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 SinkExtExpand(CastInst *I);
133 bool MoveExtToFormExtLoad(Instruction *I);
134 bool OptimizeExtUses(Instruction *I);
135 bool OptimizeSelectInst(SelectInst *SI);
136 bool OptimizeShuffleVectorInst(ShuffleVectorInst *SI);
137 bool DupRetToEnableTailCallOpts(BasicBlock *BB);
138 bool PlaceDbgValues(Function &F);
142 char CodeGenPrepare::ID = 0;
143 static void *initializeCodeGenPreparePassOnce(PassRegistry &Registry) {
144 initializeTargetLibraryInfoPass(Registry);
145 PassInfo *PI = new PassInfo(
146 "Optimize for code generation", "codegenprepare", &CodeGenPrepare::ID,
147 PassInfo::NormalCtor_t(callDefaultCtor<CodeGenPrepare>), false, false,
148 PassInfo::TargetMachineCtor_t(callTargetMachineCtor<CodeGenPrepare>));
149 Registry.registerPass(*PI, true);
153 void llvm::initializeCodeGenPreparePass(PassRegistry &Registry) {
154 CALL_ONCE_INITIALIZATION(initializeCodeGenPreparePassOnce)
157 FunctionPass *llvm::createCodeGenPreparePass(const TargetMachine *TM) {
158 return new CodeGenPrepare(TM);
161 bool CodeGenPrepare::runOnFunction(Function &F) {
162 bool EverMadeChange = false;
163 // Clear per function information.
164 InsertedTruncsSet.clear();
165 PromotedInsts.clear();
168 if (TM) TLI = TM->getTargetLowering();
169 TLInfo = &getAnalysis<TargetLibraryInfo>();
170 DominatorTreeWrapperPass *DTWP =
171 getAnalysisIfAvailable<DominatorTreeWrapperPass>();
172 DT = DTWP ? &DTWP->getDomTree() : 0;
173 OptSize = F.getAttributes().hasAttribute(AttributeSet::FunctionIndex,
174 Attribute::OptimizeForSize);
176 /// This optimization identifies DIV instructions that can be
177 /// profitably bypassed and carried out with a shorter, faster divide.
178 if (!OptSize && TLI && TLI->isSlowDivBypassed()) {
179 const DenseMap<unsigned int, unsigned int> &BypassWidths =
180 TLI->getBypassSlowDivWidths();
181 for (Function::iterator I = F.begin(); I != F.end(); I++)
182 EverMadeChange |= bypassSlowDivision(F, I, BypassWidths);
185 // Eliminate blocks that contain only PHI nodes and an
186 // unconditional branch.
187 EverMadeChange |= EliminateMostlyEmptyBlocks(F);
189 // llvm.dbg.value is far away from the value then iSel may not be able
190 // handle it properly. iSel will drop llvm.dbg.value if it can not
191 // find a node corresponding to the value.
192 EverMadeChange |= PlaceDbgValues(F);
194 bool MadeChange = true;
197 for (Function::iterator I = F.begin(); I != F.end(); ) {
198 BasicBlock *BB = I++;
199 MadeChange |= OptimizeBlock(*BB);
201 EverMadeChange |= MadeChange;
206 if (!DisableBranchOpts) {
208 SmallPtrSet<BasicBlock*, 8> WorkList;
209 for (Function::iterator BB = F.begin(), E = F.end(); BB != E; ++BB) {
210 SmallVector<BasicBlock*, 2> Successors(succ_begin(BB), succ_end(BB));
211 MadeChange |= ConstantFoldTerminator(BB, true);
212 if (!MadeChange) continue;
214 for (SmallVectorImpl<BasicBlock*>::iterator
215 II = Successors.begin(), IE = Successors.end(); II != IE; ++II)
216 if (pred_begin(*II) == pred_end(*II))
217 WorkList.insert(*II);
220 // Delete the dead blocks and any of their dead successors.
221 MadeChange |= !WorkList.empty();
222 while (!WorkList.empty()) {
223 BasicBlock *BB = *WorkList.begin();
225 SmallVector<BasicBlock*, 2> Successors(succ_begin(BB), succ_end(BB));
229 for (SmallVectorImpl<BasicBlock*>::iterator
230 II = Successors.begin(), IE = Successors.end(); II != IE; ++II)
231 if (pred_begin(*II) == pred_end(*II))
232 WorkList.insert(*II);
235 // Merge pairs of basic blocks with unconditional branches, connected by
237 if (EverMadeChange || MadeChange)
238 MadeChange |= EliminateFallThrough(F);
242 EverMadeChange |= MadeChange;
245 if (ModifiedDT && DT)
248 return EverMadeChange;
251 /// EliminateFallThrough - Merge basic blocks which are connected
252 /// by a single edge, where one of the basic blocks has a single successor
253 /// pointing to the other basic block, which has a single predecessor.
254 bool CodeGenPrepare::EliminateFallThrough(Function &F) {
255 bool Changed = false;
256 // Scan all of the blocks in the function, except for the entry block.
257 for (Function::iterator I = std::next(F.begin()), E = F.end(); I != E;) {
258 BasicBlock *BB = I++;
259 // If the destination block has a single pred, then this is a trivial
260 // edge, just collapse it.
261 BasicBlock *SinglePred = BB->getSinglePredecessor();
263 // Don't merge if BB's address is taken.
264 if (!SinglePred || SinglePred == BB || BB->hasAddressTaken()) continue;
266 BranchInst *Term = dyn_cast<BranchInst>(SinglePred->getTerminator());
267 if (Term && !Term->isConditional()) {
269 DEBUG(dbgs() << "To merge:\n"<< *SinglePred << "\n\n\n");
270 // Remember if SinglePred was the entry block of the function.
271 // If so, we will need to move BB back to the entry position.
272 bool isEntry = SinglePred == &SinglePred->getParent()->getEntryBlock();
273 MergeBasicBlockIntoOnlyPred(BB, this);
275 if (isEntry && BB != &BB->getParent()->getEntryBlock())
276 BB->moveBefore(&BB->getParent()->getEntryBlock());
278 // We have erased a block. Update the iterator.
285 /// EliminateMostlyEmptyBlocks - eliminate blocks that contain only PHI nodes,
286 /// debug info directives, and an unconditional branch. Passes before isel
287 /// (e.g. LSR/loopsimplify) often split edges in ways that are non-optimal for
288 /// isel. Start by eliminating these blocks so we can split them the way we
290 bool CodeGenPrepare::EliminateMostlyEmptyBlocks(Function &F) {
291 bool MadeChange = false;
292 // Note that this intentionally skips the entry block.
293 for (Function::iterator I = std::next(F.begin()), E = F.end(); I != E;) {
294 BasicBlock *BB = I++;
296 // If this block doesn't end with an uncond branch, ignore it.
297 BranchInst *BI = dyn_cast<BranchInst>(BB->getTerminator());
298 if (!BI || !BI->isUnconditional())
301 // If the instruction before the branch (skipping debug info) isn't a phi
302 // node, then other stuff is happening here.
303 BasicBlock::iterator BBI = BI;
304 if (BBI != BB->begin()) {
306 while (isa<DbgInfoIntrinsic>(BBI)) {
307 if (BBI == BB->begin())
311 if (!isa<DbgInfoIntrinsic>(BBI) && !isa<PHINode>(BBI))
315 // Do not break infinite loops.
316 BasicBlock *DestBB = BI->getSuccessor(0);
320 if (!CanMergeBlocks(BB, DestBB))
323 EliminateMostlyEmptyBlock(BB);
329 /// CanMergeBlocks - Return true if we can merge BB into DestBB if there is a
330 /// single uncond branch between them, and BB contains no other non-phi
332 bool CodeGenPrepare::CanMergeBlocks(const BasicBlock *BB,
333 const BasicBlock *DestBB) const {
334 // We only want to eliminate blocks whose phi nodes are used by phi nodes in
335 // the successor. If there are more complex condition (e.g. preheaders),
336 // don't mess around with them.
337 BasicBlock::const_iterator BBI = BB->begin();
338 while (const PHINode *PN = dyn_cast<PHINode>(BBI++)) {
339 for (const User *U : PN->users()) {
340 const Instruction *UI = cast<Instruction>(U);
341 if (UI->getParent() != DestBB || !isa<PHINode>(UI))
343 // If User is inside DestBB block and it is a PHINode then check
344 // incoming value. If incoming value is not from BB then this is
345 // a complex condition (e.g. preheaders) we want to avoid here.
346 if (UI->getParent() == DestBB) {
347 if (const PHINode *UPN = dyn_cast<PHINode>(UI))
348 for (unsigned I = 0, E = UPN->getNumIncomingValues(); I != E; ++I) {
349 Instruction *Insn = dyn_cast<Instruction>(UPN->getIncomingValue(I));
350 if (Insn && Insn->getParent() == BB &&
351 Insn->getParent() != UPN->getIncomingBlock(I))
358 // If BB and DestBB contain any common predecessors, then the phi nodes in BB
359 // and DestBB may have conflicting incoming values for the block. If so, we
360 // can't merge the block.
361 const PHINode *DestBBPN = dyn_cast<PHINode>(DestBB->begin());
362 if (!DestBBPN) return true; // no conflict.
364 // Collect the preds of BB.
365 SmallPtrSet<const BasicBlock*, 16> BBPreds;
366 if (const PHINode *BBPN = dyn_cast<PHINode>(BB->begin())) {
367 // It is faster to get preds from a PHI than with pred_iterator.
368 for (unsigned i = 0, e = BBPN->getNumIncomingValues(); i != e; ++i)
369 BBPreds.insert(BBPN->getIncomingBlock(i));
371 BBPreds.insert(pred_begin(BB), pred_end(BB));
374 // Walk the preds of DestBB.
375 for (unsigned i = 0, e = DestBBPN->getNumIncomingValues(); i != e; ++i) {
376 BasicBlock *Pred = DestBBPN->getIncomingBlock(i);
377 if (BBPreds.count(Pred)) { // Common predecessor?
378 BBI = DestBB->begin();
379 while (const PHINode *PN = dyn_cast<PHINode>(BBI++)) {
380 const Value *V1 = PN->getIncomingValueForBlock(Pred);
381 const Value *V2 = PN->getIncomingValueForBlock(BB);
383 // If V2 is a phi node in BB, look up what the mapped value will be.
384 if (const PHINode *V2PN = dyn_cast<PHINode>(V2))
385 if (V2PN->getParent() == BB)
386 V2 = V2PN->getIncomingValueForBlock(Pred);
388 // If there is a conflict, bail out.
389 if (V1 != V2) return false;
398 /// EliminateMostlyEmptyBlock - Eliminate a basic block that have only phi's and
399 /// an unconditional branch in it.
400 void CodeGenPrepare::EliminateMostlyEmptyBlock(BasicBlock *BB) {
401 BranchInst *BI = cast<BranchInst>(BB->getTerminator());
402 BasicBlock *DestBB = BI->getSuccessor(0);
404 DEBUG(dbgs() << "MERGING MOSTLY EMPTY BLOCKS - BEFORE:\n" << *BB << *DestBB);
406 // If the destination block has a single pred, then this is a trivial edge,
408 if (BasicBlock *SinglePred = DestBB->getSinglePredecessor()) {
409 if (SinglePred != DestBB) {
410 // Remember if SinglePred was the entry block of the function. If so, we
411 // will need to move BB back to the entry position.
412 bool isEntry = SinglePred == &SinglePred->getParent()->getEntryBlock();
413 MergeBasicBlockIntoOnlyPred(DestBB, this);
415 if (isEntry && BB != &BB->getParent()->getEntryBlock())
416 BB->moveBefore(&BB->getParent()->getEntryBlock());
418 DEBUG(dbgs() << "AFTER:\n" << *DestBB << "\n\n\n");
423 // Otherwise, we have multiple predecessors of BB. Update the PHIs in DestBB
424 // to handle the new incoming edges it is about to have.
426 for (BasicBlock::iterator BBI = DestBB->begin();
427 (PN = dyn_cast<PHINode>(BBI)); ++BBI) {
428 // Remove the incoming value for BB, and remember it.
429 Value *InVal = PN->removeIncomingValue(BB, false);
431 // Two options: either the InVal is a phi node defined in BB or it is some
432 // value that dominates BB.
433 PHINode *InValPhi = dyn_cast<PHINode>(InVal);
434 if (InValPhi && InValPhi->getParent() == BB) {
435 // Add all of the input values of the input PHI as inputs of this phi.
436 for (unsigned i = 0, e = InValPhi->getNumIncomingValues(); i != e; ++i)
437 PN->addIncoming(InValPhi->getIncomingValue(i),
438 InValPhi->getIncomingBlock(i));
440 // Otherwise, add one instance of the dominating value for each edge that
441 // we will be adding.
442 if (PHINode *BBPN = dyn_cast<PHINode>(BB->begin())) {
443 for (unsigned i = 0, e = BBPN->getNumIncomingValues(); i != e; ++i)
444 PN->addIncoming(InVal, BBPN->getIncomingBlock(i));
446 for (pred_iterator PI = pred_begin(BB), E = pred_end(BB); PI != E; ++PI)
447 PN->addIncoming(InVal, *PI);
452 // The PHIs are now updated, change everything that refers to BB to use
453 // DestBB and remove BB.
454 BB->replaceAllUsesWith(DestBB);
455 if (DT && !ModifiedDT) {
456 BasicBlock *BBIDom = DT->getNode(BB)->getIDom()->getBlock();
457 BasicBlock *DestBBIDom = DT->getNode(DestBB)->getIDom()->getBlock();
458 BasicBlock *NewIDom = DT->findNearestCommonDominator(BBIDom, DestBBIDom);
459 DT->changeImmediateDominator(DestBB, NewIDom);
462 BB->eraseFromParent();
465 DEBUG(dbgs() << "AFTER:\n" << *DestBB << "\n\n\n");
468 /// SinkCast - Sink the specified cast instruction into its user blocks
469 static bool SinkCast(CastInst *CI) {
470 BasicBlock *DefBB = CI->getParent();
472 /// InsertedCasts - Only insert a cast in each block once.
473 DenseMap<BasicBlock*, CastInst*> InsertedCasts;
475 bool MadeChange = false;
476 for (Value::user_iterator UI = CI->user_begin(), E = CI->user_end();
478 Use &TheUse = UI.getUse();
479 Instruction *User = cast<Instruction>(*UI);
481 // Figure out which BB this cast is used in. For PHI's this is the
482 // appropriate predecessor block.
483 BasicBlock *UserBB = User->getParent();
484 if (PHINode *PN = dyn_cast<PHINode>(User)) {
485 UserBB = PN->getIncomingBlock(TheUse);
488 // Preincrement use iterator so we don't invalidate it.
491 // If this user is in the same block as the cast, don't change the cast.
492 if (UserBB == DefBB) continue;
494 // If we have already inserted a cast into this block, use it.
495 CastInst *&InsertedCast = InsertedCasts[UserBB];
498 BasicBlock::iterator InsertPt = UserBB->getFirstInsertionPt();
500 CastInst::Create(CI->getOpcode(), CI->getOperand(0), CI->getType(), "",
505 // Replace a use of the cast with a use of the new cast.
506 TheUse = InsertedCast;
510 // If we removed all uses, nuke the cast.
511 if (CI->use_empty()) {
512 CI->eraseFromParent();
519 /// OptimizeNoopCopyExpression - If the specified cast instruction is a noop
520 /// copy (e.g. it's casting from one pointer type to another, i32->i8 on PPC),
521 /// sink it into user blocks to reduce the number of virtual
522 /// registers that must be created and coalesced.
524 /// Return true if any changes are made.
526 static bool OptimizeNoopCopyExpression(CastInst *CI, const TargetLowering &TLI){
527 // If this is a noop copy,
528 EVT SrcVT = TLI.getValueType(CI->getOperand(0)->getType());
529 EVT DstVT = TLI.getValueType(CI->getType());
531 // This is an fp<->int conversion?
532 if (SrcVT.isInteger() != DstVT.isInteger())
535 // If this is an extension, it will be a zero or sign extension, which
537 if (SrcVT.bitsLT(DstVT)) return false;
539 // If these values will be promoted, find out what they will be promoted
540 // to. This helps us consider truncates on PPC as noop copies when they
542 if (TLI.getTypeAction(CI->getContext(), SrcVT) ==
543 TargetLowering::TypePromoteInteger)
544 SrcVT = TLI.getTypeToTransformTo(CI->getContext(), SrcVT);
545 if (TLI.getTypeAction(CI->getContext(), DstVT) ==
546 TargetLowering::TypePromoteInteger)
547 DstVT = TLI.getTypeToTransformTo(CI->getContext(), DstVT);
549 // If, after promotion, these are the same types, this is a noop copy.
556 /// OptimizeCmpExpression - sink the given CmpInst into user blocks to reduce
557 /// the number of virtual registers that must be created and coalesced. This is
558 /// a clear win except on targets with multiple condition code registers
559 /// (PowerPC), where it might lose; some adjustment may be wanted there.
561 /// Return true if any changes are made.
562 static bool OptimizeCmpExpression(CmpInst *CI) {
563 BasicBlock *DefBB = CI->getParent();
565 /// InsertedCmp - Only insert a cmp in each block once.
566 DenseMap<BasicBlock*, CmpInst*> InsertedCmps;
568 bool MadeChange = false;
569 for (Value::user_iterator UI = CI->user_begin(), E = CI->user_end();
571 Use &TheUse = UI.getUse();
572 Instruction *User = cast<Instruction>(*UI);
574 // Preincrement use iterator so we don't invalidate it.
577 // Don't bother for PHI nodes.
578 if (isa<PHINode>(User))
581 // Figure out which BB this cmp is used in.
582 BasicBlock *UserBB = User->getParent();
584 // If this user is in the same block as the cmp, don't change the cmp.
585 if (UserBB == DefBB) continue;
587 // If we have already inserted a cmp into this block, use it.
588 CmpInst *&InsertedCmp = InsertedCmps[UserBB];
591 BasicBlock::iterator InsertPt = UserBB->getFirstInsertionPt();
593 CmpInst::Create(CI->getOpcode(),
594 CI->getPredicate(), CI->getOperand(0),
595 CI->getOperand(1), "", InsertPt);
599 // Replace a use of the cmp with a use of the new cmp.
600 TheUse = InsertedCmp;
604 // If we removed all uses, nuke the cmp.
606 CI->eraseFromParent();
612 class CodeGenPrepareFortifiedLibCalls : public SimplifyFortifiedLibCalls {
614 void replaceCall(Value *With) override {
615 CI->replaceAllUsesWith(With);
616 CI->eraseFromParent();
618 bool isFoldable(unsigned SizeCIOp, unsigned, bool) const override {
619 if (ConstantInt *SizeCI =
620 dyn_cast<ConstantInt>(CI->getArgOperand(SizeCIOp)))
621 return SizeCI->isAllOnesValue();
625 } // end anonymous namespace
627 bool CodeGenPrepare::OptimizeCallInst(CallInst *CI) {
628 BasicBlock *BB = CI->getParent();
630 // Lower inline assembly if we can.
631 // If we found an inline asm expession, and if the target knows how to
632 // lower it to normal LLVM code, do so now.
633 if (TLI && isa<InlineAsm>(CI->getCalledValue())) {
634 if (TLI->ExpandInlineAsm(CI)) {
635 // Avoid invalidating the iterator.
636 CurInstIterator = BB->begin();
637 // Avoid processing instructions out of order, which could cause
638 // reuse before a value is defined.
642 // Sink address computing for memory operands into the block.
643 if (OptimizeInlineAsmInst(CI))
647 // Lower all uses of llvm.objectsize.*
648 IntrinsicInst *II = dyn_cast<IntrinsicInst>(CI);
649 if (II && II->getIntrinsicID() == Intrinsic::objectsize) {
650 bool Min = (cast<ConstantInt>(II->getArgOperand(1))->getZExtValue() == 1);
651 Type *ReturnTy = CI->getType();
652 Constant *RetVal = ConstantInt::get(ReturnTy, Min ? 0 : -1ULL);
654 // Substituting this can cause recursive simplifications, which can
655 // invalidate our iterator. Use a WeakVH to hold onto it in case this
657 WeakVH IterHandle(CurInstIterator);
659 replaceAndRecursivelySimplify(CI, RetVal, TLI ? TLI->getDataLayout() : 0,
660 TLInfo, ModifiedDT ? 0 : DT);
662 // If the iterator instruction was recursively deleted, start over at the
663 // start of the block.
664 if (IterHandle != CurInstIterator) {
665 CurInstIterator = BB->begin();
672 SmallVector<Value*, 2> PtrOps;
674 if (TLI->GetAddrModeArguments(II, PtrOps, AccessTy))
675 while (!PtrOps.empty())
676 if (OptimizeMemoryInst(II, PtrOps.pop_back_val(), AccessTy))
680 // From here on out we're working with named functions.
681 if (CI->getCalledFunction() == 0) return false;
683 // We'll need DataLayout from here on out.
684 const DataLayout *TD = TLI ? TLI->getDataLayout() : 0;
685 if (!TD) return false;
687 // Lower all default uses of _chk calls. This is very similar
688 // to what InstCombineCalls does, but here we are only lowering calls
689 // that have the default "don't know" as the objectsize. Anything else
690 // should be left alone.
691 CodeGenPrepareFortifiedLibCalls Simplifier;
692 return Simplifier.fold(CI, TD, TLInfo);
695 /// DupRetToEnableTailCallOpts - Look for opportunities to duplicate return
696 /// instructions to the predecessor to enable tail call optimizations. The
697 /// case it is currently looking for is:
700 /// %tmp0 = tail call i32 @f0()
703 /// %tmp1 = tail call i32 @f1()
706 /// %tmp2 = tail call i32 @f2()
709 /// %retval = phi i32 [ %tmp0, %bb0 ], [ %tmp1, %bb1 ], [ %tmp2, %bb2 ]
717 /// %tmp0 = tail call i32 @f0()
720 /// %tmp1 = tail call i32 @f1()
723 /// %tmp2 = tail call i32 @f2()
726 bool CodeGenPrepare::DupRetToEnableTailCallOpts(BasicBlock *BB) {
730 ReturnInst *RI = dyn_cast<ReturnInst>(BB->getTerminator());
735 BitCastInst *BCI = 0;
736 Value *V = RI->getReturnValue();
738 BCI = dyn_cast<BitCastInst>(V);
740 V = BCI->getOperand(0);
742 PN = dyn_cast<PHINode>(V);
747 if (PN && PN->getParent() != BB)
750 // It's not safe to eliminate the sign / zero extension of the return value.
751 // See llvm::isInTailCallPosition().
752 const Function *F = BB->getParent();
753 AttributeSet CallerAttrs = F->getAttributes();
754 if (CallerAttrs.hasAttribute(AttributeSet::ReturnIndex, Attribute::ZExt) ||
755 CallerAttrs.hasAttribute(AttributeSet::ReturnIndex, Attribute::SExt))
758 // Make sure there are no instructions between the PHI and return, or that the
759 // return is the first instruction in the block.
761 BasicBlock::iterator BI = BB->begin();
762 do { ++BI; } while (isa<DbgInfoIntrinsic>(BI));
764 // Also skip over the bitcast.
769 BasicBlock::iterator BI = BB->begin();
770 while (isa<DbgInfoIntrinsic>(BI)) ++BI;
775 /// Only dup the ReturnInst if the CallInst is likely to be emitted as a tail
777 SmallVector<CallInst*, 4> TailCalls;
779 for (unsigned I = 0, E = PN->getNumIncomingValues(); I != E; ++I) {
780 CallInst *CI = dyn_cast<CallInst>(PN->getIncomingValue(I));
781 // Make sure the phi value is indeed produced by the tail call.
782 if (CI && CI->hasOneUse() && CI->getParent() == PN->getIncomingBlock(I) &&
783 TLI->mayBeEmittedAsTailCall(CI))
784 TailCalls.push_back(CI);
787 SmallPtrSet<BasicBlock*, 4> VisitedBBs;
788 for (pred_iterator PI = pred_begin(BB), PE = pred_end(BB); PI != PE; ++PI) {
789 if (!VisitedBBs.insert(*PI))
792 BasicBlock::InstListType &InstList = (*PI)->getInstList();
793 BasicBlock::InstListType::reverse_iterator RI = InstList.rbegin();
794 BasicBlock::InstListType::reverse_iterator RE = InstList.rend();
795 do { ++RI; } while (RI != RE && isa<DbgInfoIntrinsic>(&*RI));
799 CallInst *CI = dyn_cast<CallInst>(&*RI);
800 if (CI && CI->use_empty() && TLI->mayBeEmittedAsTailCall(CI))
801 TailCalls.push_back(CI);
805 bool Changed = false;
806 for (unsigned i = 0, e = TailCalls.size(); i != e; ++i) {
807 CallInst *CI = TailCalls[i];
810 // Conservatively require the attributes of the call to match those of the
811 // return. Ignore noalias because it doesn't affect the call sequence.
812 AttributeSet CalleeAttrs = CS.getAttributes();
813 if (AttrBuilder(CalleeAttrs, AttributeSet::ReturnIndex).
814 removeAttribute(Attribute::NoAlias) !=
815 AttrBuilder(CalleeAttrs, AttributeSet::ReturnIndex).
816 removeAttribute(Attribute::NoAlias))
819 // Make sure the call instruction is followed by an unconditional branch to
821 BasicBlock *CallBB = CI->getParent();
822 BranchInst *BI = dyn_cast<BranchInst>(CallBB->getTerminator());
823 if (!BI || !BI->isUnconditional() || BI->getSuccessor(0) != BB)
826 // Duplicate the return into CallBB.
827 (void)FoldReturnIntoUncondBranch(RI, BB, CallBB);
828 ModifiedDT = Changed = true;
832 // If we eliminated all predecessors of the block, delete the block now.
833 if (Changed && !BB->hasAddressTaken() && pred_begin(BB) == pred_end(BB))
834 BB->eraseFromParent();
839 //===----------------------------------------------------------------------===//
840 // Memory Optimization
841 //===----------------------------------------------------------------------===//
845 /// ExtAddrMode - This is an extended version of TargetLowering::AddrMode
846 /// which holds actual Value*'s for register values.
847 struct ExtAddrMode : public TargetLowering::AddrMode {
850 ExtAddrMode() : BaseReg(0), ScaledReg(0) {}
851 void print(raw_ostream &OS) const;
854 bool operator==(const ExtAddrMode& O) const {
855 return (BaseReg == O.BaseReg) && (ScaledReg == O.ScaledReg) &&
856 (BaseGV == O.BaseGV) && (BaseOffs == O.BaseOffs) &&
857 (HasBaseReg == O.HasBaseReg) && (Scale == O.Scale);
862 static inline raw_ostream &operator<<(raw_ostream &OS, const ExtAddrMode &AM) {
868 void ExtAddrMode::print(raw_ostream &OS) const {
869 bool NeedPlus = false;
872 OS << (NeedPlus ? " + " : "")
874 BaseGV->printAsOperand(OS, /*PrintType=*/false);
879 OS << (NeedPlus ? " + " : "") << BaseOffs, NeedPlus = true;
882 OS << (NeedPlus ? " + " : "")
884 BaseReg->printAsOperand(OS, /*PrintType=*/false);
888 OS << (NeedPlus ? " + " : "")
890 ScaledReg->printAsOperand(OS, /*PrintType=*/false);
896 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
897 void ExtAddrMode::dump() const {
903 /// \brief This class provides transaction based operation on the IR.
904 /// Every change made through this class is recorded in the internal state and
905 /// can be undone (rollback) until commit is called.
906 class TypePromotionTransaction {
908 /// \brief This represents the common interface of the individual transaction.
909 /// Each class implements the logic for doing one specific modification on
910 /// the IR via the TypePromotionTransaction.
911 class TypePromotionAction {
913 /// The Instruction modified.
917 /// \brief Constructor of the action.
918 /// The constructor performs the related action on the IR.
919 TypePromotionAction(Instruction *Inst) : Inst(Inst) {}
921 virtual ~TypePromotionAction() {}
923 /// \brief Undo the modification done by this action.
924 /// When this method is called, the IR must be in the same state as it was
925 /// before this action was applied.
926 /// \pre Undoing the action works if and only if the IR is in the exact same
927 /// state as it was directly after this action was applied.
928 virtual void undo() = 0;
930 /// \brief Advocate every change made by this action.
931 /// When the results on the IR of the action are to be kept, it is important
932 /// to call this function, otherwise hidden information may be kept forever.
933 virtual void commit() {
934 // Nothing to be done, this action is not doing anything.
938 /// \brief Utility to remember the position of an instruction.
939 class InsertionHandler {
940 /// Position of an instruction.
941 /// Either an instruction:
942 /// - Is the first in a basic block: BB is used.
943 /// - Has a previous instructon: PrevInst is used.
945 Instruction *PrevInst;
948 /// Remember whether or not the instruction had a previous instruction.
949 bool HasPrevInstruction;
952 /// \brief Record the position of \p Inst.
953 InsertionHandler(Instruction *Inst) {
954 BasicBlock::iterator It = Inst;
955 HasPrevInstruction = (It != (Inst->getParent()->begin()));
956 if (HasPrevInstruction)
957 Point.PrevInst = --It;
959 Point.BB = Inst->getParent();
962 /// \brief Insert \p Inst at the recorded position.
963 void insert(Instruction *Inst) {
964 if (HasPrevInstruction) {
965 if (Inst->getParent())
966 Inst->removeFromParent();
967 Inst->insertAfter(Point.PrevInst);
969 Instruction *Position = Point.BB->getFirstInsertionPt();
970 if (Inst->getParent())
971 Inst->moveBefore(Position);
973 Inst->insertBefore(Position);
978 /// \brief Move an instruction before another.
979 class InstructionMoveBefore : public TypePromotionAction {
980 /// Original position of the instruction.
981 InsertionHandler Position;
984 /// \brief Move \p Inst before \p Before.
985 InstructionMoveBefore(Instruction *Inst, Instruction *Before)
986 : TypePromotionAction(Inst), Position(Inst) {
987 DEBUG(dbgs() << "Do: move: " << *Inst << "\nbefore: " << *Before << "\n");
988 Inst->moveBefore(Before);
991 /// \brief Move the instruction back to its original position.
992 void undo() override {
993 DEBUG(dbgs() << "Undo: moveBefore: " << *Inst << "\n");
994 Position.insert(Inst);
998 /// \brief Set the operand of an instruction with a new value.
999 class OperandSetter : public TypePromotionAction {
1000 /// Original operand of the instruction.
1002 /// Index of the modified instruction.
1006 /// \brief Set \p Idx operand of \p Inst with \p NewVal.
1007 OperandSetter(Instruction *Inst, unsigned Idx, Value *NewVal)
1008 : TypePromotionAction(Inst), Idx(Idx) {
1009 DEBUG(dbgs() << "Do: setOperand: " << Idx << "\n"
1010 << "for:" << *Inst << "\n"
1011 << "with:" << *NewVal << "\n");
1012 Origin = Inst->getOperand(Idx);
1013 Inst->setOperand(Idx, NewVal);
1016 /// \brief Restore the original value of the instruction.
1017 void undo() override {
1018 DEBUG(dbgs() << "Undo: setOperand:" << Idx << "\n"
1019 << "for: " << *Inst << "\n"
1020 << "with: " << *Origin << "\n");
1021 Inst->setOperand(Idx, Origin);
1025 /// \brief Hide the operands of an instruction.
1026 /// Do as if this instruction was not using any of its operands.
1027 class OperandsHider : public TypePromotionAction {
1028 /// The list of original operands.
1029 SmallVector<Value *, 4> OriginalValues;
1032 /// \brief Remove \p Inst from the uses of the operands of \p Inst.
1033 OperandsHider(Instruction *Inst) : TypePromotionAction(Inst) {
1034 DEBUG(dbgs() << "Do: OperandsHider: " << *Inst << "\n");
1035 unsigned NumOpnds = Inst->getNumOperands();
1036 OriginalValues.reserve(NumOpnds);
1037 for (unsigned It = 0; It < NumOpnds; ++It) {
1038 // Save the current operand.
1039 Value *Val = Inst->getOperand(It);
1040 OriginalValues.push_back(Val);
1042 // We could use OperandSetter here, but that would implied an overhead
1043 // that we are not willing to pay.
1044 Inst->setOperand(It, UndefValue::get(Val->getType()));
1048 /// \brief Restore the original list of uses.
1049 void undo() override {
1050 DEBUG(dbgs() << "Undo: OperandsHider: " << *Inst << "\n");
1051 for (unsigned It = 0, EndIt = OriginalValues.size(); It != EndIt; ++It)
1052 Inst->setOperand(It, OriginalValues[It]);
1056 /// \brief Build a truncate instruction.
1057 class TruncBuilder : public TypePromotionAction {
1059 /// \brief Build a truncate instruction of \p Opnd producing a \p Ty
1061 /// trunc Opnd to Ty.
1062 TruncBuilder(Instruction *Opnd, Type *Ty) : TypePromotionAction(Opnd) {
1063 IRBuilder<> Builder(Opnd);
1064 Inst = cast<Instruction>(Builder.CreateTrunc(Opnd, Ty, "promoted"));
1065 DEBUG(dbgs() << "Do: TruncBuilder: " << *Inst << "\n");
1068 /// \brief Get the built instruction.
1069 Instruction *getBuiltInstruction() { return Inst; }
1071 /// \brief Remove the built instruction.
1072 void undo() override {
1073 DEBUG(dbgs() << "Undo: TruncBuilder: " << *Inst << "\n");
1074 Inst->eraseFromParent();
1078 /// \brief Build a sign extension instruction.
1079 class SExtBuilder : public TypePromotionAction {
1081 /// \brief Build a sign extension instruction of \p Opnd producing a \p Ty
1083 /// sext Opnd to Ty.
1084 SExtBuilder(Instruction *InsertPt, Value *Opnd, Type *Ty)
1085 : TypePromotionAction(Inst) {
1086 IRBuilder<> Builder(InsertPt);
1087 Inst = cast<Instruction>(Builder.CreateSExt(Opnd, Ty, "promoted"));
1088 DEBUG(dbgs() << "Do: SExtBuilder: " << *Inst << "\n");
1091 /// \brief Get the built instruction.
1092 Instruction *getBuiltInstruction() { return Inst; }
1094 /// \brief Remove the built instruction.
1095 void undo() override {
1096 DEBUG(dbgs() << "Undo: SExtBuilder: " << *Inst << "\n");
1097 Inst->eraseFromParent();
1101 /// \brief Mutate an instruction to another type.
1102 class TypeMutator : public TypePromotionAction {
1103 /// Record the original type.
1107 /// \brief Mutate the type of \p Inst into \p NewTy.
1108 TypeMutator(Instruction *Inst, Type *NewTy)
1109 : TypePromotionAction(Inst), OrigTy(Inst->getType()) {
1110 DEBUG(dbgs() << "Do: MutateType: " << *Inst << " with " << *NewTy
1112 Inst->mutateType(NewTy);
1115 /// \brief Mutate the instruction back to its original type.
1116 void undo() override {
1117 DEBUG(dbgs() << "Undo: MutateType: " << *Inst << " with " << *OrigTy
1119 Inst->mutateType(OrigTy);
1123 /// \brief Replace the uses of an instruction by another instruction.
1124 class UsesReplacer : public TypePromotionAction {
1125 /// Helper structure to keep track of the replaced uses.
1126 struct InstructionAndIdx {
1127 /// The instruction using the instruction.
1129 /// The index where this instruction is used for Inst.
1131 InstructionAndIdx(Instruction *Inst, unsigned Idx)
1132 : Inst(Inst), Idx(Idx) {}
1135 /// Keep track of the original uses (pair Instruction, Index).
1136 SmallVector<InstructionAndIdx, 4> OriginalUses;
1137 typedef SmallVectorImpl<InstructionAndIdx>::iterator use_iterator;
1140 /// \brief Replace all the use of \p Inst by \p New.
1141 UsesReplacer(Instruction *Inst, Value *New) : TypePromotionAction(Inst) {
1142 DEBUG(dbgs() << "Do: UsersReplacer: " << *Inst << " with " << *New
1144 // Record the original uses.
1145 for (Use &U : Inst->uses()) {
1146 Instruction *UserI = cast<Instruction>(U.getUser());
1147 OriginalUses.push_back(InstructionAndIdx(UserI, U.getOperandNo()));
1149 // Now, we can replace the uses.
1150 Inst->replaceAllUsesWith(New);
1153 /// \brief Reassign the original uses of Inst to Inst.
1154 void undo() override {
1155 DEBUG(dbgs() << "Undo: UsersReplacer: " << *Inst << "\n");
1156 for (use_iterator UseIt = OriginalUses.begin(),
1157 EndIt = OriginalUses.end();
1158 UseIt != EndIt; ++UseIt) {
1159 UseIt->Inst->setOperand(UseIt->Idx, Inst);
1164 /// \brief Remove an instruction from the IR.
1165 class InstructionRemover : public TypePromotionAction {
1166 /// Original position of the instruction.
1167 InsertionHandler Inserter;
1168 /// Helper structure to hide all the link to the instruction. In other
1169 /// words, this helps to do as if the instruction was removed.
1170 OperandsHider Hider;
1171 /// Keep track of the uses replaced, if any.
1172 UsesReplacer *Replacer;
1175 /// \brief Remove all reference of \p Inst and optinally replace all its
1177 /// \pre If !Inst->use_empty(), then New != NULL
1178 InstructionRemover(Instruction *Inst, Value *New = NULL)
1179 : TypePromotionAction(Inst), Inserter(Inst), Hider(Inst),
1182 Replacer = new UsesReplacer(Inst, New);
1183 DEBUG(dbgs() << "Do: InstructionRemover: " << *Inst << "\n");
1184 Inst->removeFromParent();
1187 ~InstructionRemover() { delete Replacer; }
1189 /// \brief Really remove the instruction.
1190 void commit() override { delete Inst; }
1192 /// \brief Resurrect the instruction and reassign it to the proper uses if
1193 /// new value was provided when build this action.
1194 void undo() override {
1195 DEBUG(dbgs() << "Undo: InstructionRemover: " << *Inst << "\n");
1196 Inserter.insert(Inst);
1204 /// Restoration point.
1205 /// The restoration point is a pointer to an action instead of an iterator
1206 /// because the iterator may be invalidated but not the pointer.
1207 typedef const TypePromotionAction *ConstRestorationPt;
1208 /// Advocate every changes made in that transaction.
1210 /// Undo all the changes made after the given point.
1211 void rollback(ConstRestorationPt Point);
1212 /// Get the current restoration point.
1213 ConstRestorationPt getRestorationPoint() const;
1215 /// \name API for IR modification with state keeping to support rollback.
1217 /// Same as Instruction::setOperand.
1218 void setOperand(Instruction *Inst, unsigned Idx, Value *NewVal);
1219 /// Same as Instruction::eraseFromParent.
1220 void eraseInstruction(Instruction *Inst, Value *NewVal = NULL);
1221 /// Same as Value::replaceAllUsesWith.
1222 void replaceAllUsesWith(Instruction *Inst, Value *New);
1223 /// Same as Value::mutateType.
1224 void mutateType(Instruction *Inst, Type *NewTy);
1225 /// Same as IRBuilder::createTrunc.
1226 Instruction *createTrunc(Instruction *Opnd, Type *Ty);
1227 /// Same as IRBuilder::createSExt.
1228 Instruction *createSExt(Instruction *Inst, Value *Opnd, Type *Ty);
1229 /// Same as Instruction::moveBefore.
1230 void moveBefore(Instruction *Inst, Instruction *Before);
1233 ~TypePromotionTransaction();
1236 /// The ordered list of actions made so far.
1237 SmallVector<TypePromotionAction *, 16> Actions;
1238 typedef SmallVectorImpl<TypePromotionAction *>::iterator CommitPt;
1241 void TypePromotionTransaction::setOperand(Instruction *Inst, unsigned Idx,
1244 new TypePromotionTransaction::OperandSetter(Inst, Idx, NewVal));
1247 void TypePromotionTransaction::eraseInstruction(Instruction *Inst,
1250 new TypePromotionTransaction::InstructionRemover(Inst, NewVal));
1253 void TypePromotionTransaction::replaceAllUsesWith(Instruction *Inst,
1255 Actions.push_back(new TypePromotionTransaction::UsesReplacer(Inst, New));
1258 void TypePromotionTransaction::mutateType(Instruction *Inst, Type *NewTy) {
1259 Actions.push_back(new TypePromotionTransaction::TypeMutator(Inst, NewTy));
1262 Instruction *TypePromotionTransaction::createTrunc(Instruction *Opnd,
1264 TruncBuilder *TB = new TruncBuilder(Opnd, Ty);
1265 Actions.push_back(TB);
1266 return TB->getBuiltInstruction();
1269 Instruction *TypePromotionTransaction::createSExt(Instruction *Inst,
1270 Value *Opnd, Type *Ty) {
1271 SExtBuilder *SB = new SExtBuilder(Inst, Opnd, Ty);
1272 Actions.push_back(SB);
1273 return SB->getBuiltInstruction();
1276 void TypePromotionTransaction::moveBefore(Instruction *Inst,
1277 Instruction *Before) {
1279 new TypePromotionTransaction::InstructionMoveBefore(Inst, Before));
1282 TypePromotionTransaction::ConstRestorationPt
1283 TypePromotionTransaction::getRestorationPoint() const {
1284 return Actions.rbegin() != Actions.rend() ? *Actions.rbegin() : NULL;
1287 void TypePromotionTransaction::commit() {
1288 for (CommitPt It = Actions.begin(), EndIt = Actions.end(); It != EndIt;
1296 void TypePromotionTransaction::rollback(
1297 TypePromotionTransaction::ConstRestorationPt Point) {
1298 while (!Actions.empty() && Point != (*Actions.rbegin())) {
1299 TypePromotionAction *Curr = Actions.pop_back_val();
1305 TypePromotionTransaction::~TypePromotionTransaction() {
1306 for (CommitPt It = Actions.begin(), EndIt = Actions.end(); It != EndIt; ++It)
1311 /// \brief A helper class for matching addressing modes.
1313 /// This encapsulates the logic for matching the target-legal addressing modes.
1314 class AddressingModeMatcher {
1315 SmallVectorImpl<Instruction*> &AddrModeInsts;
1316 const TargetLowering &TLI;
1318 /// AccessTy/MemoryInst - This is the type for the access (e.g. double) and
1319 /// the memory instruction that we're computing this address for.
1321 Instruction *MemoryInst;
1323 /// AddrMode - This is the addressing mode that we're building up. This is
1324 /// part of the return value of this addressing mode matching stuff.
1325 ExtAddrMode &AddrMode;
1327 /// The truncate instruction inserted by other CodeGenPrepare optimizations.
1328 const SetOfInstrs &InsertedTruncs;
1329 /// A map from the instructions to their type before promotion.
1330 InstrToOrigTy &PromotedInsts;
1331 /// The ongoing transaction where every action should be registered.
1332 TypePromotionTransaction &TPT;
1334 /// IgnoreProfitability - This is set to true when we should not do
1335 /// profitability checks. When true, IsProfitableToFoldIntoAddressingMode
1336 /// always returns true.
1337 bool IgnoreProfitability;
1339 AddressingModeMatcher(SmallVectorImpl<Instruction*> &AMI,
1340 const TargetLowering &T, Type *AT,
1341 Instruction *MI, ExtAddrMode &AM,
1342 const SetOfInstrs &InsertedTruncs,
1343 InstrToOrigTy &PromotedInsts,
1344 TypePromotionTransaction &TPT)
1345 : AddrModeInsts(AMI), TLI(T), AccessTy(AT), MemoryInst(MI), AddrMode(AM),
1346 InsertedTruncs(InsertedTruncs), PromotedInsts(PromotedInsts), TPT(TPT) {
1347 IgnoreProfitability = false;
1351 /// Match - Find the maximal addressing mode that a load/store of V can fold,
1352 /// give an access type of AccessTy. This returns a list of involved
1353 /// instructions in AddrModeInsts.
1354 /// \p InsertedTruncs The truncate instruction inserted by other
1357 /// \p PromotedInsts maps the instructions to their type before promotion.
1358 /// \p The ongoing transaction where every action should be registered.
1359 static ExtAddrMode Match(Value *V, Type *AccessTy,
1360 Instruction *MemoryInst,
1361 SmallVectorImpl<Instruction*> &AddrModeInsts,
1362 const TargetLowering &TLI,
1363 const SetOfInstrs &InsertedTruncs,
1364 InstrToOrigTy &PromotedInsts,
1365 TypePromotionTransaction &TPT) {
1368 bool Success = AddressingModeMatcher(AddrModeInsts, TLI, AccessTy,
1369 MemoryInst, Result, InsertedTruncs,
1370 PromotedInsts, TPT).MatchAddr(V, 0);
1371 (void)Success; assert(Success && "Couldn't select *anything*?");
1375 bool MatchScaledValue(Value *ScaleReg, int64_t Scale, unsigned Depth);
1376 bool MatchAddr(Value *V, unsigned Depth);
1377 bool MatchOperationAddr(User *Operation, unsigned Opcode, unsigned Depth,
1378 bool *MovedAway = NULL);
1379 bool IsProfitableToFoldIntoAddressingMode(Instruction *I,
1380 ExtAddrMode &AMBefore,
1381 ExtAddrMode &AMAfter);
1382 bool ValueAlreadyLiveAtInst(Value *Val, Value *KnownLive1, Value *KnownLive2);
1383 bool IsPromotionProfitable(unsigned MatchedSize, unsigned SizeWithPromotion,
1384 Value *PromotedOperand) const;
1387 /// MatchScaledValue - Try adding ScaleReg*Scale to the current addressing mode.
1388 /// Return true and update AddrMode if this addr mode is legal for the target,
1390 bool AddressingModeMatcher::MatchScaledValue(Value *ScaleReg, int64_t Scale,
1392 // If Scale is 1, then this is the same as adding ScaleReg to the addressing
1393 // mode. Just process that directly.
1395 return MatchAddr(ScaleReg, Depth);
1397 // If the scale is 0, it takes nothing to add this.
1401 // If we already have a scale of this value, we can add to it, otherwise, we
1402 // need an available scale field.
1403 if (AddrMode.Scale != 0 && AddrMode.ScaledReg != ScaleReg)
1406 ExtAddrMode TestAddrMode = AddrMode;
1408 // Add scale to turn X*4+X*3 -> X*7. This could also do things like
1409 // [A+B + A*7] -> [B+A*8].
1410 TestAddrMode.Scale += Scale;
1411 TestAddrMode.ScaledReg = ScaleReg;
1413 // If the new address isn't legal, bail out.
1414 if (!TLI.isLegalAddressingMode(TestAddrMode, AccessTy))
1417 // It was legal, so commit it.
1418 AddrMode = TestAddrMode;
1420 // Okay, we decided that we can add ScaleReg+Scale to AddrMode. Check now
1421 // to see if ScaleReg is actually X+C. If so, we can turn this into adding
1422 // X*Scale + C*Scale to addr mode.
1423 ConstantInt *CI = 0; Value *AddLHS = 0;
1424 if (isa<Instruction>(ScaleReg) && // not a constant expr.
1425 match(ScaleReg, m_Add(m_Value(AddLHS), m_ConstantInt(CI)))) {
1426 TestAddrMode.ScaledReg = AddLHS;
1427 TestAddrMode.BaseOffs += CI->getSExtValue()*TestAddrMode.Scale;
1429 // If this addressing mode is legal, commit it and remember that we folded
1430 // this instruction.
1431 if (TLI.isLegalAddressingMode(TestAddrMode, AccessTy)) {
1432 AddrModeInsts.push_back(cast<Instruction>(ScaleReg));
1433 AddrMode = TestAddrMode;
1438 // Otherwise, not (x+c)*scale, just return what we have.
1442 /// MightBeFoldableInst - This is a little filter, which returns true if an
1443 /// addressing computation involving I might be folded into a load/store
1444 /// accessing it. This doesn't need to be perfect, but needs to accept at least
1445 /// the set of instructions that MatchOperationAddr can.
1446 static bool MightBeFoldableInst(Instruction *I) {
1447 switch (I->getOpcode()) {
1448 case Instruction::BitCast:
1449 // Don't touch identity bitcasts.
1450 if (I->getType() == I->getOperand(0)->getType())
1452 return I->getType()->isPointerTy() || I->getType()->isIntegerTy();
1453 case Instruction::PtrToInt:
1454 // PtrToInt is always a noop, as we know that the int type is pointer sized.
1456 case Instruction::IntToPtr:
1457 // We know the input is intptr_t, so this is foldable.
1459 case Instruction::Add:
1461 case Instruction::Mul:
1462 case Instruction::Shl:
1463 // Can only handle X*C and X << C.
1464 return isa<ConstantInt>(I->getOperand(1));
1465 case Instruction::GetElementPtr:
1472 /// \brief Hepler class to perform type promotion.
1473 class TypePromotionHelper {
1474 /// \brief Utility function to check whether or not a sign extension of
1475 /// \p Inst with \p ConsideredSExtType can be moved through \p Inst by either
1476 /// using the operands of \p Inst or promoting \p Inst.
1477 /// In other words, check if:
1478 /// sext (Ty Inst opnd1 opnd2 ... opndN) to ConsideredSExtType.
1479 /// #1 Promotion applies:
1480 /// ConsideredSExtType Inst (sext opnd1 to ConsideredSExtType, ...).
1481 /// #2 Operand reuses:
1482 /// sext opnd1 to ConsideredSExtType.
1483 /// \p PromotedInsts maps the instructions to their type before promotion.
1484 static bool canGetThrough(const Instruction *Inst, Type *ConsideredSExtType,
1485 const InstrToOrigTy &PromotedInsts);
1487 /// \brief Utility function to determine if \p OpIdx should be promoted when
1488 /// promoting \p Inst.
1489 static bool shouldSExtOperand(const Instruction *Inst, int OpIdx) {
1490 if (isa<SelectInst>(Inst) && OpIdx == 0)
1495 /// \brief Utility function to promote the operand of \p SExt when this
1496 /// operand is a promotable trunc or sext.
1497 /// \p PromotedInsts maps the instructions to their type before promotion.
1498 /// \p CreatedInsts[out] contains how many non-free instructions have been
1499 /// created to promote the operand of SExt.
1500 /// Should never be called directly.
1501 /// \return The promoted value which is used instead of SExt.
1502 static Value *promoteOperandForTruncAndSExt(Instruction *SExt,
1503 TypePromotionTransaction &TPT,
1504 InstrToOrigTy &PromotedInsts,
1505 unsigned &CreatedInsts);
1507 /// \brief Utility function to promote the operand of \p SExt when this
1508 /// operand is promotable and is not a supported trunc or sext.
1509 /// \p PromotedInsts maps the instructions to their type before promotion.
1510 /// \p CreatedInsts[out] contains how many non-free instructions have been
1511 /// created to promote the operand of SExt.
1512 /// Should never be called directly.
1513 /// \return The promoted value which is used instead of SExt.
1514 static Value *promoteOperandForOther(Instruction *SExt,
1515 TypePromotionTransaction &TPT,
1516 InstrToOrigTy &PromotedInsts,
1517 unsigned &CreatedInsts);
1520 /// Type for the utility function that promotes the operand of SExt.
1521 typedef Value *(*Action)(Instruction *SExt, TypePromotionTransaction &TPT,
1522 InstrToOrigTy &PromotedInsts,
1523 unsigned &CreatedInsts);
1524 /// \brief Given a sign extend instruction \p SExt, return the approriate
1525 /// action to promote the operand of \p SExt instead of using SExt.
1526 /// \return NULL if no promotable action is possible with the current
1528 /// \p InsertedTruncs keeps track of all the truncate instructions inserted by
1529 /// the others CodeGenPrepare optimizations. This information is important
1530 /// because we do not want to promote these instructions as CodeGenPrepare
1531 /// will reinsert them later. Thus creating an infinite loop: create/remove.
1532 /// \p PromotedInsts maps the instructions to their type before promotion.
1533 static Action getAction(Instruction *SExt, const SetOfInstrs &InsertedTruncs,
1534 const TargetLowering &TLI,
1535 const InstrToOrigTy &PromotedInsts);
1538 bool TypePromotionHelper::canGetThrough(const Instruction *Inst,
1539 Type *ConsideredSExtType,
1540 const InstrToOrigTy &PromotedInsts) {
1541 // We can always get through sext.
1542 if (isa<SExtInst>(Inst))
1545 // We can get through binary operator, if it is legal. In other words, the
1546 // binary operator must have a nuw or nsw flag.
1547 const BinaryOperator *BinOp = dyn_cast<BinaryOperator>(Inst);
1548 if (BinOp && isa<OverflowingBinaryOperator>(BinOp) &&
1549 (BinOp->hasNoUnsignedWrap() || BinOp->hasNoSignedWrap()))
1552 // Check if we can do the following simplification.
1553 // sext(trunc(sext)) --> sext
1554 if (!isa<TruncInst>(Inst))
1557 Value *OpndVal = Inst->getOperand(0);
1558 // Check if we can use this operand in the sext.
1559 // If the type is larger than the result type of the sign extension,
1561 if (OpndVal->getType()->getIntegerBitWidth() >
1562 ConsideredSExtType->getIntegerBitWidth())
1565 // If the operand of the truncate is not an instruction, we will not have
1566 // any information on the dropped bits.
1567 // (Actually we could for constant but it is not worth the extra logic).
1568 Instruction *Opnd = dyn_cast<Instruction>(OpndVal);
1572 // Check if the source of the type is narrow enough.
1573 // I.e., check that trunc just drops sign extended bits.
1574 // #1 get the type of the operand.
1575 const Type *OpndType;
1576 InstrToOrigTy::const_iterator It = PromotedInsts.find(Opnd);
1577 if (It != PromotedInsts.end())
1578 OpndType = It->second;
1579 else if (isa<SExtInst>(Opnd))
1580 OpndType = cast<Instruction>(Opnd)->getOperand(0)->getType();
1584 // #2 check that the truncate just drop sign extended bits.
1585 if (Inst->getType()->getIntegerBitWidth() >= OpndType->getIntegerBitWidth())
1591 TypePromotionHelper::Action TypePromotionHelper::getAction(
1592 Instruction *SExt, const SetOfInstrs &InsertedTruncs,
1593 const TargetLowering &TLI, const InstrToOrigTy &PromotedInsts) {
1594 Instruction *SExtOpnd = dyn_cast<Instruction>(SExt->getOperand(0));
1595 Type *SExtTy = SExt->getType();
1596 // If the operand of the sign extension is not an instruction, we cannot
1598 // If it, check we can get through.
1599 if (!SExtOpnd || !canGetThrough(SExtOpnd, SExtTy, PromotedInsts))
1602 // Do not promote if the operand has been added by codegenprepare.
1603 // Otherwise, it means we are undoing an optimization that is likely to be
1604 // redone, thus causing potential infinite loop.
1605 if (isa<TruncInst>(SExtOpnd) && InsertedTruncs.count(SExtOpnd))
1608 // SExt or Trunc instructions.
1609 // Return the related handler.
1610 if (isa<SExtInst>(SExtOpnd) || isa<TruncInst>(SExtOpnd))
1611 return promoteOperandForTruncAndSExt;
1613 // Regular instruction.
1614 // Abort early if we will have to insert non-free instructions.
1615 if (!SExtOpnd->hasOneUse() &&
1616 !TLI.isTruncateFree(SExtTy, SExtOpnd->getType()))
1618 return promoteOperandForOther;
1621 Value *TypePromotionHelper::promoteOperandForTruncAndSExt(
1622 llvm::Instruction *SExt, TypePromotionTransaction &TPT,
1623 InstrToOrigTy &PromotedInsts, unsigned &CreatedInsts) {
1624 // By construction, the operand of SExt is an instruction. Otherwise we cannot
1625 // get through it and this method should not be called.
1626 Instruction *SExtOpnd = cast<Instruction>(SExt->getOperand(0));
1627 // Replace sext(trunc(opnd)) or sext(sext(opnd))
1629 TPT.setOperand(SExt, 0, SExtOpnd->getOperand(0));
1632 // Remove dead code.
1633 if (SExtOpnd->use_empty())
1634 TPT.eraseInstruction(SExtOpnd);
1636 // Check if the sext is still needed.
1637 if (SExt->getType() != SExt->getOperand(0)->getType())
1640 // At this point we have: sext ty opnd to ty.
1641 // Reassign the uses of SExt to the opnd and remove SExt.
1642 Value *NextVal = SExt->getOperand(0);
1643 TPT.eraseInstruction(SExt, NextVal);
1648 TypePromotionHelper::promoteOperandForOther(Instruction *SExt,
1649 TypePromotionTransaction &TPT,
1650 InstrToOrigTy &PromotedInsts,
1651 unsigned &CreatedInsts) {
1652 // By construction, the operand of SExt is an instruction. Otherwise we cannot
1653 // get through it and this method should not be called.
1654 Instruction *SExtOpnd = cast<Instruction>(SExt->getOperand(0));
1656 if (!SExtOpnd->hasOneUse()) {
1657 // SExtOpnd will be promoted.
1658 // All its uses, but SExt, will need to use a truncated value of the
1659 // promoted version.
1660 // Create the truncate now.
1661 Instruction *Trunc = TPT.createTrunc(SExt, SExtOpnd->getType());
1662 Trunc->removeFromParent();
1663 // Insert it just after the definition.
1664 Trunc->insertAfter(SExtOpnd);
1666 TPT.replaceAllUsesWith(SExtOpnd, Trunc);
1667 // Restore the operand of SExt (which has been replace by the previous call
1668 // to replaceAllUsesWith) to avoid creating a cycle trunc <-> sext.
1669 TPT.setOperand(SExt, 0, SExtOpnd);
1672 // Get through the Instruction:
1673 // 1. Update its type.
1674 // 2. Replace the uses of SExt by Inst.
1675 // 3. Sign extend each operand that needs to be sign extended.
1677 // Remember the original type of the instruction before promotion.
1678 // This is useful to know that the high bits are sign extended bits.
1679 PromotedInsts.insert(
1680 std::pair<Instruction *, Type *>(SExtOpnd, SExtOpnd->getType()));
1682 TPT.mutateType(SExtOpnd, SExt->getType());
1684 TPT.replaceAllUsesWith(SExt, SExtOpnd);
1686 Instruction *SExtForOpnd = SExt;
1688 DEBUG(dbgs() << "Propagate SExt to operands\n");
1689 for (int OpIdx = 0, EndOpIdx = SExtOpnd->getNumOperands(); OpIdx != EndOpIdx;
1691 DEBUG(dbgs() << "Operand:\n" << *(SExtOpnd->getOperand(OpIdx)) << '\n');
1692 if (SExtOpnd->getOperand(OpIdx)->getType() == SExt->getType() ||
1693 !shouldSExtOperand(SExtOpnd, OpIdx)) {
1694 DEBUG(dbgs() << "No need to propagate\n");
1697 // Check if we can statically sign extend the operand.
1698 Value *Opnd = SExtOpnd->getOperand(OpIdx);
1699 if (const ConstantInt *Cst = dyn_cast<ConstantInt>(Opnd)) {
1700 DEBUG(dbgs() << "Statically sign extend\n");
1703 ConstantInt::getSigned(SExt->getType(), Cst->getSExtValue()));
1706 // UndefValue are typed, so we have to statically sign extend them.
1707 if (isa<UndefValue>(Opnd)) {
1708 DEBUG(dbgs() << "Statically sign extend\n");
1709 TPT.setOperand(SExtOpnd, OpIdx, UndefValue::get(SExt->getType()));
1713 // Otherwise we have to explicity sign extend the operand.
1714 // Check if SExt was reused to sign extend an operand.
1716 // If yes, create a new one.
1717 DEBUG(dbgs() << "More operands to sext\n");
1718 SExtForOpnd = TPT.createSExt(SExt, Opnd, SExt->getType());
1722 TPT.setOperand(SExtForOpnd, 0, Opnd);
1724 // Move the sign extension before the insertion point.
1725 TPT.moveBefore(SExtForOpnd, SExtOpnd);
1726 TPT.setOperand(SExtOpnd, OpIdx, SExtForOpnd);
1727 // If more sext are required, new instructions will have to be created.
1730 if (SExtForOpnd == SExt) {
1731 DEBUG(dbgs() << "Sign extension is useless now\n");
1732 TPT.eraseInstruction(SExt);
1737 /// IsPromotionProfitable - Check whether or not promoting an instruction
1738 /// to a wider type was profitable.
1739 /// \p MatchedSize gives the number of instructions that have been matched
1740 /// in the addressing mode after the promotion was applied.
1741 /// \p SizeWithPromotion gives the number of created instructions for
1742 /// the promotion plus the number of instructions that have been
1743 /// matched in the addressing mode before the promotion.
1744 /// \p PromotedOperand is the value that has been promoted.
1745 /// \return True if the promotion is profitable, false otherwise.
1747 AddressingModeMatcher::IsPromotionProfitable(unsigned MatchedSize,
1748 unsigned SizeWithPromotion,
1749 Value *PromotedOperand) const {
1750 // We folded less instructions than what we created to promote the operand.
1751 // This is not profitable.
1752 if (MatchedSize < SizeWithPromotion)
1754 if (MatchedSize > SizeWithPromotion)
1756 // The promotion is neutral but it may help folding the sign extension in
1757 // loads for instance.
1758 // Check that we did not create an illegal instruction.
1759 Instruction *PromotedInst = dyn_cast<Instruction>(PromotedOperand);
1762 int ISDOpcode = TLI.InstructionOpcodeToISD(PromotedInst->getOpcode());
1763 // If the ISDOpcode is undefined, it was undefined before the promotion.
1766 // Otherwise, check if the promoted instruction is legal or not.
1767 return TLI.isOperationLegalOrCustom(ISDOpcode,
1768 EVT::getEVT(PromotedInst->getType()));
1771 /// MatchOperationAddr - Given an instruction or constant expr, see if we can
1772 /// fold the operation into the addressing mode. If so, update the addressing
1773 /// mode and return true, otherwise return false without modifying AddrMode.
1774 /// If \p MovedAway is not NULL, it contains the information of whether or
1775 /// not AddrInst has to be folded into the addressing mode on success.
1776 /// If \p MovedAway == true, \p AddrInst will not be part of the addressing
1777 /// because it has been moved away.
1778 /// Thus AddrInst must not be added in the matched instructions.
1779 /// This state can happen when AddrInst is a sext, since it may be moved away.
1780 /// Therefore, AddrInst may not be valid when MovedAway is true and it must
1781 /// not be referenced anymore.
1782 bool AddressingModeMatcher::MatchOperationAddr(User *AddrInst, unsigned Opcode,
1785 // Avoid exponential behavior on extremely deep expression trees.
1786 if (Depth >= 5) return false;
1788 // By default, all matched instructions stay in place.
1793 case Instruction::PtrToInt:
1794 // PtrToInt is always a noop, as we know that the int type is pointer sized.
1795 return MatchAddr(AddrInst->getOperand(0), Depth);
1796 case Instruction::IntToPtr:
1797 // This inttoptr is a no-op if the integer type is pointer sized.
1798 if (TLI.getValueType(AddrInst->getOperand(0)->getType()) ==
1799 TLI.getPointerTy(AddrInst->getType()->getPointerAddressSpace()))
1800 return MatchAddr(AddrInst->getOperand(0), Depth);
1802 case Instruction::BitCast:
1803 // BitCast is always a noop, and we can handle it as long as it is
1804 // int->int or pointer->pointer (we don't want int<->fp or something).
1805 if ((AddrInst->getOperand(0)->getType()->isPointerTy() ||
1806 AddrInst->getOperand(0)->getType()->isIntegerTy()) &&
1807 // Don't touch identity bitcasts. These were probably put here by LSR,
1808 // and we don't want to mess around with them. Assume it knows what it
1810 AddrInst->getOperand(0)->getType() != AddrInst->getType())
1811 return MatchAddr(AddrInst->getOperand(0), Depth);
1813 case Instruction::Add: {
1814 // Check to see if we can merge in the RHS then the LHS. If so, we win.
1815 ExtAddrMode BackupAddrMode = AddrMode;
1816 unsigned OldSize = AddrModeInsts.size();
1817 // Start a transaction at this point.
1818 // The LHS may match but not the RHS.
1819 // Therefore, we need a higher level restoration point to undo partially
1820 // matched operation.
1821 TypePromotionTransaction::ConstRestorationPt LastKnownGood =
1822 TPT.getRestorationPoint();
1824 if (MatchAddr(AddrInst->getOperand(1), Depth+1) &&
1825 MatchAddr(AddrInst->getOperand(0), Depth+1))
1828 // Restore the old addr mode info.
1829 AddrMode = BackupAddrMode;
1830 AddrModeInsts.resize(OldSize);
1831 TPT.rollback(LastKnownGood);
1833 // Otherwise this was over-aggressive. Try merging in the LHS then the RHS.
1834 if (MatchAddr(AddrInst->getOperand(0), Depth+1) &&
1835 MatchAddr(AddrInst->getOperand(1), Depth+1))
1838 // Otherwise we definitely can't merge the ADD in.
1839 AddrMode = BackupAddrMode;
1840 AddrModeInsts.resize(OldSize);
1841 TPT.rollback(LastKnownGood);
1844 //case Instruction::Or:
1845 // TODO: We can handle "Or Val, Imm" iff this OR is equivalent to an ADD.
1847 case Instruction::Mul:
1848 case Instruction::Shl: {
1849 // Can only handle X*C and X << C.
1850 ConstantInt *RHS = dyn_cast<ConstantInt>(AddrInst->getOperand(1));
1851 if (!RHS) return false;
1852 int64_t Scale = RHS->getSExtValue();
1853 if (Opcode == Instruction::Shl)
1854 Scale = 1LL << Scale;
1856 return MatchScaledValue(AddrInst->getOperand(0), Scale, Depth);
1858 case Instruction::GetElementPtr: {
1859 // Scan the GEP. We check it if it contains constant offsets and at most
1860 // one variable offset.
1861 int VariableOperand = -1;
1862 unsigned VariableScale = 0;
1864 int64_t ConstantOffset = 0;
1865 const DataLayout *TD = TLI.getDataLayout();
1866 gep_type_iterator GTI = gep_type_begin(AddrInst);
1867 for (unsigned i = 1, e = AddrInst->getNumOperands(); i != e; ++i, ++GTI) {
1868 if (StructType *STy = dyn_cast<StructType>(*GTI)) {
1869 const StructLayout *SL = TD->getStructLayout(STy);
1871 cast<ConstantInt>(AddrInst->getOperand(i))->getZExtValue();
1872 ConstantOffset += SL->getElementOffset(Idx);
1874 uint64_t TypeSize = TD->getTypeAllocSize(GTI.getIndexedType());
1875 if (ConstantInt *CI = dyn_cast<ConstantInt>(AddrInst->getOperand(i))) {
1876 ConstantOffset += CI->getSExtValue()*TypeSize;
1877 } else if (TypeSize) { // Scales of zero don't do anything.
1878 // We only allow one variable index at the moment.
1879 if (VariableOperand != -1)
1882 // Remember the variable index.
1883 VariableOperand = i;
1884 VariableScale = TypeSize;
1889 // A common case is for the GEP to only do a constant offset. In this case,
1890 // just add it to the disp field and check validity.
1891 if (VariableOperand == -1) {
1892 AddrMode.BaseOffs += ConstantOffset;
1893 if (ConstantOffset == 0 || TLI.isLegalAddressingMode(AddrMode, AccessTy)){
1894 // Check to see if we can fold the base pointer in too.
1895 if (MatchAddr(AddrInst->getOperand(0), Depth+1))
1898 AddrMode.BaseOffs -= ConstantOffset;
1902 // Save the valid addressing mode in case we can't match.
1903 ExtAddrMode BackupAddrMode = AddrMode;
1904 unsigned OldSize = AddrModeInsts.size();
1906 // See if the scale and offset amount is valid for this target.
1907 AddrMode.BaseOffs += ConstantOffset;
1909 // Match the base operand of the GEP.
1910 if (!MatchAddr(AddrInst->getOperand(0), Depth+1)) {
1911 // If it couldn't be matched, just stuff the value in a register.
1912 if (AddrMode.HasBaseReg) {
1913 AddrMode = BackupAddrMode;
1914 AddrModeInsts.resize(OldSize);
1917 AddrMode.HasBaseReg = true;
1918 AddrMode.BaseReg = AddrInst->getOperand(0);
1921 // Match the remaining variable portion of the GEP.
1922 if (!MatchScaledValue(AddrInst->getOperand(VariableOperand), VariableScale,
1924 // If it couldn't be matched, try stuffing the base into a register
1925 // instead of matching it, and retrying the match of the scale.
1926 AddrMode = BackupAddrMode;
1927 AddrModeInsts.resize(OldSize);
1928 if (AddrMode.HasBaseReg)
1930 AddrMode.HasBaseReg = true;
1931 AddrMode.BaseReg = AddrInst->getOperand(0);
1932 AddrMode.BaseOffs += ConstantOffset;
1933 if (!MatchScaledValue(AddrInst->getOperand(VariableOperand),
1934 VariableScale, Depth)) {
1935 // If even that didn't work, bail.
1936 AddrMode = BackupAddrMode;
1937 AddrModeInsts.resize(OldSize);
1944 case Instruction::SExt: {
1945 // Try to move this sext out of the way of the addressing mode.
1946 Instruction *SExt = cast<Instruction>(AddrInst);
1947 // Ask for a method for doing so.
1948 TypePromotionHelper::Action TPH = TypePromotionHelper::getAction(
1949 SExt, InsertedTruncs, TLI, PromotedInsts);
1953 TypePromotionTransaction::ConstRestorationPt LastKnownGood =
1954 TPT.getRestorationPoint();
1955 unsigned CreatedInsts = 0;
1956 Value *PromotedOperand = TPH(SExt, TPT, PromotedInsts, CreatedInsts);
1957 // SExt has been moved away.
1958 // Thus either it will be rematched later in the recursive calls or it is
1959 // gone. Anyway, we must not fold it into the addressing mode at this point.
1963 // addr = gep base, idx
1965 // promotedOpnd = sext opnd <- no match here
1966 // op = promoted_add promotedOpnd, 1 <- match (later in recursive calls)
1967 // addr = gep base, op <- match
1971 assert(PromotedOperand &&
1972 "TypePromotionHelper should have filtered out those cases");
1974 ExtAddrMode BackupAddrMode = AddrMode;
1975 unsigned OldSize = AddrModeInsts.size();
1977 if (!MatchAddr(PromotedOperand, Depth) ||
1978 !IsPromotionProfitable(AddrModeInsts.size(), OldSize + CreatedInsts,
1980 AddrMode = BackupAddrMode;
1981 AddrModeInsts.resize(OldSize);
1982 DEBUG(dbgs() << "Sign extension does not pay off: rollback\n");
1983 TPT.rollback(LastKnownGood);
1992 /// MatchAddr - If we can, try to add the value of 'Addr' into the current
1993 /// addressing mode. If Addr can't be added to AddrMode this returns false and
1994 /// leaves AddrMode unmodified. This assumes that Addr is either a pointer type
1995 /// or intptr_t for the target.
1997 bool AddressingModeMatcher::MatchAddr(Value *Addr, unsigned Depth) {
1998 // Start a transaction at this point that we will rollback if the matching
2000 TypePromotionTransaction::ConstRestorationPt LastKnownGood =
2001 TPT.getRestorationPoint();
2002 if (ConstantInt *CI = dyn_cast<ConstantInt>(Addr)) {
2003 // Fold in immediates if legal for the target.
2004 AddrMode.BaseOffs += CI->getSExtValue();
2005 if (TLI.isLegalAddressingMode(AddrMode, AccessTy))
2007 AddrMode.BaseOffs -= CI->getSExtValue();
2008 } else if (GlobalValue *GV = dyn_cast<GlobalValue>(Addr)) {
2009 // If this is a global variable, try to fold it into the addressing mode.
2010 if (AddrMode.BaseGV == 0) {
2011 AddrMode.BaseGV = GV;
2012 if (TLI.isLegalAddressingMode(AddrMode, AccessTy))
2014 AddrMode.BaseGV = 0;
2016 } else if (Instruction *I = dyn_cast<Instruction>(Addr)) {
2017 ExtAddrMode BackupAddrMode = AddrMode;
2018 unsigned OldSize = AddrModeInsts.size();
2020 // Check to see if it is possible to fold this operation.
2021 bool MovedAway = false;
2022 if (MatchOperationAddr(I, I->getOpcode(), Depth, &MovedAway)) {
2023 // This instruction may have been move away. If so, there is nothing
2027 // Okay, it's possible to fold this. Check to see if it is actually
2028 // *profitable* to do so. We use a simple cost model to avoid increasing
2029 // register pressure too much.
2030 if (I->hasOneUse() ||
2031 IsProfitableToFoldIntoAddressingMode(I, BackupAddrMode, AddrMode)) {
2032 AddrModeInsts.push_back(I);
2036 // It isn't profitable to do this, roll back.
2037 //cerr << "NOT FOLDING: " << *I;
2038 AddrMode = BackupAddrMode;
2039 AddrModeInsts.resize(OldSize);
2040 TPT.rollback(LastKnownGood);
2042 } else if (ConstantExpr *CE = dyn_cast<ConstantExpr>(Addr)) {
2043 if (MatchOperationAddr(CE, CE->getOpcode(), Depth))
2045 TPT.rollback(LastKnownGood);
2046 } else if (isa<ConstantPointerNull>(Addr)) {
2047 // Null pointer gets folded without affecting the addressing mode.
2051 // Worse case, the target should support [reg] addressing modes. :)
2052 if (!AddrMode.HasBaseReg) {
2053 AddrMode.HasBaseReg = true;
2054 AddrMode.BaseReg = Addr;
2055 // Still check for legality in case the target supports [imm] but not [i+r].
2056 if (TLI.isLegalAddressingMode(AddrMode, AccessTy))
2058 AddrMode.HasBaseReg = false;
2059 AddrMode.BaseReg = 0;
2062 // If the base register is already taken, see if we can do [r+r].
2063 if (AddrMode.Scale == 0) {
2065 AddrMode.ScaledReg = Addr;
2066 if (TLI.isLegalAddressingMode(AddrMode, AccessTy))
2069 AddrMode.ScaledReg = 0;
2072 TPT.rollback(LastKnownGood);
2076 /// IsOperandAMemoryOperand - Check to see if all uses of OpVal by the specified
2077 /// inline asm call are due to memory operands. If so, return true, otherwise
2079 static bool IsOperandAMemoryOperand(CallInst *CI, InlineAsm *IA, Value *OpVal,
2080 const TargetLowering &TLI) {
2081 TargetLowering::AsmOperandInfoVector TargetConstraints = TLI.ParseConstraints(ImmutableCallSite(CI));
2082 for (unsigned i = 0, e = TargetConstraints.size(); i != e; ++i) {
2083 TargetLowering::AsmOperandInfo &OpInfo = TargetConstraints[i];
2085 // Compute the constraint code and ConstraintType to use.
2086 TLI.ComputeConstraintToUse(OpInfo, SDValue());
2088 // If this asm operand is our Value*, and if it isn't an indirect memory
2089 // operand, we can't fold it!
2090 if (OpInfo.CallOperandVal == OpVal &&
2091 (OpInfo.ConstraintType != TargetLowering::C_Memory ||
2092 !OpInfo.isIndirect))
2099 /// FindAllMemoryUses - Recursively walk all the uses of I until we find a
2100 /// memory use. If we find an obviously non-foldable instruction, return true.
2101 /// Add the ultimately found memory instructions to MemoryUses.
2102 static bool FindAllMemoryUses(Instruction *I,
2103 SmallVectorImpl<std::pair<Instruction*,unsigned> > &MemoryUses,
2104 SmallPtrSet<Instruction*, 16> &ConsideredInsts,
2105 const TargetLowering &TLI) {
2106 // If we already considered this instruction, we're done.
2107 if (!ConsideredInsts.insert(I))
2110 // If this is an obviously unfoldable instruction, bail out.
2111 if (!MightBeFoldableInst(I))
2114 // Loop over all the uses, recursively processing them.
2115 for (Use &U : I->uses()) {
2116 Instruction *UserI = cast<Instruction>(U.getUser());
2118 if (LoadInst *LI = dyn_cast<LoadInst>(UserI)) {
2119 MemoryUses.push_back(std::make_pair(LI, U.getOperandNo()));
2123 if (StoreInst *SI = dyn_cast<StoreInst>(UserI)) {
2124 unsigned opNo = U.getOperandNo();
2125 if (opNo == 0) return true; // Storing addr, not into addr.
2126 MemoryUses.push_back(std::make_pair(SI, opNo));
2130 if (CallInst *CI = dyn_cast<CallInst>(UserI)) {
2131 InlineAsm *IA = dyn_cast<InlineAsm>(CI->getCalledValue());
2132 if (!IA) return true;
2134 // If this is a memory operand, we're cool, otherwise bail out.
2135 if (!IsOperandAMemoryOperand(CI, IA, I, TLI))
2140 if (FindAllMemoryUses(UserI, MemoryUses, ConsideredInsts, TLI))
2147 /// ValueAlreadyLiveAtInst - Retrn true if Val is already known to be live at
2148 /// the use site that we're folding it into. If so, there is no cost to
2149 /// include it in the addressing mode. KnownLive1 and KnownLive2 are two values
2150 /// that we know are live at the instruction already.
2151 bool AddressingModeMatcher::ValueAlreadyLiveAtInst(Value *Val,Value *KnownLive1,
2152 Value *KnownLive2) {
2153 // If Val is either of the known-live values, we know it is live!
2154 if (Val == 0 || Val == KnownLive1 || Val == KnownLive2)
2157 // All values other than instructions and arguments (e.g. constants) are live.
2158 if (!isa<Instruction>(Val) && !isa<Argument>(Val)) return true;
2160 // If Val is a constant sized alloca in the entry block, it is live, this is
2161 // true because it is just a reference to the stack/frame pointer, which is
2162 // live for the whole function.
2163 if (AllocaInst *AI = dyn_cast<AllocaInst>(Val))
2164 if (AI->isStaticAlloca())
2167 // Check to see if this value is already used in the memory instruction's
2168 // block. If so, it's already live into the block at the very least, so we
2169 // can reasonably fold it.
2170 return Val->isUsedInBasicBlock(MemoryInst->getParent());
2173 /// IsProfitableToFoldIntoAddressingMode - It is possible for the addressing
2174 /// mode of the machine to fold the specified instruction into a load or store
2175 /// that ultimately uses it. However, the specified instruction has multiple
2176 /// uses. Given this, it may actually increase register pressure to fold it
2177 /// into the load. For example, consider this code:
2181 /// use(Y) -> nonload/store
2185 /// In this case, Y has multiple uses, and can be folded into the load of Z
2186 /// (yielding load [X+2]). However, doing this will cause both "X" and "X+1" to
2187 /// be live at the use(Y) line. If we don't fold Y into load Z, we use one
2188 /// fewer register. Since Y can't be folded into "use(Y)" we don't increase the
2189 /// number of computations either.
2191 /// Note that this (like most of CodeGenPrepare) is just a rough heuristic. If
2192 /// X was live across 'load Z' for other reasons, we actually *would* want to
2193 /// fold the addressing mode in the Z case. This would make Y die earlier.
2194 bool AddressingModeMatcher::
2195 IsProfitableToFoldIntoAddressingMode(Instruction *I, ExtAddrMode &AMBefore,
2196 ExtAddrMode &AMAfter) {
2197 if (IgnoreProfitability) return true;
2199 // AMBefore is the addressing mode before this instruction was folded into it,
2200 // and AMAfter is the addressing mode after the instruction was folded. Get
2201 // the set of registers referenced by AMAfter and subtract out those
2202 // referenced by AMBefore: this is the set of values which folding in this
2203 // address extends the lifetime of.
2205 // Note that there are only two potential values being referenced here,
2206 // BaseReg and ScaleReg (global addresses are always available, as are any
2207 // folded immediates).
2208 Value *BaseReg = AMAfter.BaseReg, *ScaledReg = AMAfter.ScaledReg;
2210 // If the BaseReg or ScaledReg was referenced by the previous addrmode, their
2211 // lifetime wasn't extended by adding this instruction.
2212 if (ValueAlreadyLiveAtInst(BaseReg, AMBefore.BaseReg, AMBefore.ScaledReg))
2214 if (ValueAlreadyLiveAtInst(ScaledReg, AMBefore.BaseReg, AMBefore.ScaledReg))
2217 // If folding this instruction (and it's subexprs) didn't extend any live
2218 // ranges, we're ok with it.
2219 if (BaseReg == 0 && ScaledReg == 0)
2222 // If all uses of this instruction are ultimately load/store/inlineasm's,
2223 // check to see if their addressing modes will include this instruction. If
2224 // so, we can fold it into all uses, so it doesn't matter if it has multiple
2226 SmallVector<std::pair<Instruction*,unsigned>, 16> MemoryUses;
2227 SmallPtrSet<Instruction*, 16> ConsideredInsts;
2228 if (FindAllMemoryUses(I, MemoryUses, ConsideredInsts, TLI))
2229 return false; // Has a non-memory, non-foldable use!
2231 // Now that we know that all uses of this instruction are part of a chain of
2232 // computation involving only operations that could theoretically be folded
2233 // into a memory use, loop over each of these uses and see if they could
2234 // *actually* fold the instruction.
2235 SmallVector<Instruction*, 32> MatchedAddrModeInsts;
2236 for (unsigned i = 0, e = MemoryUses.size(); i != e; ++i) {
2237 Instruction *User = MemoryUses[i].first;
2238 unsigned OpNo = MemoryUses[i].second;
2240 // Get the access type of this use. If the use isn't a pointer, we don't
2241 // know what it accesses.
2242 Value *Address = User->getOperand(OpNo);
2243 if (!Address->getType()->isPointerTy())
2245 Type *AddressAccessTy = Address->getType()->getPointerElementType();
2247 // Do a match against the root of this address, ignoring profitability. This
2248 // will tell us if the addressing mode for the memory operation will
2249 // *actually* cover the shared instruction.
2251 TypePromotionTransaction::ConstRestorationPt LastKnownGood =
2252 TPT.getRestorationPoint();
2253 AddressingModeMatcher Matcher(MatchedAddrModeInsts, TLI, AddressAccessTy,
2254 MemoryInst, Result, InsertedTruncs,
2255 PromotedInsts, TPT);
2256 Matcher.IgnoreProfitability = true;
2257 bool Success = Matcher.MatchAddr(Address, 0);
2258 (void)Success; assert(Success && "Couldn't select *anything*?");
2260 // The match was to check the profitability, the changes made are not
2261 // part of the original matcher. Therefore, they should be dropped
2262 // otherwise the original matcher will not present the right state.
2263 TPT.rollback(LastKnownGood);
2265 // If the match didn't cover I, then it won't be shared by it.
2266 if (std::find(MatchedAddrModeInsts.begin(), MatchedAddrModeInsts.end(),
2267 I) == MatchedAddrModeInsts.end())
2270 MatchedAddrModeInsts.clear();
2276 } // end anonymous namespace
2278 /// IsNonLocalValue - Return true if the specified values are defined in a
2279 /// different basic block than BB.
2280 static bool IsNonLocalValue(Value *V, BasicBlock *BB) {
2281 if (Instruction *I = dyn_cast<Instruction>(V))
2282 return I->getParent() != BB;
2286 /// OptimizeMemoryInst - Load and Store Instructions often have
2287 /// addressing modes that can do significant amounts of computation. As such,
2288 /// instruction selection will try to get the load or store to do as much
2289 /// computation as possible for the program. The problem is that isel can only
2290 /// see within a single block. As such, we sink as much legal addressing mode
2291 /// stuff into the block as possible.
2293 /// This method is used to optimize both load/store and inline asms with memory
2295 bool CodeGenPrepare::OptimizeMemoryInst(Instruction *MemoryInst, Value *Addr,
2299 // Try to collapse single-value PHI nodes. This is necessary to undo
2300 // unprofitable PRE transformations.
2301 SmallVector<Value*, 8> worklist;
2302 SmallPtrSet<Value*, 16> Visited;
2303 worklist.push_back(Addr);
2305 // Use a worklist to iteratively look through PHI nodes, and ensure that
2306 // the addressing mode obtained from the non-PHI roots of the graph
2308 Value *Consensus = 0;
2309 unsigned NumUsesConsensus = 0;
2310 bool IsNumUsesConsensusValid = false;
2311 SmallVector<Instruction*, 16> AddrModeInsts;
2312 ExtAddrMode AddrMode;
2313 TypePromotionTransaction TPT;
2314 TypePromotionTransaction::ConstRestorationPt LastKnownGood =
2315 TPT.getRestorationPoint();
2316 while (!worklist.empty()) {
2317 Value *V = worklist.back();
2318 worklist.pop_back();
2320 // Break use-def graph loops.
2321 if (!Visited.insert(V)) {
2326 // For a PHI node, push all of its incoming values.
2327 if (PHINode *P = dyn_cast<PHINode>(V)) {
2328 for (unsigned i = 0, e = P->getNumIncomingValues(); i != e; ++i)
2329 worklist.push_back(P->getIncomingValue(i));
2333 // For non-PHIs, determine the addressing mode being computed.
2334 SmallVector<Instruction*, 16> NewAddrModeInsts;
2335 ExtAddrMode NewAddrMode = AddressingModeMatcher::Match(
2336 V, AccessTy, MemoryInst, NewAddrModeInsts, *TLI, InsertedTruncsSet,
2337 PromotedInsts, TPT);
2339 // This check is broken into two cases with very similar code to avoid using
2340 // getNumUses() as much as possible. Some values have a lot of uses, so
2341 // calling getNumUses() unconditionally caused a significant compile-time
2345 AddrMode = NewAddrMode;
2346 AddrModeInsts = NewAddrModeInsts;
2348 } else if (NewAddrMode == AddrMode) {
2349 if (!IsNumUsesConsensusValid) {
2350 NumUsesConsensus = Consensus->getNumUses();
2351 IsNumUsesConsensusValid = true;
2354 // Ensure that the obtained addressing mode is equivalent to that obtained
2355 // for all other roots of the PHI traversal. Also, when choosing one
2356 // such root as representative, select the one with the most uses in order
2357 // to keep the cost modeling heuristics in AddressingModeMatcher
2359 unsigned NumUses = V->getNumUses();
2360 if (NumUses > NumUsesConsensus) {
2362 NumUsesConsensus = NumUses;
2363 AddrModeInsts = NewAddrModeInsts;
2372 // If the addressing mode couldn't be determined, or if multiple different
2373 // ones were determined, bail out now.
2375 TPT.rollback(LastKnownGood);
2380 // Check to see if any of the instructions supersumed by this addr mode are
2381 // non-local to I's BB.
2382 bool AnyNonLocal = false;
2383 for (unsigned i = 0, e = AddrModeInsts.size(); i != e; ++i) {
2384 if (IsNonLocalValue(AddrModeInsts[i], MemoryInst->getParent())) {
2390 // If all the instructions matched are already in this BB, don't do anything.
2392 DEBUG(dbgs() << "CGP: Found local addrmode: " << AddrMode << "\n");
2396 // Insert this computation right after this user. Since our caller is
2397 // scanning from the top of the BB to the bottom, reuse of the expr are
2398 // guaranteed to happen later.
2399 IRBuilder<> Builder(MemoryInst);
2401 // Now that we determined the addressing expression we want to use and know
2402 // that we have to sink it into this block. Check to see if we have already
2403 // done this for some other load/store instr in this block. If so, reuse the
2405 Value *&SunkAddr = SunkAddrs[Addr];
2407 DEBUG(dbgs() << "CGP: Reusing nonlocal addrmode: " << AddrMode << " for "
2409 if (SunkAddr->getType() != Addr->getType())
2410 SunkAddr = Builder.CreateBitCast(SunkAddr, Addr->getType());
2412 DEBUG(dbgs() << "CGP: SINKING nonlocal addrmode: " << AddrMode << " for "
2414 Type *IntPtrTy = TLI->getDataLayout()->getIntPtrType(Addr->getType());
2417 // Start with the base register. Do this first so that subsequent address
2418 // matching finds it last, which will prevent it from trying to match it
2419 // as the scaled value in case it happens to be a mul. That would be
2420 // problematic if we've sunk a different mul for the scale, because then
2421 // we'd end up sinking both muls.
2422 if (AddrMode.BaseReg) {
2423 Value *V = AddrMode.BaseReg;
2424 if (V->getType()->isPointerTy())
2425 V = Builder.CreatePtrToInt(V, IntPtrTy, "sunkaddr");
2426 if (V->getType() != IntPtrTy)
2427 V = Builder.CreateIntCast(V, IntPtrTy, /*isSigned=*/true, "sunkaddr");
2431 // Add the scale value.
2432 if (AddrMode.Scale) {
2433 Value *V = AddrMode.ScaledReg;
2434 if (V->getType() == IntPtrTy) {
2436 } else if (V->getType()->isPointerTy()) {
2437 V = Builder.CreatePtrToInt(V, IntPtrTy, "sunkaddr");
2438 } else if (cast<IntegerType>(IntPtrTy)->getBitWidth() <
2439 cast<IntegerType>(V->getType())->getBitWidth()) {
2440 V = Builder.CreateTrunc(V, IntPtrTy, "sunkaddr");
2442 V = Builder.CreateSExt(V, IntPtrTy, "sunkaddr");
2444 if (AddrMode.Scale != 1)
2445 V = Builder.CreateMul(V, ConstantInt::get(IntPtrTy, AddrMode.Scale),
2448 Result = Builder.CreateAdd(Result, V, "sunkaddr");
2453 // Add in the BaseGV if present.
2454 if (AddrMode.BaseGV) {
2455 Value *V = Builder.CreatePtrToInt(AddrMode.BaseGV, IntPtrTy, "sunkaddr");
2457 Result = Builder.CreateAdd(Result, V, "sunkaddr");
2462 // Add in the Base Offset if present.
2463 if (AddrMode.BaseOffs) {
2464 Value *V = ConstantInt::get(IntPtrTy, AddrMode.BaseOffs);
2466 Result = Builder.CreateAdd(Result, V, "sunkaddr");
2472 SunkAddr = Constant::getNullValue(Addr->getType());
2474 SunkAddr = Builder.CreateIntToPtr(Result, Addr->getType(), "sunkaddr");
2477 MemoryInst->replaceUsesOfWith(Repl, SunkAddr);
2479 // If we have no uses, recursively delete the value and all dead instructions
2481 if (Repl->use_empty()) {
2482 // This can cause recursive deletion, which can invalidate our iterator.
2483 // Use a WeakVH to hold onto it in case this happens.
2484 WeakVH IterHandle(CurInstIterator);
2485 BasicBlock *BB = CurInstIterator->getParent();
2487 RecursivelyDeleteTriviallyDeadInstructions(Repl, TLInfo);
2489 if (IterHandle != CurInstIterator) {
2490 // If the iterator instruction was recursively deleted, start over at the
2491 // start of the block.
2492 CurInstIterator = BB->begin();
2500 /// OptimizeInlineAsmInst - If there are any memory operands, use
2501 /// OptimizeMemoryInst to sink their address computing into the block when
2502 /// possible / profitable.
2503 bool CodeGenPrepare::OptimizeInlineAsmInst(CallInst *CS) {
2504 bool MadeChange = false;
2506 TargetLowering::AsmOperandInfoVector
2507 TargetConstraints = TLI->ParseConstraints(CS);
2509 for (unsigned i = 0, e = TargetConstraints.size(); i != e; ++i) {
2510 TargetLowering::AsmOperandInfo &OpInfo = TargetConstraints[i];
2512 // Compute the constraint code and ConstraintType to use.
2513 TLI->ComputeConstraintToUse(OpInfo, SDValue());
2515 if (OpInfo.ConstraintType == TargetLowering::C_Memory &&
2516 OpInfo.isIndirect) {
2517 Value *OpVal = CS->getArgOperand(ArgNo++);
2518 MadeChange |= OptimizeMemoryInst(CS, OpVal, OpVal->getType());
2519 } else if (OpInfo.Type == InlineAsm::isInput)
2526 /// SinkExtExpand - Sink a zext or sext into its user blocks if the target type
2527 /// doesn't fit in one register
2528 bool CodeGenPrepare::SinkExtExpand(CastInst *CI) {
2530 TLI->getTypeAction(CI->getContext(), TLI->getValueType(CI->getType())) ==
2531 TargetLowering::TypeExpandInteger)
2532 return SinkCast(CI);
2536 /// MoveExtToFormExtLoad - Move a zext or sext fed by a load into the same
2537 /// basic block as the load, unless conditions are unfavorable. This allows
2538 /// SelectionDAG to fold the extend into the load.
2540 bool CodeGenPrepare::MoveExtToFormExtLoad(Instruction *I) {
2541 // Look for a load being extended.
2542 LoadInst *LI = dyn_cast<LoadInst>(I->getOperand(0));
2543 if (!LI) return false;
2545 // If they're already in the same block, there's nothing to do.
2546 if (LI->getParent() == I->getParent())
2549 // Do not undo the optimization in SinkExtExpand
2551 TLI->getTypeAction(I->getContext(), TLI->getValueType(I->getType())) ==
2552 TargetLowering::TypeExpandInteger)
2555 // If the load has other users and the truncate is not free, this probably
2556 // isn't worthwhile.
2557 if (!LI->hasOneUse() &&
2558 TLI && (TLI->isTypeLegal(TLI->getValueType(LI->getType())) ||
2559 !TLI->isTypeLegal(TLI->getValueType(I->getType()))) &&
2560 !TLI->isTruncateFree(I->getType(), LI->getType()))
2563 // Check whether the target supports casts folded into loads.
2565 if (isa<ZExtInst>(I))
2566 LType = ISD::ZEXTLOAD;
2568 assert(isa<SExtInst>(I) && "Unexpected ext type!");
2569 LType = ISD::SEXTLOAD;
2571 if (TLI && !TLI->isLoadExtLegal(LType, TLI->getValueType(LI->getType())))
2574 // Move the extend into the same block as the load, so that SelectionDAG
2576 I->removeFromParent();
2582 bool CodeGenPrepare::OptimizeExtUses(Instruction *I) {
2583 BasicBlock *DefBB = I->getParent();
2585 // If the result of a {s|z}ext and its source are both live out, rewrite all
2586 // other uses of the source with result of extension.
2587 Value *Src = I->getOperand(0);
2588 if (Src->hasOneUse())
2591 // Only do this xform if truncating is free.
2592 if (TLI && !TLI->isTruncateFree(I->getType(), Src->getType()))
2595 // Only safe to perform the optimization if the source is also defined in
2597 if (!isa<Instruction>(Src) || DefBB != cast<Instruction>(Src)->getParent())
2600 bool DefIsLiveOut = false;
2601 for (User *U : I->users()) {
2602 Instruction *UI = cast<Instruction>(U);
2604 // Figure out which BB this ext is used in.
2605 BasicBlock *UserBB = UI->getParent();
2606 if (UserBB == DefBB) continue;
2607 DefIsLiveOut = true;
2613 // Make sure none of the uses are PHI nodes.
2614 for (User *U : Src->users()) {
2615 Instruction *UI = cast<Instruction>(U);
2616 BasicBlock *UserBB = UI->getParent();
2617 if (UserBB == DefBB) continue;
2618 // Be conservative. We don't want this xform to end up introducing
2619 // reloads just before load / store instructions.
2620 if (isa<PHINode>(UI) || isa<LoadInst>(UI) || isa<StoreInst>(UI))
2624 // InsertedTruncs - Only insert one trunc in each block once.
2625 DenseMap<BasicBlock*, Instruction*> InsertedTruncs;
2627 bool MadeChange = false;
2628 for (Use &U : Src->uses()) {
2629 Instruction *User = cast<Instruction>(U.getUser());
2631 // Figure out which BB this ext is used in.
2632 BasicBlock *UserBB = User->getParent();
2633 if (UserBB == DefBB) continue;
2635 // Both src and def are live in this block. Rewrite the use.
2636 Instruction *&InsertedTrunc = InsertedTruncs[UserBB];
2638 if (!InsertedTrunc) {
2639 BasicBlock::iterator InsertPt = UserBB->getFirstInsertionPt();
2640 InsertedTrunc = new TruncInst(I, Src->getType(), "", InsertPt);
2641 InsertedTruncsSet.insert(InsertedTrunc);
2644 // Replace a use of the {s|z}ext source with a use of the result.
2653 /// isFormingBranchFromSelectProfitable - Returns true if a SelectInst should be
2654 /// turned into an explicit branch.
2655 static bool isFormingBranchFromSelectProfitable(SelectInst *SI) {
2656 // FIXME: This should use the same heuristics as IfConversion to determine
2657 // whether a select is better represented as a branch. This requires that
2658 // branch probability metadata is preserved for the select, which is not the
2661 CmpInst *Cmp = dyn_cast<CmpInst>(SI->getCondition());
2663 // If the branch is predicted right, an out of order CPU can avoid blocking on
2664 // the compare. Emit cmovs on compares with a memory operand as branches to
2665 // avoid stalls on the load from memory. If the compare has more than one use
2666 // there's probably another cmov or setcc around so it's not worth emitting a
2671 Value *CmpOp0 = Cmp->getOperand(0);
2672 Value *CmpOp1 = Cmp->getOperand(1);
2674 // We check that the memory operand has one use to avoid uses of the loaded
2675 // value directly after the compare, making branches unprofitable.
2676 return Cmp->hasOneUse() &&
2677 ((isa<LoadInst>(CmpOp0) && CmpOp0->hasOneUse()) ||
2678 (isa<LoadInst>(CmpOp1) && CmpOp1->hasOneUse()));
2682 /// If we have a SelectInst that will likely profit from branch prediction,
2683 /// turn it into a branch.
2684 bool CodeGenPrepare::OptimizeSelectInst(SelectInst *SI) {
2685 bool VectorCond = !SI->getCondition()->getType()->isIntegerTy(1);
2687 // Can we convert the 'select' to CF ?
2688 if (DisableSelectToBranch || OptSize || !TLI || VectorCond)
2691 TargetLowering::SelectSupportKind SelectKind;
2693 SelectKind = TargetLowering::VectorMaskSelect;
2694 else if (SI->getType()->isVectorTy())
2695 SelectKind = TargetLowering::ScalarCondVectorVal;
2697 SelectKind = TargetLowering::ScalarValSelect;
2699 // Do we have efficient codegen support for this kind of 'selects' ?
2700 if (TLI->isSelectSupported(SelectKind)) {
2701 // We have efficient codegen support for the select instruction.
2702 // Check if it is profitable to keep this 'select'.
2703 if (!TLI->isPredictableSelectExpensive() ||
2704 !isFormingBranchFromSelectProfitable(SI))
2710 // First, we split the block containing the select into 2 blocks.
2711 BasicBlock *StartBlock = SI->getParent();
2712 BasicBlock::iterator SplitPt = ++(BasicBlock::iterator(SI));
2713 BasicBlock *NextBlock = StartBlock->splitBasicBlock(SplitPt, "select.end");
2715 // Create a new block serving as the landing pad for the branch.
2716 BasicBlock *SmallBlock = BasicBlock::Create(SI->getContext(), "select.mid",
2717 NextBlock->getParent(), NextBlock);
2719 // Move the unconditional branch from the block with the select in it into our
2720 // landing pad block.
2721 StartBlock->getTerminator()->eraseFromParent();
2722 BranchInst::Create(NextBlock, SmallBlock);
2724 // Insert the real conditional branch based on the original condition.
2725 BranchInst::Create(NextBlock, SmallBlock, SI->getCondition(), SI);
2727 // The select itself is replaced with a PHI Node.
2728 PHINode *PN = PHINode::Create(SI->getType(), 2, "", NextBlock->begin());
2730 PN->addIncoming(SI->getTrueValue(), StartBlock);
2731 PN->addIncoming(SI->getFalseValue(), SmallBlock);
2732 SI->replaceAllUsesWith(PN);
2733 SI->eraseFromParent();
2735 // Instruct OptimizeBlock to skip to the next block.
2736 CurInstIterator = StartBlock->end();
2737 ++NumSelectsExpanded;
2741 static bool isBroadcastShuffle(ShuffleVectorInst *SVI) {
2742 SmallVector<int, 16> Mask(SVI->getShuffleMask());
2744 for (unsigned i = 0; i < Mask.size(); ++i) {
2745 if (SplatElem != -1 && Mask[i] != -1 && Mask[i] != SplatElem)
2747 SplatElem = Mask[i];
2753 /// Some targets have expensive vector shifts if the lanes aren't all the same
2754 /// (e.g. x86 only introduced "vpsllvd" and friends with AVX2). In these cases
2755 /// it's often worth sinking a shufflevector splat down to its use so that
2756 /// codegen can spot all lanes are identical.
2757 bool CodeGenPrepare::OptimizeShuffleVectorInst(ShuffleVectorInst *SVI) {
2758 BasicBlock *DefBB = SVI->getParent();
2760 // Only do this xform if variable vector shifts are particularly expensive.
2761 if (!TLI || !TLI->isVectorShiftByScalarCheap(SVI->getType()))
2764 // We only expect better codegen by sinking a shuffle if we can recognise a
2766 if (!isBroadcastShuffle(SVI))
2769 // InsertedShuffles - Only insert a shuffle in each block once.
2770 DenseMap<BasicBlock*, Instruction*> InsertedShuffles;
2772 bool MadeChange = false;
2773 for (User *U : SVI->users()) {
2774 Instruction *UI = cast<Instruction>(U);
2776 // Figure out which BB this ext is used in.
2777 BasicBlock *UserBB = UI->getParent();
2778 if (UserBB == DefBB) continue;
2780 // For now only apply this when the splat is used by a shift instruction.
2781 if (!UI->isShift()) continue;
2783 // Everything checks out, sink the shuffle if the user's block doesn't
2784 // already have a copy.
2785 Instruction *&InsertedShuffle = InsertedShuffles[UserBB];
2787 if (!InsertedShuffle) {
2788 BasicBlock::iterator InsertPt = UserBB->getFirstInsertionPt();
2789 InsertedShuffle = new ShuffleVectorInst(SVI->getOperand(0),
2791 SVI->getOperand(2), "", InsertPt);
2794 UI->replaceUsesOfWith(SVI, InsertedShuffle);
2798 // If we removed all uses, nuke the shuffle.
2799 if (SVI->use_empty()) {
2800 SVI->eraseFromParent();
2807 bool CodeGenPrepare::OptimizeInst(Instruction *I) {
2808 if (PHINode *P = dyn_cast<PHINode>(I)) {
2809 // It is possible for very late stage optimizations (such as SimplifyCFG)
2810 // to introduce PHI nodes too late to be cleaned up. If we detect such a
2811 // trivial PHI, go ahead and zap it here.
2812 if (Value *V = SimplifyInstruction(P, TLI ? TLI->getDataLayout() : 0,
2814 P->replaceAllUsesWith(V);
2815 P->eraseFromParent();
2822 if (CastInst *CI = dyn_cast<CastInst>(I)) {
2823 // If the source of the cast is a constant, then this should have
2824 // already been constant folded. The only reason NOT to constant fold
2825 // it is if something (e.g. LSR) was careful to place the constant
2826 // evaluation in a block other than then one that uses it (e.g. to hoist
2827 // the address of globals out of a loop). If this is the case, we don't
2828 // want to forward-subst the cast.
2829 if (isa<Constant>(CI->getOperand(0)))
2832 if (TLI && OptimizeNoopCopyExpression(CI, *TLI))
2835 if (isa<ZExtInst>(I) || isa<SExtInst>(I)) {
2836 if (SinkExtExpand(CI))
2838 bool MadeChange = MoveExtToFormExtLoad(I);
2839 return MadeChange | OptimizeExtUses(I);
2844 if (CmpInst *CI = dyn_cast<CmpInst>(I))
2845 if (!TLI || !TLI->hasMultipleConditionRegisters())
2846 return OptimizeCmpExpression(CI);
2848 if (LoadInst *LI = dyn_cast<LoadInst>(I)) {
2850 return OptimizeMemoryInst(I, I->getOperand(0), LI->getType());
2854 if (StoreInst *SI = dyn_cast<StoreInst>(I)) {
2856 return OptimizeMemoryInst(I, SI->getOperand(1),
2857 SI->getOperand(0)->getType());
2861 if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(I)) {
2862 if (GEPI->hasAllZeroIndices()) {
2863 /// The GEP operand must be a pointer, so must its result -> BitCast
2864 Instruction *NC = new BitCastInst(GEPI->getOperand(0), GEPI->getType(),
2865 GEPI->getName(), GEPI);
2866 GEPI->replaceAllUsesWith(NC);
2867 GEPI->eraseFromParent();
2875 if (CallInst *CI = dyn_cast<CallInst>(I))
2876 return OptimizeCallInst(CI);
2878 if (SelectInst *SI = dyn_cast<SelectInst>(I))
2879 return OptimizeSelectInst(SI);
2881 if (ShuffleVectorInst *SVI = dyn_cast<ShuffleVectorInst>(I))
2882 return OptimizeShuffleVectorInst(SVI);
2887 // In this pass we look for GEP and cast instructions that are used
2888 // across basic blocks and rewrite them to improve basic-block-at-a-time
2890 bool CodeGenPrepare::OptimizeBlock(BasicBlock &BB) {
2892 bool MadeChange = false;
2894 CurInstIterator = BB.begin();
2895 while (CurInstIterator != BB.end())
2896 MadeChange |= OptimizeInst(CurInstIterator++);
2898 MadeChange |= DupRetToEnableTailCallOpts(&BB);
2903 // llvm.dbg.value is far away from the value then iSel may not be able
2904 // handle it properly. iSel will drop llvm.dbg.value if it can not
2905 // find a node corresponding to the value.
2906 bool CodeGenPrepare::PlaceDbgValues(Function &F) {
2907 bool MadeChange = false;
2908 for (Function::iterator I = F.begin(), E = F.end(); I != E; ++I) {
2909 Instruction *PrevNonDbgInst = NULL;
2910 for (BasicBlock::iterator BI = I->begin(), BE = I->end(); BI != BE;) {
2911 Instruction *Insn = BI; ++BI;
2912 DbgValueInst *DVI = dyn_cast<DbgValueInst>(Insn);
2914 PrevNonDbgInst = Insn;
2918 Instruction *VI = dyn_cast_or_null<Instruction>(DVI->getValue());
2919 if (VI && VI != PrevNonDbgInst && !VI->isTerminator()) {
2920 DEBUG(dbgs() << "Moving Debug Value before :\n" << *DVI << ' ' << *VI);
2921 DVI->removeFromParent();
2922 if (isa<PHINode>(VI))
2923 DVI->insertBefore(VI->getParent()->getFirstInsertionPt());
2925 DVI->insertAfter(VI);