1 //===- CodeGenPrepare.cpp - Prepare a function for code generation --------===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // This pass munges the code in the input function to better prepare it for
11 // SelectionDAG-based code generation. This works around limitations in it's
12 // basic-block-at-a-time approach. It should eventually be removed.
14 //===----------------------------------------------------------------------===//
16 #define DEBUG_TYPE "codegenprepare"
17 #include "llvm/Transforms/Scalar.h"
18 #include "llvm/ADT/DenseMap.h"
19 #include "llvm/ADT/SmallSet.h"
20 #include "llvm/ADT/Statistic.h"
21 #include "llvm/Analysis/DominatorInternals.h"
22 #include "llvm/Analysis/Dominators.h"
23 #include "llvm/Analysis/InstructionSimplify.h"
24 #include "llvm/Analysis/ProfileInfo.h"
25 #include "llvm/Assembly/Writer.h"
26 #include "llvm/Constants.h"
27 #include "llvm/DataLayout.h"
28 #include "llvm/DerivedTypes.h"
29 #include "llvm/Function.h"
30 #include "llvm/IRBuilder.h"
31 #include "llvm/InlineAsm.h"
32 #include "llvm/Instructions.h"
33 #include "llvm/IntrinsicInst.h"
34 #include "llvm/Pass.h"
35 #include "llvm/Support/CallSite.h"
36 #include "llvm/Support/CommandLine.h"
37 #include "llvm/Support/Debug.h"
38 #include "llvm/Support/GetElementPtrTypeIterator.h"
39 #include "llvm/Support/PatternMatch.h"
40 #include "llvm/Support/ValueHandle.h"
41 #include "llvm/Support/raw_ostream.h"
42 #include "llvm/Target/TargetLibraryInfo.h"
43 #include "llvm/Target/TargetLowering.h"
44 #include "llvm/Transforms/Utils/AddrModeMatcher.h"
45 #include "llvm/Transforms/Utils/BasicBlockUtils.h"
46 #include "llvm/Transforms/Utils/BuildLibCalls.h"
47 #include "llvm/Transforms/Utils/BypassSlowDivision.h"
48 #include "llvm/Transforms/Utils/Local.h"
50 using namespace llvm::PatternMatch;
52 STATISTIC(NumBlocksElim, "Number of blocks eliminated");
53 STATISTIC(NumPHIsElim, "Number of trivial PHIs eliminated");
54 STATISTIC(NumGEPsElim, "Number of GEPs converted to casts");
55 STATISTIC(NumCmpUses, "Number of uses of Cmp expressions replaced with uses of "
57 STATISTIC(NumCastUses, "Number of uses of Cast expressions replaced with uses "
59 STATISTIC(NumMemoryInsts, "Number of memory instructions whose address "
60 "computations were sunk");
61 STATISTIC(NumExtsMoved, "Number of [s|z]ext instructions combined with loads");
62 STATISTIC(NumExtUses, "Number of uses of [s|z]ext instructions optimized");
63 STATISTIC(NumRetsDup, "Number of return instructions duplicated");
64 STATISTIC(NumDbgValueMoved, "Number of debug value instructions moved");
65 STATISTIC(NumSelectsExpanded, "Number of selects turned 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."));
76 class CodeGenPrepare : public FunctionPass {
77 /// TLI - Keep a pointer of a TargetLowering to consult for determining
78 /// transformation profitability.
79 const TargetLowering *TLI;
80 const TargetLibraryInfo *TLInfo;
84 /// CurInstIterator - As we scan instructions optimizing them, this is the
85 /// next instruction to optimize. Xforms that can invalidate this should
87 BasicBlock::iterator CurInstIterator;
89 /// Keeps track of non-local addresses that have been sunk into a block.
90 /// This allows us to avoid inserting duplicate code for blocks with
91 /// multiple load/stores of the same address.
92 DenseMap<Value*, Value*> SunkAddrs;
94 /// ModifiedDT - If CFG is modified in anyway, dominator tree may need to
98 /// OptSize - True if optimizing for size.
102 static char ID; // Pass identification, replacement for typeid
103 explicit CodeGenPrepare(const TargetLowering *tli = 0)
104 : FunctionPass(ID), TLI(tli) {
105 initializeCodeGenPreparePass(*PassRegistry::getPassRegistry());
107 bool runOnFunction(Function &F);
109 virtual void getAnalysisUsage(AnalysisUsage &AU) const {
110 AU.addPreserved<DominatorTree>();
111 AU.addPreserved<ProfileInfo>();
112 AU.addRequired<TargetLibraryInfo>();
116 bool EliminateFallThrough(Function &F);
117 bool EliminateMostlyEmptyBlocks(Function &F);
118 bool CanMergeBlocks(const BasicBlock *BB, const BasicBlock *DestBB) const;
119 void EliminateMostlyEmptyBlock(BasicBlock *BB);
120 bool OptimizeBlock(BasicBlock &BB);
121 bool OptimizeInst(Instruction *I);
122 bool OptimizeMemoryInst(Instruction *I, Value *Addr, Type *AccessTy);
123 bool OptimizeInlineAsmInst(CallInst *CS);
124 bool OptimizeCallInst(CallInst *CI);
125 bool MoveExtToFormExtLoad(Instruction *I);
126 bool OptimizeExtUses(Instruction *I);
127 bool OptimizeSelectInst(SelectInst *SI);
128 bool DupRetToEnableTailCallOpts(BasicBlock *BB);
129 bool PlaceDbgValues(Function &F);
133 char CodeGenPrepare::ID = 0;
134 INITIALIZE_PASS_BEGIN(CodeGenPrepare, "codegenprepare",
135 "Optimize for code generation", false, false)
136 INITIALIZE_PASS_DEPENDENCY(TargetLibraryInfo)
137 INITIALIZE_PASS_END(CodeGenPrepare, "codegenprepare",
138 "Optimize for code generation", false, false)
140 FunctionPass *llvm::createCodeGenPreparePass(const TargetLowering *TLI) {
141 return new CodeGenPrepare(TLI);
144 bool CodeGenPrepare::runOnFunction(Function &F) {
145 bool EverMadeChange = false;
148 TLInfo = &getAnalysis<TargetLibraryInfo>();
149 DT = getAnalysisIfAvailable<DominatorTree>();
150 PFI = getAnalysisIfAvailable<ProfileInfo>();
151 OptSize = F.getFnAttributes().hasAttribute(Attributes::OptimizeForSize);
153 /// This optimization identifies DIV instructions that can be
154 /// profitably bypassed and carried out with a shorter, faster divide.
155 if (TLI && TLI->isSlowDivBypassed()) {
156 const DenseMap<unsigned int, unsigned int> &BypassWidths =
157 TLI->getBypassSlowDivWidths();
158 for (Function::iterator I = F.begin(); I != F.end(); I++)
159 EverMadeChange |= bypassSlowDivision(F, I, BypassWidths);
162 // Eliminate blocks that contain only PHI nodes and an
163 // unconditional branch.
164 EverMadeChange |= EliminateMostlyEmptyBlocks(F);
166 // llvm.dbg.value is far away from the value then iSel may not be able
167 // handle it properly. iSel will drop llvm.dbg.value if it can not
168 // find a node corresponding to the value.
169 EverMadeChange |= PlaceDbgValues(F);
171 bool MadeChange = true;
174 for (Function::iterator I = F.begin(); I != F.end(); ) {
175 BasicBlock *BB = I++;
176 MadeChange |= OptimizeBlock(*BB);
178 EverMadeChange |= MadeChange;
183 if (!DisableBranchOpts) {
185 SmallPtrSet<BasicBlock*, 8> WorkList;
186 for (Function::iterator BB = F.begin(), E = F.end(); BB != E; ++BB) {
187 SmallVector<BasicBlock*, 2> Successors(succ_begin(BB), succ_end(BB));
188 MadeChange |= ConstantFoldTerminator(BB, true);
189 if (!MadeChange) continue;
191 for (SmallVectorImpl<BasicBlock*>::iterator
192 II = Successors.begin(), IE = Successors.end(); II != IE; ++II)
193 if (pred_begin(*II) == pred_end(*II))
194 WorkList.insert(*II);
197 // Delete the dead blocks and any of their dead successors.
198 while (!WorkList.empty()) {
199 BasicBlock *BB = *WorkList.begin();
201 SmallVector<BasicBlock*, 2> Successors(succ_begin(BB), succ_end(BB));
205 for (SmallVectorImpl<BasicBlock*>::iterator
206 II = Successors.begin(), IE = Successors.end(); II != IE; ++II)
207 if (pred_begin(*II) == pred_end(*II))
208 WorkList.insert(*II);
211 // Merge pairs of basic blocks with unconditional branches, connected by
213 if (EverMadeChange || MadeChange)
214 MadeChange |= EliminateFallThrough(F);
218 EverMadeChange |= MadeChange;
221 if (ModifiedDT && DT)
222 DT->DT->recalculate(F);
224 return EverMadeChange;
227 /// EliminateFallThrough - Merge basic blocks which are connected
228 /// by a single edge, where one of the basic blocks has a single successor
229 /// pointing to the other basic block, which has a single predecessor.
230 bool CodeGenPrepare::EliminateFallThrough(Function &F) {
231 bool Changed = false;
232 // Scan all of the blocks in the function, except for the entry block.
233 for (Function::iterator I = ++F.begin(), E = F.end(); I != E; ) {
234 BasicBlock *BB = I++;
235 // If the destination block has a single pred, then this is a trivial
236 // edge, just collapse it.
237 BasicBlock *SinglePred = BB->getSinglePredecessor();
239 // Don't merge if BB's address is taken.
240 if (!SinglePred || SinglePred == BB || BB->hasAddressTaken()) continue;
242 BranchInst *Term = dyn_cast<BranchInst>(SinglePred->getTerminator());
243 if (Term && !Term->isConditional()) {
245 DEBUG(dbgs() << "To merge:\n"<< *SinglePred << "\n\n\n");
246 // Remember if SinglePred was the entry block of the function.
247 // If so, we will need to move BB back to the entry position.
248 bool isEntry = SinglePred == &SinglePred->getParent()->getEntryBlock();
249 MergeBasicBlockIntoOnlyPred(BB, this);
251 if (isEntry && BB != &BB->getParent()->getEntryBlock())
252 BB->moveBefore(&BB->getParent()->getEntryBlock());
254 // We have erased a block. Update the iterator.
261 /// EliminateMostlyEmptyBlocks - eliminate blocks that contain only PHI nodes,
262 /// debug info directives, and an unconditional branch. Passes before isel
263 /// (e.g. LSR/loopsimplify) often split edges in ways that are non-optimal for
264 /// isel. Start by eliminating these blocks so we can split them the way we
266 bool CodeGenPrepare::EliminateMostlyEmptyBlocks(Function &F) {
267 bool MadeChange = false;
268 // Note that this intentionally skips the entry block.
269 for (Function::iterator I = ++F.begin(), E = F.end(); I != E; ) {
270 BasicBlock *BB = I++;
272 // If this block doesn't end with an uncond branch, ignore it.
273 BranchInst *BI = dyn_cast<BranchInst>(BB->getTerminator());
274 if (!BI || !BI->isUnconditional())
277 // If the instruction before the branch (skipping debug info) isn't a phi
278 // node, then other stuff is happening here.
279 BasicBlock::iterator BBI = BI;
280 if (BBI != BB->begin()) {
282 while (isa<DbgInfoIntrinsic>(BBI)) {
283 if (BBI == BB->begin())
287 if (!isa<DbgInfoIntrinsic>(BBI) && !isa<PHINode>(BBI))
291 // Do not break infinite loops.
292 BasicBlock *DestBB = BI->getSuccessor(0);
296 if (!CanMergeBlocks(BB, DestBB))
299 EliminateMostlyEmptyBlock(BB);
305 /// CanMergeBlocks - Return true if we can merge BB into DestBB if there is a
306 /// single uncond branch between them, and BB contains no other non-phi
308 bool CodeGenPrepare::CanMergeBlocks(const BasicBlock *BB,
309 const BasicBlock *DestBB) const {
310 // We only want to eliminate blocks whose phi nodes are used by phi nodes in
311 // the successor. If there are more complex condition (e.g. preheaders),
312 // don't mess around with them.
313 BasicBlock::const_iterator BBI = BB->begin();
314 while (const PHINode *PN = dyn_cast<PHINode>(BBI++)) {
315 for (Value::const_use_iterator UI = PN->use_begin(), E = PN->use_end();
317 const Instruction *User = cast<Instruction>(*UI);
318 if (User->getParent() != DestBB || !isa<PHINode>(User))
320 // If User is inside DestBB block and it is a PHINode then check
321 // incoming value. If incoming value is not from BB then this is
322 // a complex condition (e.g. preheaders) we want to avoid here.
323 if (User->getParent() == DestBB) {
324 if (const PHINode *UPN = dyn_cast<PHINode>(User))
325 for (unsigned I = 0, E = UPN->getNumIncomingValues(); I != E; ++I) {
326 Instruction *Insn = dyn_cast<Instruction>(UPN->getIncomingValue(I));
327 if (Insn && Insn->getParent() == BB &&
328 Insn->getParent() != UPN->getIncomingBlock(I))
335 // If BB and DestBB contain any common predecessors, then the phi nodes in BB
336 // and DestBB may have conflicting incoming values for the block. If so, we
337 // can't merge the block.
338 const PHINode *DestBBPN = dyn_cast<PHINode>(DestBB->begin());
339 if (!DestBBPN) return true; // no conflict.
341 // Collect the preds of BB.
342 SmallPtrSet<const BasicBlock*, 16> BBPreds;
343 if (const PHINode *BBPN = dyn_cast<PHINode>(BB->begin())) {
344 // It is faster to get preds from a PHI than with pred_iterator.
345 for (unsigned i = 0, e = BBPN->getNumIncomingValues(); i != e; ++i)
346 BBPreds.insert(BBPN->getIncomingBlock(i));
348 BBPreds.insert(pred_begin(BB), pred_end(BB));
351 // Walk the preds of DestBB.
352 for (unsigned i = 0, e = DestBBPN->getNumIncomingValues(); i != e; ++i) {
353 BasicBlock *Pred = DestBBPN->getIncomingBlock(i);
354 if (BBPreds.count(Pred)) { // Common predecessor?
355 BBI = DestBB->begin();
356 while (const PHINode *PN = dyn_cast<PHINode>(BBI++)) {
357 const Value *V1 = PN->getIncomingValueForBlock(Pred);
358 const Value *V2 = PN->getIncomingValueForBlock(BB);
360 // If V2 is a phi node in BB, look up what the mapped value will be.
361 if (const PHINode *V2PN = dyn_cast<PHINode>(V2))
362 if (V2PN->getParent() == BB)
363 V2 = V2PN->getIncomingValueForBlock(Pred);
365 // If there is a conflict, bail out.
366 if (V1 != V2) return false;
375 /// EliminateMostlyEmptyBlock - Eliminate a basic block that have only phi's and
376 /// an unconditional branch in it.
377 void CodeGenPrepare::EliminateMostlyEmptyBlock(BasicBlock *BB) {
378 BranchInst *BI = cast<BranchInst>(BB->getTerminator());
379 BasicBlock *DestBB = BI->getSuccessor(0);
381 DEBUG(dbgs() << "MERGING MOSTLY EMPTY BLOCKS - BEFORE:\n" << *BB << *DestBB);
383 // If the destination block has a single pred, then this is a trivial edge,
385 if (BasicBlock *SinglePred = DestBB->getSinglePredecessor()) {
386 if (SinglePred != DestBB) {
387 // Remember if SinglePred was the entry block of the function. If so, we
388 // will need to move BB back to the entry position.
389 bool isEntry = SinglePred == &SinglePred->getParent()->getEntryBlock();
390 MergeBasicBlockIntoOnlyPred(DestBB, this);
392 if (isEntry && BB != &BB->getParent()->getEntryBlock())
393 BB->moveBefore(&BB->getParent()->getEntryBlock());
395 DEBUG(dbgs() << "AFTER:\n" << *DestBB << "\n\n\n");
400 // Otherwise, we have multiple predecessors of BB. Update the PHIs in DestBB
401 // to handle the new incoming edges it is about to have.
403 for (BasicBlock::iterator BBI = DestBB->begin();
404 (PN = dyn_cast<PHINode>(BBI)); ++BBI) {
405 // Remove the incoming value for BB, and remember it.
406 Value *InVal = PN->removeIncomingValue(BB, false);
408 // Two options: either the InVal is a phi node defined in BB or it is some
409 // value that dominates BB.
410 PHINode *InValPhi = dyn_cast<PHINode>(InVal);
411 if (InValPhi && InValPhi->getParent() == BB) {
412 // Add all of the input values of the input PHI as inputs of this phi.
413 for (unsigned i = 0, e = InValPhi->getNumIncomingValues(); i != e; ++i)
414 PN->addIncoming(InValPhi->getIncomingValue(i),
415 InValPhi->getIncomingBlock(i));
417 // Otherwise, add one instance of the dominating value for each edge that
418 // we will be adding.
419 if (PHINode *BBPN = dyn_cast<PHINode>(BB->begin())) {
420 for (unsigned i = 0, e = BBPN->getNumIncomingValues(); i != e; ++i)
421 PN->addIncoming(InVal, BBPN->getIncomingBlock(i));
423 for (pred_iterator PI = pred_begin(BB), E = pred_end(BB); PI != E; ++PI)
424 PN->addIncoming(InVal, *PI);
429 // The PHIs are now updated, change everything that refers to BB to use
430 // DestBB and remove BB.
431 BB->replaceAllUsesWith(DestBB);
432 if (DT && !ModifiedDT) {
433 BasicBlock *BBIDom = DT->getNode(BB)->getIDom()->getBlock();
434 BasicBlock *DestBBIDom = DT->getNode(DestBB)->getIDom()->getBlock();
435 BasicBlock *NewIDom = DT->findNearestCommonDominator(BBIDom, DestBBIDom);
436 DT->changeImmediateDominator(DestBB, NewIDom);
440 PFI->replaceAllUses(BB, DestBB);
441 PFI->removeEdge(ProfileInfo::getEdge(BB, DestBB));
443 BB->eraseFromParent();
446 DEBUG(dbgs() << "AFTER:\n" << *DestBB << "\n\n\n");
449 /// OptimizeNoopCopyExpression - If the specified cast instruction is a noop
450 /// copy (e.g. it's casting from one pointer type to another, i32->i8 on PPC),
451 /// sink it into user blocks to reduce the number of virtual
452 /// registers that must be created and coalesced.
454 /// Return true if any changes are made.
456 static bool OptimizeNoopCopyExpression(CastInst *CI, const TargetLowering &TLI){
457 // If this is a noop copy,
458 EVT SrcVT = TLI.getValueType(CI->getOperand(0)->getType());
459 EVT DstVT = TLI.getValueType(CI->getType());
461 // This is an fp<->int conversion?
462 if (SrcVT.isInteger() != DstVT.isInteger())
465 // If this is an extension, it will be a zero or sign extension, which
467 if (SrcVT.bitsLT(DstVT)) return false;
469 // If these values will be promoted, find out what they will be promoted
470 // to. This helps us consider truncates on PPC as noop copies when they
472 if (TLI.getTypeAction(CI->getContext(), SrcVT) ==
473 TargetLowering::TypePromoteInteger)
474 SrcVT = TLI.getTypeToTransformTo(CI->getContext(), SrcVT);
475 if (TLI.getTypeAction(CI->getContext(), DstVT) ==
476 TargetLowering::TypePromoteInteger)
477 DstVT = TLI.getTypeToTransformTo(CI->getContext(), DstVT);
479 // If, after promotion, these are the same types, this is a noop copy.
483 BasicBlock *DefBB = CI->getParent();
485 /// InsertedCasts - Only insert a cast in each block once.
486 DenseMap<BasicBlock*, CastInst*> InsertedCasts;
488 bool MadeChange = false;
489 for (Value::use_iterator UI = CI->use_begin(), E = CI->use_end();
491 Use &TheUse = UI.getUse();
492 Instruction *User = cast<Instruction>(*UI);
494 // Figure out which BB this cast is used in. For PHI's this is the
495 // appropriate predecessor block.
496 BasicBlock *UserBB = User->getParent();
497 if (PHINode *PN = dyn_cast<PHINode>(User)) {
498 UserBB = PN->getIncomingBlock(UI);
501 // Preincrement use iterator so we don't invalidate it.
504 // If this user is in the same block as the cast, don't change the cast.
505 if (UserBB == DefBB) continue;
507 // If we have already inserted a cast into this block, use it.
508 CastInst *&InsertedCast = InsertedCasts[UserBB];
511 BasicBlock::iterator InsertPt = UserBB->getFirstInsertionPt();
513 CastInst::Create(CI->getOpcode(), CI->getOperand(0), CI->getType(), "",
518 // Replace a use of the cast with a use of the new cast.
519 TheUse = InsertedCast;
523 // If we removed all uses, nuke the cast.
524 if (CI->use_empty()) {
525 CI->eraseFromParent();
532 /// OptimizeCmpExpression - sink the given CmpInst into user blocks to reduce
533 /// the number of virtual registers that must be created and coalesced. This is
534 /// a clear win except on targets with multiple condition code registers
535 /// (PowerPC), where it might lose; some adjustment may be wanted there.
537 /// Return true if any changes are made.
538 static bool OptimizeCmpExpression(CmpInst *CI) {
539 BasicBlock *DefBB = CI->getParent();
541 /// InsertedCmp - Only insert a cmp in each block once.
542 DenseMap<BasicBlock*, CmpInst*> InsertedCmps;
544 bool MadeChange = false;
545 for (Value::use_iterator UI = CI->use_begin(), E = CI->use_end();
547 Use &TheUse = UI.getUse();
548 Instruction *User = cast<Instruction>(*UI);
550 // Preincrement use iterator so we don't invalidate it.
553 // Don't bother for PHI nodes.
554 if (isa<PHINode>(User))
557 // Figure out which BB this cmp is used in.
558 BasicBlock *UserBB = User->getParent();
560 // If this user is in the same block as the cmp, don't change the cmp.
561 if (UserBB == DefBB) continue;
563 // If we have already inserted a cmp into this block, use it.
564 CmpInst *&InsertedCmp = InsertedCmps[UserBB];
567 BasicBlock::iterator InsertPt = UserBB->getFirstInsertionPt();
569 CmpInst::Create(CI->getOpcode(),
570 CI->getPredicate(), CI->getOperand(0),
571 CI->getOperand(1), "", InsertPt);
575 // Replace a use of the cmp with a use of the new cmp.
576 TheUse = InsertedCmp;
580 // If we removed all uses, nuke the cmp.
582 CI->eraseFromParent();
588 class CodeGenPrepareFortifiedLibCalls : public SimplifyFortifiedLibCalls {
590 void replaceCall(Value *With) {
591 CI->replaceAllUsesWith(With);
592 CI->eraseFromParent();
594 bool isFoldable(unsigned SizeCIOp, unsigned, bool) const {
595 if (ConstantInt *SizeCI =
596 dyn_cast<ConstantInt>(CI->getArgOperand(SizeCIOp)))
597 return SizeCI->isAllOnesValue();
601 } // end anonymous namespace
603 bool CodeGenPrepare::OptimizeCallInst(CallInst *CI) {
604 BasicBlock *BB = CI->getParent();
606 // Lower inline assembly if we can.
607 // If we found an inline asm expession, and if the target knows how to
608 // lower it to normal LLVM code, do so now.
609 if (TLI && isa<InlineAsm>(CI->getCalledValue())) {
610 if (TLI->ExpandInlineAsm(CI)) {
611 // Avoid invalidating the iterator.
612 CurInstIterator = BB->begin();
613 // Avoid processing instructions out of order, which could cause
614 // reuse before a value is defined.
618 // Sink address computing for memory operands into the block.
619 if (OptimizeInlineAsmInst(CI))
623 // Lower all uses of llvm.objectsize.*
624 IntrinsicInst *II = dyn_cast<IntrinsicInst>(CI);
625 if (II && II->getIntrinsicID() == Intrinsic::objectsize) {
626 bool Min = (cast<ConstantInt>(II->getArgOperand(1))->getZExtValue() == 1);
627 Type *ReturnTy = CI->getType();
628 Constant *RetVal = ConstantInt::get(ReturnTy, Min ? 0 : -1ULL);
630 // Substituting this can cause recursive simplifications, which can
631 // invalidate our iterator. Use a WeakVH to hold onto it in case this
633 WeakVH IterHandle(CurInstIterator);
635 replaceAndRecursivelySimplify(CI, RetVal, TLI ? TLI->getDataLayout() : 0,
636 TLInfo, ModifiedDT ? 0 : DT);
638 // If the iterator instruction was recursively deleted, start over at the
639 // start of the block.
640 if (IterHandle != CurInstIterator) {
641 CurInstIterator = BB->begin();
648 SmallVector<Value*, 2> PtrOps;
650 if (TLI->GetAddrModeArguments(II, PtrOps, AccessTy))
651 while (!PtrOps.empty())
652 if (OptimizeMemoryInst(II, PtrOps.pop_back_val(), AccessTy))
656 // From here on out we're working with named functions.
657 if (CI->getCalledFunction() == 0) return false;
659 // We'll need DataLayout from here on out.
660 const DataLayout *TD = TLI ? TLI->getDataLayout() : 0;
661 if (!TD) return false;
663 // Lower all default uses of _chk calls. This is very similar
664 // to what InstCombineCalls does, but here we are only lowering calls
665 // that have the default "don't know" as the objectsize. Anything else
666 // should be left alone.
667 CodeGenPrepareFortifiedLibCalls Simplifier;
668 return Simplifier.fold(CI, TD, TLInfo);
671 /// DupRetToEnableTailCallOpts - Look for opportunities to duplicate return
672 /// instructions to the predecessor to enable tail call optimizations. The
673 /// case it is currently looking for is:
676 /// %tmp0 = tail call i32 @f0()
679 /// %tmp1 = tail call i32 @f1()
682 /// %tmp2 = tail call i32 @f2()
685 /// %retval = phi i32 [ %tmp0, %bb0 ], [ %tmp1, %bb1 ], [ %tmp2, %bb2 ]
693 /// %tmp0 = tail call i32 @f0()
696 /// %tmp1 = tail call i32 @f1()
699 /// %tmp2 = tail call i32 @f2()
702 bool CodeGenPrepare::DupRetToEnableTailCallOpts(BasicBlock *BB) {
706 ReturnInst *RI = dyn_cast<ReturnInst>(BB->getTerminator());
711 BitCastInst *BCI = 0;
712 Value *V = RI->getReturnValue();
714 BCI = dyn_cast<BitCastInst>(V);
716 V = BCI->getOperand(0);
718 PN = dyn_cast<PHINode>(V);
723 if (PN && PN->getParent() != BB)
726 // It's not safe to eliminate the sign / zero extension of the return value.
727 // See llvm::isInTailCallPosition().
728 const Function *F = BB->getParent();
729 Attributes CallerRetAttr = F->getAttributes().getRetAttributes();
730 if (CallerRetAttr.hasAttribute(Attributes::ZExt) ||
731 CallerRetAttr.hasAttribute(Attributes::SExt))
734 // Make sure there are no instructions between the PHI and return, or that the
735 // return is the first instruction in the block.
737 BasicBlock::iterator BI = BB->begin();
738 do { ++BI; } while (isa<DbgInfoIntrinsic>(BI));
740 // Also skip over the bitcast.
745 BasicBlock::iterator BI = BB->begin();
746 while (isa<DbgInfoIntrinsic>(BI)) ++BI;
751 /// Only dup the ReturnInst if the CallInst is likely to be emitted as a tail
753 SmallVector<CallInst*, 4> TailCalls;
755 for (unsigned I = 0, E = PN->getNumIncomingValues(); I != E; ++I) {
756 CallInst *CI = dyn_cast<CallInst>(PN->getIncomingValue(I));
757 // Make sure the phi value is indeed produced by the tail call.
758 if (CI && CI->hasOneUse() && CI->getParent() == PN->getIncomingBlock(I) &&
759 TLI->mayBeEmittedAsTailCall(CI))
760 TailCalls.push_back(CI);
763 SmallPtrSet<BasicBlock*, 4> VisitedBBs;
764 for (pred_iterator PI = pred_begin(BB), PE = pred_end(BB); PI != PE; ++PI) {
765 if (!VisitedBBs.insert(*PI))
768 BasicBlock::InstListType &InstList = (*PI)->getInstList();
769 BasicBlock::InstListType::reverse_iterator RI = InstList.rbegin();
770 BasicBlock::InstListType::reverse_iterator RE = InstList.rend();
771 do { ++RI; } while (RI != RE && isa<DbgInfoIntrinsic>(&*RI));
775 CallInst *CI = dyn_cast<CallInst>(&*RI);
776 if (CI && CI->use_empty() && TLI->mayBeEmittedAsTailCall(CI))
777 TailCalls.push_back(CI);
781 bool Changed = false;
782 for (unsigned i = 0, e = TailCalls.size(); i != e; ++i) {
783 CallInst *CI = TailCalls[i];
786 // Conservatively require the attributes of the call to match those of the
787 // return. Ignore noalias because it doesn't affect the call sequence.
788 Attributes CalleeRetAttr = CS.getAttributes().getRetAttributes();
789 if (AttrBuilder(CalleeRetAttr).
790 removeAttribute(Attributes::NoAlias) !=
791 AttrBuilder(CallerRetAttr).
792 removeAttribute(Attributes::NoAlias))
795 // Make sure the call instruction is followed by an unconditional branch to
797 BasicBlock *CallBB = CI->getParent();
798 BranchInst *BI = dyn_cast<BranchInst>(CallBB->getTerminator());
799 if (!BI || !BI->isUnconditional() || BI->getSuccessor(0) != BB)
802 // Duplicate the return into CallBB.
803 (void)FoldReturnIntoUncondBranch(RI, BB, CallBB);
804 ModifiedDT = Changed = true;
808 // If we eliminated all predecessors of the block, delete the block now.
809 if (Changed && !BB->hasAddressTaken() && pred_begin(BB) == pred_end(BB))
810 BB->eraseFromParent();
815 //===----------------------------------------------------------------------===//
816 // Memory Optimization
817 //===----------------------------------------------------------------------===//
819 /// IsNonLocalValue - Return true if the specified values are defined in a
820 /// different basic block than BB.
821 static bool IsNonLocalValue(Value *V, BasicBlock *BB) {
822 if (Instruction *I = dyn_cast<Instruction>(V))
823 return I->getParent() != BB;
827 /// OptimizeMemoryInst - Load and Store Instructions often have
828 /// addressing modes that can do significant amounts of computation. As such,
829 /// instruction selection will try to get the load or store to do as much
830 /// computation as possible for the program. The problem is that isel can only
831 /// see within a single block. As such, we sink as much legal addressing mode
832 /// stuff into the block as possible.
834 /// This method is used to optimize both load/store and inline asms with memory
836 bool CodeGenPrepare::OptimizeMemoryInst(Instruction *MemoryInst, Value *Addr,
840 // Try to collapse single-value PHI nodes. This is necessary to undo
841 // unprofitable PRE transformations.
842 SmallVector<Value*, 8> worklist;
843 SmallPtrSet<Value*, 16> Visited;
844 worklist.push_back(Addr);
846 // Use a worklist to iteratively look through PHI nodes, and ensure that
847 // the addressing mode obtained from the non-PHI roots of the graph
849 Value *Consensus = 0;
850 unsigned NumUsesConsensus = 0;
851 bool IsNumUsesConsensusValid = false;
852 SmallVector<Instruction*, 16> AddrModeInsts;
853 ExtAddrMode AddrMode;
854 while (!worklist.empty()) {
855 Value *V = worklist.back();
858 // Break use-def graph loops.
859 if (!Visited.insert(V)) {
864 // For a PHI node, push all of its incoming values.
865 if (PHINode *P = dyn_cast<PHINode>(V)) {
866 for (unsigned i = 0, e = P->getNumIncomingValues(); i != e; ++i)
867 worklist.push_back(P->getIncomingValue(i));
871 // For non-PHIs, determine the addressing mode being computed.
872 SmallVector<Instruction*, 16> NewAddrModeInsts;
873 ExtAddrMode NewAddrMode =
874 AddressingModeMatcher::Match(V, AccessTy, MemoryInst,
875 NewAddrModeInsts, *TLI);
877 // This check is broken into two cases with very similar code to avoid using
878 // getNumUses() as much as possible. Some values have a lot of uses, so
879 // calling getNumUses() unconditionally caused a significant compile-time
883 AddrMode = NewAddrMode;
884 AddrModeInsts = NewAddrModeInsts;
886 } else if (NewAddrMode == AddrMode) {
887 if (!IsNumUsesConsensusValid) {
888 NumUsesConsensus = Consensus->getNumUses();
889 IsNumUsesConsensusValid = true;
892 // Ensure that the obtained addressing mode is equivalent to that obtained
893 // for all other roots of the PHI traversal. Also, when choosing one
894 // such root as representative, select the one with the most uses in order
895 // to keep the cost modeling heuristics in AddressingModeMatcher
897 unsigned NumUses = V->getNumUses();
898 if (NumUses > NumUsesConsensus) {
900 NumUsesConsensus = NumUses;
901 AddrModeInsts = NewAddrModeInsts;
910 // If the addressing mode couldn't be determined, or if multiple different
911 // ones were determined, bail out now.
912 if (!Consensus) return false;
914 // Check to see if any of the instructions supersumed by this addr mode are
915 // non-local to I's BB.
916 bool AnyNonLocal = false;
917 for (unsigned i = 0, e = AddrModeInsts.size(); i != e; ++i) {
918 if (IsNonLocalValue(AddrModeInsts[i], MemoryInst->getParent())) {
924 // If all the instructions matched are already in this BB, don't do anything.
926 DEBUG(dbgs() << "CGP: Found local addrmode: " << AddrMode << "\n");
930 // Insert this computation right after this user. Since our caller is
931 // scanning from the top of the BB to the bottom, reuse of the expr are
932 // guaranteed to happen later.
933 IRBuilder<> Builder(MemoryInst);
935 // Now that we determined the addressing expression we want to use and know
936 // that we have to sink it into this block. Check to see if we have already
937 // done this for some other load/store instr in this block. If so, reuse the
939 Value *&SunkAddr = SunkAddrs[Addr];
941 DEBUG(dbgs() << "CGP: Reusing nonlocal addrmode: " << AddrMode << " for "
943 if (SunkAddr->getType() != Addr->getType())
944 SunkAddr = Builder.CreateBitCast(SunkAddr, Addr->getType());
946 DEBUG(dbgs() << "CGP: SINKING nonlocal addrmode: " << AddrMode << " for "
949 TLI->getDataLayout()->getIntPtrType(AccessTy->getContext());
953 // Start with the base register. Do this first so that subsequent address
954 // matching finds it last, which will prevent it from trying to match it
955 // as the scaled value in case it happens to be a mul. That would be
956 // problematic if we've sunk a different mul for the scale, because then
957 // we'd end up sinking both muls.
958 if (AddrMode.BaseReg) {
959 Value *V = AddrMode.BaseReg;
960 if (V->getType()->isPointerTy())
961 V = Builder.CreatePtrToInt(V, IntPtrTy, "sunkaddr");
962 if (V->getType() != IntPtrTy)
963 V = Builder.CreateIntCast(V, IntPtrTy, /*isSigned=*/true, "sunkaddr");
967 // Add the scale value.
968 if (AddrMode.Scale) {
969 Value *V = AddrMode.ScaledReg;
970 if (V->getType() == IntPtrTy) {
972 } else if (V->getType()->isPointerTy()) {
973 V = Builder.CreatePtrToInt(V, IntPtrTy, "sunkaddr");
974 } else if (cast<IntegerType>(IntPtrTy)->getBitWidth() <
975 cast<IntegerType>(V->getType())->getBitWidth()) {
976 V = Builder.CreateTrunc(V, IntPtrTy, "sunkaddr");
978 V = Builder.CreateSExt(V, IntPtrTy, "sunkaddr");
980 if (AddrMode.Scale != 1)
981 V = Builder.CreateMul(V, ConstantInt::get(IntPtrTy, AddrMode.Scale),
984 Result = Builder.CreateAdd(Result, V, "sunkaddr");
989 // Add in the BaseGV if present.
990 if (AddrMode.BaseGV) {
991 Value *V = Builder.CreatePtrToInt(AddrMode.BaseGV, IntPtrTy, "sunkaddr");
993 Result = Builder.CreateAdd(Result, V, "sunkaddr");
998 // Add in the Base Offset if present.
999 if (AddrMode.BaseOffs) {
1000 Value *V = ConstantInt::get(IntPtrTy, AddrMode.BaseOffs);
1002 Result = Builder.CreateAdd(Result, V, "sunkaddr");
1008 SunkAddr = Constant::getNullValue(Addr->getType());
1010 SunkAddr = Builder.CreateIntToPtr(Result, Addr->getType(), "sunkaddr");
1013 MemoryInst->replaceUsesOfWith(Repl, SunkAddr);
1015 // If we have no uses, recursively delete the value and all dead instructions
1017 if (Repl->use_empty()) {
1018 // This can cause recursive deletion, which can invalidate our iterator.
1019 // Use a WeakVH to hold onto it in case this happens.
1020 WeakVH IterHandle(CurInstIterator);
1021 BasicBlock *BB = CurInstIterator->getParent();
1023 RecursivelyDeleteTriviallyDeadInstructions(Repl, TLInfo);
1025 if (IterHandle != CurInstIterator) {
1026 // If the iterator instruction was recursively deleted, start over at the
1027 // start of the block.
1028 CurInstIterator = BB->begin();
1031 // This address is now available for reassignment, so erase the table
1032 // entry; we don't want to match some completely different instruction.
1033 SunkAddrs[Addr] = 0;
1040 /// OptimizeInlineAsmInst - If there are any memory operands, use
1041 /// OptimizeMemoryInst to sink their address computing into the block when
1042 /// possible / profitable.
1043 bool CodeGenPrepare::OptimizeInlineAsmInst(CallInst *CS) {
1044 bool MadeChange = false;
1046 TargetLowering::AsmOperandInfoVector
1047 TargetConstraints = TLI->ParseConstraints(CS);
1049 for (unsigned i = 0, e = TargetConstraints.size(); i != e; ++i) {
1050 TargetLowering::AsmOperandInfo &OpInfo = TargetConstraints[i];
1052 // Compute the constraint code and ConstraintType to use.
1053 TLI->ComputeConstraintToUse(OpInfo, SDValue());
1055 if (OpInfo.ConstraintType == TargetLowering::C_Memory &&
1056 OpInfo.isIndirect) {
1057 Value *OpVal = CS->getArgOperand(ArgNo++);
1058 MadeChange |= OptimizeMemoryInst(CS, OpVal, OpVal->getType());
1059 } else if (OpInfo.Type == InlineAsm::isInput)
1066 /// MoveExtToFormExtLoad - Move a zext or sext fed by a load into the same
1067 /// basic block as the load, unless conditions are unfavorable. This allows
1068 /// SelectionDAG to fold the extend into the load.
1070 bool CodeGenPrepare::MoveExtToFormExtLoad(Instruction *I) {
1071 // Look for a load being extended.
1072 LoadInst *LI = dyn_cast<LoadInst>(I->getOperand(0));
1073 if (!LI) return false;
1075 // If they're already in the same block, there's nothing to do.
1076 if (LI->getParent() == I->getParent())
1079 // If the load has other users and the truncate is not free, this probably
1080 // isn't worthwhile.
1081 if (!LI->hasOneUse() &&
1082 TLI && (TLI->isTypeLegal(TLI->getValueType(LI->getType())) ||
1083 !TLI->isTypeLegal(TLI->getValueType(I->getType()))) &&
1084 !TLI->isTruncateFree(I->getType(), LI->getType()))
1087 // Check whether the target supports casts folded into loads.
1089 if (isa<ZExtInst>(I))
1090 LType = ISD::ZEXTLOAD;
1092 assert(isa<SExtInst>(I) && "Unexpected ext type!");
1093 LType = ISD::SEXTLOAD;
1095 if (TLI && !TLI->isLoadExtLegal(LType, TLI->getValueType(LI->getType())))
1098 // Move the extend into the same block as the load, so that SelectionDAG
1100 I->removeFromParent();
1106 bool CodeGenPrepare::OptimizeExtUses(Instruction *I) {
1107 BasicBlock *DefBB = I->getParent();
1109 // If the result of a {s|z}ext and its source are both live out, rewrite all
1110 // other uses of the source with result of extension.
1111 Value *Src = I->getOperand(0);
1112 if (Src->hasOneUse())
1115 // Only do this xform if truncating is free.
1116 if (TLI && !TLI->isTruncateFree(I->getType(), Src->getType()))
1119 // Only safe to perform the optimization if the source is also defined in
1121 if (!isa<Instruction>(Src) || DefBB != cast<Instruction>(Src)->getParent())
1124 bool DefIsLiveOut = false;
1125 for (Value::use_iterator UI = I->use_begin(), E = I->use_end();
1127 Instruction *User = cast<Instruction>(*UI);
1129 // Figure out which BB this ext is used in.
1130 BasicBlock *UserBB = User->getParent();
1131 if (UserBB == DefBB) continue;
1132 DefIsLiveOut = true;
1138 // Make sure non of the uses are PHI nodes.
1139 for (Value::use_iterator UI = Src->use_begin(), E = Src->use_end();
1141 Instruction *User = cast<Instruction>(*UI);
1142 BasicBlock *UserBB = User->getParent();
1143 if (UserBB == DefBB) continue;
1144 // Be conservative. We don't want this xform to end up introducing
1145 // reloads just before load / store instructions.
1146 if (isa<PHINode>(User) || isa<LoadInst>(User) || isa<StoreInst>(User))
1150 // InsertedTruncs - Only insert one trunc in each block once.
1151 DenseMap<BasicBlock*, Instruction*> InsertedTruncs;
1153 bool MadeChange = false;
1154 for (Value::use_iterator UI = Src->use_begin(), E = Src->use_end();
1156 Use &TheUse = UI.getUse();
1157 Instruction *User = cast<Instruction>(*UI);
1159 // Figure out which BB this ext is used in.
1160 BasicBlock *UserBB = User->getParent();
1161 if (UserBB == DefBB) continue;
1163 // Both src and def are live in this block. Rewrite the use.
1164 Instruction *&InsertedTrunc = InsertedTruncs[UserBB];
1166 if (!InsertedTrunc) {
1167 BasicBlock::iterator InsertPt = UserBB->getFirstInsertionPt();
1168 InsertedTrunc = new TruncInst(I, Src->getType(), "", InsertPt);
1171 // Replace a use of the {s|z}ext source with a use of the result.
1172 TheUse = InsertedTrunc;
1180 /// isFormingBranchFromSelectProfitable - Returns true if a SelectInst should be
1181 /// turned into an explicit branch.
1182 static bool isFormingBranchFromSelectProfitable(SelectInst *SI) {
1183 // FIXME: This should use the same heuristics as IfConversion to determine
1184 // whether a select is better represented as a branch. This requires that
1185 // branch probability metadata is preserved for the select, which is not the
1188 CmpInst *Cmp = dyn_cast<CmpInst>(SI->getCondition());
1190 // If the branch is predicted right, an out of order CPU can avoid blocking on
1191 // the compare. Emit cmovs on compares with a memory operand as branches to
1192 // avoid stalls on the load from memory. If the compare has more than one use
1193 // there's probably another cmov or setcc around so it's not worth emitting a
1198 Value *CmpOp0 = Cmp->getOperand(0);
1199 Value *CmpOp1 = Cmp->getOperand(1);
1201 // We check that the memory operand has one use to avoid uses of the loaded
1202 // value directly after the compare, making branches unprofitable.
1203 return Cmp->hasOneUse() &&
1204 ((isa<LoadInst>(CmpOp0) && CmpOp0->hasOneUse()) ||
1205 (isa<LoadInst>(CmpOp1) && CmpOp1->hasOneUse()));
1209 /// If we have a SelectInst that will likely profit from branch prediction,
1210 /// turn it into a branch.
1211 bool CodeGenPrepare::OptimizeSelectInst(SelectInst *SI) {
1212 bool VectorCond = !SI->getCondition()->getType()->isIntegerTy(1);
1214 // Can we convert the 'select' to CF ?
1215 if (DisableSelectToBranch || OptSize || !TLI || VectorCond)
1218 TargetLowering::SelectSupportKind SelectKind;
1220 SelectKind = TargetLowering::VectorMaskSelect;
1221 else if (SI->getType()->isVectorTy())
1222 SelectKind = TargetLowering::ScalarCondVectorVal;
1224 SelectKind = TargetLowering::ScalarValSelect;
1226 // Do we have efficient codegen support for this kind of 'selects' ?
1227 if (TLI->isSelectSupported(SelectKind)) {
1228 // We have efficient codegen support for the select instruction.
1229 // Check if it is profitable to keep this 'select'.
1230 if (!TLI->isPredictableSelectExpensive() ||
1231 !isFormingBranchFromSelectProfitable(SI))
1237 // First, we split the block containing the select into 2 blocks.
1238 BasicBlock *StartBlock = SI->getParent();
1239 BasicBlock::iterator SplitPt = ++(BasicBlock::iterator(SI));
1240 BasicBlock *NextBlock = StartBlock->splitBasicBlock(SplitPt, "select.end");
1242 // Create a new block serving as the landing pad for the branch.
1243 BasicBlock *SmallBlock = BasicBlock::Create(SI->getContext(), "select.mid",
1244 NextBlock->getParent(), NextBlock);
1246 // Move the unconditional branch from the block with the select in it into our
1247 // landing pad block.
1248 StartBlock->getTerminator()->eraseFromParent();
1249 BranchInst::Create(NextBlock, SmallBlock);
1251 // Insert the real conditional branch based on the original condition.
1252 BranchInst::Create(NextBlock, SmallBlock, SI->getCondition(), SI);
1254 // The select itself is replaced with a PHI Node.
1255 PHINode *PN = PHINode::Create(SI->getType(), 2, "", NextBlock->begin());
1257 PN->addIncoming(SI->getTrueValue(), StartBlock);
1258 PN->addIncoming(SI->getFalseValue(), SmallBlock);
1259 SI->replaceAllUsesWith(PN);
1260 SI->eraseFromParent();
1262 // Instruct OptimizeBlock to skip to the next block.
1263 CurInstIterator = StartBlock->end();
1264 ++NumSelectsExpanded;
1268 bool CodeGenPrepare::OptimizeInst(Instruction *I) {
1269 if (PHINode *P = dyn_cast<PHINode>(I)) {
1270 // It is possible for very late stage optimizations (such as SimplifyCFG)
1271 // to introduce PHI nodes too late to be cleaned up. If we detect such a
1272 // trivial PHI, go ahead and zap it here.
1273 if (Value *V = SimplifyInstruction(P)) {
1274 P->replaceAllUsesWith(V);
1275 P->eraseFromParent();
1282 if (CastInst *CI = dyn_cast<CastInst>(I)) {
1283 // If the source of the cast is a constant, then this should have
1284 // already been constant folded. The only reason NOT to constant fold
1285 // it is if something (e.g. LSR) was careful to place the constant
1286 // evaluation in a block other than then one that uses it (e.g. to hoist
1287 // the address of globals out of a loop). If this is the case, we don't
1288 // want to forward-subst the cast.
1289 if (isa<Constant>(CI->getOperand(0)))
1292 if (TLI && OptimizeNoopCopyExpression(CI, *TLI))
1295 if (isa<ZExtInst>(I) || isa<SExtInst>(I)) {
1296 bool MadeChange = MoveExtToFormExtLoad(I);
1297 return MadeChange | OptimizeExtUses(I);
1302 if (CmpInst *CI = dyn_cast<CmpInst>(I))
1303 return OptimizeCmpExpression(CI);
1305 if (LoadInst *LI = dyn_cast<LoadInst>(I)) {
1307 return OptimizeMemoryInst(I, I->getOperand(0), LI->getType());
1311 if (StoreInst *SI = dyn_cast<StoreInst>(I)) {
1313 return OptimizeMemoryInst(I, SI->getOperand(1),
1314 SI->getOperand(0)->getType());
1318 if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(I)) {
1319 if (GEPI->hasAllZeroIndices()) {
1320 /// The GEP operand must be a pointer, so must its result -> BitCast
1321 Instruction *NC = new BitCastInst(GEPI->getOperand(0), GEPI->getType(),
1322 GEPI->getName(), GEPI);
1323 GEPI->replaceAllUsesWith(NC);
1324 GEPI->eraseFromParent();
1332 if (CallInst *CI = dyn_cast<CallInst>(I))
1333 return OptimizeCallInst(CI);
1335 if (SelectInst *SI = dyn_cast<SelectInst>(I))
1336 return OptimizeSelectInst(SI);
1341 // In this pass we look for GEP and cast instructions that are used
1342 // across basic blocks and rewrite them to improve basic-block-at-a-time
1344 bool CodeGenPrepare::OptimizeBlock(BasicBlock &BB) {
1346 bool MadeChange = false;
1348 CurInstIterator = BB.begin();
1349 while (CurInstIterator != BB.end())
1350 MadeChange |= OptimizeInst(CurInstIterator++);
1352 MadeChange |= DupRetToEnableTailCallOpts(&BB);
1357 // llvm.dbg.value is far away from the value then iSel may not be able
1358 // handle it properly. iSel will drop llvm.dbg.value if it can not
1359 // find a node corresponding to the value.
1360 bool CodeGenPrepare::PlaceDbgValues(Function &F) {
1361 bool MadeChange = false;
1362 for (Function::iterator I = F.begin(), E = F.end(); I != E; ++I) {
1363 Instruction *PrevNonDbgInst = NULL;
1364 for (BasicBlock::iterator BI = I->begin(), BE = I->end(); BI != BE;) {
1365 Instruction *Insn = BI; ++BI;
1366 DbgValueInst *DVI = dyn_cast<DbgValueInst>(Insn);
1368 PrevNonDbgInst = Insn;
1372 Instruction *VI = dyn_cast_or_null<Instruction>(DVI->getValue());
1373 if (VI && VI != PrevNonDbgInst && !VI->isTerminator()) {
1374 DEBUG(dbgs() << "Moving Debug Value before :\n" << *DVI << ' ' << *VI);
1375 DVI->removeFromParent();
1376 if (isa<PHINode>(VI))
1377 DVI->insertBefore(VI->getParent()->getFirstInsertionPt());
1379 DVI->insertAfter(VI);