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 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 ValueMap<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 const char *getPassName() const { return "CodeGen Prepare"; }
111 virtual void getAnalysisUsage(AnalysisUsage &AU) const {
112 AU.addPreserved<DominatorTree>();
113 AU.addPreserved<ProfileInfo>();
114 AU.addRequired<TargetLibraryInfo>();
118 bool EliminateFallThrough(Function &F);
119 bool EliminateMostlyEmptyBlocks(Function &F);
120 bool CanMergeBlocks(const BasicBlock *BB, const BasicBlock *DestBB) const;
121 void EliminateMostlyEmptyBlock(BasicBlock *BB);
122 bool OptimizeBlock(BasicBlock &BB);
123 bool OptimizeInst(Instruction *I);
124 bool OptimizeMemoryInst(Instruction *I, Value *Addr, Type *AccessTy);
125 bool OptimizeInlineAsmInst(CallInst *CS);
126 bool OptimizeCallInst(CallInst *CI);
127 bool MoveExtToFormExtLoad(Instruction *I);
128 bool OptimizeExtUses(Instruction *I);
129 bool OptimizeSelectInst(SelectInst *SI);
130 bool DupRetToEnableTailCallOpts(BasicBlock *BB);
131 bool PlaceDbgValues(Function &F);
135 char CodeGenPrepare::ID = 0;
136 INITIALIZE_PASS_BEGIN(CodeGenPrepare, "codegenprepare",
137 "Optimize for code generation", false, false)
138 INITIALIZE_PASS_DEPENDENCY(TargetLibraryInfo)
139 INITIALIZE_PASS_END(CodeGenPrepare, "codegenprepare",
140 "Optimize for code generation", false, false)
142 FunctionPass *llvm::createCodeGenPreparePass(const TargetLowering *TLI) {
143 return new CodeGenPrepare(TLI);
146 bool CodeGenPrepare::runOnFunction(Function &F) {
147 bool EverMadeChange = false;
150 TLInfo = &getAnalysis<TargetLibraryInfo>();
151 DT = getAnalysisIfAvailable<DominatorTree>();
152 PFI = getAnalysisIfAvailable<ProfileInfo>();
153 OptSize = F.getAttributes().hasAttribute(AttributeSet::FunctionIndex,
154 Attribute::OptimizeForSize);
156 /// This optimization identifies DIV instructions that can be
157 /// profitably bypassed and carried out with a shorter, faster divide.
158 if (!OptSize && TLI && TLI->isSlowDivBypassed()) {
159 const DenseMap<unsigned int, unsigned int> &BypassWidths =
160 TLI->getBypassSlowDivWidths();
161 for (Function::iterator I = F.begin(); I != F.end(); I++)
162 EverMadeChange |= bypassSlowDivision(F, I, BypassWidths);
165 // Eliminate blocks that contain only PHI nodes and an
166 // unconditional branch.
167 EverMadeChange |= EliminateMostlyEmptyBlocks(F);
169 // llvm.dbg.value is far away from the value then iSel may not be able
170 // handle it properly. iSel will drop llvm.dbg.value if it can not
171 // find a node corresponding to the value.
172 EverMadeChange |= PlaceDbgValues(F);
174 bool MadeChange = true;
177 for (Function::iterator I = F.begin(); I != F.end(); ) {
178 BasicBlock *BB = I++;
179 MadeChange |= OptimizeBlock(*BB);
181 EverMadeChange |= MadeChange;
186 if (!DisableBranchOpts) {
188 SmallPtrSet<BasicBlock*, 8> WorkList;
189 for (Function::iterator BB = F.begin(), E = F.end(); BB != E; ++BB) {
190 SmallVector<BasicBlock*, 2> Successors(succ_begin(BB), succ_end(BB));
191 MadeChange |= ConstantFoldTerminator(BB, true);
192 if (!MadeChange) continue;
194 for (SmallVectorImpl<BasicBlock*>::iterator
195 II = Successors.begin(), IE = Successors.end(); II != IE; ++II)
196 if (pred_begin(*II) == pred_end(*II))
197 WorkList.insert(*II);
200 // Delete the dead blocks and any of their dead successors.
201 MadeChange |= !WorkList.empty();
202 while (!WorkList.empty()) {
203 BasicBlock *BB = *WorkList.begin();
205 SmallVector<BasicBlock*, 2> Successors(succ_begin(BB), succ_end(BB));
209 for (SmallVectorImpl<BasicBlock*>::iterator
210 II = Successors.begin(), IE = Successors.end(); II != IE; ++II)
211 if (pred_begin(*II) == pred_end(*II))
212 WorkList.insert(*II);
215 // Merge pairs of basic blocks with unconditional branches, connected by
217 if (EverMadeChange || MadeChange)
218 MadeChange |= EliminateFallThrough(F);
222 EverMadeChange |= MadeChange;
225 if (ModifiedDT && DT)
226 DT->DT->recalculate(F);
228 return EverMadeChange;
231 /// EliminateFallThrough - Merge basic blocks which are connected
232 /// by a single edge, where one of the basic blocks has a single successor
233 /// pointing to the other basic block, which has a single predecessor.
234 bool CodeGenPrepare::EliminateFallThrough(Function &F) {
235 bool Changed = false;
236 // Scan all of the blocks in the function, except for the entry block.
237 for (Function::iterator I = ++F.begin(), E = F.end(); I != E; ) {
238 BasicBlock *BB = I++;
239 // If the destination block has a single pred, then this is a trivial
240 // edge, just collapse it.
241 BasicBlock *SinglePred = BB->getSinglePredecessor();
243 // Don't merge if BB's address is taken.
244 if (!SinglePred || SinglePred == BB || BB->hasAddressTaken()) continue;
246 BranchInst *Term = dyn_cast<BranchInst>(SinglePred->getTerminator());
247 if (Term && !Term->isConditional()) {
249 DEBUG(dbgs() << "To merge:\n"<< *SinglePred << "\n\n\n");
250 // Remember if SinglePred was the entry block of the function.
251 // If so, we will need to move BB back to the entry position.
252 bool isEntry = SinglePred == &SinglePred->getParent()->getEntryBlock();
253 MergeBasicBlockIntoOnlyPred(BB, this);
255 if (isEntry && BB != &BB->getParent()->getEntryBlock())
256 BB->moveBefore(&BB->getParent()->getEntryBlock());
258 // We have erased a block. Update the iterator.
265 /// EliminateMostlyEmptyBlocks - eliminate blocks that contain only PHI nodes,
266 /// debug info directives, and an unconditional branch. Passes before isel
267 /// (e.g. LSR/loopsimplify) often split edges in ways that are non-optimal for
268 /// isel. Start by eliminating these blocks so we can split them the way we
270 bool CodeGenPrepare::EliminateMostlyEmptyBlocks(Function &F) {
271 bool MadeChange = false;
272 // Note that this intentionally skips the entry block.
273 for (Function::iterator I = ++F.begin(), E = F.end(); I != E; ) {
274 BasicBlock *BB = I++;
276 // If this block doesn't end with an uncond branch, ignore it.
277 BranchInst *BI = dyn_cast<BranchInst>(BB->getTerminator());
278 if (!BI || !BI->isUnconditional())
281 // If the instruction before the branch (skipping debug info) isn't a phi
282 // node, then other stuff is happening here.
283 BasicBlock::iterator BBI = BI;
284 if (BBI != BB->begin()) {
286 while (isa<DbgInfoIntrinsic>(BBI)) {
287 if (BBI == BB->begin())
291 if (!isa<DbgInfoIntrinsic>(BBI) && !isa<PHINode>(BBI))
295 // Do not break infinite loops.
296 BasicBlock *DestBB = BI->getSuccessor(0);
300 if (!CanMergeBlocks(BB, DestBB))
303 EliminateMostlyEmptyBlock(BB);
309 /// CanMergeBlocks - Return true if we can merge BB into DestBB if there is a
310 /// single uncond branch between them, and BB contains no other non-phi
312 bool CodeGenPrepare::CanMergeBlocks(const BasicBlock *BB,
313 const BasicBlock *DestBB) const {
314 // We only want to eliminate blocks whose phi nodes are used by phi nodes in
315 // the successor. If there are more complex condition (e.g. preheaders),
316 // don't mess around with them.
317 BasicBlock::const_iterator BBI = BB->begin();
318 while (const PHINode *PN = dyn_cast<PHINode>(BBI++)) {
319 for (Value::const_use_iterator UI = PN->use_begin(), E = PN->use_end();
321 const Instruction *User = cast<Instruction>(*UI);
322 if (User->getParent() != DestBB || !isa<PHINode>(User))
324 // If User is inside DestBB block and it is a PHINode then check
325 // incoming value. If incoming value is not from BB then this is
326 // a complex condition (e.g. preheaders) we want to avoid here.
327 if (User->getParent() == DestBB) {
328 if (const PHINode *UPN = dyn_cast<PHINode>(User))
329 for (unsigned I = 0, E = UPN->getNumIncomingValues(); I != E; ++I) {
330 Instruction *Insn = dyn_cast<Instruction>(UPN->getIncomingValue(I));
331 if (Insn && Insn->getParent() == BB &&
332 Insn->getParent() != UPN->getIncomingBlock(I))
339 // If BB and DestBB contain any common predecessors, then the phi nodes in BB
340 // and DestBB may have conflicting incoming values for the block. If so, we
341 // can't merge the block.
342 const PHINode *DestBBPN = dyn_cast<PHINode>(DestBB->begin());
343 if (!DestBBPN) return true; // no conflict.
345 // Collect the preds of BB.
346 SmallPtrSet<const BasicBlock*, 16> BBPreds;
347 if (const PHINode *BBPN = dyn_cast<PHINode>(BB->begin())) {
348 // It is faster to get preds from a PHI than with pred_iterator.
349 for (unsigned i = 0, e = BBPN->getNumIncomingValues(); i != e; ++i)
350 BBPreds.insert(BBPN->getIncomingBlock(i));
352 BBPreds.insert(pred_begin(BB), pred_end(BB));
355 // Walk the preds of DestBB.
356 for (unsigned i = 0, e = DestBBPN->getNumIncomingValues(); i != e; ++i) {
357 BasicBlock *Pred = DestBBPN->getIncomingBlock(i);
358 if (BBPreds.count(Pred)) { // Common predecessor?
359 BBI = DestBB->begin();
360 while (const PHINode *PN = dyn_cast<PHINode>(BBI++)) {
361 const Value *V1 = PN->getIncomingValueForBlock(Pred);
362 const Value *V2 = PN->getIncomingValueForBlock(BB);
364 // If V2 is a phi node in BB, look up what the mapped value will be.
365 if (const PHINode *V2PN = dyn_cast<PHINode>(V2))
366 if (V2PN->getParent() == BB)
367 V2 = V2PN->getIncomingValueForBlock(Pred);
369 // If there is a conflict, bail out.
370 if (V1 != V2) return false;
379 /// EliminateMostlyEmptyBlock - Eliminate a basic block that have only phi's and
380 /// an unconditional branch in it.
381 void CodeGenPrepare::EliminateMostlyEmptyBlock(BasicBlock *BB) {
382 BranchInst *BI = cast<BranchInst>(BB->getTerminator());
383 BasicBlock *DestBB = BI->getSuccessor(0);
385 DEBUG(dbgs() << "MERGING MOSTLY EMPTY BLOCKS - BEFORE:\n" << *BB << *DestBB);
387 // If the destination block has a single pred, then this is a trivial edge,
389 if (BasicBlock *SinglePred = DestBB->getSinglePredecessor()) {
390 if (SinglePred != DestBB) {
391 // Remember if SinglePred was the entry block of the function. If so, we
392 // will need to move BB back to the entry position.
393 bool isEntry = SinglePred == &SinglePred->getParent()->getEntryBlock();
394 MergeBasicBlockIntoOnlyPred(DestBB, this);
396 if (isEntry && BB != &BB->getParent()->getEntryBlock())
397 BB->moveBefore(&BB->getParent()->getEntryBlock());
399 DEBUG(dbgs() << "AFTER:\n" << *DestBB << "\n\n\n");
404 // Otherwise, we have multiple predecessors of BB. Update the PHIs in DestBB
405 // to handle the new incoming edges it is about to have.
407 for (BasicBlock::iterator BBI = DestBB->begin();
408 (PN = dyn_cast<PHINode>(BBI)); ++BBI) {
409 // Remove the incoming value for BB, and remember it.
410 Value *InVal = PN->removeIncomingValue(BB, false);
412 // Two options: either the InVal is a phi node defined in BB or it is some
413 // value that dominates BB.
414 PHINode *InValPhi = dyn_cast<PHINode>(InVal);
415 if (InValPhi && InValPhi->getParent() == BB) {
416 // Add all of the input values of the input PHI as inputs of this phi.
417 for (unsigned i = 0, e = InValPhi->getNumIncomingValues(); i != e; ++i)
418 PN->addIncoming(InValPhi->getIncomingValue(i),
419 InValPhi->getIncomingBlock(i));
421 // Otherwise, add one instance of the dominating value for each edge that
422 // we will be adding.
423 if (PHINode *BBPN = dyn_cast<PHINode>(BB->begin())) {
424 for (unsigned i = 0, e = BBPN->getNumIncomingValues(); i != e; ++i)
425 PN->addIncoming(InVal, BBPN->getIncomingBlock(i));
427 for (pred_iterator PI = pred_begin(BB), E = pred_end(BB); PI != E; ++PI)
428 PN->addIncoming(InVal, *PI);
433 // The PHIs are now updated, change everything that refers to BB to use
434 // DestBB and remove BB.
435 BB->replaceAllUsesWith(DestBB);
436 if (DT && !ModifiedDT) {
437 BasicBlock *BBIDom = DT->getNode(BB)->getIDom()->getBlock();
438 BasicBlock *DestBBIDom = DT->getNode(DestBB)->getIDom()->getBlock();
439 BasicBlock *NewIDom = DT->findNearestCommonDominator(BBIDom, DestBBIDom);
440 DT->changeImmediateDominator(DestBB, NewIDom);
444 PFI->replaceAllUses(BB, DestBB);
445 PFI->removeEdge(ProfileInfo::getEdge(BB, DestBB));
447 BB->eraseFromParent();
450 DEBUG(dbgs() << "AFTER:\n" << *DestBB << "\n\n\n");
453 /// OptimizeNoopCopyExpression - If the specified cast instruction is a noop
454 /// copy (e.g. it's casting from one pointer type to another, i32->i8 on PPC),
455 /// sink it into user blocks to reduce the number of virtual
456 /// registers that must be created and coalesced.
458 /// Return true if any changes are made.
460 static bool OptimizeNoopCopyExpression(CastInst *CI, const TargetLowering &TLI){
461 // If this is a noop copy,
462 EVT SrcVT = TLI.getValueType(CI->getOperand(0)->getType());
463 EVT DstVT = TLI.getValueType(CI->getType());
465 // This is an fp<->int conversion?
466 if (SrcVT.isInteger() != DstVT.isInteger())
469 // If this is an extension, it will be a zero or sign extension, which
471 if (SrcVT.bitsLT(DstVT)) return false;
473 // If these values will be promoted, find out what they will be promoted
474 // to. This helps us consider truncates on PPC as noop copies when they
476 if (TLI.getTypeAction(CI->getContext(), SrcVT) ==
477 TargetLowering::TypePromoteInteger)
478 SrcVT = TLI.getTypeToTransformTo(CI->getContext(), SrcVT);
479 if (TLI.getTypeAction(CI->getContext(), DstVT) ==
480 TargetLowering::TypePromoteInteger)
481 DstVT = TLI.getTypeToTransformTo(CI->getContext(), DstVT);
483 // If, after promotion, these are the same types, this is a noop copy.
487 BasicBlock *DefBB = CI->getParent();
489 /// InsertedCasts - Only insert a cast in each block once.
490 DenseMap<BasicBlock*, CastInst*> InsertedCasts;
492 bool MadeChange = false;
493 for (Value::use_iterator UI = CI->use_begin(), E = CI->use_end();
495 Use &TheUse = UI.getUse();
496 Instruction *User = cast<Instruction>(*UI);
498 // Figure out which BB this cast is used in. For PHI's this is the
499 // appropriate predecessor block.
500 BasicBlock *UserBB = User->getParent();
501 if (PHINode *PN = dyn_cast<PHINode>(User)) {
502 UserBB = PN->getIncomingBlock(UI);
505 // Preincrement use iterator so we don't invalidate it.
508 // If this user is in the same block as the cast, don't change the cast.
509 if (UserBB == DefBB) continue;
511 // If we have already inserted a cast into this block, use it.
512 CastInst *&InsertedCast = InsertedCasts[UserBB];
515 BasicBlock::iterator InsertPt = UserBB->getFirstInsertionPt();
517 CastInst::Create(CI->getOpcode(), CI->getOperand(0), CI->getType(), "",
522 // Replace a use of the cast with a use of the new cast.
523 TheUse = InsertedCast;
527 // If we removed all uses, nuke the cast.
528 if (CI->use_empty()) {
529 CI->eraseFromParent();
536 /// OptimizeCmpExpression - sink the given CmpInst into user blocks to reduce
537 /// the number of virtual registers that must be created and coalesced. This is
538 /// a clear win except on targets with multiple condition code registers
539 /// (PowerPC), where it might lose; some adjustment may be wanted there.
541 /// Return true if any changes are made.
542 static bool OptimizeCmpExpression(CmpInst *CI) {
543 BasicBlock *DefBB = CI->getParent();
545 /// InsertedCmp - Only insert a cmp in each block once.
546 DenseMap<BasicBlock*, CmpInst*> InsertedCmps;
548 bool MadeChange = false;
549 for (Value::use_iterator UI = CI->use_begin(), E = CI->use_end();
551 Use &TheUse = UI.getUse();
552 Instruction *User = cast<Instruction>(*UI);
554 // Preincrement use iterator so we don't invalidate it.
557 // Don't bother for PHI nodes.
558 if (isa<PHINode>(User))
561 // Figure out which BB this cmp is used in.
562 BasicBlock *UserBB = User->getParent();
564 // If this user is in the same block as the cmp, don't change the cmp.
565 if (UserBB == DefBB) continue;
567 // If we have already inserted a cmp into this block, use it.
568 CmpInst *&InsertedCmp = InsertedCmps[UserBB];
571 BasicBlock::iterator InsertPt = UserBB->getFirstInsertionPt();
573 CmpInst::Create(CI->getOpcode(),
574 CI->getPredicate(), CI->getOperand(0),
575 CI->getOperand(1), "", InsertPt);
579 // Replace a use of the cmp with a use of the new cmp.
580 TheUse = InsertedCmp;
584 // If we removed all uses, nuke the cmp.
586 CI->eraseFromParent();
592 class CodeGenPrepareFortifiedLibCalls : public SimplifyFortifiedLibCalls {
594 void replaceCall(Value *With) {
595 CI->replaceAllUsesWith(With);
596 CI->eraseFromParent();
598 bool isFoldable(unsigned SizeCIOp, unsigned, bool) const {
599 if (ConstantInt *SizeCI =
600 dyn_cast<ConstantInt>(CI->getArgOperand(SizeCIOp)))
601 return SizeCI->isAllOnesValue();
605 } // end anonymous namespace
607 bool CodeGenPrepare::OptimizeCallInst(CallInst *CI) {
608 BasicBlock *BB = CI->getParent();
610 // Lower inline assembly if we can.
611 // If we found an inline asm expession, and if the target knows how to
612 // lower it to normal LLVM code, do so now.
613 if (TLI && isa<InlineAsm>(CI->getCalledValue())) {
614 if (TLI->ExpandInlineAsm(CI)) {
615 // Avoid invalidating the iterator.
616 CurInstIterator = BB->begin();
617 // Avoid processing instructions out of order, which could cause
618 // reuse before a value is defined.
622 // Sink address computing for memory operands into the block.
623 if (OptimizeInlineAsmInst(CI))
627 // Lower all uses of llvm.objectsize.*
628 IntrinsicInst *II = dyn_cast<IntrinsicInst>(CI);
629 if (II && II->getIntrinsicID() == Intrinsic::objectsize) {
630 bool Min = (cast<ConstantInt>(II->getArgOperand(1))->getZExtValue() == 1);
631 Type *ReturnTy = CI->getType();
632 Constant *RetVal = ConstantInt::get(ReturnTy, Min ? 0 : -1ULL);
634 // Substituting this can cause recursive simplifications, which can
635 // invalidate our iterator. Use a WeakVH to hold onto it in case this
637 WeakVH IterHandle(CurInstIterator);
639 replaceAndRecursivelySimplify(CI, RetVal, TLI ? TLI->getDataLayout() : 0,
640 TLInfo, ModifiedDT ? 0 : DT);
642 // If the iterator instruction was recursively deleted, start over at the
643 // start of the block.
644 if (IterHandle != CurInstIterator) {
645 CurInstIterator = BB->begin();
652 SmallVector<Value*, 2> PtrOps;
654 if (TLI->GetAddrModeArguments(II, PtrOps, AccessTy))
655 while (!PtrOps.empty())
656 if (OptimizeMemoryInst(II, PtrOps.pop_back_val(), AccessTy))
660 // From here on out we're working with named functions.
661 if (CI->getCalledFunction() == 0) return false;
663 // We'll need DataLayout from here on out.
664 const DataLayout *TD = TLI ? TLI->getDataLayout() : 0;
665 if (!TD) return false;
667 // Lower all default uses of _chk calls. This is very similar
668 // to what InstCombineCalls does, but here we are only lowering calls
669 // that have the default "don't know" as the objectsize. Anything else
670 // should be left alone.
671 CodeGenPrepareFortifiedLibCalls Simplifier;
672 return Simplifier.fold(CI, TD, TLInfo);
675 /// DupRetToEnableTailCallOpts - Look for opportunities to duplicate return
676 /// instructions to the predecessor to enable tail call optimizations. The
677 /// case it is currently looking for is:
680 /// %tmp0 = tail call i32 @f0()
683 /// %tmp1 = tail call i32 @f1()
686 /// %tmp2 = tail call i32 @f2()
689 /// %retval = phi i32 [ %tmp0, %bb0 ], [ %tmp1, %bb1 ], [ %tmp2, %bb2 ]
697 /// %tmp0 = tail call i32 @f0()
700 /// %tmp1 = tail call i32 @f1()
703 /// %tmp2 = tail call i32 @f2()
706 bool CodeGenPrepare::DupRetToEnableTailCallOpts(BasicBlock *BB) {
710 ReturnInst *RI = dyn_cast<ReturnInst>(BB->getTerminator());
715 BitCastInst *BCI = 0;
716 Value *V = RI->getReturnValue();
718 BCI = dyn_cast<BitCastInst>(V);
720 V = BCI->getOperand(0);
722 PN = dyn_cast<PHINode>(V);
727 if (PN && PN->getParent() != BB)
730 // It's not safe to eliminate the sign / zero extension of the return value.
731 // See llvm::isInTailCallPosition().
732 const Function *F = BB->getParent();
733 AttributeSet CallerAttrs = F->getAttributes();
734 if (CallerAttrs.hasAttribute(AttributeSet::ReturnIndex, Attribute::ZExt) ||
735 CallerAttrs.hasAttribute(AttributeSet::ReturnIndex, Attribute::SExt))
738 // Make sure there are no instructions between the PHI and return, or that the
739 // return is the first instruction in the block.
741 BasicBlock::iterator BI = BB->begin();
742 do { ++BI; } while (isa<DbgInfoIntrinsic>(BI));
744 // Also skip over the bitcast.
749 BasicBlock::iterator BI = BB->begin();
750 while (isa<DbgInfoIntrinsic>(BI)) ++BI;
755 /// Only dup the ReturnInst if the CallInst is likely to be emitted as a tail
757 SmallVector<CallInst*, 4> TailCalls;
759 for (unsigned I = 0, E = PN->getNumIncomingValues(); I != E; ++I) {
760 CallInst *CI = dyn_cast<CallInst>(PN->getIncomingValue(I));
761 // Make sure the phi value is indeed produced by the tail call.
762 if (CI && CI->hasOneUse() && CI->getParent() == PN->getIncomingBlock(I) &&
763 TLI->mayBeEmittedAsTailCall(CI))
764 TailCalls.push_back(CI);
767 SmallPtrSet<BasicBlock*, 4> VisitedBBs;
768 for (pred_iterator PI = pred_begin(BB), PE = pred_end(BB); PI != PE; ++PI) {
769 if (!VisitedBBs.insert(*PI))
772 BasicBlock::InstListType &InstList = (*PI)->getInstList();
773 BasicBlock::InstListType::reverse_iterator RI = InstList.rbegin();
774 BasicBlock::InstListType::reverse_iterator RE = InstList.rend();
775 do { ++RI; } while (RI != RE && isa<DbgInfoIntrinsic>(&*RI));
779 CallInst *CI = dyn_cast<CallInst>(&*RI);
780 if (CI && CI->use_empty() && TLI->mayBeEmittedAsTailCall(CI))
781 TailCalls.push_back(CI);
785 bool Changed = false;
786 for (unsigned i = 0, e = TailCalls.size(); i != e; ++i) {
787 CallInst *CI = TailCalls[i];
790 // Conservatively require the attributes of the call to match those of the
791 // return. Ignore noalias because it doesn't affect the call sequence.
792 AttributeSet CalleeAttrs = CS.getAttributes();
793 if (AttrBuilder(CalleeAttrs, AttributeSet::ReturnIndex).
794 removeAttribute(Attribute::NoAlias) !=
795 AttrBuilder(CalleeAttrs, AttributeSet::ReturnIndex).
796 removeAttribute(Attribute::NoAlias))
799 // Make sure the call instruction is followed by an unconditional branch to
801 BasicBlock *CallBB = CI->getParent();
802 BranchInst *BI = dyn_cast<BranchInst>(CallBB->getTerminator());
803 if (!BI || !BI->isUnconditional() || BI->getSuccessor(0) != BB)
806 // Duplicate the return into CallBB.
807 (void)FoldReturnIntoUncondBranch(RI, BB, CallBB);
808 ModifiedDT = Changed = true;
812 // If we eliminated all predecessors of the block, delete the block now.
813 if (Changed && !BB->hasAddressTaken() && pred_begin(BB) == pred_end(BB))
814 BB->eraseFromParent();
819 //===----------------------------------------------------------------------===//
820 // Memory Optimization
821 //===----------------------------------------------------------------------===//
825 /// ExtAddrMode - This is an extended version of TargetLowering::AddrMode
826 /// which holds actual Value*'s for register values.
827 struct ExtAddrMode : public TargetLowering::AddrMode {
830 ExtAddrMode() : BaseReg(0), ScaledReg(0) {}
831 void print(raw_ostream &OS) const;
834 bool operator==(const ExtAddrMode& O) const {
835 return (BaseReg == O.BaseReg) && (ScaledReg == O.ScaledReg) &&
836 (BaseGV == O.BaseGV) && (BaseOffs == O.BaseOffs) &&
837 (HasBaseReg == O.HasBaseReg) && (Scale == O.Scale);
841 static inline raw_ostream &operator<<(raw_ostream &OS, const ExtAddrMode &AM) {
846 void ExtAddrMode::print(raw_ostream &OS) const {
847 bool NeedPlus = false;
850 OS << (NeedPlus ? " + " : "")
852 WriteAsOperand(OS, BaseGV, /*PrintType=*/false);
857 OS << (NeedPlus ? " + " : "") << BaseOffs, NeedPlus = true;
860 OS << (NeedPlus ? " + " : "")
862 WriteAsOperand(OS, BaseReg, /*PrintType=*/false);
866 OS << (NeedPlus ? " + " : "")
868 WriteAsOperand(OS, ScaledReg, /*PrintType=*/false);
875 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
876 void ExtAddrMode::dump() const {
883 /// \brief A helper class for matching addressing modes.
885 /// This encapsulates the logic for matching the target-legal addressing modes.
886 class AddressingModeMatcher {
887 SmallVectorImpl<Instruction*> &AddrModeInsts;
888 const TargetLowering &TLI;
890 /// AccessTy/MemoryInst - This is the type for the access (e.g. double) and
891 /// the memory instruction that we're computing this address for.
893 Instruction *MemoryInst;
895 /// AddrMode - This is the addressing mode that we're building up. This is
896 /// part of the return value of this addressing mode matching stuff.
897 ExtAddrMode &AddrMode;
899 /// IgnoreProfitability - This is set to true when we should not do
900 /// profitability checks. When true, IsProfitableToFoldIntoAddressingMode
901 /// always returns true.
902 bool IgnoreProfitability;
904 AddressingModeMatcher(SmallVectorImpl<Instruction*> &AMI,
905 const TargetLowering &T, Type *AT,
906 Instruction *MI, ExtAddrMode &AM)
907 : AddrModeInsts(AMI), TLI(T), AccessTy(AT), MemoryInst(MI), AddrMode(AM) {
908 IgnoreProfitability = false;
912 /// Match - Find the maximal addressing mode that a load/store of V can fold,
913 /// give an access type of AccessTy. This returns a list of involved
914 /// instructions in AddrModeInsts.
915 static ExtAddrMode Match(Value *V, Type *AccessTy,
916 Instruction *MemoryInst,
917 SmallVectorImpl<Instruction*> &AddrModeInsts,
918 const TargetLowering &TLI) {
922 AddressingModeMatcher(AddrModeInsts, TLI, AccessTy,
923 MemoryInst, Result).MatchAddr(V, 0);
924 (void)Success; assert(Success && "Couldn't select *anything*?");
928 bool MatchScaledValue(Value *ScaleReg, int64_t Scale, unsigned Depth);
929 bool MatchAddr(Value *V, unsigned Depth);
930 bool MatchOperationAddr(User *Operation, unsigned Opcode, unsigned Depth);
931 bool IsProfitableToFoldIntoAddressingMode(Instruction *I,
932 ExtAddrMode &AMBefore,
933 ExtAddrMode &AMAfter);
934 bool ValueAlreadyLiveAtInst(Value *Val, Value *KnownLive1, Value *KnownLive2);
937 /// MatchScaledValue - Try adding ScaleReg*Scale to the current addressing mode.
938 /// Return true and update AddrMode if this addr mode is legal for the target,
940 bool AddressingModeMatcher::MatchScaledValue(Value *ScaleReg, int64_t Scale,
942 // If Scale is 1, then this is the same as adding ScaleReg to the addressing
943 // mode. Just process that directly.
945 return MatchAddr(ScaleReg, Depth);
947 // If the scale is 0, it takes nothing to add this.
951 // If we already have a scale of this value, we can add to it, otherwise, we
952 // need an available scale field.
953 if (AddrMode.Scale != 0 && AddrMode.ScaledReg != ScaleReg)
956 ExtAddrMode TestAddrMode = AddrMode;
958 // Add scale to turn X*4+X*3 -> X*7. This could also do things like
959 // [A+B + A*7] -> [B+A*8].
960 TestAddrMode.Scale += Scale;
961 TestAddrMode.ScaledReg = ScaleReg;
963 // If the new address isn't legal, bail out.
964 if (!TLI.isLegalAddressingMode(TestAddrMode, AccessTy))
967 // It was legal, so commit it.
968 AddrMode = TestAddrMode;
970 // Okay, we decided that we can add ScaleReg+Scale to AddrMode. Check now
971 // to see if ScaleReg is actually X+C. If so, we can turn this into adding
972 // X*Scale + C*Scale to addr mode.
973 ConstantInt *CI = 0; Value *AddLHS = 0;
974 if (isa<Instruction>(ScaleReg) && // not a constant expr.
975 match(ScaleReg, m_Add(m_Value(AddLHS), m_ConstantInt(CI)))) {
976 TestAddrMode.ScaledReg = AddLHS;
977 TestAddrMode.BaseOffs += CI->getSExtValue()*TestAddrMode.Scale;
979 // If this addressing mode is legal, commit it and remember that we folded
981 if (TLI.isLegalAddressingMode(TestAddrMode, AccessTy)) {
982 AddrModeInsts.push_back(cast<Instruction>(ScaleReg));
983 AddrMode = TestAddrMode;
988 // Otherwise, not (x+c)*scale, just return what we have.
992 /// MightBeFoldableInst - This is a little filter, which returns true if an
993 /// addressing computation involving I might be folded into a load/store
994 /// accessing it. This doesn't need to be perfect, but needs to accept at least
995 /// the set of instructions that MatchOperationAddr can.
996 static bool MightBeFoldableInst(Instruction *I) {
997 switch (I->getOpcode()) {
998 case Instruction::BitCast:
999 // Don't touch identity bitcasts.
1000 if (I->getType() == I->getOperand(0)->getType())
1002 return I->getType()->isPointerTy() || I->getType()->isIntegerTy();
1003 case Instruction::PtrToInt:
1004 // PtrToInt is always a noop, as we know that the int type is pointer sized.
1006 case Instruction::IntToPtr:
1007 // We know the input is intptr_t, so this is foldable.
1009 case Instruction::Add:
1011 case Instruction::Mul:
1012 case Instruction::Shl:
1013 // Can only handle X*C and X << C.
1014 return isa<ConstantInt>(I->getOperand(1));
1015 case Instruction::GetElementPtr:
1022 /// MatchOperationAddr - Given an instruction or constant expr, see if we can
1023 /// fold the operation into the addressing mode. If so, update the addressing
1024 /// mode and return true, otherwise return false without modifying AddrMode.
1025 bool AddressingModeMatcher::MatchOperationAddr(User *AddrInst, unsigned Opcode,
1027 // Avoid exponential behavior on extremely deep expression trees.
1028 if (Depth >= 5) return false;
1031 case Instruction::PtrToInt:
1032 // PtrToInt is always a noop, as we know that the int type is pointer sized.
1033 return MatchAddr(AddrInst->getOperand(0), Depth);
1034 case Instruction::IntToPtr:
1035 // This inttoptr is a no-op if the integer type is pointer sized.
1036 if (TLI.getValueType(AddrInst->getOperand(0)->getType()) ==
1038 return MatchAddr(AddrInst->getOperand(0), Depth);
1040 case Instruction::BitCast:
1041 // BitCast is always a noop, and we can handle it as long as it is
1042 // int->int or pointer->pointer (we don't want int<->fp or something).
1043 if ((AddrInst->getOperand(0)->getType()->isPointerTy() ||
1044 AddrInst->getOperand(0)->getType()->isIntegerTy()) &&
1045 // Don't touch identity bitcasts. These were probably put here by LSR,
1046 // and we don't want to mess around with them. Assume it knows what it
1048 AddrInst->getOperand(0)->getType() != AddrInst->getType())
1049 return MatchAddr(AddrInst->getOperand(0), Depth);
1051 case Instruction::Add: {
1052 // Check to see if we can merge in the RHS then the LHS. If so, we win.
1053 ExtAddrMode BackupAddrMode = AddrMode;
1054 unsigned OldSize = AddrModeInsts.size();
1055 if (MatchAddr(AddrInst->getOperand(1), Depth+1) &&
1056 MatchAddr(AddrInst->getOperand(0), Depth+1))
1059 // Restore the old addr mode info.
1060 AddrMode = BackupAddrMode;
1061 AddrModeInsts.resize(OldSize);
1063 // Otherwise this was over-aggressive. Try merging in the LHS then the RHS.
1064 if (MatchAddr(AddrInst->getOperand(0), Depth+1) &&
1065 MatchAddr(AddrInst->getOperand(1), Depth+1))
1068 // Otherwise we definitely can't merge the ADD in.
1069 AddrMode = BackupAddrMode;
1070 AddrModeInsts.resize(OldSize);
1073 //case Instruction::Or:
1074 // TODO: We can handle "Or Val, Imm" iff this OR is equivalent to an ADD.
1076 case Instruction::Mul:
1077 case Instruction::Shl: {
1078 // Can only handle X*C and X << C.
1079 ConstantInt *RHS = dyn_cast<ConstantInt>(AddrInst->getOperand(1));
1080 if (!RHS) return false;
1081 int64_t Scale = RHS->getSExtValue();
1082 if (Opcode == Instruction::Shl)
1083 Scale = 1LL << Scale;
1085 return MatchScaledValue(AddrInst->getOperand(0), Scale, Depth);
1087 case Instruction::GetElementPtr: {
1088 // Scan the GEP. We check it if it contains constant offsets and at most
1089 // one variable offset.
1090 int VariableOperand = -1;
1091 unsigned VariableScale = 0;
1093 int64_t ConstantOffset = 0;
1094 const DataLayout *TD = TLI.getDataLayout();
1095 gep_type_iterator GTI = gep_type_begin(AddrInst);
1096 for (unsigned i = 1, e = AddrInst->getNumOperands(); i != e; ++i, ++GTI) {
1097 if (StructType *STy = dyn_cast<StructType>(*GTI)) {
1098 const StructLayout *SL = TD->getStructLayout(STy);
1100 cast<ConstantInt>(AddrInst->getOperand(i))->getZExtValue();
1101 ConstantOffset += SL->getElementOffset(Idx);
1103 uint64_t TypeSize = TD->getTypeAllocSize(GTI.getIndexedType());
1104 if (ConstantInt *CI = dyn_cast<ConstantInt>(AddrInst->getOperand(i))) {
1105 ConstantOffset += CI->getSExtValue()*TypeSize;
1106 } else if (TypeSize) { // Scales of zero don't do anything.
1107 // We only allow one variable index at the moment.
1108 if (VariableOperand != -1)
1111 // Remember the variable index.
1112 VariableOperand = i;
1113 VariableScale = TypeSize;
1118 // A common case is for the GEP to only do a constant offset. In this case,
1119 // just add it to the disp field and check validity.
1120 if (VariableOperand == -1) {
1121 AddrMode.BaseOffs += ConstantOffset;
1122 if (ConstantOffset == 0 || TLI.isLegalAddressingMode(AddrMode, AccessTy)){
1123 // Check to see if we can fold the base pointer in too.
1124 if (MatchAddr(AddrInst->getOperand(0), Depth+1))
1127 AddrMode.BaseOffs -= ConstantOffset;
1131 // Save the valid addressing mode in case we can't match.
1132 ExtAddrMode BackupAddrMode = AddrMode;
1133 unsigned OldSize = AddrModeInsts.size();
1135 // See if the scale and offset amount is valid for this target.
1136 AddrMode.BaseOffs += ConstantOffset;
1138 // Match the base operand of the GEP.
1139 if (!MatchAddr(AddrInst->getOperand(0), Depth+1)) {
1140 // If it couldn't be matched, just stuff the value in a register.
1141 if (AddrMode.HasBaseReg) {
1142 AddrMode = BackupAddrMode;
1143 AddrModeInsts.resize(OldSize);
1146 AddrMode.HasBaseReg = true;
1147 AddrMode.BaseReg = AddrInst->getOperand(0);
1150 // Match the remaining variable portion of the GEP.
1151 if (!MatchScaledValue(AddrInst->getOperand(VariableOperand), VariableScale,
1153 // If it couldn't be matched, try stuffing the base into a register
1154 // instead of matching it, and retrying the match of the scale.
1155 AddrMode = BackupAddrMode;
1156 AddrModeInsts.resize(OldSize);
1157 if (AddrMode.HasBaseReg)
1159 AddrMode.HasBaseReg = true;
1160 AddrMode.BaseReg = AddrInst->getOperand(0);
1161 AddrMode.BaseOffs += ConstantOffset;
1162 if (!MatchScaledValue(AddrInst->getOperand(VariableOperand),
1163 VariableScale, Depth)) {
1164 // If even that didn't work, bail.
1165 AddrMode = BackupAddrMode;
1166 AddrModeInsts.resize(OldSize);
1177 /// MatchAddr - If we can, try to add the value of 'Addr' into the current
1178 /// addressing mode. If Addr can't be added to AddrMode this returns false and
1179 /// leaves AddrMode unmodified. This assumes that Addr is either a pointer type
1180 /// or intptr_t for the target.
1182 bool AddressingModeMatcher::MatchAddr(Value *Addr, unsigned Depth) {
1183 if (ConstantInt *CI = dyn_cast<ConstantInt>(Addr)) {
1184 // Fold in immediates if legal for the target.
1185 AddrMode.BaseOffs += CI->getSExtValue();
1186 if (TLI.isLegalAddressingMode(AddrMode, AccessTy))
1188 AddrMode.BaseOffs -= CI->getSExtValue();
1189 } else if (GlobalValue *GV = dyn_cast<GlobalValue>(Addr)) {
1190 // If this is a global variable, try to fold it into the addressing mode.
1191 if (AddrMode.BaseGV == 0) {
1192 AddrMode.BaseGV = GV;
1193 if (TLI.isLegalAddressingMode(AddrMode, AccessTy))
1195 AddrMode.BaseGV = 0;
1197 } else if (Instruction *I = dyn_cast<Instruction>(Addr)) {
1198 ExtAddrMode BackupAddrMode = AddrMode;
1199 unsigned OldSize = AddrModeInsts.size();
1201 // Check to see if it is possible to fold this operation.
1202 if (MatchOperationAddr(I, I->getOpcode(), Depth)) {
1203 // Okay, it's possible to fold this. Check to see if it is actually
1204 // *profitable* to do so. We use a simple cost model to avoid increasing
1205 // register pressure too much.
1206 if (I->hasOneUse() ||
1207 IsProfitableToFoldIntoAddressingMode(I, BackupAddrMode, AddrMode)) {
1208 AddrModeInsts.push_back(I);
1212 // It isn't profitable to do this, roll back.
1213 //cerr << "NOT FOLDING: " << *I;
1214 AddrMode = BackupAddrMode;
1215 AddrModeInsts.resize(OldSize);
1217 } else if (ConstantExpr *CE = dyn_cast<ConstantExpr>(Addr)) {
1218 if (MatchOperationAddr(CE, CE->getOpcode(), Depth))
1220 } else if (isa<ConstantPointerNull>(Addr)) {
1221 // Null pointer gets folded without affecting the addressing mode.
1225 // Worse case, the target should support [reg] addressing modes. :)
1226 if (!AddrMode.HasBaseReg) {
1227 AddrMode.HasBaseReg = true;
1228 AddrMode.BaseReg = Addr;
1229 // Still check for legality in case the target supports [imm] but not [i+r].
1230 if (TLI.isLegalAddressingMode(AddrMode, AccessTy))
1232 AddrMode.HasBaseReg = false;
1233 AddrMode.BaseReg = 0;
1236 // If the base register is already taken, see if we can do [r+r].
1237 if (AddrMode.Scale == 0) {
1239 AddrMode.ScaledReg = Addr;
1240 if (TLI.isLegalAddressingMode(AddrMode, AccessTy))
1243 AddrMode.ScaledReg = 0;
1249 /// IsOperandAMemoryOperand - Check to see if all uses of OpVal by the specified
1250 /// inline asm call are due to memory operands. If so, return true, otherwise
1252 static bool IsOperandAMemoryOperand(CallInst *CI, InlineAsm *IA, Value *OpVal,
1253 const TargetLowering &TLI) {
1254 TargetLowering::AsmOperandInfoVector TargetConstraints = TLI.ParseConstraints(ImmutableCallSite(CI));
1255 for (unsigned i = 0, e = TargetConstraints.size(); i != e; ++i) {
1256 TargetLowering::AsmOperandInfo &OpInfo = TargetConstraints[i];
1258 // Compute the constraint code and ConstraintType to use.
1259 TLI.ComputeConstraintToUse(OpInfo, SDValue());
1261 // If this asm operand is our Value*, and if it isn't an indirect memory
1262 // operand, we can't fold it!
1263 if (OpInfo.CallOperandVal == OpVal &&
1264 (OpInfo.ConstraintType != TargetLowering::C_Memory ||
1265 !OpInfo.isIndirect))
1272 /// FindAllMemoryUses - Recursively walk all the uses of I until we find a
1273 /// memory use. If we find an obviously non-foldable instruction, return true.
1274 /// Add the ultimately found memory instructions to MemoryUses.
1275 static bool FindAllMemoryUses(Instruction *I,
1276 SmallVectorImpl<std::pair<Instruction*,unsigned> > &MemoryUses,
1277 SmallPtrSet<Instruction*, 16> &ConsideredInsts,
1278 const TargetLowering &TLI) {
1279 // If we already considered this instruction, we're done.
1280 if (!ConsideredInsts.insert(I))
1283 // If this is an obviously unfoldable instruction, bail out.
1284 if (!MightBeFoldableInst(I))
1287 // Loop over all the uses, recursively processing them.
1288 for (Value::use_iterator UI = I->use_begin(), E = I->use_end();
1292 if (LoadInst *LI = dyn_cast<LoadInst>(U)) {
1293 MemoryUses.push_back(std::make_pair(LI, UI.getOperandNo()));
1297 if (StoreInst *SI = dyn_cast<StoreInst>(U)) {
1298 unsigned opNo = UI.getOperandNo();
1299 if (opNo == 0) return true; // Storing addr, not into addr.
1300 MemoryUses.push_back(std::make_pair(SI, opNo));
1304 if (CallInst *CI = dyn_cast<CallInst>(U)) {
1305 InlineAsm *IA = dyn_cast<InlineAsm>(CI->getCalledValue());
1306 if (!IA) return true;
1308 // If this is a memory operand, we're cool, otherwise bail out.
1309 if (!IsOperandAMemoryOperand(CI, IA, I, TLI))
1314 if (FindAllMemoryUses(cast<Instruction>(U), MemoryUses, ConsideredInsts,
1322 /// ValueAlreadyLiveAtInst - Retrn true if Val is already known to be live at
1323 /// the use site that we're folding it into. If so, there is no cost to
1324 /// include it in the addressing mode. KnownLive1 and KnownLive2 are two values
1325 /// that we know are live at the instruction already.
1326 bool AddressingModeMatcher::ValueAlreadyLiveAtInst(Value *Val,Value *KnownLive1,
1327 Value *KnownLive2) {
1328 // If Val is either of the known-live values, we know it is live!
1329 if (Val == 0 || Val == KnownLive1 || Val == KnownLive2)
1332 // All values other than instructions and arguments (e.g. constants) are live.
1333 if (!isa<Instruction>(Val) && !isa<Argument>(Val)) return true;
1335 // If Val is a constant sized alloca in the entry block, it is live, this is
1336 // true because it is just a reference to the stack/frame pointer, which is
1337 // live for the whole function.
1338 if (AllocaInst *AI = dyn_cast<AllocaInst>(Val))
1339 if (AI->isStaticAlloca())
1342 // Check to see if this value is already used in the memory instruction's
1343 // block. If so, it's already live into the block at the very least, so we
1344 // can reasonably fold it.
1345 return Val->isUsedInBasicBlock(MemoryInst->getParent());
1348 /// IsProfitableToFoldIntoAddressingMode - It is possible for the addressing
1349 /// mode of the machine to fold the specified instruction into a load or store
1350 /// that ultimately uses it. However, the specified instruction has multiple
1351 /// uses. Given this, it may actually increase register pressure to fold it
1352 /// into the load. For example, consider this code:
1356 /// use(Y) -> nonload/store
1360 /// In this case, Y has multiple uses, and can be folded into the load of Z
1361 /// (yielding load [X+2]). However, doing this will cause both "X" and "X+1" to
1362 /// be live at the use(Y) line. If we don't fold Y into load Z, we use one
1363 /// fewer register. Since Y can't be folded into "use(Y)" we don't increase the
1364 /// number of computations either.
1366 /// Note that this (like most of CodeGenPrepare) is just a rough heuristic. If
1367 /// X was live across 'load Z' for other reasons, we actually *would* want to
1368 /// fold the addressing mode in the Z case. This would make Y die earlier.
1369 bool AddressingModeMatcher::
1370 IsProfitableToFoldIntoAddressingMode(Instruction *I, ExtAddrMode &AMBefore,
1371 ExtAddrMode &AMAfter) {
1372 if (IgnoreProfitability) return true;
1374 // AMBefore is the addressing mode before this instruction was folded into it,
1375 // and AMAfter is the addressing mode after the instruction was folded. Get
1376 // the set of registers referenced by AMAfter and subtract out those
1377 // referenced by AMBefore: this is the set of values which folding in this
1378 // address extends the lifetime of.
1380 // Note that there are only two potential values being referenced here,
1381 // BaseReg and ScaleReg (global addresses are always available, as are any
1382 // folded immediates).
1383 Value *BaseReg = AMAfter.BaseReg, *ScaledReg = AMAfter.ScaledReg;
1385 // If the BaseReg or ScaledReg was referenced by the previous addrmode, their
1386 // lifetime wasn't extended by adding this instruction.
1387 if (ValueAlreadyLiveAtInst(BaseReg, AMBefore.BaseReg, AMBefore.ScaledReg))
1389 if (ValueAlreadyLiveAtInst(ScaledReg, AMBefore.BaseReg, AMBefore.ScaledReg))
1392 // If folding this instruction (and it's subexprs) didn't extend any live
1393 // ranges, we're ok with it.
1394 if (BaseReg == 0 && ScaledReg == 0)
1397 // If all uses of this instruction are ultimately load/store/inlineasm's,
1398 // check to see if their addressing modes will include this instruction. If
1399 // so, we can fold it into all uses, so it doesn't matter if it has multiple
1401 SmallVector<std::pair<Instruction*,unsigned>, 16> MemoryUses;
1402 SmallPtrSet<Instruction*, 16> ConsideredInsts;
1403 if (FindAllMemoryUses(I, MemoryUses, ConsideredInsts, TLI))
1404 return false; // Has a non-memory, non-foldable use!
1406 // Now that we know that all uses of this instruction are part of a chain of
1407 // computation involving only operations that could theoretically be folded
1408 // into a memory use, loop over each of these uses and see if they could
1409 // *actually* fold the instruction.
1410 SmallVector<Instruction*, 32> MatchedAddrModeInsts;
1411 for (unsigned i = 0, e = MemoryUses.size(); i != e; ++i) {
1412 Instruction *User = MemoryUses[i].first;
1413 unsigned OpNo = MemoryUses[i].second;
1415 // Get the access type of this use. If the use isn't a pointer, we don't
1416 // know what it accesses.
1417 Value *Address = User->getOperand(OpNo);
1418 if (!Address->getType()->isPointerTy())
1420 Type *AddressAccessTy =
1421 cast<PointerType>(Address->getType())->getElementType();
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 "
1576 TLI->getDataLayout()->getIntPtrType(AccessTy->getContext());
1580 // Start with the base register. Do this first so that subsequent address
1581 // matching finds it last, which will prevent it from trying to match it
1582 // as the scaled value in case it happens to be a mul. That would be
1583 // problematic if we've sunk a different mul for the scale, because then
1584 // we'd end up sinking both muls.
1585 if (AddrMode.BaseReg) {
1586 Value *V = AddrMode.BaseReg;
1587 if (V->getType()->isPointerTy())
1588 V = Builder.CreatePtrToInt(V, IntPtrTy, "sunkaddr");
1589 if (V->getType() != IntPtrTy)
1590 V = Builder.CreateIntCast(V, IntPtrTy, /*isSigned=*/true, "sunkaddr");
1594 // Add the scale value.
1595 if (AddrMode.Scale) {
1596 Value *V = AddrMode.ScaledReg;
1597 if (V->getType() == IntPtrTy) {
1599 } else if (V->getType()->isPointerTy()) {
1600 V = Builder.CreatePtrToInt(V, IntPtrTy, "sunkaddr");
1601 } else if (cast<IntegerType>(IntPtrTy)->getBitWidth() <
1602 cast<IntegerType>(V->getType())->getBitWidth()) {
1603 V = Builder.CreateTrunc(V, IntPtrTy, "sunkaddr");
1605 V = Builder.CreateSExt(V, IntPtrTy, "sunkaddr");
1607 if (AddrMode.Scale != 1)
1608 V = Builder.CreateMul(V, ConstantInt::get(IntPtrTy, AddrMode.Scale),
1611 Result = Builder.CreateAdd(Result, V, "sunkaddr");
1616 // Add in the BaseGV if present.
1617 if (AddrMode.BaseGV) {
1618 Value *V = Builder.CreatePtrToInt(AddrMode.BaseGV, IntPtrTy, "sunkaddr");
1620 Result = Builder.CreateAdd(Result, V, "sunkaddr");
1625 // Add in the Base Offset if present.
1626 if (AddrMode.BaseOffs) {
1627 Value *V = ConstantInt::get(IntPtrTy, AddrMode.BaseOffs);
1629 Result = Builder.CreateAdd(Result, V, "sunkaddr");
1635 SunkAddr = Constant::getNullValue(Addr->getType());
1637 SunkAddr = Builder.CreateIntToPtr(Result, Addr->getType(), "sunkaddr");
1640 MemoryInst->replaceUsesOfWith(Repl, SunkAddr);
1642 // If we have no uses, recursively delete the value and all dead instructions
1644 if (Repl->use_empty()) {
1645 // This can cause recursive deletion, which can invalidate our iterator.
1646 // Use a WeakVH to hold onto it in case this happens.
1647 WeakVH IterHandle(CurInstIterator);
1648 BasicBlock *BB = CurInstIterator->getParent();
1650 RecursivelyDeleteTriviallyDeadInstructions(Repl, TLInfo);
1652 if (IterHandle != CurInstIterator) {
1653 // If the iterator instruction was recursively deleted, start over at the
1654 // start of the block.
1655 CurInstIterator = BB->begin();
1663 /// OptimizeInlineAsmInst - If there are any memory operands, use
1664 /// OptimizeMemoryInst to sink their address computing into the block when
1665 /// possible / profitable.
1666 bool CodeGenPrepare::OptimizeInlineAsmInst(CallInst *CS) {
1667 bool MadeChange = false;
1669 TargetLowering::AsmOperandInfoVector
1670 TargetConstraints = TLI->ParseConstraints(CS);
1672 for (unsigned i = 0, e = TargetConstraints.size(); i != e; ++i) {
1673 TargetLowering::AsmOperandInfo &OpInfo = TargetConstraints[i];
1675 // Compute the constraint code and ConstraintType to use.
1676 TLI->ComputeConstraintToUse(OpInfo, SDValue());
1678 if (OpInfo.ConstraintType == TargetLowering::C_Memory &&
1679 OpInfo.isIndirect) {
1680 Value *OpVal = CS->getArgOperand(ArgNo++);
1681 MadeChange |= OptimizeMemoryInst(CS, OpVal, OpVal->getType());
1682 } else if (OpInfo.Type == InlineAsm::isInput)
1689 /// MoveExtToFormExtLoad - Move a zext or sext fed by a load into the same
1690 /// basic block as the load, unless conditions are unfavorable. This allows
1691 /// SelectionDAG to fold the extend into the load.
1693 bool CodeGenPrepare::MoveExtToFormExtLoad(Instruction *I) {
1694 // Look for a load being extended.
1695 LoadInst *LI = dyn_cast<LoadInst>(I->getOperand(0));
1696 if (!LI) return false;
1698 // If they're already in the same block, there's nothing to do.
1699 if (LI->getParent() == I->getParent())
1702 // If the load has other users and the truncate is not free, this probably
1703 // isn't worthwhile.
1704 if (!LI->hasOneUse() &&
1705 TLI && (TLI->isTypeLegal(TLI->getValueType(LI->getType())) ||
1706 !TLI->isTypeLegal(TLI->getValueType(I->getType()))) &&
1707 !TLI->isTruncateFree(I->getType(), LI->getType()))
1710 // Check whether the target supports casts folded into loads.
1712 if (isa<ZExtInst>(I))
1713 LType = ISD::ZEXTLOAD;
1715 assert(isa<SExtInst>(I) && "Unexpected ext type!");
1716 LType = ISD::SEXTLOAD;
1718 if (TLI && !TLI->isLoadExtLegal(LType, TLI->getValueType(LI->getType())))
1721 // Move the extend into the same block as the load, so that SelectionDAG
1723 I->removeFromParent();
1729 bool CodeGenPrepare::OptimizeExtUses(Instruction *I) {
1730 BasicBlock *DefBB = I->getParent();
1732 // If the result of a {s|z}ext and its source are both live out, rewrite all
1733 // other uses of the source with result of extension.
1734 Value *Src = I->getOperand(0);
1735 if (Src->hasOneUse())
1738 // Only do this xform if truncating is free.
1739 if (TLI && !TLI->isTruncateFree(I->getType(), Src->getType()))
1742 // Only safe to perform the optimization if the source is also defined in
1744 if (!isa<Instruction>(Src) || DefBB != cast<Instruction>(Src)->getParent())
1747 bool DefIsLiveOut = false;
1748 for (Value::use_iterator UI = I->use_begin(), E = I->use_end();
1750 Instruction *User = cast<Instruction>(*UI);
1752 // Figure out which BB this ext is used in.
1753 BasicBlock *UserBB = User->getParent();
1754 if (UserBB == DefBB) continue;
1755 DefIsLiveOut = true;
1761 // Make sure none of the uses are PHI nodes.
1762 for (Value::use_iterator UI = Src->use_begin(), E = Src->use_end();
1764 Instruction *User = cast<Instruction>(*UI);
1765 BasicBlock *UserBB = User->getParent();
1766 if (UserBB == DefBB) continue;
1767 // Be conservative. We don't want this xform to end up introducing
1768 // reloads just before load / store instructions.
1769 if (isa<PHINode>(User) || isa<LoadInst>(User) || isa<StoreInst>(User))
1773 // InsertedTruncs - Only insert one trunc in each block once.
1774 DenseMap<BasicBlock*, Instruction*> InsertedTruncs;
1776 bool MadeChange = false;
1777 for (Value::use_iterator UI = Src->use_begin(), E = Src->use_end();
1779 Use &TheUse = UI.getUse();
1780 Instruction *User = cast<Instruction>(*UI);
1782 // Figure out which BB this ext is used in.
1783 BasicBlock *UserBB = User->getParent();
1784 if (UserBB == DefBB) continue;
1786 // Both src and def are live in this block. Rewrite the use.
1787 Instruction *&InsertedTrunc = InsertedTruncs[UserBB];
1789 if (!InsertedTrunc) {
1790 BasicBlock::iterator InsertPt = UserBB->getFirstInsertionPt();
1791 InsertedTrunc = new TruncInst(I, Src->getType(), "", InsertPt);
1794 // Replace a use of the {s|z}ext source with a use of the result.
1795 TheUse = InsertedTrunc;
1803 /// isFormingBranchFromSelectProfitable - Returns true if a SelectInst should be
1804 /// turned into an explicit branch.
1805 static bool isFormingBranchFromSelectProfitable(SelectInst *SI) {
1806 // FIXME: This should use the same heuristics as IfConversion to determine
1807 // whether a select is better represented as a branch. This requires that
1808 // branch probability metadata is preserved for the select, which is not the
1811 CmpInst *Cmp = dyn_cast<CmpInst>(SI->getCondition());
1813 // If the branch is predicted right, an out of order CPU can avoid blocking on
1814 // the compare. Emit cmovs on compares with a memory operand as branches to
1815 // avoid stalls on the load from memory. If the compare has more than one use
1816 // there's probably another cmov or setcc around so it's not worth emitting a
1821 Value *CmpOp0 = Cmp->getOperand(0);
1822 Value *CmpOp1 = Cmp->getOperand(1);
1824 // We check that the memory operand has one use to avoid uses of the loaded
1825 // value directly after the compare, making branches unprofitable.
1826 return Cmp->hasOneUse() &&
1827 ((isa<LoadInst>(CmpOp0) && CmpOp0->hasOneUse()) ||
1828 (isa<LoadInst>(CmpOp1) && CmpOp1->hasOneUse()));
1832 /// If we have a SelectInst that will likely profit from branch prediction,
1833 /// turn it into a branch.
1834 bool CodeGenPrepare::OptimizeSelectInst(SelectInst *SI) {
1835 bool VectorCond = !SI->getCondition()->getType()->isIntegerTy(1);
1837 // Can we convert the 'select' to CF ?
1838 if (DisableSelectToBranch || OptSize || !TLI || VectorCond)
1841 TargetLowering::SelectSupportKind SelectKind;
1843 SelectKind = TargetLowering::VectorMaskSelect;
1844 else if (SI->getType()->isVectorTy())
1845 SelectKind = TargetLowering::ScalarCondVectorVal;
1847 SelectKind = TargetLowering::ScalarValSelect;
1849 // Do we have efficient codegen support for this kind of 'selects' ?
1850 if (TLI->isSelectSupported(SelectKind)) {
1851 // We have efficient codegen support for the select instruction.
1852 // Check if it is profitable to keep this 'select'.
1853 if (!TLI->isPredictableSelectExpensive() ||
1854 !isFormingBranchFromSelectProfitable(SI))
1860 // First, we split the block containing the select into 2 blocks.
1861 BasicBlock *StartBlock = SI->getParent();
1862 BasicBlock::iterator SplitPt = ++(BasicBlock::iterator(SI));
1863 BasicBlock *NextBlock = StartBlock->splitBasicBlock(SplitPt, "select.end");
1865 // Create a new block serving as the landing pad for the branch.
1866 BasicBlock *SmallBlock = BasicBlock::Create(SI->getContext(), "select.mid",
1867 NextBlock->getParent(), NextBlock);
1869 // Move the unconditional branch from the block with the select in it into our
1870 // landing pad block.
1871 StartBlock->getTerminator()->eraseFromParent();
1872 BranchInst::Create(NextBlock, SmallBlock);
1874 // Insert the real conditional branch based on the original condition.
1875 BranchInst::Create(NextBlock, SmallBlock, SI->getCondition(), SI);
1877 // The select itself is replaced with a PHI Node.
1878 PHINode *PN = PHINode::Create(SI->getType(), 2, "", NextBlock->begin());
1880 PN->addIncoming(SI->getTrueValue(), StartBlock);
1881 PN->addIncoming(SI->getFalseValue(), SmallBlock);
1882 SI->replaceAllUsesWith(PN);
1883 SI->eraseFromParent();
1885 // Instruct OptimizeBlock to skip to the next block.
1886 CurInstIterator = StartBlock->end();
1887 ++NumSelectsExpanded;
1891 bool CodeGenPrepare::OptimizeInst(Instruction *I) {
1892 if (PHINode *P = dyn_cast<PHINode>(I)) {
1893 // It is possible for very late stage optimizations (such as SimplifyCFG)
1894 // to introduce PHI nodes too late to be cleaned up. If we detect such a
1895 // trivial PHI, go ahead and zap it here.
1896 if (Value *V = SimplifyInstruction(P)) {
1897 P->replaceAllUsesWith(V);
1898 P->eraseFromParent();
1905 if (CastInst *CI = dyn_cast<CastInst>(I)) {
1906 // If the source of the cast is a constant, then this should have
1907 // already been constant folded. The only reason NOT to constant fold
1908 // it is if something (e.g. LSR) was careful to place the constant
1909 // evaluation in a block other than then one that uses it (e.g. to hoist
1910 // the address of globals out of a loop). If this is the case, we don't
1911 // want to forward-subst the cast.
1912 if (isa<Constant>(CI->getOperand(0)))
1915 if (TLI && OptimizeNoopCopyExpression(CI, *TLI))
1918 if (isa<ZExtInst>(I) || isa<SExtInst>(I)) {
1919 bool MadeChange = MoveExtToFormExtLoad(I);
1920 return MadeChange | OptimizeExtUses(I);
1925 if (CmpInst *CI = dyn_cast<CmpInst>(I))
1926 return OptimizeCmpExpression(CI);
1928 if (LoadInst *LI = dyn_cast<LoadInst>(I)) {
1930 return OptimizeMemoryInst(I, I->getOperand(0), LI->getType());
1934 if (StoreInst *SI = dyn_cast<StoreInst>(I)) {
1936 return OptimizeMemoryInst(I, SI->getOperand(1),
1937 SI->getOperand(0)->getType());
1941 if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(I)) {
1942 if (GEPI->hasAllZeroIndices()) {
1943 /// The GEP operand must be a pointer, so must its result -> BitCast
1944 Instruction *NC = new BitCastInst(GEPI->getOperand(0), GEPI->getType(),
1945 GEPI->getName(), GEPI);
1946 GEPI->replaceAllUsesWith(NC);
1947 GEPI->eraseFromParent();
1955 if (CallInst *CI = dyn_cast<CallInst>(I))
1956 return OptimizeCallInst(CI);
1958 if (SelectInst *SI = dyn_cast<SelectInst>(I))
1959 return OptimizeSelectInst(SI);
1964 // In this pass we look for GEP and cast instructions that are used
1965 // across basic blocks and rewrite them to improve basic-block-at-a-time
1967 bool CodeGenPrepare::OptimizeBlock(BasicBlock &BB) {
1969 bool MadeChange = false;
1971 CurInstIterator = BB.begin();
1972 while (CurInstIterator != BB.end())
1973 MadeChange |= OptimizeInst(CurInstIterator++);
1975 MadeChange |= DupRetToEnableTailCallOpts(&BB);
1980 // llvm.dbg.value is far away from the value then iSel may not be able
1981 // handle it properly. iSel will drop llvm.dbg.value if it can not
1982 // find a node corresponding to the value.
1983 bool CodeGenPrepare::PlaceDbgValues(Function &F) {
1984 bool MadeChange = false;
1985 for (Function::iterator I = F.begin(), E = F.end(); I != E; ++I) {
1986 Instruction *PrevNonDbgInst = NULL;
1987 for (BasicBlock::iterator BI = I->begin(), BE = I->end(); BI != BE;) {
1988 Instruction *Insn = BI; ++BI;
1989 DbgValueInst *DVI = dyn_cast<DbgValueInst>(Insn);
1991 PrevNonDbgInst = Insn;
1995 Instruction *VI = dyn_cast_or_null<Instruction>(DVI->getValue());
1996 if (VI && VI != PrevNonDbgInst && !VI->isTerminator()) {
1997 DEBUG(dbgs() << "Moving Debug Value before :\n" << *DVI << ' ' << *VI);
1998 DVI->removeFromParent();
1999 if (isa<PHINode>(VI))
2000 DVI->insertBefore(VI->getParent()->getFirstInsertionPt());
2002 DVI->insertAfter(VI);