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/CodeGen/Passes.h"
18 #include "llvm/ADT/DenseMap.h"
19 #include "llvm/ADT/SmallSet.h"
20 #include "llvm/ADT/Statistic.h"
21 #include "llvm/Analysis/InstructionSimplify.h"
22 #include "llvm/IR/CallSite.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/GetElementPtrTypeIterator.h"
29 #include "llvm/IR/IRBuilder.h"
30 #include "llvm/IR/InlineAsm.h"
31 #include "llvm/IR/Instructions.h"
32 #include "llvm/IR/IntrinsicInst.h"
33 #include "llvm/IR/PatternMatch.h"
34 #include "llvm/IR/ValueHandle.h"
35 #include "llvm/IR/ValueMap.h"
36 #include "llvm/Pass.h"
37 #include "llvm/Support/CommandLine.h"
38 #include "llvm/Support/Debug.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) override;
115 const char *getPassName() const override { return "CodeGen Prepare"; }
117 void getAnalysisUsage(AnalysisUsage &AU) const override {
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 OptimizeShuffleVectorInst(ShuffleVectorInst *SI);
136 bool DupRetToEnableTailCallOpts(BasicBlock *BB);
137 bool PlaceDbgValues(Function &F);
141 char CodeGenPrepare::ID = 0;
142 static void *initializeCodeGenPreparePassOnce(PassRegistry &Registry) {
143 initializeTargetLibraryInfoPass(Registry);
144 PassInfo *PI = new PassInfo(
145 "Optimize for code generation", "codegenprepare", &CodeGenPrepare::ID,
146 PassInfo::NormalCtor_t(callDefaultCtor<CodeGenPrepare>), false, false,
147 PassInfo::TargetMachineCtor_t(callTargetMachineCtor<CodeGenPrepare>));
148 Registry.registerPass(*PI, true);
152 void llvm::initializeCodeGenPreparePass(PassRegistry &Registry) {
153 CALL_ONCE_INITIALIZATION(initializeCodeGenPreparePassOnce)
156 FunctionPass *llvm::createCodeGenPreparePass(const TargetMachine *TM) {
157 return new CodeGenPrepare(TM);
160 bool CodeGenPrepare::runOnFunction(Function &F) {
161 bool EverMadeChange = false;
162 // Clear per function information.
163 InsertedTruncsSet.clear();
164 PromotedInsts.clear();
167 if (TM) TLI = TM->getTargetLowering();
168 TLInfo = &getAnalysis<TargetLibraryInfo>();
169 DominatorTreeWrapperPass *DTWP =
170 getAnalysisIfAvailable<DominatorTreeWrapperPass>();
171 DT = DTWP ? &DTWP->getDomTree() : 0;
172 OptSize = F.getAttributes().hasAttribute(AttributeSet::FunctionIndex,
173 Attribute::OptimizeForSize);
175 /// This optimization identifies DIV instructions that can be
176 /// profitably bypassed and carried out with a shorter, faster divide.
177 if (!OptSize && TLI && TLI->isSlowDivBypassed()) {
178 const DenseMap<unsigned int, unsigned int> &BypassWidths =
179 TLI->getBypassSlowDivWidths();
180 for (Function::iterator I = F.begin(); I != F.end(); I++)
181 EverMadeChange |= bypassSlowDivision(F, I, BypassWidths);
184 // Eliminate blocks that contain only PHI nodes and an
185 // unconditional branch.
186 EverMadeChange |= EliminateMostlyEmptyBlocks(F);
188 // llvm.dbg.value is far away from the value then iSel may not be able
189 // handle it properly. iSel will drop llvm.dbg.value if it can not
190 // find a node corresponding to the value.
191 EverMadeChange |= PlaceDbgValues(F);
193 bool MadeChange = true;
196 for (Function::iterator I = F.begin(); I != F.end(); ) {
197 BasicBlock *BB = I++;
198 MadeChange |= OptimizeBlock(*BB);
200 EverMadeChange |= MadeChange;
205 if (!DisableBranchOpts) {
207 SmallPtrSet<BasicBlock*, 8> WorkList;
208 for (Function::iterator BB = F.begin(), E = F.end(); BB != E; ++BB) {
209 SmallVector<BasicBlock*, 2> Successors(succ_begin(BB), succ_end(BB));
210 MadeChange |= ConstantFoldTerminator(BB, true);
211 if (!MadeChange) continue;
213 for (SmallVectorImpl<BasicBlock*>::iterator
214 II = Successors.begin(), IE = Successors.end(); II != IE; ++II)
215 if (pred_begin(*II) == pred_end(*II))
216 WorkList.insert(*II);
219 // Delete the dead blocks and any of their dead successors.
220 MadeChange |= !WorkList.empty();
221 while (!WorkList.empty()) {
222 BasicBlock *BB = *WorkList.begin();
224 SmallVector<BasicBlock*, 2> Successors(succ_begin(BB), succ_end(BB));
228 for (SmallVectorImpl<BasicBlock*>::iterator
229 II = Successors.begin(), IE = Successors.end(); II != IE; ++II)
230 if (pred_begin(*II) == pred_end(*II))
231 WorkList.insert(*II);
234 // Merge pairs of basic blocks with unconditional branches, connected by
236 if (EverMadeChange || MadeChange)
237 MadeChange |= EliminateFallThrough(F);
241 EverMadeChange |= MadeChange;
244 if (ModifiedDT && DT)
247 return EverMadeChange;
250 /// EliminateFallThrough - Merge basic blocks which are connected
251 /// by a single edge, where one of the basic blocks has a single successor
252 /// pointing to the other basic block, which has a single predecessor.
253 bool CodeGenPrepare::EliminateFallThrough(Function &F) {
254 bool Changed = false;
255 // Scan all of the blocks in the function, except for the entry block.
256 for (Function::iterator I = std::next(F.begin()), E = F.end(); I != E;) {
257 BasicBlock *BB = I++;
258 // If the destination block has a single pred, then this is a trivial
259 // edge, just collapse it.
260 BasicBlock *SinglePred = BB->getSinglePredecessor();
262 // Don't merge if BB's address is taken.
263 if (!SinglePred || SinglePred == BB || BB->hasAddressTaken()) continue;
265 BranchInst *Term = dyn_cast<BranchInst>(SinglePred->getTerminator());
266 if (Term && !Term->isConditional()) {
268 DEBUG(dbgs() << "To merge:\n"<< *SinglePred << "\n\n\n");
269 // Remember if SinglePred was the entry block of the function.
270 // If so, we will need to move BB back to the entry position.
271 bool isEntry = SinglePred == &SinglePred->getParent()->getEntryBlock();
272 MergeBasicBlockIntoOnlyPred(BB, this);
274 if (isEntry && BB != &BB->getParent()->getEntryBlock())
275 BB->moveBefore(&BB->getParent()->getEntryBlock());
277 // We have erased a block. Update the iterator.
284 /// EliminateMostlyEmptyBlocks - eliminate blocks that contain only PHI nodes,
285 /// debug info directives, and an unconditional branch. Passes before isel
286 /// (e.g. LSR/loopsimplify) often split edges in ways that are non-optimal for
287 /// isel. Start by eliminating these blocks so we can split them the way we
289 bool CodeGenPrepare::EliminateMostlyEmptyBlocks(Function &F) {
290 bool MadeChange = false;
291 // Note that this intentionally skips the entry block.
292 for (Function::iterator I = std::next(F.begin()), E = F.end(); I != E;) {
293 BasicBlock *BB = I++;
295 // If this block doesn't end with an uncond branch, ignore it.
296 BranchInst *BI = dyn_cast<BranchInst>(BB->getTerminator());
297 if (!BI || !BI->isUnconditional())
300 // If the instruction before the branch (skipping debug info) isn't a phi
301 // node, then other stuff is happening here.
302 BasicBlock::iterator BBI = BI;
303 if (BBI != BB->begin()) {
305 while (isa<DbgInfoIntrinsic>(BBI)) {
306 if (BBI == BB->begin())
310 if (!isa<DbgInfoIntrinsic>(BBI) && !isa<PHINode>(BBI))
314 // Do not break infinite loops.
315 BasicBlock *DestBB = BI->getSuccessor(0);
319 if (!CanMergeBlocks(BB, DestBB))
322 EliminateMostlyEmptyBlock(BB);
328 /// CanMergeBlocks - Return true if we can merge BB into DestBB if there is a
329 /// single uncond branch between them, and BB contains no other non-phi
331 bool CodeGenPrepare::CanMergeBlocks(const BasicBlock *BB,
332 const BasicBlock *DestBB) const {
333 // We only want to eliminate blocks whose phi nodes are used by phi nodes in
334 // the successor. If there are more complex condition (e.g. preheaders),
335 // don't mess around with them.
336 BasicBlock::const_iterator BBI = BB->begin();
337 while (const PHINode *PN = dyn_cast<PHINode>(BBI++)) {
338 for (const User *U : PN->users()) {
339 const Instruction *UI = cast<Instruction>(U);
340 if (UI->getParent() != DestBB || !isa<PHINode>(UI))
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 (UI->getParent() == DestBB) {
346 if (const PHINode *UPN = dyn_cast<PHINode>(UI))
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 /// SinkCast - Sink the specified cast instruction into its user blocks
468 static bool SinkCast(CastInst *CI) {
469 BasicBlock *DefBB = CI->getParent();
471 /// InsertedCasts - Only insert a cast in each block once.
472 DenseMap<BasicBlock*, CastInst*> InsertedCasts;
474 bool MadeChange = false;
475 for (Value::user_iterator UI = CI->user_begin(), E = CI->user_end();
477 Use &TheUse = UI.getUse();
478 Instruction *User = cast<Instruction>(*UI);
480 // Figure out which BB this cast is used in. For PHI's this is the
481 // appropriate predecessor block.
482 BasicBlock *UserBB = User->getParent();
483 if (PHINode *PN = dyn_cast<PHINode>(User)) {
484 UserBB = PN->getIncomingBlock(TheUse);
487 // Preincrement use iterator so we don't invalidate it.
490 // If this user is in the same block as the cast, don't change the cast.
491 if (UserBB == DefBB) continue;
493 // If we have already inserted a cast into this block, use it.
494 CastInst *&InsertedCast = InsertedCasts[UserBB];
497 BasicBlock::iterator InsertPt = UserBB->getFirstInsertionPt();
499 CastInst::Create(CI->getOpcode(), CI->getOperand(0), CI->getType(), "",
504 // Replace a use of the cast with a use of the new cast.
505 TheUse = InsertedCast;
509 // If we removed all uses, nuke the cast.
510 if (CI->use_empty()) {
511 CI->eraseFromParent();
518 /// OptimizeNoopCopyExpression - If the specified cast instruction is a noop
519 /// copy (e.g. it's casting from one pointer type to another, i32->i8 on PPC),
520 /// sink it into user blocks to reduce the number of virtual
521 /// registers that must be created and coalesced.
523 /// Return true if any changes are made.
525 static bool OptimizeNoopCopyExpression(CastInst *CI, const TargetLowering &TLI){
526 // If this is a noop copy,
527 EVT SrcVT = TLI.getValueType(CI->getOperand(0)->getType());
528 EVT DstVT = TLI.getValueType(CI->getType());
530 // This is an fp<->int conversion?
531 if (SrcVT.isInteger() != DstVT.isInteger())
534 // If this is an extension, it will be a zero or sign extension, which
536 if (SrcVT.bitsLT(DstVT)) return false;
538 // If these values will be promoted, find out what they will be promoted
539 // to. This helps us consider truncates on PPC as noop copies when they
541 if (TLI.getTypeAction(CI->getContext(), SrcVT) ==
542 TargetLowering::TypePromoteInteger)
543 SrcVT = TLI.getTypeToTransformTo(CI->getContext(), SrcVT);
544 if (TLI.getTypeAction(CI->getContext(), DstVT) ==
545 TargetLowering::TypePromoteInteger)
546 DstVT = TLI.getTypeToTransformTo(CI->getContext(), DstVT);
548 // If, after promotion, these are the same types, this is a noop copy.
555 /// OptimizeCmpExpression - sink the given CmpInst into user blocks to reduce
556 /// the number of virtual registers that must be created and coalesced. This is
557 /// a clear win except on targets with multiple condition code registers
558 /// (PowerPC), where it might lose; some adjustment may be wanted there.
560 /// Return true if any changes are made.
561 static bool OptimizeCmpExpression(CmpInst *CI) {
562 BasicBlock *DefBB = CI->getParent();
564 /// InsertedCmp - Only insert a cmp in each block once.
565 DenseMap<BasicBlock*, CmpInst*> InsertedCmps;
567 bool MadeChange = false;
568 for (Value::user_iterator UI = CI->user_begin(), E = CI->user_end();
570 Use &TheUse = UI.getUse();
571 Instruction *User = cast<Instruction>(*UI);
573 // Preincrement use iterator so we don't invalidate it.
576 // Don't bother for PHI nodes.
577 if (isa<PHINode>(User))
580 // Figure out which BB this cmp is used in.
581 BasicBlock *UserBB = User->getParent();
583 // If this user is in the same block as the cmp, don't change the cmp.
584 if (UserBB == DefBB) continue;
586 // If we have already inserted a cmp into this block, use it.
587 CmpInst *&InsertedCmp = InsertedCmps[UserBB];
590 BasicBlock::iterator InsertPt = UserBB->getFirstInsertionPt();
592 CmpInst::Create(CI->getOpcode(),
593 CI->getPredicate(), CI->getOperand(0),
594 CI->getOperand(1), "", InsertPt);
598 // Replace a use of the cmp with a use of the new cmp.
599 TheUse = InsertedCmp;
603 // If we removed all uses, nuke the cmp.
605 CI->eraseFromParent();
611 class CodeGenPrepareFortifiedLibCalls : public SimplifyFortifiedLibCalls {
613 void replaceCall(Value *With) override {
614 CI->replaceAllUsesWith(With);
615 CI->eraseFromParent();
617 bool isFoldable(unsigned SizeCIOp, unsigned, bool) const override {
618 if (ConstantInt *SizeCI =
619 dyn_cast<ConstantInt>(CI->getArgOperand(SizeCIOp)))
620 return SizeCI->isAllOnesValue();
624 } // end anonymous namespace
626 bool CodeGenPrepare::OptimizeCallInst(CallInst *CI) {
627 BasicBlock *BB = CI->getParent();
629 // Lower inline assembly if we can.
630 // If we found an inline asm expession, and if the target knows how to
631 // lower it to normal LLVM code, do so now.
632 if (TLI && isa<InlineAsm>(CI->getCalledValue())) {
633 if (TLI->ExpandInlineAsm(CI)) {
634 // Avoid invalidating the iterator.
635 CurInstIterator = BB->begin();
636 // Avoid processing instructions out of order, which could cause
637 // reuse before a value is defined.
641 // Sink address computing for memory operands into the block.
642 if (OptimizeInlineAsmInst(CI))
646 // Lower all uses of llvm.objectsize.*
647 IntrinsicInst *II = dyn_cast<IntrinsicInst>(CI);
648 if (II && II->getIntrinsicID() == Intrinsic::objectsize) {
649 bool Min = (cast<ConstantInt>(II->getArgOperand(1))->getZExtValue() == 1);
650 Type *ReturnTy = CI->getType();
651 Constant *RetVal = ConstantInt::get(ReturnTy, Min ? 0 : -1ULL);
653 // Substituting this can cause recursive simplifications, which can
654 // invalidate our iterator. Use a WeakVH to hold onto it in case this
656 WeakVH IterHandle(CurInstIterator);
658 replaceAndRecursivelySimplify(CI, RetVal, TLI ? TLI->getDataLayout() : 0,
659 TLInfo, ModifiedDT ? 0 : DT);
661 // If the iterator instruction was recursively deleted, start over at the
662 // start of the block.
663 if (IterHandle != CurInstIterator) {
664 CurInstIterator = BB->begin();
671 SmallVector<Value*, 2> PtrOps;
673 if (TLI->GetAddrModeArguments(II, PtrOps, AccessTy))
674 while (!PtrOps.empty())
675 if (OptimizeMemoryInst(II, PtrOps.pop_back_val(), AccessTy))
679 // From here on out we're working with named functions.
680 if (CI->getCalledFunction() == 0) return false;
682 // We'll need DataLayout from here on out.
683 const DataLayout *TD = TLI ? TLI->getDataLayout() : 0;
684 if (!TD) return false;
686 // Lower all default uses of _chk calls. This is very similar
687 // to what InstCombineCalls does, but here we are only lowering calls
688 // that have the default "don't know" as the objectsize. Anything else
689 // should be left alone.
690 CodeGenPrepareFortifiedLibCalls Simplifier;
691 return Simplifier.fold(CI, TD, TLInfo);
694 /// DupRetToEnableTailCallOpts - Look for opportunities to duplicate return
695 /// instructions to the predecessor to enable tail call optimizations. The
696 /// case it is currently looking for is:
699 /// %tmp0 = tail call i32 @f0()
702 /// %tmp1 = tail call i32 @f1()
705 /// %tmp2 = tail call i32 @f2()
708 /// %retval = phi i32 [ %tmp0, %bb0 ], [ %tmp1, %bb1 ], [ %tmp2, %bb2 ]
716 /// %tmp0 = tail call i32 @f0()
719 /// %tmp1 = tail call i32 @f1()
722 /// %tmp2 = tail call i32 @f2()
725 bool CodeGenPrepare::DupRetToEnableTailCallOpts(BasicBlock *BB) {
729 ReturnInst *RI = dyn_cast<ReturnInst>(BB->getTerminator());
734 BitCastInst *BCI = 0;
735 Value *V = RI->getReturnValue();
737 BCI = dyn_cast<BitCastInst>(V);
739 V = BCI->getOperand(0);
741 PN = dyn_cast<PHINode>(V);
746 if (PN && PN->getParent() != BB)
749 // It's not safe to eliminate the sign / zero extension of the return value.
750 // See llvm::isInTailCallPosition().
751 const Function *F = BB->getParent();
752 AttributeSet CallerAttrs = F->getAttributes();
753 if (CallerAttrs.hasAttribute(AttributeSet::ReturnIndex, Attribute::ZExt) ||
754 CallerAttrs.hasAttribute(AttributeSet::ReturnIndex, Attribute::SExt))
757 // Make sure there are no instructions between the PHI and return, or that the
758 // return is the first instruction in the block.
760 BasicBlock::iterator BI = BB->begin();
761 do { ++BI; } while (isa<DbgInfoIntrinsic>(BI));
763 // Also skip over the bitcast.
768 BasicBlock::iterator BI = BB->begin();
769 while (isa<DbgInfoIntrinsic>(BI)) ++BI;
774 /// Only dup the ReturnInst if the CallInst is likely to be emitted as a tail
776 SmallVector<CallInst*, 4> TailCalls;
778 for (unsigned I = 0, E = PN->getNumIncomingValues(); I != E; ++I) {
779 CallInst *CI = dyn_cast<CallInst>(PN->getIncomingValue(I));
780 // Make sure the phi value is indeed produced by the tail call.
781 if (CI && CI->hasOneUse() && CI->getParent() == PN->getIncomingBlock(I) &&
782 TLI->mayBeEmittedAsTailCall(CI))
783 TailCalls.push_back(CI);
786 SmallPtrSet<BasicBlock*, 4> VisitedBBs;
787 for (pred_iterator PI = pred_begin(BB), PE = pred_end(BB); PI != PE; ++PI) {
788 if (!VisitedBBs.insert(*PI))
791 BasicBlock::InstListType &InstList = (*PI)->getInstList();
792 BasicBlock::InstListType::reverse_iterator RI = InstList.rbegin();
793 BasicBlock::InstListType::reverse_iterator RE = InstList.rend();
794 do { ++RI; } while (RI != RE && isa<DbgInfoIntrinsic>(&*RI));
798 CallInst *CI = dyn_cast<CallInst>(&*RI);
799 if (CI && CI->use_empty() && TLI->mayBeEmittedAsTailCall(CI))
800 TailCalls.push_back(CI);
804 bool Changed = false;
805 for (unsigned i = 0, e = TailCalls.size(); i != e; ++i) {
806 CallInst *CI = TailCalls[i];
809 // Conservatively require the attributes of the call to match those of the
810 // return. Ignore noalias because it doesn't affect the call sequence.
811 AttributeSet CalleeAttrs = CS.getAttributes();
812 if (AttrBuilder(CalleeAttrs, AttributeSet::ReturnIndex).
813 removeAttribute(Attribute::NoAlias) !=
814 AttrBuilder(CalleeAttrs, AttributeSet::ReturnIndex).
815 removeAttribute(Attribute::NoAlias))
818 // Make sure the call instruction is followed by an unconditional branch to
820 BasicBlock *CallBB = CI->getParent();
821 BranchInst *BI = dyn_cast<BranchInst>(CallBB->getTerminator());
822 if (!BI || !BI->isUnconditional() || BI->getSuccessor(0) != BB)
825 // Duplicate the return into CallBB.
826 (void)FoldReturnIntoUncondBranch(RI, BB, CallBB);
827 ModifiedDT = Changed = true;
831 // If we eliminated all predecessors of the block, delete the block now.
832 if (Changed && !BB->hasAddressTaken() && pred_begin(BB) == pred_end(BB))
833 BB->eraseFromParent();
838 //===----------------------------------------------------------------------===//
839 // Memory Optimization
840 //===----------------------------------------------------------------------===//
844 /// ExtAddrMode - This is an extended version of TargetLowering::AddrMode
845 /// which holds actual Value*'s for register values.
846 struct ExtAddrMode : public TargetLowering::AddrMode {
849 ExtAddrMode() : BaseReg(0), ScaledReg(0) {}
850 void print(raw_ostream &OS) const;
853 bool operator==(const ExtAddrMode& O) const {
854 return (BaseReg == O.BaseReg) && (ScaledReg == O.ScaledReg) &&
855 (BaseGV == O.BaseGV) && (BaseOffs == O.BaseOffs) &&
856 (HasBaseReg == O.HasBaseReg) && (Scale == O.Scale);
861 static inline raw_ostream &operator<<(raw_ostream &OS, const ExtAddrMode &AM) {
867 void ExtAddrMode::print(raw_ostream &OS) const {
868 bool NeedPlus = false;
871 OS << (NeedPlus ? " + " : "")
873 BaseGV->printAsOperand(OS, /*PrintType=*/false);
878 OS << (NeedPlus ? " + " : "") << BaseOffs, NeedPlus = true;
881 OS << (NeedPlus ? " + " : "")
883 BaseReg->printAsOperand(OS, /*PrintType=*/false);
887 OS << (NeedPlus ? " + " : "")
889 ScaledReg->printAsOperand(OS, /*PrintType=*/false);
895 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
896 void ExtAddrMode::dump() const {
902 /// \brief This class provides transaction based operation on the IR.
903 /// Every change made through this class is recorded in the internal state and
904 /// can be undone (rollback) until commit is called.
905 class TypePromotionTransaction {
907 /// \brief This represents the common interface of the individual transaction.
908 /// Each class implements the logic for doing one specific modification on
909 /// the IR via the TypePromotionTransaction.
910 class TypePromotionAction {
912 /// The Instruction modified.
916 /// \brief Constructor of the action.
917 /// The constructor performs the related action on the IR.
918 TypePromotionAction(Instruction *Inst) : Inst(Inst) {}
920 virtual ~TypePromotionAction() {}
922 /// \brief Undo the modification done by this action.
923 /// When this method is called, the IR must be in the same state as it was
924 /// before this action was applied.
925 /// \pre Undoing the action works if and only if the IR is in the exact same
926 /// state as it was directly after this action was applied.
927 virtual void undo() = 0;
929 /// \brief Advocate every change made by this action.
930 /// When the results on the IR of the action are to be kept, it is important
931 /// to call this function, otherwise hidden information may be kept forever.
932 virtual void commit() {
933 // Nothing to be done, this action is not doing anything.
937 /// \brief Utility to remember the position of an instruction.
938 class InsertionHandler {
939 /// Position of an instruction.
940 /// Either an instruction:
941 /// - Is the first in a basic block: BB is used.
942 /// - Has a previous instructon: PrevInst is used.
944 Instruction *PrevInst;
947 /// Remember whether or not the instruction had a previous instruction.
948 bool HasPrevInstruction;
951 /// \brief Record the position of \p Inst.
952 InsertionHandler(Instruction *Inst) {
953 BasicBlock::iterator It = Inst;
954 HasPrevInstruction = (It != (Inst->getParent()->begin()));
955 if (HasPrevInstruction)
956 Point.PrevInst = --It;
958 Point.BB = Inst->getParent();
961 /// \brief Insert \p Inst at the recorded position.
962 void insert(Instruction *Inst) {
963 if (HasPrevInstruction) {
964 if (Inst->getParent())
965 Inst->removeFromParent();
966 Inst->insertAfter(Point.PrevInst);
968 Instruction *Position = Point.BB->getFirstInsertionPt();
969 if (Inst->getParent())
970 Inst->moveBefore(Position);
972 Inst->insertBefore(Position);
977 /// \brief Move an instruction before another.
978 class InstructionMoveBefore : public TypePromotionAction {
979 /// Original position of the instruction.
980 InsertionHandler Position;
983 /// \brief Move \p Inst before \p Before.
984 InstructionMoveBefore(Instruction *Inst, Instruction *Before)
985 : TypePromotionAction(Inst), Position(Inst) {
986 DEBUG(dbgs() << "Do: move: " << *Inst << "\nbefore: " << *Before << "\n");
987 Inst->moveBefore(Before);
990 /// \brief Move the instruction back to its original position.
991 void undo() override {
992 DEBUG(dbgs() << "Undo: moveBefore: " << *Inst << "\n");
993 Position.insert(Inst);
997 /// \brief Set the operand of an instruction with a new value.
998 class OperandSetter : public TypePromotionAction {
999 /// Original operand of the instruction.
1001 /// Index of the modified instruction.
1005 /// \brief Set \p Idx operand of \p Inst with \p NewVal.
1006 OperandSetter(Instruction *Inst, unsigned Idx, Value *NewVal)
1007 : TypePromotionAction(Inst), Idx(Idx) {
1008 DEBUG(dbgs() << "Do: setOperand: " << Idx << "\n"
1009 << "for:" << *Inst << "\n"
1010 << "with:" << *NewVal << "\n");
1011 Origin = Inst->getOperand(Idx);
1012 Inst->setOperand(Idx, NewVal);
1015 /// \brief Restore the original value of the instruction.
1016 void undo() override {
1017 DEBUG(dbgs() << "Undo: setOperand:" << Idx << "\n"
1018 << "for: " << *Inst << "\n"
1019 << "with: " << *Origin << "\n");
1020 Inst->setOperand(Idx, Origin);
1024 /// \brief Hide the operands of an instruction.
1025 /// Do as if this instruction was not using any of its operands.
1026 class OperandsHider : public TypePromotionAction {
1027 /// The list of original operands.
1028 SmallVector<Value *, 4> OriginalValues;
1031 /// \brief Remove \p Inst from the uses of the operands of \p Inst.
1032 OperandsHider(Instruction *Inst) : TypePromotionAction(Inst) {
1033 DEBUG(dbgs() << "Do: OperandsHider: " << *Inst << "\n");
1034 unsigned NumOpnds = Inst->getNumOperands();
1035 OriginalValues.reserve(NumOpnds);
1036 for (unsigned It = 0; It < NumOpnds; ++It) {
1037 // Save the current operand.
1038 Value *Val = Inst->getOperand(It);
1039 OriginalValues.push_back(Val);
1041 // We could use OperandSetter here, but that would implied an overhead
1042 // that we are not willing to pay.
1043 Inst->setOperand(It, UndefValue::get(Val->getType()));
1047 /// \brief Restore the original list of uses.
1048 void undo() override {
1049 DEBUG(dbgs() << "Undo: OperandsHider: " << *Inst << "\n");
1050 for (unsigned It = 0, EndIt = OriginalValues.size(); It != EndIt; ++It)
1051 Inst->setOperand(It, OriginalValues[It]);
1055 /// \brief Build a truncate instruction.
1056 class TruncBuilder : public TypePromotionAction {
1058 /// \brief Build a truncate instruction of \p Opnd producing a \p Ty
1060 /// trunc Opnd to Ty.
1061 TruncBuilder(Instruction *Opnd, Type *Ty) : TypePromotionAction(Opnd) {
1062 IRBuilder<> Builder(Opnd);
1063 Inst = cast<Instruction>(Builder.CreateTrunc(Opnd, Ty, "promoted"));
1064 DEBUG(dbgs() << "Do: TruncBuilder: " << *Inst << "\n");
1067 /// \brief Get the built instruction.
1068 Instruction *getBuiltInstruction() { return Inst; }
1070 /// \brief Remove the built instruction.
1071 void undo() override {
1072 DEBUG(dbgs() << "Undo: TruncBuilder: " << *Inst << "\n");
1073 Inst->eraseFromParent();
1077 /// \brief Build a sign extension instruction.
1078 class SExtBuilder : public TypePromotionAction {
1080 /// \brief Build a sign extension instruction of \p Opnd producing a \p Ty
1082 /// sext Opnd to Ty.
1083 SExtBuilder(Instruction *InsertPt, Value *Opnd, Type *Ty)
1084 : TypePromotionAction(Inst) {
1085 IRBuilder<> Builder(InsertPt);
1086 Inst = cast<Instruction>(Builder.CreateSExt(Opnd, Ty, "promoted"));
1087 DEBUG(dbgs() << "Do: SExtBuilder: " << *Inst << "\n");
1090 /// \brief Get the built instruction.
1091 Instruction *getBuiltInstruction() { return Inst; }
1093 /// \brief Remove the built instruction.
1094 void undo() override {
1095 DEBUG(dbgs() << "Undo: SExtBuilder: " << *Inst << "\n");
1096 Inst->eraseFromParent();
1100 /// \brief Mutate an instruction to another type.
1101 class TypeMutator : public TypePromotionAction {
1102 /// Record the original type.
1106 /// \brief Mutate the type of \p Inst into \p NewTy.
1107 TypeMutator(Instruction *Inst, Type *NewTy)
1108 : TypePromotionAction(Inst), OrigTy(Inst->getType()) {
1109 DEBUG(dbgs() << "Do: MutateType: " << *Inst << " with " << *NewTy
1111 Inst->mutateType(NewTy);
1114 /// \brief Mutate the instruction back to its original type.
1115 void undo() override {
1116 DEBUG(dbgs() << "Undo: MutateType: " << *Inst << " with " << *OrigTy
1118 Inst->mutateType(OrigTy);
1122 /// \brief Replace the uses of an instruction by another instruction.
1123 class UsesReplacer : public TypePromotionAction {
1124 /// Helper structure to keep track of the replaced uses.
1125 struct InstructionAndIdx {
1126 /// The instruction using the instruction.
1128 /// The index where this instruction is used for Inst.
1130 InstructionAndIdx(Instruction *Inst, unsigned Idx)
1131 : Inst(Inst), Idx(Idx) {}
1134 /// Keep track of the original uses (pair Instruction, Index).
1135 SmallVector<InstructionAndIdx, 4> OriginalUses;
1136 typedef SmallVectorImpl<InstructionAndIdx>::iterator use_iterator;
1139 /// \brief Replace all the use of \p Inst by \p New.
1140 UsesReplacer(Instruction *Inst, Value *New) : TypePromotionAction(Inst) {
1141 DEBUG(dbgs() << "Do: UsersReplacer: " << *Inst << " with " << *New
1143 // Record the original uses.
1144 for (Use &U : Inst->uses()) {
1145 Instruction *UserI = cast<Instruction>(U.getUser());
1146 OriginalUses.push_back(InstructionAndIdx(UserI, U.getOperandNo()));
1148 // Now, we can replace the uses.
1149 Inst->replaceAllUsesWith(New);
1152 /// \brief Reassign the original uses of Inst to Inst.
1153 void undo() override {
1154 DEBUG(dbgs() << "Undo: UsersReplacer: " << *Inst << "\n");
1155 for (use_iterator UseIt = OriginalUses.begin(),
1156 EndIt = OriginalUses.end();
1157 UseIt != EndIt; ++UseIt) {
1158 UseIt->Inst->setOperand(UseIt->Idx, Inst);
1163 /// \brief Remove an instruction from the IR.
1164 class InstructionRemover : public TypePromotionAction {
1165 /// Original position of the instruction.
1166 InsertionHandler Inserter;
1167 /// Helper structure to hide all the link to the instruction. In other
1168 /// words, this helps to do as if the instruction was removed.
1169 OperandsHider Hider;
1170 /// Keep track of the uses replaced, if any.
1171 UsesReplacer *Replacer;
1174 /// \brief Remove all reference of \p Inst and optinally replace all its
1176 /// \pre If !Inst->use_empty(), then New != NULL
1177 InstructionRemover(Instruction *Inst, Value *New = NULL)
1178 : TypePromotionAction(Inst), Inserter(Inst), Hider(Inst),
1181 Replacer = new UsesReplacer(Inst, New);
1182 DEBUG(dbgs() << "Do: InstructionRemover: " << *Inst << "\n");
1183 Inst->removeFromParent();
1186 ~InstructionRemover() { delete Replacer; }
1188 /// \brief Really remove the instruction.
1189 void commit() override { delete Inst; }
1191 /// \brief Resurrect the instruction and reassign it to the proper uses if
1192 /// new value was provided when build this action.
1193 void undo() override {
1194 DEBUG(dbgs() << "Undo: InstructionRemover: " << *Inst << "\n");
1195 Inserter.insert(Inst);
1203 /// Restoration point.
1204 /// The restoration point is a pointer to an action instead of an iterator
1205 /// because the iterator may be invalidated but not the pointer.
1206 typedef const TypePromotionAction *ConstRestorationPt;
1207 /// Advocate every changes made in that transaction.
1209 /// Undo all the changes made after the given point.
1210 void rollback(ConstRestorationPt Point);
1211 /// Get the current restoration point.
1212 ConstRestorationPt getRestorationPoint() const;
1214 /// \name API for IR modification with state keeping to support rollback.
1216 /// Same as Instruction::setOperand.
1217 void setOperand(Instruction *Inst, unsigned Idx, Value *NewVal);
1218 /// Same as Instruction::eraseFromParent.
1219 void eraseInstruction(Instruction *Inst, Value *NewVal = NULL);
1220 /// Same as Value::replaceAllUsesWith.
1221 void replaceAllUsesWith(Instruction *Inst, Value *New);
1222 /// Same as Value::mutateType.
1223 void mutateType(Instruction *Inst, Type *NewTy);
1224 /// Same as IRBuilder::createTrunc.
1225 Instruction *createTrunc(Instruction *Opnd, Type *Ty);
1226 /// Same as IRBuilder::createSExt.
1227 Instruction *createSExt(Instruction *Inst, Value *Opnd, Type *Ty);
1228 /// Same as Instruction::moveBefore.
1229 void moveBefore(Instruction *Inst, Instruction *Before);
1232 ~TypePromotionTransaction();
1235 /// The ordered list of actions made so far.
1236 SmallVector<TypePromotionAction *, 16> Actions;
1237 typedef SmallVectorImpl<TypePromotionAction *>::iterator CommitPt;
1240 void TypePromotionTransaction::setOperand(Instruction *Inst, unsigned Idx,
1243 new TypePromotionTransaction::OperandSetter(Inst, Idx, NewVal));
1246 void TypePromotionTransaction::eraseInstruction(Instruction *Inst,
1249 new TypePromotionTransaction::InstructionRemover(Inst, NewVal));
1252 void TypePromotionTransaction::replaceAllUsesWith(Instruction *Inst,
1254 Actions.push_back(new TypePromotionTransaction::UsesReplacer(Inst, New));
1257 void TypePromotionTransaction::mutateType(Instruction *Inst, Type *NewTy) {
1258 Actions.push_back(new TypePromotionTransaction::TypeMutator(Inst, NewTy));
1261 Instruction *TypePromotionTransaction::createTrunc(Instruction *Opnd,
1263 TruncBuilder *TB = new TruncBuilder(Opnd, Ty);
1264 Actions.push_back(TB);
1265 return TB->getBuiltInstruction();
1268 Instruction *TypePromotionTransaction::createSExt(Instruction *Inst,
1269 Value *Opnd, Type *Ty) {
1270 SExtBuilder *SB = new SExtBuilder(Inst, Opnd, Ty);
1271 Actions.push_back(SB);
1272 return SB->getBuiltInstruction();
1275 void TypePromotionTransaction::moveBefore(Instruction *Inst,
1276 Instruction *Before) {
1278 new TypePromotionTransaction::InstructionMoveBefore(Inst, Before));
1281 TypePromotionTransaction::ConstRestorationPt
1282 TypePromotionTransaction::getRestorationPoint() const {
1283 return Actions.rbegin() != Actions.rend() ? *Actions.rbegin() : NULL;
1286 void TypePromotionTransaction::commit() {
1287 for (CommitPt It = Actions.begin(), EndIt = Actions.end(); It != EndIt;
1295 void TypePromotionTransaction::rollback(
1296 TypePromotionTransaction::ConstRestorationPt Point) {
1297 while (!Actions.empty() && Point != (*Actions.rbegin())) {
1298 TypePromotionAction *Curr = Actions.pop_back_val();
1304 TypePromotionTransaction::~TypePromotionTransaction() {
1305 for (CommitPt It = Actions.begin(), EndIt = Actions.end(); It != EndIt; ++It)
1310 /// \brief A helper class for matching addressing modes.
1312 /// This encapsulates the logic for matching the target-legal addressing modes.
1313 class AddressingModeMatcher {
1314 SmallVectorImpl<Instruction*> &AddrModeInsts;
1315 const TargetLowering &TLI;
1317 /// AccessTy/MemoryInst - This is the type for the access (e.g. double) and
1318 /// the memory instruction that we're computing this address for.
1320 Instruction *MemoryInst;
1322 /// AddrMode - This is the addressing mode that we're building up. This is
1323 /// part of the return value of this addressing mode matching stuff.
1324 ExtAddrMode &AddrMode;
1326 /// The truncate instruction inserted by other CodeGenPrepare optimizations.
1327 const SetOfInstrs &InsertedTruncs;
1328 /// A map from the instructions to their type before promotion.
1329 InstrToOrigTy &PromotedInsts;
1330 /// The ongoing transaction where every action should be registered.
1331 TypePromotionTransaction &TPT;
1333 /// IgnoreProfitability - This is set to true when we should not do
1334 /// profitability checks. When true, IsProfitableToFoldIntoAddressingMode
1335 /// always returns true.
1336 bool IgnoreProfitability;
1338 AddressingModeMatcher(SmallVectorImpl<Instruction*> &AMI,
1339 const TargetLowering &T, Type *AT,
1340 Instruction *MI, ExtAddrMode &AM,
1341 const SetOfInstrs &InsertedTruncs,
1342 InstrToOrigTy &PromotedInsts,
1343 TypePromotionTransaction &TPT)
1344 : AddrModeInsts(AMI), TLI(T), AccessTy(AT), MemoryInst(MI), AddrMode(AM),
1345 InsertedTruncs(InsertedTruncs), PromotedInsts(PromotedInsts), TPT(TPT) {
1346 IgnoreProfitability = false;
1350 /// Match - Find the maximal addressing mode that a load/store of V can fold,
1351 /// give an access type of AccessTy. This returns a list of involved
1352 /// instructions in AddrModeInsts.
1353 /// \p InsertedTruncs The truncate instruction inserted by other
1356 /// \p PromotedInsts maps the instructions to their type before promotion.
1357 /// \p The ongoing transaction where every action should be registered.
1358 static ExtAddrMode Match(Value *V, Type *AccessTy,
1359 Instruction *MemoryInst,
1360 SmallVectorImpl<Instruction*> &AddrModeInsts,
1361 const TargetLowering &TLI,
1362 const SetOfInstrs &InsertedTruncs,
1363 InstrToOrigTy &PromotedInsts,
1364 TypePromotionTransaction &TPT) {
1367 bool Success = AddressingModeMatcher(AddrModeInsts, TLI, AccessTy,
1368 MemoryInst, Result, InsertedTruncs,
1369 PromotedInsts, TPT).MatchAddr(V, 0);
1370 (void)Success; assert(Success && "Couldn't select *anything*?");
1374 bool MatchScaledValue(Value *ScaleReg, int64_t Scale, unsigned Depth);
1375 bool MatchAddr(Value *V, unsigned Depth);
1376 bool MatchOperationAddr(User *Operation, unsigned Opcode, unsigned Depth,
1377 bool *MovedAway = NULL);
1378 bool IsProfitableToFoldIntoAddressingMode(Instruction *I,
1379 ExtAddrMode &AMBefore,
1380 ExtAddrMode &AMAfter);
1381 bool ValueAlreadyLiveAtInst(Value *Val, Value *KnownLive1, Value *KnownLive2);
1382 bool IsPromotionProfitable(unsigned MatchedSize, unsigned SizeWithPromotion,
1383 Value *PromotedOperand) const;
1386 /// MatchScaledValue - Try adding ScaleReg*Scale to the current addressing mode.
1387 /// Return true and update AddrMode if this addr mode is legal for the target,
1389 bool AddressingModeMatcher::MatchScaledValue(Value *ScaleReg, int64_t Scale,
1391 // If Scale is 1, then this is the same as adding ScaleReg to the addressing
1392 // mode. Just process that directly.
1394 return MatchAddr(ScaleReg, Depth);
1396 // If the scale is 0, it takes nothing to add this.
1400 // If we already have a scale of this value, we can add to it, otherwise, we
1401 // need an available scale field.
1402 if (AddrMode.Scale != 0 && AddrMode.ScaledReg != ScaleReg)
1405 ExtAddrMode TestAddrMode = AddrMode;
1407 // Add scale to turn X*4+X*3 -> X*7. This could also do things like
1408 // [A+B + A*7] -> [B+A*8].
1409 TestAddrMode.Scale += Scale;
1410 TestAddrMode.ScaledReg = ScaleReg;
1412 // If the new address isn't legal, bail out.
1413 if (!TLI.isLegalAddressingMode(TestAddrMode, AccessTy))
1416 // It was legal, so commit it.
1417 AddrMode = TestAddrMode;
1419 // Okay, we decided that we can add ScaleReg+Scale to AddrMode. Check now
1420 // to see if ScaleReg is actually X+C. If so, we can turn this into adding
1421 // X*Scale + C*Scale to addr mode.
1422 ConstantInt *CI = 0; Value *AddLHS = 0;
1423 if (isa<Instruction>(ScaleReg) && // not a constant expr.
1424 match(ScaleReg, m_Add(m_Value(AddLHS), m_ConstantInt(CI)))) {
1425 TestAddrMode.ScaledReg = AddLHS;
1426 TestAddrMode.BaseOffs += CI->getSExtValue()*TestAddrMode.Scale;
1428 // If this addressing mode is legal, commit it and remember that we folded
1429 // this instruction.
1430 if (TLI.isLegalAddressingMode(TestAddrMode, AccessTy)) {
1431 AddrModeInsts.push_back(cast<Instruction>(ScaleReg));
1432 AddrMode = TestAddrMode;
1437 // Otherwise, not (x+c)*scale, just return what we have.
1441 /// MightBeFoldableInst - This is a little filter, which returns true if an
1442 /// addressing computation involving I might be folded into a load/store
1443 /// accessing it. This doesn't need to be perfect, but needs to accept at least
1444 /// the set of instructions that MatchOperationAddr can.
1445 static bool MightBeFoldableInst(Instruction *I) {
1446 switch (I->getOpcode()) {
1447 case Instruction::BitCast:
1448 // Don't touch identity bitcasts.
1449 if (I->getType() == I->getOperand(0)->getType())
1451 return I->getType()->isPointerTy() || I->getType()->isIntegerTy();
1452 case Instruction::PtrToInt:
1453 // PtrToInt is always a noop, as we know that the int type is pointer sized.
1455 case Instruction::IntToPtr:
1456 // We know the input is intptr_t, so this is foldable.
1458 case Instruction::Add:
1460 case Instruction::Mul:
1461 case Instruction::Shl:
1462 // Can only handle X*C and X << C.
1463 return isa<ConstantInt>(I->getOperand(1));
1464 case Instruction::GetElementPtr:
1471 /// \brief Hepler class to perform type promotion.
1472 class TypePromotionHelper {
1473 /// \brief Utility function to check whether or not a sign extension of
1474 /// \p Inst with \p ConsideredSExtType can be moved through \p Inst by either
1475 /// using the operands of \p Inst or promoting \p Inst.
1476 /// In other words, check if:
1477 /// sext (Ty Inst opnd1 opnd2 ... opndN) to ConsideredSExtType.
1478 /// #1 Promotion applies:
1479 /// ConsideredSExtType Inst (sext opnd1 to ConsideredSExtType, ...).
1480 /// #2 Operand reuses:
1481 /// sext opnd1 to ConsideredSExtType.
1482 /// \p PromotedInsts maps the instructions to their type before promotion.
1483 static bool canGetThrough(const Instruction *Inst, Type *ConsideredSExtType,
1484 const InstrToOrigTy &PromotedInsts);
1486 /// \brief Utility function to determine if \p OpIdx should be promoted when
1487 /// promoting \p Inst.
1488 static bool shouldSExtOperand(const Instruction *Inst, int OpIdx) {
1489 if (isa<SelectInst>(Inst) && OpIdx == 0)
1494 /// \brief Utility function to promote the operand of \p SExt when this
1495 /// operand is a promotable trunc or sext.
1496 /// \p PromotedInsts maps the instructions to their type before promotion.
1497 /// \p CreatedInsts[out] contains how many non-free instructions have been
1498 /// created to promote the operand of SExt.
1499 /// Should never be called directly.
1500 /// \return The promoted value which is used instead of SExt.
1501 static Value *promoteOperandForTruncAndSExt(Instruction *SExt,
1502 TypePromotionTransaction &TPT,
1503 InstrToOrigTy &PromotedInsts,
1504 unsigned &CreatedInsts);
1506 /// \brief Utility function to promote the operand of \p SExt when this
1507 /// operand is promotable and is not a supported trunc or sext.
1508 /// \p PromotedInsts maps the instructions to their type before promotion.
1509 /// \p CreatedInsts[out] contains how many non-free instructions have been
1510 /// created to promote the operand of SExt.
1511 /// Should never be called directly.
1512 /// \return The promoted value which is used instead of SExt.
1513 static Value *promoteOperandForOther(Instruction *SExt,
1514 TypePromotionTransaction &TPT,
1515 InstrToOrigTy &PromotedInsts,
1516 unsigned &CreatedInsts);
1519 /// Type for the utility function that promotes the operand of SExt.
1520 typedef Value *(*Action)(Instruction *SExt, TypePromotionTransaction &TPT,
1521 InstrToOrigTy &PromotedInsts,
1522 unsigned &CreatedInsts);
1523 /// \brief Given a sign extend instruction \p SExt, return the approriate
1524 /// action to promote the operand of \p SExt instead of using SExt.
1525 /// \return NULL if no promotable action is possible with the current
1527 /// \p InsertedTruncs keeps track of all the truncate instructions inserted by
1528 /// the others CodeGenPrepare optimizations. This information is important
1529 /// because we do not want to promote these instructions as CodeGenPrepare
1530 /// will reinsert them later. Thus creating an infinite loop: create/remove.
1531 /// \p PromotedInsts maps the instructions to their type before promotion.
1532 static Action getAction(Instruction *SExt, const SetOfInstrs &InsertedTruncs,
1533 const TargetLowering &TLI,
1534 const InstrToOrigTy &PromotedInsts);
1537 bool TypePromotionHelper::canGetThrough(const Instruction *Inst,
1538 Type *ConsideredSExtType,
1539 const InstrToOrigTy &PromotedInsts) {
1540 // We can always get through sext.
1541 if (isa<SExtInst>(Inst))
1544 // We can get through binary operator, if it is legal. In other words, the
1545 // binary operator must have a nuw or nsw flag.
1546 const BinaryOperator *BinOp = dyn_cast<BinaryOperator>(Inst);
1547 if (BinOp && isa<OverflowingBinaryOperator>(BinOp) &&
1548 (BinOp->hasNoUnsignedWrap() || BinOp->hasNoSignedWrap()))
1551 // Check if we can do the following simplification.
1552 // sext(trunc(sext)) --> sext
1553 if (!isa<TruncInst>(Inst))
1556 Value *OpndVal = Inst->getOperand(0);
1557 // Check if we can use this operand in the sext.
1558 // If the type is larger than the result type of the sign extension,
1560 if (OpndVal->getType()->getIntegerBitWidth() >
1561 ConsideredSExtType->getIntegerBitWidth())
1564 // If the operand of the truncate is not an instruction, we will not have
1565 // any information on the dropped bits.
1566 // (Actually we could for constant but it is not worth the extra logic).
1567 Instruction *Opnd = dyn_cast<Instruction>(OpndVal);
1571 // Check if the source of the type is narrow enough.
1572 // I.e., check that trunc just drops sign extended bits.
1573 // #1 get the type of the operand.
1574 const Type *OpndType;
1575 InstrToOrigTy::const_iterator It = PromotedInsts.find(Opnd);
1576 if (It != PromotedInsts.end())
1577 OpndType = It->second;
1578 else if (isa<SExtInst>(Opnd))
1579 OpndType = cast<Instruction>(Opnd)->getOperand(0)->getType();
1583 // #2 check that the truncate just drop sign extended bits.
1584 if (Inst->getType()->getIntegerBitWidth() >= OpndType->getIntegerBitWidth())
1590 TypePromotionHelper::Action TypePromotionHelper::getAction(
1591 Instruction *SExt, const SetOfInstrs &InsertedTruncs,
1592 const TargetLowering &TLI, const InstrToOrigTy &PromotedInsts) {
1593 Instruction *SExtOpnd = dyn_cast<Instruction>(SExt->getOperand(0));
1594 Type *SExtTy = SExt->getType();
1595 // If the operand of the sign extension is not an instruction, we cannot
1597 // If it, check we can get through.
1598 if (!SExtOpnd || !canGetThrough(SExtOpnd, SExtTy, PromotedInsts))
1601 // Do not promote if the operand has been added by codegenprepare.
1602 // Otherwise, it means we are undoing an optimization that is likely to be
1603 // redone, thus causing potential infinite loop.
1604 if (isa<TruncInst>(SExtOpnd) && InsertedTruncs.count(SExtOpnd))
1607 // SExt or Trunc instructions.
1608 // Return the related handler.
1609 if (isa<SExtInst>(SExtOpnd) || isa<TruncInst>(SExtOpnd))
1610 return promoteOperandForTruncAndSExt;
1612 // Regular instruction.
1613 // Abort early if we will have to insert non-free instructions.
1614 if (!SExtOpnd->hasOneUse() &&
1615 !TLI.isTruncateFree(SExtTy, SExtOpnd->getType()))
1617 return promoteOperandForOther;
1620 Value *TypePromotionHelper::promoteOperandForTruncAndSExt(
1621 llvm::Instruction *SExt, TypePromotionTransaction &TPT,
1622 InstrToOrigTy &PromotedInsts, unsigned &CreatedInsts) {
1623 // By construction, the operand of SExt is an instruction. Otherwise we cannot
1624 // get through it and this method should not be called.
1625 Instruction *SExtOpnd = cast<Instruction>(SExt->getOperand(0));
1626 // Replace sext(trunc(opnd)) or sext(sext(opnd))
1628 TPT.setOperand(SExt, 0, SExtOpnd->getOperand(0));
1631 // Remove dead code.
1632 if (SExtOpnd->use_empty())
1633 TPT.eraseInstruction(SExtOpnd);
1635 // Check if the sext is still needed.
1636 if (SExt->getType() != SExt->getOperand(0)->getType())
1639 // At this point we have: sext ty opnd to ty.
1640 // Reassign the uses of SExt to the opnd and remove SExt.
1641 Value *NextVal = SExt->getOperand(0);
1642 TPT.eraseInstruction(SExt, NextVal);
1647 TypePromotionHelper::promoteOperandForOther(Instruction *SExt,
1648 TypePromotionTransaction &TPT,
1649 InstrToOrigTy &PromotedInsts,
1650 unsigned &CreatedInsts) {
1651 // By construction, the operand of SExt is an instruction. Otherwise we cannot
1652 // get through it and this method should not be called.
1653 Instruction *SExtOpnd = cast<Instruction>(SExt->getOperand(0));
1655 if (!SExtOpnd->hasOneUse()) {
1656 // SExtOpnd will be promoted.
1657 // All its uses, but SExt, will need to use a truncated value of the
1658 // promoted version.
1659 // Create the truncate now.
1660 Instruction *Trunc = TPT.createTrunc(SExt, SExtOpnd->getType());
1661 Trunc->removeFromParent();
1662 // Insert it just after the definition.
1663 Trunc->insertAfter(SExtOpnd);
1665 TPT.replaceAllUsesWith(SExtOpnd, Trunc);
1666 // Restore the operand of SExt (which has been replace by the previous call
1667 // to replaceAllUsesWith) to avoid creating a cycle trunc <-> sext.
1668 TPT.setOperand(SExt, 0, SExtOpnd);
1671 // Get through the Instruction:
1672 // 1. Update its type.
1673 // 2. Replace the uses of SExt by Inst.
1674 // 3. Sign extend each operand that needs to be sign extended.
1676 // Remember the original type of the instruction before promotion.
1677 // This is useful to know that the high bits are sign extended bits.
1678 PromotedInsts.insert(
1679 std::pair<Instruction *, Type *>(SExtOpnd, SExtOpnd->getType()));
1681 TPT.mutateType(SExtOpnd, SExt->getType());
1683 TPT.replaceAllUsesWith(SExt, SExtOpnd);
1685 Instruction *SExtForOpnd = SExt;
1687 DEBUG(dbgs() << "Propagate SExt to operands\n");
1688 for (int OpIdx = 0, EndOpIdx = SExtOpnd->getNumOperands(); OpIdx != EndOpIdx;
1690 DEBUG(dbgs() << "Operand:\n" << *(SExtOpnd->getOperand(OpIdx)) << '\n');
1691 if (SExtOpnd->getOperand(OpIdx)->getType() == SExt->getType() ||
1692 !shouldSExtOperand(SExtOpnd, OpIdx)) {
1693 DEBUG(dbgs() << "No need to propagate\n");
1696 // Check if we can statically sign extend the operand.
1697 Value *Opnd = SExtOpnd->getOperand(OpIdx);
1698 if (const ConstantInt *Cst = dyn_cast<ConstantInt>(Opnd)) {
1699 DEBUG(dbgs() << "Statically sign extend\n");
1702 ConstantInt::getSigned(SExt->getType(), Cst->getSExtValue()));
1705 // UndefValue are typed, so we have to statically sign extend them.
1706 if (isa<UndefValue>(Opnd)) {
1707 DEBUG(dbgs() << "Statically sign extend\n");
1708 TPT.setOperand(SExtOpnd, OpIdx, UndefValue::get(SExt->getType()));
1712 // Otherwise we have to explicity sign extend the operand.
1713 // Check if SExt was reused to sign extend an operand.
1715 // If yes, create a new one.
1716 DEBUG(dbgs() << "More operands to sext\n");
1717 SExtForOpnd = TPT.createSExt(SExt, Opnd, SExt->getType());
1721 TPT.setOperand(SExtForOpnd, 0, Opnd);
1723 // Move the sign extension before the insertion point.
1724 TPT.moveBefore(SExtForOpnd, SExtOpnd);
1725 TPT.setOperand(SExtOpnd, OpIdx, SExtForOpnd);
1726 // If more sext are required, new instructions will have to be created.
1729 if (SExtForOpnd == SExt) {
1730 DEBUG(dbgs() << "Sign extension is useless now\n");
1731 TPT.eraseInstruction(SExt);
1736 /// IsPromotionProfitable - Check whether or not promoting an instruction
1737 /// to a wider type was profitable.
1738 /// \p MatchedSize gives the number of instructions that have been matched
1739 /// in the addressing mode after the promotion was applied.
1740 /// \p SizeWithPromotion gives the number of created instructions for
1741 /// the promotion plus the number of instructions that have been
1742 /// matched in the addressing mode before the promotion.
1743 /// \p PromotedOperand is the value that has been promoted.
1744 /// \return True if the promotion is profitable, false otherwise.
1746 AddressingModeMatcher::IsPromotionProfitable(unsigned MatchedSize,
1747 unsigned SizeWithPromotion,
1748 Value *PromotedOperand) const {
1749 // We folded less instructions than what we created to promote the operand.
1750 // This is not profitable.
1751 if (MatchedSize < SizeWithPromotion)
1753 if (MatchedSize > SizeWithPromotion)
1755 // The promotion is neutral but it may help folding the sign extension in
1756 // loads for instance.
1757 // Check that we did not create an illegal instruction.
1758 Instruction *PromotedInst = dyn_cast<Instruction>(PromotedOperand);
1761 int ISDOpcode = TLI.InstructionOpcodeToISD(PromotedInst->getOpcode());
1762 // If the ISDOpcode is undefined, it was undefined before the promotion.
1765 // Otherwise, check if the promoted instruction is legal or not.
1766 return TLI.isOperationLegalOrCustom(ISDOpcode,
1767 EVT::getEVT(PromotedInst->getType()));
1770 /// MatchOperationAddr - Given an instruction or constant expr, see if we can
1771 /// fold the operation into the addressing mode. If so, update the addressing
1772 /// mode and return true, otherwise return false without modifying AddrMode.
1773 /// If \p MovedAway is not NULL, it contains the information of whether or
1774 /// not AddrInst has to be folded into the addressing mode on success.
1775 /// If \p MovedAway == true, \p AddrInst will not be part of the addressing
1776 /// because it has been moved away.
1777 /// Thus AddrInst must not be added in the matched instructions.
1778 /// This state can happen when AddrInst is a sext, since it may be moved away.
1779 /// Therefore, AddrInst may not be valid when MovedAway is true and it must
1780 /// not be referenced anymore.
1781 bool AddressingModeMatcher::MatchOperationAddr(User *AddrInst, unsigned Opcode,
1784 // Avoid exponential behavior on extremely deep expression trees.
1785 if (Depth >= 5) return false;
1787 // By default, all matched instructions stay in place.
1792 case Instruction::PtrToInt:
1793 // PtrToInt is always a noop, as we know that the int type is pointer sized.
1794 return MatchAddr(AddrInst->getOperand(0), Depth);
1795 case Instruction::IntToPtr:
1796 // This inttoptr is a no-op if the integer type is pointer sized.
1797 if (TLI.getValueType(AddrInst->getOperand(0)->getType()) ==
1798 TLI.getPointerTy(AddrInst->getType()->getPointerAddressSpace()))
1799 return MatchAddr(AddrInst->getOperand(0), Depth);
1801 case Instruction::BitCast:
1802 // BitCast is always a noop, and we can handle it as long as it is
1803 // int->int or pointer->pointer (we don't want int<->fp or something).
1804 if ((AddrInst->getOperand(0)->getType()->isPointerTy() ||
1805 AddrInst->getOperand(0)->getType()->isIntegerTy()) &&
1806 // Don't touch identity bitcasts. These were probably put here by LSR,
1807 // and we don't want to mess around with them. Assume it knows what it
1809 AddrInst->getOperand(0)->getType() != AddrInst->getType())
1810 return MatchAddr(AddrInst->getOperand(0), Depth);
1812 case Instruction::Add: {
1813 // Check to see if we can merge in the RHS then the LHS. If so, we win.
1814 ExtAddrMode BackupAddrMode = AddrMode;
1815 unsigned OldSize = AddrModeInsts.size();
1816 // Start a transaction at this point.
1817 // The LHS may match but not the RHS.
1818 // Therefore, we need a higher level restoration point to undo partially
1819 // matched operation.
1820 TypePromotionTransaction::ConstRestorationPt LastKnownGood =
1821 TPT.getRestorationPoint();
1823 if (MatchAddr(AddrInst->getOperand(1), Depth+1) &&
1824 MatchAddr(AddrInst->getOperand(0), Depth+1))
1827 // Restore the old addr mode info.
1828 AddrMode = BackupAddrMode;
1829 AddrModeInsts.resize(OldSize);
1830 TPT.rollback(LastKnownGood);
1832 // Otherwise this was over-aggressive. Try merging in the LHS then the RHS.
1833 if (MatchAddr(AddrInst->getOperand(0), Depth+1) &&
1834 MatchAddr(AddrInst->getOperand(1), Depth+1))
1837 // Otherwise we definitely can't merge the ADD in.
1838 AddrMode = BackupAddrMode;
1839 AddrModeInsts.resize(OldSize);
1840 TPT.rollback(LastKnownGood);
1843 //case Instruction::Or:
1844 // TODO: We can handle "Or Val, Imm" iff this OR is equivalent to an ADD.
1846 case Instruction::Mul:
1847 case Instruction::Shl: {
1848 // Can only handle X*C and X << C.
1849 ConstantInt *RHS = dyn_cast<ConstantInt>(AddrInst->getOperand(1));
1850 if (!RHS) return false;
1851 int64_t Scale = RHS->getSExtValue();
1852 if (Opcode == Instruction::Shl)
1853 Scale = 1LL << Scale;
1855 return MatchScaledValue(AddrInst->getOperand(0), Scale, Depth);
1857 case Instruction::GetElementPtr: {
1858 // Scan the GEP. We check it if it contains constant offsets and at most
1859 // one variable offset.
1860 int VariableOperand = -1;
1861 unsigned VariableScale = 0;
1863 int64_t ConstantOffset = 0;
1864 const DataLayout *TD = TLI.getDataLayout();
1865 gep_type_iterator GTI = gep_type_begin(AddrInst);
1866 for (unsigned i = 1, e = AddrInst->getNumOperands(); i != e; ++i, ++GTI) {
1867 if (StructType *STy = dyn_cast<StructType>(*GTI)) {
1868 const StructLayout *SL = TD->getStructLayout(STy);
1870 cast<ConstantInt>(AddrInst->getOperand(i))->getZExtValue();
1871 ConstantOffset += SL->getElementOffset(Idx);
1873 uint64_t TypeSize = TD->getTypeAllocSize(GTI.getIndexedType());
1874 if (ConstantInt *CI = dyn_cast<ConstantInt>(AddrInst->getOperand(i))) {
1875 ConstantOffset += CI->getSExtValue()*TypeSize;
1876 } else if (TypeSize) { // Scales of zero don't do anything.
1877 // We only allow one variable index at the moment.
1878 if (VariableOperand != -1)
1881 // Remember the variable index.
1882 VariableOperand = i;
1883 VariableScale = TypeSize;
1888 // A common case is for the GEP to only do a constant offset. In this case,
1889 // just add it to the disp field and check validity.
1890 if (VariableOperand == -1) {
1891 AddrMode.BaseOffs += ConstantOffset;
1892 if (ConstantOffset == 0 || TLI.isLegalAddressingMode(AddrMode, AccessTy)){
1893 // Check to see if we can fold the base pointer in too.
1894 if (MatchAddr(AddrInst->getOperand(0), Depth+1))
1897 AddrMode.BaseOffs -= ConstantOffset;
1901 // Save the valid addressing mode in case we can't match.
1902 ExtAddrMode BackupAddrMode = AddrMode;
1903 unsigned OldSize = AddrModeInsts.size();
1905 // See if the scale and offset amount is valid for this target.
1906 AddrMode.BaseOffs += ConstantOffset;
1908 // Match the base operand of the GEP.
1909 if (!MatchAddr(AddrInst->getOperand(0), Depth+1)) {
1910 // If it couldn't be matched, just stuff the value in a register.
1911 if (AddrMode.HasBaseReg) {
1912 AddrMode = BackupAddrMode;
1913 AddrModeInsts.resize(OldSize);
1916 AddrMode.HasBaseReg = true;
1917 AddrMode.BaseReg = AddrInst->getOperand(0);
1920 // Match the remaining variable portion of the GEP.
1921 if (!MatchScaledValue(AddrInst->getOperand(VariableOperand), VariableScale,
1923 // If it couldn't be matched, try stuffing the base into a register
1924 // instead of matching it, and retrying the match of the scale.
1925 AddrMode = BackupAddrMode;
1926 AddrModeInsts.resize(OldSize);
1927 if (AddrMode.HasBaseReg)
1929 AddrMode.HasBaseReg = true;
1930 AddrMode.BaseReg = AddrInst->getOperand(0);
1931 AddrMode.BaseOffs += ConstantOffset;
1932 if (!MatchScaledValue(AddrInst->getOperand(VariableOperand),
1933 VariableScale, Depth)) {
1934 // If even that didn't work, bail.
1935 AddrMode = BackupAddrMode;
1936 AddrModeInsts.resize(OldSize);
1943 case Instruction::SExt: {
1944 // Try to move this sext out of the way of the addressing mode.
1945 Instruction *SExt = cast<Instruction>(AddrInst);
1946 // Ask for a method for doing so.
1947 TypePromotionHelper::Action TPH = TypePromotionHelper::getAction(
1948 SExt, InsertedTruncs, TLI, PromotedInsts);
1952 TypePromotionTransaction::ConstRestorationPt LastKnownGood =
1953 TPT.getRestorationPoint();
1954 unsigned CreatedInsts = 0;
1955 Value *PromotedOperand = TPH(SExt, TPT, PromotedInsts, CreatedInsts);
1956 // SExt has been moved away.
1957 // Thus either it will be rematched later in the recursive calls or it is
1958 // gone. Anyway, we must not fold it into the addressing mode at this point.
1962 // addr = gep base, idx
1964 // promotedOpnd = sext opnd <- no match here
1965 // op = promoted_add promotedOpnd, 1 <- match (later in recursive calls)
1966 // addr = gep base, op <- match
1970 assert(PromotedOperand &&
1971 "TypePromotionHelper should have filtered out those cases");
1973 ExtAddrMode BackupAddrMode = AddrMode;
1974 unsigned OldSize = AddrModeInsts.size();
1976 if (!MatchAddr(PromotedOperand, Depth) ||
1977 !IsPromotionProfitable(AddrModeInsts.size(), OldSize + CreatedInsts,
1979 AddrMode = BackupAddrMode;
1980 AddrModeInsts.resize(OldSize);
1981 DEBUG(dbgs() << "Sign extension does not pay off: rollback\n");
1982 TPT.rollback(LastKnownGood);
1991 /// MatchAddr - If we can, try to add the value of 'Addr' into the current
1992 /// addressing mode. If Addr can't be added to AddrMode this returns false and
1993 /// leaves AddrMode unmodified. This assumes that Addr is either a pointer type
1994 /// or intptr_t for the target.
1996 bool AddressingModeMatcher::MatchAddr(Value *Addr, unsigned Depth) {
1997 // Start a transaction at this point that we will rollback if the matching
1999 TypePromotionTransaction::ConstRestorationPt LastKnownGood =
2000 TPT.getRestorationPoint();
2001 if (ConstantInt *CI = dyn_cast<ConstantInt>(Addr)) {
2002 // Fold in immediates if legal for the target.
2003 AddrMode.BaseOffs += CI->getSExtValue();
2004 if (TLI.isLegalAddressingMode(AddrMode, AccessTy))
2006 AddrMode.BaseOffs -= CI->getSExtValue();
2007 } else if (GlobalValue *GV = dyn_cast<GlobalValue>(Addr)) {
2008 // If this is a global variable, try to fold it into the addressing mode.
2009 if (AddrMode.BaseGV == 0) {
2010 AddrMode.BaseGV = GV;
2011 if (TLI.isLegalAddressingMode(AddrMode, AccessTy))
2013 AddrMode.BaseGV = 0;
2015 } else if (Instruction *I = dyn_cast<Instruction>(Addr)) {
2016 ExtAddrMode BackupAddrMode = AddrMode;
2017 unsigned OldSize = AddrModeInsts.size();
2019 // Check to see if it is possible to fold this operation.
2020 bool MovedAway = false;
2021 if (MatchOperationAddr(I, I->getOpcode(), Depth, &MovedAway)) {
2022 // This instruction may have been move away. If so, there is nothing
2026 // Okay, it's possible to fold this. Check to see if it is actually
2027 // *profitable* to do so. We use a simple cost model to avoid increasing
2028 // register pressure too much.
2029 if (I->hasOneUse() ||
2030 IsProfitableToFoldIntoAddressingMode(I, BackupAddrMode, AddrMode)) {
2031 AddrModeInsts.push_back(I);
2035 // It isn't profitable to do this, roll back.
2036 //cerr << "NOT FOLDING: " << *I;
2037 AddrMode = BackupAddrMode;
2038 AddrModeInsts.resize(OldSize);
2039 TPT.rollback(LastKnownGood);
2041 } else if (ConstantExpr *CE = dyn_cast<ConstantExpr>(Addr)) {
2042 if (MatchOperationAddr(CE, CE->getOpcode(), Depth))
2044 TPT.rollback(LastKnownGood);
2045 } else if (isa<ConstantPointerNull>(Addr)) {
2046 // Null pointer gets folded without affecting the addressing mode.
2050 // Worse case, the target should support [reg] addressing modes. :)
2051 if (!AddrMode.HasBaseReg) {
2052 AddrMode.HasBaseReg = true;
2053 AddrMode.BaseReg = Addr;
2054 // Still check for legality in case the target supports [imm] but not [i+r].
2055 if (TLI.isLegalAddressingMode(AddrMode, AccessTy))
2057 AddrMode.HasBaseReg = false;
2058 AddrMode.BaseReg = 0;
2061 // If the base register is already taken, see if we can do [r+r].
2062 if (AddrMode.Scale == 0) {
2064 AddrMode.ScaledReg = Addr;
2065 if (TLI.isLegalAddressingMode(AddrMode, AccessTy))
2068 AddrMode.ScaledReg = 0;
2071 TPT.rollback(LastKnownGood);
2075 /// IsOperandAMemoryOperand - Check to see if all uses of OpVal by the specified
2076 /// inline asm call are due to memory operands. If so, return true, otherwise
2078 static bool IsOperandAMemoryOperand(CallInst *CI, InlineAsm *IA, Value *OpVal,
2079 const TargetLowering &TLI) {
2080 TargetLowering::AsmOperandInfoVector TargetConstraints = TLI.ParseConstraints(ImmutableCallSite(CI));
2081 for (unsigned i = 0, e = TargetConstraints.size(); i != e; ++i) {
2082 TargetLowering::AsmOperandInfo &OpInfo = TargetConstraints[i];
2084 // Compute the constraint code and ConstraintType to use.
2085 TLI.ComputeConstraintToUse(OpInfo, SDValue());
2087 // If this asm operand is our Value*, and if it isn't an indirect memory
2088 // operand, we can't fold it!
2089 if (OpInfo.CallOperandVal == OpVal &&
2090 (OpInfo.ConstraintType != TargetLowering::C_Memory ||
2091 !OpInfo.isIndirect))
2098 /// FindAllMemoryUses - Recursively walk all the uses of I until we find a
2099 /// memory use. If we find an obviously non-foldable instruction, return true.
2100 /// Add the ultimately found memory instructions to MemoryUses.
2101 static bool FindAllMemoryUses(Instruction *I,
2102 SmallVectorImpl<std::pair<Instruction*,unsigned> > &MemoryUses,
2103 SmallPtrSet<Instruction*, 16> &ConsideredInsts,
2104 const TargetLowering &TLI) {
2105 // If we already considered this instruction, we're done.
2106 if (!ConsideredInsts.insert(I))
2109 // If this is an obviously unfoldable instruction, bail out.
2110 if (!MightBeFoldableInst(I))
2113 // Loop over all the uses, recursively processing them.
2114 for (Use &U : I->uses()) {
2115 Instruction *UserI = cast<Instruction>(U.getUser());
2117 if (LoadInst *LI = dyn_cast<LoadInst>(UserI)) {
2118 MemoryUses.push_back(std::make_pair(LI, U.getOperandNo()));
2122 if (StoreInst *SI = dyn_cast<StoreInst>(UserI)) {
2123 unsigned opNo = U.getOperandNo();
2124 if (opNo == 0) return true; // Storing addr, not into addr.
2125 MemoryUses.push_back(std::make_pair(SI, opNo));
2129 if (CallInst *CI = dyn_cast<CallInst>(UserI)) {
2130 InlineAsm *IA = dyn_cast<InlineAsm>(CI->getCalledValue());
2131 if (!IA) return true;
2133 // If this is a memory operand, we're cool, otherwise bail out.
2134 if (!IsOperandAMemoryOperand(CI, IA, I, TLI))
2139 if (FindAllMemoryUses(UserI, MemoryUses, ConsideredInsts, TLI))
2146 /// ValueAlreadyLiveAtInst - Retrn true if Val is already known to be live at
2147 /// the use site that we're folding it into. If so, there is no cost to
2148 /// include it in the addressing mode. KnownLive1 and KnownLive2 are two values
2149 /// that we know are live at the instruction already.
2150 bool AddressingModeMatcher::ValueAlreadyLiveAtInst(Value *Val,Value *KnownLive1,
2151 Value *KnownLive2) {
2152 // If Val is either of the known-live values, we know it is live!
2153 if (Val == 0 || Val == KnownLive1 || Val == KnownLive2)
2156 // All values other than instructions and arguments (e.g. constants) are live.
2157 if (!isa<Instruction>(Val) && !isa<Argument>(Val)) return true;
2159 // If Val is a constant sized alloca in the entry block, it is live, this is
2160 // true because it is just a reference to the stack/frame pointer, which is
2161 // live for the whole function.
2162 if (AllocaInst *AI = dyn_cast<AllocaInst>(Val))
2163 if (AI->isStaticAlloca())
2166 // Check to see if this value is already used in the memory instruction's
2167 // block. If so, it's already live into the block at the very least, so we
2168 // can reasonably fold it.
2169 return Val->isUsedInBasicBlock(MemoryInst->getParent());
2172 /// IsProfitableToFoldIntoAddressingMode - It is possible for the addressing
2173 /// mode of the machine to fold the specified instruction into a load or store
2174 /// that ultimately uses it. However, the specified instruction has multiple
2175 /// uses. Given this, it may actually increase register pressure to fold it
2176 /// into the load. For example, consider this code:
2180 /// use(Y) -> nonload/store
2184 /// In this case, Y has multiple uses, and can be folded into the load of Z
2185 /// (yielding load [X+2]). However, doing this will cause both "X" and "X+1" to
2186 /// be live at the use(Y) line. If we don't fold Y into load Z, we use one
2187 /// fewer register. Since Y can't be folded into "use(Y)" we don't increase the
2188 /// number of computations either.
2190 /// Note that this (like most of CodeGenPrepare) is just a rough heuristic. If
2191 /// X was live across 'load Z' for other reasons, we actually *would* want to
2192 /// fold the addressing mode in the Z case. This would make Y die earlier.
2193 bool AddressingModeMatcher::
2194 IsProfitableToFoldIntoAddressingMode(Instruction *I, ExtAddrMode &AMBefore,
2195 ExtAddrMode &AMAfter) {
2196 if (IgnoreProfitability) return true;
2198 // AMBefore is the addressing mode before this instruction was folded into it,
2199 // and AMAfter is the addressing mode after the instruction was folded. Get
2200 // the set of registers referenced by AMAfter and subtract out those
2201 // referenced by AMBefore: this is the set of values which folding in this
2202 // address extends the lifetime of.
2204 // Note that there are only two potential values being referenced here,
2205 // BaseReg and ScaleReg (global addresses are always available, as are any
2206 // folded immediates).
2207 Value *BaseReg = AMAfter.BaseReg, *ScaledReg = AMAfter.ScaledReg;
2209 // If the BaseReg or ScaledReg was referenced by the previous addrmode, their
2210 // lifetime wasn't extended by adding this instruction.
2211 if (ValueAlreadyLiveAtInst(BaseReg, AMBefore.BaseReg, AMBefore.ScaledReg))
2213 if (ValueAlreadyLiveAtInst(ScaledReg, AMBefore.BaseReg, AMBefore.ScaledReg))
2216 // If folding this instruction (and it's subexprs) didn't extend any live
2217 // ranges, we're ok with it.
2218 if (BaseReg == 0 && ScaledReg == 0)
2221 // If all uses of this instruction are ultimately load/store/inlineasm's,
2222 // check to see if their addressing modes will include this instruction. If
2223 // so, we can fold it into all uses, so it doesn't matter if it has multiple
2225 SmallVector<std::pair<Instruction*,unsigned>, 16> MemoryUses;
2226 SmallPtrSet<Instruction*, 16> ConsideredInsts;
2227 if (FindAllMemoryUses(I, MemoryUses, ConsideredInsts, TLI))
2228 return false; // Has a non-memory, non-foldable use!
2230 // Now that we know that all uses of this instruction are part of a chain of
2231 // computation involving only operations that could theoretically be folded
2232 // into a memory use, loop over each of these uses and see if they could
2233 // *actually* fold the instruction.
2234 SmallVector<Instruction*, 32> MatchedAddrModeInsts;
2235 for (unsigned i = 0, e = MemoryUses.size(); i != e; ++i) {
2236 Instruction *User = MemoryUses[i].first;
2237 unsigned OpNo = MemoryUses[i].second;
2239 // Get the access type of this use. If the use isn't a pointer, we don't
2240 // know what it accesses.
2241 Value *Address = User->getOperand(OpNo);
2242 if (!Address->getType()->isPointerTy())
2244 Type *AddressAccessTy = Address->getType()->getPointerElementType();
2246 // Do a match against the root of this address, ignoring profitability. This
2247 // will tell us if the addressing mode for the memory operation will
2248 // *actually* cover the shared instruction.
2250 TypePromotionTransaction::ConstRestorationPt LastKnownGood =
2251 TPT.getRestorationPoint();
2252 AddressingModeMatcher Matcher(MatchedAddrModeInsts, TLI, AddressAccessTy,
2253 MemoryInst, Result, InsertedTruncs,
2254 PromotedInsts, TPT);
2255 Matcher.IgnoreProfitability = true;
2256 bool Success = Matcher.MatchAddr(Address, 0);
2257 (void)Success; assert(Success && "Couldn't select *anything*?");
2259 // The match was to check the profitability, the changes made are not
2260 // part of the original matcher. Therefore, they should be dropped
2261 // otherwise the original matcher will not present the right state.
2262 TPT.rollback(LastKnownGood);
2264 // If the match didn't cover I, then it won't be shared by it.
2265 if (std::find(MatchedAddrModeInsts.begin(), MatchedAddrModeInsts.end(),
2266 I) == MatchedAddrModeInsts.end())
2269 MatchedAddrModeInsts.clear();
2275 } // end anonymous namespace
2277 /// IsNonLocalValue - Return true if the specified values are defined in a
2278 /// different basic block than BB.
2279 static bool IsNonLocalValue(Value *V, BasicBlock *BB) {
2280 if (Instruction *I = dyn_cast<Instruction>(V))
2281 return I->getParent() != BB;
2285 /// OptimizeMemoryInst - Load and Store Instructions often have
2286 /// addressing modes that can do significant amounts of computation. As such,
2287 /// instruction selection will try to get the load or store to do as much
2288 /// computation as possible for the program. The problem is that isel can only
2289 /// see within a single block. As such, we sink as much legal addressing mode
2290 /// stuff into the block as possible.
2292 /// This method is used to optimize both load/store and inline asms with memory
2294 bool CodeGenPrepare::OptimizeMemoryInst(Instruction *MemoryInst, Value *Addr,
2298 // Try to collapse single-value PHI nodes. This is necessary to undo
2299 // unprofitable PRE transformations.
2300 SmallVector<Value*, 8> worklist;
2301 SmallPtrSet<Value*, 16> Visited;
2302 worklist.push_back(Addr);
2304 // Use a worklist to iteratively look through PHI nodes, and ensure that
2305 // the addressing mode obtained from the non-PHI roots of the graph
2307 Value *Consensus = 0;
2308 unsigned NumUsesConsensus = 0;
2309 bool IsNumUsesConsensusValid = false;
2310 SmallVector<Instruction*, 16> AddrModeInsts;
2311 ExtAddrMode AddrMode;
2312 TypePromotionTransaction TPT;
2313 TypePromotionTransaction::ConstRestorationPt LastKnownGood =
2314 TPT.getRestorationPoint();
2315 while (!worklist.empty()) {
2316 Value *V = worklist.back();
2317 worklist.pop_back();
2319 // Break use-def graph loops.
2320 if (!Visited.insert(V)) {
2325 // For a PHI node, push all of its incoming values.
2326 if (PHINode *P = dyn_cast<PHINode>(V)) {
2327 for (unsigned i = 0, e = P->getNumIncomingValues(); i != e; ++i)
2328 worklist.push_back(P->getIncomingValue(i));
2332 // For non-PHIs, determine the addressing mode being computed.
2333 SmallVector<Instruction*, 16> NewAddrModeInsts;
2334 ExtAddrMode NewAddrMode = AddressingModeMatcher::Match(
2335 V, AccessTy, MemoryInst, NewAddrModeInsts, *TLI, InsertedTruncsSet,
2336 PromotedInsts, TPT);
2338 // This check is broken into two cases with very similar code to avoid using
2339 // getNumUses() as much as possible. Some values have a lot of uses, so
2340 // calling getNumUses() unconditionally caused a significant compile-time
2344 AddrMode = NewAddrMode;
2345 AddrModeInsts = NewAddrModeInsts;
2347 } else if (NewAddrMode == AddrMode) {
2348 if (!IsNumUsesConsensusValid) {
2349 NumUsesConsensus = Consensus->getNumUses();
2350 IsNumUsesConsensusValid = true;
2353 // Ensure that the obtained addressing mode is equivalent to that obtained
2354 // for all other roots of the PHI traversal. Also, when choosing one
2355 // such root as representative, select the one with the most uses in order
2356 // to keep the cost modeling heuristics in AddressingModeMatcher
2358 unsigned NumUses = V->getNumUses();
2359 if (NumUses > NumUsesConsensus) {
2361 NumUsesConsensus = NumUses;
2362 AddrModeInsts = NewAddrModeInsts;
2371 // If the addressing mode couldn't be determined, or if multiple different
2372 // ones were determined, bail out now.
2374 TPT.rollback(LastKnownGood);
2379 // Check to see if any of the instructions supersumed by this addr mode are
2380 // non-local to I's BB.
2381 bool AnyNonLocal = false;
2382 for (unsigned i = 0, e = AddrModeInsts.size(); i != e; ++i) {
2383 if (IsNonLocalValue(AddrModeInsts[i], MemoryInst->getParent())) {
2389 // If all the instructions matched are already in this BB, don't do anything.
2391 DEBUG(dbgs() << "CGP: Found local addrmode: " << AddrMode << "\n");
2395 // Insert this computation right after this user. Since our caller is
2396 // scanning from the top of the BB to the bottom, reuse of the expr are
2397 // guaranteed to happen later.
2398 IRBuilder<> Builder(MemoryInst);
2400 // Now that we determined the addressing expression we want to use and know
2401 // that we have to sink it into this block. Check to see if we have already
2402 // done this for some other load/store instr in this block. If so, reuse the
2404 Value *&SunkAddr = SunkAddrs[Addr];
2406 DEBUG(dbgs() << "CGP: Reusing nonlocal addrmode: " << AddrMode << " for "
2408 if (SunkAddr->getType() != Addr->getType())
2409 SunkAddr = Builder.CreateBitCast(SunkAddr, Addr->getType());
2411 DEBUG(dbgs() << "CGP: SINKING nonlocal addrmode: " << AddrMode << " for "
2413 Type *IntPtrTy = TLI->getDataLayout()->getIntPtrType(Addr->getType());
2416 // Start with the base register. Do this first so that subsequent address
2417 // matching finds it last, which will prevent it from trying to match it
2418 // as the scaled value in case it happens to be a mul. That would be
2419 // problematic if we've sunk a different mul for the scale, because then
2420 // we'd end up sinking both muls.
2421 if (AddrMode.BaseReg) {
2422 Value *V = AddrMode.BaseReg;
2423 if (V->getType()->isPointerTy())
2424 V = Builder.CreatePtrToInt(V, IntPtrTy, "sunkaddr");
2425 if (V->getType() != IntPtrTy)
2426 V = Builder.CreateIntCast(V, IntPtrTy, /*isSigned=*/true, "sunkaddr");
2430 // Add the scale value.
2431 if (AddrMode.Scale) {
2432 Value *V = AddrMode.ScaledReg;
2433 if (V->getType() == IntPtrTy) {
2435 } else if (V->getType()->isPointerTy()) {
2436 V = Builder.CreatePtrToInt(V, IntPtrTy, "sunkaddr");
2437 } else if (cast<IntegerType>(IntPtrTy)->getBitWidth() <
2438 cast<IntegerType>(V->getType())->getBitWidth()) {
2439 V = Builder.CreateTrunc(V, IntPtrTy, "sunkaddr");
2441 V = Builder.CreateSExt(V, IntPtrTy, "sunkaddr");
2443 if (AddrMode.Scale != 1)
2444 V = Builder.CreateMul(V, ConstantInt::get(IntPtrTy, AddrMode.Scale),
2447 Result = Builder.CreateAdd(Result, V, "sunkaddr");
2452 // Add in the BaseGV if present.
2453 if (AddrMode.BaseGV) {
2454 Value *V = Builder.CreatePtrToInt(AddrMode.BaseGV, IntPtrTy, "sunkaddr");
2456 Result = Builder.CreateAdd(Result, V, "sunkaddr");
2461 // Add in the Base Offset if present.
2462 if (AddrMode.BaseOffs) {
2463 Value *V = ConstantInt::get(IntPtrTy, AddrMode.BaseOffs);
2465 Result = Builder.CreateAdd(Result, V, "sunkaddr");
2471 SunkAddr = Constant::getNullValue(Addr->getType());
2473 SunkAddr = Builder.CreateIntToPtr(Result, Addr->getType(), "sunkaddr");
2476 MemoryInst->replaceUsesOfWith(Repl, SunkAddr);
2478 // If we have no uses, recursively delete the value and all dead instructions
2480 if (Repl->use_empty()) {
2481 // This can cause recursive deletion, which can invalidate our iterator.
2482 // Use a WeakVH to hold onto it in case this happens.
2483 WeakVH IterHandle(CurInstIterator);
2484 BasicBlock *BB = CurInstIterator->getParent();
2486 RecursivelyDeleteTriviallyDeadInstructions(Repl, TLInfo);
2488 if (IterHandle != CurInstIterator) {
2489 // If the iterator instruction was recursively deleted, start over at the
2490 // start of the block.
2491 CurInstIterator = BB->begin();
2499 /// OptimizeInlineAsmInst - If there are any memory operands, use
2500 /// OptimizeMemoryInst to sink their address computing into the block when
2501 /// possible / profitable.
2502 bool CodeGenPrepare::OptimizeInlineAsmInst(CallInst *CS) {
2503 bool MadeChange = false;
2505 TargetLowering::AsmOperandInfoVector
2506 TargetConstraints = TLI->ParseConstraints(CS);
2508 for (unsigned i = 0, e = TargetConstraints.size(); i != e; ++i) {
2509 TargetLowering::AsmOperandInfo &OpInfo = TargetConstraints[i];
2511 // Compute the constraint code and ConstraintType to use.
2512 TLI->ComputeConstraintToUse(OpInfo, SDValue());
2514 if (OpInfo.ConstraintType == TargetLowering::C_Memory &&
2515 OpInfo.isIndirect) {
2516 Value *OpVal = CS->getArgOperand(ArgNo++);
2517 MadeChange |= OptimizeMemoryInst(CS, OpVal, OpVal->getType());
2518 } else if (OpInfo.Type == InlineAsm::isInput)
2525 /// MoveExtToFormExtLoad - Move a zext or sext fed by a load into the same
2526 /// basic block as the load, unless conditions are unfavorable. This allows
2527 /// SelectionDAG to fold the extend into the load.
2529 bool CodeGenPrepare::MoveExtToFormExtLoad(Instruction *I) {
2530 // Look for a load being extended.
2531 LoadInst *LI = dyn_cast<LoadInst>(I->getOperand(0));
2532 if (!LI) return false;
2534 // If they're already in the same block, there's nothing to do.
2535 if (LI->getParent() == I->getParent())
2538 // If the load has other users and the truncate is not free, this probably
2539 // isn't worthwhile.
2540 if (!LI->hasOneUse() &&
2541 TLI && (TLI->isTypeLegal(TLI->getValueType(LI->getType())) ||
2542 !TLI->isTypeLegal(TLI->getValueType(I->getType()))) &&
2543 !TLI->isTruncateFree(I->getType(), LI->getType()))
2546 // Check whether the target supports casts folded into loads.
2548 if (isa<ZExtInst>(I))
2549 LType = ISD::ZEXTLOAD;
2551 assert(isa<SExtInst>(I) && "Unexpected ext type!");
2552 LType = ISD::SEXTLOAD;
2554 if (TLI && !TLI->isLoadExtLegal(LType, TLI->getValueType(LI->getType())))
2557 // Move the extend into the same block as the load, so that SelectionDAG
2559 I->removeFromParent();
2565 bool CodeGenPrepare::OptimizeExtUses(Instruction *I) {
2566 BasicBlock *DefBB = I->getParent();
2568 // If the result of a {s|z}ext and its source are both live out, rewrite all
2569 // other uses of the source with result of extension.
2570 Value *Src = I->getOperand(0);
2571 if (Src->hasOneUse())
2574 // Only do this xform if truncating is free.
2575 if (TLI && !TLI->isTruncateFree(I->getType(), Src->getType()))
2578 // Only safe to perform the optimization if the source is also defined in
2580 if (!isa<Instruction>(Src) || DefBB != cast<Instruction>(Src)->getParent())
2583 bool DefIsLiveOut = false;
2584 for (User *U : I->users()) {
2585 Instruction *UI = cast<Instruction>(U);
2587 // Figure out which BB this ext is used in.
2588 BasicBlock *UserBB = UI->getParent();
2589 if (UserBB == DefBB) continue;
2590 DefIsLiveOut = true;
2596 // Make sure none of the uses are PHI nodes.
2597 for (User *U : Src->users()) {
2598 Instruction *UI = cast<Instruction>(U);
2599 BasicBlock *UserBB = UI->getParent();
2600 if (UserBB == DefBB) continue;
2601 // Be conservative. We don't want this xform to end up introducing
2602 // reloads just before load / store instructions.
2603 if (isa<PHINode>(UI) || isa<LoadInst>(UI) || isa<StoreInst>(UI))
2607 // InsertedTruncs - Only insert one trunc in each block once.
2608 DenseMap<BasicBlock*, Instruction*> InsertedTruncs;
2610 bool MadeChange = false;
2611 for (Use &U : Src->uses()) {
2612 Instruction *User = cast<Instruction>(U.getUser());
2614 // Figure out which BB this ext is used in.
2615 BasicBlock *UserBB = User->getParent();
2616 if (UserBB == DefBB) continue;
2618 // Both src and def are live in this block. Rewrite the use.
2619 Instruction *&InsertedTrunc = InsertedTruncs[UserBB];
2621 if (!InsertedTrunc) {
2622 BasicBlock::iterator InsertPt = UserBB->getFirstInsertionPt();
2623 InsertedTrunc = new TruncInst(I, Src->getType(), "", InsertPt);
2624 InsertedTruncsSet.insert(InsertedTrunc);
2627 // Replace a use of the {s|z}ext source with a use of the result.
2636 /// isFormingBranchFromSelectProfitable - Returns true if a SelectInst should be
2637 /// turned into an explicit branch.
2638 static bool isFormingBranchFromSelectProfitable(SelectInst *SI) {
2639 // FIXME: This should use the same heuristics as IfConversion to determine
2640 // whether a select is better represented as a branch. This requires that
2641 // branch probability metadata is preserved for the select, which is not the
2644 CmpInst *Cmp = dyn_cast<CmpInst>(SI->getCondition());
2646 // If the branch is predicted right, an out of order CPU can avoid blocking on
2647 // the compare. Emit cmovs on compares with a memory operand as branches to
2648 // avoid stalls on the load from memory. If the compare has more than one use
2649 // there's probably another cmov or setcc around so it's not worth emitting a
2654 Value *CmpOp0 = Cmp->getOperand(0);
2655 Value *CmpOp1 = Cmp->getOperand(1);
2657 // We check that the memory operand has one use to avoid uses of the loaded
2658 // value directly after the compare, making branches unprofitable.
2659 return Cmp->hasOneUse() &&
2660 ((isa<LoadInst>(CmpOp0) && CmpOp0->hasOneUse()) ||
2661 (isa<LoadInst>(CmpOp1) && CmpOp1->hasOneUse()));
2665 /// If we have a SelectInst that will likely profit from branch prediction,
2666 /// turn it into a branch.
2667 bool CodeGenPrepare::OptimizeSelectInst(SelectInst *SI) {
2668 bool VectorCond = !SI->getCondition()->getType()->isIntegerTy(1);
2670 // Can we convert the 'select' to CF ?
2671 if (DisableSelectToBranch || OptSize || !TLI || VectorCond)
2674 TargetLowering::SelectSupportKind SelectKind;
2676 SelectKind = TargetLowering::VectorMaskSelect;
2677 else if (SI->getType()->isVectorTy())
2678 SelectKind = TargetLowering::ScalarCondVectorVal;
2680 SelectKind = TargetLowering::ScalarValSelect;
2682 // Do we have efficient codegen support for this kind of 'selects' ?
2683 if (TLI->isSelectSupported(SelectKind)) {
2684 // We have efficient codegen support for the select instruction.
2685 // Check if it is profitable to keep this 'select'.
2686 if (!TLI->isPredictableSelectExpensive() ||
2687 !isFormingBranchFromSelectProfitable(SI))
2693 // First, we split the block containing the select into 2 blocks.
2694 BasicBlock *StartBlock = SI->getParent();
2695 BasicBlock::iterator SplitPt = ++(BasicBlock::iterator(SI));
2696 BasicBlock *NextBlock = StartBlock->splitBasicBlock(SplitPt, "select.end");
2698 // Create a new block serving as the landing pad for the branch.
2699 BasicBlock *SmallBlock = BasicBlock::Create(SI->getContext(), "select.mid",
2700 NextBlock->getParent(), NextBlock);
2702 // Move the unconditional branch from the block with the select in it into our
2703 // landing pad block.
2704 StartBlock->getTerminator()->eraseFromParent();
2705 BranchInst::Create(NextBlock, SmallBlock);
2707 // Insert the real conditional branch based on the original condition.
2708 BranchInst::Create(NextBlock, SmallBlock, SI->getCondition(), SI);
2710 // The select itself is replaced with a PHI Node.
2711 PHINode *PN = PHINode::Create(SI->getType(), 2, "", NextBlock->begin());
2713 PN->addIncoming(SI->getTrueValue(), StartBlock);
2714 PN->addIncoming(SI->getFalseValue(), SmallBlock);
2715 SI->replaceAllUsesWith(PN);
2716 SI->eraseFromParent();
2718 // Instruct OptimizeBlock to skip to the next block.
2719 CurInstIterator = StartBlock->end();
2720 ++NumSelectsExpanded;
2724 static bool isBroadcastShuffle(ShuffleVectorInst *SVI) {
2725 SmallVector<int, 16> Mask(SVI->getShuffleMask());
2727 for (unsigned i = 0; i < Mask.size(); ++i) {
2728 if (SplatElem != -1 && Mask[i] != -1 && Mask[i] != SplatElem)
2730 SplatElem = Mask[i];
2736 /// Some targets have expensive vector shifts if the lanes aren't all the same
2737 /// (e.g. x86 only introduced "vpsllvd" and friends with AVX2). In these cases
2738 /// it's often worth sinking a shufflevector splat down to its use so that
2739 /// codegen can spot all lanes are identical.
2740 bool CodeGenPrepare::OptimizeShuffleVectorInst(ShuffleVectorInst *SVI) {
2741 BasicBlock *DefBB = SVI->getParent();
2743 // Only do this xform if variable vector shifts are particularly expensive.
2744 if (!TLI || !TLI->isVectorShiftByScalarCheap(SVI->getType()))
2747 // We only expect better codegen by sinking a shuffle if we can recognise a
2749 if (!isBroadcastShuffle(SVI))
2752 // InsertedShuffles - Only insert a shuffle in each block once.
2753 DenseMap<BasicBlock*, Instruction*> InsertedShuffles;
2755 bool MadeChange = false;
2756 for (User *U : SVI->users()) {
2757 Instruction *UI = cast<Instruction>(U);
2759 // Figure out which BB this ext is used in.
2760 BasicBlock *UserBB = UI->getParent();
2761 if (UserBB == DefBB) continue;
2763 // For now only apply this when the splat is used by a shift instruction.
2764 if (!UI->isShift()) continue;
2766 // Everything checks out, sink the shuffle if the user's block doesn't
2767 // already have a copy.
2768 Instruction *&InsertedShuffle = InsertedShuffles[UserBB];
2770 if (!InsertedShuffle) {
2771 BasicBlock::iterator InsertPt = UserBB->getFirstInsertionPt();
2772 InsertedShuffle = new ShuffleVectorInst(SVI->getOperand(0),
2774 SVI->getOperand(2), "", InsertPt);
2777 UI->replaceUsesOfWith(SVI, InsertedShuffle);
2781 // If we removed all uses, nuke the shuffle.
2782 if (SVI->use_empty()) {
2783 SVI->eraseFromParent();
2790 bool CodeGenPrepare::OptimizeInst(Instruction *I) {
2791 if (PHINode *P = dyn_cast<PHINode>(I)) {
2792 // It is possible for very late stage optimizations (such as SimplifyCFG)
2793 // to introduce PHI nodes too late to be cleaned up. If we detect such a
2794 // trivial PHI, go ahead and zap it here.
2795 if (Value *V = SimplifyInstruction(P, TLI ? TLI->getDataLayout() : 0,
2797 P->replaceAllUsesWith(V);
2798 P->eraseFromParent();
2805 if (CastInst *CI = dyn_cast<CastInst>(I)) {
2806 // If the source of the cast is a constant, then this should have
2807 // already been constant folded. The only reason NOT to constant fold
2808 // it is if something (e.g. LSR) was careful to place the constant
2809 // evaluation in a block other than then one that uses it (e.g. to hoist
2810 // the address of globals out of a loop). If this is the case, we don't
2811 // want to forward-subst the cast.
2812 if (isa<Constant>(CI->getOperand(0)))
2815 if (TLI && OptimizeNoopCopyExpression(CI, *TLI))
2818 if (isa<ZExtInst>(I) || isa<SExtInst>(I)) {
2819 /// Sink a zext or sext into its user blocks if the target type doesn't
2820 /// fit in one register
2821 if (TLI && TLI->getTypeAction(CI->getContext(),
2822 TLI->getValueType(CI->getType())) ==
2823 TargetLowering::TypeExpandInteger) {
2824 return SinkCast(CI);
2826 bool MadeChange = MoveExtToFormExtLoad(I);
2827 return MadeChange | OptimizeExtUses(I);
2833 if (CmpInst *CI = dyn_cast<CmpInst>(I))
2834 if (!TLI || !TLI->hasMultipleConditionRegisters())
2835 return OptimizeCmpExpression(CI);
2837 if (LoadInst *LI = dyn_cast<LoadInst>(I)) {
2839 return OptimizeMemoryInst(I, I->getOperand(0), LI->getType());
2843 if (StoreInst *SI = dyn_cast<StoreInst>(I)) {
2845 return OptimizeMemoryInst(I, SI->getOperand(1),
2846 SI->getOperand(0)->getType());
2850 if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(I)) {
2851 if (GEPI->hasAllZeroIndices()) {
2852 /// The GEP operand must be a pointer, so must its result -> BitCast
2853 Instruction *NC = new BitCastInst(GEPI->getOperand(0), GEPI->getType(),
2854 GEPI->getName(), GEPI);
2855 GEPI->replaceAllUsesWith(NC);
2856 GEPI->eraseFromParent();
2864 if (CallInst *CI = dyn_cast<CallInst>(I))
2865 return OptimizeCallInst(CI);
2867 if (SelectInst *SI = dyn_cast<SelectInst>(I))
2868 return OptimizeSelectInst(SI);
2870 if (ShuffleVectorInst *SVI = dyn_cast<ShuffleVectorInst>(I))
2871 return OptimizeShuffleVectorInst(SVI);
2876 // In this pass we look for GEP and cast instructions that are used
2877 // across basic blocks and rewrite them to improve basic-block-at-a-time
2879 bool CodeGenPrepare::OptimizeBlock(BasicBlock &BB) {
2881 bool MadeChange = false;
2883 CurInstIterator = BB.begin();
2884 while (CurInstIterator != BB.end())
2885 MadeChange |= OptimizeInst(CurInstIterator++);
2887 MadeChange |= DupRetToEnableTailCallOpts(&BB);
2892 // llvm.dbg.value is far away from the value then iSel may not be able
2893 // handle it properly. iSel will drop llvm.dbg.value if it can not
2894 // find a node corresponding to the value.
2895 bool CodeGenPrepare::PlaceDbgValues(Function &F) {
2896 bool MadeChange = false;
2897 for (Function::iterator I = F.begin(), E = F.end(); I != E; ++I) {
2898 Instruction *PrevNonDbgInst = NULL;
2899 for (BasicBlock::iterator BI = I->begin(), BE = I->end(); BI != BE;) {
2900 Instruction *Insn = BI; ++BI;
2901 DbgValueInst *DVI = dyn_cast<DbgValueInst>(Insn);
2903 PrevNonDbgInst = Insn;
2907 Instruction *VI = dyn_cast_or_null<Instruction>(DVI->getValue());
2908 if (VI && VI != PrevNonDbgInst && !VI->isTerminator()) {
2909 DEBUG(dbgs() << "Moving Debug Value before :\n" << *DVI << ' ' << *VI);
2910 DVI->removeFromParent();
2911 if (isa<PHINode>(VI))
2912 DVI->insertBefore(VI->getParent()->getFirstInsertionPt());
2914 DVI->insertAfter(VI);