1 //===- CodeGenPrepare.cpp - Prepare a function for code generation --------===//
2 //
3 // The LLVM Compiler Infrastructure
4 //
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
7 //
8 //===----------------------------------------------------------------------===//
9 //
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.
13 //
14 //===----------------------------------------------------------------------===//
16 #include "llvm/CodeGen/Passes.h"
17 #include "llvm/ADT/DenseMap.h"
18 #include "llvm/ADT/SmallSet.h"
19 #include "llvm/ADT/Statistic.h"
20 #include "llvm/Analysis/InstructionSimplify.h"
21 #include "llvm/IR/CallSite.h"
22 #include "llvm/IR/Constants.h"
23 #include "llvm/IR/DataLayout.h"
24 #include "llvm/IR/DerivedTypes.h"
25 #include "llvm/IR/Dominators.h"
26 #include "llvm/IR/Function.h"
27 #include "llvm/IR/GetElementPtrTypeIterator.h"
28 #include "llvm/IR/IRBuilder.h"
29 #include "llvm/IR/InlineAsm.h"
30 #include "llvm/IR/Instructions.h"
31 #include "llvm/IR/IntrinsicInst.h"
32 #include "llvm/IR/PatternMatch.h"
33 #include "llvm/IR/ValueHandle.h"
34 #include "llvm/IR/ValueMap.h"
35 #include "llvm/Pass.h"
36 #include "llvm/Support/CommandLine.h"
37 #include "llvm/Support/Debug.h"
38 #include "llvm/Support/raw_ostream.h"
39 #include "llvm/Target/TargetLibraryInfo.h"
40 #include "llvm/Target/TargetLowering.h"
41 #include "llvm/Target/TargetSubtargetInfo.h"
42 #include "llvm/Transforms/Utils/BasicBlockUtils.h"
43 #include "llvm/Transforms/Utils/BuildLibCalls.h"
44 #include "llvm/Transforms/Utils/BypassSlowDivision.h"
45 #include "llvm/Transforms/Utils/Local.h"
46 using namespace llvm;
47 using namespace llvm::PatternMatch;
49 #define DEBUG_TYPE "codegenprepare"
51 STATISTIC(NumBlocksElim, "Number of blocks eliminated");
52 STATISTIC(NumPHIsElim, "Number of trivial PHIs eliminated");
53 STATISTIC(NumGEPsElim, "Number of GEPs converted to casts");
54 STATISTIC(NumCmpUses, "Number of uses of Cmp expressions replaced with uses of "
55 "sunken Cmps");
56 STATISTIC(NumCastUses, "Number of uses of Cast expressions replaced with uses "
57 "of sunken Casts");
58 STATISTIC(NumMemoryInsts, "Number of memory instructions whose address "
59 "computations were sunk");
60 STATISTIC(NumExtsMoved, "Number of [s|z]ext instructions combined with loads");
61 STATISTIC(NumExtUses, "Number of uses of [s|z]ext instructions optimized");
62 STATISTIC(NumRetsDup, "Number of return instructions duplicated");
63 STATISTIC(NumDbgValueMoved, "Number of debug value instructions moved");
64 STATISTIC(NumSelectsExpanded, "Number of selects turned into branches");
65 STATISTIC(NumAndCmpsMoved, "Number of and/cmp's pushed into branches");
67 static cl::opt<bool> DisableBranchOpts(
68 "disable-cgp-branch-opts", cl::Hidden, cl::init(false),
69 cl::desc("Disable branch optimizations in CodeGenPrepare"));
71 static cl::opt<bool> DisableSelectToBranch(
72 "disable-cgp-select2branch", cl::Hidden, cl::init(false),
73 cl::desc("Disable select to branch conversion."));
75 static cl::opt<bool> AddrSinkUsingGEPs(
76 "addr-sink-using-gep", cl::Hidden, cl::init(false),
77 cl::desc("Address sinking in CGP using GEPs."));
79 static cl::opt<bool> EnableAndCmpSinking(
80 "enable-andcmp-sinking", cl::Hidden, cl::init(true),
81 cl::desc("Enable sinkinig and/cmp into branches."));
83 namespace {
84 typedef SmallPtrSet<Instruction *, 16> SetOfInstrs;
85 typedef DenseMap<Instruction *, Type *> InstrToOrigTy;
87 class CodeGenPrepare : public FunctionPass {
88 /// TLI - Keep a pointer of a TargetLowering to consult for determining
89 /// transformation profitability.
90 const TargetMachine *TM;
91 const TargetLowering *TLI;
92 const TargetLibraryInfo *TLInfo;
93 DominatorTree *DT;
95 /// CurInstIterator - As we scan instructions optimizing them, this is the
96 /// next instruction to optimize. Xforms that can invalidate this should
97 /// update it.
98 BasicBlock::iterator CurInstIterator;
100 /// Keeps track of non-local addresses that have been sunk into a block.
101 /// This allows us to avoid inserting duplicate code for blocks with
102 /// multiple load/stores of the same address.
103 ValueMap<Value*, Value*> SunkAddrs;
105 /// Keeps track of all truncates inserted for the current function.
106 SetOfInstrs InsertedTruncsSet;
107 /// Keeps track of the type of the related instruction before their
108 /// promotion for the current function.
109 InstrToOrigTy PromotedInsts;
111 /// ModifiedDT - If CFG is modified in anyway, dominator tree may need to
112 /// be updated.
113 bool ModifiedDT;
115 /// OptSize - True if optimizing for size.
116 bool OptSize;
118 public:
119 static char ID; // Pass identification, replacement for typeid
120 explicit CodeGenPrepare(const TargetMachine *TM = nullptr)
121 : FunctionPass(ID), TM(TM), TLI(nullptr) {
122 initializeCodeGenPreparePass(*PassRegistry::getPassRegistry());
123 }
124 bool runOnFunction(Function &F) override;
126 const char *getPassName() const override { return "CodeGen Prepare"; }
128 void getAnalysisUsage(AnalysisUsage &AU) const override {
129 AU.addPreserved<DominatorTreeWrapperPass>();
130 AU.addRequired<TargetLibraryInfo>();
131 }
133 private:
134 bool EliminateFallThrough(Function &F);
135 bool EliminateMostlyEmptyBlocks(Function &F);
136 bool CanMergeBlocks(const BasicBlock *BB, const BasicBlock *DestBB) const;
137 void EliminateMostlyEmptyBlock(BasicBlock *BB);
138 bool OptimizeBlock(BasicBlock &BB);
139 bool OptimizeInst(Instruction *I);
140 bool OptimizeMemoryInst(Instruction *I, Value *Addr, Type *AccessTy);
141 bool OptimizeInlineAsmInst(CallInst *CS);
142 bool OptimizeCallInst(CallInst *CI);
143 bool MoveExtToFormExtLoad(Instruction *I);
144 bool OptimizeExtUses(Instruction *I);
145 bool OptimizeSelectInst(SelectInst *SI);
146 bool OptimizeShuffleVectorInst(ShuffleVectorInst *SI);
147 bool DupRetToEnableTailCallOpts(BasicBlock *BB);
148 bool PlaceDbgValues(Function &F);
149 bool sinkAndCmp(Function &F);
150 };
151 }
153 char CodeGenPrepare::ID = 0;
154 INITIALIZE_TM_PASS(CodeGenPrepare, "codegenprepare",
155 "Optimize for code generation", false, false)
157 FunctionPass *llvm::createCodeGenPreparePass(const TargetMachine *TM) {
158 return new CodeGenPrepare(TM);
159 }
161 bool CodeGenPrepare::runOnFunction(Function &F) {
162 if (skipOptnoneFunction(F))
163 return false;
165 bool EverMadeChange = false;
166 // Clear per function information.
167 InsertedTruncsSet.clear();
168 PromotedInsts.clear();
170 ModifiedDT = false;
171 if (TM)
172 TLI = TM->getSubtargetImpl()->getTargetLowering();
173 TLInfo = &getAnalysis<TargetLibraryInfo>();
174 DominatorTreeWrapperPass *DTWP =
175 getAnalysisIfAvailable<DominatorTreeWrapperPass>();
176 DT = DTWP ? &DTWP->getDomTree() : nullptr;
177 OptSize = F.getAttributes().hasAttribute(AttributeSet::FunctionIndex,
178 Attribute::OptimizeForSize);
180 /// This optimization identifies DIV instructions that can be
181 /// profitably bypassed and carried out with a shorter, faster divide.
182 if (!OptSize && TLI && TLI->isSlowDivBypassed()) {
183 const DenseMap<unsigned int, unsigned int> &BypassWidths =
184 TLI->getBypassSlowDivWidths();
185 for (Function::iterator I = F.begin(); I != F.end(); I++)
186 EverMadeChange |= bypassSlowDivision(F, I, BypassWidths);
187 }
189 // Eliminate blocks that contain only PHI nodes and an
190 // unconditional branch.
191 EverMadeChange |= EliminateMostlyEmptyBlocks(F);
193 // llvm.dbg.value is far away from the value then iSel may not be able
194 // handle it properly. iSel will drop llvm.dbg.value if it can not
195 // find a node corresponding to the value.
196 EverMadeChange |= PlaceDbgValues(F);
198 // If there is a mask, compare against zero, and branch that can be combined
199 // into a single target instruction, push the mask and compare into branch
200 // users. Do this before OptimizeBlock -> OptimizeInst ->
201 // OptimizeCmpExpression, which perturbs the pattern being searched for.
202 if (!DisableBranchOpts)
203 EverMadeChange |= sinkAndCmp(F);
205 bool MadeChange = true;
206 while (MadeChange) {
207 MadeChange = false;
208 for (Function::iterator I = F.begin(); I != F.end(); ) {
209 BasicBlock *BB = I++;
210 MadeChange |= OptimizeBlock(*BB);
211 }
212 EverMadeChange |= MadeChange;
213 }
215 SunkAddrs.clear();
217 if (!DisableBranchOpts) {
218 MadeChange = false;
219 SmallPtrSet<BasicBlock*, 8> WorkList;
220 for (Function::iterator BB = F.begin(), E = F.end(); BB != E; ++BB) {
221 SmallVector<BasicBlock*, 2> Successors(succ_begin(BB), succ_end(BB));
222 MadeChange |= ConstantFoldTerminator(BB, true);
223 if (!MadeChange) continue;
225 for (SmallVectorImpl<BasicBlock*>::iterator
226 II = Successors.begin(), IE = Successors.end(); II != IE; ++II)
227 if (pred_begin(*II) == pred_end(*II))
228 WorkList.insert(*II);
229 }
231 // Delete the dead blocks and any of their dead successors.
232 MadeChange |= !WorkList.empty();
233 while (!WorkList.empty()) {
234 BasicBlock *BB = *WorkList.begin();
235 WorkList.erase(BB);
236 SmallVector<BasicBlock*, 2> Successors(succ_begin(BB), succ_end(BB));
238 DeleteDeadBlock(BB);
240 for (SmallVectorImpl<BasicBlock*>::iterator
241 II = Successors.begin(), IE = Successors.end(); II != IE; ++II)
242 if (pred_begin(*II) == pred_end(*II))
243 WorkList.insert(*II);
244 }
246 // Merge pairs of basic blocks with unconditional branches, connected by
247 // a single edge.
248 if (EverMadeChange || MadeChange)
249 MadeChange |= EliminateFallThrough(F);
251 if (MadeChange)
252 ModifiedDT = true;
253 EverMadeChange |= MadeChange;
254 }
256 if (ModifiedDT && DT)
257 DT->recalculate(F);
259 return EverMadeChange;
260 }
262 /// EliminateFallThrough - Merge basic blocks which are connected
263 /// by a single edge, where one of the basic blocks has a single successor
264 /// pointing to the other basic block, which has a single predecessor.
265 bool CodeGenPrepare::EliminateFallThrough(Function &F) {
266 bool Changed = false;
267 // Scan all of the blocks in the function, except for the entry block.
268 for (Function::iterator I = std::next(F.begin()), E = F.end(); I != E;) {
269 BasicBlock *BB = I++;
270 // If the destination block has a single pred, then this is a trivial
271 // edge, just collapse it.
272 BasicBlock *SinglePred = BB->getSinglePredecessor();
274 // Don't merge if BB's address is taken.
275 if (!SinglePred || SinglePred == BB || BB->hasAddressTaken()) continue;
277 BranchInst *Term = dyn_cast<BranchInst>(SinglePred->getTerminator());
278 if (Term && !Term->isConditional()) {
279 Changed = true;
280 DEBUG(dbgs() << "To merge:\n"<< *SinglePred << "\n\n\n");
281 // Remember if SinglePred was the entry block of the function.
282 // If so, we will need to move BB back to the entry position.
283 bool isEntry = SinglePred == &SinglePred->getParent()->getEntryBlock();
284 MergeBasicBlockIntoOnlyPred(BB, this);
286 if (isEntry && BB != &BB->getParent()->getEntryBlock())
287 BB->moveBefore(&BB->getParent()->getEntryBlock());
289 // We have erased a block. Update the iterator.
290 I = BB;
291 }
292 }
293 return Changed;
294 }
296 /// EliminateMostlyEmptyBlocks - eliminate blocks that contain only PHI nodes,
297 /// debug info directives, and an unconditional branch. Passes before isel
298 /// (e.g. LSR/loopsimplify) often split edges in ways that are non-optimal for
299 /// isel. Start by eliminating these blocks so we can split them the way we
300 /// want them.
301 bool CodeGenPrepare::EliminateMostlyEmptyBlocks(Function &F) {
302 bool MadeChange = false;
303 // Note that this intentionally skips the entry block.
304 for (Function::iterator I = std::next(F.begin()), E = F.end(); I != E;) {
305 BasicBlock *BB = I++;
307 // If this block doesn't end with an uncond branch, ignore it.
308 BranchInst *BI = dyn_cast<BranchInst>(BB->getTerminator());
309 if (!BI || !BI->isUnconditional())
310 continue;
312 // If the instruction before the branch (skipping debug info) isn't a phi
313 // node, then other stuff is happening here.
314 BasicBlock::iterator BBI = BI;
315 if (BBI != BB->begin()) {
316 --BBI;
317 while (isa<DbgInfoIntrinsic>(BBI)) {
318 if (BBI == BB->begin())
319 break;
320 --BBI;
321 }
322 if (!isa<DbgInfoIntrinsic>(BBI) && !isa<PHINode>(BBI))
323 continue;
324 }
326 // Do not break infinite loops.
327 BasicBlock *DestBB = BI->getSuccessor(0);
328 if (DestBB == BB)
329 continue;
331 if (!CanMergeBlocks(BB, DestBB))
332 continue;
334 EliminateMostlyEmptyBlock(BB);
335 MadeChange = true;
336 }
337 return MadeChange;
338 }
340 /// CanMergeBlocks - Return true if we can merge BB into DestBB if there is a
341 /// single uncond branch between them, and BB contains no other non-phi
342 /// instructions.
343 bool CodeGenPrepare::CanMergeBlocks(const BasicBlock *BB,
344 const BasicBlock *DestBB) const {
345 // We only want to eliminate blocks whose phi nodes are used by phi nodes in
346 // the successor. If there are more complex condition (e.g. preheaders),
347 // don't mess around with them.
348 BasicBlock::const_iterator BBI = BB->begin();
349 while (const PHINode *PN = dyn_cast<PHINode>(BBI++)) {
350 for (const User *U : PN->users()) {
351 const Instruction *UI = cast<Instruction>(U);
352 if (UI->getParent() != DestBB || !isa<PHINode>(UI))
353 return false;
354 // If User is inside DestBB block and it is a PHINode then check
355 // incoming value. If incoming value is not from BB then this is
356 // a complex condition (e.g. preheaders) we want to avoid here.
357 if (UI->getParent() == DestBB) {
358 if (const PHINode *UPN = dyn_cast<PHINode>(UI))
359 for (unsigned I = 0, E = UPN->getNumIncomingValues(); I != E; ++I) {
360 Instruction *Insn = dyn_cast<Instruction>(UPN->getIncomingValue(I));
361 if (Insn && Insn->getParent() == BB &&
362 Insn->getParent() != UPN->getIncomingBlock(I))
363 return false;
364 }
365 }
366 }
367 }
369 // If BB and DestBB contain any common predecessors, then the phi nodes in BB
370 // and DestBB may have conflicting incoming values for the block. If so, we
371 // can't merge the block.
372 const PHINode *DestBBPN = dyn_cast<PHINode>(DestBB->begin());
373 if (!DestBBPN) return true; // no conflict.
375 // Collect the preds of BB.
376 SmallPtrSet<const BasicBlock*, 16> BBPreds;
377 if (const PHINode *BBPN = dyn_cast<PHINode>(BB->begin())) {
378 // It is faster to get preds from a PHI than with pred_iterator.
379 for (unsigned i = 0, e = BBPN->getNumIncomingValues(); i != e; ++i)
380 BBPreds.insert(BBPN->getIncomingBlock(i));
381 } else {
382 BBPreds.insert(pred_begin(BB), pred_end(BB));
383 }
385 // Walk the preds of DestBB.
386 for (unsigned i = 0, e = DestBBPN->getNumIncomingValues(); i != e; ++i) {
387 BasicBlock *Pred = DestBBPN->getIncomingBlock(i);
388 if (BBPreds.count(Pred)) { // Common predecessor?
389 BBI = DestBB->begin();
390 while (const PHINode *PN = dyn_cast<PHINode>(BBI++)) {
391 const Value *V1 = PN->getIncomingValueForBlock(Pred);
392 const Value *V2 = PN->getIncomingValueForBlock(BB);
394 // If V2 is a phi node in BB, look up what the mapped value will be.
395 if (const PHINode *V2PN = dyn_cast<PHINode>(V2))
396 if (V2PN->getParent() == BB)
397 V2 = V2PN->getIncomingValueForBlock(Pred);
399 // If there is a conflict, bail out.
400 if (V1 != V2) return false;
401 }
402 }
403 }
405 return true;
406 }
409 /// EliminateMostlyEmptyBlock - Eliminate a basic block that have only phi's and
410 /// an unconditional branch in it.
411 void CodeGenPrepare::EliminateMostlyEmptyBlock(BasicBlock *BB) {
412 BranchInst *BI = cast<BranchInst>(BB->getTerminator());
413 BasicBlock *DestBB = BI->getSuccessor(0);
415 DEBUG(dbgs() << "MERGING MOSTLY EMPTY BLOCKS - BEFORE:\n" << *BB << *DestBB);
417 // If the destination block has a single pred, then this is a trivial edge,
418 // just collapse it.
419 if (BasicBlock *SinglePred = DestBB->getSinglePredecessor()) {
420 if (SinglePred != DestBB) {
421 // Remember if SinglePred was the entry block of the function. If so, we
422 // will need to move BB back to the entry position.
423 bool isEntry = SinglePred == &SinglePred->getParent()->getEntryBlock();
424 MergeBasicBlockIntoOnlyPred(DestBB, this);
426 if (isEntry && BB != &BB->getParent()->getEntryBlock())
427 BB->moveBefore(&BB->getParent()->getEntryBlock());
429 DEBUG(dbgs() << "AFTER:\n" << *DestBB << "\n\n\n");
430 return;
431 }
432 }
434 // Otherwise, we have multiple predecessors of BB. Update the PHIs in DestBB
435 // to handle the new incoming edges it is about to have.
436 PHINode *PN;
437 for (BasicBlock::iterator BBI = DestBB->begin();
438 (PN = dyn_cast<PHINode>(BBI)); ++BBI) {
439 // Remove the incoming value for BB, and remember it.
440 Value *InVal = PN->removeIncomingValue(BB, false);
442 // Two options: either the InVal is a phi node defined in BB or it is some
443 // value that dominates BB.
444 PHINode *InValPhi = dyn_cast<PHINode>(InVal);
445 if (InValPhi && InValPhi->getParent() == BB) {
446 // Add all of the input values of the input PHI as inputs of this phi.
447 for (unsigned i = 0, e = InValPhi->getNumIncomingValues(); i != e; ++i)
448 PN->addIncoming(InValPhi->getIncomingValue(i),
449 InValPhi->getIncomingBlock(i));
450 } else {
451 // Otherwise, add one instance of the dominating value for each edge that
452 // we will be adding.
453 if (PHINode *BBPN = dyn_cast<PHINode>(BB->begin())) {
454 for (unsigned i = 0, e = BBPN->getNumIncomingValues(); i != e; ++i)
455 PN->addIncoming(InVal, BBPN->getIncomingBlock(i));
456 } else {
457 for (pred_iterator PI = pred_begin(BB), E = pred_end(BB); PI != E; ++PI)
458 PN->addIncoming(InVal, *PI);
459 }
460 }
461 }
463 // The PHIs are now updated, change everything that refers to BB to use
464 // DestBB and remove BB.
465 BB->replaceAllUsesWith(DestBB);
466 if (DT && !ModifiedDT) {
467 BasicBlock *BBIDom = DT->getNode(BB)->getIDom()->getBlock();
468 BasicBlock *DestBBIDom = DT->getNode(DestBB)->getIDom()->getBlock();
469 BasicBlock *NewIDom = DT->findNearestCommonDominator(BBIDom, DestBBIDom);
470 DT->changeImmediateDominator(DestBB, NewIDom);
471 DT->eraseNode(BB);
472 }
473 BB->eraseFromParent();
474 ++NumBlocksElim;
476 DEBUG(dbgs() << "AFTER:\n" << *DestBB << "\n\n\n");
477 }
479 /// SinkCast - Sink the specified cast instruction into its user blocks
480 static bool SinkCast(CastInst *CI) {
481 BasicBlock *DefBB = CI->getParent();
483 /// InsertedCasts - Only insert a cast in each block once.
484 DenseMap<BasicBlock*, CastInst*> InsertedCasts;
486 bool MadeChange = false;
487 for (Value::user_iterator UI = CI->user_begin(), E = CI->user_end();
488 UI != E; ) {
489 Use &TheUse = UI.getUse();
490 Instruction *User = cast<Instruction>(*UI);
492 // Figure out which BB this cast is used in. For PHI's this is the
493 // appropriate predecessor block.
494 BasicBlock *UserBB = User->getParent();
495 if (PHINode *PN = dyn_cast<PHINode>(User)) {
496 UserBB = PN->getIncomingBlock(TheUse);
497 }
499 // Preincrement use iterator so we don't invalidate it.
500 ++UI;
502 // If this user is in the same block as the cast, don't change the cast.
503 if (UserBB == DefBB) continue;
505 // If we have already inserted a cast into this block, use it.
506 CastInst *&InsertedCast = InsertedCasts[UserBB];
508 if (!InsertedCast) {
509 BasicBlock::iterator InsertPt = UserBB->getFirstInsertionPt();
510 InsertedCast =
511 CastInst::Create(CI->getOpcode(), CI->getOperand(0), CI->getType(), "",
512 InsertPt);
513 MadeChange = true;
514 }
516 // Replace a use of the cast with a use of the new cast.
517 TheUse = InsertedCast;
518 ++NumCastUses;
519 }
521 // If we removed all uses, nuke the cast.
522 if (CI->use_empty()) {
523 CI->eraseFromParent();
524 MadeChange = true;
525 }
527 return MadeChange;
528 }
530 /// OptimizeNoopCopyExpression - If the specified cast instruction is a noop
531 /// copy (e.g. it's casting from one pointer type to another, i32->i8 on PPC),
532 /// sink it into user blocks to reduce the number of virtual
533 /// registers that must be created and coalesced.
534 ///
535 /// Return true if any changes are made.
536 ///
537 static bool OptimizeNoopCopyExpression(CastInst *CI, const TargetLowering &TLI){
538 // If this is a noop copy,
539 EVT SrcVT = TLI.getValueType(CI->getOperand(0)->getType());
540 EVT DstVT = TLI.getValueType(CI->getType());
542 // This is an fp<->int conversion?
543 if (SrcVT.isInteger() != DstVT.isInteger())
544 return false;
546 // If this is an extension, it will be a zero or sign extension, which
547 // isn't a noop.
548 if (SrcVT.bitsLT(DstVT)) return false;
550 // If these values will be promoted, find out what they will be promoted
551 // to. This helps us consider truncates on PPC as noop copies when they
552 // are.
553 if (TLI.getTypeAction(CI->getContext(), SrcVT) ==
554 TargetLowering::TypePromoteInteger)
555 SrcVT = TLI.getTypeToTransformTo(CI->getContext(), SrcVT);
556 if (TLI.getTypeAction(CI->getContext(), DstVT) ==
557 TargetLowering::TypePromoteInteger)
558 DstVT = TLI.getTypeToTransformTo(CI->getContext(), DstVT);
560 // If, after promotion, these are the same types, this is a noop copy.
561 if (SrcVT != DstVT)
562 return false;
564 return SinkCast(CI);
565 }
567 /// OptimizeCmpExpression - sink the given CmpInst into user blocks to reduce
568 /// the number of virtual registers that must be created and coalesced. This is
569 /// a clear win except on targets with multiple condition code registers
570 /// (PowerPC), where it might lose; some adjustment may be wanted there.
571 ///
572 /// Return true if any changes are made.
573 static bool OptimizeCmpExpression(CmpInst *CI) {
574 BasicBlock *DefBB = CI->getParent();
576 /// InsertedCmp - Only insert a cmp in each block once.
577 DenseMap<BasicBlock*, CmpInst*> InsertedCmps;
579 bool MadeChange = false;
580 for (Value::user_iterator UI = CI->user_begin(), E = CI->user_end();
581 UI != E; ) {
582 Use &TheUse = UI.getUse();
583 Instruction *User = cast<Instruction>(*UI);
585 // Preincrement use iterator so we don't invalidate it.
586 ++UI;
588 // Don't bother for PHI nodes.
589 if (isa<PHINode>(User))
590 continue;
592 // Figure out which BB this cmp is used in.
593 BasicBlock *UserBB = User->getParent();
595 // If this user is in the same block as the cmp, don't change the cmp.
596 if (UserBB == DefBB) continue;
598 // If we have already inserted a cmp into this block, use it.
599 CmpInst *&InsertedCmp = InsertedCmps[UserBB];
601 if (!InsertedCmp) {
602 BasicBlock::iterator InsertPt = UserBB->getFirstInsertionPt();
603 InsertedCmp =
604 CmpInst::Create(CI->getOpcode(),
605 CI->getPredicate(), CI->getOperand(0),
606 CI->getOperand(1), "", InsertPt);
607 MadeChange = true;
608 }
610 // Replace a use of the cmp with a use of the new cmp.
611 TheUse = InsertedCmp;
612 ++NumCmpUses;
613 }
615 // If we removed all uses, nuke the cmp.
616 if (CI->use_empty())
617 CI->eraseFromParent();
619 return MadeChange;
620 }
622 /// isExtractBitsCandidateUse - Check if the candidates could
623 /// be combined with shift instruction, which includes:
624 /// 1. Truncate instruction
625 /// 2. And instruction and the imm is a mask of the low bits:
626 /// imm & (imm+1) == 0
627 static bool isExtractBitsCandidateUse(Instruction *User) {
628 if (!isa<TruncInst>(User)) {
629 if (User->getOpcode() != Instruction::And ||
630 !isa<ConstantInt>(User->getOperand(1)))
631 return false;
633 const APInt &Cimm = cast<ConstantInt>(User->getOperand(1))->getValue();
635 if ((Cimm & (Cimm + 1)).getBoolValue())
636 return false;
637 }
638 return true;
639 }
641 /// SinkShiftAndTruncate - sink both shift and truncate instruction
642 /// to the use of truncate's BB.
643 static bool
644 SinkShiftAndTruncate(BinaryOperator *ShiftI, Instruction *User, ConstantInt *CI,
645 DenseMap<BasicBlock *, BinaryOperator *> &InsertedShifts,
646 const TargetLowering &TLI) {
647 BasicBlock *UserBB = User->getParent();
648 DenseMap<BasicBlock *, CastInst *> InsertedTruncs;
649 TruncInst *TruncI = dyn_cast<TruncInst>(User);
650 bool MadeChange = false;
652 for (Value::user_iterator TruncUI = TruncI->user_begin(),
653 TruncE = TruncI->user_end();
654 TruncUI != TruncE;) {
656 Use &TruncTheUse = TruncUI.getUse();
657 Instruction *TruncUser = cast<Instruction>(*TruncUI);
658 // Preincrement use iterator so we don't invalidate it.
660 ++TruncUI;
662 int ISDOpcode = TLI.InstructionOpcodeToISD(TruncUser->getOpcode());
663 if (!ISDOpcode)
664 continue;
666 // If the use is actually a legal node, there will not be an
667 // implicit truncate.
668 // FIXME: always querying the result type is just an
669 // approximation; some nodes' legality is determined by the
670 // operand or other means. There's no good way to find out though.
671 if (TLI.isOperationLegalOrCustom(ISDOpcode,
672 EVT::getEVT(TruncUser->getType(), true)))
673 continue;
675 // Don't bother for PHI nodes.
676 if (isa<PHINode>(TruncUser))
677 continue;
679 BasicBlock *TruncUserBB = TruncUser->getParent();
681 if (UserBB == TruncUserBB)
682 continue;
684 BinaryOperator *&InsertedShift = InsertedShifts[TruncUserBB];
685 CastInst *&InsertedTrunc = InsertedTruncs[TruncUserBB];
687 if (!InsertedShift && !InsertedTrunc) {
688 BasicBlock::iterator InsertPt = TruncUserBB->getFirstInsertionPt();
689 // Sink the shift
690 if (ShiftI->getOpcode() == Instruction::AShr)
691 InsertedShift =
692 BinaryOperator::CreateAShr(ShiftI->getOperand(0), CI, "", InsertPt);
693 else
694 InsertedShift =
695 BinaryOperator::CreateLShr(ShiftI->getOperand(0), CI, "", InsertPt);
697 // Sink the trunc
698 BasicBlock::iterator TruncInsertPt = TruncUserBB->getFirstInsertionPt();
699 TruncInsertPt++;
701 InsertedTrunc = CastInst::Create(TruncI->getOpcode(), InsertedShift,
702 TruncI->getType(), "", TruncInsertPt);
704 MadeChange = true;
706 TruncTheUse = InsertedTrunc;
707 }
708 }
709 return MadeChange;
710 }
712 /// OptimizeExtractBits - sink the shift *right* instruction into user blocks if
713 /// the uses could potentially be combined with this shift instruction and
714 /// generate BitExtract instruction. It will only be applied if the architecture
715 /// supports BitExtract instruction. Here is an example:
716 /// BB1:
717 /// %x.extract.shift = lshr i64 %arg1, 32
718 /// BB2:
719 /// %x.extract.trunc = trunc i64 %x.extract.shift to i16
720 /// ==>
721 ///
722 /// BB2:
723 /// %x.extract.shift.1 = lshr i64 %arg1, 32
724 /// %x.extract.trunc = trunc i64 %x.extract.shift.1 to i16
725 ///
726 /// CodeGen will recoginze the pattern in BB2 and generate BitExtract
727 /// instruction.
728 /// Return true if any changes are made.
729 static bool OptimizeExtractBits(BinaryOperator *ShiftI, ConstantInt *CI,
730 const TargetLowering &TLI) {
731 BasicBlock *DefBB = ShiftI->getParent();
733 /// Only insert instructions in each block once.
734 DenseMap<BasicBlock *, BinaryOperator *> InsertedShifts;
736 bool shiftIsLegal = TLI.isTypeLegal(TLI.getValueType(ShiftI->getType()));
738 bool MadeChange = false;
739 for (Value::user_iterator UI = ShiftI->user_begin(), E = ShiftI->user_end();
740 UI != E;) {
741 Use &TheUse = UI.getUse();
742 Instruction *User = cast<Instruction>(*UI);
743 // Preincrement use iterator so we don't invalidate it.
744 ++UI;
746 // Don't bother for PHI nodes.
747 if (isa<PHINode>(User))
748 continue;
750 if (!isExtractBitsCandidateUse(User))
751 continue;
753 BasicBlock *UserBB = User->getParent();
755 if (UserBB == DefBB) {
756 // If the shift and truncate instruction are in the same BB. The use of
757 // the truncate(TruncUse) may still introduce another truncate if not
758 // legal. In this case, we would like to sink both shift and truncate
759 // instruction to the BB of TruncUse.
760 // for example:
761 // BB1:
762 // i64 shift.result = lshr i64 opnd, imm
763 // trunc.result = trunc shift.result to i16
764 //
765 // BB2:
766 // ----> We will have an implicit truncate here if the architecture does
767 // not have i16 compare.
768 // cmp i16 trunc.result, opnd2
769 //
770 if (isa<TruncInst>(User) && shiftIsLegal
771 // If the type of the truncate is legal, no trucate will be
772 // introduced in other basic blocks.
773 && (!TLI.isTypeLegal(TLI.getValueType(User->getType()))))
774 MadeChange =
775 SinkShiftAndTruncate(ShiftI, User, CI, InsertedShifts, TLI);
777 continue;
778 }
779 // If we have already inserted a shift into this block, use it.
780 BinaryOperator *&InsertedShift = InsertedShifts[UserBB];
782 if (!InsertedShift) {
783 BasicBlock::iterator InsertPt = UserBB->getFirstInsertionPt();
785 if (ShiftI->getOpcode() == Instruction::AShr)
786 InsertedShift =
787 BinaryOperator::CreateAShr(ShiftI->getOperand(0), CI, "", InsertPt);
788 else
789 InsertedShift =
790 BinaryOperator::CreateLShr(ShiftI->getOperand(0), CI, "", InsertPt);
792 MadeChange = true;
793 }
795 // Replace a use of the shift with a use of the new shift.
796 TheUse = InsertedShift;
797 }
799 // If we removed all uses, nuke the shift.
800 if (ShiftI->use_empty())
801 ShiftI->eraseFromParent();
803 return MadeChange;
804 }
806 namespace {
807 class CodeGenPrepareFortifiedLibCalls : public SimplifyFortifiedLibCalls {
808 protected:
809 void replaceCall(Value *With) override {
810 CI->replaceAllUsesWith(With);
811 CI->eraseFromParent();
812 }
813 bool isFoldable(unsigned SizeCIOp, unsigned, bool) const override {
814 if (ConstantInt *SizeCI =
815 dyn_cast<ConstantInt>(CI->getArgOperand(SizeCIOp)))
816 return SizeCI->isAllOnesValue();
817 return false;
818 }
819 };
820 } // end anonymous namespace
822 bool CodeGenPrepare::OptimizeCallInst(CallInst *CI) {
823 BasicBlock *BB = CI->getParent();
825 // Lower inline assembly if we can.
826 // If we found an inline asm expession, and if the target knows how to
827 // lower it to normal LLVM code, do so now.
828 if (TLI && isa<InlineAsm>(CI->getCalledValue())) {
829 if (TLI->ExpandInlineAsm(CI)) {
830 // Avoid invalidating the iterator.
831 CurInstIterator = BB->begin();
832 // Avoid processing instructions out of order, which could cause
833 // reuse before a value is defined.
834 SunkAddrs.clear();
835 return true;
836 }
837 // Sink address computing for memory operands into the block.
838 if (OptimizeInlineAsmInst(CI))
839 return true;
840 }
842 // Lower all uses of llvm.objectsize.*
843 IntrinsicInst *II = dyn_cast<IntrinsicInst>(CI);
844 if (II && II->getIntrinsicID() == Intrinsic::objectsize) {
845 bool Min = (cast<ConstantInt>(II->getArgOperand(1))->getZExtValue() == 1);
846 Type *ReturnTy = CI->getType();
847 Constant *RetVal = ConstantInt::get(ReturnTy, Min ? 0 : -1ULL);
849 // Substituting this can cause recursive simplifications, which can
850 // invalidate our iterator. Use a WeakVH to hold onto it in case this
851 // happens.
852 WeakVH IterHandle(CurInstIterator);
854 replaceAndRecursivelySimplify(CI, RetVal,
855 TLI ? TLI->getDataLayout() : nullptr,
856 TLInfo, ModifiedDT ? nullptr : DT);
858 // If the iterator instruction was recursively deleted, start over at the
859 // start of the block.
860 if (IterHandle != CurInstIterator) {
861 CurInstIterator = BB->begin();
862 SunkAddrs.clear();
863 }
864 return true;
865 }
867 if (II && TLI) {
868 SmallVector<Value*, 2> PtrOps;
869 Type *AccessTy;
870 if (TLI->GetAddrModeArguments(II, PtrOps, AccessTy))
871 while (!PtrOps.empty())
872 if (OptimizeMemoryInst(II, PtrOps.pop_back_val(), AccessTy))
873 return true;
874 }
876 // From here on out we're working with named functions.
877 if (!CI->getCalledFunction()) return false;
879 // We'll need DataLayout from here on out.
880 const DataLayout *TD = TLI ? TLI->getDataLayout() : nullptr;
881 if (!TD) return false;
883 // Lower all default uses of _chk calls. This is very similar
884 // to what InstCombineCalls does, but here we are only lowering calls
885 // that have the default "don't know" as the objectsize. Anything else
886 // should be left alone.
887 CodeGenPrepareFortifiedLibCalls Simplifier;
888 return Simplifier.fold(CI, TD, TLInfo);
889 }
891 /// DupRetToEnableTailCallOpts - Look for opportunities to duplicate return
892 /// instructions to the predecessor to enable tail call optimizations. The
893 /// case it is currently looking for is:
894 /// @code
895 /// bb0:
896 /// %tmp0 = tail call i32 @f0()
897 /// br label %return
898 /// bb1:
899 /// %tmp1 = tail call i32 @f1()
900 /// br label %return
901 /// bb2:
902 /// %tmp2 = tail call i32 @f2()
903 /// br label %return
904 /// return:
905 /// %retval = phi i32 [ %tmp0, %bb0 ], [ %tmp1, %bb1 ], [ %tmp2, %bb2 ]
906 /// ret i32 %retval
907 /// @endcode
908 ///
909 /// =>
910 ///
911 /// @code
912 /// bb0:
913 /// %tmp0 = tail call i32 @f0()
914 /// ret i32 %tmp0
915 /// bb1:
916 /// %tmp1 = tail call i32 @f1()
917 /// ret i32 %tmp1
918 /// bb2:
919 /// %tmp2 = tail call i32 @f2()
920 /// ret i32 %tmp2
921 /// @endcode
922 bool CodeGenPrepare::DupRetToEnableTailCallOpts(BasicBlock *BB) {
923 if (!TLI)
924 return false;
926 ReturnInst *RI = dyn_cast<ReturnInst>(BB->getTerminator());
927 if (!RI)
928 return false;
930 PHINode *PN = nullptr;
931 BitCastInst *BCI = nullptr;
932 Value *V = RI->getReturnValue();
933 if (V) {
934 BCI = dyn_cast<BitCastInst>(V);
935 if (BCI)
936 V = BCI->getOperand(0);
938 PN = dyn_cast<PHINode>(V);
939 if (!PN)
940 return false;
941 }
943 if (PN && PN->getParent() != BB)
944 return false;
946 // It's not safe to eliminate the sign / zero extension of the return value.
947 // See llvm::isInTailCallPosition().
948 const Function *F = BB->getParent();
949 AttributeSet CallerAttrs = F->getAttributes();
950 if (CallerAttrs.hasAttribute(AttributeSet::ReturnIndex, Attribute::ZExt) ||
951 CallerAttrs.hasAttribute(AttributeSet::ReturnIndex, Attribute::SExt))
952 return false;
954 // Make sure there are no instructions between the PHI and return, or that the
955 // return is the first instruction in the block.
956 if (PN) {
957 BasicBlock::iterator BI = BB->begin();
958 do { ++BI; } while (isa<DbgInfoIntrinsic>(BI));
959 if (&*BI == BCI)
960 // Also skip over the bitcast.
961 ++BI;
962 if (&*BI != RI)
963 return false;
964 } else {
965 BasicBlock::iterator BI = BB->begin();
966 while (isa<DbgInfoIntrinsic>(BI)) ++BI;
967 if (&*BI != RI)
968 return false;
969 }
971 /// Only dup the ReturnInst if the CallInst is likely to be emitted as a tail
972 /// call.
973 SmallVector<CallInst*, 4> TailCalls;
974 if (PN) {
975 for (unsigned I = 0, E = PN->getNumIncomingValues(); I != E; ++I) {
976 CallInst *CI = dyn_cast<CallInst>(PN->getIncomingValue(I));
977 // Make sure the phi value is indeed produced by the tail call.
978 if (CI && CI->hasOneUse() && CI->getParent() == PN->getIncomingBlock(I) &&
979 TLI->mayBeEmittedAsTailCall(CI))
980 TailCalls.push_back(CI);
981 }
982 } else {
983 SmallPtrSet<BasicBlock*, 4> VisitedBBs;
984 for (pred_iterator PI = pred_begin(BB), PE = pred_end(BB); PI != PE; ++PI) {
985 if (!VisitedBBs.insert(*PI))
986 continue;
988 BasicBlock::InstListType &InstList = (*PI)->getInstList();
989 BasicBlock::InstListType::reverse_iterator RI = InstList.rbegin();
990 BasicBlock::InstListType::reverse_iterator RE = InstList.rend();
991 do { ++RI; } while (RI != RE && isa<DbgInfoIntrinsic>(&*RI));
992 if (RI == RE)
993 continue;
995 CallInst *CI = dyn_cast<CallInst>(&*RI);
996 if (CI && CI->use_empty() && TLI->mayBeEmittedAsTailCall(CI))
997 TailCalls.push_back(CI);
998 }
999 }
1001 bool Changed = false;
1002 for (unsigned i = 0, e = TailCalls.size(); i != e; ++i) {
1003 CallInst *CI = TailCalls[i];
1004 CallSite CS(CI);
1006 // Conservatively require the attributes of the call to match those of the
1007 // return. Ignore noalias because it doesn't affect the call sequence.
1008 AttributeSet CalleeAttrs = CS.getAttributes();
1009 if (AttrBuilder(CalleeAttrs, AttributeSet::ReturnIndex).
1010 removeAttribute(Attribute::NoAlias) !=
1011 AttrBuilder(CalleeAttrs, AttributeSet::ReturnIndex).
1012 removeAttribute(Attribute::NoAlias))
1013 continue;
1015 // Make sure the call instruction is followed by an unconditional branch to
1016 // the return block.
1017 BasicBlock *CallBB = CI->getParent();
1018 BranchInst *BI = dyn_cast<BranchInst>(CallBB->getTerminator());
1019 if (!BI || !BI->isUnconditional() || BI->getSuccessor(0) != BB)
1020 continue;
1022 // Duplicate the return into CallBB.
1023 (void)FoldReturnIntoUncondBranch(RI, BB, CallBB);
1024 ModifiedDT = Changed = true;
1025 ++NumRetsDup;
1026 }
1028 // If we eliminated all predecessors of the block, delete the block now.
1029 if (Changed && !BB->hasAddressTaken() && pred_begin(BB) == pred_end(BB))
1030 BB->eraseFromParent();
1032 return Changed;
1033 }
1035 //===----------------------------------------------------------------------===//
1036 // Memory Optimization
1037 //===----------------------------------------------------------------------===//
1039 namespace {
1041 /// ExtAddrMode - This is an extended version of TargetLowering::AddrMode
1042 /// which holds actual Value*'s for register values.
1043 struct ExtAddrMode : public TargetLowering::AddrMode {
1044 Value *BaseReg;
1045 Value *ScaledReg;
1046 ExtAddrMode() : BaseReg(nullptr), ScaledReg(nullptr) {}
1047 void print(raw_ostream &OS) const;
1048 void dump() const;
1050 bool operator==(const ExtAddrMode& O) const {
1051 return (BaseReg == O.BaseReg) && (ScaledReg == O.ScaledReg) &&
1052 (BaseGV == O.BaseGV) && (BaseOffs == O.BaseOffs) &&
1053 (HasBaseReg == O.HasBaseReg) && (Scale == O.Scale);
1054 }
1055 };
1057 #ifndef NDEBUG
1058 static inline raw_ostream &operator<<(raw_ostream &OS, const ExtAddrMode &AM) {
1059 AM.print(OS);
1060 return OS;
1061 }
1062 #endif
1064 void ExtAddrMode::print(raw_ostream &OS) const {
1065 bool NeedPlus = false;
1066 OS << "[";
1067 if (BaseGV) {
1068 OS << (NeedPlus ? " + " : "")
1069 << "GV:";
1070 BaseGV->printAsOperand(OS, /*PrintType=*/false);
1071 NeedPlus = true;
1072 }
1074 if (BaseOffs) {
1075 OS << (NeedPlus ? " + " : "")
1076 << BaseOffs;
1077 NeedPlus = true;
1078 }
1080 if (BaseReg) {
1081 OS << (NeedPlus ? " + " : "")
1082 << "Base:";
1083 BaseReg->printAsOperand(OS, /*PrintType=*/false);
1084 NeedPlus = true;
1085 }
1086 if (Scale) {
1087 OS << (NeedPlus ? " + " : "")
1088 << Scale << "*";
1089 ScaledReg->printAsOperand(OS, /*PrintType=*/false);
1090 }
1092 OS << ']';
1093 }
1095 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
1096 void ExtAddrMode::dump() const {
1097 print(dbgs());
1098 dbgs() << '\n';
1099 }
1100 #endif
1102 /// \brief This class provides transaction based operation on the IR.
1103 /// Every change made through this class is recorded in the internal state and
1104 /// can be undone (rollback) until commit is called.
1105 class TypePromotionTransaction {
1107 /// \brief This represents the common interface of the individual transaction.
1108 /// Each class implements the logic for doing one specific modification on
1109 /// the IR via the TypePromotionTransaction.
1110 class TypePromotionAction {
1111 protected:
1112 /// The Instruction modified.
1113 Instruction *Inst;
1115 public:
1116 /// \brief Constructor of the action.
1117 /// The constructor performs the related action on the IR.
1118 TypePromotionAction(Instruction *Inst) : Inst(Inst) {}
1120 virtual ~TypePromotionAction() {}
1122 /// \brief Undo the modification done by this action.
1123 /// When this method is called, the IR must be in the same state as it was
1124 /// before this action was applied.
1125 /// \pre Undoing the action works if and only if the IR is in the exact same
1126 /// state as it was directly after this action was applied.
1127 virtual void undo() = 0;
1129 /// \brief Advocate every change made by this action.
1130 /// When the results on the IR of the action are to be kept, it is important
1131 /// to call this function, otherwise hidden information may be kept forever.
1132 virtual void commit() {
1133 // Nothing to be done, this action is not doing anything.
1134 }
1135 };
1137 /// \brief Utility to remember the position of an instruction.
1138 class InsertionHandler {
1139 /// Position of an instruction.
1140 /// Either an instruction:
1141 /// - Is the first in a basic block: BB is used.
1142 /// - Has a previous instructon: PrevInst is used.
1143 union {
1144 Instruction *PrevInst;
1145 BasicBlock *BB;
1146 } Point;
1147 /// Remember whether or not the instruction had a previous instruction.
1148 bool HasPrevInstruction;
1150 public:
1151 /// \brief Record the position of \p Inst.
1152 InsertionHandler(Instruction *Inst) {
1153 BasicBlock::iterator It = Inst;
1154 HasPrevInstruction = (It != (Inst->getParent()->begin()));
1155 if (HasPrevInstruction)
1156 Point.PrevInst = --It;
1157 else
1158 Point.BB = Inst->getParent();
1159 }
1161 /// \brief Insert \p Inst at the recorded position.
1162 void insert(Instruction *Inst) {
1163 if (HasPrevInstruction) {
1164 if (Inst->getParent())
1165 Inst->removeFromParent();
1166 Inst->insertAfter(Point.PrevInst);
1167 } else {
1168 Instruction *Position = Point.BB->getFirstInsertionPt();
1169 if (Inst->getParent())
1170 Inst->moveBefore(Position);
1171 else
1172 Inst->insertBefore(Position);
1173 }
1174 }
1175 };
1177 /// \brief Move an instruction before another.
1178 class InstructionMoveBefore : public TypePromotionAction {
1179 /// Original position of the instruction.
1180 InsertionHandler Position;
1182 public:
1183 /// \brief Move \p Inst before \p Before.
1184 InstructionMoveBefore(Instruction *Inst, Instruction *Before)
1185 : TypePromotionAction(Inst), Position(Inst) {
1186 DEBUG(dbgs() << "Do: move: " << *Inst << "\nbefore: " << *Before << "\n");
1187 Inst->moveBefore(Before);
1188 }
1190 /// \brief Move the instruction back to its original position.
1191 void undo() override {
1192 DEBUG(dbgs() << "Undo: moveBefore: " << *Inst << "\n");
1193 Position.insert(Inst);
1194 }
1195 };
1197 /// \brief Set the operand of an instruction with a new value.
1198 class OperandSetter : public TypePromotionAction {
1199 /// Original operand of the instruction.
1200 Value *Origin;
1201 /// Index of the modified instruction.
1202 unsigned Idx;
1204 public:
1205 /// \brief Set \p Idx operand of \p Inst with \p NewVal.
1206 OperandSetter(Instruction *Inst, unsigned Idx, Value *NewVal)
1207 : TypePromotionAction(Inst), Idx(Idx) {
1208 DEBUG(dbgs() << "Do: setOperand: " << Idx << "\n"
1209 << "for:" << *Inst << "\n"
1210 << "with:" << *NewVal << "\n");
1211 Origin = Inst->getOperand(Idx);
1212 Inst->setOperand(Idx, NewVal);
1213 }
1215 /// \brief Restore the original value of the instruction.
1216 void undo() override {
1217 DEBUG(dbgs() << "Undo: setOperand:" << Idx << "\n"
1218 << "for: " << *Inst << "\n"
1219 << "with: " << *Origin << "\n");
1220 Inst->setOperand(Idx, Origin);
1221 }
1222 };
1224 /// \brief Hide the operands of an instruction.
1225 /// Do as if this instruction was not using any of its operands.
1226 class OperandsHider : public TypePromotionAction {
1227 /// The list of original operands.
1228 SmallVector<Value *, 4> OriginalValues;
1230 public:
1231 /// \brief Remove \p Inst from the uses of the operands of \p Inst.
1232 OperandsHider(Instruction *Inst) : TypePromotionAction(Inst) {
1233 DEBUG(dbgs() << "Do: OperandsHider: " << *Inst << "\n");
1234 unsigned NumOpnds = Inst->getNumOperands();
1235 OriginalValues.reserve(NumOpnds);
1236 for (unsigned It = 0; It < NumOpnds; ++It) {
1237 // Save the current operand.
1238 Value *Val = Inst->getOperand(It);
1239 OriginalValues.push_back(Val);
1240 // Set a dummy one.
1241 // We could use OperandSetter here, but that would implied an overhead
1242 // that we are not willing to pay.
1243 Inst->setOperand(It, UndefValue::get(Val->getType()));
1244 }
1245 }
1247 /// \brief Restore the original list of uses.
1248 void undo() override {
1249 DEBUG(dbgs() << "Undo: OperandsHider: " << *Inst << "\n");
1250 for (unsigned It = 0, EndIt = OriginalValues.size(); It != EndIt; ++It)
1251 Inst->setOperand(It, OriginalValues[It]);
1252 }
1253 };
1255 /// \brief Build a truncate instruction.
1256 class TruncBuilder : public TypePromotionAction {
1257 public:
1258 /// \brief Build a truncate instruction of \p Opnd producing a \p Ty
1259 /// result.
1260 /// trunc Opnd to Ty.
1261 TruncBuilder(Instruction *Opnd, Type *Ty) : TypePromotionAction(Opnd) {
1262 IRBuilder<> Builder(Opnd);
1263 Inst = cast<Instruction>(Builder.CreateTrunc(Opnd, Ty, "promoted"));
1264 DEBUG(dbgs() << "Do: TruncBuilder: " << *Inst << "\n");
1265 }
1267 /// \brief Get the built instruction.
1268 Instruction *getBuiltInstruction() { return Inst; }
1270 /// \brief Remove the built instruction.
1271 void undo() override {
1272 DEBUG(dbgs() << "Undo: TruncBuilder: " << *Inst << "\n");
1273 Inst->eraseFromParent();
1274 }
1275 };
1277 /// \brief Build a sign extension instruction.
1278 class SExtBuilder : public TypePromotionAction {
1279 public:
1280 /// \brief Build a sign extension instruction of \p Opnd producing a \p Ty
1281 /// result.
1282 /// sext Opnd to Ty.
1283 SExtBuilder(Instruction *InsertPt, Value *Opnd, Type *Ty)
1284 : TypePromotionAction(Inst) {
1285 IRBuilder<> Builder(InsertPt);
1286 Inst = cast<Instruction>(Builder.CreateSExt(Opnd, Ty, "promoted"));
1287 DEBUG(dbgs() << "Do: SExtBuilder: " << *Inst << "\n");
1288 }
1290 /// \brief Get the built instruction.
1291 Instruction *getBuiltInstruction() { return Inst; }
1293 /// \brief Remove the built instruction.
1294 void undo() override {
1295 DEBUG(dbgs() << "Undo: SExtBuilder: " << *Inst << "\n");
1296 Inst->eraseFromParent();
1297 }
1298 };
1300 /// \brief Build a zero extension instruction.
1301 class ZExtBuilder : public TypePromotionAction {
1302 public:
1303 /// \brief Build a zero extension instruction of \p Opnd producing a \p Ty
1304 /// result.
1305 /// zext Opnd to Ty.
1306 ZExtBuilder(Instruction *InsertPt, Value *Opnd, Type *Ty)
1307 : TypePromotionAction(Inst) {
1308 IRBuilder<> Builder(InsertPt);
1309 Inst = cast<Instruction>(Builder.CreateZExt(Opnd, Ty, "promoted"));
1310 DEBUG(dbgs() << "Do: ZExtBuilder: " << *Inst << "\n");
1311 }
1313 /// \brief Get the built instruction.
1314 Instruction *getBuiltInstruction() { return Inst; }
1316 /// \brief Remove the built instruction.
1317 void undo() override {
1318 DEBUG(dbgs() << "Undo: ZExtBuilder: " << *Inst << "\n");
1319 Inst->eraseFromParent();
1320 }
1321 };
1323 /// \brief Mutate an instruction to another type.
1324 class TypeMutator : public TypePromotionAction {
1325 /// Record the original type.
1326 Type *OrigTy;
1328 public:
1329 /// \brief Mutate the type of \p Inst into \p NewTy.
1330 TypeMutator(Instruction *Inst, Type *NewTy)
1331 : TypePromotionAction(Inst), OrigTy(Inst->getType()) {
1332 DEBUG(dbgs() << "Do: MutateType: " << *Inst << " with " << *NewTy
1333 << "\n");
1334 Inst->mutateType(NewTy);
1335 }
1337 /// \brief Mutate the instruction back to its original type.
1338 void undo() override {
1339 DEBUG(dbgs() << "Undo: MutateType: " << *Inst << " with " << *OrigTy
1340 << "\n");
1341 Inst->mutateType(OrigTy);
1342 }
1343 };
1345 /// \brief Replace the uses of an instruction by another instruction.
1346 class UsesReplacer : public TypePromotionAction {
1347 /// Helper structure to keep track of the replaced uses.
1348 struct InstructionAndIdx {
1349 /// The instruction using the instruction.
1350 Instruction *Inst;
1351 /// The index where this instruction is used for Inst.
1352 unsigned Idx;
1353 InstructionAndIdx(Instruction *Inst, unsigned Idx)
1354 : Inst(Inst), Idx(Idx) {}
1355 };
1357 /// Keep track of the original uses (pair Instruction, Index).
1358 SmallVector<InstructionAndIdx, 4> OriginalUses;
1359 typedef SmallVectorImpl<InstructionAndIdx>::iterator use_iterator;
1361 public:
1362 /// \brief Replace all the use of \p Inst by \p New.
1363 UsesReplacer(Instruction *Inst, Value *New) : TypePromotionAction(Inst) {
1364 DEBUG(dbgs() << "Do: UsersReplacer: " << *Inst << " with " << *New
1365 << "\n");
1366 // Record the original uses.
1367 for (Use &U : Inst->uses()) {
1368 Instruction *UserI = cast<Instruction>(U.getUser());
1369 OriginalUses.push_back(InstructionAndIdx(UserI, U.getOperandNo()));
1370 }
1371 // Now, we can replace the uses.
1372 Inst->replaceAllUsesWith(New);
1373 }
1375 /// \brief Reassign the original uses of Inst to Inst.
1376 void undo() override {
1377 DEBUG(dbgs() << "Undo: UsersReplacer: " << *Inst << "\n");
1378 for (use_iterator UseIt = OriginalUses.begin(),
1379 EndIt = OriginalUses.end();
1380 UseIt != EndIt; ++UseIt) {
1381 UseIt->Inst->setOperand(UseIt->Idx, Inst);
1382 }
1383 }
1384 };
1386 /// \brief Remove an instruction from the IR.
1387 class InstructionRemover : public TypePromotionAction {
1388 /// Original position of the instruction.
1389 InsertionHandler Inserter;
1390 /// Helper structure to hide all the link to the instruction. In other
1391 /// words, this helps to do as if the instruction was removed.
1392 OperandsHider Hider;
1393 /// Keep track of the uses replaced, if any.
1394 UsesReplacer *Replacer;
1396 public:
1397 /// \brief Remove all reference of \p Inst and optinally replace all its
1398 /// uses with New.
1399 /// \pre If !Inst->use_empty(), then New != nullptr
1400 InstructionRemover(Instruction *Inst, Value *New = nullptr)
1401 : TypePromotionAction(Inst), Inserter(Inst), Hider(Inst),
1402 Replacer(nullptr) {
1403 if (New)
1404 Replacer = new UsesReplacer(Inst, New);
1405 DEBUG(dbgs() << "Do: InstructionRemover: " << *Inst << "\n");
1406 Inst->removeFromParent();
1407 }
1409 ~InstructionRemover() { delete Replacer; }
1411 /// \brief Really remove the instruction.
1412 void commit() override { delete Inst; }
1414 /// \brief Resurrect the instruction and reassign it to the proper uses if
1415 /// new value was provided when build this action.
1416 void undo() override {
1417 DEBUG(dbgs() << "Undo: InstructionRemover: " << *Inst << "\n");
1418 Inserter.insert(Inst);
1419 if (Replacer)
1420 Replacer->undo();
1421 Hider.undo();
1422 }
1423 };
1425 public:
1426 /// Restoration point.
1427 /// The restoration point is a pointer to an action instead of an iterator
1428 /// because the iterator may be invalidated but not the pointer.
1429 typedef const TypePromotionAction *ConstRestorationPt;
1430 /// Advocate every changes made in that transaction.
1431 void commit();
1432 /// Undo all the changes made after the given point.
1433 void rollback(ConstRestorationPt Point);
1434 /// Get the current restoration point.
1435 ConstRestorationPt getRestorationPoint() const;
1437 /// \name API for IR modification with state keeping to support rollback.
1438 /// @{
1439 /// Same as Instruction::setOperand.
1440 void setOperand(Instruction *Inst, unsigned Idx, Value *NewVal);
1441 /// Same as Instruction::eraseFromParent.
1442 void eraseInstruction(Instruction *Inst, Value *NewVal = nullptr);
1443 /// Same as Value::replaceAllUsesWith.
1444 void replaceAllUsesWith(Instruction *Inst, Value *New);
1445 /// Same as Value::mutateType.
1446 void mutateType(Instruction *Inst, Type *NewTy);
1447 /// Same as IRBuilder::createTrunc.
1448 Instruction *createTrunc(Instruction *Opnd, Type *Ty);
1449 /// Same as IRBuilder::createSExt.
1450 Instruction *createSExt(Instruction *Inst, Value *Opnd, Type *Ty);
1451 /// Same as IRBuilder::createZExt.
1452 Instruction *createZExt(Instruction *Inst, Value *Opnd, Type *Ty);
1453 /// Same as Instruction::moveBefore.
1454 void moveBefore(Instruction *Inst, Instruction *Before);
1455 /// @}
1457 private:
1458 /// The ordered list of actions made so far.
1459 SmallVector<std::unique_ptr<TypePromotionAction>, 16> Actions;
1460 typedef SmallVectorImpl<std::unique_ptr<TypePromotionAction>>::iterator CommitPt;
1461 };
1463 void TypePromotionTransaction::setOperand(Instruction *Inst, unsigned Idx,
1464 Value *NewVal) {
1465 Actions.push_back(
1466 make_unique<TypePromotionTransaction::OperandSetter>(Inst, Idx, NewVal));
1467 }
1469 void TypePromotionTransaction::eraseInstruction(Instruction *Inst,
1470 Value *NewVal) {
1471 Actions.push_back(
1472 make_unique<TypePromotionTransaction::InstructionRemover>(Inst, NewVal));
1473 }
1475 void TypePromotionTransaction::replaceAllUsesWith(Instruction *Inst,
1476 Value *New) {
1477 Actions.push_back(make_unique<TypePromotionTransaction::UsesReplacer>(Inst, New));
1478 }
1480 void TypePromotionTransaction::mutateType(Instruction *Inst, Type *NewTy) {
1481 Actions.push_back(make_unique<TypePromotionTransaction::TypeMutator>(Inst, NewTy));
1482 }
1484 Instruction *TypePromotionTransaction::createTrunc(Instruction *Opnd,
1485 Type *Ty) {
1486 std::unique_ptr<TruncBuilder> Ptr(new TruncBuilder(Opnd, Ty));
1487 Instruction *I = Ptr->getBuiltInstruction();
1488 Actions.push_back(std::move(Ptr));
1489 return I;
1490 }
1492 Instruction *TypePromotionTransaction::createSExt(Instruction *Inst,
1493 Value *Opnd, Type *Ty) {
1494 std::unique_ptr<SExtBuilder> Ptr(new SExtBuilder(Inst, Opnd, Ty));
1495 Instruction *I = Ptr->getBuiltInstruction();
1496 Actions.push_back(std::move(Ptr));
1497 return I;
1498 }
1500 Instruction *TypePromotionTransaction::createZExt(Instruction *Inst,
1501 Value *Opnd, Type *Ty) {
1502 std::unique_ptr<ZExtBuilder> Ptr(new ZExtBuilder(Inst, Opnd, Ty));
1503 Instruction *I = Ptr->getBuiltInstruction();
1504 Actions.push_back(std::move(Ptr));
1505 return I;
1506 }
1508 void TypePromotionTransaction::moveBefore(Instruction *Inst,
1509 Instruction *Before) {
1510 Actions.push_back(
1511 make_unique<TypePromotionTransaction::InstructionMoveBefore>(Inst, Before));
1512 }
1514 TypePromotionTransaction::ConstRestorationPt
1515 TypePromotionTransaction::getRestorationPoint() const {
1516 return !Actions.empty() ? Actions.back().get() : nullptr;
1517 }
1519 void TypePromotionTransaction::commit() {
1520 for (CommitPt It = Actions.begin(), EndIt = Actions.end(); It != EndIt;
1521 ++It)
1522 (*It)->commit();
1523 Actions.clear();
1524 }
1526 void TypePromotionTransaction::rollback(
1527 TypePromotionTransaction::ConstRestorationPt Point) {
1528 while (!Actions.empty() && Point != Actions.back().get()) {
1529 std::unique_ptr<TypePromotionAction> Curr = Actions.pop_back_val();
1530 Curr->undo();
1531 }
1532 }
1534 /// \brief A helper class for matching addressing modes.
1535 ///
1536 /// This encapsulates the logic for matching the target-legal addressing modes.
1537 class AddressingModeMatcher {
1538 SmallVectorImpl<Instruction*> &AddrModeInsts;
1539 const TargetLowering &TLI;
1541 /// AccessTy/MemoryInst - This is the type for the access (e.g. double) and
1542 /// the memory instruction that we're computing this address for.
1543 Type *AccessTy;
1544 Instruction *MemoryInst;
1546 /// AddrMode - This is the addressing mode that we're building up. This is
1547 /// part of the return value of this addressing mode matching stuff.
1548 ExtAddrMode &AddrMode;
1550 /// The truncate instruction inserted by other CodeGenPrepare optimizations.
1551 const SetOfInstrs &InsertedTruncs;
1552 /// A map from the instructions to their type before promotion.
1553 InstrToOrigTy &PromotedInsts;
1554 /// The ongoing transaction where every action should be registered.
1555 TypePromotionTransaction &TPT;
1557 /// IgnoreProfitability - This is set to true when we should not do
1558 /// profitability checks. When true, IsProfitableToFoldIntoAddressingMode
1559 /// always returns true.
1560 bool IgnoreProfitability;
1562 AddressingModeMatcher(SmallVectorImpl<Instruction*> &AMI,
1563 const TargetLowering &T, Type *AT,
1564 Instruction *MI, ExtAddrMode &AM,
1565 const SetOfInstrs &InsertedTruncs,
1566 InstrToOrigTy &PromotedInsts,
1567 TypePromotionTransaction &TPT)
1568 : AddrModeInsts(AMI), TLI(T), AccessTy(AT), MemoryInst(MI), AddrMode(AM),
1569 InsertedTruncs(InsertedTruncs), PromotedInsts(PromotedInsts), TPT(TPT) {
1570 IgnoreProfitability = false;
1571 }
1572 public:
1574 /// Match - Find the maximal addressing mode that a load/store of V can fold,
1575 /// give an access type of AccessTy. This returns a list of involved
1576 /// instructions in AddrModeInsts.
1577 /// \p InsertedTruncs The truncate instruction inserted by other
1578 /// CodeGenPrepare
1579 /// optimizations.
1580 /// \p PromotedInsts maps the instructions to their type before promotion.
1581 /// \p The ongoing transaction where every action should be registered.
1582 static ExtAddrMode Match(Value *V, Type *AccessTy,
1583 Instruction *MemoryInst,
1584 SmallVectorImpl<Instruction*> &AddrModeInsts,
1585 const TargetLowering &TLI,
1586 const SetOfInstrs &InsertedTruncs,
1587 InstrToOrigTy &PromotedInsts,
1588 TypePromotionTransaction &TPT) {
1589 ExtAddrMode Result;
1591 bool Success = AddressingModeMatcher(AddrModeInsts, TLI, AccessTy,
1592 MemoryInst, Result, InsertedTruncs,
1593 PromotedInsts, TPT).MatchAddr(V, 0);
1594 (void)Success; assert(Success && "Couldn't select *anything*?");
1595 return Result;
1596 }
1597 private:
1598 bool MatchScaledValue(Value *ScaleReg, int64_t Scale, unsigned Depth);
1599 bool MatchAddr(Value *V, unsigned Depth);
1600 bool MatchOperationAddr(User *Operation, unsigned Opcode, unsigned Depth,
1601 bool *MovedAway = nullptr);
1602 bool IsProfitableToFoldIntoAddressingMode(Instruction *I,
1603 ExtAddrMode &AMBefore,
1604 ExtAddrMode &AMAfter);
1605 bool ValueAlreadyLiveAtInst(Value *Val, Value *KnownLive1, Value *KnownLive2);
1606 bool IsPromotionProfitable(unsigned MatchedSize, unsigned SizeWithPromotion,
1607 Value *PromotedOperand) const;
1608 };
1610 /// MatchScaledValue - Try adding ScaleReg*Scale to the current addressing mode.
1611 /// Return true and update AddrMode if this addr mode is legal for the target,
1612 /// false if not.
1613 bool AddressingModeMatcher::MatchScaledValue(Value *ScaleReg, int64_t Scale,
1614 unsigned Depth) {
1615 // If Scale is 1, then this is the same as adding ScaleReg to the addressing
1616 // mode. Just process that directly.
1617 if (Scale == 1)
1618 return MatchAddr(ScaleReg, Depth);
1620 // If the scale is 0, it takes nothing to add this.
1621 if (Scale == 0)
1622 return true;
1624 // If we already have a scale of this value, we can add to it, otherwise, we
1625 // need an available scale field.
1626 if (AddrMode.Scale != 0 && AddrMode.ScaledReg != ScaleReg)
1627 return false;
1629 ExtAddrMode TestAddrMode = AddrMode;
1631 // Add scale to turn X*4+X*3 -> X*7. This could also do things like
1632 // [A+B + A*7] -> [B+A*8].
1633 TestAddrMode.Scale += Scale;
1634 TestAddrMode.ScaledReg = ScaleReg;
1636 // If the new address isn't legal, bail out.
1637 if (!TLI.isLegalAddressingMode(TestAddrMode, AccessTy))
1638 return false;
1640 // It was legal, so commit it.
1641 AddrMode = TestAddrMode;
1643 // Okay, we decided that we can add ScaleReg+Scale to AddrMode. Check now
1644 // to see if ScaleReg is actually X+C. If so, we can turn this into adding
1645 // X*Scale + C*Scale to addr mode.
1646 ConstantInt *CI = nullptr; Value *AddLHS = nullptr;
1647 if (isa<Instruction>(ScaleReg) && // not a constant expr.
1648 match(ScaleReg, m_Add(m_Value(AddLHS), m_ConstantInt(CI)))) {
1649 TestAddrMode.ScaledReg = AddLHS;
1650 TestAddrMode.BaseOffs += CI->getSExtValue()*TestAddrMode.Scale;
1652 // If this addressing mode is legal, commit it and remember that we folded
1653 // this instruction.
1654 if (TLI.isLegalAddressingMode(TestAddrMode, AccessTy)) {
1655 AddrModeInsts.push_back(cast<Instruction>(ScaleReg));
1656 AddrMode = TestAddrMode;
1657 return true;
1658 }
1659 }
1661 // Otherwise, not (x+c)*scale, just return what we have.
1662 return true;
1663 }
1665 /// MightBeFoldableInst - This is a little filter, which returns true if an
1666 /// addressing computation involving I might be folded into a load/store
1667 /// accessing it. This doesn't need to be perfect, but needs to accept at least
1668 /// the set of instructions that MatchOperationAddr can.
1669 static bool MightBeFoldableInst(Instruction *I) {
1670 switch (I->getOpcode()) {
1671 case Instruction::BitCast:
1672 case Instruction::AddrSpaceCast:
1673 // Don't touch identity bitcasts.
1674 if (I->getType() == I->getOperand(0)->getType())
1675 return false;
1676 return I->getType()->isPointerTy() || I->getType()->isIntegerTy();
1677 case Instruction::PtrToInt:
1678 // PtrToInt is always a noop, as we know that the int type is pointer sized.
1679 return true;
1680 case Instruction::IntToPtr:
1681 // We know the input is intptr_t, so this is foldable.
1682 return true;
1683 case Instruction::Add:
1684 return true;
1685 case Instruction::Mul:
1686 case Instruction::Shl:
1687 // Can only handle X*C and X << C.
1688 return isa<ConstantInt>(I->getOperand(1));
1689 case Instruction::GetElementPtr:
1690 return true;
1691 default:
1692 return false;
1693 }
1694 }
1696 /// \brief Hepler class to perform type promotion.
1697 class TypePromotionHelper {
1698 /// \brief Utility function to check whether or not a sign extension of
1699 /// \p Inst with \p ConsideredSExtType can be moved through \p Inst by either
1700 /// using the operands of \p Inst or promoting \p Inst.
1701 /// In other words, check if:
1702 /// sext (Ty Inst opnd1 opnd2 ... opndN) to ConsideredSExtType.
1703 /// #1 Promotion applies:
1704 /// ConsideredSExtType Inst (sext opnd1 to ConsideredSExtType, ...).
1705 /// #2 Operand reuses:
1706 /// sext opnd1 to ConsideredSExtType.
1707 /// \p PromotedInsts maps the instructions to their type before promotion.
1708 static bool canGetThrough(const Instruction *Inst, Type *ConsideredSExtType,
1709 const InstrToOrigTy &PromotedInsts);
1711 /// \brief Utility function to determine if \p OpIdx should be promoted when
1712 /// promoting \p Inst.
1713 static bool shouldSExtOperand(const Instruction *Inst, int OpIdx) {
1714 if (isa<SelectInst>(Inst) && OpIdx == 0)
1715 return false;
1716 return true;
1717 }
1719 /// \brief Utility function to promote the operand of \p SExt when this
1720 /// operand is a promotable trunc or sext or zext.
1721 /// \p PromotedInsts maps the instructions to their type before promotion.
1722 /// \p CreatedInsts[out] contains how many non-free instructions have been
1723 /// created to promote the operand of SExt.
1724 /// Should never be called directly.
1725 /// \return The promoted value which is used instead of SExt.
1726 static Value *promoteOperandForTruncAndAnyExt(Instruction *SExt,
1727 TypePromotionTransaction &TPT,
1728 InstrToOrigTy &PromotedInsts,
1729 unsigned &CreatedInsts);
1731 /// \brief Utility function to promote the operand of \p SExt when this
1732 /// operand is promotable and is not a supported trunc or sext.
1733 /// \p PromotedInsts maps the instructions to their type before promotion.
1734 /// \p CreatedInsts[out] contains how many non-free instructions have been
1735 /// created to promote the operand of SExt.
1736 /// Should never be called directly.
1737 /// \return The promoted value which is used instead of SExt.
1738 static Value *promoteOperandForOther(Instruction *SExt,
1739 TypePromotionTransaction &TPT,
1740 InstrToOrigTy &PromotedInsts,
1741 unsigned &CreatedInsts);
1743 public:
1744 /// Type for the utility function that promotes the operand of SExt.
1745 typedef Value *(*Action)(Instruction *SExt, TypePromotionTransaction &TPT,
1746 InstrToOrigTy &PromotedInsts,
1747 unsigned &CreatedInsts);
1748 /// \brief Given a sign extend instruction \p SExt, return the approriate
1749 /// action to promote the operand of \p SExt instead of using SExt.
1750 /// \return NULL if no promotable action is possible with the current
1751 /// sign extension.
1752 /// \p InsertedTruncs keeps track of all the truncate instructions inserted by
1753 /// the others CodeGenPrepare optimizations. This information is important
1754 /// because we do not want to promote these instructions as CodeGenPrepare
1755 /// will reinsert them later. Thus creating an infinite loop: create/remove.
1756 /// \p PromotedInsts maps the instructions to their type before promotion.
1757 static Action getAction(Instruction *SExt, const SetOfInstrs &InsertedTruncs,
1758 const TargetLowering &TLI,
1759 const InstrToOrigTy &PromotedInsts);
1760 };
1762 bool TypePromotionHelper::canGetThrough(const Instruction *Inst,
1763 Type *ConsideredSExtType,
1764 const InstrToOrigTy &PromotedInsts) {
1765 // We can always get through sext or zext.
1766 if (isa<SExtInst>(Inst) || isa<ZExtInst>(Inst))
1767 return true;
1769 // We can get through binary operator, if it is legal. In other words, the
1770 // binary operator must have a nuw or nsw flag.
1771 const BinaryOperator *BinOp = dyn_cast<BinaryOperator>(Inst);
1772 if (BinOp && isa<OverflowingBinaryOperator>(BinOp) &&
1773 (BinOp->hasNoUnsignedWrap() || BinOp->hasNoSignedWrap()))
1774 return true;
1776 // Check if we can do the following simplification.
1777 // sext(trunc(sext)) --> sext
1778 if (!isa<TruncInst>(Inst))
1779 return false;
1781 Value *OpndVal = Inst->getOperand(0);
1782 // Check if we can use this operand in the sext.
1783 // If the type is larger than the result type of the sign extension,
1784 // we cannot.
1785 if (OpndVal->getType()->getIntegerBitWidth() >
1786 ConsideredSExtType->getIntegerBitWidth())
1787 return false;
1789 // If the operand of the truncate is not an instruction, we will not have
1790 // any information on the dropped bits.
1791 // (Actually we could for constant but it is not worth the extra logic).
1792 Instruction *Opnd = dyn_cast<Instruction>(OpndVal);
1793 if (!Opnd)
1794 return false;
1796 // Check if the source of the type is narrow enough.
1797 // I.e., check that trunc just drops sign extended bits.
1798 // #1 get the type of the operand.
1799 const Type *OpndType;
1800 InstrToOrigTy::const_iterator It = PromotedInsts.find(Opnd);
1801 if (It != PromotedInsts.end())
1802 OpndType = It->second;
1803 else if (isa<SExtInst>(Opnd))
1804 OpndType = cast<Instruction>(Opnd)->getOperand(0)->getType();
1805 else
1806 return false;
1808 // #2 check that the truncate just drop sign extended bits.
1809 if (Inst->getType()->getIntegerBitWidth() >= OpndType->getIntegerBitWidth())
1810 return true;
1812 return false;
1813 }
1815 TypePromotionHelper::Action TypePromotionHelper::getAction(
1816 Instruction *SExt, const SetOfInstrs &InsertedTruncs,
1817 const TargetLowering &TLI, const InstrToOrigTy &PromotedInsts) {
1818 Instruction *SExtOpnd = dyn_cast<Instruction>(SExt->getOperand(0));
1819 Type *SExtTy = SExt->getType();
1820 // If the operand of the sign extension is not an instruction, we cannot
1821 // get through.
1822 // If it, check we can get through.
1823 if (!SExtOpnd || !canGetThrough(SExtOpnd, SExtTy, PromotedInsts))
1824 return nullptr;
1826 // Do not promote if the operand has been added by codegenprepare.
1827 // Otherwise, it means we are undoing an optimization that is likely to be
1828 // redone, thus causing potential infinite loop.
1829 if (isa<TruncInst>(SExtOpnd) && InsertedTruncs.count(SExtOpnd))
1830 return nullptr;
1832 // SExt or Trunc instructions.
1833 // Return the related handler.
1834 if (isa<SExtInst>(SExtOpnd) || isa<TruncInst>(SExtOpnd) ||
1835 isa<ZExtInst>(SExtOpnd))
1836 return promoteOperandForTruncAndAnyExt;
1838 // Regular instruction.
1839 // Abort early if we will have to insert non-free instructions.
1840 if (!SExtOpnd->hasOneUse() &&
1841 !TLI.isTruncateFree(SExtTy, SExtOpnd->getType()))
1842 return nullptr;
1843 return promoteOperandForOther;
1844 }
1846 Value *TypePromotionHelper::promoteOperandForTruncAndAnyExt(
1847 llvm::Instruction *SExt, TypePromotionTransaction &TPT,
1848 InstrToOrigTy &PromotedInsts, unsigned &CreatedInsts) {
1849 // By construction, the operand of SExt is an instruction. Otherwise we cannot
1850 // get through it and this method should not be called.
1851 Instruction *SExtOpnd = cast<Instruction>(SExt->getOperand(0));
1852 Instruction *ExtInst = SExt;
1853 if (isa<ZExtInst>(SExtOpnd)) {
1854 // Replace sext(zext(opnd))
1855 // => zext(opnd).
1856 Instruction *ZExt =
1857 TPT.createZExt(SExt, SExtOpnd->getOperand(0), SExt->getType());
1858 TPT.replaceAllUsesWith(SExt, ZExt);
1859 TPT.eraseInstruction(SExt);
1860 ExtInst = ZExt;
1861 } else {
1862 // Replace sext(trunc(opnd)) or sext(sext(opnd))
1863 // => sext(opnd).
1864 TPT.setOperand(SExt, 0, SExtOpnd->getOperand(0));
1865 }
1866 CreatedInsts = 0;
1868 // Remove dead code.
1869 if (SExtOpnd->use_empty())
1870 TPT.eraseInstruction(SExtOpnd);
1872 // Check if the extension is still needed.
1873 if (ExtInst->getType() != ExtInst->getOperand(0)->getType())
1874 return ExtInst;
1876 // At this point we have: ext ty opnd to ty.
1877 // Reassign the uses of ExtInst to the opnd and remove ExtInst.
1878 Value *NextVal = ExtInst->getOperand(0);
1879 TPT.eraseInstruction(ExtInst, NextVal);
1880 return NextVal;
1881 }
1883 Value *
1884 TypePromotionHelper::promoteOperandForOther(Instruction *SExt,
1885 TypePromotionTransaction &TPT,
1886 InstrToOrigTy &PromotedInsts,
1887 unsigned &CreatedInsts) {
1888 // By construction, the operand of SExt is an instruction. Otherwise we cannot
1889 // get through it and this method should not be called.
1890 Instruction *SExtOpnd = cast<Instruction>(SExt->getOperand(0));
1891 CreatedInsts = 0;
1892 if (!SExtOpnd->hasOneUse()) {
1893 // SExtOpnd will be promoted.
1894 // All its uses, but SExt, will need to use a truncated value of the
1895 // promoted version.
1896 // Create the truncate now.
1897 Instruction *Trunc = TPT.createTrunc(SExt, SExtOpnd->getType());
1898 Trunc->removeFromParent();
1899 // Insert it just after the definition.
1900 Trunc->insertAfter(SExtOpnd);
1902 TPT.replaceAllUsesWith(SExtOpnd, Trunc);
1903 // Restore the operand of SExt (which has been replace by the previous call
1904 // to replaceAllUsesWith) to avoid creating a cycle trunc <-> sext.
1905 TPT.setOperand(SExt, 0, SExtOpnd);
1906 }
1908 // Get through the Instruction:
1909 // 1. Update its type.
1910 // 2. Replace the uses of SExt by Inst.
1911 // 3. Sign extend each operand that needs to be sign extended.
1913 // Remember the original type of the instruction before promotion.
1914 // This is useful to know that the high bits are sign extended bits.
1915 PromotedInsts.insert(
1916 std::pair<Instruction *, Type *>(SExtOpnd, SExtOpnd->getType()));
1917 // Step #1.
1918 TPT.mutateType(SExtOpnd, SExt->getType());
1919 // Step #2.
1920 TPT.replaceAllUsesWith(SExt, SExtOpnd);
1921 // Step #3.
1922 Instruction *SExtForOpnd = SExt;
1924 DEBUG(dbgs() << "Propagate SExt to operands\n");
1925 for (int OpIdx = 0, EndOpIdx = SExtOpnd->getNumOperands(); OpIdx != EndOpIdx;
1926 ++OpIdx) {
1927 DEBUG(dbgs() << "Operand:\n" << *(SExtOpnd->getOperand(OpIdx)) << '\n');
1928 if (SExtOpnd->getOperand(OpIdx)->getType() == SExt->getType() ||
1929 !shouldSExtOperand(SExtOpnd, OpIdx)) {
1930 DEBUG(dbgs() << "No need to propagate\n");
1931 continue;
1932 }
1933 // Check if we can statically sign extend the operand.
1934 Value *Opnd = SExtOpnd->getOperand(OpIdx);
1935 if (const ConstantInt *Cst = dyn_cast<ConstantInt>(Opnd)) {
1936 DEBUG(dbgs() << "Statically sign extend\n");
1937 TPT.setOperand(
1938 SExtOpnd, OpIdx,
1939 ConstantInt::getSigned(SExt->getType(), Cst->getSExtValue()));
1940 continue;
1941 }
1942 // UndefValue are typed, so we have to statically sign extend them.
1943 if (isa<UndefValue>(Opnd)) {
1944 DEBUG(dbgs() << "Statically sign extend\n");
1945 TPT.setOperand(SExtOpnd, OpIdx, UndefValue::get(SExt->getType()));
1946 continue;
1947 }
1949 // Otherwise we have to explicity sign extend the operand.
1950 // Check if SExt was reused to sign extend an operand.
1951 if (!SExtForOpnd) {
1952 // If yes, create a new one.
1953 DEBUG(dbgs() << "More operands to sext\n");
1954 SExtForOpnd = TPT.createSExt(SExt, Opnd, SExt->getType());
1955 ++CreatedInsts;
1956 }
1958 TPT.setOperand(SExtForOpnd, 0, Opnd);
1960 // Move the sign extension before the insertion point.
1961 TPT.moveBefore(SExtForOpnd, SExtOpnd);
1962 TPT.setOperand(SExtOpnd, OpIdx, SExtForOpnd);
1963 // If more sext are required, new instructions will have to be created.
1964 SExtForOpnd = nullptr;
1965 }
1966 if (SExtForOpnd == SExt) {
1967 DEBUG(dbgs() << "Sign extension is useless now\n");
1968 TPT.eraseInstruction(SExt);
1969 }
1970 return SExtOpnd;
1971 }
1973 /// IsPromotionProfitable - Check whether or not promoting an instruction
1974 /// to a wider type was profitable.
1975 /// \p MatchedSize gives the number of instructions that have been matched
1976 /// in the addressing mode after the promotion was applied.
1977 /// \p SizeWithPromotion gives the number of created instructions for
1978 /// the promotion plus the number of instructions that have been
1979 /// matched in the addressing mode before the promotion.
1980 /// \p PromotedOperand is the value that has been promoted.
1981 /// \return True if the promotion is profitable, false otherwise.
1982 bool
1983 AddressingModeMatcher::IsPromotionProfitable(unsigned MatchedSize,
1984 unsigned SizeWithPromotion,
1985 Value *PromotedOperand) const {
1986 // We folded less instructions than what we created to promote the operand.
1987 // This is not profitable.
1988 if (MatchedSize < SizeWithPromotion)
1989 return false;
1990 if (MatchedSize > SizeWithPromotion)
1991 return true;
1992 // The promotion is neutral but it may help folding the sign extension in
1993 // loads for instance.
1994 // Check that we did not create an illegal instruction.
1995 Instruction *PromotedInst = dyn_cast<Instruction>(PromotedOperand);
1996 if (!PromotedInst)
1997 return false;
1998 int ISDOpcode = TLI.InstructionOpcodeToISD(PromotedInst->getOpcode());
1999 // If the ISDOpcode is undefined, it was undefined before the promotion.
2000 if (!ISDOpcode)
2001 return true;
2002 // Otherwise, check if the promoted instruction is legal or not.
2003 return TLI.isOperationLegalOrCustom(ISDOpcode,
2004 EVT::getEVT(PromotedInst->getType()));
2005 }
2007 /// MatchOperationAddr - Given an instruction or constant expr, see if we can
2008 /// fold the operation into the addressing mode. If so, update the addressing
2009 /// mode and return true, otherwise return false without modifying AddrMode.
2010 /// If \p MovedAway is not NULL, it contains the information of whether or
2011 /// not AddrInst has to be folded into the addressing mode on success.
2012 /// If \p MovedAway == true, \p AddrInst will not be part of the addressing
2013 /// because it has been moved away.
2014 /// Thus AddrInst must not be added in the matched instructions.
2015 /// This state can happen when AddrInst is a sext, since it may be moved away.
2016 /// Therefore, AddrInst may not be valid when MovedAway is true and it must
2017 /// not be referenced anymore.
2018 bool AddressingModeMatcher::MatchOperationAddr(User *AddrInst, unsigned Opcode,
2019 unsigned Depth,
2020 bool *MovedAway) {
2021 // Avoid exponential behavior on extremely deep expression trees.
2022 if (Depth >= 5) return false;
2024 // By default, all matched instructions stay in place.
2025 if (MovedAway)
2026 *MovedAway = false;
2028 switch (Opcode) {
2029 case Instruction::PtrToInt:
2030 // PtrToInt is always a noop, as we know that the int type is pointer sized.
2031 return MatchAddr(AddrInst->getOperand(0), Depth);
2032 case Instruction::IntToPtr:
2033 // This inttoptr is a no-op if the integer type is pointer sized.
2034 if (TLI.getValueType(AddrInst->getOperand(0)->getType()) ==
2035 TLI.getPointerTy(AddrInst->getType()->getPointerAddressSpace()))
2036 return MatchAddr(AddrInst->getOperand(0), Depth);
2037 return false;
2038 case Instruction::BitCast:
2039 case Instruction::AddrSpaceCast:
2040 // BitCast is always a noop, and we can handle it as long as it is
2041 // int->int or pointer->pointer (we don't want int<->fp or something).
2042 if ((AddrInst->getOperand(0)->getType()->isPointerTy() ||
2043 AddrInst->getOperand(0)->getType()->isIntegerTy()) &&
2044 // Don't touch identity bitcasts. These were probably put here by LSR,
2045 // and we don't want to mess around with them. Assume it knows what it
2046 // is doing.
2047 AddrInst->getOperand(0)->getType() != AddrInst->getType())
2048 return MatchAddr(AddrInst->getOperand(0), Depth);
2049 return false;
2050 case Instruction::Add: {
2051 // Check to see if we can merge in the RHS then the LHS. If so, we win.
2052 ExtAddrMode BackupAddrMode = AddrMode;
2053 unsigned OldSize = AddrModeInsts.size();
2054 // Start a transaction at this point.
2055 // The LHS may match but not the RHS.
2056 // Therefore, we need a higher level restoration point to undo partially
2057 // matched operation.
2058 TypePromotionTransaction::ConstRestorationPt LastKnownGood =
2059 TPT.getRestorationPoint();
2061 if (MatchAddr(AddrInst->getOperand(1), Depth+1) &&
2062 MatchAddr(AddrInst->getOperand(0), Depth+1))
2063 return true;
2065 // Restore the old addr mode info.
2066 AddrMode = BackupAddrMode;
2067 AddrModeInsts.resize(OldSize);
2068 TPT.rollback(LastKnownGood);
2070 // Otherwise this was over-aggressive. Try merging in the LHS then the RHS.
2071 if (MatchAddr(AddrInst->getOperand(0), Depth+1) &&
2072 MatchAddr(AddrInst->getOperand(1), Depth+1))
2073 return true;
2075 // Otherwise we definitely can't merge the ADD in.
2076 AddrMode = BackupAddrMode;
2077 AddrModeInsts.resize(OldSize);
2078 TPT.rollback(LastKnownGood);
2079 break;
2080 }
2081 //case Instruction::Or:
2082 // TODO: We can handle "Or Val, Imm" iff this OR is equivalent to an ADD.
2083 //break;
2084 case Instruction::Mul:
2085 case Instruction::Shl: {
2086 // Can only handle X*C and X << C.
2087 ConstantInt *RHS = dyn_cast<ConstantInt>(AddrInst->getOperand(1));
2088 if (!RHS)
2089 return false;
2090 int64_t Scale = RHS->getSExtValue();
2091 if (Opcode == Instruction::Shl)
2092 Scale = 1LL << Scale;
2094 return MatchScaledValue(AddrInst->getOperand(0), Scale, Depth);
2095 }
2096 case Instruction::GetElementPtr: {
2097 // Scan the GEP. We check it if it contains constant offsets and at most
2098 // one variable offset.
2099 int VariableOperand = -1;
2100 unsigned VariableScale = 0;
2102 int64_t ConstantOffset = 0;
2103 const DataLayout *TD = TLI.getDataLayout();
2104 gep_type_iterator GTI = gep_type_begin(AddrInst);
2105 for (unsigned i = 1, e = AddrInst->getNumOperands(); i != e; ++i, ++GTI) {
2106 if (StructType *STy = dyn_cast<StructType>(*GTI)) {
2107 const StructLayout *SL = TD->getStructLayout(STy);
2108 unsigned Idx =
2109 cast<ConstantInt>(AddrInst->getOperand(i))->getZExtValue();
2110 ConstantOffset += SL->getElementOffset(Idx);
2111 } else {
2112 uint64_t TypeSize = TD->getTypeAllocSize(GTI.getIndexedType());
2113 if (ConstantInt *CI = dyn_cast<ConstantInt>(AddrInst->getOperand(i))) {
2114 ConstantOffset += CI->getSExtValue()*TypeSize;
2115 } else if (TypeSize) { // Scales of zero don't do anything.
2116 // We only allow one variable index at the moment.
2117 if (VariableOperand != -1)
2118 return false;
2120 // Remember the variable index.
2121 VariableOperand = i;
2122 VariableScale = TypeSize;
2123 }
2124 }
2125 }
2127 // A common case is for the GEP to only do a constant offset. In this case,
2128 // just add it to the disp field and check validity.
2129 if (VariableOperand == -1) {
2130 AddrMode.BaseOffs += ConstantOffset;
2131 if (ConstantOffset == 0 || TLI.isLegalAddressingMode(AddrMode, AccessTy)){
2132 // Check to see if we can fold the base pointer in too.
2133 if (MatchAddr(AddrInst->getOperand(0), Depth+1))
2134 return true;
2135 }
2136 AddrMode.BaseOffs -= ConstantOffset;
2137 return false;
2138 }
2140 // Save the valid addressing mode in case we can't match.
2141 ExtAddrMode BackupAddrMode = AddrMode;
2142 unsigned OldSize = AddrModeInsts.size();
2144 // See if the scale and offset amount is valid for this target.
2145 AddrMode.BaseOffs += ConstantOffset;
2147 // Match the base operand of the GEP.
2148 if (!MatchAddr(AddrInst->getOperand(0), Depth+1)) {
2149 // If it couldn't be matched, just stuff the value in a register.
2150 if (AddrMode.HasBaseReg) {
2151 AddrMode = BackupAddrMode;
2152 AddrModeInsts.resize(OldSize);
2153 return false;
2154 }
2155 AddrMode.HasBaseReg = true;
2156 AddrMode.BaseReg = AddrInst->getOperand(0);
2157 }
2159 // Match the remaining variable portion of the GEP.
2160 if (!MatchScaledValue(AddrInst->getOperand(VariableOperand), VariableScale,
2161 Depth)) {
2162 // If it couldn't be matched, try stuffing the base into a register
2163 // instead of matching it, and retrying the match of the scale.
2164 AddrMode = BackupAddrMode;
2165 AddrModeInsts.resize(OldSize);
2166 if (AddrMode.HasBaseReg)
2167 return false;
2168 AddrMode.HasBaseReg = true;
2169 AddrMode.BaseReg = AddrInst->getOperand(0);
2170 AddrMode.BaseOffs += ConstantOffset;
2171 if (!MatchScaledValue(AddrInst->getOperand(VariableOperand),
2172 VariableScale, Depth)) {
2173 // If even that didn't work, bail.
2174 AddrMode = BackupAddrMode;
2175 AddrModeInsts.resize(OldSize);
2176 return false;
2177 }
2178 }
2180 return true;
2181 }
2182 case Instruction::SExt: {
2183 Instruction *SExt = dyn_cast<Instruction>(AddrInst);
2184 if (!SExt)
2185 return false;
2187 // Try to move this sext out of the way of the addressing mode.
2188 // Ask for a method for doing so.
2189 TypePromotionHelper::Action TPH = TypePromotionHelper::getAction(
2190 SExt, InsertedTruncs, TLI, PromotedInsts);
2191 if (!TPH)
2192 return false;
2194 TypePromotionTransaction::ConstRestorationPt LastKnownGood =
2195 TPT.getRestorationPoint();
2196 unsigned CreatedInsts = 0;
2197 Value *PromotedOperand = TPH(SExt, TPT, PromotedInsts, CreatedInsts);
2198 // SExt has been moved away.
2199 // Thus either it will be rematched later in the recursive calls or it is
2200 // gone. Anyway, we must not fold it into the addressing mode at this point.
2201 // E.g.,
2202 // op = add opnd, 1
2203 // idx = sext op
2204 // addr = gep base, idx
2205 // is now:
2206 // promotedOpnd = sext opnd <- no match here
2207 // op = promoted_add promotedOpnd, 1 <- match (later in recursive calls)
2208 // addr = gep base, op <- match
2209 if (MovedAway)
2210 *MovedAway = true;
2212 assert(PromotedOperand &&
2213 "TypePromotionHelper should have filtered out those cases");
2215 ExtAddrMode BackupAddrMode = AddrMode;
2216 unsigned OldSize = AddrModeInsts.size();
2218 if (!MatchAddr(PromotedOperand, Depth) ||
2219 !IsPromotionProfitable(AddrModeInsts.size(), OldSize + CreatedInsts,
2220 PromotedOperand)) {
2221 AddrMode = BackupAddrMode;
2222 AddrModeInsts.resize(OldSize);
2223 DEBUG(dbgs() << "Sign extension does not pay off: rollback\n");
2224 TPT.rollback(LastKnownGood);
2225 return false;
2226 }
2227 return true;
2228 }
2229 }
2230 return false;
2231 }
2233 /// MatchAddr - If we can, try to add the value of 'Addr' into the current
2234 /// addressing mode. If Addr can't be added to AddrMode this returns false and
2235 /// leaves AddrMode unmodified. This assumes that Addr is either a pointer type
2236 /// or intptr_t for the target.
2237 ///
2238 bool AddressingModeMatcher::MatchAddr(Value *Addr, unsigned Depth) {
2239 // Start a transaction at this point that we will rollback if the matching
2240 // fails.
2241 TypePromotionTransaction::ConstRestorationPt LastKnownGood =
2242 TPT.getRestorationPoint();
2243 if (ConstantInt *CI = dyn_cast<ConstantInt>(Addr)) {
2244 // Fold in immediates if legal for the target.
2245 AddrMode.BaseOffs += CI->getSExtValue();
2246 if (TLI.isLegalAddressingMode(AddrMode, AccessTy))
2247 return true;
2248 AddrMode.BaseOffs -= CI->getSExtValue();
2249 } else if (GlobalValue *GV = dyn_cast<GlobalValue>(Addr)) {
2250 // If this is a global variable, try to fold it into the addressing mode.
2251 if (!AddrMode.BaseGV) {
2252 AddrMode.BaseGV = GV;
2253 if (TLI.isLegalAddressingMode(AddrMode, AccessTy))
2254 return true;
2255 AddrMode.BaseGV = nullptr;
2256 }
2257 } else if (Instruction *I = dyn_cast<Instruction>(Addr)) {
2258 ExtAddrMode BackupAddrMode = AddrMode;
2259 unsigned OldSize = AddrModeInsts.size();
2261 // Check to see if it is possible to fold this operation.
2262 bool MovedAway = false;
2263 if (MatchOperationAddr(I, I->getOpcode(), Depth, &MovedAway)) {
2264 // This instruction may have been move away. If so, there is nothing
2265 // to check here.
2266 if (MovedAway)
2267 return true;
2268 // Okay, it's possible to fold this. Check to see if it is actually
2269 // *profitable* to do so. We use a simple cost model to avoid increasing
2270 // register pressure too much.
2271 if (I->hasOneUse() ||
2272 IsProfitableToFoldIntoAddressingMode(I, BackupAddrMode, AddrMode)) {
2273 AddrModeInsts.push_back(I);
2274 return true;
2275 }
2277 // It isn't profitable to do this, roll back.
2278 //cerr << "NOT FOLDING: " << *I;
2279 AddrMode = BackupAddrMode;
2280 AddrModeInsts.resize(OldSize);
2281 TPT.rollback(LastKnownGood);
2282 }
2283 } else if (ConstantExpr *CE = dyn_cast<ConstantExpr>(Addr)) {
2284 if (MatchOperationAddr(CE, CE->getOpcode(), Depth))
2285 return true;
2286 TPT.rollback(LastKnownGood);
2287 } else if (isa<ConstantPointerNull>(Addr)) {
2288 // Null pointer gets folded without affecting the addressing mode.
2289 return true;
2290 }
2292 // Worse case, the target should support [reg] addressing modes. :)
2293 if (!AddrMode.HasBaseReg) {
2294 AddrMode.HasBaseReg = true;
2295 AddrMode.BaseReg = Addr;
2296 // Still check for legality in case the target supports [imm] but not [i+r].
2297 if (TLI.isLegalAddressingMode(AddrMode, AccessTy))
2298 return true;
2299 AddrMode.HasBaseReg = false;
2300 AddrMode.BaseReg = nullptr;
2301 }
2303 // If the base register is already taken, see if we can do [r+r].
2304 if (AddrMode.Scale == 0) {
2305 AddrMode.Scale = 1;
2306 AddrMode.ScaledReg = Addr;
2307 if (TLI.isLegalAddressingMode(AddrMode, AccessTy))
2308 return true;
2309 AddrMode.Scale = 0;
2310 AddrMode.ScaledReg = nullptr;
2311 }
2312 // Couldn't match.
2313 TPT.rollback(LastKnownGood);
2314 return false;
2315 }
2317 /// IsOperandAMemoryOperand - Check to see if all uses of OpVal by the specified
2318 /// inline asm call are due to memory operands. If so, return true, otherwise
2319 /// return false.
2320 static bool IsOperandAMemoryOperand(CallInst *CI, InlineAsm *IA, Value *OpVal,
2321 const TargetLowering &TLI) {
2322 TargetLowering::AsmOperandInfoVector TargetConstraints = TLI.ParseConstraints(ImmutableCallSite(CI));
2323 for (unsigned i = 0, e = TargetConstraints.size(); i != e; ++i) {
2324 TargetLowering::AsmOperandInfo &OpInfo = TargetConstraints[i];
2326 // Compute the constraint code and ConstraintType to use.
2327 TLI.ComputeConstraintToUse(OpInfo, SDValue());
2329 // If this asm operand is our Value*, and if it isn't an indirect memory
2330 // operand, we can't fold it!
2331 if (OpInfo.CallOperandVal == OpVal &&
2332 (OpInfo.ConstraintType != TargetLowering::C_Memory ||
2333 !OpInfo.isIndirect))
2334 return false;
2335 }
2337 return true;
2338 }
2340 /// FindAllMemoryUses - Recursively walk all the uses of I until we find a
2341 /// memory use. If we find an obviously non-foldable instruction, return true.
2342 /// Add the ultimately found memory instructions to MemoryUses.
2343 static bool FindAllMemoryUses(Instruction *I,
2344 SmallVectorImpl<std::pair<Instruction*,unsigned> > &MemoryUses,
2345 SmallPtrSetImpl<Instruction*> &ConsideredInsts,
2346 const TargetLowering &TLI) {
2347 // If we already considered this instruction, we're done.
2348 if (!ConsideredInsts.insert(I))
2349 return false;
2351 // If this is an obviously unfoldable instruction, bail out.
2352 if (!MightBeFoldableInst(I))
2353 return true;
2355 // Loop over all the uses, recursively processing them.
2356 for (Use &U : I->uses()) {
2357 Instruction *UserI = cast<Instruction>(U.getUser());
2359 if (LoadInst *LI = dyn_cast<LoadInst>(UserI)) {
2360 MemoryUses.push_back(std::make_pair(LI, U.getOperandNo()));
2361 continue;
2362 }
2364 if (StoreInst *SI = dyn_cast<StoreInst>(UserI)) {
2365 unsigned opNo = U.getOperandNo();
2366 if (opNo == 0) return true; // Storing addr, not into addr.
2367 MemoryUses.push_back(std::make_pair(SI, opNo));
2368 continue;
2369 }
2371 if (CallInst *CI = dyn_cast<CallInst>(UserI)) {
2372 InlineAsm *IA = dyn_cast<InlineAsm>(CI->getCalledValue());
2373 if (!IA) return true;
2375 // If this is a memory operand, we're cool, otherwise bail out.
2376 if (!IsOperandAMemoryOperand(CI, IA, I, TLI))
2377 return true;
2378 continue;
2379 }
2381 if (FindAllMemoryUses(UserI, MemoryUses, ConsideredInsts, TLI))
2382 return true;
2383 }
2385 return false;
2386 }
2388 /// ValueAlreadyLiveAtInst - Retrn true if Val is already known to be live at
2389 /// the use site that we're folding it into. If so, there is no cost to
2390 /// include it in the addressing mode. KnownLive1 and KnownLive2 are two values
2391 /// that we know are live at the instruction already.
2392 bool AddressingModeMatcher::ValueAlreadyLiveAtInst(Value *Val,Value *KnownLive1,
2393 Value *KnownLive2) {
2394 // If Val is either of the known-live values, we know it is live!
2395 if (Val == nullptr || Val == KnownLive1 || Val == KnownLive2)
2396 return true;
2398 // All values other than instructions and arguments (e.g. constants) are live.
2399 if (!isa<Instruction>(Val) && !isa<Argument>(Val)) return true;
2401 // If Val is a constant sized alloca in the entry block, it is live, this is
2402 // true because it is just a reference to the stack/frame pointer, which is
2403 // live for the whole function.
2404 if (AllocaInst *AI = dyn_cast<AllocaInst>(Val))
2405 if (AI->isStaticAlloca())
2406 return true;
2408 // Check to see if this value is already used in the memory instruction's
2409 // block. If so, it's already live into the block at the very least, so we
2410 // can reasonably fold it.
2411 return Val->isUsedInBasicBlock(MemoryInst->getParent());
2412 }
2414 /// IsProfitableToFoldIntoAddressingMode - It is possible for the addressing
2415 /// mode of the machine to fold the specified instruction into a load or store
2416 /// that ultimately uses it. However, the specified instruction has multiple
2417 /// uses. Given this, it may actually increase register pressure to fold it
2418 /// into the load. For example, consider this code:
2419 ///
2420 /// X = ...
2421 /// Y = X+1
2422 /// use(Y) -> nonload/store
2423 /// Z = Y+1
2424 /// load Z
2425 ///
2426 /// In this case, Y has multiple uses, and can be folded into the load of Z
2427 /// (yielding load [X+2]). However, doing this will cause both "X" and "X+1" to
2428 /// be live at the use(Y) line. If we don't fold Y into load Z, we use one
2429 /// fewer register. Since Y can't be folded into "use(Y)" we don't increase the
2430 /// number of computations either.
2431 ///
2432 /// Note that this (like most of CodeGenPrepare) is just a rough heuristic. If
2433 /// X was live across 'load Z' for other reasons, we actually *would* want to
2434 /// fold the addressing mode in the Z case. This would make Y die earlier.
2435 bool AddressingModeMatcher::
2436 IsProfitableToFoldIntoAddressingMode(Instruction *I, ExtAddrMode &AMBefore,
2437 ExtAddrMode &AMAfter) {
2438 if (IgnoreProfitability) return true;
2440 // AMBefore is the addressing mode before this instruction was folded into it,
2441 // and AMAfter is the addressing mode after the instruction was folded. Get
2442 // the set of registers referenced by AMAfter and subtract out those
2443 // referenced by AMBefore: this is the set of values which folding in this
2444 // address extends the lifetime of.
2445 //
2446 // Note that there are only two potential values being referenced here,
2447 // BaseReg and ScaleReg (global addresses are always available, as are any
2448 // folded immediates).
2449 Value *BaseReg = AMAfter.BaseReg, *ScaledReg = AMAfter.ScaledReg;
2451 // If the BaseReg or ScaledReg was referenced by the previous addrmode, their
2452 // lifetime wasn't extended by adding this instruction.
2453 if (ValueAlreadyLiveAtInst(BaseReg, AMBefore.BaseReg, AMBefore.ScaledReg))
2454 BaseReg = nullptr;
2455 if (ValueAlreadyLiveAtInst(ScaledReg, AMBefore.BaseReg, AMBefore.ScaledReg))
2456 ScaledReg = nullptr;
2458 // If folding this instruction (and it's subexprs) didn't extend any live
2459 // ranges, we're ok with it.
2460 if (!BaseReg && !ScaledReg)
2461 return true;
2463 // If all uses of this instruction are ultimately load/store/inlineasm's,
2464 // check to see if their addressing modes will include this instruction. If
2465 // so, we can fold it into all uses, so it doesn't matter if it has multiple
2466 // uses.
2467 SmallVector<std::pair<Instruction*,unsigned>, 16> MemoryUses;
2468 SmallPtrSet<Instruction*, 16> ConsideredInsts;
2469 if (FindAllMemoryUses(I, MemoryUses, ConsideredInsts, TLI))
2470 return false; // Has a non-memory, non-foldable use!
2472 // Now that we know that all uses of this instruction are part of a chain of
2473 // computation involving only operations that could theoretically be folded
2474 // into a memory use, loop over each of these uses and see if they could
2475 // *actually* fold the instruction.
2476 SmallVector<Instruction*, 32> MatchedAddrModeInsts;
2477 for (unsigned i = 0, e = MemoryUses.size(); i != e; ++i) {
2478 Instruction *User = MemoryUses[i].first;
2479 unsigned OpNo = MemoryUses[i].second;
2481 // Get the access type of this use. If the use isn't a pointer, we don't
2482 // know what it accesses.
2483 Value *Address = User->getOperand(OpNo);
2484 if (!Address->getType()->isPointerTy())
2485 return false;
2486 Type *AddressAccessTy = Address->getType()->getPointerElementType();
2488 // Do a match against the root of this address, ignoring profitability. This
2489 // will tell us if the addressing mode for the memory operation will
2490 // *actually* cover the shared instruction.
2491 ExtAddrMode Result;
2492 TypePromotionTransaction::ConstRestorationPt LastKnownGood =
2493 TPT.getRestorationPoint();
2494 AddressingModeMatcher Matcher(MatchedAddrModeInsts, TLI, AddressAccessTy,
2495 MemoryInst, Result, InsertedTruncs,
2496 PromotedInsts, TPT);
2497 Matcher.IgnoreProfitability = true;
2498 bool Success = Matcher.MatchAddr(Address, 0);
2499 (void)Success; assert(Success && "Couldn't select *anything*?");
2501 // The match was to check the profitability, the changes made are not
2502 // part of the original matcher. Therefore, they should be dropped
2503 // otherwise the original matcher will not present the right state.
2504 TPT.rollback(LastKnownGood);
2506 // If the match didn't cover I, then it won't be shared by it.
2507 if (std::find(MatchedAddrModeInsts.begin(), MatchedAddrModeInsts.end(),
2508 I) == MatchedAddrModeInsts.end())
2509 return false;
2511 MatchedAddrModeInsts.clear();
2512 }
2514 return true;
2515 }
2517 } // end anonymous namespace
2519 /// IsNonLocalValue - Return true if the specified values are defined in a
2520 /// different basic block than BB.
2521 static bool IsNonLocalValue(Value *V, BasicBlock *BB) {
2522 if (Instruction *I = dyn_cast<Instruction>(V))
2523 return I->getParent() != BB;
2524 return false;
2525 }
2527 /// OptimizeMemoryInst - Load and Store Instructions often have
2528 /// addressing modes that can do significant amounts of computation. As such,
2529 /// instruction selection will try to get the load or store to do as much
2530 /// computation as possible for the program. The problem is that isel can only
2531 /// see within a single block. As such, we sink as much legal addressing mode
2532 /// stuff into the block as possible.
2533 ///
2534 /// This method is used to optimize both load/store and inline asms with memory
2535 /// operands.
2536 bool CodeGenPrepare::OptimizeMemoryInst(Instruction *MemoryInst, Value *Addr,
2537 Type *AccessTy) {
2538 Value *Repl = Addr;
2540 // Try to collapse single-value PHI nodes. This is necessary to undo
2541 // unprofitable PRE transformations.
2542 SmallVector<Value*, 8> worklist;
2543 SmallPtrSet<Value*, 16> Visited;
2544 worklist.push_back(Addr);
2546 // Use a worklist to iteratively look through PHI nodes, and ensure that
2547 // the addressing mode obtained from the non-PHI roots of the graph
2548 // are equivalent.
2549 Value *Consensus = nullptr;
2550 unsigned NumUsesConsensus = 0;
2551 bool IsNumUsesConsensusValid = false;
2552 SmallVector<Instruction*, 16> AddrModeInsts;
2553 ExtAddrMode AddrMode;
2554 TypePromotionTransaction TPT;
2555 TypePromotionTransaction::ConstRestorationPt LastKnownGood =
2556 TPT.getRestorationPoint();
2557 while (!worklist.empty()) {
2558 Value *V = worklist.back();
2559 worklist.pop_back();
2561 // Break use-def graph loops.
2562 if (!Visited.insert(V)) {
2563 Consensus = nullptr;
2564 break;
2565 }
2567 // For a PHI node, push all of its incoming values.
2568 if (PHINode *P = dyn_cast<PHINode>(V)) {
2569 for (unsigned i = 0, e = P->getNumIncomingValues(); i != e; ++i)
2570 worklist.push_back(P->getIncomingValue(i));
2571 continue;
2572 }
2574 // For non-PHIs, determine the addressing mode being computed.
2575 SmallVector<Instruction*, 16> NewAddrModeInsts;
2576 ExtAddrMode NewAddrMode = AddressingModeMatcher::Match(
2577 V, AccessTy, MemoryInst, NewAddrModeInsts, *TLI, InsertedTruncsSet,
2578 PromotedInsts, TPT);
2580 // This check is broken into two cases with very similar code to avoid using
2581 // getNumUses() as much as possible. Some values have a lot of uses, so
2582 // calling getNumUses() unconditionally caused a significant compile-time
2583 // regression.
2584 if (!Consensus) {
2585 Consensus = V;
2586 AddrMode = NewAddrMode;
2587 AddrModeInsts = NewAddrModeInsts;
2588 continue;
2589 } else if (NewAddrMode == AddrMode) {
2590 if (!IsNumUsesConsensusValid) {
2591 NumUsesConsensus = Consensus->getNumUses();
2592 IsNumUsesConsensusValid = true;
2593 }
2595 // Ensure that the obtained addressing mode is equivalent to that obtained
2596 // for all other roots of the PHI traversal. Also, when choosing one
2597 // such root as representative, select the one with the most uses in order
2598 // to keep the cost modeling heuristics in AddressingModeMatcher
2599 // applicable.
2600 unsigned NumUses = V->getNumUses();
2601 if (NumUses > NumUsesConsensus) {
2602 Consensus = V;
2603 NumUsesConsensus = NumUses;
2604 AddrModeInsts = NewAddrModeInsts;
2605 }
2606 continue;
2607 }
2609 Consensus = nullptr;
2610 break;
2611 }
2613 // If the addressing mode couldn't be determined, or if multiple different
2614 // ones were determined, bail out now.
2615 if (!Consensus) {
2616 TPT.rollback(LastKnownGood);
2617 return false;
2618 }
2619 TPT.commit();
2621 // Check to see if any of the instructions supersumed by this addr mode are
2622 // non-local to I's BB.
2623 bool AnyNonLocal = false;
2624 for (unsigned i = 0, e = AddrModeInsts.size(); i != e; ++i) {
2625 if (IsNonLocalValue(AddrModeInsts[i], MemoryInst->getParent())) {
2626 AnyNonLocal = true;
2627 break;
2628 }
2629 }
2631 // If all the instructions matched are already in this BB, don't do anything.
2632 if (!AnyNonLocal) {
2633 DEBUG(dbgs() << "CGP: Found local addrmode: " << AddrMode << "\n");
2634 return false;
2635 }
2637 // Insert this computation right after this user. Since our caller is
2638 // scanning from the top of the BB to the bottom, reuse of the expr are
2639 // guaranteed to happen later.
2640 IRBuilder<> Builder(MemoryInst);
2642 // Now that we determined the addressing expression we want to use and know
2643 // that we have to sink it into this block. Check to see if we have already
2644 // done this for some other load/store instr in this block. If so, reuse the
2645 // computation.
2646 Value *&SunkAddr = SunkAddrs[Addr];
2647 if (SunkAddr) {
2648 DEBUG(dbgs() << "CGP: Reusing nonlocal addrmode: " << AddrMode << " for "
2649 << *MemoryInst << "\n");
2650 if (SunkAddr->getType() != Addr->getType())
2651 SunkAddr = Builder.CreateBitCast(SunkAddr, Addr->getType());
2652 } else if (AddrSinkUsingGEPs || (!AddrSinkUsingGEPs.getNumOccurrences() &&
2653 TM && TM->getSubtarget<TargetSubtargetInfo>().useAA())) {
2654 // By default, we use the GEP-based method when AA is used later. This
2655 // prevents new inttoptr/ptrtoint pairs from degrading AA capabilities.
2656 DEBUG(dbgs() << "CGP: SINKING nonlocal addrmode: " << AddrMode << " for "
2657 << *MemoryInst << "\n");
2658 Type *IntPtrTy = TLI->getDataLayout()->getIntPtrType(Addr->getType());
2659 Value *ResultPtr = nullptr, *ResultIndex = nullptr;
2661 // First, find the pointer.
2662 if (AddrMode.BaseReg && AddrMode.BaseReg->getType()->isPointerTy()) {
2663 ResultPtr = AddrMode.BaseReg;
2664 AddrMode.BaseReg = nullptr;
2665 }
2667 if (AddrMode.Scale && AddrMode.ScaledReg->getType()->isPointerTy()) {
2668 // We can't add more than one pointer together, nor can we scale a
2669 // pointer (both of which seem meaningless).
2670 if (ResultPtr || AddrMode.Scale != 1)
2671 return false;
2673 ResultPtr = AddrMode.ScaledReg;
2674 AddrMode.Scale = 0;
2675 }
2677 if (AddrMode.BaseGV) {
2678 if (ResultPtr)
2679 return false;
2681 ResultPtr = AddrMode.BaseGV;
2682 }
2684 // If the real base value actually came from an inttoptr, then the matcher
2685 // will look through it and provide only the integer value. In that case,
2686 // use it here.
2687 if (!ResultPtr && AddrMode.BaseReg) {
2688 ResultPtr =
2689 Builder.CreateIntToPtr(AddrMode.BaseReg, Addr->getType(), "sunkaddr");
2690 AddrMode.BaseReg = nullptr;
2691 } else if (!ResultPtr && AddrMode.Scale == 1) {
2692 ResultPtr =
2693 Builder.CreateIntToPtr(AddrMode.ScaledReg, Addr->getType(), "sunkaddr");
2694 AddrMode.Scale = 0;
2695 }
2697 if (!ResultPtr &&
2698 !AddrMode.BaseReg && !AddrMode.Scale && !AddrMode.BaseOffs) {
2699 SunkAddr = Constant::getNullValue(Addr->getType());
2700 } else if (!ResultPtr) {
2701 return false;
2702 } else {
2703 Type *I8PtrTy =
2704 Builder.getInt8PtrTy(Addr->getType()->getPointerAddressSpace());
2706 // Start with the base register. Do this first so that subsequent address
2707 // matching finds it last, which will prevent it from trying to match it
2708 // as the scaled value in case it happens to be a mul. That would be
2709 // problematic if we've sunk a different mul for the scale, because then
2710 // we'd end up sinking both muls.
2711 if (AddrMode.BaseReg) {
2712 Value *V = AddrMode.BaseReg;
2713 if (V->getType() != IntPtrTy)
2714 V = Builder.CreateIntCast(V, IntPtrTy, /*isSigned=*/true, "sunkaddr");
2716 ResultIndex = V;
2717 }
2719 // Add the scale value.
2720 if (AddrMode.Scale) {
2721 Value *V = AddrMode.ScaledReg;
2722 if (V->getType() == IntPtrTy) {
2723 // done.
2724 } else if (cast<IntegerType>(IntPtrTy)->getBitWidth() <
2725 cast<IntegerType>(V->getType())->getBitWidth()) {
2726 V = Builder.CreateTrunc(V, IntPtrTy, "sunkaddr");
2727 } else {
2728 // It is only safe to sign extend the BaseReg if we know that the math
2729 // required to create it did not overflow before we extend it. Since
2730 // the original IR value was tossed in favor of a constant back when
2731 // the AddrMode was created we need to bail out gracefully if widths
2732 // do not match instead of extending it.
2733 Instruction *I = dyn_cast_or_null<Instruction>(ResultIndex);
2734 if (I && (ResultIndex != AddrMode.BaseReg))
2735 I->eraseFromParent();
2736 return false;
2737 }
2739 if (AddrMode.Scale != 1)
2740 V = Builder.CreateMul(V, ConstantInt::get(IntPtrTy, AddrMode.Scale),
2741 "sunkaddr");
2742 if (ResultIndex)
2743 ResultIndex = Builder.CreateAdd(ResultIndex, V, "sunkaddr");
2744 else
2745 ResultIndex = V;
2746 }
2748 // Add in the Base Offset if present.
2749 if (AddrMode.BaseOffs) {
2750 Value *V = ConstantInt::get(IntPtrTy, AddrMode.BaseOffs);
2751 if (ResultIndex) {
2752 // We need to add this separately from the scale above to help with
2753 // SDAG consecutive load/store merging.
2754 if (ResultPtr->getType() != I8PtrTy)
2755 ResultPtr = Builder.CreateBitCast(ResultPtr, I8PtrTy);
2756 ResultPtr = Builder.CreateGEP(ResultPtr, ResultIndex, "sunkaddr");
2757 }
2759 ResultIndex = V;
2760 }
2762 if (!ResultIndex) {
2763 SunkAddr = ResultPtr;
2764 } else {
2765 if (ResultPtr->getType() != I8PtrTy)
2766 ResultPtr = Builder.CreateBitCast(ResultPtr, I8PtrTy);
2767 SunkAddr = Builder.CreateGEP(ResultPtr, ResultIndex, "sunkaddr");
2768 }
2770 if (SunkAddr->getType() != Addr->getType())
2771 SunkAddr = Builder.CreateBitCast(SunkAddr, Addr->getType());
2772 }
2773 } else {
2774 DEBUG(dbgs() << "CGP: SINKING nonlocal addrmode: " << AddrMode << " for "
2775 << *MemoryInst << "\n");
2776 Type *IntPtrTy = TLI->getDataLayout()->getIntPtrType(Addr->getType());
2777 Value *Result = nullptr;
2779 // Start with the base register. Do this first so that subsequent address
2780 // matching finds it last, which will prevent it from trying to match it
2781 // as the scaled value in case it happens to be a mul. That would be
2782 // problematic if we've sunk a different mul for the scale, because then
2783 // we'd end up sinking both muls.
2784 if (AddrMode.BaseReg) {
2785 Value *V = AddrMode.BaseReg;
2786 if (V->getType()->isPointerTy())
2787 V = Builder.CreatePtrToInt(V, IntPtrTy, "sunkaddr");
2788 if (V->getType() != IntPtrTy)
2789 V = Builder.CreateIntCast(V, IntPtrTy, /*isSigned=*/true, "sunkaddr");
2790 Result = V;
2791 }
2793 // Add the scale value.
2794 if (AddrMode.Scale) {
2795 Value *V = AddrMode.ScaledReg;
2796 if (V->getType() == IntPtrTy) {
2797 // done.
2798 } else if (V->getType()->isPointerTy()) {
2799 V = Builder.CreatePtrToInt(V, IntPtrTy, "sunkaddr");
2800 } else if (cast<IntegerType>(IntPtrTy)->getBitWidth() <
2801 cast<IntegerType>(V->getType())->getBitWidth()) {
2802 V = Builder.CreateTrunc(V, IntPtrTy, "sunkaddr");
2803 } else {
2804 // It is only safe to sign extend the BaseReg if we know that the math
2805 // required to create it did not overflow before we extend it. Since
2806 // the original IR value was tossed in favor of a constant back when
2807 // the AddrMode was created we need to bail out gracefully if widths
2808 // do not match instead of extending it.
2809 Instruction *I = dyn_cast_or_null<Instruction>(Result);
2810 if (I && (Result != AddrMode.BaseReg))
2811 I->eraseFromParent();
2812 return false;
2813 }
2814 if (AddrMode.Scale != 1)
2815 V = Builder.CreateMul(V, ConstantInt::get(IntPtrTy, AddrMode.Scale),
2816 "sunkaddr");
2817 if (Result)
2818 Result = Builder.CreateAdd(Result, V, "sunkaddr");
2819 else
2820 Result = V;
2821 }
2823 // Add in the BaseGV if present.
2824 if (AddrMode.BaseGV) {
2825 Value *V = Builder.CreatePtrToInt(AddrMode.BaseGV, IntPtrTy, "sunkaddr");
2826 if (Result)
2827 Result = Builder.CreateAdd(Result, V, "sunkaddr");
2828 else
2829 Result = V;
2830 }
2832 // Add in the Base Offset if present.
2833 if (AddrMode.BaseOffs) {
2834 Value *V = ConstantInt::get(IntPtrTy, AddrMode.BaseOffs);
2835 if (Result)
2836 Result = Builder.CreateAdd(Result, V, "sunkaddr");
2837 else
2838 Result = V;
2839 }
2841 if (!Result)
2842 SunkAddr = Constant::getNullValue(Addr->getType());
2843 else
2844 SunkAddr = Builder.CreateIntToPtr(Result, Addr->getType(), "sunkaddr");
2845 }
2847 MemoryInst->replaceUsesOfWith(Repl, SunkAddr);
2849 // If we have no uses, recursively delete the value and all dead instructions
2850 // using it.
2851 if (Repl->use_empty()) {
2852 // This can cause recursive deletion, which can invalidate our iterator.
2853 // Use a WeakVH to hold onto it in case this happens.
2854 WeakVH IterHandle(CurInstIterator);
2855 BasicBlock *BB = CurInstIterator->getParent();
2857 RecursivelyDeleteTriviallyDeadInstructions(Repl, TLInfo);
2859 if (IterHandle != CurInstIterator) {
2860 // If the iterator instruction was recursively deleted, start over at the
2861 // start of the block.
2862 CurInstIterator = BB->begin();
2863 SunkAddrs.clear();
2864 }
2865 }
2866 ++NumMemoryInsts;
2867 return true;
2868 }
2870 /// OptimizeInlineAsmInst - If there are any memory operands, use
2871 /// OptimizeMemoryInst to sink their address computing into the block when
2872 /// possible / profitable.
2873 bool CodeGenPrepare::OptimizeInlineAsmInst(CallInst *CS) {
2874 bool MadeChange = false;
2876 TargetLowering::AsmOperandInfoVector
2877 TargetConstraints = TLI->ParseConstraints(CS);
2878 unsigned ArgNo = 0;
2879 for (unsigned i = 0, e = TargetConstraints.size(); i != e; ++i) {
2880 TargetLowering::AsmOperandInfo &OpInfo = TargetConstraints[i];
2882 // Compute the constraint code and ConstraintType to use.
2883 TLI->ComputeConstraintToUse(OpInfo, SDValue());
2885 if (OpInfo.ConstraintType == TargetLowering::C_Memory &&
2886 OpInfo.isIndirect) {
2887 Value *OpVal = CS->getArgOperand(ArgNo++);
2888 MadeChange |= OptimizeMemoryInst(CS, OpVal, OpVal->getType());
2889 } else if (OpInfo.Type == InlineAsm::isInput)
2890 ArgNo++;
2891 }
2893 return MadeChange;
2894 }
2896 /// MoveExtToFormExtLoad - Move a zext or sext fed by a load into the same
2897 /// basic block as the load, unless conditions are unfavorable. This allows
2898 /// SelectionDAG to fold the extend into the load.
2899 ///
2900 bool CodeGenPrepare::MoveExtToFormExtLoad(Instruction *I) {
2901 // Look for a load being extended.
2902 LoadInst *LI = dyn_cast<LoadInst>(I->getOperand(0));
2903 if (!LI) return false;
2905 // If they're already in the same block, there's nothing to do.
2906 if (LI->getParent() == I->getParent())
2907 return false;
2909 // If the load has other users and the truncate is not free, this probably
2910 // isn't worthwhile.
2911 if (!LI->hasOneUse() &&
2912 TLI && (TLI->isTypeLegal(TLI->getValueType(LI->getType())) ||
2913 !TLI->isTypeLegal(TLI->getValueType(I->getType()))) &&
2914 !TLI->isTruncateFree(I->getType(), LI->getType()))
2915 return false;
2917 // Check whether the target supports casts folded into loads.
2918 unsigned LType;
2919 if (isa<ZExtInst>(I))
2920 LType = ISD::ZEXTLOAD;
2921 else {
2922 assert(isa<SExtInst>(I) && "Unexpected ext type!");
2923 LType = ISD::SEXTLOAD;
2924 }
2925 if (TLI && !TLI->isLoadExtLegal(LType, TLI->getValueType(LI->getType())))
2926 return false;
2928 // Move the extend into the same block as the load, so that SelectionDAG
2929 // can fold it.
2930 I->removeFromParent();
2931 I->insertAfter(LI);
2932 ++NumExtsMoved;
2933 return true;
2934 }
2936 bool CodeGenPrepare::OptimizeExtUses(Instruction *I) {
2937 BasicBlock *DefBB = I->getParent();
2939 // If the result of a {s|z}ext and its source are both live out, rewrite all
2940 // other uses of the source with result of extension.
2941 Value *Src = I->getOperand(0);
2942 if (Src->hasOneUse())
2943 return false;
2945 // Only do this xform if truncating is free.
2946 if (TLI && !TLI->isTruncateFree(I->getType(), Src->getType()))
2947 return false;
2949 // Only safe to perform the optimization if the source is also defined in
2950 // this block.
2951 if (!isa<Instruction>(Src) || DefBB != cast<Instruction>(Src)->getParent())
2952 return false;
2954 bool DefIsLiveOut = false;
2955 for (User *U : I->users()) {
2956 Instruction *UI = cast<Instruction>(U);
2958 // Figure out which BB this ext is used in.
2959 BasicBlock *UserBB = UI->getParent();
2960 if (UserBB == DefBB) continue;
2961 DefIsLiveOut = true;
2962 break;
2963 }
2964 if (!DefIsLiveOut)
2965 return false;
2967 // Make sure none of the uses are PHI nodes.
2968 for (User *U : Src->users()) {
2969 Instruction *UI = cast<Instruction>(U);
2970 BasicBlock *UserBB = UI->getParent();
2971 if (UserBB == DefBB) continue;
2972 // Be conservative. We don't want this xform to end up introducing
2973 // reloads just before load / store instructions.
2974 if (isa<PHINode>(UI) || isa<LoadInst>(UI) || isa<StoreInst>(UI))
2975 return false;
2976 }
2978 // InsertedTruncs - Only insert one trunc in each block once.
2979 DenseMap<BasicBlock*, Instruction*> InsertedTruncs;
2981 bool MadeChange = false;
2982 for (Use &U : Src->uses()) {
2983 Instruction *User = cast<Instruction>(U.getUser());
2985 // Figure out which BB this ext is used in.
2986 BasicBlock *UserBB = User->getParent();
2987 if (UserBB == DefBB) continue;
2989 // Both src and def are live in this block. Rewrite the use.
2990 Instruction *&InsertedTrunc = InsertedTruncs[UserBB];
2992 if (!InsertedTrunc) {
2993 BasicBlock::iterator InsertPt = UserBB->getFirstInsertionPt();
2994 InsertedTrunc = new TruncInst(I, Src->getType(), "", InsertPt);
2995 InsertedTruncsSet.insert(InsertedTrunc);
2996 }
2998 // Replace a use of the {s|z}ext source with a use of the result.
2999 U = InsertedTrunc;
3000 ++NumExtUses;
3001 MadeChange = true;
3002 }
3004 return MadeChange;
3005 }
3007 /// isFormingBranchFromSelectProfitable - Returns true if a SelectInst should be
3008 /// turned into an explicit branch.
3009 static bool isFormingBranchFromSelectProfitable(SelectInst *SI) {
3010 // FIXME: This should use the same heuristics as IfConversion to determine
3011 // whether a select is better represented as a branch. This requires that
3012 // branch probability metadata is preserved for the select, which is not the
3013 // case currently.
3015 CmpInst *Cmp = dyn_cast<CmpInst>(SI->getCondition());
3017 // If the branch is predicted right, an out of order CPU can avoid blocking on
3018 // the compare. Emit cmovs on compares with a memory operand as branches to
3019 // avoid stalls on the load from memory. If the compare has more than one use
3020 // there's probably another cmov or setcc around so it's not worth emitting a
3021 // branch.
3022 if (!Cmp)
3023 return false;
3025 Value *CmpOp0 = Cmp->getOperand(0);
3026 Value *CmpOp1 = Cmp->getOperand(1);
3028 // We check that the memory operand has one use to avoid uses of the loaded
3029 // value directly after the compare, making branches unprofitable.
3030 return Cmp->hasOneUse() &&
3031 ((isa<LoadInst>(CmpOp0) && CmpOp0->hasOneUse()) ||
3032 (isa<LoadInst>(CmpOp1) && CmpOp1->hasOneUse()));
3033 }
3036 /// If we have a SelectInst that will likely profit from branch prediction,
3037 /// turn it into a branch.
3038 bool CodeGenPrepare::OptimizeSelectInst(SelectInst *SI) {
3039 bool VectorCond = !SI->getCondition()->getType()->isIntegerTy(1);
3041 // Can we convert the 'select' to CF ?
3042 if (DisableSelectToBranch || OptSize || !TLI || VectorCond)
3043 return false;
3045 TargetLowering::SelectSupportKind SelectKind;
3046 if (VectorCond)
3047 SelectKind = TargetLowering::VectorMaskSelect;
3048 else if (SI->getType()->isVectorTy())
3049 SelectKind = TargetLowering::ScalarCondVectorVal;
3050 else
3051 SelectKind = TargetLowering::ScalarValSelect;
3053 // Do we have efficient codegen support for this kind of 'selects' ?
3054 if (TLI->isSelectSupported(SelectKind)) {
3055 // We have efficient codegen support for the select instruction.
3056 // Check if it is profitable to keep this 'select'.
3057 if (!TLI->isPredictableSelectExpensive() ||
3058 !isFormingBranchFromSelectProfitable(SI))
3059 return false;
3060 }
3062 ModifiedDT = true;
3064 // First, we split the block containing the select into 2 blocks.
3065 BasicBlock *StartBlock = SI->getParent();
3066 BasicBlock::iterator SplitPt = ++(BasicBlock::iterator(SI));
3067 BasicBlock *NextBlock = StartBlock->splitBasicBlock(SplitPt, "select.end");
3069 // Create a new block serving as the landing pad for the branch.
3070 BasicBlock *SmallBlock = BasicBlock::Create(SI->getContext(), "select.mid",
3071 NextBlock->getParent(), NextBlock);
3073 // Move the unconditional branch from the block with the select in it into our
3074 // landing pad block.
3075 StartBlock->getTerminator()->eraseFromParent();
3076 BranchInst::Create(NextBlock, SmallBlock);
3078 // Insert the real conditional branch based on the original condition.
3079 BranchInst::Create(NextBlock, SmallBlock, SI->getCondition(), SI);
3081 // The select itself is replaced with a PHI Node.
3082 PHINode *PN = PHINode::Create(SI->getType(), 2, "", NextBlock->begin());
3083 PN->takeName(SI);
3084 PN->addIncoming(SI->getTrueValue(), StartBlock);
3085 PN->addIncoming(SI->getFalseValue(), SmallBlock);
3086 SI->replaceAllUsesWith(PN);
3087 SI->eraseFromParent();
3089 // Instruct OptimizeBlock to skip to the next block.
3090 CurInstIterator = StartBlock->end();
3091 ++NumSelectsExpanded;
3092 return true;
3093 }
3095 static bool isBroadcastShuffle(ShuffleVectorInst *SVI) {
3096 SmallVector<int, 16> Mask(SVI->getShuffleMask());
3097 int SplatElem = -1;
3098 for (unsigned i = 0; i < Mask.size(); ++i) {
3099 if (SplatElem != -1 && Mask[i] != -1 && Mask[i] != SplatElem)
3100 return false;
3101 SplatElem = Mask[i];
3102 }
3104 return true;
3105 }
3107 /// Some targets have expensive vector shifts if the lanes aren't all the same
3108 /// (e.g. x86 only introduced "vpsllvd" and friends with AVX2). In these cases
3109 /// it's often worth sinking a shufflevector splat down to its use so that
3110 /// codegen can spot all lanes are identical.
3111 bool CodeGenPrepare::OptimizeShuffleVectorInst(ShuffleVectorInst *SVI) {
3112 BasicBlock *DefBB = SVI->getParent();
3114 // Only do this xform if variable vector shifts are particularly expensive.
3115 if (!TLI || !TLI->isVectorShiftByScalarCheap(SVI->getType()))
3116 return false;
3118 // We only expect better codegen by sinking a shuffle if we can recognise a
3119 // constant splat.
3120 if (!isBroadcastShuffle(SVI))
3121 return false;
3123 // InsertedShuffles - Only insert a shuffle in each block once.
3124 DenseMap<BasicBlock*, Instruction*> InsertedShuffles;
3126 bool MadeChange = false;
3127 for (User *U : SVI->users()) {
3128 Instruction *UI = cast<Instruction>(U);
3130 // Figure out which BB this ext is used in.
3131 BasicBlock *UserBB = UI->getParent();
3132 if (UserBB == DefBB) continue;
3134 // For now only apply this when the splat is used by a shift instruction.
3135 if (!UI->isShift()) continue;
3137 // Everything checks out, sink the shuffle if the user's block doesn't
3138 // already have a copy.
3139 Instruction *&InsertedShuffle = InsertedShuffles[UserBB];
3141 if (!InsertedShuffle) {
3142 BasicBlock::iterator InsertPt = UserBB->getFirstInsertionPt();
3143 InsertedShuffle = new ShuffleVectorInst(SVI->getOperand(0),
3144 SVI->getOperand(1),
3145 SVI->getOperand(2), "", InsertPt);
3146 }
3148 UI->replaceUsesOfWith(SVI, InsertedShuffle);
3149 MadeChange = true;
3150 }
3152 // If we removed all uses, nuke the shuffle.
3153 if (SVI->use_empty()) {
3154 SVI->eraseFromParent();
3155 MadeChange = true;
3156 }
3158 return MadeChange;
3159 }
3161 bool CodeGenPrepare::OptimizeInst(Instruction *I) {
3162 if (PHINode *P = dyn_cast<PHINode>(I)) {
3163 // It is possible for very late stage optimizations (such as SimplifyCFG)
3164 // to introduce PHI nodes too late to be cleaned up. If we detect such a
3165 // trivial PHI, go ahead and zap it here.
3166 if (Value *V = SimplifyInstruction(P, TLI ? TLI->getDataLayout() : nullptr,
3167 TLInfo, DT)) {
3168 P->replaceAllUsesWith(V);
3169 P->eraseFromParent();
3170 ++NumPHIsElim;
3171 return true;
3172 }
3173 return false;
3174 }
3176 if (CastInst *CI = dyn_cast<CastInst>(I)) {
3177 // If the source of the cast is a constant, then this should have
3178 // already been constant folded. The only reason NOT to constant fold
3179 // it is if something (e.g. LSR) was careful to place the constant
3180 // evaluation in a block other than then one that uses it (e.g. to hoist
3181 // the address of globals out of a loop). If this is the case, we don't
3182 // want to forward-subst the cast.
3183 if (isa<Constant>(CI->getOperand(0)))
3184 return false;
3186 if (TLI && OptimizeNoopCopyExpression(CI, *TLI))
3187 return true;
3189 if (isa<ZExtInst>(I) || isa<SExtInst>(I)) {
3190 /// Sink a zext or sext into its user blocks if the target type doesn't
3191 /// fit in one register
3192 if (TLI && TLI->getTypeAction(CI->getContext(),
3193 TLI->getValueType(CI->getType())) ==
3194 TargetLowering::TypeExpandInteger) {
3195 return SinkCast(CI);
3196 } else {
3197 bool MadeChange = MoveExtToFormExtLoad(I);
3198 return MadeChange | OptimizeExtUses(I);
3199 }
3200 }
3201 return false;
3202 }
3204 if (CmpInst *CI = dyn_cast<CmpInst>(I))
3205 if (!TLI || !TLI->hasMultipleConditionRegisters())
3206 return OptimizeCmpExpression(CI);
3208 if (LoadInst *LI = dyn_cast<LoadInst>(I)) {
3209 if (TLI)
3210 return OptimizeMemoryInst(I, I->getOperand(0), LI->getType());
3211 return false;
3212 }
3214 if (StoreInst *SI = dyn_cast<StoreInst>(I)) {
3215 if (TLI)
3216 return OptimizeMemoryInst(I, SI->getOperand(1),
3217 SI->getOperand(0)->getType());
3218 return false;
3219 }
3221 BinaryOperator *BinOp = dyn_cast<BinaryOperator>(I);
3223 if (BinOp && (BinOp->getOpcode() == Instruction::AShr ||
3224 BinOp->getOpcode() == Instruction::LShr)) {
3225 ConstantInt *CI = dyn_cast<ConstantInt>(BinOp->getOperand(1));
3226 if (TLI && CI && TLI->hasExtractBitsInsn())
3227 return OptimizeExtractBits(BinOp, CI, *TLI);
3229 return false;
3230 }
3232 if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(I)) {
3233 if (GEPI->hasAllZeroIndices()) {
3234 /// The GEP operand must be a pointer, so must its result -> BitCast
3235 Instruction *NC = new BitCastInst(GEPI->getOperand(0), GEPI->getType(),
3236 GEPI->getName(), GEPI);
3237 GEPI->replaceAllUsesWith(NC);
3238 GEPI->eraseFromParent();
3239 ++NumGEPsElim;
3240 OptimizeInst(NC);
3241 return true;
3242 }
3243 return false;
3244 }
3246 if (CallInst *CI = dyn_cast<CallInst>(I))
3247 return OptimizeCallInst(CI);
3249 if (SelectInst *SI = dyn_cast<SelectInst>(I))
3250 return OptimizeSelectInst(SI);
3252 if (ShuffleVectorInst *SVI = dyn_cast<ShuffleVectorInst>(I))
3253 return OptimizeShuffleVectorInst(SVI);
3255 return false;
3256 }
3258 // In this pass we look for GEP and cast instructions that are used
3259 // across basic blocks and rewrite them to improve basic-block-at-a-time
3260 // selection.
3261 bool CodeGenPrepare::OptimizeBlock(BasicBlock &BB) {
3262 SunkAddrs.clear();
3263 bool MadeChange = false;
3265 CurInstIterator = BB.begin();
3266 while (CurInstIterator != BB.end())
3267 MadeChange |= OptimizeInst(CurInstIterator++);
3269 MadeChange |= DupRetToEnableTailCallOpts(&BB);
3271 return MadeChange;
3272 }
3274 // llvm.dbg.value is far away from the value then iSel may not be able
3275 // handle it properly. iSel will drop llvm.dbg.value if it can not
3276 // find a node corresponding to the value.
3277 bool CodeGenPrepare::PlaceDbgValues(Function &F) {
3278 bool MadeChange = false;
3279 for (Function::iterator I = F.begin(), E = F.end(); I != E; ++I) {
3280 Instruction *PrevNonDbgInst = nullptr;
3281 for (BasicBlock::iterator BI = I->begin(), BE = I->end(); BI != BE;) {
3282 Instruction *Insn = BI; ++BI;
3283 DbgValueInst *DVI = dyn_cast<DbgValueInst>(Insn);
3284 // Leave dbg.values that refer to an alloca alone. These
3285 // instrinsics describe the address of a variable (= the alloca)
3286 // being taken. They should not be moved next to the alloca
3287 // (and to the beginning of the scope), but rather stay close to
3288 // where said address is used.
3289 if (!DVI || (DVI->getValue() && isa<AllocaInst>(DVI->getValue()))) {
3290 PrevNonDbgInst = Insn;
3291 continue;
3292 }
3294 Instruction *VI = dyn_cast_or_null<Instruction>(DVI->getValue());
3295 if (VI && VI != PrevNonDbgInst && !VI->isTerminator()) {
3296 DEBUG(dbgs() << "Moving Debug Value before :\n" << *DVI << ' ' << *VI);
3297 DVI->removeFromParent();
3298 if (isa<PHINode>(VI))
3299 DVI->insertBefore(VI->getParent()->getFirstInsertionPt());
3300 else
3301 DVI->insertAfter(VI);
3302 MadeChange = true;
3303 ++NumDbgValueMoved;
3304 }
3305 }
3306 }
3307 return MadeChange;
3308 }
3310 // If there is a sequence that branches based on comparing a single bit
3311 // against zero that can be combined into a single instruction, and the
3312 // target supports folding these into a single instruction, sink the
3313 // mask and compare into the branch uses. Do this before OptimizeBlock ->
3314 // OptimizeInst -> OptimizeCmpExpression, which perturbs the pattern being
3315 // searched for.
3316 bool CodeGenPrepare::sinkAndCmp(Function &F) {
3317 if (!EnableAndCmpSinking)
3318 return false;
3319 if (!TLI || !TLI->isMaskAndBranchFoldingLegal())
3320 return false;
3321 bool MadeChange = false;
3322 for (Function::iterator I = F.begin(), E = F.end(); I != E; ) {
3323 BasicBlock *BB = I++;
3325 // Does this BB end with the following?
3326 // %andVal = and %val, #single-bit-set
3327 // %icmpVal = icmp %andResult, 0
3328 // br i1 %cmpVal label %dest1, label %dest2"
3329 BranchInst *Brcc = dyn_cast<BranchInst>(BB->getTerminator());
3330 if (!Brcc || !Brcc->isConditional())
3331 continue;
3332 ICmpInst *Cmp = dyn_cast<ICmpInst>(Brcc->getOperand(0));
3333 if (!Cmp || Cmp->getParent() != BB)
3334 continue;
3335 ConstantInt *Zero = dyn_cast<ConstantInt>(Cmp->getOperand(1));
3336 if (!Zero || !Zero->isZero())
3337 continue;
3338 Instruction *And = dyn_cast<Instruction>(Cmp->getOperand(0));
3339 if (!And || And->getOpcode() != Instruction::And || And->getParent() != BB)
3340 continue;
3341 ConstantInt* Mask = dyn_cast<ConstantInt>(And->getOperand(1));
3342 if (!Mask || !Mask->getUniqueInteger().isPowerOf2())
3343 continue;
3344 DEBUG(dbgs() << "found and; icmp ?,0; brcc\n"); DEBUG(BB->dump());
3346 // Push the "and; icmp" for any users that are conditional branches.
3347 // Since there can only be one branch use per BB, we don't need to keep
3348 // track of which BBs we insert into.
3349 for (Value::use_iterator UI = Cmp->use_begin(), E = Cmp->use_end();
3350 UI != E; ) {
3351 Use &TheUse = *UI;
3352 // Find brcc use.
3353 BranchInst *BrccUser = dyn_cast<BranchInst>(*UI);
3354 ++UI;
3355 if (!BrccUser || !BrccUser->isConditional())
3356 continue;
3357 BasicBlock *UserBB = BrccUser->getParent();
3358 if (UserBB == BB) continue;
3359 DEBUG(dbgs() << "found Brcc use\n");
3361 // Sink the "and; icmp" to use.
3362 MadeChange = true;
3363 BinaryOperator *NewAnd =
3364 BinaryOperator::CreateAnd(And->getOperand(0), And->getOperand(1), "",
3365 BrccUser);
3366 CmpInst *NewCmp =
3367 CmpInst::Create(Cmp->getOpcode(), Cmp->getPredicate(), NewAnd, Zero,
3368 "", BrccUser);
3369 TheUse = NewCmp;
3370 ++NumAndCmpsMoved;
3371 DEBUG(BrccUser->getParent()->dump());
3372 }
3373 }
3374 return MadeChange;
3375 }