1 //===- CodeGenPrepare.cpp - Prepare a function for code generation --------===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // This pass munges the code in the input function to better prepare it for
11 // SelectionDAG-based code generation. This works around limitations in it's
12 // basic-block-at-a-time approach. It should eventually be removed.
14 //===----------------------------------------------------------------------===//
16 #include "llvm/CodeGen/Passes.h"
17 #include "llvm/ADT/DenseMap.h"
18 #include "llvm/ADT/SmallSet.h"
19 #include "llvm/ADT/Statistic.h"
20 #include "llvm/Analysis/InstructionSimplify.h"
21 #include "llvm/IR/CallSite.h"
22 #include "llvm/IR/Constants.h"
23 #include "llvm/IR/DataLayout.h"
24 #include "llvm/IR/DerivedTypes.h"
25 #include "llvm/IR/Dominators.h"
26 #include "llvm/IR/Function.h"
27 #include "llvm/IR/GetElementPtrTypeIterator.h"
28 #include "llvm/IR/IRBuilder.h"
29 #include "llvm/IR/InlineAsm.h"
30 #include "llvm/IR/Instructions.h"
31 #include "llvm/IR/IntrinsicInst.h"
32 #include "llvm/IR/PatternMatch.h"
33 #include "llvm/IR/ValueHandle.h"
34 #include "llvm/IR/ValueMap.h"
35 #include "llvm/Pass.h"
36 #include "llvm/Support/CommandLine.h"
37 #include "llvm/Support/Debug.h"
38 #include "llvm/Support/raw_ostream.h"
39 #include "llvm/Target/TargetLibraryInfo.h"
40 #include "llvm/Target/TargetLowering.h"
41 #include "llvm/Target/TargetSubtargetInfo.h"
42 #include "llvm/Transforms/Utils/BasicBlockUtils.h"
43 #include "llvm/Transforms/Utils/BuildLibCalls.h"
44 #include "llvm/Transforms/Utils/BypassSlowDivision.h"
45 #include "llvm/Transforms/Utils/Local.h"
47 using namespace llvm::PatternMatch;
49 #define DEBUG_TYPE "codegenprepare"
51 STATISTIC(NumBlocksElim, "Number of blocks eliminated");
52 STATISTIC(NumPHIsElim, "Number of trivial PHIs eliminated");
53 STATISTIC(NumGEPsElim, "Number of GEPs converted to casts");
54 STATISTIC(NumCmpUses, "Number of uses of Cmp expressions replaced with uses of "
56 STATISTIC(NumCastUses, "Number of uses of Cast expressions replaced with uses "
58 STATISTIC(NumMemoryInsts, "Number of memory instructions whose address "
59 "computations were sunk");
60 STATISTIC(NumExtsMoved, "Number of [s|z]ext instructions combined with loads");
61 STATISTIC(NumExtUses, "Number of uses of [s|z]ext instructions optimized");
62 STATISTIC(NumRetsDup, "Number of return instructions duplicated");
63 STATISTIC(NumDbgValueMoved, "Number of debug value instructions moved");
64 STATISTIC(NumSelectsExpanded, "Number of selects turned into branches");
65 STATISTIC(NumAndCmpsMoved, "Number of and/cmp's pushed into branches");
67 static cl::opt<bool> DisableBranchOpts(
68 "disable-cgp-branch-opts", cl::Hidden, cl::init(false),
69 cl::desc("Disable branch optimizations in CodeGenPrepare"));
71 static cl::opt<bool> DisableSelectToBranch(
72 "disable-cgp-select2branch", cl::Hidden, cl::init(false),
73 cl::desc("Disable select to branch conversion."));
75 static cl::opt<bool> AddrSinkUsingGEPs(
76 "addr-sink-using-gep", cl::Hidden, cl::init(false),
77 cl::desc("Address sinking in CGP using GEPs."));
79 static cl::opt<bool> EnableAndCmpSinking(
80 "enable-andcmp-sinking", cl::Hidden, cl::init(true),
81 cl::desc("Enable sinkinig and/cmp into branches."));
84 typedef SmallPtrSet<Instruction *, 16> SetOfInstrs;
85 typedef DenseMap<Instruction *, Type *> InstrToOrigTy;
87 class CodeGenPrepare : public FunctionPass {
88 /// TLI - Keep a pointer of a TargetLowering to consult for determining
89 /// transformation profitability.
90 const TargetMachine *TM;
91 const TargetLowering *TLI;
92 const TargetLibraryInfo *TLInfo;
95 /// CurInstIterator - As we scan instructions optimizing them, this is the
96 /// next instruction to optimize. Xforms that can invalidate this should
98 BasicBlock::iterator CurInstIterator;
100 /// Keeps track of non-local addresses that have been sunk into a block.
101 /// This allows us to avoid inserting duplicate code for blocks with
102 /// multiple load/stores of the same address.
103 ValueMap<Value*, Value*> SunkAddrs;
105 /// Keeps track of all truncates inserted for the current function.
106 SetOfInstrs InsertedTruncsSet;
107 /// Keeps track of the type of the related instruction before their
108 /// promotion for the current function.
109 InstrToOrigTy PromotedInsts;
111 /// ModifiedDT - If CFG is modified in anyway, dominator tree may need to
115 /// OptSize - True if optimizing for size.
119 static char ID; // Pass identification, replacement for typeid
120 explicit CodeGenPrepare(const TargetMachine *TM = nullptr)
121 : FunctionPass(ID), TM(TM), TLI(nullptr) {
122 initializeCodeGenPreparePass(*PassRegistry::getPassRegistry());
124 bool runOnFunction(Function &F) override;
126 const char *getPassName() const override { return "CodeGen Prepare"; }
128 void getAnalysisUsage(AnalysisUsage &AU) const override {
129 AU.addPreserved<DominatorTreeWrapperPass>();
130 AU.addRequired<TargetLibraryInfo>();
134 bool EliminateFallThrough(Function &F);
135 bool EliminateMostlyEmptyBlocks(Function &F);
136 bool CanMergeBlocks(const BasicBlock *BB, const BasicBlock *DestBB) const;
137 void EliminateMostlyEmptyBlock(BasicBlock *BB);
138 bool OptimizeBlock(BasicBlock &BB);
139 bool OptimizeInst(Instruction *I);
140 bool OptimizeMemoryInst(Instruction *I, Value *Addr, Type *AccessTy);
141 bool OptimizeInlineAsmInst(CallInst *CS);
142 bool OptimizeCallInst(CallInst *CI);
143 bool MoveExtToFormExtLoad(Instruction *I);
144 bool OptimizeExtUses(Instruction *I);
145 bool OptimizeSelectInst(SelectInst *SI);
146 bool OptimizeShuffleVectorInst(ShuffleVectorInst *SI);
147 bool DupRetToEnableTailCallOpts(BasicBlock *BB);
148 bool PlaceDbgValues(Function &F);
149 bool sinkAndCmp(Function &F);
153 char CodeGenPrepare::ID = 0;
154 INITIALIZE_TM_PASS(CodeGenPrepare, "codegenprepare",
155 "Optimize for code generation", false, false)
157 FunctionPass *llvm::createCodeGenPreparePass(const TargetMachine *TM) {
158 return new CodeGenPrepare(TM);
161 bool CodeGenPrepare::runOnFunction(Function &F) {
162 if (skipOptnoneFunction(F))
165 bool EverMadeChange = false;
166 // Clear per function information.
167 InsertedTruncsSet.clear();
168 PromotedInsts.clear();
171 if (TM) TLI = TM->getTargetLowering();
172 TLInfo = &getAnalysis<TargetLibraryInfo>();
173 DominatorTreeWrapperPass *DTWP =
174 getAnalysisIfAvailable<DominatorTreeWrapperPass>();
175 DT = DTWP ? &DTWP->getDomTree() : nullptr;
176 OptSize = F.getAttributes().hasAttribute(AttributeSet::FunctionIndex,
177 Attribute::OptimizeForSize);
179 /// This optimization identifies DIV instructions that can be
180 /// profitably bypassed and carried out with a shorter, faster divide.
181 if (!OptSize && TLI && TLI->isSlowDivBypassed()) {
182 const DenseMap<unsigned int, unsigned int> &BypassWidths =
183 TLI->getBypassSlowDivWidths();
184 for (Function::iterator I = F.begin(); I != F.end(); I++)
185 EverMadeChange |= bypassSlowDivision(F, I, BypassWidths);
188 // Eliminate blocks that contain only PHI nodes and an
189 // unconditional branch.
190 EverMadeChange |= EliminateMostlyEmptyBlocks(F);
192 // llvm.dbg.value is far away from the value then iSel may not be able
193 // handle it properly. iSel will drop llvm.dbg.value if it can not
194 // find a node corresponding to the value.
195 EverMadeChange |= PlaceDbgValues(F);
197 // If there is a mask, compare against zero, and branch that can be combined
198 // into a single target instruction, push the mask and compare into branch
199 // users. Do this before OptimizeBlock -> OptimizeInst ->
200 // OptimizeCmpExpression, which perturbs the pattern being searched for.
201 if (!DisableBranchOpts)
202 EverMadeChange |= sinkAndCmp(F);
204 bool MadeChange = true;
207 for (Function::iterator I = F.begin(); I != F.end(); ) {
208 BasicBlock *BB = I++;
209 MadeChange |= OptimizeBlock(*BB);
211 EverMadeChange |= MadeChange;
216 if (!DisableBranchOpts) {
218 SmallPtrSet<BasicBlock*, 8> WorkList;
219 for (Function::iterator BB = F.begin(), E = F.end(); BB != E; ++BB) {
220 SmallVector<BasicBlock*, 2> Successors(succ_begin(BB), succ_end(BB));
221 MadeChange |= ConstantFoldTerminator(BB, true);
222 if (!MadeChange) continue;
224 for (SmallVectorImpl<BasicBlock*>::iterator
225 II = Successors.begin(), IE = Successors.end(); II != IE; ++II)
226 if (pred_begin(*II) == pred_end(*II))
227 WorkList.insert(*II);
230 // Delete the dead blocks and any of their dead successors.
231 MadeChange |= !WorkList.empty();
232 while (!WorkList.empty()) {
233 BasicBlock *BB = *WorkList.begin();
235 SmallVector<BasicBlock*, 2> Successors(succ_begin(BB), succ_end(BB));
239 for (SmallVectorImpl<BasicBlock*>::iterator
240 II = Successors.begin(), IE = Successors.end(); II != IE; ++II)
241 if (pred_begin(*II) == pred_end(*II))
242 WorkList.insert(*II);
245 // Merge pairs of basic blocks with unconditional branches, connected by
247 if (EverMadeChange || MadeChange)
248 MadeChange |= EliminateFallThrough(F);
252 EverMadeChange |= MadeChange;
255 if (ModifiedDT && DT)
258 return EverMadeChange;
261 /// EliminateFallThrough - Merge basic blocks which are connected
262 /// by a single edge, where one of the basic blocks has a single successor
263 /// pointing to the other basic block, which has a single predecessor.
264 bool CodeGenPrepare::EliminateFallThrough(Function &F) {
265 bool Changed = false;
266 // Scan all of the blocks in the function, except for the entry block.
267 for (Function::iterator I = std::next(F.begin()), E = F.end(); I != E;) {
268 BasicBlock *BB = I++;
269 // If the destination block has a single pred, then this is a trivial
270 // edge, just collapse it.
271 BasicBlock *SinglePred = BB->getSinglePredecessor();
273 // Don't merge if BB's address is taken.
274 if (!SinglePred || SinglePred == BB || BB->hasAddressTaken()) continue;
276 BranchInst *Term = dyn_cast<BranchInst>(SinglePred->getTerminator());
277 if (Term && !Term->isConditional()) {
279 DEBUG(dbgs() << "To merge:\n"<< *SinglePred << "\n\n\n");
280 // Remember if SinglePred was the entry block of the function.
281 // If so, we will need to move BB back to the entry position.
282 bool isEntry = SinglePred == &SinglePred->getParent()->getEntryBlock();
283 MergeBasicBlockIntoOnlyPred(BB, this);
285 if (isEntry && BB != &BB->getParent()->getEntryBlock())
286 BB->moveBefore(&BB->getParent()->getEntryBlock());
288 // We have erased a block. Update the iterator.
295 /// EliminateMostlyEmptyBlocks - eliminate blocks that contain only PHI nodes,
296 /// debug info directives, and an unconditional branch. Passes before isel
297 /// (e.g. LSR/loopsimplify) often split edges in ways that are non-optimal for
298 /// isel. Start by eliminating these blocks so we can split them the way we
300 bool CodeGenPrepare::EliminateMostlyEmptyBlocks(Function &F) {
301 bool MadeChange = false;
302 // Note that this intentionally skips the entry block.
303 for (Function::iterator I = std::next(F.begin()), E = F.end(); I != E;) {
304 BasicBlock *BB = I++;
306 // If this block doesn't end with an uncond branch, ignore it.
307 BranchInst *BI = dyn_cast<BranchInst>(BB->getTerminator());
308 if (!BI || !BI->isUnconditional())
311 // If the instruction before the branch (skipping debug info) isn't a phi
312 // node, then other stuff is happening here.
313 BasicBlock::iterator BBI = BI;
314 if (BBI != BB->begin()) {
316 while (isa<DbgInfoIntrinsic>(BBI)) {
317 if (BBI == BB->begin())
321 if (!isa<DbgInfoIntrinsic>(BBI) && !isa<PHINode>(BBI))
325 // Do not break infinite loops.
326 BasicBlock *DestBB = BI->getSuccessor(0);
330 if (!CanMergeBlocks(BB, DestBB))
333 EliminateMostlyEmptyBlock(BB);
339 /// CanMergeBlocks - Return true if we can merge BB into DestBB if there is a
340 /// single uncond branch between them, and BB contains no other non-phi
342 bool CodeGenPrepare::CanMergeBlocks(const BasicBlock *BB,
343 const BasicBlock *DestBB) const {
344 // We only want to eliminate blocks whose phi nodes are used by phi nodes in
345 // the successor. If there are more complex condition (e.g. preheaders),
346 // don't mess around with them.
347 BasicBlock::const_iterator BBI = BB->begin();
348 while (const PHINode *PN = dyn_cast<PHINode>(BBI++)) {
349 for (const User *U : PN->users()) {
350 const Instruction *UI = cast<Instruction>(U);
351 if (UI->getParent() != DestBB || !isa<PHINode>(UI))
353 // If User is inside DestBB block and it is a PHINode then check
354 // incoming value. If incoming value is not from BB then this is
355 // a complex condition (e.g. preheaders) we want to avoid here.
356 if (UI->getParent() == DestBB) {
357 if (const PHINode *UPN = dyn_cast<PHINode>(UI))
358 for (unsigned I = 0, E = UPN->getNumIncomingValues(); I != E; ++I) {
359 Instruction *Insn = dyn_cast<Instruction>(UPN->getIncomingValue(I));
360 if (Insn && Insn->getParent() == BB &&
361 Insn->getParent() != UPN->getIncomingBlock(I))
368 // If BB and DestBB contain any common predecessors, then the phi nodes in BB
369 // and DestBB may have conflicting incoming values for the block. If so, we
370 // can't merge the block.
371 const PHINode *DestBBPN = dyn_cast<PHINode>(DestBB->begin());
372 if (!DestBBPN) return true; // no conflict.
374 // Collect the preds of BB.
375 SmallPtrSet<const BasicBlock*, 16> BBPreds;
376 if (const PHINode *BBPN = dyn_cast<PHINode>(BB->begin())) {
377 // It is faster to get preds from a PHI than with pred_iterator.
378 for (unsigned i = 0, e = BBPN->getNumIncomingValues(); i != e; ++i)
379 BBPreds.insert(BBPN->getIncomingBlock(i));
381 BBPreds.insert(pred_begin(BB), pred_end(BB));
384 // Walk the preds of DestBB.
385 for (unsigned i = 0, e = DestBBPN->getNumIncomingValues(); i != e; ++i) {
386 BasicBlock *Pred = DestBBPN->getIncomingBlock(i);
387 if (BBPreds.count(Pred)) { // Common predecessor?
388 BBI = DestBB->begin();
389 while (const PHINode *PN = dyn_cast<PHINode>(BBI++)) {
390 const Value *V1 = PN->getIncomingValueForBlock(Pred);
391 const Value *V2 = PN->getIncomingValueForBlock(BB);
393 // If V2 is a phi node in BB, look up what the mapped value will be.
394 if (const PHINode *V2PN = dyn_cast<PHINode>(V2))
395 if (V2PN->getParent() == BB)
396 V2 = V2PN->getIncomingValueForBlock(Pred);
398 // If there is a conflict, bail out.
399 if (V1 != V2) return false;
408 /// EliminateMostlyEmptyBlock - Eliminate a basic block that have only phi's and
409 /// an unconditional branch in it.
410 void CodeGenPrepare::EliminateMostlyEmptyBlock(BasicBlock *BB) {
411 BranchInst *BI = cast<BranchInst>(BB->getTerminator());
412 BasicBlock *DestBB = BI->getSuccessor(0);
414 DEBUG(dbgs() << "MERGING MOSTLY EMPTY BLOCKS - BEFORE:\n" << *BB << *DestBB);
416 // If the destination block has a single pred, then this is a trivial edge,
418 if (BasicBlock *SinglePred = DestBB->getSinglePredecessor()) {
419 if (SinglePred != DestBB) {
420 // Remember if SinglePred was the entry block of the function. If so, we
421 // will need to move BB back to the entry position.
422 bool isEntry = SinglePred == &SinglePred->getParent()->getEntryBlock();
423 MergeBasicBlockIntoOnlyPred(DestBB, this);
425 if (isEntry && BB != &BB->getParent()->getEntryBlock())
426 BB->moveBefore(&BB->getParent()->getEntryBlock());
428 DEBUG(dbgs() << "AFTER:\n" << *DestBB << "\n\n\n");
433 // Otherwise, we have multiple predecessors of BB. Update the PHIs in DestBB
434 // to handle the new incoming edges it is about to have.
436 for (BasicBlock::iterator BBI = DestBB->begin();
437 (PN = dyn_cast<PHINode>(BBI)); ++BBI) {
438 // Remove the incoming value for BB, and remember it.
439 Value *InVal = PN->removeIncomingValue(BB, false);
441 // Two options: either the InVal is a phi node defined in BB or it is some
442 // value that dominates BB.
443 PHINode *InValPhi = dyn_cast<PHINode>(InVal);
444 if (InValPhi && InValPhi->getParent() == BB) {
445 // Add all of the input values of the input PHI as inputs of this phi.
446 for (unsigned i = 0, e = InValPhi->getNumIncomingValues(); i != e; ++i)
447 PN->addIncoming(InValPhi->getIncomingValue(i),
448 InValPhi->getIncomingBlock(i));
450 // Otherwise, add one instance of the dominating value for each edge that
451 // we will be adding.
452 if (PHINode *BBPN = dyn_cast<PHINode>(BB->begin())) {
453 for (unsigned i = 0, e = BBPN->getNumIncomingValues(); i != e; ++i)
454 PN->addIncoming(InVal, BBPN->getIncomingBlock(i));
456 for (pred_iterator PI = pred_begin(BB), E = pred_end(BB); PI != E; ++PI)
457 PN->addIncoming(InVal, *PI);
462 // The PHIs are now updated, change everything that refers to BB to use
463 // DestBB and remove BB.
464 BB->replaceAllUsesWith(DestBB);
465 if (DT && !ModifiedDT) {
466 BasicBlock *BBIDom = DT->getNode(BB)->getIDom()->getBlock();
467 BasicBlock *DestBBIDom = DT->getNode(DestBB)->getIDom()->getBlock();
468 BasicBlock *NewIDom = DT->findNearestCommonDominator(BBIDom, DestBBIDom);
469 DT->changeImmediateDominator(DestBB, NewIDom);
472 BB->eraseFromParent();
475 DEBUG(dbgs() << "AFTER:\n" << *DestBB << "\n\n\n");
478 /// SinkCast - Sink the specified cast instruction into its user blocks
479 static bool SinkCast(CastInst *CI) {
480 BasicBlock *DefBB = CI->getParent();
482 /// InsertedCasts - Only insert a cast in each block once.
483 DenseMap<BasicBlock*, CastInst*> InsertedCasts;
485 bool MadeChange = false;
486 for (Value::user_iterator UI = CI->user_begin(), E = CI->user_end();
488 Use &TheUse = UI.getUse();
489 Instruction *User = cast<Instruction>(*UI);
491 // Figure out which BB this cast is used in. For PHI's this is the
492 // appropriate predecessor block.
493 BasicBlock *UserBB = User->getParent();
494 if (PHINode *PN = dyn_cast<PHINode>(User)) {
495 UserBB = PN->getIncomingBlock(TheUse);
498 // Preincrement use iterator so we don't invalidate it.
501 // If this user is in the same block as the cast, don't change the cast.
502 if (UserBB == DefBB) continue;
504 // If we have already inserted a cast into this block, use it.
505 CastInst *&InsertedCast = InsertedCasts[UserBB];
508 BasicBlock::iterator InsertPt = UserBB->getFirstInsertionPt();
510 CastInst::Create(CI->getOpcode(), CI->getOperand(0), CI->getType(), "",
515 // Replace a use of the cast with a use of the new cast.
516 TheUse = InsertedCast;
520 // If we removed all uses, nuke the cast.
521 if (CI->use_empty()) {
522 CI->eraseFromParent();
529 /// OptimizeNoopCopyExpression - If the specified cast instruction is a noop
530 /// copy (e.g. it's casting from one pointer type to another, i32->i8 on PPC),
531 /// sink it into user blocks to reduce the number of virtual
532 /// registers that must be created and coalesced.
534 /// Return true if any changes are made.
536 static bool OptimizeNoopCopyExpression(CastInst *CI, const TargetLowering &TLI){
537 // If this is a noop copy,
538 EVT SrcVT = TLI.getValueType(CI->getOperand(0)->getType());
539 EVT DstVT = TLI.getValueType(CI->getType());
541 // This is an fp<->int conversion?
542 if (SrcVT.isInteger() != DstVT.isInteger())
545 // If this is an extension, it will be a zero or sign extension, which
547 if (SrcVT.bitsLT(DstVT)) return false;
549 // If these values will be promoted, find out what they will be promoted
550 // to. This helps us consider truncates on PPC as noop copies when they
552 if (TLI.getTypeAction(CI->getContext(), SrcVT) ==
553 TargetLowering::TypePromoteInteger)
554 SrcVT = TLI.getTypeToTransformTo(CI->getContext(), SrcVT);
555 if (TLI.getTypeAction(CI->getContext(), DstVT) ==
556 TargetLowering::TypePromoteInteger)
557 DstVT = TLI.getTypeToTransformTo(CI->getContext(), DstVT);
559 // If, after promotion, these are the same types, this is a noop copy.
566 /// OptimizeCmpExpression - sink the given CmpInst into user blocks to reduce
567 /// the number of virtual registers that must be created and coalesced. This is
568 /// a clear win except on targets with multiple condition code registers
569 /// (PowerPC), where it might lose; some adjustment may be wanted there.
571 /// Return true if any changes are made.
572 static bool OptimizeCmpExpression(CmpInst *CI) {
573 BasicBlock *DefBB = CI->getParent();
575 /// InsertedCmp - Only insert a cmp in each block once.
576 DenseMap<BasicBlock*, CmpInst*> InsertedCmps;
578 bool MadeChange = false;
579 for (Value::user_iterator UI = CI->user_begin(), E = CI->user_end();
581 Use &TheUse = UI.getUse();
582 Instruction *User = cast<Instruction>(*UI);
584 // Preincrement use iterator so we don't invalidate it.
587 // Don't bother for PHI nodes.
588 if (isa<PHINode>(User))
591 // Figure out which BB this cmp is used in.
592 BasicBlock *UserBB = User->getParent();
594 // If this user is in the same block as the cmp, don't change the cmp.
595 if (UserBB == DefBB) continue;
597 // If we have already inserted a cmp into this block, use it.
598 CmpInst *&InsertedCmp = InsertedCmps[UserBB];
601 BasicBlock::iterator InsertPt = UserBB->getFirstInsertionPt();
603 CmpInst::Create(CI->getOpcode(),
604 CI->getPredicate(), CI->getOperand(0),
605 CI->getOperand(1), "", InsertPt);
609 // Replace a use of the cmp with a use of the new cmp.
610 TheUse = InsertedCmp;
614 // If we removed all uses, nuke the cmp.
616 CI->eraseFromParent();
621 /// isExtractBitsCandidateUse - Check if the candidates could
622 /// be combined with shift instruction, which includes:
623 /// 1. Truncate instruction
624 /// 2. And instruction and the imm is a mask of the low bits:
625 /// imm & (imm+1) == 0
626 static bool isExtractBitsCandidateUse(Instruction *User) {
627 if (!isa<TruncInst>(User)) {
628 if (User->getOpcode() != Instruction::And ||
629 !isa<ConstantInt>(User->getOperand(1)))
632 const APInt &Cimm = cast<ConstantInt>(User->getOperand(1))->getValue();
634 if ((Cimm & (Cimm + 1)).getBoolValue())
640 /// SinkShiftAndTruncate - sink both shift and truncate instruction
641 /// to the use of truncate's BB.
643 SinkShiftAndTruncate(BinaryOperator *ShiftI, Instruction *User, ConstantInt *CI,
644 DenseMap<BasicBlock *, BinaryOperator *> &InsertedShifts,
645 const TargetLowering &TLI) {
646 BasicBlock *UserBB = User->getParent();
647 DenseMap<BasicBlock *, CastInst *> InsertedTruncs;
648 TruncInst *TruncI = dyn_cast<TruncInst>(User);
649 bool MadeChange = false;
651 for (Value::user_iterator TruncUI = TruncI->user_begin(),
652 TruncE = TruncI->user_end();
653 TruncUI != TruncE;) {
655 Use &TruncTheUse = TruncUI.getUse();
656 Instruction *TruncUser = cast<Instruction>(*TruncUI);
657 // Preincrement use iterator so we don't invalidate it.
661 int ISDOpcode = TLI.InstructionOpcodeToISD(TruncUser->getOpcode());
665 // If the use is actually a legal node, there will not be an implicit
667 if (TLI.isOperationLegalOrCustom(ISDOpcode,
668 EVT::getEVT(TruncUser->getType())))
671 // Don't bother for PHI nodes.
672 if (isa<PHINode>(TruncUser))
675 BasicBlock *TruncUserBB = TruncUser->getParent();
677 if (UserBB == TruncUserBB)
680 BinaryOperator *&InsertedShift = InsertedShifts[TruncUserBB];
681 CastInst *&InsertedTrunc = InsertedTruncs[TruncUserBB];
683 if (!InsertedShift && !InsertedTrunc) {
684 BasicBlock::iterator InsertPt = TruncUserBB->getFirstInsertionPt();
686 if (ShiftI->getOpcode() == Instruction::AShr)
688 BinaryOperator::CreateAShr(ShiftI->getOperand(0), CI, "", InsertPt);
691 BinaryOperator::CreateLShr(ShiftI->getOperand(0), CI, "", InsertPt);
694 BasicBlock::iterator TruncInsertPt = TruncUserBB->getFirstInsertionPt();
697 InsertedTrunc = CastInst::Create(TruncI->getOpcode(), InsertedShift,
698 TruncI->getType(), "", TruncInsertPt);
702 TruncTheUse = InsertedTrunc;
708 /// OptimizeExtractBits - sink the shift *right* instruction into user blocks if
709 /// the uses could potentially be combined with this shift instruction and
710 /// generate BitExtract instruction. It will only be applied if the architecture
711 /// supports BitExtract instruction. Here is an example:
713 /// %x.extract.shift = lshr i64 %arg1, 32
715 /// %x.extract.trunc = trunc i64 %x.extract.shift to i16
719 /// %x.extract.shift.1 = lshr i64 %arg1, 32
720 /// %x.extract.trunc = trunc i64 %x.extract.shift.1 to i16
722 /// CodeGen will recoginze the pattern in BB2 and generate BitExtract
724 /// Return true if any changes are made.
725 static bool OptimizeExtractBits(BinaryOperator *ShiftI, ConstantInt *CI,
726 const TargetLowering &TLI) {
727 BasicBlock *DefBB = ShiftI->getParent();
729 /// Only insert instructions in each block once.
730 DenseMap<BasicBlock *, BinaryOperator *> InsertedShifts;
732 bool shiftIsLegal = TLI.isTypeLegal(TLI.getValueType(ShiftI->getType()));
734 bool MadeChange = false;
735 for (Value::user_iterator UI = ShiftI->user_begin(), E = ShiftI->user_end();
737 Use &TheUse = UI.getUse();
738 Instruction *User = cast<Instruction>(*UI);
739 // Preincrement use iterator so we don't invalidate it.
742 // Don't bother for PHI nodes.
743 if (isa<PHINode>(User))
746 if (!isExtractBitsCandidateUse(User))
749 BasicBlock *UserBB = User->getParent();
751 if (UserBB == DefBB) {
752 // If the shift and truncate instruction are in the same BB. The use of
753 // the truncate(TruncUse) may still introduce another truncate if not
754 // legal. In this case, we would like to sink both shift and truncate
755 // instruction to the BB of TruncUse.
758 // i64 shift.result = lshr i64 opnd, imm
759 // trunc.result = trunc shift.result to i16
762 // ----> We will have an implicit truncate here if the architecture does
763 // not have i16 compare.
764 // cmp i16 trunc.result, opnd2
766 if (isa<TruncInst>(User) && shiftIsLegal
767 // If the type of the truncate is legal, no trucate will be
768 // introduced in other basic blocks.
769 && (!TLI.isTypeLegal(TLI.getValueType(User->getType()))))
771 SinkShiftAndTruncate(ShiftI, User, CI, InsertedShifts, TLI);
775 // If we have already inserted a shift into this block, use it.
776 BinaryOperator *&InsertedShift = InsertedShifts[UserBB];
778 if (!InsertedShift) {
779 BasicBlock::iterator InsertPt = UserBB->getFirstInsertionPt();
781 if (ShiftI->getOpcode() == Instruction::AShr)
783 BinaryOperator::CreateAShr(ShiftI->getOperand(0), CI, "", InsertPt);
786 BinaryOperator::CreateLShr(ShiftI->getOperand(0), CI, "", InsertPt);
791 // Replace a use of the shift with a use of the new shift.
792 TheUse = InsertedShift;
795 // If we removed all uses, nuke the shift.
796 if (ShiftI->use_empty())
797 ShiftI->eraseFromParent();
803 class CodeGenPrepareFortifiedLibCalls : public SimplifyFortifiedLibCalls {
805 void replaceCall(Value *With) override {
806 CI->replaceAllUsesWith(With);
807 CI->eraseFromParent();
809 bool isFoldable(unsigned SizeCIOp, unsigned, bool) const override {
810 if (ConstantInt *SizeCI =
811 dyn_cast<ConstantInt>(CI->getArgOperand(SizeCIOp)))
812 return SizeCI->isAllOnesValue();
816 } // end anonymous namespace
818 bool CodeGenPrepare::OptimizeCallInst(CallInst *CI) {
819 BasicBlock *BB = CI->getParent();
821 // Lower inline assembly if we can.
822 // If we found an inline asm expession, and if the target knows how to
823 // lower it to normal LLVM code, do so now.
824 if (TLI && isa<InlineAsm>(CI->getCalledValue())) {
825 if (TLI->ExpandInlineAsm(CI)) {
826 // Avoid invalidating the iterator.
827 CurInstIterator = BB->begin();
828 // Avoid processing instructions out of order, which could cause
829 // reuse before a value is defined.
833 // Sink address computing for memory operands into the block.
834 if (OptimizeInlineAsmInst(CI))
838 // Lower all uses of llvm.objectsize.*
839 IntrinsicInst *II = dyn_cast<IntrinsicInst>(CI);
840 if (II && II->getIntrinsicID() == Intrinsic::objectsize) {
841 bool Min = (cast<ConstantInt>(II->getArgOperand(1))->getZExtValue() == 1);
842 Type *ReturnTy = CI->getType();
843 Constant *RetVal = ConstantInt::get(ReturnTy, Min ? 0 : -1ULL);
845 // Substituting this can cause recursive simplifications, which can
846 // invalidate our iterator. Use a WeakVH to hold onto it in case this
848 WeakVH IterHandle(CurInstIterator);
850 replaceAndRecursivelySimplify(CI, RetVal,
851 TLI ? TLI->getDataLayout() : nullptr,
852 TLInfo, ModifiedDT ? nullptr : DT);
854 // If the iterator instruction was recursively deleted, start over at the
855 // start of the block.
856 if (IterHandle != CurInstIterator) {
857 CurInstIterator = BB->begin();
864 SmallVector<Value*, 2> PtrOps;
866 if (TLI->GetAddrModeArguments(II, PtrOps, AccessTy))
867 while (!PtrOps.empty())
868 if (OptimizeMemoryInst(II, PtrOps.pop_back_val(), AccessTy))
872 // From here on out we're working with named functions.
873 if (!CI->getCalledFunction()) return false;
875 // We'll need DataLayout from here on out.
876 const DataLayout *TD = TLI ? TLI->getDataLayout() : nullptr;
877 if (!TD) return false;
879 // Lower all default uses of _chk calls. This is very similar
880 // to what InstCombineCalls does, but here we are only lowering calls
881 // that have the default "don't know" as the objectsize. Anything else
882 // should be left alone.
883 CodeGenPrepareFortifiedLibCalls Simplifier;
884 return Simplifier.fold(CI, TD, TLInfo);
887 /// DupRetToEnableTailCallOpts - Look for opportunities to duplicate return
888 /// instructions to the predecessor to enable tail call optimizations. The
889 /// case it is currently looking for is:
892 /// %tmp0 = tail call i32 @f0()
895 /// %tmp1 = tail call i32 @f1()
898 /// %tmp2 = tail call i32 @f2()
901 /// %retval = phi i32 [ %tmp0, %bb0 ], [ %tmp1, %bb1 ], [ %tmp2, %bb2 ]
909 /// %tmp0 = tail call i32 @f0()
912 /// %tmp1 = tail call i32 @f1()
915 /// %tmp2 = tail call i32 @f2()
918 bool CodeGenPrepare::DupRetToEnableTailCallOpts(BasicBlock *BB) {
922 ReturnInst *RI = dyn_cast<ReturnInst>(BB->getTerminator());
926 PHINode *PN = nullptr;
927 BitCastInst *BCI = nullptr;
928 Value *V = RI->getReturnValue();
930 BCI = dyn_cast<BitCastInst>(V);
932 V = BCI->getOperand(0);
934 PN = dyn_cast<PHINode>(V);
939 if (PN && PN->getParent() != BB)
942 // It's not safe to eliminate the sign / zero extension of the return value.
943 // See llvm::isInTailCallPosition().
944 const Function *F = BB->getParent();
945 AttributeSet CallerAttrs = F->getAttributes();
946 if (CallerAttrs.hasAttribute(AttributeSet::ReturnIndex, Attribute::ZExt) ||
947 CallerAttrs.hasAttribute(AttributeSet::ReturnIndex, Attribute::SExt))
950 // Make sure there are no instructions between the PHI and return, or that the
951 // return is the first instruction in the block.
953 BasicBlock::iterator BI = BB->begin();
954 do { ++BI; } while (isa<DbgInfoIntrinsic>(BI));
956 // Also skip over the bitcast.
961 BasicBlock::iterator BI = BB->begin();
962 while (isa<DbgInfoIntrinsic>(BI)) ++BI;
967 /// Only dup the ReturnInst if the CallInst is likely to be emitted as a tail
969 SmallVector<CallInst*, 4> TailCalls;
971 for (unsigned I = 0, E = PN->getNumIncomingValues(); I != E; ++I) {
972 CallInst *CI = dyn_cast<CallInst>(PN->getIncomingValue(I));
973 // Make sure the phi value is indeed produced by the tail call.
974 if (CI && CI->hasOneUse() && CI->getParent() == PN->getIncomingBlock(I) &&
975 TLI->mayBeEmittedAsTailCall(CI))
976 TailCalls.push_back(CI);
979 SmallPtrSet<BasicBlock*, 4> VisitedBBs;
980 for (pred_iterator PI = pred_begin(BB), PE = pred_end(BB); PI != PE; ++PI) {
981 if (!VisitedBBs.insert(*PI))
984 BasicBlock::InstListType &InstList = (*PI)->getInstList();
985 BasicBlock::InstListType::reverse_iterator RI = InstList.rbegin();
986 BasicBlock::InstListType::reverse_iterator RE = InstList.rend();
987 do { ++RI; } while (RI != RE && isa<DbgInfoIntrinsic>(&*RI));
991 CallInst *CI = dyn_cast<CallInst>(&*RI);
992 if (CI && CI->use_empty() && TLI->mayBeEmittedAsTailCall(CI))
993 TailCalls.push_back(CI);
997 bool Changed = false;
998 for (unsigned i = 0, e = TailCalls.size(); i != e; ++i) {
999 CallInst *CI = TailCalls[i];
1002 // Conservatively require the attributes of the call to match those of the
1003 // return. Ignore noalias because it doesn't affect the call sequence.
1004 AttributeSet CalleeAttrs = CS.getAttributes();
1005 if (AttrBuilder(CalleeAttrs, AttributeSet::ReturnIndex).
1006 removeAttribute(Attribute::NoAlias) !=
1007 AttrBuilder(CalleeAttrs, AttributeSet::ReturnIndex).
1008 removeAttribute(Attribute::NoAlias))
1011 // Make sure the call instruction is followed by an unconditional branch to
1012 // the return block.
1013 BasicBlock *CallBB = CI->getParent();
1014 BranchInst *BI = dyn_cast<BranchInst>(CallBB->getTerminator());
1015 if (!BI || !BI->isUnconditional() || BI->getSuccessor(0) != BB)
1018 // Duplicate the return into CallBB.
1019 (void)FoldReturnIntoUncondBranch(RI, BB, CallBB);
1020 ModifiedDT = Changed = true;
1024 // If we eliminated all predecessors of the block, delete the block now.
1025 if (Changed && !BB->hasAddressTaken() && pred_begin(BB) == pred_end(BB))
1026 BB->eraseFromParent();
1031 //===----------------------------------------------------------------------===//
1032 // Memory Optimization
1033 //===----------------------------------------------------------------------===//
1037 /// ExtAddrMode - This is an extended version of TargetLowering::AddrMode
1038 /// which holds actual Value*'s for register values.
1039 struct ExtAddrMode : public TargetLowering::AddrMode {
1042 ExtAddrMode() : BaseReg(nullptr), ScaledReg(nullptr) {}
1043 void print(raw_ostream &OS) const;
1046 bool operator==(const ExtAddrMode& O) const {
1047 return (BaseReg == O.BaseReg) && (ScaledReg == O.ScaledReg) &&
1048 (BaseGV == O.BaseGV) && (BaseOffs == O.BaseOffs) &&
1049 (HasBaseReg == O.HasBaseReg) && (Scale == O.Scale);
1054 static inline raw_ostream &operator<<(raw_ostream &OS, const ExtAddrMode &AM) {
1060 void ExtAddrMode::print(raw_ostream &OS) const {
1061 bool NeedPlus = false;
1064 OS << (NeedPlus ? " + " : "")
1066 BaseGV->printAsOperand(OS, /*PrintType=*/false);
1071 OS << (NeedPlus ? " + " : "")
1077 OS << (NeedPlus ? " + " : "")
1079 BaseReg->printAsOperand(OS, /*PrintType=*/false);
1083 OS << (NeedPlus ? " + " : "")
1085 ScaledReg->printAsOperand(OS, /*PrintType=*/false);
1091 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
1092 void ExtAddrMode::dump() const {
1098 /// \brief This class provides transaction based operation on the IR.
1099 /// Every change made through this class is recorded in the internal state and
1100 /// can be undone (rollback) until commit is called.
1101 class TypePromotionTransaction {
1103 /// \brief This represents the common interface of the individual transaction.
1104 /// Each class implements the logic for doing one specific modification on
1105 /// the IR via the TypePromotionTransaction.
1106 class TypePromotionAction {
1108 /// The Instruction modified.
1112 /// \brief Constructor of the action.
1113 /// The constructor performs the related action on the IR.
1114 TypePromotionAction(Instruction *Inst) : Inst(Inst) {}
1116 virtual ~TypePromotionAction() {}
1118 /// \brief Undo the modification done by this action.
1119 /// When this method is called, the IR must be in the same state as it was
1120 /// before this action was applied.
1121 /// \pre Undoing the action works if and only if the IR is in the exact same
1122 /// state as it was directly after this action was applied.
1123 virtual void undo() = 0;
1125 /// \brief Advocate every change made by this action.
1126 /// When the results on the IR of the action are to be kept, it is important
1127 /// to call this function, otherwise hidden information may be kept forever.
1128 virtual void commit() {
1129 // Nothing to be done, this action is not doing anything.
1133 /// \brief Utility to remember the position of an instruction.
1134 class InsertionHandler {
1135 /// Position of an instruction.
1136 /// Either an instruction:
1137 /// - Is the first in a basic block: BB is used.
1138 /// - Has a previous instructon: PrevInst is used.
1140 Instruction *PrevInst;
1143 /// Remember whether or not the instruction had a previous instruction.
1144 bool HasPrevInstruction;
1147 /// \brief Record the position of \p Inst.
1148 InsertionHandler(Instruction *Inst) {
1149 BasicBlock::iterator It = Inst;
1150 HasPrevInstruction = (It != (Inst->getParent()->begin()));
1151 if (HasPrevInstruction)
1152 Point.PrevInst = --It;
1154 Point.BB = Inst->getParent();
1157 /// \brief Insert \p Inst at the recorded position.
1158 void insert(Instruction *Inst) {
1159 if (HasPrevInstruction) {
1160 if (Inst->getParent())
1161 Inst->removeFromParent();
1162 Inst->insertAfter(Point.PrevInst);
1164 Instruction *Position = Point.BB->getFirstInsertionPt();
1165 if (Inst->getParent())
1166 Inst->moveBefore(Position);
1168 Inst->insertBefore(Position);
1173 /// \brief Move an instruction before another.
1174 class InstructionMoveBefore : public TypePromotionAction {
1175 /// Original position of the instruction.
1176 InsertionHandler Position;
1179 /// \brief Move \p Inst before \p Before.
1180 InstructionMoveBefore(Instruction *Inst, Instruction *Before)
1181 : TypePromotionAction(Inst), Position(Inst) {
1182 DEBUG(dbgs() << "Do: move: " << *Inst << "\nbefore: " << *Before << "\n");
1183 Inst->moveBefore(Before);
1186 /// \brief Move the instruction back to its original position.
1187 void undo() override {
1188 DEBUG(dbgs() << "Undo: moveBefore: " << *Inst << "\n");
1189 Position.insert(Inst);
1193 /// \brief Set the operand of an instruction with a new value.
1194 class OperandSetter : public TypePromotionAction {
1195 /// Original operand of the instruction.
1197 /// Index of the modified instruction.
1201 /// \brief Set \p Idx operand of \p Inst with \p NewVal.
1202 OperandSetter(Instruction *Inst, unsigned Idx, Value *NewVal)
1203 : TypePromotionAction(Inst), Idx(Idx) {
1204 DEBUG(dbgs() << "Do: setOperand: " << Idx << "\n"
1205 << "for:" << *Inst << "\n"
1206 << "with:" << *NewVal << "\n");
1207 Origin = Inst->getOperand(Idx);
1208 Inst->setOperand(Idx, NewVal);
1211 /// \brief Restore the original value of the instruction.
1212 void undo() override {
1213 DEBUG(dbgs() << "Undo: setOperand:" << Idx << "\n"
1214 << "for: " << *Inst << "\n"
1215 << "with: " << *Origin << "\n");
1216 Inst->setOperand(Idx, Origin);
1220 /// \brief Hide the operands of an instruction.
1221 /// Do as if this instruction was not using any of its operands.
1222 class OperandsHider : public TypePromotionAction {
1223 /// The list of original operands.
1224 SmallVector<Value *, 4> OriginalValues;
1227 /// \brief Remove \p Inst from the uses of the operands of \p Inst.
1228 OperandsHider(Instruction *Inst) : TypePromotionAction(Inst) {
1229 DEBUG(dbgs() << "Do: OperandsHider: " << *Inst << "\n");
1230 unsigned NumOpnds = Inst->getNumOperands();
1231 OriginalValues.reserve(NumOpnds);
1232 for (unsigned It = 0; It < NumOpnds; ++It) {
1233 // Save the current operand.
1234 Value *Val = Inst->getOperand(It);
1235 OriginalValues.push_back(Val);
1237 // We could use OperandSetter here, but that would implied an overhead
1238 // that we are not willing to pay.
1239 Inst->setOperand(It, UndefValue::get(Val->getType()));
1243 /// \brief Restore the original list of uses.
1244 void undo() override {
1245 DEBUG(dbgs() << "Undo: OperandsHider: " << *Inst << "\n");
1246 for (unsigned It = 0, EndIt = OriginalValues.size(); It != EndIt; ++It)
1247 Inst->setOperand(It, OriginalValues[It]);
1251 /// \brief Build a truncate instruction.
1252 class TruncBuilder : public TypePromotionAction {
1254 /// \brief Build a truncate instruction of \p Opnd producing a \p Ty
1256 /// trunc Opnd to Ty.
1257 TruncBuilder(Instruction *Opnd, Type *Ty) : TypePromotionAction(Opnd) {
1258 IRBuilder<> Builder(Opnd);
1259 Inst = cast<Instruction>(Builder.CreateTrunc(Opnd, Ty, "promoted"));
1260 DEBUG(dbgs() << "Do: TruncBuilder: " << *Inst << "\n");
1263 /// \brief Get the built instruction.
1264 Instruction *getBuiltInstruction() { return Inst; }
1266 /// \brief Remove the built instruction.
1267 void undo() override {
1268 DEBUG(dbgs() << "Undo: TruncBuilder: " << *Inst << "\n");
1269 Inst->eraseFromParent();
1273 /// \brief Build a sign extension instruction.
1274 class SExtBuilder : public TypePromotionAction {
1276 /// \brief Build a sign extension instruction of \p Opnd producing a \p Ty
1278 /// sext Opnd to Ty.
1279 SExtBuilder(Instruction *InsertPt, Value *Opnd, Type *Ty)
1280 : TypePromotionAction(Inst) {
1281 IRBuilder<> Builder(InsertPt);
1282 Inst = cast<Instruction>(Builder.CreateSExt(Opnd, Ty, "promoted"));
1283 DEBUG(dbgs() << "Do: SExtBuilder: " << *Inst << "\n");
1286 /// \brief Get the built instruction.
1287 Instruction *getBuiltInstruction() { return Inst; }
1289 /// \brief Remove the built instruction.
1290 void undo() override {
1291 DEBUG(dbgs() << "Undo: SExtBuilder: " << *Inst << "\n");
1292 Inst->eraseFromParent();
1296 /// \brief Mutate an instruction to another type.
1297 class TypeMutator : public TypePromotionAction {
1298 /// Record the original type.
1302 /// \brief Mutate the type of \p Inst into \p NewTy.
1303 TypeMutator(Instruction *Inst, Type *NewTy)
1304 : TypePromotionAction(Inst), OrigTy(Inst->getType()) {
1305 DEBUG(dbgs() << "Do: MutateType: " << *Inst << " with " << *NewTy
1307 Inst->mutateType(NewTy);
1310 /// \brief Mutate the instruction back to its original type.
1311 void undo() override {
1312 DEBUG(dbgs() << "Undo: MutateType: " << *Inst << " with " << *OrigTy
1314 Inst->mutateType(OrigTy);
1318 /// \brief Replace the uses of an instruction by another instruction.
1319 class UsesReplacer : public TypePromotionAction {
1320 /// Helper structure to keep track of the replaced uses.
1321 struct InstructionAndIdx {
1322 /// The instruction using the instruction.
1324 /// The index where this instruction is used for Inst.
1326 InstructionAndIdx(Instruction *Inst, unsigned Idx)
1327 : Inst(Inst), Idx(Idx) {}
1330 /// Keep track of the original uses (pair Instruction, Index).
1331 SmallVector<InstructionAndIdx, 4> OriginalUses;
1332 typedef SmallVectorImpl<InstructionAndIdx>::iterator use_iterator;
1335 /// \brief Replace all the use of \p Inst by \p New.
1336 UsesReplacer(Instruction *Inst, Value *New) : TypePromotionAction(Inst) {
1337 DEBUG(dbgs() << "Do: UsersReplacer: " << *Inst << " with " << *New
1339 // Record the original uses.
1340 for (Use &U : Inst->uses()) {
1341 Instruction *UserI = cast<Instruction>(U.getUser());
1342 OriginalUses.push_back(InstructionAndIdx(UserI, U.getOperandNo()));
1344 // Now, we can replace the uses.
1345 Inst->replaceAllUsesWith(New);
1348 /// \brief Reassign the original uses of Inst to Inst.
1349 void undo() override {
1350 DEBUG(dbgs() << "Undo: UsersReplacer: " << *Inst << "\n");
1351 for (use_iterator UseIt = OriginalUses.begin(),
1352 EndIt = OriginalUses.end();
1353 UseIt != EndIt; ++UseIt) {
1354 UseIt->Inst->setOperand(UseIt->Idx, Inst);
1359 /// \brief Remove an instruction from the IR.
1360 class InstructionRemover : public TypePromotionAction {
1361 /// Original position of the instruction.
1362 InsertionHandler Inserter;
1363 /// Helper structure to hide all the link to the instruction. In other
1364 /// words, this helps to do as if the instruction was removed.
1365 OperandsHider Hider;
1366 /// Keep track of the uses replaced, if any.
1367 UsesReplacer *Replacer;
1370 /// \brief Remove all reference of \p Inst and optinally replace all its
1372 /// \pre If !Inst->use_empty(), then New != nullptr
1373 InstructionRemover(Instruction *Inst, Value *New = nullptr)
1374 : TypePromotionAction(Inst), Inserter(Inst), Hider(Inst),
1377 Replacer = new UsesReplacer(Inst, New);
1378 DEBUG(dbgs() << "Do: InstructionRemover: " << *Inst << "\n");
1379 Inst->removeFromParent();
1382 ~InstructionRemover() { delete Replacer; }
1384 /// \brief Really remove the instruction.
1385 void commit() override { delete Inst; }
1387 /// \brief Resurrect the instruction and reassign it to the proper uses if
1388 /// new value was provided when build this action.
1389 void undo() override {
1390 DEBUG(dbgs() << "Undo: InstructionRemover: " << *Inst << "\n");
1391 Inserter.insert(Inst);
1399 /// Restoration point.
1400 /// The restoration point is a pointer to an action instead of an iterator
1401 /// because the iterator may be invalidated but not the pointer.
1402 typedef const TypePromotionAction *ConstRestorationPt;
1403 /// Advocate every changes made in that transaction.
1405 /// Undo all the changes made after the given point.
1406 void rollback(ConstRestorationPt Point);
1407 /// Get the current restoration point.
1408 ConstRestorationPt getRestorationPoint() const;
1410 /// \name API for IR modification with state keeping to support rollback.
1412 /// Same as Instruction::setOperand.
1413 void setOperand(Instruction *Inst, unsigned Idx, Value *NewVal);
1414 /// Same as Instruction::eraseFromParent.
1415 void eraseInstruction(Instruction *Inst, Value *NewVal = nullptr);
1416 /// Same as Value::replaceAllUsesWith.
1417 void replaceAllUsesWith(Instruction *Inst, Value *New);
1418 /// Same as Value::mutateType.
1419 void mutateType(Instruction *Inst, Type *NewTy);
1420 /// Same as IRBuilder::createTrunc.
1421 Instruction *createTrunc(Instruction *Opnd, Type *Ty);
1422 /// Same as IRBuilder::createSExt.
1423 Instruction *createSExt(Instruction *Inst, Value *Opnd, Type *Ty);
1424 /// Same as Instruction::moveBefore.
1425 void moveBefore(Instruction *Inst, Instruction *Before);
1429 /// The ordered list of actions made so far.
1430 SmallVector<std::unique_ptr<TypePromotionAction>, 16> Actions;
1431 typedef SmallVectorImpl<std::unique_ptr<TypePromotionAction>>::iterator CommitPt;
1434 void TypePromotionTransaction::setOperand(Instruction *Inst, unsigned Idx,
1437 make_unique<TypePromotionTransaction::OperandSetter>(Inst, Idx, NewVal));
1440 void TypePromotionTransaction::eraseInstruction(Instruction *Inst,
1443 make_unique<TypePromotionTransaction::InstructionRemover>(Inst, NewVal));
1446 void TypePromotionTransaction::replaceAllUsesWith(Instruction *Inst,
1448 Actions.push_back(make_unique<TypePromotionTransaction::UsesReplacer>(Inst, New));
1451 void TypePromotionTransaction::mutateType(Instruction *Inst, Type *NewTy) {
1452 Actions.push_back(make_unique<TypePromotionTransaction::TypeMutator>(Inst, NewTy));
1455 Instruction *TypePromotionTransaction::createTrunc(Instruction *Opnd,
1457 std::unique_ptr<TruncBuilder> Ptr(new TruncBuilder(Opnd, Ty));
1458 Instruction *I = Ptr->getBuiltInstruction();
1459 Actions.push_back(std::move(Ptr));
1463 Instruction *TypePromotionTransaction::createSExt(Instruction *Inst,
1464 Value *Opnd, Type *Ty) {
1465 std::unique_ptr<SExtBuilder> Ptr(new SExtBuilder(Inst, Opnd, Ty));
1466 Instruction *I = Ptr->getBuiltInstruction();
1467 Actions.push_back(std::move(Ptr));
1471 void TypePromotionTransaction::moveBefore(Instruction *Inst,
1472 Instruction *Before) {
1474 make_unique<TypePromotionTransaction::InstructionMoveBefore>(Inst, Before));
1477 TypePromotionTransaction::ConstRestorationPt
1478 TypePromotionTransaction::getRestorationPoint() const {
1479 return !Actions.empty() ? Actions.back().get() : nullptr;
1482 void TypePromotionTransaction::commit() {
1483 for (CommitPt It = Actions.begin(), EndIt = Actions.end(); It != EndIt;
1489 void TypePromotionTransaction::rollback(
1490 TypePromotionTransaction::ConstRestorationPt Point) {
1491 while (!Actions.empty() && Point != Actions.back().get()) {
1492 std::unique_ptr<TypePromotionAction> Curr = Actions.pop_back_val();
1497 /// \brief A helper class for matching addressing modes.
1499 /// This encapsulates the logic for matching the target-legal addressing modes.
1500 class AddressingModeMatcher {
1501 SmallVectorImpl<Instruction*> &AddrModeInsts;
1502 const TargetLowering &TLI;
1504 /// AccessTy/MemoryInst - This is the type for the access (e.g. double) and
1505 /// the memory instruction that we're computing this address for.
1507 Instruction *MemoryInst;
1509 /// AddrMode - This is the addressing mode that we're building up. This is
1510 /// part of the return value of this addressing mode matching stuff.
1511 ExtAddrMode &AddrMode;
1513 /// The truncate instruction inserted by other CodeGenPrepare optimizations.
1514 const SetOfInstrs &InsertedTruncs;
1515 /// A map from the instructions to their type before promotion.
1516 InstrToOrigTy &PromotedInsts;
1517 /// The ongoing transaction where every action should be registered.
1518 TypePromotionTransaction &TPT;
1520 /// IgnoreProfitability - This is set to true when we should not do
1521 /// profitability checks. When true, IsProfitableToFoldIntoAddressingMode
1522 /// always returns true.
1523 bool IgnoreProfitability;
1525 AddressingModeMatcher(SmallVectorImpl<Instruction*> &AMI,
1526 const TargetLowering &T, Type *AT,
1527 Instruction *MI, ExtAddrMode &AM,
1528 const SetOfInstrs &InsertedTruncs,
1529 InstrToOrigTy &PromotedInsts,
1530 TypePromotionTransaction &TPT)
1531 : AddrModeInsts(AMI), TLI(T), AccessTy(AT), MemoryInst(MI), AddrMode(AM),
1532 InsertedTruncs(InsertedTruncs), PromotedInsts(PromotedInsts), TPT(TPT) {
1533 IgnoreProfitability = false;
1537 /// Match - Find the maximal addressing mode that a load/store of V can fold,
1538 /// give an access type of AccessTy. This returns a list of involved
1539 /// instructions in AddrModeInsts.
1540 /// \p InsertedTruncs The truncate instruction inserted by other
1543 /// \p PromotedInsts maps the instructions to their type before promotion.
1544 /// \p The ongoing transaction where every action should be registered.
1545 static ExtAddrMode Match(Value *V, Type *AccessTy,
1546 Instruction *MemoryInst,
1547 SmallVectorImpl<Instruction*> &AddrModeInsts,
1548 const TargetLowering &TLI,
1549 const SetOfInstrs &InsertedTruncs,
1550 InstrToOrigTy &PromotedInsts,
1551 TypePromotionTransaction &TPT) {
1554 bool Success = AddressingModeMatcher(AddrModeInsts, TLI, AccessTy,
1555 MemoryInst, Result, InsertedTruncs,
1556 PromotedInsts, TPT).MatchAddr(V, 0);
1557 (void)Success; assert(Success && "Couldn't select *anything*?");
1561 bool MatchScaledValue(Value *ScaleReg, int64_t Scale, unsigned Depth);
1562 bool MatchAddr(Value *V, unsigned Depth);
1563 bool MatchOperationAddr(User *Operation, unsigned Opcode, unsigned Depth,
1564 bool *MovedAway = nullptr);
1565 bool IsProfitableToFoldIntoAddressingMode(Instruction *I,
1566 ExtAddrMode &AMBefore,
1567 ExtAddrMode &AMAfter);
1568 bool ValueAlreadyLiveAtInst(Value *Val, Value *KnownLive1, Value *KnownLive2);
1569 bool IsPromotionProfitable(unsigned MatchedSize, unsigned SizeWithPromotion,
1570 Value *PromotedOperand) const;
1573 /// MatchScaledValue - Try adding ScaleReg*Scale to the current addressing mode.
1574 /// Return true and update AddrMode if this addr mode is legal for the target,
1576 bool AddressingModeMatcher::MatchScaledValue(Value *ScaleReg, int64_t Scale,
1578 // If Scale is 1, then this is the same as adding ScaleReg to the addressing
1579 // mode. Just process that directly.
1581 return MatchAddr(ScaleReg, Depth);
1583 // If the scale is 0, it takes nothing to add this.
1587 // If we already have a scale of this value, we can add to it, otherwise, we
1588 // need an available scale field.
1589 if (AddrMode.Scale != 0 && AddrMode.ScaledReg != ScaleReg)
1592 ExtAddrMode TestAddrMode = AddrMode;
1594 // Add scale to turn X*4+X*3 -> X*7. This could also do things like
1595 // [A+B + A*7] -> [B+A*8].
1596 TestAddrMode.Scale += Scale;
1597 TestAddrMode.ScaledReg = ScaleReg;
1599 // If the new address isn't legal, bail out.
1600 if (!TLI.isLegalAddressingMode(TestAddrMode, AccessTy))
1603 // It was legal, so commit it.
1604 AddrMode = TestAddrMode;
1606 // Okay, we decided that we can add ScaleReg+Scale to AddrMode. Check now
1607 // to see if ScaleReg is actually X+C. If so, we can turn this into adding
1608 // X*Scale + C*Scale to addr mode.
1609 ConstantInt *CI = nullptr; Value *AddLHS = nullptr;
1610 if (isa<Instruction>(ScaleReg) && // not a constant expr.
1611 match(ScaleReg, m_Add(m_Value(AddLHS), m_ConstantInt(CI)))) {
1612 TestAddrMode.ScaledReg = AddLHS;
1613 TestAddrMode.BaseOffs += CI->getSExtValue()*TestAddrMode.Scale;
1615 // If this addressing mode is legal, commit it and remember that we folded
1616 // this instruction.
1617 if (TLI.isLegalAddressingMode(TestAddrMode, AccessTy)) {
1618 AddrModeInsts.push_back(cast<Instruction>(ScaleReg));
1619 AddrMode = TestAddrMode;
1624 // Otherwise, not (x+c)*scale, just return what we have.
1628 /// MightBeFoldableInst - This is a little filter, which returns true if an
1629 /// addressing computation involving I might be folded into a load/store
1630 /// accessing it. This doesn't need to be perfect, but needs to accept at least
1631 /// the set of instructions that MatchOperationAddr can.
1632 static bool MightBeFoldableInst(Instruction *I) {
1633 switch (I->getOpcode()) {
1634 case Instruction::BitCast:
1635 case Instruction::AddrSpaceCast:
1636 // Don't touch identity bitcasts.
1637 if (I->getType() == I->getOperand(0)->getType())
1639 return I->getType()->isPointerTy() || I->getType()->isIntegerTy();
1640 case Instruction::PtrToInt:
1641 // PtrToInt is always a noop, as we know that the int type is pointer sized.
1643 case Instruction::IntToPtr:
1644 // We know the input is intptr_t, so this is foldable.
1646 case Instruction::Add:
1648 case Instruction::Mul:
1649 case Instruction::Shl:
1650 // Can only handle X*C and X << C.
1651 return isa<ConstantInt>(I->getOperand(1));
1652 case Instruction::GetElementPtr:
1659 /// \brief Hepler class to perform type promotion.
1660 class TypePromotionHelper {
1661 /// \brief Utility function to check whether or not a sign extension of
1662 /// \p Inst with \p ConsideredSExtType can be moved through \p Inst by either
1663 /// using the operands of \p Inst or promoting \p Inst.
1664 /// In other words, check if:
1665 /// sext (Ty Inst opnd1 opnd2 ... opndN) to ConsideredSExtType.
1666 /// #1 Promotion applies:
1667 /// ConsideredSExtType Inst (sext opnd1 to ConsideredSExtType, ...).
1668 /// #2 Operand reuses:
1669 /// sext opnd1 to ConsideredSExtType.
1670 /// \p PromotedInsts maps the instructions to their type before promotion.
1671 static bool canGetThrough(const Instruction *Inst, Type *ConsideredSExtType,
1672 const InstrToOrigTy &PromotedInsts);
1674 /// \brief Utility function to determine if \p OpIdx should be promoted when
1675 /// promoting \p Inst.
1676 static bool shouldSExtOperand(const Instruction *Inst, int OpIdx) {
1677 if (isa<SelectInst>(Inst) && OpIdx == 0)
1682 /// \brief Utility function to promote the operand of \p SExt when this
1683 /// operand is a promotable trunc or sext.
1684 /// \p PromotedInsts maps the instructions to their type before promotion.
1685 /// \p CreatedInsts[out] contains how many non-free instructions have been
1686 /// created to promote the operand of SExt.
1687 /// Should never be called directly.
1688 /// \return The promoted value which is used instead of SExt.
1689 static Value *promoteOperandForTruncAndSExt(Instruction *SExt,
1690 TypePromotionTransaction &TPT,
1691 InstrToOrigTy &PromotedInsts,
1692 unsigned &CreatedInsts);
1694 /// \brief Utility function to promote the operand of \p SExt when this
1695 /// operand is promotable and is not a supported trunc or sext.
1696 /// \p PromotedInsts maps the instructions to their type before promotion.
1697 /// \p CreatedInsts[out] contains how many non-free instructions have been
1698 /// created to promote the operand of SExt.
1699 /// Should never be called directly.
1700 /// \return The promoted value which is used instead of SExt.
1701 static Value *promoteOperandForOther(Instruction *SExt,
1702 TypePromotionTransaction &TPT,
1703 InstrToOrigTy &PromotedInsts,
1704 unsigned &CreatedInsts);
1707 /// Type for the utility function that promotes the operand of SExt.
1708 typedef Value *(*Action)(Instruction *SExt, TypePromotionTransaction &TPT,
1709 InstrToOrigTy &PromotedInsts,
1710 unsigned &CreatedInsts);
1711 /// \brief Given a sign extend instruction \p SExt, return the approriate
1712 /// action to promote the operand of \p SExt instead of using SExt.
1713 /// \return NULL if no promotable action is possible with the current
1715 /// \p InsertedTruncs keeps track of all the truncate instructions inserted by
1716 /// the others CodeGenPrepare optimizations. This information is important
1717 /// because we do not want to promote these instructions as CodeGenPrepare
1718 /// will reinsert them later. Thus creating an infinite loop: create/remove.
1719 /// \p PromotedInsts maps the instructions to their type before promotion.
1720 static Action getAction(Instruction *SExt, const SetOfInstrs &InsertedTruncs,
1721 const TargetLowering &TLI,
1722 const InstrToOrigTy &PromotedInsts);
1725 bool TypePromotionHelper::canGetThrough(const Instruction *Inst,
1726 Type *ConsideredSExtType,
1727 const InstrToOrigTy &PromotedInsts) {
1728 // We can always get through sext.
1729 if (isa<SExtInst>(Inst))
1732 // We can get through binary operator, if it is legal. In other words, the
1733 // binary operator must have a nuw or nsw flag.
1734 const BinaryOperator *BinOp = dyn_cast<BinaryOperator>(Inst);
1735 if (BinOp && isa<OverflowingBinaryOperator>(BinOp) &&
1736 (BinOp->hasNoUnsignedWrap() || BinOp->hasNoSignedWrap()))
1739 // Check if we can do the following simplification.
1740 // sext(trunc(sext)) --> sext
1741 if (!isa<TruncInst>(Inst))
1744 Value *OpndVal = Inst->getOperand(0);
1745 // Check if we can use this operand in the sext.
1746 // If the type is larger than the result type of the sign extension,
1748 if (OpndVal->getType()->getIntegerBitWidth() >
1749 ConsideredSExtType->getIntegerBitWidth())
1752 // If the operand of the truncate is not an instruction, we will not have
1753 // any information on the dropped bits.
1754 // (Actually we could for constant but it is not worth the extra logic).
1755 Instruction *Opnd = dyn_cast<Instruction>(OpndVal);
1759 // Check if the source of the type is narrow enough.
1760 // I.e., check that trunc just drops sign extended bits.
1761 // #1 get the type of the operand.
1762 const Type *OpndType;
1763 InstrToOrigTy::const_iterator It = PromotedInsts.find(Opnd);
1764 if (It != PromotedInsts.end())
1765 OpndType = It->second;
1766 else if (isa<SExtInst>(Opnd))
1767 OpndType = cast<Instruction>(Opnd)->getOperand(0)->getType();
1771 // #2 check that the truncate just drop sign extended bits.
1772 if (Inst->getType()->getIntegerBitWidth() >= OpndType->getIntegerBitWidth())
1778 TypePromotionHelper::Action TypePromotionHelper::getAction(
1779 Instruction *SExt, const SetOfInstrs &InsertedTruncs,
1780 const TargetLowering &TLI, const InstrToOrigTy &PromotedInsts) {
1781 Instruction *SExtOpnd = dyn_cast<Instruction>(SExt->getOperand(0));
1782 Type *SExtTy = SExt->getType();
1783 // If the operand of the sign extension is not an instruction, we cannot
1785 // If it, check we can get through.
1786 if (!SExtOpnd || !canGetThrough(SExtOpnd, SExtTy, PromotedInsts))
1789 // Do not promote if the operand has been added by codegenprepare.
1790 // Otherwise, it means we are undoing an optimization that is likely to be
1791 // redone, thus causing potential infinite loop.
1792 if (isa<TruncInst>(SExtOpnd) && InsertedTruncs.count(SExtOpnd))
1795 // SExt or Trunc instructions.
1796 // Return the related handler.
1797 if (isa<SExtInst>(SExtOpnd) || isa<TruncInst>(SExtOpnd))
1798 return promoteOperandForTruncAndSExt;
1800 // Regular instruction.
1801 // Abort early if we will have to insert non-free instructions.
1802 if (!SExtOpnd->hasOneUse() &&
1803 !TLI.isTruncateFree(SExtTy, SExtOpnd->getType()))
1805 return promoteOperandForOther;
1808 Value *TypePromotionHelper::promoteOperandForTruncAndSExt(
1809 llvm::Instruction *SExt, TypePromotionTransaction &TPT,
1810 InstrToOrigTy &PromotedInsts, unsigned &CreatedInsts) {
1811 // By construction, the operand of SExt is an instruction. Otherwise we cannot
1812 // get through it and this method should not be called.
1813 Instruction *SExtOpnd = cast<Instruction>(SExt->getOperand(0));
1814 // Replace sext(trunc(opnd)) or sext(sext(opnd))
1816 TPT.setOperand(SExt, 0, SExtOpnd->getOperand(0));
1819 // Remove dead code.
1820 if (SExtOpnd->use_empty())
1821 TPT.eraseInstruction(SExtOpnd);
1823 // Check if the sext is still needed.
1824 if (SExt->getType() != SExt->getOperand(0)->getType())
1827 // At this point we have: sext ty opnd to ty.
1828 // Reassign the uses of SExt to the opnd and remove SExt.
1829 Value *NextVal = SExt->getOperand(0);
1830 TPT.eraseInstruction(SExt, NextVal);
1835 TypePromotionHelper::promoteOperandForOther(Instruction *SExt,
1836 TypePromotionTransaction &TPT,
1837 InstrToOrigTy &PromotedInsts,
1838 unsigned &CreatedInsts) {
1839 // By construction, the operand of SExt is an instruction. Otherwise we cannot
1840 // get through it and this method should not be called.
1841 Instruction *SExtOpnd = cast<Instruction>(SExt->getOperand(0));
1843 if (!SExtOpnd->hasOneUse()) {
1844 // SExtOpnd will be promoted.
1845 // All its uses, but SExt, will need to use a truncated value of the
1846 // promoted version.
1847 // Create the truncate now.
1848 Instruction *Trunc = TPT.createTrunc(SExt, SExtOpnd->getType());
1849 Trunc->removeFromParent();
1850 // Insert it just after the definition.
1851 Trunc->insertAfter(SExtOpnd);
1853 TPT.replaceAllUsesWith(SExtOpnd, Trunc);
1854 // Restore the operand of SExt (which has been replace by the previous call
1855 // to replaceAllUsesWith) to avoid creating a cycle trunc <-> sext.
1856 TPT.setOperand(SExt, 0, SExtOpnd);
1859 // Get through the Instruction:
1860 // 1. Update its type.
1861 // 2. Replace the uses of SExt by Inst.
1862 // 3. Sign extend each operand that needs to be sign extended.
1864 // Remember the original type of the instruction before promotion.
1865 // This is useful to know that the high bits are sign extended bits.
1866 PromotedInsts.insert(
1867 std::pair<Instruction *, Type *>(SExtOpnd, SExtOpnd->getType()));
1869 TPT.mutateType(SExtOpnd, SExt->getType());
1871 TPT.replaceAllUsesWith(SExt, SExtOpnd);
1873 Instruction *SExtForOpnd = SExt;
1875 DEBUG(dbgs() << "Propagate SExt to operands\n");
1876 for (int OpIdx = 0, EndOpIdx = SExtOpnd->getNumOperands(); OpIdx != EndOpIdx;
1878 DEBUG(dbgs() << "Operand:\n" << *(SExtOpnd->getOperand(OpIdx)) << '\n');
1879 if (SExtOpnd->getOperand(OpIdx)->getType() == SExt->getType() ||
1880 !shouldSExtOperand(SExtOpnd, OpIdx)) {
1881 DEBUG(dbgs() << "No need to propagate\n");
1884 // Check if we can statically sign extend the operand.
1885 Value *Opnd = SExtOpnd->getOperand(OpIdx);
1886 if (const ConstantInt *Cst = dyn_cast<ConstantInt>(Opnd)) {
1887 DEBUG(dbgs() << "Statically sign extend\n");
1890 ConstantInt::getSigned(SExt->getType(), Cst->getSExtValue()));
1893 // UndefValue are typed, so we have to statically sign extend them.
1894 if (isa<UndefValue>(Opnd)) {
1895 DEBUG(dbgs() << "Statically sign extend\n");
1896 TPT.setOperand(SExtOpnd, OpIdx, UndefValue::get(SExt->getType()));
1900 // Otherwise we have to explicity sign extend the operand.
1901 // Check if SExt was reused to sign extend an operand.
1903 // If yes, create a new one.
1904 DEBUG(dbgs() << "More operands to sext\n");
1905 SExtForOpnd = TPT.createSExt(SExt, Opnd, SExt->getType());
1909 TPT.setOperand(SExtForOpnd, 0, Opnd);
1911 // Move the sign extension before the insertion point.
1912 TPT.moveBefore(SExtForOpnd, SExtOpnd);
1913 TPT.setOperand(SExtOpnd, OpIdx, SExtForOpnd);
1914 // If more sext are required, new instructions will have to be created.
1915 SExtForOpnd = nullptr;
1917 if (SExtForOpnd == SExt) {
1918 DEBUG(dbgs() << "Sign extension is useless now\n");
1919 TPT.eraseInstruction(SExt);
1924 /// IsPromotionProfitable - Check whether or not promoting an instruction
1925 /// to a wider type was profitable.
1926 /// \p MatchedSize gives the number of instructions that have been matched
1927 /// in the addressing mode after the promotion was applied.
1928 /// \p SizeWithPromotion gives the number of created instructions for
1929 /// the promotion plus the number of instructions that have been
1930 /// matched in the addressing mode before the promotion.
1931 /// \p PromotedOperand is the value that has been promoted.
1932 /// \return True if the promotion is profitable, false otherwise.
1934 AddressingModeMatcher::IsPromotionProfitable(unsigned MatchedSize,
1935 unsigned SizeWithPromotion,
1936 Value *PromotedOperand) const {
1937 // We folded less instructions than what we created to promote the operand.
1938 // This is not profitable.
1939 if (MatchedSize < SizeWithPromotion)
1941 if (MatchedSize > SizeWithPromotion)
1943 // The promotion is neutral but it may help folding the sign extension in
1944 // loads for instance.
1945 // Check that we did not create an illegal instruction.
1946 Instruction *PromotedInst = dyn_cast<Instruction>(PromotedOperand);
1949 int ISDOpcode = TLI.InstructionOpcodeToISD(PromotedInst->getOpcode());
1950 // If the ISDOpcode is undefined, it was undefined before the promotion.
1953 // Otherwise, check if the promoted instruction is legal or not.
1954 return TLI.isOperationLegalOrCustom(ISDOpcode,
1955 EVT::getEVT(PromotedInst->getType()));
1958 /// MatchOperationAddr - Given an instruction or constant expr, see if we can
1959 /// fold the operation into the addressing mode. If so, update the addressing
1960 /// mode and return true, otherwise return false without modifying AddrMode.
1961 /// If \p MovedAway is not NULL, it contains the information of whether or
1962 /// not AddrInst has to be folded into the addressing mode on success.
1963 /// If \p MovedAway == true, \p AddrInst will not be part of the addressing
1964 /// because it has been moved away.
1965 /// Thus AddrInst must not be added in the matched instructions.
1966 /// This state can happen when AddrInst is a sext, since it may be moved away.
1967 /// Therefore, AddrInst may not be valid when MovedAway is true and it must
1968 /// not be referenced anymore.
1969 bool AddressingModeMatcher::MatchOperationAddr(User *AddrInst, unsigned Opcode,
1972 // Avoid exponential behavior on extremely deep expression trees.
1973 if (Depth >= 5) return false;
1975 // By default, all matched instructions stay in place.
1980 case Instruction::PtrToInt:
1981 // PtrToInt is always a noop, as we know that the int type is pointer sized.
1982 return MatchAddr(AddrInst->getOperand(0), Depth);
1983 case Instruction::IntToPtr:
1984 // This inttoptr is a no-op if the integer type is pointer sized.
1985 if (TLI.getValueType(AddrInst->getOperand(0)->getType()) ==
1986 TLI.getPointerTy(AddrInst->getType()->getPointerAddressSpace()))
1987 return MatchAddr(AddrInst->getOperand(0), Depth);
1989 case Instruction::BitCast:
1990 case Instruction::AddrSpaceCast:
1991 // BitCast is always a noop, and we can handle it as long as it is
1992 // int->int or pointer->pointer (we don't want int<->fp or something).
1993 if ((AddrInst->getOperand(0)->getType()->isPointerTy() ||
1994 AddrInst->getOperand(0)->getType()->isIntegerTy()) &&
1995 // Don't touch identity bitcasts. These were probably put here by LSR,
1996 // and we don't want to mess around with them. Assume it knows what it
1998 AddrInst->getOperand(0)->getType() != AddrInst->getType())
1999 return MatchAddr(AddrInst->getOperand(0), Depth);
2001 case Instruction::Add: {
2002 // Check to see if we can merge in the RHS then the LHS. If so, we win.
2003 ExtAddrMode BackupAddrMode = AddrMode;
2004 unsigned OldSize = AddrModeInsts.size();
2005 // Start a transaction at this point.
2006 // The LHS may match but not the RHS.
2007 // Therefore, we need a higher level restoration point to undo partially
2008 // matched operation.
2009 TypePromotionTransaction::ConstRestorationPt LastKnownGood =
2010 TPT.getRestorationPoint();
2012 if (MatchAddr(AddrInst->getOperand(1), Depth+1) &&
2013 MatchAddr(AddrInst->getOperand(0), Depth+1))
2016 // Restore the old addr mode info.
2017 AddrMode = BackupAddrMode;
2018 AddrModeInsts.resize(OldSize);
2019 TPT.rollback(LastKnownGood);
2021 // Otherwise this was over-aggressive. Try merging in the LHS then the RHS.
2022 if (MatchAddr(AddrInst->getOperand(0), Depth+1) &&
2023 MatchAddr(AddrInst->getOperand(1), Depth+1))
2026 // Otherwise we definitely can't merge the ADD in.
2027 AddrMode = BackupAddrMode;
2028 AddrModeInsts.resize(OldSize);
2029 TPT.rollback(LastKnownGood);
2032 //case Instruction::Or:
2033 // TODO: We can handle "Or Val, Imm" iff this OR is equivalent to an ADD.
2035 case Instruction::Mul:
2036 case Instruction::Shl: {
2037 // Can only handle X*C and X << C.
2038 ConstantInt *RHS = dyn_cast<ConstantInt>(AddrInst->getOperand(1));
2039 if (!RHS) return false;
2040 int64_t Scale = RHS->getSExtValue();
2041 if (Opcode == Instruction::Shl)
2042 Scale = 1LL << Scale;
2044 return MatchScaledValue(AddrInst->getOperand(0), Scale, Depth);
2046 case Instruction::GetElementPtr: {
2047 // Scan the GEP. We check it if it contains constant offsets and at most
2048 // one variable offset.
2049 int VariableOperand = -1;
2050 unsigned VariableScale = 0;
2052 int64_t ConstantOffset = 0;
2053 const DataLayout *TD = TLI.getDataLayout();
2054 gep_type_iterator GTI = gep_type_begin(AddrInst);
2055 for (unsigned i = 1, e = AddrInst->getNumOperands(); i != e; ++i, ++GTI) {
2056 if (StructType *STy = dyn_cast<StructType>(*GTI)) {
2057 const StructLayout *SL = TD->getStructLayout(STy);
2059 cast<ConstantInt>(AddrInst->getOperand(i))->getZExtValue();
2060 ConstantOffset += SL->getElementOffset(Idx);
2062 uint64_t TypeSize = TD->getTypeAllocSize(GTI.getIndexedType());
2063 if (ConstantInt *CI = dyn_cast<ConstantInt>(AddrInst->getOperand(i))) {
2064 ConstantOffset += CI->getSExtValue()*TypeSize;
2065 } else if (TypeSize) { // Scales of zero don't do anything.
2066 // We only allow one variable index at the moment.
2067 if (VariableOperand != -1)
2070 // Remember the variable index.
2071 VariableOperand = i;
2072 VariableScale = TypeSize;
2077 // A common case is for the GEP to only do a constant offset. In this case,
2078 // just add it to the disp field and check validity.
2079 if (VariableOperand == -1) {
2080 AddrMode.BaseOffs += ConstantOffset;
2081 if (ConstantOffset == 0 || TLI.isLegalAddressingMode(AddrMode, AccessTy)){
2082 // Check to see if we can fold the base pointer in too.
2083 if (MatchAddr(AddrInst->getOperand(0), Depth+1))
2086 AddrMode.BaseOffs -= ConstantOffset;
2090 // Save the valid addressing mode in case we can't match.
2091 ExtAddrMode BackupAddrMode = AddrMode;
2092 unsigned OldSize = AddrModeInsts.size();
2094 // See if the scale and offset amount is valid for this target.
2095 AddrMode.BaseOffs += ConstantOffset;
2097 // Match the base operand of the GEP.
2098 if (!MatchAddr(AddrInst->getOperand(0), Depth+1)) {
2099 // If it couldn't be matched, just stuff the value in a register.
2100 if (AddrMode.HasBaseReg) {
2101 AddrMode = BackupAddrMode;
2102 AddrModeInsts.resize(OldSize);
2105 AddrMode.HasBaseReg = true;
2106 AddrMode.BaseReg = AddrInst->getOperand(0);
2109 // Match the remaining variable portion of the GEP.
2110 if (!MatchScaledValue(AddrInst->getOperand(VariableOperand), VariableScale,
2112 // If it couldn't be matched, try stuffing the base into a register
2113 // instead of matching it, and retrying the match of the scale.
2114 AddrMode = BackupAddrMode;
2115 AddrModeInsts.resize(OldSize);
2116 if (AddrMode.HasBaseReg)
2118 AddrMode.HasBaseReg = true;
2119 AddrMode.BaseReg = AddrInst->getOperand(0);
2120 AddrMode.BaseOffs += ConstantOffset;
2121 if (!MatchScaledValue(AddrInst->getOperand(VariableOperand),
2122 VariableScale, Depth)) {
2123 // If even that didn't work, bail.
2124 AddrMode = BackupAddrMode;
2125 AddrModeInsts.resize(OldSize);
2132 case Instruction::SExt: {
2133 // Make sure this isn't a ConstantExpr (PR20314).
2134 Instruction *SExt = dyn_cast<Instruction>(AddrInst);
2135 if (!SExt) return false;
2137 // Try to move this sext out of the way of the addressing mode.
2138 // Ask for a method for doing so.
2139 TypePromotionHelper::Action TPH = TypePromotionHelper::getAction(
2140 SExt, InsertedTruncs, TLI, PromotedInsts);
2144 TypePromotionTransaction::ConstRestorationPt LastKnownGood =
2145 TPT.getRestorationPoint();
2146 unsigned CreatedInsts = 0;
2147 Value *PromotedOperand = TPH(SExt, TPT, PromotedInsts, CreatedInsts);
2148 // SExt has been moved away.
2149 // Thus either it will be rematched later in the recursive calls or it is
2150 // gone. Anyway, we must not fold it into the addressing mode at this point.
2154 // addr = gep base, idx
2156 // promotedOpnd = sext opnd <- no match here
2157 // op = promoted_add promotedOpnd, 1 <- match (later in recursive calls)
2158 // addr = gep base, op <- match
2162 assert(PromotedOperand &&
2163 "TypePromotionHelper should have filtered out those cases");
2165 ExtAddrMode BackupAddrMode = AddrMode;
2166 unsigned OldSize = AddrModeInsts.size();
2168 if (!MatchAddr(PromotedOperand, Depth) ||
2169 !IsPromotionProfitable(AddrModeInsts.size(), OldSize + CreatedInsts,
2171 AddrMode = BackupAddrMode;
2172 AddrModeInsts.resize(OldSize);
2173 DEBUG(dbgs() << "Sign extension does not pay off: rollback\n");
2174 TPT.rollback(LastKnownGood);
2183 /// MatchAddr - If we can, try to add the value of 'Addr' into the current
2184 /// addressing mode. If Addr can't be added to AddrMode this returns false and
2185 /// leaves AddrMode unmodified. This assumes that Addr is either a pointer type
2186 /// or intptr_t for the target.
2188 bool AddressingModeMatcher::MatchAddr(Value *Addr, unsigned Depth) {
2189 // Start a transaction at this point that we will rollback if the matching
2191 TypePromotionTransaction::ConstRestorationPt LastKnownGood =
2192 TPT.getRestorationPoint();
2193 if (ConstantInt *CI = dyn_cast<ConstantInt>(Addr)) {
2194 // Fold in immediates if legal for the target.
2195 AddrMode.BaseOffs += CI->getSExtValue();
2196 if (TLI.isLegalAddressingMode(AddrMode, AccessTy))
2198 AddrMode.BaseOffs -= CI->getSExtValue();
2199 } else if (GlobalValue *GV = dyn_cast<GlobalValue>(Addr)) {
2200 // If this is a global variable, try to fold it into the addressing mode.
2201 if (!AddrMode.BaseGV) {
2202 AddrMode.BaseGV = GV;
2203 if (TLI.isLegalAddressingMode(AddrMode, AccessTy))
2205 AddrMode.BaseGV = nullptr;
2207 } else if (Instruction *I = dyn_cast<Instruction>(Addr)) {
2208 ExtAddrMode BackupAddrMode = AddrMode;
2209 unsigned OldSize = AddrModeInsts.size();
2211 // Check to see if it is possible to fold this operation.
2212 bool MovedAway = false;
2213 if (MatchOperationAddr(I, I->getOpcode(), Depth, &MovedAway)) {
2214 // This instruction may have been move away. If so, there is nothing
2218 // Okay, it's possible to fold this. Check to see if it is actually
2219 // *profitable* to do so. We use a simple cost model to avoid increasing
2220 // register pressure too much.
2221 if (I->hasOneUse() ||
2222 IsProfitableToFoldIntoAddressingMode(I, BackupAddrMode, AddrMode)) {
2223 AddrModeInsts.push_back(I);
2227 // It isn't profitable to do this, roll back.
2228 //cerr << "NOT FOLDING: " << *I;
2229 AddrMode = BackupAddrMode;
2230 AddrModeInsts.resize(OldSize);
2231 TPT.rollback(LastKnownGood);
2233 } else if (ConstantExpr *CE = dyn_cast<ConstantExpr>(Addr)) {
2234 if (MatchOperationAddr(CE, CE->getOpcode(), Depth))
2236 TPT.rollback(LastKnownGood);
2237 } else if (isa<ConstantPointerNull>(Addr)) {
2238 // Null pointer gets folded without affecting the addressing mode.
2242 // Worse case, the target should support [reg] addressing modes. :)
2243 if (!AddrMode.HasBaseReg) {
2244 AddrMode.HasBaseReg = true;
2245 AddrMode.BaseReg = Addr;
2246 // Still check for legality in case the target supports [imm] but not [i+r].
2247 if (TLI.isLegalAddressingMode(AddrMode, AccessTy))
2249 AddrMode.HasBaseReg = false;
2250 AddrMode.BaseReg = nullptr;
2253 // If the base register is already taken, see if we can do [r+r].
2254 if (AddrMode.Scale == 0) {
2256 AddrMode.ScaledReg = Addr;
2257 if (TLI.isLegalAddressingMode(AddrMode, AccessTy))
2260 AddrMode.ScaledReg = nullptr;
2263 TPT.rollback(LastKnownGood);
2267 /// IsOperandAMemoryOperand - Check to see if all uses of OpVal by the specified
2268 /// inline asm call are due to memory operands. If so, return true, otherwise
2270 static bool IsOperandAMemoryOperand(CallInst *CI, InlineAsm *IA, Value *OpVal,
2271 const TargetLowering &TLI) {
2272 TargetLowering::AsmOperandInfoVector TargetConstraints = TLI.ParseConstraints(ImmutableCallSite(CI));
2273 for (unsigned i = 0, e = TargetConstraints.size(); i != e; ++i) {
2274 TargetLowering::AsmOperandInfo &OpInfo = TargetConstraints[i];
2276 // Compute the constraint code and ConstraintType to use.
2277 TLI.ComputeConstraintToUse(OpInfo, SDValue());
2279 // If this asm operand is our Value*, and if it isn't an indirect memory
2280 // operand, we can't fold it!
2281 if (OpInfo.CallOperandVal == OpVal &&
2282 (OpInfo.ConstraintType != TargetLowering::C_Memory ||
2283 !OpInfo.isIndirect))
2290 /// FindAllMemoryUses - Recursively walk all the uses of I until we find a
2291 /// memory use. If we find an obviously non-foldable instruction, return true.
2292 /// Add the ultimately found memory instructions to MemoryUses.
2293 static bool FindAllMemoryUses(Instruction *I,
2294 SmallVectorImpl<std::pair<Instruction*,unsigned> > &MemoryUses,
2295 SmallPtrSet<Instruction*, 16> &ConsideredInsts,
2296 const TargetLowering &TLI) {
2297 // If we already considered this instruction, we're done.
2298 if (!ConsideredInsts.insert(I))
2301 // If this is an obviously unfoldable instruction, bail out.
2302 if (!MightBeFoldableInst(I))
2305 // Loop over all the uses, recursively processing them.
2306 for (Use &U : I->uses()) {
2307 Instruction *UserI = cast<Instruction>(U.getUser());
2309 if (LoadInst *LI = dyn_cast<LoadInst>(UserI)) {
2310 MemoryUses.push_back(std::make_pair(LI, U.getOperandNo()));
2314 if (StoreInst *SI = dyn_cast<StoreInst>(UserI)) {
2315 unsigned opNo = U.getOperandNo();
2316 if (opNo == 0) return true; // Storing addr, not into addr.
2317 MemoryUses.push_back(std::make_pair(SI, opNo));
2321 if (CallInst *CI = dyn_cast<CallInst>(UserI)) {
2322 InlineAsm *IA = dyn_cast<InlineAsm>(CI->getCalledValue());
2323 if (!IA) return true;
2325 // If this is a memory operand, we're cool, otherwise bail out.
2326 if (!IsOperandAMemoryOperand(CI, IA, I, TLI))
2331 if (FindAllMemoryUses(UserI, MemoryUses, ConsideredInsts, TLI))
2338 /// ValueAlreadyLiveAtInst - Retrn true if Val is already known to be live at
2339 /// the use site that we're folding it into. If so, there is no cost to
2340 /// include it in the addressing mode. KnownLive1 and KnownLive2 are two values
2341 /// that we know are live at the instruction already.
2342 bool AddressingModeMatcher::ValueAlreadyLiveAtInst(Value *Val,Value *KnownLive1,
2343 Value *KnownLive2) {
2344 // If Val is either of the known-live values, we know it is live!
2345 if (Val == nullptr || Val == KnownLive1 || Val == KnownLive2)
2348 // All values other than instructions and arguments (e.g. constants) are live.
2349 if (!isa<Instruction>(Val) && !isa<Argument>(Val)) return true;
2351 // If Val is a constant sized alloca in the entry block, it is live, this is
2352 // true because it is just a reference to the stack/frame pointer, which is
2353 // live for the whole function.
2354 if (AllocaInst *AI = dyn_cast<AllocaInst>(Val))
2355 if (AI->isStaticAlloca())
2358 // Check to see if this value is already used in the memory instruction's
2359 // block. If so, it's already live into the block at the very least, so we
2360 // can reasonably fold it.
2361 return Val->isUsedInBasicBlock(MemoryInst->getParent());
2364 /// IsProfitableToFoldIntoAddressingMode - It is possible for the addressing
2365 /// mode of the machine to fold the specified instruction into a load or store
2366 /// that ultimately uses it. However, the specified instruction has multiple
2367 /// uses. Given this, it may actually increase register pressure to fold it
2368 /// into the load. For example, consider this code:
2372 /// use(Y) -> nonload/store
2376 /// In this case, Y has multiple uses, and can be folded into the load of Z
2377 /// (yielding load [X+2]). However, doing this will cause both "X" and "X+1" to
2378 /// be live at the use(Y) line. If we don't fold Y into load Z, we use one
2379 /// fewer register. Since Y can't be folded into "use(Y)" we don't increase the
2380 /// number of computations either.
2382 /// Note that this (like most of CodeGenPrepare) is just a rough heuristic. If
2383 /// X was live across 'load Z' for other reasons, we actually *would* want to
2384 /// fold the addressing mode in the Z case. This would make Y die earlier.
2385 bool AddressingModeMatcher::
2386 IsProfitableToFoldIntoAddressingMode(Instruction *I, ExtAddrMode &AMBefore,
2387 ExtAddrMode &AMAfter) {
2388 if (IgnoreProfitability) return true;
2390 // AMBefore is the addressing mode before this instruction was folded into it,
2391 // and AMAfter is the addressing mode after the instruction was folded. Get
2392 // the set of registers referenced by AMAfter and subtract out those
2393 // referenced by AMBefore: this is the set of values which folding in this
2394 // address extends the lifetime of.
2396 // Note that there are only two potential values being referenced here,
2397 // BaseReg and ScaleReg (global addresses are always available, as are any
2398 // folded immediates).
2399 Value *BaseReg = AMAfter.BaseReg, *ScaledReg = AMAfter.ScaledReg;
2401 // If the BaseReg or ScaledReg was referenced by the previous addrmode, their
2402 // lifetime wasn't extended by adding this instruction.
2403 if (ValueAlreadyLiveAtInst(BaseReg, AMBefore.BaseReg, AMBefore.ScaledReg))
2405 if (ValueAlreadyLiveAtInst(ScaledReg, AMBefore.BaseReg, AMBefore.ScaledReg))
2406 ScaledReg = nullptr;
2408 // If folding this instruction (and it's subexprs) didn't extend any live
2409 // ranges, we're ok with it.
2410 if (!BaseReg && !ScaledReg)
2413 // If all uses of this instruction are ultimately load/store/inlineasm's,
2414 // check to see if their addressing modes will include this instruction. If
2415 // so, we can fold it into all uses, so it doesn't matter if it has multiple
2417 SmallVector<std::pair<Instruction*,unsigned>, 16> MemoryUses;
2418 SmallPtrSet<Instruction*, 16> ConsideredInsts;
2419 if (FindAllMemoryUses(I, MemoryUses, ConsideredInsts, TLI))
2420 return false; // Has a non-memory, non-foldable use!
2422 // Now that we know that all uses of this instruction are part of a chain of
2423 // computation involving only operations that could theoretically be folded
2424 // into a memory use, loop over each of these uses and see if they could
2425 // *actually* fold the instruction.
2426 SmallVector<Instruction*, 32> MatchedAddrModeInsts;
2427 for (unsigned i = 0, e = MemoryUses.size(); i != e; ++i) {
2428 Instruction *User = MemoryUses[i].first;
2429 unsigned OpNo = MemoryUses[i].second;
2431 // Get the access type of this use. If the use isn't a pointer, we don't
2432 // know what it accesses.
2433 Value *Address = User->getOperand(OpNo);
2434 if (!Address->getType()->isPointerTy())
2436 Type *AddressAccessTy = Address->getType()->getPointerElementType();
2438 // Do a match against the root of this address, ignoring profitability. This
2439 // will tell us if the addressing mode for the memory operation will
2440 // *actually* cover the shared instruction.
2442 TypePromotionTransaction::ConstRestorationPt LastKnownGood =
2443 TPT.getRestorationPoint();
2444 AddressingModeMatcher Matcher(MatchedAddrModeInsts, TLI, AddressAccessTy,
2445 MemoryInst, Result, InsertedTruncs,
2446 PromotedInsts, TPT);
2447 Matcher.IgnoreProfitability = true;
2448 bool Success = Matcher.MatchAddr(Address, 0);
2449 (void)Success; assert(Success && "Couldn't select *anything*?");
2451 // The match was to check the profitability, the changes made are not
2452 // part of the original matcher. Therefore, they should be dropped
2453 // otherwise the original matcher will not present the right state.
2454 TPT.rollback(LastKnownGood);
2456 // If the match didn't cover I, then it won't be shared by it.
2457 if (std::find(MatchedAddrModeInsts.begin(), MatchedAddrModeInsts.end(),
2458 I) == MatchedAddrModeInsts.end())
2461 MatchedAddrModeInsts.clear();
2467 } // end anonymous namespace
2469 /// IsNonLocalValue - Return true if the specified values are defined in a
2470 /// different basic block than BB.
2471 static bool IsNonLocalValue(Value *V, BasicBlock *BB) {
2472 if (Instruction *I = dyn_cast<Instruction>(V))
2473 return I->getParent() != BB;
2477 /// OptimizeMemoryInst - Load and Store Instructions often have
2478 /// addressing modes that can do significant amounts of computation. As such,
2479 /// instruction selection will try to get the load or store to do as much
2480 /// computation as possible for the program. The problem is that isel can only
2481 /// see within a single block. As such, we sink as much legal addressing mode
2482 /// stuff into the block as possible.
2484 /// This method is used to optimize both load/store and inline asms with memory
2486 bool CodeGenPrepare::OptimizeMemoryInst(Instruction *MemoryInst, Value *Addr,
2490 // Try to collapse single-value PHI nodes. This is necessary to undo
2491 // unprofitable PRE transformations.
2492 SmallVector<Value*, 8> worklist;
2493 SmallPtrSet<Value*, 16> Visited;
2494 worklist.push_back(Addr);
2496 // Use a worklist to iteratively look through PHI nodes, and ensure that
2497 // the addressing mode obtained from the non-PHI roots of the graph
2499 Value *Consensus = nullptr;
2500 unsigned NumUsesConsensus = 0;
2501 bool IsNumUsesConsensusValid = false;
2502 SmallVector<Instruction*, 16> AddrModeInsts;
2503 ExtAddrMode AddrMode;
2504 TypePromotionTransaction TPT;
2505 TypePromotionTransaction::ConstRestorationPt LastKnownGood =
2506 TPT.getRestorationPoint();
2507 while (!worklist.empty()) {
2508 Value *V = worklist.back();
2509 worklist.pop_back();
2511 // Break use-def graph loops.
2512 if (!Visited.insert(V)) {
2513 Consensus = nullptr;
2517 // For a PHI node, push all of its incoming values.
2518 if (PHINode *P = dyn_cast<PHINode>(V)) {
2519 for (unsigned i = 0, e = P->getNumIncomingValues(); i != e; ++i)
2520 worklist.push_back(P->getIncomingValue(i));
2524 // For non-PHIs, determine the addressing mode being computed.
2525 SmallVector<Instruction*, 16> NewAddrModeInsts;
2526 ExtAddrMode NewAddrMode = AddressingModeMatcher::Match(
2527 V, AccessTy, MemoryInst, NewAddrModeInsts, *TLI, InsertedTruncsSet,
2528 PromotedInsts, TPT);
2530 // This check is broken into two cases with very similar code to avoid using
2531 // getNumUses() as much as possible. Some values have a lot of uses, so
2532 // calling getNumUses() unconditionally caused a significant compile-time
2536 AddrMode = NewAddrMode;
2537 AddrModeInsts = NewAddrModeInsts;
2539 } else if (NewAddrMode == AddrMode) {
2540 if (!IsNumUsesConsensusValid) {
2541 NumUsesConsensus = Consensus->getNumUses();
2542 IsNumUsesConsensusValid = true;
2545 // Ensure that the obtained addressing mode is equivalent to that obtained
2546 // for all other roots of the PHI traversal. Also, when choosing one
2547 // such root as representative, select the one with the most uses in order
2548 // to keep the cost modeling heuristics in AddressingModeMatcher
2550 unsigned NumUses = V->getNumUses();
2551 if (NumUses > NumUsesConsensus) {
2553 NumUsesConsensus = NumUses;
2554 AddrModeInsts = NewAddrModeInsts;
2559 Consensus = nullptr;
2563 // If the addressing mode couldn't be determined, or if multiple different
2564 // ones were determined, bail out now.
2566 TPT.rollback(LastKnownGood);
2571 // Check to see if any of the instructions supersumed by this addr mode are
2572 // non-local to I's BB.
2573 bool AnyNonLocal = false;
2574 for (unsigned i = 0, e = AddrModeInsts.size(); i != e; ++i) {
2575 if (IsNonLocalValue(AddrModeInsts[i], MemoryInst->getParent())) {
2581 // If all the instructions matched are already in this BB, don't do anything.
2583 DEBUG(dbgs() << "CGP: Found local addrmode: " << AddrMode << "\n");
2587 // Insert this computation right after this user. Since our caller is
2588 // scanning from the top of the BB to the bottom, reuse of the expr are
2589 // guaranteed to happen later.
2590 IRBuilder<> Builder(MemoryInst);
2592 // Now that we determined the addressing expression we want to use and know
2593 // that we have to sink it into this block. Check to see if we have already
2594 // done this for some other load/store instr in this block. If so, reuse the
2596 Value *&SunkAddr = SunkAddrs[Addr];
2598 DEBUG(dbgs() << "CGP: Reusing nonlocal addrmode: " << AddrMode << " for "
2599 << *MemoryInst << "\n");
2600 if (SunkAddr->getType() != Addr->getType())
2601 SunkAddr = Builder.CreateBitCast(SunkAddr, Addr->getType());
2602 } else if (AddrSinkUsingGEPs || (!AddrSinkUsingGEPs.getNumOccurrences() &&
2603 TM && TM->getSubtarget<TargetSubtargetInfo>().useAA())) {
2604 // By default, we use the GEP-based method when AA is used later. This
2605 // prevents new inttoptr/ptrtoint pairs from degrading AA capabilities.
2606 DEBUG(dbgs() << "CGP: SINKING nonlocal addrmode: " << AddrMode << " for "
2607 << *MemoryInst << "\n");
2608 Type *IntPtrTy = TLI->getDataLayout()->getIntPtrType(Addr->getType());
2609 Value *ResultPtr = nullptr, *ResultIndex = nullptr;
2611 // First, find the pointer.
2612 if (AddrMode.BaseReg && AddrMode.BaseReg->getType()->isPointerTy()) {
2613 ResultPtr = AddrMode.BaseReg;
2614 AddrMode.BaseReg = nullptr;
2617 if (AddrMode.Scale && AddrMode.ScaledReg->getType()->isPointerTy()) {
2618 // We can't add more than one pointer together, nor can we scale a
2619 // pointer (both of which seem meaningless).
2620 if (ResultPtr || AddrMode.Scale != 1)
2623 ResultPtr = AddrMode.ScaledReg;
2627 if (AddrMode.BaseGV) {
2631 ResultPtr = AddrMode.BaseGV;
2634 // If the real base value actually came from an inttoptr, then the matcher
2635 // will look through it and provide only the integer value. In that case,
2637 if (!ResultPtr && AddrMode.BaseReg) {
2639 Builder.CreateIntToPtr(AddrMode.BaseReg, Addr->getType(), "sunkaddr");
2640 AddrMode.BaseReg = nullptr;
2641 } else if (!ResultPtr && AddrMode.Scale == 1) {
2643 Builder.CreateIntToPtr(AddrMode.ScaledReg, Addr->getType(), "sunkaddr");
2648 !AddrMode.BaseReg && !AddrMode.Scale && !AddrMode.BaseOffs) {
2649 SunkAddr = Constant::getNullValue(Addr->getType());
2650 } else if (!ResultPtr) {
2654 Builder.getInt8PtrTy(Addr->getType()->getPointerAddressSpace());
2656 // Start with the base register. Do this first so that subsequent address
2657 // matching finds it last, which will prevent it from trying to match it
2658 // as the scaled value in case it happens to be a mul. That would be
2659 // problematic if we've sunk a different mul for the scale, because then
2660 // we'd end up sinking both muls.
2661 if (AddrMode.BaseReg) {
2662 Value *V = AddrMode.BaseReg;
2663 if (V->getType() != IntPtrTy)
2664 V = Builder.CreateIntCast(V, IntPtrTy, /*isSigned=*/true, "sunkaddr");
2669 // Add the scale value.
2670 if (AddrMode.Scale) {
2671 Value *V = AddrMode.ScaledReg;
2672 if (V->getType() == IntPtrTy) {
2674 } else if (cast<IntegerType>(IntPtrTy)->getBitWidth() <
2675 cast<IntegerType>(V->getType())->getBitWidth()) {
2676 V = Builder.CreateTrunc(V, IntPtrTy, "sunkaddr");
2678 // It is only safe to sign extend the BaseReg if we know that the math
2679 // required to create it did not overflow before we extend it. Since
2680 // the original IR value was tossed in favor of a constant back when
2681 // the AddrMode was created we need to bail out gracefully if widths
2682 // do not match instead of extending it.
2683 Instruction *I = dyn_cast_or_null<Instruction>(ResultIndex);
2684 if (I && (ResultIndex != AddrMode.BaseReg))
2685 I->eraseFromParent();
2689 if (AddrMode.Scale != 1)
2690 V = Builder.CreateMul(V, ConstantInt::get(IntPtrTy, AddrMode.Scale),
2693 ResultIndex = Builder.CreateAdd(ResultIndex, V, "sunkaddr");
2698 // Add in the Base Offset if present.
2699 if (AddrMode.BaseOffs) {
2700 Value *V = ConstantInt::get(IntPtrTy, AddrMode.BaseOffs);
2702 // We need to add this separately from the scale above to help with
2703 // SDAG consecutive load/store merging.
2704 if (ResultPtr->getType() != I8PtrTy)
2705 ResultPtr = Builder.CreateBitCast(ResultPtr, I8PtrTy);
2706 ResultPtr = Builder.CreateGEP(ResultPtr, ResultIndex, "sunkaddr");
2713 SunkAddr = ResultPtr;
2715 if (ResultPtr->getType() != I8PtrTy)
2716 ResultPtr = Builder.CreateBitCast(ResultPtr, I8PtrTy);
2717 SunkAddr = Builder.CreateGEP(ResultPtr, ResultIndex, "sunkaddr");
2720 if (SunkAddr->getType() != Addr->getType())
2721 SunkAddr = Builder.CreateBitCast(SunkAddr, Addr->getType());
2724 DEBUG(dbgs() << "CGP: SINKING nonlocal addrmode: " << AddrMode << " for "
2725 << *MemoryInst << "\n");
2726 Type *IntPtrTy = TLI->getDataLayout()->getIntPtrType(Addr->getType());
2727 Value *Result = nullptr;
2729 // Start with the base register. Do this first so that subsequent address
2730 // matching finds it last, which will prevent it from trying to match it
2731 // as the scaled value in case it happens to be a mul. That would be
2732 // problematic if we've sunk a different mul for the scale, because then
2733 // we'd end up sinking both muls.
2734 if (AddrMode.BaseReg) {
2735 Value *V = AddrMode.BaseReg;
2736 if (V->getType()->isPointerTy())
2737 V = Builder.CreatePtrToInt(V, IntPtrTy, "sunkaddr");
2738 if (V->getType() != IntPtrTy)
2739 V = Builder.CreateIntCast(V, IntPtrTy, /*isSigned=*/true, "sunkaddr");
2743 // Add the scale value.
2744 if (AddrMode.Scale) {
2745 Value *V = AddrMode.ScaledReg;
2746 if (V->getType() == IntPtrTy) {
2748 } else if (V->getType()->isPointerTy()) {
2749 V = Builder.CreatePtrToInt(V, IntPtrTy, "sunkaddr");
2750 } else if (cast<IntegerType>(IntPtrTy)->getBitWidth() <
2751 cast<IntegerType>(V->getType())->getBitWidth()) {
2752 V = Builder.CreateTrunc(V, IntPtrTy, "sunkaddr");
2754 // It is only safe to sign extend the BaseReg if we know that the math
2755 // required to create it did not overflow before we extend it. Since
2756 // the original IR value was tossed in favor of a constant back when
2757 // the AddrMode was created we need to bail out gracefully if widths
2758 // do not match instead of extending it.
2759 Instruction *I = dyn_cast_or_null<Instruction>(Result);
2760 if (I && (Result != AddrMode.BaseReg))
2761 I->eraseFromParent();
2764 if (AddrMode.Scale != 1)
2765 V = Builder.CreateMul(V, ConstantInt::get(IntPtrTy, AddrMode.Scale),
2768 Result = Builder.CreateAdd(Result, V, "sunkaddr");
2773 // Add in the BaseGV if present.
2774 if (AddrMode.BaseGV) {
2775 Value *V = Builder.CreatePtrToInt(AddrMode.BaseGV, IntPtrTy, "sunkaddr");
2777 Result = Builder.CreateAdd(Result, V, "sunkaddr");
2782 // Add in the Base Offset if present.
2783 if (AddrMode.BaseOffs) {
2784 Value *V = ConstantInt::get(IntPtrTy, AddrMode.BaseOffs);
2786 Result = Builder.CreateAdd(Result, V, "sunkaddr");
2792 SunkAddr = Constant::getNullValue(Addr->getType());
2794 SunkAddr = Builder.CreateIntToPtr(Result, Addr->getType(), "sunkaddr");
2797 MemoryInst->replaceUsesOfWith(Repl, SunkAddr);
2799 // If we have no uses, recursively delete the value and all dead instructions
2801 if (Repl->use_empty()) {
2802 // This can cause recursive deletion, which can invalidate our iterator.
2803 // Use a WeakVH to hold onto it in case this happens.
2804 WeakVH IterHandle(CurInstIterator);
2805 BasicBlock *BB = CurInstIterator->getParent();
2807 RecursivelyDeleteTriviallyDeadInstructions(Repl, TLInfo);
2809 if (IterHandle != CurInstIterator) {
2810 // If the iterator instruction was recursively deleted, start over at the
2811 // start of the block.
2812 CurInstIterator = BB->begin();
2820 /// OptimizeInlineAsmInst - If there are any memory operands, use
2821 /// OptimizeMemoryInst to sink their address computing into the block when
2822 /// possible / profitable.
2823 bool CodeGenPrepare::OptimizeInlineAsmInst(CallInst *CS) {
2824 bool MadeChange = false;
2826 TargetLowering::AsmOperandInfoVector
2827 TargetConstraints = TLI->ParseConstraints(CS);
2829 for (unsigned i = 0, e = TargetConstraints.size(); i != e; ++i) {
2830 TargetLowering::AsmOperandInfo &OpInfo = TargetConstraints[i];
2832 // Compute the constraint code and ConstraintType to use.
2833 TLI->ComputeConstraintToUse(OpInfo, SDValue());
2835 if (OpInfo.ConstraintType == TargetLowering::C_Memory &&
2836 OpInfo.isIndirect) {
2837 Value *OpVal = CS->getArgOperand(ArgNo++);
2838 MadeChange |= OptimizeMemoryInst(CS, OpVal, OpVal->getType());
2839 } else if (OpInfo.Type == InlineAsm::isInput)
2846 /// MoveExtToFormExtLoad - Move a zext or sext fed by a load into the same
2847 /// basic block as the load, unless conditions are unfavorable. This allows
2848 /// SelectionDAG to fold the extend into the load.
2850 bool CodeGenPrepare::MoveExtToFormExtLoad(Instruction *I) {
2851 // Look for a load being extended.
2852 LoadInst *LI = dyn_cast<LoadInst>(I->getOperand(0));
2853 if (!LI) return false;
2855 // If they're already in the same block, there's nothing to do.
2856 if (LI->getParent() == I->getParent())
2859 // If the load has other users and the truncate is not free, this probably
2860 // isn't worthwhile.
2861 if (!LI->hasOneUse() &&
2862 TLI && (TLI->isTypeLegal(TLI->getValueType(LI->getType())) ||
2863 !TLI->isTypeLegal(TLI->getValueType(I->getType()))) &&
2864 !TLI->isTruncateFree(I->getType(), LI->getType()))
2867 // Check whether the target supports casts folded into loads.
2869 if (isa<ZExtInst>(I))
2870 LType = ISD::ZEXTLOAD;
2872 assert(isa<SExtInst>(I) && "Unexpected ext type!");
2873 LType = ISD::SEXTLOAD;
2875 if (TLI && !TLI->isLoadExtLegal(LType, TLI->getValueType(LI->getType())))
2878 // Move the extend into the same block as the load, so that SelectionDAG
2880 I->removeFromParent();
2886 bool CodeGenPrepare::OptimizeExtUses(Instruction *I) {
2887 BasicBlock *DefBB = I->getParent();
2889 // If the result of a {s|z}ext and its source are both live out, rewrite all
2890 // other uses of the source with result of extension.
2891 Value *Src = I->getOperand(0);
2892 if (Src->hasOneUse())
2895 // Only do this xform if truncating is free.
2896 if (TLI && !TLI->isTruncateFree(I->getType(), Src->getType()))
2899 // Only safe to perform the optimization if the source is also defined in
2901 if (!isa<Instruction>(Src) || DefBB != cast<Instruction>(Src)->getParent())
2904 bool DefIsLiveOut = false;
2905 for (User *U : I->users()) {
2906 Instruction *UI = cast<Instruction>(U);
2908 // Figure out which BB this ext is used in.
2909 BasicBlock *UserBB = UI->getParent();
2910 if (UserBB == DefBB) continue;
2911 DefIsLiveOut = true;
2917 // Make sure none of the uses are PHI nodes.
2918 for (User *U : Src->users()) {
2919 Instruction *UI = cast<Instruction>(U);
2920 BasicBlock *UserBB = UI->getParent();
2921 if (UserBB == DefBB) continue;
2922 // Be conservative. We don't want this xform to end up introducing
2923 // reloads just before load / store instructions.
2924 if (isa<PHINode>(UI) || isa<LoadInst>(UI) || isa<StoreInst>(UI))
2928 // InsertedTruncs - Only insert one trunc in each block once.
2929 DenseMap<BasicBlock*, Instruction*> InsertedTruncs;
2931 bool MadeChange = false;
2932 for (Use &U : Src->uses()) {
2933 Instruction *User = cast<Instruction>(U.getUser());
2935 // Figure out which BB this ext is used in.
2936 BasicBlock *UserBB = User->getParent();
2937 if (UserBB == DefBB) continue;
2939 // Both src and def are live in this block. Rewrite the use.
2940 Instruction *&InsertedTrunc = InsertedTruncs[UserBB];
2942 if (!InsertedTrunc) {
2943 BasicBlock::iterator InsertPt = UserBB->getFirstInsertionPt();
2944 InsertedTrunc = new TruncInst(I, Src->getType(), "", InsertPt);
2945 InsertedTruncsSet.insert(InsertedTrunc);
2948 // Replace a use of the {s|z}ext source with a use of the result.
2957 /// isFormingBranchFromSelectProfitable - Returns true if a SelectInst should be
2958 /// turned into an explicit branch.
2959 static bool isFormingBranchFromSelectProfitable(SelectInst *SI) {
2960 // FIXME: This should use the same heuristics as IfConversion to determine
2961 // whether a select is better represented as a branch. This requires that
2962 // branch probability metadata is preserved for the select, which is not the
2965 CmpInst *Cmp = dyn_cast<CmpInst>(SI->getCondition());
2967 // If the branch is predicted right, an out of order CPU can avoid blocking on
2968 // the compare. Emit cmovs on compares with a memory operand as branches to
2969 // avoid stalls on the load from memory. If the compare has more than one use
2970 // there's probably another cmov or setcc around so it's not worth emitting a
2975 Value *CmpOp0 = Cmp->getOperand(0);
2976 Value *CmpOp1 = Cmp->getOperand(1);
2978 // We check that the memory operand has one use to avoid uses of the loaded
2979 // value directly after the compare, making branches unprofitable.
2980 return Cmp->hasOneUse() &&
2981 ((isa<LoadInst>(CmpOp0) && CmpOp0->hasOneUse()) ||
2982 (isa<LoadInst>(CmpOp1) && CmpOp1->hasOneUse()));
2986 /// If we have a SelectInst that will likely profit from branch prediction,
2987 /// turn it into a branch.
2988 bool CodeGenPrepare::OptimizeSelectInst(SelectInst *SI) {
2989 bool VectorCond = !SI->getCondition()->getType()->isIntegerTy(1);
2991 // Can we convert the 'select' to CF ?
2992 if (DisableSelectToBranch || OptSize || !TLI || VectorCond)
2995 TargetLowering::SelectSupportKind SelectKind;
2997 SelectKind = TargetLowering::VectorMaskSelect;
2998 else if (SI->getType()->isVectorTy())
2999 SelectKind = TargetLowering::ScalarCondVectorVal;
3001 SelectKind = TargetLowering::ScalarValSelect;
3003 // Do we have efficient codegen support for this kind of 'selects' ?
3004 if (TLI->isSelectSupported(SelectKind)) {
3005 // We have efficient codegen support for the select instruction.
3006 // Check if it is profitable to keep this 'select'.
3007 if (!TLI->isPredictableSelectExpensive() ||
3008 !isFormingBranchFromSelectProfitable(SI))
3014 // First, we split the block containing the select into 2 blocks.
3015 BasicBlock *StartBlock = SI->getParent();
3016 BasicBlock::iterator SplitPt = ++(BasicBlock::iterator(SI));
3017 BasicBlock *NextBlock = StartBlock->splitBasicBlock(SplitPt, "select.end");
3019 // Create a new block serving as the landing pad for the branch.
3020 BasicBlock *SmallBlock = BasicBlock::Create(SI->getContext(), "select.mid",
3021 NextBlock->getParent(), NextBlock);
3023 // Move the unconditional branch from the block with the select in it into our
3024 // landing pad block.
3025 StartBlock->getTerminator()->eraseFromParent();
3026 BranchInst::Create(NextBlock, SmallBlock);
3028 // Insert the real conditional branch based on the original condition.
3029 BranchInst::Create(NextBlock, SmallBlock, SI->getCondition(), SI);
3031 // The select itself is replaced with a PHI Node.
3032 PHINode *PN = PHINode::Create(SI->getType(), 2, "", NextBlock->begin());
3034 PN->addIncoming(SI->getTrueValue(), StartBlock);
3035 PN->addIncoming(SI->getFalseValue(), SmallBlock);
3036 SI->replaceAllUsesWith(PN);
3037 SI->eraseFromParent();
3039 // Instruct OptimizeBlock to skip to the next block.
3040 CurInstIterator = StartBlock->end();
3041 ++NumSelectsExpanded;
3045 static bool isBroadcastShuffle(ShuffleVectorInst *SVI) {
3046 SmallVector<int, 16> Mask(SVI->getShuffleMask());
3048 for (unsigned i = 0; i < Mask.size(); ++i) {
3049 if (SplatElem != -1 && Mask[i] != -1 && Mask[i] != SplatElem)
3051 SplatElem = Mask[i];
3057 /// Some targets have expensive vector shifts if the lanes aren't all the same
3058 /// (e.g. x86 only introduced "vpsllvd" and friends with AVX2). In these cases
3059 /// it's often worth sinking a shufflevector splat down to its use so that
3060 /// codegen can spot all lanes are identical.
3061 bool CodeGenPrepare::OptimizeShuffleVectorInst(ShuffleVectorInst *SVI) {
3062 BasicBlock *DefBB = SVI->getParent();
3064 // Only do this xform if variable vector shifts are particularly expensive.
3065 if (!TLI || !TLI->isVectorShiftByScalarCheap(SVI->getType()))
3068 // We only expect better codegen by sinking a shuffle if we can recognise a
3070 if (!isBroadcastShuffle(SVI))
3073 // InsertedShuffles - Only insert a shuffle in each block once.
3074 DenseMap<BasicBlock*, Instruction*> InsertedShuffles;
3076 bool MadeChange = false;
3077 for (User *U : SVI->users()) {
3078 Instruction *UI = cast<Instruction>(U);
3080 // Figure out which BB this ext is used in.
3081 BasicBlock *UserBB = UI->getParent();
3082 if (UserBB == DefBB) continue;
3084 // For now only apply this when the splat is used by a shift instruction.
3085 if (!UI->isShift()) continue;
3087 // Everything checks out, sink the shuffle if the user's block doesn't
3088 // already have a copy.
3089 Instruction *&InsertedShuffle = InsertedShuffles[UserBB];
3091 if (!InsertedShuffle) {
3092 BasicBlock::iterator InsertPt = UserBB->getFirstInsertionPt();
3093 InsertedShuffle = new ShuffleVectorInst(SVI->getOperand(0),
3095 SVI->getOperand(2), "", InsertPt);
3098 UI->replaceUsesOfWith(SVI, InsertedShuffle);
3102 // If we removed all uses, nuke the shuffle.
3103 if (SVI->use_empty()) {
3104 SVI->eraseFromParent();
3111 bool CodeGenPrepare::OptimizeInst(Instruction *I) {
3112 if (PHINode *P = dyn_cast<PHINode>(I)) {
3113 // It is possible for very late stage optimizations (such as SimplifyCFG)
3114 // to introduce PHI nodes too late to be cleaned up. If we detect such a
3115 // trivial PHI, go ahead and zap it here.
3116 if (Value *V = SimplifyInstruction(P, TLI ? TLI->getDataLayout() : nullptr,
3118 P->replaceAllUsesWith(V);
3119 P->eraseFromParent();
3126 if (CastInst *CI = dyn_cast<CastInst>(I)) {
3127 // If the source of the cast is a constant, then this should have
3128 // already been constant folded. The only reason NOT to constant fold
3129 // it is if something (e.g. LSR) was careful to place the constant
3130 // evaluation in a block other than then one that uses it (e.g. to hoist
3131 // the address of globals out of a loop). If this is the case, we don't
3132 // want to forward-subst the cast.
3133 if (isa<Constant>(CI->getOperand(0)))
3136 if (TLI && OptimizeNoopCopyExpression(CI, *TLI))
3139 if (isa<ZExtInst>(I) || isa<SExtInst>(I)) {
3140 /// Sink a zext or sext into its user blocks if the target type doesn't
3141 /// fit in one register
3142 if (TLI && TLI->getTypeAction(CI->getContext(),
3143 TLI->getValueType(CI->getType())) ==
3144 TargetLowering::TypeExpandInteger) {
3145 return SinkCast(CI);
3147 bool MadeChange = MoveExtToFormExtLoad(I);
3148 return MadeChange | OptimizeExtUses(I);
3154 if (CmpInst *CI = dyn_cast<CmpInst>(I))
3155 if (!TLI || !TLI->hasMultipleConditionRegisters())
3156 return OptimizeCmpExpression(CI);
3158 if (LoadInst *LI = dyn_cast<LoadInst>(I)) {
3160 return OptimizeMemoryInst(I, I->getOperand(0), LI->getType());
3164 if (StoreInst *SI = dyn_cast<StoreInst>(I)) {
3166 return OptimizeMemoryInst(I, SI->getOperand(1),
3167 SI->getOperand(0)->getType());
3171 BinaryOperator *BinOp = dyn_cast<BinaryOperator>(I);
3173 if (BinOp && (BinOp->getOpcode() == Instruction::AShr ||
3174 BinOp->getOpcode() == Instruction::LShr)) {
3175 ConstantInt *CI = dyn_cast<ConstantInt>(BinOp->getOperand(1));
3176 if (TLI && CI && TLI->hasExtractBitsInsn())
3177 return OptimizeExtractBits(BinOp, CI, *TLI);
3182 if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(I)) {
3183 if (GEPI->hasAllZeroIndices()) {
3184 /// The GEP operand must be a pointer, so must its result -> BitCast
3185 Instruction *NC = new BitCastInst(GEPI->getOperand(0), GEPI->getType(),
3186 GEPI->getName(), GEPI);
3187 GEPI->replaceAllUsesWith(NC);
3188 GEPI->eraseFromParent();
3196 if (CallInst *CI = dyn_cast<CallInst>(I))
3197 return OptimizeCallInst(CI);
3199 if (SelectInst *SI = dyn_cast<SelectInst>(I))
3200 return OptimizeSelectInst(SI);
3202 if (ShuffleVectorInst *SVI = dyn_cast<ShuffleVectorInst>(I))
3203 return OptimizeShuffleVectorInst(SVI);
3208 // In this pass we look for GEP and cast instructions that are used
3209 // across basic blocks and rewrite them to improve basic-block-at-a-time
3211 bool CodeGenPrepare::OptimizeBlock(BasicBlock &BB) {
3213 bool MadeChange = false;
3215 CurInstIterator = BB.begin();
3216 while (CurInstIterator != BB.end())
3217 MadeChange |= OptimizeInst(CurInstIterator++);
3219 MadeChange |= DupRetToEnableTailCallOpts(&BB);
3224 // llvm.dbg.value is far away from the value then iSel may not be able
3225 // handle it properly. iSel will drop llvm.dbg.value if it can not
3226 // find a node corresponding to the value.
3227 bool CodeGenPrepare::PlaceDbgValues(Function &F) {
3228 bool MadeChange = false;
3229 for (Function::iterator I = F.begin(), E = F.end(); I != E; ++I) {
3230 Instruction *PrevNonDbgInst = nullptr;
3231 for (BasicBlock::iterator BI = I->begin(), BE = I->end(); BI != BE;) {
3232 Instruction *Insn = BI; ++BI;
3233 DbgValueInst *DVI = dyn_cast<DbgValueInst>(Insn);
3234 // Leave dbg.values that refer to an alloca alone. These
3235 // instrinsics describe the address of a variable (= the alloca)
3236 // being taken. They should not be moved next to the alloca
3237 // (and to the beginning of the scope), but rather stay close to
3238 // where said address is used.
3239 if (!DVI || (DVI->getValue() && isa<AllocaInst>(DVI->getValue()))) {
3240 PrevNonDbgInst = Insn;
3244 Instruction *VI = dyn_cast_or_null<Instruction>(DVI->getValue());
3245 if (VI && VI != PrevNonDbgInst && !VI->isTerminator()) {
3246 DEBUG(dbgs() << "Moving Debug Value before :\n" << *DVI << ' ' << *VI);
3247 DVI->removeFromParent();
3248 if (isa<PHINode>(VI))
3249 DVI->insertBefore(VI->getParent()->getFirstInsertionPt());
3251 DVI->insertAfter(VI);
3260 // If there is a sequence that branches based on comparing a single bit
3261 // against zero that can be combined into a single instruction, and the
3262 // target supports folding these into a single instruction, sink the
3263 // mask and compare into the branch uses. Do this before OptimizeBlock ->
3264 // OptimizeInst -> OptimizeCmpExpression, which perturbs the pattern being
3266 bool CodeGenPrepare::sinkAndCmp(Function &F) {
3267 if (!EnableAndCmpSinking)
3269 if (!TLI || !TLI->isMaskAndBranchFoldingLegal())
3271 bool MadeChange = false;
3272 for (Function::iterator I = F.begin(), E = F.end(); I != E; ) {
3273 BasicBlock *BB = I++;
3275 // Does this BB end with the following?
3276 // %andVal = and %val, #single-bit-set
3277 // %icmpVal = icmp %andResult, 0
3278 // br i1 %cmpVal label %dest1, label %dest2"
3279 BranchInst *Brcc = dyn_cast<BranchInst>(BB->getTerminator());
3280 if (!Brcc || !Brcc->isConditional())
3282 ICmpInst *Cmp = dyn_cast<ICmpInst>(Brcc->getOperand(0));
3283 if (!Cmp || Cmp->getParent() != BB)
3285 ConstantInt *Zero = dyn_cast<ConstantInt>(Cmp->getOperand(1));
3286 if (!Zero || !Zero->isZero())
3288 Instruction *And = dyn_cast<Instruction>(Cmp->getOperand(0));
3289 if (!And || And->getOpcode() != Instruction::And || And->getParent() != BB)
3291 ConstantInt* Mask = dyn_cast<ConstantInt>(And->getOperand(1));
3292 if (!Mask || !Mask->getUniqueInteger().isPowerOf2())
3294 DEBUG(dbgs() << "found and; icmp ?,0; brcc\n"); DEBUG(BB->dump());
3296 // Push the "and; icmp" for any users that are conditional branches.
3297 // Since there can only be one branch use per BB, we don't need to keep
3298 // track of which BBs we insert into.
3299 for (Value::use_iterator UI = Cmp->use_begin(), E = Cmp->use_end();
3303 BranchInst *BrccUser = dyn_cast<BranchInst>(*UI);
3305 if (!BrccUser || !BrccUser->isConditional())
3307 BasicBlock *UserBB = BrccUser->getParent();
3308 if (UserBB == BB) continue;
3309 DEBUG(dbgs() << "found Brcc use\n");
3311 // Sink the "and; icmp" to use.
3313 BinaryOperator *NewAnd =
3314 BinaryOperator::CreateAnd(And->getOperand(0), And->getOperand(1), "",
3317 CmpInst::Create(Cmp->getOpcode(), Cmp->getPredicate(), NewAnd, Zero,
3321 DEBUG(BrccUser->getParent()->dump());