI->getSuccessor(0) == EndBB &&
4436 // Try to speculate calls to intrinsic cttz/ctlz from 'ThenBB'.
4437 OptimizeBranchInst(BI, *TLI)) {
4451 // In this pass we look for GEP and cast instructions that are used
4452 // across basic blocks and rewrite them to improve basic-block-at-a-time
4454 bool CodeGenPrepare::OptimizeBlock(BasicBlock &BB, bool& ModifiedDT) {
4456 bool MadeChange = false;
4458 CurInstIterator = BB.begin();
4459 while (CurInstIterator != BB.end()) {
4460 MadeChange |= OptimizeInst(CurInstIterator++, ModifiedDT);
4464 MadeChange |= DupRetToEnableTailCallOpts(&BB);
4469 // llvm.dbg.value is far away from the value then iSel may not be able
4470 // handle it properly. iSel will drop llvm.dbg.value if it can not
4471 // find a node corresponding to the value.
4472 bool CodeGenPrepare::PlaceDbgValues(Function &F) {
4473 bool MadeChange = false;
4474 for (BasicBlock &BB : F) {
4475 Instruction *PrevNonDbgInst = nullptr;
4476 for (BasicBlock::iterator BI = BB.begin(), BE = BB.end(); BI != BE;) {
4477 Instruction *Insn = BI++;
4478 DbgValueInst *DVI = dyn_cast<DbgValueInst>(Insn);
4479 // Leave dbg.values that refer to an alloca alone. These
4480 // instrinsics describe the address of a variable (= the alloca)
4481 // being taken. They should not be moved next to the alloca
4482 // (and to the beginning of the scope), but rather stay close to
4483 // where said address is used.
4484 if (!DVI || (DVI->getValue() && isa<AllocaInst>(DVI->getValue()))) {
4485 PrevNonDbgInst = Insn;
4489 Instruction *VI = dyn_cast_or_null<Instruction>(DVI->getValue());
4490 if (VI && VI != PrevNonDbgInst && !VI->isTerminator()) {
4491 DEBUG(dbgs() << "Moving Debug Value before :\n" << *DVI << ' ' << *VI);
4492 DVI->removeFromParent();
4493 if (isa<PHINode>(VI))
4494 DVI->insertBefore(VI->getParent()->getFirstInsertionPt());
4496 DVI->insertAfter(VI);
4505 // If there is a sequence that branches based on comparing a single bit
4506 // against zero that can be combined into a single instruction, and the
4507 // target supports folding these into a single instruction, sink the
4508 // mask and compare into the branch uses. Do this before OptimizeBlock ->
4509 // OptimizeInst -> OptimizeCmpExpression, which perturbs the pattern being
4511 bool CodeGenPrepare::sinkAndCmp(Function &F) {
4512 if (!EnableAndCmpSinking)
4514 if (!TLI || !TLI->isMaskAndBranchFoldingLegal())
4516 bool MadeChange = false;
4517 for (Function::iterator I = F.begin(), E = F.end(); I != E; ) {
4518 BasicBlock *BB = I++;
4520 // Does this BB end with the following?
4521 // %andVal = and %val, #single-bit-set
4522 // %icmpVal = icmp %andResult, 0
4523 // br i1 %cmpVal label %dest1, label %dest2"
4524 BranchInst *Brcc = dyn_cast<BranchInst>(BB->getTerminator());
4525 if (!Brcc || !Brcc->isConditional())
4527 ICmpInst *Cmp = dyn_cast<ICmpInst>(Brcc->getOperand(0));
4528 if (!Cmp || Cmp->getParent() != BB)
4530 ConstantInt *Zero = dyn_cast<ConstantInt>(Cmp->getOperand(1));
4531 if (!Zero || !Zero->isZero())
4533 Instruction *And = dyn_cast<Instruction>(Cmp->getOperand(0));
4534 if (!And || And->getOpcode() != Instruction::And || And->getParent() != BB)
4536 ConstantInt* Mask = dyn_cast<ConstantInt>(And->getOperand(1));
4537 if (!Mask || !Mask->getUniqueInteger().isPowerOf2())
4539 DEBUG(dbgs() << "found and; icmp ?,0; brcc\n"); DEBUG(BB->dump());
4541 // Push the "and; icmp" for any users that are conditional branches.
4542 // Since there can only be one branch use per BB, we don't need to keep
4543 // track of which BBs we insert into.
4544 for (Value::use_iterator UI = Cmp->use_begin(), E = Cmp->use_end();
4548 BranchInst *BrccUser = dyn_cast<BranchInst>(*UI);
4550 if (!BrccUser || !BrccUser->isConditional())
4552 BasicBlock *UserBB = BrccUser->getParent();
4553 if (UserBB == BB) continue;
4554 DEBUG(dbgs() << "found Brcc use\n");
4556 // Sink the "and; icmp" to use.
4558 BinaryOperator *NewAnd =
4559 BinaryOperator::CreateAnd(And->getOperand(0), And->getOperand(1), "",
4562 CmpInst::Create(Cmp->getOpcode(), Cmp->getPredicate(), NewAnd, Zero,
4566 DEBUG(BrccUser->getParent()->dump());
4572 /// \brief Retrieve the probabilities of a conditional branch. Returns true on
4573 /// success, or returns false if no or invalid metadata was found.
4574 static bool extractBranchMetadata(BranchInst *BI,
4575 uint64_t &ProbTrue, uint64_t &ProbFalse) {
4576 assert(BI->isConditional() &&
4577 "Looking for probabilities on unconditional branch?");
4578 auto *ProfileData = BI->getMetadata(LLVMContext::MD_prof);
4579 if (!ProfileData || ProfileData->getNumOperands() != 3)
4582 const auto *CITrue =
4583 mdconst::dyn_extract<ConstantInt>(ProfileData->getOperand(1));
4584 const auto *CIFalse =
4585 mdconst::dyn_extract<ConstantInt>(ProfileData->getOperand(2));
4586 if (!CITrue || !CIFalse)
4589 ProbTrue = CITrue->getValue().getZExtValue();
4590 ProbFalse = CIFalse->getValue().getZExtValue();
4595 /// \brief Scale down both weights to fit into uint32_t.
4596 static void scaleWeights(uint64_t &NewTrue, uint64_t &NewFalse) {
4597 uint64_t NewMax = (NewTrue > NewFalse) ? NewTrue : NewFalse;
4598 uint32_t Scale = (NewMax / UINT32_MAX) + 1;
4599 NewTrue = NewTrue / Scale;
4600 NewFalse = NewFalse / Scale;
4603 /// \brief Some targets prefer to split a conditional branch like:
4605 /// %0 = icmp ne i32 %a, 0
4606 /// %1 = icmp ne i32 %b, 0
4607 /// %or.cond = or i1 %0, %1
4608 /// br i1 %or.cond, label %TrueBB, label %FalseBB
4610 /// into multiple branch instructions like:
4613 /// %0 = icmp ne i32 %a, 0
4614 /// br i1 %0, label %TrueBB, label %bb2
4616 /// %1 = icmp ne i32 %b, 0
4617 /// br i1 %1, label %TrueBB, label %FalseBB
4619 /// This usually allows instruction selection to do even further optimizations
4620 /// and combine the compare with the branch instruction. Currently this is
4621 /// applied for targets which have "cheap" jump instructions.
4623 /// FIXME: Remove the (equivalent?) implementation in SelectionDAG.
4625 bool CodeGenPrepare::splitBranchCondition(Function &F) {
4626 if (!TM || TM->Options.EnableFastISel != true ||
4627 !TLI || TLI->isJumpExpensive())
4630 bool MadeChange = false;
4631 for (auto &BB : F) {
4632 // Does this BB end with the following?
4633 // %cond1 = icmp|fcmp|binary instruction ...
4634 // %cond2 = icmp|fcmp|binary instruction ...
4635 // %cond.or = or|and i1 %cond1, cond2
4636 // br i1 %cond.or label %dest1, label %dest2"
4637 BinaryOperator *LogicOp;
4638 BasicBlock *TBB, *FBB;
4639 if (!match(BB.getTerminator(), m_Br(m_OneUse(m_BinOp(LogicOp)), TBB, FBB)))
4643 Value *Cond1, *Cond2;
4644 if (match(LogicOp, m_And(m_OneUse(m_Value(Cond1)),
4645 m_OneUse(m_Value(Cond2)))))
4646 Opc = Instruction::And;
4647 else if (match(LogicOp, m_Or(m_OneUse(m_Value(Cond1)),
4648 m_OneUse(m_Value(Cond2)))))
4649 Opc = Instruction::Or;
4653 if (!match(Cond1, m_CombineOr(m_Cmp(), m_BinOp())) ||
4654 !match(Cond2, m_CombineOr(m_Cmp(), m_BinOp())) )
4657 DEBUG(dbgs() << "Before branch condition splitting\n"; BB.dump());
4660 auto *InsertBefore = std::next(Function::iterator(BB))
4661 .getNodePtrUnchecked();
4662 auto TmpBB = BasicBlock::Create(BB.getContext(),
4663 BB.getName() + ".cond.split",
4664 BB.getParent(), InsertBefore);
4666 // Update original basic block by using the first condition directly by the
4667 // branch instruction and removing the no longer needed and/or instruction.
4668 auto *Br1 = cast<BranchInst>(BB.getTerminator());
4669 Br1->setCondition(Cond1);
4670 LogicOp->eraseFromParent();
4672 // Depending on the conditon we have to either replace the true or the false
4673 // successor of the original branch instruction.
4674 if (Opc == Instruction::And)
4675 Br1->setSuccessor(0, TmpBB);
4677 Br1->setSuccessor(1, TmpBB);
4679 // Fill in the new basic block.
4680 auto *Br2 = IRBuilder<>(TmpBB).CreateCondBr(Cond2, TBB, FBB);
4681 if (auto *I = dyn_cast<Instruction>(Cond2)) {
4682 I->removeFromParent();
4683 I->insertBefore(Br2);
4686 // Update PHI nodes in both successors. The original BB needs to be
4687 // replaced in one succesor's PHI nodes, because the branch comes now from
4688 // the newly generated BB (NewBB). In the other successor we need to add one
4689 // incoming edge to the PHI nodes, because both branch instructions target
4690 // now the same successor. Depending on the original branch condition
4691 // (and/or) we have to swap the successors (TrueDest, FalseDest), so that
4692 // we perfrom the correct update for the PHI nodes.
4693 // This doesn't change the successor order of the just created branch
4694 // instruction (or any other instruction).
4695 if (Opc == Instruction::Or)
4696 std::swap(TBB, FBB);
4698 // Replace the old BB with the new BB.
4699 for (auto &I : *TBB) {
4700 PHINode *PN = dyn_cast<PHINode>(&I);
4704 while ((i = PN->getBasicBlockIndex(&BB)) >= 0)
4705 PN->setIncomingBlock(i, TmpBB);
4708 // Add another incoming edge form the new BB.
4709 for (auto &I : *FBB) {
4710 PHINode *PN = dyn_cast<PHINode>(&I);
4713 auto *Val = PN->getIncomingValueForBlock(&BB);
4714 PN->addIncoming(Val, TmpBB);
4717 // Update the branch weights (from SelectionDAGBuilder::
4718 // FindMergedConditions).
4719 if (Opc == Instruction::Or) {
4720 // Codegen X | Y as:
4729 // We have flexibility in setting Prob for BB1 and Prob for NewBB.
4730 // The requirement is that
4731 // TrueProb for BB1 + (FalseProb for BB1 * TrueProb for TmpBB)
4732 // = TrueProb for orignal BB.
4733 // Assuming the orignal weights are A and B, one choice is to set BB1's
4734 // weights to A and A+2B, and set TmpBB's weights to A and 2B. This choice
4736 // TrueProb for BB1 == FalseProb for BB1 * TrueProb for TmpBB.
4737 // Another choice is to assume TrueProb for BB1 equals to TrueProb for
4738 // TmpBB, but the math is more complicated.
4739 uint64_t TrueWeight, FalseWeight;
4740 if (extractBranchMetadata(Br1, TrueWeight, FalseWeight)) {
4741 uint64_t NewTrueWeight = TrueWeight;
4742 uint64_t NewFalseWeight = TrueWeight + 2 * FalseWeight;
4743 scaleWeights(NewTrueWeight, NewFalseWeight);
4744 Br1->setMetadata(LLVMContext::MD_prof, MDBuilder(Br1->getContext())
4745 .createBranchWeights(TrueWeight, FalseWeight));
4747 NewTrueWeight = TrueWeight;
4748 NewFalseWeight = 2 * FalseWeight;
4749 scaleWeights(NewTrueWeight, NewFalseWeight);
4750 Br2->setMetadata(LLVMContext::MD_prof, MDBuilder(Br2->getContext())
4751 .createBranchWeights(TrueWeight, FalseWeight));
4754 // Codegen X & Y as:
4762 // This requires creation of TmpBB after CurBB.
4764 // We have flexibility in setting Prob for BB1 and Prob for TmpBB.
4765 // The requirement is that
4766 // FalseProb for BB1 + (TrueProb for BB1 * FalseProb for TmpBB)
4767 // = FalseProb for orignal BB.
4768 // Assuming the orignal weights are A and B, one choice is to set BB1's
4769 // weights to 2A+B and B, and set TmpBB's weights to 2A and B. This choice
4771 // FalseProb for BB1 == TrueProb for BB1 * FalseProb for TmpBB.
4772 uint64_t TrueWeight, FalseWeight;
4773 if (extractBranchMetadata(Br1, TrueWeight, FalseWeight)) {
4774 uint64_t NewTrueWeight = 2 * TrueWeight + FalseWeight;
4775 uint64_t NewFalseWeight = FalseWeight;
4776 scaleWeights(NewTrueWeight, NewFalseWeight);
4777 Br1->setMetadata(LLVMContext::MD_prof, MDBuilder(Br1->getContext())
4778 .createBranchWeights(TrueWeight, FalseWeight));
4780 NewTrueWeight = 2 * TrueWeight;
4781 NewFalseWeight = FalseWeight;
4782 scaleWeights(NewTrueWeight, NewFalseWeight);
4783 Br2->setMetadata(LLVMContext::MD_prof, MDBuilder(Br2->getContext())
4784 .createBranchWeights(TrueWeight, FalseWeight));
4788 // Request DOM Tree update.
4789 // Note: No point in getting fancy here, since the DT info is never
4790 // available to CodeGenPrepare and the existing update code is broken
4796 DEBUG(dbgs() << "After branch condition splitting\n"; BB.dump();