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 #define DEBUG_TYPE "codegenprepare"
17 #include "llvm/CodeGen/Passes.h"
18 #include "llvm/ADT/DenseMap.h"
19 #include "llvm/ADT/SmallSet.h"
20 #include "llvm/ADT/Statistic.h"
21 #include "llvm/Analysis/InstructionSimplify.h"
22 #include "llvm/IR/CallSite.h"
23 #include "llvm/IR/Constants.h"
24 #include "llvm/IR/DataLayout.h"
25 #include "llvm/IR/DerivedTypes.h"
26 #include "llvm/IR/Dominators.h"
27 #include "llvm/IR/Function.h"
28 #include "llvm/IR/GetElementPtrTypeIterator.h"
29 #include "llvm/IR/IRBuilder.h"
30 #include "llvm/IR/InlineAsm.h"
31 #include "llvm/IR/Instructions.h"
32 #include "llvm/IR/IntrinsicInst.h"
33 #include "llvm/IR/PatternMatch.h"
34 #include "llvm/IR/ValueHandle.h"
35 #include "llvm/IR/ValueMap.h"
36 #include "llvm/Pass.h"
37 #include "llvm/Support/CommandLine.h"
38 #include "llvm/Support/Debug.h"
39 #include "llvm/Support/raw_ostream.h"
40 #include "llvm/Target/TargetLibraryInfo.h"
41 #include "llvm/Target/TargetLowering.h"
42 #include "llvm/Target/TargetSubtargetInfo.h"
43 #include "llvm/Transforms/Utils/BasicBlockUtils.h"
44 #include "llvm/Transforms/Utils/BuildLibCalls.h"
45 #include "llvm/Transforms/Utils/BypassSlowDivision.h"
46 #include "llvm/Transforms/Utils/Local.h"
48 using namespace llvm::PatternMatch;
50 STATISTIC(NumBlocksElim, "Number of blocks eliminated");
51 STATISTIC(NumPHIsElim, "Number of trivial PHIs eliminated");
52 STATISTIC(NumGEPsElim, "Number of GEPs converted to casts");
53 STATISTIC(NumCmpUses, "Number of uses of Cmp expressions replaced with uses of "
55 STATISTIC(NumCastUses, "Number of uses of Cast expressions replaced with uses "
57 STATISTIC(NumMemoryInsts, "Number of memory instructions whose address "
58 "computations were sunk");
59 STATISTIC(NumExtsMoved, "Number of [s|z]ext instructions combined with loads");
60 STATISTIC(NumExtUses, "Number of uses of [s|z]ext instructions optimized");
61 STATISTIC(NumRetsDup, "Number of return instructions duplicated");
62 STATISTIC(NumDbgValueMoved, "Number of debug value instructions moved");
63 STATISTIC(NumSelectsExpanded, "Number of selects turned into branches");
64 STATISTIC(NumAndCmpsMoved, "Number of and/cmp's pushed into branches");
66 static cl::opt<bool> DisableBranchOpts(
67 "disable-cgp-branch-opts", cl::Hidden, cl::init(false),
68 cl::desc("Disable branch optimizations in CodeGenPrepare"));
70 static cl::opt<bool> DisableSelectToBranch(
71 "disable-cgp-select2branch", cl::Hidden, cl::init(false),
72 cl::desc("Disable select to branch conversion."));
74 static cl::opt<bool> AddrSinkUsingGEPs(
75 "addr-sink-using-gep", cl::Hidden, cl::init(false),
76 cl::desc("Address sinking in CGP using GEPs."));
78 static cl::opt<bool> EnableAndCmpSinking(
79 "enable-andcmp-sinking", cl::Hidden, cl::init(true),
80 cl::desc("Enable sinkinig and/cmp into branches."));
83 typedef SmallPtrSet<Instruction *, 16> SetOfInstrs;
84 typedef DenseMap<Instruction *, Type *> InstrToOrigTy;
86 class CodeGenPrepare : public FunctionPass {
87 /// TLI - Keep a pointer of a TargetLowering to consult for determining
88 /// transformation profitability.
89 const TargetMachine *TM;
90 const TargetLowering *TLI;
91 const TargetLibraryInfo *TLInfo;
94 /// CurInstIterator - As we scan instructions optimizing them, this is the
95 /// next instruction to optimize. Xforms that can invalidate this should
97 BasicBlock::iterator CurInstIterator;
99 /// Keeps track of non-local addresses that have been sunk into a block.
100 /// This allows us to avoid inserting duplicate code for blocks with
101 /// multiple load/stores of the same address.
102 ValueMap<Value*, Value*> SunkAddrs;
104 /// Keeps track of all truncates inserted for the current function.
105 SetOfInstrs InsertedTruncsSet;
106 /// Keeps track of the type of the related instruction before their
107 /// promotion for the current function.
108 InstrToOrigTy PromotedInsts;
110 /// ModifiedDT - If CFG is modified in anyway, dominator tree may need to
114 /// OptSize - True if optimizing for size.
118 static char ID; // Pass identification, replacement for typeid
119 explicit CodeGenPrepare(const TargetMachine *TM = nullptr)
120 : FunctionPass(ID), TM(TM), TLI(nullptr) {
121 initializeCodeGenPreparePass(*PassRegistry::getPassRegistry());
123 bool runOnFunction(Function &F) override;
125 const char *getPassName() const override { return "CodeGen Prepare"; }
127 void getAnalysisUsage(AnalysisUsage &AU) const override {
128 AU.addPreserved<DominatorTreeWrapperPass>();
129 AU.addRequired<TargetLibraryInfo>();
133 bool EliminateFallThrough(Function &F);
134 bool EliminateMostlyEmptyBlocks(Function &F);
135 bool CanMergeBlocks(const BasicBlock *BB, const BasicBlock *DestBB) const;
136 void EliminateMostlyEmptyBlock(BasicBlock *BB);
137 bool OptimizeBlock(BasicBlock &BB);
138 bool OptimizeInst(Instruction *I);
139 bool OptimizeMemoryInst(Instruction *I, Value *Addr, Type *AccessTy);
140 bool OptimizeInlineAsmInst(CallInst *CS);
141 bool OptimizeCallInst(CallInst *CI);
142 bool MoveExtToFormExtLoad(Instruction *I);
143 bool OptimizeExtUses(Instruction *I);
144 bool OptimizeSelectInst(SelectInst *SI);
145 bool OptimizeShuffleVectorInst(ShuffleVectorInst *SI);
146 bool DupRetToEnableTailCallOpts(BasicBlock *BB);
147 bool PlaceDbgValues(Function &F);
148 bool sinkAndCmp(Function &F);
152 char CodeGenPrepare::ID = 0;
153 static void *initializeCodeGenPreparePassOnce(PassRegistry &Registry) {
154 initializeTargetLibraryInfoPass(Registry);
155 PassInfo *PI = new PassInfo(
156 "Optimize for code generation", "codegenprepare", &CodeGenPrepare::ID,
157 PassInfo::NormalCtor_t(callDefaultCtor<CodeGenPrepare>), false, false,
158 PassInfo::TargetMachineCtor_t(callTargetMachineCtor<CodeGenPrepare>));
159 Registry.registerPass(*PI, true);
163 void llvm::initializeCodeGenPreparePass(PassRegistry &Registry) {
164 CALL_ONCE_INITIALIZATION(initializeCodeGenPreparePassOnce)
167 FunctionPass *llvm::createCodeGenPreparePass(const TargetMachine *TM) {
168 return new CodeGenPrepare(TM);
171 bool CodeGenPrepare::runOnFunction(Function &F) {
172 if (skipOptnoneFunction(F))
175 bool EverMadeChange = false;
176 // Clear per function information.
177 InsertedTruncsSet.clear();
178 PromotedInsts.clear();
181 if (TM) TLI = TM->getTargetLowering();
182 TLInfo = &getAnalysis<TargetLibraryInfo>();
183 DominatorTreeWrapperPass *DTWP =
184 getAnalysisIfAvailable<DominatorTreeWrapperPass>();
185 DT = DTWP ? &DTWP->getDomTree() : nullptr;
186 OptSize = F.getAttributes().hasAttribute(AttributeSet::FunctionIndex,
187 Attribute::OptimizeForSize);
189 /// This optimization identifies DIV instructions that can be
190 /// profitably bypassed and carried out with a shorter, faster divide.
191 if (!OptSize && TLI && TLI->isSlowDivBypassed()) {
192 const DenseMap<unsigned int, unsigned int> &BypassWidths =
193 TLI->getBypassSlowDivWidths();
194 for (Function::iterator I = F.begin(); I != F.end(); I++)
195 EverMadeChange |= bypassSlowDivision(F, I, BypassWidths);
198 // Eliminate blocks that contain only PHI nodes and an
199 // unconditional branch.
200 EverMadeChange |= EliminateMostlyEmptyBlocks(F);
202 // llvm.dbg.value is far away from the value then iSel may not be able
203 // handle it properly. iSel will drop llvm.dbg.value if it can not
204 // find a node corresponding to the value.
205 EverMadeChange |= PlaceDbgValues(F);
207 // If there is a mask, compare against zero, and branch that can be combined
208 // into a single target instruction, push the mask and compare into branch
209 // users. Do this before OptimizeBlock -> OptimizeInst ->
210 // OptimizeCmpExpression, which perturbs the pattern being searched for.
211 if (!DisableBranchOpts)
212 EverMadeChange |= sinkAndCmp(F);
214 bool MadeChange = true;
217 for (Function::iterator I = F.begin(); I != F.end(); ) {
218 BasicBlock *BB = I++;
219 MadeChange |= OptimizeBlock(*BB);
221 EverMadeChange |= MadeChange;
226 if (!DisableBranchOpts) {
228 SmallPtrSet<BasicBlock*, 8> WorkList;
229 for (Function::iterator BB = F.begin(), E = F.end(); BB != E; ++BB) {
230 SmallVector<BasicBlock*, 2> Successors(succ_begin(BB), succ_end(BB));
231 MadeChange |= ConstantFoldTerminator(BB, true);
232 if (!MadeChange) continue;
234 for (SmallVectorImpl<BasicBlock*>::iterator
235 II = Successors.begin(), IE = Successors.end(); II != IE; ++II)
236 if (pred_begin(*II) == pred_end(*II))
237 WorkList.insert(*II);
240 // Delete the dead blocks and any of their dead successors.
241 MadeChange |= !WorkList.empty();
242 while (!WorkList.empty()) {
243 BasicBlock *BB = *WorkList.begin();
245 SmallVector<BasicBlock*, 2> Successors(succ_begin(BB), succ_end(BB));
249 for (SmallVectorImpl<BasicBlock*>::iterator
250 II = Successors.begin(), IE = Successors.end(); II != IE; ++II)
251 if (pred_begin(*II) == pred_end(*II))
252 WorkList.insert(*II);
255 // Merge pairs of basic blocks with unconditional branches, connected by
257 if (EverMadeChange || MadeChange)
258 MadeChange |= EliminateFallThrough(F);
262 EverMadeChange |= MadeChange;
265 if (ModifiedDT && DT)
268 return EverMadeChange;
271 /// EliminateFallThrough - Merge basic blocks which are connected
272 /// by a single edge, where one of the basic blocks has a single successor
273 /// pointing to the other basic block, which has a single predecessor.
274 bool CodeGenPrepare::EliminateFallThrough(Function &F) {
275 bool Changed = false;
276 // Scan all of the blocks in the function, except for the entry block.
277 for (Function::iterator I = std::next(F.begin()), E = F.end(); I != E;) {
278 BasicBlock *BB = I++;
279 // If the destination block has a single pred, then this is a trivial
280 // edge, just collapse it.
281 BasicBlock *SinglePred = BB->getSinglePredecessor();
283 // Don't merge if BB's address is taken.
284 if (!SinglePred || SinglePred == BB || BB->hasAddressTaken()) continue;
286 BranchInst *Term = dyn_cast<BranchInst>(SinglePred->getTerminator());
287 if (Term && !Term->isConditional()) {
289 DEBUG(dbgs() << "To merge:\n"<< *SinglePred << "\n\n\n");
290 // Remember if SinglePred was the entry block of the function.
291 // If so, we will need to move BB back to the entry position.
292 bool isEntry = SinglePred == &SinglePred->getParent()->getEntryBlock();
293 MergeBasicBlockIntoOnlyPred(BB, this);
295 if (isEntry && BB != &BB->getParent()->getEntryBlock())
296 BB->moveBefore(&BB->getParent()->getEntryBlock());
298 // We have erased a block. Update the iterator.
305 /// EliminateMostlyEmptyBlocks - eliminate blocks that contain only PHI nodes,
306 /// debug info directives, and an unconditional branch. Passes before isel
307 /// (e.g. LSR/loopsimplify) often split edges in ways that are non-optimal for
308 /// isel. Start by eliminating these blocks so we can split them the way we
310 bool CodeGenPrepare::EliminateMostlyEmptyBlocks(Function &F) {
311 bool MadeChange = false;
312 // Note that this intentionally skips the entry block.
313 for (Function::iterator I = std::next(F.begin()), E = F.end(); I != E;) {
314 BasicBlock *BB = I++;
316 // If this block doesn't end with an uncond branch, ignore it.
317 BranchInst *BI = dyn_cast<BranchInst>(BB->getTerminator());
318 if (!BI || !BI->isUnconditional())
321 // If the instruction before the branch (skipping debug info) isn't a phi
322 // node, then other stuff is happening here.
323 BasicBlock::iterator BBI = BI;
324 if (BBI != BB->begin()) {
326 while (isa<DbgInfoIntrinsic>(BBI)) {
327 if (BBI == BB->begin())
331 if (!isa<DbgInfoIntrinsic>(BBI) && !isa<PHINode>(BBI))
335 // Do not break infinite loops.
336 BasicBlock *DestBB = BI->getSuccessor(0);
340 if (!CanMergeBlocks(BB, DestBB))
343 EliminateMostlyEmptyBlock(BB);
349 /// CanMergeBlocks - Return true if we can merge BB into DestBB if there is a
350 /// single uncond branch between them, and BB contains no other non-phi
352 bool CodeGenPrepare::CanMergeBlocks(const BasicBlock *BB,
353 const BasicBlock *DestBB) const {
354 // We only want to eliminate blocks whose phi nodes are used by phi nodes in
355 // the successor. If there are more complex condition (e.g. preheaders),
356 // don't mess around with them.
357 BasicBlock::const_iterator BBI = BB->begin();
358 while (const PHINode *PN = dyn_cast<PHINode>(BBI++)) {
359 for (const User *U : PN->users()) {
360 const Instruction *UI = cast<Instruction>(U);
361 if (UI->getParent() != DestBB || !isa<PHINode>(UI))
363 // If User is inside DestBB block and it is a PHINode then check
364 // incoming value. If incoming value is not from BB then this is
365 // a complex condition (e.g. preheaders) we want to avoid here.
366 if (UI->getParent() == DestBB) {
367 if (const PHINode *UPN = dyn_cast<PHINode>(UI))
368 for (unsigned I = 0, E = UPN->getNumIncomingValues(); I != E; ++I) {
369 Instruction *Insn = dyn_cast<Instruction>(UPN->getIncomingValue(I));
370 if (Insn && Insn->getParent() == BB &&
371 Insn->getParent() != UPN->getIncomingBlock(I))
378 // If BB and DestBB contain any common predecessors, then the phi nodes in BB
379 // and DestBB may have conflicting incoming values for the block. If so, we
380 // can't merge the block.
381 const PHINode *DestBBPN = dyn_cast<PHINode>(DestBB->begin());
382 if (!DestBBPN) return true; // no conflict.
384 // Collect the preds of BB.
385 SmallPtrSet<const BasicBlock*, 16> BBPreds;
386 if (const PHINode *BBPN = dyn_cast<PHINode>(BB->begin())) {
387 // It is faster to get preds from a PHI than with pred_iterator.
388 for (unsigned i = 0, e = BBPN->getNumIncomingValues(); i != e; ++i)
389 BBPreds.insert(BBPN->getIncomingBlock(i));
391 BBPreds.insert(pred_begin(BB), pred_end(BB));
394 // Walk the preds of DestBB.
395 for (unsigned i = 0, e = DestBBPN->getNumIncomingValues(); i != e; ++i) {
396 BasicBlock *Pred = DestBBPN->getIncomingBlock(i);
397 if (BBPreds.count(Pred)) { // Common predecessor?
398 BBI = DestBB->begin();
399 while (const PHINode *PN = dyn_cast<PHINode>(BBI++)) {
400 const Value *V1 = PN->getIncomingValueForBlock(Pred);
401 const Value *V2 = PN->getIncomingValueForBlock(BB);
403 // If V2 is a phi node in BB, look up what the mapped value will be.
404 if (const PHINode *V2PN = dyn_cast<PHINode>(V2))
405 if (V2PN->getParent() == BB)
406 V2 = V2PN->getIncomingValueForBlock(Pred);
408 // If there is a conflict, bail out.
409 if (V1 != V2) return false;
418 /// EliminateMostlyEmptyBlock - Eliminate a basic block that have only phi's and
419 /// an unconditional branch in it.
420 void CodeGenPrepare::EliminateMostlyEmptyBlock(BasicBlock *BB) {
421 BranchInst *BI = cast<BranchInst>(BB->getTerminator());
422 BasicBlock *DestBB = BI->getSuccessor(0);
424 DEBUG(dbgs() << "MERGING MOSTLY EMPTY BLOCKS - BEFORE:\n" << *BB << *DestBB);
426 // If the destination block has a single pred, then this is a trivial edge,
428 if (BasicBlock *SinglePred = DestBB->getSinglePredecessor()) {
429 if (SinglePred != DestBB) {
430 // Remember if SinglePred was the entry block of the function. If so, we
431 // will need to move BB back to the entry position.
432 bool isEntry = SinglePred == &SinglePred->getParent()->getEntryBlock();
433 MergeBasicBlockIntoOnlyPred(DestBB, this);
435 if (isEntry && BB != &BB->getParent()->getEntryBlock())
436 BB->moveBefore(&BB->getParent()->getEntryBlock());
438 DEBUG(dbgs() << "AFTER:\n" << *DestBB << "\n\n\n");
443 // Otherwise, we have multiple predecessors of BB. Update the PHIs in DestBB
444 // to handle the new incoming edges it is about to have.
446 for (BasicBlock::iterator BBI = DestBB->begin();
447 (PN = dyn_cast<PHINode>(BBI)); ++BBI) {
448 // Remove the incoming value for BB, and remember it.
449 Value *InVal = PN->removeIncomingValue(BB, false);
451 // Two options: either the InVal is a phi node defined in BB or it is some
452 // value that dominates BB.
453 PHINode *InValPhi = dyn_cast<PHINode>(InVal);
454 if (InValPhi && InValPhi->getParent() == BB) {
455 // Add all of the input values of the input PHI as inputs of this phi.
456 for (unsigned i = 0, e = InValPhi->getNumIncomingValues(); i != e; ++i)
457 PN->addIncoming(InValPhi->getIncomingValue(i),
458 InValPhi->getIncomingBlock(i));
460 // Otherwise, add one instance of the dominating value for each edge that
461 // we will be adding.
462 if (PHINode *BBPN = dyn_cast<PHINode>(BB->begin())) {
463 for (unsigned i = 0, e = BBPN->getNumIncomingValues(); i != e; ++i)
464 PN->addIncoming(InVal, BBPN->getIncomingBlock(i));
466 for (pred_iterator PI = pred_begin(BB), E = pred_end(BB); PI != E; ++PI)
467 PN->addIncoming(InVal, *PI);
472 // The PHIs are now updated, change everything that refers to BB to use
473 // DestBB and remove BB.
474 BB->replaceAllUsesWith(DestBB);
475 if (DT && !ModifiedDT) {
476 BasicBlock *BBIDom = DT->getNode(BB)->getIDom()->getBlock();
477 BasicBlock *DestBBIDom = DT->getNode(DestBB)->getIDom()->getBlock();
478 BasicBlock *NewIDom = DT->findNearestCommonDominator(BBIDom, DestBBIDom);
479 DT->changeImmediateDominator(DestBB, NewIDom);
482 BB->eraseFromParent();
485 DEBUG(dbgs() << "AFTER:\n" << *DestBB << "\n\n\n");
488 /// SinkCast - Sink the specified cast instruction into its user blocks
489 static bool SinkCast(CastInst *CI) {
490 BasicBlock *DefBB = CI->getParent();
492 /// InsertedCasts - Only insert a cast in each block once.
493 DenseMap<BasicBlock*, CastInst*> InsertedCasts;
495 bool MadeChange = false;
496 for (Value::user_iterator UI = CI->user_begin(), E = CI->user_end();
498 Use &TheUse = UI.getUse();
499 Instruction *User = cast<Instruction>(*UI);
501 // Figure out which BB this cast is used in. For PHI's this is the
502 // appropriate predecessor block.
503 BasicBlock *UserBB = User->getParent();
504 if (PHINode *PN = dyn_cast<PHINode>(User)) {
505 UserBB = PN->getIncomingBlock(TheUse);
508 // Preincrement use iterator so we don't invalidate it.
511 // If this user is in the same block as the cast, don't change the cast.
512 if (UserBB == DefBB) continue;
514 // If we have already inserted a cast into this block, use it.
515 CastInst *&InsertedCast = InsertedCasts[UserBB];
518 BasicBlock::iterator InsertPt = UserBB->getFirstInsertionPt();
520 CastInst::Create(CI->getOpcode(), CI->getOperand(0), CI->getType(), "",
525 // Replace a use of the cast with a use of the new cast.
526 TheUse = InsertedCast;
530 // If we removed all uses, nuke the cast.
531 if (CI->use_empty()) {
532 CI->eraseFromParent();
539 /// OptimizeNoopCopyExpression - If the specified cast instruction is a noop
540 /// copy (e.g. it's casting from one pointer type to another, i32->i8 on PPC),
541 /// sink it into user blocks to reduce the number of virtual
542 /// registers that must be created and coalesced.
544 /// Return true if any changes are made.
546 static bool OptimizeNoopCopyExpression(CastInst *CI, const TargetLowering &TLI){
547 // If this is a noop copy,
548 EVT SrcVT = TLI.getValueType(CI->getOperand(0)->getType());
549 EVT DstVT = TLI.getValueType(CI->getType());
551 // This is an fp<->int conversion?
552 if (SrcVT.isInteger() != DstVT.isInteger())
555 // If this is an extension, it will be a zero or sign extension, which
557 if (SrcVT.bitsLT(DstVT)) return false;
559 // If these values will be promoted, find out what they will be promoted
560 // to. This helps us consider truncates on PPC as noop copies when they
562 if (TLI.getTypeAction(CI->getContext(), SrcVT) ==
563 TargetLowering::TypePromoteInteger)
564 SrcVT = TLI.getTypeToTransformTo(CI->getContext(), SrcVT);
565 if (TLI.getTypeAction(CI->getContext(), DstVT) ==
566 TargetLowering::TypePromoteInteger)
567 DstVT = TLI.getTypeToTransformTo(CI->getContext(), DstVT);
569 // If, after promotion, these are the same types, this is a noop copy.
576 /// OptimizeCmpExpression - sink the given CmpInst into user blocks to reduce
577 /// the number of virtual registers that must be created and coalesced. This is
578 /// a clear win except on targets with multiple condition code registers
579 /// (PowerPC), where it might lose; some adjustment may be wanted there.
581 /// Return true if any changes are made.
582 static bool OptimizeCmpExpression(CmpInst *CI) {
583 BasicBlock *DefBB = CI->getParent();
585 /// InsertedCmp - Only insert a cmp in each block once.
586 DenseMap<BasicBlock*, CmpInst*> InsertedCmps;
588 bool MadeChange = false;
589 for (Value::user_iterator UI = CI->user_begin(), E = CI->user_end();
591 Use &TheUse = UI.getUse();
592 Instruction *User = cast<Instruction>(*UI);
594 // Preincrement use iterator so we don't invalidate it.
597 // Don't bother for PHI nodes.
598 if (isa<PHINode>(User))
601 // Figure out which BB this cmp is used in.
602 BasicBlock *UserBB = User->getParent();
604 // If this user is in the same block as the cmp, don't change the cmp.
605 if (UserBB == DefBB) continue;
607 // If we have already inserted a cmp into this block, use it.
608 CmpInst *&InsertedCmp = InsertedCmps[UserBB];
611 BasicBlock::iterator InsertPt = UserBB->getFirstInsertionPt();
613 CmpInst::Create(CI->getOpcode(),
614 CI->getPredicate(), CI->getOperand(0),
615 CI->getOperand(1), "", InsertPt);
619 // Replace a use of the cmp with a use of the new cmp.
620 TheUse = InsertedCmp;
624 // If we removed all uses, nuke the cmp.
626 CI->eraseFromParent();
632 class CodeGenPrepareFortifiedLibCalls : public SimplifyFortifiedLibCalls {
634 void replaceCall(Value *With) override {
635 CI->replaceAllUsesWith(With);
636 CI->eraseFromParent();
638 bool isFoldable(unsigned SizeCIOp, unsigned, bool) const override {
639 if (ConstantInt *SizeCI =
640 dyn_cast<ConstantInt>(CI->getArgOperand(SizeCIOp)))
641 return SizeCI->isAllOnesValue();
645 } // end anonymous namespace
647 bool CodeGenPrepare::OptimizeCallInst(CallInst *CI) {
648 BasicBlock *BB = CI->getParent();
650 // Lower inline assembly if we can.
651 // If we found an inline asm expession, and if the target knows how to
652 // lower it to normal LLVM code, do so now.
653 if (TLI && isa<InlineAsm>(CI->getCalledValue())) {
654 if (TLI->ExpandInlineAsm(CI)) {
655 // Avoid invalidating the iterator.
656 CurInstIterator = BB->begin();
657 // Avoid processing instructions out of order, which could cause
658 // reuse before a value is defined.
662 // Sink address computing for memory operands into the block.
663 if (OptimizeInlineAsmInst(CI))
667 // Lower all uses of llvm.objectsize.*
668 IntrinsicInst *II = dyn_cast<IntrinsicInst>(CI);
669 if (II && II->getIntrinsicID() == Intrinsic::objectsize) {
670 bool Min = (cast<ConstantInt>(II->getArgOperand(1))->getZExtValue() == 1);
671 Type *ReturnTy = CI->getType();
672 Constant *RetVal = ConstantInt::get(ReturnTy, Min ? 0 : -1ULL);
674 // Substituting this can cause recursive simplifications, which can
675 // invalidate our iterator. Use a WeakVH to hold onto it in case this
677 WeakVH IterHandle(CurInstIterator);
679 replaceAndRecursivelySimplify(CI, RetVal,
680 TLI ? TLI->getDataLayout() : nullptr,
681 TLInfo, ModifiedDT ? nullptr : DT);
683 // If the iterator instruction was recursively deleted, start over at the
684 // start of the block.
685 if (IterHandle != CurInstIterator) {
686 CurInstIterator = BB->begin();
691 // Lower all uses of llvm.safe.[us]{div|rem}...
693 (II->getIntrinsicID() == Intrinsic::safe_sdiv ||
694 II->getIntrinsicID() == Intrinsic::safe_udiv ||
695 II->getIntrinsicID() == Intrinsic::safe_srem ||
696 II->getIntrinsicID() == Intrinsic::safe_urem)) {
698 // result_struct = type {iN, i1}
699 // %R = call result_struct llvm.safe.sdiv.iN(iN %x, iN %y)
700 // Expand it to actual IR, which produces result to the same variable %R.
701 // First element of the result %R.1 is the result of division, second
702 // element shows whether the division was correct or not.
703 // If %y is 0, %R.1 is 0, %R.2 is 1. (1)
704 // If %x is minSignedValue and %y is -1, %R.1 is %x, %R.2 is 1. (2)
705 // In other cases %R.1 is (sdiv %x, %y), %R.2 is 0. (3)
707 // Similar applies to srem, udiv, and urem builtins, except that in unsigned
708 // variants we don't check condition (2).
711 BinaryOperator::BinaryOps Op;
712 switch (II->getIntrinsicID()) {
713 case Intrinsic::safe_sdiv:
715 Op = Instruction::SDiv;
717 case Intrinsic::safe_udiv:
719 Op = Instruction::UDiv;
721 case Intrinsic::safe_srem:
723 Op = Instruction::SRem;
725 case Intrinsic::safe_urem:
727 Op = Instruction::URem;
730 llvm_unreachable("Only Div/Rem intrinsics are handled here.");
733 Value *LHS = II->getOperand(0), *RHS = II->getOperand(1);
734 bool DivWellDefined = TLI && TLI->isDivWellDefined();
736 bool ResultNeeded[2] = {false, false};
737 SmallVector<User*, 1> ResultsUsers[2];
738 bool BadCase = false;
739 for (User *U: II->users()) {
740 ExtractValueInst *EVI = dyn_cast<ExtractValueInst>(U);
741 if (!EVI || EVI->getNumIndices() > 1 || EVI->getIndices()[0] > 1) {
745 ResultNeeded[EVI->getIndices()[0]] = true;
746 ResultsUsers[EVI->getIndices()[0]].push_back(U);
748 // Behave conservatively, if there is an unusual user of the results.
750 ResultNeeded[0] = ResultNeeded[1] = true;
752 // Early exit if non of the results is ever used.
753 if (!ResultNeeded[0] && !ResultNeeded[1]) {
754 II->eraseFromParent();
758 // Early exit if the second result (flag) isn't used and target
759 // div-instruction computes exactly what we want to get as the first result
761 if (ResultNeeded[0] && !ResultNeeded[1] && DivWellDefined) {
762 BinaryOperator *Div = BinaryOperator::Create(Op, LHS, RHS);
763 Div->insertAfter(II);
764 for (User *U: ResultsUsers[0]) {
765 Instruction *UserInst = dyn_cast<Instruction>(U);
766 assert(UserInst && "Unexpected null-instruction");
767 UserInst->replaceAllUsesWith(Div);
768 UserInst->eraseFromParent();
770 II->eraseFromParent();
771 CurInstIterator = Div;
776 Value *MinusOne = Constant::getAllOnesValue(LHS->getType());
777 Value *Zero = Constant::getNullValue(LHS->getType());
779 // Split the original BB and create other basic blocks that will be used
781 BasicBlock *StartBB = II->getParent();
782 BasicBlock::iterator SplitPt = ++(BasicBlock::iterator(II));
783 BasicBlock *NextBB = StartBB->splitBasicBlock(SplitPt, "div.end");
785 BasicBlock *DivByZeroBB;
786 DivByZeroBB = BasicBlock::Create(II->getContext(), "div.divz",
787 NextBB->getParent(), NextBB);
788 BranchInst::Create(NextBB, DivByZeroBB);
789 BasicBlock *DivBB = BasicBlock::Create(II->getContext(), "div.div",
790 NextBB->getParent(), NextBB);
791 BranchInst::Create(NextBB, DivBB);
793 // For signed variants, check the condition (2):
794 // LHS == SignedMinValue, RHS == -1.
797 BasicBlock *ChkDivMinBB;
798 BasicBlock *DivMinBB;
801 APInt SignedMinValue =
802 APInt::getSignedMinValue(LHS->getType()->getPrimitiveSizeInBits());
803 MinValue = Constant::getIntegerValue(LHS->getType(), SignedMinValue);
804 ChkDivMinBB = BasicBlock::Create(II->getContext(), "div.chkdivmin",
805 NextBB->getParent(), NextBB);
806 BranchInst::Create(NextBB, ChkDivMinBB);
807 DivMinBB = BasicBlock::Create(II->getContext(), "div.divmin",
808 NextBB->getParent(), NextBB);
809 BranchInst::Create(NextBB, DivMinBB);
810 CmpMinusOne = CmpInst::Create(Instruction::ICmp, CmpInst::ICMP_EQ,
811 RHS, MinusOne, "cmp.rhs.minus.one",
812 ChkDivMinBB->getTerminator());
813 CmpMinValue = CmpInst::Create(Instruction::ICmp, CmpInst::ICMP_EQ,
814 LHS, MinValue, "cmp.lhs.signed.min",
815 ChkDivMinBB->getTerminator());
816 BinaryOperator *CmpSignedOvf = BinaryOperator::Create(Instruction::And,
819 // Here we're interested in the case when both %x is TMin and %y is -1.
820 // In this case the result will overflow.
821 // If that's not the case, we can perform usual division. These blocks
822 // will be inserted after DivByZero, so the division will be safe.
823 CmpSignedOvf->insertBefore(ChkDivMinBB->getTerminator());
824 BranchInst::Create(DivMinBB, DivBB, CmpSignedOvf,
825 ChkDivMinBB->getTerminator());
826 ChkDivMinBB->getTerminator()->eraseFromParent();
829 // Check the condition (1):
831 Value *CmpDivZero = CmpInst::Create(Instruction::ICmp, CmpInst::ICMP_EQ,
832 RHS, Zero, "cmp.rhs.zero",
833 StartBB->getTerminator());
835 // If RHS != 0, we want to check condition (2) in signed case, or proceed
836 // to usual division in unsigned case.
837 BranchInst::Create(DivByZeroBB, IsSigned ? ChkDivMinBB : DivBB, CmpDivZero,
838 StartBB->getTerminator());
839 StartBB->getTerminator()->eraseFromParent();
841 // At the moment we have all the control flow created. We just need to
842 // insert DIV and PHI (if needed) to get the result value.
843 Instruction *DivRes, *FlagRes;
844 Instruction *InsPoint = nullptr;
845 if (ResultNeeded[0]) {
846 BinaryOperator *Div = BinaryOperator::Create(Op, LHS, RHS);
847 if (DivWellDefined) {
848 // The result value is the result of DIV operation placed right at the
849 // original place of the intrinsic.
850 Div->insertAfter(II);
853 // The result is a PHI-node.
854 Div->insertBefore(DivBB->getTerminator());
856 PHINode::Create(LHS->getType(), IsSigned ? 3 : 2, "div.res.phi",
858 DivResPN->addIncoming(Div, DivBB);
859 DivResPN->addIncoming(Zero, DivByZeroBB);
861 DivResPN->addIncoming(MinValue, DivMinBB);
867 // Prepare a value for the second result (flag) if it is needed.
868 if (ResultNeeded[1]) {
869 Type *FlagTy = II->getType()->getStructElementType(1);
871 PHINode::Create(FlagTy, IsSigned ? 3 : 2, "div.flag.phi",
873 FlagResPN->addIncoming(Constant::getNullValue(FlagTy), DivBB);
874 FlagResPN->addIncoming(Constant::getAllOnesValue(FlagTy), DivByZeroBB);
876 FlagResPN->addIncoming(Constant::getAllOnesValue(FlagTy), DivMinBB);
882 // If possible, propagate the results to the user. Otherwise, create alloca,
883 // and create a struct with the results on stack.
885 if (ResultNeeded[0]) {
886 for (User *U: ResultsUsers[0]) {
887 Instruction *UserInst = dyn_cast<Instruction>(U);
888 assert(UserInst && "Unexpected null-instruction");
889 UserInst->replaceAllUsesWith(DivRes);
890 UserInst->eraseFromParent();
893 if (ResultNeeded[1]) {
894 for (User *FlagU: ResultsUsers[1]) {
895 Instruction *FlagUInst = dyn_cast<Instruction>(FlagU);
896 FlagUInst->replaceAllUsesWith(FlagRes);
897 FlagUInst->eraseFromParent();
901 // Create alloca, store our new values to it, and then load the final
903 Constant *Idx0 = ConstantInt::get(Type::getInt32Ty(II->getContext()), 0);
904 Constant *Idx1 = ConstantInt::get(Type::getInt32Ty(II->getContext()), 1);
905 Value *Idxs_DivRes[2] = {Idx0, Idx0};
906 Value *Idxs_FlagRes[2] = {Idx0, Idx1};
907 Value *NewRes = new llvm::AllocaInst(II->getType(), 0, "div.res.ptr", II);
908 Instruction *ResDivAddr = GetElementPtrInst::Create(NewRes, Idxs_DivRes);
909 Instruction *ResFlagAddr =
910 GetElementPtrInst::Create(NewRes, Idxs_FlagRes);
911 ResDivAddr->insertAfter(InsPoint);
912 ResFlagAddr->insertAfter(ResDivAddr);
913 StoreInst *StoreResDiv = new StoreInst(DivRes, ResDivAddr);
914 StoreInst *StoreResFlag = new StoreInst(FlagRes, ResFlagAddr);
915 StoreResDiv->insertAfter(ResFlagAddr);
916 StoreResFlag->insertAfter(StoreResDiv);
917 LoadInst *LoadRes = new LoadInst(NewRes, "div.res");
918 LoadRes->insertAfter(StoreResFlag);
919 II->replaceAllUsesWith(LoadRes);
922 II->eraseFromParent();
923 CurInstIterator = StartBB->end();
929 SmallVector<Value*, 2> PtrOps;
931 if (TLI->GetAddrModeArguments(II, PtrOps, AccessTy))
932 while (!PtrOps.empty())
933 if (OptimizeMemoryInst(II, PtrOps.pop_back_val(), AccessTy))
937 // From here on out we're working with named functions.
938 if (!CI->getCalledFunction()) return false;
940 // We'll need DataLayout from here on out.
941 const DataLayout *TD = TLI ? TLI->getDataLayout() : nullptr;
942 if (!TD) return false;
944 // Lower all default uses of _chk calls. This is very similar
945 // to what InstCombineCalls does, but here we are only lowering calls
946 // that have the default "don't know" as the objectsize. Anything else
947 // should be left alone.
948 CodeGenPrepareFortifiedLibCalls Simplifier;
949 return Simplifier.fold(CI, TD, TLInfo);
952 /// DupRetToEnableTailCallOpts - Look for opportunities to duplicate return
953 /// instructions to the predecessor to enable tail call optimizations. The
954 /// case it is currently looking for is:
957 /// %tmp0 = tail call i32 @f0()
960 /// %tmp1 = tail call i32 @f1()
963 /// %tmp2 = tail call i32 @f2()
966 /// %retval = phi i32 [ %tmp0, %bb0 ], [ %tmp1, %bb1 ], [ %tmp2, %bb2 ]
974 /// %tmp0 = tail call i32 @f0()
977 /// %tmp1 = tail call i32 @f1()
980 /// %tmp2 = tail call i32 @f2()
983 bool CodeGenPrepare::DupRetToEnableTailCallOpts(BasicBlock *BB) {
987 ReturnInst *RI = dyn_cast<ReturnInst>(BB->getTerminator());
991 PHINode *PN = nullptr;
992 BitCastInst *BCI = nullptr;
993 Value *V = RI->getReturnValue();
995 BCI = dyn_cast<BitCastInst>(V);
997 V = BCI->getOperand(0);
999 PN = dyn_cast<PHINode>(V);
1004 if (PN && PN->getParent() != BB)
1007 // It's not safe to eliminate the sign / zero extension of the return value.
1008 // See llvm::isInTailCallPosition().
1009 const Function *F = BB->getParent();
1010 AttributeSet CallerAttrs = F->getAttributes();
1011 if (CallerAttrs.hasAttribute(AttributeSet::ReturnIndex, Attribute::ZExt) ||
1012 CallerAttrs.hasAttribute(AttributeSet::ReturnIndex, Attribute::SExt))
1015 // Make sure there are no instructions between the PHI and return, or that the
1016 // return is the first instruction in the block.
1018 BasicBlock::iterator BI = BB->begin();
1019 do { ++BI; } while (isa<DbgInfoIntrinsic>(BI));
1021 // Also skip over the bitcast.
1026 BasicBlock::iterator BI = BB->begin();
1027 while (isa<DbgInfoIntrinsic>(BI)) ++BI;
1032 /// Only dup the ReturnInst if the CallInst is likely to be emitted as a tail
1034 SmallVector<CallInst*, 4> TailCalls;
1036 for (unsigned I = 0, E = PN->getNumIncomingValues(); I != E; ++I) {
1037 CallInst *CI = dyn_cast<CallInst>(PN->getIncomingValue(I));
1038 // Make sure the phi value is indeed produced by the tail call.
1039 if (CI && CI->hasOneUse() && CI->getParent() == PN->getIncomingBlock(I) &&
1040 TLI->mayBeEmittedAsTailCall(CI))
1041 TailCalls.push_back(CI);
1044 SmallPtrSet<BasicBlock*, 4> VisitedBBs;
1045 for (pred_iterator PI = pred_begin(BB), PE = pred_end(BB); PI != PE; ++PI) {
1046 if (!VisitedBBs.insert(*PI))
1049 BasicBlock::InstListType &InstList = (*PI)->getInstList();
1050 BasicBlock::InstListType::reverse_iterator RI = InstList.rbegin();
1051 BasicBlock::InstListType::reverse_iterator RE = InstList.rend();
1052 do { ++RI; } while (RI != RE && isa<DbgInfoIntrinsic>(&*RI));
1056 CallInst *CI = dyn_cast<CallInst>(&*RI);
1057 if (CI && CI->use_empty() && TLI->mayBeEmittedAsTailCall(CI))
1058 TailCalls.push_back(CI);
1062 bool Changed = false;
1063 for (unsigned i = 0, e = TailCalls.size(); i != e; ++i) {
1064 CallInst *CI = TailCalls[i];
1067 // Conservatively require the attributes of the call to match those of the
1068 // return. Ignore noalias because it doesn't affect the call sequence.
1069 AttributeSet CalleeAttrs = CS.getAttributes();
1070 if (AttrBuilder(CalleeAttrs, AttributeSet::ReturnIndex).
1071 removeAttribute(Attribute::NoAlias) !=
1072 AttrBuilder(CalleeAttrs, AttributeSet::ReturnIndex).
1073 removeAttribute(Attribute::NoAlias))
1076 // Make sure the call instruction is followed by an unconditional branch to
1077 // the return block.
1078 BasicBlock *CallBB = CI->getParent();
1079 BranchInst *BI = dyn_cast<BranchInst>(CallBB->getTerminator());
1080 if (!BI || !BI->isUnconditional() || BI->getSuccessor(0) != BB)
1083 // Duplicate the return into CallBB.
1084 (void)FoldReturnIntoUncondBranch(RI, BB, CallBB);
1085 ModifiedDT = Changed = true;
1089 // If we eliminated all predecessors of the block, delete the block now.
1090 if (Changed && !BB->hasAddressTaken() && pred_begin(BB) == pred_end(BB))
1091 BB->eraseFromParent();
1096 //===----------------------------------------------------------------------===//
1097 // Memory Optimization
1098 //===----------------------------------------------------------------------===//
1102 /// ExtAddrMode - This is an extended version of TargetLowering::AddrMode
1103 /// which holds actual Value*'s for register values.
1104 struct ExtAddrMode : public TargetLowering::AddrMode {
1107 ExtAddrMode() : BaseReg(nullptr), ScaledReg(nullptr) {}
1108 void print(raw_ostream &OS) const;
1111 bool operator==(const ExtAddrMode& O) const {
1112 return (BaseReg == O.BaseReg) && (ScaledReg == O.ScaledReg) &&
1113 (BaseGV == O.BaseGV) && (BaseOffs == O.BaseOffs) &&
1114 (HasBaseReg == O.HasBaseReg) && (Scale == O.Scale);
1119 static inline raw_ostream &operator<<(raw_ostream &OS, const ExtAddrMode &AM) {
1125 void ExtAddrMode::print(raw_ostream &OS) const {
1126 bool NeedPlus = false;
1129 OS << (NeedPlus ? " + " : "")
1131 BaseGV->printAsOperand(OS, /*PrintType=*/false);
1136 OS << (NeedPlus ? " + " : "") << BaseOffs, NeedPlus = true;
1139 OS << (NeedPlus ? " + " : "")
1141 BaseReg->printAsOperand(OS, /*PrintType=*/false);
1145 OS << (NeedPlus ? " + " : "")
1147 ScaledReg->printAsOperand(OS, /*PrintType=*/false);
1153 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
1154 void ExtAddrMode::dump() const {
1160 /// \brief This class provides transaction based operation on the IR.
1161 /// Every change made through this class is recorded in the internal state and
1162 /// can be undone (rollback) until commit is called.
1163 class TypePromotionTransaction {
1165 /// \brief This represents the common interface of the individual transaction.
1166 /// Each class implements the logic for doing one specific modification on
1167 /// the IR via the TypePromotionTransaction.
1168 class TypePromotionAction {
1170 /// The Instruction modified.
1174 /// \brief Constructor of the action.
1175 /// The constructor performs the related action on the IR.
1176 TypePromotionAction(Instruction *Inst) : Inst(Inst) {}
1178 virtual ~TypePromotionAction() {}
1180 /// \brief Undo the modification done by this action.
1181 /// When this method is called, the IR must be in the same state as it was
1182 /// before this action was applied.
1183 /// \pre Undoing the action works if and only if the IR is in the exact same
1184 /// state as it was directly after this action was applied.
1185 virtual void undo() = 0;
1187 /// \brief Advocate every change made by this action.
1188 /// When the results on the IR of the action are to be kept, it is important
1189 /// to call this function, otherwise hidden information may be kept forever.
1190 virtual void commit() {
1191 // Nothing to be done, this action is not doing anything.
1195 /// \brief Utility to remember the position of an instruction.
1196 class InsertionHandler {
1197 /// Position of an instruction.
1198 /// Either an instruction:
1199 /// - Is the first in a basic block: BB is used.
1200 /// - Has a previous instructon: PrevInst is used.
1202 Instruction *PrevInst;
1205 /// Remember whether or not the instruction had a previous instruction.
1206 bool HasPrevInstruction;
1209 /// \brief Record the position of \p Inst.
1210 InsertionHandler(Instruction *Inst) {
1211 BasicBlock::iterator It = Inst;
1212 HasPrevInstruction = (It != (Inst->getParent()->begin()));
1213 if (HasPrevInstruction)
1214 Point.PrevInst = --It;
1216 Point.BB = Inst->getParent();
1219 /// \brief Insert \p Inst at the recorded position.
1220 void insert(Instruction *Inst) {
1221 if (HasPrevInstruction) {
1222 if (Inst->getParent())
1223 Inst->removeFromParent();
1224 Inst->insertAfter(Point.PrevInst);
1226 Instruction *Position = Point.BB->getFirstInsertionPt();
1227 if (Inst->getParent())
1228 Inst->moveBefore(Position);
1230 Inst->insertBefore(Position);
1235 /// \brief Move an instruction before another.
1236 class InstructionMoveBefore : public TypePromotionAction {
1237 /// Original position of the instruction.
1238 InsertionHandler Position;
1241 /// \brief Move \p Inst before \p Before.
1242 InstructionMoveBefore(Instruction *Inst, Instruction *Before)
1243 : TypePromotionAction(Inst), Position(Inst) {
1244 DEBUG(dbgs() << "Do: move: " << *Inst << "\nbefore: " << *Before << "\n");
1245 Inst->moveBefore(Before);
1248 /// \brief Move the instruction back to its original position.
1249 void undo() override {
1250 DEBUG(dbgs() << "Undo: moveBefore: " << *Inst << "\n");
1251 Position.insert(Inst);
1255 /// \brief Set the operand of an instruction with a new value.
1256 class OperandSetter : public TypePromotionAction {
1257 /// Original operand of the instruction.
1259 /// Index of the modified instruction.
1263 /// \brief Set \p Idx operand of \p Inst with \p NewVal.
1264 OperandSetter(Instruction *Inst, unsigned Idx, Value *NewVal)
1265 : TypePromotionAction(Inst), Idx(Idx) {
1266 DEBUG(dbgs() << "Do: setOperand: " << Idx << "\n"
1267 << "for:" << *Inst << "\n"
1268 << "with:" << *NewVal << "\n");
1269 Origin = Inst->getOperand(Idx);
1270 Inst->setOperand(Idx, NewVal);
1273 /// \brief Restore the original value of the instruction.
1274 void undo() override {
1275 DEBUG(dbgs() << "Undo: setOperand:" << Idx << "\n"
1276 << "for: " << *Inst << "\n"
1277 << "with: " << *Origin << "\n");
1278 Inst->setOperand(Idx, Origin);
1282 /// \brief Hide the operands of an instruction.
1283 /// Do as if this instruction was not using any of its operands.
1284 class OperandsHider : public TypePromotionAction {
1285 /// The list of original operands.
1286 SmallVector<Value *, 4> OriginalValues;
1289 /// \brief Remove \p Inst from the uses of the operands of \p Inst.
1290 OperandsHider(Instruction *Inst) : TypePromotionAction(Inst) {
1291 DEBUG(dbgs() << "Do: OperandsHider: " << *Inst << "\n");
1292 unsigned NumOpnds = Inst->getNumOperands();
1293 OriginalValues.reserve(NumOpnds);
1294 for (unsigned It = 0; It < NumOpnds; ++It) {
1295 // Save the current operand.
1296 Value *Val = Inst->getOperand(It);
1297 OriginalValues.push_back(Val);
1299 // We could use OperandSetter here, but that would implied an overhead
1300 // that we are not willing to pay.
1301 Inst->setOperand(It, UndefValue::get(Val->getType()));
1305 /// \brief Restore the original list of uses.
1306 void undo() override {
1307 DEBUG(dbgs() << "Undo: OperandsHider: " << *Inst << "\n");
1308 for (unsigned It = 0, EndIt = OriginalValues.size(); It != EndIt; ++It)
1309 Inst->setOperand(It, OriginalValues[It]);
1313 /// \brief Build a truncate instruction.
1314 class TruncBuilder : public TypePromotionAction {
1316 /// \brief Build a truncate instruction of \p Opnd producing a \p Ty
1318 /// trunc Opnd to Ty.
1319 TruncBuilder(Instruction *Opnd, Type *Ty) : TypePromotionAction(Opnd) {
1320 IRBuilder<> Builder(Opnd);
1321 Inst = cast<Instruction>(Builder.CreateTrunc(Opnd, Ty, "promoted"));
1322 DEBUG(dbgs() << "Do: TruncBuilder: " << *Inst << "\n");
1325 /// \brief Get the built instruction.
1326 Instruction *getBuiltInstruction() { return Inst; }
1328 /// \brief Remove the built instruction.
1329 void undo() override {
1330 DEBUG(dbgs() << "Undo: TruncBuilder: " << *Inst << "\n");
1331 Inst->eraseFromParent();
1335 /// \brief Build a sign extension instruction.
1336 class SExtBuilder : public TypePromotionAction {
1338 /// \brief Build a sign extension instruction of \p Opnd producing a \p Ty
1340 /// sext Opnd to Ty.
1341 SExtBuilder(Instruction *InsertPt, Value *Opnd, Type *Ty)
1342 : TypePromotionAction(Inst) {
1343 IRBuilder<> Builder(InsertPt);
1344 Inst = cast<Instruction>(Builder.CreateSExt(Opnd, Ty, "promoted"));
1345 DEBUG(dbgs() << "Do: SExtBuilder: " << *Inst << "\n");
1348 /// \brief Get the built instruction.
1349 Instruction *getBuiltInstruction() { return Inst; }
1351 /// \brief Remove the built instruction.
1352 void undo() override {
1353 DEBUG(dbgs() << "Undo: SExtBuilder: " << *Inst << "\n");
1354 Inst->eraseFromParent();
1358 /// \brief Mutate an instruction to another type.
1359 class TypeMutator : public TypePromotionAction {
1360 /// Record the original type.
1364 /// \brief Mutate the type of \p Inst into \p NewTy.
1365 TypeMutator(Instruction *Inst, Type *NewTy)
1366 : TypePromotionAction(Inst), OrigTy(Inst->getType()) {
1367 DEBUG(dbgs() << "Do: MutateType: " << *Inst << " with " << *NewTy
1369 Inst->mutateType(NewTy);
1372 /// \brief Mutate the instruction back to its original type.
1373 void undo() override {
1374 DEBUG(dbgs() << "Undo: MutateType: " << *Inst << " with " << *OrigTy
1376 Inst->mutateType(OrigTy);
1380 /// \brief Replace the uses of an instruction by another instruction.
1381 class UsesReplacer : public TypePromotionAction {
1382 /// Helper structure to keep track of the replaced uses.
1383 struct InstructionAndIdx {
1384 /// The instruction using the instruction.
1386 /// The index where this instruction is used for Inst.
1388 InstructionAndIdx(Instruction *Inst, unsigned Idx)
1389 : Inst(Inst), Idx(Idx) {}
1392 /// Keep track of the original uses (pair Instruction, Index).
1393 SmallVector<InstructionAndIdx, 4> OriginalUses;
1394 typedef SmallVectorImpl<InstructionAndIdx>::iterator use_iterator;
1397 /// \brief Replace all the use of \p Inst by \p New.
1398 UsesReplacer(Instruction *Inst, Value *New) : TypePromotionAction(Inst) {
1399 DEBUG(dbgs() << "Do: UsersReplacer: " << *Inst << " with " << *New
1401 // Record the original uses.
1402 for (Use &U : Inst->uses()) {
1403 Instruction *UserI = cast<Instruction>(U.getUser());
1404 OriginalUses.push_back(InstructionAndIdx(UserI, U.getOperandNo()));
1406 // Now, we can replace the uses.
1407 Inst->replaceAllUsesWith(New);
1410 /// \brief Reassign the original uses of Inst to Inst.
1411 void undo() override {
1412 DEBUG(dbgs() << "Undo: UsersReplacer: " << *Inst << "\n");
1413 for (use_iterator UseIt = OriginalUses.begin(),
1414 EndIt = OriginalUses.end();
1415 UseIt != EndIt; ++UseIt) {
1416 UseIt->Inst->setOperand(UseIt->Idx, Inst);
1421 /// \brief Remove an instruction from the IR.
1422 class InstructionRemover : public TypePromotionAction {
1423 /// Original position of the instruction.
1424 InsertionHandler Inserter;
1425 /// Helper structure to hide all the link to the instruction. In other
1426 /// words, this helps to do as if the instruction was removed.
1427 OperandsHider Hider;
1428 /// Keep track of the uses replaced, if any.
1429 UsesReplacer *Replacer;
1432 /// \brief Remove all reference of \p Inst and optinally replace all its
1434 /// \pre If !Inst->use_empty(), then New != nullptr
1435 InstructionRemover(Instruction *Inst, Value *New = nullptr)
1436 : TypePromotionAction(Inst), Inserter(Inst), Hider(Inst),
1439 Replacer = new UsesReplacer(Inst, New);
1440 DEBUG(dbgs() << "Do: InstructionRemover: " << *Inst << "\n");
1441 Inst->removeFromParent();
1444 ~InstructionRemover() { delete Replacer; }
1446 /// \brief Really remove the instruction.
1447 void commit() override { delete Inst; }
1449 /// \brief Resurrect the instruction and reassign it to the proper uses if
1450 /// new value was provided when build this action.
1451 void undo() override {
1452 DEBUG(dbgs() << "Undo: InstructionRemover: " << *Inst << "\n");
1453 Inserter.insert(Inst);
1461 /// Restoration point.
1462 /// The restoration point is a pointer to an action instead of an iterator
1463 /// because the iterator may be invalidated but not the pointer.
1464 typedef const TypePromotionAction *ConstRestorationPt;
1465 /// Advocate every changes made in that transaction.
1467 /// Undo all the changes made after the given point.
1468 void rollback(ConstRestorationPt Point);
1469 /// Get the current restoration point.
1470 ConstRestorationPt getRestorationPoint() const;
1472 /// \name API for IR modification with state keeping to support rollback.
1474 /// Same as Instruction::setOperand.
1475 void setOperand(Instruction *Inst, unsigned Idx, Value *NewVal);
1476 /// Same as Instruction::eraseFromParent.
1477 void eraseInstruction(Instruction *Inst, Value *NewVal = nullptr);
1478 /// Same as Value::replaceAllUsesWith.
1479 void replaceAllUsesWith(Instruction *Inst, Value *New);
1480 /// Same as Value::mutateType.
1481 void mutateType(Instruction *Inst, Type *NewTy);
1482 /// Same as IRBuilder::createTrunc.
1483 Instruction *createTrunc(Instruction *Opnd, Type *Ty);
1484 /// Same as IRBuilder::createSExt.
1485 Instruction *createSExt(Instruction *Inst, Value *Opnd, Type *Ty);
1486 /// Same as Instruction::moveBefore.
1487 void moveBefore(Instruction *Inst, Instruction *Before);
1491 /// The ordered list of actions made so far.
1492 SmallVector<std::unique_ptr<TypePromotionAction>, 16> Actions;
1493 typedef SmallVectorImpl<std::unique_ptr<TypePromotionAction>>::iterator CommitPt;
1496 void TypePromotionTransaction::setOperand(Instruction *Inst, unsigned Idx,
1499 make_unique<TypePromotionTransaction::OperandSetter>(Inst, Idx, NewVal));
1502 void TypePromotionTransaction::eraseInstruction(Instruction *Inst,
1505 make_unique<TypePromotionTransaction::InstructionRemover>(Inst, NewVal));
1508 void TypePromotionTransaction::replaceAllUsesWith(Instruction *Inst,
1510 Actions.push_back(make_unique<TypePromotionTransaction::UsesReplacer>(Inst, New));
1513 void TypePromotionTransaction::mutateType(Instruction *Inst, Type *NewTy) {
1514 Actions.push_back(make_unique<TypePromotionTransaction::TypeMutator>(Inst, NewTy));
1517 Instruction *TypePromotionTransaction::createTrunc(Instruction *Opnd,
1519 std::unique_ptr<TruncBuilder> Ptr(new TruncBuilder(Opnd, Ty));
1520 Instruction *I = Ptr->getBuiltInstruction();
1521 Actions.push_back(std::move(Ptr));
1525 Instruction *TypePromotionTransaction::createSExt(Instruction *Inst,
1526 Value *Opnd, Type *Ty) {
1527 std::unique_ptr<SExtBuilder> Ptr(new SExtBuilder(Inst, Opnd, Ty));
1528 Instruction *I = Ptr->getBuiltInstruction();
1529 Actions.push_back(std::move(Ptr));
1533 void TypePromotionTransaction::moveBefore(Instruction *Inst,
1534 Instruction *Before) {
1536 make_unique<TypePromotionTransaction::InstructionMoveBefore>(Inst, Before));
1539 TypePromotionTransaction::ConstRestorationPt
1540 TypePromotionTransaction::getRestorationPoint() const {
1541 return !Actions.empty() ? Actions.back().get() : nullptr;
1544 void TypePromotionTransaction::commit() {
1545 for (CommitPt It = Actions.begin(), EndIt = Actions.end(); It != EndIt;
1551 void TypePromotionTransaction::rollback(
1552 TypePromotionTransaction::ConstRestorationPt Point) {
1553 while (!Actions.empty() && Point != Actions.back().get()) {
1554 std::unique_ptr<TypePromotionAction> Curr = Actions.pop_back_val();
1559 /// \brief A helper class for matching addressing modes.
1561 /// This encapsulates the logic for matching the target-legal addressing modes.
1562 class AddressingModeMatcher {
1563 SmallVectorImpl<Instruction*> &AddrModeInsts;
1564 const TargetLowering &TLI;
1566 /// AccessTy/MemoryInst - This is the type for the access (e.g. double) and
1567 /// the memory instruction that we're computing this address for.
1569 Instruction *MemoryInst;
1571 /// AddrMode - This is the addressing mode that we're building up. This is
1572 /// part of the return value of this addressing mode matching stuff.
1573 ExtAddrMode &AddrMode;
1575 /// The truncate instruction inserted by other CodeGenPrepare optimizations.
1576 const SetOfInstrs &InsertedTruncs;
1577 /// A map from the instructions to their type before promotion.
1578 InstrToOrigTy &PromotedInsts;
1579 /// The ongoing transaction where every action should be registered.
1580 TypePromotionTransaction &TPT;
1582 /// IgnoreProfitability - This is set to true when we should not do
1583 /// profitability checks. When true, IsProfitableToFoldIntoAddressingMode
1584 /// always returns true.
1585 bool IgnoreProfitability;
1587 AddressingModeMatcher(SmallVectorImpl<Instruction*> &AMI,
1588 const TargetLowering &T, Type *AT,
1589 Instruction *MI, ExtAddrMode &AM,
1590 const SetOfInstrs &InsertedTruncs,
1591 InstrToOrigTy &PromotedInsts,
1592 TypePromotionTransaction &TPT)
1593 : AddrModeInsts(AMI), TLI(T), AccessTy(AT), MemoryInst(MI), AddrMode(AM),
1594 InsertedTruncs(InsertedTruncs), PromotedInsts(PromotedInsts), TPT(TPT) {
1595 IgnoreProfitability = false;
1599 /// Match - Find the maximal addressing mode that a load/store of V can fold,
1600 /// give an access type of AccessTy. This returns a list of involved
1601 /// instructions in AddrModeInsts.
1602 /// \p InsertedTruncs The truncate instruction inserted by other
1605 /// \p PromotedInsts maps the instructions to their type before promotion.
1606 /// \p The ongoing transaction where every action should be registered.
1607 static ExtAddrMode Match(Value *V, Type *AccessTy,
1608 Instruction *MemoryInst,
1609 SmallVectorImpl<Instruction*> &AddrModeInsts,
1610 const TargetLowering &TLI,
1611 const SetOfInstrs &InsertedTruncs,
1612 InstrToOrigTy &PromotedInsts,
1613 TypePromotionTransaction &TPT) {
1616 bool Success = AddressingModeMatcher(AddrModeInsts, TLI, AccessTy,
1617 MemoryInst, Result, InsertedTruncs,
1618 PromotedInsts, TPT).MatchAddr(V, 0);
1619 (void)Success; assert(Success && "Couldn't select *anything*?");
1623 bool MatchScaledValue(Value *ScaleReg, int64_t Scale, unsigned Depth);
1624 bool MatchAddr(Value *V, unsigned Depth);
1625 bool MatchOperationAddr(User *Operation, unsigned Opcode, unsigned Depth,
1626 bool *MovedAway = nullptr);
1627 bool IsProfitableToFoldIntoAddressingMode(Instruction *I,
1628 ExtAddrMode &AMBefore,
1629 ExtAddrMode &AMAfter);
1630 bool ValueAlreadyLiveAtInst(Value *Val, Value *KnownLive1, Value *KnownLive2);
1631 bool IsPromotionProfitable(unsigned MatchedSize, unsigned SizeWithPromotion,
1632 Value *PromotedOperand) const;
1635 /// MatchScaledValue - Try adding ScaleReg*Scale to the current addressing mode.
1636 /// Return true and update AddrMode if this addr mode is legal for the target,
1638 bool AddressingModeMatcher::MatchScaledValue(Value *ScaleReg, int64_t Scale,
1640 // If Scale is 1, then this is the same as adding ScaleReg to the addressing
1641 // mode. Just process that directly.
1643 return MatchAddr(ScaleReg, Depth);
1645 // If the scale is 0, it takes nothing to add this.
1649 // If we already have a scale of this value, we can add to it, otherwise, we
1650 // need an available scale field.
1651 if (AddrMode.Scale != 0 && AddrMode.ScaledReg != ScaleReg)
1654 ExtAddrMode TestAddrMode = AddrMode;
1656 // Add scale to turn X*4+X*3 -> X*7. This could also do things like
1657 // [A+B + A*7] -> [B+A*8].
1658 TestAddrMode.Scale += Scale;
1659 TestAddrMode.ScaledReg = ScaleReg;
1661 // If the new address isn't legal, bail out.
1662 if (!TLI.isLegalAddressingMode(TestAddrMode, AccessTy))
1665 // It was legal, so commit it.
1666 AddrMode = TestAddrMode;
1668 // Okay, we decided that we can add ScaleReg+Scale to AddrMode. Check now
1669 // to see if ScaleReg is actually X+C. If so, we can turn this into adding
1670 // X*Scale + C*Scale to addr mode.
1671 ConstantInt *CI = nullptr; Value *AddLHS = nullptr;
1672 if (isa<Instruction>(ScaleReg) && // not a constant expr.
1673 match(ScaleReg, m_Add(m_Value(AddLHS), m_ConstantInt(CI)))) {
1674 TestAddrMode.ScaledReg = AddLHS;
1675 TestAddrMode.BaseOffs += CI->getSExtValue()*TestAddrMode.Scale;
1677 // If this addressing mode is legal, commit it and remember that we folded
1678 // this instruction.
1679 if (TLI.isLegalAddressingMode(TestAddrMode, AccessTy)) {
1680 AddrModeInsts.push_back(cast<Instruction>(ScaleReg));
1681 AddrMode = TestAddrMode;
1686 // Otherwise, not (x+c)*scale, just return what we have.
1690 /// MightBeFoldableInst - This is a little filter, which returns true if an
1691 /// addressing computation involving I might be folded into a load/store
1692 /// accessing it. This doesn't need to be perfect, but needs to accept at least
1693 /// the set of instructions that MatchOperationAddr can.
1694 static bool MightBeFoldableInst(Instruction *I) {
1695 switch (I->getOpcode()) {
1696 case Instruction::BitCast:
1697 // Don't touch identity bitcasts.
1698 if (I->getType() == I->getOperand(0)->getType())
1700 return I->getType()->isPointerTy() || I->getType()->isIntegerTy();
1701 case Instruction::PtrToInt:
1702 // PtrToInt is always a noop, as we know that the int type is pointer sized.
1704 case Instruction::IntToPtr:
1705 // We know the input is intptr_t, so this is foldable.
1707 case Instruction::Add:
1709 case Instruction::Mul:
1710 case Instruction::Shl:
1711 // Can only handle X*C and X << C.
1712 return isa<ConstantInt>(I->getOperand(1));
1713 case Instruction::GetElementPtr:
1720 /// \brief Hepler class to perform type promotion.
1721 class TypePromotionHelper {
1722 /// \brief Utility function to check whether or not a sign extension of
1723 /// \p Inst with \p ConsideredSExtType can be moved through \p Inst by either
1724 /// using the operands of \p Inst or promoting \p Inst.
1725 /// In other words, check if:
1726 /// sext (Ty Inst opnd1 opnd2 ... opndN) to ConsideredSExtType.
1727 /// #1 Promotion applies:
1728 /// ConsideredSExtType Inst (sext opnd1 to ConsideredSExtType, ...).
1729 /// #2 Operand reuses:
1730 /// sext opnd1 to ConsideredSExtType.
1731 /// \p PromotedInsts maps the instructions to their type before promotion.
1732 static bool canGetThrough(const Instruction *Inst, Type *ConsideredSExtType,
1733 const InstrToOrigTy &PromotedInsts);
1735 /// \brief Utility function to determine if \p OpIdx should be promoted when
1736 /// promoting \p Inst.
1737 static bool shouldSExtOperand(const Instruction *Inst, int OpIdx) {
1738 if (isa<SelectInst>(Inst) && OpIdx == 0)
1743 /// \brief Utility function to promote the operand of \p SExt when this
1744 /// operand is a promotable trunc or sext.
1745 /// \p PromotedInsts maps the instructions to their type before promotion.
1746 /// \p CreatedInsts[out] contains how many non-free instructions have been
1747 /// created to promote the operand of SExt.
1748 /// Should never be called directly.
1749 /// \return The promoted value which is used instead of SExt.
1750 static Value *promoteOperandForTruncAndSExt(Instruction *SExt,
1751 TypePromotionTransaction &TPT,
1752 InstrToOrigTy &PromotedInsts,
1753 unsigned &CreatedInsts);
1755 /// \brief Utility function to promote the operand of \p SExt when this
1756 /// operand is promotable and is not a supported trunc or sext.
1757 /// \p PromotedInsts maps the instructions to their type before promotion.
1758 /// \p CreatedInsts[out] contains how many non-free instructions have been
1759 /// created to promote the operand of SExt.
1760 /// Should never be called directly.
1761 /// \return The promoted value which is used instead of SExt.
1762 static Value *promoteOperandForOther(Instruction *SExt,
1763 TypePromotionTransaction &TPT,
1764 InstrToOrigTy &PromotedInsts,
1765 unsigned &CreatedInsts);
1768 /// Type for the utility function that promotes the operand of SExt.
1769 typedef Value *(*Action)(Instruction *SExt, TypePromotionTransaction &TPT,
1770 InstrToOrigTy &PromotedInsts,
1771 unsigned &CreatedInsts);
1772 /// \brief Given a sign extend instruction \p SExt, return the approriate
1773 /// action to promote the operand of \p SExt instead of using SExt.
1774 /// \return NULL if no promotable action is possible with the current
1776 /// \p InsertedTruncs keeps track of all the truncate instructions inserted by
1777 /// the others CodeGenPrepare optimizations. This information is important
1778 /// because we do not want to promote these instructions as CodeGenPrepare
1779 /// will reinsert them later. Thus creating an infinite loop: create/remove.
1780 /// \p PromotedInsts maps the instructions to their type before promotion.
1781 static Action getAction(Instruction *SExt, const SetOfInstrs &InsertedTruncs,
1782 const TargetLowering &TLI,
1783 const InstrToOrigTy &PromotedInsts);
1786 bool TypePromotionHelper::canGetThrough(const Instruction *Inst,
1787 Type *ConsideredSExtType,
1788 const InstrToOrigTy &PromotedInsts) {
1789 // We can always get through sext.
1790 if (isa<SExtInst>(Inst))
1793 // We can get through binary operator, if it is legal. In other words, the
1794 // binary operator must have a nuw or nsw flag.
1795 const BinaryOperator *BinOp = dyn_cast<BinaryOperator>(Inst);
1796 if (BinOp && isa<OverflowingBinaryOperator>(BinOp) &&
1797 (BinOp->hasNoUnsignedWrap() || BinOp->hasNoSignedWrap()))
1800 // Check if we can do the following simplification.
1801 // sext(trunc(sext)) --> sext
1802 if (!isa<TruncInst>(Inst))
1805 Value *OpndVal = Inst->getOperand(0);
1806 // Check if we can use this operand in the sext.
1807 // If the type is larger than the result type of the sign extension,
1809 if (OpndVal->getType()->getIntegerBitWidth() >
1810 ConsideredSExtType->getIntegerBitWidth())
1813 // If the operand of the truncate is not an instruction, we will not have
1814 // any information on the dropped bits.
1815 // (Actually we could for constant but it is not worth the extra logic).
1816 Instruction *Opnd = dyn_cast<Instruction>(OpndVal);
1820 // Check if the source of the type is narrow enough.
1821 // I.e., check that trunc just drops sign extended bits.
1822 // #1 get the type of the operand.
1823 const Type *OpndType;
1824 InstrToOrigTy::const_iterator It = PromotedInsts.find(Opnd);
1825 if (It != PromotedInsts.end())
1826 OpndType = It->second;
1827 else if (isa<SExtInst>(Opnd))
1828 OpndType = cast<Instruction>(Opnd)->getOperand(0)->getType();
1832 // #2 check that the truncate just drop sign extended bits.
1833 if (Inst->getType()->getIntegerBitWidth() >= OpndType->getIntegerBitWidth())
1839 TypePromotionHelper::Action TypePromotionHelper::getAction(
1840 Instruction *SExt, const SetOfInstrs &InsertedTruncs,
1841 const TargetLowering &TLI, const InstrToOrigTy &PromotedInsts) {
1842 Instruction *SExtOpnd = dyn_cast<Instruction>(SExt->getOperand(0));
1843 Type *SExtTy = SExt->getType();
1844 // If the operand of the sign extension is not an instruction, we cannot
1846 // If it, check we can get through.
1847 if (!SExtOpnd || !canGetThrough(SExtOpnd, SExtTy, PromotedInsts))
1850 // Do not promote if the operand has been added by codegenprepare.
1851 // Otherwise, it means we are undoing an optimization that is likely to be
1852 // redone, thus causing potential infinite loop.
1853 if (isa<TruncInst>(SExtOpnd) && InsertedTruncs.count(SExtOpnd))
1856 // SExt or Trunc instructions.
1857 // Return the related handler.
1858 if (isa<SExtInst>(SExtOpnd) || isa<TruncInst>(SExtOpnd))
1859 return promoteOperandForTruncAndSExt;
1861 // Regular instruction.
1862 // Abort early if we will have to insert non-free instructions.
1863 if (!SExtOpnd->hasOneUse() &&
1864 !TLI.isTruncateFree(SExtTy, SExtOpnd->getType()))
1866 return promoteOperandForOther;
1869 Value *TypePromotionHelper::promoteOperandForTruncAndSExt(
1870 llvm::Instruction *SExt, TypePromotionTransaction &TPT,
1871 InstrToOrigTy &PromotedInsts, unsigned &CreatedInsts) {
1872 // By construction, the operand of SExt is an instruction. Otherwise we cannot
1873 // get through it and this method should not be called.
1874 Instruction *SExtOpnd = cast<Instruction>(SExt->getOperand(0));
1875 // Replace sext(trunc(opnd)) or sext(sext(opnd))
1877 TPT.setOperand(SExt, 0, SExtOpnd->getOperand(0));
1880 // Remove dead code.
1881 if (SExtOpnd->use_empty())
1882 TPT.eraseInstruction(SExtOpnd);
1884 // Check if the sext is still needed.
1885 if (SExt->getType() != SExt->getOperand(0)->getType())
1888 // At this point we have: sext ty opnd to ty.
1889 // Reassign the uses of SExt to the opnd and remove SExt.
1890 Value *NextVal = SExt->getOperand(0);
1891 TPT.eraseInstruction(SExt, NextVal);
1896 TypePromotionHelper::promoteOperandForOther(Instruction *SExt,
1897 TypePromotionTransaction &TPT,
1898 InstrToOrigTy &PromotedInsts,
1899 unsigned &CreatedInsts) {
1900 // By construction, the operand of SExt is an instruction. Otherwise we cannot
1901 // get through it and this method should not be called.
1902 Instruction *SExtOpnd = cast<Instruction>(SExt->getOperand(0));
1904 if (!SExtOpnd->hasOneUse()) {
1905 // SExtOpnd will be promoted.
1906 // All its uses, but SExt, will need to use a truncated value of the
1907 // promoted version.
1908 // Create the truncate now.
1909 Instruction *Trunc = TPT.createTrunc(SExt, SExtOpnd->getType());
1910 Trunc->removeFromParent();
1911 // Insert it just after the definition.
1912 Trunc->insertAfter(SExtOpnd);
1914 TPT.replaceAllUsesWith(SExtOpnd, Trunc);
1915 // Restore the operand of SExt (which has been replace by the previous call
1916 // to replaceAllUsesWith) to avoid creating a cycle trunc <-> sext.
1917 TPT.setOperand(SExt, 0, SExtOpnd);
1920 // Get through the Instruction:
1921 // 1. Update its type.
1922 // 2. Replace the uses of SExt by Inst.
1923 // 3. Sign extend each operand that needs to be sign extended.
1925 // Remember the original type of the instruction before promotion.
1926 // This is useful to know that the high bits are sign extended bits.
1927 PromotedInsts.insert(
1928 std::pair<Instruction *, Type *>(SExtOpnd, SExtOpnd->getType()));
1930 TPT.mutateType(SExtOpnd, SExt->getType());
1932 TPT.replaceAllUsesWith(SExt, SExtOpnd);
1934 Instruction *SExtForOpnd = SExt;
1936 DEBUG(dbgs() << "Propagate SExt to operands\n");
1937 for (int OpIdx = 0, EndOpIdx = SExtOpnd->getNumOperands(); OpIdx != EndOpIdx;
1939 DEBUG(dbgs() << "Operand:\n" << *(SExtOpnd->getOperand(OpIdx)) << '\n');
1940 if (SExtOpnd->getOperand(OpIdx)->getType() == SExt->getType() ||
1941 !shouldSExtOperand(SExtOpnd, OpIdx)) {
1942 DEBUG(dbgs() << "No need to propagate\n");
1945 // Check if we can statically sign extend the operand.
1946 Value *Opnd = SExtOpnd->getOperand(OpIdx);
1947 if (const ConstantInt *Cst = dyn_cast<ConstantInt>(Opnd)) {
1948 DEBUG(dbgs() << "Statically sign extend\n");
1951 ConstantInt::getSigned(SExt->getType(), Cst->getSExtValue()));
1954 // UndefValue are typed, so we have to statically sign extend them.
1955 if (isa<UndefValue>(Opnd)) {
1956 DEBUG(dbgs() << "Statically sign extend\n");
1957 TPT.setOperand(SExtOpnd, OpIdx, UndefValue::get(SExt->getType()));
1961 // Otherwise we have to explicity sign extend the operand.
1962 // Check if SExt was reused to sign extend an operand.
1964 // If yes, create a new one.
1965 DEBUG(dbgs() << "More operands to sext\n");
1966 SExtForOpnd = TPT.createSExt(SExt, Opnd, SExt->getType());
1970 TPT.setOperand(SExtForOpnd, 0, Opnd);
1972 // Move the sign extension before the insertion point.
1973 TPT.moveBefore(SExtForOpnd, SExtOpnd);
1974 TPT.setOperand(SExtOpnd, OpIdx, SExtForOpnd);
1975 // If more sext are required, new instructions will have to be created.
1976 SExtForOpnd = nullptr;
1978 if (SExtForOpnd == SExt) {
1979 DEBUG(dbgs() << "Sign extension is useless now\n");
1980 TPT.eraseInstruction(SExt);
1985 /// IsPromotionProfitable - Check whether or not promoting an instruction
1986 /// to a wider type was profitable.
1987 /// \p MatchedSize gives the number of instructions that have been matched
1988 /// in the addressing mode after the promotion was applied.
1989 /// \p SizeWithPromotion gives the number of created instructions for
1990 /// the promotion plus the number of instructions that have been
1991 /// matched in the addressing mode before the promotion.
1992 /// \p PromotedOperand is the value that has been promoted.
1993 /// \return True if the promotion is profitable, false otherwise.
1995 AddressingModeMatcher::IsPromotionProfitable(unsigned MatchedSize,
1996 unsigned SizeWithPromotion,
1997 Value *PromotedOperand) const {
1998 // We folded less instructions than what we created to promote the operand.
1999 // This is not profitable.
2000 if (MatchedSize < SizeWithPromotion)
2002 if (MatchedSize > SizeWithPromotion)
2004 // The promotion is neutral but it may help folding the sign extension in
2005 // loads for instance.
2006 // Check that we did not create an illegal instruction.
2007 Instruction *PromotedInst = dyn_cast<Instruction>(PromotedOperand);
2010 int ISDOpcode = TLI.InstructionOpcodeToISD(PromotedInst->getOpcode());
2011 // If the ISDOpcode is undefined, it was undefined before the promotion.
2014 // Otherwise, check if the promoted instruction is legal or not.
2015 return TLI.isOperationLegalOrCustom(ISDOpcode,
2016 EVT::getEVT(PromotedInst->getType()));
2019 /// MatchOperationAddr - Given an instruction or constant expr, see if we can
2020 /// fold the operation into the addressing mode. If so, update the addressing
2021 /// mode and return true, otherwise return false without modifying AddrMode.
2022 /// If \p MovedAway is not NULL, it contains the information of whether or
2023 /// not AddrInst has to be folded into the addressing mode on success.
2024 /// If \p MovedAway == true, \p AddrInst will not be part of the addressing
2025 /// because it has been moved away.
2026 /// Thus AddrInst must not be added in the matched instructions.
2027 /// This state can happen when AddrInst is a sext, since it may be moved away.
2028 /// Therefore, AddrInst may not be valid when MovedAway is true and it must
2029 /// not be referenced anymore.
2030 bool AddressingModeMatcher::MatchOperationAddr(User *AddrInst, unsigned Opcode,
2033 // Avoid exponential behavior on extremely deep expression trees.
2034 if (Depth >= 5) return false;
2036 // By default, all matched instructions stay in place.
2041 case Instruction::PtrToInt:
2042 // PtrToInt is always a noop, as we know that the int type is pointer sized.
2043 return MatchAddr(AddrInst->getOperand(0), Depth);
2044 case Instruction::IntToPtr:
2045 // This inttoptr is a no-op if the integer type is pointer sized.
2046 if (TLI.getValueType(AddrInst->getOperand(0)->getType()) ==
2047 TLI.getPointerTy(AddrInst->getType()->getPointerAddressSpace()))
2048 return MatchAddr(AddrInst->getOperand(0), Depth);
2050 case Instruction::BitCast:
2051 // BitCast is always a noop, and we can handle it as long as it is
2052 // int->int or pointer->pointer (we don't want int<->fp or something).
2053 if ((AddrInst->getOperand(0)->getType()->isPointerTy() ||
2054 AddrInst->getOperand(0)->getType()->isIntegerTy()) &&
2055 // Don't touch identity bitcasts. These were probably put here by LSR,
2056 // and we don't want to mess around with them. Assume it knows what it
2058 AddrInst->getOperand(0)->getType() != AddrInst->getType())
2059 return MatchAddr(AddrInst->getOperand(0), Depth);
2061 case Instruction::Add: {
2062 // Check to see if we can merge in the RHS then the LHS. If so, we win.
2063 ExtAddrMode BackupAddrMode = AddrMode;
2064 unsigned OldSize = AddrModeInsts.size();
2065 // Start a transaction at this point.
2066 // The LHS may match but not the RHS.
2067 // Therefore, we need a higher level restoration point to undo partially
2068 // matched operation.
2069 TypePromotionTransaction::ConstRestorationPt LastKnownGood =
2070 TPT.getRestorationPoint();
2072 if (MatchAddr(AddrInst->getOperand(1), Depth+1) &&
2073 MatchAddr(AddrInst->getOperand(0), Depth+1))
2076 // Restore the old addr mode info.
2077 AddrMode = BackupAddrMode;
2078 AddrModeInsts.resize(OldSize);
2079 TPT.rollback(LastKnownGood);
2081 // Otherwise this was over-aggressive. Try merging in the LHS then the RHS.
2082 if (MatchAddr(AddrInst->getOperand(0), Depth+1) &&
2083 MatchAddr(AddrInst->getOperand(1), Depth+1))
2086 // Otherwise we definitely can't merge the ADD in.
2087 AddrMode = BackupAddrMode;
2088 AddrModeInsts.resize(OldSize);
2089 TPT.rollback(LastKnownGood);
2092 //case Instruction::Or:
2093 // TODO: We can handle "Or Val, Imm" iff this OR is equivalent to an ADD.
2095 case Instruction::Mul:
2096 case Instruction::Shl: {
2097 // Can only handle X*C and X << C.
2098 ConstantInt *RHS = dyn_cast<ConstantInt>(AddrInst->getOperand(1));
2099 if (!RHS) return false;
2100 int64_t Scale = RHS->getSExtValue();
2101 if (Opcode == Instruction::Shl)
2102 Scale = 1LL << Scale;
2104 return MatchScaledValue(AddrInst->getOperand(0), Scale, Depth);
2106 case Instruction::GetElementPtr: {
2107 // Scan the GEP. We check it if it contains constant offsets and at most
2108 // one variable offset.
2109 int VariableOperand = -1;
2110 unsigned VariableScale = 0;
2112 int64_t ConstantOffset = 0;
2113 const DataLayout *TD = TLI.getDataLayout();
2114 gep_type_iterator GTI = gep_type_begin(AddrInst);
2115 for (unsigned i = 1, e = AddrInst->getNumOperands(); i != e; ++i, ++GTI) {
2116 if (StructType *STy = dyn_cast<StructType>(*GTI)) {
2117 const StructLayout *SL = TD->getStructLayout(STy);
2119 cast<ConstantInt>(AddrInst->getOperand(i))->getZExtValue();
2120 ConstantOffset += SL->getElementOffset(Idx);
2122 uint64_t TypeSize = TD->getTypeAllocSize(GTI.getIndexedType());
2123 if (ConstantInt *CI = dyn_cast<ConstantInt>(AddrInst->getOperand(i))) {
2124 ConstantOffset += CI->getSExtValue()*TypeSize;
2125 } else if (TypeSize) { // Scales of zero don't do anything.
2126 // We only allow one variable index at the moment.
2127 if (VariableOperand != -1)
2130 // Remember the variable index.
2131 VariableOperand = i;
2132 VariableScale = TypeSize;
2137 // A common case is for the GEP to only do a constant offset. In this case,
2138 // just add it to the disp field and check validity.
2139 if (VariableOperand == -1) {
2140 AddrMode.BaseOffs += ConstantOffset;
2141 if (ConstantOffset == 0 || TLI.isLegalAddressingMode(AddrMode, AccessTy)){
2142 // Check to see if we can fold the base pointer in too.
2143 if (MatchAddr(AddrInst->getOperand(0), Depth+1))
2146 AddrMode.BaseOffs -= ConstantOffset;
2150 // Save the valid addressing mode in case we can't match.
2151 ExtAddrMode BackupAddrMode = AddrMode;
2152 unsigned OldSize = AddrModeInsts.size();
2154 // See if the scale and offset amount is valid for this target.
2155 AddrMode.BaseOffs += ConstantOffset;
2157 // Match the base operand of the GEP.
2158 if (!MatchAddr(AddrInst->getOperand(0), Depth+1)) {
2159 // If it couldn't be matched, just stuff the value in a register.
2160 if (AddrMode.HasBaseReg) {
2161 AddrMode = BackupAddrMode;
2162 AddrModeInsts.resize(OldSize);
2165 AddrMode.HasBaseReg = true;
2166 AddrMode.BaseReg = AddrInst->getOperand(0);
2169 // Match the remaining variable portion of the GEP.
2170 if (!MatchScaledValue(AddrInst->getOperand(VariableOperand), VariableScale,
2172 // If it couldn't be matched, try stuffing the base into a register
2173 // instead of matching it, and retrying the match of the scale.
2174 AddrMode = BackupAddrMode;
2175 AddrModeInsts.resize(OldSize);
2176 if (AddrMode.HasBaseReg)
2178 AddrMode.HasBaseReg = true;
2179 AddrMode.BaseReg = AddrInst->getOperand(0);
2180 AddrMode.BaseOffs += ConstantOffset;
2181 if (!MatchScaledValue(AddrInst->getOperand(VariableOperand),
2182 VariableScale, Depth)) {
2183 // If even that didn't work, bail.
2184 AddrMode = BackupAddrMode;
2185 AddrModeInsts.resize(OldSize);
2192 case Instruction::SExt: {
2193 // Try to move this sext out of the way of the addressing mode.
2194 Instruction *SExt = cast<Instruction>(AddrInst);
2195 // Ask for a method for doing so.
2196 TypePromotionHelper::Action TPH = TypePromotionHelper::getAction(
2197 SExt, InsertedTruncs, TLI, PromotedInsts);
2201 TypePromotionTransaction::ConstRestorationPt LastKnownGood =
2202 TPT.getRestorationPoint();
2203 unsigned CreatedInsts = 0;
2204 Value *PromotedOperand = TPH(SExt, TPT, PromotedInsts, CreatedInsts);
2205 // SExt has been moved away.
2206 // Thus either it will be rematched later in the recursive calls or it is
2207 // gone. Anyway, we must not fold it into the addressing mode at this point.
2211 // addr = gep base, idx
2213 // promotedOpnd = sext opnd <- no match here
2214 // op = promoted_add promotedOpnd, 1 <- match (later in recursive calls)
2215 // addr = gep base, op <- match
2219 assert(PromotedOperand &&
2220 "TypePromotionHelper should have filtered out those cases");
2222 ExtAddrMode BackupAddrMode = AddrMode;
2223 unsigned OldSize = AddrModeInsts.size();
2225 if (!MatchAddr(PromotedOperand, Depth) ||
2226 !IsPromotionProfitable(AddrModeInsts.size(), OldSize + CreatedInsts,
2228 AddrMode = BackupAddrMode;
2229 AddrModeInsts.resize(OldSize);
2230 DEBUG(dbgs() << "Sign extension does not pay off: rollback\n");
2231 TPT.rollback(LastKnownGood);
2240 /// MatchAddr - If we can, try to add the value of 'Addr' into the current
2241 /// addressing mode. If Addr can't be added to AddrMode this returns false and
2242 /// leaves AddrMode unmodified. This assumes that Addr is either a pointer type
2243 /// or intptr_t for the target.
2245 bool AddressingModeMatcher::MatchAddr(Value *Addr, unsigned Depth) {
2246 // Start a transaction at this point that we will rollback if the matching
2248 TypePromotionTransaction::ConstRestorationPt LastKnownGood =
2249 TPT.getRestorationPoint();
2250 if (ConstantInt *CI = dyn_cast<ConstantInt>(Addr)) {
2251 // Fold in immediates if legal for the target.
2252 AddrMode.BaseOffs += CI->getSExtValue();
2253 if (TLI.isLegalAddressingMode(AddrMode, AccessTy))
2255 AddrMode.BaseOffs -= CI->getSExtValue();
2256 } else if (GlobalValue *GV = dyn_cast<GlobalValue>(Addr)) {
2257 // If this is a global variable, try to fold it into the addressing mode.
2258 if (!AddrMode.BaseGV) {
2259 AddrMode.BaseGV = GV;
2260 if (TLI.isLegalAddressingMode(AddrMode, AccessTy))
2262 AddrMode.BaseGV = nullptr;
2264 } else if (Instruction *I = dyn_cast<Instruction>(Addr)) {
2265 ExtAddrMode BackupAddrMode = AddrMode;
2266 unsigned OldSize = AddrModeInsts.size();
2268 // Check to see if it is possible to fold this operation.
2269 bool MovedAway = false;
2270 if (MatchOperationAddr(I, I->getOpcode(), Depth, &MovedAway)) {
2271 // This instruction may have been move away. If so, there is nothing
2275 // Okay, it's possible to fold this. Check to see if it is actually
2276 // *profitable* to do so. We use a simple cost model to avoid increasing
2277 // register pressure too much.
2278 if (I->hasOneUse() ||
2279 IsProfitableToFoldIntoAddressingMode(I, BackupAddrMode, AddrMode)) {
2280 AddrModeInsts.push_back(I);
2284 // It isn't profitable to do this, roll back.
2285 //cerr << "NOT FOLDING: " << *I;
2286 AddrMode = BackupAddrMode;
2287 AddrModeInsts.resize(OldSize);
2288 TPT.rollback(LastKnownGood);
2290 } else if (ConstantExpr *CE = dyn_cast<ConstantExpr>(Addr)) {
2291 if (MatchOperationAddr(CE, CE->getOpcode(), Depth))
2293 TPT.rollback(LastKnownGood);
2294 } else if (isa<ConstantPointerNull>(Addr)) {
2295 // Null pointer gets folded without affecting the addressing mode.
2299 // Worse case, the target should support [reg] addressing modes. :)
2300 if (!AddrMode.HasBaseReg) {
2301 AddrMode.HasBaseReg = true;
2302 AddrMode.BaseReg = Addr;
2303 // Still check for legality in case the target supports [imm] but not [i+r].
2304 if (TLI.isLegalAddressingMode(AddrMode, AccessTy))
2306 AddrMode.HasBaseReg = false;
2307 AddrMode.BaseReg = nullptr;
2310 // If the base register is already taken, see if we can do [r+r].
2311 if (AddrMode.Scale == 0) {
2313 AddrMode.ScaledReg = Addr;
2314 if (TLI.isLegalAddressingMode(AddrMode, AccessTy))
2317 AddrMode.ScaledReg = nullptr;
2320 TPT.rollback(LastKnownGood);
2324 /// IsOperandAMemoryOperand - Check to see if all uses of OpVal by the specified
2325 /// inline asm call are due to memory operands. If so, return true, otherwise
2327 static bool IsOperandAMemoryOperand(CallInst *CI, InlineAsm *IA, Value *OpVal,
2328 const TargetLowering &TLI) {
2329 TargetLowering::AsmOperandInfoVector TargetConstraints = TLI.ParseConstraints(ImmutableCallSite(CI));
2330 for (unsigned i = 0, e = TargetConstraints.size(); i != e; ++i) {
2331 TargetLowering::AsmOperandInfo &OpInfo = TargetConstraints[i];
2333 // Compute the constraint code and ConstraintType to use.
2334 TLI.ComputeConstraintToUse(OpInfo, SDValue());
2336 // If this asm operand is our Value*, and if it isn't an indirect memory
2337 // operand, we can't fold it!
2338 if (OpInfo.CallOperandVal == OpVal &&
2339 (OpInfo.ConstraintType != TargetLowering::C_Memory ||
2340 !OpInfo.isIndirect))
2347 /// FindAllMemoryUses - Recursively walk all the uses of I until we find a
2348 /// memory use. If we find an obviously non-foldable instruction, return true.
2349 /// Add the ultimately found memory instructions to MemoryUses.
2350 static bool FindAllMemoryUses(Instruction *I,
2351 SmallVectorImpl<std::pair<Instruction*,unsigned> > &MemoryUses,
2352 SmallPtrSet<Instruction*, 16> &ConsideredInsts,
2353 const TargetLowering &TLI) {
2354 // If we already considered this instruction, we're done.
2355 if (!ConsideredInsts.insert(I))
2358 // If this is an obviously unfoldable instruction, bail out.
2359 if (!MightBeFoldableInst(I))
2362 // Loop over all the uses, recursively processing them.
2363 for (Use &U : I->uses()) {
2364 Instruction *UserI = cast<Instruction>(U.getUser());
2366 if (LoadInst *LI = dyn_cast<LoadInst>(UserI)) {
2367 MemoryUses.push_back(std::make_pair(LI, U.getOperandNo()));
2371 if (StoreInst *SI = dyn_cast<StoreInst>(UserI)) {
2372 unsigned opNo = U.getOperandNo();
2373 if (opNo == 0) return true; // Storing addr, not into addr.
2374 MemoryUses.push_back(std::make_pair(SI, opNo));
2378 if (CallInst *CI = dyn_cast<CallInst>(UserI)) {
2379 InlineAsm *IA = dyn_cast<InlineAsm>(CI->getCalledValue());
2380 if (!IA) return true;
2382 // If this is a memory operand, we're cool, otherwise bail out.
2383 if (!IsOperandAMemoryOperand(CI, IA, I, TLI))
2388 if (FindAllMemoryUses(UserI, MemoryUses, ConsideredInsts, TLI))
2395 /// ValueAlreadyLiveAtInst - Retrn true if Val is already known to be live at
2396 /// the use site that we're folding it into. If so, there is no cost to
2397 /// include it in the addressing mode. KnownLive1 and KnownLive2 are two values
2398 /// that we know are live at the instruction already.
2399 bool AddressingModeMatcher::ValueAlreadyLiveAtInst(Value *Val,Value *KnownLive1,
2400 Value *KnownLive2) {
2401 // If Val is either of the known-live values, we know it is live!
2402 if (Val == nullptr || Val == KnownLive1 || Val == KnownLive2)
2405 // All values other than instructions and arguments (e.g. constants) are live.
2406 if (!isa<Instruction>(Val) && !isa<Argument>(Val)) return true;
2408 // If Val is a constant sized alloca in the entry block, it is live, this is
2409 // true because it is just a reference to the stack/frame pointer, which is
2410 // live for the whole function.
2411 if (AllocaInst *AI = dyn_cast<AllocaInst>(Val))
2412 if (AI->isStaticAlloca())
2415 // Check to see if this value is already used in the memory instruction's
2416 // block. If so, it's already live into the block at the very least, so we
2417 // can reasonably fold it.
2418 return Val->isUsedInBasicBlock(MemoryInst->getParent());
2421 /// IsProfitableToFoldIntoAddressingMode - It is possible for the addressing
2422 /// mode of the machine to fold the specified instruction into a load or store
2423 /// that ultimately uses it. However, the specified instruction has multiple
2424 /// uses. Given this, it may actually increase register pressure to fold it
2425 /// into the load. For example, consider this code:
2429 /// use(Y) -> nonload/store
2433 /// In this case, Y has multiple uses, and can be folded into the load of Z
2434 /// (yielding load [X+2]). However, doing this will cause both "X" and "X+1" to
2435 /// be live at the use(Y) line. If we don't fold Y into load Z, we use one
2436 /// fewer register. Since Y can't be folded into "use(Y)" we don't increase the
2437 /// number of computations either.
2439 /// Note that this (like most of CodeGenPrepare) is just a rough heuristic. If
2440 /// X was live across 'load Z' for other reasons, we actually *would* want to
2441 /// fold the addressing mode in the Z case. This would make Y die earlier.
2442 bool AddressingModeMatcher::
2443 IsProfitableToFoldIntoAddressingMode(Instruction *I, ExtAddrMode &AMBefore,
2444 ExtAddrMode &AMAfter) {
2445 if (IgnoreProfitability) return true;
2447 // AMBefore is the addressing mode before this instruction was folded into it,
2448 // and AMAfter is the addressing mode after the instruction was folded. Get
2449 // the set of registers referenced by AMAfter and subtract out those
2450 // referenced by AMBefore: this is the set of values which folding in this
2451 // address extends the lifetime of.
2453 // Note that there are only two potential values being referenced here,
2454 // BaseReg and ScaleReg (global addresses are always available, as are any
2455 // folded immediates).
2456 Value *BaseReg = AMAfter.BaseReg, *ScaledReg = AMAfter.ScaledReg;
2458 // If the BaseReg or ScaledReg was referenced by the previous addrmode, their
2459 // lifetime wasn't extended by adding this instruction.
2460 if (ValueAlreadyLiveAtInst(BaseReg, AMBefore.BaseReg, AMBefore.ScaledReg))
2462 if (ValueAlreadyLiveAtInst(ScaledReg, AMBefore.BaseReg, AMBefore.ScaledReg))
2463 ScaledReg = nullptr;
2465 // If folding this instruction (and it's subexprs) didn't extend any live
2466 // ranges, we're ok with it.
2467 if (!BaseReg && !ScaledReg)
2470 // If all uses of this instruction are ultimately load/store/inlineasm's,
2471 // check to see if their addressing modes will include this instruction. If
2472 // so, we can fold it into all uses, so it doesn't matter if it has multiple
2474 SmallVector<std::pair<Instruction*,unsigned>, 16> MemoryUses;
2475 SmallPtrSet<Instruction*, 16> ConsideredInsts;
2476 if (FindAllMemoryUses(I, MemoryUses, ConsideredInsts, TLI))
2477 return false; // Has a non-memory, non-foldable use!
2479 // Now that we know that all uses of this instruction are part of a chain of
2480 // computation involving only operations that could theoretically be folded
2481 // into a memory use, loop over each of these uses and see if they could
2482 // *actually* fold the instruction.
2483 SmallVector<Instruction*, 32> MatchedAddrModeInsts;
2484 for (unsigned i = 0, e = MemoryUses.size(); i != e; ++i) {
2485 Instruction *User = MemoryUses[i].first;
2486 unsigned OpNo = MemoryUses[i].second;
2488 // Get the access type of this use. If the use isn't a pointer, we don't
2489 // know what it accesses.
2490 Value *Address = User->getOperand(OpNo);
2491 if (!Address->getType()->isPointerTy())
2493 Type *AddressAccessTy = Address->getType()->getPointerElementType();
2495 // Do a match against the root of this address, ignoring profitability. This
2496 // will tell us if the addressing mode for the memory operation will
2497 // *actually* cover the shared instruction.
2499 TypePromotionTransaction::ConstRestorationPt LastKnownGood =
2500 TPT.getRestorationPoint();
2501 AddressingModeMatcher Matcher(MatchedAddrModeInsts, TLI, AddressAccessTy,
2502 MemoryInst, Result, InsertedTruncs,
2503 PromotedInsts, TPT);
2504 Matcher.IgnoreProfitability = true;
2505 bool Success = Matcher.MatchAddr(Address, 0);
2506 (void)Success; assert(Success && "Couldn't select *anything*?");
2508 // The match was to check the profitability, the changes made are not
2509 // part of the original matcher. Therefore, they should be dropped
2510 // otherwise the original matcher will not present the right state.
2511 TPT.rollback(LastKnownGood);
2513 // If the match didn't cover I, then it won't be shared by it.
2514 if (std::find(MatchedAddrModeInsts.begin(), MatchedAddrModeInsts.end(),
2515 I) == MatchedAddrModeInsts.end())
2518 MatchedAddrModeInsts.clear();
2524 } // end anonymous namespace
2526 /// IsNonLocalValue - Return true if the specified values are defined in a
2527 /// different basic block than BB.
2528 static bool IsNonLocalValue(Value *V, BasicBlock *BB) {
2529 if (Instruction *I = dyn_cast<Instruction>(V))
2530 return I->getParent() != BB;
2534 /// OptimizeMemoryInst - Load and Store Instructions often have
2535 /// addressing modes that can do significant amounts of computation. As such,
2536 /// instruction selection will try to get the load or store to do as much
2537 /// computation as possible for the program. The problem is that isel can only
2538 /// see within a single block. As such, we sink as much legal addressing mode
2539 /// stuff into the block as possible.
2541 /// This method is used to optimize both load/store and inline asms with memory
2543 bool CodeGenPrepare::OptimizeMemoryInst(Instruction *MemoryInst, Value *Addr,
2547 // Try to collapse single-value PHI nodes. This is necessary to undo
2548 // unprofitable PRE transformations.
2549 SmallVector<Value*, 8> worklist;
2550 SmallPtrSet<Value*, 16> Visited;
2551 worklist.push_back(Addr);
2553 // Use a worklist to iteratively look through PHI nodes, and ensure that
2554 // the addressing mode obtained from the non-PHI roots of the graph
2556 Value *Consensus = nullptr;
2557 unsigned NumUsesConsensus = 0;
2558 bool IsNumUsesConsensusValid = false;
2559 SmallVector<Instruction*, 16> AddrModeInsts;
2560 ExtAddrMode AddrMode;
2561 TypePromotionTransaction TPT;
2562 TypePromotionTransaction::ConstRestorationPt LastKnownGood =
2563 TPT.getRestorationPoint();
2564 while (!worklist.empty()) {
2565 Value *V = worklist.back();
2566 worklist.pop_back();
2568 // Break use-def graph loops.
2569 if (!Visited.insert(V)) {
2570 Consensus = nullptr;
2574 // For a PHI node, push all of its incoming values.
2575 if (PHINode *P = dyn_cast<PHINode>(V)) {
2576 for (unsigned i = 0, e = P->getNumIncomingValues(); i != e; ++i)
2577 worklist.push_back(P->getIncomingValue(i));
2581 // For non-PHIs, determine the addressing mode being computed.
2582 SmallVector<Instruction*, 16> NewAddrModeInsts;
2583 ExtAddrMode NewAddrMode = AddressingModeMatcher::Match(
2584 V, AccessTy, MemoryInst, NewAddrModeInsts, *TLI, InsertedTruncsSet,
2585 PromotedInsts, TPT);
2587 // This check is broken into two cases with very similar code to avoid using
2588 // getNumUses() as much as possible. Some values have a lot of uses, so
2589 // calling getNumUses() unconditionally caused a significant compile-time
2593 AddrMode = NewAddrMode;
2594 AddrModeInsts = NewAddrModeInsts;
2596 } else if (NewAddrMode == AddrMode) {
2597 if (!IsNumUsesConsensusValid) {
2598 NumUsesConsensus = Consensus->getNumUses();
2599 IsNumUsesConsensusValid = true;
2602 // Ensure that the obtained addressing mode is equivalent to that obtained
2603 // for all other roots of the PHI traversal. Also, when choosing one
2604 // such root as representative, select the one with the most uses in order
2605 // to keep the cost modeling heuristics in AddressingModeMatcher
2607 unsigned NumUses = V->getNumUses();
2608 if (NumUses > NumUsesConsensus) {
2610 NumUsesConsensus = NumUses;
2611 AddrModeInsts = NewAddrModeInsts;
2616 Consensus = nullptr;
2620 // If the addressing mode couldn't be determined, or if multiple different
2621 // ones were determined, bail out now.
2623 TPT.rollback(LastKnownGood);
2628 // Check to see if any of the instructions supersumed by this addr mode are
2629 // non-local to I's BB.
2630 bool AnyNonLocal = false;
2631 for (unsigned i = 0, e = AddrModeInsts.size(); i != e; ++i) {
2632 if (IsNonLocalValue(AddrModeInsts[i], MemoryInst->getParent())) {
2638 // If all the instructions matched are already in this BB, don't do anything.
2640 DEBUG(dbgs() << "CGP: Found local addrmode: " << AddrMode << "\n");
2644 // Insert this computation right after this user. Since our caller is
2645 // scanning from the top of the BB to the bottom, reuse of the expr are
2646 // guaranteed to happen later.
2647 IRBuilder<> Builder(MemoryInst);
2649 // Now that we determined the addressing expression we want to use and know
2650 // that we have to sink it into this block. Check to see if we have already
2651 // done this for some other load/store instr in this block. If so, reuse the
2653 Value *&SunkAddr = SunkAddrs[Addr];
2655 DEBUG(dbgs() << "CGP: Reusing nonlocal addrmode: " << AddrMode << " for "
2657 if (SunkAddr->getType() != Addr->getType())
2658 SunkAddr = Builder.CreateBitCast(SunkAddr, Addr->getType());
2659 } else if (AddrSinkUsingGEPs || (!AddrSinkUsingGEPs.getNumOccurrences() &&
2660 TM && TM->getSubtarget<TargetSubtargetInfo>().useAA())) {
2661 // By default, we use the GEP-based method when AA is used later. This
2662 // prevents new inttoptr/ptrtoint pairs from degrading AA capabilities.
2663 DEBUG(dbgs() << "CGP: SINKING nonlocal addrmode: " << AddrMode << " for "
2665 Type *IntPtrTy = TLI->getDataLayout()->getIntPtrType(Addr->getType());
2666 Value *ResultPtr = nullptr, *ResultIndex = nullptr;
2668 // First, find the pointer.
2669 if (AddrMode.BaseReg && AddrMode.BaseReg->getType()->isPointerTy()) {
2670 ResultPtr = AddrMode.BaseReg;
2671 AddrMode.BaseReg = nullptr;
2674 if (AddrMode.Scale && AddrMode.ScaledReg->getType()->isPointerTy()) {
2675 // We can't add more than one pointer together, nor can we scale a
2676 // pointer (both of which seem meaningless).
2677 if (ResultPtr || AddrMode.Scale != 1)
2680 ResultPtr = AddrMode.ScaledReg;
2684 if (AddrMode.BaseGV) {
2688 ResultPtr = AddrMode.BaseGV;
2691 // If the real base value actually came from an inttoptr, then the matcher
2692 // will look through it and provide only the integer value. In that case,
2694 if (!ResultPtr && AddrMode.BaseReg) {
2696 Builder.CreateIntToPtr(AddrMode.BaseReg, Addr->getType(), "sunkaddr");
2697 AddrMode.BaseReg = nullptr;
2698 } else if (!ResultPtr && AddrMode.Scale == 1) {
2700 Builder.CreateIntToPtr(AddrMode.ScaledReg, Addr->getType(), "sunkaddr");
2705 !AddrMode.BaseReg && !AddrMode.Scale && !AddrMode.BaseOffs) {
2706 SunkAddr = Constant::getNullValue(Addr->getType());
2707 } else if (!ResultPtr) {
2711 Builder.getInt8PtrTy(Addr->getType()->getPointerAddressSpace());
2713 // Start with the base register. Do this first so that subsequent address
2714 // matching finds it last, which will prevent it from trying to match it
2715 // as the scaled value in case it happens to be a mul. That would be
2716 // problematic if we've sunk a different mul for the scale, because then
2717 // we'd end up sinking both muls.
2718 if (AddrMode.BaseReg) {
2719 Value *V = AddrMode.BaseReg;
2720 if (V->getType() != IntPtrTy)
2721 V = Builder.CreateIntCast(V, IntPtrTy, /*isSigned=*/true, "sunkaddr");
2726 // Add the scale value.
2727 if (AddrMode.Scale) {
2728 Value *V = AddrMode.ScaledReg;
2729 if (V->getType() == IntPtrTy) {
2731 } else if (cast<IntegerType>(IntPtrTy)->getBitWidth() <
2732 cast<IntegerType>(V->getType())->getBitWidth()) {
2733 V = Builder.CreateTrunc(V, IntPtrTy, "sunkaddr");
2735 // It is only safe to sign extend the BaseReg if we know that the math
2736 // required to create it did not overflow before we extend it. Since
2737 // the original IR value was tossed in favor of a constant back when
2738 // the AddrMode was created we need to bail out gracefully if widths
2739 // do not match instead of extending it.
2740 Instruction *I = dyn_cast_or_null<Instruction>(ResultIndex);
2741 if (I && (ResultIndex != AddrMode.BaseReg))
2742 I->eraseFromParent();
2746 if (AddrMode.Scale != 1)
2747 V = Builder.CreateMul(V, ConstantInt::get(IntPtrTy, AddrMode.Scale),
2750 ResultIndex = Builder.CreateAdd(ResultIndex, V, "sunkaddr");
2755 // Add in the Base Offset if present.
2756 if (AddrMode.BaseOffs) {
2757 Value *V = ConstantInt::get(IntPtrTy, AddrMode.BaseOffs);
2759 // We need to add this separately from the scale above to help with
2760 // SDAG consecutive load/store merging.
2761 if (ResultPtr->getType() != I8PtrTy)
2762 ResultPtr = Builder.CreateBitCast(ResultPtr, I8PtrTy);
2763 ResultPtr = Builder.CreateGEP(ResultPtr, ResultIndex, "sunkaddr");
2770 SunkAddr = ResultPtr;
2772 if (ResultPtr->getType() != I8PtrTy)
2773 ResultPtr = Builder.CreateBitCast(ResultPtr, I8PtrTy);
2774 SunkAddr = Builder.CreateGEP(ResultPtr, ResultIndex, "sunkaddr");
2777 if (SunkAddr->getType() != Addr->getType())
2778 SunkAddr = Builder.CreateBitCast(SunkAddr, Addr->getType());
2781 DEBUG(dbgs() << "CGP: SINKING nonlocal addrmode: " << AddrMode << " for "
2783 Type *IntPtrTy = TLI->getDataLayout()->getIntPtrType(Addr->getType());
2784 Value *Result = nullptr;
2786 // Start with the base register. Do this first so that subsequent address
2787 // matching finds it last, which will prevent it from trying to match it
2788 // as the scaled value in case it happens to be a mul. That would be
2789 // problematic if we've sunk a different mul for the scale, because then
2790 // we'd end up sinking both muls.
2791 if (AddrMode.BaseReg) {
2792 Value *V = AddrMode.BaseReg;
2793 if (V->getType()->isPointerTy())
2794 V = Builder.CreatePtrToInt(V, IntPtrTy, "sunkaddr");
2795 if (V->getType() != IntPtrTy)
2796 V = Builder.CreateIntCast(V, IntPtrTy, /*isSigned=*/true, "sunkaddr");
2800 // Add the scale value.
2801 if (AddrMode.Scale) {
2802 Value *V = AddrMode.ScaledReg;
2803 if (V->getType() == IntPtrTy) {
2805 } else if (V->getType()->isPointerTy()) {
2806 V = Builder.CreatePtrToInt(V, IntPtrTy, "sunkaddr");
2807 } else if (cast<IntegerType>(IntPtrTy)->getBitWidth() <
2808 cast<IntegerType>(V->getType())->getBitWidth()) {
2809 V = Builder.CreateTrunc(V, IntPtrTy, "sunkaddr");
2811 // It is only safe to sign extend the BaseReg if we know that the math
2812 // required to create it did not overflow before we extend it. Since
2813 // the original IR value was tossed in favor of a constant back when
2814 // the AddrMode was created we need to bail out gracefully if widths
2815 // do not match instead of extending it.
2816 Instruction *I = dyn_cast<Instruction>(Result);
2817 if (I && (Result != AddrMode.BaseReg))
2818 I->eraseFromParent();
2821 if (AddrMode.Scale != 1)
2822 V = Builder.CreateMul(V, ConstantInt::get(IntPtrTy, AddrMode.Scale),
2825 Result = Builder.CreateAdd(Result, V, "sunkaddr");
2830 // Add in the BaseGV if present.
2831 if (AddrMode.BaseGV) {
2832 Value *V = Builder.CreatePtrToInt(AddrMode.BaseGV, IntPtrTy, "sunkaddr");
2834 Result = Builder.CreateAdd(Result, V, "sunkaddr");
2839 // Add in the Base Offset if present.
2840 if (AddrMode.BaseOffs) {
2841 Value *V = ConstantInt::get(IntPtrTy, AddrMode.BaseOffs);
2843 Result = Builder.CreateAdd(Result, V, "sunkaddr");
2849 SunkAddr = Constant::getNullValue(Addr->getType());
2851 SunkAddr = Builder.CreateIntToPtr(Result, Addr->getType(), "sunkaddr");
2854 MemoryInst->replaceUsesOfWith(Repl, SunkAddr);
2856 // If we have no uses, recursively delete the value and all dead instructions
2858 if (Repl->use_empty()) {
2859 // This can cause recursive deletion, which can invalidate our iterator.
2860 // Use a WeakVH to hold onto it in case this happens.
2861 WeakVH IterHandle(CurInstIterator);
2862 BasicBlock *BB = CurInstIterator->getParent();
2864 RecursivelyDeleteTriviallyDeadInstructions(Repl, TLInfo);
2866 if (IterHandle != CurInstIterator) {
2867 // If the iterator instruction was recursively deleted, start over at the
2868 // start of the block.
2869 CurInstIterator = BB->begin();
2877 /// OptimizeInlineAsmInst - If there are any memory operands, use
2878 /// OptimizeMemoryInst to sink their address computing into the block when
2879 /// possible / profitable.
2880 bool CodeGenPrepare::OptimizeInlineAsmInst(CallInst *CS) {
2881 bool MadeChange = false;
2883 TargetLowering::AsmOperandInfoVector
2884 TargetConstraints = TLI->ParseConstraints(CS);
2886 for (unsigned i = 0, e = TargetConstraints.size(); i != e; ++i) {
2887 TargetLowering::AsmOperandInfo &OpInfo = TargetConstraints[i];
2889 // Compute the constraint code and ConstraintType to use.
2890 TLI->ComputeConstraintToUse(OpInfo, SDValue());
2892 if (OpInfo.ConstraintType == TargetLowering::C_Memory &&
2893 OpInfo.isIndirect) {
2894 Value *OpVal = CS->getArgOperand(ArgNo++);
2895 MadeChange |= OptimizeMemoryInst(CS, OpVal, OpVal->getType());
2896 } else if (OpInfo.Type == InlineAsm::isInput)
2903 /// MoveExtToFormExtLoad - Move a zext or sext fed by a load into the same
2904 /// basic block as the load, unless conditions are unfavorable. This allows
2905 /// SelectionDAG to fold the extend into the load.
2907 bool CodeGenPrepare::MoveExtToFormExtLoad(Instruction *I) {
2908 // Look for a load being extended.
2909 LoadInst *LI = dyn_cast<LoadInst>(I->getOperand(0));
2910 if (!LI) return false;
2912 // If they're already in the same block, there's nothing to do.
2913 if (LI->getParent() == I->getParent())
2916 // If the load has other users and the truncate is not free, this probably
2917 // isn't worthwhile.
2918 if (!LI->hasOneUse() &&
2919 TLI && (TLI->isTypeLegal(TLI->getValueType(LI->getType())) ||
2920 !TLI->isTypeLegal(TLI->getValueType(I->getType()))) &&
2921 !TLI->isTruncateFree(I->getType(), LI->getType()))
2924 // Check whether the target supports casts folded into loads.
2926 if (isa<ZExtInst>(I))
2927 LType = ISD::ZEXTLOAD;
2929 assert(isa<SExtInst>(I) && "Unexpected ext type!");
2930 LType = ISD::SEXTLOAD;
2932 if (TLI && !TLI->isLoadExtLegal(LType, TLI->getValueType(LI->getType())))
2935 // Move the extend into the same block as the load, so that SelectionDAG
2937 I->removeFromParent();
2943 bool CodeGenPrepare::OptimizeExtUses(Instruction *I) {
2944 BasicBlock *DefBB = I->getParent();
2946 // If the result of a {s|z}ext and its source are both live out, rewrite all
2947 // other uses of the source with result of extension.
2948 Value *Src = I->getOperand(0);
2949 if (Src->hasOneUse())
2952 // Only do this xform if truncating is free.
2953 if (TLI && !TLI->isTruncateFree(I->getType(), Src->getType()))
2956 // Only safe to perform the optimization if the source is also defined in
2958 if (!isa<Instruction>(Src) || DefBB != cast<Instruction>(Src)->getParent())
2961 bool DefIsLiveOut = false;
2962 for (User *U : I->users()) {
2963 Instruction *UI = cast<Instruction>(U);
2965 // Figure out which BB this ext is used in.
2966 BasicBlock *UserBB = UI->getParent();
2967 if (UserBB == DefBB) continue;
2968 DefIsLiveOut = true;
2974 // Make sure none of the uses are PHI nodes.
2975 for (User *U : Src->users()) {
2976 Instruction *UI = cast<Instruction>(U);
2977 BasicBlock *UserBB = UI->getParent();
2978 if (UserBB == DefBB) continue;
2979 // Be conservative. We don't want this xform to end up introducing
2980 // reloads just before load / store instructions.
2981 if (isa<PHINode>(UI) || isa<LoadInst>(UI) || isa<StoreInst>(UI))
2985 // InsertedTruncs - Only insert one trunc in each block once.
2986 DenseMap<BasicBlock*, Instruction*> InsertedTruncs;
2988 bool MadeChange = false;
2989 for (Use &U : Src->uses()) {
2990 Instruction *User = cast<Instruction>(U.getUser());
2992 // Figure out which BB this ext is used in.
2993 BasicBlock *UserBB = User->getParent();
2994 if (UserBB == DefBB) continue;
2996 // Both src and def are live in this block. Rewrite the use.
2997 Instruction *&InsertedTrunc = InsertedTruncs[UserBB];
2999 if (!InsertedTrunc) {
3000 BasicBlock::iterator InsertPt = UserBB->getFirstInsertionPt();
3001 InsertedTrunc = new TruncInst(I, Src->getType(), "", InsertPt);
3002 InsertedTruncsSet.insert(InsertedTrunc);
3005 // Replace a use of the {s|z}ext source with a use of the result.
3014 /// isFormingBranchFromSelectProfitable - Returns true if a SelectInst should be
3015 /// turned into an explicit branch.
3016 static bool isFormingBranchFromSelectProfitable(SelectInst *SI) {
3017 // FIXME: This should use the same heuristics as IfConversion to determine
3018 // whether a select is better represented as a branch. This requires that
3019 // branch probability metadata is preserved for the select, which is not the
3022 CmpInst *Cmp = dyn_cast<CmpInst>(SI->getCondition());
3024 // If the branch is predicted right, an out of order CPU can avoid blocking on
3025 // the compare. Emit cmovs on compares with a memory operand as branches to
3026 // avoid stalls on the load from memory. If the compare has more than one use
3027 // there's probably another cmov or setcc around so it's not worth emitting a
3032 Value *CmpOp0 = Cmp->getOperand(0);
3033 Value *CmpOp1 = Cmp->getOperand(1);
3035 // We check that the memory operand has one use to avoid uses of the loaded
3036 // value directly after the compare, making branches unprofitable.
3037 return Cmp->hasOneUse() &&
3038 ((isa<LoadInst>(CmpOp0) && CmpOp0->hasOneUse()) ||
3039 (isa<LoadInst>(CmpOp1) && CmpOp1->hasOneUse()));
3043 /// If we have a SelectInst that will likely profit from branch prediction,
3044 /// turn it into a branch.
3045 bool CodeGenPrepare::OptimizeSelectInst(SelectInst *SI) {
3046 bool VectorCond = !SI->getCondition()->getType()->isIntegerTy(1);
3048 // Can we convert the 'select' to CF ?
3049 if (DisableSelectToBranch || OptSize || !TLI || VectorCond)
3052 TargetLowering::SelectSupportKind SelectKind;
3054 SelectKind = TargetLowering::VectorMaskSelect;
3055 else if (SI->getType()->isVectorTy())
3056 SelectKind = TargetLowering::ScalarCondVectorVal;
3058 SelectKind = TargetLowering::ScalarValSelect;
3060 // Do we have efficient codegen support for this kind of 'selects' ?
3061 if (TLI->isSelectSupported(SelectKind)) {
3062 // We have efficient codegen support for the select instruction.
3063 // Check if it is profitable to keep this 'select'.
3064 if (!TLI->isPredictableSelectExpensive() ||
3065 !isFormingBranchFromSelectProfitable(SI))
3071 // First, we split the block containing the select into 2 blocks.
3072 BasicBlock *StartBlock = SI->getParent();
3073 BasicBlock::iterator SplitPt = ++(BasicBlock::iterator(SI));
3074 BasicBlock *NextBlock = StartBlock->splitBasicBlock(SplitPt, "select.end");
3076 // Create a new block serving as the landing pad for the branch.
3077 BasicBlock *SmallBlock = BasicBlock::Create(SI->getContext(), "select.mid",
3078 NextBlock->getParent(), NextBlock);
3080 // Move the unconditional branch from the block with the select in it into our
3081 // landing pad block.
3082 StartBlock->getTerminator()->eraseFromParent();
3083 BranchInst::Create(NextBlock, SmallBlock);
3085 // Insert the real conditional branch based on the original condition.
3086 BranchInst::Create(NextBlock, SmallBlock, SI->getCondition(), SI);
3088 // The select itself is replaced with a PHI Node.
3089 PHINode *PN = PHINode::Create(SI->getType(), 2, "", NextBlock->begin());
3091 PN->addIncoming(SI->getTrueValue(), StartBlock);
3092 PN->addIncoming(SI->getFalseValue(), SmallBlock);
3093 SI->replaceAllUsesWith(PN);
3094 SI->eraseFromParent();
3096 // Instruct OptimizeBlock to skip to the next block.
3097 CurInstIterator = StartBlock->end();
3098 ++NumSelectsExpanded;
3102 static bool isBroadcastShuffle(ShuffleVectorInst *SVI) {
3103 SmallVector<int, 16> Mask(SVI->getShuffleMask());
3105 for (unsigned i = 0; i < Mask.size(); ++i) {
3106 if (SplatElem != -1 && Mask[i] != -1 && Mask[i] != SplatElem)
3108 SplatElem = Mask[i];
3114 /// Some targets have expensive vector shifts if the lanes aren't all the same
3115 /// (e.g. x86 only introduced "vpsllvd" and friends with AVX2). In these cases
3116 /// it's often worth sinking a shufflevector splat down to its use so that
3117 /// codegen can spot all lanes are identical.
3118 bool CodeGenPrepare::OptimizeShuffleVectorInst(ShuffleVectorInst *SVI) {
3119 BasicBlock *DefBB = SVI->getParent();
3121 // Only do this xform if variable vector shifts are particularly expensive.
3122 if (!TLI || !TLI->isVectorShiftByScalarCheap(SVI->getType()))
3125 // We only expect better codegen by sinking a shuffle if we can recognise a
3127 if (!isBroadcastShuffle(SVI))
3130 // InsertedShuffles - Only insert a shuffle in each block once.
3131 DenseMap<BasicBlock*, Instruction*> InsertedShuffles;
3133 bool MadeChange = false;
3134 for (User *U : SVI->users()) {
3135 Instruction *UI = cast<Instruction>(U);
3137 // Figure out which BB this ext is used in.
3138 BasicBlock *UserBB = UI->getParent();
3139 if (UserBB == DefBB) continue;
3141 // For now only apply this when the splat is used by a shift instruction.
3142 if (!UI->isShift()) continue;
3144 // Everything checks out, sink the shuffle if the user's block doesn't
3145 // already have a copy.
3146 Instruction *&InsertedShuffle = InsertedShuffles[UserBB];
3148 if (!InsertedShuffle) {
3149 BasicBlock::iterator InsertPt = UserBB->getFirstInsertionPt();
3150 InsertedShuffle = new ShuffleVectorInst(SVI->getOperand(0),
3152 SVI->getOperand(2), "", InsertPt);
3155 UI->replaceUsesOfWith(SVI, InsertedShuffle);
3159 // If we removed all uses, nuke the shuffle.
3160 if (SVI->use_empty()) {
3161 SVI->eraseFromParent();
3168 bool CodeGenPrepare::OptimizeInst(Instruction *I) {
3169 if (PHINode *P = dyn_cast<PHINode>(I)) {
3170 // It is possible for very late stage optimizations (such as SimplifyCFG)
3171 // to introduce PHI nodes too late to be cleaned up. If we detect such a
3172 // trivial PHI, go ahead and zap it here.
3173 if (Value *V = SimplifyInstruction(P, TLI ? TLI->getDataLayout() : nullptr,
3175 P->replaceAllUsesWith(V);
3176 P->eraseFromParent();
3183 if (CastInst *CI = dyn_cast<CastInst>(I)) {
3184 // If the source of the cast is a constant, then this should have
3185 // already been constant folded. The only reason NOT to constant fold
3186 // it is if something (e.g. LSR) was careful to place the constant
3187 // evaluation in a block other than then one that uses it (e.g. to hoist
3188 // the address of globals out of a loop). If this is the case, we don't
3189 // want to forward-subst the cast.
3190 if (isa<Constant>(CI->getOperand(0)))
3193 if (TLI && OptimizeNoopCopyExpression(CI, *TLI))
3196 if (isa<ZExtInst>(I) || isa<SExtInst>(I)) {
3197 /// Sink a zext or sext into its user blocks if the target type doesn't
3198 /// fit in one register
3199 if (TLI && TLI->getTypeAction(CI->getContext(),
3200 TLI->getValueType(CI->getType())) ==
3201 TargetLowering::TypeExpandInteger) {
3202 return SinkCast(CI);
3204 bool MadeChange = MoveExtToFormExtLoad(I);
3205 return MadeChange | OptimizeExtUses(I);
3211 if (CmpInst *CI = dyn_cast<CmpInst>(I))
3212 if (!TLI || !TLI->hasMultipleConditionRegisters())
3213 return OptimizeCmpExpression(CI);
3215 if (LoadInst *LI = dyn_cast<LoadInst>(I)) {
3217 return OptimizeMemoryInst(I, I->getOperand(0), LI->getType());
3221 if (StoreInst *SI = dyn_cast<StoreInst>(I)) {
3223 return OptimizeMemoryInst(I, SI->getOperand(1),
3224 SI->getOperand(0)->getType());
3228 if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(I)) {
3229 if (GEPI->hasAllZeroIndices()) {
3230 /// The GEP operand must be a pointer, so must its result -> BitCast
3231 Instruction *NC = new BitCastInst(GEPI->getOperand(0), GEPI->getType(),
3232 GEPI->getName(), GEPI);
3233 GEPI->replaceAllUsesWith(NC);
3234 GEPI->eraseFromParent();
3242 if (CallInst *CI = dyn_cast<CallInst>(I))
3243 return OptimizeCallInst(CI);
3245 if (SelectInst *SI = dyn_cast<SelectInst>(I))
3246 return OptimizeSelectInst(SI);
3248 if (ShuffleVectorInst *SVI = dyn_cast<ShuffleVectorInst>(I))
3249 return OptimizeShuffleVectorInst(SVI);
3254 // In this pass we look for GEP and cast instructions that are used
3255 // across basic blocks and rewrite them to improve basic-block-at-a-time
3257 bool CodeGenPrepare::OptimizeBlock(BasicBlock &BB) {
3259 bool MadeChange = false;
3261 CurInstIterator = BB.begin();
3262 while (CurInstIterator != BB.end())
3263 MadeChange |= OptimizeInst(CurInstIterator++);
3265 MadeChange |= DupRetToEnableTailCallOpts(&BB);
3270 // llvm.dbg.value is far away from the value then iSel may not be able
3271 // handle it properly. iSel will drop llvm.dbg.value if it can not
3272 // find a node corresponding to the value.
3273 bool CodeGenPrepare::PlaceDbgValues(Function &F) {
3274 bool MadeChange = false;
3275 for (Function::iterator I = F.begin(), E = F.end(); I != E; ++I) {
3276 Instruction *PrevNonDbgInst = nullptr;
3277 for (BasicBlock::iterator BI = I->begin(), BE = I->end(); BI != BE;) {
3278 Instruction *Insn = BI; ++BI;
3279 DbgValueInst *DVI = dyn_cast<DbgValueInst>(Insn);
3281 PrevNonDbgInst = Insn;
3285 Instruction *VI = dyn_cast_or_null<Instruction>(DVI->getValue());
3286 if (VI && VI != PrevNonDbgInst && !VI->isTerminator()) {
3287 DEBUG(dbgs() << "Moving Debug Value before :\n" << *DVI << ' ' << *VI);
3288 DVI->removeFromParent();
3289 if (isa<PHINode>(VI))
3290 DVI->insertBefore(VI->getParent()->getFirstInsertionPt());
3292 DVI->insertAfter(VI);
3301 // If there is a sequence that branches based on comparing a single bit
3302 // against zero that can be combined into a single instruction, and the
3303 // target supports folding these into a single instruction, sink the
3304 // mask and compare into the branch uses. Do this before OptimizeBlock ->
3305 // OptimizeInst -> OptimizeCmpExpression, which perturbs the pattern being
3307 bool CodeGenPrepare::sinkAndCmp(Function &F) {
3308 if (!EnableAndCmpSinking)
3310 if (!TLI || !TLI->isMaskAndBranchFoldingLegal())
3312 bool MadeChange = false;
3313 for (Function::iterator I = F.begin(), E = F.end(); I != E; ) {
3314 BasicBlock *BB = I++;
3316 // Does this BB end with the following?
3317 // %andVal = and %val, #single-bit-set
3318 // %icmpVal = icmp %andResult, 0
3319 // br i1 %cmpVal label %dest1, label %dest2"
3320 BranchInst *Brcc = dyn_cast<BranchInst>(BB->getTerminator());
3321 if (!Brcc || !Brcc->isConditional())
3323 ICmpInst *Cmp = dyn_cast<ICmpInst>(Brcc->getOperand(0));
3324 if (!Cmp || Cmp->getParent() != BB)
3326 ConstantInt *Zero = dyn_cast<ConstantInt>(Cmp->getOperand(1));
3327 if (!Zero || !Zero->isZero())
3329 Instruction *And = dyn_cast<Instruction>(Cmp->getOperand(0));
3330 if (!And || And->getOpcode() != Instruction::And || And->getParent() != BB)
3332 ConstantInt* Mask = dyn_cast<ConstantInt>(And->getOperand(1));
3333 if (!Mask || !Mask->getUniqueInteger().isPowerOf2())
3335 DEBUG(dbgs() << "found and; icmp ?,0; brcc\n"); DEBUG(BB->dump());
3337 // Push the "and; icmp" for any users that are conditional branches.
3338 // Since there can only be one branch use per BB, we don't need to keep
3339 // track of which BBs we insert into.
3340 for (Value::use_iterator UI = Cmp->use_begin(), E = Cmp->use_end();
3344 BranchInst *BrccUser = dyn_cast<BranchInst>(*UI);
3346 if (!BrccUser || !BrccUser->isConditional())
3348 BasicBlock *UserBB = BrccUser->getParent();
3349 if (UserBB == BB) continue;
3350 DEBUG(dbgs() << "found Brcc use\n");
3352 // Sink the "and; icmp" to use.
3354 BinaryOperator *NewAnd =
3355 BinaryOperator::CreateAnd(And->getOperand(0), And->getOperand(1), "",
3358 CmpInst::Create(Cmp->getOpcode(), Cmp->getPredicate(), NewAnd, Zero,
3362 DEBUG(BrccUser->getParent()->dump());