1 //===- CodeGenPrepare.cpp - Prepare a function for code generation --------===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // This pass munges the code in the input function to better prepare it for
11 // SelectionDAG-based code generation. This works around limitations in it's
12 // basic-block-at-a-time approach. It should eventually be removed.
14 //===----------------------------------------------------------------------===//
16 #define DEBUG_TYPE "codegenprepare"
17 #include "llvm/Transforms/Scalar.h"
18 #include "llvm/Constants.h"
19 #include "llvm/DerivedTypes.h"
20 #include "llvm/Function.h"
21 #include "llvm/InlineAsm.h"
22 #include "llvm/Instructions.h"
23 #include "llvm/IntrinsicInst.h"
24 #include "llvm/Pass.h"
25 #include "llvm/Analysis/InstructionSimplify.h"
26 #include "llvm/Analysis/ProfileInfo.h"
27 #include "llvm/Target/TargetData.h"
28 #include "llvm/Target/TargetLowering.h"
29 #include "llvm/Transforms/Utils/AddrModeMatcher.h"
30 #include "llvm/Transforms/Utils/BasicBlockUtils.h"
31 #include "llvm/Transforms/Utils/Local.h"
32 #include "llvm/Transforms/Utils/BuildLibCalls.h"
33 #include "llvm/ADT/DenseMap.h"
34 #include "llvm/ADT/SmallSet.h"
35 #include "llvm/ADT/Statistic.h"
36 #include "llvm/Assembly/Writer.h"
37 #include "llvm/Support/CallSite.h"
38 #include "llvm/Support/CommandLine.h"
39 #include "llvm/Support/Debug.h"
40 #include "llvm/Support/GetElementPtrTypeIterator.h"
41 #include "llvm/Support/PatternMatch.h"
42 #include "llvm/Support/raw_ostream.h"
43 #include "llvm/Support/IRBuilder.h"
45 using namespace llvm::PatternMatch;
47 STATISTIC(NumBlocksElim, "Number of blocks eliminated");
48 STATISTIC(NumPHIsElim, "Number of trivial PHIs eliminated");
49 STATISTIC(NumGEPsElim, "Number of GEPs converted to casts");
50 STATISTIC(NumCmpUses, "Number of uses of Cmp expressions replaced with uses of "
52 STATISTIC(NumCastUses, "Number of uses of Cast expressions replaced with uses "
54 STATISTIC(NumMemoryInsts, "Number of memory instructions whose address "
55 "computations were sunk");
56 STATISTIC(NumExtsMoved, "Number of [s|z]ext instructions combined with loads");
57 STATISTIC(NumExtUses, "Number of uses of [s|z]ext instructions optimized");
60 CriticalEdgeSplit("cgp-critical-edge-splitting",
61 cl::desc("Split critical edges during codegen prepare"),
62 cl::init(false), cl::Hidden);
65 class CodeGenPrepare : public FunctionPass {
66 /// TLI - Keep a pointer of a TargetLowering to consult for determining
67 /// transformation profitability.
68 const TargetLowering *TLI;
71 /// BackEdges - Keep a set of all the loop back edges.
73 SmallSet<std::pair<const BasicBlock*, const BasicBlock*>, 8> BackEdges;
75 // Keeps track of non-local addresses that have been sunk into a block. This
76 // allows us to avoid inserting duplicate code for blocks with multiple
77 // load/stores of the same address.
78 DenseMap<Value*, Value*> SunkAddrs;
81 static char ID; // Pass identification, replacement for typeid
82 explicit CodeGenPrepare(const TargetLowering *tli = 0)
83 : FunctionPass(ID), TLI(tli) {
84 initializeCodeGenPreparePass(*PassRegistry::getPassRegistry());
86 bool runOnFunction(Function &F);
88 virtual void getAnalysisUsage(AnalysisUsage &AU) const {
89 AU.addPreserved<ProfileInfo>();
92 virtual void releaseMemory() {
97 bool EliminateMostlyEmptyBlocks(Function &F);
98 bool CanMergeBlocks(const BasicBlock *BB, const BasicBlock *DestBB) const;
99 void EliminateMostlyEmptyBlock(BasicBlock *BB);
100 bool OptimizeBlock(BasicBlock &BB);
101 bool OptimizeMemoryInst(Instruction *I, Value *Addr, const Type *AccessTy,
102 DenseMap<Value*,Value*> &SunkAddrs);
103 bool OptimizeInlineAsmInst(Instruction *I, CallSite CS,
104 DenseMap<Value*,Value*> &SunkAddrs);
105 bool OptimizeCallInst(CallInst *CI);
106 bool MoveExtToFormExtLoad(Instruction *I);
107 bool OptimizeExtUses(Instruction *I);
108 void findLoopBackEdges(const Function &F);
112 char CodeGenPrepare::ID = 0;
113 INITIALIZE_PASS(CodeGenPrepare, "codegenprepare",
114 "Optimize for code generation", false, false)
116 FunctionPass *llvm::createCodeGenPreparePass(const TargetLowering *TLI) {
117 return new CodeGenPrepare(TLI);
120 /// findLoopBackEdges - Do a DFS walk to find loop back edges.
122 void CodeGenPrepare::findLoopBackEdges(const Function &F) {
123 SmallVector<std::pair<const BasicBlock*,const BasicBlock*>, 32> Edges;
124 FindFunctionBackedges(F, Edges);
126 BackEdges.insert(Edges.begin(), Edges.end());
130 bool CodeGenPrepare::runOnFunction(Function &F) {
131 bool EverMadeChange = false;
133 PFI = getAnalysisIfAvailable<ProfileInfo>();
134 // First pass, eliminate blocks that contain only PHI nodes and an
135 // unconditional branch.
136 EverMadeChange |= EliminateMostlyEmptyBlocks(F);
138 // Now find loop back edges, but only if they are being used to decide which
139 // critical edges to split.
140 if (CriticalEdgeSplit)
141 findLoopBackEdges(F);
143 bool MadeChange = true;
146 for (Function::iterator BB = F.begin(), E = F.end(); BB != E; ++BB)
147 MadeChange |= OptimizeBlock(*BB);
148 EverMadeChange |= MadeChange;
153 return EverMadeChange;
156 /// EliminateMostlyEmptyBlocks - eliminate blocks that contain only PHI nodes,
157 /// debug info directives, and an unconditional branch. Passes before isel
158 /// (e.g. LSR/loopsimplify) often split edges in ways that are non-optimal for
159 /// isel. Start by eliminating these blocks so we can split them the way we
161 bool CodeGenPrepare::EliminateMostlyEmptyBlocks(Function &F) {
162 bool MadeChange = false;
163 // Note that this intentionally skips the entry block.
164 for (Function::iterator I = ++F.begin(), E = F.end(); I != E; ) {
165 BasicBlock *BB = I++;
167 // If this block doesn't end with an uncond branch, ignore it.
168 BranchInst *BI = dyn_cast<BranchInst>(BB->getTerminator());
169 if (!BI || !BI->isUnconditional())
172 // If the instruction before the branch (skipping debug info) isn't a phi
173 // node, then other stuff is happening here.
174 BasicBlock::iterator BBI = BI;
175 if (BBI != BB->begin()) {
177 while (isa<DbgInfoIntrinsic>(BBI)) {
178 if (BBI == BB->begin())
182 if (!isa<DbgInfoIntrinsic>(BBI) && !isa<PHINode>(BBI))
186 // Do not break infinite loops.
187 BasicBlock *DestBB = BI->getSuccessor(0);
191 if (!CanMergeBlocks(BB, DestBB))
194 EliminateMostlyEmptyBlock(BB);
200 /// CanMergeBlocks - Return true if we can merge BB into DestBB if there is a
201 /// single uncond branch between them, and BB contains no other non-phi
203 bool CodeGenPrepare::CanMergeBlocks(const BasicBlock *BB,
204 const BasicBlock *DestBB) const {
205 // We only want to eliminate blocks whose phi nodes are used by phi nodes in
206 // the successor. If there are more complex condition (e.g. preheaders),
207 // don't mess around with them.
208 BasicBlock::const_iterator BBI = BB->begin();
209 while (const PHINode *PN = dyn_cast<PHINode>(BBI++)) {
210 for (Value::const_use_iterator UI = PN->use_begin(), E = PN->use_end();
212 const Instruction *User = cast<Instruction>(*UI);
213 if (User->getParent() != DestBB || !isa<PHINode>(User))
215 // If User is inside DestBB block and it is a PHINode then check
216 // incoming value. If incoming value is not from BB then this is
217 // a complex condition (e.g. preheaders) we want to avoid here.
218 if (User->getParent() == DestBB) {
219 if (const PHINode *UPN = dyn_cast<PHINode>(User))
220 for (unsigned I = 0, E = UPN->getNumIncomingValues(); I != E; ++I) {
221 Instruction *Insn = dyn_cast<Instruction>(UPN->getIncomingValue(I));
222 if (Insn && Insn->getParent() == BB &&
223 Insn->getParent() != UPN->getIncomingBlock(I))
230 // If BB and DestBB contain any common predecessors, then the phi nodes in BB
231 // and DestBB may have conflicting incoming values for the block. If so, we
232 // can't merge the block.
233 const PHINode *DestBBPN = dyn_cast<PHINode>(DestBB->begin());
234 if (!DestBBPN) return true; // no conflict.
236 // Collect the preds of BB.
237 SmallPtrSet<const BasicBlock*, 16> BBPreds;
238 if (const PHINode *BBPN = dyn_cast<PHINode>(BB->begin())) {
239 // It is faster to get preds from a PHI than with pred_iterator.
240 for (unsigned i = 0, e = BBPN->getNumIncomingValues(); i != e; ++i)
241 BBPreds.insert(BBPN->getIncomingBlock(i));
243 BBPreds.insert(pred_begin(BB), pred_end(BB));
246 // Walk the preds of DestBB.
247 for (unsigned i = 0, e = DestBBPN->getNumIncomingValues(); i != e; ++i) {
248 BasicBlock *Pred = DestBBPN->getIncomingBlock(i);
249 if (BBPreds.count(Pred)) { // Common predecessor?
250 BBI = DestBB->begin();
251 while (const PHINode *PN = dyn_cast<PHINode>(BBI++)) {
252 const Value *V1 = PN->getIncomingValueForBlock(Pred);
253 const Value *V2 = PN->getIncomingValueForBlock(BB);
255 // If V2 is a phi node in BB, look up what the mapped value will be.
256 if (const PHINode *V2PN = dyn_cast<PHINode>(V2))
257 if (V2PN->getParent() == BB)
258 V2 = V2PN->getIncomingValueForBlock(Pred);
260 // If there is a conflict, bail out.
261 if (V1 != V2) return false;
270 /// EliminateMostlyEmptyBlock - Eliminate a basic block that have only phi's and
271 /// an unconditional branch in it.
272 void CodeGenPrepare::EliminateMostlyEmptyBlock(BasicBlock *BB) {
273 BranchInst *BI = cast<BranchInst>(BB->getTerminator());
274 BasicBlock *DestBB = BI->getSuccessor(0);
276 DEBUG(dbgs() << "MERGING MOSTLY EMPTY BLOCKS - BEFORE:\n" << *BB << *DestBB);
278 // If the destination block has a single pred, then this is a trivial edge,
280 if (BasicBlock *SinglePred = DestBB->getSinglePredecessor()) {
281 if (SinglePred != DestBB) {
282 // Remember if SinglePred was the entry block of the function. If so, we
283 // will need to move BB back to the entry position.
284 bool isEntry = SinglePred == &SinglePred->getParent()->getEntryBlock();
285 MergeBasicBlockIntoOnlyPred(DestBB, this);
287 if (isEntry && BB != &BB->getParent()->getEntryBlock())
288 BB->moveBefore(&BB->getParent()->getEntryBlock());
290 DEBUG(dbgs() << "AFTER:\n" << *DestBB << "\n\n\n");
295 // Otherwise, we have multiple predecessors of BB. Update the PHIs in DestBB
296 // to handle the new incoming edges it is about to have.
298 for (BasicBlock::iterator BBI = DestBB->begin();
299 (PN = dyn_cast<PHINode>(BBI)); ++BBI) {
300 // Remove the incoming value for BB, and remember it.
301 Value *InVal = PN->removeIncomingValue(BB, false);
303 // Two options: either the InVal is a phi node defined in BB or it is some
304 // value that dominates BB.
305 PHINode *InValPhi = dyn_cast<PHINode>(InVal);
306 if (InValPhi && InValPhi->getParent() == BB) {
307 // Add all of the input values of the input PHI as inputs of this phi.
308 for (unsigned i = 0, e = InValPhi->getNumIncomingValues(); i != e; ++i)
309 PN->addIncoming(InValPhi->getIncomingValue(i),
310 InValPhi->getIncomingBlock(i));
312 // Otherwise, add one instance of the dominating value for each edge that
313 // we will be adding.
314 if (PHINode *BBPN = dyn_cast<PHINode>(BB->begin())) {
315 for (unsigned i = 0, e = BBPN->getNumIncomingValues(); i != e; ++i)
316 PN->addIncoming(InVal, BBPN->getIncomingBlock(i));
318 for (pred_iterator PI = pred_begin(BB), E = pred_end(BB); PI != E; ++PI)
319 PN->addIncoming(InVal, *PI);
324 // The PHIs are now updated, change everything that refers to BB to use
325 // DestBB and remove BB.
326 BB->replaceAllUsesWith(DestBB);
328 PFI->replaceAllUses(BB, DestBB);
329 PFI->removeEdge(ProfileInfo::getEdge(BB, DestBB));
331 BB->eraseFromParent();
334 DEBUG(dbgs() << "AFTER:\n" << *DestBB << "\n\n\n");
337 /// FindReusablePredBB - Check all of the predecessors of the block DestPHI
338 /// lives in to see if there is a block that we can reuse as a critical edge
340 static BasicBlock *FindReusablePredBB(PHINode *DestPHI, BasicBlock *TIBB) {
341 BasicBlock *Dest = DestPHI->getParent();
343 /// TIPHIValues - This array is lazily computed to determine the values of
344 /// PHIs in Dest that TI would provide.
345 SmallVector<Value*, 32> TIPHIValues;
347 /// TIBBEntryNo - This is a cache to speed up pred queries for TIBB.
348 unsigned TIBBEntryNo = 0;
350 // Check to see if Dest has any blocks that can be used as a split edge for
352 for (unsigned pi = 0, e = DestPHI->getNumIncomingValues(); pi != e; ++pi) {
353 BasicBlock *Pred = DestPHI->getIncomingBlock(pi);
354 // To be usable, the pred has to end with an uncond branch to the dest.
355 BranchInst *PredBr = dyn_cast<BranchInst>(Pred->getTerminator());
356 if (!PredBr || !PredBr->isUnconditional())
358 // Must be empty other than the branch and debug info.
359 BasicBlock::iterator I = Pred->begin();
360 while (isa<DbgInfoIntrinsic>(I))
364 // Cannot be the entry block; its label does not get emitted.
365 if (Pred == &Dest->getParent()->getEntryBlock())
368 // Finally, since we know that Dest has phi nodes in it, we have to make
369 // sure that jumping to Pred will have the same effect as going to Dest in
370 // terms of PHI values.
373 unsigned PredEntryNo = pi;
375 bool FoundMatch = true;
376 for (BasicBlock::iterator I = Dest->begin();
377 (PN = dyn_cast<PHINode>(I)); ++I, ++PHINo) {
378 if (PHINo == TIPHIValues.size()) {
379 if (PN->getIncomingBlock(TIBBEntryNo) != TIBB)
380 TIBBEntryNo = PN->getBasicBlockIndex(TIBB);
381 TIPHIValues.push_back(PN->getIncomingValue(TIBBEntryNo));
384 // If the PHI entry doesn't work, we can't use this pred.
385 if (PN->getIncomingBlock(PredEntryNo) != Pred)
386 PredEntryNo = PN->getBasicBlockIndex(Pred);
388 if (TIPHIValues[PHINo] != PN->getIncomingValue(PredEntryNo)) {
394 // If we found a workable predecessor, change TI to branch to Succ.
402 /// SplitEdgeNicely - Split the critical edge from TI to its specified
403 /// successor if it will improve codegen. We only do this if the successor has
404 /// phi nodes (otherwise critical edges are ok). If there is already another
405 /// predecessor of the succ that is empty (and thus has no phi nodes), use it
406 /// instead of introducing a new block.
407 static void SplitEdgeNicely(TerminatorInst *TI, unsigned SuccNum,
408 SmallSet<std::pair<const BasicBlock*,
409 const BasicBlock*>, 8> &BackEdges,
411 BasicBlock *TIBB = TI->getParent();
412 BasicBlock *Dest = TI->getSuccessor(SuccNum);
413 assert(isa<PHINode>(Dest->begin()) &&
414 "This should only be called if Dest has a PHI!");
415 PHINode *DestPHI = cast<PHINode>(Dest->begin());
417 // Do not split edges to EH landing pads.
418 if (InvokeInst *Invoke = dyn_cast<InvokeInst>(TI))
419 if (Invoke->getSuccessor(1) == Dest)
422 // As a hack, never split backedges of loops. Even though the copy for any
423 // PHIs inserted on the backedge would be dead for exits from the loop, we
424 // assume that the cost of *splitting* the backedge would be too high.
425 if (BackEdges.count(std::make_pair(TIBB, Dest)))
428 if (BasicBlock *ReuseBB = FindReusablePredBB(DestPHI, TIBB)) {
429 ProfileInfo *PFI = P->getAnalysisIfAvailable<ProfileInfo>();
431 PFI->splitEdge(TIBB, Dest, ReuseBB);
432 Dest->removePredecessor(TIBB);
433 TI->setSuccessor(SuccNum, ReuseBB);
437 SplitCriticalEdge(TI, SuccNum, P, true);
441 /// OptimizeNoopCopyExpression - If the specified cast instruction is a noop
442 /// copy (e.g. it's casting from one pointer type to another, i32->i8 on PPC),
443 /// sink it into user blocks to reduce the number of virtual
444 /// registers that must be created and coalesced.
446 /// Return true if any changes are made.
448 static bool OptimizeNoopCopyExpression(CastInst *CI, const TargetLowering &TLI){
449 // If this is a noop copy,
450 EVT SrcVT = TLI.getValueType(CI->getOperand(0)->getType());
451 EVT DstVT = TLI.getValueType(CI->getType());
453 // This is an fp<->int conversion?
454 if (SrcVT.isInteger() != DstVT.isInteger())
457 // If this is an extension, it will be a zero or sign extension, which
459 if (SrcVT.bitsLT(DstVT)) return false;
461 // If these values will be promoted, find out what they will be promoted
462 // to. This helps us consider truncates on PPC as noop copies when they
464 if (TLI.getTypeAction(SrcVT) == TargetLowering::Promote)
465 SrcVT = TLI.getTypeToTransformTo(CI->getContext(), SrcVT);
466 if (TLI.getTypeAction(DstVT) == TargetLowering::Promote)
467 DstVT = TLI.getTypeToTransformTo(CI->getContext(), DstVT);
469 // If, after promotion, these are the same types, this is a noop copy.
473 BasicBlock *DefBB = CI->getParent();
475 /// InsertedCasts - Only insert a cast in each block once.
476 DenseMap<BasicBlock*, CastInst*> InsertedCasts;
478 bool MadeChange = false;
479 for (Value::use_iterator UI = CI->use_begin(), E = CI->use_end();
481 Use &TheUse = UI.getUse();
482 Instruction *User = cast<Instruction>(*UI);
484 // Figure out which BB this cast is used in. For PHI's this is the
485 // appropriate predecessor block.
486 BasicBlock *UserBB = User->getParent();
487 if (PHINode *PN = dyn_cast<PHINode>(User)) {
488 UserBB = PN->getIncomingBlock(UI);
491 // Preincrement use iterator so we don't invalidate it.
494 // If this user is in the same block as the cast, don't change the cast.
495 if (UserBB == DefBB) continue;
497 // If we have already inserted a cast into this block, use it.
498 CastInst *&InsertedCast = InsertedCasts[UserBB];
501 BasicBlock::iterator InsertPt = UserBB->getFirstNonPHI();
504 CastInst::Create(CI->getOpcode(), CI->getOperand(0), CI->getType(), "",
509 // Replace a use of the cast with a use of the new cast.
510 TheUse = InsertedCast;
514 // If we removed all uses, nuke the cast.
515 if (CI->use_empty()) {
516 CI->eraseFromParent();
523 /// OptimizeCmpExpression - sink the given CmpInst into user blocks to reduce
524 /// the number of virtual registers that must be created and coalesced. This is
525 /// a clear win except on targets with multiple condition code registers
526 /// (PowerPC), where it might lose; some adjustment may be wanted there.
528 /// Return true if any changes are made.
529 static bool OptimizeCmpExpression(CmpInst *CI) {
530 BasicBlock *DefBB = CI->getParent();
532 /// InsertedCmp - Only insert a cmp in each block once.
533 DenseMap<BasicBlock*, CmpInst*> InsertedCmps;
535 bool MadeChange = false;
536 for (Value::use_iterator UI = CI->use_begin(), E = CI->use_end();
538 Use &TheUse = UI.getUse();
539 Instruction *User = cast<Instruction>(*UI);
541 // Preincrement use iterator so we don't invalidate it.
544 // Don't bother for PHI nodes.
545 if (isa<PHINode>(User))
548 // Figure out which BB this cmp is used in.
549 BasicBlock *UserBB = User->getParent();
551 // If this user is in the same block as the cmp, don't change the cmp.
552 if (UserBB == DefBB) continue;
554 // If we have already inserted a cmp into this block, use it.
555 CmpInst *&InsertedCmp = InsertedCmps[UserBB];
558 BasicBlock::iterator InsertPt = UserBB->getFirstNonPHI();
561 CmpInst::Create(CI->getOpcode(),
562 CI->getPredicate(), CI->getOperand(0),
563 CI->getOperand(1), "", InsertPt);
567 // Replace a use of the cmp with a use of the new cmp.
568 TheUse = InsertedCmp;
572 // If we removed all uses, nuke the cmp.
574 CI->eraseFromParent();
580 class CodeGenPrepareFortifiedLibCalls : public SimplifyFortifiedLibCalls {
582 void replaceCall(Value *With) {
583 CI->replaceAllUsesWith(With);
584 CI->eraseFromParent();
586 bool isFoldable(unsigned SizeCIOp, unsigned, bool) const {
587 if (ConstantInt *SizeCI =
588 dyn_cast<ConstantInt>(CI->getArgOperand(SizeCIOp)))
589 return SizeCI->isAllOnesValue();
593 } // end anonymous namespace
595 bool CodeGenPrepare::OptimizeCallInst(CallInst *CI) {
596 // Lower all uses of llvm.objectsize.*
597 IntrinsicInst *II = dyn_cast<IntrinsicInst>(CI);
598 if (II && II->getIntrinsicID() == Intrinsic::objectsize) {
599 bool Min = (cast<ConstantInt>(II->getArgOperand(1))->getZExtValue() == 1);
600 const Type *ReturnTy = CI->getType();
601 Constant *RetVal = ConstantInt::get(ReturnTy, Min ? 0 : -1ULL);
602 CI->replaceAllUsesWith(RetVal);
603 CI->eraseFromParent();
607 // From here on out we're working with named functions.
608 if (CI->getCalledFunction() == 0) return false;
610 // We'll need TargetData from here on out.
611 const TargetData *TD = TLI ? TLI->getTargetData() : 0;
612 if (!TD) return false;
614 // Lower all default uses of _chk calls. This is very similar
615 // to what InstCombineCalls does, but here we are only lowering calls
616 // that have the default "don't know" as the objectsize. Anything else
617 // should be left alone.
618 CodeGenPrepareFortifiedLibCalls Simplifier;
619 return Simplifier.fold(CI, TD);
621 //===----------------------------------------------------------------------===//
622 // Memory Optimization
623 //===----------------------------------------------------------------------===//
625 /// IsNonLocalValue - Return true if the specified values are defined in a
626 /// different basic block than BB.
627 static bool IsNonLocalValue(Value *V, BasicBlock *BB) {
628 if (Instruction *I = dyn_cast<Instruction>(V))
629 return I->getParent() != BB;
633 /// OptimizeMemoryInst - Load and Store Instructions often have
634 /// addressing modes that can do significant amounts of computation. As such,
635 /// instruction selection will try to get the load or store to do as much
636 /// computation as possible for the program. The problem is that isel can only
637 /// see within a single block. As such, we sink as much legal addressing mode
638 /// stuff into the block as possible.
640 /// This method is used to optimize both load/store and inline asms with memory
642 bool CodeGenPrepare::OptimizeMemoryInst(Instruction *MemoryInst, Value *Addr,
643 const Type *AccessTy,
644 DenseMap<Value*,Value*> &SunkAddrs) {
647 // Try to collapse single-value PHI nodes. This is necessary to undo
648 // unprofitable PRE transformations.
649 SmallVector<Value*, 8> worklist;
650 SmallPtrSet<Value*, 16> Visited;
651 worklist.push_back(Addr);
653 // Use a worklist to iteratively look through PHI nodes, and ensure that
654 // the addressing mode obtained from the non-PHI roots of the graph
656 Value *Consensus = 0;
657 unsigned NumUses = 0;
658 SmallVector<Instruction*, 16> AddrModeInsts;
659 ExtAddrMode AddrMode;
660 while (!worklist.empty()) {
661 Value *V = worklist.back();
664 // Break use-def graph loops.
665 if (Visited.count(V)) {
672 // For a PHI node, push all of its incoming values.
673 if (PHINode *P = dyn_cast<PHINode>(V)) {
674 for (unsigned i = 0, e = P->getNumIncomingValues(); i != e; ++i)
675 worklist.push_back(P->getIncomingValue(i));
679 // For non-PHIs, determine the addressing mode being computed.
680 SmallVector<Instruction*, 16> NewAddrModeInsts;
681 ExtAddrMode NewAddrMode =
682 AddressingModeMatcher::Match(V, AccessTy,MemoryInst,
683 NewAddrModeInsts, *TLI);
685 // Ensure that the obtained addressing mode is equivalent to that obtained
686 // for all other roots of the PHI traversal. Also, when choosing one
687 // such root as representative, select the one with the most uses in order
688 // to keep the cost modeling heuristics in AddressingModeMatcher applicable.
689 if (!Consensus || NewAddrMode == AddrMode) {
690 if (V->getNumUses() > NumUses) {
692 NumUses = V->getNumUses();
693 AddrMode = NewAddrMode;
694 AddrModeInsts = NewAddrModeInsts;
703 // If the addressing mode couldn't be determined, or if multiple different
704 // ones were determined, bail out now.
705 if (!Consensus) return false;
707 // Check to see if any of the instructions supersumed by this addr mode are
708 // non-local to I's BB.
709 bool AnyNonLocal = false;
710 for (unsigned i = 0, e = AddrModeInsts.size(); i != e; ++i) {
711 if (IsNonLocalValue(AddrModeInsts[i], MemoryInst->getParent())) {
717 // If all the instructions matched are already in this BB, don't do anything.
719 DEBUG(dbgs() << "CGP: Found local addrmode: " << AddrMode << "\n");
723 // Insert this computation right after this user. Since our caller is
724 // scanning from the top of the BB to the bottom, reuse of the expr are
725 // guaranteed to happen later.
726 BasicBlock::iterator InsertPt = MemoryInst;
728 // Now that we determined the addressing expression we want to use and know
729 // that we have to sink it into this block. Check to see if we have already
730 // done this for some other load/store instr in this block. If so, reuse the
732 Value *&SunkAddr = SunkAddrs[Addr];
734 DEBUG(dbgs() << "CGP: Reusing nonlocal addrmode: " << AddrMode << " for "
736 if (SunkAddr->getType() != Addr->getType())
737 SunkAddr = new BitCastInst(SunkAddr, Addr->getType(), "tmp", InsertPt);
739 DEBUG(dbgs() << "CGP: SINKING nonlocal addrmode: " << AddrMode << " for "
741 const Type *IntPtrTy =
742 TLI->getTargetData()->getIntPtrType(AccessTy->getContext());
746 // Start with the base register. Do this first so that subsequent address
747 // matching finds it last, which will prevent it from trying to match it
748 // as the scaled value in case it happens to be a mul. That would be
749 // problematic if we've sunk a different mul for the scale, because then
750 // we'd end up sinking both muls.
751 if (AddrMode.BaseReg) {
752 Value *V = AddrMode.BaseReg;
753 if (V->getType()->isPointerTy())
754 V = new PtrToIntInst(V, IntPtrTy, "sunkaddr", InsertPt);
755 if (V->getType() != IntPtrTy)
756 V = CastInst::CreateIntegerCast(V, IntPtrTy, /*isSigned=*/true,
757 "sunkaddr", InsertPt);
761 // Add the scale value.
762 if (AddrMode.Scale) {
763 Value *V = AddrMode.ScaledReg;
764 if (V->getType() == IntPtrTy) {
766 } else if (V->getType()->isPointerTy()) {
767 V = new PtrToIntInst(V, IntPtrTy, "sunkaddr", InsertPt);
768 } else if (cast<IntegerType>(IntPtrTy)->getBitWidth() <
769 cast<IntegerType>(V->getType())->getBitWidth()) {
770 V = new TruncInst(V, IntPtrTy, "sunkaddr", InsertPt);
772 V = new SExtInst(V, IntPtrTy, "sunkaddr", InsertPt);
774 if (AddrMode.Scale != 1)
775 V = BinaryOperator::CreateMul(V, ConstantInt::get(IntPtrTy,
777 "sunkaddr", InsertPt);
779 Result = BinaryOperator::CreateAdd(Result, V, "sunkaddr", InsertPt);
784 // Add in the BaseGV if present.
785 if (AddrMode.BaseGV) {
786 Value *V = new PtrToIntInst(AddrMode.BaseGV, IntPtrTy, "sunkaddr",
789 Result = BinaryOperator::CreateAdd(Result, V, "sunkaddr", InsertPt);
794 // Add in the Base Offset if present.
795 if (AddrMode.BaseOffs) {
796 Value *V = ConstantInt::get(IntPtrTy, AddrMode.BaseOffs);
798 Result = BinaryOperator::CreateAdd(Result, V, "sunkaddr", InsertPt);
804 SunkAddr = Constant::getNullValue(Addr->getType());
806 SunkAddr = new IntToPtrInst(Result, Addr->getType(), "sunkaddr",InsertPt);
809 MemoryInst->replaceUsesOfWith(Repl, SunkAddr);
811 if (Repl->use_empty()) {
812 RecursivelyDeleteTriviallyDeadInstructions(Repl);
813 // This address is now available for reassignment, so erase the table entry;
814 // we don't want to match some completely different instruction.
821 /// OptimizeInlineAsmInst - If there are any memory operands, use
822 /// OptimizeMemoryInst to sink their address computing into the block when
823 /// possible / profitable.
824 bool CodeGenPrepare::OptimizeInlineAsmInst(Instruction *I, CallSite CS,
825 DenseMap<Value*,Value*> &SunkAddrs) {
826 bool MadeChange = false;
828 TargetLowering::AsmOperandInfoVector TargetConstraints = TLI->ParseConstraints(CS);
830 for (unsigned i = 0, e = TargetConstraints.size(); i != e; ++i) {
831 TargetLowering::AsmOperandInfo &OpInfo = TargetConstraints[i];
833 // Compute the constraint code and ConstraintType to use.
834 TLI->ComputeConstraintToUse(OpInfo, SDValue());
836 if (OpInfo.ConstraintType == TargetLowering::C_Memory &&
838 Value *OpVal = const_cast<Value *>(CS.getArgument(ArgNo++));
839 MadeChange |= OptimizeMemoryInst(I, OpVal, OpVal->getType(), SunkAddrs);
840 } else if (OpInfo.Type == InlineAsm::isInput)
847 /// MoveExtToFormExtLoad - Move a zext or sext fed by a load into the same
848 /// basic block as the load, unless conditions are unfavorable. This allows
849 /// SelectionDAG to fold the extend into the load.
851 bool CodeGenPrepare::MoveExtToFormExtLoad(Instruction *I) {
852 // Look for a load being extended.
853 LoadInst *LI = dyn_cast<LoadInst>(I->getOperand(0));
854 if (!LI) return false;
856 // If they're already in the same block, there's nothing to do.
857 if (LI->getParent() == I->getParent())
860 // If the load has other users and the truncate is not free, this probably
862 if (!LI->hasOneUse() &&
863 TLI && (TLI->isTypeLegal(TLI->getValueType(LI->getType())) ||
864 !TLI->isTypeLegal(TLI->getValueType(I->getType()))) &&
865 !TLI->isTruncateFree(I->getType(), LI->getType()))
868 // Check whether the target supports casts folded into loads.
870 if (isa<ZExtInst>(I))
871 LType = ISD::ZEXTLOAD;
873 assert(isa<SExtInst>(I) && "Unexpected ext type!");
874 LType = ISD::SEXTLOAD;
876 if (TLI && !TLI->isLoadExtLegal(LType, TLI->getValueType(LI->getType())))
879 // Move the extend into the same block as the load, so that SelectionDAG
881 I->removeFromParent();
887 bool CodeGenPrepare::OptimizeExtUses(Instruction *I) {
888 BasicBlock *DefBB = I->getParent();
890 // If the result of a {s|z}ext and its source are both live out, rewrite all
891 // other uses of the source with result of extension.
892 Value *Src = I->getOperand(0);
893 if (Src->hasOneUse())
896 // Only do this xform if truncating is free.
897 if (TLI && !TLI->isTruncateFree(I->getType(), Src->getType()))
900 // Only safe to perform the optimization if the source is also defined in
902 if (!isa<Instruction>(Src) || DefBB != cast<Instruction>(Src)->getParent())
905 bool DefIsLiveOut = false;
906 for (Value::use_iterator UI = I->use_begin(), E = I->use_end();
908 Instruction *User = cast<Instruction>(*UI);
910 // Figure out which BB this ext is used in.
911 BasicBlock *UserBB = User->getParent();
912 if (UserBB == DefBB) continue;
919 // Make sure non of the uses are PHI nodes.
920 for (Value::use_iterator UI = Src->use_begin(), E = Src->use_end();
922 Instruction *User = cast<Instruction>(*UI);
923 BasicBlock *UserBB = User->getParent();
924 if (UserBB == DefBB) continue;
925 // Be conservative. We don't want this xform to end up introducing
926 // reloads just before load / store instructions.
927 if (isa<PHINode>(User) || isa<LoadInst>(User) || isa<StoreInst>(User))
931 // InsertedTruncs - Only insert one trunc in each block once.
932 DenseMap<BasicBlock*, Instruction*> InsertedTruncs;
934 bool MadeChange = false;
935 for (Value::use_iterator UI = Src->use_begin(), E = Src->use_end();
937 Use &TheUse = UI.getUse();
938 Instruction *User = cast<Instruction>(*UI);
940 // Figure out which BB this ext is used in.
941 BasicBlock *UserBB = User->getParent();
942 if (UserBB == DefBB) continue;
944 // Both src and def are live in this block. Rewrite the use.
945 Instruction *&InsertedTrunc = InsertedTruncs[UserBB];
947 if (!InsertedTrunc) {
948 BasicBlock::iterator InsertPt = UserBB->getFirstNonPHI();
950 InsertedTrunc = new TruncInst(I, Src->getType(), "", InsertPt);
953 // Replace a use of the {s|z}ext source with a use of the result.
954 TheUse = InsertedTrunc;
962 // In this pass we look for GEP and cast instructions that are used
963 // across basic blocks and rewrite them to improve basic-block-at-a-time
965 bool CodeGenPrepare::OptimizeBlock(BasicBlock &BB) {
966 bool MadeChange = false;
968 // Split all critical edges where the dest block has a PHI.
969 if (CriticalEdgeSplit) {
970 TerminatorInst *BBTI = BB.getTerminator();
971 if (BBTI->getNumSuccessors() > 1 && !isa<IndirectBrInst>(BBTI)) {
972 for (unsigned i = 0, e = BBTI->getNumSuccessors(); i != e; ++i) {
973 BasicBlock *SuccBB = BBTI->getSuccessor(i);
974 if (isa<PHINode>(SuccBB->begin()) && isCriticalEdge(BBTI, i, true))
975 SplitEdgeNicely(BBTI, i, BackEdges, this);
982 for (BasicBlock::iterator BBI = BB.begin(), E = BB.end(); BBI != E; ) {
983 Instruction *I = BBI++;
985 if (PHINode *P = dyn_cast<PHINode>(I)) {
986 // It is possible for very late stage optimizations (such as SimplifyCFG)
987 // to introduce PHI nodes too late to be cleaned up. If we detect such a
988 // trivial PHI, go ahead and zap it here.
989 if (Value *V = SimplifyInstruction(P)) {
990 P->replaceAllUsesWith(V);
991 P->eraseFromParent();
994 } else if (CastInst *CI = dyn_cast<CastInst>(I)) {
995 // If the source of the cast is a constant, then this should have
996 // already been constant folded. The only reason NOT to constant fold
997 // it is if something (e.g. LSR) was careful to place the constant
998 // evaluation in a block other than then one that uses it (e.g. to hoist
999 // the address of globals out of a loop). If this is the case, we don't
1000 // want to forward-subst the cast.
1001 if (isa<Constant>(CI->getOperand(0)))
1004 bool Change = false;
1006 Change = OptimizeNoopCopyExpression(CI, *TLI);
1007 MadeChange |= Change;
1010 if (!Change && (isa<ZExtInst>(I) || isa<SExtInst>(I))) {
1011 MadeChange |= MoveExtToFormExtLoad(I);
1012 MadeChange |= OptimizeExtUses(I);
1014 } else if (CmpInst *CI = dyn_cast<CmpInst>(I)) {
1015 MadeChange |= OptimizeCmpExpression(CI);
1016 } else if (LoadInst *LI = dyn_cast<LoadInst>(I)) {
1018 MadeChange |= OptimizeMemoryInst(I, I->getOperand(0), LI->getType(),
1020 } else if (StoreInst *SI = dyn_cast<StoreInst>(I)) {
1022 MadeChange |= OptimizeMemoryInst(I, SI->getOperand(1),
1023 SI->getOperand(0)->getType(),
1025 } else if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(I)) {
1026 if (GEPI->hasAllZeroIndices()) {
1027 /// The GEP operand must be a pointer, so must its result -> BitCast
1028 Instruction *NC = new BitCastInst(GEPI->getOperand(0), GEPI->getType(),
1029 GEPI->getName(), GEPI);
1030 GEPI->replaceAllUsesWith(NC);
1031 GEPI->eraseFromParent();
1036 } else if (CallInst *CI = dyn_cast<CallInst>(I)) {
1037 // If we found an inline asm expession, and if the target knows how to
1038 // lower it to normal LLVM code, do so now.
1039 if (TLI && isa<InlineAsm>(CI->getCalledValue())) {
1040 if (TLI->ExpandInlineAsm(CI)) {
1042 // Avoid processing instructions out of order, which could cause
1043 // reuse before a value is defined.
1046 // Sink address computing for memory operands into the block.
1047 MadeChange |= OptimizeInlineAsmInst(I, &(*CI), SunkAddrs);
1049 // Other CallInst optimizations that don't need to muck with the
1050 // enclosing iterator here.
1051 MadeChange |= OptimizeCallInst(CI);