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();
693 SmallVector<Value*, 2> PtrOps;
695 if (TLI->GetAddrModeArguments(II, PtrOps, AccessTy))
696 while (!PtrOps.empty())
697 if (OptimizeMemoryInst(II, PtrOps.pop_back_val(), AccessTy))
701 // From here on out we're working with named functions.
702 if (!CI->getCalledFunction()) return false;
704 // We'll need DataLayout from here on out.
705 const DataLayout *TD = TLI ? TLI->getDataLayout() : nullptr;
706 if (!TD) return false;
708 // Lower all default uses of _chk calls. This is very similar
709 // to what InstCombineCalls does, but here we are only lowering calls
710 // that have the default "don't know" as the objectsize. Anything else
711 // should be left alone.
712 CodeGenPrepareFortifiedLibCalls Simplifier;
713 return Simplifier.fold(CI, TD, TLInfo);
716 /// DupRetToEnableTailCallOpts - Look for opportunities to duplicate return
717 /// instructions to the predecessor to enable tail call optimizations. The
718 /// case it is currently looking for is:
721 /// %tmp0 = tail call i32 @f0()
724 /// %tmp1 = tail call i32 @f1()
727 /// %tmp2 = tail call i32 @f2()
730 /// %retval = phi i32 [ %tmp0, %bb0 ], [ %tmp1, %bb1 ], [ %tmp2, %bb2 ]
738 /// %tmp0 = tail call i32 @f0()
741 /// %tmp1 = tail call i32 @f1()
744 /// %tmp2 = tail call i32 @f2()
747 bool CodeGenPrepare::DupRetToEnableTailCallOpts(BasicBlock *BB) {
751 ReturnInst *RI = dyn_cast<ReturnInst>(BB->getTerminator());
755 PHINode *PN = nullptr;
756 BitCastInst *BCI = nullptr;
757 Value *V = RI->getReturnValue();
759 BCI = dyn_cast<BitCastInst>(V);
761 V = BCI->getOperand(0);
763 PN = dyn_cast<PHINode>(V);
768 if (PN && PN->getParent() != BB)
771 // It's not safe to eliminate the sign / zero extension of the return value.
772 // See llvm::isInTailCallPosition().
773 const Function *F = BB->getParent();
774 AttributeSet CallerAttrs = F->getAttributes();
775 if (CallerAttrs.hasAttribute(AttributeSet::ReturnIndex, Attribute::ZExt) ||
776 CallerAttrs.hasAttribute(AttributeSet::ReturnIndex, Attribute::SExt))
779 // Make sure there are no instructions between the PHI and return, or that the
780 // return is the first instruction in the block.
782 BasicBlock::iterator BI = BB->begin();
783 do { ++BI; } while (isa<DbgInfoIntrinsic>(BI));
785 // Also skip over the bitcast.
790 BasicBlock::iterator BI = BB->begin();
791 while (isa<DbgInfoIntrinsic>(BI)) ++BI;
796 /// Only dup the ReturnInst if the CallInst is likely to be emitted as a tail
798 SmallVector<CallInst*, 4> TailCalls;
800 for (unsigned I = 0, E = PN->getNumIncomingValues(); I != E; ++I) {
801 CallInst *CI = dyn_cast<CallInst>(PN->getIncomingValue(I));
802 // Make sure the phi value is indeed produced by the tail call.
803 if (CI && CI->hasOneUse() && CI->getParent() == PN->getIncomingBlock(I) &&
804 TLI->mayBeEmittedAsTailCall(CI))
805 TailCalls.push_back(CI);
808 SmallPtrSet<BasicBlock*, 4> VisitedBBs;
809 for (pred_iterator PI = pred_begin(BB), PE = pred_end(BB); PI != PE; ++PI) {
810 if (!VisitedBBs.insert(*PI))
813 BasicBlock::InstListType &InstList = (*PI)->getInstList();
814 BasicBlock::InstListType::reverse_iterator RI = InstList.rbegin();
815 BasicBlock::InstListType::reverse_iterator RE = InstList.rend();
816 do { ++RI; } while (RI != RE && isa<DbgInfoIntrinsic>(&*RI));
820 CallInst *CI = dyn_cast<CallInst>(&*RI);
821 if (CI && CI->use_empty() && TLI->mayBeEmittedAsTailCall(CI))
822 TailCalls.push_back(CI);
826 bool Changed = false;
827 for (unsigned i = 0, e = TailCalls.size(); i != e; ++i) {
828 CallInst *CI = TailCalls[i];
831 // Conservatively require the attributes of the call to match those of the
832 // return. Ignore noalias because it doesn't affect the call sequence.
833 AttributeSet CalleeAttrs = CS.getAttributes();
834 if (AttrBuilder(CalleeAttrs, AttributeSet::ReturnIndex).
835 removeAttribute(Attribute::NoAlias) !=
836 AttrBuilder(CalleeAttrs, AttributeSet::ReturnIndex).
837 removeAttribute(Attribute::NoAlias))
840 // Make sure the call instruction is followed by an unconditional branch to
842 BasicBlock *CallBB = CI->getParent();
843 BranchInst *BI = dyn_cast<BranchInst>(CallBB->getTerminator());
844 if (!BI || !BI->isUnconditional() || BI->getSuccessor(0) != BB)
847 // Duplicate the return into CallBB.
848 (void)FoldReturnIntoUncondBranch(RI, BB, CallBB);
849 ModifiedDT = Changed = true;
853 // If we eliminated all predecessors of the block, delete the block now.
854 if (Changed && !BB->hasAddressTaken() && pred_begin(BB) == pred_end(BB))
855 BB->eraseFromParent();
860 //===----------------------------------------------------------------------===//
861 // Memory Optimization
862 //===----------------------------------------------------------------------===//
866 /// ExtAddrMode - This is an extended version of TargetLowering::AddrMode
867 /// which holds actual Value*'s for register values.
868 struct ExtAddrMode : public TargetLowering::AddrMode {
871 ExtAddrMode() : BaseReg(nullptr), ScaledReg(nullptr) {}
872 void print(raw_ostream &OS) const;
875 bool operator==(const ExtAddrMode& O) const {
876 return (BaseReg == O.BaseReg) && (ScaledReg == O.ScaledReg) &&
877 (BaseGV == O.BaseGV) && (BaseOffs == O.BaseOffs) &&
878 (HasBaseReg == O.HasBaseReg) && (Scale == O.Scale);
883 static inline raw_ostream &operator<<(raw_ostream &OS, const ExtAddrMode &AM) {
889 void ExtAddrMode::print(raw_ostream &OS) const {
890 bool NeedPlus = false;
893 OS << (NeedPlus ? " + " : "")
895 BaseGV->printAsOperand(OS, /*PrintType=*/false);
900 OS << (NeedPlus ? " + " : "") << BaseOffs, NeedPlus = true;
903 OS << (NeedPlus ? " + " : "")
905 BaseReg->printAsOperand(OS, /*PrintType=*/false);
909 OS << (NeedPlus ? " + " : "")
911 ScaledReg->printAsOperand(OS, /*PrintType=*/false);
917 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
918 void ExtAddrMode::dump() const {
924 /// \brief This class provides transaction based operation on the IR.
925 /// Every change made through this class is recorded in the internal state and
926 /// can be undone (rollback) until commit is called.
927 class TypePromotionTransaction {
929 /// \brief This represents the common interface of the individual transaction.
930 /// Each class implements the logic for doing one specific modification on
931 /// the IR via the TypePromotionTransaction.
932 class TypePromotionAction {
934 /// The Instruction modified.
938 /// \brief Constructor of the action.
939 /// The constructor performs the related action on the IR.
940 TypePromotionAction(Instruction *Inst) : Inst(Inst) {}
942 virtual ~TypePromotionAction() {}
944 /// \brief Undo the modification done by this action.
945 /// When this method is called, the IR must be in the same state as it was
946 /// before this action was applied.
947 /// \pre Undoing the action works if and only if the IR is in the exact same
948 /// state as it was directly after this action was applied.
949 virtual void undo() = 0;
951 /// \brief Advocate every change made by this action.
952 /// When the results on the IR of the action are to be kept, it is important
953 /// to call this function, otherwise hidden information may be kept forever.
954 virtual void commit() {
955 // Nothing to be done, this action is not doing anything.
959 /// \brief Utility to remember the position of an instruction.
960 class InsertionHandler {
961 /// Position of an instruction.
962 /// Either an instruction:
963 /// - Is the first in a basic block: BB is used.
964 /// - Has a previous instructon: PrevInst is used.
966 Instruction *PrevInst;
969 /// Remember whether or not the instruction had a previous instruction.
970 bool HasPrevInstruction;
973 /// \brief Record the position of \p Inst.
974 InsertionHandler(Instruction *Inst) {
975 BasicBlock::iterator It = Inst;
976 HasPrevInstruction = (It != (Inst->getParent()->begin()));
977 if (HasPrevInstruction)
978 Point.PrevInst = --It;
980 Point.BB = Inst->getParent();
983 /// \brief Insert \p Inst at the recorded position.
984 void insert(Instruction *Inst) {
985 if (HasPrevInstruction) {
986 if (Inst->getParent())
987 Inst->removeFromParent();
988 Inst->insertAfter(Point.PrevInst);
990 Instruction *Position = Point.BB->getFirstInsertionPt();
991 if (Inst->getParent())
992 Inst->moveBefore(Position);
994 Inst->insertBefore(Position);
999 /// \brief Move an instruction before another.
1000 class InstructionMoveBefore : public TypePromotionAction {
1001 /// Original position of the instruction.
1002 InsertionHandler Position;
1005 /// \brief Move \p Inst before \p Before.
1006 InstructionMoveBefore(Instruction *Inst, Instruction *Before)
1007 : TypePromotionAction(Inst), Position(Inst) {
1008 DEBUG(dbgs() << "Do: move: " << *Inst << "\nbefore: " << *Before << "\n");
1009 Inst->moveBefore(Before);
1012 /// \brief Move the instruction back to its original position.
1013 void undo() override {
1014 DEBUG(dbgs() << "Undo: moveBefore: " << *Inst << "\n");
1015 Position.insert(Inst);
1019 /// \brief Set the operand of an instruction with a new value.
1020 class OperandSetter : public TypePromotionAction {
1021 /// Original operand of the instruction.
1023 /// Index of the modified instruction.
1027 /// \brief Set \p Idx operand of \p Inst with \p NewVal.
1028 OperandSetter(Instruction *Inst, unsigned Idx, Value *NewVal)
1029 : TypePromotionAction(Inst), Idx(Idx) {
1030 DEBUG(dbgs() << "Do: setOperand: " << Idx << "\n"
1031 << "for:" << *Inst << "\n"
1032 << "with:" << *NewVal << "\n");
1033 Origin = Inst->getOperand(Idx);
1034 Inst->setOperand(Idx, NewVal);
1037 /// \brief Restore the original value of the instruction.
1038 void undo() override {
1039 DEBUG(dbgs() << "Undo: setOperand:" << Idx << "\n"
1040 << "for: " << *Inst << "\n"
1041 << "with: " << *Origin << "\n");
1042 Inst->setOperand(Idx, Origin);
1046 /// \brief Hide the operands of an instruction.
1047 /// Do as if this instruction was not using any of its operands.
1048 class OperandsHider : public TypePromotionAction {
1049 /// The list of original operands.
1050 SmallVector<Value *, 4> OriginalValues;
1053 /// \brief Remove \p Inst from the uses of the operands of \p Inst.
1054 OperandsHider(Instruction *Inst) : TypePromotionAction(Inst) {
1055 DEBUG(dbgs() << "Do: OperandsHider: " << *Inst << "\n");
1056 unsigned NumOpnds = Inst->getNumOperands();
1057 OriginalValues.reserve(NumOpnds);
1058 for (unsigned It = 0; It < NumOpnds; ++It) {
1059 // Save the current operand.
1060 Value *Val = Inst->getOperand(It);
1061 OriginalValues.push_back(Val);
1063 // We could use OperandSetter here, but that would implied an overhead
1064 // that we are not willing to pay.
1065 Inst->setOperand(It, UndefValue::get(Val->getType()));
1069 /// \brief Restore the original list of uses.
1070 void undo() override {
1071 DEBUG(dbgs() << "Undo: OperandsHider: " << *Inst << "\n");
1072 for (unsigned It = 0, EndIt = OriginalValues.size(); It != EndIt; ++It)
1073 Inst->setOperand(It, OriginalValues[It]);
1077 /// \brief Build a truncate instruction.
1078 class TruncBuilder : public TypePromotionAction {
1080 /// \brief Build a truncate instruction of \p Opnd producing a \p Ty
1082 /// trunc Opnd to Ty.
1083 TruncBuilder(Instruction *Opnd, Type *Ty) : TypePromotionAction(Opnd) {
1084 IRBuilder<> Builder(Opnd);
1085 Inst = cast<Instruction>(Builder.CreateTrunc(Opnd, Ty, "promoted"));
1086 DEBUG(dbgs() << "Do: TruncBuilder: " << *Inst << "\n");
1089 /// \brief Get the built instruction.
1090 Instruction *getBuiltInstruction() { return Inst; }
1092 /// \brief Remove the built instruction.
1093 void undo() override {
1094 DEBUG(dbgs() << "Undo: TruncBuilder: " << *Inst << "\n");
1095 Inst->eraseFromParent();
1099 /// \brief Build a sign extension instruction.
1100 class SExtBuilder : public TypePromotionAction {
1102 /// \brief Build a sign extension instruction of \p Opnd producing a \p Ty
1104 /// sext Opnd to Ty.
1105 SExtBuilder(Instruction *InsertPt, Value *Opnd, Type *Ty)
1106 : TypePromotionAction(Inst) {
1107 IRBuilder<> Builder(InsertPt);
1108 Inst = cast<Instruction>(Builder.CreateSExt(Opnd, Ty, "promoted"));
1109 DEBUG(dbgs() << "Do: SExtBuilder: " << *Inst << "\n");
1112 /// \brief Get the built instruction.
1113 Instruction *getBuiltInstruction() { return Inst; }
1115 /// \brief Remove the built instruction.
1116 void undo() override {
1117 DEBUG(dbgs() << "Undo: SExtBuilder: " << *Inst << "\n");
1118 Inst->eraseFromParent();
1122 /// \brief Mutate an instruction to another type.
1123 class TypeMutator : public TypePromotionAction {
1124 /// Record the original type.
1128 /// \brief Mutate the type of \p Inst into \p NewTy.
1129 TypeMutator(Instruction *Inst, Type *NewTy)
1130 : TypePromotionAction(Inst), OrigTy(Inst->getType()) {
1131 DEBUG(dbgs() << "Do: MutateType: " << *Inst << " with " << *NewTy
1133 Inst->mutateType(NewTy);
1136 /// \brief Mutate the instruction back to its original type.
1137 void undo() override {
1138 DEBUG(dbgs() << "Undo: MutateType: " << *Inst << " with " << *OrigTy
1140 Inst->mutateType(OrigTy);
1144 /// \brief Replace the uses of an instruction by another instruction.
1145 class UsesReplacer : public TypePromotionAction {
1146 /// Helper structure to keep track of the replaced uses.
1147 struct InstructionAndIdx {
1148 /// The instruction using the instruction.
1150 /// The index where this instruction is used for Inst.
1152 InstructionAndIdx(Instruction *Inst, unsigned Idx)
1153 : Inst(Inst), Idx(Idx) {}
1156 /// Keep track of the original uses (pair Instruction, Index).
1157 SmallVector<InstructionAndIdx, 4> OriginalUses;
1158 typedef SmallVectorImpl<InstructionAndIdx>::iterator use_iterator;
1161 /// \brief Replace all the use of \p Inst by \p New.
1162 UsesReplacer(Instruction *Inst, Value *New) : TypePromotionAction(Inst) {
1163 DEBUG(dbgs() << "Do: UsersReplacer: " << *Inst << " with " << *New
1165 // Record the original uses.
1166 for (Use &U : Inst->uses()) {
1167 Instruction *UserI = cast<Instruction>(U.getUser());
1168 OriginalUses.push_back(InstructionAndIdx(UserI, U.getOperandNo()));
1170 // Now, we can replace the uses.
1171 Inst->replaceAllUsesWith(New);
1174 /// \brief Reassign the original uses of Inst to Inst.
1175 void undo() override {
1176 DEBUG(dbgs() << "Undo: UsersReplacer: " << *Inst << "\n");
1177 for (use_iterator UseIt = OriginalUses.begin(),
1178 EndIt = OriginalUses.end();
1179 UseIt != EndIt; ++UseIt) {
1180 UseIt->Inst->setOperand(UseIt->Idx, Inst);
1185 /// \brief Remove an instruction from the IR.
1186 class InstructionRemover : public TypePromotionAction {
1187 /// Original position of the instruction.
1188 InsertionHandler Inserter;
1189 /// Helper structure to hide all the link to the instruction. In other
1190 /// words, this helps to do as if the instruction was removed.
1191 OperandsHider Hider;
1192 /// Keep track of the uses replaced, if any.
1193 UsesReplacer *Replacer;
1196 /// \brief Remove all reference of \p Inst and optinally replace all its
1198 /// \pre If !Inst->use_empty(), then New != nullptr
1199 InstructionRemover(Instruction *Inst, Value *New = nullptr)
1200 : TypePromotionAction(Inst), Inserter(Inst), Hider(Inst),
1203 Replacer = new UsesReplacer(Inst, New);
1204 DEBUG(dbgs() << "Do: InstructionRemover: " << *Inst << "\n");
1205 Inst->removeFromParent();
1208 ~InstructionRemover() { delete Replacer; }
1210 /// \brief Really remove the instruction.
1211 void commit() override { delete Inst; }
1213 /// \brief Resurrect the instruction and reassign it to the proper uses if
1214 /// new value was provided when build this action.
1215 void undo() override {
1216 DEBUG(dbgs() << "Undo: InstructionRemover: " << *Inst << "\n");
1217 Inserter.insert(Inst);
1225 /// Restoration point.
1226 /// The restoration point is a pointer to an action instead of an iterator
1227 /// because the iterator may be invalidated but not the pointer.
1228 typedef const TypePromotionAction *ConstRestorationPt;
1229 /// Advocate every changes made in that transaction.
1231 /// Undo all the changes made after the given point.
1232 void rollback(ConstRestorationPt Point);
1233 /// Get the current restoration point.
1234 ConstRestorationPt getRestorationPoint() const;
1236 /// \name API for IR modification with state keeping to support rollback.
1238 /// Same as Instruction::setOperand.
1239 void setOperand(Instruction *Inst, unsigned Idx, Value *NewVal);
1240 /// Same as Instruction::eraseFromParent.
1241 void eraseInstruction(Instruction *Inst, Value *NewVal = nullptr);
1242 /// Same as Value::replaceAllUsesWith.
1243 void replaceAllUsesWith(Instruction *Inst, Value *New);
1244 /// Same as Value::mutateType.
1245 void mutateType(Instruction *Inst, Type *NewTy);
1246 /// Same as IRBuilder::createTrunc.
1247 Instruction *createTrunc(Instruction *Opnd, Type *Ty);
1248 /// Same as IRBuilder::createSExt.
1249 Instruction *createSExt(Instruction *Inst, Value *Opnd, Type *Ty);
1250 /// Same as Instruction::moveBefore.
1251 void moveBefore(Instruction *Inst, Instruction *Before);
1255 /// The ordered list of actions made so far.
1256 SmallVector<std::unique_ptr<TypePromotionAction>, 16> Actions;
1257 typedef SmallVectorImpl<std::unique_ptr<TypePromotionAction>>::iterator CommitPt;
1260 void TypePromotionTransaction::setOperand(Instruction *Inst, unsigned Idx,
1263 make_unique<TypePromotionTransaction::OperandSetter>(Inst, Idx, NewVal));
1266 void TypePromotionTransaction::eraseInstruction(Instruction *Inst,
1269 make_unique<TypePromotionTransaction::InstructionRemover>(Inst, NewVal));
1272 void TypePromotionTransaction::replaceAllUsesWith(Instruction *Inst,
1274 Actions.push_back(make_unique<TypePromotionTransaction::UsesReplacer>(Inst, New));
1277 void TypePromotionTransaction::mutateType(Instruction *Inst, Type *NewTy) {
1278 Actions.push_back(make_unique<TypePromotionTransaction::TypeMutator>(Inst, NewTy));
1281 Instruction *TypePromotionTransaction::createTrunc(Instruction *Opnd,
1283 std::unique_ptr<TruncBuilder> Ptr(new TruncBuilder(Opnd, Ty));
1284 Instruction *I = Ptr->getBuiltInstruction();
1285 Actions.push_back(std::move(Ptr));
1289 Instruction *TypePromotionTransaction::createSExt(Instruction *Inst,
1290 Value *Opnd, Type *Ty) {
1291 std::unique_ptr<SExtBuilder> Ptr(new SExtBuilder(Inst, Opnd, Ty));
1292 Instruction *I = Ptr->getBuiltInstruction();
1293 Actions.push_back(std::move(Ptr));
1297 void TypePromotionTransaction::moveBefore(Instruction *Inst,
1298 Instruction *Before) {
1300 make_unique<TypePromotionTransaction::InstructionMoveBefore>(Inst, Before));
1303 TypePromotionTransaction::ConstRestorationPt
1304 TypePromotionTransaction::getRestorationPoint() const {
1305 return !Actions.empty() ? Actions.back().get() : nullptr;
1308 void TypePromotionTransaction::commit() {
1309 for (CommitPt It = Actions.begin(), EndIt = Actions.end(); It != EndIt;
1315 void TypePromotionTransaction::rollback(
1316 TypePromotionTransaction::ConstRestorationPt Point) {
1317 while (!Actions.empty() && Point != Actions.back().get()) {
1318 std::unique_ptr<TypePromotionAction> Curr = Actions.pop_back_val();
1323 /// \brief A helper class for matching addressing modes.
1325 /// This encapsulates the logic for matching the target-legal addressing modes.
1326 class AddressingModeMatcher {
1327 SmallVectorImpl<Instruction*> &AddrModeInsts;
1328 const TargetLowering &TLI;
1330 /// AccessTy/MemoryInst - This is the type for the access (e.g. double) and
1331 /// the memory instruction that we're computing this address for.
1333 Instruction *MemoryInst;
1335 /// AddrMode - This is the addressing mode that we're building up. This is
1336 /// part of the return value of this addressing mode matching stuff.
1337 ExtAddrMode &AddrMode;
1339 /// The truncate instruction inserted by other CodeGenPrepare optimizations.
1340 const SetOfInstrs &InsertedTruncs;
1341 /// A map from the instructions to their type before promotion.
1342 InstrToOrigTy &PromotedInsts;
1343 /// The ongoing transaction where every action should be registered.
1344 TypePromotionTransaction &TPT;
1346 /// IgnoreProfitability - This is set to true when we should not do
1347 /// profitability checks. When true, IsProfitableToFoldIntoAddressingMode
1348 /// always returns true.
1349 bool IgnoreProfitability;
1351 AddressingModeMatcher(SmallVectorImpl<Instruction*> &AMI,
1352 const TargetLowering &T, Type *AT,
1353 Instruction *MI, ExtAddrMode &AM,
1354 const SetOfInstrs &InsertedTruncs,
1355 InstrToOrigTy &PromotedInsts,
1356 TypePromotionTransaction &TPT)
1357 : AddrModeInsts(AMI), TLI(T), AccessTy(AT), MemoryInst(MI), AddrMode(AM),
1358 InsertedTruncs(InsertedTruncs), PromotedInsts(PromotedInsts), TPT(TPT) {
1359 IgnoreProfitability = false;
1363 /// Match - Find the maximal addressing mode that a load/store of V can fold,
1364 /// give an access type of AccessTy. This returns a list of involved
1365 /// instructions in AddrModeInsts.
1366 /// \p InsertedTruncs The truncate instruction inserted by other
1369 /// \p PromotedInsts maps the instructions to their type before promotion.
1370 /// \p The ongoing transaction where every action should be registered.
1371 static ExtAddrMode Match(Value *V, Type *AccessTy,
1372 Instruction *MemoryInst,
1373 SmallVectorImpl<Instruction*> &AddrModeInsts,
1374 const TargetLowering &TLI,
1375 const SetOfInstrs &InsertedTruncs,
1376 InstrToOrigTy &PromotedInsts,
1377 TypePromotionTransaction &TPT) {
1380 bool Success = AddressingModeMatcher(AddrModeInsts, TLI, AccessTy,
1381 MemoryInst, Result, InsertedTruncs,
1382 PromotedInsts, TPT).MatchAddr(V, 0);
1383 (void)Success; assert(Success && "Couldn't select *anything*?");
1387 bool MatchScaledValue(Value *ScaleReg, int64_t Scale, unsigned Depth);
1388 bool MatchAddr(Value *V, unsigned Depth);
1389 bool MatchOperationAddr(User *Operation, unsigned Opcode, unsigned Depth,
1390 bool *MovedAway = nullptr);
1391 bool IsProfitableToFoldIntoAddressingMode(Instruction *I,
1392 ExtAddrMode &AMBefore,
1393 ExtAddrMode &AMAfter);
1394 bool ValueAlreadyLiveAtInst(Value *Val, Value *KnownLive1, Value *KnownLive2);
1395 bool IsPromotionProfitable(unsigned MatchedSize, unsigned SizeWithPromotion,
1396 Value *PromotedOperand) const;
1399 /// MatchScaledValue - Try adding ScaleReg*Scale to the current addressing mode.
1400 /// Return true and update AddrMode if this addr mode is legal for the target,
1402 bool AddressingModeMatcher::MatchScaledValue(Value *ScaleReg, int64_t Scale,
1404 // If Scale is 1, then this is the same as adding ScaleReg to the addressing
1405 // mode. Just process that directly.
1407 return MatchAddr(ScaleReg, Depth);
1409 // If the scale is 0, it takes nothing to add this.
1413 // If we already have a scale of this value, we can add to it, otherwise, we
1414 // need an available scale field.
1415 if (AddrMode.Scale != 0 && AddrMode.ScaledReg != ScaleReg)
1418 ExtAddrMode TestAddrMode = AddrMode;
1420 // Add scale to turn X*4+X*3 -> X*7. This could also do things like
1421 // [A+B + A*7] -> [B+A*8].
1422 TestAddrMode.Scale += Scale;
1423 TestAddrMode.ScaledReg = ScaleReg;
1425 // If the new address isn't legal, bail out.
1426 if (!TLI.isLegalAddressingMode(TestAddrMode, AccessTy))
1429 // It was legal, so commit it.
1430 AddrMode = TestAddrMode;
1432 // Okay, we decided that we can add ScaleReg+Scale to AddrMode. Check now
1433 // to see if ScaleReg is actually X+C. If so, we can turn this into adding
1434 // X*Scale + C*Scale to addr mode.
1435 ConstantInt *CI = nullptr; Value *AddLHS = nullptr;
1436 if (isa<Instruction>(ScaleReg) && // not a constant expr.
1437 match(ScaleReg, m_Add(m_Value(AddLHS), m_ConstantInt(CI)))) {
1438 TestAddrMode.ScaledReg = AddLHS;
1439 TestAddrMode.BaseOffs += CI->getSExtValue()*TestAddrMode.Scale;
1441 // If this addressing mode is legal, commit it and remember that we folded
1442 // this instruction.
1443 if (TLI.isLegalAddressingMode(TestAddrMode, AccessTy)) {
1444 AddrModeInsts.push_back(cast<Instruction>(ScaleReg));
1445 AddrMode = TestAddrMode;
1450 // Otherwise, not (x+c)*scale, just return what we have.
1454 /// MightBeFoldableInst - This is a little filter, which returns true if an
1455 /// addressing computation involving I might be folded into a load/store
1456 /// accessing it. This doesn't need to be perfect, but needs to accept at least
1457 /// the set of instructions that MatchOperationAddr can.
1458 static bool MightBeFoldableInst(Instruction *I) {
1459 switch (I->getOpcode()) {
1460 case Instruction::BitCast:
1461 // Don't touch identity bitcasts.
1462 if (I->getType() == I->getOperand(0)->getType())
1464 return I->getType()->isPointerTy() || I->getType()->isIntegerTy();
1465 case Instruction::PtrToInt:
1466 // PtrToInt is always a noop, as we know that the int type is pointer sized.
1468 case Instruction::IntToPtr:
1469 // We know the input is intptr_t, so this is foldable.
1471 case Instruction::Add:
1473 case Instruction::Mul:
1474 case Instruction::Shl:
1475 // Can only handle X*C and X << C.
1476 return isa<ConstantInt>(I->getOperand(1));
1477 case Instruction::GetElementPtr:
1484 /// \brief Hepler class to perform type promotion.
1485 class TypePromotionHelper {
1486 /// \brief Utility function to check whether or not a sign extension of
1487 /// \p Inst with \p ConsideredSExtType can be moved through \p Inst by either
1488 /// using the operands of \p Inst or promoting \p Inst.
1489 /// In other words, check if:
1490 /// sext (Ty Inst opnd1 opnd2 ... opndN) to ConsideredSExtType.
1491 /// #1 Promotion applies:
1492 /// ConsideredSExtType Inst (sext opnd1 to ConsideredSExtType, ...).
1493 /// #2 Operand reuses:
1494 /// sext opnd1 to ConsideredSExtType.
1495 /// \p PromotedInsts maps the instructions to their type before promotion.
1496 static bool canGetThrough(const Instruction *Inst, Type *ConsideredSExtType,
1497 const InstrToOrigTy &PromotedInsts);
1499 /// \brief Utility function to determine if \p OpIdx should be promoted when
1500 /// promoting \p Inst.
1501 static bool shouldSExtOperand(const Instruction *Inst, int OpIdx) {
1502 if (isa<SelectInst>(Inst) && OpIdx == 0)
1507 /// \brief Utility function to promote the operand of \p SExt when this
1508 /// operand is a promotable trunc or sext.
1509 /// \p PromotedInsts maps the instructions to their type before promotion.
1510 /// \p CreatedInsts[out] contains how many non-free instructions have been
1511 /// created to promote the operand of SExt.
1512 /// Should never be called directly.
1513 /// \return The promoted value which is used instead of SExt.
1514 static Value *promoteOperandForTruncAndSExt(Instruction *SExt,
1515 TypePromotionTransaction &TPT,
1516 InstrToOrigTy &PromotedInsts,
1517 unsigned &CreatedInsts);
1519 /// \brief Utility function to promote the operand of \p SExt when this
1520 /// operand is promotable and is not a supported trunc or sext.
1521 /// \p PromotedInsts maps the instructions to their type before promotion.
1522 /// \p CreatedInsts[out] contains how many non-free instructions have been
1523 /// created to promote the operand of SExt.
1524 /// Should never be called directly.
1525 /// \return The promoted value which is used instead of SExt.
1526 static Value *promoteOperandForOther(Instruction *SExt,
1527 TypePromotionTransaction &TPT,
1528 InstrToOrigTy &PromotedInsts,
1529 unsigned &CreatedInsts);
1532 /// Type for the utility function that promotes the operand of SExt.
1533 typedef Value *(*Action)(Instruction *SExt, TypePromotionTransaction &TPT,
1534 InstrToOrigTy &PromotedInsts,
1535 unsigned &CreatedInsts);
1536 /// \brief Given a sign extend instruction \p SExt, return the approriate
1537 /// action to promote the operand of \p SExt instead of using SExt.
1538 /// \return NULL if no promotable action is possible with the current
1540 /// \p InsertedTruncs keeps track of all the truncate instructions inserted by
1541 /// the others CodeGenPrepare optimizations. This information is important
1542 /// because we do not want to promote these instructions as CodeGenPrepare
1543 /// will reinsert them later. Thus creating an infinite loop: create/remove.
1544 /// \p PromotedInsts maps the instructions to their type before promotion.
1545 static Action getAction(Instruction *SExt, const SetOfInstrs &InsertedTruncs,
1546 const TargetLowering &TLI,
1547 const InstrToOrigTy &PromotedInsts);
1550 bool TypePromotionHelper::canGetThrough(const Instruction *Inst,
1551 Type *ConsideredSExtType,
1552 const InstrToOrigTy &PromotedInsts) {
1553 // We can always get through sext.
1554 if (isa<SExtInst>(Inst))
1557 // We can get through binary operator, if it is legal. In other words, the
1558 // binary operator must have a nuw or nsw flag.
1559 const BinaryOperator *BinOp = dyn_cast<BinaryOperator>(Inst);
1560 if (BinOp && isa<OverflowingBinaryOperator>(BinOp) &&
1561 (BinOp->hasNoUnsignedWrap() || BinOp->hasNoSignedWrap()))
1564 // Check if we can do the following simplification.
1565 // sext(trunc(sext)) --> sext
1566 if (!isa<TruncInst>(Inst))
1569 Value *OpndVal = Inst->getOperand(0);
1570 // Check if we can use this operand in the sext.
1571 // If the type is larger than the result type of the sign extension,
1573 if (OpndVal->getType()->getIntegerBitWidth() >
1574 ConsideredSExtType->getIntegerBitWidth())
1577 // If the operand of the truncate is not an instruction, we will not have
1578 // any information on the dropped bits.
1579 // (Actually we could for constant but it is not worth the extra logic).
1580 Instruction *Opnd = dyn_cast<Instruction>(OpndVal);
1584 // Check if the source of the type is narrow enough.
1585 // I.e., check that trunc just drops sign extended bits.
1586 // #1 get the type of the operand.
1587 const Type *OpndType;
1588 InstrToOrigTy::const_iterator It = PromotedInsts.find(Opnd);
1589 if (It != PromotedInsts.end())
1590 OpndType = It->second;
1591 else if (isa<SExtInst>(Opnd))
1592 OpndType = cast<Instruction>(Opnd)->getOperand(0)->getType();
1596 // #2 check that the truncate just drop sign extended bits.
1597 if (Inst->getType()->getIntegerBitWidth() >= OpndType->getIntegerBitWidth())
1603 TypePromotionHelper::Action TypePromotionHelper::getAction(
1604 Instruction *SExt, const SetOfInstrs &InsertedTruncs,
1605 const TargetLowering &TLI, const InstrToOrigTy &PromotedInsts) {
1606 Instruction *SExtOpnd = dyn_cast<Instruction>(SExt->getOperand(0));
1607 Type *SExtTy = SExt->getType();
1608 // If the operand of the sign extension is not an instruction, we cannot
1610 // If it, check we can get through.
1611 if (!SExtOpnd || !canGetThrough(SExtOpnd, SExtTy, PromotedInsts))
1614 // Do not promote if the operand has been added by codegenprepare.
1615 // Otherwise, it means we are undoing an optimization that is likely to be
1616 // redone, thus causing potential infinite loop.
1617 if (isa<TruncInst>(SExtOpnd) && InsertedTruncs.count(SExtOpnd))
1620 // SExt or Trunc instructions.
1621 // Return the related handler.
1622 if (isa<SExtInst>(SExtOpnd) || isa<TruncInst>(SExtOpnd))
1623 return promoteOperandForTruncAndSExt;
1625 // Regular instruction.
1626 // Abort early if we will have to insert non-free instructions.
1627 if (!SExtOpnd->hasOneUse() &&
1628 !TLI.isTruncateFree(SExtTy, SExtOpnd->getType()))
1630 return promoteOperandForOther;
1633 Value *TypePromotionHelper::promoteOperandForTruncAndSExt(
1634 llvm::Instruction *SExt, TypePromotionTransaction &TPT,
1635 InstrToOrigTy &PromotedInsts, unsigned &CreatedInsts) {
1636 // By construction, the operand of SExt is an instruction. Otherwise we cannot
1637 // get through it and this method should not be called.
1638 Instruction *SExtOpnd = cast<Instruction>(SExt->getOperand(0));
1639 // Replace sext(trunc(opnd)) or sext(sext(opnd))
1641 TPT.setOperand(SExt, 0, SExtOpnd->getOperand(0));
1644 // Remove dead code.
1645 if (SExtOpnd->use_empty())
1646 TPT.eraseInstruction(SExtOpnd);
1648 // Check if the sext is still needed.
1649 if (SExt->getType() != SExt->getOperand(0)->getType())
1652 // At this point we have: sext ty opnd to ty.
1653 // Reassign the uses of SExt to the opnd and remove SExt.
1654 Value *NextVal = SExt->getOperand(0);
1655 TPT.eraseInstruction(SExt, NextVal);
1660 TypePromotionHelper::promoteOperandForOther(Instruction *SExt,
1661 TypePromotionTransaction &TPT,
1662 InstrToOrigTy &PromotedInsts,
1663 unsigned &CreatedInsts) {
1664 // By construction, the operand of SExt is an instruction. Otherwise we cannot
1665 // get through it and this method should not be called.
1666 Instruction *SExtOpnd = cast<Instruction>(SExt->getOperand(0));
1668 if (!SExtOpnd->hasOneUse()) {
1669 // SExtOpnd will be promoted.
1670 // All its uses, but SExt, will need to use a truncated value of the
1671 // promoted version.
1672 // Create the truncate now.
1673 Instruction *Trunc = TPT.createTrunc(SExt, SExtOpnd->getType());
1674 Trunc->removeFromParent();
1675 // Insert it just after the definition.
1676 Trunc->insertAfter(SExtOpnd);
1678 TPT.replaceAllUsesWith(SExtOpnd, Trunc);
1679 // Restore the operand of SExt (which has been replace by the previous call
1680 // to replaceAllUsesWith) to avoid creating a cycle trunc <-> sext.
1681 TPT.setOperand(SExt, 0, SExtOpnd);
1684 // Get through the Instruction:
1685 // 1. Update its type.
1686 // 2. Replace the uses of SExt by Inst.
1687 // 3. Sign extend each operand that needs to be sign extended.
1689 // Remember the original type of the instruction before promotion.
1690 // This is useful to know that the high bits are sign extended bits.
1691 PromotedInsts.insert(
1692 std::pair<Instruction *, Type *>(SExtOpnd, SExtOpnd->getType()));
1694 TPT.mutateType(SExtOpnd, SExt->getType());
1696 TPT.replaceAllUsesWith(SExt, SExtOpnd);
1698 Instruction *SExtForOpnd = SExt;
1700 DEBUG(dbgs() << "Propagate SExt to operands\n");
1701 for (int OpIdx = 0, EndOpIdx = SExtOpnd->getNumOperands(); OpIdx != EndOpIdx;
1703 DEBUG(dbgs() << "Operand:\n" << *(SExtOpnd->getOperand(OpIdx)) << '\n');
1704 if (SExtOpnd->getOperand(OpIdx)->getType() == SExt->getType() ||
1705 !shouldSExtOperand(SExtOpnd, OpIdx)) {
1706 DEBUG(dbgs() << "No need to propagate\n");
1709 // Check if we can statically sign extend the operand.
1710 Value *Opnd = SExtOpnd->getOperand(OpIdx);
1711 if (const ConstantInt *Cst = dyn_cast<ConstantInt>(Opnd)) {
1712 DEBUG(dbgs() << "Statically sign extend\n");
1715 ConstantInt::getSigned(SExt->getType(), Cst->getSExtValue()));
1718 // UndefValue are typed, so we have to statically sign extend them.
1719 if (isa<UndefValue>(Opnd)) {
1720 DEBUG(dbgs() << "Statically sign extend\n");
1721 TPT.setOperand(SExtOpnd, OpIdx, UndefValue::get(SExt->getType()));
1725 // Otherwise we have to explicity sign extend the operand.
1726 // Check if SExt was reused to sign extend an operand.
1728 // If yes, create a new one.
1729 DEBUG(dbgs() << "More operands to sext\n");
1730 SExtForOpnd = TPT.createSExt(SExt, Opnd, SExt->getType());
1734 TPT.setOperand(SExtForOpnd, 0, Opnd);
1736 // Move the sign extension before the insertion point.
1737 TPT.moveBefore(SExtForOpnd, SExtOpnd);
1738 TPT.setOperand(SExtOpnd, OpIdx, SExtForOpnd);
1739 // If more sext are required, new instructions will have to be created.
1740 SExtForOpnd = nullptr;
1742 if (SExtForOpnd == SExt) {
1743 DEBUG(dbgs() << "Sign extension is useless now\n");
1744 TPT.eraseInstruction(SExt);
1749 /// IsPromotionProfitable - Check whether or not promoting an instruction
1750 /// to a wider type was profitable.
1751 /// \p MatchedSize gives the number of instructions that have been matched
1752 /// in the addressing mode after the promotion was applied.
1753 /// \p SizeWithPromotion gives the number of created instructions for
1754 /// the promotion plus the number of instructions that have been
1755 /// matched in the addressing mode before the promotion.
1756 /// \p PromotedOperand is the value that has been promoted.
1757 /// \return True if the promotion is profitable, false otherwise.
1759 AddressingModeMatcher::IsPromotionProfitable(unsigned MatchedSize,
1760 unsigned SizeWithPromotion,
1761 Value *PromotedOperand) const {
1762 // We folded less instructions than what we created to promote the operand.
1763 // This is not profitable.
1764 if (MatchedSize < SizeWithPromotion)
1766 if (MatchedSize > SizeWithPromotion)
1768 // The promotion is neutral but it may help folding the sign extension in
1769 // loads for instance.
1770 // Check that we did not create an illegal instruction.
1771 Instruction *PromotedInst = dyn_cast<Instruction>(PromotedOperand);
1774 int ISDOpcode = TLI.InstructionOpcodeToISD(PromotedInst->getOpcode());
1775 // If the ISDOpcode is undefined, it was undefined before the promotion.
1778 // Otherwise, check if the promoted instruction is legal or not.
1779 return TLI.isOperationLegalOrCustom(ISDOpcode,
1780 EVT::getEVT(PromotedInst->getType()));
1783 /// MatchOperationAddr - Given an instruction or constant expr, see if we can
1784 /// fold the operation into the addressing mode. If so, update the addressing
1785 /// mode and return true, otherwise return false without modifying AddrMode.
1786 /// If \p MovedAway is not NULL, it contains the information of whether or
1787 /// not AddrInst has to be folded into the addressing mode on success.
1788 /// If \p MovedAway == true, \p AddrInst will not be part of the addressing
1789 /// because it has been moved away.
1790 /// Thus AddrInst must not be added in the matched instructions.
1791 /// This state can happen when AddrInst is a sext, since it may be moved away.
1792 /// Therefore, AddrInst may not be valid when MovedAway is true and it must
1793 /// not be referenced anymore.
1794 bool AddressingModeMatcher::MatchOperationAddr(User *AddrInst, unsigned Opcode,
1797 // Avoid exponential behavior on extremely deep expression trees.
1798 if (Depth >= 5) return false;
1800 // By default, all matched instructions stay in place.
1805 case Instruction::PtrToInt:
1806 // PtrToInt is always a noop, as we know that the int type is pointer sized.
1807 return MatchAddr(AddrInst->getOperand(0), Depth);
1808 case Instruction::IntToPtr:
1809 // This inttoptr is a no-op if the integer type is pointer sized.
1810 if (TLI.getValueType(AddrInst->getOperand(0)->getType()) ==
1811 TLI.getPointerTy(AddrInst->getType()->getPointerAddressSpace()))
1812 return MatchAddr(AddrInst->getOperand(0), Depth);
1814 case Instruction::BitCast:
1815 // BitCast is always a noop, and we can handle it as long as it is
1816 // int->int or pointer->pointer (we don't want int<->fp or something).
1817 if ((AddrInst->getOperand(0)->getType()->isPointerTy() ||
1818 AddrInst->getOperand(0)->getType()->isIntegerTy()) &&
1819 // Don't touch identity bitcasts. These were probably put here by LSR,
1820 // and we don't want to mess around with them. Assume it knows what it
1822 AddrInst->getOperand(0)->getType() != AddrInst->getType())
1823 return MatchAddr(AddrInst->getOperand(0), Depth);
1825 case Instruction::Add: {
1826 // Check to see if we can merge in the RHS then the LHS. If so, we win.
1827 ExtAddrMode BackupAddrMode = AddrMode;
1828 unsigned OldSize = AddrModeInsts.size();
1829 // Start a transaction at this point.
1830 // The LHS may match but not the RHS.
1831 // Therefore, we need a higher level restoration point to undo partially
1832 // matched operation.
1833 TypePromotionTransaction::ConstRestorationPt LastKnownGood =
1834 TPT.getRestorationPoint();
1836 if (MatchAddr(AddrInst->getOperand(1), Depth+1) &&
1837 MatchAddr(AddrInst->getOperand(0), Depth+1))
1840 // Restore the old addr mode info.
1841 AddrMode = BackupAddrMode;
1842 AddrModeInsts.resize(OldSize);
1843 TPT.rollback(LastKnownGood);
1845 // Otherwise this was over-aggressive. Try merging in the LHS then the RHS.
1846 if (MatchAddr(AddrInst->getOperand(0), Depth+1) &&
1847 MatchAddr(AddrInst->getOperand(1), Depth+1))
1850 // Otherwise we definitely can't merge the ADD in.
1851 AddrMode = BackupAddrMode;
1852 AddrModeInsts.resize(OldSize);
1853 TPT.rollback(LastKnownGood);
1856 //case Instruction::Or:
1857 // TODO: We can handle "Or Val, Imm" iff this OR is equivalent to an ADD.
1859 case Instruction::Mul:
1860 case Instruction::Shl: {
1861 // Can only handle X*C and X << C.
1862 ConstantInt *RHS = dyn_cast<ConstantInt>(AddrInst->getOperand(1));
1863 if (!RHS) return false;
1864 int64_t Scale = RHS->getSExtValue();
1865 if (Opcode == Instruction::Shl)
1866 Scale = 1LL << Scale;
1868 return MatchScaledValue(AddrInst->getOperand(0), Scale, Depth);
1870 case Instruction::GetElementPtr: {
1871 // Scan the GEP. We check it if it contains constant offsets and at most
1872 // one variable offset.
1873 int VariableOperand = -1;
1874 unsigned VariableScale = 0;
1876 int64_t ConstantOffset = 0;
1877 const DataLayout *TD = TLI.getDataLayout();
1878 gep_type_iterator GTI = gep_type_begin(AddrInst);
1879 for (unsigned i = 1, e = AddrInst->getNumOperands(); i != e; ++i, ++GTI) {
1880 if (StructType *STy = dyn_cast<StructType>(*GTI)) {
1881 const StructLayout *SL = TD->getStructLayout(STy);
1883 cast<ConstantInt>(AddrInst->getOperand(i))->getZExtValue();
1884 ConstantOffset += SL->getElementOffset(Idx);
1886 uint64_t TypeSize = TD->getTypeAllocSize(GTI.getIndexedType());
1887 if (ConstantInt *CI = dyn_cast<ConstantInt>(AddrInst->getOperand(i))) {
1888 ConstantOffset += CI->getSExtValue()*TypeSize;
1889 } else if (TypeSize) { // Scales of zero don't do anything.
1890 // We only allow one variable index at the moment.
1891 if (VariableOperand != -1)
1894 // Remember the variable index.
1895 VariableOperand = i;
1896 VariableScale = TypeSize;
1901 // A common case is for the GEP to only do a constant offset. In this case,
1902 // just add it to the disp field and check validity.
1903 if (VariableOperand == -1) {
1904 AddrMode.BaseOffs += ConstantOffset;
1905 if (ConstantOffset == 0 || TLI.isLegalAddressingMode(AddrMode, AccessTy)){
1906 // Check to see if we can fold the base pointer in too.
1907 if (MatchAddr(AddrInst->getOperand(0), Depth+1))
1910 AddrMode.BaseOffs -= ConstantOffset;
1914 // Save the valid addressing mode in case we can't match.
1915 ExtAddrMode BackupAddrMode = AddrMode;
1916 unsigned OldSize = AddrModeInsts.size();
1918 // See if the scale and offset amount is valid for this target.
1919 AddrMode.BaseOffs += ConstantOffset;
1921 // Match the base operand of the GEP.
1922 if (!MatchAddr(AddrInst->getOperand(0), Depth+1)) {
1923 // If it couldn't be matched, just stuff the value in a register.
1924 if (AddrMode.HasBaseReg) {
1925 AddrMode = BackupAddrMode;
1926 AddrModeInsts.resize(OldSize);
1929 AddrMode.HasBaseReg = true;
1930 AddrMode.BaseReg = AddrInst->getOperand(0);
1933 // Match the remaining variable portion of the GEP.
1934 if (!MatchScaledValue(AddrInst->getOperand(VariableOperand), VariableScale,
1936 // If it couldn't be matched, try stuffing the base into a register
1937 // instead of matching it, and retrying the match of the scale.
1938 AddrMode = BackupAddrMode;
1939 AddrModeInsts.resize(OldSize);
1940 if (AddrMode.HasBaseReg)
1942 AddrMode.HasBaseReg = true;
1943 AddrMode.BaseReg = AddrInst->getOperand(0);
1944 AddrMode.BaseOffs += ConstantOffset;
1945 if (!MatchScaledValue(AddrInst->getOperand(VariableOperand),
1946 VariableScale, Depth)) {
1947 // If even that didn't work, bail.
1948 AddrMode = BackupAddrMode;
1949 AddrModeInsts.resize(OldSize);
1956 case Instruction::SExt: {
1957 // Try to move this sext out of the way of the addressing mode.
1958 Instruction *SExt = cast<Instruction>(AddrInst);
1959 // Ask for a method for doing so.
1960 TypePromotionHelper::Action TPH = TypePromotionHelper::getAction(
1961 SExt, InsertedTruncs, TLI, PromotedInsts);
1965 TypePromotionTransaction::ConstRestorationPt LastKnownGood =
1966 TPT.getRestorationPoint();
1967 unsigned CreatedInsts = 0;
1968 Value *PromotedOperand = TPH(SExt, TPT, PromotedInsts, CreatedInsts);
1969 // SExt has been moved away.
1970 // Thus either it will be rematched later in the recursive calls or it is
1971 // gone. Anyway, we must not fold it into the addressing mode at this point.
1975 // addr = gep base, idx
1977 // promotedOpnd = sext opnd <- no match here
1978 // op = promoted_add promotedOpnd, 1 <- match (later in recursive calls)
1979 // addr = gep base, op <- match
1983 assert(PromotedOperand &&
1984 "TypePromotionHelper should have filtered out those cases");
1986 ExtAddrMode BackupAddrMode = AddrMode;
1987 unsigned OldSize = AddrModeInsts.size();
1989 if (!MatchAddr(PromotedOperand, Depth) ||
1990 !IsPromotionProfitable(AddrModeInsts.size(), OldSize + CreatedInsts,
1992 AddrMode = BackupAddrMode;
1993 AddrModeInsts.resize(OldSize);
1994 DEBUG(dbgs() << "Sign extension does not pay off: rollback\n");
1995 TPT.rollback(LastKnownGood);
2004 /// MatchAddr - If we can, try to add the value of 'Addr' into the current
2005 /// addressing mode. If Addr can't be added to AddrMode this returns false and
2006 /// leaves AddrMode unmodified. This assumes that Addr is either a pointer type
2007 /// or intptr_t for the target.
2009 bool AddressingModeMatcher::MatchAddr(Value *Addr, unsigned Depth) {
2010 // Start a transaction at this point that we will rollback if the matching
2012 TypePromotionTransaction::ConstRestorationPt LastKnownGood =
2013 TPT.getRestorationPoint();
2014 if (ConstantInt *CI = dyn_cast<ConstantInt>(Addr)) {
2015 // Fold in immediates if legal for the target.
2016 AddrMode.BaseOffs += CI->getSExtValue();
2017 if (TLI.isLegalAddressingMode(AddrMode, AccessTy))
2019 AddrMode.BaseOffs -= CI->getSExtValue();
2020 } else if (GlobalValue *GV = dyn_cast<GlobalValue>(Addr)) {
2021 // If this is a global variable, try to fold it into the addressing mode.
2022 if (!AddrMode.BaseGV) {
2023 AddrMode.BaseGV = GV;
2024 if (TLI.isLegalAddressingMode(AddrMode, AccessTy))
2026 AddrMode.BaseGV = nullptr;
2028 } else if (Instruction *I = dyn_cast<Instruction>(Addr)) {
2029 ExtAddrMode BackupAddrMode = AddrMode;
2030 unsigned OldSize = AddrModeInsts.size();
2032 // Check to see if it is possible to fold this operation.
2033 bool MovedAway = false;
2034 if (MatchOperationAddr(I, I->getOpcode(), Depth, &MovedAway)) {
2035 // This instruction may have been move away. If so, there is nothing
2039 // Okay, it's possible to fold this. Check to see if it is actually
2040 // *profitable* to do so. We use a simple cost model to avoid increasing
2041 // register pressure too much.
2042 if (I->hasOneUse() ||
2043 IsProfitableToFoldIntoAddressingMode(I, BackupAddrMode, AddrMode)) {
2044 AddrModeInsts.push_back(I);
2048 // It isn't profitable to do this, roll back.
2049 //cerr << "NOT FOLDING: " << *I;
2050 AddrMode = BackupAddrMode;
2051 AddrModeInsts.resize(OldSize);
2052 TPT.rollback(LastKnownGood);
2054 } else if (ConstantExpr *CE = dyn_cast<ConstantExpr>(Addr)) {
2055 if (MatchOperationAddr(CE, CE->getOpcode(), Depth))
2057 TPT.rollback(LastKnownGood);
2058 } else if (isa<ConstantPointerNull>(Addr)) {
2059 // Null pointer gets folded without affecting the addressing mode.
2063 // Worse case, the target should support [reg] addressing modes. :)
2064 if (!AddrMode.HasBaseReg) {
2065 AddrMode.HasBaseReg = true;
2066 AddrMode.BaseReg = Addr;
2067 // Still check for legality in case the target supports [imm] but not [i+r].
2068 if (TLI.isLegalAddressingMode(AddrMode, AccessTy))
2070 AddrMode.HasBaseReg = false;
2071 AddrMode.BaseReg = nullptr;
2074 // If the base register is already taken, see if we can do [r+r].
2075 if (AddrMode.Scale == 0) {
2077 AddrMode.ScaledReg = Addr;
2078 if (TLI.isLegalAddressingMode(AddrMode, AccessTy))
2081 AddrMode.ScaledReg = nullptr;
2084 TPT.rollback(LastKnownGood);
2088 /// IsOperandAMemoryOperand - Check to see if all uses of OpVal by the specified
2089 /// inline asm call are due to memory operands. If so, return true, otherwise
2091 static bool IsOperandAMemoryOperand(CallInst *CI, InlineAsm *IA, Value *OpVal,
2092 const TargetLowering &TLI) {
2093 TargetLowering::AsmOperandInfoVector TargetConstraints = TLI.ParseConstraints(ImmutableCallSite(CI));
2094 for (unsigned i = 0, e = TargetConstraints.size(); i != e; ++i) {
2095 TargetLowering::AsmOperandInfo &OpInfo = TargetConstraints[i];
2097 // Compute the constraint code and ConstraintType to use.
2098 TLI.ComputeConstraintToUse(OpInfo, SDValue());
2100 // If this asm operand is our Value*, and if it isn't an indirect memory
2101 // operand, we can't fold it!
2102 if (OpInfo.CallOperandVal == OpVal &&
2103 (OpInfo.ConstraintType != TargetLowering::C_Memory ||
2104 !OpInfo.isIndirect))
2111 /// FindAllMemoryUses - Recursively walk all the uses of I until we find a
2112 /// memory use. If we find an obviously non-foldable instruction, return true.
2113 /// Add the ultimately found memory instructions to MemoryUses.
2114 static bool FindAllMemoryUses(Instruction *I,
2115 SmallVectorImpl<std::pair<Instruction*,unsigned> > &MemoryUses,
2116 SmallPtrSet<Instruction*, 16> &ConsideredInsts,
2117 const TargetLowering &TLI) {
2118 // If we already considered this instruction, we're done.
2119 if (!ConsideredInsts.insert(I))
2122 // If this is an obviously unfoldable instruction, bail out.
2123 if (!MightBeFoldableInst(I))
2126 // Loop over all the uses, recursively processing them.
2127 for (Use &U : I->uses()) {
2128 Instruction *UserI = cast<Instruction>(U.getUser());
2130 if (LoadInst *LI = dyn_cast<LoadInst>(UserI)) {
2131 MemoryUses.push_back(std::make_pair(LI, U.getOperandNo()));
2135 if (StoreInst *SI = dyn_cast<StoreInst>(UserI)) {
2136 unsigned opNo = U.getOperandNo();
2137 if (opNo == 0) return true; // Storing addr, not into addr.
2138 MemoryUses.push_back(std::make_pair(SI, opNo));
2142 if (CallInst *CI = dyn_cast<CallInst>(UserI)) {
2143 InlineAsm *IA = dyn_cast<InlineAsm>(CI->getCalledValue());
2144 if (!IA) return true;
2146 // If this is a memory operand, we're cool, otherwise bail out.
2147 if (!IsOperandAMemoryOperand(CI, IA, I, TLI))
2152 if (FindAllMemoryUses(UserI, MemoryUses, ConsideredInsts, TLI))
2159 /// ValueAlreadyLiveAtInst - Retrn true if Val is already known to be live at
2160 /// the use site that we're folding it into. If so, there is no cost to
2161 /// include it in the addressing mode. KnownLive1 and KnownLive2 are two values
2162 /// that we know are live at the instruction already.
2163 bool AddressingModeMatcher::ValueAlreadyLiveAtInst(Value *Val,Value *KnownLive1,
2164 Value *KnownLive2) {
2165 // If Val is either of the known-live values, we know it is live!
2166 if (Val == nullptr || Val == KnownLive1 || Val == KnownLive2)
2169 // All values other than instructions and arguments (e.g. constants) are live.
2170 if (!isa<Instruction>(Val) && !isa<Argument>(Val)) return true;
2172 // If Val is a constant sized alloca in the entry block, it is live, this is
2173 // true because it is just a reference to the stack/frame pointer, which is
2174 // live for the whole function.
2175 if (AllocaInst *AI = dyn_cast<AllocaInst>(Val))
2176 if (AI->isStaticAlloca())
2179 // Check to see if this value is already used in the memory instruction's
2180 // block. If so, it's already live into the block at the very least, so we
2181 // can reasonably fold it.
2182 return Val->isUsedInBasicBlock(MemoryInst->getParent());
2185 /// IsProfitableToFoldIntoAddressingMode - It is possible for the addressing
2186 /// mode of the machine to fold the specified instruction into a load or store
2187 /// that ultimately uses it. However, the specified instruction has multiple
2188 /// uses. Given this, it may actually increase register pressure to fold it
2189 /// into the load. For example, consider this code:
2193 /// use(Y) -> nonload/store
2197 /// In this case, Y has multiple uses, and can be folded into the load of Z
2198 /// (yielding load [X+2]). However, doing this will cause both "X" and "X+1" to
2199 /// be live at the use(Y) line. If we don't fold Y into load Z, we use one
2200 /// fewer register. Since Y can't be folded into "use(Y)" we don't increase the
2201 /// number of computations either.
2203 /// Note that this (like most of CodeGenPrepare) is just a rough heuristic. If
2204 /// X was live across 'load Z' for other reasons, we actually *would* want to
2205 /// fold the addressing mode in the Z case. This would make Y die earlier.
2206 bool AddressingModeMatcher::
2207 IsProfitableToFoldIntoAddressingMode(Instruction *I, ExtAddrMode &AMBefore,
2208 ExtAddrMode &AMAfter) {
2209 if (IgnoreProfitability) return true;
2211 // AMBefore is the addressing mode before this instruction was folded into it,
2212 // and AMAfter is the addressing mode after the instruction was folded. Get
2213 // the set of registers referenced by AMAfter and subtract out those
2214 // referenced by AMBefore: this is the set of values which folding in this
2215 // address extends the lifetime of.
2217 // Note that there are only two potential values being referenced here,
2218 // BaseReg and ScaleReg (global addresses are always available, as are any
2219 // folded immediates).
2220 Value *BaseReg = AMAfter.BaseReg, *ScaledReg = AMAfter.ScaledReg;
2222 // If the BaseReg or ScaledReg was referenced by the previous addrmode, their
2223 // lifetime wasn't extended by adding this instruction.
2224 if (ValueAlreadyLiveAtInst(BaseReg, AMBefore.BaseReg, AMBefore.ScaledReg))
2226 if (ValueAlreadyLiveAtInst(ScaledReg, AMBefore.BaseReg, AMBefore.ScaledReg))
2227 ScaledReg = nullptr;
2229 // If folding this instruction (and it's subexprs) didn't extend any live
2230 // ranges, we're ok with it.
2231 if (!BaseReg && !ScaledReg)
2234 // If all uses of this instruction are ultimately load/store/inlineasm's,
2235 // check to see if their addressing modes will include this instruction. If
2236 // so, we can fold it into all uses, so it doesn't matter if it has multiple
2238 SmallVector<std::pair<Instruction*,unsigned>, 16> MemoryUses;
2239 SmallPtrSet<Instruction*, 16> ConsideredInsts;
2240 if (FindAllMemoryUses(I, MemoryUses, ConsideredInsts, TLI))
2241 return false; // Has a non-memory, non-foldable use!
2243 // Now that we know that all uses of this instruction are part of a chain of
2244 // computation involving only operations that could theoretically be folded
2245 // into a memory use, loop over each of these uses and see if they could
2246 // *actually* fold the instruction.
2247 SmallVector<Instruction*, 32> MatchedAddrModeInsts;
2248 for (unsigned i = 0, e = MemoryUses.size(); i != e; ++i) {
2249 Instruction *User = MemoryUses[i].first;
2250 unsigned OpNo = MemoryUses[i].second;
2252 // Get the access type of this use. If the use isn't a pointer, we don't
2253 // know what it accesses.
2254 Value *Address = User->getOperand(OpNo);
2255 if (!Address->getType()->isPointerTy())
2257 Type *AddressAccessTy = Address->getType()->getPointerElementType();
2259 // Do a match against the root of this address, ignoring profitability. This
2260 // will tell us if the addressing mode for the memory operation will
2261 // *actually* cover the shared instruction.
2263 TypePromotionTransaction::ConstRestorationPt LastKnownGood =
2264 TPT.getRestorationPoint();
2265 AddressingModeMatcher Matcher(MatchedAddrModeInsts, TLI, AddressAccessTy,
2266 MemoryInst, Result, InsertedTruncs,
2267 PromotedInsts, TPT);
2268 Matcher.IgnoreProfitability = true;
2269 bool Success = Matcher.MatchAddr(Address, 0);
2270 (void)Success; assert(Success && "Couldn't select *anything*?");
2272 // The match was to check the profitability, the changes made are not
2273 // part of the original matcher. Therefore, they should be dropped
2274 // otherwise the original matcher will not present the right state.
2275 TPT.rollback(LastKnownGood);
2277 // If the match didn't cover I, then it won't be shared by it.
2278 if (std::find(MatchedAddrModeInsts.begin(), MatchedAddrModeInsts.end(),
2279 I) == MatchedAddrModeInsts.end())
2282 MatchedAddrModeInsts.clear();
2288 } // end anonymous namespace
2290 /// IsNonLocalValue - Return true if the specified values are defined in a
2291 /// different basic block than BB.
2292 static bool IsNonLocalValue(Value *V, BasicBlock *BB) {
2293 if (Instruction *I = dyn_cast<Instruction>(V))
2294 return I->getParent() != BB;
2298 /// OptimizeMemoryInst - Load and Store Instructions often have
2299 /// addressing modes that can do significant amounts of computation. As such,
2300 /// instruction selection will try to get the load or store to do as much
2301 /// computation as possible for the program. The problem is that isel can only
2302 /// see within a single block. As such, we sink as much legal addressing mode
2303 /// stuff into the block as possible.
2305 /// This method is used to optimize both load/store and inline asms with memory
2307 bool CodeGenPrepare::OptimizeMemoryInst(Instruction *MemoryInst, Value *Addr,
2311 // Try to collapse single-value PHI nodes. This is necessary to undo
2312 // unprofitable PRE transformations.
2313 SmallVector<Value*, 8> worklist;
2314 SmallPtrSet<Value*, 16> Visited;
2315 worklist.push_back(Addr);
2317 // Use a worklist to iteratively look through PHI nodes, and ensure that
2318 // the addressing mode obtained from the non-PHI roots of the graph
2320 Value *Consensus = nullptr;
2321 unsigned NumUsesConsensus = 0;
2322 bool IsNumUsesConsensusValid = false;
2323 SmallVector<Instruction*, 16> AddrModeInsts;
2324 ExtAddrMode AddrMode;
2325 TypePromotionTransaction TPT;
2326 TypePromotionTransaction::ConstRestorationPt LastKnownGood =
2327 TPT.getRestorationPoint();
2328 while (!worklist.empty()) {
2329 Value *V = worklist.back();
2330 worklist.pop_back();
2332 // Break use-def graph loops.
2333 if (!Visited.insert(V)) {
2334 Consensus = nullptr;
2338 // For a PHI node, push all of its incoming values.
2339 if (PHINode *P = dyn_cast<PHINode>(V)) {
2340 for (unsigned i = 0, e = P->getNumIncomingValues(); i != e; ++i)
2341 worklist.push_back(P->getIncomingValue(i));
2345 // For non-PHIs, determine the addressing mode being computed.
2346 SmallVector<Instruction*, 16> NewAddrModeInsts;
2347 ExtAddrMode NewAddrMode = AddressingModeMatcher::Match(
2348 V, AccessTy, MemoryInst, NewAddrModeInsts, *TLI, InsertedTruncsSet,
2349 PromotedInsts, TPT);
2351 // This check is broken into two cases with very similar code to avoid using
2352 // getNumUses() as much as possible. Some values have a lot of uses, so
2353 // calling getNumUses() unconditionally caused a significant compile-time
2357 AddrMode = NewAddrMode;
2358 AddrModeInsts = NewAddrModeInsts;
2360 } else if (NewAddrMode == AddrMode) {
2361 if (!IsNumUsesConsensusValid) {
2362 NumUsesConsensus = Consensus->getNumUses();
2363 IsNumUsesConsensusValid = true;
2366 // Ensure that the obtained addressing mode is equivalent to that obtained
2367 // for all other roots of the PHI traversal. Also, when choosing one
2368 // such root as representative, select the one with the most uses in order
2369 // to keep the cost modeling heuristics in AddressingModeMatcher
2371 unsigned NumUses = V->getNumUses();
2372 if (NumUses > NumUsesConsensus) {
2374 NumUsesConsensus = NumUses;
2375 AddrModeInsts = NewAddrModeInsts;
2380 Consensus = nullptr;
2384 // If the addressing mode couldn't be determined, or if multiple different
2385 // ones were determined, bail out now.
2387 TPT.rollback(LastKnownGood);
2392 // Check to see if any of the instructions supersumed by this addr mode are
2393 // non-local to I's BB.
2394 bool AnyNonLocal = false;
2395 for (unsigned i = 0, e = AddrModeInsts.size(); i != e; ++i) {
2396 if (IsNonLocalValue(AddrModeInsts[i], MemoryInst->getParent())) {
2402 // If all the instructions matched are already in this BB, don't do anything.
2404 DEBUG(dbgs() << "CGP: Found local addrmode: " << AddrMode << "\n");
2408 // Insert this computation right after this user. Since our caller is
2409 // scanning from the top of the BB to the bottom, reuse of the expr are
2410 // guaranteed to happen later.
2411 IRBuilder<> Builder(MemoryInst);
2413 // Now that we determined the addressing expression we want to use and know
2414 // that we have to sink it into this block. Check to see if we have already
2415 // done this for some other load/store instr in this block. If so, reuse the
2417 Value *&SunkAddr = SunkAddrs[Addr];
2419 DEBUG(dbgs() << "CGP: Reusing nonlocal addrmode: " << AddrMode << " for "
2421 if (SunkAddr->getType() != Addr->getType())
2422 SunkAddr = Builder.CreateBitCast(SunkAddr, Addr->getType());
2423 } else if (AddrSinkUsingGEPs || (!AddrSinkUsingGEPs.getNumOccurrences() &&
2424 TM && TM->getSubtarget<TargetSubtargetInfo>().useAA())) {
2425 // By default, we use the GEP-based method when AA is used later. This
2426 // prevents new inttoptr/ptrtoint pairs from degrading AA capabilities.
2427 DEBUG(dbgs() << "CGP: SINKING nonlocal addrmode: " << AddrMode << " for "
2429 Type *IntPtrTy = TLI->getDataLayout()->getIntPtrType(Addr->getType());
2430 Value *ResultPtr = nullptr, *ResultIndex = nullptr;
2432 // First, find the pointer.
2433 if (AddrMode.BaseReg && AddrMode.BaseReg->getType()->isPointerTy()) {
2434 ResultPtr = AddrMode.BaseReg;
2435 AddrMode.BaseReg = nullptr;
2438 if (AddrMode.Scale && AddrMode.ScaledReg->getType()->isPointerTy()) {
2439 // We can't add more than one pointer together, nor can we scale a
2440 // pointer (both of which seem meaningless).
2441 if (ResultPtr || AddrMode.Scale != 1)
2444 ResultPtr = AddrMode.ScaledReg;
2448 if (AddrMode.BaseGV) {
2452 ResultPtr = AddrMode.BaseGV;
2455 // If the real base value actually came from an inttoptr, then the matcher
2456 // will look through it and provide only the integer value. In that case,
2458 if (!ResultPtr && AddrMode.BaseReg) {
2460 Builder.CreateIntToPtr(AddrMode.BaseReg, Addr->getType(), "sunkaddr");
2461 AddrMode.BaseReg = nullptr;
2462 } else if (!ResultPtr && AddrMode.Scale == 1) {
2464 Builder.CreateIntToPtr(AddrMode.ScaledReg, Addr->getType(), "sunkaddr");
2469 !AddrMode.BaseReg && !AddrMode.Scale && !AddrMode.BaseOffs) {
2470 SunkAddr = Constant::getNullValue(Addr->getType());
2471 } else if (!ResultPtr) {
2475 Builder.getInt8PtrTy(Addr->getType()->getPointerAddressSpace());
2477 // Start with the base register. Do this first so that subsequent address
2478 // matching finds it last, which will prevent it from trying to match it
2479 // as the scaled value in case it happens to be a mul. That would be
2480 // problematic if we've sunk a different mul for the scale, because then
2481 // we'd end up sinking both muls.
2482 if (AddrMode.BaseReg) {
2483 Value *V = AddrMode.BaseReg;
2484 if (V->getType() != IntPtrTy)
2485 V = Builder.CreateIntCast(V, IntPtrTy, /*isSigned=*/true, "sunkaddr");
2490 // Add the scale value.
2491 if (AddrMode.Scale) {
2492 Value *V = AddrMode.ScaledReg;
2493 if (V->getType() == IntPtrTy) {
2495 } else if (cast<IntegerType>(IntPtrTy)->getBitWidth() <
2496 cast<IntegerType>(V->getType())->getBitWidth()) {
2497 V = Builder.CreateTrunc(V, IntPtrTy, "sunkaddr");
2499 // It is only safe to sign extend the BaseReg if we know that the math
2500 // required to create it did not overflow before we extend it. Since
2501 // the original IR value was tossed in favor of a constant back when
2502 // the AddrMode was created we need to bail out gracefully if widths
2503 // do not match instead of extending it.
2504 Instruction *I = dyn_cast_or_null<Instruction>(ResultIndex);
2505 if (I && (ResultIndex != AddrMode.BaseReg))
2506 I->eraseFromParent();
2510 if (AddrMode.Scale != 1)
2511 V = Builder.CreateMul(V, ConstantInt::get(IntPtrTy, AddrMode.Scale),
2514 ResultIndex = Builder.CreateAdd(ResultIndex, V, "sunkaddr");
2519 // Add in the Base Offset if present.
2520 if (AddrMode.BaseOffs) {
2521 Value *V = ConstantInt::get(IntPtrTy, AddrMode.BaseOffs);
2523 // We need to add this separately from the scale above to help with
2524 // SDAG consecutive load/store merging.
2525 if (ResultPtr->getType() != I8PtrTy)
2526 ResultPtr = Builder.CreateBitCast(ResultPtr, I8PtrTy);
2527 ResultPtr = Builder.CreateGEP(ResultPtr, ResultIndex, "sunkaddr");
2534 SunkAddr = ResultPtr;
2536 if (ResultPtr->getType() != I8PtrTy)
2537 ResultPtr = Builder.CreateBitCast(ResultPtr, I8PtrTy);
2538 SunkAddr = Builder.CreateGEP(ResultPtr, ResultIndex, "sunkaddr");
2541 if (SunkAddr->getType() != Addr->getType())
2542 SunkAddr = Builder.CreateBitCast(SunkAddr, Addr->getType());
2545 DEBUG(dbgs() << "CGP: SINKING nonlocal addrmode: " << AddrMode << " for "
2547 Type *IntPtrTy = TLI->getDataLayout()->getIntPtrType(Addr->getType());
2548 Value *Result = nullptr;
2550 // Start with the base register. Do this first so that subsequent address
2551 // matching finds it last, which will prevent it from trying to match it
2552 // as the scaled value in case it happens to be a mul. That would be
2553 // problematic if we've sunk a different mul for the scale, because then
2554 // we'd end up sinking both muls.
2555 if (AddrMode.BaseReg) {
2556 Value *V = AddrMode.BaseReg;
2557 if (V->getType()->isPointerTy())
2558 V = Builder.CreatePtrToInt(V, IntPtrTy, "sunkaddr");
2559 if (V->getType() != IntPtrTy)
2560 V = Builder.CreateIntCast(V, IntPtrTy, /*isSigned=*/true, "sunkaddr");
2564 // Add the scale value.
2565 if (AddrMode.Scale) {
2566 Value *V = AddrMode.ScaledReg;
2567 if (V->getType() == IntPtrTy) {
2569 } else if (V->getType()->isPointerTy()) {
2570 V = Builder.CreatePtrToInt(V, IntPtrTy, "sunkaddr");
2571 } else if (cast<IntegerType>(IntPtrTy)->getBitWidth() <
2572 cast<IntegerType>(V->getType())->getBitWidth()) {
2573 V = Builder.CreateTrunc(V, IntPtrTy, "sunkaddr");
2575 // It is only safe to sign extend the BaseReg if we know that the math
2576 // required to create it did not overflow before we extend it. Since
2577 // the original IR value was tossed in favor of a constant back when
2578 // the AddrMode was created we need to bail out gracefully if widths
2579 // do not match instead of extending it.
2580 Instruction *I = dyn_cast<Instruction>(Result);
2581 if (I && (Result != AddrMode.BaseReg))
2582 I->eraseFromParent();
2585 if (AddrMode.Scale != 1)
2586 V = Builder.CreateMul(V, ConstantInt::get(IntPtrTy, AddrMode.Scale),
2589 Result = Builder.CreateAdd(Result, V, "sunkaddr");
2594 // Add in the BaseGV if present.
2595 if (AddrMode.BaseGV) {
2596 Value *V = Builder.CreatePtrToInt(AddrMode.BaseGV, IntPtrTy, "sunkaddr");
2598 Result = Builder.CreateAdd(Result, V, "sunkaddr");
2603 // Add in the Base Offset if present.
2604 if (AddrMode.BaseOffs) {
2605 Value *V = ConstantInt::get(IntPtrTy, AddrMode.BaseOffs);
2607 Result = Builder.CreateAdd(Result, V, "sunkaddr");
2613 SunkAddr = Constant::getNullValue(Addr->getType());
2615 SunkAddr = Builder.CreateIntToPtr(Result, Addr->getType(), "sunkaddr");
2618 MemoryInst->replaceUsesOfWith(Repl, SunkAddr);
2620 // If we have no uses, recursively delete the value and all dead instructions
2622 if (Repl->use_empty()) {
2623 // This can cause recursive deletion, which can invalidate our iterator.
2624 // Use a WeakVH to hold onto it in case this happens.
2625 WeakVH IterHandle(CurInstIterator);
2626 BasicBlock *BB = CurInstIterator->getParent();
2628 RecursivelyDeleteTriviallyDeadInstructions(Repl, TLInfo);
2630 if (IterHandle != CurInstIterator) {
2631 // If the iterator instruction was recursively deleted, start over at the
2632 // start of the block.
2633 CurInstIterator = BB->begin();
2641 /// OptimizeInlineAsmInst - If there are any memory operands, use
2642 /// OptimizeMemoryInst to sink their address computing into the block when
2643 /// possible / profitable.
2644 bool CodeGenPrepare::OptimizeInlineAsmInst(CallInst *CS) {
2645 bool MadeChange = false;
2647 TargetLowering::AsmOperandInfoVector
2648 TargetConstraints = TLI->ParseConstraints(CS);
2650 for (unsigned i = 0, e = TargetConstraints.size(); i != e; ++i) {
2651 TargetLowering::AsmOperandInfo &OpInfo = TargetConstraints[i];
2653 // Compute the constraint code and ConstraintType to use.
2654 TLI->ComputeConstraintToUse(OpInfo, SDValue());
2656 if (OpInfo.ConstraintType == TargetLowering::C_Memory &&
2657 OpInfo.isIndirect) {
2658 Value *OpVal = CS->getArgOperand(ArgNo++);
2659 MadeChange |= OptimizeMemoryInst(CS, OpVal, OpVal->getType());
2660 } else if (OpInfo.Type == InlineAsm::isInput)
2667 /// MoveExtToFormExtLoad - Move a zext or sext fed by a load into the same
2668 /// basic block as the load, unless conditions are unfavorable. This allows
2669 /// SelectionDAG to fold the extend into the load.
2671 bool CodeGenPrepare::MoveExtToFormExtLoad(Instruction *I) {
2672 // Look for a load being extended.
2673 LoadInst *LI = dyn_cast<LoadInst>(I->getOperand(0));
2674 if (!LI) return false;
2676 // If they're already in the same block, there's nothing to do.
2677 if (LI->getParent() == I->getParent())
2680 // If the load has other users and the truncate is not free, this probably
2681 // isn't worthwhile.
2682 if (!LI->hasOneUse() &&
2683 TLI && (TLI->isTypeLegal(TLI->getValueType(LI->getType())) ||
2684 !TLI->isTypeLegal(TLI->getValueType(I->getType()))) &&
2685 !TLI->isTruncateFree(I->getType(), LI->getType()))
2688 // Check whether the target supports casts folded into loads.
2690 if (isa<ZExtInst>(I))
2691 LType = ISD::ZEXTLOAD;
2693 assert(isa<SExtInst>(I) && "Unexpected ext type!");
2694 LType = ISD::SEXTLOAD;
2696 if (TLI && !TLI->isLoadExtLegal(LType, TLI->getValueType(LI->getType())))
2699 // Move the extend into the same block as the load, so that SelectionDAG
2701 I->removeFromParent();
2707 bool CodeGenPrepare::OptimizeExtUses(Instruction *I) {
2708 BasicBlock *DefBB = I->getParent();
2710 // If the result of a {s|z}ext and its source are both live out, rewrite all
2711 // other uses of the source with result of extension.
2712 Value *Src = I->getOperand(0);
2713 if (Src->hasOneUse())
2716 // Only do this xform if truncating is free.
2717 if (TLI && !TLI->isTruncateFree(I->getType(), Src->getType()))
2720 // Only safe to perform the optimization if the source is also defined in
2722 if (!isa<Instruction>(Src) || DefBB != cast<Instruction>(Src)->getParent())
2725 bool DefIsLiveOut = false;
2726 for (User *U : I->users()) {
2727 Instruction *UI = cast<Instruction>(U);
2729 // Figure out which BB this ext is used in.
2730 BasicBlock *UserBB = UI->getParent();
2731 if (UserBB == DefBB) continue;
2732 DefIsLiveOut = true;
2738 // Make sure none of the uses are PHI nodes.
2739 for (User *U : Src->users()) {
2740 Instruction *UI = cast<Instruction>(U);
2741 BasicBlock *UserBB = UI->getParent();
2742 if (UserBB == DefBB) continue;
2743 // Be conservative. We don't want this xform to end up introducing
2744 // reloads just before load / store instructions.
2745 if (isa<PHINode>(UI) || isa<LoadInst>(UI) || isa<StoreInst>(UI))
2749 // InsertedTruncs - Only insert one trunc in each block once.
2750 DenseMap<BasicBlock*, Instruction*> InsertedTruncs;
2752 bool MadeChange = false;
2753 for (Use &U : Src->uses()) {
2754 Instruction *User = cast<Instruction>(U.getUser());
2756 // Figure out which BB this ext is used in.
2757 BasicBlock *UserBB = User->getParent();
2758 if (UserBB == DefBB) continue;
2760 // Both src and def are live in this block. Rewrite the use.
2761 Instruction *&InsertedTrunc = InsertedTruncs[UserBB];
2763 if (!InsertedTrunc) {
2764 BasicBlock::iterator InsertPt = UserBB->getFirstInsertionPt();
2765 InsertedTrunc = new TruncInst(I, Src->getType(), "", InsertPt);
2766 InsertedTruncsSet.insert(InsertedTrunc);
2769 // Replace a use of the {s|z}ext source with a use of the result.
2778 /// isFormingBranchFromSelectProfitable - Returns true if a SelectInst should be
2779 /// turned into an explicit branch.
2780 static bool isFormingBranchFromSelectProfitable(SelectInst *SI) {
2781 // FIXME: This should use the same heuristics as IfConversion to determine
2782 // whether a select is better represented as a branch. This requires that
2783 // branch probability metadata is preserved for the select, which is not the
2786 CmpInst *Cmp = dyn_cast<CmpInst>(SI->getCondition());
2788 // If the branch is predicted right, an out of order CPU can avoid blocking on
2789 // the compare. Emit cmovs on compares with a memory operand as branches to
2790 // avoid stalls on the load from memory. If the compare has more than one use
2791 // there's probably another cmov or setcc around so it's not worth emitting a
2796 Value *CmpOp0 = Cmp->getOperand(0);
2797 Value *CmpOp1 = Cmp->getOperand(1);
2799 // We check that the memory operand has one use to avoid uses of the loaded
2800 // value directly after the compare, making branches unprofitable.
2801 return Cmp->hasOneUse() &&
2802 ((isa<LoadInst>(CmpOp0) && CmpOp0->hasOneUse()) ||
2803 (isa<LoadInst>(CmpOp1) && CmpOp1->hasOneUse()));
2807 /// If we have a SelectInst that will likely profit from branch prediction,
2808 /// turn it into a branch.
2809 bool CodeGenPrepare::OptimizeSelectInst(SelectInst *SI) {
2810 bool VectorCond = !SI->getCondition()->getType()->isIntegerTy(1);
2812 // Can we convert the 'select' to CF ?
2813 if (DisableSelectToBranch || OptSize || !TLI || VectorCond)
2816 TargetLowering::SelectSupportKind SelectKind;
2818 SelectKind = TargetLowering::VectorMaskSelect;
2819 else if (SI->getType()->isVectorTy())
2820 SelectKind = TargetLowering::ScalarCondVectorVal;
2822 SelectKind = TargetLowering::ScalarValSelect;
2824 // Do we have efficient codegen support for this kind of 'selects' ?
2825 if (TLI->isSelectSupported(SelectKind)) {
2826 // We have efficient codegen support for the select instruction.
2827 // Check if it is profitable to keep this 'select'.
2828 if (!TLI->isPredictableSelectExpensive() ||
2829 !isFormingBranchFromSelectProfitable(SI))
2835 // First, we split the block containing the select into 2 blocks.
2836 BasicBlock *StartBlock = SI->getParent();
2837 BasicBlock::iterator SplitPt = ++(BasicBlock::iterator(SI));
2838 BasicBlock *NextBlock = StartBlock->splitBasicBlock(SplitPt, "select.end");
2840 // Create a new block serving as the landing pad for the branch.
2841 BasicBlock *SmallBlock = BasicBlock::Create(SI->getContext(), "select.mid",
2842 NextBlock->getParent(), NextBlock);
2844 // Move the unconditional branch from the block with the select in it into our
2845 // landing pad block.
2846 StartBlock->getTerminator()->eraseFromParent();
2847 BranchInst::Create(NextBlock, SmallBlock);
2849 // Insert the real conditional branch based on the original condition.
2850 BranchInst::Create(NextBlock, SmallBlock, SI->getCondition(), SI);
2852 // The select itself is replaced with a PHI Node.
2853 PHINode *PN = PHINode::Create(SI->getType(), 2, "", NextBlock->begin());
2855 PN->addIncoming(SI->getTrueValue(), StartBlock);
2856 PN->addIncoming(SI->getFalseValue(), SmallBlock);
2857 SI->replaceAllUsesWith(PN);
2858 SI->eraseFromParent();
2860 // Instruct OptimizeBlock to skip to the next block.
2861 CurInstIterator = StartBlock->end();
2862 ++NumSelectsExpanded;
2866 static bool isBroadcastShuffle(ShuffleVectorInst *SVI) {
2867 SmallVector<int, 16> Mask(SVI->getShuffleMask());
2869 for (unsigned i = 0; i < Mask.size(); ++i) {
2870 if (SplatElem != -1 && Mask[i] != -1 && Mask[i] != SplatElem)
2872 SplatElem = Mask[i];
2878 /// Some targets have expensive vector shifts if the lanes aren't all the same
2879 /// (e.g. x86 only introduced "vpsllvd" and friends with AVX2). In these cases
2880 /// it's often worth sinking a shufflevector splat down to its use so that
2881 /// codegen can spot all lanes are identical.
2882 bool CodeGenPrepare::OptimizeShuffleVectorInst(ShuffleVectorInst *SVI) {
2883 BasicBlock *DefBB = SVI->getParent();
2885 // Only do this xform if variable vector shifts are particularly expensive.
2886 if (!TLI || !TLI->isVectorShiftByScalarCheap(SVI->getType()))
2889 // We only expect better codegen by sinking a shuffle if we can recognise a
2891 if (!isBroadcastShuffle(SVI))
2894 // InsertedShuffles - Only insert a shuffle in each block once.
2895 DenseMap<BasicBlock*, Instruction*> InsertedShuffles;
2897 bool MadeChange = false;
2898 for (User *U : SVI->users()) {
2899 Instruction *UI = cast<Instruction>(U);
2901 // Figure out which BB this ext is used in.
2902 BasicBlock *UserBB = UI->getParent();
2903 if (UserBB == DefBB) continue;
2905 // For now only apply this when the splat is used by a shift instruction.
2906 if (!UI->isShift()) continue;
2908 // Everything checks out, sink the shuffle if the user's block doesn't
2909 // already have a copy.
2910 Instruction *&InsertedShuffle = InsertedShuffles[UserBB];
2912 if (!InsertedShuffle) {
2913 BasicBlock::iterator InsertPt = UserBB->getFirstInsertionPt();
2914 InsertedShuffle = new ShuffleVectorInst(SVI->getOperand(0),
2916 SVI->getOperand(2), "", InsertPt);
2919 UI->replaceUsesOfWith(SVI, InsertedShuffle);
2923 // If we removed all uses, nuke the shuffle.
2924 if (SVI->use_empty()) {
2925 SVI->eraseFromParent();
2932 bool CodeGenPrepare::OptimizeInst(Instruction *I) {
2933 if (PHINode *P = dyn_cast<PHINode>(I)) {
2934 // It is possible for very late stage optimizations (such as SimplifyCFG)
2935 // to introduce PHI nodes too late to be cleaned up. If we detect such a
2936 // trivial PHI, go ahead and zap it here.
2937 if (Value *V = SimplifyInstruction(P, TLI ? TLI->getDataLayout() : nullptr,
2939 P->replaceAllUsesWith(V);
2940 P->eraseFromParent();
2947 if (CastInst *CI = dyn_cast<CastInst>(I)) {
2948 // If the source of the cast is a constant, then this should have
2949 // already been constant folded. The only reason NOT to constant fold
2950 // it is if something (e.g. LSR) was careful to place the constant
2951 // evaluation in a block other than then one that uses it (e.g. to hoist
2952 // the address of globals out of a loop). If this is the case, we don't
2953 // want to forward-subst the cast.
2954 if (isa<Constant>(CI->getOperand(0)))
2957 if (TLI && OptimizeNoopCopyExpression(CI, *TLI))
2960 if (isa<ZExtInst>(I) || isa<SExtInst>(I)) {
2961 /// Sink a zext or sext into its user blocks if the target type doesn't
2962 /// fit in one register
2963 if (TLI && TLI->getTypeAction(CI->getContext(),
2964 TLI->getValueType(CI->getType())) ==
2965 TargetLowering::TypeExpandInteger) {
2966 return SinkCast(CI);
2968 bool MadeChange = MoveExtToFormExtLoad(I);
2969 return MadeChange | OptimizeExtUses(I);
2975 if (CmpInst *CI = dyn_cast<CmpInst>(I))
2976 if (!TLI || !TLI->hasMultipleConditionRegisters())
2977 return OptimizeCmpExpression(CI);
2979 if (LoadInst *LI = dyn_cast<LoadInst>(I)) {
2981 return OptimizeMemoryInst(I, I->getOperand(0), LI->getType());
2985 if (StoreInst *SI = dyn_cast<StoreInst>(I)) {
2987 return OptimizeMemoryInst(I, SI->getOperand(1),
2988 SI->getOperand(0)->getType());
2992 if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(I)) {
2993 if (GEPI->hasAllZeroIndices()) {
2994 /// The GEP operand must be a pointer, so must its result -> BitCast
2995 Instruction *NC = new BitCastInst(GEPI->getOperand(0), GEPI->getType(),
2996 GEPI->getName(), GEPI);
2997 GEPI->replaceAllUsesWith(NC);
2998 GEPI->eraseFromParent();
3006 if (CallInst *CI = dyn_cast<CallInst>(I))
3007 return OptimizeCallInst(CI);
3009 if (SelectInst *SI = dyn_cast<SelectInst>(I))
3010 return OptimizeSelectInst(SI);
3012 if (ShuffleVectorInst *SVI = dyn_cast<ShuffleVectorInst>(I))
3013 return OptimizeShuffleVectorInst(SVI);
3018 // In this pass we look for GEP and cast instructions that are used
3019 // across basic blocks and rewrite them to improve basic-block-at-a-time
3021 bool CodeGenPrepare::OptimizeBlock(BasicBlock &BB) {
3023 bool MadeChange = false;
3025 CurInstIterator = BB.begin();
3026 while (CurInstIterator != BB.end())
3027 MadeChange |= OptimizeInst(CurInstIterator++);
3029 MadeChange |= DupRetToEnableTailCallOpts(&BB);
3034 // llvm.dbg.value is far away from the value then iSel may not be able
3035 // handle it properly. iSel will drop llvm.dbg.value if it can not
3036 // find a node corresponding to the value.
3037 bool CodeGenPrepare::PlaceDbgValues(Function &F) {
3038 bool MadeChange = false;
3039 for (Function::iterator I = F.begin(), E = F.end(); I != E; ++I) {
3040 Instruction *PrevNonDbgInst = nullptr;
3041 for (BasicBlock::iterator BI = I->begin(), BE = I->end(); BI != BE;) {
3042 Instruction *Insn = BI; ++BI;
3043 DbgValueInst *DVI = dyn_cast<DbgValueInst>(Insn);
3045 PrevNonDbgInst = Insn;
3049 Instruction *VI = dyn_cast_or_null<Instruction>(DVI->getValue());
3050 if (VI && VI != PrevNonDbgInst && !VI->isTerminator()) {
3051 DEBUG(dbgs() << "Moving Debug Value before :\n" << *DVI << ' ' << *VI);
3052 DVI->removeFromParent();
3053 if (isa<PHINode>(VI))
3054 DVI->insertBefore(VI->getParent()->getFirstInsertionPt());
3056 DVI->insertAfter(VI);
3065 // If there is a sequence that branches based on comparing a single bit
3066 // against zero that can be combined into a single instruction, and the
3067 // target supports folding these into a single instruction, sink the
3068 // mask and compare into the branch uses. Do this before OptimizeBlock ->
3069 // OptimizeInst -> OptimizeCmpExpression, which perturbs the pattern being
3071 bool CodeGenPrepare::sinkAndCmp(Function &F) {
3072 if (!EnableAndCmpSinking)
3074 if (!TLI || !TLI->isMaskAndBranchFoldingLegal())
3076 bool MadeChange = false;
3077 for (Function::iterator I = F.begin(), E = F.end(); I != E; ) {
3078 BasicBlock *BB = I++;
3080 // Does this BB end with the following?
3081 // %andVal = and %val, #single-bit-set
3082 // %icmpVal = icmp %andResult, 0
3083 // br i1 %cmpVal label %dest1, label %dest2"
3084 BranchInst *Brcc = dyn_cast<BranchInst>(BB->getTerminator());
3085 if (!Brcc || !Brcc->isConditional())
3087 ICmpInst *Cmp = dyn_cast<ICmpInst>(Brcc->getOperand(0));
3088 if (!Cmp || Cmp->getParent() != BB)
3090 ConstantInt *Zero = dyn_cast<ConstantInt>(Cmp->getOperand(1));
3091 if (!Zero || !Zero->isZero())
3093 Instruction *And = dyn_cast<Instruction>(Cmp->getOperand(0));
3094 if (!And || And->getOpcode() != Instruction::And || And->getParent() != BB)
3096 ConstantInt* Mask = dyn_cast<ConstantInt>(And->getOperand(1));
3097 if (!Mask || !Mask->getUniqueInteger().isPowerOf2())
3099 DEBUG(dbgs() << "found and; icmp ?,0; brcc\n"); DEBUG(BB->dump());
3101 // Push the "and; icmp" for any users that are conditional branches.
3102 // Since there can only be one branch use per BB, we don't need to keep
3103 // track of which BBs we insert into.
3104 for (Value::use_iterator UI = Cmp->use_begin(), E = Cmp->use_end();
3108 BranchInst *BrccUser = dyn_cast<BranchInst>(*UI);
3110 if (!BrccUser || !BrccUser->isConditional())
3112 BasicBlock *UserBB = BrccUser->getParent();
3113 if (UserBB == BB) continue;
3114 DEBUG(dbgs() << "found Brcc use\n");
3116 // Sink the "and; icmp" to use.
3118 BinaryOperator *NewAnd =
3119 BinaryOperator::CreateAnd(And->getOperand(0), And->getOperand(1), "",
3122 CmpInst::Create(Cmp->getOpcode(), Cmp->getPredicate(), NewAnd, Zero,
3126 DEBUG(BrccUser->getParent()->dump());