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/ADT/ValueMap.h"
22 #include "llvm/Analysis/DominatorInternals.h"
23 #include "llvm/Analysis/Dominators.h"
24 #include "llvm/Analysis/InstructionSimplify.h"
25 #include "llvm/Analysis/ProfileInfo.h"
26 #include "llvm/Assembly/Writer.h"
27 #include "llvm/IR/Constants.h"
28 #include "llvm/IR/DataLayout.h"
29 #include "llvm/IR/DerivedTypes.h"
30 #include "llvm/IR/Function.h"
31 #include "llvm/IR/IRBuilder.h"
32 #include "llvm/IR/InlineAsm.h"
33 #include "llvm/IR/Instructions.h"
34 #include "llvm/IR/IntrinsicInst.h"
35 #include "llvm/Pass.h"
36 #include "llvm/Support/CallSite.h"
37 #include "llvm/Support/CommandLine.h"
38 #include "llvm/Support/Debug.h"
39 #include "llvm/Support/GetElementPtrTypeIterator.h"
40 #include "llvm/Support/PatternMatch.h"
41 #include "llvm/Support/ValueHandle.h"
42 #include "llvm/Support/raw_ostream.h"
43 #include "llvm/Target/TargetLibraryInfo.h"
44 #include "llvm/Target/TargetLowering.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 TargetMachine *TM;
80 const TargetLowering *TLI;
81 const TargetLibraryInfo *TLInfo;
85 /// CurInstIterator - As we scan instructions optimizing them, this is the
86 /// next instruction to optimize. Xforms that can invalidate this should
88 BasicBlock::iterator CurInstIterator;
90 /// Keeps track of non-local addresses that have been sunk into a block.
91 /// This allows us to avoid inserting duplicate code for blocks with
92 /// multiple load/stores of the same address.
93 ValueMap<Value*, Value*> SunkAddrs;
95 /// ModifiedDT - If CFG is modified in anyway, dominator tree may need to
99 /// OptSize - True if optimizing for size.
103 static char ID; // Pass identification, replacement for typeid
104 explicit CodeGenPrepare(const TargetMachine *TM = 0)
105 : FunctionPass(ID), TM(TM), TLI(0) {
106 initializeCodeGenPreparePass(*PassRegistry::getPassRegistry());
108 bool runOnFunction(Function &F);
110 const char *getPassName() const { return "CodeGen Prepare"; }
112 virtual void getAnalysisUsage(AnalysisUsage &AU) const {
113 AU.addPreserved<DominatorTree>();
114 AU.addPreserved<ProfileInfo>();
115 AU.addRequired<TargetLibraryInfo>();
119 bool EliminateFallThrough(Function &F);
120 bool EliminateMostlyEmptyBlocks(Function &F);
121 bool CanMergeBlocks(const BasicBlock *BB, const BasicBlock *DestBB) const;
122 void EliminateMostlyEmptyBlock(BasicBlock *BB);
123 bool OptimizeBlock(BasicBlock &BB);
124 bool OptimizeInst(Instruction *I);
125 bool OptimizeMemoryInst(Instruction *I, Value *Addr, Type *AccessTy);
126 bool OptimizeInlineAsmInst(CallInst *CS);
127 bool OptimizeCallInst(CallInst *CI);
128 bool MoveExtToFormExtLoad(Instruction *I);
129 bool OptimizeExtUses(Instruction *I);
130 bool OptimizeSelectInst(SelectInst *SI);
131 bool DupRetToEnableTailCallOpts(BasicBlock *BB);
132 bool PlaceDbgValues(Function &F);
136 char CodeGenPrepare::ID = 0;
137 INITIALIZE_PASS_BEGIN(CodeGenPrepare, "codegenprepare",
138 "Optimize for code generation", false, false)
139 INITIALIZE_PASS_DEPENDENCY(TargetLibraryInfo)
140 INITIALIZE_PASS_END(CodeGenPrepare, "codegenprepare",
141 "Optimize for code generation", false, false)
143 FunctionPass *llvm::createCodeGenPreparePass(const TargetMachine *TM) {
144 return new CodeGenPrepare(TM);
147 bool CodeGenPrepare::runOnFunction(Function &F) {
148 bool EverMadeChange = false;
151 if (TM) TLI = TM->getTargetLowering();
152 TLInfo = &getAnalysis<TargetLibraryInfo>();
153 DT = getAnalysisIfAvailable<DominatorTree>();
154 PFI = getAnalysisIfAvailable<ProfileInfo>();
155 OptSize = F.getAttributes().hasAttribute(AttributeSet::FunctionIndex,
156 Attribute::OptimizeForSize);
158 /// This optimization identifies DIV instructions that can be
159 /// profitably bypassed and carried out with a shorter, faster divide.
160 if (!OptSize && TLI && TLI->isSlowDivBypassed()) {
161 const DenseMap<unsigned int, unsigned int> &BypassWidths =
162 TLI->getBypassSlowDivWidths();
163 for (Function::iterator I = F.begin(); I != F.end(); I++)
164 EverMadeChange |= bypassSlowDivision(F, I, BypassWidths);
167 // Eliminate blocks that contain only PHI nodes and an
168 // unconditional branch.
169 EverMadeChange |= EliminateMostlyEmptyBlocks(F);
171 // llvm.dbg.value is far away from the value then iSel may not be able
172 // handle it properly. iSel will drop llvm.dbg.value if it can not
173 // find a node corresponding to the value.
174 EverMadeChange |= PlaceDbgValues(F);
176 bool MadeChange = true;
179 for (Function::iterator I = F.begin(); I != F.end(); ) {
180 BasicBlock *BB = I++;
181 MadeChange |= OptimizeBlock(*BB);
183 EverMadeChange |= MadeChange;
188 if (!DisableBranchOpts) {
190 SmallPtrSet<BasicBlock*, 8> WorkList;
191 for (Function::iterator BB = F.begin(), E = F.end(); BB != E; ++BB) {
192 SmallVector<BasicBlock*, 2> Successors(succ_begin(BB), succ_end(BB));
193 MadeChange |= ConstantFoldTerminator(BB, true);
194 if (!MadeChange) continue;
196 for (SmallVectorImpl<BasicBlock*>::iterator
197 II = Successors.begin(), IE = Successors.end(); II != IE; ++II)
198 if (pred_begin(*II) == pred_end(*II))
199 WorkList.insert(*II);
202 // Delete the dead blocks and any of their dead successors.
203 MadeChange |= !WorkList.empty();
204 while (!WorkList.empty()) {
205 BasicBlock *BB = *WorkList.begin();
207 SmallVector<BasicBlock*, 2> Successors(succ_begin(BB), succ_end(BB));
211 for (SmallVectorImpl<BasicBlock*>::iterator
212 II = Successors.begin(), IE = Successors.end(); II != IE; ++II)
213 if (pred_begin(*II) == pred_end(*II))
214 WorkList.insert(*II);
217 // Merge pairs of basic blocks with unconditional branches, connected by
219 if (EverMadeChange || MadeChange)
220 MadeChange |= EliminateFallThrough(F);
224 EverMadeChange |= MadeChange;
227 if (ModifiedDT && DT)
228 DT->DT->recalculate(F);
230 return EverMadeChange;
233 /// EliminateFallThrough - Merge basic blocks which are connected
234 /// by a single edge, where one of the basic blocks has a single successor
235 /// pointing to the other basic block, which has a single predecessor.
236 bool CodeGenPrepare::EliminateFallThrough(Function &F) {
237 bool Changed = false;
238 // Scan all of the blocks in the function, except for the entry block.
239 for (Function::iterator I = ++F.begin(), E = F.end(); I != E; ) {
240 BasicBlock *BB = I++;
241 // If the destination block has a single pred, then this is a trivial
242 // edge, just collapse it.
243 BasicBlock *SinglePred = BB->getSinglePredecessor();
245 // Don't merge if BB's address is taken.
246 if (!SinglePred || SinglePred == BB || BB->hasAddressTaken()) continue;
248 BranchInst *Term = dyn_cast<BranchInst>(SinglePred->getTerminator());
249 if (Term && !Term->isConditional()) {
251 DEBUG(dbgs() << "To merge:\n"<< *SinglePred << "\n\n\n");
252 // Remember if SinglePred was the entry block of the function.
253 // If so, we will need to move BB back to the entry position.
254 bool isEntry = SinglePred == &SinglePred->getParent()->getEntryBlock();
255 MergeBasicBlockIntoOnlyPred(BB, this);
257 if (isEntry && BB != &BB->getParent()->getEntryBlock())
258 BB->moveBefore(&BB->getParent()->getEntryBlock());
260 // We have erased a block. Update the iterator.
267 /// EliminateMostlyEmptyBlocks - eliminate blocks that contain only PHI nodes,
268 /// debug info directives, and an unconditional branch. Passes before isel
269 /// (e.g. LSR/loopsimplify) often split edges in ways that are non-optimal for
270 /// isel. Start by eliminating these blocks so we can split them the way we
272 bool CodeGenPrepare::EliminateMostlyEmptyBlocks(Function &F) {
273 bool MadeChange = false;
274 // Note that this intentionally skips the entry block.
275 for (Function::iterator I = ++F.begin(), E = F.end(); I != E; ) {
276 BasicBlock *BB = I++;
278 // If this block doesn't end with an uncond branch, ignore it.
279 BranchInst *BI = dyn_cast<BranchInst>(BB->getTerminator());
280 if (!BI || !BI->isUnconditional())
283 // If the instruction before the branch (skipping debug info) isn't a phi
284 // node, then other stuff is happening here.
285 BasicBlock::iterator BBI = BI;
286 if (BBI != BB->begin()) {
288 while (isa<DbgInfoIntrinsic>(BBI)) {
289 if (BBI == BB->begin())
293 if (!isa<DbgInfoIntrinsic>(BBI) && !isa<PHINode>(BBI))
297 // Do not break infinite loops.
298 BasicBlock *DestBB = BI->getSuccessor(0);
302 if (!CanMergeBlocks(BB, DestBB))
305 EliminateMostlyEmptyBlock(BB);
311 /// CanMergeBlocks - Return true if we can merge BB into DestBB if there is a
312 /// single uncond branch between them, and BB contains no other non-phi
314 bool CodeGenPrepare::CanMergeBlocks(const BasicBlock *BB,
315 const BasicBlock *DestBB) const {
316 // We only want to eliminate blocks whose phi nodes are used by phi nodes in
317 // the successor. If there are more complex condition (e.g. preheaders),
318 // don't mess around with them.
319 BasicBlock::const_iterator BBI = BB->begin();
320 while (const PHINode *PN = dyn_cast<PHINode>(BBI++)) {
321 for (Value::const_use_iterator UI = PN->use_begin(), E = PN->use_end();
323 const Instruction *User = cast<Instruction>(*UI);
324 if (User->getParent() != DestBB || !isa<PHINode>(User))
326 // If User is inside DestBB block and it is a PHINode then check
327 // incoming value. If incoming value is not from BB then this is
328 // a complex condition (e.g. preheaders) we want to avoid here.
329 if (User->getParent() == DestBB) {
330 if (const PHINode *UPN = dyn_cast<PHINode>(User))
331 for (unsigned I = 0, E = UPN->getNumIncomingValues(); I != E; ++I) {
332 Instruction *Insn = dyn_cast<Instruction>(UPN->getIncomingValue(I));
333 if (Insn && Insn->getParent() == BB &&
334 Insn->getParent() != UPN->getIncomingBlock(I))
341 // If BB and DestBB contain any common predecessors, then the phi nodes in BB
342 // and DestBB may have conflicting incoming values for the block. If so, we
343 // can't merge the block.
344 const PHINode *DestBBPN = dyn_cast<PHINode>(DestBB->begin());
345 if (!DestBBPN) return true; // no conflict.
347 // Collect the preds of BB.
348 SmallPtrSet<const BasicBlock*, 16> BBPreds;
349 if (const PHINode *BBPN = dyn_cast<PHINode>(BB->begin())) {
350 // It is faster to get preds from a PHI than with pred_iterator.
351 for (unsigned i = 0, e = BBPN->getNumIncomingValues(); i != e; ++i)
352 BBPreds.insert(BBPN->getIncomingBlock(i));
354 BBPreds.insert(pred_begin(BB), pred_end(BB));
357 // Walk the preds of DestBB.
358 for (unsigned i = 0, e = DestBBPN->getNumIncomingValues(); i != e; ++i) {
359 BasicBlock *Pred = DestBBPN->getIncomingBlock(i);
360 if (BBPreds.count(Pred)) { // Common predecessor?
361 BBI = DestBB->begin();
362 while (const PHINode *PN = dyn_cast<PHINode>(BBI++)) {
363 const Value *V1 = PN->getIncomingValueForBlock(Pred);
364 const Value *V2 = PN->getIncomingValueForBlock(BB);
366 // If V2 is a phi node in BB, look up what the mapped value will be.
367 if (const PHINode *V2PN = dyn_cast<PHINode>(V2))
368 if (V2PN->getParent() == BB)
369 V2 = V2PN->getIncomingValueForBlock(Pred);
371 // If there is a conflict, bail out.
372 if (V1 != V2) return false;
381 /// EliminateMostlyEmptyBlock - Eliminate a basic block that have only phi's and
382 /// an unconditional branch in it.
383 void CodeGenPrepare::EliminateMostlyEmptyBlock(BasicBlock *BB) {
384 BranchInst *BI = cast<BranchInst>(BB->getTerminator());
385 BasicBlock *DestBB = BI->getSuccessor(0);
387 DEBUG(dbgs() << "MERGING MOSTLY EMPTY BLOCKS - BEFORE:\n" << *BB << *DestBB);
389 // If the destination block has a single pred, then this is a trivial edge,
391 if (BasicBlock *SinglePred = DestBB->getSinglePredecessor()) {
392 if (SinglePred != DestBB) {
393 // Remember if SinglePred was the entry block of the function. If so, we
394 // will need to move BB back to the entry position.
395 bool isEntry = SinglePred == &SinglePred->getParent()->getEntryBlock();
396 MergeBasicBlockIntoOnlyPred(DestBB, this);
398 if (isEntry && BB != &BB->getParent()->getEntryBlock())
399 BB->moveBefore(&BB->getParent()->getEntryBlock());
401 DEBUG(dbgs() << "AFTER:\n" << *DestBB << "\n\n\n");
406 // Otherwise, we have multiple predecessors of BB. Update the PHIs in DestBB
407 // to handle the new incoming edges it is about to have.
409 for (BasicBlock::iterator BBI = DestBB->begin();
410 (PN = dyn_cast<PHINode>(BBI)); ++BBI) {
411 // Remove the incoming value for BB, and remember it.
412 Value *InVal = PN->removeIncomingValue(BB, false);
414 // Two options: either the InVal is a phi node defined in BB or it is some
415 // value that dominates BB.
416 PHINode *InValPhi = dyn_cast<PHINode>(InVal);
417 if (InValPhi && InValPhi->getParent() == BB) {
418 // Add all of the input values of the input PHI as inputs of this phi.
419 for (unsigned i = 0, e = InValPhi->getNumIncomingValues(); i != e; ++i)
420 PN->addIncoming(InValPhi->getIncomingValue(i),
421 InValPhi->getIncomingBlock(i));
423 // Otherwise, add one instance of the dominating value for each edge that
424 // we will be adding.
425 if (PHINode *BBPN = dyn_cast<PHINode>(BB->begin())) {
426 for (unsigned i = 0, e = BBPN->getNumIncomingValues(); i != e; ++i)
427 PN->addIncoming(InVal, BBPN->getIncomingBlock(i));
429 for (pred_iterator PI = pred_begin(BB), E = pred_end(BB); PI != E; ++PI)
430 PN->addIncoming(InVal, *PI);
435 // The PHIs are now updated, change everything that refers to BB to use
436 // DestBB and remove BB.
437 BB->replaceAllUsesWith(DestBB);
438 if (DT && !ModifiedDT) {
439 BasicBlock *BBIDom = DT->getNode(BB)->getIDom()->getBlock();
440 BasicBlock *DestBBIDom = DT->getNode(DestBB)->getIDom()->getBlock();
441 BasicBlock *NewIDom = DT->findNearestCommonDominator(BBIDom, DestBBIDom);
442 DT->changeImmediateDominator(DestBB, NewIDom);
446 PFI->replaceAllUses(BB, DestBB);
447 PFI->removeEdge(ProfileInfo::getEdge(BB, DestBB));
449 BB->eraseFromParent();
452 DEBUG(dbgs() << "AFTER:\n" << *DestBB << "\n\n\n");
455 /// OptimizeNoopCopyExpression - If the specified cast instruction is a noop
456 /// copy (e.g. it's casting from one pointer type to another, i32->i8 on PPC),
457 /// sink it into user blocks to reduce the number of virtual
458 /// registers that must be created and coalesced.
460 /// Return true if any changes are made.
462 static bool OptimizeNoopCopyExpression(CastInst *CI, const TargetLowering &TLI){
463 // If this is a noop copy,
464 EVT SrcVT = TLI.getValueType(CI->getOperand(0)->getType());
465 EVT DstVT = TLI.getValueType(CI->getType());
467 // This is an fp<->int conversion?
468 if (SrcVT.isInteger() != DstVT.isInteger())
471 // If this is an extension, it will be a zero or sign extension, which
473 if (SrcVT.bitsLT(DstVT)) return false;
475 // If these values will be promoted, find out what they will be promoted
476 // to. This helps us consider truncates on PPC as noop copies when they
478 if (TLI.getTypeAction(CI->getContext(), SrcVT) ==
479 TargetLowering::TypePromoteInteger)
480 SrcVT = TLI.getTypeToTransformTo(CI->getContext(), SrcVT);
481 if (TLI.getTypeAction(CI->getContext(), DstVT) ==
482 TargetLowering::TypePromoteInteger)
483 DstVT = TLI.getTypeToTransformTo(CI->getContext(), DstVT);
485 // If, after promotion, these are the same types, this is a noop copy.
489 BasicBlock *DefBB = CI->getParent();
491 /// InsertedCasts - Only insert a cast in each block once.
492 DenseMap<BasicBlock*, CastInst*> InsertedCasts;
494 bool MadeChange = false;
495 for (Value::use_iterator UI = CI->use_begin(), E = CI->use_end();
497 Use &TheUse = UI.getUse();
498 Instruction *User = cast<Instruction>(*UI);
500 // Figure out which BB this cast is used in. For PHI's this is the
501 // appropriate predecessor block.
502 BasicBlock *UserBB = User->getParent();
503 if (PHINode *PN = dyn_cast<PHINode>(User)) {
504 UserBB = PN->getIncomingBlock(UI);
507 // Preincrement use iterator so we don't invalidate it.
510 // If this user is in the same block as the cast, don't change the cast.
511 if (UserBB == DefBB) continue;
513 // If we have already inserted a cast into this block, use it.
514 CastInst *&InsertedCast = InsertedCasts[UserBB];
517 BasicBlock::iterator InsertPt = UserBB->getFirstInsertionPt();
519 CastInst::Create(CI->getOpcode(), CI->getOperand(0), CI->getType(), "",
524 // Replace a use of the cast with a use of the new cast.
525 TheUse = InsertedCast;
529 // If we removed all uses, nuke the cast.
530 if (CI->use_empty()) {
531 CI->eraseFromParent();
538 /// OptimizeCmpExpression - sink the given CmpInst into user blocks to reduce
539 /// the number of virtual registers that must be created and coalesced. This is
540 /// a clear win except on targets with multiple condition code registers
541 /// (PowerPC), where it might lose; some adjustment may be wanted there.
543 /// Return true if any changes are made.
544 static bool OptimizeCmpExpression(CmpInst *CI) {
545 BasicBlock *DefBB = CI->getParent();
547 /// InsertedCmp - Only insert a cmp in each block once.
548 DenseMap<BasicBlock*, CmpInst*> InsertedCmps;
550 bool MadeChange = false;
551 for (Value::use_iterator UI = CI->use_begin(), E = CI->use_end();
553 Use &TheUse = UI.getUse();
554 Instruction *User = cast<Instruction>(*UI);
556 // Preincrement use iterator so we don't invalidate it.
559 // Don't bother for PHI nodes.
560 if (isa<PHINode>(User))
563 // Figure out which BB this cmp is used in.
564 BasicBlock *UserBB = User->getParent();
566 // If this user is in the same block as the cmp, don't change the cmp.
567 if (UserBB == DefBB) continue;
569 // If we have already inserted a cmp into this block, use it.
570 CmpInst *&InsertedCmp = InsertedCmps[UserBB];
573 BasicBlock::iterator InsertPt = UserBB->getFirstInsertionPt();
575 CmpInst::Create(CI->getOpcode(),
576 CI->getPredicate(), CI->getOperand(0),
577 CI->getOperand(1), "", InsertPt);
581 // Replace a use of the cmp with a use of the new cmp.
582 TheUse = InsertedCmp;
586 // If we removed all uses, nuke the cmp.
588 CI->eraseFromParent();
594 class CodeGenPrepareFortifiedLibCalls : public SimplifyFortifiedLibCalls {
596 void replaceCall(Value *With) {
597 CI->replaceAllUsesWith(With);
598 CI->eraseFromParent();
600 bool isFoldable(unsigned SizeCIOp, unsigned, bool) const {
601 if (ConstantInt *SizeCI =
602 dyn_cast<ConstantInt>(CI->getArgOperand(SizeCIOp)))
603 return SizeCI->isAllOnesValue();
607 } // end anonymous namespace
609 bool CodeGenPrepare::OptimizeCallInst(CallInst *CI) {
610 BasicBlock *BB = CI->getParent();
612 // Lower inline assembly if we can.
613 // If we found an inline asm expession, and if the target knows how to
614 // lower it to normal LLVM code, do so now.
615 if (TLI && isa<InlineAsm>(CI->getCalledValue())) {
616 if (TLI->ExpandInlineAsm(CI)) {
617 // Avoid invalidating the iterator.
618 CurInstIterator = BB->begin();
619 // Avoid processing instructions out of order, which could cause
620 // reuse before a value is defined.
624 // Sink address computing for memory operands into the block.
625 if (OptimizeInlineAsmInst(CI))
629 // Lower all uses of llvm.objectsize.*
630 IntrinsicInst *II = dyn_cast<IntrinsicInst>(CI);
631 if (II && II->getIntrinsicID() == Intrinsic::objectsize) {
632 bool Min = (cast<ConstantInt>(II->getArgOperand(1))->getZExtValue() == 1);
633 Type *ReturnTy = CI->getType();
634 Constant *RetVal = ConstantInt::get(ReturnTy, Min ? 0 : -1ULL);
636 // Substituting this can cause recursive simplifications, which can
637 // invalidate our iterator. Use a WeakVH to hold onto it in case this
639 WeakVH IterHandle(CurInstIterator);
641 replaceAndRecursivelySimplify(CI, RetVal, TLI ? TLI->getDataLayout() : 0,
642 TLInfo, ModifiedDT ? 0 : DT);
644 // If the iterator instruction was recursively deleted, start over at the
645 // start of the block.
646 if (IterHandle != CurInstIterator) {
647 CurInstIterator = BB->begin();
654 SmallVector<Value*, 2> PtrOps;
656 if (TLI->GetAddrModeArguments(II, PtrOps, AccessTy))
657 while (!PtrOps.empty())
658 if (OptimizeMemoryInst(II, PtrOps.pop_back_val(), AccessTy))
662 // From here on out we're working with named functions.
663 if (CI->getCalledFunction() == 0) return false;
665 // We'll need DataLayout from here on out.
666 const DataLayout *TD = TLI ? TLI->getDataLayout() : 0;
667 if (!TD) return false;
669 // Lower all default uses of _chk calls. This is very similar
670 // to what InstCombineCalls does, but here we are only lowering calls
671 // that have the default "don't know" as the objectsize. Anything else
672 // should be left alone.
673 CodeGenPrepareFortifiedLibCalls Simplifier;
674 return Simplifier.fold(CI, TD, TLInfo);
677 /// DupRetToEnableTailCallOpts - Look for opportunities to duplicate return
678 /// instructions to the predecessor to enable tail call optimizations. The
679 /// case it is currently looking for is:
682 /// %tmp0 = tail call i32 @f0()
685 /// %tmp1 = tail call i32 @f1()
688 /// %tmp2 = tail call i32 @f2()
691 /// %retval = phi i32 [ %tmp0, %bb0 ], [ %tmp1, %bb1 ], [ %tmp2, %bb2 ]
699 /// %tmp0 = tail call i32 @f0()
702 /// %tmp1 = tail call i32 @f1()
705 /// %tmp2 = tail call i32 @f2()
708 bool CodeGenPrepare::DupRetToEnableTailCallOpts(BasicBlock *BB) {
712 ReturnInst *RI = dyn_cast<ReturnInst>(BB->getTerminator());
717 BitCastInst *BCI = 0;
718 Value *V = RI->getReturnValue();
720 BCI = dyn_cast<BitCastInst>(V);
722 V = BCI->getOperand(0);
724 PN = dyn_cast<PHINode>(V);
729 if (PN && PN->getParent() != BB)
732 // It's not safe to eliminate the sign / zero extension of the return value.
733 // See llvm::isInTailCallPosition().
734 const Function *F = BB->getParent();
735 AttributeSet CallerAttrs = F->getAttributes();
736 if (CallerAttrs.hasAttribute(AttributeSet::ReturnIndex, Attribute::ZExt) ||
737 CallerAttrs.hasAttribute(AttributeSet::ReturnIndex, Attribute::SExt))
740 // Make sure there are no instructions between the PHI and return, or that the
741 // return is the first instruction in the block.
743 BasicBlock::iterator BI = BB->begin();
744 do { ++BI; } while (isa<DbgInfoIntrinsic>(BI));
746 // Also skip over the bitcast.
751 BasicBlock::iterator BI = BB->begin();
752 while (isa<DbgInfoIntrinsic>(BI)) ++BI;
757 /// Only dup the ReturnInst if the CallInst is likely to be emitted as a tail
759 SmallVector<CallInst*, 4> TailCalls;
761 for (unsigned I = 0, E = PN->getNumIncomingValues(); I != E; ++I) {
762 CallInst *CI = dyn_cast<CallInst>(PN->getIncomingValue(I));
763 // Make sure the phi value is indeed produced by the tail call.
764 if (CI && CI->hasOneUse() && CI->getParent() == PN->getIncomingBlock(I) &&
765 TLI->mayBeEmittedAsTailCall(CI))
766 TailCalls.push_back(CI);
769 SmallPtrSet<BasicBlock*, 4> VisitedBBs;
770 for (pred_iterator PI = pred_begin(BB), PE = pred_end(BB); PI != PE; ++PI) {
771 if (!VisitedBBs.insert(*PI))
774 BasicBlock::InstListType &InstList = (*PI)->getInstList();
775 BasicBlock::InstListType::reverse_iterator RI = InstList.rbegin();
776 BasicBlock::InstListType::reverse_iterator RE = InstList.rend();
777 do { ++RI; } while (RI != RE && isa<DbgInfoIntrinsic>(&*RI));
781 CallInst *CI = dyn_cast<CallInst>(&*RI);
782 if (CI && CI->use_empty() && TLI->mayBeEmittedAsTailCall(CI))
783 TailCalls.push_back(CI);
787 bool Changed = false;
788 for (unsigned i = 0, e = TailCalls.size(); i != e; ++i) {
789 CallInst *CI = TailCalls[i];
792 // Conservatively require the attributes of the call to match those of the
793 // return. Ignore noalias because it doesn't affect the call sequence.
794 AttributeSet CalleeAttrs = CS.getAttributes();
795 if (AttrBuilder(CalleeAttrs, AttributeSet::ReturnIndex).
796 removeAttribute(Attribute::NoAlias) !=
797 AttrBuilder(CalleeAttrs, AttributeSet::ReturnIndex).
798 removeAttribute(Attribute::NoAlias))
801 // Make sure the call instruction is followed by an unconditional branch to
803 BasicBlock *CallBB = CI->getParent();
804 BranchInst *BI = dyn_cast<BranchInst>(CallBB->getTerminator());
805 if (!BI || !BI->isUnconditional() || BI->getSuccessor(0) != BB)
808 // Duplicate the return into CallBB.
809 (void)FoldReturnIntoUncondBranch(RI, BB, CallBB);
810 ModifiedDT = Changed = true;
814 // If we eliminated all predecessors of the block, delete the block now.
815 if (Changed && !BB->hasAddressTaken() && pred_begin(BB) == pred_end(BB))
816 BB->eraseFromParent();
821 //===----------------------------------------------------------------------===//
822 // Memory Optimization
823 //===----------------------------------------------------------------------===//
827 /// ExtAddrMode - This is an extended version of TargetLowering::AddrMode
828 /// which holds actual Value*'s for register values.
829 struct ExtAddrMode : public TargetLowering::AddrMode {
832 ExtAddrMode() : BaseReg(0), ScaledReg(0) {}
833 void print(raw_ostream &OS) const;
836 bool operator==(const ExtAddrMode& O) const {
837 return (BaseReg == O.BaseReg) && (ScaledReg == O.ScaledReg) &&
838 (BaseGV == O.BaseGV) && (BaseOffs == O.BaseOffs) &&
839 (HasBaseReg == O.HasBaseReg) && (Scale == O.Scale);
843 static inline raw_ostream &operator<<(raw_ostream &OS, const ExtAddrMode &AM) {
848 void ExtAddrMode::print(raw_ostream &OS) const {
849 bool NeedPlus = false;
852 OS << (NeedPlus ? " + " : "")
854 WriteAsOperand(OS, BaseGV, /*PrintType=*/false);
859 OS << (NeedPlus ? " + " : "") << BaseOffs, NeedPlus = true;
862 OS << (NeedPlus ? " + " : "")
864 WriteAsOperand(OS, BaseReg, /*PrintType=*/false);
868 OS << (NeedPlus ? " + " : "")
870 WriteAsOperand(OS, ScaledReg, /*PrintType=*/false);
876 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
877 void ExtAddrMode::dump() const {
884 /// \brief A helper class for matching addressing modes.
886 /// This encapsulates the logic for matching the target-legal addressing modes.
887 class AddressingModeMatcher {
888 SmallVectorImpl<Instruction*> &AddrModeInsts;
889 const TargetLowering &TLI;
891 /// AccessTy/MemoryInst - This is the type for the access (e.g. double) and
892 /// the memory instruction that we're computing this address for.
894 Instruction *MemoryInst;
896 /// AddrMode - This is the addressing mode that we're building up. This is
897 /// part of the return value of this addressing mode matching stuff.
898 ExtAddrMode &AddrMode;
900 /// IgnoreProfitability - This is set to true when we should not do
901 /// profitability checks. When true, IsProfitableToFoldIntoAddressingMode
902 /// always returns true.
903 bool IgnoreProfitability;
905 AddressingModeMatcher(SmallVectorImpl<Instruction*> &AMI,
906 const TargetLowering &T, Type *AT,
907 Instruction *MI, ExtAddrMode &AM)
908 : AddrModeInsts(AMI), TLI(T), AccessTy(AT), MemoryInst(MI), AddrMode(AM) {
909 IgnoreProfitability = false;
913 /// Match - Find the maximal addressing mode that a load/store of V can fold,
914 /// give an access type of AccessTy. This returns a list of involved
915 /// instructions in AddrModeInsts.
916 static ExtAddrMode Match(Value *V, Type *AccessTy,
917 Instruction *MemoryInst,
918 SmallVectorImpl<Instruction*> &AddrModeInsts,
919 const TargetLowering &TLI) {
923 AddressingModeMatcher(AddrModeInsts, TLI, AccessTy,
924 MemoryInst, Result).MatchAddr(V, 0);
925 (void)Success; assert(Success && "Couldn't select *anything*?");
929 bool MatchScaledValue(Value *ScaleReg, int64_t Scale, unsigned Depth);
930 bool MatchAddr(Value *V, unsigned Depth);
931 bool MatchOperationAddr(User *Operation, unsigned Opcode, unsigned Depth);
932 bool IsProfitableToFoldIntoAddressingMode(Instruction *I,
933 ExtAddrMode &AMBefore,
934 ExtAddrMode &AMAfter);
935 bool ValueAlreadyLiveAtInst(Value *Val, Value *KnownLive1, Value *KnownLive2);
938 /// MatchScaledValue - Try adding ScaleReg*Scale to the current addressing mode.
939 /// Return true and update AddrMode if this addr mode is legal for the target,
941 bool AddressingModeMatcher::MatchScaledValue(Value *ScaleReg, int64_t Scale,
943 // If Scale is 1, then this is the same as adding ScaleReg to the addressing
944 // mode. Just process that directly.
946 return MatchAddr(ScaleReg, Depth);
948 // If the scale is 0, it takes nothing to add this.
952 // If we already have a scale of this value, we can add to it, otherwise, we
953 // need an available scale field.
954 if (AddrMode.Scale != 0 && AddrMode.ScaledReg != ScaleReg)
957 ExtAddrMode TestAddrMode = AddrMode;
959 // Add scale to turn X*4+X*3 -> X*7. This could also do things like
960 // [A+B + A*7] -> [B+A*8].
961 TestAddrMode.Scale += Scale;
962 TestAddrMode.ScaledReg = ScaleReg;
964 // If the new address isn't legal, bail out.
965 if (!TLI.isLegalAddressingMode(TestAddrMode, AccessTy))
968 // It was legal, so commit it.
969 AddrMode = TestAddrMode;
971 // Okay, we decided that we can add ScaleReg+Scale to AddrMode. Check now
972 // to see if ScaleReg is actually X+C. If so, we can turn this into adding
973 // X*Scale + C*Scale to addr mode.
974 ConstantInt *CI = 0; Value *AddLHS = 0;
975 if (isa<Instruction>(ScaleReg) && // not a constant expr.
976 match(ScaleReg, m_Add(m_Value(AddLHS), m_ConstantInt(CI)))) {
977 TestAddrMode.ScaledReg = AddLHS;
978 TestAddrMode.BaseOffs += CI->getSExtValue()*TestAddrMode.Scale;
980 // If this addressing mode is legal, commit it and remember that we folded
982 if (TLI.isLegalAddressingMode(TestAddrMode, AccessTy)) {
983 AddrModeInsts.push_back(cast<Instruction>(ScaleReg));
984 AddrMode = TestAddrMode;
989 // Otherwise, not (x+c)*scale, just return what we have.
993 /// MightBeFoldableInst - This is a little filter, which returns true if an
994 /// addressing computation involving I might be folded into a load/store
995 /// accessing it. This doesn't need to be perfect, but needs to accept at least
996 /// the set of instructions that MatchOperationAddr can.
997 static bool MightBeFoldableInst(Instruction *I) {
998 switch (I->getOpcode()) {
999 case Instruction::BitCast:
1000 // Don't touch identity bitcasts.
1001 if (I->getType() == I->getOperand(0)->getType())
1003 return I->getType()->isPointerTy() || I->getType()->isIntegerTy();
1004 case Instruction::PtrToInt:
1005 // PtrToInt is always a noop, as we know that the int type is pointer sized.
1007 case Instruction::IntToPtr:
1008 // We know the input is intptr_t, so this is foldable.
1010 case Instruction::Add:
1012 case Instruction::Mul:
1013 case Instruction::Shl:
1014 // Can only handle X*C and X << C.
1015 return isa<ConstantInt>(I->getOperand(1));
1016 case Instruction::GetElementPtr:
1023 /// MatchOperationAddr - Given an instruction or constant expr, see if we can
1024 /// fold the operation into the addressing mode. If so, update the addressing
1025 /// mode and return true, otherwise return false without modifying AddrMode.
1026 bool AddressingModeMatcher::MatchOperationAddr(User *AddrInst, unsigned Opcode,
1028 // Avoid exponential behavior on extremely deep expression trees.
1029 if (Depth >= 5) return false;
1032 case Instruction::PtrToInt:
1033 // PtrToInt is always a noop, as we know that the int type is pointer sized.
1034 return MatchAddr(AddrInst->getOperand(0), Depth);
1035 case Instruction::IntToPtr:
1036 // This inttoptr is a no-op if the integer type is pointer sized.
1037 if (TLI.getValueType(AddrInst->getOperand(0)->getType()) ==
1038 TLI.getPointerTy(AddrInst->getType()->getPointerAddressSpace()))
1039 return MatchAddr(AddrInst->getOperand(0), Depth);
1041 case Instruction::BitCast:
1042 // BitCast is always a noop, and we can handle it as long as it is
1043 // int->int or pointer->pointer (we don't want int<->fp or something).
1044 if ((AddrInst->getOperand(0)->getType()->isPointerTy() ||
1045 AddrInst->getOperand(0)->getType()->isIntegerTy()) &&
1046 // Don't touch identity bitcasts. These were probably put here by LSR,
1047 // and we don't want to mess around with them. Assume it knows what it
1049 AddrInst->getOperand(0)->getType() != AddrInst->getType())
1050 return MatchAddr(AddrInst->getOperand(0), Depth);
1052 case Instruction::Add: {
1053 // Check to see if we can merge in the RHS then the LHS. If so, we win.
1054 ExtAddrMode BackupAddrMode = AddrMode;
1055 unsigned OldSize = AddrModeInsts.size();
1056 if (MatchAddr(AddrInst->getOperand(1), Depth+1) &&
1057 MatchAddr(AddrInst->getOperand(0), Depth+1))
1060 // Restore the old addr mode info.
1061 AddrMode = BackupAddrMode;
1062 AddrModeInsts.resize(OldSize);
1064 // Otherwise this was over-aggressive. Try merging in the LHS then the RHS.
1065 if (MatchAddr(AddrInst->getOperand(0), Depth+1) &&
1066 MatchAddr(AddrInst->getOperand(1), Depth+1))
1069 // Otherwise we definitely can't merge the ADD in.
1070 AddrMode = BackupAddrMode;
1071 AddrModeInsts.resize(OldSize);
1074 //case Instruction::Or:
1075 // TODO: We can handle "Or Val, Imm" iff this OR is equivalent to an ADD.
1077 case Instruction::Mul:
1078 case Instruction::Shl: {
1079 // Can only handle X*C and X << C.
1080 ConstantInt *RHS = dyn_cast<ConstantInt>(AddrInst->getOperand(1));
1081 if (!RHS) return false;
1082 int64_t Scale = RHS->getSExtValue();
1083 if (Opcode == Instruction::Shl)
1084 Scale = 1LL << Scale;
1086 return MatchScaledValue(AddrInst->getOperand(0), Scale, Depth);
1088 case Instruction::GetElementPtr: {
1089 // Scan the GEP. We check it if it contains constant offsets and at most
1090 // one variable offset.
1091 int VariableOperand = -1;
1092 unsigned VariableScale = 0;
1094 int64_t ConstantOffset = 0;
1095 const DataLayout *TD = TLI.getDataLayout();
1096 gep_type_iterator GTI = gep_type_begin(AddrInst);
1097 for (unsigned i = 1, e = AddrInst->getNumOperands(); i != e; ++i, ++GTI) {
1098 if (StructType *STy = dyn_cast<StructType>(*GTI)) {
1099 const StructLayout *SL = TD->getStructLayout(STy);
1101 cast<ConstantInt>(AddrInst->getOperand(i))->getZExtValue();
1102 ConstantOffset += SL->getElementOffset(Idx);
1104 uint64_t TypeSize = TD->getTypeAllocSize(GTI.getIndexedType());
1105 if (ConstantInt *CI = dyn_cast<ConstantInt>(AddrInst->getOperand(i))) {
1106 ConstantOffset += CI->getSExtValue()*TypeSize;
1107 } else if (TypeSize) { // Scales of zero don't do anything.
1108 // We only allow one variable index at the moment.
1109 if (VariableOperand != -1)
1112 // Remember the variable index.
1113 VariableOperand = i;
1114 VariableScale = TypeSize;
1119 // A common case is for the GEP to only do a constant offset. In this case,
1120 // just add it to the disp field and check validity.
1121 if (VariableOperand == -1) {
1122 AddrMode.BaseOffs += ConstantOffset;
1123 if (ConstantOffset == 0 || TLI.isLegalAddressingMode(AddrMode, AccessTy)){
1124 // Check to see if we can fold the base pointer in too.
1125 if (MatchAddr(AddrInst->getOperand(0), Depth+1))
1128 AddrMode.BaseOffs -= ConstantOffset;
1132 // Save the valid addressing mode in case we can't match.
1133 ExtAddrMode BackupAddrMode = AddrMode;
1134 unsigned OldSize = AddrModeInsts.size();
1136 // See if the scale and offset amount is valid for this target.
1137 AddrMode.BaseOffs += ConstantOffset;
1139 // Match the base operand of the GEP.
1140 if (!MatchAddr(AddrInst->getOperand(0), Depth+1)) {
1141 // If it couldn't be matched, just stuff the value in a register.
1142 if (AddrMode.HasBaseReg) {
1143 AddrMode = BackupAddrMode;
1144 AddrModeInsts.resize(OldSize);
1147 AddrMode.HasBaseReg = true;
1148 AddrMode.BaseReg = AddrInst->getOperand(0);
1151 // Match the remaining variable portion of the GEP.
1152 if (!MatchScaledValue(AddrInst->getOperand(VariableOperand), VariableScale,
1154 // If it couldn't be matched, try stuffing the base into a register
1155 // instead of matching it, and retrying the match of the scale.
1156 AddrMode = BackupAddrMode;
1157 AddrModeInsts.resize(OldSize);
1158 if (AddrMode.HasBaseReg)
1160 AddrMode.HasBaseReg = true;
1161 AddrMode.BaseReg = AddrInst->getOperand(0);
1162 AddrMode.BaseOffs += ConstantOffset;
1163 if (!MatchScaledValue(AddrInst->getOperand(VariableOperand),
1164 VariableScale, Depth)) {
1165 // If even that didn't work, bail.
1166 AddrMode = BackupAddrMode;
1167 AddrModeInsts.resize(OldSize);
1178 /// MatchAddr - If we can, try to add the value of 'Addr' into the current
1179 /// addressing mode. If Addr can't be added to AddrMode this returns false and
1180 /// leaves AddrMode unmodified. This assumes that Addr is either a pointer type
1181 /// or intptr_t for the target.
1183 bool AddressingModeMatcher::MatchAddr(Value *Addr, unsigned Depth) {
1184 if (ConstantInt *CI = dyn_cast<ConstantInt>(Addr)) {
1185 // Fold in immediates if legal for the target.
1186 AddrMode.BaseOffs += CI->getSExtValue();
1187 if (TLI.isLegalAddressingMode(AddrMode, AccessTy))
1189 AddrMode.BaseOffs -= CI->getSExtValue();
1190 } else if (GlobalValue *GV = dyn_cast<GlobalValue>(Addr)) {
1191 // If this is a global variable, try to fold it into the addressing mode.
1192 if (AddrMode.BaseGV == 0) {
1193 AddrMode.BaseGV = GV;
1194 if (TLI.isLegalAddressingMode(AddrMode, AccessTy))
1196 AddrMode.BaseGV = 0;
1198 } else if (Instruction *I = dyn_cast<Instruction>(Addr)) {
1199 ExtAddrMode BackupAddrMode = AddrMode;
1200 unsigned OldSize = AddrModeInsts.size();
1202 // Check to see if it is possible to fold this operation.
1203 if (MatchOperationAddr(I, I->getOpcode(), Depth)) {
1204 // Okay, it's possible to fold this. Check to see if it is actually
1205 // *profitable* to do so. We use a simple cost model to avoid increasing
1206 // register pressure too much.
1207 if (I->hasOneUse() ||
1208 IsProfitableToFoldIntoAddressingMode(I, BackupAddrMode, AddrMode)) {
1209 AddrModeInsts.push_back(I);
1213 // It isn't profitable to do this, roll back.
1214 //cerr << "NOT FOLDING: " << *I;
1215 AddrMode = BackupAddrMode;
1216 AddrModeInsts.resize(OldSize);
1218 } else if (ConstantExpr *CE = dyn_cast<ConstantExpr>(Addr)) {
1219 if (MatchOperationAddr(CE, CE->getOpcode(), Depth))
1221 } else if (isa<ConstantPointerNull>(Addr)) {
1222 // Null pointer gets folded without affecting the addressing mode.
1226 // Worse case, the target should support [reg] addressing modes. :)
1227 if (!AddrMode.HasBaseReg) {
1228 AddrMode.HasBaseReg = true;
1229 AddrMode.BaseReg = Addr;
1230 // Still check for legality in case the target supports [imm] but not [i+r].
1231 if (TLI.isLegalAddressingMode(AddrMode, AccessTy))
1233 AddrMode.HasBaseReg = false;
1234 AddrMode.BaseReg = 0;
1237 // If the base register is already taken, see if we can do [r+r].
1238 if (AddrMode.Scale == 0) {
1240 AddrMode.ScaledReg = Addr;
1241 if (TLI.isLegalAddressingMode(AddrMode, AccessTy))
1244 AddrMode.ScaledReg = 0;
1250 /// IsOperandAMemoryOperand - Check to see if all uses of OpVal by the specified
1251 /// inline asm call are due to memory operands. If so, return true, otherwise
1253 static bool IsOperandAMemoryOperand(CallInst *CI, InlineAsm *IA, Value *OpVal,
1254 const TargetLowering &TLI) {
1255 TargetLowering::AsmOperandInfoVector TargetConstraints = TLI.ParseConstraints(ImmutableCallSite(CI));
1256 for (unsigned i = 0, e = TargetConstraints.size(); i != e; ++i) {
1257 TargetLowering::AsmOperandInfo &OpInfo = TargetConstraints[i];
1259 // Compute the constraint code and ConstraintType to use.
1260 TLI.ComputeConstraintToUse(OpInfo, SDValue());
1262 // If this asm operand is our Value*, and if it isn't an indirect memory
1263 // operand, we can't fold it!
1264 if (OpInfo.CallOperandVal == OpVal &&
1265 (OpInfo.ConstraintType != TargetLowering::C_Memory ||
1266 !OpInfo.isIndirect))
1273 /// FindAllMemoryUses - Recursively walk all the uses of I until we find a
1274 /// memory use. If we find an obviously non-foldable instruction, return true.
1275 /// Add the ultimately found memory instructions to MemoryUses.
1276 static bool FindAllMemoryUses(Instruction *I,
1277 SmallVectorImpl<std::pair<Instruction*,unsigned> > &MemoryUses,
1278 SmallPtrSet<Instruction*, 16> &ConsideredInsts,
1279 const TargetLowering &TLI) {
1280 // If we already considered this instruction, we're done.
1281 if (!ConsideredInsts.insert(I))
1284 // If this is an obviously unfoldable instruction, bail out.
1285 if (!MightBeFoldableInst(I))
1288 // Loop over all the uses, recursively processing them.
1289 for (Value::use_iterator UI = I->use_begin(), E = I->use_end();
1293 if (LoadInst *LI = dyn_cast<LoadInst>(U)) {
1294 MemoryUses.push_back(std::make_pair(LI, UI.getOperandNo()));
1298 if (StoreInst *SI = dyn_cast<StoreInst>(U)) {
1299 unsigned opNo = UI.getOperandNo();
1300 if (opNo == 0) return true; // Storing addr, not into addr.
1301 MemoryUses.push_back(std::make_pair(SI, opNo));
1305 if (CallInst *CI = dyn_cast<CallInst>(U)) {
1306 InlineAsm *IA = dyn_cast<InlineAsm>(CI->getCalledValue());
1307 if (!IA) return true;
1309 // If this is a memory operand, we're cool, otherwise bail out.
1310 if (!IsOperandAMemoryOperand(CI, IA, I, TLI))
1315 if (FindAllMemoryUses(cast<Instruction>(U), MemoryUses, ConsideredInsts,
1323 /// ValueAlreadyLiveAtInst - Retrn true if Val is already known to be live at
1324 /// the use site that we're folding it into. If so, there is no cost to
1325 /// include it in the addressing mode. KnownLive1 and KnownLive2 are two values
1326 /// that we know are live at the instruction already.
1327 bool AddressingModeMatcher::ValueAlreadyLiveAtInst(Value *Val,Value *KnownLive1,
1328 Value *KnownLive2) {
1329 // If Val is either of the known-live values, we know it is live!
1330 if (Val == 0 || Val == KnownLive1 || Val == KnownLive2)
1333 // All values other than instructions and arguments (e.g. constants) are live.
1334 if (!isa<Instruction>(Val) && !isa<Argument>(Val)) return true;
1336 // If Val is a constant sized alloca in the entry block, it is live, this is
1337 // true because it is just a reference to the stack/frame pointer, which is
1338 // live for the whole function.
1339 if (AllocaInst *AI = dyn_cast<AllocaInst>(Val))
1340 if (AI->isStaticAlloca())
1343 // Check to see if this value is already used in the memory instruction's
1344 // block. If so, it's already live into the block at the very least, so we
1345 // can reasonably fold it.
1346 return Val->isUsedInBasicBlock(MemoryInst->getParent());
1349 /// IsProfitableToFoldIntoAddressingMode - It is possible for the addressing
1350 /// mode of the machine to fold the specified instruction into a load or store
1351 /// that ultimately uses it. However, the specified instruction has multiple
1352 /// uses. Given this, it may actually increase register pressure to fold it
1353 /// into the load. For example, consider this code:
1357 /// use(Y) -> nonload/store
1361 /// In this case, Y has multiple uses, and can be folded into the load of Z
1362 /// (yielding load [X+2]). However, doing this will cause both "X" and "X+1" to
1363 /// be live at the use(Y) line. If we don't fold Y into load Z, we use one
1364 /// fewer register. Since Y can't be folded into "use(Y)" we don't increase the
1365 /// number of computations either.
1367 /// Note that this (like most of CodeGenPrepare) is just a rough heuristic. If
1368 /// X was live across 'load Z' for other reasons, we actually *would* want to
1369 /// fold the addressing mode in the Z case. This would make Y die earlier.
1370 bool AddressingModeMatcher::
1371 IsProfitableToFoldIntoAddressingMode(Instruction *I, ExtAddrMode &AMBefore,
1372 ExtAddrMode &AMAfter) {
1373 if (IgnoreProfitability) return true;
1375 // AMBefore is the addressing mode before this instruction was folded into it,
1376 // and AMAfter is the addressing mode after the instruction was folded. Get
1377 // the set of registers referenced by AMAfter and subtract out those
1378 // referenced by AMBefore: this is the set of values which folding in this
1379 // address extends the lifetime of.
1381 // Note that there are only two potential values being referenced here,
1382 // BaseReg and ScaleReg (global addresses are always available, as are any
1383 // folded immediates).
1384 Value *BaseReg = AMAfter.BaseReg, *ScaledReg = AMAfter.ScaledReg;
1386 // If the BaseReg or ScaledReg was referenced by the previous addrmode, their
1387 // lifetime wasn't extended by adding this instruction.
1388 if (ValueAlreadyLiveAtInst(BaseReg, AMBefore.BaseReg, AMBefore.ScaledReg))
1390 if (ValueAlreadyLiveAtInst(ScaledReg, AMBefore.BaseReg, AMBefore.ScaledReg))
1393 // If folding this instruction (and it's subexprs) didn't extend any live
1394 // ranges, we're ok with it.
1395 if (BaseReg == 0 && ScaledReg == 0)
1398 // If all uses of this instruction are ultimately load/store/inlineasm's,
1399 // check to see if their addressing modes will include this instruction. If
1400 // so, we can fold it into all uses, so it doesn't matter if it has multiple
1402 SmallVector<std::pair<Instruction*,unsigned>, 16> MemoryUses;
1403 SmallPtrSet<Instruction*, 16> ConsideredInsts;
1404 if (FindAllMemoryUses(I, MemoryUses, ConsideredInsts, TLI))
1405 return false; // Has a non-memory, non-foldable use!
1407 // Now that we know that all uses of this instruction are part of a chain of
1408 // computation involving only operations that could theoretically be folded
1409 // into a memory use, loop over each of these uses and see if they could
1410 // *actually* fold the instruction.
1411 SmallVector<Instruction*, 32> MatchedAddrModeInsts;
1412 for (unsigned i = 0, e = MemoryUses.size(); i != e; ++i) {
1413 Instruction *User = MemoryUses[i].first;
1414 unsigned OpNo = MemoryUses[i].second;
1416 // Get the access type of this use. If the use isn't a pointer, we don't
1417 // know what it accesses.
1418 Value *Address = User->getOperand(OpNo);
1419 if (!Address->getType()->isPointerTy())
1421 Type *AddressAccessTy = Address->getType()->getPointerElementType();
1423 // Do a match against the root of this address, ignoring profitability. This
1424 // will tell us if the addressing mode for the memory operation will
1425 // *actually* cover the shared instruction.
1427 AddressingModeMatcher Matcher(MatchedAddrModeInsts, TLI, AddressAccessTy,
1428 MemoryInst, Result);
1429 Matcher.IgnoreProfitability = true;
1430 bool Success = Matcher.MatchAddr(Address, 0);
1431 (void)Success; assert(Success && "Couldn't select *anything*?");
1433 // If the match didn't cover I, then it won't be shared by it.
1434 if (std::find(MatchedAddrModeInsts.begin(), MatchedAddrModeInsts.end(),
1435 I) == MatchedAddrModeInsts.end())
1438 MatchedAddrModeInsts.clear();
1444 } // end anonymous namespace
1446 /// IsNonLocalValue - Return true if the specified values are defined in a
1447 /// different basic block than BB.
1448 static bool IsNonLocalValue(Value *V, BasicBlock *BB) {
1449 if (Instruction *I = dyn_cast<Instruction>(V))
1450 return I->getParent() != BB;
1454 /// OptimizeMemoryInst - Load and Store Instructions often have
1455 /// addressing modes that can do significant amounts of computation. As such,
1456 /// instruction selection will try to get the load or store to do as much
1457 /// computation as possible for the program. The problem is that isel can only
1458 /// see within a single block. As such, we sink as much legal addressing mode
1459 /// stuff into the block as possible.
1461 /// This method is used to optimize both load/store and inline asms with memory
1463 bool CodeGenPrepare::OptimizeMemoryInst(Instruction *MemoryInst, Value *Addr,
1467 // Try to collapse single-value PHI nodes. This is necessary to undo
1468 // unprofitable PRE transformations.
1469 SmallVector<Value*, 8> worklist;
1470 SmallPtrSet<Value*, 16> Visited;
1471 worklist.push_back(Addr);
1473 // Use a worklist to iteratively look through PHI nodes, and ensure that
1474 // the addressing mode obtained from the non-PHI roots of the graph
1476 Value *Consensus = 0;
1477 unsigned NumUsesConsensus = 0;
1478 bool IsNumUsesConsensusValid = false;
1479 SmallVector<Instruction*, 16> AddrModeInsts;
1480 ExtAddrMode AddrMode;
1481 while (!worklist.empty()) {
1482 Value *V = worklist.back();
1483 worklist.pop_back();
1485 // Break use-def graph loops.
1486 if (!Visited.insert(V)) {
1491 // For a PHI node, push all of its incoming values.
1492 if (PHINode *P = dyn_cast<PHINode>(V)) {
1493 for (unsigned i = 0, e = P->getNumIncomingValues(); i != e; ++i)
1494 worklist.push_back(P->getIncomingValue(i));
1498 // For non-PHIs, determine the addressing mode being computed.
1499 SmallVector<Instruction*, 16> NewAddrModeInsts;
1500 ExtAddrMode NewAddrMode =
1501 AddressingModeMatcher::Match(V, AccessTy, MemoryInst,
1502 NewAddrModeInsts, *TLI);
1504 // This check is broken into two cases with very similar code to avoid using
1505 // getNumUses() as much as possible. Some values have a lot of uses, so
1506 // calling getNumUses() unconditionally caused a significant compile-time
1510 AddrMode = NewAddrMode;
1511 AddrModeInsts = NewAddrModeInsts;
1513 } else if (NewAddrMode == AddrMode) {
1514 if (!IsNumUsesConsensusValid) {
1515 NumUsesConsensus = Consensus->getNumUses();
1516 IsNumUsesConsensusValid = true;
1519 // Ensure that the obtained addressing mode is equivalent to that obtained
1520 // for all other roots of the PHI traversal. Also, when choosing one
1521 // such root as representative, select the one with the most uses in order
1522 // to keep the cost modeling heuristics in AddressingModeMatcher
1524 unsigned NumUses = V->getNumUses();
1525 if (NumUses > NumUsesConsensus) {
1527 NumUsesConsensus = NumUses;
1528 AddrModeInsts = NewAddrModeInsts;
1537 // If the addressing mode couldn't be determined, or if multiple different
1538 // ones were determined, bail out now.
1539 if (!Consensus) return false;
1541 // Check to see if any of the instructions supersumed by this addr mode are
1542 // non-local to I's BB.
1543 bool AnyNonLocal = false;
1544 for (unsigned i = 0, e = AddrModeInsts.size(); i != e; ++i) {
1545 if (IsNonLocalValue(AddrModeInsts[i], MemoryInst->getParent())) {
1551 // If all the instructions matched are already in this BB, don't do anything.
1553 DEBUG(dbgs() << "CGP: Found local addrmode: " << AddrMode << "\n");
1557 // Insert this computation right after this user. Since our caller is
1558 // scanning from the top of the BB to the bottom, reuse of the expr are
1559 // guaranteed to happen later.
1560 IRBuilder<> Builder(MemoryInst);
1562 // Now that we determined the addressing expression we want to use and know
1563 // that we have to sink it into this block. Check to see if we have already
1564 // done this for some other load/store instr in this block. If so, reuse the
1566 Value *&SunkAddr = SunkAddrs[Addr];
1568 DEBUG(dbgs() << "CGP: Reusing nonlocal addrmode: " << AddrMode << " for "
1570 if (SunkAddr->getType() != Addr->getType())
1571 SunkAddr = Builder.CreateBitCast(SunkAddr, Addr->getType());
1573 DEBUG(dbgs() << "CGP: SINKING nonlocal addrmode: " << AddrMode << " for "
1575 Type *IntPtrTy = TLI->getDataLayout()->getIntPtrType(Addr->getType());
1578 // Start with the base register. Do this first so that subsequent address
1579 // matching finds it last, which will prevent it from trying to match it
1580 // as the scaled value in case it happens to be a mul. That would be
1581 // problematic if we've sunk a different mul for the scale, because then
1582 // we'd end up sinking both muls.
1583 if (AddrMode.BaseReg) {
1584 Value *V = AddrMode.BaseReg;
1585 if (V->getType()->isPointerTy())
1586 V = Builder.CreatePtrToInt(V, IntPtrTy, "sunkaddr");
1587 if (V->getType() != IntPtrTy)
1588 V = Builder.CreateIntCast(V, IntPtrTy, /*isSigned=*/true, "sunkaddr");
1592 // Add the scale value.
1593 if (AddrMode.Scale) {
1594 Value *V = AddrMode.ScaledReg;
1595 if (V->getType() == IntPtrTy) {
1597 } else if (V->getType()->isPointerTy()) {
1598 V = Builder.CreatePtrToInt(V, IntPtrTy, "sunkaddr");
1599 } else if (cast<IntegerType>(IntPtrTy)->getBitWidth() <
1600 cast<IntegerType>(V->getType())->getBitWidth()) {
1601 V = Builder.CreateTrunc(V, IntPtrTy, "sunkaddr");
1603 V = Builder.CreateSExt(V, IntPtrTy, "sunkaddr");
1605 if (AddrMode.Scale != 1)
1606 V = Builder.CreateMul(V, ConstantInt::get(IntPtrTy, AddrMode.Scale),
1609 Result = Builder.CreateAdd(Result, V, "sunkaddr");
1614 // Add in the BaseGV if present.
1615 if (AddrMode.BaseGV) {
1616 Value *V = Builder.CreatePtrToInt(AddrMode.BaseGV, IntPtrTy, "sunkaddr");
1618 Result = Builder.CreateAdd(Result, V, "sunkaddr");
1623 // Add in the Base Offset if present.
1624 if (AddrMode.BaseOffs) {
1625 Value *V = ConstantInt::get(IntPtrTy, AddrMode.BaseOffs);
1627 Result = Builder.CreateAdd(Result, V, "sunkaddr");
1633 SunkAddr = Constant::getNullValue(Addr->getType());
1635 SunkAddr = Builder.CreateIntToPtr(Result, Addr->getType(), "sunkaddr");
1638 MemoryInst->replaceUsesOfWith(Repl, SunkAddr);
1640 // If we have no uses, recursively delete the value and all dead instructions
1642 if (Repl->use_empty()) {
1643 // This can cause recursive deletion, which can invalidate our iterator.
1644 // Use a WeakVH to hold onto it in case this happens.
1645 WeakVH IterHandle(CurInstIterator);
1646 BasicBlock *BB = CurInstIterator->getParent();
1648 RecursivelyDeleteTriviallyDeadInstructions(Repl, TLInfo);
1650 if (IterHandle != CurInstIterator) {
1651 // If the iterator instruction was recursively deleted, start over at the
1652 // start of the block.
1653 CurInstIterator = BB->begin();
1661 /// OptimizeInlineAsmInst - If there are any memory operands, use
1662 /// OptimizeMemoryInst to sink their address computing into the block when
1663 /// possible / profitable.
1664 bool CodeGenPrepare::OptimizeInlineAsmInst(CallInst *CS) {
1665 bool MadeChange = false;
1667 TargetLowering::AsmOperandInfoVector
1668 TargetConstraints = TLI->ParseConstraints(CS);
1670 for (unsigned i = 0, e = TargetConstraints.size(); i != e; ++i) {
1671 TargetLowering::AsmOperandInfo &OpInfo = TargetConstraints[i];
1673 // Compute the constraint code and ConstraintType to use.
1674 TLI->ComputeConstraintToUse(OpInfo, SDValue());
1676 if (OpInfo.ConstraintType == TargetLowering::C_Memory &&
1677 OpInfo.isIndirect) {
1678 Value *OpVal = CS->getArgOperand(ArgNo++);
1679 MadeChange |= OptimizeMemoryInst(CS, OpVal, OpVal->getType());
1680 } else if (OpInfo.Type == InlineAsm::isInput)
1687 /// MoveExtToFormExtLoad - Move a zext or sext fed by a load into the same
1688 /// basic block as the load, unless conditions are unfavorable. This allows
1689 /// SelectionDAG to fold the extend into the load.
1691 bool CodeGenPrepare::MoveExtToFormExtLoad(Instruction *I) {
1692 // Look for a load being extended.
1693 LoadInst *LI = dyn_cast<LoadInst>(I->getOperand(0));
1694 if (!LI) return false;
1696 // If they're already in the same block, there's nothing to do.
1697 if (LI->getParent() == I->getParent())
1700 // If the load has other users and the truncate is not free, this probably
1701 // isn't worthwhile.
1702 if (!LI->hasOneUse() &&
1703 TLI && (TLI->isTypeLegal(TLI->getValueType(LI->getType())) ||
1704 !TLI->isTypeLegal(TLI->getValueType(I->getType()))) &&
1705 !TLI->isTruncateFree(I->getType(), LI->getType()))
1708 // Check whether the target supports casts folded into loads.
1710 if (isa<ZExtInst>(I))
1711 LType = ISD::ZEXTLOAD;
1713 assert(isa<SExtInst>(I) && "Unexpected ext type!");
1714 LType = ISD::SEXTLOAD;
1716 if (TLI && !TLI->isLoadExtLegal(LType, TLI->getValueType(LI->getType())))
1719 // Move the extend into the same block as the load, so that SelectionDAG
1721 I->removeFromParent();
1727 bool CodeGenPrepare::OptimizeExtUses(Instruction *I) {
1728 BasicBlock *DefBB = I->getParent();
1730 // If the result of a {s|z}ext and its source are both live out, rewrite all
1731 // other uses of the source with result of extension.
1732 Value *Src = I->getOperand(0);
1733 if (Src->hasOneUse())
1736 // Only do this xform if truncating is free.
1737 if (TLI && !TLI->isTruncateFree(I->getType(), Src->getType()))
1740 // Only safe to perform the optimization if the source is also defined in
1742 if (!isa<Instruction>(Src) || DefBB != cast<Instruction>(Src)->getParent())
1745 bool DefIsLiveOut = false;
1746 for (Value::use_iterator UI = I->use_begin(), E = I->use_end();
1748 Instruction *User = cast<Instruction>(*UI);
1750 // Figure out which BB this ext is used in.
1751 BasicBlock *UserBB = User->getParent();
1752 if (UserBB == DefBB) continue;
1753 DefIsLiveOut = true;
1759 // Make sure none of the uses are PHI nodes.
1760 for (Value::use_iterator UI = Src->use_begin(), E = Src->use_end();
1762 Instruction *User = cast<Instruction>(*UI);
1763 BasicBlock *UserBB = User->getParent();
1764 if (UserBB == DefBB) continue;
1765 // Be conservative. We don't want this xform to end up introducing
1766 // reloads just before load / store instructions.
1767 if (isa<PHINode>(User) || isa<LoadInst>(User) || isa<StoreInst>(User))
1771 // InsertedTruncs - Only insert one trunc in each block once.
1772 DenseMap<BasicBlock*, Instruction*> InsertedTruncs;
1774 bool MadeChange = false;
1775 for (Value::use_iterator UI = Src->use_begin(), E = Src->use_end();
1777 Use &TheUse = UI.getUse();
1778 Instruction *User = cast<Instruction>(*UI);
1780 // Figure out which BB this ext is used in.
1781 BasicBlock *UserBB = User->getParent();
1782 if (UserBB == DefBB) continue;
1784 // Both src and def are live in this block. Rewrite the use.
1785 Instruction *&InsertedTrunc = InsertedTruncs[UserBB];
1787 if (!InsertedTrunc) {
1788 BasicBlock::iterator InsertPt = UserBB->getFirstInsertionPt();
1789 InsertedTrunc = new TruncInst(I, Src->getType(), "", InsertPt);
1792 // Replace a use of the {s|z}ext source with a use of the result.
1793 TheUse = InsertedTrunc;
1801 /// isFormingBranchFromSelectProfitable - Returns true if a SelectInst should be
1802 /// turned into an explicit branch.
1803 static bool isFormingBranchFromSelectProfitable(SelectInst *SI) {
1804 // FIXME: This should use the same heuristics as IfConversion to determine
1805 // whether a select is better represented as a branch. This requires that
1806 // branch probability metadata is preserved for the select, which is not the
1809 CmpInst *Cmp = dyn_cast<CmpInst>(SI->getCondition());
1811 // If the branch is predicted right, an out of order CPU can avoid blocking on
1812 // the compare. Emit cmovs on compares with a memory operand as branches to
1813 // avoid stalls on the load from memory. If the compare has more than one use
1814 // there's probably another cmov or setcc around so it's not worth emitting a
1819 Value *CmpOp0 = Cmp->getOperand(0);
1820 Value *CmpOp1 = Cmp->getOperand(1);
1822 // We check that the memory operand has one use to avoid uses of the loaded
1823 // value directly after the compare, making branches unprofitable.
1824 return Cmp->hasOneUse() &&
1825 ((isa<LoadInst>(CmpOp0) && CmpOp0->hasOneUse()) ||
1826 (isa<LoadInst>(CmpOp1) && CmpOp1->hasOneUse()));
1830 /// If we have a SelectInst that will likely profit from branch prediction,
1831 /// turn it into a branch.
1832 bool CodeGenPrepare::OptimizeSelectInst(SelectInst *SI) {
1833 bool VectorCond = !SI->getCondition()->getType()->isIntegerTy(1);
1835 // Can we convert the 'select' to CF ?
1836 if (DisableSelectToBranch || OptSize || !TLI || VectorCond)
1839 TargetLowering::SelectSupportKind SelectKind;
1841 SelectKind = TargetLowering::VectorMaskSelect;
1842 else if (SI->getType()->isVectorTy())
1843 SelectKind = TargetLowering::ScalarCondVectorVal;
1845 SelectKind = TargetLowering::ScalarValSelect;
1847 // Do we have efficient codegen support for this kind of 'selects' ?
1848 if (TLI->isSelectSupported(SelectKind)) {
1849 // We have efficient codegen support for the select instruction.
1850 // Check if it is profitable to keep this 'select'.
1851 if (!TLI->isPredictableSelectExpensive() ||
1852 !isFormingBranchFromSelectProfitable(SI))
1858 // First, we split the block containing the select into 2 blocks.
1859 BasicBlock *StartBlock = SI->getParent();
1860 BasicBlock::iterator SplitPt = ++(BasicBlock::iterator(SI));
1861 BasicBlock *NextBlock = StartBlock->splitBasicBlock(SplitPt, "select.end");
1863 // Create a new block serving as the landing pad for the branch.
1864 BasicBlock *SmallBlock = BasicBlock::Create(SI->getContext(), "select.mid",
1865 NextBlock->getParent(), NextBlock);
1867 // Move the unconditional branch from the block with the select in it into our
1868 // landing pad block.
1869 StartBlock->getTerminator()->eraseFromParent();
1870 BranchInst::Create(NextBlock, SmallBlock);
1872 // Insert the real conditional branch based on the original condition.
1873 BranchInst::Create(NextBlock, SmallBlock, SI->getCondition(), SI);
1875 // The select itself is replaced with a PHI Node.
1876 PHINode *PN = PHINode::Create(SI->getType(), 2, "", NextBlock->begin());
1878 PN->addIncoming(SI->getTrueValue(), StartBlock);
1879 PN->addIncoming(SI->getFalseValue(), SmallBlock);
1880 SI->replaceAllUsesWith(PN);
1881 SI->eraseFromParent();
1883 // Instruct OptimizeBlock to skip to the next block.
1884 CurInstIterator = StartBlock->end();
1885 ++NumSelectsExpanded;
1889 bool CodeGenPrepare::OptimizeInst(Instruction *I) {
1890 if (PHINode *P = dyn_cast<PHINode>(I)) {
1891 // It is possible for very late stage optimizations (such as SimplifyCFG)
1892 // to introduce PHI nodes too late to be cleaned up. If we detect such a
1893 // trivial PHI, go ahead and zap it here.
1894 if (Value *V = SimplifyInstruction(P)) {
1895 P->replaceAllUsesWith(V);
1896 P->eraseFromParent();
1903 if (CastInst *CI = dyn_cast<CastInst>(I)) {
1904 // If the source of the cast is a constant, then this should have
1905 // already been constant folded. The only reason NOT to constant fold
1906 // it is if something (e.g. LSR) was careful to place the constant
1907 // evaluation in a block other than then one that uses it (e.g. to hoist
1908 // the address of globals out of a loop). If this is the case, we don't
1909 // want to forward-subst the cast.
1910 if (isa<Constant>(CI->getOperand(0)))
1913 if (TLI && OptimizeNoopCopyExpression(CI, *TLI))
1916 if (isa<ZExtInst>(I) || isa<SExtInst>(I)) {
1917 bool MadeChange = MoveExtToFormExtLoad(I);
1918 return MadeChange | OptimizeExtUses(I);
1923 if (CmpInst *CI = dyn_cast<CmpInst>(I))
1924 return OptimizeCmpExpression(CI);
1926 if (LoadInst *LI = dyn_cast<LoadInst>(I)) {
1928 return OptimizeMemoryInst(I, I->getOperand(0), LI->getType());
1932 if (StoreInst *SI = dyn_cast<StoreInst>(I)) {
1934 return OptimizeMemoryInst(I, SI->getOperand(1),
1935 SI->getOperand(0)->getType());
1939 if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(I)) {
1940 if (GEPI->hasAllZeroIndices()) {
1941 /// The GEP operand must be a pointer, so must its result -> BitCast
1942 Instruction *NC = new BitCastInst(GEPI->getOperand(0), GEPI->getType(),
1943 GEPI->getName(), GEPI);
1944 GEPI->replaceAllUsesWith(NC);
1945 GEPI->eraseFromParent();
1953 if (CallInst *CI = dyn_cast<CallInst>(I))
1954 return OptimizeCallInst(CI);
1956 if (SelectInst *SI = dyn_cast<SelectInst>(I))
1957 return OptimizeSelectInst(SI);
1962 // In this pass we look for GEP and cast instructions that are used
1963 // across basic blocks and rewrite them to improve basic-block-at-a-time
1965 bool CodeGenPrepare::OptimizeBlock(BasicBlock &BB) {
1967 bool MadeChange = false;
1969 CurInstIterator = BB.begin();
1970 while (CurInstIterator != BB.end())
1971 MadeChange |= OptimizeInst(CurInstIterator++);
1973 MadeChange |= DupRetToEnableTailCallOpts(&BB);
1978 // llvm.dbg.value is far away from the value then iSel may not be able
1979 // handle it properly. iSel will drop llvm.dbg.value if it can not
1980 // find a node corresponding to the value.
1981 bool CodeGenPrepare::PlaceDbgValues(Function &F) {
1982 bool MadeChange = false;
1983 for (Function::iterator I = F.begin(), E = F.end(); I != E; ++I) {
1984 Instruction *PrevNonDbgInst = NULL;
1985 for (BasicBlock::iterator BI = I->begin(), BE = I->end(); BI != BE;) {
1986 Instruction *Insn = BI; ++BI;
1987 DbgValueInst *DVI = dyn_cast<DbgValueInst>(Insn);
1989 PrevNonDbgInst = Insn;
1993 Instruction *VI = dyn_cast_or_null<Instruction>(DVI->getValue());
1994 if (VI && VI != PrevNonDbgInst && !VI->isTerminator()) {
1995 DEBUG(dbgs() << "Moving Debug Value before :\n" << *DVI << ' ' << *VI);
1996 DVI->removeFromParent();
1997 if (isa<PHINode>(VI))
1998 DVI->insertBefore(VI->getParent()->getFirstInsertionPt());
2000 DVI->insertAfter(VI);