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/Target/TargetAsmInfo.h"
27 #include "llvm/Target/TargetData.h"
28 #include "llvm/Target/TargetLowering.h"
29 #include "llvm/Target/TargetMachine.h"
30 #include "llvm/Transforms/Utils/AddrModeMatcher.h"
31 #include "llvm/Transforms/Utils/BasicBlockUtils.h"
32 #include "llvm/Transforms/Utils/Local.h"
33 #include "llvm/ADT/DenseMap.h"
34 #include "llvm/ADT/SmallSet.h"
35 #include "llvm/Assembly/Writer.h"
36 #include "llvm/Support/CallSite.h"
37 #include "llvm/Support/CommandLine.h"
38 #include "llvm/Support/Compiler.h"
39 #include "llvm/Support/Debug.h"
40 #include "llvm/Support/GetElementPtrTypeIterator.h"
41 #include "llvm/Support/PatternMatch.h"
43 using namespace llvm::PatternMatch;
45 static cl::opt<bool> FactorCommonPreds("split-critical-paths-tweak",
46 cl::init(false), cl::Hidden);
49 class VISIBILITY_HIDDEN CodeGenPrepare : public FunctionPass {
50 /// TLI - Keep a pointer of a TargetLowering to consult for determining
51 /// transformation profitability.
52 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);
64 bool EliminateMostlyEmptyBlocks(Function &F);
65 bool CanMergeBlocks(const BasicBlock *BB, const BasicBlock *DestBB) const;
66 void EliminateMostlyEmptyBlock(BasicBlock *BB);
67 bool OptimizeBlock(BasicBlock &BB);
68 bool OptimizeMemoryInst(Instruction *I, Value *Addr, const Type *AccessTy,
69 DenseMap<Value*,Value*> &SunkAddrs);
70 bool OptimizeInlineAsmInst(Instruction *I, CallSite CS,
71 DenseMap<Value*,Value*> &SunkAddrs);
72 bool OptimizeExtUses(Instruction *I);
73 void findLoopBackEdges(const Function &F);
77 char CodeGenPrepare::ID = 0;
78 static RegisterPass<CodeGenPrepare> X("codegenprepare",
79 "Optimize for code generation");
81 FunctionPass *llvm::createCodeGenPreparePass(const TargetLowering *TLI) {
82 return new CodeGenPrepare(TLI);
85 /// findLoopBackEdges - Do a DFS walk to find loop back edges.
87 void CodeGenPrepare::findLoopBackEdges(const Function &F) {
88 SmallVector<std::pair<const BasicBlock*,const BasicBlock*>, 32> Edges;
89 FindFunctionBackedges(F, Edges);
91 BackEdges.insert(Edges.begin(), Edges.end());
95 bool CodeGenPrepare::runOnFunction(Function &F) {
96 bool EverMadeChange = false;
98 // First pass, eliminate blocks that contain only PHI nodes and an
99 // unconditional branch.
100 EverMadeChange |= EliminateMostlyEmptyBlocks(F);
102 // Now find loop back edges.
103 findLoopBackEdges(F);
105 bool MadeChange = true;
108 for (Function::iterator BB = F.begin(), E = F.end(); BB != E; ++BB)
109 MadeChange |= OptimizeBlock(*BB);
110 EverMadeChange |= MadeChange;
112 return EverMadeChange;
115 /// EliminateMostlyEmptyBlocks - eliminate blocks that contain only PHI nodes,
116 /// debug info directives, and an unconditional branch. Passes before isel
117 /// (e.g. LSR/loopsimplify) often split edges in ways that are non-optimal for
118 /// isel. Start by eliminating these blocks so we can split them the way we
120 bool CodeGenPrepare::EliminateMostlyEmptyBlocks(Function &F) {
121 bool MadeChange = false;
122 // Note that this intentionally skips the entry block.
123 for (Function::iterator I = ++F.begin(), E = F.end(); I != E; ) {
124 BasicBlock *BB = I++;
126 // If this block doesn't end with an uncond branch, ignore it.
127 BranchInst *BI = dyn_cast<BranchInst>(BB->getTerminator());
128 if (!BI || !BI->isUnconditional())
131 // If the instruction before the branch (skipping debug info) isn't a phi
132 // node, then other stuff is happening here.
133 BasicBlock::iterator BBI = BI;
134 if (BBI != BB->begin()) {
136 while (isa<DbgInfoIntrinsic>(BBI)) {
137 if (BBI == BB->begin())
141 if (!isa<DbgInfoIntrinsic>(BBI) && !isa<PHINode>(BBI))
145 // Do not break infinite loops.
146 BasicBlock *DestBB = BI->getSuccessor(0);
150 if (!CanMergeBlocks(BB, DestBB))
153 EliminateMostlyEmptyBlock(BB);
159 /// CanMergeBlocks - Return true if we can merge BB into DestBB if there is a
160 /// single uncond branch between them, and BB contains no other non-phi
162 bool CodeGenPrepare::CanMergeBlocks(const BasicBlock *BB,
163 const BasicBlock *DestBB) const {
164 // We only want to eliminate blocks whose phi nodes are used by phi nodes in
165 // the successor. If there are more complex condition (e.g. preheaders),
166 // don't mess around with them.
167 BasicBlock::const_iterator BBI = BB->begin();
168 while (const PHINode *PN = dyn_cast<PHINode>(BBI++)) {
169 for (Value::use_const_iterator UI = PN->use_begin(), E = PN->use_end();
171 const Instruction *User = cast<Instruction>(*UI);
172 if (User->getParent() != DestBB || !isa<PHINode>(User))
174 // If User is inside DestBB block and it is a PHINode then check
175 // incoming value. If incoming value is not from BB then this is
176 // a complex condition (e.g. preheaders) we want to avoid here.
177 if (User->getParent() == DestBB) {
178 if (const PHINode *UPN = dyn_cast<PHINode>(User))
179 for (unsigned I = 0, E = UPN->getNumIncomingValues(); I != E; ++I) {
180 Instruction *Insn = dyn_cast<Instruction>(UPN->getIncomingValue(I));
181 if (Insn && Insn->getParent() == BB &&
182 Insn->getParent() != UPN->getIncomingBlock(I))
189 // If BB and DestBB contain any common predecessors, then the phi nodes in BB
190 // and DestBB may have conflicting incoming values for the block. If so, we
191 // can't merge the block.
192 const PHINode *DestBBPN = dyn_cast<PHINode>(DestBB->begin());
193 if (!DestBBPN) return true; // no conflict.
195 // Collect the preds of BB.
196 SmallPtrSet<const BasicBlock*, 16> BBPreds;
197 if (const PHINode *BBPN = dyn_cast<PHINode>(BB->begin())) {
198 // It is faster to get preds from a PHI than with pred_iterator.
199 for (unsigned i = 0, e = BBPN->getNumIncomingValues(); i != e; ++i)
200 BBPreds.insert(BBPN->getIncomingBlock(i));
202 BBPreds.insert(pred_begin(BB), pred_end(BB));
205 // Walk the preds of DestBB.
206 for (unsigned i = 0, e = DestBBPN->getNumIncomingValues(); i != e; ++i) {
207 BasicBlock *Pred = DestBBPN->getIncomingBlock(i);
208 if (BBPreds.count(Pred)) { // Common predecessor?
209 BBI = DestBB->begin();
210 while (const PHINode *PN = dyn_cast<PHINode>(BBI++)) {
211 const Value *V1 = PN->getIncomingValueForBlock(Pred);
212 const Value *V2 = PN->getIncomingValueForBlock(BB);
214 // If V2 is a phi node in BB, look up what the mapped value will be.
215 if (const PHINode *V2PN = dyn_cast<PHINode>(V2))
216 if (V2PN->getParent() == BB)
217 V2 = V2PN->getIncomingValueForBlock(Pred);
219 // If there is a conflict, bail out.
220 if (V1 != V2) return false;
229 /// EliminateMostlyEmptyBlock - Eliminate a basic block that have only phi's and
230 /// an unconditional branch in it.
231 void CodeGenPrepare::EliminateMostlyEmptyBlock(BasicBlock *BB) {
232 BranchInst *BI = cast<BranchInst>(BB->getTerminator());
233 BasicBlock *DestBB = BI->getSuccessor(0);
235 DOUT << "MERGING MOSTLY EMPTY BLOCKS - BEFORE:\n" << *BB << *DestBB;
237 // If the destination block has a single pred, then this is a trivial edge,
239 if (BasicBlock *SinglePred = DestBB->getSinglePredecessor()) {
240 if (SinglePred != DestBB) {
241 // Remember if SinglePred was the entry block of the function. If so, we
242 // will need to move BB back to the entry position.
243 bool isEntry = SinglePred == &SinglePred->getParent()->getEntryBlock();
244 MergeBasicBlockIntoOnlyPred(DestBB);
246 if (isEntry && BB != &BB->getParent()->getEntryBlock())
247 BB->moveBefore(&BB->getParent()->getEntryBlock());
249 DOUT << "AFTER:\n" << *DestBB << "\n\n\n";
254 // Otherwise, we have multiple predecessors of BB. Update the PHIs in DestBB
255 // to handle the new incoming edges it is about to have.
257 for (BasicBlock::iterator BBI = DestBB->begin();
258 (PN = dyn_cast<PHINode>(BBI)); ++BBI) {
259 // Remove the incoming value for BB, and remember it.
260 Value *InVal = PN->removeIncomingValue(BB, false);
262 // Two options: either the InVal is a phi node defined in BB or it is some
263 // value that dominates BB.
264 PHINode *InValPhi = dyn_cast<PHINode>(InVal);
265 if (InValPhi && InValPhi->getParent() == BB) {
266 // Add all of the input values of the input PHI as inputs of this phi.
267 for (unsigned i = 0, e = InValPhi->getNumIncomingValues(); i != e; ++i)
268 PN->addIncoming(InValPhi->getIncomingValue(i),
269 InValPhi->getIncomingBlock(i));
271 // Otherwise, add one instance of the dominating value for each edge that
272 // we will be adding.
273 if (PHINode *BBPN = dyn_cast<PHINode>(BB->begin())) {
274 for (unsigned i = 0, e = BBPN->getNumIncomingValues(); i != e; ++i)
275 PN->addIncoming(InVal, BBPN->getIncomingBlock(i));
277 for (pred_iterator PI = pred_begin(BB), E = pred_end(BB); PI != E; ++PI)
278 PN->addIncoming(InVal, *PI);
283 // The PHIs are now updated, change everything that refers to BB to use
284 // DestBB and remove BB.
285 BB->replaceAllUsesWith(DestBB);
286 BB->eraseFromParent();
288 DOUT << "AFTER:\n" << *DestBB << "\n\n\n";
292 /// SplitEdgeNicely - Split the critical edge from TI to its specified
293 /// successor if it will improve codegen. We only do this if the successor has
294 /// phi nodes (otherwise critical edges are ok). If there is already another
295 /// predecessor of the succ that is empty (and thus has no phi nodes), use it
296 /// instead of introducing a new block.
297 static void SplitEdgeNicely(TerminatorInst *TI, unsigned SuccNum,
298 SmallSet<std::pair<const BasicBlock*,
299 const BasicBlock*>, 8> &BackEdges,
301 BasicBlock *TIBB = TI->getParent();
302 BasicBlock *Dest = TI->getSuccessor(SuccNum);
303 assert(isa<PHINode>(Dest->begin()) &&
304 "This should only be called if Dest has a PHI!");
306 // Do not split edges to EH landing pads.
307 if (InvokeInst *Invoke = dyn_cast<InvokeInst>(TI)) {
308 if (Invoke->getSuccessor(1) == Dest)
312 // As a hack, never split backedges of loops. Even though the copy for any
313 // PHIs inserted on the backedge would be dead for exits from the loop, we
314 // assume that the cost of *splitting* the backedge would be too high.
315 if (BackEdges.count(std::make_pair(TIBB, Dest)))
318 if (!FactorCommonPreds) {
319 /// TIPHIValues - This array is lazily computed to determine the values of
320 /// PHIs in Dest that TI would provide.
321 SmallVector<Value*, 32> TIPHIValues;
323 // Check to see if Dest has any blocks that can be used as a split edge for
325 for (pred_iterator PI = pred_begin(Dest), E = pred_end(Dest); PI != E; ++PI) {
326 BasicBlock *Pred = *PI;
327 // To be usable, the pred has to end with an uncond branch to the dest.
328 BranchInst *PredBr = dyn_cast<BranchInst>(Pred->getTerminator());
329 if (!PredBr || !PredBr->isUnconditional())
331 // Must be empty other than the branch and debug info.
332 BasicBlock::iterator I = Pred->begin();
333 while (isa<DbgInfoIntrinsic>(I))
335 if (dyn_cast<Instruction>(I) != PredBr)
337 // Cannot be the entry block; its label does not get emitted.
338 if (Pred == &(Dest->getParent()->getEntryBlock()))
341 // Finally, since we know that Dest has phi nodes in it, we have to make
342 // sure that jumping to Pred will have the same effect as going to Dest in
343 // terms of PHI values.
346 bool FoundMatch = true;
347 for (BasicBlock::iterator I = Dest->begin();
348 (PN = dyn_cast<PHINode>(I)); ++I, ++PHINo) {
349 if (PHINo == TIPHIValues.size())
350 TIPHIValues.push_back(PN->getIncomingValueForBlock(TIBB));
352 // If the PHI entry doesn't work, we can't use this pred.
353 if (TIPHIValues[PHINo] != PN->getIncomingValueForBlock(Pred)) {
359 // If we found a workable predecessor, change TI to branch to Succ.
361 Dest->removePredecessor(TIBB);
362 TI->setSuccessor(SuccNum, Pred);
367 SplitCriticalEdge(TI, SuccNum, P, true);
372 SmallVector<Value*, 8> TIPHIValues;
373 for (BasicBlock::iterator I = Dest->begin();
374 (PN = dyn_cast<PHINode>(I)); ++I)
375 TIPHIValues.push_back(PN->getIncomingValueForBlock(TIBB));
377 SmallVector<BasicBlock*, 8> IdenticalPreds;
378 for (pred_iterator PI = pred_begin(Dest), E = pred_end(Dest); PI != E; ++PI) {
379 BasicBlock *Pred = *PI;
380 if (BackEdges.count(std::make_pair(Pred, Dest)))
383 IdenticalPreds.push_back(Pred);
385 bool Identical = true;
387 for (BasicBlock::iterator I = Dest->begin();
388 (PN = dyn_cast<PHINode>(I)); ++I, ++PHINo)
389 if (TIPHIValues[PHINo] != PN->getIncomingValueForBlock(Pred)) {
394 IdenticalPreds.push_back(Pred);
398 assert(!IdenticalPreds.empty());
399 SplitBlockPredecessors(Dest, &IdenticalPreds[0], IdenticalPreds.size(),
404 /// OptimizeNoopCopyExpression - If the specified cast instruction is a noop
405 /// copy (e.g. it's casting from one pointer type to another, i32->i8 on PPC),
406 /// sink it into user blocks to reduce the number of virtual
407 /// registers that must be created and coalesced.
409 /// Return true if any changes are made.
411 static bool OptimizeNoopCopyExpression(CastInst *CI, const TargetLowering &TLI){
412 // If this is a noop copy,
413 MVT SrcVT = TLI.getValueType(CI->getOperand(0)->getType());
414 MVT DstVT = TLI.getValueType(CI->getType());
416 // This is an fp<->int conversion?
417 if (SrcVT.isInteger() != DstVT.isInteger())
420 // If this is an extension, it will be a zero or sign extension, which
422 if (SrcVT.bitsLT(DstVT)) return false;
424 // If these values will be promoted, find out what they will be promoted
425 // to. This helps us consider truncates on PPC as noop copies when they
427 if (TLI.getTypeAction(SrcVT) == TargetLowering::Promote)
428 SrcVT = TLI.getTypeToTransformTo(SrcVT);
429 if (TLI.getTypeAction(DstVT) == TargetLowering::Promote)
430 DstVT = TLI.getTypeToTransformTo(DstVT);
432 // If, after promotion, these are the same types, this is a noop copy.
436 BasicBlock *DefBB = CI->getParent();
438 /// InsertedCasts - Only insert a cast in each block once.
439 DenseMap<BasicBlock*, CastInst*> InsertedCasts;
441 bool MadeChange = false;
442 for (Value::use_iterator UI = CI->use_begin(), E = CI->use_end();
444 Use &TheUse = UI.getUse();
445 Instruction *User = cast<Instruction>(*UI);
447 // Figure out which BB this cast is used in. For PHI's this is the
448 // appropriate predecessor block.
449 BasicBlock *UserBB = User->getParent();
450 if (PHINode *PN = dyn_cast<PHINode>(User)) {
451 UserBB = PN->getIncomingBlock(UI);
454 // Preincrement use iterator so we don't invalidate it.
457 // If this user is in the same block as the cast, don't change the cast.
458 if (UserBB == DefBB) continue;
460 // If we have already inserted a cast into this block, use it.
461 CastInst *&InsertedCast = InsertedCasts[UserBB];
464 BasicBlock::iterator InsertPt = UserBB->getFirstNonPHI();
467 CastInst::Create(CI->getOpcode(), CI->getOperand(0), CI->getType(), "",
472 // Replace a use of the cast with a use of the new cast.
473 TheUse = InsertedCast;
476 // If we removed all uses, nuke the cast.
477 if (CI->use_empty()) {
478 CI->eraseFromParent();
485 /// OptimizeCmpExpression - sink the given CmpInst into user blocks to reduce
486 /// the number of virtual registers that must be created and coalesced. This is
487 /// a clear win except on targets with multiple condition code registers
488 /// (PowerPC), where it might lose; some adjustment may be wanted there.
490 /// Return true if any changes are made.
491 static bool OptimizeCmpExpression(CmpInst *CI) {
492 BasicBlock *DefBB = CI->getParent();
494 /// InsertedCmp - Only insert a cmp in each block once.
495 DenseMap<BasicBlock*, CmpInst*> InsertedCmps;
497 bool MadeChange = false;
498 for (Value::use_iterator UI = CI->use_begin(), E = CI->use_end();
500 Use &TheUse = UI.getUse();
501 Instruction *User = cast<Instruction>(*UI);
503 // Preincrement use iterator so we don't invalidate it.
506 // Don't bother for PHI nodes.
507 if (isa<PHINode>(User))
510 // Figure out which BB this cmp is used in.
511 BasicBlock *UserBB = User->getParent();
513 // If this user is in the same block as the cmp, don't change the cmp.
514 if (UserBB == DefBB) continue;
516 // If we have already inserted a cmp into this block, use it.
517 CmpInst *&InsertedCmp = InsertedCmps[UserBB];
520 BasicBlock::iterator InsertPt = UserBB->getFirstNonPHI();
523 CmpInst::Create(CI->getOpcode(), CI->getPredicate(), CI->getOperand(0),
524 CI->getOperand(1), "", InsertPt);
528 // Replace a use of the cmp with a use of the new cmp.
529 TheUse = InsertedCmp;
532 // If we removed all uses, nuke the cmp.
534 CI->eraseFromParent();
539 //===----------------------------------------------------------------------===//
540 // Memory Optimization
541 //===----------------------------------------------------------------------===//
543 /// IsNonLocalValue - Return true if the specified values are defined in a
544 /// different basic block than BB.
545 static bool IsNonLocalValue(Value *V, BasicBlock *BB) {
546 if (Instruction *I = dyn_cast<Instruction>(V))
547 return I->getParent() != BB;
551 /// OptimizeMemoryInst - Load and Store Instructions have often have
552 /// addressing modes that can do significant amounts of computation. As such,
553 /// instruction selection will try to get the load or store to do as much
554 /// computation as possible for the program. The problem is that isel can only
555 /// see within a single block. As such, we sink as much legal addressing mode
556 /// stuff into the block as possible.
558 /// This method is used to optimize both load/store and inline asms with memory
560 bool CodeGenPrepare::OptimizeMemoryInst(Instruction *MemoryInst, Value *Addr,
561 const Type *AccessTy,
562 DenseMap<Value*,Value*> &SunkAddrs) {
563 // Figure out what addressing mode will be built up for this operation.
564 SmallVector<Instruction*, 16> AddrModeInsts;
565 ExtAddrMode AddrMode = AddressingModeMatcher::Match(Addr, AccessTy,MemoryInst,
566 AddrModeInsts, *TLI);
568 // Check to see if any of the instructions supersumed by this addr mode are
569 // non-local to I's BB.
570 bool AnyNonLocal = false;
571 for (unsigned i = 0, e = AddrModeInsts.size(); i != e; ++i) {
572 if (IsNonLocalValue(AddrModeInsts[i], MemoryInst->getParent())) {
578 // If all the instructions matched are already in this BB, don't do anything.
580 DEBUG(cerr << "CGP: Found local addrmode: " << AddrMode << "\n");
584 // Insert this computation right after this user. Since our caller is
585 // scanning from the top of the BB to the bottom, reuse of the expr are
586 // guaranteed to happen later.
587 BasicBlock::iterator InsertPt = MemoryInst;
589 // Now that we determined the addressing expression we want to use and know
590 // that we have to sink it into this block. Check to see if we have already
591 // done this for some other load/store instr in this block. If so, reuse the
593 Value *&SunkAddr = SunkAddrs[Addr];
595 DEBUG(cerr << "CGP: Reusing nonlocal addrmode: " << AddrMode << " for "
597 if (SunkAddr->getType() != Addr->getType())
598 SunkAddr = new BitCastInst(SunkAddr, Addr->getType(), "tmp", InsertPt);
600 DEBUG(cerr << "CGP: SINKING nonlocal addrmode: " << AddrMode << " for "
602 const Type *IntPtrTy = TLI->getTargetData()->getIntPtrType();
605 // Start with the scale value.
606 if (AddrMode.Scale) {
607 Value *V = AddrMode.ScaledReg;
608 if (V->getType() == IntPtrTy) {
610 } else if (isa<PointerType>(V->getType())) {
611 V = new PtrToIntInst(V, IntPtrTy, "sunkaddr", InsertPt);
612 } else if (cast<IntegerType>(IntPtrTy)->getBitWidth() <
613 cast<IntegerType>(V->getType())->getBitWidth()) {
614 V = new TruncInst(V, IntPtrTy, "sunkaddr", InsertPt);
616 V = new SExtInst(V, IntPtrTy, "sunkaddr", InsertPt);
618 if (AddrMode.Scale != 1)
619 V = BinaryOperator::CreateMul(V, Context->getConstantInt(IntPtrTy,
621 "sunkaddr", InsertPt);
625 // Add in the base register.
626 if (AddrMode.BaseReg) {
627 Value *V = AddrMode.BaseReg;
628 if (isa<PointerType>(V->getType()))
629 V = new PtrToIntInst(V, IntPtrTy, "sunkaddr", InsertPt);
630 if (V->getType() != IntPtrTy)
631 V = CastInst::CreateIntegerCast(V, IntPtrTy, /*isSigned=*/true,
632 "sunkaddr", InsertPt);
634 Result = BinaryOperator::CreateAdd(Result, V, "sunkaddr", InsertPt);
639 // Add in the BaseGV if present.
640 if (AddrMode.BaseGV) {
641 Value *V = new PtrToIntInst(AddrMode.BaseGV, IntPtrTy, "sunkaddr",
644 Result = BinaryOperator::CreateAdd(Result, V, "sunkaddr", InsertPt);
649 // Add in the Base Offset if present.
650 if (AddrMode.BaseOffs) {
651 Value *V = Context->getConstantInt(IntPtrTy, AddrMode.BaseOffs);
653 Result = BinaryOperator::CreateAdd(Result, V, "sunkaddr", InsertPt);
659 SunkAddr = Context->getNullValue(Addr->getType());
661 SunkAddr = new IntToPtrInst(Result, Addr->getType(), "sunkaddr",InsertPt);
664 MemoryInst->replaceUsesOfWith(Addr, SunkAddr);
666 if (Addr->use_empty())
667 RecursivelyDeleteTriviallyDeadInstructions(Addr);
671 /// OptimizeInlineAsmInst - If there are any memory operands, use
672 /// OptimizeMemoryInst to sink their address computing into the block when
673 /// possible / profitable.
674 bool CodeGenPrepare::OptimizeInlineAsmInst(Instruction *I, CallSite CS,
675 DenseMap<Value*,Value*> &SunkAddrs) {
676 bool MadeChange = false;
677 InlineAsm *IA = cast<InlineAsm>(CS.getCalledValue());
679 // Do a prepass over the constraints, canonicalizing them, and building up the
680 // ConstraintOperands list.
681 std::vector<InlineAsm::ConstraintInfo>
682 ConstraintInfos = IA->ParseConstraints();
684 /// ConstraintOperands - Information about all of the constraints.
685 std::vector<TargetLowering::AsmOperandInfo> ConstraintOperands;
686 unsigned ArgNo = 0; // ArgNo - The argument of the CallInst.
687 for (unsigned i = 0, e = ConstraintInfos.size(); i != e; ++i) {
689 push_back(TargetLowering::AsmOperandInfo(ConstraintInfos[i]));
690 TargetLowering::AsmOperandInfo &OpInfo = ConstraintOperands.back();
692 // Compute the value type for each operand.
693 switch (OpInfo.Type) {
694 case InlineAsm::isOutput:
695 if (OpInfo.isIndirect)
696 OpInfo.CallOperandVal = CS.getArgument(ArgNo++);
698 case InlineAsm::isInput:
699 OpInfo.CallOperandVal = CS.getArgument(ArgNo++);
701 case InlineAsm::isClobber:
706 // Compute the constraint code and ConstraintType to use.
707 TLI->ComputeConstraintToUse(OpInfo, SDValue(),
708 OpInfo.ConstraintType == TargetLowering::C_Memory);
710 if (OpInfo.ConstraintType == TargetLowering::C_Memory &&
712 Value *OpVal = OpInfo.CallOperandVal;
713 MadeChange |= OptimizeMemoryInst(I, OpVal, OpVal->getType(), SunkAddrs);
720 bool CodeGenPrepare::OptimizeExtUses(Instruction *I) {
721 BasicBlock *DefBB = I->getParent();
723 // If both result of the {s|z}xt and its source are live out, rewrite all
724 // other uses of the source with result of extension.
725 Value *Src = I->getOperand(0);
726 if (Src->hasOneUse())
729 // Only do this xform if truncating is free.
730 if (TLI && !TLI->isTruncateFree(I->getType(), Src->getType()))
733 // Only safe to perform the optimization if the source is also defined in
735 if (!isa<Instruction>(Src) || DefBB != cast<Instruction>(Src)->getParent())
738 bool DefIsLiveOut = false;
739 for (Value::use_iterator UI = I->use_begin(), E = I->use_end();
741 Instruction *User = cast<Instruction>(*UI);
743 // Figure out which BB this ext is used in.
744 BasicBlock *UserBB = User->getParent();
745 if (UserBB == DefBB) continue;
752 // Make sure non of the uses are PHI nodes.
753 for (Value::use_iterator UI = Src->use_begin(), E = Src->use_end();
755 Instruction *User = cast<Instruction>(*UI);
756 BasicBlock *UserBB = User->getParent();
757 if (UserBB == DefBB) continue;
758 // Be conservative. We don't want this xform to end up introducing
759 // reloads just before load / store instructions.
760 if (isa<PHINode>(User) || isa<LoadInst>(User) || isa<StoreInst>(User))
764 // InsertedTruncs - Only insert one trunc in each block once.
765 DenseMap<BasicBlock*, Instruction*> InsertedTruncs;
767 bool MadeChange = false;
768 for (Value::use_iterator UI = Src->use_begin(), E = Src->use_end();
770 Use &TheUse = UI.getUse();
771 Instruction *User = cast<Instruction>(*UI);
773 // Figure out which BB this ext is used in.
774 BasicBlock *UserBB = User->getParent();
775 if (UserBB == DefBB) continue;
777 // Both src and def are live in this block. Rewrite the use.
778 Instruction *&InsertedTrunc = InsertedTruncs[UserBB];
780 if (!InsertedTrunc) {
781 BasicBlock::iterator InsertPt = UserBB->getFirstNonPHI();
783 InsertedTrunc = new TruncInst(I, Src->getType(), "", InsertPt);
786 // Replace a use of the {s|z}ext source with a use of the result.
787 TheUse = InsertedTrunc;
795 // In this pass we look for GEP and cast instructions that are used
796 // across basic blocks and rewrite them to improve basic-block-at-a-time
798 bool CodeGenPrepare::OptimizeBlock(BasicBlock &BB) {
799 bool MadeChange = false;
801 // Split all critical edges where the dest block has a PHI.
802 TerminatorInst *BBTI = BB.getTerminator();
803 if (BBTI->getNumSuccessors() > 1) {
804 for (unsigned i = 0, e = BBTI->getNumSuccessors(); i != e; ++i) {
805 BasicBlock *SuccBB = BBTI->getSuccessor(i);
806 if (isa<PHINode>(SuccBB->begin()) && isCriticalEdge(BBTI, i, true))
807 SplitEdgeNicely(BBTI, i, BackEdges, this);
811 // Keep track of non-local addresses that have been sunk into this block.
812 // This allows us to avoid inserting duplicate code for blocks with multiple
813 // load/stores of the same address.
814 DenseMap<Value*, Value*> SunkAddrs;
816 for (BasicBlock::iterator BBI = BB.begin(), E = BB.end(); BBI != E; ) {
817 Instruction *I = BBI++;
819 if (CastInst *CI = dyn_cast<CastInst>(I)) {
820 // If the source of the cast is a constant, then this should have
821 // already been constant folded. The only reason NOT to constant fold
822 // it is if something (e.g. LSR) was careful to place the constant
823 // evaluation in a block other than then one that uses it (e.g. to hoist
824 // the address of globals out of a loop). If this is the case, we don't
825 // want to forward-subst the cast.
826 if (isa<Constant>(CI->getOperand(0)))
831 Change = OptimizeNoopCopyExpression(CI, *TLI);
832 MadeChange |= Change;
835 if (!Change && (isa<ZExtInst>(I) || isa<SExtInst>(I)))
836 MadeChange |= OptimizeExtUses(I);
837 } else if (CmpInst *CI = dyn_cast<CmpInst>(I)) {
838 MadeChange |= OptimizeCmpExpression(CI);
839 } else if (LoadInst *LI = dyn_cast<LoadInst>(I)) {
841 MadeChange |= OptimizeMemoryInst(I, I->getOperand(0), LI->getType(),
843 } else if (StoreInst *SI = dyn_cast<StoreInst>(I)) {
845 MadeChange |= OptimizeMemoryInst(I, SI->getOperand(1),
846 SI->getOperand(0)->getType(),
848 } else if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(I)) {
849 if (GEPI->hasAllZeroIndices()) {
850 /// The GEP operand must be a pointer, so must its result -> BitCast
851 Instruction *NC = new BitCastInst(GEPI->getOperand(0), GEPI->getType(),
852 GEPI->getName(), GEPI);
853 GEPI->replaceAllUsesWith(NC);
854 GEPI->eraseFromParent();
858 } else if (CallInst *CI = dyn_cast<CallInst>(I)) {
859 // If we found an inline asm expession, and if the target knows how to
860 // lower it to normal LLVM code, do so now.
861 if (TLI && isa<InlineAsm>(CI->getCalledValue()))
862 if (const TargetAsmInfo *TAI =
863 TLI->getTargetMachine().getTargetAsmInfo()) {
864 if (TAI->ExpandInlineAsm(CI)) {
866 // Avoid processing instructions out of order, which could cause
867 // reuse before a value is defined.
870 // Sink address computing for memory operands into the block.
871 MadeChange |= OptimizeInlineAsmInst(I, &(*CI), SunkAddrs);