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.getFnAttributes().hasOptimizeForSizeAttr();
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 // Don't merge if BB's address is taken.
230 if (!SinglePred || SinglePred == BB || BB->hasAddressTaken()) continue;
232 BranchInst *Term = dyn_cast<BranchInst>(SinglePred->getTerminator());
233 if (Term && !Term->isConditional()) {
235 DEBUG(dbgs() << "To merge:\n"<< *SinglePred << "\n\n\n");
236 // Remember if SinglePred was the entry block of the function.
237 // If so, we will need to move BB back to the entry position.
238 bool isEntry = SinglePred == &SinglePred->getParent()->getEntryBlock();
239 MergeBasicBlockIntoOnlyPred(BB, this);
241 if (isEntry && BB != &BB->getParent()->getEntryBlock())
242 BB->moveBefore(&BB->getParent()->getEntryBlock());
244 // We have erased a block. Update the iterator.
251 /// EliminateMostlyEmptyBlocks - eliminate blocks that contain only PHI nodes,
252 /// debug info directives, and an unconditional branch. Passes before isel
253 /// (e.g. LSR/loopsimplify) often split edges in ways that are non-optimal for
254 /// isel. Start by eliminating these blocks so we can split them the way we
256 bool CodeGenPrepare::EliminateMostlyEmptyBlocks(Function &F) {
257 bool MadeChange = false;
258 // Note that this intentionally skips the entry block.
259 for (Function::iterator I = ++F.begin(), E = F.end(); I != E; ) {
260 BasicBlock *BB = I++;
262 // If this block doesn't end with an uncond branch, ignore it.
263 BranchInst *BI = dyn_cast<BranchInst>(BB->getTerminator());
264 if (!BI || !BI->isUnconditional())
267 // If the instruction before the branch (skipping debug info) isn't a phi
268 // node, then other stuff is happening here.
269 BasicBlock::iterator BBI = BI;
270 if (BBI != BB->begin()) {
272 while (isa<DbgInfoIntrinsic>(BBI)) {
273 if (BBI == BB->begin())
277 if (!isa<DbgInfoIntrinsic>(BBI) && !isa<PHINode>(BBI))
281 // Do not break infinite loops.
282 BasicBlock *DestBB = BI->getSuccessor(0);
286 if (!CanMergeBlocks(BB, DestBB))
289 EliminateMostlyEmptyBlock(BB);
295 /// CanMergeBlocks - Return true if we can merge BB into DestBB if there is a
296 /// single uncond branch between them, and BB contains no other non-phi
298 bool CodeGenPrepare::CanMergeBlocks(const BasicBlock *BB,
299 const BasicBlock *DestBB) const {
300 // We only want to eliminate blocks whose phi nodes are used by phi nodes in
301 // the successor. If there are more complex condition (e.g. preheaders),
302 // don't mess around with them.
303 BasicBlock::const_iterator BBI = BB->begin();
304 while (const PHINode *PN = dyn_cast<PHINode>(BBI++)) {
305 for (Value::const_use_iterator UI = PN->use_begin(), E = PN->use_end();
307 const Instruction *User = cast<Instruction>(*UI);
308 if (User->getParent() != DestBB || !isa<PHINode>(User))
310 // If User is inside DestBB block and it is a PHINode then check
311 // incoming value. If incoming value is not from BB then this is
312 // a complex condition (e.g. preheaders) we want to avoid here.
313 if (User->getParent() == DestBB) {
314 if (const PHINode *UPN = dyn_cast<PHINode>(User))
315 for (unsigned I = 0, E = UPN->getNumIncomingValues(); I != E; ++I) {
316 Instruction *Insn = dyn_cast<Instruction>(UPN->getIncomingValue(I));
317 if (Insn && Insn->getParent() == BB &&
318 Insn->getParent() != UPN->getIncomingBlock(I))
325 // If BB and DestBB contain any common predecessors, then the phi nodes in BB
326 // and DestBB may have conflicting incoming values for the block. If so, we
327 // can't merge the block.
328 const PHINode *DestBBPN = dyn_cast<PHINode>(DestBB->begin());
329 if (!DestBBPN) return true; // no conflict.
331 // Collect the preds of BB.
332 SmallPtrSet<const BasicBlock*, 16> BBPreds;
333 if (const PHINode *BBPN = dyn_cast<PHINode>(BB->begin())) {
334 // It is faster to get preds from a PHI than with pred_iterator.
335 for (unsigned i = 0, e = BBPN->getNumIncomingValues(); i != e; ++i)
336 BBPreds.insert(BBPN->getIncomingBlock(i));
338 BBPreds.insert(pred_begin(BB), pred_end(BB));
341 // Walk the preds of DestBB.
342 for (unsigned i = 0, e = DestBBPN->getNumIncomingValues(); i != e; ++i) {
343 BasicBlock *Pred = DestBBPN->getIncomingBlock(i);
344 if (BBPreds.count(Pred)) { // Common predecessor?
345 BBI = DestBB->begin();
346 while (const PHINode *PN = dyn_cast<PHINode>(BBI++)) {
347 const Value *V1 = PN->getIncomingValueForBlock(Pred);
348 const Value *V2 = PN->getIncomingValueForBlock(BB);
350 // If V2 is a phi node in BB, look up what the mapped value will be.
351 if (const PHINode *V2PN = dyn_cast<PHINode>(V2))
352 if (V2PN->getParent() == BB)
353 V2 = V2PN->getIncomingValueForBlock(Pred);
355 // If there is a conflict, bail out.
356 if (V1 != V2) return false;
365 /// EliminateMostlyEmptyBlock - Eliminate a basic block that have only phi's and
366 /// an unconditional branch in it.
367 void CodeGenPrepare::EliminateMostlyEmptyBlock(BasicBlock *BB) {
368 BranchInst *BI = cast<BranchInst>(BB->getTerminator());
369 BasicBlock *DestBB = BI->getSuccessor(0);
371 DEBUG(dbgs() << "MERGING MOSTLY EMPTY BLOCKS - BEFORE:\n" << *BB << *DestBB);
373 // If the destination block has a single pred, then this is a trivial edge,
375 if (BasicBlock *SinglePred = DestBB->getSinglePredecessor()) {
376 if (SinglePred != DestBB) {
377 // Remember if SinglePred was the entry block of the function. If so, we
378 // will need to move BB back to the entry position.
379 bool isEntry = SinglePred == &SinglePred->getParent()->getEntryBlock();
380 MergeBasicBlockIntoOnlyPred(DestBB, this);
382 if (isEntry && BB != &BB->getParent()->getEntryBlock())
383 BB->moveBefore(&BB->getParent()->getEntryBlock());
385 DEBUG(dbgs() << "AFTER:\n" << *DestBB << "\n\n\n");
390 // Otherwise, we have multiple predecessors of BB. Update the PHIs in DestBB
391 // to handle the new incoming edges it is about to have.
393 for (BasicBlock::iterator BBI = DestBB->begin();
394 (PN = dyn_cast<PHINode>(BBI)); ++BBI) {
395 // Remove the incoming value for BB, and remember it.
396 Value *InVal = PN->removeIncomingValue(BB, false);
398 // Two options: either the InVal is a phi node defined in BB or it is some
399 // value that dominates BB.
400 PHINode *InValPhi = dyn_cast<PHINode>(InVal);
401 if (InValPhi && InValPhi->getParent() == BB) {
402 // Add all of the input values of the input PHI as inputs of this phi.
403 for (unsigned i = 0, e = InValPhi->getNumIncomingValues(); i != e; ++i)
404 PN->addIncoming(InValPhi->getIncomingValue(i),
405 InValPhi->getIncomingBlock(i));
407 // Otherwise, add one instance of the dominating value for each edge that
408 // we will be adding.
409 if (PHINode *BBPN = dyn_cast<PHINode>(BB->begin())) {
410 for (unsigned i = 0, e = BBPN->getNumIncomingValues(); i != e; ++i)
411 PN->addIncoming(InVal, BBPN->getIncomingBlock(i));
413 for (pred_iterator PI = pred_begin(BB), E = pred_end(BB); PI != E; ++PI)
414 PN->addIncoming(InVal, *PI);
419 // The PHIs are now updated, change everything that refers to BB to use
420 // DestBB and remove BB.
421 BB->replaceAllUsesWith(DestBB);
422 if (DT && !ModifiedDT) {
423 BasicBlock *BBIDom = DT->getNode(BB)->getIDom()->getBlock();
424 BasicBlock *DestBBIDom = DT->getNode(DestBB)->getIDom()->getBlock();
425 BasicBlock *NewIDom = DT->findNearestCommonDominator(BBIDom, DestBBIDom);
426 DT->changeImmediateDominator(DestBB, NewIDom);
430 PFI->replaceAllUses(BB, DestBB);
431 PFI->removeEdge(ProfileInfo::getEdge(BB, DestBB));
433 BB->eraseFromParent();
436 DEBUG(dbgs() << "AFTER:\n" << *DestBB << "\n\n\n");
439 /// OptimizeNoopCopyExpression - If the specified cast instruction is a noop
440 /// copy (e.g. it's casting from one pointer type to another, i32->i8 on PPC),
441 /// sink it into user blocks to reduce the number of virtual
442 /// registers that must be created and coalesced.
444 /// Return true if any changes are made.
446 static bool OptimizeNoopCopyExpression(CastInst *CI, const TargetLowering &TLI){
447 // If this is a noop copy,
448 EVT SrcVT = TLI.getValueType(CI->getOperand(0)->getType());
449 EVT DstVT = TLI.getValueType(CI->getType());
451 // This is an fp<->int conversion?
452 if (SrcVT.isInteger() != DstVT.isInteger())
455 // If this is an extension, it will be a zero or sign extension, which
457 if (SrcVT.bitsLT(DstVT)) return false;
459 // If these values will be promoted, find out what they will be promoted
460 // to. This helps us consider truncates on PPC as noop copies when they
462 if (TLI.getTypeAction(CI->getContext(), SrcVT) ==
463 TargetLowering::TypePromoteInteger)
464 SrcVT = TLI.getTypeToTransformTo(CI->getContext(), SrcVT);
465 if (TLI.getTypeAction(CI->getContext(), DstVT) ==
466 TargetLowering::TypePromoteInteger)
467 DstVT = TLI.getTypeToTransformTo(CI->getContext(), DstVT);
469 // If, after promotion, these are the same types, this is a noop copy.
473 BasicBlock *DefBB = CI->getParent();
475 /// InsertedCasts - Only insert a cast in each block once.
476 DenseMap<BasicBlock*, CastInst*> InsertedCasts;
478 bool MadeChange = false;
479 for (Value::use_iterator UI = CI->use_begin(), E = CI->use_end();
481 Use &TheUse = UI.getUse();
482 Instruction *User = cast<Instruction>(*UI);
484 // Figure out which BB this cast is used in. For PHI's this is the
485 // appropriate predecessor block.
486 BasicBlock *UserBB = User->getParent();
487 if (PHINode *PN = dyn_cast<PHINode>(User)) {
488 UserBB = PN->getIncomingBlock(UI);
491 // Preincrement use iterator so we don't invalidate it.
494 // If this user is in the same block as the cast, don't change the cast.
495 if (UserBB == DefBB) continue;
497 // If we have already inserted a cast into this block, use it.
498 CastInst *&InsertedCast = InsertedCasts[UserBB];
501 BasicBlock::iterator InsertPt = UserBB->getFirstInsertionPt();
503 CastInst::Create(CI->getOpcode(), CI->getOperand(0), CI->getType(), "",
508 // Replace a use of the cast with a use of the new cast.
509 TheUse = InsertedCast;
513 // If we removed all uses, nuke the cast.
514 if (CI->use_empty()) {
515 CI->eraseFromParent();
522 /// OptimizeCmpExpression - sink the given CmpInst into user blocks to reduce
523 /// the number of virtual registers that must be created and coalesced. This is
524 /// a clear win except on targets with multiple condition code registers
525 /// (PowerPC), where it might lose; some adjustment may be wanted there.
527 /// Return true if any changes are made.
528 static bool OptimizeCmpExpression(CmpInst *CI) {
529 BasicBlock *DefBB = CI->getParent();
531 /// InsertedCmp - Only insert a cmp in each block once.
532 DenseMap<BasicBlock*, CmpInst*> InsertedCmps;
534 bool MadeChange = false;
535 for (Value::use_iterator UI = CI->use_begin(), E = CI->use_end();
537 Use &TheUse = UI.getUse();
538 Instruction *User = cast<Instruction>(*UI);
540 // Preincrement use iterator so we don't invalidate it.
543 // Don't bother for PHI nodes.
544 if (isa<PHINode>(User))
547 // Figure out which BB this cmp is used in.
548 BasicBlock *UserBB = User->getParent();
550 // If this user is in the same block as the cmp, don't change the cmp.
551 if (UserBB == DefBB) continue;
553 // If we have already inserted a cmp into this block, use it.
554 CmpInst *&InsertedCmp = InsertedCmps[UserBB];
557 BasicBlock::iterator InsertPt = UserBB->getFirstInsertionPt();
559 CmpInst::Create(CI->getOpcode(),
560 CI->getPredicate(), CI->getOperand(0),
561 CI->getOperand(1), "", InsertPt);
565 // Replace a use of the cmp with a use of the new cmp.
566 TheUse = InsertedCmp;
570 // If we removed all uses, nuke the cmp.
572 CI->eraseFromParent();
578 class CodeGenPrepareFortifiedLibCalls : public SimplifyFortifiedLibCalls {
580 void replaceCall(Value *With) {
581 CI->replaceAllUsesWith(With);
582 CI->eraseFromParent();
584 bool isFoldable(unsigned SizeCIOp, unsigned, bool) const {
585 if (ConstantInt *SizeCI =
586 dyn_cast<ConstantInt>(CI->getArgOperand(SizeCIOp)))
587 return SizeCI->isAllOnesValue();
591 } // end anonymous namespace
593 bool CodeGenPrepare::OptimizeCallInst(CallInst *CI) {
594 BasicBlock *BB = CI->getParent();
596 // Lower inline assembly if we can.
597 // If we found an inline asm expession, and if the target knows how to
598 // lower it to normal LLVM code, do so now.
599 if (TLI && isa<InlineAsm>(CI->getCalledValue())) {
600 if (TLI->ExpandInlineAsm(CI)) {
601 // Avoid invalidating the iterator.
602 CurInstIterator = BB->begin();
603 // Avoid processing instructions out of order, which could cause
604 // reuse before a value is defined.
608 // Sink address computing for memory operands into the block.
609 if (OptimizeInlineAsmInst(CI))
613 // Lower all uses of llvm.objectsize.*
614 IntrinsicInst *II = dyn_cast<IntrinsicInst>(CI);
615 if (II && II->getIntrinsicID() == Intrinsic::objectsize) {
616 bool Min = (cast<ConstantInt>(II->getArgOperand(1))->getZExtValue() == 1);
617 Type *ReturnTy = CI->getType();
618 Constant *RetVal = ConstantInt::get(ReturnTy, Min ? 0 : -1ULL);
620 // Substituting this can cause recursive simplifications, which can
621 // invalidate our iterator. Use a WeakVH to hold onto it in case this
623 WeakVH IterHandle(CurInstIterator);
625 replaceAndRecursivelySimplify(CI, RetVal, TLI ? TLI->getTargetData() : 0,
626 TLInfo, ModifiedDT ? 0 : DT);
628 // If the iterator instruction was recursively deleted, start over at the
629 // start of the block.
630 if (IterHandle != CurInstIterator) {
631 CurInstIterator = BB->begin();
638 SmallVector<Value*, 2> PtrOps;
640 if (TLI->GetAddrModeArguments(II, PtrOps, AccessTy))
641 while (!PtrOps.empty())
642 if (OptimizeMemoryInst(II, PtrOps.pop_back_val(), AccessTy))
646 // From here on out we're working with named functions.
647 if (CI->getCalledFunction() == 0) return false;
649 // We'll need TargetData from here on out.
650 const TargetData *TD = TLI ? TLI->getTargetData() : 0;
651 if (!TD) return false;
653 // Lower all default uses of _chk calls. This is very similar
654 // to what InstCombineCalls does, but here we are only lowering calls
655 // that have the default "don't know" as the objectsize. Anything else
656 // should be left alone.
657 CodeGenPrepareFortifiedLibCalls Simplifier;
658 return Simplifier.fold(CI, TD, TLInfo);
661 /// DupRetToEnableTailCallOpts - Look for opportunities to duplicate return
662 /// instructions to the predecessor to enable tail call optimizations. The
663 /// case it is currently looking for is:
666 /// %tmp0 = tail call i32 @f0()
669 /// %tmp1 = tail call i32 @f1()
672 /// %tmp2 = tail call i32 @f2()
675 /// %retval = phi i32 [ %tmp0, %bb0 ], [ %tmp1, %bb1 ], [ %tmp2, %bb2 ]
683 /// %tmp0 = tail call i32 @f0()
686 /// %tmp1 = tail call i32 @f1()
689 /// %tmp2 = tail call i32 @f2()
692 bool CodeGenPrepare::DupRetToEnableTailCallOpts(ReturnInst *RI) {
697 BitCastInst *BCI = 0;
698 Value *V = RI->getReturnValue();
700 BCI = dyn_cast<BitCastInst>(V);
702 V = BCI->getOperand(0);
704 PN = dyn_cast<PHINode>(V);
709 BasicBlock *BB = RI->getParent();
710 if (PN && PN->getParent() != BB)
713 // It's not safe to eliminate the sign / zero extension of the return value.
714 // See llvm::isInTailCallPosition().
715 const Function *F = BB->getParent();
716 Attributes CallerRetAttr = F->getAttributes().getRetAttributes();
717 if ((CallerRetAttr & Attribute::ZExt) || (CallerRetAttr & Attribute::SExt))
720 // Make sure there are no instructions between the PHI and return, or that the
721 // return is the first instruction in the block.
723 BasicBlock::iterator BI = BB->begin();
724 do { ++BI; } while (isa<DbgInfoIntrinsic>(BI));
726 // Also skip over the bitcast.
731 BasicBlock::iterator BI = BB->begin();
732 while (isa<DbgInfoIntrinsic>(BI)) ++BI;
737 /// Only dup the ReturnInst if the CallInst is likely to be emitted as a tail
739 SmallVector<CallInst*, 4> TailCalls;
741 for (unsigned I = 0, E = PN->getNumIncomingValues(); I != E; ++I) {
742 CallInst *CI = dyn_cast<CallInst>(PN->getIncomingValue(I));
743 // Make sure the phi value is indeed produced by the tail call.
744 if (CI && CI->hasOneUse() && CI->getParent() == PN->getIncomingBlock(I) &&
745 TLI->mayBeEmittedAsTailCall(CI))
746 TailCalls.push_back(CI);
749 SmallPtrSet<BasicBlock*, 4> VisitedBBs;
750 for (pred_iterator PI = pred_begin(BB), PE = pred_end(BB); PI != PE; ++PI) {
751 if (!VisitedBBs.insert(*PI))
754 BasicBlock::InstListType &InstList = (*PI)->getInstList();
755 BasicBlock::InstListType::reverse_iterator RI = InstList.rbegin();
756 BasicBlock::InstListType::reverse_iterator RE = InstList.rend();
757 do { ++RI; } while (RI != RE && isa<DbgInfoIntrinsic>(&*RI));
761 CallInst *CI = dyn_cast<CallInst>(&*RI);
762 if (CI && CI->use_empty() && TLI->mayBeEmittedAsTailCall(CI))
763 TailCalls.push_back(CI);
767 bool Changed = false;
768 for (unsigned i = 0, e = TailCalls.size(); i != e; ++i) {
769 CallInst *CI = TailCalls[i];
772 // Conservatively require the attributes of the call to match those of the
773 // return. Ignore noalias because it doesn't affect the call sequence.
774 Attributes CalleeRetAttr = CS.getAttributes().getRetAttributes();
775 if ((CalleeRetAttr ^ CallerRetAttr) & ~Attribute::NoAlias)
778 // Make sure the call instruction is followed by an unconditional branch to
780 BasicBlock *CallBB = CI->getParent();
781 BranchInst *BI = dyn_cast<BranchInst>(CallBB->getTerminator());
782 if (!BI || !BI->isUnconditional() || BI->getSuccessor(0) != BB)
785 // Duplicate the return into CallBB.
786 (void)FoldReturnIntoUncondBranch(RI, BB, CallBB);
787 ModifiedDT = Changed = true;
791 // If we eliminated all predecessors of the block, delete the block now.
792 if (Changed && !BB->hasAddressTaken() && pred_begin(BB) == pred_end(BB))
793 BB->eraseFromParent();
798 //===----------------------------------------------------------------------===//
799 // Memory Optimization
800 //===----------------------------------------------------------------------===//
802 /// IsNonLocalValue - Return true if the specified values are defined in a
803 /// different basic block than BB.
804 static bool IsNonLocalValue(Value *V, BasicBlock *BB) {
805 if (Instruction *I = dyn_cast<Instruction>(V))
806 return I->getParent() != BB;
810 /// OptimizeMemoryInst - Load and Store Instructions often have
811 /// addressing modes that can do significant amounts of computation. As such,
812 /// instruction selection will try to get the load or store to do as much
813 /// computation as possible for the program. The problem is that isel can only
814 /// see within a single block. As such, we sink as much legal addressing mode
815 /// stuff into the block as possible.
817 /// This method is used to optimize both load/store and inline asms with memory
819 bool CodeGenPrepare::OptimizeMemoryInst(Instruction *MemoryInst, Value *Addr,
823 // Try to collapse single-value PHI nodes. This is necessary to undo
824 // unprofitable PRE transformations.
825 SmallVector<Value*, 8> worklist;
826 SmallPtrSet<Value*, 16> Visited;
827 worklist.push_back(Addr);
829 // Use a worklist to iteratively look through PHI nodes, and ensure that
830 // the addressing mode obtained from the non-PHI roots of the graph
832 Value *Consensus = 0;
833 unsigned NumUsesConsensus = 0;
834 bool IsNumUsesConsensusValid = false;
835 SmallVector<Instruction*, 16> AddrModeInsts;
836 ExtAddrMode AddrMode;
837 while (!worklist.empty()) {
838 Value *V = worklist.back();
841 // Break use-def graph loops.
842 if (!Visited.insert(V)) {
847 // For a PHI node, push all of its incoming values.
848 if (PHINode *P = dyn_cast<PHINode>(V)) {
849 for (unsigned i = 0, e = P->getNumIncomingValues(); i != e; ++i)
850 worklist.push_back(P->getIncomingValue(i));
854 // For non-PHIs, determine the addressing mode being computed.
855 SmallVector<Instruction*, 16> NewAddrModeInsts;
856 ExtAddrMode NewAddrMode =
857 AddressingModeMatcher::Match(V, AccessTy, MemoryInst,
858 NewAddrModeInsts, *TLI);
860 // This check is broken into two cases with very similar code to avoid using
861 // getNumUses() as much as possible. Some values have a lot of uses, so
862 // calling getNumUses() unconditionally caused a significant compile-time
866 AddrMode = NewAddrMode;
867 AddrModeInsts = NewAddrModeInsts;
869 } else if (NewAddrMode == AddrMode) {
870 if (!IsNumUsesConsensusValid) {
871 NumUsesConsensus = Consensus->getNumUses();
872 IsNumUsesConsensusValid = true;
875 // Ensure that the obtained addressing mode is equivalent to that obtained
876 // for all other roots of the PHI traversal. Also, when choosing one
877 // such root as representative, select the one with the most uses in order
878 // to keep the cost modeling heuristics in AddressingModeMatcher
880 unsigned NumUses = V->getNumUses();
881 if (NumUses > NumUsesConsensus) {
883 NumUsesConsensus = NumUses;
884 AddrModeInsts = NewAddrModeInsts;
893 // If the addressing mode couldn't be determined, or if multiple different
894 // ones were determined, bail out now.
895 if (!Consensus) return false;
897 // Check to see if any of the instructions supersumed by this addr mode are
898 // non-local to I's BB.
899 bool AnyNonLocal = false;
900 for (unsigned i = 0, e = AddrModeInsts.size(); i != e; ++i) {
901 if (IsNonLocalValue(AddrModeInsts[i], MemoryInst->getParent())) {
907 // If all the instructions matched are already in this BB, don't do anything.
909 DEBUG(dbgs() << "CGP: Found local addrmode: " << AddrMode << "\n");
913 // Insert this computation right after this user. Since our caller is
914 // scanning from the top of the BB to the bottom, reuse of the expr are
915 // guaranteed to happen later.
916 IRBuilder<> Builder(MemoryInst);
918 // Now that we determined the addressing expression we want to use and know
919 // that we have to sink it into this block. Check to see if we have already
920 // done this for some other load/store instr in this block. If so, reuse the
922 Value *&SunkAddr = SunkAddrs[Addr];
924 DEBUG(dbgs() << "CGP: Reusing nonlocal addrmode: " << AddrMode << " for "
926 if (SunkAddr->getType() != Addr->getType())
927 SunkAddr = Builder.CreateBitCast(SunkAddr, Addr->getType());
929 DEBUG(dbgs() << "CGP: SINKING nonlocal addrmode: " << AddrMode << " for "
932 TLI->getTargetData()->getIntPtrType(AccessTy->getContext());
936 // Start with the base register. Do this first so that subsequent address
937 // matching finds it last, which will prevent it from trying to match it
938 // as the scaled value in case it happens to be a mul. That would be
939 // problematic if we've sunk a different mul for the scale, because then
940 // we'd end up sinking both muls.
941 if (AddrMode.BaseReg) {
942 Value *V = AddrMode.BaseReg;
943 if (V->getType()->isPointerTy())
944 V = Builder.CreatePtrToInt(V, IntPtrTy, "sunkaddr");
945 if (V->getType() != IntPtrTy)
946 V = Builder.CreateIntCast(V, IntPtrTy, /*isSigned=*/true, "sunkaddr");
950 // Add the scale value.
951 if (AddrMode.Scale) {
952 Value *V = AddrMode.ScaledReg;
953 if (V->getType() == IntPtrTy) {
955 } else if (V->getType()->isPointerTy()) {
956 V = Builder.CreatePtrToInt(V, IntPtrTy, "sunkaddr");
957 } else if (cast<IntegerType>(IntPtrTy)->getBitWidth() <
958 cast<IntegerType>(V->getType())->getBitWidth()) {
959 V = Builder.CreateTrunc(V, IntPtrTy, "sunkaddr");
961 V = Builder.CreateSExt(V, IntPtrTy, "sunkaddr");
963 if (AddrMode.Scale != 1)
964 V = Builder.CreateMul(V, ConstantInt::get(IntPtrTy, AddrMode.Scale),
967 Result = Builder.CreateAdd(Result, V, "sunkaddr");
972 // Add in the BaseGV if present.
973 if (AddrMode.BaseGV) {
974 Value *V = Builder.CreatePtrToInt(AddrMode.BaseGV, IntPtrTy, "sunkaddr");
976 Result = Builder.CreateAdd(Result, V, "sunkaddr");
981 // Add in the Base Offset if present.
982 if (AddrMode.BaseOffs) {
983 Value *V = ConstantInt::get(IntPtrTy, AddrMode.BaseOffs);
985 Result = Builder.CreateAdd(Result, V, "sunkaddr");
991 SunkAddr = Constant::getNullValue(Addr->getType());
993 SunkAddr = Builder.CreateIntToPtr(Result, Addr->getType(), "sunkaddr");
996 MemoryInst->replaceUsesOfWith(Repl, SunkAddr);
998 // If we have no uses, recursively delete the value and all dead instructions
1000 if (Repl->use_empty()) {
1001 // This can cause recursive deletion, which can invalidate our iterator.
1002 // Use a WeakVH to hold onto it in case this happens.
1003 WeakVH IterHandle(CurInstIterator);
1004 BasicBlock *BB = CurInstIterator->getParent();
1006 RecursivelyDeleteTriviallyDeadInstructions(Repl, TLInfo);
1008 if (IterHandle != CurInstIterator) {
1009 // If the iterator instruction was recursively deleted, start over at the
1010 // start of the block.
1011 CurInstIterator = BB->begin();
1014 // This address is now available for reassignment, so erase the table
1015 // entry; we don't want to match some completely different instruction.
1016 SunkAddrs[Addr] = 0;
1023 /// OptimizeInlineAsmInst - If there are any memory operands, use
1024 /// OptimizeMemoryInst to sink their address computing into the block when
1025 /// possible / profitable.
1026 bool CodeGenPrepare::OptimizeInlineAsmInst(CallInst *CS) {
1027 bool MadeChange = false;
1029 TargetLowering::AsmOperandInfoVector
1030 TargetConstraints = TLI->ParseConstraints(CS);
1032 for (unsigned i = 0, e = TargetConstraints.size(); i != e; ++i) {
1033 TargetLowering::AsmOperandInfo &OpInfo = TargetConstraints[i];
1035 // Compute the constraint code and ConstraintType to use.
1036 TLI->ComputeConstraintToUse(OpInfo, SDValue());
1038 if (OpInfo.ConstraintType == TargetLowering::C_Memory &&
1039 OpInfo.isIndirect) {
1040 Value *OpVal = CS->getArgOperand(ArgNo++);
1041 MadeChange |= OptimizeMemoryInst(CS, OpVal, OpVal->getType());
1042 } else if (OpInfo.Type == InlineAsm::isInput)
1049 /// MoveExtToFormExtLoad - Move a zext or sext fed by a load into the same
1050 /// basic block as the load, unless conditions are unfavorable. This allows
1051 /// SelectionDAG to fold the extend into the load.
1053 bool CodeGenPrepare::MoveExtToFormExtLoad(Instruction *I) {
1054 // Look for a load being extended.
1055 LoadInst *LI = dyn_cast<LoadInst>(I->getOperand(0));
1056 if (!LI) return false;
1058 // If they're already in the same block, there's nothing to do.
1059 if (LI->getParent() == I->getParent())
1062 // If the load has other users and the truncate is not free, this probably
1063 // isn't worthwhile.
1064 if (!LI->hasOneUse() &&
1065 TLI && (TLI->isTypeLegal(TLI->getValueType(LI->getType())) ||
1066 !TLI->isTypeLegal(TLI->getValueType(I->getType()))) &&
1067 !TLI->isTruncateFree(I->getType(), LI->getType()))
1070 // Check whether the target supports casts folded into loads.
1072 if (isa<ZExtInst>(I))
1073 LType = ISD::ZEXTLOAD;
1075 assert(isa<SExtInst>(I) && "Unexpected ext type!");
1076 LType = ISD::SEXTLOAD;
1078 if (TLI && !TLI->isLoadExtLegal(LType, TLI->getValueType(LI->getType())))
1081 // Move the extend into the same block as the load, so that SelectionDAG
1083 I->removeFromParent();
1089 bool CodeGenPrepare::OptimizeExtUses(Instruction *I) {
1090 BasicBlock *DefBB = I->getParent();
1092 // If the result of a {s|z}ext and its source are both live out, rewrite all
1093 // other uses of the source with result of extension.
1094 Value *Src = I->getOperand(0);
1095 if (Src->hasOneUse())
1098 // Only do this xform if truncating is free.
1099 if (TLI && !TLI->isTruncateFree(I->getType(), Src->getType()))
1102 // Only safe to perform the optimization if the source is also defined in
1104 if (!isa<Instruction>(Src) || DefBB != cast<Instruction>(Src)->getParent())
1107 bool DefIsLiveOut = false;
1108 for (Value::use_iterator UI = I->use_begin(), E = I->use_end();
1110 Instruction *User = cast<Instruction>(*UI);
1112 // Figure out which BB this ext is used in.
1113 BasicBlock *UserBB = User->getParent();
1114 if (UserBB == DefBB) continue;
1115 DefIsLiveOut = true;
1121 // Make sure non of the uses are PHI nodes.
1122 for (Value::use_iterator UI = Src->use_begin(), E = Src->use_end();
1124 Instruction *User = cast<Instruction>(*UI);
1125 BasicBlock *UserBB = User->getParent();
1126 if (UserBB == DefBB) continue;
1127 // Be conservative. We don't want this xform to end up introducing
1128 // reloads just before load / store instructions.
1129 if (isa<PHINode>(User) || isa<LoadInst>(User) || isa<StoreInst>(User))
1133 // InsertedTruncs - Only insert one trunc in each block once.
1134 DenseMap<BasicBlock*, Instruction*> InsertedTruncs;
1136 bool MadeChange = false;
1137 for (Value::use_iterator UI = Src->use_begin(), E = Src->use_end();
1139 Use &TheUse = UI.getUse();
1140 Instruction *User = cast<Instruction>(*UI);
1142 // Figure out which BB this ext is used in.
1143 BasicBlock *UserBB = User->getParent();
1144 if (UserBB == DefBB) continue;
1146 // Both src and def are live in this block. Rewrite the use.
1147 Instruction *&InsertedTrunc = InsertedTruncs[UserBB];
1149 if (!InsertedTrunc) {
1150 BasicBlock::iterator InsertPt = UserBB->getFirstInsertionPt();
1151 InsertedTrunc = new TruncInst(I, Src->getType(), "", InsertPt);
1154 // Replace a use of the {s|z}ext source with a use of the result.
1155 TheUse = InsertedTrunc;
1163 /// isFormingBranchFromSelectProfitable - Returns true if a SelectInst should be
1164 /// turned into an explicit branch.
1165 static bool isFormingBranchFromSelectProfitable(SelectInst *SI) {
1166 // FIXME: This should use the same heuristics as IfConversion to determine
1167 // whether a select is better represented as a branch. This requires that
1168 // branch probability metadata is preserved for the select, which is not the
1171 CmpInst *Cmp = dyn_cast<CmpInst>(SI->getCondition());
1173 // If the branch is predicted right, an out of order CPU can avoid blocking on
1174 // the compare. Emit cmovs on compares with a memory operand as branches to
1175 // avoid stalls on the load from memory. If the compare has more than one use
1176 // there's probably another cmov or setcc around so it's not worth emitting a
1181 Value *CmpOp0 = Cmp->getOperand(0);
1182 Value *CmpOp1 = Cmp->getOperand(1);
1184 // We check that the memory operand has one use to avoid uses of the loaded
1185 // value directly after the compare, making branches unprofitable.
1186 return Cmp->hasOneUse() &&
1187 ((isa<LoadInst>(CmpOp0) && CmpOp0->hasOneUse()) ||
1188 (isa<LoadInst>(CmpOp1) && CmpOp1->hasOneUse()));
1192 /// If we have a SelectInst that will likely profit from branch prediction,
1193 /// turn it into a branch.
1194 bool CodeGenPrepare::OptimizeSelectInst(SelectInst *SI) {
1195 bool VectorCond = !SI->getCondition()->getType()->isIntegerTy(1);
1197 // Can we convert the 'select' to CF ?
1198 if (DisableSelectToBranch || OptSize || !TLI || VectorCond)
1201 TargetLowering::SelectSupportKind SelectKind;
1203 SelectKind = TargetLowering::VectorMaskSelect;
1204 else if (SI->getType()->isVectorTy())
1205 SelectKind = TargetLowering::ScalarCondVectorVal;
1207 SelectKind = TargetLowering::ScalarValSelect;
1209 // Do we have efficient codegen support for this kind of 'selects' ?
1210 if (TLI->isSelectSupported(SelectKind)) {
1211 // We have efficient codegen support for the select instruction.
1212 // Check if it is profitable to keep this 'select'.
1213 if (!TLI->isPredictableSelectExpensive() ||
1214 !isFormingBranchFromSelectProfitable(SI))
1220 // First, we split the block containing the select into 2 blocks.
1221 BasicBlock *StartBlock = SI->getParent();
1222 BasicBlock::iterator SplitPt = ++(BasicBlock::iterator(SI));
1223 BasicBlock *NextBlock = StartBlock->splitBasicBlock(SplitPt, "select.end");
1225 // Create a new block serving as the landing pad for the branch.
1226 BasicBlock *SmallBlock = BasicBlock::Create(SI->getContext(), "select.mid",
1227 NextBlock->getParent(), NextBlock);
1229 // Move the unconditional branch from the block with the select in it into our
1230 // landing pad block.
1231 StartBlock->getTerminator()->eraseFromParent();
1232 BranchInst::Create(NextBlock, SmallBlock);
1234 // Insert the real conditional branch based on the original condition.
1235 BranchInst::Create(NextBlock, SmallBlock, SI->getCondition(), SI);
1237 // The select itself is replaced with a PHI Node.
1238 PHINode *PN = PHINode::Create(SI->getType(), 2, "", NextBlock->begin());
1240 PN->addIncoming(SI->getTrueValue(), StartBlock);
1241 PN->addIncoming(SI->getFalseValue(), SmallBlock);
1242 SI->replaceAllUsesWith(PN);
1243 SI->eraseFromParent();
1245 // Instruct OptimizeBlock to skip to the next block.
1246 CurInstIterator = StartBlock->end();
1247 ++NumSelectsExpanded;
1251 bool CodeGenPrepare::OptimizeInst(Instruction *I) {
1252 if (PHINode *P = dyn_cast<PHINode>(I)) {
1253 // It is possible for very late stage optimizations (such as SimplifyCFG)
1254 // to introduce PHI nodes too late to be cleaned up. If we detect such a
1255 // trivial PHI, go ahead and zap it here.
1256 if (Value *V = SimplifyInstruction(P)) {
1257 P->replaceAllUsesWith(V);
1258 P->eraseFromParent();
1265 if (CastInst *CI = dyn_cast<CastInst>(I)) {
1266 // If the source of the cast is a constant, then this should have
1267 // already been constant folded. The only reason NOT to constant fold
1268 // it is if something (e.g. LSR) was careful to place the constant
1269 // evaluation in a block other than then one that uses it (e.g. to hoist
1270 // the address of globals out of a loop). If this is the case, we don't
1271 // want to forward-subst the cast.
1272 if (isa<Constant>(CI->getOperand(0)))
1275 if (TLI && OptimizeNoopCopyExpression(CI, *TLI))
1278 if (isa<ZExtInst>(I) || isa<SExtInst>(I)) {
1279 bool MadeChange = MoveExtToFormExtLoad(I);
1280 return MadeChange | OptimizeExtUses(I);
1285 if (CmpInst *CI = dyn_cast<CmpInst>(I))
1286 return OptimizeCmpExpression(CI);
1288 if (LoadInst *LI = dyn_cast<LoadInst>(I)) {
1289 bool Changed = false;
1291 Changed |= OptimizeMemoryInst(I, I->getOperand(0), LI->getType());
1292 Changed |= ConvertLoadToSwitch(LI);
1296 if (StoreInst *SI = dyn_cast<StoreInst>(I)) {
1298 return OptimizeMemoryInst(I, SI->getOperand(1),
1299 SI->getOperand(0)->getType());
1303 if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(I)) {
1304 if (GEPI->hasAllZeroIndices()) {
1305 /// The GEP operand must be a pointer, so must its result -> BitCast
1306 Instruction *NC = new BitCastInst(GEPI->getOperand(0), GEPI->getType(),
1307 GEPI->getName(), GEPI);
1308 GEPI->replaceAllUsesWith(NC);
1309 GEPI->eraseFromParent();
1317 if (CallInst *CI = dyn_cast<CallInst>(I))
1318 return OptimizeCallInst(CI);
1320 if (ReturnInst *RI = dyn_cast<ReturnInst>(I))
1321 return DupRetToEnableTailCallOpts(RI);
1323 if (SelectInst *SI = dyn_cast<SelectInst>(I))
1324 return OptimizeSelectInst(SI);
1329 // In this pass we look for GEP and cast instructions that are used
1330 // across basic blocks and rewrite them to improve basic-block-at-a-time
1332 bool CodeGenPrepare::OptimizeBlock(BasicBlock &BB) {
1334 bool MadeChange = false;
1336 CurInstIterator = BB.begin();
1337 while (CurInstIterator != BB.end())
1338 MadeChange |= OptimizeInst(CurInstIterator++);
1343 // llvm.dbg.value is far away from the value then iSel may not be able
1344 // handle it properly. iSel will drop llvm.dbg.value if it can not
1345 // find a node corresponding to the value.
1346 bool CodeGenPrepare::PlaceDbgValues(Function &F) {
1347 bool MadeChange = false;
1348 for (Function::iterator I = F.begin(), E = F.end(); I != E; ++I) {
1349 Instruction *PrevNonDbgInst = NULL;
1350 for (BasicBlock::iterator BI = I->begin(), BE = I->end(); BI != BE;) {
1351 Instruction *Insn = BI; ++BI;
1352 DbgValueInst *DVI = dyn_cast<DbgValueInst>(Insn);
1354 PrevNonDbgInst = Insn;
1358 Instruction *VI = dyn_cast_or_null<Instruction>(DVI->getValue());
1359 if (VI && VI != PrevNonDbgInst && !VI->isTerminator()) {
1360 DEBUG(dbgs() << "Moving Debug Value before :\n" << *DVI << ' ' << *VI);
1361 DVI->removeFromParent();
1362 if (isa<PHINode>(VI))
1363 DVI->insertBefore(VI->getParent()->getFirstInsertionPt());
1365 DVI->insertAfter(VI);
1374 static bool TargetSupportsJumpTables(const TargetLowering &TLI) {
1375 return TLI.supportJumpTables() &&
1376 (TLI.isOperationLegalOrCustom(ISD::BR_JT, MVT::Other) ||
1377 TLI.isOperationLegalOrCustom(ISD::BRIND, MVT::Other));
1380 /// ConvertLoadToSwitch - Convert loads from constant lookup tables into
1381 /// switches. This undos the switch-to-lookup table transformation in
1382 /// SimplifyCFG for targets where that is inprofitable.
1383 bool CodeGenPrepare::ConvertLoadToSwitch(LoadInst *LI) {
1384 // This only applies to targets that don't support jump tables.
1385 if (!TLI || TargetSupportsJumpTables(*TLI))
1388 // FIXME: In the future, it would be desirable to have enough target
1389 // information in SimplifyCFG, so we could decide at that stage whether to
1390 // transform the switch to a lookup table or not, and this
1391 // reverse-transformation could be removed.
1393 GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(LI->getPointerOperand());
1394 if (!GEP || !GEP->isInBounds() || GEP->getPointerAddressSpace())
1396 if (GEP->getNumIndices() != 2)
1398 Value *FirstIndex = GEP->idx_begin()[0];
1399 ConstantInt *FirstIndexInt = dyn_cast<ConstantInt>(FirstIndex);
1400 if (!FirstIndexInt || !FirstIndexInt->isZero())
1403 Value *TableIndex = GEP->idx_begin()[1];
1404 IntegerType *TableIndexTy = cast<IntegerType>(TableIndex->getType());
1406 GlobalVariable *GV = dyn_cast<GlobalVariable>(GEP->getPointerOperand());
1407 if (!GV || !GV->isConstant() || !GV->hasDefinitiveInitializer())
1410 Constant *Arr = GV->getInitializer();
1411 uint64_t NumElements;
1412 if (ConstantArray *CA = dyn_cast<ConstantArray>(Arr))
1413 NumElements = CA->getType()->getNumElements();
1414 else if (ConstantDataArray *CDA = dyn_cast<ConstantDataArray>(Arr))
1415 NumElements = CDA->getNumElements();
1418 if (NumElements < 2)
1422 BasicBlock *OriginalBB = LI->getParent();
1423 BasicBlock *PostSwitchBB = OriginalBB->splitBasicBlock(LI);
1425 // Replace OriginalBB's terminator with a switch.
1426 IRBuilder<> Builder(OriginalBB->getTerminator());
1427 SwitchInst *Switch = Builder.CreateSwitch(TableIndex, PostSwitchBB,
1429 OriginalBB->getTerminator()->eraseFromParent();
1431 // Count the frequency of each value to decide which to use as default.
1432 SmallDenseMap<Constant*, uint64_t> ValueFreq;
1433 for (uint64_t I = 0; I < NumElements; ++I)
1434 ++ValueFreq[Arr->getAggregateElement(I)];
1435 uint64_t MaxCount = 0;
1436 Constant *DefaultValue = NULL;
1437 for (SmallDenseMap<Constant*, uint64_t>::iterator I = ValueFreq.begin(),
1438 E = ValueFreq.end(); I != E; ++I) {
1439 if (I->second > MaxCount) {
1440 MaxCount = I->second;
1441 DefaultValue = I->first;
1444 assert(DefaultValue && "No values in the array?");
1446 // Create the phi node in PostSwitchBB, which will replace the load.
1447 Builder.SetInsertPoint(PostSwitchBB->begin());
1448 PHINode *PHI = Builder.CreatePHI(LI->getType(), NumElements);
1449 PHI->addIncoming(DefaultValue, OriginalBB);
1451 // Build basic blocks to target with the switch.
1452 for (uint64_t I = 0; I < NumElements; ++I) {
1453 Constant *C = Arr->getAggregateElement(I);
1454 if (C == DefaultValue) continue; // Already covered by the default case.
1456 BasicBlock *BB = BasicBlock::Create(PostSwitchBB->getContext(),
1458 PostSwitchBB->getParent(),
1460 Switch->addCase(ConstantInt::get(TableIndexTy, I), BB);
1461 Builder.SetInsertPoint(BB);
1462 Builder.CreateBr(PostSwitchBB);
1463 PHI->addIncoming(C, BB);
1467 LI->replaceAllUsesWith(PHI);
1468 LI->eraseFromParent();
1471 if (GEP->use_empty())
1472 GEP->eraseFromParent();
1473 if (GV->hasUnnamedAddr() && GV->hasPrivateLinkage() && GV->use_empty())
1474 GV->eraseFromParent();
1476 CurInstIterator = Switch;