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/LLVMContext.h"
25 #include "llvm/Pass.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/ADT/DenseMap.h"
33 #include "llvm/ADT/SmallSet.h"
34 #include "llvm/Assembly/Writer.h"
35 #include "llvm/Support/CallSite.h"
36 #include "llvm/Support/CommandLine.h"
37 #include "llvm/Support/Debug.h"
38 #include "llvm/Support/GetElementPtrTypeIterator.h"
39 #include "llvm/Support/PatternMatch.h"
40 #include "llvm/Support/raw_ostream.h"
42 using namespace llvm::PatternMatch;
44 static cl::opt<bool> FactorCommonPreds("split-critical-paths-tweak",
45 cl::init(false), cl::Hidden);
48 class CodeGenPrepare : public FunctionPass {
49 /// TLI - Keep a pointer of a TargetLowering to consult for determining
50 /// transformation profitability.
51 const TargetLowering *TLI;
54 /// BackEdges - Keep a set of all the loop back edges.
56 SmallSet<std::pair<const BasicBlock*, const BasicBlock*>, 8> BackEdges;
58 static char ID; // Pass identification, replacement for typeid
59 explicit CodeGenPrepare(const TargetLowering *tli = 0)
60 : FunctionPass(&ID), TLI(tli) {}
61 bool runOnFunction(Function &F);
63 virtual void getAnalysisUsage(AnalysisUsage &AU) const {
64 AU.addPreserved<ProfileInfo>();
68 bool EliminateMostlyEmptyBlocks(Function &F);
69 bool CanMergeBlocks(const BasicBlock *BB, const BasicBlock *DestBB) const;
70 void EliminateMostlyEmptyBlock(BasicBlock *BB);
71 bool OptimizeBlock(BasicBlock &BB);
72 bool OptimizeMemoryInst(Instruction *I, Value *Addr, const Type *AccessTy,
73 DenseMap<Value*,Value*> &SunkAddrs);
74 bool OptimizeInlineAsmInst(Instruction *I, CallSite CS,
75 DenseMap<Value*,Value*> &SunkAddrs);
76 bool MoveExtToFormExtLoad(Instruction *I);
77 bool OptimizeExtUses(Instruction *I);
78 void findLoopBackEdges(const Function &F);
82 char CodeGenPrepare::ID = 0;
83 static RegisterPass<CodeGenPrepare> X("codegenprepare",
84 "Optimize for code generation");
86 FunctionPass *llvm::createCodeGenPreparePass(const TargetLowering *TLI) {
87 return new CodeGenPrepare(TLI);
90 /// findLoopBackEdges - Do a DFS walk to find loop back edges.
92 void CodeGenPrepare::findLoopBackEdges(const Function &F) {
93 SmallVector<std::pair<const BasicBlock*,const BasicBlock*>, 32> Edges;
94 FindFunctionBackedges(F, Edges);
96 BackEdges.insert(Edges.begin(), Edges.end());
100 bool CodeGenPrepare::runOnFunction(Function &F) {
101 bool EverMadeChange = false;
103 PI = getAnalysisIfAvailable<ProfileInfo>();
104 // First pass, eliminate blocks that contain only PHI nodes and an
105 // unconditional branch.
106 EverMadeChange |= EliminateMostlyEmptyBlocks(F);
108 // Now find loop back edges.
109 findLoopBackEdges(F);
111 bool MadeChange = true;
114 for (Function::iterator BB = F.begin(), E = F.end(); BB != E; ++BB)
115 MadeChange |= OptimizeBlock(*BB);
116 EverMadeChange |= MadeChange;
118 return EverMadeChange;
121 /// EliminateMostlyEmptyBlocks - eliminate blocks that contain only PHI nodes,
122 /// debug info directives, and an unconditional branch. Passes before isel
123 /// (e.g. LSR/loopsimplify) often split edges in ways that are non-optimal for
124 /// isel. Start by eliminating these blocks so we can split them the way we
126 bool CodeGenPrepare::EliminateMostlyEmptyBlocks(Function &F) {
127 bool MadeChange = false;
128 // Note that this intentionally skips the entry block.
129 for (Function::iterator I = ++F.begin(), E = F.end(); I != E; ) {
130 BasicBlock *BB = I++;
132 // If this block doesn't end with an uncond branch, ignore it.
133 BranchInst *BI = dyn_cast<BranchInst>(BB->getTerminator());
134 if (!BI || !BI->isUnconditional())
137 // If the instruction before the branch (skipping debug info) isn't a phi
138 // node, then other stuff is happening here.
139 BasicBlock::iterator BBI = BI;
140 if (BBI != BB->begin()) {
142 while (isa<DbgInfoIntrinsic>(BBI)) {
143 if (BBI == BB->begin())
147 if (!isa<DbgInfoIntrinsic>(BBI) && !isa<PHINode>(BBI))
151 // Do not break infinite loops.
152 BasicBlock *DestBB = BI->getSuccessor(0);
156 if (!CanMergeBlocks(BB, DestBB))
159 EliminateMostlyEmptyBlock(BB);
165 /// CanMergeBlocks - Return true if we can merge BB into DestBB if there is a
166 /// single uncond branch between them, and BB contains no other non-phi
168 bool CodeGenPrepare::CanMergeBlocks(const BasicBlock *BB,
169 const BasicBlock *DestBB) const {
170 // We only want to eliminate blocks whose phi nodes are used by phi nodes in
171 // the successor. If there are more complex condition (e.g. preheaders),
172 // don't mess around with them.
173 BasicBlock::const_iterator BBI = BB->begin();
174 while (const PHINode *PN = dyn_cast<PHINode>(BBI++)) {
175 for (Value::use_const_iterator UI = PN->use_begin(), E = PN->use_end();
177 const Instruction *User = cast<Instruction>(*UI);
178 if (User->getParent() != DestBB || !isa<PHINode>(User))
180 // If User is inside DestBB block and it is a PHINode then check
181 // incoming value. If incoming value is not from BB then this is
182 // a complex condition (e.g. preheaders) we want to avoid here.
183 if (User->getParent() == DestBB) {
184 if (const PHINode *UPN = dyn_cast<PHINode>(User))
185 for (unsigned I = 0, E = UPN->getNumIncomingValues(); I != E; ++I) {
186 Instruction *Insn = dyn_cast<Instruction>(UPN->getIncomingValue(I));
187 if (Insn && Insn->getParent() == BB &&
188 Insn->getParent() != UPN->getIncomingBlock(I))
195 // If BB and DestBB contain any common predecessors, then the phi nodes in BB
196 // and DestBB may have conflicting incoming values for the block. If so, we
197 // can't merge the block.
198 const PHINode *DestBBPN = dyn_cast<PHINode>(DestBB->begin());
199 if (!DestBBPN) return true; // no conflict.
201 // Collect the preds of BB.
202 SmallPtrSet<const BasicBlock*, 16> BBPreds;
203 if (const PHINode *BBPN = dyn_cast<PHINode>(BB->begin())) {
204 // It is faster to get preds from a PHI than with pred_iterator.
205 for (unsigned i = 0, e = BBPN->getNumIncomingValues(); i != e; ++i)
206 BBPreds.insert(BBPN->getIncomingBlock(i));
208 BBPreds.insert(pred_begin(BB), pred_end(BB));
211 // Walk the preds of DestBB.
212 for (unsigned i = 0, e = DestBBPN->getNumIncomingValues(); i != e; ++i) {
213 BasicBlock *Pred = DestBBPN->getIncomingBlock(i);
214 if (BBPreds.count(Pred)) { // Common predecessor?
215 BBI = DestBB->begin();
216 while (const PHINode *PN = dyn_cast<PHINode>(BBI++)) {
217 const Value *V1 = PN->getIncomingValueForBlock(Pred);
218 const Value *V2 = PN->getIncomingValueForBlock(BB);
220 // If V2 is a phi node in BB, look up what the mapped value will be.
221 if (const PHINode *V2PN = dyn_cast<PHINode>(V2))
222 if (V2PN->getParent() == BB)
223 V2 = V2PN->getIncomingValueForBlock(Pred);
225 // If there is a conflict, bail out.
226 if (V1 != V2) return false;
235 /// EliminateMostlyEmptyBlock - Eliminate a basic block that have only phi's and
236 /// an unconditional branch in it.
237 void CodeGenPrepare::EliminateMostlyEmptyBlock(BasicBlock *BB) {
238 BranchInst *BI = cast<BranchInst>(BB->getTerminator());
239 BasicBlock *DestBB = BI->getSuccessor(0);
241 DEBUG(errs() << "MERGING MOSTLY EMPTY BLOCKS - BEFORE:\n" << *BB << *DestBB);
243 // If the destination block has a single pred, then this is a trivial edge,
245 if (BasicBlock *SinglePred = DestBB->getSinglePredecessor()) {
246 if (SinglePred != DestBB) {
247 // Remember if SinglePred was the entry block of the function. If so, we
248 // will need to move BB back to the entry position.
249 bool isEntry = SinglePred == &SinglePred->getParent()->getEntryBlock();
250 MergeBasicBlockIntoOnlyPred(DestBB, this);
252 if (isEntry && BB != &BB->getParent()->getEntryBlock())
253 BB->moveBefore(&BB->getParent()->getEntryBlock());
255 DEBUG(errs() << "AFTER:\n" << *DestBB << "\n\n\n");
260 // Otherwise, we have multiple predecessors of BB. Update the PHIs in DestBB
261 // to handle the new incoming edges it is about to have.
263 for (BasicBlock::iterator BBI = DestBB->begin();
264 (PN = dyn_cast<PHINode>(BBI)); ++BBI) {
265 // Remove the incoming value for BB, and remember it.
266 Value *InVal = PN->removeIncomingValue(BB, false);
268 // Two options: either the InVal is a phi node defined in BB or it is some
269 // value that dominates BB.
270 PHINode *InValPhi = dyn_cast<PHINode>(InVal);
271 if (InValPhi && InValPhi->getParent() == BB) {
272 // Add all of the input values of the input PHI as inputs of this phi.
273 for (unsigned i = 0, e = InValPhi->getNumIncomingValues(); i != e; ++i)
274 PN->addIncoming(InValPhi->getIncomingValue(i),
275 InValPhi->getIncomingBlock(i));
277 // Otherwise, add one instance of the dominating value for each edge that
278 // we will be adding.
279 if (PHINode *BBPN = dyn_cast<PHINode>(BB->begin())) {
280 for (unsigned i = 0, e = BBPN->getNumIncomingValues(); i != e; ++i)
281 PN->addIncoming(InVal, BBPN->getIncomingBlock(i));
283 for (pred_iterator PI = pred_begin(BB), E = pred_end(BB); PI != E; ++PI)
284 PN->addIncoming(InVal, *PI);
289 // The PHIs are now updated, change everything that refers to BB to use
290 // DestBB and remove BB.
291 BB->replaceAllUsesWith(DestBB);
293 PI->replaceAllUses(BB, DestBB);
294 PI->removeEdge(ProfileInfo::getEdge(BB, DestBB));
296 BB->eraseFromParent();
298 DEBUG(errs() << "AFTER:\n" << *DestBB << "\n\n\n");
302 /// SplitEdgeNicely - Split the critical edge from TI to its specified
303 /// successor if it will improve codegen. We only do this if the successor has
304 /// phi nodes (otherwise critical edges are ok). If there is already another
305 /// predecessor of the succ that is empty (and thus has no phi nodes), use it
306 /// instead of introducing a new block.
307 static void SplitEdgeNicely(TerminatorInst *TI, unsigned SuccNum,
308 SmallSet<std::pair<const BasicBlock*,
309 const BasicBlock*>, 8> &BackEdges,
311 BasicBlock *TIBB = TI->getParent();
312 BasicBlock *Dest = TI->getSuccessor(SuccNum);
313 assert(isa<PHINode>(Dest->begin()) &&
314 "This should only be called if Dest has a PHI!");
316 // Do not split edges to EH landing pads.
317 if (InvokeInst *Invoke = dyn_cast<InvokeInst>(TI)) {
318 if (Invoke->getSuccessor(1) == Dest)
323 // As a hack, never split backedges of loops. Even though the copy for any
324 // PHIs inserted on the backedge would be dead for exits from the loop, we
325 // assume that the cost of *splitting* the backedge would be too high.
326 if (BackEdges.count(std::make_pair(TIBB, Dest)))
329 if (!FactorCommonPreds) {
330 /// TIPHIValues - This array is lazily computed to determine the values of
331 /// PHIs in Dest that TI would provide.
332 SmallVector<Value*, 32> TIPHIValues;
334 // Check to see if Dest has any blocks that can be used as a split edge for
336 for (pred_iterator PI = pred_begin(Dest), E = pred_end(Dest); PI != E; ++PI) {
337 BasicBlock *Pred = *PI;
338 // To be usable, the pred has to end with an uncond branch to the dest.
339 BranchInst *PredBr = dyn_cast<BranchInst>(Pred->getTerminator());
340 if (!PredBr || !PredBr->isUnconditional())
342 // Must be empty other than the branch and debug info.
343 BasicBlock::iterator I = Pred->begin();
344 while (isa<DbgInfoIntrinsic>(I))
346 if (dyn_cast<Instruction>(I) != PredBr)
348 // Cannot be the entry block; its label does not get emitted.
349 if (Pred == &(Dest->getParent()->getEntryBlock()))
352 // Finally, since we know that Dest has phi nodes in it, we have to make
353 // sure that jumping to Pred will have the same effect as going to Dest in
354 // terms of PHI values.
357 bool FoundMatch = true;
358 for (BasicBlock::iterator I = Dest->begin();
359 (PN = dyn_cast<PHINode>(I)); ++I, ++PHINo) {
360 if (PHINo == TIPHIValues.size())
361 TIPHIValues.push_back(PN->getIncomingValueForBlock(TIBB));
363 // If the PHI entry doesn't work, we can't use this pred.
364 if (TIPHIValues[PHINo] != PN->getIncomingValueForBlock(Pred)) {
370 // If we found a workable predecessor, change TI to branch to Succ.
372 ProfileInfo *PI = P->getAnalysisIfAvailable<ProfileInfo>();
374 PI->splitEdge(TIBB, Dest, Pred);
375 Dest->removePredecessor(TIBB);
376 TI->setSuccessor(SuccNum, Pred);
381 SplitCriticalEdge(TI, SuccNum, P, true);
386 SmallVector<Value*, 8> TIPHIValues;
387 for (BasicBlock::iterator I = Dest->begin();
388 (PN = dyn_cast<PHINode>(I)); ++I)
389 TIPHIValues.push_back(PN->getIncomingValueForBlock(TIBB));
391 SmallVector<BasicBlock*, 8> IdenticalPreds;
392 for (pred_iterator PI = pred_begin(Dest), E = pred_end(Dest); PI != E; ++PI) {
393 BasicBlock *Pred = *PI;
394 if (BackEdges.count(std::make_pair(Pred, Dest)))
397 IdenticalPreds.push_back(Pred);
399 bool Identical = true;
401 for (BasicBlock::iterator I = Dest->begin();
402 (PN = dyn_cast<PHINode>(I)); ++I, ++PHINo)
403 if (TIPHIValues[PHINo] != PN->getIncomingValueForBlock(Pred)) {
408 IdenticalPreds.push_back(Pred);
412 assert(!IdenticalPreds.empty());
413 SplitBlockPredecessors(Dest, &IdenticalPreds[0], IdenticalPreds.size(),
418 /// OptimizeNoopCopyExpression - If the specified cast instruction is a noop
419 /// copy (e.g. it's casting from one pointer type to another, i32->i8 on PPC),
420 /// sink it into user blocks to reduce the number of virtual
421 /// registers that must be created and coalesced.
423 /// Return true if any changes are made.
425 static bool OptimizeNoopCopyExpression(CastInst *CI, const TargetLowering &TLI){
426 // If this is a noop copy,
427 EVT SrcVT = TLI.getValueType(CI->getOperand(0)->getType());
428 EVT DstVT = TLI.getValueType(CI->getType());
430 // This is an fp<->int conversion?
431 if (SrcVT.isInteger() != DstVT.isInteger())
434 // If this is an extension, it will be a zero or sign extension, which
436 if (SrcVT.bitsLT(DstVT)) return false;
438 // If these values will be promoted, find out what they will be promoted
439 // to. This helps us consider truncates on PPC as noop copies when they
441 if (TLI.getTypeAction(CI->getContext(), SrcVT) == TargetLowering::Promote)
442 SrcVT = TLI.getTypeToTransformTo(CI->getContext(), SrcVT);
443 if (TLI.getTypeAction(CI->getContext(), DstVT) == TargetLowering::Promote)
444 DstVT = TLI.getTypeToTransformTo(CI->getContext(), DstVT);
446 // If, after promotion, these are the same types, this is a noop copy.
450 BasicBlock *DefBB = CI->getParent();
452 /// InsertedCasts - Only insert a cast in each block once.
453 DenseMap<BasicBlock*, CastInst*> InsertedCasts;
455 bool MadeChange = false;
456 for (Value::use_iterator UI = CI->use_begin(), E = CI->use_end();
458 Use &TheUse = UI.getUse();
459 Instruction *User = cast<Instruction>(*UI);
461 // Figure out which BB this cast is used in. For PHI's this is the
462 // appropriate predecessor block.
463 BasicBlock *UserBB = User->getParent();
464 if (PHINode *PN = dyn_cast<PHINode>(User)) {
465 UserBB = PN->getIncomingBlock(UI);
468 // Preincrement use iterator so we don't invalidate it.
471 // If this user is in the same block as the cast, don't change the cast.
472 if (UserBB == DefBB) continue;
474 // If we have already inserted a cast into this block, use it.
475 CastInst *&InsertedCast = InsertedCasts[UserBB];
478 BasicBlock::iterator InsertPt = UserBB->getFirstNonPHI();
481 CastInst::Create(CI->getOpcode(), CI->getOperand(0), CI->getType(), "",
486 // Replace a use of the cast with a use of the new cast.
487 TheUse = InsertedCast;
490 // If we removed all uses, nuke the cast.
491 if (CI->use_empty()) {
492 CI->eraseFromParent();
499 /// OptimizeCmpExpression - sink the given CmpInst into user blocks to reduce
500 /// the number of virtual registers that must be created and coalesced. This is
501 /// a clear win except on targets with multiple condition code registers
502 /// (PowerPC), where it might lose; some adjustment may be wanted there.
504 /// Return true if any changes are made.
505 static bool OptimizeCmpExpression(CmpInst *CI) {
506 BasicBlock *DefBB = CI->getParent();
508 /// InsertedCmp - Only insert a cmp in each block once.
509 DenseMap<BasicBlock*, CmpInst*> InsertedCmps;
511 bool MadeChange = false;
512 for (Value::use_iterator UI = CI->use_begin(), E = CI->use_end();
514 Use &TheUse = UI.getUse();
515 Instruction *User = cast<Instruction>(*UI);
517 // Preincrement use iterator so we don't invalidate it.
520 // Don't bother for PHI nodes.
521 if (isa<PHINode>(User))
524 // Figure out which BB this cmp is used in.
525 BasicBlock *UserBB = User->getParent();
527 // If this user is in the same block as the cmp, don't change the cmp.
528 if (UserBB == DefBB) continue;
530 // If we have already inserted a cmp into this block, use it.
531 CmpInst *&InsertedCmp = InsertedCmps[UserBB];
534 BasicBlock::iterator InsertPt = UserBB->getFirstNonPHI();
537 CmpInst::Create(CI->getOpcode(),
538 CI->getPredicate(), CI->getOperand(0),
539 CI->getOperand(1), "", InsertPt);
543 // Replace a use of the cmp with a use of the new cmp.
544 TheUse = InsertedCmp;
547 // If we removed all uses, nuke the cmp.
549 CI->eraseFromParent();
554 //===----------------------------------------------------------------------===//
555 // Memory Optimization
556 //===----------------------------------------------------------------------===//
558 /// IsNonLocalValue - Return true if the specified values are defined in a
559 /// different basic block than BB.
560 static bool IsNonLocalValue(Value *V, BasicBlock *BB) {
561 if (Instruction *I = dyn_cast<Instruction>(V))
562 return I->getParent() != BB;
566 /// OptimizeMemoryInst - Load and Store Instructions have often have
567 /// addressing modes that can do significant amounts of computation. As such,
568 /// instruction selection will try to get the load or store to do as much
569 /// computation as possible for the program. The problem is that isel can only
570 /// see within a single block. As such, we sink as much legal addressing mode
571 /// stuff into the block as possible.
573 /// This method is used to optimize both load/store and inline asms with memory
575 bool CodeGenPrepare::OptimizeMemoryInst(Instruction *MemoryInst, Value *Addr,
576 const Type *AccessTy,
577 DenseMap<Value*,Value*> &SunkAddrs) {
578 // Figure out what addressing mode will be built up for this operation.
579 SmallVector<Instruction*, 16> AddrModeInsts;
580 ExtAddrMode AddrMode = AddressingModeMatcher::Match(Addr, AccessTy,MemoryInst,
581 AddrModeInsts, *TLI);
583 // Check to see if any of the instructions supersumed by this addr mode are
584 // non-local to I's BB.
585 bool AnyNonLocal = false;
586 for (unsigned i = 0, e = AddrModeInsts.size(); i != e; ++i) {
587 if (IsNonLocalValue(AddrModeInsts[i], MemoryInst->getParent())) {
593 // If all the instructions matched are already in this BB, don't do anything.
595 DEBUG(errs() << "CGP: Found local addrmode: " << AddrMode << "\n");
599 // Insert this computation right after this user. Since our caller is
600 // scanning from the top of the BB to the bottom, reuse of the expr are
601 // guaranteed to happen later.
602 BasicBlock::iterator InsertPt = MemoryInst;
604 // Now that we determined the addressing expression we want to use and know
605 // that we have to sink it into this block. Check to see if we have already
606 // done this for some other load/store instr in this block. If so, reuse the
608 Value *&SunkAddr = SunkAddrs[Addr];
610 DEBUG(errs() << "CGP: Reusing nonlocal addrmode: " << AddrMode << " for "
612 if (SunkAddr->getType() != Addr->getType())
613 SunkAddr = new BitCastInst(SunkAddr, Addr->getType(), "tmp", InsertPt);
615 DEBUG(errs() << "CGP: SINKING nonlocal addrmode: " << AddrMode << " for "
617 const Type *IntPtrTy =
618 TLI->getTargetData()->getIntPtrType(AccessTy->getContext());
621 // Start with the scale value.
622 if (AddrMode.Scale) {
623 Value *V = AddrMode.ScaledReg;
624 if (V->getType() == IntPtrTy) {
626 } else if (isa<PointerType>(V->getType())) {
627 V = new PtrToIntInst(V, IntPtrTy, "sunkaddr", InsertPt);
628 } else if (cast<IntegerType>(IntPtrTy)->getBitWidth() <
629 cast<IntegerType>(V->getType())->getBitWidth()) {
630 V = new TruncInst(V, IntPtrTy, "sunkaddr", InsertPt);
632 V = new SExtInst(V, IntPtrTy, "sunkaddr", InsertPt);
634 if (AddrMode.Scale != 1)
635 V = BinaryOperator::CreateMul(V, ConstantInt::get(IntPtrTy,
637 "sunkaddr", InsertPt);
641 // Add in the base register.
642 if (AddrMode.BaseReg) {
643 Value *V = AddrMode.BaseReg;
644 if (isa<PointerType>(V->getType()))
645 V = new PtrToIntInst(V, IntPtrTy, "sunkaddr", InsertPt);
646 if (V->getType() != IntPtrTy)
647 V = CastInst::CreateIntegerCast(V, IntPtrTy, /*isSigned=*/true,
648 "sunkaddr", InsertPt);
650 Result = BinaryOperator::CreateAdd(Result, V, "sunkaddr", InsertPt);
655 // Add in the BaseGV if present.
656 if (AddrMode.BaseGV) {
657 Value *V = new PtrToIntInst(AddrMode.BaseGV, IntPtrTy, "sunkaddr",
660 Result = BinaryOperator::CreateAdd(Result, V, "sunkaddr", InsertPt);
665 // Add in the Base Offset if present.
666 if (AddrMode.BaseOffs) {
667 Value *V = ConstantInt::get(IntPtrTy, AddrMode.BaseOffs);
669 Result = BinaryOperator::CreateAdd(Result, V, "sunkaddr", InsertPt);
675 SunkAddr = Constant::getNullValue(Addr->getType());
677 SunkAddr = new IntToPtrInst(Result, Addr->getType(), "sunkaddr",InsertPt);
680 MemoryInst->replaceUsesOfWith(Addr, SunkAddr);
682 if (Addr->use_empty())
683 RecursivelyDeleteTriviallyDeadInstructions(Addr);
687 /// OptimizeInlineAsmInst - If there are any memory operands, use
688 /// OptimizeMemoryInst to sink their address computing into the block when
689 /// possible / profitable.
690 bool CodeGenPrepare::OptimizeInlineAsmInst(Instruction *I, CallSite CS,
691 DenseMap<Value*,Value*> &SunkAddrs) {
692 bool MadeChange = false;
693 InlineAsm *IA = cast<InlineAsm>(CS.getCalledValue());
695 // Do a prepass over the constraints, canonicalizing them, and building up the
696 // ConstraintOperands list.
697 std::vector<InlineAsm::ConstraintInfo>
698 ConstraintInfos = IA->ParseConstraints();
700 /// ConstraintOperands - Information about all of the constraints.
701 std::vector<TargetLowering::AsmOperandInfo> ConstraintOperands;
702 unsigned ArgNo = 0; // ArgNo - The argument of the CallInst.
703 for (unsigned i = 0, e = ConstraintInfos.size(); i != e; ++i) {
705 push_back(TargetLowering::AsmOperandInfo(ConstraintInfos[i]));
706 TargetLowering::AsmOperandInfo &OpInfo = ConstraintOperands.back();
708 // Compute the value type for each operand.
709 switch (OpInfo.Type) {
710 case InlineAsm::isOutput:
711 if (OpInfo.isIndirect)
712 OpInfo.CallOperandVal = CS.getArgument(ArgNo++);
714 case InlineAsm::isInput:
715 OpInfo.CallOperandVal = CS.getArgument(ArgNo++);
717 case InlineAsm::isClobber:
722 // Compute the constraint code and ConstraintType to use.
723 TLI->ComputeConstraintToUse(OpInfo, SDValue(),
724 OpInfo.ConstraintType == TargetLowering::C_Memory);
726 if (OpInfo.ConstraintType == TargetLowering::C_Memory &&
728 Value *OpVal = OpInfo.CallOperandVal;
729 MadeChange |= OptimizeMemoryInst(I, OpVal, OpVal->getType(), SunkAddrs);
736 /// MoveExtToFormExtLoad - Move a zext or sext fed by a load into the same
737 /// basic block as the load, unless conditions are unfavorable. This allows
738 /// SelectionDAG to fold the extend into the load.
740 bool CodeGenPrepare::MoveExtToFormExtLoad(Instruction *I) {
741 // Look for a load being extended.
742 LoadInst *LI = dyn_cast<LoadInst>(I->getOperand(0));
743 if (!LI) return false;
745 // If they're already in the same block, there's nothing to do.
746 if (LI->getParent() == I->getParent())
749 // If the load has other users and the truncate is not free, this probably
751 if (!LI->hasOneUse() &&
752 TLI && !TLI->isTruncateFree(I->getType(), LI->getType()))
755 // Check whether the target supports casts folded into loads.
757 if (isa<ZExtInst>(I))
758 LType = ISD::ZEXTLOAD;
760 assert(isa<SExtInst>(I) && "Unexpected ext type!");
761 LType = ISD::SEXTLOAD;
763 if (TLI && !TLI->isLoadExtLegal(LType, TLI->getValueType(LI->getType())))
766 // Move the extend into the same block as the load, so that SelectionDAG
768 I->removeFromParent();
773 bool CodeGenPrepare::OptimizeExtUses(Instruction *I) {
774 BasicBlock *DefBB = I->getParent();
776 // If both result of the {s|z}xt and its source are live out, rewrite all
777 // other uses of the source with result of extension.
778 Value *Src = I->getOperand(0);
779 if (Src->hasOneUse())
782 // Only do this xform if truncating is free.
783 if (TLI && !TLI->isTruncateFree(I->getType(), Src->getType()))
786 // Only safe to perform the optimization if the source is also defined in
788 if (!isa<Instruction>(Src) || DefBB != cast<Instruction>(Src)->getParent())
791 bool DefIsLiveOut = false;
792 for (Value::use_iterator UI = I->use_begin(), E = I->use_end();
794 Instruction *User = cast<Instruction>(*UI);
796 // Figure out which BB this ext is used in.
797 BasicBlock *UserBB = User->getParent();
798 if (UserBB == DefBB) continue;
805 // Make sure non of the uses are PHI nodes.
806 for (Value::use_iterator UI = Src->use_begin(), E = Src->use_end();
808 Instruction *User = cast<Instruction>(*UI);
809 BasicBlock *UserBB = User->getParent();
810 if (UserBB == DefBB) continue;
811 // Be conservative. We don't want this xform to end up introducing
812 // reloads just before load / store instructions.
813 if (isa<PHINode>(User) || isa<LoadInst>(User) || isa<StoreInst>(User))
817 // InsertedTruncs - Only insert one trunc in each block once.
818 DenseMap<BasicBlock*, Instruction*> InsertedTruncs;
820 bool MadeChange = false;
821 for (Value::use_iterator UI = Src->use_begin(), E = Src->use_end();
823 Use &TheUse = UI.getUse();
824 Instruction *User = cast<Instruction>(*UI);
826 // Figure out which BB this ext is used in.
827 BasicBlock *UserBB = User->getParent();
828 if (UserBB == DefBB) continue;
830 // Both src and def are live in this block. Rewrite the use.
831 Instruction *&InsertedTrunc = InsertedTruncs[UserBB];
833 if (!InsertedTrunc) {
834 BasicBlock::iterator InsertPt = UserBB->getFirstNonPHI();
836 InsertedTrunc = new TruncInst(I, Src->getType(), "", InsertPt);
839 // Replace a use of the {s|z}ext source with a use of the result.
840 TheUse = InsertedTrunc;
848 // In this pass we look for GEP and cast instructions that are used
849 // across basic blocks and rewrite them to improve basic-block-at-a-time
851 bool CodeGenPrepare::OptimizeBlock(BasicBlock &BB) {
852 bool MadeChange = false;
854 // Split all critical edges where the dest block has a PHI.
855 TerminatorInst *BBTI = BB.getTerminator();
856 if (BBTI->getNumSuccessors() > 1 && !isa<IndirectBrInst>(BBTI)) {
857 for (unsigned i = 0, e = BBTI->getNumSuccessors(); i != e; ++i) {
858 BasicBlock *SuccBB = BBTI->getSuccessor(i);
859 if (isa<PHINode>(SuccBB->begin()) && isCriticalEdge(BBTI, i, true))
860 SplitEdgeNicely(BBTI, i, BackEdges, this);
864 // Keep track of non-local addresses that have been sunk into this block.
865 // This allows us to avoid inserting duplicate code for blocks with multiple
866 // load/stores of the same address.
867 DenseMap<Value*, Value*> SunkAddrs;
869 for (BasicBlock::iterator BBI = BB.begin(), E = BB.end(); BBI != E; ) {
870 Instruction *I = BBI++;
872 if (CastInst *CI = dyn_cast<CastInst>(I)) {
873 // If the source of the cast is a constant, then this should have
874 // already been constant folded. The only reason NOT to constant fold
875 // it is if something (e.g. LSR) was careful to place the constant
876 // evaluation in a block other than then one that uses it (e.g. to hoist
877 // the address of globals out of a loop). If this is the case, we don't
878 // want to forward-subst the cast.
879 if (isa<Constant>(CI->getOperand(0)))
884 Change = OptimizeNoopCopyExpression(CI, *TLI);
885 MadeChange |= Change;
888 if (!Change && (isa<ZExtInst>(I) || isa<SExtInst>(I))) {
889 MadeChange |= MoveExtToFormExtLoad(I);
890 MadeChange |= OptimizeExtUses(I);
892 } else if (CmpInst *CI = dyn_cast<CmpInst>(I)) {
893 MadeChange |= OptimizeCmpExpression(CI);
894 } else if (LoadInst *LI = dyn_cast<LoadInst>(I)) {
896 MadeChange |= OptimizeMemoryInst(I, I->getOperand(0), LI->getType(),
898 } else if (StoreInst *SI = dyn_cast<StoreInst>(I)) {
900 MadeChange |= OptimizeMemoryInst(I, SI->getOperand(1),
901 SI->getOperand(0)->getType(),
903 } else if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(I)) {
904 if (GEPI->hasAllZeroIndices()) {
905 /// The GEP operand must be a pointer, so must its result -> BitCast
906 Instruction *NC = new BitCastInst(GEPI->getOperand(0), GEPI->getType(),
907 GEPI->getName(), GEPI);
908 GEPI->replaceAllUsesWith(NC);
909 GEPI->eraseFromParent();
913 } else if (CallInst *CI = dyn_cast<CallInst>(I)) {
914 // If we found an inline asm expession, and if the target knows how to
915 // lower it to normal LLVM code, do so now.
916 if (TLI && isa<InlineAsm>(CI->getCalledValue())) {
917 if (TLI->ExpandInlineAsm(CI)) {
919 // Avoid processing instructions out of order, which could cause
920 // reuse before a value is defined.
923 // Sink address computing for memory operands into the block.
924 MadeChange |= OptimizeInlineAsmInst(I, &(*CI), SunkAddrs);