1 //===- CodeGenPrepare.cpp - Prepare a function for code generation --------===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // This pass munges the code in the input function to better prepare it for
11 // SelectionDAG-based code generation. This works around limitations in it's
12 // basic-block-at-a-time approach. It should eventually be removed.
14 //===----------------------------------------------------------------------===//
16 #define DEBUG_TYPE "codegenprepare"
17 #include "llvm/Transforms/Scalar.h"
18 #include "llvm/ADT/DenseMap.h"
19 #include "llvm/ADT/SmallSet.h"
20 #include "llvm/ADT/Statistic.h"
21 #include "llvm/ADT/ValueMap.h"
22 #include "llvm/Analysis/InstructionSimplify.h"
23 #include "llvm/IR/Constants.h"
24 #include "llvm/IR/DataLayout.h"
25 #include "llvm/IR/DerivedTypes.h"
26 #include "llvm/IR/Dominators.h"
27 #include "llvm/IR/Function.h"
28 #include "llvm/IR/IRBuilder.h"
29 #include "llvm/IR/InlineAsm.h"
30 #include "llvm/IR/Instructions.h"
31 #include "llvm/IR/IntrinsicInst.h"
32 #include "llvm/Pass.h"
33 #include "llvm/Support/CallSite.h"
34 #include "llvm/Support/CommandLine.h"
35 #include "llvm/Support/Debug.h"
36 #include "llvm/Support/GetElementPtrTypeIterator.h"
37 #include "llvm/Support/PatternMatch.h"
38 #include "llvm/Support/ValueHandle.h"
39 #include "llvm/Support/raw_ostream.h"
40 #include "llvm/Target/TargetLibraryInfo.h"
41 #include "llvm/Target/TargetLowering.h"
42 #include "llvm/Transforms/Utils/BasicBlockUtils.h"
43 #include "llvm/Transforms/Utils/BuildLibCalls.h"
44 #include "llvm/Transforms/Utils/BypassSlowDivision.h"
45 #include "llvm/Transforms/Utils/Local.h"
47 using namespace llvm::PatternMatch;
49 STATISTIC(NumBlocksElim, "Number of blocks eliminated");
50 STATISTIC(NumPHIsElim, "Number of trivial PHIs eliminated");
51 STATISTIC(NumGEPsElim, "Number of GEPs converted to casts");
52 STATISTIC(NumCmpUses, "Number of uses of Cmp expressions replaced with uses of "
54 STATISTIC(NumCastUses, "Number of uses of Cast expressions replaced with uses "
56 STATISTIC(NumMemoryInsts, "Number of memory instructions whose address "
57 "computations were sunk");
58 STATISTIC(NumExtsMoved, "Number of [s|z]ext instructions combined with loads");
59 STATISTIC(NumExtUses, "Number of uses of [s|z]ext instructions optimized");
60 STATISTIC(NumRetsDup, "Number of return instructions duplicated");
61 STATISTIC(NumDbgValueMoved, "Number of debug value instructions moved");
62 STATISTIC(NumSelectsExpanded, "Number of selects turned into branches");
64 static cl::opt<bool> DisableBranchOpts(
65 "disable-cgp-branch-opts", cl::Hidden, cl::init(false),
66 cl::desc("Disable branch optimizations in CodeGenPrepare"));
68 static cl::opt<bool> DisableSelectToBranch(
69 "disable-cgp-select2branch", cl::Hidden, cl::init(false),
70 cl::desc("Disable select to branch conversion."));
73 typedef SmallPtrSet<Instruction *, 16> SetOfInstrs;
74 typedef DenseMap<Instruction *, Type *> InstrToOrigTy;
76 class CodeGenPrepare : public FunctionPass {
77 /// TLI - Keep a pointer of a TargetLowering to consult for determining
78 /// transformation profitability.
79 const TargetMachine *TM;
80 const TargetLowering *TLI;
81 const TargetLibraryInfo *TLInfo;
84 /// CurInstIterator - As we scan instructions optimizing them, this is the
85 /// next instruction to optimize. Xforms that can invalidate this should
87 BasicBlock::iterator CurInstIterator;
89 /// Keeps track of non-local addresses that have been sunk into a block.
90 /// This allows us to avoid inserting duplicate code for blocks with
91 /// multiple load/stores of the same address.
92 ValueMap<Value*, Value*> SunkAddrs;
94 /// Keeps track of all truncates inserted for the current function.
95 SetOfInstrs InsertedTruncsSet;
96 /// Keeps track of the type of the related instruction before their
97 /// promotion for the current function.
98 InstrToOrigTy PromotedInsts;
100 /// ModifiedDT - If CFG is modified in anyway, dominator tree may need to
104 /// OptSize - True if optimizing for size.
108 static char ID; // Pass identification, replacement for typeid
109 explicit CodeGenPrepare(const TargetMachine *TM = 0)
110 : FunctionPass(ID), TM(TM), TLI(0) {
111 initializeCodeGenPreparePass(*PassRegistry::getPassRegistry());
113 bool runOnFunction(Function &F);
115 const char *getPassName() const { return "CodeGen Prepare"; }
117 virtual void getAnalysisUsage(AnalysisUsage &AU) const {
118 AU.addPreserved<DominatorTreeWrapperPass>();
119 AU.addRequired<TargetLibraryInfo>();
123 bool EliminateFallThrough(Function &F);
124 bool EliminateMostlyEmptyBlocks(Function &F);
125 bool CanMergeBlocks(const BasicBlock *BB, const BasicBlock *DestBB) const;
126 void EliminateMostlyEmptyBlock(BasicBlock *BB);
127 bool OptimizeBlock(BasicBlock &BB);
128 bool OptimizeInst(Instruction *I);
129 bool OptimizeMemoryInst(Instruction *I, Value *Addr, Type *AccessTy);
130 bool OptimizeInlineAsmInst(CallInst *CS);
131 bool OptimizeCallInst(CallInst *CI);
132 bool MoveExtToFormExtLoad(Instruction *I);
133 bool OptimizeExtUses(Instruction *I);
134 bool OptimizeSelectInst(SelectInst *SI);
135 bool DupRetToEnableTailCallOpts(BasicBlock *BB);
136 bool PlaceDbgValues(Function &F);
140 char CodeGenPrepare::ID = 0;
141 static void *initializeCodeGenPreparePassOnce(PassRegistry &Registry) {
142 initializeTargetLibraryInfoPass(Registry);
143 PassInfo *PI = new PassInfo(
144 "Optimize for code generation", "codegenprepare", &CodeGenPrepare::ID,
145 PassInfo::NormalCtor_t(callDefaultCtor<CodeGenPrepare>), false, false,
146 PassInfo::TargetMachineCtor_t(callTargetMachineCtor<CodeGenPrepare>));
147 Registry.registerPass(*PI, true);
151 void llvm::initializeCodeGenPreparePass(PassRegistry &Registry) {
152 CALL_ONCE_INITIALIZATION(initializeCodeGenPreparePassOnce)
155 FunctionPass *llvm::createCodeGenPreparePass(const TargetMachine *TM) {
156 return new CodeGenPrepare(TM);
159 bool CodeGenPrepare::runOnFunction(Function &F) {
160 bool EverMadeChange = false;
161 // Clear per function information.
162 InsertedTruncsSet.clear();
163 PromotedInsts.clear();
166 if (TM) TLI = TM->getTargetLowering();
167 TLInfo = &getAnalysis<TargetLibraryInfo>();
168 DominatorTreeWrapperPass *DTWP =
169 getAnalysisIfAvailable<DominatorTreeWrapperPass>();
170 DT = DTWP ? &DTWP->getDomTree() : 0;
171 OptSize = F.getAttributes().hasAttribute(AttributeSet::FunctionIndex,
172 Attribute::OptimizeForSize);
174 /// This optimization identifies DIV instructions that can be
175 /// profitably bypassed and carried out with a shorter, faster divide.
176 if (!OptSize && TLI && TLI->isSlowDivBypassed()) {
177 const DenseMap<unsigned int, unsigned int> &BypassWidths =
178 TLI->getBypassSlowDivWidths();
179 for (Function::iterator I = F.begin(); I != F.end(); I++)
180 EverMadeChange |= bypassSlowDivision(F, I, BypassWidths);
183 // Eliminate blocks that contain only PHI nodes and an
184 // unconditional branch.
185 EverMadeChange |= EliminateMostlyEmptyBlocks(F);
187 // llvm.dbg.value is far away from the value then iSel may not be able
188 // handle it properly. iSel will drop llvm.dbg.value if it can not
189 // find a node corresponding to the value.
190 EverMadeChange |= PlaceDbgValues(F);
192 bool MadeChange = true;
195 for (Function::iterator I = F.begin(); I != F.end(); ) {
196 BasicBlock *BB = I++;
197 MadeChange |= OptimizeBlock(*BB);
199 EverMadeChange |= MadeChange;
204 if (!DisableBranchOpts) {
206 SmallPtrSet<BasicBlock*, 8> WorkList;
207 for (Function::iterator BB = F.begin(), E = F.end(); BB != E; ++BB) {
208 SmallVector<BasicBlock*, 2> Successors(succ_begin(BB), succ_end(BB));
209 MadeChange |= ConstantFoldTerminator(BB, true);
210 if (!MadeChange) continue;
212 for (SmallVectorImpl<BasicBlock*>::iterator
213 II = Successors.begin(), IE = Successors.end(); II != IE; ++II)
214 if (pred_begin(*II) == pred_end(*II))
215 WorkList.insert(*II);
218 // Delete the dead blocks and any of their dead successors.
219 MadeChange |= !WorkList.empty();
220 while (!WorkList.empty()) {
221 BasicBlock *BB = *WorkList.begin();
223 SmallVector<BasicBlock*, 2> Successors(succ_begin(BB), succ_end(BB));
227 for (SmallVectorImpl<BasicBlock*>::iterator
228 II = Successors.begin(), IE = Successors.end(); II != IE; ++II)
229 if (pred_begin(*II) == pred_end(*II))
230 WorkList.insert(*II);
233 // Merge pairs of basic blocks with unconditional branches, connected by
235 if (EverMadeChange || MadeChange)
236 MadeChange |= EliminateFallThrough(F);
240 EverMadeChange |= MadeChange;
243 if (ModifiedDT && DT)
246 return EverMadeChange;
249 /// EliminateFallThrough - Merge basic blocks which are connected
250 /// by a single edge, where one of the basic blocks has a single successor
251 /// pointing to the other basic block, which has a single predecessor.
252 bool CodeGenPrepare::EliminateFallThrough(Function &F) {
253 bool Changed = false;
254 // Scan all of the blocks in the function, except for the entry block.
255 for (Function::iterator I = llvm::next(F.begin()), E = F.end(); I != E; ) {
256 BasicBlock *BB = I++;
257 // If the destination block has a single pred, then this is a trivial
258 // edge, just collapse it.
259 BasicBlock *SinglePred = BB->getSinglePredecessor();
261 // Don't merge if BB's address is taken.
262 if (!SinglePred || SinglePred == BB || BB->hasAddressTaken()) continue;
264 BranchInst *Term = dyn_cast<BranchInst>(SinglePred->getTerminator());
265 if (Term && !Term->isConditional()) {
267 DEBUG(dbgs() << "To merge:\n"<< *SinglePred << "\n\n\n");
268 // Remember if SinglePred was the entry block of the function.
269 // If so, we will need to move BB back to the entry position.
270 bool isEntry = SinglePred == &SinglePred->getParent()->getEntryBlock();
271 MergeBasicBlockIntoOnlyPred(BB, this);
273 if (isEntry && BB != &BB->getParent()->getEntryBlock())
274 BB->moveBefore(&BB->getParent()->getEntryBlock());
276 // We have erased a block. Update the iterator.
283 /// EliminateMostlyEmptyBlocks - eliminate blocks that contain only PHI nodes,
284 /// debug info directives, and an unconditional branch. Passes before isel
285 /// (e.g. LSR/loopsimplify) often split edges in ways that are non-optimal for
286 /// isel. Start by eliminating these blocks so we can split them the way we
288 bool CodeGenPrepare::EliminateMostlyEmptyBlocks(Function &F) {
289 bool MadeChange = false;
290 // Note that this intentionally skips the entry block.
291 for (Function::iterator I = llvm::next(F.begin()), E = F.end(); I != E; ) {
292 BasicBlock *BB = I++;
294 // If this block doesn't end with an uncond branch, ignore it.
295 BranchInst *BI = dyn_cast<BranchInst>(BB->getTerminator());
296 if (!BI || !BI->isUnconditional())
299 // If the instruction before the branch (skipping debug info) isn't a phi
300 // node, then other stuff is happening here.
301 BasicBlock::iterator BBI = BI;
302 if (BBI != BB->begin()) {
304 while (isa<DbgInfoIntrinsic>(BBI)) {
305 if (BBI == BB->begin())
309 if (!isa<DbgInfoIntrinsic>(BBI) && !isa<PHINode>(BBI))
313 // Do not break infinite loops.
314 BasicBlock *DestBB = BI->getSuccessor(0);
318 if (!CanMergeBlocks(BB, DestBB))
321 EliminateMostlyEmptyBlock(BB);
327 /// CanMergeBlocks - Return true if we can merge BB into DestBB if there is a
328 /// single uncond branch between them, and BB contains no other non-phi
330 bool CodeGenPrepare::CanMergeBlocks(const BasicBlock *BB,
331 const BasicBlock *DestBB) const {
332 // We only want to eliminate blocks whose phi nodes are used by phi nodes in
333 // the successor. If there are more complex condition (e.g. preheaders),
334 // don't mess around with them.
335 BasicBlock::const_iterator BBI = BB->begin();
336 while (const PHINode *PN = dyn_cast<PHINode>(BBI++)) {
337 for (Value::const_use_iterator UI = PN->use_begin(), E = PN->use_end();
339 const Instruction *User = cast<Instruction>(*UI);
340 if (User->getParent() != DestBB || !isa<PHINode>(User))
342 // If User is inside DestBB block and it is a PHINode then check
343 // incoming value. If incoming value is not from BB then this is
344 // a complex condition (e.g. preheaders) we want to avoid here.
345 if (User->getParent() == DestBB) {
346 if (const PHINode *UPN = dyn_cast<PHINode>(User))
347 for (unsigned I = 0, E = UPN->getNumIncomingValues(); I != E; ++I) {
348 Instruction *Insn = dyn_cast<Instruction>(UPN->getIncomingValue(I));
349 if (Insn && Insn->getParent() == BB &&
350 Insn->getParent() != UPN->getIncomingBlock(I))
357 // If BB and DestBB contain any common predecessors, then the phi nodes in BB
358 // and DestBB may have conflicting incoming values for the block. If so, we
359 // can't merge the block.
360 const PHINode *DestBBPN = dyn_cast<PHINode>(DestBB->begin());
361 if (!DestBBPN) return true; // no conflict.
363 // Collect the preds of BB.
364 SmallPtrSet<const BasicBlock*, 16> BBPreds;
365 if (const PHINode *BBPN = dyn_cast<PHINode>(BB->begin())) {
366 // It is faster to get preds from a PHI than with pred_iterator.
367 for (unsigned i = 0, e = BBPN->getNumIncomingValues(); i != e; ++i)
368 BBPreds.insert(BBPN->getIncomingBlock(i));
370 BBPreds.insert(pred_begin(BB), pred_end(BB));
373 // Walk the preds of DestBB.
374 for (unsigned i = 0, e = DestBBPN->getNumIncomingValues(); i != e; ++i) {
375 BasicBlock *Pred = DestBBPN->getIncomingBlock(i);
376 if (BBPreds.count(Pred)) { // Common predecessor?
377 BBI = DestBB->begin();
378 while (const PHINode *PN = dyn_cast<PHINode>(BBI++)) {
379 const Value *V1 = PN->getIncomingValueForBlock(Pred);
380 const Value *V2 = PN->getIncomingValueForBlock(BB);
382 // If V2 is a phi node in BB, look up what the mapped value will be.
383 if (const PHINode *V2PN = dyn_cast<PHINode>(V2))
384 if (V2PN->getParent() == BB)
385 V2 = V2PN->getIncomingValueForBlock(Pred);
387 // If there is a conflict, bail out.
388 if (V1 != V2) return false;
397 /// EliminateMostlyEmptyBlock - Eliminate a basic block that have only phi's and
398 /// an unconditional branch in it.
399 void CodeGenPrepare::EliminateMostlyEmptyBlock(BasicBlock *BB) {
400 BranchInst *BI = cast<BranchInst>(BB->getTerminator());
401 BasicBlock *DestBB = BI->getSuccessor(0);
403 DEBUG(dbgs() << "MERGING MOSTLY EMPTY BLOCKS - BEFORE:\n" << *BB << *DestBB);
405 // If the destination block has a single pred, then this is a trivial edge,
407 if (BasicBlock *SinglePred = DestBB->getSinglePredecessor()) {
408 if (SinglePred != DestBB) {
409 // Remember if SinglePred was the entry block of the function. If so, we
410 // will need to move BB back to the entry position.
411 bool isEntry = SinglePred == &SinglePred->getParent()->getEntryBlock();
412 MergeBasicBlockIntoOnlyPred(DestBB, this);
414 if (isEntry && BB != &BB->getParent()->getEntryBlock())
415 BB->moveBefore(&BB->getParent()->getEntryBlock());
417 DEBUG(dbgs() << "AFTER:\n" << *DestBB << "\n\n\n");
422 // Otherwise, we have multiple predecessors of BB. Update the PHIs in DestBB
423 // to handle the new incoming edges it is about to have.
425 for (BasicBlock::iterator BBI = DestBB->begin();
426 (PN = dyn_cast<PHINode>(BBI)); ++BBI) {
427 // Remove the incoming value for BB, and remember it.
428 Value *InVal = PN->removeIncomingValue(BB, false);
430 // Two options: either the InVal is a phi node defined in BB or it is some
431 // value that dominates BB.
432 PHINode *InValPhi = dyn_cast<PHINode>(InVal);
433 if (InValPhi && InValPhi->getParent() == BB) {
434 // Add all of the input values of the input PHI as inputs of this phi.
435 for (unsigned i = 0, e = InValPhi->getNumIncomingValues(); i != e; ++i)
436 PN->addIncoming(InValPhi->getIncomingValue(i),
437 InValPhi->getIncomingBlock(i));
439 // Otherwise, add one instance of the dominating value for each edge that
440 // we will be adding.
441 if (PHINode *BBPN = dyn_cast<PHINode>(BB->begin())) {
442 for (unsigned i = 0, e = BBPN->getNumIncomingValues(); i != e; ++i)
443 PN->addIncoming(InVal, BBPN->getIncomingBlock(i));
445 for (pred_iterator PI = pred_begin(BB), E = pred_end(BB); PI != E; ++PI)
446 PN->addIncoming(InVal, *PI);
451 // The PHIs are now updated, change everything that refers to BB to use
452 // DestBB and remove BB.
453 BB->replaceAllUsesWith(DestBB);
454 if (DT && !ModifiedDT) {
455 BasicBlock *BBIDom = DT->getNode(BB)->getIDom()->getBlock();
456 BasicBlock *DestBBIDom = DT->getNode(DestBB)->getIDom()->getBlock();
457 BasicBlock *NewIDom = DT->findNearestCommonDominator(BBIDom, DestBBIDom);
458 DT->changeImmediateDominator(DestBB, NewIDom);
461 BB->eraseFromParent();
464 DEBUG(dbgs() << "AFTER:\n" << *DestBB << "\n\n\n");
467 /// OptimizeNoopCopyExpression - If the specified cast instruction is a noop
468 /// copy (e.g. it's casting from one pointer type to another, i32->i8 on PPC),
469 /// sink it into user blocks to reduce the number of virtual
470 /// registers that must be created and coalesced.
472 /// Return true if any changes are made.
474 static bool OptimizeNoopCopyExpression(CastInst *CI, const TargetLowering &TLI){
475 // If this is a noop copy,
476 EVT SrcVT = TLI.getValueType(CI->getOperand(0)->getType());
477 EVT DstVT = TLI.getValueType(CI->getType());
479 // This is an fp<->int conversion?
480 if (SrcVT.isInteger() != DstVT.isInteger())
483 // If this is an extension, it will be a zero or sign extension, which
485 if (SrcVT.bitsLT(DstVT)) return false;
487 // If these values will be promoted, find out what they will be promoted
488 // to. This helps us consider truncates on PPC as noop copies when they
490 if (TLI.getTypeAction(CI->getContext(), SrcVT) ==
491 TargetLowering::TypePromoteInteger)
492 SrcVT = TLI.getTypeToTransformTo(CI->getContext(), SrcVT);
493 if (TLI.getTypeAction(CI->getContext(), DstVT) ==
494 TargetLowering::TypePromoteInteger)
495 DstVT = TLI.getTypeToTransformTo(CI->getContext(), DstVT);
497 // If, after promotion, these are the same types, this is a noop copy.
501 BasicBlock *DefBB = CI->getParent();
503 /// InsertedCasts - Only insert a cast in each block once.
504 DenseMap<BasicBlock*, CastInst*> InsertedCasts;
506 bool MadeChange = false;
507 for (Value::use_iterator UI = CI->use_begin(), E = CI->use_end();
509 Use &TheUse = UI.getUse();
510 Instruction *User = cast<Instruction>(*UI);
512 // Figure out which BB this cast is used in. For PHI's this is the
513 // appropriate predecessor block.
514 BasicBlock *UserBB = User->getParent();
515 if (PHINode *PN = dyn_cast<PHINode>(User)) {
516 UserBB = PN->getIncomingBlock(UI);
519 // Preincrement use iterator so we don't invalidate it.
522 // If this user is in the same block as the cast, don't change the cast.
523 if (UserBB == DefBB) continue;
525 // If we have already inserted a cast into this block, use it.
526 CastInst *&InsertedCast = InsertedCasts[UserBB];
529 BasicBlock::iterator InsertPt = UserBB->getFirstInsertionPt();
531 CastInst::Create(CI->getOpcode(), CI->getOperand(0), CI->getType(), "",
536 // Replace a use of the cast with a use of the new cast.
537 TheUse = InsertedCast;
541 // If we removed all uses, nuke the cast.
542 if (CI->use_empty()) {
543 CI->eraseFromParent();
550 /// OptimizeCmpExpression - sink the given CmpInst into user blocks to reduce
551 /// the number of virtual registers that must be created and coalesced. This is
552 /// a clear win except on targets with multiple condition code registers
553 /// (PowerPC), where it might lose; some adjustment may be wanted there.
555 /// Return true if any changes are made.
556 static bool OptimizeCmpExpression(CmpInst *CI) {
557 BasicBlock *DefBB = CI->getParent();
559 /// InsertedCmp - Only insert a cmp in each block once.
560 DenseMap<BasicBlock*, CmpInst*> InsertedCmps;
562 bool MadeChange = false;
563 for (Value::use_iterator UI = CI->use_begin(), E = CI->use_end();
565 Use &TheUse = UI.getUse();
566 Instruction *User = cast<Instruction>(*UI);
568 // Preincrement use iterator so we don't invalidate it.
571 // Don't bother for PHI nodes.
572 if (isa<PHINode>(User))
575 // Figure out which BB this cmp is used in.
576 BasicBlock *UserBB = User->getParent();
578 // If this user is in the same block as the cmp, don't change the cmp.
579 if (UserBB == DefBB) continue;
581 // If we have already inserted a cmp into this block, use it.
582 CmpInst *&InsertedCmp = InsertedCmps[UserBB];
585 BasicBlock::iterator InsertPt = UserBB->getFirstInsertionPt();
587 CmpInst::Create(CI->getOpcode(),
588 CI->getPredicate(), CI->getOperand(0),
589 CI->getOperand(1), "", InsertPt);
593 // Replace a use of the cmp with a use of the new cmp.
594 TheUse = InsertedCmp;
598 // If we removed all uses, nuke the cmp.
600 CI->eraseFromParent();
606 class CodeGenPrepareFortifiedLibCalls : public SimplifyFortifiedLibCalls {
608 void replaceCall(Value *With) {
609 CI->replaceAllUsesWith(With);
610 CI->eraseFromParent();
612 bool isFoldable(unsigned SizeCIOp, unsigned, bool) const {
613 if (ConstantInt *SizeCI =
614 dyn_cast<ConstantInt>(CI->getArgOperand(SizeCIOp)))
615 return SizeCI->isAllOnesValue();
619 } // end anonymous namespace
621 bool CodeGenPrepare::OptimizeCallInst(CallInst *CI) {
622 BasicBlock *BB = CI->getParent();
624 // Lower inline assembly if we can.
625 // If we found an inline asm expession, and if the target knows how to
626 // lower it to normal LLVM code, do so now.
627 if (TLI && isa<InlineAsm>(CI->getCalledValue())) {
628 if (TLI->ExpandInlineAsm(CI)) {
629 // Avoid invalidating the iterator.
630 CurInstIterator = BB->begin();
631 // Avoid processing instructions out of order, which could cause
632 // reuse before a value is defined.
636 // Sink address computing for memory operands into the block.
637 if (OptimizeInlineAsmInst(CI))
641 // Lower all uses of llvm.objectsize.*
642 IntrinsicInst *II = dyn_cast<IntrinsicInst>(CI);
643 if (II && II->getIntrinsicID() == Intrinsic::objectsize) {
644 bool Min = (cast<ConstantInt>(II->getArgOperand(1))->getZExtValue() == 1);
645 Type *ReturnTy = CI->getType();
646 Constant *RetVal = ConstantInt::get(ReturnTy, Min ? 0 : -1ULL);
648 // Substituting this can cause recursive simplifications, which can
649 // invalidate our iterator. Use a WeakVH to hold onto it in case this
651 WeakVH IterHandle(CurInstIterator);
653 replaceAndRecursivelySimplify(CI, RetVal, TLI ? TLI->getDataLayout() : 0,
654 TLInfo, ModifiedDT ? 0 : DT);
656 // If the iterator instruction was recursively deleted, start over at the
657 // start of the block.
658 if (IterHandle != CurInstIterator) {
659 CurInstIterator = BB->begin();
666 SmallVector<Value*, 2> PtrOps;
668 if (TLI->GetAddrModeArguments(II, PtrOps, AccessTy))
669 while (!PtrOps.empty())
670 if (OptimizeMemoryInst(II, PtrOps.pop_back_val(), AccessTy))
674 // From here on out we're working with named functions.
675 if (CI->getCalledFunction() == 0) return false;
677 // We'll need DataLayout from here on out.
678 const DataLayout *TD = TLI ? TLI->getDataLayout() : 0;
679 if (!TD) return false;
681 // Lower all default uses of _chk calls. This is very similar
682 // to what InstCombineCalls does, but here we are only lowering calls
683 // that have the default "don't know" as the objectsize. Anything else
684 // should be left alone.
685 CodeGenPrepareFortifiedLibCalls Simplifier;
686 return Simplifier.fold(CI, TD, TLInfo);
689 /// DupRetToEnableTailCallOpts - Look for opportunities to duplicate return
690 /// instructions to the predecessor to enable tail call optimizations. The
691 /// case it is currently looking for is:
694 /// %tmp0 = tail call i32 @f0()
697 /// %tmp1 = tail call i32 @f1()
700 /// %tmp2 = tail call i32 @f2()
703 /// %retval = phi i32 [ %tmp0, %bb0 ], [ %tmp1, %bb1 ], [ %tmp2, %bb2 ]
711 /// %tmp0 = tail call i32 @f0()
714 /// %tmp1 = tail call i32 @f1()
717 /// %tmp2 = tail call i32 @f2()
720 bool CodeGenPrepare::DupRetToEnableTailCallOpts(BasicBlock *BB) {
724 ReturnInst *RI = dyn_cast<ReturnInst>(BB->getTerminator());
729 BitCastInst *BCI = 0;
730 Value *V = RI->getReturnValue();
732 BCI = dyn_cast<BitCastInst>(V);
734 V = BCI->getOperand(0);
736 PN = dyn_cast<PHINode>(V);
741 if (PN && PN->getParent() != BB)
744 // It's not safe to eliminate the sign / zero extension of the return value.
745 // See llvm::isInTailCallPosition().
746 const Function *F = BB->getParent();
747 AttributeSet CallerAttrs = F->getAttributes();
748 if (CallerAttrs.hasAttribute(AttributeSet::ReturnIndex, Attribute::ZExt) ||
749 CallerAttrs.hasAttribute(AttributeSet::ReturnIndex, Attribute::SExt))
752 // Make sure there are no instructions between the PHI and return, or that the
753 // return is the first instruction in the block.
755 BasicBlock::iterator BI = BB->begin();
756 do { ++BI; } while (isa<DbgInfoIntrinsic>(BI));
758 // Also skip over the bitcast.
763 BasicBlock::iterator BI = BB->begin();
764 while (isa<DbgInfoIntrinsic>(BI)) ++BI;
769 /// Only dup the ReturnInst if the CallInst is likely to be emitted as a tail
771 SmallVector<CallInst*, 4> TailCalls;
773 for (unsigned I = 0, E = PN->getNumIncomingValues(); I != E; ++I) {
774 CallInst *CI = dyn_cast<CallInst>(PN->getIncomingValue(I));
775 // Make sure the phi value is indeed produced by the tail call.
776 if (CI && CI->hasOneUse() && CI->getParent() == PN->getIncomingBlock(I) &&
777 TLI->mayBeEmittedAsTailCall(CI))
778 TailCalls.push_back(CI);
781 SmallPtrSet<BasicBlock*, 4> VisitedBBs;
782 for (pred_iterator PI = pred_begin(BB), PE = pred_end(BB); PI != PE; ++PI) {
783 if (!VisitedBBs.insert(*PI))
786 BasicBlock::InstListType &InstList = (*PI)->getInstList();
787 BasicBlock::InstListType::reverse_iterator RI = InstList.rbegin();
788 BasicBlock::InstListType::reverse_iterator RE = InstList.rend();
789 do { ++RI; } while (RI != RE && isa<DbgInfoIntrinsic>(&*RI));
793 CallInst *CI = dyn_cast<CallInst>(&*RI);
794 if (CI && CI->use_empty() && TLI->mayBeEmittedAsTailCall(CI))
795 TailCalls.push_back(CI);
799 bool Changed = false;
800 for (unsigned i = 0, e = TailCalls.size(); i != e; ++i) {
801 CallInst *CI = TailCalls[i];
804 // Conservatively require the attributes of the call to match those of the
805 // return. Ignore noalias because it doesn't affect the call sequence.
806 AttributeSet CalleeAttrs = CS.getAttributes();
807 if (AttrBuilder(CalleeAttrs, AttributeSet::ReturnIndex).
808 removeAttribute(Attribute::NoAlias) !=
809 AttrBuilder(CalleeAttrs, AttributeSet::ReturnIndex).
810 removeAttribute(Attribute::NoAlias))
813 // Make sure the call instruction is followed by an unconditional branch to
815 BasicBlock *CallBB = CI->getParent();
816 BranchInst *BI = dyn_cast<BranchInst>(CallBB->getTerminator());
817 if (!BI || !BI->isUnconditional() || BI->getSuccessor(0) != BB)
820 // Duplicate the return into CallBB.
821 (void)FoldReturnIntoUncondBranch(RI, BB, CallBB);
822 ModifiedDT = Changed = true;
826 // If we eliminated all predecessors of the block, delete the block now.
827 if (Changed && !BB->hasAddressTaken() && pred_begin(BB) == pred_end(BB))
828 BB->eraseFromParent();
833 //===----------------------------------------------------------------------===//
834 // Memory Optimization
835 //===----------------------------------------------------------------------===//
839 /// ExtAddrMode - This is an extended version of TargetLowering::AddrMode
840 /// which holds actual Value*'s for register values.
841 struct ExtAddrMode : public TargetLowering::AddrMode {
844 ExtAddrMode() : BaseReg(0), ScaledReg(0) {}
845 void print(raw_ostream &OS) const;
848 bool operator==(const ExtAddrMode& O) const {
849 return (BaseReg == O.BaseReg) && (ScaledReg == O.ScaledReg) &&
850 (BaseGV == O.BaseGV) && (BaseOffs == O.BaseOffs) &&
851 (HasBaseReg == O.HasBaseReg) && (Scale == O.Scale);
856 static inline raw_ostream &operator<<(raw_ostream &OS, const ExtAddrMode &AM) {
862 void ExtAddrMode::print(raw_ostream &OS) const {
863 bool NeedPlus = false;
866 OS << (NeedPlus ? " + " : "")
868 BaseGV->printAsOperand(OS, /*PrintType=*/false);
873 OS << (NeedPlus ? " + " : "") << BaseOffs, NeedPlus = true;
876 OS << (NeedPlus ? " + " : "")
878 BaseReg->printAsOperand(OS, /*PrintType=*/false);
882 OS << (NeedPlus ? " + " : "")
884 ScaledReg->printAsOperand(OS, /*PrintType=*/false);
890 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
891 void ExtAddrMode::dump() const {
897 /// \brief This class provides transaction based operation on the IR.
898 /// Every change made through this class is recorded in the internal state and
899 /// can be undone (rollback) until commit is called.
900 class TypePromotionTransaction {
902 /// \brief This represents the common interface of the individual transaction.
903 /// Each class implements the logic for doing one specific modification on
904 /// the IR via the TypePromotionTransaction.
905 class TypePromotionAction {
907 /// The Instruction modified.
911 /// \brief Constructor of the action.
912 /// The constructor performs the related action on the IR.
913 TypePromotionAction(Instruction *Inst) : Inst(Inst) {}
915 virtual ~TypePromotionAction() {}
917 /// \brief Undo the modification done by this action.
918 /// When this method is called, the IR must be in the same state as it was
919 /// before this action was applied.
920 /// \pre Undoing the action works if and only if the IR is in the exact same
921 /// state as it was directly after this action was applied.
922 virtual void undo() = 0;
924 /// \brief Advocate every change made by this action.
925 /// When the results on the IR of the action are to be kept, it is important
926 /// to call this function, otherwise hidden information may be kept forever.
927 virtual void commit() {
928 // Nothing to be done, this action is not doing anything.
932 /// \brief Utility to remember the position of an instruction.
933 class InsertionHandler {
934 /// Position of an instruction.
935 /// Either an instruction:
936 /// - Is the first in a basic block: BB is used.
937 /// - Has a previous instructon: PrevInst is used.
939 Instruction *PrevInst;
942 /// Remember whether or not the instruction had a previous instruction.
943 bool HasPrevInstruction;
946 /// \brief Record the position of \p Inst.
947 InsertionHandler(Instruction *Inst) {
948 BasicBlock::iterator It = Inst;
949 HasPrevInstruction = (It != (Inst->getParent()->begin()));
950 if (HasPrevInstruction)
951 Point.PrevInst = --It;
953 Point.BB = Inst->getParent();
956 /// \brief Insert \p Inst at the recorded position.
957 void insert(Instruction *Inst) {
958 if (HasPrevInstruction) {
959 if (Inst->getParent())
960 Inst->removeFromParent();
961 Inst->insertAfter(Point.PrevInst);
963 Instruction *Position = Point.BB->getFirstInsertionPt();
964 if (Inst->getParent())
965 Inst->moveBefore(Position);
967 Inst->insertBefore(Position);
972 /// \brief Move an instruction before another.
973 class InstructionMoveBefore : public TypePromotionAction {
974 /// Original position of the instruction.
975 InsertionHandler Position;
978 /// \brief Move \p Inst before \p Before.
979 InstructionMoveBefore(Instruction *Inst, Instruction *Before)
980 : TypePromotionAction(Inst), Position(Inst) {
981 DEBUG(dbgs() << "Do: move: " << *Inst << "\nbefore: " << *Before << "\n");
982 Inst->moveBefore(Before);
985 /// \brief Move the instruction back to its original position.
987 DEBUG(dbgs() << "Undo: moveBefore: " << *Inst << "\n");
988 Position.insert(Inst);
992 /// \brief Set the operand of an instruction with a new value.
993 class OperandSetter : public TypePromotionAction {
994 /// Original operand of the instruction.
996 /// Index of the modified instruction.
1000 /// \brief Set \p Idx operand of \p Inst with \p NewVal.
1001 OperandSetter(Instruction *Inst, unsigned Idx, Value *NewVal)
1002 : TypePromotionAction(Inst), Idx(Idx) {
1003 DEBUG(dbgs() << "Do: setOperand: " << Idx << "\n"
1004 << "for:" << *Inst << "\n"
1005 << "with:" << *NewVal << "\n");
1006 Origin = Inst->getOperand(Idx);
1007 Inst->setOperand(Idx, NewVal);
1010 /// \brief Restore the original value of the instruction.
1012 DEBUG(dbgs() << "Undo: setOperand:" << Idx << "\n"
1013 << "for: " << *Inst << "\n"
1014 << "with: " << *Origin << "\n");
1015 Inst->setOperand(Idx, Origin);
1019 /// \brief Hide the operands of an instruction.
1020 /// Do as if this instruction was not using any of its operands.
1021 class OperandsHider : public TypePromotionAction {
1022 /// The list of original operands.
1023 SmallVector<Value *, 4> OriginalValues;
1026 /// \brief Remove \p Inst from the uses of the operands of \p Inst.
1027 OperandsHider(Instruction *Inst) : TypePromotionAction(Inst) {
1028 DEBUG(dbgs() << "Do: OperandsHider: " << *Inst << "\n");
1029 unsigned NumOpnds = Inst->getNumOperands();
1030 OriginalValues.reserve(NumOpnds);
1031 for (unsigned It = 0; It < NumOpnds; ++It) {
1032 // Save the current operand.
1033 Value *Val = Inst->getOperand(It);
1034 OriginalValues.push_back(Val);
1036 // We could use OperandSetter here, but that would implied an overhead
1037 // that we are not willing to pay.
1038 Inst->setOperand(It, UndefValue::get(Val->getType()));
1042 /// \brief Restore the original list of uses.
1044 DEBUG(dbgs() << "Undo: OperandsHider: " << *Inst << "\n");
1045 for (unsigned It = 0, EndIt = OriginalValues.size(); It != EndIt; ++It)
1046 Inst->setOperand(It, OriginalValues[It]);
1050 /// \brief Build a truncate instruction.
1051 class TruncBuilder : public TypePromotionAction {
1053 /// \brief Build a truncate instruction of \p Opnd producing a \p Ty
1055 /// trunc Opnd to Ty.
1056 TruncBuilder(Instruction *Opnd, Type *Ty) : TypePromotionAction(Opnd) {
1057 IRBuilder<> Builder(Opnd);
1058 Inst = cast<Instruction>(Builder.CreateTrunc(Opnd, Ty, "promoted"));
1059 DEBUG(dbgs() << "Do: TruncBuilder: " << *Inst << "\n");
1062 /// \brief Get the built instruction.
1063 Instruction *getBuiltInstruction() { return Inst; }
1065 /// \brief Remove the built instruction.
1067 DEBUG(dbgs() << "Undo: TruncBuilder: " << *Inst << "\n");
1068 Inst->eraseFromParent();
1072 /// \brief Build a sign extension instruction.
1073 class SExtBuilder : public TypePromotionAction {
1075 /// \brief Build a sign extension instruction of \p Opnd producing a \p Ty
1077 /// sext Opnd to Ty.
1078 SExtBuilder(Instruction *InsertPt, Value *Opnd, Type *Ty)
1079 : TypePromotionAction(Inst) {
1080 IRBuilder<> Builder(InsertPt);
1081 Inst = cast<Instruction>(Builder.CreateSExt(Opnd, Ty, "promoted"));
1082 DEBUG(dbgs() << "Do: SExtBuilder: " << *Inst << "\n");
1085 /// \brief Get the built instruction.
1086 Instruction *getBuiltInstruction() { return Inst; }
1088 /// \brief Remove the built instruction.
1090 DEBUG(dbgs() << "Undo: SExtBuilder: " << *Inst << "\n");
1091 Inst->eraseFromParent();
1095 /// \brief Mutate an instruction to another type.
1096 class TypeMutator : public TypePromotionAction {
1097 /// Record the original type.
1101 /// \brief Mutate the type of \p Inst into \p NewTy.
1102 TypeMutator(Instruction *Inst, Type *NewTy)
1103 : TypePromotionAction(Inst), OrigTy(Inst->getType()) {
1104 DEBUG(dbgs() << "Do: MutateType: " << *Inst << " with " << *NewTy
1106 Inst->mutateType(NewTy);
1109 /// \brief Mutate the instruction back to its original type.
1111 DEBUG(dbgs() << "Undo: MutateType: " << *Inst << " with " << *OrigTy
1113 Inst->mutateType(OrigTy);
1117 /// \brief Replace the uses of an instruction by another instruction.
1118 class UsesReplacer : public TypePromotionAction {
1119 /// Helper structure to keep track of the replaced uses.
1120 struct InstructionAndIdx {
1121 /// The instruction using the instruction.
1123 /// The index where this instruction is used for Inst.
1125 InstructionAndIdx(Instruction *Inst, unsigned Idx)
1126 : Inst(Inst), Idx(Idx) {}
1129 /// Keep track of the original uses (pair Instruction, Index).
1130 SmallVector<InstructionAndIdx, 4> OriginalUses;
1131 typedef SmallVectorImpl<InstructionAndIdx>::iterator use_iterator;
1134 /// \brief Replace all the use of \p Inst by \p New.
1135 UsesReplacer(Instruction *Inst, Value *New) : TypePromotionAction(Inst) {
1136 DEBUG(dbgs() << "Do: UsersReplacer: " << *Inst << " with " << *New
1138 // Record the original uses.
1139 for (Value::use_iterator UseIt = Inst->use_begin(),
1140 EndIt = Inst->use_end();
1141 UseIt != EndIt; ++UseIt) {
1142 Instruction *Use = cast<Instruction>(*UseIt);
1143 OriginalUses.push_back(InstructionAndIdx(Use, UseIt.getOperandNo()));
1145 // Now, we can replace the uses.
1146 Inst->replaceAllUsesWith(New);
1149 /// \brief Reassign the original uses of Inst to Inst.
1151 DEBUG(dbgs() << "Undo: UsersReplacer: " << *Inst << "\n");
1152 for (use_iterator UseIt = OriginalUses.begin(),
1153 EndIt = OriginalUses.end();
1154 UseIt != EndIt; ++UseIt) {
1155 UseIt->Inst->setOperand(UseIt->Idx, Inst);
1160 /// \brief Remove an instruction from the IR.
1161 class InstructionRemover : public TypePromotionAction {
1162 /// Original position of the instruction.
1163 InsertionHandler Inserter;
1164 /// Helper structure to hide all the link to the instruction. In other
1165 /// words, this helps to do as if the instruction was removed.
1166 OperandsHider Hider;
1167 /// Keep track of the uses replaced, if any.
1168 UsesReplacer *Replacer;
1171 /// \brief Remove all reference of \p Inst and optinally replace all its
1173 /// \pre If !Inst->use_empty(), then New != NULL
1174 InstructionRemover(Instruction *Inst, Value *New = NULL)
1175 : TypePromotionAction(Inst), Inserter(Inst), Hider(Inst),
1178 Replacer = new UsesReplacer(Inst, New);
1179 DEBUG(dbgs() << "Do: InstructionRemover: " << *Inst << "\n");
1180 Inst->removeFromParent();
1183 ~InstructionRemover() { delete Replacer; }
1185 /// \brief Really remove the instruction.
1186 void commit() { delete Inst; }
1188 /// \brief Resurrect the instruction and reassign it to the proper uses if
1189 /// new value was provided when build this action.
1191 DEBUG(dbgs() << "Undo: InstructionRemover: " << *Inst << "\n");
1192 Inserter.insert(Inst);
1200 /// Restoration point.
1201 /// The restoration point is a pointer to an action instead of an iterator
1202 /// because the iterator may be invalidated but not the pointer.
1203 typedef const TypePromotionAction *ConstRestorationPt;
1204 /// Advocate every changes made in that transaction.
1206 /// Undo all the changes made after the given point.
1207 void rollback(ConstRestorationPt Point);
1208 /// Get the current restoration point.
1209 ConstRestorationPt getRestorationPoint() const;
1211 /// \name API for IR modification with state keeping to support rollback.
1213 /// Same as Instruction::setOperand.
1214 void setOperand(Instruction *Inst, unsigned Idx, Value *NewVal);
1215 /// Same as Instruction::eraseFromParent.
1216 void eraseInstruction(Instruction *Inst, Value *NewVal = NULL);
1217 /// Same as Value::replaceAllUsesWith.
1218 void replaceAllUsesWith(Instruction *Inst, Value *New);
1219 /// Same as Value::mutateType.
1220 void mutateType(Instruction *Inst, Type *NewTy);
1221 /// Same as IRBuilder::createTrunc.
1222 Instruction *createTrunc(Instruction *Opnd, Type *Ty);
1223 /// Same as IRBuilder::createSExt.
1224 Instruction *createSExt(Instruction *Inst, Value *Opnd, Type *Ty);
1225 /// Same as Instruction::moveBefore.
1226 void moveBefore(Instruction *Inst, Instruction *Before);
1229 ~TypePromotionTransaction();
1232 /// The ordered list of actions made so far.
1233 SmallVector<TypePromotionAction *, 16> Actions;
1234 typedef SmallVectorImpl<TypePromotionAction *>::iterator CommitPt;
1237 void TypePromotionTransaction::setOperand(Instruction *Inst, unsigned Idx,
1240 new TypePromotionTransaction::OperandSetter(Inst, Idx, NewVal));
1243 void TypePromotionTransaction::eraseInstruction(Instruction *Inst,
1246 new TypePromotionTransaction::InstructionRemover(Inst, NewVal));
1249 void TypePromotionTransaction::replaceAllUsesWith(Instruction *Inst,
1251 Actions.push_back(new TypePromotionTransaction::UsesReplacer(Inst, New));
1254 void TypePromotionTransaction::mutateType(Instruction *Inst, Type *NewTy) {
1255 Actions.push_back(new TypePromotionTransaction::TypeMutator(Inst, NewTy));
1258 Instruction *TypePromotionTransaction::createTrunc(Instruction *Opnd,
1260 TruncBuilder *TB = new TruncBuilder(Opnd, Ty);
1261 Actions.push_back(TB);
1262 return TB->getBuiltInstruction();
1265 Instruction *TypePromotionTransaction::createSExt(Instruction *Inst,
1266 Value *Opnd, Type *Ty) {
1267 SExtBuilder *SB = new SExtBuilder(Inst, Opnd, Ty);
1268 Actions.push_back(SB);
1269 return SB->getBuiltInstruction();
1272 void TypePromotionTransaction::moveBefore(Instruction *Inst,
1273 Instruction *Before) {
1275 new TypePromotionTransaction::InstructionMoveBefore(Inst, Before));
1278 TypePromotionTransaction::ConstRestorationPt
1279 TypePromotionTransaction::getRestorationPoint() const {
1280 return Actions.rbegin() != Actions.rend() ? *Actions.rbegin() : NULL;
1283 void TypePromotionTransaction::commit() {
1284 for (CommitPt It = Actions.begin(), EndIt = Actions.end(); It != EndIt;
1292 void TypePromotionTransaction::rollback(
1293 TypePromotionTransaction::ConstRestorationPt Point) {
1294 while (!Actions.empty() && Point != (*Actions.rbegin())) {
1295 TypePromotionAction *Curr = Actions.pop_back_val();
1301 TypePromotionTransaction::~TypePromotionTransaction() {
1302 for (CommitPt It = Actions.begin(), EndIt = Actions.end(); It != EndIt; ++It)
1307 /// \brief A helper class for matching addressing modes.
1309 /// This encapsulates the logic for matching the target-legal addressing modes.
1310 class AddressingModeMatcher {
1311 SmallVectorImpl<Instruction*> &AddrModeInsts;
1312 const TargetLowering &TLI;
1314 /// AccessTy/MemoryInst - This is the type for the access (e.g. double) and
1315 /// the memory instruction that we're computing this address for.
1317 Instruction *MemoryInst;
1319 /// AddrMode - This is the addressing mode that we're building up. This is
1320 /// part of the return value of this addressing mode matching stuff.
1321 ExtAddrMode &AddrMode;
1323 /// The truncate instruction inserted by other CodeGenPrepare optimizations.
1324 const SetOfInstrs &InsertedTruncs;
1325 /// A map from the instructions to their type before promotion.
1326 InstrToOrigTy &PromotedInsts;
1327 /// The ongoing transaction where every action should be registered.
1328 TypePromotionTransaction &TPT;
1330 /// IgnoreProfitability - This is set to true when we should not do
1331 /// profitability checks. When true, IsProfitableToFoldIntoAddressingMode
1332 /// always returns true.
1333 bool IgnoreProfitability;
1335 AddressingModeMatcher(SmallVectorImpl<Instruction*> &AMI,
1336 const TargetLowering &T, Type *AT,
1337 Instruction *MI, ExtAddrMode &AM,
1338 const SetOfInstrs &InsertedTruncs,
1339 InstrToOrigTy &PromotedInsts,
1340 TypePromotionTransaction &TPT)
1341 : AddrModeInsts(AMI), TLI(T), AccessTy(AT), MemoryInst(MI), AddrMode(AM),
1342 InsertedTruncs(InsertedTruncs), PromotedInsts(PromotedInsts), TPT(TPT) {
1343 IgnoreProfitability = false;
1347 /// Match - Find the maximal addressing mode that a load/store of V can fold,
1348 /// give an access type of AccessTy. This returns a list of involved
1349 /// instructions in AddrModeInsts.
1350 /// \p InsertedTruncs The truncate instruction inserted by other
1353 /// \p PromotedInsts maps the instructions to their type before promotion.
1354 /// \p The ongoing transaction where every action should be registered.
1355 static ExtAddrMode Match(Value *V, Type *AccessTy,
1356 Instruction *MemoryInst,
1357 SmallVectorImpl<Instruction*> &AddrModeInsts,
1358 const TargetLowering &TLI,
1359 const SetOfInstrs &InsertedTruncs,
1360 InstrToOrigTy &PromotedInsts,
1361 TypePromotionTransaction &TPT) {
1364 bool Success = AddressingModeMatcher(AddrModeInsts, TLI, AccessTy,
1365 MemoryInst, Result, InsertedTruncs,
1366 PromotedInsts, TPT).MatchAddr(V, 0);
1367 (void)Success; assert(Success && "Couldn't select *anything*?");
1371 bool MatchScaledValue(Value *ScaleReg, int64_t Scale, unsigned Depth);
1372 bool MatchAddr(Value *V, unsigned Depth);
1373 bool MatchOperationAddr(User *Operation, unsigned Opcode, unsigned Depth,
1374 bool *MovedAway = NULL);
1375 bool IsProfitableToFoldIntoAddressingMode(Instruction *I,
1376 ExtAddrMode &AMBefore,
1377 ExtAddrMode &AMAfter);
1378 bool ValueAlreadyLiveAtInst(Value *Val, Value *KnownLive1, Value *KnownLive2);
1381 /// MatchScaledValue - Try adding ScaleReg*Scale to the current addressing mode.
1382 /// Return true and update AddrMode if this addr mode is legal for the target,
1384 bool AddressingModeMatcher::MatchScaledValue(Value *ScaleReg, int64_t Scale,
1386 // If Scale is 1, then this is the same as adding ScaleReg to the addressing
1387 // mode. Just process that directly.
1389 return MatchAddr(ScaleReg, Depth);
1391 // If the scale is 0, it takes nothing to add this.
1395 // If we already have a scale of this value, we can add to it, otherwise, we
1396 // need an available scale field.
1397 if (AddrMode.Scale != 0 && AddrMode.ScaledReg != ScaleReg)
1400 ExtAddrMode TestAddrMode = AddrMode;
1402 // Add scale to turn X*4+X*3 -> X*7. This could also do things like
1403 // [A+B + A*7] -> [B+A*8].
1404 TestAddrMode.Scale += Scale;
1405 TestAddrMode.ScaledReg = ScaleReg;
1407 // If the new address isn't legal, bail out.
1408 if (!TLI.isLegalAddressingMode(TestAddrMode, AccessTy))
1411 // It was legal, so commit it.
1412 AddrMode = TestAddrMode;
1414 // Okay, we decided that we can add ScaleReg+Scale to AddrMode. Check now
1415 // to see if ScaleReg is actually X+C. If so, we can turn this into adding
1416 // X*Scale + C*Scale to addr mode.
1417 ConstantInt *CI = 0; Value *AddLHS = 0;
1418 if (isa<Instruction>(ScaleReg) && // not a constant expr.
1419 match(ScaleReg, m_Add(m_Value(AddLHS), m_ConstantInt(CI)))) {
1420 TestAddrMode.ScaledReg = AddLHS;
1421 TestAddrMode.BaseOffs += CI->getSExtValue()*TestAddrMode.Scale;
1423 // If this addressing mode is legal, commit it and remember that we folded
1424 // this instruction.
1425 if (TLI.isLegalAddressingMode(TestAddrMode, AccessTy)) {
1426 AddrModeInsts.push_back(cast<Instruction>(ScaleReg));
1427 AddrMode = TestAddrMode;
1432 // Otherwise, not (x+c)*scale, just return what we have.
1436 /// MightBeFoldableInst - This is a little filter, which returns true if an
1437 /// addressing computation involving I might be folded into a load/store
1438 /// accessing it. This doesn't need to be perfect, but needs to accept at least
1439 /// the set of instructions that MatchOperationAddr can.
1440 static bool MightBeFoldableInst(Instruction *I) {
1441 switch (I->getOpcode()) {
1442 case Instruction::BitCast:
1443 // Don't touch identity bitcasts.
1444 if (I->getType() == I->getOperand(0)->getType())
1446 return I->getType()->isPointerTy() || I->getType()->isIntegerTy();
1447 case Instruction::PtrToInt:
1448 // PtrToInt is always a noop, as we know that the int type is pointer sized.
1450 case Instruction::IntToPtr:
1451 // We know the input is intptr_t, so this is foldable.
1453 case Instruction::Add:
1455 case Instruction::Mul:
1456 case Instruction::Shl:
1457 // Can only handle X*C and X << C.
1458 return isa<ConstantInt>(I->getOperand(1));
1459 case Instruction::GetElementPtr:
1466 /// \brief Hepler class to perform type promotion.
1467 class TypePromotionHelper {
1468 /// \brief Utility function to check whether or not a sign extension of
1469 /// \p Inst with \p ConsideredSExtType can be moved through \p Inst by either
1470 /// using the operands of \p Inst or promoting \p Inst.
1471 /// In other words, check if:
1472 /// sext (Ty Inst opnd1 opnd2 ... opndN) to ConsideredSExtType.
1473 /// #1 Promotion applies:
1474 /// ConsideredSExtType Inst (sext opnd1 to ConsideredSExtType, ...).
1475 /// #2 Operand reuses:
1476 /// sext opnd1 to ConsideredSExtType.
1477 /// \p PromotedInsts maps the instructions to their type before promotion.
1478 static bool canGetThrough(const Instruction *Inst, Type *ConsideredSExtType,
1479 const InstrToOrigTy &PromotedInsts);
1481 /// \brief Utility function to determine if \p OpIdx should be promoted when
1482 /// promoting \p Inst.
1483 static bool shouldSExtOperand(const Instruction *Inst, int OpIdx) {
1484 if (isa<SelectInst>(Inst) && OpIdx == 0)
1489 /// \brief Utility function to promote the operand of \p SExt when this
1490 /// operand is a promotable trunc or sext.
1491 /// \p PromotedInsts maps the instructions to their type before promotion.
1492 /// \p CreatedInsts[out] contains how many non-free instructions have been
1493 /// created to promote the operand of SExt.
1494 /// Should never be called directly.
1495 /// \return The promoted value which is used instead of SExt.
1496 static Value *promoteOperandForTruncAndSExt(Instruction *SExt,
1497 TypePromotionTransaction &TPT,
1498 InstrToOrigTy &PromotedInsts,
1499 unsigned &CreatedInsts);
1501 /// \brief Utility function to promote the operand of \p SExt when this
1502 /// operand is promotable and is not a supported trunc or sext.
1503 /// \p PromotedInsts maps the instructions to their type before promotion.
1504 /// \p CreatedInsts[out] contains how many non-free instructions have been
1505 /// created to promote the operand of SExt.
1506 /// Should never be called directly.
1507 /// \return The promoted value which is used instead of SExt.
1508 static Value *promoteOperandForOther(Instruction *SExt,
1509 TypePromotionTransaction &TPT,
1510 InstrToOrigTy &PromotedInsts,
1511 unsigned &CreatedInsts);
1514 /// Type for the utility function that promotes the operand of SExt.
1515 typedef Value *(*Action)(Instruction *SExt, TypePromotionTransaction &TPT,
1516 InstrToOrigTy &PromotedInsts,
1517 unsigned &CreatedInsts);
1518 /// \brief Given a sign extend instruction \p SExt, return the approriate
1519 /// action to promote the operand of \p SExt instead of using SExt.
1520 /// \return NULL if no promotable action is possible with the current
1522 /// \p InsertedTruncs keeps track of all the truncate instructions inserted by
1523 /// the others CodeGenPrepare optimizations. This information is important
1524 /// because we do not want to promote these instructions as CodeGenPrepare
1525 /// will reinsert them later. Thus creating an infinite loop: create/remove.
1526 /// \p PromotedInsts maps the instructions to their type before promotion.
1527 static Action getAction(Instruction *SExt, const SetOfInstrs &InsertedTruncs,
1528 const TargetLowering &TLI,
1529 const InstrToOrigTy &PromotedInsts);
1532 bool TypePromotionHelper::canGetThrough(const Instruction *Inst,
1533 Type *ConsideredSExtType,
1534 const InstrToOrigTy &PromotedInsts) {
1535 // We can always get through sext.
1536 if (isa<SExtInst>(Inst))
1539 // We can get through binary operator, if it is legal. In other words, the
1540 // binary operator must have a nuw or nsw flag.
1541 const BinaryOperator *BinOp = dyn_cast<BinaryOperator>(Inst);
1542 if (BinOp && isa<OverflowingBinaryOperator>(BinOp) &&
1543 (BinOp->hasNoUnsignedWrap() || BinOp->hasNoSignedWrap()))
1546 // Check if we can do the following simplification.
1547 // sext(trunc(sext)) --> sext
1548 if (!isa<TruncInst>(Inst))
1551 Value *OpndVal = Inst->getOperand(0);
1552 // Check if we can use this operand in the sext.
1553 // If the type is larger than the result type of the sign extension,
1555 if (OpndVal->getType()->getIntegerBitWidth() >
1556 ConsideredSExtType->getIntegerBitWidth())
1559 // If the operand of the truncate is not an instruction, we will not have
1560 // any information on the dropped bits.
1561 // (Actually we could for constant but it is not worth the extra logic).
1562 Instruction *Opnd = dyn_cast<Instruction>(OpndVal);
1566 // Check if the source of the type is narrow enough.
1567 // I.e., check that trunc just drops sign extended bits.
1568 // #1 get the type of the operand.
1569 const Type *OpndType;
1570 InstrToOrigTy::const_iterator It = PromotedInsts.find(Opnd);
1571 if (It != PromotedInsts.end())
1572 OpndType = It->second;
1573 else if (isa<SExtInst>(Opnd))
1574 OpndType = cast<Instruction>(Opnd)->getOperand(0)->getType();
1578 // #2 check that the truncate just drop sign extended bits.
1579 if (Inst->getType()->getIntegerBitWidth() >= OpndType->getIntegerBitWidth())
1585 TypePromotionHelper::Action TypePromotionHelper::getAction(
1586 Instruction *SExt, const SetOfInstrs &InsertedTruncs,
1587 const TargetLowering &TLI, const InstrToOrigTy &PromotedInsts) {
1588 Instruction *SExtOpnd = dyn_cast<Instruction>(SExt->getOperand(0));
1589 Type *SExtTy = SExt->getType();
1590 // If the operand of the sign extension is not an instruction, we cannot
1592 // If it, check we can get through.
1593 if (!SExtOpnd || !canGetThrough(SExtOpnd, SExtTy, PromotedInsts))
1596 // Do not promote if the operand has been added by codegenprepare.
1597 // Otherwise, it means we are undoing an optimization that is likely to be
1598 // redone, thus causing potential infinite loop.
1599 if (isa<TruncInst>(SExtOpnd) && InsertedTruncs.count(SExtOpnd))
1602 // SExt or Trunc instructions.
1603 // Return the related handler.
1604 if (isa<SExtInst>(SExtOpnd) || isa<TruncInst>(SExtOpnd))
1605 return promoteOperandForTruncAndSExt;
1607 // Regular instruction.
1608 // Abort early if we will have to insert non-free instructions.
1609 if (!SExtOpnd->hasOneUse() &&
1610 !TLI.isTruncateFree(SExtTy, SExtOpnd->getType()))
1612 return promoteOperandForOther;
1615 Value *TypePromotionHelper::promoteOperandForTruncAndSExt(
1616 llvm::Instruction *SExt, TypePromotionTransaction &TPT,
1617 InstrToOrigTy &PromotedInsts, unsigned &CreatedInsts) {
1618 // By construction, the operand of SExt is an instruction. Otherwise we cannot
1619 // get through it and this method should not be called.
1620 Instruction *SExtOpnd = cast<Instruction>(SExt->getOperand(0));
1621 // Replace sext(trunc(opnd)) or sext(sext(opnd))
1623 TPT.setOperand(SExt, 0, SExtOpnd->getOperand(0));
1626 // Remove dead code.
1627 if (SExtOpnd->use_empty())
1628 TPT.eraseInstruction(SExtOpnd);
1630 // Check if the sext is still needed.
1631 if (SExt->getType() != SExt->getOperand(0)->getType())
1634 // At this point we have: sext ty opnd to ty.
1635 // Reassign the uses of SExt to the opnd and remove SExt.
1636 Value *NextVal = SExt->getOperand(0);
1637 TPT.eraseInstruction(SExt, NextVal);
1642 TypePromotionHelper::promoteOperandForOther(Instruction *SExt,
1643 TypePromotionTransaction &TPT,
1644 InstrToOrigTy &PromotedInsts,
1645 unsigned &CreatedInsts) {
1646 // By construction, the operand of SExt is an instruction. Otherwise we cannot
1647 // get through it and this method should not be called.
1648 Instruction *SExtOpnd = cast<Instruction>(SExt->getOperand(0));
1650 if (!SExtOpnd->hasOneUse()) {
1651 // SExtOpnd will be promoted.
1652 // All its uses, but SExt, will need to use a truncated value of the
1653 // promoted version.
1654 // Create the truncate now.
1655 Instruction *Trunc = TPT.createTrunc(SExt, SExtOpnd->getType());
1656 Trunc->removeFromParent();
1657 // Insert it just after the definition.
1658 Trunc->insertAfter(SExtOpnd);
1660 TPT.replaceAllUsesWith(SExtOpnd, Trunc);
1661 // Restore the operand of SExt (which has been replace by the previous call
1662 // to replaceAllUsesWith) to avoid creating a cycle trunc <-> sext.
1663 TPT.setOperand(SExt, 0, SExtOpnd);
1666 // Get through the Instruction:
1667 // 1. Update its type.
1668 // 2. Replace the uses of SExt by Inst.
1669 // 3. Sign extend each operand that needs to be sign extended.
1671 // Remember the original type of the instruction before promotion.
1672 // This is useful to know that the high bits are sign extended bits.
1673 PromotedInsts.insert(
1674 std::pair<Instruction *, Type *>(SExtOpnd, SExtOpnd->getType()));
1676 TPT.mutateType(SExtOpnd, SExt->getType());
1678 TPT.replaceAllUsesWith(SExt, SExtOpnd);
1680 Instruction *SExtForOpnd = SExt;
1682 DEBUG(dbgs() << "Propagate SExt to operands\n");
1683 for (int OpIdx = 0, EndOpIdx = SExtOpnd->getNumOperands(); OpIdx != EndOpIdx;
1685 DEBUG(dbgs() << "Operand:\n" << *(SExtOpnd->getOperand(OpIdx)) << '\n');
1686 if (SExtOpnd->getOperand(OpIdx)->getType() == SExt->getType() ||
1687 !shouldSExtOperand(SExtOpnd, OpIdx)) {
1688 DEBUG(dbgs() << "No need to propagate\n");
1691 // Check if we can statically sign extend the operand.
1692 Value *Opnd = SExtOpnd->getOperand(OpIdx);
1693 if (const ConstantInt *Cst = dyn_cast<ConstantInt>(Opnd)) {
1694 DEBUG(dbgs() << "Statically sign extend\n");
1697 ConstantInt::getSigned(SExt->getType(), Cst->getSExtValue()));
1700 // UndefValue are typed, so we have to statically sign extend them.
1701 if (isa<UndefValue>(Opnd)) {
1702 DEBUG(dbgs() << "Statically sign extend\n");
1703 TPT.setOperand(SExtOpnd, OpIdx, UndefValue::get(SExt->getType()));
1707 // Otherwise we have to explicity sign extend the operand.
1708 // Check if SExt was reused to sign extend an operand.
1710 // If yes, create a new one.
1711 DEBUG(dbgs() << "More operands to sext\n");
1712 SExtForOpnd = TPT.createSExt(SExt, Opnd, SExt->getType());
1716 TPT.setOperand(SExtForOpnd, 0, Opnd);
1718 // Move the sign extension before the insertion point.
1719 TPT.moveBefore(SExtForOpnd, SExtOpnd);
1720 TPT.setOperand(SExtOpnd, OpIdx, SExtForOpnd);
1721 // If more sext are required, new instructions will have to be created.
1724 if (SExtForOpnd == SExt) {
1725 DEBUG(dbgs() << "Sign extension is useless now\n");
1726 TPT.eraseInstruction(SExt);
1731 /// MatchOperationAddr - Given an instruction or constant expr, see if we can
1732 /// fold the operation into the addressing mode. If so, update the addressing
1733 /// mode and return true, otherwise return false without modifying AddrMode.
1734 /// If \p MovedAway is not NULL, it contains the information of whether or
1735 /// not AddrInst has to be folded into the addressing mode on success.
1736 /// If \p MovedAway == true, \p AddrInst will not be part of the addressing
1737 /// because it has been moved away.
1738 /// Thus AddrInst must not be added in the matched instructions.
1739 /// This state can happen when AddrInst is a sext, since it may be moved away.
1740 /// Therefore, AddrInst may not be valid when MovedAway is true and it must
1741 /// not be referenced anymore.
1742 bool AddressingModeMatcher::MatchOperationAddr(User *AddrInst, unsigned Opcode,
1745 // Avoid exponential behavior on extremely deep expression trees.
1746 if (Depth >= 5) return false;
1748 // By default, all matched instructions stay in place.
1753 case Instruction::PtrToInt:
1754 // PtrToInt is always a noop, as we know that the int type is pointer sized.
1755 return MatchAddr(AddrInst->getOperand(0), Depth);
1756 case Instruction::IntToPtr:
1757 // This inttoptr is a no-op if the integer type is pointer sized.
1758 if (TLI.getValueType(AddrInst->getOperand(0)->getType()) ==
1759 TLI.getPointerTy(AddrInst->getType()->getPointerAddressSpace()))
1760 return MatchAddr(AddrInst->getOperand(0), Depth);
1762 case Instruction::BitCast:
1763 // BitCast is always a noop, and we can handle it as long as it is
1764 // int->int or pointer->pointer (we don't want int<->fp or something).
1765 if ((AddrInst->getOperand(0)->getType()->isPointerTy() ||
1766 AddrInst->getOperand(0)->getType()->isIntegerTy()) &&
1767 // Don't touch identity bitcasts. These were probably put here by LSR,
1768 // and we don't want to mess around with them. Assume it knows what it
1770 AddrInst->getOperand(0)->getType() != AddrInst->getType())
1771 return MatchAddr(AddrInst->getOperand(0), Depth);
1773 case Instruction::Add: {
1774 // Check to see if we can merge in the RHS then the LHS. If so, we win.
1775 ExtAddrMode BackupAddrMode = AddrMode;
1776 unsigned OldSize = AddrModeInsts.size();
1777 // Start a transaction at this point.
1778 // The LHS may match but not the RHS.
1779 // Therefore, we need a higher level restoration point to undo partially
1780 // matched operation.
1781 TypePromotionTransaction::ConstRestorationPt LastKnownGood =
1782 TPT.getRestorationPoint();
1784 if (MatchAddr(AddrInst->getOperand(1), Depth+1) &&
1785 MatchAddr(AddrInst->getOperand(0), Depth+1))
1788 // Restore the old addr mode info.
1789 AddrMode = BackupAddrMode;
1790 AddrModeInsts.resize(OldSize);
1791 TPT.rollback(LastKnownGood);
1793 // Otherwise this was over-aggressive. Try merging in the LHS then the RHS.
1794 if (MatchAddr(AddrInst->getOperand(0), Depth+1) &&
1795 MatchAddr(AddrInst->getOperand(1), Depth+1))
1798 // Otherwise we definitely can't merge the ADD in.
1799 AddrMode = BackupAddrMode;
1800 AddrModeInsts.resize(OldSize);
1801 TPT.rollback(LastKnownGood);
1804 //case Instruction::Or:
1805 // TODO: We can handle "Or Val, Imm" iff this OR is equivalent to an ADD.
1807 case Instruction::Mul:
1808 case Instruction::Shl: {
1809 // Can only handle X*C and X << C.
1810 ConstantInt *RHS = dyn_cast<ConstantInt>(AddrInst->getOperand(1));
1811 if (!RHS) return false;
1812 int64_t Scale = RHS->getSExtValue();
1813 if (Opcode == Instruction::Shl)
1814 Scale = 1LL << Scale;
1816 return MatchScaledValue(AddrInst->getOperand(0), Scale, Depth);
1818 case Instruction::GetElementPtr: {
1819 // Scan the GEP. We check it if it contains constant offsets and at most
1820 // one variable offset.
1821 int VariableOperand = -1;
1822 unsigned VariableScale = 0;
1824 int64_t ConstantOffset = 0;
1825 const DataLayout *TD = TLI.getDataLayout();
1826 gep_type_iterator GTI = gep_type_begin(AddrInst);
1827 for (unsigned i = 1, e = AddrInst->getNumOperands(); i != e; ++i, ++GTI) {
1828 if (StructType *STy = dyn_cast<StructType>(*GTI)) {
1829 const StructLayout *SL = TD->getStructLayout(STy);
1831 cast<ConstantInt>(AddrInst->getOperand(i))->getZExtValue();
1832 ConstantOffset += SL->getElementOffset(Idx);
1834 uint64_t TypeSize = TD->getTypeAllocSize(GTI.getIndexedType());
1835 if (ConstantInt *CI = dyn_cast<ConstantInt>(AddrInst->getOperand(i))) {
1836 ConstantOffset += CI->getSExtValue()*TypeSize;
1837 } else if (TypeSize) { // Scales of zero don't do anything.
1838 // We only allow one variable index at the moment.
1839 if (VariableOperand != -1)
1842 // Remember the variable index.
1843 VariableOperand = i;
1844 VariableScale = TypeSize;
1849 // A common case is for the GEP to only do a constant offset. In this case,
1850 // just add it to the disp field and check validity.
1851 if (VariableOperand == -1) {
1852 AddrMode.BaseOffs += ConstantOffset;
1853 if (ConstantOffset == 0 || TLI.isLegalAddressingMode(AddrMode, AccessTy)){
1854 // Check to see if we can fold the base pointer in too.
1855 if (MatchAddr(AddrInst->getOperand(0), Depth+1))
1858 AddrMode.BaseOffs -= ConstantOffset;
1862 // Save the valid addressing mode in case we can't match.
1863 ExtAddrMode BackupAddrMode = AddrMode;
1864 unsigned OldSize = AddrModeInsts.size();
1866 // See if the scale and offset amount is valid for this target.
1867 AddrMode.BaseOffs += ConstantOffset;
1869 // Match the base operand of the GEP.
1870 if (!MatchAddr(AddrInst->getOperand(0), Depth+1)) {
1871 // If it couldn't be matched, just stuff the value in a register.
1872 if (AddrMode.HasBaseReg) {
1873 AddrMode = BackupAddrMode;
1874 AddrModeInsts.resize(OldSize);
1877 AddrMode.HasBaseReg = true;
1878 AddrMode.BaseReg = AddrInst->getOperand(0);
1881 // Match the remaining variable portion of the GEP.
1882 if (!MatchScaledValue(AddrInst->getOperand(VariableOperand), VariableScale,
1884 // If it couldn't be matched, try stuffing the base into a register
1885 // instead of matching it, and retrying the match of the scale.
1886 AddrMode = BackupAddrMode;
1887 AddrModeInsts.resize(OldSize);
1888 if (AddrMode.HasBaseReg)
1890 AddrMode.HasBaseReg = true;
1891 AddrMode.BaseReg = AddrInst->getOperand(0);
1892 AddrMode.BaseOffs += ConstantOffset;
1893 if (!MatchScaledValue(AddrInst->getOperand(VariableOperand),
1894 VariableScale, Depth)) {
1895 // If even that didn't work, bail.
1896 AddrMode = BackupAddrMode;
1897 AddrModeInsts.resize(OldSize);
1904 case Instruction::SExt: {
1905 // Try to move this sext out of the way of the addressing mode.
1906 Instruction *SExt = cast<Instruction>(AddrInst);
1907 // Ask for a method for doing so.
1908 TypePromotionHelper::Action TPH = TypePromotionHelper::getAction(
1909 SExt, InsertedTruncs, TLI, PromotedInsts);
1913 TypePromotionTransaction::ConstRestorationPt LastKnownGood =
1914 TPT.getRestorationPoint();
1915 unsigned CreatedInsts = 0;
1916 Value *PromotedOperand = TPH(SExt, TPT, PromotedInsts, CreatedInsts);
1917 // SExt has been moved away.
1918 // Thus either it will be rematched later in the recursive calls or it is
1919 // gone. Anyway, we must not fold it into the addressing mode at this point.
1923 // addr = gep base, idx
1925 // promotedOpnd = sext opnd <- no match here
1926 // op = promoted_add promotedOpnd, 1 <- match (later in recursive calls)
1927 // addr = gep base, op <- match
1931 assert(PromotedOperand &&
1932 "TypePromotionHelper should have filtered out those cases");
1934 ExtAddrMode BackupAddrMode = AddrMode;
1935 unsigned OldSize = AddrModeInsts.size();
1937 if (!MatchAddr(PromotedOperand, Depth) ||
1938 // We fold less instructions than what we created.
1939 // Undo at this point.
1940 (OldSize + CreatedInsts > AddrModeInsts.size())) {
1941 AddrMode = BackupAddrMode;
1942 AddrModeInsts.resize(OldSize);
1943 DEBUG(dbgs() << "Sign extension does not pay off: rollback\n");
1944 TPT.rollback(LastKnownGood);
1953 /// MatchAddr - If we can, try to add the value of 'Addr' into the current
1954 /// addressing mode. If Addr can't be added to AddrMode this returns false and
1955 /// leaves AddrMode unmodified. This assumes that Addr is either a pointer type
1956 /// or intptr_t for the target.
1958 bool AddressingModeMatcher::MatchAddr(Value *Addr, unsigned Depth) {
1959 // Start a transaction at this point that we will rollback if the matching
1961 TypePromotionTransaction::ConstRestorationPt LastKnownGood =
1962 TPT.getRestorationPoint();
1963 if (ConstantInt *CI = dyn_cast<ConstantInt>(Addr)) {
1964 // Fold in immediates if legal for the target.
1965 AddrMode.BaseOffs += CI->getSExtValue();
1966 if (TLI.isLegalAddressingMode(AddrMode, AccessTy))
1968 AddrMode.BaseOffs -= CI->getSExtValue();
1969 } else if (GlobalValue *GV = dyn_cast<GlobalValue>(Addr)) {
1970 // If this is a global variable, try to fold it into the addressing mode.
1971 if (AddrMode.BaseGV == 0) {
1972 AddrMode.BaseGV = GV;
1973 if (TLI.isLegalAddressingMode(AddrMode, AccessTy))
1975 AddrMode.BaseGV = 0;
1977 } else if (Instruction *I = dyn_cast<Instruction>(Addr)) {
1978 ExtAddrMode BackupAddrMode = AddrMode;
1979 unsigned OldSize = AddrModeInsts.size();
1981 // Check to see if it is possible to fold this operation.
1982 bool MovedAway = false;
1983 if (MatchOperationAddr(I, I->getOpcode(), Depth, &MovedAway)) {
1984 // This instruction may have been move away. If so, there is nothing
1988 // Okay, it's possible to fold this. Check to see if it is actually
1989 // *profitable* to do so. We use a simple cost model to avoid increasing
1990 // register pressure too much.
1991 if (I->hasOneUse() ||
1992 IsProfitableToFoldIntoAddressingMode(I, BackupAddrMode, AddrMode)) {
1993 AddrModeInsts.push_back(I);
1997 // It isn't profitable to do this, roll back.
1998 //cerr << "NOT FOLDING: " << *I;
1999 AddrMode = BackupAddrMode;
2000 AddrModeInsts.resize(OldSize);
2001 TPT.rollback(LastKnownGood);
2003 } else if (ConstantExpr *CE = dyn_cast<ConstantExpr>(Addr)) {
2004 if (MatchOperationAddr(CE, CE->getOpcode(), Depth))
2006 TPT.rollback(LastKnownGood);
2007 } else if (isa<ConstantPointerNull>(Addr)) {
2008 // Null pointer gets folded without affecting the addressing mode.
2012 // Worse case, the target should support [reg] addressing modes. :)
2013 if (!AddrMode.HasBaseReg) {
2014 AddrMode.HasBaseReg = true;
2015 AddrMode.BaseReg = Addr;
2016 // Still check for legality in case the target supports [imm] but not [i+r].
2017 if (TLI.isLegalAddressingMode(AddrMode, AccessTy))
2019 AddrMode.HasBaseReg = false;
2020 AddrMode.BaseReg = 0;
2023 // If the base register is already taken, see if we can do [r+r].
2024 if (AddrMode.Scale == 0) {
2026 AddrMode.ScaledReg = Addr;
2027 if (TLI.isLegalAddressingMode(AddrMode, AccessTy))
2030 AddrMode.ScaledReg = 0;
2033 TPT.rollback(LastKnownGood);
2037 /// IsOperandAMemoryOperand - Check to see if all uses of OpVal by the specified
2038 /// inline asm call are due to memory operands. If so, return true, otherwise
2040 static bool IsOperandAMemoryOperand(CallInst *CI, InlineAsm *IA, Value *OpVal,
2041 const TargetLowering &TLI) {
2042 TargetLowering::AsmOperandInfoVector TargetConstraints = TLI.ParseConstraints(ImmutableCallSite(CI));
2043 for (unsigned i = 0, e = TargetConstraints.size(); i != e; ++i) {
2044 TargetLowering::AsmOperandInfo &OpInfo = TargetConstraints[i];
2046 // Compute the constraint code and ConstraintType to use.
2047 TLI.ComputeConstraintToUse(OpInfo, SDValue());
2049 // If this asm operand is our Value*, and if it isn't an indirect memory
2050 // operand, we can't fold it!
2051 if (OpInfo.CallOperandVal == OpVal &&
2052 (OpInfo.ConstraintType != TargetLowering::C_Memory ||
2053 !OpInfo.isIndirect))
2060 /// FindAllMemoryUses - Recursively walk all the uses of I until we find a
2061 /// memory use. If we find an obviously non-foldable instruction, return true.
2062 /// Add the ultimately found memory instructions to MemoryUses.
2063 static bool FindAllMemoryUses(Instruction *I,
2064 SmallVectorImpl<std::pair<Instruction*,unsigned> > &MemoryUses,
2065 SmallPtrSet<Instruction*, 16> &ConsideredInsts,
2066 const TargetLowering &TLI) {
2067 // If we already considered this instruction, we're done.
2068 if (!ConsideredInsts.insert(I))
2071 // If this is an obviously unfoldable instruction, bail out.
2072 if (!MightBeFoldableInst(I))
2075 // Loop over all the uses, recursively processing them.
2076 for (Value::use_iterator UI = I->use_begin(), E = I->use_end();
2080 if (LoadInst *LI = dyn_cast<LoadInst>(U)) {
2081 MemoryUses.push_back(std::make_pair(LI, UI.getOperandNo()));
2085 if (StoreInst *SI = dyn_cast<StoreInst>(U)) {
2086 unsigned opNo = UI.getOperandNo();
2087 if (opNo == 0) return true; // Storing addr, not into addr.
2088 MemoryUses.push_back(std::make_pair(SI, opNo));
2092 if (CallInst *CI = dyn_cast<CallInst>(U)) {
2093 InlineAsm *IA = dyn_cast<InlineAsm>(CI->getCalledValue());
2094 if (!IA) return true;
2096 // If this is a memory operand, we're cool, otherwise bail out.
2097 if (!IsOperandAMemoryOperand(CI, IA, I, TLI))
2102 if (FindAllMemoryUses(cast<Instruction>(U), MemoryUses, ConsideredInsts,
2110 /// ValueAlreadyLiveAtInst - Retrn true if Val is already known to be live at
2111 /// the use site that we're folding it into. If so, there is no cost to
2112 /// include it in the addressing mode. KnownLive1 and KnownLive2 are two values
2113 /// that we know are live at the instruction already.
2114 bool AddressingModeMatcher::ValueAlreadyLiveAtInst(Value *Val,Value *KnownLive1,
2115 Value *KnownLive2) {
2116 // If Val is either of the known-live values, we know it is live!
2117 if (Val == 0 || Val == KnownLive1 || Val == KnownLive2)
2120 // All values other than instructions and arguments (e.g. constants) are live.
2121 if (!isa<Instruction>(Val) && !isa<Argument>(Val)) return true;
2123 // If Val is a constant sized alloca in the entry block, it is live, this is
2124 // true because it is just a reference to the stack/frame pointer, which is
2125 // live for the whole function.
2126 if (AllocaInst *AI = dyn_cast<AllocaInst>(Val))
2127 if (AI->isStaticAlloca())
2130 // Check to see if this value is already used in the memory instruction's
2131 // block. If so, it's already live into the block at the very least, so we
2132 // can reasonably fold it.
2133 return Val->isUsedInBasicBlock(MemoryInst->getParent());
2136 /// IsProfitableToFoldIntoAddressingMode - It is possible for the addressing
2137 /// mode of the machine to fold the specified instruction into a load or store
2138 /// that ultimately uses it. However, the specified instruction has multiple
2139 /// uses. Given this, it may actually increase register pressure to fold it
2140 /// into the load. For example, consider this code:
2144 /// use(Y) -> nonload/store
2148 /// In this case, Y has multiple uses, and can be folded into the load of Z
2149 /// (yielding load [X+2]). However, doing this will cause both "X" and "X+1" to
2150 /// be live at the use(Y) line. If we don't fold Y into load Z, we use one
2151 /// fewer register. Since Y can't be folded into "use(Y)" we don't increase the
2152 /// number of computations either.
2154 /// Note that this (like most of CodeGenPrepare) is just a rough heuristic. If
2155 /// X was live across 'load Z' for other reasons, we actually *would* want to
2156 /// fold the addressing mode in the Z case. This would make Y die earlier.
2157 bool AddressingModeMatcher::
2158 IsProfitableToFoldIntoAddressingMode(Instruction *I, ExtAddrMode &AMBefore,
2159 ExtAddrMode &AMAfter) {
2160 if (IgnoreProfitability) return true;
2162 // AMBefore is the addressing mode before this instruction was folded into it,
2163 // and AMAfter is the addressing mode after the instruction was folded. Get
2164 // the set of registers referenced by AMAfter and subtract out those
2165 // referenced by AMBefore: this is the set of values which folding in this
2166 // address extends the lifetime of.
2168 // Note that there are only two potential values being referenced here,
2169 // BaseReg and ScaleReg (global addresses are always available, as are any
2170 // folded immediates).
2171 Value *BaseReg = AMAfter.BaseReg, *ScaledReg = AMAfter.ScaledReg;
2173 // If the BaseReg or ScaledReg was referenced by the previous addrmode, their
2174 // lifetime wasn't extended by adding this instruction.
2175 if (ValueAlreadyLiveAtInst(BaseReg, AMBefore.BaseReg, AMBefore.ScaledReg))
2177 if (ValueAlreadyLiveAtInst(ScaledReg, AMBefore.BaseReg, AMBefore.ScaledReg))
2180 // If folding this instruction (and it's subexprs) didn't extend any live
2181 // ranges, we're ok with it.
2182 if (BaseReg == 0 && ScaledReg == 0)
2185 // If all uses of this instruction are ultimately load/store/inlineasm's,
2186 // check to see if their addressing modes will include this instruction. If
2187 // so, we can fold it into all uses, so it doesn't matter if it has multiple
2189 SmallVector<std::pair<Instruction*,unsigned>, 16> MemoryUses;
2190 SmallPtrSet<Instruction*, 16> ConsideredInsts;
2191 if (FindAllMemoryUses(I, MemoryUses, ConsideredInsts, TLI))
2192 return false; // Has a non-memory, non-foldable use!
2194 // Now that we know that all uses of this instruction are part of a chain of
2195 // computation involving only operations that could theoretically be folded
2196 // into a memory use, loop over each of these uses and see if they could
2197 // *actually* fold the instruction.
2198 SmallVector<Instruction*, 32> MatchedAddrModeInsts;
2199 for (unsigned i = 0, e = MemoryUses.size(); i != e; ++i) {
2200 Instruction *User = MemoryUses[i].first;
2201 unsigned OpNo = MemoryUses[i].second;
2203 // Get the access type of this use. If the use isn't a pointer, we don't
2204 // know what it accesses.
2205 Value *Address = User->getOperand(OpNo);
2206 if (!Address->getType()->isPointerTy())
2208 Type *AddressAccessTy = Address->getType()->getPointerElementType();
2210 // Do a match against the root of this address, ignoring profitability. This
2211 // will tell us if the addressing mode for the memory operation will
2212 // *actually* cover the shared instruction.
2214 TypePromotionTransaction::ConstRestorationPt LastKnownGood =
2215 TPT.getRestorationPoint();
2216 AddressingModeMatcher Matcher(MatchedAddrModeInsts, TLI, AddressAccessTy,
2217 MemoryInst, Result, InsertedTruncs,
2218 PromotedInsts, TPT);
2219 Matcher.IgnoreProfitability = true;
2220 bool Success = Matcher.MatchAddr(Address, 0);
2221 (void)Success; assert(Success && "Couldn't select *anything*?");
2223 // The match was to check the profitability, the changes made are not
2224 // part of the original matcher. Therefore, they should be dropped
2225 // otherwise the original matcher will not present the right state.
2226 TPT.rollback(LastKnownGood);
2228 // If the match didn't cover I, then it won't be shared by it.
2229 if (std::find(MatchedAddrModeInsts.begin(), MatchedAddrModeInsts.end(),
2230 I) == MatchedAddrModeInsts.end())
2233 MatchedAddrModeInsts.clear();
2239 } // end anonymous namespace
2241 /// IsNonLocalValue - Return true if the specified values are defined in a
2242 /// different basic block than BB.
2243 static bool IsNonLocalValue(Value *V, BasicBlock *BB) {
2244 if (Instruction *I = dyn_cast<Instruction>(V))
2245 return I->getParent() != BB;
2249 /// OptimizeMemoryInst - Load and Store Instructions often have
2250 /// addressing modes that can do significant amounts of computation. As such,
2251 /// instruction selection will try to get the load or store to do as much
2252 /// computation as possible for the program. The problem is that isel can only
2253 /// see within a single block. As such, we sink as much legal addressing mode
2254 /// stuff into the block as possible.
2256 /// This method is used to optimize both load/store and inline asms with memory
2258 bool CodeGenPrepare::OptimizeMemoryInst(Instruction *MemoryInst, Value *Addr,
2262 // Try to collapse single-value PHI nodes. This is necessary to undo
2263 // unprofitable PRE transformations.
2264 SmallVector<Value*, 8> worklist;
2265 SmallPtrSet<Value*, 16> Visited;
2266 worklist.push_back(Addr);
2268 // Use a worklist to iteratively look through PHI nodes, and ensure that
2269 // the addressing mode obtained from the non-PHI roots of the graph
2271 Value *Consensus = 0;
2272 unsigned NumUsesConsensus = 0;
2273 bool IsNumUsesConsensusValid = false;
2274 SmallVector<Instruction*, 16> AddrModeInsts;
2275 ExtAddrMode AddrMode;
2276 TypePromotionTransaction TPT;
2277 TypePromotionTransaction::ConstRestorationPt LastKnownGood =
2278 TPT.getRestorationPoint();
2279 while (!worklist.empty()) {
2280 Value *V = worklist.back();
2281 worklist.pop_back();
2283 // Break use-def graph loops.
2284 if (!Visited.insert(V)) {
2289 // For a PHI node, push all of its incoming values.
2290 if (PHINode *P = dyn_cast<PHINode>(V)) {
2291 for (unsigned i = 0, e = P->getNumIncomingValues(); i != e; ++i)
2292 worklist.push_back(P->getIncomingValue(i));
2296 // For non-PHIs, determine the addressing mode being computed.
2297 SmallVector<Instruction*, 16> NewAddrModeInsts;
2298 ExtAddrMode NewAddrMode = AddressingModeMatcher::Match(
2299 V, AccessTy, MemoryInst, NewAddrModeInsts, *TLI, InsertedTruncsSet,
2300 PromotedInsts, TPT);
2302 // This check is broken into two cases with very similar code to avoid using
2303 // getNumUses() as much as possible. Some values have a lot of uses, so
2304 // calling getNumUses() unconditionally caused a significant compile-time
2308 AddrMode = NewAddrMode;
2309 AddrModeInsts = NewAddrModeInsts;
2311 } else if (NewAddrMode == AddrMode) {
2312 if (!IsNumUsesConsensusValid) {
2313 NumUsesConsensus = Consensus->getNumUses();
2314 IsNumUsesConsensusValid = true;
2317 // Ensure that the obtained addressing mode is equivalent to that obtained
2318 // for all other roots of the PHI traversal. Also, when choosing one
2319 // such root as representative, select the one with the most uses in order
2320 // to keep the cost modeling heuristics in AddressingModeMatcher
2322 unsigned NumUses = V->getNumUses();
2323 if (NumUses > NumUsesConsensus) {
2325 NumUsesConsensus = NumUses;
2326 AddrModeInsts = NewAddrModeInsts;
2335 // If the addressing mode couldn't be determined, or if multiple different
2336 // ones were determined, bail out now.
2338 TPT.rollback(LastKnownGood);
2343 // Check to see if any of the instructions supersumed by this addr mode are
2344 // non-local to I's BB.
2345 bool AnyNonLocal = false;
2346 for (unsigned i = 0, e = AddrModeInsts.size(); i != e; ++i) {
2347 if (IsNonLocalValue(AddrModeInsts[i], MemoryInst->getParent())) {
2353 // If all the instructions matched are already in this BB, don't do anything.
2355 DEBUG(dbgs() << "CGP: Found local addrmode: " << AddrMode << "\n");
2359 // Insert this computation right after this user. Since our caller is
2360 // scanning from the top of the BB to the bottom, reuse of the expr are
2361 // guaranteed to happen later.
2362 IRBuilder<> Builder(MemoryInst);
2364 // Now that we determined the addressing expression we want to use and know
2365 // that we have to sink it into this block. Check to see if we have already
2366 // done this for some other load/store instr in this block. If so, reuse the
2368 Value *&SunkAddr = SunkAddrs[Addr];
2370 DEBUG(dbgs() << "CGP: Reusing nonlocal addrmode: " << AddrMode << " for "
2372 if (SunkAddr->getType() != Addr->getType())
2373 SunkAddr = Builder.CreateBitCast(SunkAddr, Addr->getType());
2375 DEBUG(dbgs() << "CGP: SINKING nonlocal addrmode: " << AddrMode << " for "
2377 Type *IntPtrTy = TLI->getDataLayout()->getIntPtrType(Addr->getType());
2380 // Start with the base register. Do this first so that subsequent address
2381 // matching finds it last, which will prevent it from trying to match it
2382 // as the scaled value in case it happens to be a mul. That would be
2383 // problematic if we've sunk a different mul for the scale, because then
2384 // we'd end up sinking both muls.
2385 if (AddrMode.BaseReg) {
2386 Value *V = AddrMode.BaseReg;
2387 if (V->getType()->isPointerTy())
2388 V = Builder.CreatePtrToInt(V, IntPtrTy, "sunkaddr");
2389 if (V->getType() != IntPtrTy)
2390 V = Builder.CreateIntCast(V, IntPtrTy, /*isSigned=*/true, "sunkaddr");
2394 // Add the scale value.
2395 if (AddrMode.Scale) {
2396 Value *V = AddrMode.ScaledReg;
2397 if (V->getType() == IntPtrTy) {
2399 } else if (V->getType()->isPointerTy()) {
2400 V = Builder.CreatePtrToInt(V, IntPtrTy, "sunkaddr");
2401 } else if (cast<IntegerType>(IntPtrTy)->getBitWidth() <
2402 cast<IntegerType>(V->getType())->getBitWidth()) {
2403 V = Builder.CreateTrunc(V, IntPtrTy, "sunkaddr");
2405 V = Builder.CreateSExt(V, IntPtrTy, "sunkaddr");
2407 if (AddrMode.Scale != 1)
2408 V = Builder.CreateMul(V, ConstantInt::get(IntPtrTy, AddrMode.Scale),
2411 Result = Builder.CreateAdd(Result, V, "sunkaddr");
2416 // Add in the BaseGV if present.
2417 if (AddrMode.BaseGV) {
2418 Value *V = Builder.CreatePtrToInt(AddrMode.BaseGV, IntPtrTy, "sunkaddr");
2420 Result = Builder.CreateAdd(Result, V, "sunkaddr");
2425 // Add in the Base Offset if present.
2426 if (AddrMode.BaseOffs) {
2427 Value *V = ConstantInt::get(IntPtrTy, AddrMode.BaseOffs);
2429 Result = Builder.CreateAdd(Result, V, "sunkaddr");
2435 SunkAddr = Constant::getNullValue(Addr->getType());
2437 SunkAddr = Builder.CreateIntToPtr(Result, Addr->getType(), "sunkaddr");
2440 MemoryInst->replaceUsesOfWith(Repl, SunkAddr);
2442 // If we have no uses, recursively delete the value and all dead instructions
2444 if (Repl->use_empty()) {
2445 // This can cause recursive deletion, which can invalidate our iterator.
2446 // Use a WeakVH to hold onto it in case this happens.
2447 WeakVH IterHandle(CurInstIterator);
2448 BasicBlock *BB = CurInstIterator->getParent();
2450 RecursivelyDeleteTriviallyDeadInstructions(Repl, TLInfo);
2452 if (IterHandle != CurInstIterator) {
2453 // If the iterator instruction was recursively deleted, start over at the
2454 // start of the block.
2455 CurInstIterator = BB->begin();
2463 /// OptimizeInlineAsmInst - If there are any memory operands, use
2464 /// OptimizeMemoryInst to sink their address computing into the block when
2465 /// possible / profitable.
2466 bool CodeGenPrepare::OptimizeInlineAsmInst(CallInst *CS) {
2467 bool MadeChange = false;
2469 TargetLowering::AsmOperandInfoVector
2470 TargetConstraints = TLI->ParseConstraints(CS);
2472 for (unsigned i = 0, e = TargetConstraints.size(); i != e; ++i) {
2473 TargetLowering::AsmOperandInfo &OpInfo = TargetConstraints[i];
2475 // Compute the constraint code and ConstraintType to use.
2476 TLI->ComputeConstraintToUse(OpInfo, SDValue());
2478 if (OpInfo.ConstraintType == TargetLowering::C_Memory &&
2479 OpInfo.isIndirect) {
2480 Value *OpVal = CS->getArgOperand(ArgNo++);
2481 MadeChange |= OptimizeMemoryInst(CS, OpVal, OpVal->getType());
2482 } else if (OpInfo.Type == InlineAsm::isInput)
2489 /// MoveExtToFormExtLoad - Move a zext or sext fed by a load into the same
2490 /// basic block as the load, unless conditions are unfavorable. This allows
2491 /// SelectionDAG to fold the extend into the load.
2493 bool CodeGenPrepare::MoveExtToFormExtLoad(Instruction *I) {
2494 // Look for a load being extended.
2495 LoadInst *LI = dyn_cast<LoadInst>(I->getOperand(0));
2496 if (!LI) return false;
2498 // If they're already in the same block, there's nothing to do.
2499 if (LI->getParent() == I->getParent())
2502 // If the load has other users and the truncate is not free, this probably
2503 // isn't worthwhile.
2504 if (!LI->hasOneUse() &&
2505 TLI && (TLI->isTypeLegal(TLI->getValueType(LI->getType())) ||
2506 !TLI->isTypeLegal(TLI->getValueType(I->getType()))) &&
2507 !TLI->isTruncateFree(I->getType(), LI->getType()))
2510 // Check whether the target supports casts folded into loads.
2512 if (isa<ZExtInst>(I))
2513 LType = ISD::ZEXTLOAD;
2515 assert(isa<SExtInst>(I) && "Unexpected ext type!");
2516 LType = ISD::SEXTLOAD;
2518 if (TLI && !TLI->isLoadExtLegal(LType, TLI->getValueType(LI->getType())))
2521 // Move the extend into the same block as the load, so that SelectionDAG
2523 I->removeFromParent();
2529 bool CodeGenPrepare::OptimizeExtUses(Instruction *I) {
2530 BasicBlock *DefBB = I->getParent();
2532 // If the result of a {s|z}ext and its source are both live out, rewrite all
2533 // other uses of the source with result of extension.
2534 Value *Src = I->getOperand(0);
2535 if (Src->hasOneUse())
2538 // Only do this xform if truncating is free.
2539 if (TLI && !TLI->isTruncateFree(I->getType(), Src->getType()))
2542 // Only safe to perform the optimization if the source is also defined in
2544 if (!isa<Instruction>(Src) || DefBB != cast<Instruction>(Src)->getParent())
2547 bool DefIsLiveOut = false;
2548 for (Value::use_iterator UI = I->use_begin(), E = I->use_end();
2550 Instruction *User = cast<Instruction>(*UI);
2552 // Figure out which BB this ext is used in.
2553 BasicBlock *UserBB = User->getParent();
2554 if (UserBB == DefBB) continue;
2555 DefIsLiveOut = true;
2561 // Make sure none of the uses are PHI nodes.
2562 for (Value::use_iterator UI = Src->use_begin(), E = Src->use_end();
2564 Instruction *User = cast<Instruction>(*UI);
2565 BasicBlock *UserBB = User->getParent();
2566 if (UserBB == DefBB) continue;
2567 // Be conservative. We don't want this xform to end up introducing
2568 // reloads just before load / store instructions.
2569 if (isa<PHINode>(User) || isa<LoadInst>(User) || isa<StoreInst>(User))
2573 // InsertedTruncs - Only insert one trunc in each block once.
2574 DenseMap<BasicBlock*, Instruction*> InsertedTruncs;
2576 bool MadeChange = false;
2577 for (Value::use_iterator UI = Src->use_begin(), E = Src->use_end();
2579 Use &TheUse = UI.getUse();
2580 Instruction *User = cast<Instruction>(*UI);
2582 // Figure out which BB this ext is used in.
2583 BasicBlock *UserBB = User->getParent();
2584 if (UserBB == DefBB) continue;
2586 // Both src and def are live in this block. Rewrite the use.
2587 Instruction *&InsertedTrunc = InsertedTruncs[UserBB];
2589 if (!InsertedTrunc) {
2590 BasicBlock::iterator InsertPt = UserBB->getFirstInsertionPt();
2591 InsertedTrunc = new TruncInst(I, Src->getType(), "", InsertPt);
2592 InsertedTruncsSet.insert(InsertedTrunc);
2595 // Replace a use of the {s|z}ext source with a use of the result.
2596 TheUse = InsertedTrunc;
2604 /// isFormingBranchFromSelectProfitable - Returns true if a SelectInst should be
2605 /// turned into an explicit branch.
2606 static bool isFormingBranchFromSelectProfitable(SelectInst *SI) {
2607 // FIXME: This should use the same heuristics as IfConversion to determine
2608 // whether a select is better represented as a branch. This requires that
2609 // branch probability metadata is preserved for the select, which is not the
2612 CmpInst *Cmp = dyn_cast<CmpInst>(SI->getCondition());
2614 // If the branch is predicted right, an out of order CPU can avoid blocking on
2615 // the compare. Emit cmovs on compares with a memory operand as branches to
2616 // avoid stalls on the load from memory. If the compare has more than one use
2617 // there's probably another cmov or setcc around so it's not worth emitting a
2622 Value *CmpOp0 = Cmp->getOperand(0);
2623 Value *CmpOp1 = Cmp->getOperand(1);
2625 // We check that the memory operand has one use to avoid uses of the loaded
2626 // value directly after the compare, making branches unprofitable.
2627 return Cmp->hasOneUse() &&
2628 ((isa<LoadInst>(CmpOp0) && CmpOp0->hasOneUse()) ||
2629 (isa<LoadInst>(CmpOp1) && CmpOp1->hasOneUse()));
2633 /// If we have a SelectInst that will likely profit from branch prediction,
2634 /// turn it into a branch.
2635 bool CodeGenPrepare::OptimizeSelectInst(SelectInst *SI) {
2636 bool VectorCond = !SI->getCondition()->getType()->isIntegerTy(1);
2638 // Can we convert the 'select' to CF ?
2639 if (DisableSelectToBranch || OptSize || !TLI || VectorCond)
2642 TargetLowering::SelectSupportKind SelectKind;
2644 SelectKind = TargetLowering::VectorMaskSelect;
2645 else if (SI->getType()->isVectorTy())
2646 SelectKind = TargetLowering::ScalarCondVectorVal;
2648 SelectKind = TargetLowering::ScalarValSelect;
2650 // Do we have efficient codegen support for this kind of 'selects' ?
2651 if (TLI->isSelectSupported(SelectKind)) {
2652 // We have efficient codegen support for the select instruction.
2653 // Check if it is profitable to keep this 'select'.
2654 if (!TLI->isPredictableSelectExpensive() ||
2655 !isFormingBranchFromSelectProfitable(SI))
2661 // First, we split the block containing the select into 2 blocks.
2662 BasicBlock *StartBlock = SI->getParent();
2663 BasicBlock::iterator SplitPt = ++(BasicBlock::iterator(SI));
2664 BasicBlock *NextBlock = StartBlock->splitBasicBlock(SplitPt, "select.end");
2666 // Create a new block serving as the landing pad for the branch.
2667 BasicBlock *SmallBlock = BasicBlock::Create(SI->getContext(), "select.mid",
2668 NextBlock->getParent(), NextBlock);
2670 // Move the unconditional branch from the block with the select in it into our
2671 // landing pad block.
2672 StartBlock->getTerminator()->eraseFromParent();
2673 BranchInst::Create(NextBlock, SmallBlock);
2675 // Insert the real conditional branch based on the original condition.
2676 BranchInst::Create(NextBlock, SmallBlock, SI->getCondition(), SI);
2678 // The select itself is replaced with a PHI Node.
2679 PHINode *PN = PHINode::Create(SI->getType(), 2, "", NextBlock->begin());
2681 PN->addIncoming(SI->getTrueValue(), StartBlock);
2682 PN->addIncoming(SI->getFalseValue(), SmallBlock);
2683 SI->replaceAllUsesWith(PN);
2684 SI->eraseFromParent();
2686 // Instruct OptimizeBlock to skip to the next block.
2687 CurInstIterator = StartBlock->end();
2688 ++NumSelectsExpanded;
2692 bool CodeGenPrepare::OptimizeInst(Instruction *I) {
2693 if (PHINode *P = dyn_cast<PHINode>(I)) {
2694 // It is possible for very late stage optimizations (such as SimplifyCFG)
2695 // to introduce PHI nodes too late to be cleaned up. If we detect such a
2696 // trivial PHI, go ahead and zap it here.
2697 if (Value *V = SimplifyInstruction(P, TLI ? TLI->getDataLayout() : 0,
2699 P->replaceAllUsesWith(V);
2700 P->eraseFromParent();
2707 if (CastInst *CI = dyn_cast<CastInst>(I)) {
2708 // If the source of the cast is a constant, then this should have
2709 // already been constant folded. The only reason NOT to constant fold
2710 // it is if something (e.g. LSR) was careful to place the constant
2711 // evaluation in a block other than then one that uses it (e.g. to hoist
2712 // the address of globals out of a loop). If this is the case, we don't
2713 // want to forward-subst the cast.
2714 if (isa<Constant>(CI->getOperand(0)))
2717 if (TLI && OptimizeNoopCopyExpression(CI, *TLI))
2720 if (isa<ZExtInst>(I) || isa<SExtInst>(I)) {
2721 bool MadeChange = MoveExtToFormExtLoad(I);
2722 return MadeChange | OptimizeExtUses(I);
2727 if (CmpInst *CI = dyn_cast<CmpInst>(I))
2728 if (!TLI || !TLI->hasMultipleConditionRegisters())
2729 return OptimizeCmpExpression(CI);
2731 if (LoadInst *LI = dyn_cast<LoadInst>(I)) {
2733 return OptimizeMemoryInst(I, I->getOperand(0), LI->getType());
2737 if (StoreInst *SI = dyn_cast<StoreInst>(I)) {
2739 return OptimizeMemoryInst(I, SI->getOperand(1),
2740 SI->getOperand(0)->getType());
2744 if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(I)) {
2745 if (GEPI->hasAllZeroIndices()) {
2746 /// The GEP operand must be a pointer, so must its result -> BitCast
2747 Instruction *NC = new BitCastInst(GEPI->getOperand(0), GEPI->getType(),
2748 GEPI->getName(), GEPI);
2749 GEPI->replaceAllUsesWith(NC);
2750 GEPI->eraseFromParent();
2758 if (CallInst *CI = dyn_cast<CallInst>(I))
2759 return OptimizeCallInst(CI);
2761 if (SelectInst *SI = dyn_cast<SelectInst>(I))
2762 return OptimizeSelectInst(SI);
2767 // In this pass we look for GEP and cast instructions that are used
2768 // across basic blocks and rewrite them to improve basic-block-at-a-time
2770 bool CodeGenPrepare::OptimizeBlock(BasicBlock &BB) {
2772 bool MadeChange = false;
2774 CurInstIterator = BB.begin();
2775 while (CurInstIterator != BB.end())
2776 MadeChange |= OptimizeInst(CurInstIterator++);
2778 MadeChange |= DupRetToEnableTailCallOpts(&BB);
2783 // llvm.dbg.value is far away from the value then iSel may not be able
2784 // handle it properly. iSel will drop llvm.dbg.value if it can not
2785 // find a node corresponding to the value.
2786 bool CodeGenPrepare::PlaceDbgValues(Function &F) {
2787 bool MadeChange = false;
2788 for (Function::iterator I = F.begin(), E = F.end(); I != E; ++I) {
2789 Instruction *PrevNonDbgInst = NULL;
2790 for (BasicBlock::iterator BI = I->begin(), BE = I->end(); BI != BE;) {
2791 Instruction *Insn = BI; ++BI;
2792 DbgValueInst *DVI = dyn_cast<DbgValueInst>(Insn);
2794 PrevNonDbgInst = Insn;
2798 Instruction *VI = dyn_cast_or_null<Instruction>(DVI->getValue());
2799 if (VI && VI != PrevNonDbgInst && !VI->isTerminator()) {
2800 DEBUG(dbgs() << "Moving Debug Value before :\n" << *DVI << ' ' << *VI);
2801 DVI->removeFromParent();
2802 if (isa<PHINode>(VI))
2803 DVI->insertBefore(VI->getParent()->getFirstInsertionPt());
2805 DVI->insertAfter(VI);