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(NumCmpUses, "Number of uses of Cmp expressions replaced with uses of "
50 STATISTIC(NumCastUses, "Number of uses of Cast expressions replaced with uses "
52 STATISTIC(NumMemoryInsts, "Number of memory instructions whose address "
53 "computations were sunk");
54 STATISTIC(NumExtsMoved, "Number of [s|z]ext instructions combined with loads");
55 STATISTIC(NumExtUses, "Number of uses of [s|z]ext instructions optimized");
58 CriticalEdgeSplit("cgp-critical-edge-splitting",
59 cl::desc("Split critical edges during codegen prepare"),
60 cl::init(false), cl::Hidden);
63 class CodeGenPrepare : public FunctionPass {
64 /// TLI - Keep a pointer of a TargetLowering to consult for determining
65 /// transformation profitability.
66 const TargetLowering *TLI;
69 /// BackEdges - Keep a set of all the loop back edges.
71 SmallSet<std::pair<const BasicBlock*, const BasicBlock*>, 8> BackEdges;
73 static char ID; // Pass identification, replacement for typeid
74 explicit CodeGenPrepare(const TargetLowering *tli = 0)
75 : FunctionPass(ID), TLI(tli) {
76 initializeCodeGenPreparePass(*PassRegistry::getPassRegistry());
78 bool runOnFunction(Function &F);
80 virtual void getAnalysisUsage(AnalysisUsage &AU) const {
81 AU.addPreserved<ProfileInfo>();
84 virtual void releaseMemory() {
89 bool EliminateMostlyEmptyBlocks(Function &F);
90 bool CanMergeBlocks(const BasicBlock *BB, const BasicBlock *DestBB) const;
91 void EliminateMostlyEmptyBlock(BasicBlock *BB);
92 bool OptimizeBlock(BasicBlock &BB);
93 bool OptimizeMemoryInst(Instruction *I, Value *Addr, const Type *AccessTy,
94 DenseMap<Value*,Value*> &SunkAddrs);
95 bool OptimizeInlineAsmInst(Instruction *I, CallSite CS,
96 DenseMap<Value*,Value*> &SunkAddrs);
97 bool OptimizeCallInst(CallInst *CI);
98 bool MoveExtToFormExtLoad(Instruction *I);
99 bool OptimizeExtUses(Instruction *I);
100 void findLoopBackEdges(const Function &F);
104 char CodeGenPrepare::ID = 0;
105 INITIALIZE_PASS(CodeGenPrepare, "codegenprepare",
106 "Optimize for code generation", false, false)
108 FunctionPass *llvm::createCodeGenPreparePass(const TargetLowering *TLI) {
109 return new CodeGenPrepare(TLI);
112 /// findLoopBackEdges - Do a DFS walk to find loop back edges.
114 void CodeGenPrepare::findLoopBackEdges(const Function &F) {
115 SmallVector<std::pair<const BasicBlock*,const BasicBlock*>, 32> Edges;
116 FindFunctionBackedges(F, Edges);
118 BackEdges.insert(Edges.begin(), Edges.end());
122 bool CodeGenPrepare::runOnFunction(Function &F) {
123 bool EverMadeChange = false;
125 PFI = getAnalysisIfAvailable<ProfileInfo>();
126 // First pass, eliminate blocks that contain only PHI nodes and an
127 // unconditional branch.
128 EverMadeChange |= EliminateMostlyEmptyBlocks(F);
130 // Now find loop back edges, but only if they are being used to decide which
131 // critical edges to split.
132 if (CriticalEdgeSplit)
133 findLoopBackEdges(F);
135 bool MadeChange = true;
138 for (Function::iterator BB = F.begin(), E = F.end(); BB != E; ++BB)
139 MadeChange |= OptimizeBlock(*BB);
140 EverMadeChange |= MadeChange;
142 return EverMadeChange;
145 /// EliminateMostlyEmptyBlocks - eliminate blocks that contain only PHI nodes,
146 /// debug info directives, and an unconditional branch. Passes before isel
147 /// (e.g. LSR/loopsimplify) often split edges in ways that are non-optimal for
148 /// isel. Start by eliminating these blocks so we can split them the way we
150 bool CodeGenPrepare::EliminateMostlyEmptyBlocks(Function &F) {
151 bool MadeChange = false;
152 // Note that this intentionally skips the entry block.
153 for (Function::iterator I = ++F.begin(), E = F.end(); I != E; ) {
154 BasicBlock *BB = I++;
156 // If this block doesn't end with an uncond branch, ignore it.
157 BranchInst *BI = dyn_cast<BranchInst>(BB->getTerminator());
158 if (!BI || !BI->isUnconditional())
161 // If the instruction before the branch (skipping debug info) isn't a phi
162 // node, then other stuff is happening here.
163 BasicBlock::iterator BBI = BI;
164 if (BBI != BB->begin()) {
166 while (isa<DbgInfoIntrinsic>(BBI)) {
167 if (BBI == BB->begin())
171 if (!isa<DbgInfoIntrinsic>(BBI) && !isa<PHINode>(BBI))
175 // Do not break infinite loops.
176 BasicBlock *DestBB = BI->getSuccessor(0);
180 if (!CanMergeBlocks(BB, DestBB))
183 EliminateMostlyEmptyBlock(BB);
189 /// CanMergeBlocks - Return true if we can merge BB into DestBB if there is a
190 /// single uncond branch between them, and BB contains no other non-phi
192 bool CodeGenPrepare::CanMergeBlocks(const BasicBlock *BB,
193 const BasicBlock *DestBB) const {
194 // We only want to eliminate blocks whose phi nodes are used by phi nodes in
195 // the successor. If there are more complex condition (e.g. preheaders),
196 // don't mess around with them.
197 BasicBlock::const_iterator BBI = BB->begin();
198 while (const PHINode *PN = dyn_cast<PHINode>(BBI++)) {
199 for (Value::const_use_iterator UI = PN->use_begin(), E = PN->use_end();
201 const Instruction *User = cast<Instruction>(*UI);
202 if (User->getParent() != DestBB || !isa<PHINode>(User))
204 // If User is inside DestBB block and it is a PHINode then check
205 // incoming value. If incoming value is not from BB then this is
206 // a complex condition (e.g. preheaders) we want to avoid here.
207 if (User->getParent() == DestBB) {
208 if (const PHINode *UPN = dyn_cast<PHINode>(User))
209 for (unsigned I = 0, E = UPN->getNumIncomingValues(); I != E; ++I) {
210 Instruction *Insn = dyn_cast<Instruction>(UPN->getIncomingValue(I));
211 if (Insn && Insn->getParent() == BB &&
212 Insn->getParent() != UPN->getIncomingBlock(I))
219 // If BB and DestBB contain any common predecessors, then the phi nodes in BB
220 // and DestBB may have conflicting incoming values for the block. If so, we
221 // can't merge the block.
222 const PHINode *DestBBPN = dyn_cast<PHINode>(DestBB->begin());
223 if (!DestBBPN) return true; // no conflict.
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;
503 // If we removed all uses, nuke the cast.
504 if (CI->use_empty()) {
505 CI->eraseFromParent();
512 /// OptimizeCmpExpression - sink the given CmpInst into user blocks to reduce
513 /// the number of virtual registers that must be created and coalesced. This is
514 /// a clear win except on targets with multiple condition code registers
515 /// (PowerPC), where it might lose; some adjustment may be wanted there.
517 /// Return true if any changes are made.
518 static bool OptimizeCmpExpression(CmpInst *CI) {
519 BasicBlock *DefBB = CI->getParent();
521 /// InsertedCmp - Only insert a cmp in each block once.
522 DenseMap<BasicBlock*, CmpInst*> InsertedCmps;
524 bool MadeChange = false;
525 for (Value::use_iterator UI = CI->use_begin(), E = CI->use_end();
527 Use &TheUse = UI.getUse();
528 Instruction *User = cast<Instruction>(*UI);
530 // Preincrement use iterator so we don't invalidate it.
533 // Don't bother for PHI nodes.
534 if (isa<PHINode>(User))
537 // Figure out which BB this cmp is used in.
538 BasicBlock *UserBB = User->getParent();
540 // If this user is in the same block as the cmp, don't change the cmp.
541 if (UserBB == DefBB) continue;
543 // If we have already inserted a cmp into this block, use it.
544 CmpInst *&InsertedCmp = InsertedCmps[UserBB];
547 BasicBlock::iterator InsertPt = UserBB->getFirstNonPHI();
550 CmpInst::Create(CI->getOpcode(),
551 CI->getPredicate(), CI->getOperand(0),
552 CI->getOperand(1), "", InsertPt);
556 // Replace a use of the cmp with a use of the new cmp.
557 TheUse = InsertedCmp;
561 // If we removed all uses, nuke the cmp.
563 CI->eraseFromParent();
569 class CodeGenPrepareFortifiedLibCalls : public SimplifyFortifiedLibCalls {
571 void replaceCall(Value *With) {
572 CI->replaceAllUsesWith(With);
573 CI->eraseFromParent();
575 bool isFoldable(unsigned SizeCIOp, unsigned, bool) const {
576 if (ConstantInt *SizeCI =
577 dyn_cast<ConstantInt>(CI->getArgOperand(SizeCIOp)))
578 return SizeCI->isAllOnesValue();
582 } // end anonymous namespace
584 bool CodeGenPrepare::OptimizeCallInst(CallInst *CI) {
585 // Lower all uses of llvm.objectsize.*
586 IntrinsicInst *II = dyn_cast<IntrinsicInst>(CI);
587 if (II && II->getIntrinsicID() == Intrinsic::objectsize) {
588 bool Min = (cast<ConstantInt>(II->getArgOperand(1))->getZExtValue() == 1);
589 const Type *ReturnTy = CI->getType();
590 Constant *RetVal = ConstantInt::get(ReturnTy, Min ? 0 : -1ULL);
591 CI->replaceAllUsesWith(RetVal);
592 CI->eraseFromParent();
596 // From here on out we're working with named functions.
597 if (CI->getCalledFunction() == 0) return false;
599 // We'll need TargetData from here on out.
600 const TargetData *TD = TLI ? TLI->getTargetData() : 0;
601 if (!TD) return false;
603 // Lower all default uses of _chk calls. This is very similar
604 // to what InstCombineCalls does, but here we are only lowering calls
605 // that have the default "don't know" as the objectsize. Anything else
606 // should be left alone.
607 CodeGenPrepareFortifiedLibCalls Simplifier;
608 return Simplifier.fold(CI, TD);
610 //===----------------------------------------------------------------------===//
611 // Memory Optimization
612 //===----------------------------------------------------------------------===//
614 /// IsNonLocalValue - Return true if the specified values are defined in a
615 /// different basic block than BB.
616 static bool IsNonLocalValue(Value *V, BasicBlock *BB) {
617 if (Instruction *I = dyn_cast<Instruction>(V))
618 return I->getParent() != BB;
622 /// OptimizeMemoryInst - Load and Store Instructions often have
623 /// addressing modes that can do significant amounts of computation. As such,
624 /// instruction selection will try to get the load or store to do as much
625 /// computation as possible for the program. The problem is that isel can only
626 /// see within a single block. As such, we sink as much legal addressing mode
627 /// stuff into the block as possible.
629 /// This method is used to optimize both load/store and inline asms with memory
631 bool CodeGenPrepare::OptimizeMemoryInst(Instruction *MemoryInst, Value *Addr,
632 const Type *AccessTy,
633 DenseMap<Value*,Value*> &SunkAddrs) {
636 // Try to collapse single-value PHI nodes. This is necessary to undo
637 // unprofitable PRE transformations.
638 SmallVector<Value*, 8> worklist;
639 SmallPtrSet<Value*, 16> Visited;
640 worklist.push_back(Addr);
642 // Use a worklist to iteratively look through PHI nodes, and ensure that
643 // the addressing mode obtained from the non-PHI roots of the graph
645 Value *Consensus = 0;
646 unsigned NumUses = 0;
647 SmallVector<Instruction*, 16> AddrModeInsts;
648 ExtAddrMode AddrMode;
649 while (!worklist.empty()) {
650 Value *V = worklist.back();
653 // Break use-def graph loops.
654 if (Visited.count(V)) {
661 // For a PHI node, push all of its incoming values.
662 if (PHINode *P = dyn_cast<PHINode>(V)) {
663 for (unsigned i = 0, e = P->getNumIncomingValues(); i != e; ++i)
664 worklist.push_back(P->getIncomingValue(i));
668 // For non-PHIs, determine the addressing mode being computed.
669 SmallVector<Instruction*, 16> NewAddrModeInsts;
670 ExtAddrMode NewAddrMode =
671 AddressingModeMatcher::Match(V, AccessTy,MemoryInst,
672 NewAddrModeInsts, *TLI);
674 // Ensure that the obtained addressing mode is equivalent to that obtained
675 // for all other roots of the PHI traversal. Also, when choosing one
676 // such root as representative, select the one with the most uses in order
677 // to keep the cost modeling heuristics in AddressingModeMatcher applicable.
678 if (!Consensus || NewAddrMode == AddrMode) {
679 if (V->getNumUses() > NumUses) {
681 NumUses = V->getNumUses();
682 AddrMode = NewAddrMode;
683 AddrModeInsts = NewAddrModeInsts;
692 // If the addressing mode couldn't be determined, or if multiple different
693 // ones were determined, bail out now.
694 if (!Consensus) return false;
696 // Check to see if any of the instructions supersumed by this addr mode are
697 // non-local to I's BB.
698 bool AnyNonLocal = false;
699 for (unsigned i = 0, e = AddrModeInsts.size(); i != e; ++i) {
700 if (IsNonLocalValue(AddrModeInsts[i], MemoryInst->getParent())) {
706 // If all the instructions matched are already in this BB, don't do anything.
708 DEBUG(dbgs() << "CGP: Found local addrmode: " << AddrMode << "\n");
712 // Insert this computation right after this user. Since our caller is
713 // scanning from the top of the BB to the bottom, reuse of the expr are
714 // guaranteed to happen later.
715 BasicBlock::iterator InsertPt = MemoryInst;
717 // Now that we determined the addressing expression we want to use and know
718 // that we have to sink it into this block. Check to see if we have already
719 // done this for some other load/store instr in this block. If so, reuse the
721 Value *&SunkAddr = SunkAddrs[Addr];
723 DEBUG(dbgs() << "CGP: Reusing nonlocal addrmode: " << AddrMode << " for "
725 if (SunkAddr->getType() != Addr->getType())
726 SunkAddr = new BitCastInst(SunkAddr, Addr->getType(), "tmp", InsertPt);
728 DEBUG(dbgs() << "CGP: SINKING nonlocal addrmode: " << AddrMode << " for "
730 const Type *IntPtrTy =
731 TLI->getTargetData()->getIntPtrType(AccessTy->getContext());
735 // Start with the base register. Do this first so that subsequent address
736 // matching finds it last, which will prevent it from trying to match it
737 // as the scaled value in case it happens to be a mul. That would be
738 // problematic if we've sunk a different mul for the scale, because then
739 // we'd end up sinking both muls.
740 if (AddrMode.BaseReg) {
741 Value *V = AddrMode.BaseReg;
742 if (V->getType()->isPointerTy())
743 V = new PtrToIntInst(V, IntPtrTy, "sunkaddr", InsertPt);
744 if (V->getType() != IntPtrTy)
745 V = CastInst::CreateIntegerCast(V, IntPtrTy, /*isSigned=*/true,
746 "sunkaddr", InsertPt);
750 // Add the scale value.
751 if (AddrMode.Scale) {
752 Value *V = AddrMode.ScaledReg;
753 if (V->getType() == IntPtrTy) {
755 } else if (V->getType()->isPointerTy()) {
756 V = new PtrToIntInst(V, IntPtrTy, "sunkaddr", InsertPt);
757 } else if (cast<IntegerType>(IntPtrTy)->getBitWidth() <
758 cast<IntegerType>(V->getType())->getBitWidth()) {
759 V = new TruncInst(V, IntPtrTy, "sunkaddr", InsertPt);
761 V = new SExtInst(V, IntPtrTy, "sunkaddr", InsertPt);
763 if (AddrMode.Scale != 1)
764 V = BinaryOperator::CreateMul(V, ConstantInt::get(IntPtrTy,
766 "sunkaddr", InsertPt);
768 Result = BinaryOperator::CreateAdd(Result, V, "sunkaddr", InsertPt);
773 // Add in the BaseGV if present.
774 if (AddrMode.BaseGV) {
775 Value *V = new PtrToIntInst(AddrMode.BaseGV, IntPtrTy, "sunkaddr",
778 Result = BinaryOperator::CreateAdd(Result, V, "sunkaddr", InsertPt);
783 // Add in the Base Offset if present.
784 if (AddrMode.BaseOffs) {
785 Value *V = ConstantInt::get(IntPtrTy, AddrMode.BaseOffs);
787 Result = BinaryOperator::CreateAdd(Result, V, "sunkaddr", InsertPt);
793 SunkAddr = Constant::getNullValue(Addr->getType());
795 SunkAddr = new IntToPtrInst(Result, Addr->getType(), "sunkaddr",InsertPt);
798 MemoryInst->replaceUsesOfWith(Repl, SunkAddr);
800 if (Repl->use_empty()) {
801 RecursivelyDeleteTriviallyDeadInstructions(Repl);
802 // This address is now available for reassignment, so erase the table entry;
803 // we don't want to match some completely different instruction.
810 /// OptimizeInlineAsmInst - If there are any memory operands, use
811 /// OptimizeMemoryInst to sink their address computing into the block when
812 /// possible / profitable.
813 bool CodeGenPrepare::OptimizeInlineAsmInst(Instruction *I, CallSite CS,
814 DenseMap<Value*,Value*> &SunkAddrs) {
815 bool MadeChange = false;
817 TargetLowering::AsmOperandInfoVector TargetConstraints = TLI->ParseConstraints(CS);
819 for (unsigned i = 0, e = TargetConstraints.size(); i != e; ++i) {
820 TargetLowering::AsmOperandInfo &OpInfo = TargetConstraints[i];
822 // Compute the constraint code and ConstraintType to use.
823 TLI->ComputeConstraintToUse(OpInfo, SDValue());
825 if (OpInfo.ConstraintType == TargetLowering::C_Memory &&
827 Value *OpVal = const_cast<Value *>(CS.getArgument(ArgNo++));
828 MadeChange |= OptimizeMemoryInst(I, OpVal, OpVal->getType(), SunkAddrs);
829 } else if (OpInfo.Type == InlineAsm::isInput)
836 /// MoveExtToFormExtLoad - Move a zext or sext fed by a load into the same
837 /// basic block as the load, unless conditions are unfavorable. This allows
838 /// SelectionDAG to fold the extend into the load.
840 bool CodeGenPrepare::MoveExtToFormExtLoad(Instruction *I) {
841 // Look for a load being extended.
842 LoadInst *LI = dyn_cast<LoadInst>(I->getOperand(0));
843 if (!LI) return false;
845 // If they're already in the same block, there's nothing to do.
846 if (LI->getParent() == I->getParent())
849 // If the load has other users and the truncate is not free, this probably
851 if (!LI->hasOneUse() &&
852 TLI && (TLI->isTypeLegal(TLI->getValueType(LI->getType())) ||
853 !TLI->isTypeLegal(TLI->getValueType(I->getType()))) &&
854 !TLI->isTruncateFree(I->getType(), LI->getType()))
857 // Check whether the target supports casts folded into loads.
859 if (isa<ZExtInst>(I))
860 LType = ISD::ZEXTLOAD;
862 assert(isa<SExtInst>(I) && "Unexpected ext type!");
863 LType = ISD::SEXTLOAD;
865 if (TLI && !TLI->isLoadExtLegal(LType, TLI->getValueType(LI->getType())))
868 // Move the extend into the same block as the load, so that SelectionDAG
870 I->removeFromParent();
876 bool CodeGenPrepare::OptimizeExtUses(Instruction *I) {
877 BasicBlock *DefBB = I->getParent();
879 // If the result of a {s|z}ext and its source are both live out, rewrite all
880 // other uses of the source with result of extension.
881 Value *Src = I->getOperand(0);
882 if (Src->hasOneUse())
885 // Only do this xform if truncating is free.
886 if (TLI && !TLI->isTruncateFree(I->getType(), Src->getType()))
889 // Only safe to perform the optimization if the source is also defined in
891 if (!isa<Instruction>(Src) || DefBB != cast<Instruction>(Src)->getParent())
894 bool DefIsLiveOut = false;
895 for (Value::use_iterator UI = I->use_begin(), E = I->use_end();
897 Instruction *User = cast<Instruction>(*UI);
899 // Figure out which BB this ext is used in.
900 BasicBlock *UserBB = User->getParent();
901 if (UserBB == DefBB) continue;
908 // Make sure non of the uses are PHI nodes.
909 for (Value::use_iterator UI = Src->use_begin(), E = Src->use_end();
911 Instruction *User = cast<Instruction>(*UI);
912 BasicBlock *UserBB = User->getParent();
913 if (UserBB == DefBB) continue;
914 // Be conservative. We don't want this xform to end up introducing
915 // reloads just before load / store instructions.
916 if (isa<PHINode>(User) || isa<LoadInst>(User) || isa<StoreInst>(User))
920 // InsertedTruncs - Only insert one trunc in each block once.
921 DenseMap<BasicBlock*, Instruction*> InsertedTruncs;
923 bool MadeChange = false;
924 for (Value::use_iterator UI = Src->use_begin(), E = Src->use_end();
926 Use &TheUse = UI.getUse();
927 Instruction *User = cast<Instruction>(*UI);
929 // Figure out which BB this ext is used in.
930 BasicBlock *UserBB = User->getParent();
931 if (UserBB == DefBB) continue;
933 // Both src and def are live in this block. Rewrite the use.
934 Instruction *&InsertedTrunc = InsertedTruncs[UserBB];
936 if (!InsertedTrunc) {
937 BasicBlock::iterator InsertPt = UserBB->getFirstNonPHI();
939 InsertedTrunc = new TruncInst(I, Src->getType(), "", InsertPt);
942 // Replace a use of the {s|z}ext source with a use of the result.
943 TheUse = InsertedTrunc;
951 // In this pass we look for GEP and cast instructions that are used
952 // across basic blocks and rewrite them to improve basic-block-at-a-time
954 bool CodeGenPrepare::OptimizeBlock(BasicBlock &BB) {
955 bool MadeChange = false;
957 // Split all critical edges where the dest block has a PHI.
958 if (CriticalEdgeSplit) {
959 TerminatorInst *BBTI = BB.getTerminator();
960 if (BBTI->getNumSuccessors() > 1 && !isa<IndirectBrInst>(BBTI)) {
961 for (unsigned i = 0, e = BBTI->getNumSuccessors(); i != e; ++i) {
962 BasicBlock *SuccBB = BBTI->getSuccessor(i);
963 if (isa<PHINode>(SuccBB->begin()) && isCriticalEdge(BBTI, i, true))
964 SplitEdgeNicely(BBTI, i, BackEdges, this);
969 // Keep track of non-local addresses that have been sunk into this block.
970 // This allows us to avoid inserting duplicate code for blocks with multiple
971 // load/stores of the same address.
972 DenseMap<Value*, Value*> SunkAddrs;
974 for (BasicBlock::iterator BBI = BB.begin(), E = BB.end(); BBI != E; ) {
975 Instruction *I = BBI++;
977 if (PHINode *P = dyn_cast<PHINode>(I)) {
978 // It is possible for very late stage optimizations (such as SimplifyCFG)
979 // to introduce PHI nodes too late to be cleaned up. If we detect such a
980 // trivial PHI, go ahead and zap it here.
981 if (Value *V = SimplifyInstruction(P)) {
982 P->replaceAllUsesWith(V);
983 P->eraseFromParent();
985 } else if (CastInst *CI = dyn_cast<CastInst>(I)) {
986 // If the source of the cast is a constant, then this should have
987 // already been constant folded. The only reason NOT to constant fold
988 // it is if something (e.g. LSR) was careful to place the constant
989 // evaluation in a block other than then one that uses it (e.g. to hoist
990 // the address of globals out of a loop). If this is the case, we don't
991 // want to forward-subst the cast.
992 if (isa<Constant>(CI->getOperand(0)))
997 Change = OptimizeNoopCopyExpression(CI, *TLI);
998 MadeChange |= Change;
1001 if (!Change && (isa<ZExtInst>(I) || isa<SExtInst>(I))) {
1002 MadeChange |= MoveExtToFormExtLoad(I);
1003 MadeChange |= OptimizeExtUses(I);
1005 } else if (CmpInst *CI = dyn_cast<CmpInst>(I)) {
1006 MadeChange |= OptimizeCmpExpression(CI);
1007 } else if (LoadInst *LI = dyn_cast<LoadInst>(I)) {
1009 MadeChange |= OptimizeMemoryInst(I, I->getOperand(0), LI->getType(),
1011 } else if (StoreInst *SI = dyn_cast<StoreInst>(I)) {
1013 MadeChange |= OptimizeMemoryInst(I, SI->getOperand(1),
1014 SI->getOperand(0)->getType(),
1016 } else if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(I)) {
1017 if (GEPI->hasAllZeroIndices()) {
1018 /// The GEP operand must be a pointer, so must its result -> BitCast
1019 Instruction *NC = new BitCastInst(GEPI->getOperand(0), GEPI->getType(),
1020 GEPI->getName(), GEPI);
1021 GEPI->replaceAllUsesWith(NC);
1022 GEPI->eraseFromParent();
1026 } else if (CallInst *CI = dyn_cast<CallInst>(I)) {
1027 // If we found an inline asm expession, and if the target knows how to
1028 // lower it to normal LLVM code, do so now.
1029 if (TLI && isa<InlineAsm>(CI->getCalledValue())) {
1030 if (TLI->ExpandInlineAsm(CI)) {
1032 // Avoid processing instructions out of order, which could cause
1033 // reuse before a value is defined.
1036 // Sink address computing for memory operands into the block.
1037 MadeChange |= OptimizeInlineAsmInst(I, &(*CI), SunkAddrs);
1039 // Other CallInst optimizations that don't need to muck with the
1040 // enclosing iterator here.
1041 MadeChange |= OptimizeCallInst(CI);