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/Constants.h"
19 #include "llvm/DerivedTypes.h"
20 #include "llvm/Function.h"
21 #include "llvm/GlobalVariable.h"
22 #include "llvm/IRBuilder.h"
23 #include "llvm/InlineAsm.h"
24 #include "llvm/Instructions.h"
25 #include "llvm/IntrinsicInst.h"
26 #include "llvm/Pass.h"
27 #include "llvm/ADT/DenseMap.h"
28 #include "llvm/ADT/SmallSet.h"
29 #include "llvm/ADT/Statistic.h"
30 #include "llvm/Analysis/Dominators.h"
31 #include "llvm/Analysis/InstructionSimplify.h"
32 #include "llvm/Analysis/ProfileInfo.h"
33 #include "llvm/Assembly/Writer.h"
34 #include "llvm/Support/CallSite.h"
35 #include "llvm/Support/CommandLine.h"
36 #include "llvm/Support/Debug.h"
37 #include "llvm/Support/GetElementPtrTypeIterator.h"
38 #include "llvm/Support/PatternMatch.h"
39 #include "llvm/Support/ValueHandle.h"
40 #include "llvm/Support/raw_ostream.h"
41 #include "llvm/Target/TargetData.h"
42 #include "llvm/Target/TargetLibraryInfo.h"
43 #include "llvm/Target/TargetLowering.h"
44 #include "llvm/Transforms/Utils/AddrModeMatcher.h"
45 #include "llvm/Transforms/Utils/BasicBlockUtils.h"
46 #include "llvm/Transforms/Utils/BuildLibCalls.h"
47 #include "llvm/Transforms/Utils/BypassSlowDivision.h"
48 #include "llvm/Transforms/Utils/Local.h"
50 using namespace llvm::PatternMatch;
52 STATISTIC(NumBlocksElim, "Number of blocks eliminated");
53 STATISTIC(NumPHIsElim, "Number of trivial PHIs eliminated");
54 STATISTIC(NumGEPsElim, "Number of GEPs converted to casts");
55 STATISTIC(NumCmpUses, "Number of uses of Cmp expressions replaced with uses of "
57 STATISTIC(NumCastUses, "Number of uses of Cast expressions replaced with uses "
59 STATISTIC(NumMemoryInsts, "Number of memory instructions whose address "
60 "computations were sunk");
61 STATISTIC(NumExtsMoved, "Number of [s|z]ext instructions combined with loads");
62 STATISTIC(NumExtUses, "Number of uses of [s|z]ext instructions optimized");
63 STATISTIC(NumRetsDup, "Number of return instructions duplicated");
64 STATISTIC(NumDbgValueMoved, "Number of debug value instructions moved");
65 STATISTIC(NumSelectsExpanded, "Number of selects turned into branches");
67 static cl::opt<bool> DisableBranchOpts(
68 "disable-cgp-branch-opts", cl::Hidden, cl::init(false),
69 cl::desc("Disable branch optimizations in CodeGenPrepare"));
71 static cl::opt<bool> DisableSelectToBranch(
72 "disable-cgp-select2branch", cl::Hidden, cl::init(false),
73 cl::desc("Disable select to branch conversion."));
76 class CodeGenPrepare : public FunctionPass {
77 /// TLI - Keep a pointer of a TargetLowering to consult for determining
78 /// transformation profitability.
79 const TargetLowering *TLI;
80 const TargetLibraryInfo *TLInfo;
84 /// CurInstIterator - As we scan instructions optimizing them, this is the
85 /// next instruction to optimize. Xforms that can invalidate this should
87 BasicBlock::iterator CurInstIterator;
89 /// Keeps track of non-local addresses that have been sunk into a block.
90 /// This allows us to avoid inserting duplicate code for blocks with
91 /// multiple load/stores of the same address.
92 DenseMap<Value*, Value*> SunkAddrs;
94 /// ModifiedDT - If CFG is modified in anyway, dominator tree may need to
98 /// OptSize - True if optimizing for size.
102 static char ID; // Pass identification, replacement for typeid
103 explicit CodeGenPrepare(const TargetLowering *tli = 0)
104 : FunctionPass(ID), TLI(tli) {
105 initializeCodeGenPreparePass(*PassRegistry::getPassRegistry());
107 bool runOnFunction(Function &F);
109 virtual void getAnalysisUsage(AnalysisUsage &AU) const {
110 AU.addPreserved<DominatorTree>();
111 AU.addPreserved<ProfileInfo>();
112 AU.addRequired<TargetLibraryInfo>();
116 bool EliminateFallThrough(Function &F);
117 bool EliminateMostlyEmptyBlocks(Function &F);
118 bool CanMergeBlocks(const BasicBlock *BB, const BasicBlock *DestBB) const;
119 void EliminateMostlyEmptyBlock(BasicBlock *BB);
120 bool OptimizeBlock(BasicBlock &BB);
121 bool OptimizeInst(Instruction *I);
122 bool OptimizeMemoryInst(Instruction *I, Value *Addr, Type *AccessTy);
123 bool OptimizeInlineAsmInst(CallInst *CS);
124 bool OptimizeCallInst(CallInst *CI);
125 bool MoveExtToFormExtLoad(Instruction *I);
126 bool OptimizeExtUses(Instruction *I);
127 bool OptimizeSelectInst(SelectInst *SI);
128 bool DupRetToEnableTailCallOpts(ReturnInst *RI);
129 bool PlaceDbgValues(Function &F);
130 bool ConvertLoadToSwitch(LoadInst *LI);
134 char CodeGenPrepare::ID = 0;
135 INITIALIZE_PASS_BEGIN(CodeGenPrepare, "codegenprepare",
136 "Optimize for code generation", false, false)
137 INITIALIZE_PASS_DEPENDENCY(TargetLibraryInfo)
138 INITIALIZE_PASS_END(CodeGenPrepare, "codegenprepare",
139 "Optimize for code generation", false, false)
141 FunctionPass *llvm::createCodeGenPreparePass(const TargetLowering *TLI) {
142 return new CodeGenPrepare(TLI);
145 bool CodeGenPrepare::runOnFunction(Function &F) {
146 bool EverMadeChange = false;
149 TLInfo = &getAnalysis<TargetLibraryInfo>();
150 DT = getAnalysisIfAvailable<DominatorTree>();
151 PFI = getAnalysisIfAvailable<ProfileInfo>();
152 OptSize = F.hasFnAttr(Attribute::OptimizeForSize);
154 /// This optimization identifies DIV instructions that can be
155 /// profitably bypassed and carried out with a shorter, faster divide.
156 if (TLI && TLI->isSlowDivBypassed()) {
157 const DenseMap<Type*, Type*> &BypassTypeMap = TLI->getBypassSlowDivTypes();
158 for (Function::iterator I = F.begin(); I != F.end(); I++)
159 EverMadeChange |= bypassSlowDivision(F, I, BypassTypeMap);
162 // Eliminate blocks that contain only PHI nodes and an
163 // unconditional branch.
164 EverMadeChange |= EliminateMostlyEmptyBlocks(F);
166 // llvm.dbg.value is far away from the value then iSel may not be able
167 // handle it properly. iSel will drop llvm.dbg.value if it can not
168 // find a node corresponding to the value.
169 EverMadeChange |= PlaceDbgValues(F);
171 bool MadeChange = true;
174 for (Function::iterator I = F.begin(); I != F.end(); ) {
175 BasicBlock *BB = I++;
176 MadeChange |= OptimizeBlock(*BB);
178 EverMadeChange |= MadeChange;
183 if (!DisableBranchOpts) {
185 SmallPtrSet<BasicBlock*, 8> WorkList;
186 for (Function::iterator BB = F.begin(), E = F.end(); BB != E; ++BB) {
187 SmallVector<BasicBlock*, 2> Successors(succ_begin(BB), succ_end(BB));
188 MadeChange |= ConstantFoldTerminator(BB, true);
189 if (!MadeChange) continue;
191 for (SmallVectorImpl<BasicBlock*>::iterator
192 II = Successors.begin(), IE = Successors.end(); II != IE; ++II)
193 if (pred_begin(*II) == pred_end(*II))
194 WorkList.insert(*II);
197 for (SmallPtrSet<BasicBlock*, 8>::iterator
198 I = WorkList.begin(), E = WorkList.end(); I != E; ++I)
201 // Merge pairs of basic blocks with unconditional branches, connected by
203 if (EverMadeChange || MadeChange)
204 MadeChange |= EliminateFallThrough(F);
208 EverMadeChange |= MadeChange;
211 if (ModifiedDT && DT)
212 DT->DT->recalculate(F);
214 return EverMadeChange;
217 /// EliminateFallThrough - Merge basic blocks which are connected
218 /// by a single edge, where one of the basic blocks has a single successor
219 /// pointing to the other basic block, which has a single predecessor.
220 bool CodeGenPrepare::EliminateFallThrough(Function &F) {
221 bool Changed = false;
222 // Scan all of the blocks in the function, except for the entry block.
223 for (Function::iterator I = ++F.begin(), E = F.end(); I != E; ) {
224 BasicBlock *BB = I++;
225 // If the destination block has a single pred, then this is a trivial
226 // edge, just collapse it.
227 BasicBlock *SinglePred = BB->getSinglePredecessor();
229 if (!SinglePred || SinglePred == BB) continue;
231 BranchInst *Term = dyn_cast<BranchInst>(SinglePred->getTerminator());
232 if (Term && !Term->isConditional()) {
234 DEBUG(dbgs() << "To merge:\n"<< *SinglePred << "\n\n\n");
235 // Remember if SinglePred was the entry block of the function.
236 // If so, we will need to move BB back to the entry position.
237 bool isEntry = SinglePred == &SinglePred->getParent()->getEntryBlock();
238 MergeBasicBlockIntoOnlyPred(BB, this);
240 if (isEntry && BB != &BB->getParent()->getEntryBlock())
241 BB->moveBefore(&BB->getParent()->getEntryBlock());
243 // We have erased a block. Update the iterator.
250 /// EliminateMostlyEmptyBlocks - eliminate blocks that contain only PHI nodes,
251 /// debug info directives, and an unconditional branch. Passes before isel
252 /// (e.g. LSR/loopsimplify) often split edges in ways that are non-optimal for
253 /// isel. Start by eliminating these blocks so we can split them the way we
255 bool CodeGenPrepare::EliminateMostlyEmptyBlocks(Function &F) {
256 bool MadeChange = false;
257 // Note that this intentionally skips the entry block.
258 for (Function::iterator I = ++F.begin(), E = F.end(); I != E; ) {
259 BasicBlock *BB = I++;
261 // If this block doesn't end with an uncond branch, ignore it.
262 BranchInst *BI = dyn_cast<BranchInst>(BB->getTerminator());
263 if (!BI || !BI->isUnconditional())
266 // If the instruction before the branch (skipping debug info) isn't a phi
267 // node, then other stuff is happening here.
268 BasicBlock::iterator BBI = BI;
269 if (BBI != BB->begin()) {
271 while (isa<DbgInfoIntrinsic>(BBI)) {
272 if (BBI == BB->begin())
276 if (!isa<DbgInfoIntrinsic>(BBI) && !isa<PHINode>(BBI))
280 // Do not break infinite loops.
281 BasicBlock *DestBB = BI->getSuccessor(0);
285 if (!CanMergeBlocks(BB, DestBB))
288 EliminateMostlyEmptyBlock(BB);
294 /// CanMergeBlocks - Return true if we can merge BB into DestBB if there is a
295 /// single uncond branch between them, and BB contains no other non-phi
297 bool CodeGenPrepare::CanMergeBlocks(const BasicBlock *BB,
298 const BasicBlock *DestBB) const {
299 // We only want to eliminate blocks whose phi nodes are used by phi nodes in
300 // the successor. If there are more complex condition (e.g. preheaders),
301 // don't mess around with them.
302 BasicBlock::const_iterator BBI = BB->begin();
303 while (const PHINode *PN = dyn_cast<PHINode>(BBI++)) {
304 for (Value::const_use_iterator UI = PN->use_begin(), E = PN->use_end();
306 const Instruction *User = cast<Instruction>(*UI);
307 if (User->getParent() != DestBB || !isa<PHINode>(User))
309 // If User is inside DestBB block and it is a PHINode then check
310 // incoming value. If incoming value is not from BB then this is
311 // a complex condition (e.g. preheaders) we want to avoid here.
312 if (User->getParent() == DestBB) {
313 if (const PHINode *UPN = dyn_cast<PHINode>(User))
314 for (unsigned I = 0, E = UPN->getNumIncomingValues(); I != E; ++I) {
315 Instruction *Insn = dyn_cast<Instruction>(UPN->getIncomingValue(I));
316 if (Insn && Insn->getParent() == BB &&
317 Insn->getParent() != UPN->getIncomingBlock(I))
324 // If BB and DestBB contain any common predecessors, then the phi nodes in BB
325 // and DestBB may have conflicting incoming values for the block. If so, we
326 // can't merge the block.
327 const PHINode *DestBBPN = dyn_cast<PHINode>(DestBB->begin());
328 if (!DestBBPN) return true; // no conflict.
330 // Collect the preds of BB.
331 SmallPtrSet<const BasicBlock*, 16> BBPreds;
332 if (const PHINode *BBPN = dyn_cast<PHINode>(BB->begin())) {
333 // It is faster to get preds from a PHI than with pred_iterator.
334 for (unsigned i = 0, e = BBPN->getNumIncomingValues(); i != e; ++i)
335 BBPreds.insert(BBPN->getIncomingBlock(i));
337 BBPreds.insert(pred_begin(BB), pred_end(BB));
340 // Walk the preds of DestBB.
341 for (unsigned i = 0, e = DestBBPN->getNumIncomingValues(); i != e; ++i) {
342 BasicBlock *Pred = DestBBPN->getIncomingBlock(i);
343 if (BBPreds.count(Pred)) { // Common predecessor?
344 BBI = DestBB->begin();
345 while (const PHINode *PN = dyn_cast<PHINode>(BBI++)) {
346 const Value *V1 = PN->getIncomingValueForBlock(Pred);
347 const Value *V2 = PN->getIncomingValueForBlock(BB);
349 // If V2 is a phi node in BB, look up what the mapped value will be.
350 if (const PHINode *V2PN = dyn_cast<PHINode>(V2))
351 if (V2PN->getParent() == BB)
352 V2 = V2PN->getIncomingValueForBlock(Pred);
354 // If there is a conflict, bail out.
355 if (V1 != V2) return false;
364 /// EliminateMostlyEmptyBlock - Eliminate a basic block that have only phi's and
365 /// an unconditional branch in it.
366 void CodeGenPrepare::EliminateMostlyEmptyBlock(BasicBlock *BB) {
367 BranchInst *BI = cast<BranchInst>(BB->getTerminator());
368 BasicBlock *DestBB = BI->getSuccessor(0);
370 DEBUG(dbgs() << "MERGING MOSTLY EMPTY BLOCKS - BEFORE:\n" << *BB << *DestBB);
372 // If the destination block has a single pred, then this is a trivial edge,
374 if (BasicBlock *SinglePred = DestBB->getSinglePredecessor()) {
375 if (SinglePred != DestBB) {
376 // Remember if SinglePred was the entry block of the function. If so, we
377 // will need to move BB back to the entry position.
378 bool isEntry = SinglePred == &SinglePred->getParent()->getEntryBlock();
379 MergeBasicBlockIntoOnlyPred(DestBB, this);
381 if (isEntry && BB != &BB->getParent()->getEntryBlock())
382 BB->moveBefore(&BB->getParent()->getEntryBlock());
384 DEBUG(dbgs() << "AFTER:\n" << *DestBB << "\n\n\n");
389 // Otherwise, we have multiple predecessors of BB. Update the PHIs in DestBB
390 // to handle the new incoming edges it is about to have.
392 for (BasicBlock::iterator BBI = DestBB->begin();
393 (PN = dyn_cast<PHINode>(BBI)); ++BBI) {
394 // Remove the incoming value for BB, and remember it.
395 Value *InVal = PN->removeIncomingValue(BB, false);
397 // Two options: either the InVal is a phi node defined in BB or it is some
398 // value that dominates BB.
399 PHINode *InValPhi = dyn_cast<PHINode>(InVal);
400 if (InValPhi && InValPhi->getParent() == BB) {
401 // Add all of the input values of the input PHI as inputs of this phi.
402 for (unsigned i = 0, e = InValPhi->getNumIncomingValues(); i != e; ++i)
403 PN->addIncoming(InValPhi->getIncomingValue(i),
404 InValPhi->getIncomingBlock(i));
406 // Otherwise, add one instance of the dominating value for each edge that
407 // we will be adding.
408 if (PHINode *BBPN = dyn_cast<PHINode>(BB->begin())) {
409 for (unsigned i = 0, e = BBPN->getNumIncomingValues(); i != e; ++i)
410 PN->addIncoming(InVal, BBPN->getIncomingBlock(i));
412 for (pred_iterator PI = pred_begin(BB), E = pred_end(BB); PI != E; ++PI)
413 PN->addIncoming(InVal, *PI);
418 // The PHIs are now updated, change everything that refers to BB to use
419 // DestBB and remove BB.
420 BB->replaceAllUsesWith(DestBB);
421 if (DT && !ModifiedDT) {
422 BasicBlock *BBIDom = DT->getNode(BB)->getIDom()->getBlock();
423 BasicBlock *DestBBIDom = DT->getNode(DestBB)->getIDom()->getBlock();
424 BasicBlock *NewIDom = DT->findNearestCommonDominator(BBIDom, DestBBIDom);
425 DT->changeImmediateDominator(DestBB, NewIDom);
429 PFI->replaceAllUses(BB, DestBB);
430 PFI->removeEdge(ProfileInfo::getEdge(BB, DestBB));
432 BB->eraseFromParent();
435 DEBUG(dbgs() << "AFTER:\n" << *DestBB << "\n\n\n");
438 /// OptimizeNoopCopyExpression - If the specified cast instruction is a noop
439 /// copy (e.g. it's casting from one pointer type to another, i32->i8 on PPC),
440 /// sink it into user blocks to reduce the number of virtual
441 /// registers that must be created and coalesced.
443 /// Return true if any changes are made.
445 static bool OptimizeNoopCopyExpression(CastInst *CI, const TargetLowering &TLI){
446 // If this is a noop copy,
447 EVT SrcVT = TLI.getValueType(CI->getOperand(0)->getType());
448 EVT DstVT = TLI.getValueType(CI->getType());
450 // This is an fp<->int conversion?
451 if (SrcVT.isInteger() != DstVT.isInteger())
454 // If this is an extension, it will be a zero or sign extension, which
456 if (SrcVT.bitsLT(DstVT)) return false;
458 // If these values will be promoted, find out what they will be promoted
459 // to. This helps us consider truncates on PPC as noop copies when they
461 if (TLI.getTypeAction(CI->getContext(), SrcVT) ==
462 TargetLowering::TypePromoteInteger)
463 SrcVT = TLI.getTypeToTransformTo(CI->getContext(), SrcVT);
464 if (TLI.getTypeAction(CI->getContext(), DstVT) ==
465 TargetLowering::TypePromoteInteger)
466 DstVT = TLI.getTypeToTransformTo(CI->getContext(), DstVT);
468 // If, after promotion, these are the same types, this is a noop copy.
472 BasicBlock *DefBB = CI->getParent();
474 /// InsertedCasts - Only insert a cast in each block once.
475 DenseMap<BasicBlock*, CastInst*> InsertedCasts;
477 bool MadeChange = false;
478 for (Value::use_iterator UI = CI->use_begin(), E = CI->use_end();
480 Use &TheUse = UI.getUse();
481 Instruction *User = cast<Instruction>(*UI);
483 // Figure out which BB this cast is used in. For PHI's this is the
484 // appropriate predecessor block.
485 BasicBlock *UserBB = User->getParent();
486 if (PHINode *PN = dyn_cast<PHINode>(User)) {
487 UserBB = PN->getIncomingBlock(UI);
490 // Preincrement use iterator so we don't invalidate it.
493 // If this user is in the same block as the cast, don't change the cast.
494 if (UserBB == DefBB) continue;
496 // If we have already inserted a cast into this block, use it.
497 CastInst *&InsertedCast = InsertedCasts[UserBB];
500 BasicBlock::iterator InsertPt = UserBB->getFirstInsertionPt();
502 CastInst::Create(CI->getOpcode(), CI->getOperand(0), CI->getType(), "",
507 // Replace a use of the cast with a use of the new cast.
508 TheUse = InsertedCast;
512 // If we removed all uses, nuke the cast.
513 if (CI->use_empty()) {
514 CI->eraseFromParent();
521 /// OptimizeCmpExpression - sink the given CmpInst into user blocks to reduce
522 /// the number of virtual registers that must be created and coalesced. This is
523 /// a clear win except on targets with multiple condition code registers
524 /// (PowerPC), where it might lose; some adjustment may be wanted there.
526 /// Return true if any changes are made.
527 static bool OptimizeCmpExpression(CmpInst *CI) {
528 BasicBlock *DefBB = CI->getParent();
530 /// InsertedCmp - Only insert a cmp in each block once.
531 DenseMap<BasicBlock*, CmpInst*> InsertedCmps;
533 bool MadeChange = false;
534 for (Value::use_iterator UI = CI->use_begin(), E = CI->use_end();
536 Use &TheUse = UI.getUse();
537 Instruction *User = cast<Instruction>(*UI);
539 // Preincrement use iterator so we don't invalidate it.
542 // Don't bother for PHI nodes.
543 if (isa<PHINode>(User))
546 // Figure out which BB this cmp is used in.
547 BasicBlock *UserBB = User->getParent();
549 // If this user is in the same block as the cmp, don't change the cmp.
550 if (UserBB == DefBB) continue;
552 // If we have already inserted a cmp into this block, use it.
553 CmpInst *&InsertedCmp = InsertedCmps[UserBB];
556 BasicBlock::iterator InsertPt = UserBB->getFirstInsertionPt();
558 CmpInst::Create(CI->getOpcode(),
559 CI->getPredicate(), CI->getOperand(0),
560 CI->getOperand(1), "", InsertPt);
564 // Replace a use of the cmp with a use of the new cmp.
565 TheUse = InsertedCmp;
569 // If we removed all uses, nuke the cmp.
571 CI->eraseFromParent();
577 class CodeGenPrepareFortifiedLibCalls : public SimplifyFortifiedLibCalls {
579 void replaceCall(Value *With) {
580 CI->replaceAllUsesWith(With);
581 CI->eraseFromParent();
583 bool isFoldable(unsigned SizeCIOp, unsigned, bool) const {
584 if (ConstantInt *SizeCI =
585 dyn_cast<ConstantInt>(CI->getArgOperand(SizeCIOp)))
586 return SizeCI->isAllOnesValue();
590 } // end anonymous namespace
592 bool CodeGenPrepare::OptimizeCallInst(CallInst *CI) {
593 BasicBlock *BB = CI->getParent();
595 // Lower inline assembly if we can.
596 // If we found an inline asm expession, and if the target knows how to
597 // lower it to normal LLVM code, do so now.
598 if (TLI && isa<InlineAsm>(CI->getCalledValue())) {
599 if (TLI->ExpandInlineAsm(CI)) {
600 // Avoid invalidating the iterator.
601 CurInstIterator = BB->begin();
602 // Avoid processing instructions out of order, which could cause
603 // reuse before a value is defined.
607 // Sink address computing for memory operands into the block.
608 if (OptimizeInlineAsmInst(CI))
612 // Lower all uses of llvm.objectsize.*
613 IntrinsicInst *II = dyn_cast<IntrinsicInst>(CI);
614 if (II && II->getIntrinsicID() == Intrinsic::objectsize) {
615 bool Min = (cast<ConstantInt>(II->getArgOperand(1))->getZExtValue() == 1);
616 Type *ReturnTy = CI->getType();
617 Constant *RetVal = ConstantInt::get(ReturnTy, Min ? 0 : -1ULL);
619 // Substituting this can cause recursive simplifications, which can
620 // invalidate our iterator. Use a WeakVH to hold onto it in case this
622 WeakVH IterHandle(CurInstIterator);
624 replaceAndRecursivelySimplify(CI, RetVal, TLI ? TLI->getTargetData() : 0,
625 TLInfo, ModifiedDT ? 0 : DT);
627 // If the iterator instruction was recursively deleted, start over at the
628 // start of the block.
629 if (IterHandle != CurInstIterator) {
630 CurInstIterator = BB->begin();
637 SmallVector<Value*, 2> PtrOps;
639 if (TLI->GetAddrModeArguments(II, PtrOps, AccessTy))
640 while (!PtrOps.empty())
641 if (OptimizeMemoryInst(II, PtrOps.pop_back_val(), AccessTy))
645 // From here on out we're working with named functions.
646 if (CI->getCalledFunction() == 0) return false;
648 // We'll need TargetData from here on out.
649 const TargetData *TD = TLI ? TLI->getTargetData() : 0;
650 if (!TD) return false;
652 // Lower all default uses of _chk calls. This is very similar
653 // to what InstCombineCalls does, but here we are only lowering calls
654 // that have the default "don't know" as the objectsize. Anything else
655 // should be left alone.
656 CodeGenPrepareFortifiedLibCalls Simplifier;
657 return Simplifier.fold(CI, TD, TLInfo);
660 /// DupRetToEnableTailCallOpts - Look for opportunities to duplicate return
661 /// instructions to the predecessor to enable tail call optimizations. The
662 /// case it is currently looking for is:
665 /// %tmp0 = tail call i32 @f0()
668 /// %tmp1 = tail call i32 @f1()
671 /// %tmp2 = tail call i32 @f2()
674 /// %retval = phi i32 [ %tmp0, %bb0 ], [ %tmp1, %bb1 ], [ %tmp2, %bb2 ]
682 /// %tmp0 = tail call i32 @f0()
685 /// %tmp1 = tail call i32 @f1()
688 /// %tmp2 = tail call i32 @f2()
691 bool CodeGenPrepare::DupRetToEnableTailCallOpts(ReturnInst *RI) {
696 BitCastInst *BCI = 0;
697 Value *V = RI->getReturnValue();
699 BCI = dyn_cast<BitCastInst>(V);
701 V = BCI->getOperand(0);
703 PN = dyn_cast<PHINode>(V);
708 BasicBlock *BB = RI->getParent();
709 if (PN && PN->getParent() != BB)
712 // It's not safe to eliminate the sign / zero extension of the return value.
713 // See llvm::isInTailCallPosition().
714 const Function *F = BB->getParent();
715 Attributes CallerRetAttr = F->getAttributes().getRetAttributes();
716 if ((CallerRetAttr & Attribute::ZExt) || (CallerRetAttr & Attribute::SExt))
719 // Make sure there are no instructions between the PHI and return, or that the
720 // return is the first instruction in the block.
722 BasicBlock::iterator BI = BB->begin();
723 do { ++BI; } while (isa<DbgInfoIntrinsic>(BI));
725 // Also skip over the bitcast.
730 BasicBlock::iterator BI = BB->begin();
731 while (isa<DbgInfoIntrinsic>(BI)) ++BI;
736 /// Only dup the ReturnInst if the CallInst is likely to be emitted as a tail
738 SmallVector<CallInst*, 4> TailCalls;
740 for (unsigned I = 0, E = PN->getNumIncomingValues(); I != E; ++I) {
741 CallInst *CI = dyn_cast<CallInst>(PN->getIncomingValue(I));
742 // Make sure the phi value is indeed produced by the tail call.
743 if (CI && CI->hasOneUse() && CI->getParent() == PN->getIncomingBlock(I) &&
744 TLI->mayBeEmittedAsTailCall(CI))
745 TailCalls.push_back(CI);
748 SmallPtrSet<BasicBlock*, 4> VisitedBBs;
749 for (pred_iterator PI = pred_begin(BB), PE = pred_end(BB); PI != PE; ++PI) {
750 if (!VisitedBBs.insert(*PI))
753 BasicBlock::InstListType &InstList = (*PI)->getInstList();
754 BasicBlock::InstListType::reverse_iterator RI = InstList.rbegin();
755 BasicBlock::InstListType::reverse_iterator RE = InstList.rend();
756 do { ++RI; } while (RI != RE && isa<DbgInfoIntrinsic>(&*RI));
760 CallInst *CI = dyn_cast<CallInst>(&*RI);
761 if (CI && CI->use_empty() && TLI->mayBeEmittedAsTailCall(CI))
762 TailCalls.push_back(CI);
766 bool Changed = false;
767 for (unsigned i = 0, e = TailCalls.size(); i != e; ++i) {
768 CallInst *CI = TailCalls[i];
771 // Conservatively require the attributes of the call to match those of the
772 // return. Ignore noalias because it doesn't affect the call sequence.
773 Attributes CalleeRetAttr = CS.getAttributes().getRetAttributes();
774 if ((CalleeRetAttr ^ CallerRetAttr) & ~Attribute::NoAlias)
777 // Make sure the call instruction is followed by an unconditional branch to
779 BasicBlock *CallBB = CI->getParent();
780 BranchInst *BI = dyn_cast<BranchInst>(CallBB->getTerminator());
781 if (!BI || !BI->isUnconditional() || BI->getSuccessor(0) != BB)
784 // Duplicate the return into CallBB.
785 (void)FoldReturnIntoUncondBranch(RI, BB, CallBB);
786 ModifiedDT = Changed = true;
790 // If we eliminated all predecessors of the block, delete the block now.
791 if (Changed && pred_begin(BB) == pred_end(BB))
792 BB->eraseFromParent();
797 //===----------------------------------------------------------------------===//
798 // Memory Optimization
799 //===----------------------------------------------------------------------===//
801 /// IsNonLocalValue - Return true if the specified values are defined in a
802 /// different basic block than BB.
803 static bool IsNonLocalValue(Value *V, BasicBlock *BB) {
804 if (Instruction *I = dyn_cast<Instruction>(V))
805 return I->getParent() != BB;
809 /// OptimizeMemoryInst - Load and Store Instructions often have
810 /// addressing modes that can do significant amounts of computation. As such,
811 /// instruction selection will try to get the load or store to do as much
812 /// computation as possible for the program. The problem is that isel can only
813 /// see within a single block. As such, we sink as much legal addressing mode
814 /// stuff into the block as possible.
816 /// This method is used to optimize both load/store and inline asms with memory
818 bool CodeGenPrepare::OptimizeMemoryInst(Instruction *MemoryInst, Value *Addr,
822 // Try to collapse single-value PHI nodes. This is necessary to undo
823 // unprofitable PRE transformations.
824 SmallVector<Value*, 8> worklist;
825 SmallPtrSet<Value*, 16> Visited;
826 worklist.push_back(Addr);
828 // Use a worklist to iteratively look through PHI nodes, and ensure that
829 // the addressing mode obtained from the non-PHI roots of the graph
831 Value *Consensus = 0;
832 unsigned NumUsesConsensus = 0;
833 bool IsNumUsesConsensusValid = false;
834 SmallVector<Instruction*, 16> AddrModeInsts;
835 ExtAddrMode AddrMode;
836 while (!worklist.empty()) {
837 Value *V = worklist.back();
840 // Break use-def graph loops.
841 if (!Visited.insert(V)) {
846 // For a PHI node, push all of its incoming values.
847 if (PHINode *P = dyn_cast<PHINode>(V)) {
848 for (unsigned i = 0, e = P->getNumIncomingValues(); i != e; ++i)
849 worklist.push_back(P->getIncomingValue(i));
853 // For non-PHIs, determine the addressing mode being computed.
854 SmallVector<Instruction*, 16> NewAddrModeInsts;
855 ExtAddrMode NewAddrMode =
856 AddressingModeMatcher::Match(V, AccessTy, MemoryInst,
857 NewAddrModeInsts, *TLI);
859 // This check is broken into two cases with very similar code to avoid using
860 // getNumUses() as much as possible. Some values have a lot of uses, so
861 // calling getNumUses() unconditionally caused a significant compile-time
865 AddrMode = NewAddrMode;
866 AddrModeInsts = NewAddrModeInsts;
868 } else if (NewAddrMode == AddrMode) {
869 if (!IsNumUsesConsensusValid) {
870 NumUsesConsensus = Consensus->getNumUses();
871 IsNumUsesConsensusValid = true;
874 // Ensure that the obtained addressing mode is equivalent to that obtained
875 // for all other roots of the PHI traversal. Also, when choosing one
876 // such root as representative, select the one with the most uses in order
877 // to keep the cost modeling heuristics in AddressingModeMatcher
879 unsigned NumUses = V->getNumUses();
880 if (NumUses > NumUsesConsensus) {
882 NumUsesConsensus = NumUses;
883 AddrModeInsts = NewAddrModeInsts;
892 // If the addressing mode couldn't be determined, or if multiple different
893 // ones were determined, bail out now.
894 if (!Consensus) return false;
896 // Check to see if any of the instructions supersumed by this addr mode are
897 // non-local to I's BB.
898 bool AnyNonLocal = false;
899 for (unsigned i = 0, e = AddrModeInsts.size(); i != e; ++i) {
900 if (IsNonLocalValue(AddrModeInsts[i], MemoryInst->getParent())) {
906 // If all the instructions matched are already in this BB, don't do anything.
908 DEBUG(dbgs() << "CGP: Found local addrmode: " << AddrMode << "\n");
912 // Insert this computation right after this user. Since our caller is
913 // scanning from the top of the BB to the bottom, reuse of the expr are
914 // guaranteed to happen later.
915 IRBuilder<> Builder(MemoryInst);
917 // Now that we determined the addressing expression we want to use and know
918 // that we have to sink it into this block. Check to see if we have already
919 // done this for some other load/store instr in this block. If so, reuse the
921 Value *&SunkAddr = SunkAddrs[Addr];
923 DEBUG(dbgs() << "CGP: Reusing nonlocal addrmode: " << AddrMode << " for "
925 if (SunkAddr->getType() != Addr->getType())
926 SunkAddr = Builder.CreateBitCast(SunkAddr, Addr->getType());
928 DEBUG(dbgs() << "CGP: SINKING nonlocal addrmode: " << AddrMode << " for "
931 TLI->getTargetData()->getIntPtrType(AccessTy->getContext());
935 // Start with the base register. Do this first so that subsequent address
936 // matching finds it last, which will prevent it from trying to match it
937 // as the scaled value in case it happens to be a mul. That would be
938 // problematic if we've sunk a different mul for the scale, because then
939 // we'd end up sinking both muls.
940 if (AddrMode.BaseReg) {
941 Value *V = AddrMode.BaseReg;
942 if (V->getType()->isPointerTy())
943 V = Builder.CreatePtrToInt(V, IntPtrTy, "sunkaddr");
944 if (V->getType() != IntPtrTy)
945 V = Builder.CreateIntCast(V, IntPtrTy, /*isSigned=*/true, "sunkaddr");
949 // Add the scale value.
950 if (AddrMode.Scale) {
951 Value *V = AddrMode.ScaledReg;
952 if (V->getType() == IntPtrTy) {
954 } else if (V->getType()->isPointerTy()) {
955 V = Builder.CreatePtrToInt(V, IntPtrTy, "sunkaddr");
956 } else if (cast<IntegerType>(IntPtrTy)->getBitWidth() <
957 cast<IntegerType>(V->getType())->getBitWidth()) {
958 V = Builder.CreateTrunc(V, IntPtrTy, "sunkaddr");
960 V = Builder.CreateSExt(V, IntPtrTy, "sunkaddr");
962 if (AddrMode.Scale != 1)
963 V = Builder.CreateMul(V, ConstantInt::get(IntPtrTy, AddrMode.Scale),
966 Result = Builder.CreateAdd(Result, V, "sunkaddr");
971 // Add in the BaseGV if present.
972 if (AddrMode.BaseGV) {
973 Value *V = Builder.CreatePtrToInt(AddrMode.BaseGV, IntPtrTy, "sunkaddr");
975 Result = Builder.CreateAdd(Result, V, "sunkaddr");
980 // Add in the Base Offset if present.
981 if (AddrMode.BaseOffs) {
982 Value *V = ConstantInt::get(IntPtrTy, AddrMode.BaseOffs);
984 Result = Builder.CreateAdd(Result, V, "sunkaddr");
990 SunkAddr = Constant::getNullValue(Addr->getType());
992 SunkAddr = Builder.CreateIntToPtr(Result, Addr->getType(), "sunkaddr");
995 MemoryInst->replaceUsesOfWith(Repl, SunkAddr);
997 // If we have no uses, recursively delete the value and all dead instructions
999 if (Repl->use_empty()) {
1000 // This can cause recursive deletion, which can invalidate our iterator.
1001 // Use a WeakVH to hold onto it in case this happens.
1002 WeakVH IterHandle(CurInstIterator);
1003 BasicBlock *BB = CurInstIterator->getParent();
1005 RecursivelyDeleteTriviallyDeadInstructions(Repl, TLInfo);
1007 if (IterHandle != CurInstIterator) {
1008 // If the iterator instruction was recursively deleted, start over at the
1009 // start of the block.
1010 CurInstIterator = BB->begin();
1013 // This address is now available for reassignment, so erase the table
1014 // entry; we don't want to match some completely different instruction.
1015 SunkAddrs[Addr] = 0;
1022 /// OptimizeInlineAsmInst - If there are any memory operands, use
1023 /// OptimizeMemoryInst to sink their address computing into the block when
1024 /// possible / profitable.
1025 bool CodeGenPrepare::OptimizeInlineAsmInst(CallInst *CS) {
1026 bool MadeChange = false;
1028 TargetLowering::AsmOperandInfoVector
1029 TargetConstraints = TLI->ParseConstraints(CS);
1031 for (unsigned i = 0, e = TargetConstraints.size(); i != e; ++i) {
1032 TargetLowering::AsmOperandInfo &OpInfo = TargetConstraints[i];
1034 // Compute the constraint code and ConstraintType to use.
1035 TLI->ComputeConstraintToUse(OpInfo, SDValue());
1037 if (OpInfo.ConstraintType == TargetLowering::C_Memory &&
1038 OpInfo.isIndirect) {
1039 Value *OpVal = CS->getArgOperand(ArgNo++);
1040 MadeChange |= OptimizeMemoryInst(CS, OpVal, OpVal->getType());
1041 } else if (OpInfo.Type == InlineAsm::isInput)
1048 /// MoveExtToFormExtLoad - Move a zext or sext fed by a load into the same
1049 /// basic block as the load, unless conditions are unfavorable. This allows
1050 /// SelectionDAG to fold the extend into the load.
1052 bool CodeGenPrepare::MoveExtToFormExtLoad(Instruction *I) {
1053 // Look for a load being extended.
1054 LoadInst *LI = dyn_cast<LoadInst>(I->getOperand(0));
1055 if (!LI) return false;
1057 // If they're already in the same block, there's nothing to do.
1058 if (LI->getParent() == I->getParent())
1061 // If the load has other users and the truncate is not free, this probably
1062 // isn't worthwhile.
1063 if (!LI->hasOneUse() &&
1064 TLI && (TLI->isTypeLegal(TLI->getValueType(LI->getType())) ||
1065 !TLI->isTypeLegal(TLI->getValueType(I->getType()))) &&
1066 !TLI->isTruncateFree(I->getType(), LI->getType()))
1069 // Check whether the target supports casts folded into loads.
1071 if (isa<ZExtInst>(I))
1072 LType = ISD::ZEXTLOAD;
1074 assert(isa<SExtInst>(I) && "Unexpected ext type!");
1075 LType = ISD::SEXTLOAD;
1077 if (TLI && !TLI->isLoadExtLegal(LType, TLI->getValueType(LI->getType())))
1080 // Move the extend into the same block as the load, so that SelectionDAG
1082 I->removeFromParent();
1088 bool CodeGenPrepare::OptimizeExtUses(Instruction *I) {
1089 BasicBlock *DefBB = I->getParent();
1091 // If the result of a {s|z}ext and its source are both live out, rewrite all
1092 // other uses of the source with result of extension.
1093 Value *Src = I->getOperand(0);
1094 if (Src->hasOneUse())
1097 // Only do this xform if truncating is free.
1098 if (TLI && !TLI->isTruncateFree(I->getType(), Src->getType()))
1101 // Only safe to perform the optimization if the source is also defined in
1103 if (!isa<Instruction>(Src) || DefBB != cast<Instruction>(Src)->getParent())
1106 bool DefIsLiveOut = false;
1107 for (Value::use_iterator UI = I->use_begin(), E = I->use_end();
1109 Instruction *User = cast<Instruction>(*UI);
1111 // Figure out which BB this ext is used in.
1112 BasicBlock *UserBB = User->getParent();
1113 if (UserBB == DefBB) continue;
1114 DefIsLiveOut = true;
1120 // Make sure non of the uses are PHI nodes.
1121 for (Value::use_iterator UI = Src->use_begin(), E = Src->use_end();
1123 Instruction *User = cast<Instruction>(*UI);
1124 BasicBlock *UserBB = User->getParent();
1125 if (UserBB == DefBB) continue;
1126 // Be conservative. We don't want this xform to end up introducing
1127 // reloads just before load / store instructions.
1128 if (isa<PHINode>(User) || isa<LoadInst>(User) || isa<StoreInst>(User))
1132 // InsertedTruncs - Only insert one trunc in each block once.
1133 DenseMap<BasicBlock*, Instruction*> InsertedTruncs;
1135 bool MadeChange = false;
1136 for (Value::use_iterator UI = Src->use_begin(), E = Src->use_end();
1138 Use &TheUse = UI.getUse();
1139 Instruction *User = cast<Instruction>(*UI);
1141 // Figure out which BB this ext is used in.
1142 BasicBlock *UserBB = User->getParent();
1143 if (UserBB == DefBB) continue;
1145 // Both src and def are live in this block. Rewrite the use.
1146 Instruction *&InsertedTrunc = InsertedTruncs[UserBB];
1148 if (!InsertedTrunc) {
1149 BasicBlock::iterator InsertPt = UserBB->getFirstInsertionPt();
1150 InsertedTrunc = new TruncInst(I, Src->getType(), "", InsertPt);
1153 // Replace a use of the {s|z}ext source with a use of the result.
1154 TheUse = InsertedTrunc;
1162 /// isFormingBranchFromSelectProfitable - Returns true if a SelectInst should be
1163 /// turned into an explicit branch.
1164 static bool isFormingBranchFromSelectProfitable(SelectInst *SI) {
1165 // FIXME: This should use the same heuristics as IfConversion to determine
1166 // whether a select is better represented as a branch. This requires that
1167 // branch probability metadata is preserved for the select, which is not the
1170 CmpInst *Cmp = dyn_cast<CmpInst>(SI->getCondition());
1172 // If the branch is predicted right, an out of order CPU can avoid blocking on
1173 // the compare. Emit cmovs on compares with a memory operand as branches to
1174 // avoid stalls on the load from memory. If the compare has more than one use
1175 // there's probably another cmov or setcc around so it's not worth emitting a
1180 Value *CmpOp0 = Cmp->getOperand(0);
1181 Value *CmpOp1 = Cmp->getOperand(1);
1183 // We check that the memory operand has one use to avoid uses of the loaded
1184 // value directly after the compare, making branches unprofitable.
1185 return Cmp->hasOneUse() &&
1186 ((isa<LoadInst>(CmpOp0) && CmpOp0->hasOneUse()) ||
1187 (isa<LoadInst>(CmpOp1) && CmpOp1->hasOneUse()));
1191 /// If we have a SelectInst that will likely profit from branch prediction,
1192 /// turn it into a branch.
1193 bool CodeGenPrepare::OptimizeSelectInst(SelectInst *SI) {
1194 bool VectorCond = !SI->getCondition()->getType()->isIntegerTy(1);
1196 // Can we convert the 'select' to CF ?
1197 if (DisableSelectToBranch || OptSize || !TLI || VectorCond)
1200 TargetLowering::SelectSupportKind SelectKind;
1202 SelectKind = TargetLowering::VectorMaskSelect;
1203 else if (SI->getType()->isVectorTy())
1204 SelectKind = TargetLowering::ScalarCondVectorVal;
1206 SelectKind = TargetLowering::ScalarValSelect;
1208 // Do we have efficient codegen support for this kind of 'selects' ?
1209 if (TLI->isSelectSupported(SelectKind)) {
1210 // We have efficient codegen support for the select instruction.
1211 // Check if it is profitable to keep this 'select'.
1212 if (!TLI->isPredictableSelectExpensive() ||
1213 !isFormingBranchFromSelectProfitable(SI))
1219 // First, we split the block containing the select into 2 blocks.
1220 BasicBlock *StartBlock = SI->getParent();
1221 BasicBlock::iterator SplitPt = ++(BasicBlock::iterator(SI));
1222 BasicBlock *NextBlock = StartBlock->splitBasicBlock(SplitPt, "select.end");
1224 // Create a new block serving as the landing pad for the branch.
1225 BasicBlock *SmallBlock = BasicBlock::Create(SI->getContext(), "select.mid",
1226 NextBlock->getParent(), NextBlock);
1228 // Move the unconditional branch from the block with the select in it into our
1229 // landing pad block.
1230 StartBlock->getTerminator()->eraseFromParent();
1231 BranchInst::Create(NextBlock, SmallBlock);
1233 // Insert the real conditional branch based on the original condition.
1234 BranchInst::Create(NextBlock, SmallBlock, SI->getCondition(), SI);
1236 // The select itself is replaced with a PHI Node.
1237 PHINode *PN = PHINode::Create(SI->getType(), 2, "", NextBlock->begin());
1239 PN->addIncoming(SI->getTrueValue(), StartBlock);
1240 PN->addIncoming(SI->getFalseValue(), SmallBlock);
1241 SI->replaceAllUsesWith(PN);
1242 SI->eraseFromParent();
1244 // Instruct OptimizeBlock to skip to the next block.
1245 CurInstIterator = StartBlock->end();
1246 ++NumSelectsExpanded;
1250 bool CodeGenPrepare::OptimizeInst(Instruction *I) {
1251 if (PHINode *P = dyn_cast<PHINode>(I)) {
1252 // It is possible for very late stage optimizations (such as SimplifyCFG)
1253 // to introduce PHI nodes too late to be cleaned up. If we detect such a
1254 // trivial PHI, go ahead and zap it here.
1255 if (Value *V = SimplifyInstruction(P)) {
1256 P->replaceAllUsesWith(V);
1257 P->eraseFromParent();
1264 if (CastInst *CI = dyn_cast<CastInst>(I)) {
1265 // If the source of the cast is a constant, then this should have
1266 // already been constant folded. The only reason NOT to constant fold
1267 // it is if something (e.g. LSR) was careful to place the constant
1268 // evaluation in a block other than then one that uses it (e.g. to hoist
1269 // the address of globals out of a loop). If this is the case, we don't
1270 // want to forward-subst the cast.
1271 if (isa<Constant>(CI->getOperand(0)))
1274 if (TLI && OptimizeNoopCopyExpression(CI, *TLI))
1277 if (isa<ZExtInst>(I) || isa<SExtInst>(I)) {
1278 bool MadeChange = MoveExtToFormExtLoad(I);
1279 return MadeChange | OptimizeExtUses(I);
1284 if (CmpInst *CI = dyn_cast<CmpInst>(I))
1285 return OptimizeCmpExpression(CI);
1287 if (LoadInst *LI = dyn_cast<LoadInst>(I)) {
1288 bool Changed = false;
1290 Changed |= OptimizeMemoryInst(I, I->getOperand(0), LI->getType());
1291 Changed |= ConvertLoadToSwitch(LI);
1295 if (StoreInst *SI = dyn_cast<StoreInst>(I)) {
1297 return OptimizeMemoryInst(I, SI->getOperand(1),
1298 SI->getOperand(0)->getType());
1302 if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(I)) {
1303 if (GEPI->hasAllZeroIndices()) {
1304 /// The GEP operand must be a pointer, so must its result -> BitCast
1305 Instruction *NC = new BitCastInst(GEPI->getOperand(0), GEPI->getType(),
1306 GEPI->getName(), GEPI);
1307 GEPI->replaceAllUsesWith(NC);
1308 GEPI->eraseFromParent();
1316 if (CallInst *CI = dyn_cast<CallInst>(I))
1317 return OptimizeCallInst(CI);
1319 if (ReturnInst *RI = dyn_cast<ReturnInst>(I))
1320 return DupRetToEnableTailCallOpts(RI);
1322 if (SelectInst *SI = dyn_cast<SelectInst>(I))
1323 return OptimizeSelectInst(SI);
1328 // In this pass we look for GEP and cast instructions that are used
1329 // across basic blocks and rewrite them to improve basic-block-at-a-time
1331 bool CodeGenPrepare::OptimizeBlock(BasicBlock &BB) {
1333 bool MadeChange = false;
1335 CurInstIterator = BB.begin();
1336 while (CurInstIterator != BB.end())
1337 MadeChange |= OptimizeInst(CurInstIterator++);
1342 // llvm.dbg.value is far away from the value then iSel may not be able
1343 // handle it properly. iSel will drop llvm.dbg.value if it can not
1344 // find a node corresponding to the value.
1345 bool CodeGenPrepare::PlaceDbgValues(Function &F) {
1346 bool MadeChange = false;
1347 for (Function::iterator I = F.begin(), E = F.end(); I != E; ++I) {
1348 Instruction *PrevNonDbgInst = NULL;
1349 for (BasicBlock::iterator BI = I->begin(), BE = I->end(); BI != BE;) {
1350 Instruction *Insn = BI; ++BI;
1351 DbgValueInst *DVI = dyn_cast<DbgValueInst>(Insn);
1353 PrevNonDbgInst = Insn;
1357 Instruction *VI = dyn_cast_or_null<Instruction>(DVI->getValue());
1358 if (VI && VI != PrevNonDbgInst && !VI->isTerminator()) {
1359 DEBUG(dbgs() << "Moving Debug Value before :\n" << *DVI << ' ' << *VI);
1360 DVI->removeFromParent();
1361 if (isa<PHINode>(VI))
1362 DVI->insertBefore(VI->getParent()->getFirstInsertionPt());
1364 DVI->insertAfter(VI);
1373 static bool TargetSupportsJumpTables(const TargetLowering &TLI) {
1374 return TLI.supportJumpTables() &&
1375 (TLI.isOperationLegalOrCustom(ISD::BR_JT, MVT::Other) ||
1376 TLI.isOperationLegalOrCustom(ISD::BRIND, MVT::Other));
1379 /// ConvertLoadToSwitch - Convert loads from constant lookup tables into
1380 /// switches. This undos the switch-to-lookup table transformation in
1381 /// SimplifyCFG for targets where that is inprofitable.
1382 bool CodeGenPrepare::ConvertLoadToSwitch(LoadInst *LI) {
1383 // This only applies to targets that don't support jump tables.
1384 if (!TLI || TargetSupportsJumpTables(*TLI))
1387 // FIXME: In the future, it would be desirable to have enough target
1388 // information in SimplifyCFG, so we could decide at that stage whether to
1389 // transform the switch to a lookup table or not, and this
1390 // reverse-transformation could be removed.
1392 GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(LI->getPointerOperand());
1393 if (!GEP || !GEP->isInBounds() || GEP->getPointerAddressSpace())
1395 if (GEP->getNumIndices() != 2)
1397 Value *FirstIndex = GEP->idx_begin()[0];
1398 ConstantInt *FirstIndexInt = dyn_cast<ConstantInt>(FirstIndex);
1399 if (!FirstIndexInt || !FirstIndexInt->isZero())
1402 Value *TableIndex = GEP->idx_begin()[1];
1403 IntegerType *TableIndexTy = cast<IntegerType>(TableIndex->getType());
1405 GlobalVariable *GV = dyn_cast<GlobalVariable>(GEP->getPointerOperand());
1406 if (!GV || !GV->isConstant() || !GV->hasDefinitiveInitializer())
1409 Constant *Arr = GV->getInitializer();
1410 uint64_t NumElements;
1411 if (ConstantArray *CA = dyn_cast<ConstantArray>(Arr))
1412 NumElements = CA->getType()->getNumElements();
1413 else if (ConstantDataArray *CDA = dyn_cast<ConstantDataArray>(Arr))
1414 NumElements = CDA->getNumElements();
1417 if (NumElements < 2)
1421 BasicBlock *OriginalBB = LI->getParent();
1422 BasicBlock *PostSwitchBB = OriginalBB->splitBasicBlock(LI);
1424 // Replace OriginalBB's terminator with a switch.
1425 IRBuilder<> Builder(OriginalBB->getTerminator());
1426 SwitchInst *Switch = Builder.CreateSwitch(TableIndex, PostSwitchBB,
1428 OriginalBB->getTerminator()->eraseFromParent();
1430 // Count the frequency of each value to decide which to use as default.
1431 SmallDenseMap<Constant*, uint64_t> ValueFreq;
1432 for (uint64_t I = 0; I < NumElements; ++I)
1433 ++ValueFreq[Arr->getAggregateElement(I)];
1434 uint64_t MaxCount = 0;
1435 Constant *DefaultValue = NULL;
1436 for (SmallDenseMap<Constant*, uint64_t>::iterator I = ValueFreq.begin(),
1437 E = ValueFreq.end(); I != E; ++I) {
1438 if (I->second > MaxCount) {
1439 MaxCount = I->second;
1440 DefaultValue = I->first;
1443 assert(DefaultValue && "No values in the array?");
1445 // Create the phi node in PostSwitchBB, which will replace the load.
1446 Builder.SetInsertPoint(PostSwitchBB->begin());
1447 PHINode *PHI = Builder.CreatePHI(LI->getType(), NumElements);
1448 PHI->addIncoming(DefaultValue, OriginalBB);
1450 // Build basic blocks to target with the switch.
1451 for (uint64_t I = 0; I < NumElements; ++I) {
1452 Constant *C = Arr->getAggregateElement(I);
1453 if (C == DefaultValue) continue; // Already covered by the default case.
1455 BasicBlock *BB = BasicBlock::Create(PostSwitchBB->getContext(),
1457 PostSwitchBB->getParent(),
1459 Switch->addCase(ConstantInt::get(TableIndexTy, I), BB);
1460 Builder.SetInsertPoint(BB);
1461 Builder.CreateBr(PostSwitchBB);
1462 PHI->addIncoming(C, BB);
1466 LI->replaceAllUsesWith(PHI);
1467 LI->eraseFromParent();
1470 if (GEP->use_empty())
1471 GEP->eraseFromParent();
1472 if (GV->hasUnnamedAddr() && GV->hasPrivateLinkage() && GV->use_empty())
1473 GV->eraseFromParent();
1475 CurInstIterator = Switch;