LLVM  6.0.0svn
CodeGenPrepare.cpp
Go to the documentation of this file.
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 //===----------------------------------------------------------------------===//
15 
16 #include "llvm/ADT/APInt.h"
17 #include "llvm/ADT/ArrayRef.h"
18 #include "llvm/ADT/DenseMap.h"
20 #include "llvm/ADT/STLExtras.h"
21 #include "llvm/ADT/SetVector.h"
22 #include "llvm/ADT/SmallPtrSet.h"
23 #include "llvm/ADT/SmallVector.h"
24 #include "llvm/ADT/Statistic.h"
29 #include "llvm/Analysis/LoopInfo.h"
35 #include "llvm/CodeGen/Analysis.h"
43 #include "llvm/IR/Argument.h"
44 #include "llvm/IR/Attributes.h"
45 #include "llvm/IR/BasicBlock.h"
46 #include "llvm/IR/CallSite.h"
47 #include "llvm/IR/Constant.h"
48 #include "llvm/IR/Constants.h"
49 #include "llvm/IR/DataLayout.h"
50 #include "llvm/IR/DerivedTypes.h"
51 #include "llvm/IR/Dominators.h"
52 #include "llvm/IR/Function.h"
54 #include "llvm/IR/GlobalValue.h"
55 #include "llvm/IR/GlobalVariable.h"
56 #include "llvm/IR/IRBuilder.h"
57 #include "llvm/IR/InlineAsm.h"
58 #include "llvm/IR/InstrTypes.h"
59 #include "llvm/IR/Instruction.h"
60 #include "llvm/IR/Instructions.h"
61 #include "llvm/IR/IntrinsicInst.h"
62 #include "llvm/IR/Intrinsics.h"
63 #include "llvm/IR/LLVMContext.h"
64 #include "llvm/IR/MDBuilder.h"
65 #include "llvm/IR/Module.h"
66 #include "llvm/IR/Operator.h"
67 #include "llvm/IR/PatternMatch.h"
68 #include "llvm/IR/Statepoint.h"
69 #include "llvm/IR/Type.h"
70 #include "llvm/IR/Use.h"
71 #include "llvm/IR/User.h"
72 #include "llvm/IR/Value.h"
73 #include "llvm/IR/ValueHandle.h"
74 #include "llvm/IR/ValueMap.h"
75 #include "llvm/Pass.h"
78 #include "llvm/Support/Casting.h"
80 #include "llvm/Support/Compiler.h"
81 #include "llvm/Support/Debug.h"
93 #include <algorithm>
94 #include <cassert>
95 #include <cstdint>
96 #include <iterator>
97 #include <limits>
98 #include <memory>
99 #include <utility>
100 #include <vector>
101 
102 using namespace llvm;
103 using namespace llvm::PatternMatch;
104 
105 #define DEBUG_TYPE "codegenprepare"
106 
107 STATISTIC(NumBlocksElim, "Number of blocks eliminated");
108 STATISTIC(NumPHIsElim, "Number of trivial PHIs eliminated");
109 STATISTIC(NumGEPsElim, "Number of GEPs converted to casts");
110 STATISTIC(NumCmpUses, "Number of uses of Cmp expressions replaced with uses of "
111  "sunken Cmps");
112 STATISTIC(NumCastUses, "Number of uses of Cast expressions replaced with uses "
113  "of sunken Casts");
114 STATISTIC(NumMemoryInsts, "Number of memory instructions whose address "
115  "computations were sunk");
116 STATISTIC(NumMemoryInstsPhiCreated,
117  "Number of phis created when address "
118  "computations were sunk to memory instructions");
119 STATISTIC(NumMemoryInstsSelectCreated,
120  "Number of select created when address "
121  "computations were sunk to memory instructions");
122 STATISTIC(NumExtsMoved, "Number of [s|z]ext instructions combined with loads");
123 STATISTIC(NumExtUses, "Number of uses of [s|z]ext instructions optimized");
124 STATISTIC(NumAndsAdded,
125  "Number of and mask instructions added to form ext loads");
126 STATISTIC(NumAndUses, "Number of uses of and mask instructions optimized");
127 STATISTIC(NumRetsDup, "Number of return instructions duplicated");
128 STATISTIC(NumDbgValueMoved, "Number of debug value instructions moved");
129 STATISTIC(NumSelectsExpanded, "Number of selects turned into branches");
130 STATISTIC(NumStoreExtractExposed, "Number of store(extractelement) exposed");
131 
133  "disable-cgp-branch-opts", cl::Hidden, cl::init(false),
134  cl::desc("Disable branch optimizations in CodeGenPrepare"));
135 
136 static cl::opt<bool>
137  DisableGCOpts("disable-cgp-gc-opts", cl::Hidden, cl::init(false),
138  cl::desc("Disable GC optimizations in CodeGenPrepare"));
139 
141  "disable-cgp-select2branch", cl::Hidden, cl::init(false),
142  cl::desc("Disable select to branch conversion."));
143 
145  "addr-sink-using-gep", cl::Hidden, cl::init(true),
146  cl::desc("Address sinking in CGP using GEPs."));
147 
149  "enable-andcmp-sinking", cl::Hidden, cl::init(true),
150  cl::desc("Enable sinkinig and/cmp into branches."));
151 
153  "disable-cgp-store-extract", cl::Hidden, cl::init(false),
154  cl::desc("Disable store(extract) optimizations in CodeGenPrepare"));
155 
157  "stress-cgp-store-extract", cl::Hidden, cl::init(false),
158  cl::desc("Stress test store(extract) optimizations in CodeGenPrepare"));
159 
161  "disable-cgp-ext-ld-promotion", cl::Hidden, cl::init(false),
162  cl::desc("Disable ext(promotable(ld)) -> promoted(ext(ld)) optimization in "
163  "CodeGenPrepare"));
164 
166  "stress-cgp-ext-ld-promotion", cl::Hidden, cl::init(false),
167  cl::desc("Stress test ext(promotable(ld)) -> promoted(ext(ld)) "
168  "optimization in CodeGenPrepare"));
169 
171  "disable-preheader-prot", cl::Hidden, cl::init(false),
172  cl::desc("Disable protection against removing loop preheaders"));
173 
175  "profile-guided-section-prefix", cl::Hidden, cl::init(true), cl::ZeroOrMore,
176  cl::desc("Use profile info to add section prefix for hot/cold functions"));
177 
179  "cgp-freq-ratio-to-skip-merge", cl::Hidden, cl::init(2),
180  cl::desc("Skip merging empty blocks if (frequency of empty block) / "
181  "(frequency of destination block) is greater than this ratio"));
182 
184  "force-split-store", cl::Hidden, cl::init(false),
185  cl::desc("Force store splitting no matter what the target query says."));
186 
187 static cl::opt<bool>
188 EnableTypePromotionMerge("cgp-type-promotion-merge", cl::Hidden,
189  cl::desc("Enable merging of redundant sexts when one is dominating"
190  " the other."), cl::init(true));
191 
193  "disable-complex-addr-modes", cl::Hidden, cl::init(true),
194  cl::desc("Disables combining addressing modes with different parts "
195  "in optimizeMemoryInst."));
196 
197 static cl::opt<bool>
198 AddrSinkNewPhis("addr-sink-new-phis", cl::Hidden, cl::init(false),
199  cl::desc("Allow creation of Phis in Address sinking."));
200 
201 static cl::opt<bool>
202 AddrSinkNewSelects("addr-sink-new-select", cl::Hidden, cl::init(false),
203  cl::desc("Allow creation of selects in Address sinking."));
204 
205 namespace {
206 
207 using SetOfInstrs = SmallPtrSet<Instruction *, 16>;
208 using TypeIsSExt = PointerIntPair<Type *, 1, bool>;
209 using InstrToOrigTy = DenseMap<Instruction *, TypeIsSExt>;
210 using SExts = SmallVector<Instruction *, 16>;
211 using ValueToSExts = DenseMap<Value *, SExts>;
212 
213 class TypePromotionTransaction;
214 
215  class CodeGenPrepare : public FunctionPass {
216  const TargetMachine *TM = nullptr;
217  const TargetSubtargetInfo *SubtargetInfo;
218  const TargetLowering *TLI = nullptr;
219  const TargetRegisterInfo *TRI;
220  const TargetTransformInfo *TTI = nullptr;
221  const TargetLibraryInfo *TLInfo;
222  const LoopInfo *LI;
223  std::unique_ptr<BlockFrequencyInfo> BFI;
224  std::unique_ptr<BranchProbabilityInfo> BPI;
225 
226  /// As we scan instructions optimizing them, this is the next instruction
227  /// to optimize. Transforms that can invalidate this should update it.
228  BasicBlock::iterator CurInstIterator;
229 
230  /// Keeps track of non-local addresses that have been sunk into a block.
231  /// This allows us to avoid inserting duplicate code for blocks with
232  /// multiple load/stores of the same address.
233  ValueMap<Value*, Value*> SunkAddrs;
234 
235  /// Keeps track of all instructions inserted for the current function.
236  SetOfInstrs InsertedInsts;
237 
238  /// Keeps track of the type of the related instruction before their
239  /// promotion for the current function.
240  InstrToOrigTy PromotedInsts;
241 
242  /// Keep track of instructions removed during promotion.
243  SetOfInstrs RemovedInsts;
244 
245  /// Keep track of sext chains based on their initial value.
246  DenseMap<Value *, Instruction *> SeenChainsForSExt;
247 
248  /// Keep track of SExt promoted.
249  ValueToSExts ValToSExtendedUses;
250 
251  /// True if CFG is modified in any way.
252  bool ModifiedDT;
253 
254  /// True if optimizing for size.
255  bool OptSize;
256 
257  /// DataLayout for the Function being processed.
258  const DataLayout *DL = nullptr;
259 
260  public:
261  static char ID; // Pass identification, replacement for typeid
262 
263  CodeGenPrepare() : FunctionPass(ID) {
265  }
266 
267  bool runOnFunction(Function &F) override;
268 
269  StringRef getPassName() const override { return "CodeGen Prepare"; }
270 
271  void getAnalysisUsage(AnalysisUsage &AU) const override {
272  // FIXME: When we can selectively preserve passes, preserve the domtree.
277  }
278 
279  private:
280  bool eliminateFallThrough(Function &F);
281  bool eliminateMostlyEmptyBlocks(Function &F);
282  BasicBlock *findDestBlockOfMergeableEmptyBlock(BasicBlock *BB);
283  bool canMergeBlocks(const BasicBlock *BB, const BasicBlock *DestBB) const;
284  void eliminateMostlyEmptyBlock(BasicBlock *BB);
285  bool isMergingEmptyBlockProfitable(BasicBlock *BB, BasicBlock *DestBB,
286  bool isPreheader);
287  bool optimizeBlock(BasicBlock &BB, bool &ModifiedDT);
288  bool optimizeInst(Instruction *I, bool &ModifiedDT);
289  bool optimizeMemoryInst(Instruction *I, Value *Addr,
290  Type *AccessTy, unsigned AS);
291  bool optimizeInlineAsmInst(CallInst *CS);
292  bool optimizeCallInst(CallInst *CI, bool &ModifiedDT);
293  bool optimizeExt(Instruction *&I);
294  bool optimizeExtUses(Instruction *I);
295  bool optimizeLoadExt(LoadInst *I);
296  bool optimizeSelectInst(SelectInst *SI);
297  bool optimizeShuffleVectorInst(ShuffleVectorInst *SI);
298  bool optimizeSwitchInst(SwitchInst *CI);
299  bool optimizeExtractElementInst(Instruction *Inst);
300  bool dupRetToEnableTailCallOpts(BasicBlock *BB);
301  bool placeDbgValues(Function &F);
302  bool canFormExtLd(const SmallVectorImpl<Instruction *> &MovedExts,
303  LoadInst *&LI, Instruction *&Inst, bool HasPromoted);
304  bool tryToPromoteExts(TypePromotionTransaction &TPT,
305  const SmallVectorImpl<Instruction *> &Exts,
306  SmallVectorImpl<Instruction *> &ProfitablyMovedExts,
307  unsigned CreatedInstsCost = 0);
308  bool mergeSExts(Function &F);
309  bool performAddressTypePromotion(
310  Instruction *&Inst,
311  bool AllowPromotionWithoutCommonHeader,
312  bool HasPromoted, TypePromotionTransaction &TPT,
313  SmallVectorImpl<Instruction *> &SpeculativelyMovedExts);
314  bool splitBranchCondition(Function &F);
315  bool simplifyOffsetableRelocate(Instruction &I);
316  bool splitIndirectCriticalEdges(Function &F);
317  };
318 
319 } // end anonymous namespace
320 
321 char CodeGenPrepare::ID = 0;
322 
323 INITIALIZE_PASS_BEGIN(CodeGenPrepare, DEBUG_TYPE,
324  "Optimize for code generation", false, false)
327  "Optimize for code generation", false, false)
328 
329 FunctionPass *llvm::createCodeGenPreparePass() { return new CodeGenPrepare(); }
330 
332  if (skipFunction(F))
333  return false;
334 
335  DL = &F.getParent()->getDataLayout();
336 
337  bool EverMadeChange = false;
338  // Clear per function information.
339  InsertedInsts.clear();
340  PromotedInsts.clear();
341  BFI.reset();
342  BPI.reset();
343 
344  ModifiedDT = false;
345  if (auto *TPC = getAnalysisIfAvailable<TargetPassConfig>()) {
346  TM = &TPC->getTM<TargetMachine>();
347  SubtargetInfo = TM->getSubtargetImpl(F);
348  TLI = SubtargetInfo->getTargetLowering();
349  TRI = SubtargetInfo->getRegisterInfo();
350  }
351  TLInfo = &getAnalysis<TargetLibraryInfoWrapperPass>().getTLI();
352  TTI = &getAnalysis<TargetTransformInfoWrapperPass>().getTTI(F);
353  LI = &getAnalysis<LoopInfoWrapperPass>().getLoopInfo();
354  OptSize = F.optForSize();
355 
356  ProfileSummaryInfo *PSI =
357  getAnalysis<ProfileSummaryInfoWrapperPass>().getPSI();
359  if (PSI->isFunctionHotInCallGraph(&F))
360  F.setSectionPrefix(".hot");
361  else if (PSI->isFunctionColdInCallGraph(&F))
362  F.setSectionPrefix(".unlikely");
363  }
364 
365  /// This optimization identifies DIV instructions that can be
366  /// profitably bypassed and carried out with a shorter, faster divide.
367  if (!OptSize && !PSI->hasHugeWorkingSetSize() && TLI &&
368  TLI->isSlowDivBypassed()) {
369  const DenseMap<unsigned int, unsigned int> &BypassWidths =
370  TLI->getBypassSlowDivWidths();
371  BasicBlock* BB = &*F.begin();
372  while (BB != nullptr) {
373  // bypassSlowDivision may create new BBs, but we don't want to reapply the
374  // optimization to those blocks.
375  BasicBlock* Next = BB->getNextNode();
376  EverMadeChange |= bypassSlowDivision(BB, BypassWidths);
377  BB = Next;
378  }
379  }
380 
381  // Eliminate blocks that contain only PHI nodes and an
382  // unconditional branch.
383  EverMadeChange |= eliminateMostlyEmptyBlocks(F);
384 
385  // llvm.dbg.value is far away from the value then iSel may not be able
386  // handle it properly. iSel will drop llvm.dbg.value if it can not
387  // find a node corresponding to the value.
388  EverMadeChange |= placeDbgValues(F);
389 
390  if (!DisableBranchOpts)
391  EverMadeChange |= splitBranchCondition(F);
392 
393  // Split some critical edges where one of the sources is an indirect branch,
394  // to help generate sane code for PHIs involving such edges.
395  EverMadeChange |= splitIndirectCriticalEdges(F);
396 
397  bool MadeChange = true;
398  while (MadeChange) {
399  MadeChange = false;
400  SeenChainsForSExt.clear();
401  ValToSExtendedUses.clear();
402  RemovedInsts.clear();
403  for (Function::iterator I = F.begin(); I != F.end(); ) {
404  BasicBlock *BB = &*I++;
405  bool ModifiedDTOnIteration = false;
406  MadeChange |= optimizeBlock(*BB, ModifiedDTOnIteration);
407 
408  // Restart BB iteration if the dominator tree of the Function was changed
409  if (ModifiedDTOnIteration)
410  break;
411  }
412  if (EnableTypePromotionMerge && !ValToSExtendedUses.empty())
413  MadeChange |= mergeSExts(F);
414 
415  // Really free removed instructions during promotion.
416  for (Instruction *I : RemovedInsts)
417  I->deleteValue();
418 
419  EverMadeChange |= MadeChange;
420  }
421 
422  SunkAddrs.clear();
423 
424  if (!DisableBranchOpts) {
425  MadeChange = false;
427  for (BasicBlock &BB : F) {
428  SmallVector<BasicBlock *, 2> Successors(succ_begin(&BB), succ_end(&BB));
429  MadeChange |= ConstantFoldTerminator(&BB, true);
430  if (!MadeChange) continue;
431 
433  II = Successors.begin(), IE = Successors.end(); II != IE; ++II)
434  if (pred_begin(*II) == pred_end(*II))
435  WorkList.insert(*II);
436  }
437 
438  // Delete the dead blocks and any of their dead successors.
439  MadeChange |= !WorkList.empty();
440  while (!WorkList.empty()) {
441  BasicBlock *BB = *WorkList.begin();
442  WorkList.erase(BB);
443  SmallVector<BasicBlock*, 2> Successors(succ_begin(BB), succ_end(BB));
444 
445  DeleteDeadBlock(BB);
446 
448  II = Successors.begin(), IE = Successors.end(); II != IE; ++II)
449  if (pred_begin(*II) == pred_end(*II))
450  WorkList.insert(*II);
451  }
452 
453  // Merge pairs of basic blocks with unconditional branches, connected by
454  // a single edge.
455  if (EverMadeChange || MadeChange)
456  MadeChange |= eliminateFallThrough(F);
457 
458  EverMadeChange |= MadeChange;
459  }
460 
461  if (!DisableGCOpts) {
462  SmallVector<Instruction *, 2> Statepoints;
463  for (BasicBlock &BB : F)
464  for (Instruction &I : BB)
465  if (isStatepoint(I))
466  Statepoints.push_back(&I);
467  for (auto &I : Statepoints)
468  EverMadeChange |= simplifyOffsetableRelocate(*I);
469  }
470 
471  return EverMadeChange;
472 }
473 
474 /// Merge basic blocks which are connected by a single edge, where one of the
475 /// basic blocks has a single successor pointing to the other basic block,
476 /// which has a single predecessor.
477 bool CodeGenPrepare::eliminateFallThrough(Function &F) {
478  bool Changed = false;
479  // Scan all of the blocks in the function, except for the entry block.
480  for (Function::iterator I = std::next(F.begin()), E = F.end(); I != E;) {
481  BasicBlock *BB = &*I++;
482  // If the destination block has a single pred, then this is a trivial
483  // edge, just collapse it.
484  BasicBlock *SinglePred = BB->getSinglePredecessor();
485 
486  // Don't merge if BB's address is taken.
487  if (!SinglePred || SinglePred == BB || BB->hasAddressTaken()) continue;
488 
489  BranchInst *Term = dyn_cast<BranchInst>(SinglePred->getTerminator());
490  if (Term && !Term->isConditional()) {
491  Changed = true;
492  DEBUG(dbgs() << "To merge:\n"<< *SinglePred << "\n\n\n");
493  // Remember if SinglePred was the entry block of the function.
494  // If so, we will need to move BB back to the entry position.
495  bool isEntry = SinglePred == &SinglePred->getParent()->getEntryBlock();
496  MergeBasicBlockIntoOnlyPred(BB, nullptr);
497 
498  if (isEntry && BB != &BB->getParent()->getEntryBlock())
499  BB->moveBefore(&BB->getParent()->getEntryBlock());
500 
501  // We have erased a block. Update the iterator.
502  I = BB->getIterator();
503  }
504  }
505  return Changed;
506 }
507 
508 /// Find a destination block from BB if BB is mergeable empty block.
509 BasicBlock *CodeGenPrepare::findDestBlockOfMergeableEmptyBlock(BasicBlock *BB) {
510  // If this block doesn't end with an uncond branch, ignore it.
512  if (!BI || !BI->isUnconditional())
513  return nullptr;
514 
515  // If the instruction before the branch (skipping debug info) isn't a phi
516  // node, then other stuff is happening here.
517  BasicBlock::iterator BBI = BI->getIterator();
518  if (BBI != BB->begin()) {
519  --BBI;
520  while (isa<DbgInfoIntrinsic>(BBI)) {
521  if (BBI == BB->begin())
522  break;
523  --BBI;
524  }
525  if (!isa<DbgInfoIntrinsic>(BBI) && !isa<PHINode>(BBI))
526  return nullptr;
527  }
528 
529  // Do not break infinite loops.
530  BasicBlock *DestBB = BI->getSuccessor(0);
531  if (DestBB == BB)
532  return nullptr;
533 
534  if (!canMergeBlocks(BB, DestBB))
535  DestBB = nullptr;
536 
537  return DestBB;
538 }
539 
540 // Return the unique indirectbr predecessor of a block. This may return null
541 // even if such a predecessor exists, if it's not useful for splitting.
542 // If a predecessor is found, OtherPreds will contain all other (non-indirectbr)
543 // predecessors of BB.
544 static BasicBlock *
546  // If the block doesn't have any PHIs, we don't care about it, since there's
547  // no point in splitting it.
548  PHINode *PN = dyn_cast<PHINode>(BB->begin());
549  if (!PN)
550  return nullptr;
551 
552  // Verify we have exactly one IBR predecessor.
553  // Conservatively bail out if one of the other predecessors is not a "regular"
554  // terminator (that is, not a switch or a br).
555  BasicBlock *IBB = nullptr;
556  for (unsigned Pred = 0, E = PN->getNumIncomingValues(); Pred != E; ++Pred) {
557  BasicBlock *PredBB = PN->getIncomingBlock(Pred);
558  TerminatorInst *PredTerm = PredBB->getTerminator();
559  switch (PredTerm->getOpcode()) {
560  case Instruction::IndirectBr:
561  if (IBB)
562  return nullptr;
563  IBB = PredBB;
564  break;
565  case Instruction::Br:
566  case Instruction::Switch:
567  OtherPreds.push_back(PredBB);
568  continue;
569  default:
570  return nullptr;
571  }
572  }
573 
574  return IBB;
575 }
576 
577 // Split critical edges where the source of the edge is an indirectbr
578 // instruction. This isn't always possible, but we can handle some easy cases.
579 // This is useful because MI is unable to split such critical edges,
580 // which means it will not be able to sink instructions along those edges.
581 // This is especially painful for indirect branches with many successors, where
582 // we end up having to prepare all outgoing values in the origin block.
583 //
584 // Our normal algorithm for splitting critical edges requires us to update
585 // the outgoing edges of the edge origin block, but for an indirectbr this
586 // is hard, since it would require finding and updating the block addresses
587 // the indirect branch uses. But if a block only has a single indirectbr
588 // predecessor, with the others being regular branches, we can do it in a
589 // different way.
590 // Say we have A -> D, B -> D, I -> D where only I -> D is an indirectbr.
591 // We can split D into D0 and D1, where D0 contains only the PHIs from D,
592 // and D1 is the D block body. We can then duplicate D0 as D0A and D0B, and
593 // create the following structure:
594 // A -> D0A, B -> D0A, I -> D0B, D0A -> D1, D0B -> D1
595 bool CodeGenPrepare::splitIndirectCriticalEdges(Function &F) {
596  // Check whether the function has any indirectbrs, and collect which blocks
597  // they may jump to. Since most functions don't have indirect branches,
598  // this lowers the common case's overhead to O(Blocks) instead of O(Edges).
600  for (auto &BB : F) {
601  auto *IBI = dyn_cast<IndirectBrInst>(BB.getTerminator());
602  if (!IBI)
603  continue;
604 
605  for (unsigned Succ = 0, E = IBI->getNumSuccessors(); Succ != E; ++Succ)
606  Targets.insert(IBI->getSuccessor(Succ));
607  }
608 
609  if (Targets.empty())
610  return false;
611 
612  bool Changed = false;
613  for (BasicBlock *Target : Targets) {
615  BasicBlock *IBRPred = findIBRPredecessor(Target, OtherPreds);
616  // If we did not found an indirectbr, or the indirectbr is the only
617  // incoming edge, this isn't the kind of edge we're looking for.
618  if (!IBRPred || OtherPreds.empty())
619  continue;
620 
621  // Don't even think about ehpads/landingpads.
622  Instruction *FirstNonPHI = Target->getFirstNonPHI();
623  if (FirstNonPHI->isEHPad() || Target->isLandingPad())
624  continue;
625 
626  BasicBlock *BodyBlock = Target->splitBasicBlock(FirstNonPHI, ".split");
627  // It's possible Target was its own successor through an indirectbr.
628  // In this case, the indirectbr now comes from BodyBlock.
629  if (IBRPred == Target)
630  IBRPred = BodyBlock;
631 
632  // At this point Target only has PHIs, and BodyBlock has the rest of the
633  // block's body. Create a copy of Target that will be used by the "direct"
634  // preds.
635  ValueToValueMapTy VMap;
636  BasicBlock *DirectSucc = CloneBasicBlock(Target, VMap, ".clone", &F);
637 
638  for (BasicBlock *Pred : OtherPreds) {
639  // If the target is a loop to itself, then the terminator of the split
640  // block needs to be updated.
641  if (Pred == Target)
642  BodyBlock->getTerminator()->replaceUsesOfWith(Target, DirectSucc);
643  else
644  Pred->getTerminator()->replaceUsesOfWith(Target, DirectSucc);
645  }
646 
647  // Ok, now fix up the PHIs. We know the two blocks only have PHIs, and that
648  // they are clones, so the number of PHIs are the same.
649  // (a) Remove the edge coming from IBRPred from the "Direct" PHI
650  // (b) Leave that as the only edge in the "Indirect" PHI.
651  // (c) Merge the two in the body block.
652  BasicBlock::iterator Indirect = Target->begin(),
653  End = Target->getFirstNonPHI()->getIterator();
654  BasicBlock::iterator Direct = DirectSucc->begin();
655  BasicBlock::iterator MergeInsert = BodyBlock->getFirstInsertionPt();
656 
657  assert(&*End == Target->getTerminator() &&
658  "Block was expected to only contain PHIs");
659 
660  while (Indirect != End) {
661  PHINode *DirPHI = cast<PHINode>(Direct);
662  PHINode *IndPHI = cast<PHINode>(Indirect);
663 
664  // Now, clean up - the direct block shouldn't get the indirect value,
665  // and vice versa.
666  DirPHI->removeIncomingValue(IBRPred);
667  Direct++;
668 
669  // Advance the pointer here, to avoid invalidation issues when the old
670  // PHI is erased.
671  Indirect++;
672 
673  PHINode *NewIndPHI = PHINode::Create(IndPHI->getType(), 1, "ind", IndPHI);
674  NewIndPHI->addIncoming(IndPHI->getIncomingValueForBlock(IBRPred),
675  IBRPred);
676 
677  // Create a PHI in the body block, to merge the direct and indirect
678  // predecessors.
679  PHINode *MergePHI =
680  PHINode::Create(IndPHI->getType(), 2, "merge", &*MergeInsert);
681  MergePHI->addIncoming(NewIndPHI, Target);
682  MergePHI->addIncoming(DirPHI, DirectSucc);
683 
684  IndPHI->replaceAllUsesWith(MergePHI);
685  IndPHI->eraseFromParent();
686  }
687 
688  Changed = true;
689  }
690 
691  return Changed;
692 }
693 
694 /// Eliminate blocks that contain only PHI nodes, debug info directives, and an
695 /// unconditional branch. Passes before isel (e.g. LSR/loopsimplify) often split
696 /// edges in ways that are non-optimal for isel. Start by eliminating these
697 /// blocks so we can split them the way we want them.
698 bool CodeGenPrepare::eliminateMostlyEmptyBlocks(Function &F) {
700  SmallVector<Loop *, 16> LoopList(LI->begin(), LI->end());
701  while (!LoopList.empty()) {
702  Loop *L = LoopList.pop_back_val();
703  LoopList.insert(LoopList.end(), L->begin(), L->end());
704  if (BasicBlock *Preheader = L->getLoopPreheader())
705  Preheaders.insert(Preheader);
706  }
707 
708  bool MadeChange = false;
709  // Note that this intentionally skips the entry block.
710  for (Function::iterator I = std::next(F.begin()), E = F.end(); I != E;) {
711  BasicBlock *BB = &*I++;
712  BasicBlock *DestBB = findDestBlockOfMergeableEmptyBlock(BB);
713  if (!DestBB ||
714  !isMergingEmptyBlockProfitable(BB, DestBB, Preheaders.count(BB)))
715  continue;
716 
717  eliminateMostlyEmptyBlock(BB);
718  MadeChange = true;
719  }
720  return MadeChange;
721 }
722 
723 bool CodeGenPrepare::isMergingEmptyBlockProfitable(BasicBlock *BB,
724  BasicBlock *DestBB,
725  bool isPreheader) {
726  // Do not delete loop preheaders if doing so would create a critical edge.
727  // Loop preheaders can be good locations to spill registers. If the
728  // preheader is deleted and we create a critical edge, registers may be
729  // spilled in the loop body instead.
730  if (!DisablePreheaderProtect && isPreheader &&
731  !(BB->getSinglePredecessor() &&
733  return false;
734 
735  // Try to skip merging if the unique predecessor of BB is terminated by a
736  // switch or indirect branch instruction, and BB is used as an incoming block
737  // of PHIs in DestBB. In such case, merging BB and DestBB would cause ISel to
738  // add COPY instructions in the predecessor of BB instead of BB (if it is not
739  // merged). Note that the critical edge created by merging such blocks wont be
740  // split in MachineSink because the jump table is not analyzable. By keeping
741  // such empty block (BB), ISel will place COPY instructions in BB, not in the
742  // predecessor of BB.
743  BasicBlock *Pred = BB->getUniquePredecessor();
744  if (!Pred ||
745  !(isa<SwitchInst>(Pred->getTerminator()) ||
746  isa<IndirectBrInst>(Pred->getTerminator())))
747  return true;
748 
749  if (BB->getTerminator() != BB->getFirstNonPHI())
750  return true;
751 
752  // We use a simple cost heuristic which determine skipping merging is
753  // profitable if the cost of skipping merging is less than the cost of
754  // merging : Cost(skipping merging) < Cost(merging BB), where the
755  // Cost(skipping merging) is Freq(BB) * (Cost(Copy) + Cost(Branch)), and
756  // the Cost(merging BB) is Freq(Pred) * Cost(Copy).
757  // Assuming Cost(Copy) == Cost(Branch), we could simplify it to :
758  // Freq(Pred) / Freq(BB) > 2.
759  // Note that if there are multiple empty blocks sharing the same incoming
760  // value for the PHIs in the DestBB, we consider them together. In such
761  // case, Cost(merging BB) will be the sum of their frequencies.
762 
763  if (!isa<PHINode>(DestBB->begin()))
764  return true;
765 
766  SmallPtrSet<BasicBlock *, 16> SameIncomingValueBBs;
767 
768  // Find all other incoming blocks from which incoming values of all PHIs in
769  // DestBB are the same as the ones from BB.
770  for (pred_iterator PI = pred_begin(DestBB), E = pred_end(DestBB); PI != E;
771  ++PI) {
772  BasicBlock *DestBBPred = *PI;
773  if (DestBBPred == BB)
774  continue;
775 
776  bool HasAllSameValue = true;
777  BasicBlock::const_iterator DestBBI = DestBB->begin();
778  while (const PHINode *DestPN = dyn_cast<PHINode>(DestBBI++)) {
779  if (DestPN->getIncomingValueForBlock(BB) !=
780  DestPN->getIncomingValueForBlock(DestBBPred)) {
781  HasAllSameValue = false;
782  break;
783  }
784  }
785  if (HasAllSameValue)
786  SameIncomingValueBBs.insert(DestBBPred);
787  }
788 
789  // See if all BB's incoming values are same as the value from Pred. In this
790  // case, no reason to skip merging because COPYs are expected to be place in
791  // Pred already.
792  if (SameIncomingValueBBs.count(Pred))
793  return true;
794 
795  if (!BFI) {
796  Function &F = *BB->getParent();
797  LoopInfo LI{DominatorTree(F)};
798  BPI.reset(new BranchProbabilityInfo(F, LI));
799  BFI.reset(new BlockFrequencyInfo(F, *BPI, LI));
800  }
801 
802  BlockFrequency PredFreq = BFI->getBlockFreq(Pred);
803  BlockFrequency BBFreq = BFI->getBlockFreq(BB);
804 
805  for (auto SameValueBB : SameIncomingValueBBs)
806  if (SameValueBB->getUniquePredecessor() == Pred &&
807  DestBB == findDestBlockOfMergeableEmptyBlock(SameValueBB))
808  BBFreq += BFI->getBlockFreq(SameValueBB);
809 
810  return PredFreq.getFrequency() <=
812 }
813 
814 /// Return true if we can merge BB into DestBB if there is a single
815 /// unconditional branch between them, and BB contains no other non-phi
816 /// instructions.
817 bool CodeGenPrepare::canMergeBlocks(const BasicBlock *BB,
818  const BasicBlock *DestBB) const {
819  // We only want to eliminate blocks whose phi nodes are used by phi nodes in
820  // the successor. If there are more complex condition (e.g. preheaders),
821  // don't mess around with them.
822  BasicBlock::const_iterator BBI = BB->begin();
823  while (const PHINode *PN = dyn_cast<PHINode>(BBI++)) {
824  for (const User *U : PN->users()) {
825  const Instruction *UI = cast<Instruction>(U);
826  if (UI->getParent() != DestBB || !isa<PHINode>(UI))
827  return false;
828  // If User is inside DestBB block and it is a PHINode then check
829  // incoming value. If incoming value is not from BB then this is
830  // a complex condition (e.g. preheaders) we want to avoid here.
831  if (UI->getParent() == DestBB) {
832  if (const PHINode *UPN = dyn_cast<PHINode>(UI))
833  for (unsigned I = 0, E = UPN->getNumIncomingValues(); I != E; ++I) {
834  Instruction *Insn = dyn_cast<Instruction>(UPN->getIncomingValue(I));
835  if (Insn && Insn->getParent() == BB &&
836  Insn->getParent() != UPN->getIncomingBlock(I))
837  return false;
838  }
839  }
840  }
841  }
842 
843  // If BB and DestBB contain any common predecessors, then the phi nodes in BB
844  // and DestBB may have conflicting incoming values for the block. If so, we
845  // can't merge the block.
846  const PHINode *DestBBPN = dyn_cast<PHINode>(DestBB->begin());
847  if (!DestBBPN) return true; // no conflict.
848 
849  // Collect the preds of BB.
851  if (const PHINode *BBPN = dyn_cast<PHINode>(BB->begin())) {
852  // It is faster to get preds from a PHI than with pred_iterator.
853  for (unsigned i = 0, e = BBPN->getNumIncomingValues(); i != e; ++i)
854  BBPreds.insert(BBPN->getIncomingBlock(i));
855  } else {
856  BBPreds.insert(pred_begin(BB), pred_end(BB));
857  }
858 
859  // Walk the preds of DestBB.
860  for (unsigned i = 0, e = DestBBPN->getNumIncomingValues(); i != e; ++i) {
861  BasicBlock *Pred = DestBBPN->getIncomingBlock(i);
862  if (BBPreds.count(Pred)) { // Common predecessor?
863  BBI = DestBB->begin();
864  while (const PHINode *PN = dyn_cast<PHINode>(BBI++)) {
865  const Value *V1 = PN->getIncomingValueForBlock(Pred);
866  const Value *V2 = PN->getIncomingValueForBlock(BB);
867 
868  // If V2 is a phi node in BB, look up what the mapped value will be.
869  if (const PHINode *V2PN = dyn_cast<PHINode>(V2))
870  if (V2PN->getParent() == BB)
871  V2 = V2PN->getIncomingValueForBlock(Pred);
872 
873  // If there is a conflict, bail out.
874  if (V1 != V2) return false;
875  }
876  }
877  }
878 
879  return true;
880 }
881 
882 /// Eliminate a basic block that has only phi's and an unconditional branch in
883 /// it.
884 void CodeGenPrepare::eliminateMostlyEmptyBlock(BasicBlock *BB) {
885  BranchInst *BI = cast<BranchInst>(BB->getTerminator());
886  BasicBlock *DestBB = BI->getSuccessor(0);
887 
888  DEBUG(dbgs() << "MERGING MOSTLY EMPTY BLOCKS - BEFORE:\n" << *BB << *DestBB);
889 
890  // If the destination block has a single pred, then this is a trivial edge,
891  // just collapse it.
892  if (BasicBlock *SinglePred = DestBB->getSinglePredecessor()) {
893  if (SinglePred != DestBB) {
894  // Remember if SinglePred was the entry block of the function. If so, we
895  // will need to move BB back to the entry position.
896  bool isEntry = SinglePred == &SinglePred->getParent()->getEntryBlock();
897  MergeBasicBlockIntoOnlyPred(DestBB, nullptr);
898 
899  if (isEntry && BB != &BB->getParent()->getEntryBlock())
900  BB->moveBefore(&BB->getParent()->getEntryBlock());
901 
902  DEBUG(dbgs() << "AFTER:\n" << *DestBB << "\n\n\n");
903  return;
904  }
905  }
906 
907  // Otherwise, we have multiple predecessors of BB. Update the PHIs in DestBB
908  // to handle the new incoming edges it is about to have.
909  PHINode *PN;
910  for (BasicBlock::iterator BBI = DestBB->begin();
911  (PN = dyn_cast<PHINode>(BBI)); ++BBI) {
912  // Remove the incoming value for BB, and remember it.
913  Value *InVal = PN->removeIncomingValue(BB, false);
914 
915  // Two options: either the InVal is a phi node defined in BB or it is some
916  // value that dominates BB.
917  PHINode *InValPhi = dyn_cast<PHINode>(InVal);
918  if (InValPhi && InValPhi->getParent() == BB) {
919  // Add all of the input values of the input PHI as inputs of this phi.
920  for (unsigned i = 0, e = InValPhi->getNumIncomingValues(); i != e; ++i)
921  PN->addIncoming(InValPhi->getIncomingValue(i),
922  InValPhi->getIncomingBlock(i));
923  } else {
924  // Otherwise, add one instance of the dominating value for each edge that
925  // we will be adding.
926  if (PHINode *BBPN = dyn_cast<PHINode>(BB->begin())) {
927  for (unsigned i = 0, e = BBPN->getNumIncomingValues(); i != e; ++i)
928  PN->addIncoming(InVal, BBPN->getIncomingBlock(i));
929  } else {
930  for (pred_iterator PI = pred_begin(BB), E = pred_end(BB); PI != E; ++PI)
931  PN->addIncoming(InVal, *PI);
932  }
933  }
934  }
935 
936  // The PHIs are now updated, change everything that refers to BB to use
937  // DestBB and remove BB.
938  BB->replaceAllUsesWith(DestBB);
939  BB->eraseFromParent();
940  ++NumBlocksElim;
941 
942  DEBUG(dbgs() << "AFTER:\n" << *DestBB << "\n\n\n");
943 }
944 
945 // Computes a map of base pointer relocation instructions to corresponding
946 // derived pointer relocation instructions given a vector of all relocate calls
948  const SmallVectorImpl<GCRelocateInst *> &AllRelocateCalls,
950  &RelocateInstMap) {
951  // Collect information in two maps: one primarily for locating the base object
952  // while filling the second map; the second map is the final structure holding
953  // a mapping between Base and corresponding Derived relocate calls
955  for (auto *ThisRelocate : AllRelocateCalls) {
956  auto K = std::make_pair(ThisRelocate->getBasePtrIndex(),
957  ThisRelocate->getDerivedPtrIndex());
958  RelocateIdxMap.insert(std::make_pair(K, ThisRelocate));
959  }
960  for (auto &Item : RelocateIdxMap) {
961  std::pair<unsigned, unsigned> Key = Item.first;
962  if (Key.first == Key.second)
963  // Base relocation: nothing to insert
964  continue;
965 
966  GCRelocateInst *I = Item.second;
967  auto BaseKey = std::make_pair(Key.first, Key.first);
968 
969  // We're iterating over RelocateIdxMap so we cannot modify it.
970  auto MaybeBase = RelocateIdxMap.find(BaseKey);
971  if (MaybeBase == RelocateIdxMap.end())
972  // TODO: We might want to insert a new base object relocate and gep off
973  // that, if there are enough derived object relocates.
974  continue;
975 
976  RelocateInstMap[MaybeBase->second].push_back(I);
977  }
978 }
979 
980 // Accepts a GEP and extracts the operands into a vector provided they're all
981 // small integer constants
983  SmallVectorImpl<Value *> &OffsetV) {
984  for (unsigned i = 1; i < GEP->getNumOperands(); i++) {
985  // Only accept small constant integer operands
986  auto Op = dyn_cast<ConstantInt>(GEP->getOperand(i));
987  if (!Op || Op->getZExtValue() > 20)
988  return false;
989  }
990 
991  for (unsigned i = 1; i < GEP->getNumOperands(); i++)
992  OffsetV.push_back(GEP->getOperand(i));
993  return true;
994 }
995 
996 // Takes a RelocatedBase (base pointer relocation instruction) and Targets to
997 // replace, computes a replacement, and affects it.
998 static bool
1000  const SmallVectorImpl<GCRelocateInst *> &Targets) {
1001  bool MadeChange = false;
1002  // We must ensure the relocation of derived pointer is defined after
1003  // relocation of base pointer. If we find a relocation corresponding to base
1004  // defined earlier than relocation of base then we move relocation of base
1005  // right before found relocation. We consider only relocation in the same
1006  // basic block as relocation of base. Relocations from other basic block will
1007  // be skipped by optimization and we do not care about them.
1008  for (auto R = RelocatedBase->getParent()->getFirstInsertionPt();
1009  &*R != RelocatedBase; ++R)
1010  if (auto RI = dyn_cast<GCRelocateInst>(R))
1011  if (RI->getStatepoint() == RelocatedBase->getStatepoint())
1012  if (RI->getBasePtrIndex() == RelocatedBase->getBasePtrIndex()) {
1013  RelocatedBase->moveBefore(RI);
1014  break;
1015  }
1016 
1017  for (GCRelocateInst *ToReplace : Targets) {
1018  assert(ToReplace->getBasePtrIndex() == RelocatedBase->getBasePtrIndex() &&
1019  "Not relocating a derived object of the original base object");
1020  if (ToReplace->getBasePtrIndex() == ToReplace->getDerivedPtrIndex()) {
1021  // A duplicate relocate call. TODO: coalesce duplicates.
1022  continue;
1023  }
1024 
1025  if (RelocatedBase->getParent() != ToReplace->getParent()) {
1026  // Base and derived relocates are in different basic blocks.
1027  // In this case transform is only valid when base dominates derived
1028  // relocate. However it would be too expensive to check dominance
1029  // for each such relocate, so we skip the whole transformation.
1030  continue;
1031  }
1032 
1033  Value *Base = ToReplace->getBasePtr();
1034  auto Derived = dyn_cast<GetElementPtrInst>(ToReplace->getDerivedPtr());
1035  if (!Derived || Derived->getPointerOperand() != Base)
1036  continue;
1037 
1038  SmallVector<Value *, 2> OffsetV;
1039  if (!getGEPSmallConstantIntOffsetV(Derived, OffsetV))
1040  continue;
1041 
1042  // Create a Builder and replace the target callsite with a gep
1043  assert(RelocatedBase->getNextNode() &&
1044  "Should always have one since it's not a terminator");
1045 
1046  // Insert after RelocatedBase
1047  IRBuilder<> Builder(RelocatedBase->getNextNode());
1048  Builder.SetCurrentDebugLocation(ToReplace->getDebugLoc());
1049 
1050  // If gc_relocate does not match the actual type, cast it to the right type.
1051  // In theory, there must be a bitcast after gc_relocate if the type does not
1052  // match, and we should reuse it to get the derived pointer. But it could be
1053  // cases like this:
1054  // bb1:
1055  // ...
1056  // %g1 = call coldcc i8 addrspace(1)* @llvm.experimental.gc.relocate.p1i8(...)
1057  // br label %merge
1058  //
1059  // bb2:
1060  // ...
1061  // %g2 = call coldcc i8 addrspace(1)* @llvm.experimental.gc.relocate.p1i8(...)
1062  // br label %merge
1063  //
1064  // merge:
1065  // %p1 = phi i8 addrspace(1)* [ %g1, %bb1 ], [ %g2, %bb2 ]
1066  // %cast = bitcast i8 addrspace(1)* %p1 in to i32 addrspace(1)*
1067  //
1068  // In this case, we can not find the bitcast any more. So we insert a new bitcast
1069  // no matter there is already one or not. In this way, we can handle all cases, and
1070  // the extra bitcast should be optimized away in later passes.
1071  Value *ActualRelocatedBase = RelocatedBase;
1072  if (RelocatedBase->getType() != Base->getType()) {
1073  ActualRelocatedBase =
1074  Builder.CreateBitCast(RelocatedBase, Base->getType());
1075  }
1076  Value *Replacement = Builder.CreateGEP(
1077  Derived->getSourceElementType(), ActualRelocatedBase, makeArrayRef(OffsetV));
1078  Replacement->takeName(ToReplace);
1079  // If the newly generated derived pointer's type does not match the original derived
1080  // pointer's type, cast the new derived pointer to match it. Same reasoning as above.
1081  Value *ActualReplacement = Replacement;
1082  if (Replacement->getType() != ToReplace->getType()) {
1083  ActualReplacement =
1084  Builder.CreateBitCast(Replacement, ToReplace->getType());
1085  }
1086  ToReplace->replaceAllUsesWith(ActualReplacement);
1087  ToReplace->eraseFromParent();
1088 
1089  MadeChange = true;
1090  }
1091  return MadeChange;
1092 }
1093 
1094 // Turns this:
1095 //
1096 // %base = ...
1097 // %ptr = gep %base + 15
1098 // %tok = statepoint (%fun, i32 0, i32 0, i32 0, %base, %ptr)
1099 // %base' = relocate(%tok, i32 4, i32 4)
1100 // %ptr' = relocate(%tok, i32 4, i32 5)
1101 // %val = load %ptr'
1102 //
1103 // into this:
1104 //
1105 // %base = ...
1106 // %ptr = gep %base + 15
1107 // %tok = statepoint (%fun, i32 0, i32 0, i32 0, %base, %ptr)
1108 // %base' = gc.relocate(%tok, i32 4, i32 4)
1109 // %ptr' = gep %base' + 15
1110 // %val = load %ptr'
1111 bool CodeGenPrepare::simplifyOffsetableRelocate(Instruction &I) {
1112  bool MadeChange = false;
1113  SmallVector<GCRelocateInst *, 2> AllRelocateCalls;
1114 
1115  for (auto *U : I.users())
1116  if (GCRelocateInst *Relocate = dyn_cast<GCRelocateInst>(U))
1117  // Collect all the relocate calls associated with a statepoint
1118  AllRelocateCalls.push_back(Relocate);
1119 
1120  // We need atleast one base pointer relocation + one derived pointer
1121  // relocation to mangle
1122  if (AllRelocateCalls.size() < 2)
1123  return false;
1124 
1125  // RelocateInstMap is a mapping from the base relocate instruction to the
1126  // corresponding derived relocate instructions
1128  computeBaseDerivedRelocateMap(AllRelocateCalls, RelocateInstMap);
1129  if (RelocateInstMap.empty())
1130  return false;
1131 
1132  for (auto &Item : RelocateInstMap)
1133  // Item.first is the RelocatedBase to offset against
1134  // Item.second is the vector of Targets to replace
1135  MadeChange = simplifyRelocatesOffABase(Item.first, Item.second);
1136  return MadeChange;
1137 }
1138 
1139 /// SinkCast - Sink the specified cast instruction into its user blocks
1140 static bool SinkCast(CastInst *CI) {
1141  BasicBlock *DefBB = CI->getParent();
1142 
1143  /// InsertedCasts - Only insert a cast in each block once.
1144  DenseMap<BasicBlock*, CastInst*> InsertedCasts;
1145 
1146  bool MadeChange = false;
1147  for (Value::user_iterator UI = CI->user_begin(), E = CI->user_end();
1148  UI != E; ) {
1149  Use &TheUse = UI.getUse();
1150  Instruction *User = cast<Instruction>(*UI);
1151 
1152  // Figure out which BB this cast is used in. For PHI's this is the
1153  // appropriate predecessor block.
1154  BasicBlock *UserBB = User->getParent();
1155  if (PHINode *PN = dyn_cast<PHINode>(User)) {
1156  UserBB = PN->getIncomingBlock(TheUse);
1157  }
1158 
1159  // Preincrement use iterator so we don't invalidate it.
1160  ++UI;
1161 
1162  // The first insertion point of a block containing an EH pad is after the
1163  // pad. If the pad is the user, we cannot sink the cast past the pad.
1164  if (User->isEHPad())
1165  continue;
1166 
1167  // If the block selected to receive the cast is an EH pad that does not
1168  // allow non-PHI instructions before the terminator, we can't sink the
1169  // cast.
1170  if (UserBB->getTerminator()->isEHPad())
1171  continue;
1172 
1173  // If this user is in the same block as the cast, don't change the cast.
1174  if (UserBB == DefBB) continue;
1175 
1176  // If we have already inserted a cast into this block, use it.
1177  CastInst *&InsertedCast = InsertedCasts[UserBB];
1178 
1179  if (!InsertedCast) {
1180  BasicBlock::iterator InsertPt = UserBB->getFirstInsertionPt();
1181  assert(InsertPt != UserBB->end());
1182  InsertedCast = CastInst::Create(CI->getOpcode(), CI->getOperand(0),
1183  CI->getType(), "", &*InsertPt);
1184  }
1185 
1186  // Replace a use of the cast with a use of the new cast.
1187  TheUse = InsertedCast;
1188  MadeChange = true;
1189  ++NumCastUses;
1190  }
1191 
1192  // If we removed all uses, nuke the cast.
1193  if (CI->use_empty()) {
1194  salvageDebugInfo(*CI);
1195  CI->eraseFromParent();
1196  MadeChange = true;
1197  }
1198 
1199  return MadeChange;
1200 }
1201 
1202 /// If the specified cast instruction is a noop copy (e.g. it's casting from
1203 /// one pointer type to another, i32->i8 on PPC), sink it into user blocks to
1204 /// reduce the number of virtual registers that must be created and coalesced.
1205 ///
1206 /// Return true if any changes are made.
1208  const DataLayout &DL) {
1209  // Sink only "cheap" (or nop) address-space casts. This is a weaker condition
1210  // than sinking only nop casts, but is helpful on some platforms.
1211  if (auto *ASC = dyn_cast<AddrSpaceCastInst>(CI)) {
1212  if (!TLI.isCheapAddrSpaceCast(ASC->getSrcAddressSpace(),
1213  ASC->getDestAddressSpace()))
1214  return false;
1215  }
1216 
1217  // If this is a noop copy,
1218  EVT SrcVT = TLI.getValueType(DL, CI->getOperand(0)->getType());
1219  EVT DstVT = TLI.getValueType(DL, CI->getType());
1220 
1221  // This is an fp<->int conversion?
1222  if (SrcVT.isInteger() != DstVT.isInteger())
1223  return false;
1224 
1225  // If this is an extension, it will be a zero or sign extension, which
1226  // isn't a noop.
1227  if (SrcVT.bitsLT(DstVT)) return false;
1228 
1229  // If these values will be promoted, find out what they will be promoted
1230  // to. This helps us consider truncates on PPC as noop copies when they
1231  // are.
1232  if (TLI.getTypeAction(CI->getContext(), SrcVT) ==
1234  SrcVT = TLI.getTypeToTransformTo(CI->getContext(), SrcVT);
1235  if (TLI.getTypeAction(CI->getContext(), DstVT) ==
1237  DstVT = TLI.getTypeToTransformTo(CI->getContext(), DstVT);
1238 
1239  // If, after promotion, these are the same types, this is a noop copy.
1240  if (SrcVT != DstVT)
1241  return false;
1242 
1243  return SinkCast(CI);
1244 }
1245 
1246 /// Try to combine CI into a call to the llvm.uadd.with.overflow intrinsic if
1247 /// possible.
1248 ///
1249 /// Return true if any changes were made.
1251  Value *A, *B;
1252  Instruction *AddI;
1253  if (!match(CI,
1255  return false;
1256 
1257  Type *Ty = AddI->getType();
1258  if (!isa<IntegerType>(Ty))
1259  return false;
1260 
1261  // We don't want to move around uses of condition values this late, so we we
1262  // check if it is legal to create the call to the intrinsic in the basic
1263  // block containing the icmp:
1264 
1265  if (AddI->getParent() != CI->getParent() && !AddI->hasOneUse())
1266  return false;
1267 
1268 #ifndef NDEBUG
1269  // Someday m_UAddWithOverflow may get smarter, but this is a safe assumption
1270  // for now:
1271  if (AddI->hasOneUse())
1272  assert(*AddI->user_begin() == CI && "expected!");
1273 #endif
1274 
1275  Module *M = CI->getModule();
1276  Value *F = Intrinsic::getDeclaration(M, Intrinsic::uadd_with_overflow, Ty);
1277 
1278  auto *InsertPt = AddI->hasOneUse() ? CI : AddI;
1279 
1280  auto *UAddWithOverflow =
1281  CallInst::Create(F, {A, B}, "uadd.overflow", InsertPt);
1282  auto *UAdd = ExtractValueInst::Create(UAddWithOverflow, 0, "uadd", InsertPt);
1283  auto *Overflow =
1284  ExtractValueInst::Create(UAddWithOverflow, 1, "overflow", InsertPt);
1285 
1286  CI->replaceAllUsesWith(Overflow);
1287  AddI->replaceAllUsesWith(UAdd);
1288  CI->eraseFromParent();
1289  AddI->eraseFromParent();
1290  return true;
1291 }
1292 
1293 /// Sink the given CmpInst into user blocks to reduce the number of virtual
1294 /// registers that must be created and coalesced. This is a clear win except on
1295 /// targets with multiple condition code registers (PowerPC), where it might
1296 /// lose; some adjustment may be wanted there.
1297 ///
1298 /// Return true if any changes are made.
1299 static bool SinkCmpExpression(CmpInst *CI, const TargetLowering *TLI) {
1300  BasicBlock *DefBB = CI->getParent();
1301 
1302  // Avoid sinking soft-FP comparisons, since this can move them into a loop.
1303  if (TLI && TLI->useSoftFloat() && isa<FCmpInst>(CI))
1304  return false;
1305 
1306  // Only insert a cmp in each block once.
1307  DenseMap<BasicBlock*, CmpInst*> InsertedCmps;
1308 
1309  bool MadeChange = false;
1310  for (Value::user_iterator UI = CI->user_begin(), E = CI->user_end();
1311  UI != E; ) {
1312  Use &TheUse = UI.getUse();
1313  Instruction *User = cast<Instruction>(*UI);
1314 
1315  // Preincrement use iterator so we don't invalidate it.
1316  ++UI;
1317 
1318  // Don't bother for PHI nodes.
1319  if (isa<PHINode>(User))
1320  continue;
1321 
1322  // Figure out which BB this cmp is used in.
1323  BasicBlock *UserBB = User->getParent();
1324 
1325  // If this user is in the same block as the cmp, don't change the cmp.
1326  if (UserBB == DefBB) continue;
1327 
1328  // If we have already inserted a cmp into this block, use it.
1329  CmpInst *&InsertedCmp = InsertedCmps[UserBB];
1330 
1331  if (!InsertedCmp) {
1332  BasicBlock::iterator InsertPt = UserBB->getFirstInsertionPt();
1333  assert(InsertPt != UserBB->end());
1334  InsertedCmp =
1335  CmpInst::Create(CI->getOpcode(), CI->getPredicate(),
1336  CI->getOperand(0), CI->getOperand(1), "", &*InsertPt);
1337  // Propagate the debug info.
1338  InsertedCmp->setDebugLoc(CI->getDebugLoc());
1339  }
1340 
1341  // Replace a use of the cmp with a use of the new cmp.
1342  TheUse = InsertedCmp;
1343  MadeChange = true;
1344  ++NumCmpUses;
1345  }
1346 
1347  // If we removed all uses, nuke the cmp.
1348  if (CI->use_empty()) {
1349  CI->eraseFromParent();
1350  MadeChange = true;
1351  }
1352 
1353  return MadeChange;
1354 }
1355 
1356 static bool OptimizeCmpExpression(CmpInst *CI, const TargetLowering *TLI) {
1357  if (SinkCmpExpression(CI, TLI))
1358  return true;
1359 
1360  if (CombineUAddWithOverflow(CI))
1361  return true;
1362 
1363  return false;
1364 }
1365 
1366 /// Duplicate and sink the given 'and' instruction into user blocks where it is
1367 /// used in a compare to allow isel to generate better code for targets where
1368 /// this operation can be combined.
1369 ///
1370 /// Return true if any changes are made.
1372  const TargetLowering &TLI,
1373  SetOfInstrs &InsertedInsts) {
1374  // Double-check that we're not trying to optimize an instruction that was
1375  // already optimized by some other part of this pass.
1376  assert(!InsertedInsts.count(AndI) &&
1377  "Attempting to optimize already optimized and instruction");
1378  (void) InsertedInsts;
1379 
1380  // Nothing to do for single use in same basic block.
1381  if (AndI->hasOneUse() &&
1382  AndI->getParent() == cast<Instruction>(*AndI->user_begin())->getParent())
1383  return false;
1384 
1385  // Try to avoid cases where sinking/duplicating is likely to increase register
1386  // pressure.
1387  if (!isa<ConstantInt>(AndI->getOperand(0)) &&
1388  !isa<ConstantInt>(AndI->getOperand(1)) &&
1389  AndI->getOperand(0)->hasOneUse() && AndI->getOperand(1)->hasOneUse())
1390  return false;
1391 
1392  for (auto *U : AndI->users()) {
1393  Instruction *User = cast<Instruction>(U);
1394 
1395  // Only sink for and mask feeding icmp with 0.
1396  if (!isa<ICmpInst>(User))
1397  return false;
1398 
1399  auto *CmpC = dyn_cast<ConstantInt>(User->getOperand(1));
1400  if (!CmpC || !CmpC->isZero())
1401  return false;
1402  }
1403 
1404  if (!TLI.isMaskAndCmp0FoldingBeneficial(*AndI))
1405  return false;
1406 
1407  DEBUG(dbgs() << "found 'and' feeding only icmp 0;\n");
1408  DEBUG(AndI->getParent()->dump());
1409 
1410  // Push the 'and' into the same block as the icmp 0. There should only be
1411  // one (icmp (and, 0)) in each block, since CSE/GVN should have removed any
1412  // others, so we don't need to keep track of which BBs we insert into.
1413  for (Value::user_iterator UI = AndI->user_begin(), E = AndI->user_end();
1414  UI != E; ) {
1415  Use &TheUse = UI.getUse();
1416  Instruction *User = cast<Instruction>(*UI);
1417 
1418  // Preincrement use iterator so we don't invalidate it.
1419  ++UI;
1420 
1421  DEBUG(dbgs() << "sinking 'and' use: " << *User << "\n");
1422 
1423  // Keep the 'and' in the same place if the use is already in the same block.
1424  Instruction *InsertPt =
1425  User->getParent() == AndI->getParent() ? AndI : User;
1426  Instruction *InsertedAnd =
1427  BinaryOperator::Create(Instruction::And, AndI->getOperand(0),
1428  AndI->getOperand(1), "", InsertPt);
1429  // Propagate the debug info.
1430  InsertedAnd->setDebugLoc(AndI->getDebugLoc());
1431 
1432  // Replace a use of the 'and' with a use of the new 'and'.
1433  TheUse = InsertedAnd;
1434  ++NumAndUses;
1435  DEBUG(User->getParent()->dump());
1436  }
1437 
1438  // We removed all uses, nuke the and.
1439  AndI->eraseFromParent();
1440  return true;
1441 }
1442 
1443 /// Check if the candidates could be combined with a shift instruction, which
1444 /// includes:
1445 /// 1. Truncate instruction
1446 /// 2. And instruction and the imm is a mask of the low bits:
1447 /// imm & (imm+1) == 0
1449  if (!isa<TruncInst>(User)) {
1450  if (User->getOpcode() != Instruction::And ||
1451  !isa<ConstantInt>(User->getOperand(1)))
1452  return false;
1453 
1454  const APInt &Cimm = cast<ConstantInt>(User->getOperand(1))->getValue();
1455 
1456  if ((Cimm & (Cimm + 1)).getBoolValue())
1457  return false;
1458  }
1459  return true;
1460 }
1461 
1462 /// Sink both shift and truncate instruction to the use of truncate's BB.
1463 static bool
1466  const TargetLowering &TLI, const DataLayout &DL) {
1467  BasicBlock *UserBB = User->getParent();
1468  DenseMap<BasicBlock *, CastInst *> InsertedTruncs;
1469  TruncInst *TruncI = dyn_cast<TruncInst>(User);
1470  bool MadeChange = false;
1471 
1472  for (Value::user_iterator TruncUI = TruncI->user_begin(),
1473  TruncE = TruncI->user_end();
1474  TruncUI != TruncE;) {
1475 
1476  Use &TruncTheUse = TruncUI.getUse();
1477  Instruction *TruncUser = cast<Instruction>(*TruncUI);
1478  // Preincrement use iterator so we don't invalidate it.
1479 
1480  ++TruncUI;
1481 
1482  int ISDOpcode = TLI.InstructionOpcodeToISD(TruncUser->getOpcode());
1483  if (!ISDOpcode)
1484  continue;
1485 
1486  // If the use is actually a legal node, there will not be an
1487  // implicit truncate.
1488  // FIXME: always querying the result type is just an
1489  // approximation; some nodes' legality is determined by the
1490  // operand or other means. There's no good way to find out though.
1491  if (TLI.isOperationLegalOrCustom(
1492  ISDOpcode, TLI.getValueType(DL, TruncUser->getType(), true)))
1493  continue;
1494 
1495  // Don't bother for PHI nodes.
1496  if (isa<PHINode>(TruncUser))
1497  continue;
1498 
1499  BasicBlock *TruncUserBB = TruncUser->getParent();
1500 
1501  if (UserBB == TruncUserBB)
1502  continue;
1503 
1504  BinaryOperator *&InsertedShift = InsertedShifts[TruncUserBB];
1505  CastInst *&InsertedTrunc = InsertedTruncs[TruncUserBB];
1506 
1507  if (!InsertedShift && !InsertedTrunc) {
1508  BasicBlock::iterator InsertPt = TruncUserBB->getFirstInsertionPt();
1509  assert(InsertPt != TruncUserBB->end());
1510  // Sink the shift
1511  if (ShiftI->getOpcode() == Instruction::AShr)
1512  InsertedShift = BinaryOperator::CreateAShr(ShiftI->getOperand(0), CI,
1513  "", &*InsertPt);
1514  else
1515  InsertedShift = BinaryOperator::CreateLShr(ShiftI->getOperand(0), CI,
1516  "", &*InsertPt);
1517 
1518  // Sink the trunc
1519  BasicBlock::iterator TruncInsertPt = TruncUserBB->getFirstInsertionPt();
1520  TruncInsertPt++;
1521  assert(TruncInsertPt != TruncUserBB->end());
1522 
1523  InsertedTrunc = CastInst::Create(TruncI->getOpcode(), InsertedShift,
1524  TruncI->getType(), "", &*TruncInsertPt);
1525 
1526  MadeChange = true;
1527 
1528  TruncTheUse = InsertedTrunc;
1529  }
1530  }
1531  return MadeChange;
1532 }
1533 
1534 /// Sink the shift *right* instruction into user blocks if the uses could
1535 /// potentially be combined with this shift instruction and generate BitExtract
1536 /// instruction. It will only be applied if the architecture supports BitExtract
1537 /// instruction. Here is an example:
1538 /// BB1:
1539 /// %x.extract.shift = lshr i64 %arg1, 32
1540 /// BB2:
1541 /// %x.extract.trunc = trunc i64 %x.extract.shift to i16
1542 /// ==>
1543 ///
1544 /// BB2:
1545 /// %x.extract.shift.1 = lshr i64 %arg1, 32
1546 /// %x.extract.trunc = trunc i64 %x.extract.shift.1 to i16
1547 ///
1548 /// CodeGen will recoginze the pattern in BB2 and generate BitExtract
1549 /// instruction.
1550 /// Return true if any changes are made.
1552  const TargetLowering &TLI,
1553  const DataLayout &DL) {
1554  BasicBlock *DefBB = ShiftI->getParent();
1555 
1556  /// Only insert instructions in each block once.
1558 
1559  bool shiftIsLegal = TLI.isTypeLegal(TLI.getValueType(DL, ShiftI->getType()));
1560 
1561  bool MadeChange = false;
1562  for (Value::user_iterator UI = ShiftI->user_begin(), E = ShiftI->user_end();
1563  UI != E;) {
1564  Use &TheUse = UI.getUse();
1565  Instruction *User = cast<Instruction>(*UI);
1566  // Preincrement use iterator so we don't invalidate it.
1567  ++UI;
1568 
1569  // Don't bother for PHI nodes.
1570  if (isa<PHINode>(User))
1571  continue;
1572 
1573  if (!isExtractBitsCandidateUse(User))
1574  continue;
1575 
1576  BasicBlock *UserBB = User->getParent();
1577 
1578  if (UserBB == DefBB) {
1579  // If the shift and truncate instruction are in the same BB. The use of
1580  // the truncate(TruncUse) may still introduce another truncate if not
1581  // legal. In this case, we would like to sink both shift and truncate
1582  // instruction to the BB of TruncUse.
1583  // for example:
1584  // BB1:
1585  // i64 shift.result = lshr i64 opnd, imm
1586  // trunc.result = trunc shift.result to i16
1587  //
1588  // BB2:
1589  // ----> We will have an implicit truncate here if the architecture does
1590  // not have i16 compare.
1591  // cmp i16 trunc.result, opnd2
1592  //
1593  if (isa<TruncInst>(User) && shiftIsLegal
1594  // If the type of the truncate is legal, no trucate will be
1595  // introduced in other basic blocks.
1596  &&
1597  (!TLI.isTypeLegal(TLI.getValueType(DL, User->getType()))))
1598  MadeChange =
1599  SinkShiftAndTruncate(ShiftI, User, CI, InsertedShifts, TLI, DL);
1600 
1601  continue;
1602  }
1603  // If we have already inserted a shift into this block, use it.
1604  BinaryOperator *&InsertedShift = InsertedShifts[UserBB];
1605 
1606  if (!InsertedShift) {
1607  BasicBlock::iterator InsertPt = UserBB->getFirstInsertionPt();
1608  assert(InsertPt != UserBB->end());
1609 
1610  if (ShiftI->getOpcode() == Instruction::AShr)
1611  InsertedShift = BinaryOperator::CreateAShr(ShiftI->getOperand(0), CI,
1612  "", &*InsertPt);
1613  else
1614  InsertedShift = BinaryOperator::CreateLShr(ShiftI->getOperand(0), CI,
1615  "", &*InsertPt);
1616 
1617  MadeChange = true;
1618  }
1619 
1620  // Replace a use of the shift with a use of the new shift.
1621  TheUse = InsertedShift;
1622  }
1623 
1624  // If we removed all uses, nuke the shift.
1625  if (ShiftI->use_empty())
1626  ShiftI->eraseFromParent();
1627 
1628  return MadeChange;
1629 }
1630 
1631 /// If counting leading or trailing zeros is an expensive operation and a zero
1632 /// input is defined, add a check for zero to avoid calling the intrinsic.
1633 ///
1634 /// We want to transform:
1635 /// %z = call i64 @llvm.cttz.i64(i64 %A, i1 false)
1636 ///
1637 /// into:
1638 /// entry:
1639 /// %cmpz = icmp eq i64 %A, 0
1640 /// br i1 %cmpz, label %cond.end, label %cond.false
1641 /// cond.false:
1642 /// %z = call i64 @llvm.cttz.i64(i64 %A, i1 true)
1643 /// br label %cond.end
1644 /// cond.end:
1645 /// %ctz = phi i64 [ 64, %entry ], [ %z, %cond.false ]
1646 ///
1647 /// If the transform is performed, return true and set ModifiedDT to true.
1648 static bool despeculateCountZeros(IntrinsicInst *CountZeros,
1649  const TargetLowering *TLI,
1650  const DataLayout *DL,
1651  bool &ModifiedDT) {
1652  if (!TLI || !DL)
1653  return false;
1654 
1655  // If a zero input is undefined, it doesn't make sense to despeculate that.
1656  if (match(CountZeros->getOperand(1), m_One()))
1657  return false;
1658 
1659  // If it's cheap to speculate, there's nothing to do.
1660  auto IntrinsicID = CountZeros->getIntrinsicID();
1661  if ((IntrinsicID == Intrinsic::cttz && TLI->isCheapToSpeculateCttz()) ||
1662  (IntrinsicID == Intrinsic::ctlz && TLI->isCheapToSpeculateCtlz()))
1663  return false;
1664 
1665  // Only handle legal scalar cases. Anything else requires too much work.
1666  Type *Ty = CountZeros->getType();
1667  unsigned SizeInBits = Ty->getPrimitiveSizeInBits();
1668  if (Ty->isVectorTy() || SizeInBits > DL->getLargestLegalIntTypeSizeInBits())
1669  return false;
1670 
1671  // The intrinsic will be sunk behind a compare against zero and branch.
1672  BasicBlock *StartBlock = CountZeros->getParent();
1673  BasicBlock *CallBlock = StartBlock->splitBasicBlock(CountZeros, "cond.false");
1674 
1675  // Create another block after the count zero intrinsic. A PHI will be added
1676  // in this block to select the result of the intrinsic or the bit-width
1677  // constant if the input to the intrinsic is zero.
1678  BasicBlock::iterator SplitPt = ++(BasicBlock::iterator(CountZeros));
1679  BasicBlock *EndBlock = CallBlock->splitBasicBlock(SplitPt, "cond.end");
1680 
1681  // Set up a builder to create a compare, conditional branch, and PHI.
1682  IRBuilder<> Builder(CountZeros->getContext());
1683  Builder.SetInsertPoint(StartBlock->getTerminator());
1684  Builder.SetCurrentDebugLocation(CountZeros->getDebugLoc());
1685 
1686  // Replace the unconditional branch that was created by the first split with
1687  // a compare against zero and a conditional branch.
1688  Value *Zero = Constant::getNullValue(Ty);
1689  Value *Cmp = Builder.CreateICmpEQ(CountZeros->getOperand(0), Zero, "cmpz");
1690  Builder.CreateCondBr(Cmp, EndBlock, CallBlock);
1691  StartBlock->getTerminator()->eraseFromParent();
1692 
1693  // Create a PHI in the end block to select either the output of the intrinsic
1694  // or the bit width of the operand.
1695  Builder.SetInsertPoint(&EndBlock->front());
1696  PHINode *PN = Builder.CreatePHI(Ty, 2, "ctz");
1697  CountZeros->replaceAllUsesWith(PN);
1698  Value *BitWidth = Builder.getInt(APInt(SizeInBits, SizeInBits));
1699  PN->addIncoming(BitWidth, StartBlock);
1700  PN->addIncoming(CountZeros, CallBlock);
1701 
1702  // We are explicitly handling the zero case, so we can set the intrinsic's
1703  // undefined zero argument to 'true'. This will also prevent reprocessing the
1704  // intrinsic; we only despeculate when a zero input is defined.
1705  CountZeros->setArgOperand(1, Builder.getTrue());
1706  ModifiedDT = true;
1707  return true;
1708 }
1709 
1710 bool CodeGenPrepare::optimizeCallInst(CallInst *CI, bool &ModifiedDT) {
1711  BasicBlock *BB = CI->getParent();
1712 
1713  // Lower inline assembly if we can.
1714  // If we found an inline asm expession, and if the target knows how to
1715  // lower it to normal LLVM code, do so now.
1716  if (TLI && isa<InlineAsm>(CI->getCalledValue())) {
1717  if (TLI->ExpandInlineAsm(CI)) {
1718  // Avoid invalidating the iterator.
1719  CurInstIterator = BB->begin();
1720  // Avoid processing instructions out of order, which could cause
1721  // reuse before a value is defined.
1722  SunkAddrs.clear();
1723  return true;
1724  }
1725  // Sink address computing for memory operands into the block.
1726  if (optimizeInlineAsmInst(CI))
1727  return true;
1728  }
1729 
1730  // Align the pointer arguments to this call if the target thinks it's a good
1731  // idea
1732  unsigned MinSize, PrefAlign;
1733  if (TLI && TLI->shouldAlignPointerArgs(CI, MinSize, PrefAlign)) {
1734  for (auto &Arg : CI->arg_operands()) {
1735  // We want to align both objects whose address is used directly and
1736  // objects whose address is used in casts and GEPs, though it only makes
1737  // sense for GEPs if the offset is a multiple of the desired alignment and
1738  // if size - offset meets the size threshold.
1739  if (!Arg->getType()->isPointerTy())
1740  continue;
1741  APInt Offset(DL->getPointerSizeInBits(
1742  cast<PointerType>(Arg->getType())->getAddressSpace()),
1743  0);
1745  uint64_t Offset2 = Offset.getLimitedValue();
1746  if ((Offset2 & (PrefAlign-1)) != 0)
1747  continue;
1748  AllocaInst *AI;
1749  if ((AI = dyn_cast<AllocaInst>(Val)) && AI->getAlignment() < PrefAlign &&
1750  DL->getTypeAllocSize(AI->getAllocatedType()) >= MinSize + Offset2)
1751  AI->setAlignment(PrefAlign);
1752  // Global variables can only be aligned if they are defined in this
1753  // object (i.e. they are uniquely initialized in this object), and
1754  // over-aligning global variables that have an explicit section is
1755  // forbidden.
1756  GlobalVariable *GV;
1757  if ((GV = dyn_cast<GlobalVariable>(Val)) && GV->canIncreaseAlignment() &&
1758  GV->getPointerAlignment(*DL) < PrefAlign &&
1759  DL->getTypeAllocSize(GV->getValueType()) >=
1760  MinSize + Offset2)
1761  GV->setAlignment(PrefAlign);
1762  }
1763  // If this is a memcpy (or similar) then we may be able to improve the
1764  // alignment
1765  if (MemIntrinsic *MI = dyn_cast<MemIntrinsic>(CI)) {
1766  unsigned Align = getKnownAlignment(MI->getDest(), *DL);
1767  if (MemTransferInst *MTI = dyn_cast<MemTransferInst>(MI))
1768  Align = std::min(Align, getKnownAlignment(MTI->getSource(), *DL));
1769  if (Align > MI->getAlignment())
1770  MI->setAlignment(ConstantInt::get(MI->getAlignmentType(), Align));
1771  }
1772  }
1773 
1774  // If we have a cold call site, try to sink addressing computation into the
1775  // cold block. This interacts with our handling for loads and stores to
1776  // ensure that we can fold all uses of a potential addressing computation
1777  // into their uses. TODO: generalize this to work over profiling data
1778  if (!OptSize && CI->hasFnAttr(Attribute::Cold))
1779  for (auto &Arg : CI->arg_operands()) {
1780  if (!Arg->getType()->isPointerTy())
1781  continue;
1782  unsigned AS = Arg->getType()->getPointerAddressSpace();
1783  return optimizeMemoryInst(CI, Arg, Arg->getType(), AS);
1784  }
1785 
1787  if (II) {
1788  switch (II->getIntrinsicID()) {
1789  default: break;
1790  case Intrinsic::objectsize: {
1791  // Lower all uses of llvm.objectsize.*
1792  ConstantInt *RetVal =
1793  lowerObjectSizeCall(II, *DL, TLInfo, /*MustSucceed=*/true);
1794  // Substituting this can cause recursive simplifications, which can
1795  // invalidate our iterator. Use a WeakTrackingVH to hold onto it in case
1796  // this
1797  // happens.
1798  Value *CurValue = &*CurInstIterator;
1799  WeakTrackingVH IterHandle(CurValue);
1800 
1801  replaceAndRecursivelySimplify(CI, RetVal, TLInfo, nullptr);
1802 
1803  // If the iterator instruction was recursively deleted, start over at the
1804  // start of the block.
1805  if (IterHandle != CurValue) {
1806  CurInstIterator = BB->begin();
1807  SunkAddrs.clear();
1808  }
1809  return true;
1810  }
1811  case Intrinsic::aarch64_stlxr:
1812  case Intrinsic::aarch64_stxr: {
1813  ZExtInst *ExtVal = dyn_cast<ZExtInst>(CI->getArgOperand(0));
1814  if (!ExtVal || !ExtVal->hasOneUse() ||
1815  ExtVal->getParent() == CI->getParent())
1816  return false;
1817  // Sink a zext feeding stlxr/stxr before it, so it can be folded into it.
1818  ExtVal->moveBefore(CI);
1819  // Mark this instruction as "inserted by CGP", so that other
1820  // optimizations don't touch it.
1821  InsertedInsts.insert(ExtVal);
1822  return true;
1823  }
1824  case Intrinsic::invariant_group_barrier:
1825  II->replaceAllUsesWith(II->getArgOperand(0));
1826  II->eraseFromParent();
1827  return true;
1828 
1829  case Intrinsic::cttz:
1830  case Intrinsic::ctlz:
1831  // If counting zeros is expensive, try to avoid it.
1832  return despeculateCountZeros(II, TLI, DL, ModifiedDT);
1833  }
1834 
1835  if (TLI) {
1836  SmallVector<Value*, 2> PtrOps;
1837  Type *AccessTy;
1838  if (TLI->getAddrModeArguments(II, PtrOps, AccessTy))
1839  while (!PtrOps.empty()) {
1840  Value *PtrVal = PtrOps.pop_back_val();
1841  unsigned AS = PtrVal->getType()->getPointerAddressSpace();
1842  if (optimizeMemoryInst(II, PtrVal, AccessTy, AS))
1843  return true;
1844  }
1845  }
1846  }
1847 
1848  // From here on out we're working with named functions.
1849  if (!CI->getCalledFunction()) return false;
1850 
1851  // Lower all default uses of _chk calls. This is very similar
1852  // to what InstCombineCalls does, but here we are only lowering calls
1853  // to fortified library functions (e.g. __memcpy_chk) that have the default
1854  // "don't know" as the objectsize. Anything else should be left alone.
1855  FortifiedLibCallSimplifier Simplifier(TLInfo, true);
1856  if (Value *V = Simplifier.optimizeCall(CI)) {
1857  CI->replaceAllUsesWith(V);
1858  CI->eraseFromParent();
1859  return true;
1860  }
1861 
1862  return false;
1863 }
1864 
1865 /// Look for opportunities to duplicate return instructions to the predecessor
1866 /// to enable tail call optimizations. The case it is currently looking for is:
1867 /// @code
1868 /// bb0:
1869 /// %tmp0 = tail call i32 @f0()
1870 /// br label %return
1871 /// bb1:
1872 /// %tmp1 = tail call i32 @f1()
1873 /// br label %return
1874 /// bb2:
1875 /// %tmp2 = tail call i32 @f2()
1876 /// br label %return
1877 /// return:
1878 /// %retval = phi i32 [ %tmp0, %bb0 ], [ %tmp1, %bb1 ], [ %tmp2, %bb2 ]
1879 /// ret i32 %retval
1880 /// @endcode
1881 ///
1882 /// =>
1883 ///
1884 /// @code
1885 /// bb0:
1886 /// %tmp0 = tail call i32 @f0()
1887 /// ret i32 %tmp0
1888 /// bb1:
1889 /// %tmp1 = tail call i32 @f1()
1890 /// ret i32 %tmp1
1891 /// bb2:
1892 /// %tmp2 = tail call i32 @f2()
1893 /// ret i32 %tmp2
1894 /// @endcode
1895 bool CodeGenPrepare::dupRetToEnableTailCallOpts(BasicBlock *BB) {
1896  if (!TLI)
1897  return false;
1898 
1899  ReturnInst *RetI = dyn_cast<ReturnInst>(BB->getTerminator());
1900  if (!RetI)
1901  return false;
1902 
1903  PHINode *PN = nullptr;
1904  BitCastInst *BCI = nullptr;
1905  Value *V = RetI->getReturnValue();
1906  if (V) {
1907  BCI = dyn_cast<BitCastInst>(V);
1908  if (BCI)
1909  V = BCI->getOperand(0);
1910 
1911  PN = dyn_cast<PHINode>(V);
1912  if (!PN)
1913  return false;
1914  }
1915 
1916  if (PN && PN->getParent() != BB)
1917  return false;
1918 
1919  // Make sure there are no instructions between the PHI and return, or that the
1920  // return is the first instruction in the block.
1921  if (PN) {
1922  BasicBlock::iterator BI = BB->begin();
1923  do { ++BI; } while (isa<DbgInfoIntrinsic>(BI));
1924  if (&*BI == BCI)
1925  // Also skip over the bitcast.
1926  ++BI;
1927  if (&*BI != RetI)
1928  return false;
1929  } else {
1930  BasicBlock::iterator BI = BB->begin();
1931  while (isa<DbgInfoIntrinsic>(BI)) ++BI;
1932  if (&*BI != RetI)
1933  return false;
1934  }
1935 
1936  /// Only dup the ReturnInst if the CallInst is likely to be emitted as a tail
1937  /// call.
1938  const Function *F = BB->getParent();
1939  SmallVector<CallInst*, 4> TailCalls;
1940  if (PN) {
1941  for (unsigned I = 0, E = PN->getNumIncomingValues(); I != E; ++I) {
1942  CallInst *CI = dyn_cast<CallInst>(PN->getIncomingValue(I));
1943  // Make sure the phi value is indeed produced by the tail call.
1944  if (CI && CI->hasOneUse() && CI->getParent() == PN->getIncomingBlock(I) &&
1945  TLI->mayBeEmittedAsTailCall(CI) &&
1946  attributesPermitTailCall(F, CI, RetI, *TLI))
1947  TailCalls.push_back(CI);
1948  }
1949  } else {
1950  SmallPtrSet<BasicBlock*, 4> VisitedBBs;
1951  for (pred_iterator PI = pred_begin(BB), PE = pred_end(BB); PI != PE; ++PI) {
1952  if (!VisitedBBs.insert(*PI).second)
1953  continue;
1954 
1955  BasicBlock::InstListType &InstList = (*PI)->getInstList();
1956  BasicBlock::InstListType::reverse_iterator RI = InstList.rbegin();
1957  BasicBlock::InstListType::reverse_iterator RE = InstList.rend();
1958  do { ++RI; } while (RI != RE && isa<DbgInfoIntrinsic>(&*RI));
1959  if (RI == RE)
1960  continue;
1961 
1962  CallInst *CI = dyn_cast<CallInst>(&*RI);
1963  if (CI && CI->use_empty() && TLI->mayBeEmittedAsTailCall(CI) &&
1964  attributesPermitTailCall(F, CI, RetI, *TLI))
1965  TailCalls.push_back(CI);
1966  }
1967  }
1968 
1969  bool Changed = false;
1970  for (unsigned i = 0, e = TailCalls.size(); i != e; ++i) {
1971  CallInst *CI = TailCalls[i];
1972  CallSite CS(CI);
1973 
1974  // Conservatively require the attributes of the call to match those of the
1975  // return. Ignore noalias because it doesn't affect the call sequence.
1976  AttributeList CalleeAttrs = CS.getAttributes();
1977  if (AttrBuilder(CalleeAttrs, AttributeList::ReturnIndex)
1978  .removeAttribute(Attribute::NoAlias) !=
1980  .removeAttribute(Attribute::NoAlias))
1981  continue;
1982 
1983  // Make sure the call instruction is followed by an unconditional branch to
1984  // the return block.
1985  BasicBlock *CallBB = CI->getParent();
1986  BranchInst *BI = dyn_cast<BranchInst>(CallBB->getTerminator());
1987  if (!BI || !BI->isUnconditional() || BI->getSuccessor(0) != BB)
1988  continue;
1989 
1990  // Duplicate the return into CallBB.
1991  (void)FoldReturnIntoUncondBranch(RetI, BB, CallBB);
1992  ModifiedDT = Changed = true;
1993  ++NumRetsDup;
1994  }
1995 
1996  // If we eliminated all predecessors of the block, delete the block now.
1997  if (Changed && !BB->hasAddressTaken() && pred_begin(BB) == pred_end(BB))
1998  BB->eraseFromParent();
1999 
2000  return Changed;
2001 }
2002 
2003 //===----------------------------------------------------------------------===//
2004 // Memory Optimization
2005 //===----------------------------------------------------------------------===//
2006 
2007 namespace {
2008 
2009 /// This is an extended version of TargetLowering::AddrMode
2010 /// which holds actual Value*'s for register values.
2011 struct ExtAddrMode : public TargetLowering::AddrMode {
2012  Value *BaseReg = nullptr;
2013  Value *ScaledReg = nullptr;
2014  Value *OriginalValue = nullptr;
2015 
2016  enum FieldName {
2017  NoField = 0x00,
2018  BaseRegField = 0x01,
2019  BaseGVField = 0x02,
2020  BaseOffsField = 0x04,
2021  ScaledRegField = 0x08,
2022  ScaleField = 0x10,
2023  MultipleFields = 0xff
2024  };
2025 
2026  ExtAddrMode() = default;
2027 
2028  void print(raw_ostream &OS) const;
2029  void dump() const;
2030 
2031  FieldName compare(const ExtAddrMode &other) {
2032  // First check that the types are the same on each field, as differing types
2033  // is something we can't cope with later on.
2034  if (BaseReg && other.BaseReg &&
2035  BaseReg->getType() != other.BaseReg->getType())
2036  return MultipleFields;
2037  if (BaseGV && other.BaseGV &&
2038  BaseGV->getType() != other.BaseGV->getType())
2039  return MultipleFields;
2040  if (ScaledReg && other.ScaledReg &&
2041  ScaledReg->getType() != other.ScaledReg->getType())
2042  return MultipleFields;
2043 
2044  // Check each field to see if it differs.
2045  unsigned Result = NoField;
2046  if (BaseReg != other.BaseReg)
2047  Result |= BaseRegField;
2048  if (BaseGV != other.BaseGV)
2049  Result |= BaseGVField;
2050  if (BaseOffs != other.BaseOffs)
2051  Result |= BaseOffsField;
2052  if (ScaledReg != other.ScaledReg)
2053  Result |= ScaledRegField;
2054  // Don't count 0 as being a different scale, because that actually means
2055  // unscaled (which will already be counted by having no ScaledReg).
2056  if (Scale && other.Scale && Scale != other.Scale)
2057  Result |= ScaleField;
2058 
2059  if (countPopulation(Result) > 1)
2060  return MultipleFields;
2061  else
2062  return static_cast<FieldName>(Result);
2063  }
2064 
2065  // AddrModes with a baseReg or gv where the reg/gv is
2066  // the only populated field are trivial.
2067  bool isTrivial() {
2068  if (BaseGV && !BaseOffs && !Scale && !BaseReg)
2069  return true;
2070 
2071  if (!BaseGV && !BaseOffs && !Scale && BaseReg)
2072  return true;
2073 
2074  return false;
2075  }
2076 };
2077 
2078 } // end anonymous namespace
2079 
2080 #ifndef NDEBUG
2081 static inline raw_ostream &operator<<(raw_ostream &OS, const ExtAddrMode &AM) {
2082  AM.print(OS);
2083  return OS;
2084 }
2085 #endif
2086 
2087 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
2088 void ExtAddrMode::print(raw_ostream &OS) const {
2089  bool NeedPlus = false;
2090  OS << "[";
2091  if (BaseGV) {
2092  OS << (NeedPlus ? " + " : "")
2093  << "GV:";
2094  BaseGV->printAsOperand(OS, /*PrintType=*/false);
2095  NeedPlus = true;
2096  }
2097 
2098  if (BaseOffs) {
2099  OS << (NeedPlus ? " + " : "")
2100  << BaseOffs;
2101  NeedPlus = true;
2102  }
2103 
2104  if (BaseReg) {
2105  OS << (NeedPlus ? " + " : "")
2106  << "Base:";
2107  BaseReg->printAsOperand(OS, /*PrintType=*/false);
2108  NeedPlus = true;
2109  }
2110  if (Scale) {
2111  OS << (NeedPlus ? " + " : "")
2112  << Scale << "*";
2113  ScaledReg->printAsOperand(OS, /*PrintType=*/false);
2114  }
2115 
2116  OS << ']';
2117 }
2118 
2119 LLVM_DUMP_METHOD void ExtAddrMode::dump() const {
2120  print(dbgs());
2121  dbgs() << '\n';
2122 }
2123 #endif
2124 
2125 namespace {
2126 
2127 /// \brief This class provides transaction based operation on the IR.
2128 /// Every change made through this class is recorded in the internal state and
2129 /// can be undone (rollback) until commit is called.
2130 class TypePromotionTransaction {
2131  /// \brief This represents the common interface of the individual transaction.
2132  /// Each class implements the logic for doing one specific modification on
2133  /// the IR via the TypePromotionTransaction.
2134  class TypePromotionAction {
2135  protected:
2136  /// The Instruction modified.
2137  Instruction *Inst;
2138 
2139  public:
2140  /// \brief Constructor of the action.
2141  /// The constructor performs the related action on the IR.
2142  TypePromotionAction(Instruction *Inst) : Inst(Inst) {}
2143 
2144  virtual ~TypePromotionAction() = default;
2145 
2146  /// \brief Undo the modification done by this action.
2147  /// When this method is called, the IR must be in the same state as it was
2148  /// before this action was applied.
2149  /// \pre Undoing the action works if and only if the IR is in the exact same
2150  /// state as it was directly after this action was applied.
2151  virtual void undo() = 0;
2152 
2153  /// \brief Advocate every change made by this action.
2154  /// When the results on the IR of the action are to be kept, it is important
2155  /// to call this function, otherwise hidden information may be kept forever.
2156  virtual void commit() {
2157  // Nothing to be done, this action is not doing anything.
2158  }
2159  };
2160 
2161  /// \brief Utility to remember the position of an instruction.
2162  class InsertionHandler {
2163  /// Position of an instruction.
2164  /// Either an instruction:
2165  /// - Is the first in a basic block: BB is used.
2166  /// - Has a previous instructon: PrevInst is used.
2167  union {
2168  Instruction *PrevInst;
2169  BasicBlock *BB;
2170  } Point;
2171 
2172  /// Remember whether or not the instruction had a previous instruction.
2173  bool HasPrevInstruction;
2174 
2175  public:
2176  /// \brief Record the position of \p Inst.
2177  InsertionHandler(Instruction *Inst) {
2178  BasicBlock::iterator It = Inst->getIterator();
2179  HasPrevInstruction = (It != (Inst->getParent()->begin()));
2180  if (HasPrevInstruction)
2181  Point.PrevInst = &*--It;
2182  else
2183  Point.BB = Inst->getParent();
2184  }
2185 
2186  /// \brief Insert \p Inst at the recorded position.
2187  void insert(Instruction *Inst) {
2188  if (HasPrevInstruction) {
2189  if (Inst->getParent())
2190  Inst->removeFromParent();
2191  Inst->insertAfter(Point.PrevInst);
2192  } else {
2193  Instruction *Position = &*Point.BB->getFirstInsertionPt();
2194  if (Inst->getParent())
2195  Inst->moveBefore(Position);
2196  else
2197  Inst->insertBefore(Position);
2198  }
2199  }
2200  };
2201 
2202  /// \brief Move an instruction before another.
2203  class InstructionMoveBefore : public TypePromotionAction {
2204  /// Original position of the instruction.
2205  InsertionHandler Position;
2206 
2207  public:
2208  /// \brief Move \p Inst before \p Before.
2209  InstructionMoveBefore(Instruction *Inst, Instruction *Before)
2210  : TypePromotionAction(Inst), Position(Inst) {
2211  DEBUG(dbgs() << "Do: move: " << *Inst << "\nbefore: " << *Before << "\n");
2212  Inst->moveBefore(Before);
2213  }
2214 
2215  /// \brief Move the instruction back to its original position.
2216  void undo() override {
2217  DEBUG(dbgs() << "Undo: moveBefore: " << *Inst << "\n");
2218  Position.insert(Inst);
2219  }
2220  };
2221 
2222  /// \brief Set the operand of an instruction with a new value.
2223  class OperandSetter : public TypePromotionAction {
2224  /// Original operand of the instruction.
2225  Value *Origin;
2226 
2227  /// Index of the modified instruction.
2228  unsigned Idx;
2229 
2230  public:
2231  /// \brief Set \p Idx operand of \p Inst with \p NewVal.
2232  OperandSetter(Instruction *Inst, unsigned Idx, Value *NewVal)
2233  : TypePromotionAction(Inst), Idx(Idx) {
2234  DEBUG(dbgs() << "Do: setOperand: " << Idx << "\n"
2235  << "for:" << *Inst << "\n"
2236  << "with:" << *NewVal << "\n");
2237  Origin = Inst->getOperand(Idx);
2238  Inst->setOperand(Idx, NewVal);
2239  }
2240 
2241  /// \brief Restore the original value of the instruction.
2242  void undo() override {
2243  DEBUG(dbgs() << "Undo: setOperand:" << Idx << "\n"
2244  << "for: " << *Inst << "\n"
2245  << "with: " << *Origin << "\n");
2246  Inst->setOperand(Idx, Origin);
2247  }
2248  };
2249 
2250  /// \brief Hide the operands of an instruction.
2251  /// Do as if this instruction was not using any of its operands.
2252  class OperandsHider : public TypePromotionAction {
2253  /// The list of original operands.
2254  SmallVector<Value *, 4> OriginalValues;
2255 
2256  public:
2257  /// \brief Remove \p Inst from the uses of the operands of \p Inst.
2258  OperandsHider(Instruction *Inst) : TypePromotionAction(Inst) {
2259  DEBUG(dbgs() << "Do: OperandsHider: " << *Inst << "\n");
2260  unsigned NumOpnds = Inst->getNumOperands();
2261  OriginalValues.reserve(NumOpnds);
2262  for (unsigned It = 0; It < NumOpnds; ++It) {
2263  // Save the current operand.
2264  Value *Val = Inst->getOperand(It);
2265  OriginalValues.push_back(Val);
2266  // Set a dummy one.
2267  // We could use OperandSetter here, but that would imply an overhead
2268  // that we are not willing to pay.
2269  Inst->setOperand(It, UndefValue::get(Val->getType()));
2270  }
2271  }
2272 
2273  /// \brief Restore the original list of uses.
2274  void undo() override {
2275  DEBUG(dbgs() << "Undo: OperandsHider: " << *Inst << "\n");
2276  for (unsigned It = 0, EndIt = OriginalValues.size(); It != EndIt; ++It)
2277  Inst->setOperand(It, OriginalValues[It]);
2278  }
2279  };
2280 
2281  /// \brief Build a truncate instruction.
2282  class TruncBuilder : public TypePromotionAction {
2283  Value *Val;
2284 
2285  public:
2286  /// \brief Build a truncate instruction of \p Opnd producing a \p Ty
2287  /// result.
2288  /// trunc Opnd to Ty.
2289  TruncBuilder(Instruction *Opnd, Type *Ty) : TypePromotionAction(Opnd) {
2290  IRBuilder<> Builder(Opnd);
2291  Val = Builder.CreateTrunc(Opnd, Ty, "promoted");
2292  DEBUG(dbgs() << "Do: TruncBuilder: " << *Val << "\n");
2293  }
2294 
2295  /// \brief Get the built value.
2296  Value *getBuiltValue() { return Val; }
2297 
2298  /// \brief Remove the built instruction.
2299  void undo() override {
2300  DEBUG(dbgs() << "Undo: TruncBuilder: " << *Val << "\n");
2301  if (Instruction *IVal = dyn_cast<Instruction>(Val))
2302  IVal->eraseFromParent();
2303  }
2304  };
2305 
2306  /// \brief Build a sign extension instruction.
2307  class SExtBuilder : public TypePromotionAction {
2308  Value *Val;
2309 
2310  public:
2311  /// \brief Build a sign extension instruction of \p Opnd producing a \p Ty
2312  /// result.
2313  /// sext Opnd to Ty.
2314  SExtBuilder(Instruction *InsertPt, Value *Opnd, Type *Ty)
2315  : TypePromotionAction(InsertPt) {
2316  IRBuilder<> Builder(InsertPt);
2317  Val = Builder.CreateSExt(Opnd, Ty, "promoted");
2318  DEBUG(dbgs() << "Do: SExtBuilder: " << *Val << "\n");
2319  }
2320 
2321  /// \brief Get the built value.
2322  Value *getBuiltValue() { return Val; }
2323 
2324  /// \brief Remove the built instruction.
2325  void undo() override {
2326  DEBUG(dbgs() << "Undo: SExtBuilder: " << *Val << "\n");
2327  if (Instruction *IVal = dyn_cast<Instruction>(Val))
2328  IVal->eraseFromParent();
2329  }
2330  };
2331 
2332  /// \brief Build a zero extension instruction.
2333  class ZExtBuilder : public TypePromotionAction {
2334  Value *Val;
2335 
2336  public:
2337  /// \brief Build a zero extension instruction of \p Opnd producing a \p Ty
2338  /// result.
2339  /// zext Opnd to Ty.
2340  ZExtBuilder(Instruction *InsertPt, Value *Opnd, Type *Ty)
2341  : TypePromotionAction(InsertPt) {
2342  IRBuilder<> Builder(InsertPt);
2343  Val = Builder.CreateZExt(Opnd, Ty, "promoted");
2344  DEBUG(dbgs() << "Do: ZExtBuilder: " << *Val << "\n");
2345  }
2346 
2347  /// \brief Get the built value.
2348  Value *getBuiltValue() { return Val; }
2349 
2350  /// \brief Remove the built instruction.
2351  void undo() override {
2352  DEBUG(dbgs() << "Undo: ZExtBuilder: " << *Val << "\n");
2353  if (Instruction *IVal = dyn_cast<Instruction>(Val))
2354  IVal->eraseFromParent();
2355  }
2356  };
2357 
2358  /// \brief Mutate an instruction to another type.
2359  class TypeMutator : public TypePromotionAction {
2360  /// Record the original type.
2361  Type *OrigTy;
2362 
2363  public:
2364  /// \brief Mutate the type of \p Inst into \p NewTy.
2365  TypeMutator(Instruction *Inst, Type *NewTy)
2366  : TypePromotionAction(Inst), OrigTy(Inst->getType()) {
2367  DEBUG(dbgs() << "Do: MutateType: " << *Inst << " with " << *NewTy
2368  << "\n");
2369  Inst->mutateType(NewTy);
2370  }
2371 
2372  /// \brief Mutate the instruction back to its original type.
2373  void undo() override {
2374  DEBUG(dbgs() << "Undo: MutateType: " << *Inst << " with " << *OrigTy
2375  << "\n");
2376  Inst->mutateType(OrigTy);
2377  }
2378  };
2379 
2380  /// \brief Replace the uses of an instruction by another instruction.
2381  class UsesReplacer : public TypePromotionAction {
2382  /// Helper structure to keep track of the replaced uses.
2383  struct InstructionAndIdx {
2384  /// The instruction using the instruction.
2385  Instruction *Inst;
2386 
2387  /// The index where this instruction is used for Inst.
2388  unsigned Idx;
2389 
2390  InstructionAndIdx(Instruction *Inst, unsigned Idx)
2391  : Inst(Inst), Idx(Idx) {}
2392  };
2393 
2394  /// Keep track of the original uses (pair Instruction, Index).
2395  SmallVector<InstructionAndIdx, 4> OriginalUses;
2396 
2397  using use_iterator = SmallVectorImpl<InstructionAndIdx>::iterator;
2398 
2399  public:
2400  /// \brief Replace all the use of \p Inst by \p New.
2401  UsesReplacer(Instruction *Inst, Value *New) : TypePromotionAction(Inst) {
2402  DEBUG(dbgs() << "Do: UsersReplacer: " << *Inst << " with " << *New
2403  << "\n");
2404  // Record the original uses.
2405  for (Use &U : Inst->uses()) {
2406  Instruction *UserI = cast<Instruction>(U.getUser());
2407  OriginalUses.push_back(InstructionAndIdx(UserI, U.getOperandNo()));
2408  }
2409  // Now, we can replace the uses.
2410  Inst->replaceAllUsesWith(New);
2411  }
2412 
2413  /// \brief Reassign the original uses of Inst to Inst.
2414  void undo() override {
2415  DEBUG(dbgs() << "Undo: UsersReplacer: " << *Inst << "\n");
2416  for (use_iterator UseIt = OriginalUses.begin(),
2417  EndIt = OriginalUses.end();
2418  UseIt != EndIt; ++UseIt) {
2419  UseIt->Inst->setOperand(UseIt->Idx, Inst);
2420  }
2421  }
2422  };
2423 
2424  /// \brief Remove an instruction from the IR.
2425  class InstructionRemover : public TypePromotionAction {
2426  /// Original position of the instruction.
2427  InsertionHandler Inserter;
2428 
2429  /// Helper structure to hide all the link to the instruction. In other
2430  /// words, this helps to do as if the instruction was removed.
2431  OperandsHider Hider;
2432 
2433  /// Keep track of the uses replaced, if any.
2434  UsesReplacer *Replacer = nullptr;
2435 
2436  /// Keep track of instructions removed.
2437  SetOfInstrs &RemovedInsts;
2438 
2439  public:
2440  /// \brief Remove all reference of \p Inst and optinally replace all its
2441  /// uses with New.
2442  /// \p RemovedInsts Keep track of the instructions removed by this Action.
2443  /// \pre If !Inst->use_empty(), then New != nullptr
2444  InstructionRemover(Instruction *Inst, SetOfInstrs &RemovedInsts,
2445  Value *New = nullptr)
2446  : TypePromotionAction(Inst), Inserter(Inst), Hider(Inst),
2447  RemovedInsts(RemovedInsts) {
2448  if (New)
2449  Replacer = new UsesReplacer(Inst, New);
2450  DEBUG(dbgs() << "Do: InstructionRemover: " << *Inst << "\n");
2451  RemovedInsts.insert(Inst);
2452  /// The instructions removed here will be freed after completing
2453  /// optimizeBlock() for all blocks as we need to keep track of the
2454  /// removed instructions during promotion.
2455  Inst->removeFromParent();
2456  }
2457 
2458  ~InstructionRemover() override { delete Replacer; }
2459 
2460  /// \brief Resurrect the instruction and reassign it to the proper uses if
2461  /// new value was provided when build this action.
2462  void undo() override {
2463  DEBUG(dbgs() << "Undo: InstructionRemover: " << *Inst << "\n");
2464  Inserter.insert(Inst);
2465  if (Replacer)
2466  Replacer->undo();
2467  Hider.undo();
2468  RemovedInsts.erase(Inst);
2469  }
2470  };
2471 
2472 public:
2473  /// Restoration point.
2474  /// The restoration point is a pointer to an action instead of an iterator
2475  /// because the iterator may be invalidated but not the pointer.
2476  using ConstRestorationPt = const TypePromotionAction *;
2477 
2478  TypePromotionTransaction(SetOfInstrs &RemovedInsts)
2479  : RemovedInsts(RemovedInsts) {}
2480 
2481  /// Advocate every changes made in that transaction.
2482  void commit();
2483 
2484  /// Undo all the changes made after the given point.
2485  void rollback(ConstRestorationPt Point);
2486 
2487  /// Get the current restoration point.
2488  ConstRestorationPt getRestorationPoint() const;
2489 
2490  /// \name API for IR modification with state keeping to support rollback.
2491  /// @{
2492  /// Same as Instruction::setOperand.
2493  void setOperand(Instruction *Inst, unsigned Idx, Value *NewVal);
2494 
2495  /// Same as Instruction::eraseFromParent.
2496  void eraseInstruction(Instruction *Inst, Value *NewVal = nullptr);
2497 
2498  /// Same as Value::replaceAllUsesWith.
2499  void replaceAllUsesWith(Instruction *Inst, Value *New);
2500 
2501  /// Same as Value::mutateType.
2502  void mutateType(Instruction *Inst, Type *NewTy);
2503 
2504  /// Same as IRBuilder::createTrunc.
2505  Value *createTrunc(Instruction *Opnd, Type *Ty);
2506 
2507  /// Same as IRBuilder::createSExt.
2508  Value *createSExt(Instruction *Inst, Value *Opnd, Type *Ty);
2509 
2510  /// Same as IRBuilder::createZExt.
2511  Value *createZExt(Instruction *Inst, Value *Opnd, Type *Ty);
2512 
2513  /// Same as Instruction::moveBefore.
2514  void moveBefore(Instruction *Inst, Instruction *Before);
2515  /// @}
2516 
2517 private:
2518  /// The ordered list of actions made so far.
2520 
2521  using CommitPt = SmallVectorImpl<std::unique_ptr<TypePromotionAction>>::iterator;
2522 
2523  SetOfInstrs &RemovedInsts;
2524 };
2525 
2526 } // end anonymous namespace
2527 
2528 void TypePromotionTransaction::setOperand(Instruction *Inst, unsigned Idx,
2529  Value *NewVal) {
2530  Actions.push_back(llvm::make_unique<TypePromotionTransaction::OperandSetter>(
2531  Inst, Idx, NewVal));
2532 }
2533 
2534 void TypePromotionTransaction::eraseInstruction(Instruction *Inst,
2535  Value *NewVal) {
2536  Actions.push_back(
2537  llvm::make_unique<TypePromotionTransaction::InstructionRemover>(
2538  Inst, RemovedInsts, NewVal));
2539 }
2540 
2541 void TypePromotionTransaction::replaceAllUsesWith(Instruction *Inst,
2542  Value *New) {
2543  Actions.push_back(
2544  llvm::make_unique<TypePromotionTransaction::UsesReplacer>(Inst, New));
2545 }
2546 
2547 void TypePromotionTransaction::mutateType(Instruction *Inst, Type *NewTy) {
2548  Actions.push_back(
2549  llvm::make_unique<TypePromotionTransaction::TypeMutator>(Inst, NewTy));
2550 }
2551 
2552 Value *TypePromotionTransaction::createTrunc(Instruction *Opnd,
2553  Type *Ty) {
2554  std::unique_ptr<TruncBuilder> Ptr(new TruncBuilder(Opnd, Ty));
2555  Value *Val = Ptr->getBuiltValue();
2556  Actions.push_back(std::move(Ptr));
2557  return Val;
2558 }
2559 
2560 Value *TypePromotionTransaction::createSExt(Instruction *Inst,
2561  Value *Opnd, Type *Ty) {
2562  std::unique_ptr<SExtBuilder> Ptr(new SExtBuilder(Inst, Opnd, Ty));
2563  Value *Val = Ptr->getBuiltValue();
2564  Actions.push_back(std::move(Ptr));
2565  return Val;
2566 }
2567 
2568 Value *TypePromotionTransaction::createZExt(Instruction *Inst,
2569  Value *Opnd, Type *Ty) {
2570  std::unique_ptr<ZExtBuilder> Ptr(new ZExtBuilder(Inst, Opnd, Ty));
2571  Value *Val = Ptr->getBuiltValue();
2572  Actions.push_back(std::move(Ptr));
2573  return Val;
2574 }
2575 
2576 void TypePromotionTransaction::moveBefore(Instruction *Inst,
2577  Instruction *Before) {
2578  Actions.push_back(
2579  llvm::make_unique<TypePromotionTransaction::InstructionMoveBefore>(
2580  Inst, Before));
2581 }
2582 
2583 TypePromotionTransaction::ConstRestorationPt
2584 TypePromotionTransaction::getRestorationPoint() const {
2585  return !Actions.empty() ? Actions.back().get() : nullptr;
2586 }
2587 
2588 void TypePromotionTransaction::commit() {
2589  for (CommitPt It = Actions.begin(), EndIt = Actions.end(); It != EndIt;
2590  ++It)
2591  (*It)->commit();
2592  Actions.clear();
2593 }
2594 
2595 void TypePromotionTransaction::rollback(
2596  TypePromotionTransaction::ConstRestorationPt Point) {
2597  while (!Actions.empty() && Point != Actions.back().get()) {
2598  std::unique_ptr<TypePromotionAction> Curr = Actions.pop_back_val();
2599  Curr->undo();
2600  }
2601 }
2602 
2603 namespace {
2604 
2605 /// \brief A helper class for matching addressing modes.
2606 ///
2607 /// This encapsulates the logic for matching the target-legal addressing modes.
2608 class AddressingModeMatcher {
2609  SmallVectorImpl<Instruction*> &AddrModeInsts;
2610  const TargetLowering &TLI;
2611  const TargetRegisterInfo &TRI;
2612  const DataLayout &DL;
2613 
2614  /// AccessTy/MemoryInst - This is the type for the access (e.g. double) and
2615  /// the memory instruction that we're computing this address for.
2616  Type *AccessTy;
2617  unsigned AddrSpace;
2618  Instruction *MemoryInst;
2619 
2620  /// This is the addressing mode that we're building up. This is
2621  /// part of the return value of this addressing mode matching stuff.
2622  ExtAddrMode &AddrMode;
2623 
2624  /// The instructions inserted by other CodeGenPrepare optimizations.
2625  const SetOfInstrs &InsertedInsts;
2626 
2627  /// A map from the instructions to their type before promotion.
2628  InstrToOrigTy &PromotedInsts;
2629 
2630  /// The ongoing transaction where every action should be registered.
2631  TypePromotionTransaction &TPT;
2632 
2633  /// This is set to true when we should not do profitability checks.
2634  /// When true, IsProfitableToFoldIntoAddressingMode always returns true.
2635  bool IgnoreProfitability;
2636 
2637  AddressingModeMatcher(SmallVectorImpl<Instruction *> &AMI,
2638  const TargetLowering &TLI,
2639  const TargetRegisterInfo &TRI,
2640  Type *AT, unsigned AS,
2641  Instruction *MI, ExtAddrMode &AM,
2642  const SetOfInstrs &InsertedInsts,
2643  InstrToOrigTy &PromotedInsts,
2644  TypePromotionTransaction &TPT)
2645  : AddrModeInsts(AMI), TLI(TLI), TRI(TRI),
2646  DL(MI->getModule()->getDataLayout()), AccessTy(AT), AddrSpace(AS),
2647  MemoryInst(MI), AddrMode(AM), InsertedInsts(InsertedInsts),
2648  PromotedInsts(PromotedInsts), TPT(TPT) {
2649  IgnoreProfitability = false;
2650  }
2651 
2652 public:
2653  /// Find the maximal addressing mode that a load/store of V can fold,
2654  /// give an access type of AccessTy. This returns a list of involved
2655  /// instructions in AddrModeInsts.
2656  /// \p InsertedInsts The instructions inserted by other CodeGenPrepare
2657  /// optimizations.
2658  /// \p PromotedInsts maps the instructions to their type before promotion.
2659  /// \p The ongoing transaction where every action should be registered.
2660  static ExtAddrMode Match(Value *V, Type *AccessTy, unsigned AS,
2661  Instruction *MemoryInst,
2662  SmallVectorImpl<Instruction*> &AddrModeInsts,
2663  const TargetLowering &TLI,
2664  const TargetRegisterInfo &TRI,
2665  const SetOfInstrs &InsertedInsts,
2666  InstrToOrigTy &PromotedInsts,
2667  TypePromotionTransaction &TPT) {
2668  ExtAddrMode Result;
2669 
2670  bool Success = AddressingModeMatcher(AddrModeInsts, TLI, TRI,
2671  AccessTy, AS,
2672  MemoryInst, Result, InsertedInsts,
2673  PromotedInsts, TPT).matchAddr(V, 0);
2674  (void)Success; assert(Success && "Couldn't select *anything*?");
2675  return Result;
2676  }
2677 
2678 private:
2679  bool matchScaledValue(Value *ScaleReg, int64_t Scale, unsigned Depth);
2680  bool matchAddr(Value *V, unsigned Depth);
2681  bool matchOperationAddr(User *Operation, unsigned Opcode, unsigned Depth,
2682  bool *MovedAway = nullptr);
2683  bool isProfitableToFoldIntoAddressingMode(Instruction *I,
2684  ExtAddrMode &AMBefore,
2685  ExtAddrMode &AMAfter);
2686  bool valueAlreadyLiveAtInst(Value *Val, Value *KnownLive1, Value *KnownLive2);
2687  bool isPromotionProfitable(unsigned NewCost, unsigned OldCost,
2688  Value *PromotedOperand) const;
2689 };
2690 
2691 /// \brief Keep track of simplification of Phi nodes.
2692 /// Accept the set of all phi nodes and erase phi node from this set
2693 /// if it is simplified.
2694 class SimplificationTracker {
2696  const SimplifyQuery &SQ;
2697  SmallPtrSetImpl<PHINode *> &AllPhiNodes;
2698  SmallPtrSetImpl<SelectInst *> &AllSelectNodes;
2699 
2700 public:
2701  SimplificationTracker(const SimplifyQuery &sq,
2704  : SQ(sq), AllPhiNodes(APN), AllSelectNodes(ASN) {}
2705 
2706  Value *Get(Value *V) {
2707  do {
2708  auto SV = Storage.find(V);
2709  if (SV == Storage.end())
2710  return V;
2711  V = SV->second;
2712  } while (true);
2713  }
2714 
2715  Value *Simplify(Value *Val) {
2716  SmallVector<Value *, 32> WorkList;
2717  SmallPtrSet<Value *, 32> Visited;
2718  WorkList.push_back(Val);
2719  while (!WorkList.empty()) {
2720  auto P = WorkList.pop_back_val();
2721  if (!Visited.insert(P).second)
2722  continue;
2723  if (auto *PI = dyn_cast<Instruction>(P))
2724  if (Value *V = SimplifyInstruction(cast<Instruction>(PI), SQ)) {
2725  for (auto *U : PI->users())
2726  WorkList.push_back(cast<Value>(U));
2727  Put(PI, V);
2728  PI->replaceAllUsesWith(V);
2729  if (auto *PHI = dyn_cast<PHINode>(PI))
2730  AllPhiNodes.erase(PHI);
2731  if (auto *Select = dyn_cast<SelectInst>(PI))
2732  AllSelectNodes.erase(Select);
2733  PI->eraseFromParent();
2734  }
2735  }
2736  return Get(Val);
2737  }
2738 
2739  void Put(Value *From, Value *To) {
2740  Storage.insert({ From, To });
2741  }
2742 };
2743 
2744 /// \brief A helper class for combining addressing modes.
2745 class AddressingModeCombiner {
2746  typedef std::pair<Value *, BasicBlock *> ValueInBB;
2747  typedef DenseMap<ValueInBB, Value *> FoldAddrToValueMapping;
2748  typedef std::pair<PHINode *, PHINode *> PHIPair;
2749 
2750 private:
2751  /// The addressing modes we've collected.
2752  SmallVector<ExtAddrMode, 16> AddrModes;
2753 
2754  /// The field in which the AddrModes differ, when we have more than one.
2755  ExtAddrMode::FieldName DifferentField = ExtAddrMode::NoField;
2756 
2757  /// Are the AddrModes that we have all just equal to their original values?
2758  bool AllAddrModesTrivial = true;
2759 
2760  /// Common Type for all different fields in addressing modes.
2761  Type *CommonType;
2762 
2763  /// SimplifyQuery for simplifyInstruction utility.
2764  const SimplifyQuery &SQ;
2765 
2766  /// Original Address.
2767  ValueInBB Original;
2768 
2769 public:
2770  AddressingModeCombiner(const SimplifyQuery &_SQ, ValueInBB OriginalValue)
2771  : CommonType(nullptr), SQ(_SQ), Original(OriginalValue) {}
2772 
2773  /// \brief Get the combined AddrMode
2774  const ExtAddrMode &getAddrMode() const {
2775  return AddrModes[0];
2776  }
2777 
2778  /// \brief Add a new AddrMode if it's compatible with the AddrModes we already
2779  /// have.
2780  /// \return True iff we succeeded in doing so.
2781  bool addNewAddrMode(ExtAddrMode &NewAddrMode) {
2782  // Take note of if we have any non-trivial AddrModes, as we need to detect
2783  // when all AddrModes are trivial as then we would introduce a phi or select
2784  // which just duplicates what's already there.
2785  AllAddrModesTrivial = AllAddrModesTrivial && NewAddrMode.isTrivial();
2786 
2787  // If this is the first addrmode then everything is fine.
2788  if (AddrModes.empty()) {
2789  AddrModes.emplace_back(NewAddrMode);
2790  return true;
2791  }
2792 
2793  // Figure out how different this is from the other address modes, which we
2794  // can do just by comparing against the first one given that we only care
2795  // about the cumulative difference.
2796  ExtAddrMode::FieldName ThisDifferentField =
2797  AddrModes[0].compare(NewAddrMode);
2798  if (DifferentField == ExtAddrMode::NoField)
2799  DifferentField = ThisDifferentField;
2800  else if (DifferentField != ThisDifferentField)
2801  DifferentField = ExtAddrMode::MultipleFields;
2802 
2803  // If NewAddrMode differs in only one dimension then we can handle it by
2804  // inserting a phi/select later on. Even if NewAddMode is the same
2805  // we still need to collect it due to original value is different.
2806  // And later we will need all original values as anchors during
2807  // finding the common Phi node.
2808  if (DifferentField != ExtAddrMode::MultipleFields) {
2809  AddrModes.emplace_back(NewAddrMode);
2810  return true;
2811  }
2812 
2813  // We couldn't combine NewAddrMode with the rest, so return failure.
2814  AddrModes.clear();
2815  return false;
2816  }
2817 
2818  /// \brief Combine the addressing modes we've collected into a single
2819  /// addressing mode.
2820  /// \return True iff we successfully combined them or we only had one so
2821  /// didn't need to combine them anyway.
2822  bool combineAddrModes() {
2823  // If we have no AddrModes then they can't be combined.
2824  if (AddrModes.size() == 0)
2825  return false;
2826 
2827  // A single AddrMode can trivially be combined.
2828  if (AddrModes.size() == 1 || DifferentField == ExtAddrMode::NoField)
2829  return true;
2830 
2831  // If the AddrModes we collected are all just equal to the value they are
2832  // derived from then combining them wouldn't do anything useful.
2833  if (AllAddrModesTrivial)
2834  return false;
2835 
2837  return false;
2838 
2839  // For now we support only different base registers.
2840  // TODO: enable others.
2841  if (DifferentField != ExtAddrMode::BaseRegField)
2842  return false;
2843 
2844  // Build a map between <original value, basic block where we saw it> to
2845  // value of base register.
2846  FoldAddrToValueMapping Map;
2847  initializeMap(Map);
2848 
2849  Value *CommonValue = findCommon(Map);
2850  if (CommonValue)
2851  AddrModes[0].BaseReg = CommonValue;
2852  return CommonValue != nullptr;
2853  }
2854 
2855 private:
2856  /// \brief Initialize Map with anchor values. For address seen in some BB
2857  /// we set the value of different field saw in this address.
2858  /// If address is not an instruction than basic block is set to null.
2859  /// At the same time we find a common type for different field we will
2860  /// use to create new Phi/Select nodes. Keep it in CommonType field.
2861  void initializeMap(FoldAddrToValueMapping &Map) {
2862  // Keep track of keys where the value is null. We will need to replace it
2863  // with constant null when we know the common type.
2864  SmallVector<ValueInBB, 2> NullValue;
2865  for (auto &AM : AddrModes) {
2866  BasicBlock *BB = nullptr;
2867  if (Instruction *I = dyn_cast<Instruction>(AM.OriginalValue))
2868  BB = I->getParent();
2869 
2870  // For now we support only base register as different field.
2871  // TODO: Enable others.
2872  Value *DV = AM.BaseReg;
2873  if (DV) {
2874  if (CommonType)
2875  assert(CommonType == DV->getType() && "Different types detected!");
2876  else
2877  CommonType = DV->getType();
2878  Map[{ AM.OriginalValue, BB }] = DV;
2879  } else {
2880  NullValue.push_back({ AM.OriginalValue, BB });
2881  }
2882  }
2883  assert(CommonType && "At least one non-null value must be!");
2884  for (auto VIBB : NullValue)
2885  Map[VIBB] = Constant::getNullValue(CommonType);
2886  }
2887 
2888  /// \brief We have mapping between value A and basic block where value A
2889  /// seen to other value B where B was a field in addressing mode represented
2890  /// by A. Also we have an original value C representin an address in some
2891  /// basic block. Traversing from C through phi and selects we ended up with
2892  /// A's in a map. This utility function tries to find a value V which is a
2893  /// field in addressing mode C and traversing through phi nodes and selects
2894  /// we will end up in corresponded values B in a map.
2895  /// The utility will create a new Phi/Selects if needed.
2896  // The simple example looks as follows:
2897  // BB1:
2898  // p1 = b1 + 40
2899  // br cond BB2, BB3
2900  // BB2:
2901  // p2 = b2 + 40
2902  // br BB3
2903  // BB3:
2904  // p = phi [p1, BB1], [p2, BB2]
2905  // v = load p
2906  // Map is
2907  // <p1, BB1> -> b1
2908  // <p2, BB2> -> b2
2909  // Request is
2910  // <p, BB3> -> ?
2911  // The function tries to find or build phi [b1, BB1], [b2, BB2] in BB3
2912  Value *findCommon(FoldAddrToValueMapping &Map) {
2913  // Tracks of new created Phi nodes.
2914  SmallPtrSet<PHINode *, 32> NewPhiNodes;
2915  // Tracks of new created Select nodes.
2916  SmallPtrSet<SelectInst *, 32> NewSelectNodes;
2917  // Tracks the simplification of new created phi nodes. The reason we use
2918  // this mapping is because we will add new created Phi nodes in AddrToBase.
2919  // Simplification of Phi nodes is recursive, so some Phi node may
2920  // be simplified after we added it to AddrToBase.
2921  // Using this mapping we can find the current value in AddrToBase.
2922  SimplificationTracker ST(SQ, NewPhiNodes, NewSelectNodes);
2923 
2924  // First step, DFS to create PHI nodes for all intermediate blocks.
2925  // Also fill traverse order for the second step.
2926  SmallVector<ValueInBB, 32> TraverseOrder;
2927  InsertPlaceholders(Map, TraverseOrder, NewPhiNodes, NewSelectNodes);
2928 
2929  // Second Step, fill new nodes by merged values and simplify if possible.
2930  FillPlaceholders(Map, TraverseOrder, ST);
2931 
2932  if (!AddrSinkNewSelects && NewSelectNodes.size() > 0) {
2933  DestroyNodes(NewPhiNodes);
2934  DestroyNodes(NewSelectNodes);
2935  return nullptr;
2936  }
2937 
2938  // Now we'd like to match New Phi nodes to existed ones.
2939  unsigned PhiNotMatchedCount = 0;
2940  if (!MatchPhiSet(NewPhiNodes, ST, AddrSinkNewPhis, PhiNotMatchedCount)) {
2941  DestroyNodes(NewPhiNodes);
2942  DestroyNodes(NewSelectNodes);
2943  return nullptr;
2944  }
2945 
2946  auto *Result = ST.Get(Map.find(Original)->second);
2947  if (Result) {
2948  NumMemoryInstsPhiCreated += NewPhiNodes.size() + PhiNotMatchedCount;
2949  NumMemoryInstsSelectCreated += NewSelectNodes.size();
2950  }
2951  return Result;
2952  }
2953 
2954  /// \brief Destroy nodes from a set.
2955  template <typename T> void DestroyNodes(SmallPtrSetImpl<T *> &Instructions) {
2956  // For safe erasing, replace the Phi with dummy value first.
2957  auto Dummy = UndefValue::get(CommonType);
2958  for (auto I : Instructions) {
2960  I->eraseFromParent();
2961  }
2962  }
2963 
2964  /// \brief Try to match PHI node to Candidate.
2965  /// Matcher tracks the matched Phi nodes.
2966  bool MatchPhiNode(PHINode *PHI, PHINode *Candidate,
2967  DenseSet<PHIPair> &Matcher,
2968  SmallPtrSetImpl<PHINode *> &PhiNodesToMatch) {
2969  SmallVector<PHIPair, 8> WorkList;
2970  Matcher.insert({ PHI, Candidate });
2971  WorkList.push_back({ PHI, Candidate });
2972  SmallSet<PHIPair, 8> Visited;
2973  while (!WorkList.empty()) {
2974  auto Item = WorkList.pop_back_val();
2975  if (!Visited.insert(Item).second)
2976  continue;
2977  // We iterate over all incoming values to Phi to compare them.
2978  // If values are different and both of them Phi and the first one is a
2979  // Phi we added (subject to match) and both of them is in the same basic
2980  // block then we can match our pair if values match. So we state that
2981  // these values match and add it to work list to verify that.
2982  for (auto B : Item.first->blocks()) {
2983  Value *FirstValue = Item.first->getIncomingValueForBlock(B);
2984  Value *SecondValue = Item.second->getIncomingValueForBlock(B);
2985  if (FirstValue == SecondValue)
2986  continue;
2987 
2988  PHINode *FirstPhi = dyn_cast<PHINode>(FirstValue);
2989  PHINode *SecondPhi = dyn_cast<PHINode>(SecondValue);
2990 
2991  // One of them is not Phi or
2992  // The first one is not Phi node from the set we'd like to match or
2993  // Phi nodes from different basic blocks then
2994  // we will not be able to match.
2995  if (!FirstPhi || !SecondPhi || !PhiNodesToMatch.count(FirstPhi) ||
2996  FirstPhi->getParent() != SecondPhi->getParent())
2997  return false;
2998 
2999  // If we already matched them then continue.
3000  if (Matcher.count({ FirstPhi, SecondPhi }))
3001  continue;
3002  // So the values are different and does not match. So we need them to
3003  // match.
3004  Matcher.insert({ FirstPhi, SecondPhi });
3005  // But me must check it.
3006  WorkList.push_back({ FirstPhi, SecondPhi });
3007  }
3008  }
3009  return true;
3010  }
3011 
3012  /// \brief For the given set of PHI nodes try to find their equivalents.
3013  /// Returns false if this matching fails and creation of new Phi is disabled.
3014  bool MatchPhiSet(SmallPtrSetImpl<PHINode *> &PhiNodesToMatch,
3015  SimplificationTracker &ST, bool AllowNewPhiNodes,
3016  unsigned &PhiNotMatchedCount) {
3017  DenseSet<PHIPair> Matched;
3018  SmallPtrSet<PHINode *, 8> WillNotMatch;
3019  while (PhiNodesToMatch.size()) {
3020  PHINode *PHI = *PhiNodesToMatch.begin();
3021 
3022  // Add us, if no Phi nodes in the basic block we do not match.
3023  WillNotMatch.clear();
3024  WillNotMatch.insert(PHI);
3025 
3026  // Traverse all Phis until we found equivalent or fail to do that.
3027  bool IsMatched = false;
3028  for (auto &P : PHI->getParent()->phis()) {
3029  if (&P == PHI)
3030  continue;
3031  if ((IsMatched = MatchPhiNode(PHI, &P, Matched, PhiNodesToMatch)))
3032  break;
3033  // If it does not match, collect all Phi nodes from matcher.
3034  // if we end up with no match, them all these Phi nodes will not match
3035  // later.
3036  for (auto M : Matched)
3037  WillNotMatch.insert(M.first);
3038  Matched.clear();
3039  }
3040  if (IsMatched) {
3041  // Replace all matched values and erase them.
3042  for (auto MV : Matched) {
3043  MV.first->replaceAllUsesWith(MV.second);
3044  PhiNodesToMatch.erase(MV.first);
3045  ST.Put(MV.first, MV.second);
3046  MV.first->eraseFromParent();
3047  }
3048  Matched.clear();
3049  continue;
3050  }
3051  // If we are not allowed to create new nodes then bail out.
3052  if (!AllowNewPhiNodes)
3053  return false;
3054  // Just remove all seen values in matcher. They will not match anything.
3055  PhiNotMatchedCount += WillNotMatch.size();
3056  for (auto *P : WillNotMatch)
3057  PhiNodesToMatch.erase(P);
3058  }
3059  return true;
3060  }
3061  /// \brief Fill the placeholder with values from predecessors and simplify it.
3062  void FillPlaceholders(FoldAddrToValueMapping &Map,
3063  SmallVectorImpl<ValueInBB> &TraverseOrder,
3064  SimplificationTracker &ST) {
3065  while (!TraverseOrder.empty()) {
3066  auto Current = TraverseOrder.pop_back_val();
3067  assert(Map.find(Current) != Map.end() && "No node to fill!!!");
3068  Value *CurrentValue = Current.first;
3069  BasicBlock *CurrentBlock = Current.second;
3070  Value *V = Map[Current];
3071 
3072  if (SelectInst *Select = dyn_cast<SelectInst>(V)) {
3073  // CurrentValue also must be Select.
3074  auto *CurrentSelect = cast<SelectInst>(CurrentValue);
3075  auto *TrueValue = CurrentSelect->getTrueValue();
3076  ValueInBB TrueItem = { TrueValue, isa<Instruction>(TrueValue)
3077  ? CurrentBlock
3078  : nullptr };
3079  assert(Map.find(TrueItem) != Map.end() && "No True Value!");
3080  Select->setTrueValue(Map[TrueItem]);
3081  auto *FalseValue = CurrentSelect->getFalseValue();
3082  ValueInBB FalseItem = { FalseValue, isa<Instruction>(FalseValue)
3083  ? CurrentBlock
3084  : nullptr };
3085  assert(Map.find(FalseItem) != Map.end() && "No False Value!");
3086  Select->setFalseValue(Map[FalseItem]);
3087  } else {
3088  // Must be a Phi node then.
3089  PHINode *PHI = cast<PHINode>(V);
3090  // Fill the Phi node with values from predecessors.
3091  bool IsDefinedInThisBB =
3092  cast<Instruction>(CurrentValue)->getParent() == CurrentBlock;
3093  auto *CurrentPhi = dyn_cast<PHINode>(CurrentValue);
3094  for (auto B : predecessors(CurrentBlock)) {
3095  Value *PV = IsDefinedInThisBB
3096  ? CurrentPhi->getIncomingValueForBlock(B)
3097  : CurrentValue;
3098  ValueInBB item = { PV, isa<Instruction>(PV) ? B : nullptr };
3099  assert(Map.find(item) != Map.end() && "No predecessor Value!");
3100  PHI->addIncoming(ST.Get(Map[item]), B);
3101  }
3102  }
3103  // Simplify if possible.
3104  Map[Current] = ST.Simplify(V);
3105  }
3106  }
3107 
3108  /// Starting from value recursively iterates over predecessors up to known
3109  /// ending values represented in a map. For each traversed block inserts
3110  /// a placeholder Phi or Select.
3111  /// Reports all new created Phi/Select nodes by adding them to set.
3112  /// Also reports and order in what basic blocks have been traversed.
3113  void InsertPlaceholders(FoldAddrToValueMapping &Map,
3114  SmallVectorImpl<ValueInBB> &TraverseOrder,
3115  SmallPtrSetImpl<PHINode *> &NewPhiNodes,
3116  SmallPtrSetImpl<SelectInst *> &NewSelectNodes) {
3117  SmallVector<ValueInBB, 32> Worklist;
3118  assert((isa<PHINode>(Original.first) || isa<SelectInst>(Original.first)) &&
3119  "Address must be a Phi or Select node");
3120  auto *Dummy = UndefValue::get(CommonType);
3121  Worklist.push_back(Original);
3122  while (!Worklist.empty()) {
3123  auto Current = Worklist.pop_back_val();
3124  // If value is not an instruction it is something global, constant,
3125  // parameter and we can say that this value is observable in any block.
3126  // Set block to null to denote it.
3127  // Also please take into account that it is how we build anchors.
3128  if (!isa<Instruction>(Current.first))
3129  Current.second = nullptr;
3130  // if it is already visited or it is an ending value then skip it.
3131  if (Map.find(Current) != Map.end())
3132  continue;
3133  TraverseOrder.push_back(Current);
3134 
3135  Value *CurrentValue = Current.first;
3136  BasicBlock *CurrentBlock = Current.second;
3137  // CurrentValue must be a Phi node or select. All others must be covered
3138  // by anchors.
3139  Instruction *CurrentI = cast<Instruction>(CurrentValue);
3140  bool IsDefinedInThisBB = CurrentI->getParent() == CurrentBlock;
3141 
3142  unsigned PredCount =
3143  std::distance(pred_begin(CurrentBlock), pred_end(CurrentBlock));
3144  // if Current Value is not defined in this basic block we are interested
3145  // in values in predecessors.
3146  if (!IsDefinedInThisBB) {
3147  assert(PredCount && "Unreachable block?!");
3148  PHINode *PHI = PHINode::Create(CommonType, PredCount, "sunk_phi",
3149  &CurrentBlock->front());
3150  Map[Current] = PHI;
3151  NewPhiNodes.insert(PHI);
3152  // Add all predecessors in work list.
3153  for (auto B : predecessors(CurrentBlock))
3154  Worklist.push_back({ CurrentValue, B });
3155  continue;
3156  }
3157  // Value is defined in this basic block.
3158  if (SelectInst *OrigSelect = dyn_cast<SelectInst>(CurrentI)) {
3159  // Is it OK to get metadata from OrigSelect?!
3160  // Create a Select placeholder with dummy value.
3161  SelectInst *Select =
3162  SelectInst::Create(OrigSelect->getCondition(), Dummy, Dummy,
3163  OrigSelect->getName(), OrigSelect, OrigSelect);
3164  Map[Current] = Select;
3165  NewSelectNodes.insert(Select);
3166  // We are interested in True and False value in this basic block.
3167  Worklist.push_back({ OrigSelect->getTrueValue(), CurrentBlock });
3168  Worklist.push_back({ OrigSelect->getFalseValue(), CurrentBlock });
3169  } else {
3170  // It must be a Phi node then.
3171  auto *CurrentPhi = cast<PHINode>(CurrentI);
3172  // Create new Phi node for merge of bases.
3173  assert(PredCount && "Unreachable block?!");
3174  PHINode *PHI = PHINode::Create(CommonType, PredCount, "sunk_phi",
3175  &CurrentBlock->front());
3176  Map[Current] = PHI;
3177  NewPhiNodes.insert(PHI);
3178 
3179  // Add all predecessors in work list.
3180  for (auto B : predecessors(CurrentBlock))
3181  Worklist.push_back({ CurrentPhi->getIncomingValueForBlock(B), B });
3182  }
3183  }
3184  }
3185 };
3186 } // end anonymous namespace
3187 
3188 /// Try adding ScaleReg*Scale to the current addressing mode.
3189 /// Return true and update AddrMode if this addr mode is legal for the target,
3190 /// false if not.
3191 bool AddressingModeMatcher::matchScaledValue(Value *ScaleReg, int64_t Scale,
3192  unsigned Depth) {
3193  // If Scale is 1, then this is the same as adding ScaleReg to the addressing
3194  // mode. Just process that directly.
3195  if (Scale == 1)
3196  return matchAddr(ScaleReg, Depth);
3197 
3198  // If the scale is 0, it takes nothing to add this.
3199  if (Scale == 0)
3200  return true;
3201 
3202  // If we already have a scale of this value, we can add to it, otherwise, we
3203  // need an available scale field.
3204  if (AddrMode.Scale != 0 && AddrMode.ScaledReg != ScaleReg)
3205  return false;
3206 
3207  ExtAddrMode TestAddrMode = AddrMode;
3208 
3209  // Add scale to turn X*4+X*3 -> X*7. This could also do things like
3210  // [A+B + A*7] -> [B+A*8].
3211  TestAddrMode.Scale += Scale;
3212  TestAddrMode.ScaledReg = ScaleReg;
3213 
3214  // If the new address isn't legal, bail out.
3215  if (!TLI.isLegalAddressingMode(DL, TestAddrMode, AccessTy, AddrSpace))
3216  return false;
3217 
3218  // It was legal, so commit it.
3219  AddrMode = TestAddrMode;
3220 
3221  // Okay, we decided that we can add ScaleReg+Scale to AddrMode. Check now
3222  // to see if ScaleReg is actually X+C. If so, we can turn this into adding
3223  // X*Scale + C*Scale to addr mode.
3224  ConstantInt *CI = nullptr; Value *AddLHS = nullptr;
3225  if (isa<Instruction>(ScaleReg) && // not a constant expr.
3226  match(ScaleReg, m_Add(m_Value(AddLHS), m_ConstantInt(CI)))) {
3227  TestAddrMode.ScaledReg = AddLHS;
3228  TestAddrMode.BaseOffs += CI->getSExtValue()*TestAddrMode.Scale;
3229 
3230  // If this addressing mode is legal, commit it and remember that we folded
3231  // this instruction.
3232  if (TLI.isLegalAddressingMode(DL, TestAddrMode, AccessTy, AddrSpace)) {
3233  AddrModeInsts.push_back(cast<Instruction>(ScaleReg));
3234  AddrMode = TestAddrMode;
3235  return true;
3236  }
3237  }
3238 
3239  // Otherwise, not (x+c)*scale, just return what we have.
3240  return true;
3241 }
3242 
3243 /// This is a little filter, which returns true if an addressing computation
3244 /// involving I might be folded into a load/store accessing it.
3245 /// This doesn't need to be perfect, but needs to accept at least
3246 /// the set of instructions that MatchOperationAddr can.
3248  switch (I->getOpcode()) {
3249  case Instruction::BitCast:
3250  case Instruction::AddrSpaceCast:
3251  // Don't touch identity bitcasts.
3252  if (I->getType() == I->getOperand(0)->getType())
3253  return false;
3254  return I->getType()->isPointerTy() || I->getType()->isIntegerTy();
3255  case Instruction::PtrToInt:
3256  // PtrToInt is always a noop, as we know that the int type is pointer sized.
3257  return true;
3258  case Instruction::IntToPtr:
3259  // We know the input is intptr_t, so this is foldable.
3260  return true;
3261  case Instruction::Add:
3262  return true;
3263  case Instruction::Mul:
3264  case Instruction::Shl:
3265  // Can only handle X*C and X << C.
3266  return isa<ConstantInt>(I->getOperand(1));
3267  case Instruction::GetElementPtr:
3268  return true;
3269  default:
3270  return false;
3271  }
3272 }
3273 
3274 /// \brief Check whether or not \p Val is a legal instruction for \p TLI.
3275 /// \note \p Val is assumed to be the product of some type promotion.
3276 /// Therefore if \p Val has an undefined state in \p TLI, this is assumed
3277 /// to be legal, as the non-promoted value would have had the same state.
3279  const DataLayout &DL, Value *Val) {
3280  Instruction *PromotedInst = dyn_cast<Instruction>(Val);
3281  if (!PromotedInst)
3282  return false;
3283  int ISDOpcode = TLI.InstructionOpcodeToISD(PromotedInst->getOpcode());
3284  // If the ISDOpcode is undefined, it was undefined before the promotion.
3285  if (!ISDOpcode)
3286  return true;
3287  // Otherwise, check if the promoted instruction is legal or not.
3288  return TLI.isOperationLegalOrCustom(
3289  ISDOpcode, TLI.getValueType(DL, PromotedInst->getType()));
3290 }
3291 
3292 namespace {
3293 
3294 /// \brief Hepler class to perform type promotion.
3295 class TypePromotionHelper {
3296  /// \brief Utility function to check whether or not a sign or zero extension
3297  /// of \p Inst with \p ConsideredExtType can be moved through \p Inst by
3298  /// either using the operands of \p Inst or promoting \p Inst.
3299  /// The type of the extension is defined by \p IsSExt.
3300  /// In other words, check if:
3301  /// ext (Ty Inst opnd1 opnd2 ... opndN) to ConsideredExtType.
3302  /// #1 Promotion applies:
3303  /// ConsideredExtType Inst (ext opnd1 to ConsideredExtType, ...).
3304  /// #2 Operand reuses:
3305  /// ext opnd1 to ConsideredExtType.
3306  /// \p PromotedInsts maps the instructions to their type before promotion.
3307  static bool canGetThrough(const Instruction *Inst, Type *ConsideredExtType,
3308  const InstrToOrigTy &PromotedInsts, bool IsSExt);
3309 
3310  /// \brief Utility function to determine if \p OpIdx should be promoted when
3311  /// promoting \p Inst.
3312  static bool shouldExtOperand(const Instruction *Inst, int OpIdx) {
3313  return !(isa<SelectInst>(Inst) && OpIdx == 0);
3314  }
3315 
3316  /// \brief Utility function to promote the operand of \p Ext when this
3317  /// operand is a promotable trunc or sext or zext.
3318  /// \p PromotedInsts maps the instructions to their type before promotion.
3319  /// \p CreatedInstsCost[out] contains the cost of all instructions
3320  /// created to promote the operand of Ext.
3321  /// Newly added extensions are inserted in \p Exts.
3322  /// Newly added truncates are inserted in \p Truncs.
3323  /// Should never be called directly.
3324  /// \return The promoted value which is used instead of Ext.
3325  static Value *promoteOperandForTruncAndAnyExt(
3326  Instruction *Ext, TypePromotionTransaction &TPT,
3327  InstrToOrigTy &PromotedInsts, unsigned &CreatedInstsCost,
3329  SmallVectorImpl<Instruction *> *Truncs, const TargetLowering &TLI);
3330 
3331  /// \brief Utility function to promote the operand of \p Ext when this
3332  /// operand is promotable and is not a supported trunc or sext.
3333  /// \p PromotedInsts maps the instructions to their type before promotion.
3334  /// \p CreatedInstsCost[out] contains the cost of all the instructions
3335  /// created to promote the operand of Ext.
3336  /// Newly added extensions are inserted in \p Exts.
3337  /// Newly added truncates are inserted in \p Truncs.
3338  /// Should never be called directly.
3339  /// \return The promoted value which is used instead of Ext.
3340  static Value *promoteOperandForOther(Instruction *Ext,
3341  TypePromotionTransaction &TPT,
3342  InstrToOrigTy &PromotedInsts,
3343  unsigned &CreatedInstsCost,
3346  const TargetLowering &TLI, bool IsSExt);
3347 
3348  /// \see promoteOperandForOther.
3349  static Value *signExtendOperandForOther(
3350  Instruction *Ext, TypePromotionTransaction &TPT,
3351  InstrToOrigTy &PromotedInsts, unsigned &CreatedInstsCost,
3353  SmallVectorImpl<Instruction *> *Truncs, const TargetLowering &TLI) {
3354  return promoteOperandForOther(Ext, TPT, PromotedInsts, CreatedInstsCost,
3355  Exts, Truncs, TLI, true);
3356  }
3357 
3358  /// \see promoteOperandForOther.
3359  static Value *zeroExtendOperandForOther(
3360  Instruction *Ext, TypePromotionTransaction &TPT,
3361  InstrToOrigTy &PromotedInsts, unsigned &CreatedInstsCost,
3363  SmallVectorImpl<Instruction *> *Truncs, const TargetLowering &TLI) {
3364  return promoteOperandForOther(Ext, TPT, PromotedInsts, CreatedInstsCost,
3365  Exts, Truncs, TLI, false);
3366  }
3367 
3368 public:
3369  /// Type for the utility function that promotes the operand of Ext.
3370  using Action = Value *(*)(Instruction *Ext, TypePromotionTransaction &TPT,
3371  InstrToOrigTy &PromotedInsts,
3372  unsigned &CreatedInstsCost,
3375  const TargetLowering &TLI);
3376 
3377  /// \brief Given a sign/zero extend instruction \p Ext, return the approriate
3378  /// action to promote the operand of \p Ext instead of using Ext.
3379  /// \return NULL if no promotable action is possible with the current
3380  /// sign extension.
3381  /// \p InsertedInsts keeps track of all the instructions inserted by the
3382  /// other CodeGenPrepare optimizations. This information is important
3383  /// because we do not want to promote these instructions as CodeGenPrepare
3384  /// will reinsert them later. Thus creating an infinite loop: create/remove.
3385  /// \p PromotedInsts maps the instructions to their type before promotion.
3386  static Action getAction(Instruction *Ext, const SetOfInstrs &InsertedInsts,
3387  const TargetLowering &TLI,
3388  const InstrToOrigTy &PromotedInsts);
3389 };
3390 
3391 } // end anonymous namespace
3392 
3393 bool TypePromotionHelper::canGetThrough(const Instruction *Inst,
3394  Type *ConsideredExtType,
3395  const InstrToOrigTy &PromotedInsts,
3396  bool IsSExt) {
3397  // The promotion helper does not know how to deal with vector types yet.
3398  // To be able to fix that, we would need to fix the places where we
3399  // statically extend, e.g., constants and such.
3400  if (Inst->getType()->isVectorTy())
3401  return false;
3402 
3403  // We can always get through zext.
3404  if (isa<ZExtInst>(Inst))
3405  return true;
3406 
3407  // sext(sext) is ok too.
3408  if (IsSExt && isa<SExtInst>(Inst))
3409  return true;
3410 
3411  // We can get through binary operator, if it is legal. In other words, the
3412  // binary operator must have a nuw or nsw flag.
3413  const BinaryOperator *BinOp = dyn_cast<BinaryOperator>(Inst);
3414  if (BinOp && isa<OverflowingBinaryOperator>(BinOp) &&
3415  ((!IsSExt && BinOp->hasNoUnsignedWrap()) ||
3416  (IsSExt && BinOp->hasNoSignedWrap())))
3417  return true;
3418 
3419  // Check if we can do the following simplification.
3420  // ext(trunc(opnd)) --> ext(opnd)
3421  if (!isa<TruncInst>(Inst))
3422  return false;
3423 
3424  Value *OpndVal = Inst->getOperand(0);
3425  // Check if we can use this operand in the extension.
3426  // If the type is larger than the result type of the extension, we cannot.
3427  if (!OpndVal->getType()->isIntegerTy() ||
3428  OpndVal->getType()->getIntegerBitWidth() >
3429  ConsideredExtType->getIntegerBitWidth())
3430  return false;
3431 
3432  // If the operand of the truncate is not an instruction, we will not have
3433  // any information on the dropped bits.
3434  // (Actually we could for constant but it is not worth the extra logic).
3435  Instruction *Opnd = dyn_cast<Instruction>(OpndVal);
3436  if (!Opnd)
3437  return false;
3438 
3439  // Check if the source of the type is narrow enough.
3440  // I.e., check that trunc just drops extended bits of the same kind of
3441  // the extension.
3442  // #1 get the type of the operand and check the kind of the extended bits.
3443  const Type *OpndType;
3444  InstrToOrigTy::const_iterator It = PromotedInsts.find(Opnd);
3445  if (It != PromotedInsts.end() && It->second.getInt() == IsSExt)
3446  OpndType = It->second.getPointer();
3447  else if ((IsSExt && isa<SExtInst>(Opnd)) || (!IsSExt && isa<ZExtInst>(Opnd)))
3448  OpndType = Opnd->getOperand(0)->getType();
3449  else
3450  return false;
3451 
3452  // #2 check that the truncate just drops extended bits.
3453  return Inst->getType()->getIntegerBitWidth() >=
3454  OpndType->getIntegerBitWidth();
3455 }
3456 
3457 TypePromotionHelper::Action TypePromotionHelper::getAction(
3458  Instruction *Ext, const SetOfInstrs &InsertedInsts,
3459  const TargetLowering &TLI, const InstrToOrigTy &PromotedInsts) {
3460  assert((isa<SExtInst>(Ext) || isa<ZExtInst>(Ext)) &&
3461  "Unexpected instruction type");
3462  Instruction *ExtOpnd = dyn_cast<Instruction>(Ext->getOperand(0));
3463  Type *ExtTy = Ext->getType();
3464  bool IsSExt = isa<SExtInst>(Ext);
3465  // If the operand of the extension is not an instruction, we cannot
3466  // get through.
3467  // If it, check we can get through.
3468  if (!ExtOpnd || !canGetThrough(ExtOpnd, ExtTy, PromotedInsts, IsSExt))
3469  return nullptr;
3470 
3471  // Do not promote if the operand has been added by codegenprepare.
3472  // Otherwise, it means we are undoing an optimization that is likely to be
3473  // redone, thus causing potential infinite loop.
3474  if (isa<TruncInst>(ExtOpnd) && InsertedInsts.count(ExtOpnd))
3475  return nullptr;
3476 
3477  // SExt or Trunc instructions.
3478  // Return the related handler.
3479  if (isa<SExtInst>(ExtOpnd) || isa<TruncInst>(ExtOpnd) ||
3480  isa<ZExtInst>(ExtOpnd))
3481  return promoteOperandForTruncAndAnyExt;
3482 
3483  // Regular instruction.
3484  // Abort early if we will have to insert non-free instructions.
3485  if (!ExtOpnd->hasOneUse() && !TLI.isTruncateFree(ExtTy, ExtOpnd->getType()))
3486  return nullptr;
3487  return IsSExt ? signExtendOperandForOther : zeroExtendOperandForOther;
3488 }
3489 
3490 Value *TypePromotionHelper::promoteOperandForTruncAndAnyExt(
3491  Instruction *SExt, TypePromotionTransaction &TPT,
3492  InstrToOrigTy &PromotedInsts, unsigned &CreatedInstsCost,
3494  SmallVectorImpl<Instruction *> *Truncs, const TargetLowering &TLI) {
3495  // By construction, the operand of SExt is an instruction. Otherwise we cannot
3496  // get through it and this method should not be called.
3497  Instruction *SExtOpnd = cast<Instruction>(SExt->getOperand(0));
3498  Value *ExtVal = SExt;
3499  bool HasMergedNonFreeExt = false;
3500  if (isa<ZExtInst>(SExtOpnd)) {
3501  // Replace s|zext(zext(opnd))
3502  // => zext(opnd).
3503  HasMergedNonFreeExt = !TLI.isExtFree(SExtOpnd);
3504  Value *ZExt =
3505  TPT.createZExt(SExt, SExtOpnd->getOperand(0), SExt->getType());
3506  TPT.replaceAllUsesWith(SExt, ZExt);
3507  TPT.eraseInstruction(SExt);
3508  ExtVal = ZExt;
3509  } else {
3510  // Replace z|sext(trunc(opnd)) or sext(sext(opnd))
3511  // => z|sext(opnd).
3512  TPT.setOperand(SExt, 0, SExtOpnd->getOperand(0));
3513  }
3514  CreatedInstsCost = 0;
3515 
3516  // Remove dead code.
3517  if (SExtOpnd->use_empty())
3518  TPT.eraseInstruction(SExtOpnd);
3519 
3520  // Check if the extension is still needed.
3521  Instruction *ExtInst = dyn_cast<Instruction>(ExtVal);
3522  if (!ExtInst || ExtInst->getType() != ExtInst->getOperand(0)->getType()) {
3523  if (ExtInst) {
3524  if (Exts)
3525  Exts->push_back(ExtInst);
3526  CreatedInstsCost = !TLI.isExtFree(ExtInst) && !HasMergedNonFreeExt;
3527  }
3528  return ExtVal;
3529  }
3530 
3531  // At this point we have: ext ty opnd to ty.
3532  // Reassign the uses of ExtInst to the opnd and remove ExtInst.
3533  Value *NextVal = ExtInst->getOperand(0);
3534  TPT.eraseInstruction(ExtInst, NextVal);
3535  return NextVal;
3536 }
3537 
3538 Value *TypePromotionHelper::promoteOperandForOther(
3539  Instruction *Ext, TypePromotionTransaction &TPT,
3540  InstrToOrigTy &PromotedInsts, unsigned &CreatedInstsCost,
3542  SmallVectorImpl<Instruction *> *Truncs, const TargetLowering &TLI,
3543  bool IsSExt) {
3544  // By construction, the operand of Ext is an instruction. Otherwise we cannot
3545  // get through it and this method should not be called.
3546  Instruction *ExtOpnd = cast<Instruction>(Ext->getOperand(0));
3547  CreatedInstsCost = 0;
3548  if (!ExtOpnd->hasOneUse()) {
3549  // ExtOpnd will be promoted.
3550  // All its uses, but Ext, will need to use a truncated value of the
3551  // promoted version.
3552  // Create the truncate now.
3553  Value *Trunc = TPT.createTrunc(Ext, ExtOpnd->getType());
3554  if (Instruction *ITrunc = dyn_cast<Instruction>(Trunc)) {
3555  // Insert it just after the definition.
3556  ITrunc->moveAfter(ExtOpnd);
3557  if (Truncs)
3558  Truncs->push_back(ITrunc);
3559  }
3560 
3561  TPT.replaceAllUsesWith(ExtOpnd, Trunc);
3562  // Restore the operand of Ext (which has been replaced by the previous call
3563  // to replaceAllUsesWith) to avoid creating a cycle trunc <-> sext.
3564  TPT.setOperand(Ext, 0, ExtOpnd);
3565  }
3566 
3567  // Get through the Instruction:
3568  // 1. Update its type.
3569  // 2. Replace the uses of Ext by Inst.
3570  // 3. Extend each operand that needs to be extended.
3571 
3572  // Remember the original type of the instruction before promotion.
3573  // This is useful to know that the high bits are sign extended bits.
3574  PromotedInsts.insert(std::pair<Instruction *, TypeIsSExt>(
3575  ExtOpnd, TypeIsSExt(ExtOpnd->getType(), IsSExt)));
3576  // Step #1.
3577  TPT.mutateType(ExtOpnd, Ext->getType());
3578  // Step #2.
3579  TPT.replaceAllUsesWith(Ext, ExtOpnd);
3580  // Step #3.
3581  Instruction *ExtForOpnd = Ext;
3582 
3583  DEBUG(dbgs() << "Propagate Ext to operands\n");
3584  for (int OpIdx = 0, EndOpIdx = ExtOpnd->getNumOperands(); OpIdx != EndOpIdx;
3585  ++OpIdx) {
3586  DEBUG(dbgs() << "Operand:\n" << *(ExtOpnd->getOperand(OpIdx)) << '\n');
3587  if (ExtOpnd->getOperand(OpIdx)->getType() == Ext->getType() ||
3588  !shouldExtOperand(ExtOpnd, OpIdx)) {
3589  DEBUG(dbgs() << "No need to propagate\n");
3590  continue;
3591  }
3592  // Check if we can statically extend the operand.
3593  Value *Opnd = ExtOpnd->getOperand(OpIdx);
3594  if (const ConstantInt *Cst = dyn_cast<ConstantInt>(Opnd)) {
3595  DEBUG(dbgs() << "Statically extend\n");
3596  unsigned BitWidth = Ext->getType()->getIntegerBitWidth();
3597  APInt CstVal = IsSExt ? Cst->getValue().sext(BitWidth)
3598  : Cst->getValue().zext(BitWidth);
3599  TPT.setOperand(ExtOpnd, OpIdx, ConstantInt::get(Ext->getType(), CstVal));
3600  continue;
3601  }
3602  // UndefValue are typed, so we have to statically sign extend them.
3603  if (isa<UndefValue>(Opnd)) {
3604  DEBUG(dbgs() << "Statically extend\n");
3605  TPT.setOperand(ExtOpnd, OpIdx, UndefValue::get(Ext->getType()));
3606  continue;
3607  }
3608 
3609  // Otherwise we have to explicity sign extend the operand.
3610  // Check if Ext was reused to extend an operand.
3611  if (!ExtForOpnd) {
3612  // If yes, create a new one.
3613  DEBUG(dbgs() << "More operands to ext\n");
3614  Value *ValForExtOpnd = IsSExt ? TPT.createSExt(Ext, Opnd, Ext->getType())
3615  : TPT.createZExt(Ext, Opnd, Ext->getType());
3616  if (!isa<Instruction>(ValForExtOpnd)) {
3617  TPT.setOperand(ExtOpnd, OpIdx, ValForExtOpnd);
3618  continue;
3619  }
3620  ExtForOpnd = cast<Instruction>(ValForExtOpnd);
3621  }
3622  if (Exts)
3623  Exts->push_back(ExtForOpnd);
3624  TPT.setOperand(ExtForOpnd, 0, Opnd);
3625 
3626  // Move the sign extension before the insertion point.
3627  TPT.moveBefore(ExtForOpnd, ExtOpnd);
3628  TPT.setOperand(ExtOpnd, OpIdx, ExtForOpnd);
3629  CreatedInstsCost += !TLI.isExtFree(ExtForOpnd);
3630  // If more sext are required, new instructions will have to be created.
3631  ExtForOpnd = nullptr;
3632  }
3633  if (ExtForOpnd == Ext) {
3634  DEBUG(dbgs() << "Extension is useless now\n");
3635  TPT.eraseInstruction(Ext);
3636  }
3637  return ExtOpnd;
3638 }
3639 
3640 /// Check whether or not promoting an instruction to a wider type is profitable.
3641 /// \p NewCost gives the cost of extension instructions created by the
3642 /// promotion.
3643 /// \p OldCost gives the cost of extension instructions before the promotion
3644 /// plus the number of instructions that have been
3645 /// matched in the addressing mode the promotion.
3646 /// \p PromotedOperand is the value that has been promoted.
3647 /// \return True if the promotion is profitable, false otherwise.
3648 bool AddressingModeMatcher::isPromotionProfitable(
3649  unsigned NewCost, unsigned OldCost, Value *PromotedOperand) const {
3650  DEBUG(dbgs() << "OldCost: " << OldCost << "\tNewCost: " << NewCost << '\n');
3651  // The cost of the new extensions is greater than the cost of the
3652  // old extension plus what we folded.
3653  // This is not profitable.
3654  if (NewCost > OldCost)
3655  return false;
3656  if (NewCost < OldCost)
3657  return true;
3658  // The promotion is neutral but it may help folding the sign extension in
3659  // loads for instance.
3660  // Check that we did not create an illegal instruction.
3661  return isPromotedInstructionLegal(TLI, DL, PromotedOperand);
3662 }
3663 
3664 /// Given an instruction or constant expr, see if we can fold the operation
3665 /// into the addressing mode. If so, update the addressing mode and return
3666 /// true, otherwise return false without modifying AddrMode.
3667 /// If \p MovedAway is not NULL, it contains the information of whether or
3668 /// not AddrInst has to be folded into the addressing mode on success.
3669 /// If \p MovedAway == true, \p AddrInst will not be part of the addressing
3670 /// because it has been moved away.
3671 /// Thus AddrInst must not be added in the matched instructions.
3672 /// This state can happen when AddrInst is a sext, since it may be moved away.
3673 /// Therefore, AddrInst may not be valid when MovedAway is true and it must
3674 /// not be referenced anymore.
3675 bool AddressingModeMatcher::matchOperationAddr(User *AddrInst, unsigned Opcode,
3676  unsigned Depth,
3677  bool *MovedAway) {
3678  // Avoid exponential behavior on extremely deep expression trees.
3679  if (Depth >= 5) return false;
3680 
3681  // By default, all matched instructions stay in place.
3682  if (MovedAway)
3683  *MovedAway = false;
3684 
3685  switch (Opcode) {
3686  case Instruction::PtrToInt:
3687  // PtrToInt is always a noop, as we know that the int type is pointer sized.
3688  return matchAddr(AddrInst->getOperand(0), Depth);
3689  case Instruction::IntToPtr: {
3690  auto AS = AddrInst->getType()->getPointerAddressSpace();
3691  auto PtrTy = MVT::getIntegerVT(DL.getPointerSizeInBits(AS));
3692  // This inttoptr is a no-op if the integer type is pointer sized.
3693  if (TLI.getValueType(DL, AddrInst->getOperand(0)->getType()) == PtrTy)
3694  return matchAddr(AddrInst->getOperand(0), Depth);
3695  return false;
3696  }
3697  case Instruction::BitCast:
3698  // BitCast is always a noop, and we can handle it as long as it is
3699  // int->int or pointer->pointer (we don't want int<->fp or something).
3700  if ((AddrInst->getOperand(0)->getType()->isPointerTy() ||
3701  AddrInst->getOperand(0)->getType()->isIntegerTy()) &&
3702  // Don't touch identity bitcasts. These were probably put here by LSR,
3703  // and we don't want to mess around with them. Assume it knows what it
3704  // is doing.
3705  AddrInst->getOperand(0)->getType() != AddrInst->getType())
3706  return matchAddr(AddrInst->getOperand(0), Depth);
3707  return false;
3708  case Instruction::AddrSpaceCast: {
3709  unsigned SrcAS
3710  = AddrInst->getOperand(0)->getType()->getPointerAddressSpace();
3711  unsigned DestAS = AddrInst->getType()->getPointerAddressSpace();
3712  if (TLI.isNoopAddrSpaceCast(SrcAS, DestAS))
3713  return matchAddr(AddrInst->getOperand(0), Depth);
3714  return false;
3715  }
3716  case Instruction::Add: {
3717  // Check to see if we can merge in the RHS then the LHS. If so, we win.
3718  ExtAddrMode BackupAddrMode = AddrMode;
3719  unsigned OldSize = AddrModeInsts.size();
3720  // Start a transaction at this point.
3721  // The LHS may match but not the RHS.
3722  // Therefore, we need a higher level restoration point to undo partially
3723  // matched operation.
3724  TypePromotionTransaction::ConstRestorationPt LastKnownGood =
3725  TPT.getRestorationPoint();
3726 
3727  if (matchAddr(AddrInst->getOperand(1), Depth+1) &&
3728  matchAddr(AddrInst->getOperand(0), Depth+1))
3729  return true;
3730 
3731  // Restore the old addr mode info.
3732  AddrMode = BackupAddrMode;
3733  AddrModeInsts.resize(OldSize);
3734  TPT.rollback(LastKnownGood);
3735 
3736  // Otherwise this was over-aggressive. Try merging in the LHS then the RHS.
3737  if (matchAddr(AddrInst->getOperand(0), Depth+1) &&
3738  matchAddr(AddrInst->getOperand(1), Depth+1))
3739  return true;
3740 
3741  // Otherwise we definitely can't merge the ADD in.
3742  AddrMode = BackupAddrMode;
3743  AddrModeInsts.resize(OldSize);
3744  TPT.rollback(LastKnownGood);
3745  break;
3746  }
3747  //case Instruction::Or:
3748  // TODO: We can handle "Or Val, Imm" iff this OR is equivalent to an ADD.
3749  //break;
3750  case Instruction::Mul:
3751  case Instruction::Shl: {
3752  // Can only handle X*C and X << C.
3753  ConstantInt *RHS = dyn_cast<ConstantInt>(AddrInst->getOperand(1));
3754  if (!RHS || RHS->getBitWidth() > 64)
3755  return false;
3756  int64_t Scale = RHS->getSExtValue();
3757  if (Opcode == Instruction::Shl)
3758  Scale = 1LL << Scale;
3759 
3760  return matchScaledValue(AddrInst->getOperand(0), Scale, Depth);
3761  }
3762  case Instruction::GetElementPtr: {
3763  // Scan the GEP. We check it if it contains constant offsets and at most
3764  // one variable offset.
3765  int VariableOperand = -1;
3766  unsigned VariableScale = 0;
3767 
3768  int64_t ConstantOffset = 0;
3769  gep_type_iterator GTI = gep_type_begin(AddrInst);
3770  for (unsigned i = 1, e = AddrInst->getNumOperands(); i != e; ++i, ++GTI) {
3771  if (StructType *STy = GTI.getStructTypeOrNull()) {
3772  const StructLayout *SL = DL.getStructLayout(STy);
3773  unsigned Idx =
3774  cast<ConstantInt>(AddrInst->getOperand(i))->getZExtValue();
3775  ConstantOffset += SL->getElementOffset(Idx);
3776  } else {
3777  uint64_t TypeSize = DL.getTypeAllocSize(GTI.getIndexedType());
3778  if (ConstantInt *CI = dyn_cast<ConstantInt>(AddrInst->getOperand(i))) {
3779  ConstantOffset += CI->getSExtValue()*TypeSize;
3780  } else if (TypeSize) { // Scales of zero don't do anything.
3781  // We only allow one variable index at the moment.
3782  if (VariableOperand != -1)
3783  return false;
3784 
3785  // Remember the variable index.
3786  VariableOperand = i;
3787  VariableScale = TypeSize;
3788  }
3789  }
3790  }
3791 
3792  // A common case is for the GEP to only do a constant offset. In this case,
3793  // just add it to the disp field and check validity.
3794  if (VariableOperand == -1) {
3795  AddrMode.BaseOffs += ConstantOffset;
3796  if (ConstantOffset == 0 ||
3797  TLI.isLegalAddressingMode(DL, AddrMode, AccessTy, AddrSpace)) {
3798  // Check to see if we can fold the base pointer in too.
3799  if (matchAddr(AddrInst->getOperand(0), Depth+1))
3800  return true;
3801  }
3802  AddrMode.BaseOffs -= ConstantOffset;
3803  return false;
3804  }
3805 
3806  // Save the valid addressing mode in case we can't match.
3807  ExtAddrMode BackupAddrMode = AddrMode;
3808  unsigned OldSize = AddrModeInsts.size();
3809 
3810  // See if the scale and offset amount is valid for this target.
3811  AddrMode.BaseOffs += ConstantOffset;
3812 
3813  // Match the base operand of the GEP.
3814  if (!matchAddr(AddrInst->getOperand(0), Depth+1)) {
3815  // If it couldn't be matched, just stuff the value in a register.
3816  if (AddrMode.HasBaseReg) {
3817  AddrMode = BackupAddrMode;
3818  AddrModeInsts.resize(OldSize);
3819  return false;
3820  }
3821  AddrMode.HasBaseReg = true;
3822  AddrMode.BaseReg = AddrInst->getOperand(0);
3823  }
3824 
3825  // Match the remaining variable portion of the GEP.
3826  if (!matchScaledValue(AddrInst->getOperand(VariableOperand), VariableScale,
3827  Depth)) {
3828  // If it couldn't be matched, try stuffing the base into a register
3829  // instead of matching it, and retrying the match of the scale.
3830  AddrMode = BackupAddrMode;
3831  AddrModeInsts.resize(OldSize);
3832  if (AddrMode.HasBaseReg)
3833  return false;
3834  AddrMode.HasBaseReg = true;
3835  AddrMode.BaseReg = AddrInst->getOperand(0);
3836  AddrMode.BaseOffs += ConstantOffset;
3837  if (!matchScaledValue(AddrInst->getOperand(VariableOperand),
3838  VariableScale, Depth)) {
3839  // If even that didn't work, bail.
3840  AddrMode = BackupAddrMode;
3841  AddrModeInsts.resize(OldSize);
3842  return false;
3843  }
3844  }
3845 
3846  return true;
3847  }
3848  case Instruction::SExt:
3849  case Instruction::ZExt: {
3850  Instruction *Ext = dyn_cast<Instruction>(AddrInst);
3851  if (!Ext)
3852  return false;
3853 
3854  // Try to move this ext out of the way of the addressing mode.
3855  // Ask for a method for doing so.
3856  TypePromotionHelper::Action TPH =
3857  TypePromotionHelper::getAction(Ext, InsertedInsts, TLI, PromotedInsts);
3858  if (!TPH)
3859  return false;
3860 
3861  TypePromotionTransaction::ConstRestorationPt LastKnownGood =
3862  TPT.getRestorationPoint();
3863  unsigned CreatedInstsCost = 0;
3864  unsigned ExtCost = !TLI.isExtFree(Ext);
3865  Value *PromotedOperand =
3866  TPH(Ext, TPT, PromotedInsts, CreatedInstsCost, nullptr, nullptr, TLI);
3867  // SExt has been moved away.
3868  // Thus either it will be rematched later in the recursive calls or it is
3869  // gone. Anyway, we must not fold it into the addressing mode at this point.
3870  // E.g.,
3871  // op = add opnd, 1
3872  // idx = ext op
3873  // addr = gep base, idx
3874  // is now:
3875  // promotedOpnd = ext opnd <- no match here
3876  // op = promoted_add promotedOpnd, 1 <- match (later in recursive calls)
3877  // addr = gep base, op <- match
3878  if (MovedAway)
3879  *MovedAway = true;
3880 
3881  assert(PromotedOperand &&
3882  "TypePromotionHelper should have filtered out those cases");
3883 
3884  ExtAddrMode BackupAddrMode = AddrMode;
3885  unsigned OldSize = AddrModeInsts.size();
3886 
3887  if (!matchAddr(PromotedOperand, Depth) ||
3888  // The total of the new cost is equal to the cost of the created
3889  // instructions.
3890  // The total of the old cost is equal to the cost of the extension plus
3891  // what we have saved in the addressing mode.
3892  !isPromotionProfitable(CreatedInstsCost,
3893  ExtCost + (AddrModeInsts.size() - OldSize),
3894  PromotedOperand)) {
3895  AddrMode = BackupAddrMode;
3896  AddrModeInsts.resize(OldSize);
3897  DEBUG(dbgs() << "Sign extension does not pay off: rollback\n");
3898  TPT.rollback(LastKnownGood);
3899  return false;
3900  }
3901  return true;
3902  }
3903  }
3904  return false;
3905 }
3906 
3907 /// If we can, try to add the value of 'Addr' into the current addressing mode.
3908 /// If Addr can't be added to AddrMode this returns false and leaves AddrMode
3909 /// unmodified. This assumes that Addr is either a pointer type or intptr_t
3910 /// for the target.
3911 ///
3912 bool AddressingModeMatcher::matchAddr(Value *Addr, unsigned Depth) {
3913  // Start a transaction at this point that we will rollback if the matching
3914  // fails.
3915  TypePromotionTransaction::ConstRestorationPt LastKnownGood =
3916  TPT.getRestorationPoint();
3917  if (ConstantInt *CI = dyn_cast<ConstantInt>(Addr)) {
3918  // Fold in immediates if legal for the target.
3919  AddrMode.BaseOffs += CI->getSExtValue();
3920  if (TLI.isLegalAddressingMode(DL, AddrMode, AccessTy, AddrSpace))
3921  return true;
3922  AddrMode.BaseOffs -= CI->getSExtValue();
3923  } else if (GlobalValue *GV = dyn_cast<GlobalValue>(Addr)) {
3924  // If this is a global variable, try to fold it into the addressing mode.
3925  if (!AddrMode.BaseGV) {
3926  AddrMode.BaseGV = GV;
3927  if (TLI.isLegalAddressingMode(DL, AddrMode, AccessTy, AddrSpace))
3928  return true;
3929  AddrMode.BaseGV = nullptr;
3930  }
3931  } else if (Instruction *I = dyn_cast<Instruction>(Addr)) {
3932  ExtAddrMode BackupAddrMode = AddrMode;
3933  unsigned OldSize = AddrModeInsts.size();
3934 
3935  // Check to see if it is possible to fold this operation.
3936  bool MovedAway = false;
3937  if (matchOperationAddr(I, I->getOpcode(), Depth, &MovedAway)) {
3938  // This instruction may have been moved away. If so, there is nothing
3939  // to check here.
3940  if (MovedAway)
3941  return true;
3942  // Okay, it's possible to fold this. Check to see if it is actually
3943  // *profitable* to do so. We use a simple cost model to avoid increasing
3944  // register pressure too much.
3945  if (I->hasOneUse() ||
3946  isProfitableToFoldIntoAddressingMode(I, BackupAddrMode, AddrMode)) {
3947  AddrModeInsts.push_back(I);
3948  return true;
3949  }
3950 
3951  // It isn't profitable to do this, roll back.
3952  //cerr << "NOT FOLDING: " << *I;
3953  AddrMode = BackupAddrMode;
3954  AddrModeInsts.resize(OldSize);
3955  TPT.rollback(LastKnownGood);
3956  }
3957  } else if (ConstantExpr *CE = dyn_cast<ConstantExpr>(Addr)) {
3958  if (matchOperationAddr(CE, CE->getOpcode(), Depth))
3959  return true;
3960  TPT.rollback(LastKnownGood);
3961  } else if (isa<ConstantPointerNull>(Addr)) {
3962  // Null pointer gets folded without affecting the addressing mode.
3963  return true;
3964  }
3965 
3966  // Worse case, the target should support [reg] addressing modes. :)
3967  if (!AddrMode.HasBaseReg) {
3968  AddrMode.HasBaseReg = true;
3969  AddrMode.BaseReg = Addr;
3970  // Still check for legality in case the target supports [imm] but not [i+r].
3971  if (TLI.isLegalAddressingMode(DL, AddrMode, AccessTy, AddrSpace))
3972  return true;
3973  AddrMode.HasBaseReg = false;
3974  AddrMode.BaseReg = nullptr;
3975  }
3976 
3977  // If the base register is already taken, see if we can do [r+r].
3978  if (AddrMode.Scale == 0) {
3979  AddrMode.Scale = 1;
3980  AddrMode.ScaledReg = Addr;
3981  if (TLI.isLegalAddressingMode(DL, AddrMode, AccessTy, AddrSpace))
3982  return true;
3983  AddrMode.Scale = 0;
3984  AddrMode.ScaledReg = nullptr;
3985  }
3986  // Couldn't match.
3987  TPT.rollback(LastKnownGood);
3988  return false;
3989 }
3990 
3991 /// Check to see if all uses of OpVal by the specified inline asm call are due
3992 /// to memory operands. If so, return true, otherwise return false.
3993 static bool IsOperandAMemoryOperand(CallInst *CI, InlineAsm *IA, Value *OpVal,
3994  const TargetLowering &TLI,
3995  const TargetRegisterInfo &TRI) {
3996  const Function *F = CI->getFunction();
3997  TargetLowering::AsmOperandInfoVector TargetConstraints =
3998  TLI.ParseConstraints(F->getParent()->getDataLayout(), &TRI,
3999  ImmutableCallSite(CI));
4000 
4001  for (unsigned i = 0, e = TargetConstraints.size(); i != e; ++i) {
4002  TargetLowering::AsmOperandInfo &OpInfo = TargetConstraints[i];
4003 
4004  // Compute the constraint code and ConstraintType to use.
4005  TLI.ComputeConstraintToUse(OpInfo, SDValue());
4006 
4007  // If this asm operand is our Value*, and if it isn't an indirect memory
4008  // operand, we can't fold it!
4009  if (OpInfo.CallOperandVal == OpVal &&
4011  !OpInfo.isIndirect))
4012  return false;
4013  }
4014 
4015  return true;
4016 }
4017 
4018 // Max number of memory uses to look at before aborting the search to conserve
4019 // compile time.
4020 static constexpr int MaxMemoryUsesToScan = 20;
4021 
4022 /// Recursively walk all the uses of I until we find a memory use.
4023 /// If we find an obviously non-foldable instruction, return true.
4024 /// Add the ultimately found memory instructions to MemoryUses.
4025 static bool FindAllMemoryUses(
4026  Instruction *I,
4027  SmallVectorImpl<std::pair<Instruction *, unsigned>> &MemoryUses,
4028  SmallPtrSetImpl<Instruction *> &ConsideredInsts, const TargetLowering &TLI,
4029  const TargetRegisterInfo &TRI, int SeenInsts = 0) {
4030  // If we already considered this instruction, we're done.
4031  if (!ConsideredInsts.insert(I).second)
4032  return false;
4033 
4034  // If this is an obviously unfoldable instruction, bail out.
4035  if (!MightBeFoldableInst(I))
4036  return true;
4037 
4038  const bool OptSize = I->getFunction()->optForSize();
4039 
4040  // Loop over all the uses, recursively processing them.
4041  for (Use &U : I->uses()) {
4042  // Conservatively return true if we're seeing a large number or a deep chain
4043  // of users. This avoids excessive compilation times in pathological cases.
4044  if (SeenInsts++ >= MaxMemoryUsesToScan)
4045  return true;
4046 
4047  Instruction *UserI = cast<Instruction>(U.getUser());
4048  if (LoadInst *LI = dyn_cast<LoadInst>(UserI)) {
4049  MemoryUses.push_back(std::make_pair(LI, U.getOperandNo()));
4050  continue;
4051  }
4052 
4053  if (StoreInst *SI = dyn_cast<StoreInst>(UserI)) {
4054  unsigned opNo = U.getOperandNo();
4055  if (opNo != StoreInst::getPointerOperandIndex())
4056  return true; // Storing addr, not into addr.
4057  MemoryUses.push_back(std::make_pair(SI, opNo));
4058  continue;
4059  }
4060 
4061  if (AtomicRMWInst *RMW = dyn_cast<AtomicRMWInst>(UserI)) {
4062  unsigned opNo = U.getOperandNo();
4064  return true; // Storing addr, not into addr.
4065  MemoryUses.push_back(std::make_pair(RMW, opNo));
4066  continue;
4067  }
4068 
4069  if (AtomicCmpXchgInst *CmpX = dyn_cast<AtomicCmpXchgInst>(UserI)) {
4070  unsigned opNo = U.getOperandNo();
4072  return true; // Storing addr, not into addr.
4073  MemoryUses.push_back(std::make_pair(CmpX, opNo));
4074  continue;
4075  }
4076 
4077  if (CallInst *CI = dyn_cast<CallInst>(UserI)) {
4078  // If this is a cold call, we can sink the addressing calculation into
4079  // the cold path. See optimizeCallInst
4080  if (!OptSize && CI->hasFnAttr(Attribute::Cold))
4081  continue;
4082 
4084  if (!IA) return true;
4085 
4086  // If this is a memory operand, we're cool, otherwise bail out.
4087  if (!IsOperandAMemoryOperand(CI, IA, I, TLI, TRI))
4088  return true;
4089  continue;
4090  }
4091 
4092  if (FindAllMemoryUses(UserI, MemoryUses, ConsideredInsts, TLI, TRI,
4093  SeenInsts))
4094  return true;
4095  }
4096 
4097  return false;
4098 }
4099 
4100 /// Return true if Val is already known to be live at the use site that we're
4101 /// folding it into. If so, there is no cost to include it in the addressing
4102 /// mode. KnownLive1 and KnownLive2 are two values that we know are live at the
4103 /// instruction already.
4104 bool AddressingModeMatcher::valueAlreadyLiveAtInst(Value *Val,Value *KnownLive1,
4105  Value *KnownLive2) {
4106  // If Val is either of the known-live values, we know it is live!
4107  if (Val == nullptr || Val == KnownLive1 || Val == KnownLive2)
4108  return true;
4109 
4110  // All values other than instructions and arguments (e.g. constants) are live.
4111  if (!isa<Instruction>(Val) && !isa<Argument>(Val)) return true;
4112 
4113  // If Val is a constant sized alloca in the entry block, it is live, this is
4114  // true because it is just a reference to the stack/frame pointer, which is
4115  // live for the whole function.
4116  if (AllocaInst *AI = dyn_cast<AllocaInst>(Val))
4117  if (AI->isStaticAlloca())
4118  return true;
4119 
4120  // Check to see if this value is already used in the memory instruction's
4121  // block. If so, it's already live into the block at the very least, so we
4122  // can reasonably fold it.
4123  return Val->isUsedInBasicBlock(MemoryInst->getParent());
4124 }
4125 
4126 /// It is possible for the addressing mode of the machine to fold the specified
4127 /// instruction into a load or store that ultimately uses it.
4128 /// However, the specified instruction has multiple uses.
4129 /// Given this, it may actually increase register pressure to fold it
4130 /// into the load. For example, consider this code:
4131 ///
4132 /// X = ...
4133 /// Y = X+1
4134 /// use(Y) -> nonload/store
4135 /// Z = Y+1
4136 /// load Z
4137 ///
4138 /// In this case, Y has multiple uses, and can be folded into the load of Z
4139 /// (yielding load [X+2]). However, doing this will cause both "X" and "X+1" to
4140 /// be live at the use(Y) line. If we don't fold Y into load Z, we use one
4141 /// fewer register. Since Y can't be folded into "use(Y)" we don't increase the
4142 /// number of computations either.
4143 ///
4144 /// Note that this (like most of CodeGenPrepare) is just a rough heuristic. If
4145 /// X was live across 'load Z' for other reasons, we actually *would* want to
4146 /// fold the addressing mode in the Z case. This would make Y die earlier.
4147 bool AddressingModeMatcher::
4148 isProfitableToFoldIntoAddressingMode(Instruction *I, ExtAddrMode &AMBefore,
4149  ExtAddrMode &AMAfter) {
4150  if (IgnoreProfitability) return true;
4151 
4152  // AMBefore is the addressing mode before this instruction was folded into it,
4153  // and AMAfter is the addressing mode after the instruction was folded. Get
4154  // the set of registers referenced by AMAfter and subtract out those
4155  // referenced by AMBefore: this is the set of values which folding in this
4156  // address extends the lifetime of.
4157  //
4158  // Note that there are only two potential values being referenced here,
4159  // BaseReg and ScaleReg (global addresses are always available, as are any
4160  // folded immediates).
4161  Value *BaseReg = AMAfter.BaseReg, *ScaledReg = AMAfter.ScaledReg;
4162 
4163  // If the BaseReg or ScaledReg was referenced by the previous addrmode, their
4164  // lifetime wasn't extended by adding this instruction.
4165  if (valueAlreadyLiveAtInst(BaseReg, AMBefore.BaseReg, AMBefore.ScaledReg))
4166  BaseReg = nullptr;
4167  if (valueAlreadyLiveAtInst(ScaledReg, AMBefore.BaseReg, AMBefore.ScaledReg))
4168  ScaledReg = nullptr;
4169 
4170  // If folding this instruction (and it's subexprs) didn't extend any live
4171  // ranges, we're ok with it.
4172  if (!BaseReg && !ScaledReg)
4173  return true;
4174 
4175  // If all uses of this instruction can have the address mode sunk into them,
4176  // we can remove the addressing mode and effectively trade one live register
4177  // for another (at worst.) In this context, folding an addressing mode into
4178  // the use is just a particularly nice way of sinking it.
4180  SmallPtrSet<Instruction*, 16> ConsideredInsts;
4181  if (FindAllMemoryUses(I, MemoryUses, ConsideredInsts, TLI, TRI))
4182  return false; // Has a non-memory, non-foldable use!
4183 
4184  // Now that we know that all uses of this instruction are part of a chain of
4185  // computation involving only operations that could theoretically be folded
4186  // into a memory use, loop over each of these memory operation uses and see
4187  // if they could *actually* fold the instruction. The assumption is that
4188  // addressing modes are cheap and that duplicating the computation involved
4189  // many times is worthwhile, even on a fastpath. For sinking candidates
4190  // (i.e. cold call sites), this serves as a way to prevent excessive code
4191  // growth since most architectures have some reasonable small and fast way to
4192  // compute an effective address. (i.e LEA on x86)
4193  SmallVector<Instruction*, 32> MatchedAddrModeInsts;
4194  for (unsigned i = 0, e = MemoryUses.size(); i != e; ++i) {
4195  Instruction *User = MemoryUses[i].first;
4196  unsigned OpNo = MemoryUses[i].second;
4197 
4198  // Get the access type of this use. If the use isn't a pointer, we don't
4199  // know what it accesses.
4200  Value *Address = User->getOperand(OpNo);
4201  PointerType *AddrTy = dyn_cast<PointerType>(Address->getType());
4202  if (!AddrTy)
4203  return false;
4204  Type *AddressAccessTy = AddrTy->getElementType();
4205  unsigned AS = AddrTy->getAddressSpace();
4206 
4207  // Do a match against the root of this address, ignoring profitability. This
4208  // will tell us if the addressing mode for the memory operation will
4209  // *actually* cover the shared instruction.
4210  ExtAddrMode Result;
4211  TypePromotionTransaction::ConstRestorationPt LastKnownGood =
4212  TPT.getRestorationPoint();
4213  AddressingModeMatcher Matcher(MatchedAddrModeInsts, TLI, TRI,
4214  AddressAccessTy, AS,
4215  MemoryInst, Result, InsertedInsts,
4216  PromotedInsts, TPT);
4217  Matcher.IgnoreProfitability = true;
4218  bool Success = Matcher.matchAddr(Address, 0);
4219  (void)Success; assert(Success && "Couldn't select *anything*?");
4220 
4221  // The match was to check the profitability, the changes made are not
4222  // part of the original matcher. Therefore, they should be dropped
4223  // otherwise the original matcher will not present the right state.
4224  TPT.rollback(LastKnownGood);
4225 
4226  // If the match didn't cover I, then it won't be shared by it.
4227  if (!is_contained(MatchedAddrModeInsts, I))
4228  return false;
4229 
4230  MatchedAddrModeInsts.clear();
4231  }
4232 
4233  return true;
4234 }
4235 
4236 /// Return true if the specified values are defined in a
4237 /// different basic block than BB.
4238 static bool IsNonLocalValue(Value *V, BasicBlock *BB) {
4239  if (Instruction *I = dyn_cast<Instruction>(V))
4240  return I->getParent() != BB;
4241  return false;
4242 }
4243 
4244 /// Sink addressing mode computation immediate before MemoryInst if doing so
4245 /// can be done without increasing register pressure. The need for the
4246 /// register pressure constraint means this can end up being an all or nothing
4247 /// decision for all uses of the same addressing computation.
4248 ///
4249 /// Load and Store Instructions often have addressing modes that can do
4250 /// significant amounts of computation. As such, instruction selection will try
4251 /// to get the load or store to do as much computation as possible for the
4252 /// program. The problem is that isel can only see within a single block. As
4253 /// such, we sink as much legal addressing mode work into the block as possible.
4254 ///
4255 /// This method is used to optimize both load/store and inline asms with memory
4256 /// operands. It's also used to sink addressing computations feeding into cold
4257 /// call sites into their (cold) basic block.
4258 ///
4259 /// The motivation for handling sinking into cold blocks is that doing so can
4260 /// both enable other address mode sinking (by satisfying the register pressure
4261 /// constraint above), and reduce register pressure globally (by removing the
4262 /// addressing mode computation from the fast path entirely.).
4263 bool CodeGenPrepare::optimizeMemoryInst(Instruction *MemoryInst, Value *Addr,
4264  Type *AccessTy, unsigned AddrSpace) {
4265  Value *Repl = Addr;
4266 
4267  // Try to collapse single-value PHI nodes. This is necessary to undo
4268  // unprofitable PRE transformations.
4269  SmallVector<Value*, 8> worklist;
4270  SmallPtrSet<Value*, 16> Visited;
4271  worklist.push_back(Addr);
4272 
4273  // Use a worklist to iteratively look through PHI and select nodes, and
4274  // ensure that the addressing mode obtained from the non-PHI/select roots of
4275  // the graph are compatible.
4276  bool PhiOrSelectSeen = false;
4277  SmallVector<Instruction*, 16> AddrModeInsts;
4278  const SimplifyQuery SQ(*DL, TLInfo);
4279  AddressingModeCombiner AddrModes(SQ, { Addr, MemoryInst->getParent() });
4280  TypePromotionTransaction TPT(RemovedInsts);
4281  TypePromotionTransaction::ConstRestorationPt LastKnownGood =
4282  TPT.getRestorationPoint();
4283  while (!worklist.empty()) {
4284  Value *V = worklist.back();
4285  worklist.pop_back();
4286 
4287  // We allow traversing cyclic Phi nodes.
4288  // In case of success after this loop we ensure that traversing through
4289  // Phi nodes ends up with all cases to compute address of the form
4290  // BaseGV + Base + Scale * Index + Offset
4291  // where Scale and Offset are constans and BaseGV, Base and Index
4292  // are exactly the same Values in all cases.
4293  // It means that BaseGV, Scale and Offset dominate our memory instruction
4294  // and have the same value as they had in address computation represented
4295  // as Phi. So we can safely sink address computation to memory instruction.
4296  if (!Visited.insert(V).second)
4297  continue;
4298 
4299  // For a PHI node, push all of its incoming values.
4300  if (PHINode *P = dyn_cast<PHINode>(V)) {
4301  for (Value *IncValue : P->incoming_values())
4302  worklist.push_back(IncValue);
4303  PhiOrSelectSeen = true;
4304  continue;
4305  }
4306  // Similar for select.
4307  if (SelectInst *SI = dyn_cast<SelectInst>(V)) {
4308  worklist.push_back(SI->getFalseValue());
4309  worklist.push_back(SI->getTrueValue());
4310  PhiOrSelectSeen = true;
4311  continue;
4312  }
4313 
4314  // For non-PHIs, determine the addressing mode being computed. Note that
4315  // the result may differ depending on what other uses our candidate
4316  // addressing instructions might have.
4317  AddrModeInsts.clear();
4318  ExtAddrMode NewAddrMode = AddressingModeMatcher::Match(
4319  V, AccessTy, AddrSpace, MemoryInst, AddrModeInsts, *TLI, *TRI,
4320  InsertedInsts, PromotedInsts, TPT);
4321  NewAddrMode.OriginalValue = V;
4322 
4323  if (!AddrModes.addNewAddrMode(NewAddrMode))
4324  break;
4325  }
4326 
4327  // Try to combine the AddrModes we've collected. If we couldn't collect any,
4328  // or we have multiple but either couldn't combine them or combining them
4329  // wouldn't do anything useful, bail out now.
4330  if (!AddrModes.combineAddrModes()) {
4331  TPT.rollback(LastKnownGood);
4332  return false;
4333  }
4334  TPT.commit();
4335 
4336  // Get the combined AddrMode (or the only AddrMode, if we only had one).
4337  ExtAddrMode AddrMode = AddrModes.getAddrMode();
4338 
4339  // If all the instructions matched are already in this BB, don't do anything.
4340  // If we saw a Phi node then it is not local definitely, and if we saw a select
4341  // then we want to push the address calculation past it even if it's already
4342  // in this BB.
4343  if (!PhiOrSelectSeen && none_of(AddrModeInsts, [&](Value *V) {
4344  return IsNonLocalValue(V, MemoryInst->getParent());
4345  })) {
4346  DEBUG(dbgs() << "CGP: Found local addrmode: " << AddrMode << "\n");
4347  return false;
4348  }
4349 
4350  // Insert this computation right after this user. Since our caller is
4351  // scanning from the top of the BB to the bottom, reuse of the expr are
4352  // guaranteed to happen later.
4353  IRBuilder<> Builder(MemoryInst);
4354 
4355  // Now that we determined the addressing expression we want to use and know
4356  // that we have to sink it into this block. Check to see if we have already
4357  // done this for some other load/store instr in this block. If so, reuse the
4358  // computation.
4359  Value *&SunkAddr = SunkAddrs[Addr];
4360  if (SunkAddr) {
4361  DEBUG(dbgs() << "CGP: Reusing nonlocal addrmode: " << AddrMode << " for "
4362  << *MemoryInst << "\n");
4363  if (SunkAddr->getType() != Addr->getType())
4364  SunkAddr = Builder.CreatePointerCast(SunkAddr, Addr->getType());
4365  } else if (AddrSinkUsingGEPs ||
4366  (!AddrSinkUsingGEPs.getNumOccurrences() && TM &&
4367  SubtargetInfo->useAA())) {
4368  // By default, we use the GEP-based method when AA is used later. This
4369  // prevents new inttoptr/ptrtoint pairs from degrading AA capabilities.
4370  DEBUG(dbgs() << "CGP: SINKING nonlocal addrmode: " << AddrMode << " for "
4371  << *MemoryInst << "\n");
4372  Type *IntPtrTy = DL->getIntPtrType(Addr->getType());
4373  Value *ResultPtr = nullptr, *ResultIndex = nullptr;
4374 
4375  // First, find the pointer.
4376  if (AddrMode.BaseReg && AddrMode.BaseReg->getType()->isPointerTy()) {
4377  ResultPtr = AddrMode.BaseReg;
4378  AddrMode.BaseReg = nullptr;
4379  }
4380 
4381  if (AddrMode.Scale && AddrMode.ScaledReg->getType()->isPointerTy()) {
4382  // We can't add more than one pointer together, nor can we scale a
4383  // pointer (both of which seem meaningless).
4384  if (ResultPtr || AddrMode.Scale != 1)
4385  return false;
4386 
4387  ResultPtr = AddrMode.ScaledReg;
4388  AddrMode.Scale = 0;
4389  }
4390 
4391  // It is only safe to sign extend the BaseReg if we know that the math
4392  // required to create it did not overflow before we extend it. Since
4393  // the original IR value was tossed in favor of a constant back when
4394  // the AddrMode was created we need to bail out gracefully if widths
4395  // do not match instead of extending it.
4396  //
4397  // (See below for code to add the scale.)
4398  if (AddrMode.Scale) {
4399  Type *ScaledRegTy = AddrMode.ScaledReg->getType();
4400  if (cast<IntegerType>(IntPtrTy)->getBitWidth() >
4401  cast<IntegerType>(ScaledRegTy)->getBitWidth())
4402  return false;
4403  }
4404 
4405  if (AddrMode.BaseGV) {
4406  if (ResultPtr)
4407  return false;
4408 
4409  ResultPtr = AddrMode.BaseGV;
4410  }
4411 
4412  // If the real base value actually came from an inttoptr, then the matcher
4413  // will look through it and provide only the integer value. In that case,
4414  // use it here.
4415  if (!DL->isNonIntegralPointerType(Addr->getType())) {
4416  if (!ResultPtr && AddrMode.BaseReg) {
4417  ResultPtr = Builder.CreateIntToPtr(AddrMode.BaseReg, Addr->getType(),
4418  "sunkaddr");
4419  AddrMode.BaseReg = nullptr;
4420  } else if (!ResultPtr && AddrMode.Scale == 1) {
4421  ResultPtr = Builder.CreateIntToPtr(AddrMode.ScaledReg, Addr->getType(),
4422  "sunkaddr");
4423  AddrMode.Scale = 0;
4424  }
4425  }
4426 
4427  if (!ResultPtr &&
4428  !AddrMode.BaseReg && !AddrMode.Scale && !AddrMode.BaseOffs) {
4429  SunkAddr = Constant::getNullValue(Addr->getType());
4430  } else if (!ResultPtr) {
4431  return false;
4432  } else {
4433  Type *I8PtrTy =
4434  Builder.getInt8PtrTy(Addr->getType()->getPointerAddressSpace());
4435  Type *I8Ty = Builder.getInt8Ty();
4436 
4437  // Start with the base register. Do this first so that subsequent address
4438  // matching finds it last, which will prevent it from trying to match it
4439  // as the scaled value in case it happens to be a mul. That would be
4440  // problematic if we've sunk a different mul for the scale, because then
4441  // we'd end up sinking both muls.
4442  if (AddrMode.BaseReg) {
4443  Value *V = AddrMode.BaseReg;
4444  if (V->getType() != IntPtrTy)
4445  V = Builder.CreateIntCast(V, IntPtrTy, /*isSigned=*/true, "sunkaddr");
4446 
4447  ResultIndex = V;
4448  }
4449 
4450  // Add the scale value.
4451  if (AddrMode.Scale) {
4452  Value *V = AddrMode.ScaledReg;
4453  if (V->getType() == IntPtrTy) {
4454  // done.
4455  } else {
4456  assert(cast<IntegerType>(IntPtrTy)->getBitWidth() <
4457  cast<IntegerType>(V->getType())->getBitWidth() &&
4458  "We can't transform if ScaledReg is too narrow");
4459  V = Builder.CreateTrunc(V, IntPtrTy, "sunkaddr");
4460  }
4461 
4462  if (AddrMode.Scale != 1)
4463  V = Builder.CreateMul(V, ConstantInt::get(IntPtrTy, AddrMode.Scale),
4464  "sunkaddr");
4465  if (ResultIndex)
4466  ResultIndex = Builder.CreateAdd(ResultIndex, V, "sunkaddr");
4467  else
4468  ResultIndex = V;
4469  }
4470 
4471  // Add in the Base Offset if present.
4472  if (AddrMode.BaseOffs) {
4473  Value *V = ConstantInt::get(IntPtrTy, AddrMode.BaseOffs);
4474  if (ResultIndex) {
4475  // We need to add this separately from the scale above to help with
4476  // SDAG consecutive load/store merging.
4477  if (ResultPtr->getType() != I8PtrTy)
4478  ResultPtr = Builder.CreatePointerCast(ResultPtr, I8PtrTy);
4479  ResultPtr = Builder.CreateGEP(I8Ty, ResultPtr, ResultIndex, "sunkaddr");
4480  }
4481 
4482  ResultIndex = V;
4483  }
4484 
4485  if (!ResultIndex) {
4486  SunkAddr = ResultPtr;
4487  } else {
4488  if (ResultPtr->getType() != I8PtrTy)
4489  ResultPtr = Builder.CreatePointerCast(ResultPtr, I8PtrTy);
4490  SunkAddr = Builder.CreateGEP(I8Ty, ResultPtr, ResultIndex, "sunkaddr");
4491  }
4492 
4493  if (SunkAddr->getType() != Addr->getType())
4494  SunkAddr = Builder.CreatePointerCast(SunkAddr, Addr->getType());
4495  }
4496  } else {
4497  // We'd require a ptrtoint/inttoptr down the line, which we can't do for
4498  // non-integral pointers, so in that case bail out now.
4499  Type *BaseTy = AddrMode.BaseReg ? AddrMode.BaseReg->getType() : nullptr;
4500  Type *ScaleTy = AddrMode.Scale ? AddrMode.ScaledReg->getType() : nullptr;
4501  PointerType *BasePtrTy = dyn_cast_or_null<PointerType>(BaseTy);
4502  PointerType *ScalePtrTy = dyn_cast_or_null<PointerType>(ScaleTy);
4503  if (DL->isNonIntegralPointerType(Addr->getType()) ||
4504  (BasePtrTy && DL->isNonIntegralPointerType(BasePtrTy)) ||
4505  (ScalePtrTy && DL->isNonIntegralPointerType(ScalePtrTy)) ||
4506  (AddrMode.BaseGV &&
4507  DL->isNonIntegralPointerType(AddrMode.BaseGV->getType())))
4508  return false;
4509 
4510  DEBUG(dbgs() << "CGP: SINKING nonlocal addrmode: " << AddrMode << " for "
4511  << *MemoryInst << "\n");
4512  Type *IntPtrTy = DL->getIntPtrType(Addr->getType());
4513  Value *Result = nullptr;
4514 
4515  // Start with the base register. Do this first so that subsequent address
4516  // matching finds it last, which will prevent it from trying to match it
4517  // as the scaled value in case it happens to be a mul. That would be
4518  // problematic if we've sunk a different mul for the scale, because then
4519  // we'd end up sinking both muls.
4520  if (AddrMode.BaseReg) {
4521  Value *V = AddrMode.BaseReg;
4522  if (V->getType()->isPointerTy())
4523  V = Builder.CreatePtrToInt(V, IntPtrTy, "sunkaddr");
4524  if (V->getType() != IntPtrTy)
4525  V = Builder.CreateIntCast(V, IntPtrTy, /*isSigned=*/true, "sunkaddr");
4526  Result = V;
4527  }
4528 
4529  // Add the scale value.
4530  if (AddrMode.Scale) {
4531  Value *V = AddrMode.ScaledReg;
4532  if (V->getType() == IntPtrTy) {
4533  // done.
4534  } else if (V->getType()->isPointerTy()) {
4535  V = Builder.CreatePtrToInt(V, IntPtrTy, "sunkaddr");
4536  } else if (cast<IntegerType>(IntPtrTy)->getBitWidth() <
4537  cast<IntegerType>(V->getType())->getBitWidth()) {
4538  V = Builder.CreateTrunc(V, IntPtrTy, "sunkaddr");
4539  } else {
4540  // It is only safe to sign extend the BaseReg if we know that the math
4541  // required to create it did not overflow before we extend it. Since
4542  // the original IR value was tossed in favor of a constant back when
4543  // the AddrMode was created we need to bail out gracefully if widths
4544  // do not match instead of extending it.
4545  Instruction *I = dyn_cast_or_null<Instruction>(Result);
4546  if (I && (Result != AddrMode.BaseReg))
4547  I->eraseFromParent();
4548  return false;
4549  }
4550  if (AddrMode.Scale != 1)
4551  V = Builder.CreateMul(V, ConstantInt::get(IntPtrTy, AddrMode.Scale),
4552  "sunkaddr");
4553  if (Result)
4554  Result = Builder.CreateAdd(Result, V, "sunkaddr");
4555  else
4556  Result = V;
4557  }
4558 
4559  // Add in the BaseGV if present.
4560  if (AddrMode.BaseGV) {
4561  Value *V = Builder.CreatePtrToInt(AddrMode.BaseGV, IntPtrTy, "sunkaddr");
4562  if (Result)
4563  Result = Builder.CreateAdd(Result, V, "sunkaddr");
4564  else
4565  Result = V;
4566  }
4567 
4568  // Add in the Base Offset if present.
4569  if (AddrMode.BaseOffs) {
4570  Value *V = ConstantInt::get(IntPtrTy, AddrMode.BaseOffs);
4571  if (Result)
4572  Result = Builder.CreateAdd(Result, V, "sunkaddr");
4573  else
4574  Result = V;
4575  }
4576 
4577  if (!Result)
4578  SunkAddr = Constant::getNullValue(Addr->getType());
4579  else
4580  SunkAddr = Builder.CreateIntToPtr(Result, Addr->getType(), "sunkaddr");
4581  }
4582 
4583  MemoryInst->replaceUsesOfWith(Repl, SunkAddr);
4584 
4585  // If we have no uses, recursively delete the value and all dead instructions
4586  // using it.
4587  if (Repl->use_empty()) {
4588  // This can cause recursive deletion, which can invalidate our iterator.
4589  // Use a WeakTrackingVH to hold onto it in case this happens.
4590  Value *CurValue = &*CurInstIterator;
4591  WeakTrackingVH IterHandle(CurValue);
4592  BasicBlock *BB = CurInstIterator->getParent();
4593 
4595 
4596  if (IterHandle != CurValue) {
4597  // If the iterator instruction was recursively deleted, start over at the
4598  // start of the block.
4599  CurInstIterator = BB->begin();
4600  SunkAddrs.clear();
4601  }
4602  }
4603  ++NumMemoryInsts;
4604  return true;
4605 }
4606 
4607 /// If there are any memory operands, use OptimizeMemoryInst to sink their
4608 /// address computing into the block when possible / profitable.
4609 bool CodeGenPrepare::optimizeInlineAsmInst(CallInst *CS) {
4610  bool MadeChange = false;
4611 
4612  const TargetRegisterInfo *TRI =
4613  TM->getSubtargetImpl(*CS->getFunction())->getRegisterInfo();
4614  TargetLowering::AsmOperandInfoVector TargetConstraints =
4615  TLI->ParseConstraints(*DL, TRI, CS);
4616  unsigned ArgNo = 0;
4617  for (unsigned i = 0, e = TargetConstraints.size(); i != e; ++i) {
4618  TargetLowering::AsmOperandInfo &OpInfo = TargetConstraints[i];
4619 
4620  // Compute the constraint code and ConstraintType to use.
4621  TLI->ComputeConstraintToUse(OpInfo, SDValue());
4622 
4623  if (OpInfo.ConstraintType == TargetLowering::C_Memory &&
4624  OpInfo.isIndirect) {
4625  Value *OpVal = CS->getArgOperand(ArgNo++);
4626  MadeChange |= optimizeMemoryInst(CS, OpVal, OpVal->getType(), ~0u);
4627  } else if (OpInfo.Type == InlineAsm::isInput)
4628  ArgNo++;
4629  }
4630 
4631  return MadeChange;
4632 }
4633 
4634 /// \brief Check if all the uses of \p Val are equivalent (or free) zero or
4635 /// sign extensions.
4636 static bool hasSameExtUse(Value *Val, const TargetLowering &TLI) {
4637  assert(!Val->use_empty() && "Input must have at least one use");
4638  const Instruction *FirstUser = cast<Instruction>(*Val->user_begin());
4639  bool IsSExt = isa<SExtInst>(FirstUser);
4640  Type *ExtTy = FirstUser->getType();
4641  for (const User *U : Val->users()) {
4642  const Instruction *UI = cast<Instruction>(U);
4643  if ((IsSExt && !isa<SExtInst>(UI)) || (!IsSExt && !isa<ZExtInst>(UI)))
4644  return false;
4645  Type *CurTy = UI->getType();
4646  // Same input and output types: Same instruction after CSE.
4647  if (CurTy == ExtTy)
4648  continue;
4649 
4650  // If IsSExt is true, we are in this situation:
4651  // a = Val
4652  // b = sext ty1 a to ty2
4653  // c = sext ty1 a to ty3
4654  // Assuming ty2 is shorter than ty3, this could be turned into:
4655  // a = Val
4656  // b = sext ty1 a to ty2
4657  // c = sext ty2 b to ty3
4658  // However, the last sext is not free.
4659  if (IsSExt)
4660  return false;
4661 
4662  // This is a ZExt, maybe this is free to extend from one type to another.
4663  // In that case, we would not account for a different use.
4664  Type *NarrowTy;
4665  Type *LargeTy;
4666  if (ExtTy->getScalarType()->getIntegerBitWidth() >
4667  CurTy->getScalarType()->getIntegerBitWidth()) {
4668  NarrowTy = CurTy;
4669  LargeTy = ExtTy;
4670  } else {
4671  NarrowTy = ExtTy;
4672  LargeTy = CurTy;
4673  }
4674 
4675  if (!TLI.isZExtFree(NarrowTy, LargeTy))
4676  return false;
4677  }
4678  // All uses are the same or can be derived from one another for free.
4679  return true;
4680 }
4681 
4682 /// \brief Try to speculatively promote extensions in \p Exts and continue
4683 /// promoting through newly promoted operands recursively as far as doing so is
4684 /// profitable. Save extensions profitably moved up, in \p ProfitablyMovedExts.
4685 /// When some promotion happened, \p TPT contains the proper state to revert
4686 /// them.
4687 ///
4688 /// \return true if some promotion happened, false otherwise.
4689 bool CodeGenPrepare::tryToPromoteExts(
4690  TypePromotionTransaction &TPT, const SmallVectorImpl<Instruction *> &Exts,
4691  SmallVectorImpl<Instruction *> &ProfitablyMovedExts,
4692  unsigned CreatedInstsCost) {
4693  bool Promoted = false;
4694 
4695  // Iterate over all the extensions to try to promote them.
4696  for (auto I : Exts) {
4697  // Early check if we directly have ext(load).
4698  if (isa<LoadInst>(I->getOperand(0))) {
4699  ProfitablyMovedExts.push_back(I);
4700  continue;
4701  }
4702 
4703  // Check whether or not we want to do any promotion. The reason we have
4704  // this check inside the for loop is to catch the case where an extension
4705  // is directly fed by a load because in such case the extension can be moved
4706  // up without any promotion on its operands.
4707  if (!TLI || !TLI->enableExtLdPromotion() || DisableExtLdPromotion)
4708  return false;
4709 
4710  // Get the action to perform the promotion.
4711  TypePromotionHelper::Action TPH =
4712  TypePromotionHelper::getAction(I, InsertedInsts, *TLI, PromotedInsts);
4713  // Check if we can promote.
4714  if (!TPH) {
4715  // Save the current extension as we cannot move up through its operand.
4716  ProfitablyMovedExts.push_back(I);
4717  continue;
4718  }
4719 
4720  // Save the current state.
4721  TypePromotionTransaction::ConstRestorationPt LastKnownGood =
4722  TPT.getRestorationPoint();
4724  unsigned NewCreatedInstsCost = 0;
4725  unsigned ExtCost = !TLI->isExtFree(I);
4726  // Promote.
4727  Value *PromotedVal = TPH(I, TPT, PromotedInsts, NewCreatedInstsCost,
4728  &NewExts, nullptr, *TLI);
4729  assert(PromotedVal &&
4730  "TypePromotionHelper should have filtered out those cases");
4731 
4732  // We would be able to merge only one extension in a load.
4733  // Therefore, if we have more than 1 new extension we heuristically
4734  // cut this search path, because it means we degrade the code quality.
4735  // With exactly 2, the transformation is neutral, because we will merge
4736  // one extension but leave one. However, we optimistically keep going,
4737  // because the new extension may be removed too.
4738  long long TotalCreatedInstsCost = CreatedInstsCost + NewCreatedInstsCost;
4739  // FIXME: It would be possible to propagate a negative value instead of
4740  // conservatively ceiling it to 0.
4741  TotalCreatedInstsCost =
4742  std::max((long long)0, (TotalCreatedInstsCost - ExtCost));
4743  if (!StressExtLdPromotion &&
4744  (TotalCreatedInstsCost > 1 ||
4745  !isPromotedInstructionLegal(*TLI, *DL, PromotedVal))) {
4746  // This promotion is not profitable, rollback to the previous state, and
4747  // save the current extension in ProfitablyMovedExts as the latest
4748  // speculative promotion turned out to be unprofitable.
4749  TPT.rollback(LastKnownGood);
4750  ProfitablyMovedExts.push_back(I);
4751  continue;
4752  }
4753  // Continue promoting NewExts as far as doing so is profitable.
4754  SmallVector<Instruction *, 2> NewlyMovedExts;
4755  (void)tryToPromoteExts(TPT, NewExts, NewlyMovedExts, TotalCreatedInstsCost);
4756  bool NewPromoted = false;
4757  for (auto ExtInst : NewlyMovedExts) {
4758  Instruction *MovedExt = cast<Instruction>(ExtInst);
4759  Value *ExtOperand = MovedExt->getOperand(0);
4760  // If we have reached to a load, we need this extra profitability check
4761  // as it could potentially be merged into an ext(load).
4762  if (isa<LoadInst>(ExtOperand) &&
4763  !(StressExtLdPromotion || NewCreatedInstsCost <= ExtCost ||
4764  (ExtOperand->hasOneUse() || hasSameExtUse(ExtOperand, *TLI))))
4765  continue;
4766 
4767  ProfitablyMovedExts.push_back(MovedExt);
4768  NewPromoted = true;
4769  }
4770 
4771  // If none of speculative promotions for NewExts is profitable, rollback
4772  // and save the current extension (I) as the last profitable extension.
4773  if (!NewPromoted) {
4774  TPT.rollback(LastKnownGood);
4775  ProfitablyMovedExts.push_back(I);
4776  continue;
4777  }
4778  // The promotion is profitable.
4779  Promoted = true;
4780  }
4781  return Promoted;
4782 }
4783 
4784 /// Merging redundant sexts when one is dominating the other.
4785 bool CodeGenPrepare::mergeSExts(Function &F) {
4786  DominatorTree DT(F);
4787  bool Changed = false;
4788  for (auto &Entry : ValToSExtendedUses) {
4789  SExts &Insts = Entry.second;
4790  SExts CurPts;
4791  for (Instruction *Inst : Insts) {
4792  if (RemovedInsts.count(Inst) || !isa<SExtInst>(Inst) ||
4793  Inst->getOperand(0) != Entry.first)
4794  continue;
4795  bool inserted = false;
4796  for (auto &Pt : CurPts) {
4797  if (DT.dominates(Inst, Pt)) {
4798  Pt->replaceAllUsesWith(Inst);
4799  RemovedInsts.insert(Pt);
4800  Pt->removeFromParent();
4801  Pt = Inst;
4802  inserted = true;
4803  Changed = true;
4804  break;
4805  }
4806  if (!DT.dominates(Pt, Inst))
4807  // Give up if we need to merge in a common dominator as the
4808  // expermients show it is not profitable.
4809  continue;
4810  Inst->replaceAllUsesWith(Pt);
4811  RemovedInsts.insert(Inst);
4812  Inst->removeFromParent();
4813  inserted = true;
4814  Changed = true;
4815  break;
4816  }
4817  if (!inserted)
4818  CurPts.push_back(Inst);
4819  }
4820  }
4821  return Changed;
4822 }
4823 
4824 /// Return true, if an ext(load) can be formed from an extension in
4825 /// \p MovedExts.
4826 bool CodeGenPrepare::canFormExtLd(
4827  const SmallVectorImpl<Instruction *> &MovedExts, LoadInst *&LI,
4828  Instruction *&Inst, bool HasPromoted) {
4829  for (auto *MovedExtInst : MovedExts) {
4830  if (isa<LoadInst>(MovedExtInst->getOperand(0))) {
4831  LI = cast<LoadInst>(MovedExtInst->getOperand(0));
4832  Inst = MovedExtInst;
4833  break;
4834  }
4835  }
4836  if (!LI)
4837  return false;
4838 
4839  // If they're already in the same block, there's nothing to do.
4840  // Make the cheap checks first if we did not promote.
4841  // If we promoted, we need to check if it is indeed profitable.
4842  if (!HasPromoted && LI->getParent() == Inst->getParent())
4843  return false;
4844 
4845  return TLI->isExtLoad(LI, Inst, *DL);
4846 }
4847 
4848 /// Move a zext or sext fed by a load into the same basic block as the load,
4849 /// unless conditions are unfavorable. This allows SelectionDAG to fold the
4850 /// extend into the load.
4851 ///
4852 /// E.g.,
4853 /// \code
4854 /// %ld = load i32* %addr
4855 /// %add = add nuw i32 %ld, 4
4856 /// %zext = zext i32 %add to i64
4857 // \endcode
4858 /// =>
4859 /// \code
4860 /// %ld = load i32* %addr
4861 /// %zext = zext i32 %ld to i64
4862 /// %add = add nuw i64 %zext, 4
4863 /// \encode
4864 /// Note that the promotion in %add to i64 is done in tryToPromoteExts(), which
4865 /// allow us to match zext(load i32*) to i64.
4866 ///
4867 /// Also, try to promote the computations used to obtain a sign extended
4868 /// value used into memory accesses.
4869 /// E.g.,
4870 /// \code
4871 /// a = add nsw i32 b, 3
4872 /// d = sext i32 a to i64
4873 /// e = getelementptr ..., i64 d
4874 /// \endcode
4875 /// =>
4876 /// \code
4877 /// f = sext i32 b to i64
4878 /// a = add nsw i64 f, 3
4879 /// e = getelementptr ..., i64 a
4880 /// \endcode
4881 ///
4882 /// \p Inst[in/out] the extension may be modified during the process if some
4883 /// promotions apply.
4884 bool CodeGenPrepare::optimizeExt(Instruction *&Inst) {
4885  // ExtLoad formation and address type promotion infrastructure requires TLI to
4886  // be effective.
4887  if (!TLI)
4888  return false;
4889 
4890  bool AllowPromotionWithoutCommonHeader = false;
4891  /// See if it is an interesting sext operations for the address type
4892  /// promotion before trying to promote it, e.g., the ones with the right
4893  /// type and used in memory accesses.
4894  bool ATPConsiderable = TTI->shouldConsiderAddressTypePromotion(
4895  *Inst, AllowPromotionWithoutCommonHeader);
4896  TypePromotionTransaction TPT(RemovedInsts);
4897  TypePromotionTransaction::ConstRestorationPt LastKnownGood =
4898  TPT.getRestorationPoint();
4900  SmallVector<Instruction *, 2> SpeculativelyMovedExts;
4901  Exts.push_back(Inst);
4902 
4903  bool HasPromoted = tryToPromoteExts(TPT, Exts, SpeculativelyMovedExts);
4904 
4905  // Look for a load being extended.
4906  LoadInst *LI = nullptr;
4907  Instruction *ExtFedByLoad;
4908 
4909  // Try to promote a chain of computation if it allows to form an extended
4910  // load.
4911  if (canFormExtLd(SpeculativelyMovedExts, LI, ExtFedByLoad, HasPromoted)) {
4912  assert(LI && ExtFedByLoad && "Expect a valid load and extension");
4913  TPT.commit();
4914  // Move the extend into the same block as the load
4915  ExtFedByLoad->moveAfter(LI);
4916  // CGP does not check if the zext would be speculatively executed when moved
4917  // to the same basic block as the load. Preserving its original location
4918  // would pessimize the debugging experience, as well as negatively impact
4919  // the quality of sample pgo. We don't want to use "line 0" as that has a
4920  // size cost in the line-table section and logically the zext can be seen as
4921  // part of the load. Therefore we conservatively reuse the same debug
4922  // location for the load and the zext.
4923  ExtFedByLoad->setDebugLoc(LI->getDebugLoc());
4924  ++NumExtsMoved;
4925  Inst = ExtFedByLoad;
4926  return true;
4927  }
4928 
4929  // Continue promoting SExts if known as considerable depending on targets.
4930  if (ATPConsiderable &&
4931  performAddressTypePromotion(Inst, AllowPromotionWithoutCommonHeader,
4932  HasPromoted, TPT, SpeculativelyMovedExts))
4933  return true;
4934 
4935  TPT.rollback(LastKnownGood);
4936  return false;
4937 }
4938 
4939 // Perform address type promotion if doing so is profitable.
4940 // If AllowPromotionWithoutCommonHeader == false, we should find other sext
4941 // instructions that sign extended the same initial value. However, if
4942 // AllowPromotionWithoutCommonHeader == true, we expect promoting the
4943 // extension is just profitable.
4944 bool CodeGenPrepare::performAddressTypePromotion(
4945  Instruction *&Inst, bool AllowPromotionWithoutCommonHeader,
4946  bool HasPromoted, TypePromotionTransaction &TPT,
4947  SmallVectorImpl<Instruction *> &SpeculativelyMovedExts) {
4948  bool Promoted = false;
4949  SmallPtrSet<Instruction *, 1> UnhandledExts;
4950  bool AllSeenFirst = true;
4951  for (auto I : SpeculativelyMovedExts) {
4952  Value *HeadOfChain = I->getOperand(0);
4954  SeenChainsForSExt.find(HeadOfChain);
4955  // If there is an unhandled SExt which has the same header, try to promote
4956  // it as well.
4957  if (AlreadySeen != SeenChainsForSExt.end()) {
4958  if (AlreadySeen->second != nullptr)
4959  UnhandledExts.insert(AlreadySeen->second);
4960  AllSeenFirst = false;
4961  }
4962  }
4963 
4964  if (!AllSeenFirst || (AllowPromotionWithoutCommonHeader &&
4965  SpeculativelyMovedExts.size() == 1)) {
4966  TPT.commit();
4967  if (HasPromoted)
4968  Promoted = true;
4969  for (auto I : SpeculativelyMovedExts) {
4970  Value *HeadOfChain = I->getOperand(0);
4971  SeenChainsForSExt[HeadOfChain] = nullptr;
4972  ValToSExtendedUses[HeadOfChain].push_back(I);
4973  }
4974  // Update Inst as promotion happen.
4975  Inst = SpeculativelyMovedExts.pop_back_val();
4976  } else {
4977  // This is the first chain visited from the header, keep the current chain
4978  // as unhandled. Defer to promote this until we encounter another SExt
4979  // chain derived from the same header.
4980  for (auto I : SpeculativelyMovedExts) {
4981  Value *HeadOfChain = I->getOperand(0);
4982  SeenChainsForSExt[HeadOfChain] = Inst;
4983  }
4984  return false;
4985  }
4986 
4987  if (!AllSeenFirst && !UnhandledExts.empty())
4988  for (auto VisitedSExt : UnhandledExts) {
4989  if (RemovedInsts.count(VisitedSExt))
4990  continue;
4991  TypePromotionTransaction TPT(RemovedInsts);
4994  Exts.push_back(VisitedSExt);
4995  bool HasPromoted = tryToPromoteExts(TPT, Exts, Chains);
4996  TPT.commit();
4997  if (HasPromoted)
4998  Promoted = true;
4999  for (auto I : Chains) {
5000  Value *HeadOfChain = I->getOperand(0);
5001  // Mark this as handled.
5002  SeenChainsForSExt[HeadOfChain] = nullptr;
5003  ValToSExtendedUses[HeadOfChain].push_back(I);
5004  }
5005  }
5006  return Promoted;
5007 }
5008 
5009 bool CodeGenPrepare::optimizeExtUses(Instruction *I) {
5010  BasicBlock *DefBB = I->getParent();
5011 
5012  // If the result of a {s|z}ext and its source are both live out, rewrite all
5013  // other uses of the source with result of extension.
5014  Value *Src = I->getOperand(0);
5015  if (Src->hasOneUse())
5016  return false;
5017 
5018  // Only do this xform if truncating is free.
5019  if (TLI && !TLI->isTruncateFree(I->getType(), Src->getType()))
5020  return false;
5021 
5022  // Only safe to perform the optimization if the source is also defined in
5023  // this block.
5024  if (!isa<Instruction>(Src) || DefBB != cast<Instruction>(Src)->getParent())
5025  return false;
5026 
5027  bool DefIsLiveOut = false;
5028  for (User *U : I->users()) {
5029  Instruction *UI = cast<Instruction>(U);
5030 
5031  // Figure out which BB this ext is used in.
5032  BasicBlock *UserBB = UI->getParent();
5033  if (UserBB == DefBB) continue;
5034  DefIsLiveOut = true;
5035  break;
5036  }
5037  if (!DefIsLiveOut)
5038  return false;
5039 
5040  // Make sure none of the uses are PHI nodes.
5041  for (User *U : Src->users()) {
5042  Instruction *UI = cast<Instruction>(U);
5043  BasicBlock *UserBB = UI->getParent();
5044  if (UserBB == DefBB) continue;
5045  // Be conservative. We don't want this xform to end up introducing
5046  // reloads just before load / store instructions.
5047  if (isa<PHINode>(UI) || isa<LoadInst>(UI) || isa<StoreInst>(UI))
5048  return false;
5049  }
5050 
5051  // InsertedTruncs - Only insert one trunc in each block once.
5052  DenseMap<BasicBlock*, Instruction*> InsertedTruncs;
5053 
5054  bool MadeChange = false;
5055  for (Use &U : Src->uses()) {
5056  Instruction *User = cast<Instruction>(U.getUser());
5057 
5058  // Figure out which BB this ext is used in.
5059  BasicBlock *UserBB = User->getParent();
5060  if (UserBB == DefBB) continue;
5061 
5062  // Both src and def are live in this block. Rewrite the use.
5063  Instruction *&InsertedTrunc = InsertedTruncs[UserBB];
5064 
5065  if (!InsertedTrunc) {
5066  BasicBlock::iterator InsertPt = UserBB->getFirstInsertionPt();
5067  assert(InsertPt != UserBB->end());
5068  InsertedTrunc = new TruncInst(I, Src->getType(), "", &*InsertPt);
5069  InsertedInsts.insert(InsertedTrunc);
5070  }
5071 
5072  // Replace a use of the {s|z}ext source with a use of the result.
5073  U = InsertedTrunc;
5074  ++NumExtUses;
5075  MadeChange = true;
5076  }
5077 
5078  return MadeChange;
5079 }
5080 
5081 // Find loads whose uses only use some of the loaded value's bits. Add an "and"
5082 // just after the load if the target can fold this into one extload instruction,
5083 // with the hope of eliminating some of the other later "and" instructions using
5084 // the loaded value. "and"s that are made trivially redundant by the insertion
5085 // of the new "and" are removed by this function, while others (e.g. those whose
5086 // path from the load goes through a phi) are left for isel to potentially
5087 // remove.
5088 //
5089 // For example:
5090 //
5091 // b0:
5092 // x = load i32
5093 // ...
5094 // b1:
5095 // y = and x, 0xff
5096 // z = use y
5097 //
5098 // becomes:
5099 //
5100 // b0:
5101 // x = load i32
5102 // x' = and x, 0xff
5103 // ...
5104 // b1:
5105 // z = use x'
5106 //
5107 // whereas:
5108 //
5109 // b0:
5110 // x1 = load i32
5111 // ...
5112 // b1:
5113 // x2 = load i32
5114 // ...
5115 // b2:
5116 // x = phi x1, x2
5117 // y = and x, 0xff
5118 //
5119 // becomes (after a call to optimizeLoadExt for each load):
5120 //
5121 // b0:
5122 // x1 = load i32
5123 // x1' = and x1, 0xff
5124 // ...
5125 // b1:
5126 // x2 = load i32
5127 // x2' = and x2, 0xff
5128 // ...
5129 // b2:
5130 // x = phi x1', x2'
5131 // y = and x, 0xff
5132 bool CodeGenPrepare::optimizeLoadExt(LoadInst *Load) {
5133  if (!Load->isSimple() ||
5134  !(Load->getType()->isIntegerTy() || Load->getType()->isPointerTy()))
5135  return false;
5136 
5137  // Skip loads we've already transformed.
5138  if (Load->hasOneUse() &&
5139  InsertedInsts.count(cast<Instruction>(*Load->user_begin())))
5140  return false;
5141 
5142  // Look at all uses of Load, looking through phis, to determine how many bits
5143  // of the loaded value are needed.
5146  SmallVector<Instruction *, 8> AndsToMaybeRemove;
5147  for (auto *U : Load->users())
5148  WorkList.push_back(cast<Instruction>(U));
5149 
5150  EVT LoadResultVT = TLI->getValueType(*DL, Load->getType());
5151  unsigned BitWidth = LoadResultVT.getSizeInBits();
5152  APInt DemandBits(BitWidth, 0);
5153  APInt WidestAndBits(BitWidth, 0);
5154 
5155  while (!WorkList.empty()) {
5156  Instruction *I = WorkList.back();
5157  WorkList.pop_back();
5158 
5159  // Break use-def graph loops.
5160  if (!Visited.insert(I).second)
5161  continue;
5162 
5163  // For a PHI node, push all of its users.
5164  if (auto *Phi = dyn_cast<PHINode>(I)) {
5165  for (auto *U : Phi->users())
5166  WorkList.push_back(cast<Instruction>(U));
5167  continue;
5168  }
5169 
5170  switch (I->getOpcode()) {
5171  case Instruction::And: {
5172  auto *AndC = dyn_cast<ConstantInt>(I->getOperand(1));
5173  if (!AndC)
5174  return false;
5175  APInt AndBits = AndC->getValue();
5176  DemandBits |= AndBits;
5177  // Keep track of the widest and mask we see.
5178  if (AndBits.ugt(WidestAndBits))
5179  WidestAndBits = AndBits;
5180  if (AndBits == WidestAndBits && I->getOperand(0) == Load)
5181  AndsToMaybeRemove.push_back(I);
5182  break;
5183  }
5184 
5185  case Instruction::Shl: {
5186  auto *ShlC = dyn_cast<ConstantInt>(I->getOperand(1));
5187  if (!ShlC)
5188  return false;
5189  uint64_t ShiftAmt = ShlC->getLimitedValue(BitWidth - 1);
5190  DemandBits.setLowBits(BitWidth - ShiftAmt);
5191  break;
5192  }
5193 
5194  case Instruction::Trunc: {
5195  EVT TruncVT = TLI->getValueType(*DL, I->getType());
5196  unsigned TruncBitWidth = TruncVT.getSizeInBits();
5197  DemandBits.setLowBits(TruncBitWidth);
5198  break;
5199  }
5200 
5201  default:
5202  return false;
5203  }
5204  }
5205 
5206  uint32_t ActiveBits = DemandBits.getActiveBits();
5207  // Avoid hoisting (and (load x) 1) since it is unlikely to be folded by the
5208  // target even if isLoadExtLegal says an i1 EXTLOAD is valid. For example,
5209  // for the AArch64 target isLoadExtLegal(ZEXTLOAD, i32, i1) returns true, but
5210  // (and (load x) 1) is not matched as a single instruction, rather as a LDR
5211  // followed by an AND.
5212  // TODO: Look into removing this restriction by fixing backends to either
5213  // return false for isLoadExtLegal for i1 or have them select this pattern to
5214  // a single instruction.
5215  //
5216  // Also avoid hoisting if we didn't see any ands with the exact DemandBits
5217  // mask, since these are the only ands that will be removed by isel.
5218  if (ActiveBits <= 1 || !DemandBits.isMask(ActiveBits) ||
5219  WidestAndBits != DemandBits)
5220  return false;
5221 
5222  LLVMContext &Ctx = Load->getType()->getContext();
5223  Type *TruncTy = Type::getIntNTy(Ctx, ActiveBits);
5224  EVT TruncVT = TLI->getValueType(*DL, TruncTy);
5225 
5226  // Reject cases that won't be matched as extloads.
5227  if (!LoadResultVT.bitsGT(TruncVT) || !TruncVT.isRound() ||
5228  !TLI->isLoadExtLegal(ISD::ZEXTLOAD, LoadResultVT, TruncVT))
5229  return false;
5230 
5231  IRBuilder<> Builder(Load->getNextNode());
5232  auto *NewAnd = dyn_cast<Instruction>(
5233  Builder.CreateAnd(Load, ConstantInt::get(Ctx, DemandBits)));
5234  // Mark this instruction as "inserted by CGP", so that other
5235  // optimizations don't touch it.
5236  InsertedInsts.insert(NewAnd);
5237 
5238  // Replace all uses of load with new and (except for the use of load in the
5239  // new and itself).
5240  Load->replaceAllUsesWith(NewAnd);
5241  NewAnd->setOperand(0, Load);
5242 
5243  // Remove any and instructions that are now redundant.
5244  for (auto *And : AndsToMaybeRemove)
5245  // Check that the and mask is the same as the one we decided to put on the
5246  // new and.
5247  if (cast<ConstantInt>(And->getOperand(1))->getValue() == DemandBits) {
5248  And->replaceAllUsesWith(NewAnd);
5249  if (&*CurInstIterator == And)
5250  CurInstIterator = std::next(And->getIterator());
5251  And->eraseFromParent();
5252  ++NumAndUses;
5253  }
5254 
5255  ++NumAndsAdded;
5256  return true;
5257 }
5258 
5259 /// Check if V (an operand of a select instruction) is an expensive instruction
5260 /// that is only used once.
5261 static bool sinkSelectOperand(const TargetTransformInfo *TTI, Value *V) {
5262  auto *I = dyn_cast<Instruction>(V);
5263  // If it's safe to speculatively execute, then it should not have side
5264  // effects; therefore, it's safe to sink and possibly *not* execute.
5265  return I && I->hasOneUse() && isSafeToSpeculativelyExecute(I) &&
5267 }
5268 
5269 /// Returns true if a SelectInst should be turned into an explicit branch.
5271  const TargetLowering *TLI,
5272  SelectInst *SI) {
5273  // If even a predictable select is cheap, then a branch can't be cheaper.
5274  if (!TLI->isPredictableSelectExpensive())
5275  return false;
5276 
5277  // FIXME: This should use the same heuristics as IfConversion to determine
5278  // whether a select is better represented as a branch.
5279 
5280  // If metadata tells us that the select condition is obviously predictable,
5281  // then we want to replace the select with a branch.
5282  uint64_t TrueWeight, FalseWeight;
5283  if (SI->extractProfMetadata(TrueWeight, FalseWeight)) {
5284  uint64_t Max = std::max(TrueWeight, FalseWeight);
5285  uint64_t Sum = TrueWeight + FalseWeight;
5286  if (Sum != 0) {
5287  auto Probability = BranchProbability::getBranchProbability(Max, Sum);
5288  if (Probability > TLI->getPredictableBranchThreshold())
5289  return true;
5290  }
5291  }
5292 
5293  CmpInst *Cmp = dyn_cast<CmpInst>(SI->getCondition());
5294 
5295  // If a branch is predictable, an out-of-order CPU can avoid blocking on its
5296  // comparison condition. If the compare has more than one use, there's
5297  // probably another cmov or setcc around, so it's not worth emitting a branch.
5298  if (!Cmp || !Cmp->hasOneUse())
5299  return false;
5300 
5301  // If either operand of the select is expensive and only needed on one side
5302  // of the select, we should form a branch.
5303  if (sinkSelectOperand(TTI, SI->getTrueValue()) ||
5304  sinkSelectOperand(TTI, SI->getFalseValue()))
5305  return true;
5306 
5307  return false;
5308 }
5309 
5310 /// If \p isTrue is true, return the true value of \p SI, otherwise return
5311 /// false value of \p SI. If the true/false value of \p SI is defined by any
5312 /// select instructions in \p Selects, look through the defining select
5313 /// instruction until the true/false value is not defined in \p Selects.
5315  SelectInst *SI, bool isTrue,
5316  const SmallPtrSet<const Instruction *, 2> &Selects) {
5317  Value *V;
5318 
5319  for (SelectInst *DefSI = SI; DefSI != nullptr && Selects.count(DefSI);
5320  DefSI = dyn_cast<SelectInst>(V)) {
5321  assert(DefSI->getCondition() == SI->getCondition() &&
5322  "The condition of DefSI does not match with SI");
5323  V = (isTrue ? DefSI->getTrueValue() : DefSI->getFalseValue());
5324  }
5325  return V;
5326 }
5327 
5328 /// If we have a SelectInst that will likely profit from branch prediction,
5329 /// turn it into a branch.
5330 bool CodeGenPrepare::optimizeSelectInst(SelectInst *SI) {
5331  // Find all consecutive select instructions that share the same condition.
5333  ASI.push_back(SI);
5335  It != SI->getParent()->end(); ++It) {
5336  SelectInst *I = dyn_cast<SelectInst>(&*It);
5337  if (I && SI->getCondition() == I->getCondition()) {
5338  ASI.push_back(I);
5339  } else {
5340  break;
5341  }
5342  }
5343 
5344  SelectInst *LastSI = ASI.back();
5345  // Increment the current iterator to skip all the rest of select instructions
5346  // because they will be either "not lowered" or "all lowered" to branch.
5347  CurInstIterator = std::next(LastSI->getIterator());
5348 
5349  bool VectorCond = !SI->getCondition()->getType()->isIntegerTy(1);
5350 
5351  // Can we convert the 'select' to CF ?
5352  if (DisableSelectToBranch || OptSize || !TLI || VectorCond ||
5354  return false;
5355 
5357  if (VectorCond)
5358  SelectKind = TargetLowering::VectorMaskSelect;
5359  else if (SI->getType()->isVectorTy())
5361  else
5362  SelectKind = TargetLowering::ScalarValSelect;
5363 
5364  if (TLI->isSelectSupported(SelectKind) &&
5365  !isFormingBranchFromSelectProfitable(TTI, TLI, SI))
5366  return false;
5367 
5368  ModifiedDT = true;
5369 
5370  // Transform a sequence like this:
5371  // start:
5372  // %cmp = cmp uge i32 %a, %b
5373  // %sel = select i1 %cmp, i32 %c, i32 %d
5374  //
5375  // Into:
5376  // start:
5377  // %cmp = cmp uge i32 %a, %b
5378  // br i1 %cmp, label %select.true, label %select.false
5379  // select.true:
5380  // br label %select.end
5381  // select.false:
5382  // br label %select.end
5383  // select.end:
5384  // %sel = phi i32 [ %c, %select.true ], [ %d, %select.false ]
5385  //
5386  // In addition, we may sink instructions that produce %c or %d from
5387  // the entry block into the destination(s) of the new branch.
5388  // If the true or false blocks do not contain a sunken instruction, that
5389  // block and its branch may be optimized away. In that case, one side of the
5390  // first branch will point directly to select.end, and the corresponding PHI
5391  // predecessor block will be the start block.
5392 
5393  // First, we split the block containing the select into 2 blocks.
5394  BasicBlock *StartBlock = SI->getParent();
5395  BasicBlock::iterator SplitPt = ++(BasicBlock::iterator(LastSI));
5396  BasicBlock *EndBlock = StartBlock->splitBasicBlock(SplitPt, "select.end");
5397 
5398  // Delete the unconditional branch that was just created by the split.
5399  StartBlock->getTerminator()->eraseFromParent();
5400 
5401  // These are the new basic blocks for the conditional branch.
5402  // At least one will become an actual new basic block.
5403  BasicBlock *TrueBlock = nullptr;
5404  BasicBlock *FalseBlock = nullptr;
5405  BranchInst *TrueBranch = nullptr;
5406  BranchInst *FalseBranch = nullptr;
5407 
5408  // Sink expensive instructions into the conditional blocks to avoid executing
5409  // them speculatively.
5410  for (SelectInst *SI : ASI) {
5411  if (sinkSelectOperand(TTI, SI->getTrueValue())) {
5412  if (TrueBlock == nullptr) {
5413  TrueBlock = BasicBlock::Create(SI->getContext(), "select.true.sink",
5414  EndBlock->getParent(), EndBlock);
5415  TrueBranch = BranchInst::Create(EndBlock, TrueBlock);
5416  }
5417  auto *TrueInst = cast<Instruction>(SI->getTrueValue());
5418  TrueInst->moveBefore(TrueBranch);
5419  }
5420  if (sinkSelectOperand(TTI, SI->getFalseValue())) {
5421  if (FalseBlock == nullptr) {
5422  FalseBlock = BasicBlock::Create(SI->getContext(), "select.false.sink",
5423  EndBlock->getParent(), EndBlock);
5424  FalseBranch = BranchInst::Create(EndBlock, FalseBlock);
5425  }
5426  auto *FalseInst = cast<Instruction>(SI->getFalseValue());
5427  FalseInst->moveBefore(FalseBranch);
5428  }
5429  }
5430 
5431  // If there was nothing to sink, then arbitrarily choose the 'false' side
5432  // for a new input value to the PHI.
5433  if (TrueBlock == FalseBlock) {
5434  assert(TrueBlock == nullptr &&
5435  "Unexpected basic block transform while optimizing select");
5436 
5437  FalseBlock = BasicBlock::Create(SI->getContext(), "select.false",
5438  EndBlock->getParent(), EndBlock);
5439  BranchInst::Create(EndBlock, FalseBlock);
5440  }
5441 
5442  // Insert the real conditional branch based on the original condition.
5443  // If we did not create a new block for one of the 'true' or 'false' paths
5444  // of the condition, it means that side of the branch goes to the end block
5445  // directly and the path originates from the start block from the point of
5446  // view of the new PHI.
5447  BasicBlock *TT, *FT;
5448  if (TrueBlock == nullptr) {
5449  TT = EndBlock;
5450  FT = FalseBlock;
5451  TrueBlock = StartBlock;
5452  } else if (FalseBlock == nullptr) {
5453  TT = TrueBlock;
5454  FT = EndBlock;
5455  FalseBlock = StartBlock;
5456  } else {
5457  TT = TrueBlock;
5458  FT = FalseBlock;
5459  }
5460  IRBuilder<>(SI).CreateCondBr(SI->getCondition(), TT, FT, SI);
5461 
5463  INS.insert(ASI.begin(), ASI.end());
5464  // Use reverse iterator because later select may use the value of the
5465  // earlier select, and we need to propagate value through earlier select
5466  // to get the PHI operand.
5467  for (auto It = ASI.rbegin(); It != ASI.rend(); ++It) {
5468  SelectInst *SI = *It;
5469  // The select itself is replaced with a PHI Node.
5470  PHINode *PN = PHINode::Create(SI->getType(), 2, "", &EndBlock->front());
5471  PN->takeName(SI);
5472  PN->addIncoming(getTrueOrFalseValue(SI, true, INS), TrueBlock);
5473  PN->addIncoming(getTrueOrFalseValue(SI, false, INS), FalseBlock);
5474 
5475  SI->replaceAllUsesWith(PN);
5476  SI->eraseFromParent();
5477  INS.erase(SI);
5478  ++NumSelectsExpanded;
5479  }
5480 
5481  // Instruct OptimizeBlock to skip to the next block.
5482  CurInstIterator = StartBlock->end();
5483  return true;
5484 }
5485 
5488  int SplatElem = -1;
5489  for (unsigned i = 0; i < Mask.size(); ++i) {
5490  if (SplatElem != -1 && Mask[i] != -1 && Mask[i] != SplatElem)
5491  return false;
5492  SplatElem = Mask[i];
5493  }
5494 
5495  return true;
5496 }
5497 
5498 /// Some targets have expensive vector shifts if the lanes aren't all the same
5499 /// (e.g. x86 only introduced "vpsllvd" and friends with AVX2). In these cases
5500 /// it's often worth sinking a shufflevector splat down to its use so that
5501 /// codegen can spot all lanes are identical.
5502 bool CodeGenPrepare::optimizeShuffleVectorInst(ShuffleVectorInst *SVI) {
5503  BasicBlock *DefBB = SVI->getParent();
5504 
5505  // Only do this xform if variable vector shifts are particularly expensive.
5506  if (!TLI || !TLI->isVectorShiftByScalarCheap(SVI->getType()))
5507  return false;
5508 
5509  // We only expect better codegen by sinking a shuffle if we can recognise a
5510  // constant splat.
5511  if (!isBroadcastShuffle(SVI))
5512  return false;
5513 
5514  // InsertedShuffles - Only insert a shuffle in each block once.
5515  DenseMap<BasicBlock*, Instruction*> InsertedShuffles;
5516 
5517  bool MadeChange = false;
5518  for (User *U : SVI->users()) {
5519  Instruction *UI = cast<Instruction>(U);
5520 
5521  // Figure out which BB this ext is used in.
5522  BasicBlock *UserBB = UI->getParent();
5523  if (UserBB == DefBB) continue;
5524 
5525  // For now only apply this when the splat is used by a shift instruction.
5526  if (!UI->isShift()) continue;
5527 
5528  // Everything checks out, sink the shuffle if the user's block doesn't
5529  // already have a copy.
5530  Instruction *&InsertedShuffle = InsertedShuffles[UserBB];
5531 
5532  if (!InsertedShuffle) {
5533  BasicBlock::iterator InsertPt = UserBB->getFirstInsertionPt();
5534  assert(InsertPt != UserBB->end());
5535  InsertedShuffle =
5536  new ShuffleVectorInst(SVI->getOperand(0), SVI->getOperand(1),
5537  SVI->getOperand(2), "", &*InsertPt);
5538  }
5539 
5540  UI->replaceUsesOfWith(SVI, InsertedShuffle);
5541  MadeChange = true;
5542  }
5543 
5544  // If we removed all uses, nuke the shuffle.
5545  if (SVI->use_empty()) {
5546  SVI->eraseFromParent();
5547  MadeChange = true;
5548  }
5549 
5550  return MadeChange;
5551 }
5552 
5553 bool CodeGenPrepare::optimizeSwitchInst(SwitchInst *SI) {
5554  if (!TLI || !DL)
5555  return false;
5556 
5557  Value *Cond = SI->getCondition();
5558  Type *OldType = Cond->getType();
5559  LLVMContext &Context = Cond->getContext();
5560  MVT RegType = TLI->getRegisterType(Context, TLI->getValueType(*DL, OldType));
5561  unsigned RegWidth = RegType.getSizeInBits();
5562 
5563  if (RegWidth <= cast<IntegerType>(OldType)->getBitWidth())
5564  return false;
5565 
5566  // If the register width is greater than the type width, expand the condition
5567  // of the switch instruction and each case constant to the width of the
5568  // register. By widening the type of the switch condition, subsequent
5569  // comparisons (for case comparisons) will not need to be extended to the
5570  // preferred register width, so we will potentially eliminate N-1 extends,
5571  // where N is the number of cases in the switch.
5572  auto *NewType = Type::getIntNTy(Context, RegWidth);
5573 
5574  // Zero-extend the switch condition and case constants unless the switch
5575  // condition is a function argument that is already being sign-extended.
5576  // In that case, we can avoid an unnecessary mask/extension by sign-extending
5577  // everything instead.
5578  Instruction::CastOps ExtType = Instruction::ZExt;
5579  if (auto *Arg = dyn_cast<Argument>(Cond))
5580  if (Arg->hasSExtAttr())
5581  ExtType = Instruction::SExt;
5582 
5583  auto *ExtInst = CastInst::Create(ExtType, Cond, NewType);
5584  ExtInst->insertBefore(SI);
5585  SI->setCondition(ExtInst);
5586  for (auto Case : SI->cases()) {
5587  APInt NarrowConst = Case.getCaseValue()->getValue();
5588  APInt WideConst = (ExtType == Instruction::ZExt) ?
5589  NarrowConst.zext(RegWidth) : NarrowConst.sext(RegWidth);
5590  Case.setValue(ConstantInt::get(Context, WideConst));
5591  }
5592 
5593  return true;
5594 }
5595 
5596 
5597 namespace {
5598 
5599 /// \brief Helper class to promote a scalar operation to a vector one.
5600 /// This class is used to move downward extractelement transition.
5601 /// E.g.,
5602 /// a = vector_op <2 x i32>
5603 /// b = extractelement <2 x i32> a, i32 0
5604 /// c = scalar_op b
5605 /// store c
5606 ///
5607 /// =>
5608 /// a = vector_op <2 x i32>
5609 /// c = vector_op a (equivalent to scalar_op on the related lane)
5610 /// * d = extractelement <2 x i32> c, i32 0
5611 /// * store d
5612 /// Assuming both extractelement and store can be combine, we get rid of the
5613 /// transition.
5614 class VectorPromoteHelper {
5615  /// DataLayout associated with the current module.
5616  const DataLayout &DL;
5617 
5618  /// Used to perform some checks on the legality of vector operations.
5619  const TargetLowering &TLI;
5620 
5621  /// Used to estimated the cost of the promoted chain.
5622  const TargetTransformInfo &TTI;
5623 
5624  /// The transition being moved downwards.
5625  Instruction *Transition;
5626 
5627  /// The sequence of instructions to be promoted.
5628  SmallVector<Instruction *, 4> InstsToBePromoted;
5629 
5630  /// Cost of combining a store and an extract.
5631  unsigned StoreExtractCombineCost;
5632 
5633  /// Instruction that will be combined with the transition.
5634  Instruction *CombineInst = nullptr;
5635 
5636  /// \brief The instruction that represents the current end of the transition.
5637  /// Since we are faking the promotion until we reach the end of the chain
5638  /// of computation, we need a way to get the current end of the transition.
5639  Instruction *getEndOfTransition() const {
5640  if (InstsToBePromoted.empty())
5641  return Transition;
5642  return InstsToBePromoted.back();
5643  }
5644 
5645  /// \brief Return the index of the original value in the transition.
5646  /// E.g., for "extractelement <2 x i32> c, i32 1" the original value,
5647  /// c, is at index 0.
5648  unsigned getTransitionOriginalValueIdx() const {
5649  assert(isa<ExtractElementInst>(Transition) &&
5650  "Other kind of transitions are not supported yet");
5651  return 0;
5652  }
5653 
5654  /// \brief Return the index of the index in the transition.
5655  /// E.g., for "extractelement <2 x i32> c, i32 0" the index
5656  /// is at index 1.
5657  unsigned getTransitionIdx() const {
5658  assert(isa<ExtractElementInst>(Transition) &&
5659  "Other kind of transitions are not supported yet");
5660  return 1;
5661  }
5662 
5663  /// \brief Get the type of the transition.
5664  /// This is the type of the original value.
5665  /// E.g., for "extractelement <2 x i32> c, i32 1" the type of the
5666  /// transition is <2 x i32>.
5667  Type *getTransitionType() const {
5668  return Transition->getOperand(getTransitionOriginalValueIdx())->getType();
5669  }
5670 
5671  /// \brief Promote \p ToBePromoted by moving \p Def downward through.
5672  /// I.e., we have the following sequence:
5673  /// Def = Transition <ty1> a to <ty2>
5674  /// b = ToBePromoted <ty2> Def, ...
5675  /// =>
5676  /// b = ToBePromoted <ty1> a, ...
5677  /// Def = Transition <ty1> ToBePromoted to <ty2>
5678  void promoteImpl(Instruction *ToBePromoted);
5679 
5680  /// \brief Check whether or not it is profitable to promote all the
5681  /// instructions enqueued to be promoted.
5682  bool isProfitableToPromote() {
5683  Value *ValIdx = Transition->getOperand(getTransitionOriginalValueIdx());
5684  unsigned Index = isa<ConstantInt>(ValIdx)
5685  ? cast<ConstantInt>(ValIdx)->getZExtValue()
5686  : -1;
5687  Type *PromotedType = getTransitionType();
5688 
5689  StoreInst *ST = cast<StoreInst>(CombineInst);
5690  unsigned AS = ST->getPointerAddressSpace();
5691  unsigned Align = ST->getAlignment();
5692  // Check if this store is supported.
5694  TLI.getValueType(DL, ST->getValueOperand()->getType()), AS,
5695  Align)) {
5696  // If this is not supported, there is no way we can combine
5697  // the extract with the store.
5698  return false;
5699  }
5700 
5701  // The scalar chain of computation has to pay for the transition
5702  // scalar to vector.
5703  // The vector chain has to account for the combining cost.
5704  uint64_t ScalarCost =
5705  TTI.getVectorInstrCost(Transition->getOpcode(), PromotedType, Index);
5706  uint64_t VectorCost = StoreExtractCombineCost;
5707  for (const auto &Inst : InstsToBePromoted) {
5708  // Compute the cost.
5709  // By construction, all instructions being promoted are arithmetic ones.
5710  // Moreover, one argument is a constant that can be viewed as a splat
5711  // constant.
5712  Value *Arg0 = Inst->getOperand(0);
5713  bool IsArg0Constant = isa<UndefValue>(Arg0) || isa<ConstantInt>(Arg0) ||
5714  isa<ConstantFP>(Arg0);
5721  ScalarCost += TTI.getArithmeticInstrCost(
5722  Inst->getOpcode(), Inst->getType(), Arg0OVK, Arg1OVK);
5723  VectorCost += TTI.getArithmeticInstrCost(Inst->getOpcode(), PromotedType,
5724  Arg0OVK, Arg1OVK);
5725  }
5726  DEBUG(dbgs() << "Estimated cost of computation to be promoted:\nScalar: "
5727  << ScalarCost << "\nVector: " << VectorCost << '\n');
5728  return ScalarCost > VectorCost;
5729  }
5730 
5731  /// \brief Generate a constant vector with \p Val with the same
5732  /// number of elements as the transition.
5733  /// \p UseSplat defines whether or not \p Val should be replicated
5734  /// across the whole vector.
5735  /// In other words, if UseSplat == true, we generate <Val, Val, ..., Val>,
5736  /// otherwise we generate a vector with as many undef as possible:
5737  /// <undef, ..., undef, Val, undef, ..., undef> where \p Val is only
5738  /// used at the index of the extract.
5739  Value *getConstantVector(Constant *Val, bool UseSplat) const {
5740  unsigned ExtractIdx = std::numeric_limits<unsigned>::max();
5741  if (!UseSplat) {
5742  // If we cannot determine where the constant must be, we have to
5743  // use a splat constant.
5744  Value *ValExtractIdx = Transition->getOperand(getTransitionIdx());
5745  if (ConstantInt *CstVal = dyn_cast<ConstantInt>(ValExtractIdx))
5746  ExtractIdx = CstVal->getSExtValue();
5747  else
5748  UseSplat = true;
5749  }
5750 
5751  unsigned End = getTransitionType()->getVectorNumElements();
5752  if (UseSplat)
5753  return ConstantVector::getSplat(End, Val);
5754 
5755  SmallVector<Constant *, 4> ConstVec;
5756  UndefValue *UndefVal = UndefValue::get(Val->getType());
5757  for (unsigned Idx = 0; Idx != End; ++Idx) {
5758  if (Idx == ExtractIdx)
5759  ConstVec.push_back(Val);
5760  else
5761  ConstVec.push_back(UndefVal);
5762  }
5763  return ConstantVector::get(ConstVec);
5764  }
5765 
5766  /// \brief Check if promoting to a vector type an operand at \p OperandIdx
5767  /// in \p Use can trigger undefined behavior.
5768  static bool canCauseUndefinedBehavior(const Instruction *Use,
5769  unsigned OperandIdx) {
5770  // This is not safe to introduce undef when the operand is on
5771  // the right hand side of a division-like instruction.
5772  if (OperandIdx != 1)
5773  return false;
5774  switch (Use->getOpcode()) {
5775  default:
5776  return false;
5777  case Instruction::SDiv:
5778  case Instruction::UDiv:
5779  case Instruction::SRem:
5780  case Instruction::URem:
5781  return true;
5782  case Instruction::FDiv:
5783  case Instruction::FRem:
5784  return !Use->hasNoNaNs();
5785  }
5786  llvm_unreachable(nullptr);
5787  }
5788 
5789 public:
5790  VectorPromoteHelper(const Dat