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/ProfileInfo.h"
26 #include "llvm/Target/TargetData.h"
27 #include "llvm/Target/TargetLowering.h"
28 #include "llvm/Transforms/Utils/AddrModeMatcher.h"
29 #include "llvm/Transforms/Utils/BasicBlockUtils.h"
30 #include "llvm/Transforms/Utils/Local.h"
31 #include "llvm/Transforms/Utils/BuildLibCalls.h"
32 #include "llvm/ADT/DenseMap.h"
33 #include "llvm/ADT/SmallSet.h"
34 #include "llvm/ADT/Statistic.h"
35 #include "llvm/Assembly/Writer.h"
36 #include "llvm/Support/CallSite.h"
37 #include "llvm/Support/CommandLine.h"
38 #include "llvm/Support/Debug.h"
39 #include "llvm/Support/GetElementPtrTypeIterator.h"
40 #include "llvm/Support/PatternMatch.h"
41 #include "llvm/Support/raw_ostream.h"
42 #include "llvm/Support/IRBuilder.h"
44 using namespace llvm::PatternMatch;
46 STATISTIC(NumElim, "Number of blocks eliminated");
49 CriticalEdgeSplit("cgp-critical-edge-splitting",
50 cl::desc("Split critical edges during codegen prepare"),
51 cl::init(false), cl::Hidden);
54 class CodeGenPrepare : public FunctionPass {
55 /// TLI - Keep a pointer of a TargetLowering to consult for determining
56 /// transformation profitability.
57 const TargetLowering *TLI;
60 /// BackEdges - Keep a set of all the loop back edges.
62 SmallSet<std::pair<const BasicBlock*, const BasicBlock*>, 8> BackEdges;
64 static char ID; // Pass identification, replacement for typeid
65 explicit CodeGenPrepare(const TargetLowering *tli = 0)
66 : FunctionPass(ID), TLI(tli) {
67 initializeCodeGenPreparePass(*PassRegistry::getPassRegistry());
69 bool runOnFunction(Function &F);
71 virtual void getAnalysisUsage(AnalysisUsage &AU) const {
72 AU.addPreserved<ProfileInfo>();
75 virtual void releaseMemory() {
80 bool EliminateMostlyEmptyBlocks(Function &F);
81 bool CanMergeBlocks(const BasicBlock *BB, const BasicBlock *DestBB) const;
82 void EliminateMostlyEmptyBlock(BasicBlock *BB);
83 bool OptimizeBlock(BasicBlock &BB);
84 bool OptimizeMemoryInst(Instruction *I, Value *Addr, const Type *AccessTy,
85 DenseMap<Value*,Value*> &SunkAddrs);
86 bool OptimizeInlineAsmInst(Instruction *I, CallSite CS,
87 DenseMap<Value*,Value*> &SunkAddrs);
88 bool OptimizeCallInst(CallInst *CI);
89 bool MoveExtToFormExtLoad(Instruction *I);
90 bool OptimizeExtUses(Instruction *I);
91 void findLoopBackEdges(const Function &F);
95 char CodeGenPrepare::ID = 0;
96 INITIALIZE_PASS(CodeGenPrepare, "codegenprepare",
97 "Optimize for code generation", false, false)
99 FunctionPass *llvm::createCodeGenPreparePass(const TargetLowering *TLI) {
100 return new CodeGenPrepare(TLI);
103 /// findLoopBackEdges - Do a DFS walk to find loop back edges.
105 void CodeGenPrepare::findLoopBackEdges(const Function &F) {
106 SmallVector<std::pair<const BasicBlock*,const BasicBlock*>, 32> Edges;
107 FindFunctionBackedges(F, Edges);
109 BackEdges.insert(Edges.begin(), Edges.end());
113 bool CodeGenPrepare::runOnFunction(Function &F) {
114 bool EverMadeChange = false;
116 PFI = getAnalysisIfAvailable<ProfileInfo>();
117 // First pass, eliminate blocks that contain only PHI nodes and an
118 // unconditional branch.
119 EverMadeChange |= EliminateMostlyEmptyBlocks(F);
121 // Now find loop back edges.
122 findLoopBackEdges(F);
124 bool MadeChange = true;
127 for (Function::iterator BB = F.begin(), E = F.end(); BB != E; ++BB)
128 MadeChange |= OptimizeBlock(*BB);
129 EverMadeChange |= MadeChange;
131 return EverMadeChange;
134 /// EliminateMostlyEmptyBlocks - eliminate blocks that contain only PHI nodes,
135 /// debug info directives, and an unconditional branch. Passes before isel
136 /// (e.g. LSR/loopsimplify) often split edges in ways that are non-optimal for
137 /// isel. Start by eliminating these blocks so we can split them the way we
139 bool CodeGenPrepare::EliminateMostlyEmptyBlocks(Function &F) {
140 bool MadeChange = false;
141 // Note that this intentionally skips the entry block.
142 for (Function::iterator I = ++F.begin(), E = F.end(); I != E; ) {
143 BasicBlock *BB = I++;
145 // If this block doesn't end with an uncond branch, ignore it.
146 BranchInst *BI = dyn_cast<BranchInst>(BB->getTerminator());
147 if (!BI || !BI->isUnconditional())
150 // If the instruction before the branch (skipping debug info) isn't a phi
151 // node, then other stuff is happening here.
152 BasicBlock::iterator BBI = BI;
153 if (BBI != BB->begin()) {
155 while (isa<DbgInfoIntrinsic>(BBI)) {
156 if (BBI == BB->begin())
160 if (!isa<DbgInfoIntrinsic>(BBI) && !isa<PHINode>(BBI))
164 // Do not break infinite loops.
165 BasicBlock *DestBB = BI->getSuccessor(0);
169 if (!CanMergeBlocks(BB, DestBB))
172 EliminateMostlyEmptyBlock(BB);
178 /// CanMergeBlocks - Return true if we can merge BB into DestBB if there is a
179 /// single uncond branch between them, and BB contains no other non-phi
181 bool CodeGenPrepare::CanMergeBlocks(const BasicBlock *BB,
182 const BasicBlock *DestBB) const {
183 // We only want to eliminate blocks whose phi nodes are used by phi nodes in
184 // the successor. If there are more complex condition (e.g. preheaders),
185 // don't mess around with them.
186 BasicBlock::const_iterator BBI = BB->begin();
187 while (const PHINode *PN = dyn_cast<PHINode>(BBI++)) {
188 for (Value::const_use_iterator UI = PN->use_begin(), E = PN->use_end();
190 const Instruction *User = cast<Instruction>(*UI);
191 if (User->getParent() != DestBB || !isa<PHINode>(User))
193 // If User is inside DestBB block and it is a PHINode then check
194 // incoming value. If incoming value is not from BB then this is
195 // a complex condition (e.g. preheaders) we want to avoid here.
196 if (User->getParent() == DestBB) {
197 if (const PHINode *UPN = dyn_cast<PHINode>(User))
198 for (unsigned I = 0, E = UPN->getNumIncomingValues(); I != E; ++I) {
199 Instruction *Insn = dyn_cast<Instruction>(UPN->getIncomingValue(I));
200 if (Insn && Insn->getParent() == BB &&
201 Insn->getParent() != UPN->getIncomingBlock(I))
208 // If BB and DestBB contain any common predecessors, then the phi nodes in BB
209 // and DestBB may have conflicting incoming values for the block. If so, we
210 // can't merge the block.
211 const PHINode *DestBBPN = dyn_cast<PHINode>(DestBB->begin());
212 if (!DestBBPN) return true; // no conflict.
214 // Walk all the PHI nodes in DestBB. If any of the input values to the PHI
215 // are trapping constant exprs, then merging this block would introduce the
216 // possible trap into new control flow if we have any critical predecessor
218 for (BasicBlock::const_iterator I = DestBB->begin(); isa<PHINode>(I); ++I) {
219 const PHINode *PN = cast<PHINode>(I);
220 if (const Constant *C =dyn_cast<Constant>(PN->getIncomingValueForBlock(BB)))
225 // Collect the preds of BB.
226 SmallPtrSet<const BasicBlock*, 16> BBPreds;
227 if (const PHINode *BBPN = dyn_cast<PHINode>(BB->begin())) {
228 // It is faster to get preds from a PHI than with pred_iterator.
229 for (unsigned i = 0, e = BBPN->getNumIncomingValues(); i != e; ++i)
230 BBPreds.insert(BBPN->getIncomingBlock(i));
232 BBPreds.insert(pred_begin(BB), pred_end(BB));
235 // Walk the preds of DestBB.
236 for (unsigned i = 0, e = DestBBPN->getNumIncomingValues(); i != e; ++i) {
237 BasicBlock *Pred = DestBBPN->getIncomingBlock(i);
238 if (BBPreds.count(Pred)) { // Common predecessor?
239 BBI = DestBB->begin();
240 while (const PHINode *PN = dyn_cast<PHINode>(BBI++)) {
241 const Value *V1 = PN->getIncomingValueForBlock(Pred);
242 const Value *V2 = PN->getIncomingValueForBlock(BB);
244 // If V2 is a phi node in BB, look up what the mapped value will be.
245 if (const PHINode *V2PN = dyn_cast<PHINode>(V2))
246 if (V2PN->getParent() == BB)
247 V2 = V2PN->getIncomingValueForBlock(Pred);
249 // If there is a conflict, bail out.
250 if (V1 != V2) return false;
259 /// EliminateMostlyEmptyBlock - Eliminate a basic block that have only phi's and
260 /// an unconditional branch in it.
261 void CodeGenPrepare::EliminateMostlyEmptyBlock(BasicBlock *BB) {
262 BranchInst *BI = cast<BranchInst>(BB->getTerminator());
263 BasicBlock *DestBB = BI->getSuccessor(0);
265 DEBUG(dbgs() << "MERGING MOSTLY EMPTY BLOCKS - BEFORE:\n" << *BB << *DestBB);
267 // If the destination block has a single pred, then this is a trivial edge,
269 if (BasicBlock *SinglePred = DestBB->getSinglePredecessor()) {
270 if (SinglePred != DestBB) {
271 // Remember if SinglePred was the entry block of the function. If so, we
272 // will need to move BB back to the entry position.
273 bool isEntry = SinglePred == &SinglePred->getParent()->getEntryBlock();
274 MergeBasicBlockIntoOnlyPred(DestBB, this);
276 if (isEntry && BB != &BB->getParent()->getEntryBlock())
277 BB->moveBefore(&BB->getParent()->getEntryBlock());
279 DEBUG(dbgs() << "AFTER:\n" << *DestBB << "\n\n\n");
284 // Otherwise, we have multiple predecessors of BB. Update the PHIs in DestBB
285 // to handle the new incoming edges it is about to have.
287 for (BasicBlock::iterator BBI = DestBB->begin();
288 (PN = dyn_cast<PHINode>(BBI)); ++BBI) {
289 // Remove the incoming value for BB, and remember it.
290 Value *InVal = PN->removeIncomingValue(BB, false);
292 // Two options: either the InVal is a phi node defined in BB or it is some
293 // value that dominates BB.
294 PHINode *InValPhi = dyn_cast<PHINode>(InVal);
295 if (InValPhi && InValPhi->getParent() == BB) {
296 // Add all of the input values of the input PHI as inputs of this phi.
297 for (unsigned i = 0, e = InValPhi->getNumIncomingValues(); i != e; ++i)
298 PN->addIncoming(InValPhi->getIncomingValue(i),
299 InValPhi->getIncomingBlock(i));
301 // Otherwise, add one instance of the dominating value for each edge that
302 // we will be adding.
303 if (PHINode *BBPN = dyn_cast<PHINode>(BB->begin())) {
304 for (unsigned i = 0, e = BBPN->getNumIncomingValues(); i != e; ++i)
305 PN->addIncoming(InVal, BBPN->getIncomingBlock(i));
307 for (pred_iterator PI = pred_begin(BB), E = pred_end(BB); PI != E; ++PI)
308 PN->addIncoming(InVal, *PI);
313 // The PHIs are now updated, change everything that refers to BB to use
314 // DestBB and remove BB.
315 BB->replaceAllUsesWith(DestBB);
317 PFI->replaceAllUses(BB, DestBB);
318 PFI->removeEdge(ProfileInfo::getEdge(BB, DestBB));
320 BB->eraseFromParent();
323 DEBUG(dbgs() << "AFTER:\n" << *DestBB << "\n\n\n");
326 /// FindReusablePredBB - Check all of the predecessors of the block DestPHI
327 /// lives in to see if there is a block that we can reuse as a critical edge
329 static BasicBlock *FindReusablePredBB(PHINode *DestPHI, BasicBlock *TIBB) {
330 BasicBlock *Dest = DestPHI->getParent();
332 /// TIPHIValues - This array is lazily computed to determine the values of
333 /// PHIs in Dest that TI would provide.
334 SmallVector<Value*, 32> TIPHIValues;
336 /// TIBBEntryNo - This is a cache to speed up pred queries for TIBB.
337 unsigned TIBBEntryNo = 0;
339 // Check to see if Dest has any blocks that can be used as a split edge for
341 for (unsigned pi = 0, e = DestPHI->getNumIncomingValues(); pi != e; ++pi) {
342 BasicBlock *Pred = DestPHI->getIncomingBlock(pi);
343 // To be usable, the pred has to end with an uncond branch to the dest.
344 BranchInst *PredBr = dyn_cast<BranchInst>(Pred->getTerminator());
345 if (!PredBr || !PredBr->isUnconditional())
347 // Must be empty other than the branch and debug info.
348 BasicBlock::iterator I = Pred->begin();
349 while (isa<DbgInfoIntrinsic>(I))
353 // Cannot be the entry block; its label does not get emitted.
354 if (Pred == &Dest->getParent()->getEntryBlock())
357 // Finally, since we know that Dest has phi nodes in it, we have to make
358 // sure that jumping to Pred will have the same effect as going to Dest in
359 // terms of PHI values.
362 unsigned PredEntryNo = pi;
364 bool FoundMatch = true;
365 for (BasicBlock::iterator I = Dest->begin();
366 (PN = dyn_cast<PHINode>(I)); ++I, ++PHINo) {
367 if (PHINo == TIPHIValues.size()) {
368 if (PN->getIncomingBlock(TIBBEntryNo) != TIBB)
369 TIBBEntryNo = PN->getBasicBlockIndex(TIBB);
370 TIPHIValues.push_back(PN->getIncomingValue(TIBBEntryNo));
373 // If the PHI entry doesn't work, we can't use this pred.
374 if (PN->getIncomingBlock(PredEntryNo) != Pred)
375 PredEntryNo = PN->getBasicBlockIndex(Pred);
377 if (TIPHIValues[PHINo] != PN->getIncomingValue(PredEntryNo)) {
383 // If we found a workable predecessor, change TI to branch to Succ.
391 /// SplitEdgeNicely - Split the critical edge from TI to its specified
392 /// successor if it will improve codegen. We only do this if the successor has
393 /// phi nodes (otherwise critical edges are ok). If there is already another
394 /// predecessor of the succ that is empty (and thus has no phi nodes), use it
395 /// instead of introducing a new block.
396 static void SplitEdgeNicely(TerminatorInst *TI, unsigned SuccNum,
397 SmallSet<std::pair<const BasicBlock*,
398 const BasicBlock*>, 8> &BackEdges,
400 BasicBlock *TIBB = TI->getParent();
401 BasicBlock *Dest = TI->getSuccessor(SuccNum);
402 assert(isa<PHINode>(Dest->begin()) &&
403 "This should only be called if Dest has a PHI!");
404 PHINode *DestPHI = cast<PHINode>(Dest->begin());
406 // Do not split edges to EH landing pads.
407 if (InvokeInst *Invoke = dyn_cast<InvokeInst>(TI))
408 if (Invoke->getSuccessor(1) == Dest)
411 // As a hack, never split backedges of loops. Even though the copy for any
412 // PHIs inserted on the backedge would be dead for exits from the loop, we
413 // assume that the cost of *splitting* the backedge would be too high.
414 if (BackEdges.count(std::make_pair(TIBB, Dest)))
417 if (BasicBlock *ReuseBB = FindReusablePredBB(DestPHI, TIBB)) {
418 ProfileInfo *PFI = P->getAnalysisIfAvailable<ProfileInfo>();
420 PFI->splitEdge(TIBB, Dest, ReuseBB);
421 Dest->removePredecessor(TIBB);
422 TI->setSuccessor(SuccNum, ReuseBB);
426 SplitCriticalEdge(TI, SuccNum, P, true);
430 /// OptimizeNoopCopyExpression - If the specified cast instruction is a noop
431 /// copy (e.g. it's casting from one pointer type to another, i32->i8 on PPC),
432 /// sink it into user blocks to reduce the number of virtual
433 /// registers that must be created and coalesced.
435 /// Return true if any changes are made.
437 static bool OptimizeNoopCopyExpression(CastInst *CI, const TargetLowering &TLI){
438 // If this is a noop copy,
439 EVT SrcVT = TLI.getValueType(CI->getOperand(0)->getType());
440 EVT DstVT = TLI.getValueType(CI->getType());
442 // This is an fp<->int conversion?
443 if (SrcVT.isInteger() != DstVT.isInteger())
446 // If this is an extension, it will be a zero or sign extension, which
448 if (SrcVT.bitsLT(DstVT)) return false;
450 // If these values will be promoted, find out what they will be promoted
451 // to. This helps us consider truncates on PPC as noop copies when they
453 if (TLI.getTypeAction(SrcVT) == TargetLowering::Promote)
454 SrcVT = TLI.getTypeToTransformTo(CI->getContext(), SrcVT);
455 if (TLI.getTypeAction(DstVT) == TargetLowering::Promote)
456 DstVT = TLI.getTypeToTransformTo(CI->getContext(), DstVT);
458 // If, after promotion, these are the same types, this is a noop copy.
462 BasicBlock *DefBB = CI->getParent();
464 /// InsertedCasts - Only insert a cast in each block once.
465 DenseMap<BasicBlock*, CastInst*> InsertedCasts;
467 bool MadeChange = false;
468 for (Value::use_iterator UI = CI->use_begin(), E = CI->use_end();
470 Use &TheUse = UI.getUse();
471 Instruction *User = cast<Instruction>(*UI);
473 // Figure out which BB this cast is used in. For PHI's this is the
474 // appropriate predecessor block.
475 BasicBlock *UserBB = User->getParent();
476 if (PHINode *PN = dyn_cast<PHINode>(User)) {
477 UserBB = PN->getIncomingBlock(UI);
480 // Preincrement use iterator so we don't invalidate it.
483 // If this user is in the same block as the cast, don't change the cast.
484 if (UserBB == DefBB) continue;
486 // If we have already inserted a cast into this block, use it.
487 CastInst *&InsertedCast = InsertedCasts[UserBB];
490 BasicBlock::iterator InsertPt = UserBB->getFirstNonPHI();
493 CastInst::Create(CI->getOpcode(), CI->getOperand(0), CI->getType(), "",
498 // Replace a use of the cast with a use of the new cast.
499 TheUse = InsertedCast;
502 // If we removed all uses, nuke the cast.
503 if (CI->use_empty()) {
504 CI->eraseFromParent();
511 /// OptimizeCmpExpression - sink the given CmpInst into user blocks to reduce
512 /// the number of virtual registers that must be created and coalesced. This is
513 /// a clear win except on targets with multiple condition code registers
514 /// (PowerPC), where it might lose; some adjustment may be wanted there.
516 /// Return true if any changes are made.
517 static bool OptimizeCmpExpression(CmpInst *CI) {
518 BasicBlock *DefBB = CI->getParent();
520 /// InsertedCmp - Only insert a cmp in each block once.
521 DenseMap<BasicBlock*, CmpInst*> InsertedCmps;
523 bool MadeChange = false;
524 for (Value::use_iterator UI = CI->use_begin(), E = CI->use_end();
526 Use &TheUse = UI.getUse();
527 Instruction *User = cast<Instruction>(*UI);
529 // Preincrement use iterator so we don't invalidate it.
532 // Don't bother for PHI nodes.
533 if (isa<PHINode>(User))
536 // Figure out which BB this cmp is used in.
537 BasicBlock *UserBB = User->getParent();
539 // If this user is in the same block as the cmp, don't change the cmp.
540 if (UserBB == DefBB) continue;
542 // If we have already inserted a cmp into this block, use it.
543 CmpInst *&InsertedCmp = InsertedCmps[UserBB];
546 BasicBlock::iterator InsertPt = UserBB->getFirstNonPHI();
549 CmpInst::Create(CI->getOpcode(),
550 CI->getPredicate(), CI->getOperand(0),
551 CI->getOperand(1), "", InsertPt);
555 // Replace a use of the cmp with a use of the new cmp.
556 TheUse = InsertedCmp;
559 // If we removed all uses, nuke the cmp.
561 CI->eraseFromParent();
567 class CodeGenPrepareFortifiedLibCalls : public SimplifyFortifiedLibCalls {
569 void replaceCall(Value *With) {
570 CI->replaceAllUsesWith(With);
571 CI->eraseFromParent();
573 bool isFoldable(unsigned SizeCIOp, unsigned, bool) const {
574 if (ConstantInt *SizeCI =
575 dyn_cast<ConstantInt>(CI->getArgOperand(SizeCIOp)))
576 return SizeCI->isAllOnesValue();
580 } // end anonymous namespace
582 bool CodeGenPrepare::OptimizeCallInst(CallInst *CI) {
583 // Lower all uses of llvm.objectsize.*
584 IntrinsicInst *II = dyn_cast<IntrinsicInst>(CI);
585 if (II && II->getIntrinsicID() == Intrinsic::objectsize) {
586 bool Min = (cast<ConstantInt>(II->getArgOperand(1))->getZExtValue() == 1);
587 const Type *ReturnTy = CI->getType();
588 Constant *RetVal = ConstantInt::get(ReturnTy, Min ? 0 : -1ULL);
589 CI->replaceAllUsesWith(RetVal);
590 CI->eraseFromParent();
594 // From here on out we're working with named functions.
595 if (CI->getCalledFunction() == 0) return false;
597 // We'll need TargetData from here on out.
598 const TargetData *TD = TLI ? TLI->getTargetData() : 0;
599 if (!TD) return false;
601 // Lower all default uses of _chk calls. This is very similar
602 // to what InstCombineCalls does, but here we are only lowering calls
603 // that have the default "don't know" as the objectsize. Anything else
604 // should be left alone.
605 CodeGenPrepareFortifiedLibCalls Simplifier;
606 return Simplifier.fold(CI, TD);
608 //===----------------------------------------------------------------------===//
609 // Memory Optimization
610 //===----------------------------------------------------------------------===//
612 /// IsNonLocalValue - Return true if the specified values are defined in a
613 /// different basic block than BB.
614 static bool IsNonLocalValue(Value *V, BasicBlock *BB) {
615 if (Instruction *I = dyn_cast<Instruction>(V))
616 return I->getParent() != BB;
620 /// OptimizeMemoryInst - Load and Store Instructions often have
621 /// addressing modes that can do significant amounts of computation. As such,
622 /// instruction selection will try to get the load or store to do as much
623 /// computation as possible for the program. The problem is that isel can only
624 /// see within a single block. As such, we sink as much legal addressing mode
625 /// stuff into the block as possible.
627 /// This method is used to optimize both load/store and inline asms with memory
629 bool CodeGenPrepare::OptimizeMemoryInst(Instruction *MemoryInst, Value *Addr,
630 const Type *AccessTy,
631 DenseMap<Value*,Value*> &SunkAddrs) {
634 // Try to collapse single-value PHI nodes. This is necessary to undo
635 // unprofitable PRE transformations.
636 std::vector<Value*> worklist;
637 SmallPtrSet<Value*, 4> Visited;
638 worklist.push_back(Addr);
640 // Use a worklist to iteratively look through PHI nodes, and ensure that
641 // the addressing mode obtained from the non-PHI roots of the graph
643 Value *Consensus = 0;
644 unsigned NumUses = 0;
645 SmallVector<Instruction*, 16> AddrModeInsts;
646 ExtAddrMode AddrMode;
647 while (!worklist.empty()) {
648 Value *V = worklist.back();
651 // Break use-def graph loops.
652 if (Visited.count(V)) {
659 // For a PHI node, push all of its incoming values.
660 if (PHINode *P = dyn_cast<PHINode>(V)) {
661 for (unsigned i = 0, e = P->getNumIncomingValues(); i != e; ++i)
662 worklist.push_back(P->getIncomingValue(i));
666 // For non-PHIs, determine the addressing mode being computed.
667 SmallVector<Instruction*, 16> NewAddrModeInsts;
668 ExtAddrMode NewAddrMode =
669 AddressingModeMatcher::Match(V, AccessTy,MemoryInst,
670 NewAddrModeInsts, *TLI);
672 // Ensure that the obtained addressing mode is equivalent to that obtained
673 // for all other roots of the PHI traversal. Also, when choosing one
674 // such root as representative, select the one with the most uses in order
675 // to keep the cost modeling heuristics in AddressingModeMatcher applicable.
676 if (!Consensus || NewAddrMode == AddrMode) {
677 if (V->getNumUses() > NumUses) {
679 NumUses = V->getNumUses();
680 AddrMode = NewAddrMode;
681 AddrModeInsts = NewAddrModeInsts;
690 // If the addressing mode couldn't be determined, or if multiple different
691 // ones were determined, bail out now.
692 if (!Consensus) return false;
694 // Check to see if any of the instructions supersumed by this addr mode are
695 // non-local to I's BB.
696 bool AnyNonLocal = false;
697 for (unsigned i = 0, e = AddrModeInsts.size(); i != e; ++i) {
698 if (IsNonLocalValue(AddrModeInsts[i], MemoryInst->getParent())) {
704 // If all the instructions matched are already in this BB, don't do anything.
706 DEBUG(dbgs() << "CGP: Found local addrmode: " << AddrMode << "\n");
710 // Insert this computation right after this user. Since our caller is
711 // scanning from the top of the BB to the bottom, reuse of the expr are
712 // guaranteed to happen later.
713 BasicBlock::iterator InsertPt = MemoryInst;
715 // Now that we determined the addressing expression we want to use and know
716 // that we have to sink it into this block. Check to see if we have already
717 // done this for some other load/store instr in this block. If so, reuse the
719 Value *&SunkAddr = SunkAddrs[Addr];
721 DEBUG(dbgs() << "CGP: Reusing nonlocal addrmode: " << AddrMode << " for "
723 if (SunkAddr->getType() != Addr->getType())
724 SunkAddr = new BitCastInst(SunkAddr, Addr->getType(), "tmp", InsertPt);
726 DEBUG(dbgs() << "CGP: SINKING nonlocal addrmode: " << AddrMode << " for "
728 const Type *IntPtrTy =
729 TLI->getTargetData()->getIntPtrType(AccessTy->getContext());
733 // Start with the base register. Do this first so that subsequent address
734 // matching finds it last, which will prevent it from trying to match it
735 // as the scaled value in case it happens to be a mul. That would be
736 // problematic if we've sunk a different mul for the scale, because then
737 // we'd end up sinking both muls.
738 if (AddrMode.BaseReg) {
739 Value *V = AddrMode.BaseReg;
740 if (V->getType()->isPointerTy())
741 V = new PtrToIntInst(V, IntPtrTy, "sunkaddr", InsertPt);
742 if (V->getType() != IntPtrTy)
743 V = CastInst::CreateIntegerCast(V, IntPtrTy, /*isSigned=*/true,
744 "sunkaddr", InsertPt);
748 // Add the scale value.
749 if (AddrMode.Scale) {
750 Value *V = AddrMode.ScaledReg;
751 if (V->getType() == IntPtrTy) {
753 } else if (V->getType()->isPointerTy()) {
754 V = new PtrToIntInst(V, IntPtrTy, "sunkaddr", InsertPt);
755 } else if (cast<IntegerType>(IntPtrTy)->getBitWidth() <
756 cast<IntegerType>(V->getType())->getBitWidth()) {
757 V = new TruncInst(V, IntPtrTy, "sunkaddr", InsertPt);
759 V = new SExtInst(V, IntPtrTy, "sunkaddr", InsertPt);
761 if (AddrMode.Scale != 1)
762 V = BinaryOperator::CreateMul(V, ConstantInt::get(IntPtrTy,
764 "sunkaddr", InsertPt);
766 Result = BinaryOperator::CreateAdd(Result, V, "sunkaddr", InsertPt);
771 // Add in the BaseGV if present.
772 if (AddrMode.BaseGV) {
773 Value *V = new PtrToIntInst(AddrMode.BaseGV, IntPtrTy, "sunkaddr",
776 Result = BinaryOperator::CreateAdd(Result, V, "sunkaddr", InsertPt);
781 // Add in the Base Offset if present.
782 if (AddrMode.BaseOffs) {
783 Value *V = ConstantInt::get(IntPtrTy, AddrMode.BaseOffs);
785 Result = BinaryOperator::CreateAdd(Result, V, "sunkaddr", InsertPt);
791 SunkAddr = Constant::getNullValue(Addr->getType());
793 SunkAddr = new IntToPtrInst(Result, Addr->getType(), "sunkaddr",InsertPt);
796 MemoryInst->replaceUsesOfWith(Repl, SunkAddr);
798 if (Repl->use_empty()) {
799 RecursivelyDeleteTriviallyDeadInstructions(Repl);
800 // This address is now available for reassignment, so erase the table entry;
801 // we don't want to match some completely different instruction.
807 /// OptimizeInlineAsmInst - If there are any memory operands, use
808 /// OptimizeMemoryInst to sink their address computing into the block when
809 /// possible / profitable.
810 bool CodeGenPrepare::OptimizeInlineAsmInst(Instruction *I, CallSite CS,
811 DenseMap<Value*,Value*> &SunkAddrs) {
812 bool MadeChange = false;
814 TargetLowering::AsmOperandInfoVector TargetConstraints = TLI->ParseConstraints(CS);
816 for (unsigned i = 0, e = TargetConstraints.size(); i != e; ++i) {
817 TargetLowering::AsmOperandInfo &OpInfo = TargetConstraints[i];
819 // Compute the constraint code and ConstraintType to use.
820 TLI->ComputeConstraintToUse(OpInfo, SDValue());
822 if (OpInfo.ConstraintType == TargetLowering::C_Memory &&
824 Value *OpVal = const_cast<Value *>(CS.getArgument(ArgNo++));
825 MadeChange |= OptimizeMemoryInst(I, OpVal, OpVal->getType(), SunkAddrs);
826 } else if (OpInfo.Type == InlineAsm::isInput)
833 /// MoveExtToFormExtLoad - Move a zext or sext fed by a load into the same
834 /// basic block as the load, unless conditions are unfavorable. This allows
835 /// SelectionDAG to fold the extend into the load.
837 bool CodeGenPrepare::MoveExtToFormExtLoad(Instruction *I) {
838 // Look for a load being extended.
839 LoadInst *LI = dyn_cast<LoadInst>(I->getOperand(0));
840 if (!LI) return false;
842 // If they're already in the same block, there's nothing to do.
843 if (LI->getParent() == I->getParent())
846 // If the load has other users and the truncate is not free, this probably
848 if (!LI->hasOneUse() &&
849 TLI && (TLI->isTypeLegal(TLI->getValueType(LI->getType())) ||
850 !TLI->isTypeLegal(TLI->getValueType(I->getType()))) &&
851 !TLI->isTruncateFree(I->getType(), LI->getType()))
854 // Check whether the target supports casts folded into loads.
856 if (isa<ZExtInst>(I))
857 LType = ISD::ZEXTLOAD;
859 assert(isa<SExtInst>(I) && "Unexpected ext type!");
860 LType = ISD::SEXTLOAD;
862 if (TLI && !TLI->isLoadExtLegal(LType, TLI->getValueType(LI->getType())))
865 // Move the extend into the same block as the load, so that SelectionDAG
867 I->removeFromParent();
872 bool CodeGenPrepare::OptimizeExtUses(Instruction *I) {
873 BasicBlock *DefBB = I->getParent();
875 // If the result of a {s|z}ext and its source are both live out, rewrite all
876 // other uses of the source with result of extension.
877 Value *Src = I->getOperand(0);
878 if (Src->hasOneUse())
881 // Only do this xform if truncating is free.
882 if (TLI && !TLI->isTruncateFree(I->getType(), Src->getType()))
885 // Only safe to perform the optimization if the source is also defined in
887 if (!isa<Instruction>(Src) || DefBB != cast<Instruction>(Src)->getParent())
890 bool DefIsLiveOut = false;
891 for (Value::use_iterator UI = I->use_begin(), E = I->use_end();
893 Instruction *User = cast<Instruction>(*UI);
895 // Figure out which BB this ext is used in.
896 BasicBlock *UserBB = User->getParent();
897 if (UserBB == DefBB) continue;
904 // Make sure non of the uses are PHI nodes.
905 for (Value::use_iterator UI = Src->use_begin(), E = Src->use_end();
907 Instruction *User = cast<Instruction>(*UI);
908 BasicBlock *UserBB = User->getParent();
909 if (UserBB == DefBB) continue;
910 // Be conservative. We don't want this xform to end up introducing
911 // reloads just before load / store instructions.
912 if (isa<PHINode>(User) || isa<LoadInst>(User) || isa<StoreInst>(User))
916 // InsertedTruncs - Only insert one trunc in each block once.
917 DenseMap<BasicBlock*, Instruction*> InsertedTruncs;
919 bool MadeChange = false;
920 for (Value::use_iterator UI = Src->use_begin(), E = Src->use_end();
922 Use &TheUse = UI.getUse();
923 Instruction *User = cast<Instruction>(*UI);
925 // Figure out which BB this ext is used in.
926 BasicBlock *UserBB = User->getParent();
927 if (UserBB == DefBB) continue;
929 // Both src and def are live in this block. Rewrite the use.
930 Instruction *&InsertedTrunc = InsertedTruncs[UserBB];
932 if (!InsertedTrunc) {
933 BasicBlock::iterator InsertPt = UserBB->getFirstNonPHI();
935 InsertedTrunc = new TruncInst(I, Src->getType(), "", InsertPt);
938 // Replace a use of the {s|z}ext source with a use of the result.
939 TheUse = InsertedTrunc;
947 // In this pass we look for GEP and cast instructions that are used
948 // across basic blocks and rewrite them to improve basic-block-at-a-time
950 bool CodeGenPrepare::OptimizeBlock(BasicBlock &BB) {
951 bool MadeChange = false;
953 // Split all critical edges where the dest block has a PHI.
954 if (CriticalEdgeSplit) {
955 TerminatorInst *BBTI = BB.getTerminator();
956 if (BBTI->getNumSuccessors() > 1 && !isa<IndirectBrInst>(BBTI)) {
957 for (unsigned i = 0, e = BBTI->getNumSuccessors(); i != e; ++i) {
958 BasicBlock *SuccBB = BBTI->getSuccessor(i);
959 if (isa<PHINode>(SuccBB->begin()) && isCriticalEdge(BBTI, i, true))
960 SplitEdgeNicely(BBTI, i, BackEdges, this);
965 // Keep track of non-local addresses that have been sunk into this block.
966 // This allows us to avoid inserting duplicate code for blocks with multiple
967 // load/stores of the same address.
968 DenseMap<Value*, Value*> SunkAddrs;
970 for (BasicBlock::iterator BBI = BB.begin(), E = BB.end(); BBI != E; ) {
971 Instruction *I = BBI++;
973 if (CastInst *CI = dyn_cast<CastInst>(I)) {
974 // If the source of the cast is a constant, then this should have
975 // already been constant folded. The only reason NOT to constant fold
976 // it is if something (e.g. LSR) was careful to place the constant
977 // evaluation in a block other than then one that uses it (e.g. to hoist
978 // the address of globals out of a loop). If this is the case, we don't
979 // want to forward-subst the cast.
980 if (isa<Constant>(CI->getOperand(0)))
985 Change = OptimizeNoopCopyExpression(CI, *TLI);
986 MadeChange |= Change;
989 if (!Change && (isa<ZExtInst>(I) || isa<SExtInst>(I))) {
990 MadeChange |= MoveExtToFormExtLoad(I);
991 MadeChange |= OptimizeExtUses(I);
993 } else if (CmpInst *CI = dyn_cast<CmpInst>(I)) {
994 MadeChange |= OptimizeCmpExpression(CI);
995 } else if (LoadInst *LI = dyn_cast<LoadInst>(I)) {
997 MadeChange |= OptimizeMemoryInst(I, I->getOperand(0), LI->getType(),
999 } else if (StoreInst *SI = dyn_cast<StoreInst>(I)) {
1001 MadeChange |= OptimizeMemoryInst(I, SI->getOperand(1),
1002 SI->getOperand(0)->getType(),
1004 } else if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(I)) {
1005 if (GEPI->hasAllZeroIndices()) {
1006 /// The GEP operand must be a pointer, so must its result -> BitCast
1007 Instruction *NC = new BitCastInst(GEPI->getOperand(0), GEPI->getType(),
1008 GEPI->getName(), GEPI);
1009 GEPI->replaceAllUsesWith(NC);
1010 GEPI->eraseFromParent();
1014 } else if (CallInst *CI = dyn_cast<CallInst>(I)) {
1015 // If we found an inline asm expession, and if the target knows how to
1016 // lower it to normal LLVM code, do so now.
1017 if (TLI && isa<InlineAsm>(CI->getCalledValue())) {
1018 if (TLI->ExpandInlineAsm(CI)) {
1020 // Avoid processing instructions out of order, which could cause
1021 // reuse before a value is defined.
1024 // Sink address computing for memory operands into the block.
1025 MadeChange |= OptimizeInlineAsmInst(I, &(*CI), SunkAddrs);
1027 // Other CallInst optimizations that don't need to muck with the
1028 // enclosing iterator here.
1029 MadeChange |= OptimizeCallInst(CI);