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/SmallPtrSet.h"
22 #include "llvm/ADT/SmallVector.h"
23 #include "llvm/ADT/Statistic.h"
28 #include "llvm/Analysis/LoopInfo.h"
34 #include "llvm/CodeGen/Analysis.h"
42 #include "llvm/IR/Argument.h"
43 #include "llvm/IR/Attributes.h"
44 #include "llvm/IR/BasicBlock.h"
45 #include "llvm/IR/CallSite.h"
46 #include "llvm/IR/Constant.h"
47 #include "llvm/IR/Constants.h"
48 #include "llvm/IR/DataLayout.h"
49 #include "llvm/IR/DerivedTypes.h"
50 #include "llvm/IR/Dominators.h"
51 #include "llvm/IR/Function.h"
53 #include "llvm/IR/GlobalValue.h"
54 #include "llvm/IR/GlobalVariable.h"
55 #include "llvm/IR/IRBuilder.h"
56 #include "llvm/IR/InlineAsm.h"
57 #include "llvm/IR/InstrTypes.h"
58 #include "llvm/IR/Instruction.h"
59 #include "llvm/IR/Instructions.h"
60 #include "llvm/IR/IntrinsicInst.h"
61 #include "llvm/IR/Intrinsics.h"
62 #include "llvm/IR/LLVMContext.h"
63 #include "llvm/IR/MDBuilder.h"
64 #include "llvm/IR/Module.h"
65 #include "llvm/IR/Operator.h"
66 #include "llvm/IR/PatternMatch.h"
67 #include "llvm/IR/Statepoint.h"
68 #include "llvm/IR/Type.h"
69 #include "llvm/IR/Use.h"
70 #include "llvm/IR/User.h"
71 #include "llvm/IR/Value.h"
72 #include "llvm/IR/ValueHandle.h"
73 #include "llvm/IR/ValueMap.h"
74 #include "llvm/Pass.h"
77 #include "llvm/Support/Casting.h"
79 #include "llvm/Support/Compiler.h"
80 #include "llvm/Support/Debug.h"
90 #include <algorithm>
91 #include <cassert>
92 #include <cstdint>
93 #include <iterator>
94 #include <limits>
95 #include <memory>
96 #include <utility>
97 #include <vector>
98 
99 using namespace llvm;
100 using namespace llvm::PatternMatch;
101 
102 #define DEBUG_TYPE "codegenprepare"
103 
104 STATISTIC(NumBlocksElim, "Number of blocks eliminated");
105 STATISTIC(NumPHIsElim, "Number of trivial PHIs eliminated");
106 STATISTIC(NumGEPsElim, "Number of GEPs converted to casts");
107 STATISTIC(NumCmpUses, "Number of uses of Cmp expressions replaced with uses of "
108  "sunken Cmps");
109 STATISTIC(NumCastUses, "Number of uses of Cast expressions replaced with uses "
110  "of sunken Casts");
111 STATISTIC(NumMemoryInsts, "Number of memory instructions whose address "
112  "computations were sunk");
113 STATISTIC(NumMemoryInstsPhiCreated,
114  "Number of phis created when address "
115  "computations were sunk to memory instructions");
116 STATISTIC(NumMemoryInstsSelectCreated,
117  "Number of select created when address "
118  "computations were sunk to memory instructions");
119 STATISTIC(NumExtsMoved, "Number of [s|z]ext instructions combined with loads");
120 STATISTIC(NumExtUses, "Number of uses of [s|z]ext instructions optimized");
121 STATISTIC(NumAndsAdded,
122  "Number of and mask instructions added to form ext loads");
123 STATISTIC(NumAndUses, "Number of uses of and mask instructions optimized");
124 STATISTIC(NumRetsDup, "Number of return instructions duplicated");
125 STATISTIC(NumDbgValueMoved, "Number of debug value instructions moved");
126 STATISTIC(NumSelectsExpanded, "Number of selects turned into branches");
127 STATISTIC(NumStoreExtractExposed, "Number of store(extractelement) exposed");
128 
130  "disable-cgp-branch-opts", cl::Hidden, cl::init(false),
131  cl::desc("Disable branch optimizations in CodeGenPrepare"));
132 
133 static cl::opt<bool>
134  DisableGCOpts("disable-cgp-gc-opts", cl::Hidden, cl::init(false),
135  cl::desc("Disable GC optimizations in CodeGenPrepare"));
136 
138  "disable-cgp-select2branch", cl::Hidden, cl::init(false),
139  cl::desc("Disable select to branch conversion."));
140 
142  "addr-sink-using-gep", cl::Hidden, cl::init(true),
143  cl::desc("Address sinking in CGP using GEPs."));
144 
146  "enable-andcmp-sinking", cl::Hidden, cl::init(true),
147  cl::desc("Enable sinkinig and/cmp into branches."));
148 
150  "disable-cgp-store-extract", cl::Hidden, cl::init(false),
151  cl::desc("Disable store(extract) optimizations in CodeGenPrepare"));
152 
154  "stress-cgp-store-extract", cl::Hidden, cl::init(false),
155  cl::desc("Stress test store(extract) optimizations in CodeGenPrepare"));
156 
158  "disable-cgp-ext-ld-promotion", cl::Hidden, cl::init(false),
159  cl::desc("Disable ext(promotable(ld)) -> promoted(ext(ld)) optimization in "
160  "CodeGenPrepare"));
161 
163  "stress-cgp-ext-ld-promotion", cl::Hidden, cl::init(false),
164  cl::desc("Stress test ext(promotable(ld)) -> promoted(ext(ld)) "
165  "optimization in CodeGenPrepare"));
166 
168  "disable-preheader-prot", cl::Hidden, cl::init(false),
169  cl::desc("Disable protection against removing loop preheaders"));
170 
172  "profile-guided-section-prefix", cl::Hidden, cl::init(true), cl::ZeroOrMore,
173  cl::desc("Use profile info to add section prefix for hot/cold functions"));
174 
176  "cgp-freq-ratio-to-skip-merge", cl::Hidden, cl::init(2),
177  cl::desc("Skip merging empty blocks if (frequency of empty block) / "
178  "(frequency of destination block) is greater than this ratio"));
179 
181  "force-split-store", cl::Hidden, cl::init(false),
182  cl::desc("Force store splitting no matter what the target query says."));
183 
184 static cl::opt<bool>
185 EnableTypePromotionMerge("cgp-type-promotion-merge", cl::Hidden,
186  cl::desc("Enable merging of redundant sexts when one is dominating"
187  " the other."), cl::init(true));
188 
190  "disable-complex-addr-modes", cl::Hidden, cl::init(false),
191  cl::desc("Disables combining addressing modes with different parts "
192  "in optimizeMemoryInst."));
193 
194 static cl::opt<bool>
195 AddrSinkNewPhis("addr-sink-new-phis", cl::Hidden, cl::init(false),
196  cl::desc("Allow creation of Phis in Address sinking."));
197 
198 static cl::opt<bool>
199 AddrSinkNewSelects("addr-sink-new-select", cl::Hidden, cl::init(false),
200  cl::desc("Allow creation of selects in Address sinking."));
201 
203  "addr-sink-combine-base-reg", cl::Hidden, cl::init(true),
204  cl::desc("Allow combining of BaseReg field in Address sinking."));
205 
207  "addr-sink-combine-base-gv", cl::Hidden, cl::init(true),
208  cl::desc("Allow combining of BaseGV field in Address sinking."));
209 
211  "addr-sink-combine-base-offs", cl::Hidden, cl::init(true),
212  cl::desc("Allow combining of BaseOffs field in Address sinking."));
213 
215  "addr-sink-combine-scaled-reg", cl::Hidden, cl::init(true),
216  cl::desc("Allow combining of ScaledReg field in Address sinking."));
217 
218 namespace {
219 
220 using SetOfInstrs = SmallPtrSet<Instruction *, 16>;
221 using TypeIsSExt = PointerIntPair<Type *, 1, bool>;
222 using InstrToOrigTy = DenseMap<Instruction *, TypeIsSExt>;
223 using SExts = SmallVector<Instruction *, 16>;
224 using ValueToSExts = DenseMap<Value *, SExts>;
225 
226 class TypePromotionTransaction;
227 
228  class CodeGenPrepare : public FunctionPass {
229  const TargetMachine *TM = nullptr;
230  const TargetSubtargetInfo *SubtargetInfo;
231  const TargetLowering *TLI = nullptr;
232  const TargetRegisterInfo *TRI;
233  const TargetTransformInfo *TTI = nullptr;
234  const TargetLibraryInfo *TLInfo;
235  const LoopInfo *LI;
236  std::unique_ptr<BlockFrequencyInfo> BFI;
237  std::unique_ptr<BranchProbabilityInfo> BPI;
238 
239  /// As we scan instructions optimizing them, this is the next instruction
240  /// to optimize. Transforms that can invalidate this should update it.
241  BasicBlock::iterator CurInstIterator;
242 
243  /// Keeps track of non-local addresses that have been sunk into a block.
244  /// This allows us to avoid inserting duplicate code for blocks with
245  /// multiple load/stores of the same address. The usage of WeakTrackingVH
246  /// enables SunkAddrs to be treated as a cache whose entries can be
247  /// invalidated if a sunken address computation has been erased.
249 
250  /// Keeps track of all instructions inserted for the current function.
251  SetOfInstrs InsertedInsts;
252 
253  /// Keeps track of the type of the related instruction before their
254  /// promotion for the current function.
255  InstrToOrigTy PromotedInsts;
256 
257  /// Keep track of instructions removed during promotion.
258  SetOfInstrs RemovedInsts;
259 
260  /// Keep track of sext chains based on their initial value.
261  DenseMap<Value *, Instruction *> SeenChainsForSExt;
262 
263  /// Keep track of SExt promoted.
264  ValueToSExts ValToSExtendedUses;
265 
266  /// True if CFG is modified in any way.
267  bool ModifiedDT;
268 
269  /// True if optimizing for size.
270  bool OptSize;
271 
272  /// DataLayout for the Function being processed.
273  const DataLayout *DL = nullptr;
274 
275  public:
276  static char ID; // Pass identification, replacement for typeid
277 
278  CodeGenPrepare() : FunctionPass(ID) {
280  }
281 
282  bool runOnFunction(Function &F) override;
283 
284  StringRef getPassName() const override { return "CodeGen Prepare"; }
285 
286  void getAnalysisUsage(AnalysisUsage &AU) const override {
287  // FIXME: When we can selectively preserve passes, preserve the domtree.
292  }
293 
294  private:
295  bool eliminateFallThrough(Function &F);
296  bool eliminateMostlyEmptyBlocks(Function &F);
297  BasicBlock *findDestBlockOfMergeableEmptyBlock(BasicBlock *BB);
298  bool canMergeBlocks(const BasicBlock *BB, const BasicBlock *DestBB) const;
299  void eliminateMostlyEmptyBlock(BasicBlock *BB);
300  bool isMergingEmptyBlockProfitable(BasicBlock *BB, BasicBlock *DestBB,
301  bool isPreheader);
302  bool optimizeBlock(BasicBlock &BB, bool &ModifiedDT);
303  bool optimizeInst(Instruction *I, bool &ModifiedDT);
304  bool optimizeMemoryInst(Instruction *I, Value *Addr,
305  Type *AccessTy, unsigned AS);
306  bool optimizeInlineAsmInst(CallInst *CS);
307  bool optimizeCallInst(CallInst *CI, bool &ModifiedDT);
308  bool optimizeExt(Instruction *&I);
309  bool optimizeExtUses(Instruction *I);
310  bool optimizeLoadExt(LoadInst *I);
311  bool optimizeSelectInst(SelectInst *SI);
312  bool optimizeShuffleVectorInst(ShuffleVectorInst *SI);
313  bool optimizeSwitchInst(SwitchInst *CI);
314  bool optimizeExtractElementInst(Instruction *Inst);
315  bool dupRetToEnableTailCallOpts(BasicBlock *BB);
316  bool placeDbgValues(Function &F);
317  bool canFormExtLd(const SmallVectorImpl<Instruction *> &MovedExts,
318  LoadInst *&LI, Instruction *&Inst, bool HasPromoted);
319  bool tryToPromoteExts(TypePromotionTransaction &TPT,
320  const SmallVectorImpl<Instruction *> &Exts,
321  SmallVectorImpl<Instruction *> &ProfitablyMovedExts,
322  unsigned CreatedInstsCost = 0);
323  bool mergeSExts(Function &F);
324  bool performAddressTypePromotion(
325  Instruction *&Inst,
326  bool AllowPromotionWithoutCommonHeader,
327  bool HasPromoted, TypePromotionTransaction &TPT,
328  SmallVectorImpl<Instruction *> &SpeculativelyMovedExts);
329  bool splitBranchCondition(Function &F);
330  bool simplifyOffsetableRelocate(Instruction &I);
331  };
332 
333 } // end anonymous namespace
334 
335 char CodeGenPrepare::ID = 0;
336 
337 INITIALIZE_PASS_BEGIN(CodeGenPrepare, DEBUG_TYPE,
338  "Optimize for code generation", false, false)
341  "Optimize for code generation", false, false)
342 
343 FunctionPass *llvm::createCodeGenPreparePass() { return new CodeGenPrepare(); }
344 
346  if (skipFunction(F))
347  return false;
348 
349  DL = &F.getParent()->getDataLayout();
350 
351  bool EverMadeChange = false;
352  // Clear per function information.
353  InsertedInsts.clear();
354  PromotedInsts.clear();
355  BFI.reset();
356  BPI.reset();
357 
358  ModifiedDT = false;
359  if (auto *TPC = getAnalysisIfAvailable<TargetPassConfig>()) {
360  TM = &TPC->getTM<TargetMachine>();
361  SubtargetInfo = TM->getSubtargetImpl(F);
362  TLI = SubtargetInfo->getTargetLowering();
363  TRI = SubtargetInfo->getRegisterInfo();
364  }
365  TLInfo = &getAnalysis<TargetLibraryInfoWrapperPass>().getTLI();
366  TTI = &getAnalysis<TargetTransformInfoWrapperPass>().getTTI(F);
367  LI = &getAnalysis<LoopInfoWrapperPass>().getLoopInfo();
368  OptSize = F.optForSize();
369 
370  ProfileSummaryInfo *PSI =
371  getAnalysis<ProfileSummaryInfoWrapperPass>().getPSI();
373  if (PSI->isFunctionHotInCallGraph(&F))
374  F.setSectionPrefix(".hot");
375  else if (PSI->isFunctionColdInCallGraph(&F))
376  F.setSectionPrefix(".unlikely");
377  }
378 
379  /// This optimization identifies DIV instructions that can be
380  /// profitably bypassed and carried out with a shorter, faster divide.
381  if (!OptSize && !PSI->hasHugeWorkingSetSize() && TLI &&
382  TLI->isSlowDivBypassed()) {
383  const DenseMap<unsigned int, unsigned int> &BypassWidths =
384  TLI->getBypassSlowDivWidths();
385  BasicBlock* BB = &*F.begin();
386  while (BB != nullptr) {
387  // bypassSlowDivision may create new BBs, but we don't want to reapply the
388  // optimization to those blocks.
389  BasicBlock* Next = BB->getNextNode();
390  EverMadeChange |= bypassSlowDivision(BB, BypassWidths);
391  BB = Next;
392  }
393  }
394 
395  // Eliminate blocks that contain only PHI nodes and an
396  // unconditional branch.
397  EverMadeChange |= eliminateMostlyEmptyBlocks(F);
398 
399  // llvm.dbg.value is far away from the value then iSel may not be able
400  // handle it properly. iSel will drop llvm.dbg.value if it can not
401  // find a node corresponding to the value.
402  EverMadeChange |= placeDbgValues(F);
403 
404  if (!DisableBranchOpts)
405  EverMadeChange |= splitBranchCondition(F);
406 
407  // Split some critical edges where one of the sources is an indirect branch,
408  // to help generate sane code for PHIs involving such edges.
409  EverMadeChange |= SplitIndirectBrCriticalEdges(F);
410 
411  bool MadeChange = true;
412  while (MadeChange) {
413  MadeChange = false;
414  SeenChainsForSExt.clear();
415  ValToSExtendedUses.clear();
416  RemovedInsts.clear();
417  for (Function::iterator I = F.begin(); I != F.end(); ) {
418  BasicBlock *BB = &*I++;
419  bool ModifiedDTOnIteration = false;
420  MadeChange |= optimizeBlock(*BB, ModifiedDTOnIteration);
421 
422  // Restart BB iteration if the dominator tree of the Function was changed
423  if (ModifiedDTOnIteration)
424  break;
425  }
426  if (EnableTypePromotionMerge && !ValToSExtendedUses.empty())
427  MadeChange |= mergeSExts(F);
428 
429  // Really free removed instructions during promotion.
430  for (Instruction *I : RemovedInsts)
431  I->deleteValue();
432 
433  EverMadeChange |= MadeChange;
434  }
435 
436  SunkAddrs.clear();
437 
438  if (!DisableBranchOpts) {
439  MadeChange = false;
441  for (BasicBlock &BB : F) {
442  SmallVector<BasicBlock *, 2> Successors(succ_begin(&BB), succ_end(&BB));
443  MadeChange |= ConstantFoldTerminator(&BB, true);
444  if (!MadeChange) continue;
445 
447  II = Successors.begin(), IE = Successors.end(); II != IE; ++II)
448  if (pred_begin(*II) == pred_end(*II))
449  WorkList.insert(*II);
450  }
451 
452  // Delete the dead blocks and any of their dead successors.
453  MadeChange |= !WorkList.empty();
454  while (!WorkList.empty()) {
455  BasicBlock *BB = *WorkList.begin();
456  WorkList.erase(BB);
457  SmallVector<BasicBlock*, 2> Successors(succ_begin(BB), succ_end(BB));
458 
459  DeleteDeadBlock(BB);
460 
462  II = Successors.begin(), IE = Successors.end(); II != IE; ++II)
463  if (pred_begin(*II) == pred_end(*II))
464  WorkList.insert(*II);
465  }
466 
467  // Merge pairs of basic blocks with unconditional branches, connected by
468  // a single edge.
469  if (EverMadeChange || MadeChange)
470  MadeChange |= eliminateFallThrough(F);
471 
472  EverMadeChange |= MadeChange;
473  }
474 
475  if (!DisableGCOpts) {
476  SmallVector<Instruction *, 2> Statepoints;
477  for (BasicBlock &BB : F)
478  for (Instruction &I : BB)
479  if (isStatepoint(I))
480  Statepoints.push_back(&I);
481  for (auto &I : Statepoints)
482  EverMadeChange |= simplifyOffsetableRelocate(*I);
483  }
484 
485  return EverMadeChange;
486 }
487 
488 /// Merge basic blocks which are connected by a single edge, where one of the
489 /// basic blocks has a single successor pointing to the other basic block,
490 /// which has a single predecessor.
491 bool CodeGenPrepare::eliminateFallThrough(Function &F) {
492  bool Changed = false;
493  // Scan all of the blocks in the function, except for the entry block.
494  for (Function::iterator I = std::next(F.begin()), E = F.end(); I != E;) {
495  BasicBlock *BB = &*I++;
496  // If the destination block has a single pred, then this is a trivial
497  // edge, just collapse it.
498  BasicBlock *SinglePred = BB->getSinglePredecessor();
499 
500  // Don't merge if BB's address is taken.
501  if (!SinglePred || SinglePred == BB || BB->hasAddressTaken()) continue;
502 
503  BranchInst *Term = dyn_cast<BranchInst>(SinglePred->getTerminator());
504  if (Term && !Term->isConditional()) {
505  Changed = true;
506  DEBUG(dbgs() << "To merge:\n"<< *SinglePred << "\n\n\n");
507  // Remember if SinglePred was the entry block of the function.
508  // If so, we will need to move BB back to the entry position.
509  bool isEntry = SinglePred == &SinglePred->getParent()->getEntryBlock();
510  MergeBasicBlockIntoOnlyPred(BB, nullptr);
511 
512  if (isEntry && BB != &BB->getParent()->getEntryBlock())
513  BB->moveBefore(&BB->getParent()->getEntryBlock());
514 
515  // We have erased a block. Update the iterator.
516  I = BB->getIterator();
517  }
518  }
519  return Changed;
520 }
521 
522 /// Find a destination block from BB if BB is mergeable empty block.
523 BasicBlock *CodeGenPrepare::findDestBlockOfMergeableEmptyBlock(BasicBlock *BB) {
524  // If this block doesn't end with an uncond branch, ignore it.
526  if (!BI || !BI->isUnconditional())
527  return nullptr;
528 
529  // If the instruction before the branch (skipping debug info) isn't a phi
530  // node, then other stuff is happening here.
531  BasicBlock::iterator BBI = BI->getIterator();
532  if (BBI != BB->begin()) {
533  --BBI;
534  while (isa<DbgInfoIntrinsic>(BBI)) {
535  if (BBI == BB->begin())
536  break;
537  --BBI;
538  }
539  if (!isa<DbgInfoIntrinsic>(BBI) && !isa<PHINode>(BBI))
540  return nullptr;
541  }
542 
543  // Do not break infinite loops.
544  BasicBlock *DestBB = BI->getSuccessor(0);
545  if (DestBB == BB)
546  return nullptr;
547 
548  if (!canMergeBlocks(BB, DestBB))
549  DestBB = nullptr;
550 
551  return DestBB;
552 }
553 
554 /// Eliminate blocks that contain only PHI nodes, debug info directives, and an
555 /// unconditional branch. Passes before isel (e.g. LSR/loopsimplify) often split
556 /// edges in ways that are non-optimal for isel. Start by eliminating these
557 /// blocks so we can split them the way we want them.
558 bool CodeGenPrepare::eliminateMostlyEmptyBlocks(Function &F) {
560  SmallVector<Loop *, 16> LoopList(LI->begin(), LI->end());
561  while (!LoopList.empty()) {
562  Loop *L = LoopList.pop_back_val();
563  LoopList.insert(LoopList.end(), L->begin(), L->end());
564  if (BasicBlock *Preheader = L->getLoopPreheader())
565  Preheaders.insert(Preheader);
566  }
567 
568  bool MadeChange = false;
569  // Note that this intentionally skips the entry block.
570  for (Function::iterator I = std::next(F.begin()), E = F.end(); I != E;) {
571  BasicBlock *BB = &*I++;
572  BasicBlock *DestBB = findDestBlockOfMergeableEmptyBlock(BB);
573  if (!DestBB ||
574  !isMergingEmptyBlockProfitable(BB, DestBB, Preheaders.count(BB)))
575  continue;
576 
577  eliminateMostlyEmptyBlock(BB);
578  MadeChange = true;
579  }
580  return MadeChange;
581 }
582 
583 bool CodeGenPrepare::isMergingEmptyBlockProfitable(BasicBlock *BB,
584  BasicBlock *DestBB,
585  bool isPreheader) {
586  // Do not delete loop preheaders if doing so would create a critical edge.
587  // Loop preheaders can be good locations to spill registers. If the
588  // preheader is deleted and we create a critical edge, registers may be
589  // spilled in the loop body instead.
590  if (!DisablePreheaderProtect && isPreheader &&
591  !(BB->getSinglePredecessor() &&
593  return false;
594 
595  // Try to skip merging if the unique predecessor of BB is terminated by a
596  // switch or indirect branch instruction, and BB is used as an incoming block
597  // of PHIs in DestBB. In such case, merging BB and DestBB would cause ISel to
598  // add COPY instructions in the predecessor of BB instead of BB (if it is not
599  // merged). Note that the critical edge created by merging such blocks wont be
600  // split in MachineSink because the jump table is not analyzable. By keeping
601  // such empty block (BB), ISel will place COPY instructions in BB, not in the
602  // predecessor of BB.
603  BasicBlock *Pred = BB->getUniquePredecessor();
604  if (!Pred ||
605  !(isa<SwitchInst>(Pred->getTerminator()) ||
606  isa<IndirectBrInst>(Pred->getTerminator())))
607  return true;
608 
609  if (BB->getTerminator() != BB->getFirstNonPHI())
610  return true;
611 
612  // We use a simple cost heuristic which determine skipping merging is
613  // profitable if the cost of skipping merging is less than the cost of
614  // merging : Cost(skipping merging) < Cost(merging BB), where the
615  // Cost(skipping merging) is Freq(BB) * (Cost(Copy) + Cost(Branch)), and
616  // the Cost(merging BB) is Freq(Pred) * Cost(Copy).
617  // Assuming Cost(Copy) == Cost(Branch), we could simplify it to :
618  // Freq(Pred) / Freq(BB) > 2.
619  // Note that if there are multiple empty blocks sharing the same incoming
620  // value for the PHIs in the DestBB, we consider them together. In such
621  // case, Cost(merging BB) will be the sum of their frequencies.
622 
623  if (!isa<PHINode>(DestBB->begin()))
624  return true;
625 
626  SmallPtrSet<BasicBlock *, 16> SameIncomingValueBBs;
627 
628  // Find all other incoming blocks from which incoming values of all PHIs in
629  // DestBB are the same as the ones from BB.
630  for (pred_iterator PI = pred_begin(DestBB), E = pred_end(DestBB); PI != E;
631  ++PI) {
632  BasicBlock *DestBBPred = *PI;
633  if (DestBBPred == BB)
634  continue;
635 
636  bool HasAllSameValue = true;
637  BasicBlock::const_iterator DestBBI = DestBB->begin();
638  while (const PHINode *DestPN = dyn_cast<PHINode>(DestBBI++)) {
639  if (DestPN->getIncomingValueForBlock(BB) !=
640  DestPN->getIncomingValueForBlock(DestBBPred)) {
641  HasAllSameValue = false;
642  break;
643  }
644  }
645  if (HasAllSameValue)
646  SameIncomingValueBBs.insert(DestBBPred);
647  }
648 
649  // See if all BB's incoming values are same as the value from Pred. In this
650  // case, no reason to skip merging because COPYs are expected to be place in
651  // Pred already.
652  if (SameIncomingValueBBs.count(Pred))
653  return true;
654 
655  if (!BFI) {
656  Function &F = *BB->getParent();
657  LoopInfo LI{DominatorTree(F)};
658  BPI.reset(new BranchProbabilityInfo(F, LI));
659  BFI.reset(new BlockFrequencyInfo(F, *BPI, LI));
660  }
661 
662  BlockFrequency PredFreq = BFI->getBlockFreq(Pred);
663  BlockFrequency BBFreq = BFI->getBlockFreq(BB);
664 
665  for (auto SameValueBB : SameIncomingValueBBs)
666  if (SameValueBB->getUniquePredecessor() == Pred &&
667  DestBB == findDestBlockOfMergeableEmptyBlock(SameValueBB))
668  BBFreq += BFI->getBlockFreq(SameValueBB);
669 
670  return PredFreq.getFrequency() <=
672 }
673 
674 /// Return true if we can merge BB into DestBB if there is a single
675 /// unconditional branch between them, and BB contains no other non-phi
676 /// instructions.
677 bool CodeGenPrepare::canMergeBlocks(const BasicBlock *BB,
678  const BasicBlock *DestBB) const {
679  // We only want to eliminate blocks whose phi nodes are used by phi nodes in
680  // the successor. If there are more complex condition (e.g. preheaders),
681  // don't mess around with them.
682  BasicBlock::const_iterator BBI = BB->begin();
683  while (const PHINode *PN = dyn_cast<PHINode>(BBI++)) {
684  for (const User *U : PN->users()) {
685  const Instruction *UI = cast<Instruction>(U);
686  if (UI->getParent() != DestBB || !isa<PHINode>(UI))
687  return false;
688  // If User is inside DestBB block and it is a PHINode then check
689  // incoming value. If incoming value is not from BB then this is
690  // a complex condition (e.g. preheaders) we want to avoid here.
691  if (UI->getParent() == DestBB) {
692  if (const PHINode *UPN = dyn_cast<PHINode>(UI))
693  for (unsigned I = 0, E = UPN->getNumIncomingValues(); I != E; ++I) {
694  Instruction *Insn = dyn_cast<Instruction>(UPN->getIncomingValue(I));
695  if (Insn && Insn->getParent() == BB &&
696  Insn->getParent() != UPN->getIncomingBlock(I))
697  return false;
698  }
699  }
700  }
701  }
702 
703  // If BB and DestBB contain any common predecessors, then the phi nodes in BB
704  // and DestBB may have conflicting incoming values for the block. If so, we
705  // can't merge the block.
706  const PHINode *DestBBPN = dyn_cast<PHINode>(DestBB->begin());
707  if (!DestBBPN) return true; // no conflict.
708 
709  // Collect the preds of BB.
711  if (const PHINode *BBPN = dyn_cast<PHINode>(BB->begin())) {
712  // It is faster to get preds from a PHI than with pred_iterator.
713  for (unsigned i = 0, e = BBPN->getNumIncomingValues(); i != e; ++i)
714  BBPreds.insert(BBPN->getIncomingBlock(i));
715  } else {
716  BBPreds.insert(pred_begin(BB), pred_end(BB));
717  }
718 
719  // Walk the preds of DestBB.
720  for (unsigned i = 0, e = DestBBPN->getNumIncomingValues(); i != e; ++i) {
721  BasicBlock *Pred = DestBBPN->getIncomingBlock(i);
722  if (BBPreds.count(Pred)) { // Common predecessor?
723  BBI = DestBB->begin();
724  while (const PHINode *PN = dyn_cast<PHINode>(BBI++)) {
725  const Value *V1 = PN->getIncomingValueForBlock(Pred);
726  const Value *V2 = PN->getIncomingValueForBlock(BB);
727 
728  // If V2 is a phi node in BB, look up what the mapped value will be.
729  if (const PHINode *V2PN = dyn_cast<PHINode>(V2))
730  if (V2PN->getParent() == BB)
731  V2 = V2PN->getIncomingValueForBlock(Pred);
732 
733  // If there is a conflict, bail out.
734  if (V1 != V2) return false;
735  }
736  }
737  }
738 
739  return true;
740 }
741 
742 /// Eliminate a basic block that has only phi's and an unconditional branch in
743 /// it.
744 void CodeGenPrepare::eliminateMostlyEmptyBlock(BasicBlock *BB) {
745  BranchInst *BI = cast<BranchInst>(BB->getTerminator());
746  BasicBlock *DestBB = BI->getSuccessor(0);
747 
748  DEBUG(dbgs() << "MERGING MOSTLY EMPTY BLOCKS - BEFORE:\n" << *BB << *DestBB);
749 
750  // If the destination block has a single pred, then this is a trivial edge,
751  // just collapse it.
752  if (BasicBlock *SinglePred = DestBB->getSinglePredecessor()) {
753  if (SinglePred != DestBB) {
754  // Remember if SinglePred was the entry block of the function. If so, we
755  // will need to move BB back to the entry position.
756  bool isEntry = SinglePred == &SinglePred->getParent()->getEntryBlock();
757  MergeBasicBlockIntoOnlyPred(DestBB, nullptr);
758 
759  if (isEntry && BB != &BB->getParent()->getEntryBlock())
760  BB->moveBefore(&BB->getParent()->getEntryBlock());
761 
762  DEBUG(dbgs() << "AFTER:\n" << *DestBB << "\n\n\n");
763  return;
764  }
765  }
766 
767  // Otherwise, we have multiple predecessors of BB. Update the PHIs in DestBB
768  // to handle the new incoming edges it is about to have.
769  PHINode *PN;
770  for (BasicBlock::iterator BBI = DestBB->begin();
771  (PN = dyn_cast<PHINode>(BBI)); ++BBI) {
772  // Remove the incoming value for BB, and remember it.
773  Value *InVal = PN->removeIncomingValue(BB, false);
774 
775  // Two options: either the InVal is a phi node defined in BB or it is some
776  // value that dominates BB.
777  PHINode *InValPhi = dyn_cast<PHINode>(InVal);
778  if (InValPhi && InValPhi->getParent() == BB) {
779  // Add all of the input values of the input PHI as inputs of this phi.
780  for (unsigned i = 0, e = InValPhi->getNumIncomingValues(); i != e; ++i)
781  PN->addIncoming(InValPhi->getIncomingValue(i),
782  InValPhi->getIncomingBlock(i));
783  } else {
784  // Otherwise, add one instance of the dominating value for each edge that
785  // we will be adding.
786  if (PHINode *BBPN = dyn_cast<PHINode>(BB->begin())) {
787  for (unsigned i = 0, e = BBPN->getNumIncomingValues(); i != e; ++i)
788  PN->addIncoming(InVal, BBPN->getIncomingBlock(i));
789  } else {
790  for (pred_iterator PI = pred_begin(BB), E = pred_end(BB); PI != E; ++PI)
791  PN->addIncoming(InVal, *PI);
792  }
793  }
794  }
795 
796  // The PHIs are now updated, change everything that refers to BB to use
797  // DestBB and remove BB.
798  BB->replaceAllUsesWith(DestBB);
799  BB->eraseFromParent();
800  ++NumBlocksElim;
801 
802  DEBUG(dbgs() << "AFTER:\n" << *DestBB << "\n\n\n");
803 }
804 
805 // Computes a map of base pointer relocation instructions to corresponding
806 // derived pointer relocation instructions given a vector of all relocate calls
808  const SmallVectorImpl<GCRelocateInst *> &AllRelocateCalls,
810  &RelocateInstMap) {
811  // Collect information in two maps: one primarily for locating the base object
812  // while filling the second map; the second map is the final structure holding
813  // a mapping between Base and corresponding Derived relocate calls
815  for (auto *ThisRelocate : AllRelocateCalls) {
816  auto K = std::make_pair(ThisRelocate->getBasePtrIndex(),
817  ThisRelocate->getDerivedPtrIndex());
818  RelocateIdxMap.insert(std::make_pair(K, ThisRelocate));
819  }
820  for (auto &Item : RelocateIdxMap) {
821  std::pair<unsigned, unsigned> Key = Item.first;
822  if (Key.first == Key.second)
823  // Base relocation: nothing to insert
824  continue;
825 
826  GCRelocateInst *I = Item.second;
827  auto BaseKey = std::make_pair(Key.first, Key.first);
828 
829  // We're iterating over RelocateIdxMap so we cannot modify it.
830  auto MaybeBase = RelocateIdxMap.find(BaseKey);
831  if (MaybeBase == RelocateIdxMap.end())
832  // TODO: We might want to insert a new base object relocate and gep off
833  // that, if there are enough derived object relocates.
834  continue;
835 
836  RelocateInstMap[MaybeBase->second].push_back(I);
837  }
838 }
839 
840 // Accepts a GEP and extracts the operands into a vector provided they're all
841 // small integer constants
843  SmallVectorImpl<Value *> &OffsetV) {
844  for (unsigned i = 1; i < GEP->getNumOperands(); i++) {
845  // Only accept small constant integer operands
846  auto Op = dyn_cast<ConstantInt>(GEP->getOperand(i));
847  if (!Op || Op->getZExtValue() > 20)
848  return false;
849  }
850 
851  for (unsigned i = 1; i < GEP->getNumOperands(); i++)
852  OffsetV.push_back(GEP->getOperand(i));
853  return true;
854 }
855 
856 // Takes a RelocatedBase (base pointer relocation instruction) and Targets to
857 // replace, computes a replacement, and affects it.
858 static bool
860  const SmallVectorImpl<GCRelocateInst *> &Targets) {
861  bool MadeChange = false;
862  // We must ensure the relocation of derived pointer is defined after
863  // relocation of base pointer. If we find a relocation corresponding to base
864  // defined earlier than relocation of base then we move relocation of base
865  // right before found relocation. We consider only relocation in the same
866  // basic block as relocation of base. Relocations from other basic block will
867  // be skipped by optimization and we do not care about them.
868  for (auto R = RelocatedBase->getParent()->getFirstInsertionPt();
869  &*R != RelocatedBase; ++R)
870  if (auto RI = dyn_cast<GCRelocateInst>(R))
871  if (RI->getStatepoint() == RelocatedBase->getStatepoint())
872  if (RI->getBasePtrIndex() == RelocatedBase->getBasePtrIndex()) {
873  RelocatedBase->moveBefore(RI);
874  break;
875  }
876 
877  for (GCRelocateInst *ToReplace : Targets) {
878  assert(ToReplace->getBasePtrIndex() == RelocatedBase->getBasePtrIndex() &&
879  "Not relocating a derived object of the original base object");
880  if (ToReplace->getBasePtrIndex() == ToReplace->getDerivedPtrIndex()) {
881  // A duplicate relocate call. TODO: coalesce duplicates.
882  continue;
883  }
884 
885  if (RelocatedBase->getParent() != ToReplace->getParent()) {
886  // Base and derived relocates are in different basic blocks.
887  // In this case transform is only valid when base dominates derived
888  // relocate. However it would be too expensive to check dominance
889  // for each such relocate, so we skip the whole transformation.
890  continue;
891  }
892 
893  Value *Base = ToReplace->getBasePtr();
894  auto Derived = dyn_cast<GetElementPtrInst>(ToReplace->getDerivedPtr());
895  if (!Derived || Derived->getPointerOperand() != Base)
896  continue;
897 
898  SmallVector<Value *, 2> OffsetV;
899  if (!getGEPSmallConstantIntOffsetV(Derived, OffsetV))
900  continue;
901 
902  // Create a Builder and replace the target callsite with a gep
903  assert(RelocatedBase->getNextNode() &&
904  "Should always have one since it's not a terminator");
905 
906  // Insert after RelocatedBase
907  IRBuilder<> Builder(RelocatedBase->getNextNode());
908  Builder.SetCurrentDebugLocation(ToReplace->getDebugLoc());
909 
910  // If gc_relocate does not match the actual type, cast it to the right type.
911  // In theory, there must be a bitcast after gc_relocate if the type does not
912  // match, and we should reuse it to get the derived pointer. But it could be
913  // cases like this:
914  // bb1:
915  // ...
916  // %g1 = call coldcc i8 addrspace(1)* @llvm.experimental.gc.relocate.p1i8(...)
917  // br label %merge
918  //
919  // bb2:
920  // ...
921  // %g2 = call coldcc i8 addrspace(1)* @llvm.experimental.gc.relocate.p1i8(...)
922  // br label %merge
923  //
924  // merge:
925  // %p1 = phi i8 addrspace(1)* [ %g1, %bb1 ], [ %g2, %bb2 ]
926  // %cast = bitcast i8 addrspace(1)* %p1 in to i32 addrspace(1)*
927  //
928  // In this case, we can not find the bitcast any more. So we insert a new bitcast
929  // no matter there is already one or not. In this way, we can handle all cases, and
930  // the extra bitcast should be optimized away in later passes.
931  Value *ActualRelocatedBase = RelocatedBase;
932  if (RelocatedBase->getType() != Base->getType()) {
933  ActualRelocatedBase =
934  Builder.CreateBitCast(RelocatedBase, Base->getType());
935  }
936  Value *Replacement = Builder.CreateGEP(
937  Derived->getSourceElementType(), ActualRelocatedBase, makeArrayRef(OffsetV));
938  Replacement->takeName(ToReplace);
939  // If the newly generated derived pointer's type does not match the original derived
940  // pointer's type, cast the new derived pointer to match it. Same reasoning as above.
941  Value *ActualReplacement = Replacement;
942  if (Replacement->getType() != ToReplace->getType()) {
943  ActualReplacement =
944  Builder.CreateBitCast(Replacement, ToReplace->getType());
945  }
946  ToReplace->replaceAllUsesWith(ActualReplacement);
947  ToReplace->eraseFromParent();
948 
949  MadeChange = true;
950  }
951  return MadeChange;
952 }
953 
954 // Turns this:
955 //
956 // %base = ...
957 // %ptr = gep %base + 15
958 // %tok = statepoint (%fun, i32 0, i32 0, i32 0, %base, %ptr)
959 // %base' = relocate(%tok, i32 4, i32 4)
960 // %ptr' = relocate(%tok, i32 4, i32 5)
961 // %val = load %ptr'
962 //
963 // into this:
964 //
965 // %base = ...
966 // %ptr = gep %base + 15
967 // %tok = statepoint (%fun, i32 0, i32 0, i32 0, %base, %ptr)
968 // %base' = gc.relocate(%tok, i32 4, i32 4)
969 // %ptr' = gep %base' + 15
970 // %val = load %ptr'
971 bool CodeGenPrepare::simplifyOffsetableRelocate(Instruction &I) {
972  bool MadeChange = false;
973  SmallVector<GCRelocateInst *, 2> AllRelocateCalls;
974 
975  for (auto *U : I.users())
976  if (GCRelocateInst *Relocate = dyn_cast<GCRelocateInst>(U))
977  // Collect all the relocate calls associated with a statepoint
978  AllRelocateCalls.push_back(Relocate);
979 
980  // We need atleast one base pointer relocation + one derived pointer
981  // relocation to mangle
982  if (AllRelocateCalls.size() < 2)
983  return false;
984 
985  // RelocateInstMap is a mapping from the base relocate instruction to the
986  // corresponding derived relocate instructions
988  computeBaseDerivedRelocateMap(AllRelocateCalls, RelocateInstMap);
989  if (RelocateInstMap.empty())
990  return false;
991 
992  for (auto &Item : RelocateInstMap)
993  // Item.first is the RelocatedBase to offset against
994  // Item.second is the vector of Targets to replace
995  MadeChange = simplifyRelocatesOffABase(Item.first, Item.second);
996  return MadeChange;
997 }
998 
999 /// SinkCast - Sink the specified cast instruction into its user blocks
1000 static bool SinkCast(CastInst *CI) {
1001  BasicBlock *DefBB = CI->getParent();
1002 
1003  /// InsertedCasts - Only insert a cast in each block once.
1004  DenseMap<BasicBlock*, CastInst*> InsertedCasts;
1005 
1006  bool MadeChange = false;
1007  for (Value::user_iterator UI = CI->user_begin(), E = CI->user_end();
1008  UI != E; ) {
1009  Use &TheUse = UI.getUse();
1010  Instruction *User = cast<Instruction>(*UI);
1011 
1012  // Figure out which BB this cast is used in. For PHI's this is the
1013  // appropriate predecessor block.
1014  BasicBlock *UserBB = User->getParent();
1015  if (PHINode *PN = dyn_cast<PHINode>(User)) {
1016  UserBB = PN->getIncomingBlock(TheUse);
1017  }
1018 
1019  // Preincrement use iterator so we don't invalidate it.
1020  ++UI;
1021 
1022  // The first insertion point of a block containing an EH pad is after the
1023  // pad. If the pad is the user, we cannot sink the cast past the pad.
1024  if (User->isEHPad())
1025  continue;
1026 
1027  // If the block selected to receive the cast is an EH pad that does not
1028  // allow non-PHI instructions before the terminator, we can't sink the
1029  // cast.
1030  if (UserBB->getTerminator()->isEHPad())
1031  continue;
1032 
1033  // If this user is in the same block as the cast, don't change the cast.
1034  if (UserBB == DefBB) continue;
1035 
1036  // If we have already inserted a cast into this block, use it.
1037  CastInst *&InsertedCast = InsertedCasts[UserBB];
1038 
1039  if (!InsertedCast) {
1040  BasicBlock::iterator InsertPt = UserBB->getFirstInsertionPt();
1041  assert(InsertPt != UserBB->end());
1042  InsertedCast = CastInst::Create(CI->getOpcode(), CI->getOperand(0),
1043  CI->getType(), "", &*InsertPt);
1044  }
1045 
1046  // Replace a use of the cast with a use of the new cast.
1047  TheUse = InsertedCast;
1048  MadeChange = true;
1049  ++NumCastUses;
1050  }
1051 
1052  // If we removed all uses, nuke the cast.
1053  if (CI->use_empty()) {
1054  salvageDebugInfo(*CI);
1055  CI->eraseFromParent();
1056  MadeChange = true;
1057  }
1058 
1059  return MadeChange;
1060 }
1061 
1062 /// If the specified cast instruction is a noop copy (e.g. it's casting from
1063 /// one pointer type to another, i32->i8 on PPC), sink it into user blocks to
1064 /// reduce the number of virtual registers that must be created and coalesced.
1065 ///
1066 /// Return true if any changes are made.
1068  const DataLayout &DL) {
1069  // Sink only "cheap" (or nop) address-space casts. This is a weaker condition
1070  // than sinking only nop casts, but is helpful on some platforms.
1071  if (auto *ASC = dyn_cast<AddrSpaceCastInst>(CI)) {
1072  if (!TLI.isCheapAddrSpaceCast(ASC->getSrcAddressSpace(),
1073  ASC->getDestAddressSpace()))
1074  return false;
1075  }
1076 
1077  // If this is a noop copy,
1078  EVT SrcVT = TLI.getValueType(DL, CI->getOperand(0)->getType());
1079  EVT DstVT = TLI.getValueType(DL, CI->getType());
1080 
1081  // This is an fp<->int conversion?
1082  if (SrcVT.isInteger() != DstVT.isInteger())
1083  return false;
1084 
1085  // If this is an extension, it will be a zero or sign extension, which
1086  // isn't a noop.
1087  if (SrcVT.bitsLT(DstVT)) return false;
1088 
1089  // If these values will be promoted, find out what they will be promoted
1090  // to. This helps us consider truncates on PPC as noop copies when they
1091  // are.
1092  if (TLI.getTypeAction(CI->getContext(), SrcVT) ==
1094  SrcVT = TLI.getTypeToTransformTo(CI->getContext(), SrcVT);
1095  if (TLI.getTypeAction(CI->getContext(), DstVT) ==
1097  DstVT = TLI.getTypeToTransformTo(CI->getContext(), DstVT);
1098 
1099  // If, after promotion, these are the same types, this is a noop copy.
1100  if (SrcVT != DstVT)
1101  return false;
1102 
1103  return SinkCast(CI);
1104 }
1105 
1106 /// Try to combine CI into a call to the llvm.uadd.with.overflow intrinsic if
1107 /// possible.
1108 ///
1109 /// Return true if any changes were made.
1111  Value *A, *B;
1112  Instruction *AddI;
1113  if (!match(CI,
1115  return false;
1116 
1117  Type *Ty = AddI->getType();
1118  if (!isa<IntegerType>(Ty))
1119  return false;
1120 
1121  // We don't want to move around uses of condition values this late, so we we
1122  // check if it is legal to create the call to the intrinsic in the basic
1123  // block containing the icmp:
1124 
1125  if (AddI->getParent() != CI->getParent() && !AddI->hasOneUse())
1126  return false;
1127 
1128 #ifndef NDEBUG
1129  // Someday m_UAddWithOverflow may get smarter, but this is a safe assumption
1130  // for now:
1131  if (AddI->hasOneUse())
1132  assert(*AddI->user_begin() == CI && "expected!");
1133 #endif
1134 
1135  Module *M = CI->getModule();
1136  Value *F = Intrinsic::getDeclaration(M, Intrinsic::uadd_with_overflow, Ty);
1137 
1138  auto *InsertPt = AddI->hasOneUse() ? CI : AddI;
1139 
1140  auto *UAddWithOverflow =
1141  CallInst::Create(F, {A, B}, "uadd.overflow", InsertPt);
1142  auto *UAdd = ExtractValueInst::Create(UAddWithOverflow, 0, "uadd", InsertPt);
1143  auto *Overflow =
1144  ExtractValueInst::Create(UAddWithOverflow, 1, "overflow", InsertPt);
1145 
1146  CI->replaceAllUsesWith(Overflow);
1147  AddI->replaceAllUsesWith(UAdd);
1148  CI->eraseFromParent();
1149  AddI->eraseFromParent();
1150  return true;
1151 }
1152 
1153 /// Sink the given CmpInst into user blocks to reduce the number of virtual
1154 /// registers that must be created and coalesced. This is a clear win except on
1155 /// targets with multiple condition code registers (PowerPC), where it might
1156 /// lose; some adjustment may be wanted there.
1157 ///
1158 /// Return true if any changes are made.
1159 static bool SinkCmpExpression(CmpInst *CI, const TargetLowering *TLI) {
1160  BasicBlock *DefBB = CI->getParent();
1161 
1162  // Avoid sinking soft-FP comparisons, since this can move them into a loop.
1163  if (TLI && TLI->useSoftFloat() && isa<FCmpInst>(CI))
1164  return false;
1165 
1166  // Only insert a cmp in each block once.
1167  DenseMap<BasicBlock*, CmpInst*> InsertedCmps;
1168 
1169  bool MadeChange = false;
1170  for (Value::user_iterator UI = CI->user_begin(), E = CI->user_end();
1171  UI != E; ) {
1172  Use &TheUse = UI.getUse();
1173  Instruction *User = cast<Instruction>(*UI);
1174 
1175  // Preincrement use iterator so we don't invalidate it.
1176  ++UI;
1177 
1178  // Don't bother for PHI nodes.
1179  if (isa<PHINode>(User))
1180  continue;
1181 
1182  // Figure out which BB this cmp is used in.
1183  BasicBlock *UserBB = User->getParent();
1184 
1185  // If this user is in the same block as the cmp, don't change the cmp.
1186  if (UserBB == DefBB) continue;
1187 
1188  // If we have already inserted a cmp into this block, use it.
1189  CmpInst *&InsertedCmp = InsertedCmps[UserBB];
1190 
1191  if (!InsertedCmp) {
1192  BasicBlock::iterator InsertPt = UserBB->getFirstInsertionPt();
1193  assert(InsertPt != UserBB->end());
1194  InsertedCmp =
1195  CmpInst::Create(CI->getOpcode(), CI->getPredicate(),
1196  CI->getOperand(0), CI->getOperand(1), "", &*InsertPt);
1197  // Propagate the debug info.
1198  InsertedCmp->setDebugLoc(CI->getDebugLoc());
1199  }
1200 
1201  // Replace a use of the cmp with a use of the new cmp.
1202  TheUse = InsertedCmp;
1203  MadeChange = true;
1204  ++NumCmpUses;
1205  }
1206 
1207  // If we removed all uses, nuke the cmp.
1208  if (CI->use_empty()) {
1209  CI->eraseFromParent();
1210  MadeChange = true;
1211  }
1212 
1213  return MadeChange;
1214 }
1215 
1216 static bool OptimizeCmpExpression(CmpInst *CI, const TargetLowering *TLI) {
1217  if (SinkCmpExpression(CI, TLI))
1218  return true;
1219 
1220  if (CombineUAddWithOverflow(CI))
1221  return true;
1222 
1223  return false;
1224 }
1225 
1226 /// Duplicate and sink the given 'and' instruction into user blocks where it is
1227 /// used in a compare to allow isel to generate better code for targets where
1228 /// this operation can be combined.
1229 ///
1230 /// Return true if any changes are made.
1232  const TargetLowering &TLI,
1233  SetOfInstrs &InsertedInsts) {
1234  // Double-check that we're not trying to optimize an instruction that was
1235  // already optimized by some other part of this pass.
1236  assert(!InsertedInsts.count(AndI) &&
1237  "Attempting to optimize already optimized and instruction");
1238  (void) InsertedInsts;
1239 
1240  // Nothing to do for single use in same basic block.
1241  if (AndI->hasOneUse() &&
1242  AndI->getParent() == cast<Instruction>(*AndI->user_begin())->getParent())
1243  return false;
1244 
1245  // Try to avoid cases where sinking/duplicating is likely to increase register
1246  // pressure.
1247  if (!isa<ConstantInt>(AndI->getOperand(0)) &&
1248  !isa<ConstantInt>(AndI->getOperand(1)) &&
1249  AndI->getOperand(0)->hasOneUse() && AndI->getOperand(1)->hasOneUse())
1250  return false;
1251 
1252  for (auto *U : AndI->users()) {
1253  Instruction *User = cast<Instruction>(U);
1254 
1255  // Only sink for and mask feeding icmp with 0.
1256  if (!isa<ICmpInst>(User))
1257  return false;
1258 
1259  auto *CmpC = dyn_cast<ConstantInt>(User->getOperand(1));
1260  if (!CmpC || !CmpC->isZero())
1261  return false;
1262  }
1263 
1264  if (!TLI.isMaskAndCmp0FoldingBeneficial(*AndI))
1265  return false;
1266 
1267  DEBUG(dbgs() << "found 'and' feeding only icmp 0;\n");
1268  DEBUG(AndI->getParent()->dump());
1269 
1270  // Push the 'and' into the same block as the icmp 0. There should only be
1271  // one (icmp (and, 0)) in each block, since CSE/GVN should have removed any
1272  // others, so we don't need to keep track of which BBs we insert into.
1273  for (Value::user_iterator UI = AndI->user_begin(), E = AndI->user_end();
1274  UI != E; ) {
1275  Use &TheUse = UI.getUse();
1276  Instruction *User = cast<Instruction>(*UI);
1277 
1278  // Preincrement use iterator so we don't invalidate it.
1279  ++UI;
1280 
1281  DEBUG(dbgs() << "sinking 'and' use: " << *User << "\n");
1282 
1283  // Keep the 'and' in the same place if the use is already in the same block.
1284  Instruction *InsertPt =
1285  User->getParent() == AndI->getParent() ? AndI : User;
1286  Instruction *InsertedAnd =
1287  BinaryOperator::Create(Instruction::And, AndI->getOperand(0),
1288  AndI->getOperand(1), "", InsertPt);
1289  // Propagate the debug info.
1290  InsertedAnd->setDebugLoc(AndI->getDebugLoc());
1291 
1292  // Replace a use of the 'and' with a use of the new 'and'.
1293  TheUse = InsertedAnd;
1294  ++NumAndUses;
1295  DEBUG(User->getParent()->dump());
1296  }
1297 
1298  // We removed all uses, nuke the and.
1299  AndI->eraseFromParent();
1300  return true;
1301 }
1302 
1303 /// Check if the candidates could be combined with a shift instruction, which
1304 /// includes:
1305 /// 1. Truncate instruction
1306 /// 2. And instruction and the imm is a mask of the low bits:
1307 /// imm & (imm+1) == 0
1309  if (!isa<TruncInst>(User)) {
1310  if (User->getOpcode() != Instruction::And ||
1311  !isa<ConstantInt>(User->getOperand(1)))
1312  return false;
1313 
1314  const APInt &Cimm = cast<ConstantInt>(User->getOperand(1))->getValue();
1315 
1316  if ((Cimm & (Cimm + 1)).getBoolValue())
1317  return false;
1318  }
1319  return true;
1320 }
1321 
1322 /// Sink both shift and truncate instruction to the use of truncate's BB.
1323 static bool
1326  const TargetLowering &TLI, const DataLayout &DL) {
1327  BasicBlock *UserBB = User->getParent();
1328  DenseMap<BasicBlock *, CastInst *> InsertedTruncs;
1329  TruncInst *TruncI = dyn_cast<TruncInst>(User);
1330  bool MadeChange = false;
1331 
1332  for (Value::user_iterator TruncUI = TruncI->user_begin(),
1333  TruncE = TruncI->user_end();
1334  TruncUI != TruncE;) {
1335 
1336  Use &TruncTheUse = TruncUI.getUse();
1337  Instruction *TruncUser = cast<Instruction>(*TruncUI);
1338  // Preincrement use iterator so we don't invalidate it.
1339 
1340  ++TruncUI;
1341 
1342  int ISDOpcode = TLI.InstructionOpcodeToISD(TruncUser->getOpcode());
1343  if (!ISDOpcode)
1344  continue;
1345 
1346  // If the use is actually a legal node, there will not be an
1347  // implicit truncate.
1348  // FIXME: always querying the result type is just an
1349  // approximation; some nodes' legality is determined by the
1350  // operand or other means. There's no good way to find out though.
1351  if (TLI.isOperationLegalOrCustom(
1352  ISDOpcode, TLI.getValueType(DL, TruncUser->getType(), true)))
1353  continue;
1354 
1355  // Don't bother for PHI nodes.
1356  if (isa<PHINode>(TruncUser))
1357  continue;
1358 
1359  BasicBlock *TruncUserBB = TruncUser->getParent();
1360 
1361  if (UserBB == TruncUserBB)
1362  continue;
1363 
1364  BinaryOperator *&InsertedShift = InsertedShifts[TruncUserBB];
1365  CastInst *&InsertedTrunc = InsertedTruncs[TruncUserBB];
1366 
1367  if (!InsertedShift && !InsertedTrunc) {
1368  BasicBlock::iterator InsertPt = TruncUserBB->getFirstInsertionPt();
1369  assert(InsertPt != TruncUserBB->end());
1370  // Sink the shift
1371  if (ShiftI->getOpcode() == Instruction::AShr)
1372  InsertedShift = BinaryOperator::CreateAShr(ShiftI->getOperand(0), CI,
1373  "", &*InsertPt);
1374  else
1375  InsertedShift = BinaryOperator::CreateLShr(ShiftI->getOperand(0), CI,
1376  "", &*InsertPt);
1377 
1378  // Sink the trunc
1379  BasicBlock::iterator TruncInsertPt = TruncUserBB->getFirstInsertionPt();
1380  TruncInsertPt++;
1381  assert(TruncInsertPt != TruncUserBB->end());
1382 
1383  InsertedTrunc = CastInst::Create(TruncI->getOpcode(), InsertedShift,
1384  TruncI->getType(), "", &*TruncInsertPt);
1385 
1386  MadeChange = true;
1387 
1388  TruncTheUse = InsertedTrunc;
1389  }
1390  }
1391  return MadeChange;
1392 }
1393 
1394 /// Sink the shift *right* instruction into user blocks if the uses could
1395 /// potentially be combined with this shift instruction and generate BitExtract
1396 /// instruction. It will only be applied if the architecture supports BitExtract
1397 /// instruction. Here is an example:
1398 /// BB1:
1399 /// %x.extract.shift = lshr i64 %arg1, 32
1400 /// BB2:
1401 /// %x.extract.trunc = trunc i64 %x.extract.shift to i16
1402 /// ==>
1403 ///
1404 /// BB2:
1405 /// %x.extract.shift.1 = lshr i64 %arg1, 32
1406 /// %x.extract.trunc = trunc i64 %x.extract.shift.1 to i16
1407 ///
1408 /// CodeGen will recoginze the pattern in BB2 and generate BitExtract
1409 /// instruction.
1410 /// Return true if any changes are made.
1412  const TargetLowering &TLI,
1413  const DataLayout &DL) {
1414  BasicBlock *DefBB = ShiftI->getParent();
1415 
1416  /// Only insert instructions in each block once.
1418 
1419  bool shiftIsLegal = TLI.isTypeLegal(TLI.getValueType(DL, ShiftI->getType()));
1420 
1421  bool MadeChange = false;
1422  for (Value::user_iterator UI = ShiftI->user_begin(), E = ShiftI->user_end();
1423  UI != E;) {
1424  Use &TheUse = UI.getUse();
1425  Instruction *User = cast<Instruction>(*UI);
1426  // Preincrement use iterator so we don't invalidate it.
1427  ++UI;
1428 
1429  // Don't bother for PHI nodes.
1430  if (isa<PHINode>(User))
1431  continue;
1432 
1433  if (!isExtractBitsCandidateUse(User))
1434  continue;
1435 
1436  BasicBlock *UserBB = User->getParent();
1437 
1438  if (UserBB == DefBB) {
1439  // If the shift and truncate instruction are in the same BB. The use of
1440  // the truncate(TruncUse) may still introduce another truncate if not
1441  // legal. In this case, we would like to sink both shift and truncate
1442  // instruction to the BB of TruncUse.
1443  // for example:
1444  // BB1:
1445  // i64 shift.result = lshr i64 opnd, imm
1446  // trunc.result = trunc shift.result to i16
1447  //
1448  // BB2:
1449  // ----> We will have an implicit truncate here if the architecture does
1450  // not have i16 compare.
1451  // cmp i16 trunc.result, opnd2
1452  //
1453  if (isa<TruncInst>(User) && shiftIsLegal
1454  // If the type of the truncate is legal, no trucate will be
1455  // introduced in other basic blocks.
1456  &&
1457  (!TLI.isTypeLegal(TLI.getValueType(DL, User->getType()))))
1458  MadeChange =
1459  SinkShiftAndTruncate(ShiftI, User, CI, InsertedShifts, TLI, DL);
1460 
1461  continue;
1462  }
1463  // If we have already inserted a shift into this block, use it.
1464  BinaryOperator *&InsertedShift = InsertedShifts[UserBB];
1465 
1466  if (!InsertedShift) {
1467  BasicBlock::iterator InsertPt = UserBB->getFirstInsertionPt();
1468  assert(InsertPt != UserBB->end());
1469 
1470  if (ShiftI->getOpcode() == Instruction::AShr)
1471  InsertedShift = BinaryOperator::CreateAShr(ShiftI->getOperand(0), CI,
1472  "", &*InsertPt);
1473  else
1474  InsertedShift = BinaryOperator::CreateLShr(ShiftI->getOperand(0), CI,
1475  "", &*InsertPt);
1476 
1477  MadeChange = true;
1478  }
1479 
1480  // Replace a use of the shift with a use of the new shift.
1481  TheUse = InsertedShift;
1482  }
1483 
1484  // If we removed all uses, nuke the shift.
1485  if (ShiftI->use_empty())
1486  ShiftI->eraseFromParent();
1487 
1488  return MadeChange;
1489 }
1490 
1491 /// If counting leading or trailing zeros is an expensive operation and a zero
1492 /// input is defined, add a check for zero to avoid calling the intrinsic.
1493 ///
1494 /// We want to transform:
1495 /// %z = call i64 @llvm.cttz.i64(i64 %A, i1 false)
1496 ///
1497 /// into:
1498 /// entry:
1499 /// %cmpz = icmp eq i64 %A, 0
1500 /// br i1 %cmpz, label %cond.end, label %cond.false
1501 /// cond.false:
1502 /// %z = call i64 @llvm.cttz.i64(i64 %A, i1 true)
1503 /// br label %cond.end
1504 /// cond.end:
1505 /// %ctz = phi i64 [ 64, %entry ], [ %z, %cond.false ]
1506 ///
1507 /// If the transform is performed, return true and set ModifiedDT to true.
1508 static bool despeculateCountZeros(IntrinsicInst *CountZeros,
1509  const TargetLowering *TLI,
1510  const DataLayout *DL,
1511  bool &ModifiedDT) {
1512  if (!TLI || !DL)
1513  return false;
1514 
1515  // If a zero input is undefined, it doesn't make sense to despeculate that.
1516  if (match(CountZeros->getOperand(1), m_One()))
1517  return false;
1518 
1519  // If it's cheap to speculate, there's nothing to do.
1520  auto IntrinsicID = CountZeros->getIntrinsicID();
1521  if ((IntrinsicID == Intrinsic::cttz && TLI->isCheapToSpeculateCttz()) ||
1522  (IntrinsicID == Intrinsic::ctlz && TLI->isCheapToSpeculateCtlz()))
1523  return false;
1524 
1525  // Only handle legal scalar cases. Anything else requires too much work.
1526  Type *Ty = CountZeros->getType();
1527  unsigned SizeInBits = Ty->getPrimitiveSizeInBits();
1528  if (Ty->isVectorTy() || SizeInBits > DL->getLargestLegalIntTypeSizeInBits())
1529  return false;
1530 
1531  // The intrinsic will be sunk behind a compare against zero and branch.
1532  BasicBlock *StartBlock = CountZeros->getParent();
1533  BasicBlock *CallBlock = StartBlock->splitBasicBlock(CountZeros, "cond.false");
1534 
1535  // Create another block after the count zero intrinsic. A PHI will be added
1536  // in this block to select the result of the intrinsic or the bit-width
1537  // constant if the input to the intrinsic is zero.
1538  BasicBlock::iterator SplitPt = ++(BasicBlock::iterator(CountZeros));
1539  BasicBlock *EndBlock = CallBlock->splitBasicBlock(SplitPt, "cond.end");
1540 
1541  // Set up a builder to create a compare, conditional branch, and PHI.
1542  IRBuilder<> Builder(CountZeros->getContext());
1543  Builder.SetInsertPoint(StartBlock->getTerminator());
1544  Builder.SetCurrentDebugLocation(CountZeros->getDebugLoc());
1545 
1546  // Replace the unconditional branch that was created by the first split with
1547  // a compare against zero and a conditional branch.
1548  Value *Zero = Constant::getNullValue(Ty);
1549  Value *Cmp = Builder.CreateICmpEQ(CountZeros->getOperand(0), Zero, "cmpz");
1550  Builder.CreateCondBr(Cmp, EndBlock, CallBlock);
1551  StartBlock->getTerminator()->eraseFromParent();
1552 
1553  // Create a PHI in the end block to select either the output of the intrinsic
1554  // or the bit width of the operand.
1555  Builder.SetInsertPoint(&EndBlock->front());
1556  PHINode *PN = Builder.CreatePHI(Ty, 2, "ctz");
1557  CountZeros->replaceAllUsesWith(PN);
1558  Value *BitWidth = Builder.getInt(APInt(SizeInBits, SizeInBits));
1559  PN->addIncoming(BitWidth, StartBlock);
1560  PN->addIncoming(CountZeros, CallBlock);
1561 
1562  // We are explicitly handling the zero case, so we can set the intrinsic's
1563  // undefined zero argument to 'true'. This will also prevent reprocessing the
1564  // intrinsic; we only despeculate when a zero input is defined.
1565  CountZeros->setArgOperand(1, Builder.getTrue());
1566  ModifiedDT = true;
1567  return true;
1568 }
1569 
1570 bool CodeGenPrepare::optimizeCallInst(CallInst *CI, bool &ModifiedDT) {
1571  BasicBlock *BB = CI->getParent();
1572 
1573  // Lower inline assembly if we can.
1574  // If we found an inline asm expession, and if the target knows how to
1575  // lower it to normal LLVM code, do so now.
1576  if (TLI && isa<InlineAsm>(CI->getCalledValue())) {
1577  if (TLI->ExpandInlineAsm(CI)) {
1578  // Avoid invalidating the iterator.
1579  CurInstIterator = BB->begin();
1580  // Avoid processing instructions out of order, which could cause
1581  // reuse before a value is defined.
1582  SunkAddrs.clear();
1583  return true;
1584  }
1585  // Sink address computing for memory operands into the block.
1586  if (optimizeInlineAsmInst(CI))
1587  return true;
1588  }
1589 
1590  // Align the pointer arguments to this call if the target thinks it's a good
1591  // idea
1592  unsigned MinSize, PrefAlign;
1593  if (TLI && TLI->shouldAlignPointerArgs(CI, MinSize, PrefAlign)) {
1594  for (auto &Arg : CI->arg_operands()) {
1595  // We want to align both objects whose address is used directly and
1596  // objects whose address is used in casts and GEPs, though it only makes
1597  // sense for GEPs if the offset is a multiple of the desired alignment and
1598  // if size - offset meets the size threshold.
1599  if (!Arg->getType()->isPointerTy())
1600  continue;
1601  APInt Offset(DL->getPointerSizeInBits(
1602  cast<PointerType>(Arg->getType())->getAddressSpace()),
1603  0);
1605  uint64_t Offset2 = Offset.getLimitedValue();
1606  if ((Offset2 & (PrefAlign-1)) != 0)
1607  continue;
1608  AllocaInst *AI;
1609  if ((AI = dyn_cast<AllocaInst>(Val)) && AI->getAlignment() < PrefAlign &&
1610  DL->getTypeAllocSize(AI->getAllocatedType()) >= MinSize + Offset2)
1611  AI->setAlignment(PrefAlign);
1612  // Global variables can only be aligned if they are defined in this
1613  // object (i.e. they are uniquely initialized in this object), and
1614  // over-aligning global variables that have an explicit section is
1615  // forbidden.
1616  GlobalVariable *GV;
1617  if ((GV = dyn_cast<GlobalVariable>(Val)) && GV->canIncreaseAlignment() &&
1618  GV->getPointerAlignment(*DL) < PrefAlign &&
1619  DL->getTypeAllocSize(GV->getValueType()) >=
1620  MinSize + Offset2)
1621  GV->setAlignment(PrefAlign);
1622  }
1623  // If this is a memcpy (or similar) then we may be able to improve the
1624  // alignment
1625  if (MemIntrinsic *MI = dyn_cast<MemIntrinsic>(CI)) {
1626  unsigned Align = getKnownAlignment(MI->getDest(), *DL);
1627  if (MemTransferInst *MTI = dyn_cast<MemTransferInst>(MI))
1628  Align = std::min(Align, getKnownAlignment(MTI->getSource(), *DL));
1629  if (Align > MI->getAlignment())
1630  MI->setAlignment(ConstantInt::get(MI->getAlignmentType(), Align));
1631  }
1632  }
1633 
1634  // If we have a cold call site, try to sink addressing computation into the
1635  // cold block. This interacts with our handling for loads and stores to
1636  // ensure that we can fold all uses of a potential addressing computation
1637  // into their uses. TODO: generalize this to work over profiling data
1638  if (!OptSize && CI->hasFnAttr(Attribute::Cold))
1639  for (auto &Arg : CI->arg_operands()) {
1640  if (!Arg->getType()->isPointerTy())
1641  continue;
1642  unsigned AS = Arg->getType()->getPointerAddressSpace();
1643  return optimizeMemoryInst(CI, Arg, Arg->getType(), AS);
1644  }
1645 
1647  if (II) {
1648  switch (II->getIntrinsicID()) {
1649  default: break;
1650  case Intrinsic::objectsize: {
1651  // Lower all uses of llvm.objectsize.*
1652  ConstantInt *RetVal =
1653  lowerObjectSizeCall(II, *DL, TLInfo, /*MustSucceed=*/true);
1654  // Substituting this can cause recursive simplifications, which can
1655  // invalidate our iterator. Use a WeakTrackingVH to hold onto it in case
1656  // this
1657  // happens.
1658  Value *CurValue = &*CurInstIterator;
1659  WeakTrackingVH IterHandle(CurValue);
1660 
1661  replaceAndRecursivelySimplify(CI, RetVal, TLInfo, nullptr);
1662 
1663  // If the iterator instruction was recursively deleted, start over at the
1664  // start of the block.
1665  if (IterHandle != CurValue) {
1666  CurInstIterator = BB->begin();
1667  SunkAddrs.clear();
1668  }
1669  return true;
1670  }
1671  case Intrinsic::aarch64_stlxr:
1672  case Intrinsic::aarch64_stxr: {
1673  ZExtInst *ExtVal = dyn_cast<ZExtInst>(CI->getArgOperand(0));
1674  if (!ExtVal || !ExtVal->hasOneUse() ||
1675  ExtVal->getParent() == CI->getParent())
1676  return false;
1677  // Sink a zext feeding stlxr/stxr before it, so it can be folded into it.
1678  ExtVal->moveBefore(CI);
1679  // Mark this instruction as "inserted by CGP", so that other
1680  // optimizations don't touch it.
1681  InsertedInsts.insert(ExtVal);
1682  return true;
1683  }
1684  case Intrinsic::invariant_group_barrier:
1685  II->replaceAllUsesWith(II->getArgOperand(0));
1686  II->eraseFromParent();
1687  return true;
1688 
1689  case Intrinsic::cttz:
1690  case Intrinsic::ctlz:
1691  // If counting zeros is expensive, try to avoid it.
1692  return despeculateCountZeros(II, TLI, DL, ModifiedDT);
1693  }
1694 
1695  if (TLI) {
1696  SmallVector<Value*, 2> PtrOps;
1697  Type *AccessTy;
1698  if (TLI->getAddrModeArguments(II, PtrOps, AccessTy))
1699  while (!PtrOps.empty()) {
1700  Value *PtrVal = PtrOps.pop_back_val();
1701  unsigned AS = PtrVal->getType()->getPointerAddressSpace();
1702  if (optimizeMemoryInst(II, PtrVal, AccessTy, AS))
1703  return true;
1704  }
1705  }
1706  }
1707 
1708  // From here on out we're working with named functions.
1709  if (!CI->getCalledFunction()) return false;
1710 
1711  // Lower all default uses of _chk calls. This is very similar
1712  // to what InstCombineCalls does, but here we are only lowering calls
1713  // to fortified library functions (e.g. __memcpy_chk) that have the default
1714  // "don't know" as the objectsize. Anything else should be left alone.
1715  FortifiedLibCallSimplifier Simplifier(TLInfo, true);
1716  if (Value *V = Simplifier.optimizeCall(CI)) {
1717  CI->replaceAllUsesWith(V);
1718  CI->eraseFromParent();
1719  return true;
1720  }
1721 
1722  return false;
1723 }
1724 
1725 /// Look for opportunities to duplicate return instructions to the predecessor
1726 /// to enable tail call optimizations. The case it is currently looking for is:
1727 /// @code
1728 /// bb0:
1729 /// %tmp0 = tail call i32 @f0()
1730 /// br label %return
1731 /// bb1:
1732 /// %tmp1 = tail call i32 @f1()
1733 /// br label %return
1734 /// bb2:
1735 /// %tmp2 = tail call i32 @f2()
1736 /// br label %return
1737 /// return:
1738 /// %retval = phi i32 [ %tmp0, %bb0 ], [ %tmp1, %bb1 ], [ %tmp2, %bb2 ]
1739 /// ret i32 %retval
1740 /// @endcode
1741 ///
1742 /// =>
1743 ///
1744 /// @code
1745 /// bb0:
1746 /// %tmp0 = tail call i32 @f0()
1747 /// ret i32 %tmp0
1748 /// bb1:
1749 /// %tmp1 = tail call i32 @f1()
1750 /// ret i32 %tmp1
1751 /// bb2:
1752 /// %tmp2 = tail call i32 @f2()
1753 /// ret i32 %tmp2
1754 /// @endcode
1755 bool CodeGenPrepare::dupRetToEnableTailCallOpts(BasicBlock *BB) {
1756  if (!TLI)
1757  return false;
1758 
1759  ReturnInst *RetI = dyn_cast<ReturnInst>(BB->getTerminator());
1760  if (!RetI)
1761  return false;
1762 
1763  PHINode *PN = nullptr;
1764  BitCastInst *BCI = nullptr;
1765  Value *V = RetI->getReturnValue();
1766  if (V) {
1767  BCI = dyn_cast<BitCastInst>(V);
1768  if (BCI)
1769  V = BCI->getOperand(0);
1770 
1771  PN = dyn_cast<PHINode>(V);
1772  if (!PN)
1773  return false;
1774  }
1775 
1776  if (PN && PN->getParent() != BB)
1777  return false;
1778 
1779  // Make sure there are no instructions between the PHI and return, or that the
1780  // return is the first instruction in the block.
1781  if (PN) {
1782  BasicBlock::iterator BI = BB->begin();
1783  do { ++BI; } while (isa<DbgInfoIntrinsic>(BI));
1784  if (&*BI == BCI)
1785  // Also skip over the bitcast.
1786  ++BI;
1787  if (&*BI != RetI)
1788  return false;
1789  } else {
1790  BasicBlock::iterator BI = BB->begin();
1791  while (isa<DbgInfoIntrinsic>(BI)) ++BI;
1792  if (&*BI != RetI)
1793  return false;
1794  }
1795 
1796  /// Only dup the ReturnInst if the CallInst is likely to be emitted as a tail
1797  /// call.
1798  const Function *F = BB->getParent();
1799  SmallVector<CallInst*, 4> TailCalls;
1800  if (PN) {
1801  for (unsigned I = 0, E = PN->getNumIncomingValues(); I != E; ++I) {
1802  CallInst *CI = dyn_cast<CallInst>(PN->getIncomingValue(I));
1803  // Make sure the phi value is indeed produced by the tail call.
1804  if (CI && CI->hasOneUse() && CI->getParent() == PN->getIncomingBlock(I) &&
1805  TLI->mayBeEmittedAsTailCall(CI) &&
1806  attributesPermitTailCall(F, CI, RetI, *TLI))
1807  TailCalls.push_back(CI);
1808  }
1809  } else {
1810  SmallPtrSet<BasicBlock*, 4> VisitedBBs;
1811  for (pred_iterator PI = pred_begin(BB), PE = pred_end(BB); PI != PE; ++PI) {
1812  if (!VisitedBBs.insert(*PI).second)
1813  continue;
1814 
1815  BasicBlock::InstListType &InstList = (*PI)->getInstList();
1816  BasicBlock::InstListType::reverse_iterator RI = InstList.rbegin();
1817  BasicBlock::InstListType::reverse_iterator RE = InstList.rend();
1818  do { ++RI; } while (RI != RE && isa<DbgInfoIntrinsic>(&*RI));
1819  if (RI == RE)
1820  continue;
1821 
1822  CallInst *CI = dyn_cast<CallInst>(&*RI);
1823  if (CI && CI->use_empty() && TLI->mayBeEmittedAsTailCall(CI) &&
1824  attributesPermitTailCall(F, CI, RetI, *TLI))
1825  TailCalls.push_back(CI);
1826  }
1827  }
1828 
1829  bool Changed = false;
1830  for (unsigned i = 0, e = TailCalls.size(); i != e; ++i) {
1831  CallInst *CI = TailCalls[i];
1832  CallSite CS(CI);
1833 
1834  // Conservatively require the attributes of the call to match those of the
1835  // return. Ignore noalias because it doesn't affect the call sequence.
1836  AttributeList CalleeAttrs = CS.getAttributes();
1837  if (AttrBuilder(CalleeAttrs, AttributeList::ReturnIndex)
1838  .removeAttribute(Attribute::NoAlias) !=
1840  .removeAttribute(Attribute::NoAlias))
1841  continue;
1842 
1843  // Make sure the call instruction is followed by an unconditional branch to
1844  // the return block.
1845  BasicBlock *CallBB = CI->getParent();
1846  BranchInst *BI = dyn_cast<BranchInst>(CallBB->getTerminator());
1847  if (!BI || !BI->isUnconditional() || BI->getSuccessor(0) != BB)
1848  continue;
1849 
1850  // Duplicate the return into CallBB.
1851  (void)FoldReturnIntoUncondBranch(RetI, BB, CallBB);
1852  ModifiedDT = Changed = true;
1853  ++NumRetsDup;
1854  }
1855 
1856  // If we eliminated all predecessors of the block, delete the block now.
1857  if (Changed && !BB->hasAddressTaken() && pred_begin(BB) == pred_end(BB))
1858  BB->eraseFromParent();
1859 
1860  return Changed;
1861 }
1862 
1863 //===----------------------------------------------------------------------===//
1864 // Memory Optimization
1865 //===----------------------------------------------------------------------===//
1866 
1867 namespace {
1868 
1869 /// This is an extended version of TargetLowering::AddrMode
1870 /// which holds actual Value*'s for register values.
1871 struct ExtAddrMode : public TargetLowering::AddrMode {
1872  Value *BaseReg = nullptr;
1873  Value *ScaledReg = nullptr;
1874  Value *OriginalValue = nullptr;
1875 
1876  enum FieldName {
1877  NoField = 0x00,
1878  BaseRegField = 0x01,
1879  BaseGVField = 0x02,
1880  BaseOffsField = 0x04,
1881  ScaledRegField = 0x08,
1882  ScaleField = 0x10,
1883  MultipleFields = 0xff
1884  };
1885 
1886  ExtAddrMode() = default;
1887 
1888  void print(raw_ostream &OS) const;
1889  void dump() const;
1890 
1891  FieldName compare(const ExtAddrMode &other) {
1892  // First check that the types are the same on each field, as differing types
1893  // is something we can't cope with later on.
1894  if (BaseReg && other.BaseReg &&
1895  BaseReg->getType() != other.BaseReg->getType())
1896  return MultipleFields;
1897  if (BaseGV && other.BaseGV &&
1898  BaseGV->getType() != other.BaseGV->getType())
1899  return MultipleFields;
1900  if (ScaledReg && other.ScaledReg &&
1901  ScaledReg->getType() != other.ScaledReg->getType())
1902  return MultipleFields;
1903 
1904  // Check each field to see if it differs.
1905  unsigned Result = NoField;
1906  if (BaseReg != other.BaseReg)
1907  Result |= BaseRegField;
1908  if (BaseGV != other.BaseGV)
1909  Result |= BaseGVField;
1910  if (BaseOffs != other.BaseOffs)
1911  Result |= BaseOffsField;
1912  if (ScaledReg != other.ScaledReg)
1913  Result |= ScaledRegField;
1914  // Don't count 0 as being a different scale, because that actually means
1915  // unscaled (which will already be counted by having no ScaledReg).
1916  if (Scale && other.Scale && Scale != other.Scale)
1917  Result |= ScaleField;
1918 
1919  if (countPopulation(Result) > 1)
1920  return MultipleFields;
1921  else
1922  return static_cast<FieldName>(Result);
1923  }
1924 
1925  // An AddrMode is trivial if it involves no calculation i.e. it is just a base
1926  // with no offset.
1927  bool isTrivial() {
1928  // An AddrMode is (BaseGV + BaseReg + BaseOffs + ScaleReg * Scale) so it is
1929  // trivial if at most one of these terms is nonzero, except that BaseGV and
1930  // BaseReg both being zero actually means a null pointer value, which we
1931  // consider to be 'non-zero' here.
1932  return !BaseOffs && !Scale && !(BaseGV && BaseReg);
1933  }
1934 
1935  Value *GetFieldAsValue(FieldName Field, Type *IntPtrTy) {
1936  switch (Field) {
1937  default:
1938  return nullptr;
1939  case BaseRegField:
1940  return BaseReg;
1941  case BaseGVField:
1942  return BaseGV;
1943  case ScaledRegField:
1944  return ScaledReg;
1945  case BaseOffsField:
1946  return ConstantInt::get(IntPtrTy, BaseOffs);
1947  }
1948  }
1949 
1950  void SetCombinedField(FieldName Field, Value *V,
1951  const SmallVectorImpl<ExtAddrMode> &AddrModes) {
1952  switch (Field) {
1953  default:
1954  llvm_unreachable("Unhandled fields are expected to be rejected earlier");
1955  break;
1956  case ExtAddrMode::BaseRegField:
1957  BaseReg = V;
1958  break;
1959  case ExtAddrMode::BaseGVField:
1960  // A combined BaseGV is an Instruction, not a GlobalValue, so it goes
1961  // in the BaseReg field.
1962  assert(BaseReg == nullptr);
1963  BaseReg = V;
1964  BaseGV = nullptr;
1965  break;
1966  case ExtAddrMode::ScaledRegField:
1967  ScaledReg = V;
1968  // If we have a mix of scaled and unscaled addrmodes then we want scale
1969  // to be the scale and not zero.
1970  if (!Scale)
1971  for (const ExtAddrMode &AM : AddrModes)
1972  if (AM.Scale) {
1973  Scale = AM.Scale;
1974  break;
1975  }
1976  break;
1977  case ExtAddrMode::BaseOffsField:
1978  // The offset is no longer a constant, so it goes in ScaledReg with a
1979  // scale of 1.
1980  assert(ScaledReg == nullptr);
1981  ScaledReg = V;
1982  Scale = 1;
1983  BaseOffs = 0;
1984  break;
1985  }
1986  }
1987 };
1988 
1989 } // end anonymous namespace
1990 
1991 #ifndef NDEBUG
1992 static inline raw_ostream &operator<<(raw_ostream &OS, const ExtAddrMode &AM) {
1993  AM.print(OS);
1994  return OS;
1995 }
1996 #endif
1997 
1998 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
1999 void ExtAddrMode::print(raw_ostream &OS) const {
2000  bool NeedPlus = false;
2001  OS << "[";
2002  if (BaseGV) {
2003  OS << (NeedPlus ? " + " : "")
2004  << "GV:";
2005  BaseGV->printAsOperand(OS, /*PrintType=*/false);
2006  NeedPlus = true;
2007  }
2008 
2009  if (BaseOffs) {
2010  OS << (NeedPlus ? " + " : "")
2011  << BaseOffs;
2012  NeedPlus = true;
2013  }
2014 
2015  if (BaseReg) {
2016  OS << (NeedPlus ? " + " : "")
2017  << "Base:";
2018  BaseReg->printAsOperand(OS, /*PrintType=*/false);
2019  NeedPlus = true;
2020  }
2021  if (Scale) {
2022  OS << (NeedPlus ? " + " : "")
2023  << Scale << "*";
2024  ScaledReg->printAsOperand(OS, /*PrintType=*/false);
2025  }
2026 
2027  OS << ']';
2028 }
2029 
2030 LLVM_DUMP_METHOD void ExtAddrMode::dump() const {
2031  print(dbgs());
2032  dbgs() << '\n';
2033 }
2034 #endif
2035 
2036 namespace {
2037 
2038 /// \brief This class provides transaction based operation on the IR.
2039 /// Every change made through this class is recorded in the internal state and
2040 /// can be undone (rollback) until commit is called.
2041 class TypePromotionTransaction {
2042  /// \brief This represents the common interface of the individual transaction.
2043  /// Each class implements the logic for doing one specific modification on
2044  /// the IR via the TypePromotionTransaction.
2045  class TypePromotionAction {
2046  protected:
2047  /// The Instruction modified.
2048  Instruction *Inst;
2049 
2050  public:
2051  /// \brief Constructor of the action.
2052  /// The constructor performs the related action on the IR.
2053  TypePromotionAction(Instruction *Inst) : Inst(Inst) {}
2054 
2055  virtual ~TypePromotionAction() = default;
2056 
2057  /// \brief Undo the modification done by this action.
2058  /// When this method is called, the IR must be in the same state as it was
2059  /// before this action was applied.
2060  /// \pre Undoing the action works if and only if the IR is in the exact same
2061  /// state as it was directly after this action was applied.
2062  virtual void undo() = 0;
2063 
2064  /// \brief Advocate every change made by this action.
2065  /// When the results on the IR of the action are to be kept, it is important
2066  /// to call this function, otherwise hidden information may be kept forever.
2067  virtual void commit() {
2068  // Nothing to be done, this action is not doing anything.
2069  }
2070  };
2071 
2072  /// \brief Utility to remember the position of an instruction.
2073  class InsertionHandler {
2074  /// Position of an instruction.
2075  /// Either an instruction:
2076  /// - Is the first in a basic block: BB is used.
2077  /// - Has a previous instructon: PrevInst is used.
2078  union {
2079  Instruction *PrevInst;
2080  BasicBlock *BB;
2081  } Point;
2082 
2083  /// Remember whether or not the instruction had a previous instruction.
2084  bool HasPrevInstruction;
2085 
2086  public:
2087  /// \brief Record the position of \p Inst.
2088  InsertionHandler(Instruction *Inst) {
2089  BasicBlock::iterator It = Inst->getIterator();
2090  HasPrevInstruction = (It != (Inst->getParent()->begin()));
2091  if (HasPrevInstruction)
2092  Point.PrevInst = &*--It;
2093  else
2094  Point.BB = Inst->getParent();
2095  }
2096 
2097  /// \brief Insert \p Inst at the recorded position.
2098  void insert(Instruction *Inst) {
2099  if (HasPrevInstruction) {
2100  if (Inst->getParent())
2101  Inst->removeFromParent();
2102  Inst->insertAfter(Point.PrevInst);
2103  } else {
2104  Instruction *Position = &*Point.BB->getFirstInsertionPt();
2105  if (Inst->getParent())
2106  Inst->moveBefore(Position);
2107  else
2108  Inst->insertBefore(Position);
2109  }
2110  }
2111  };
2112 
2113  /// \brief Move an instruction before another.
2114  class InstructionMoveBefore : public TypePromotionAction {
2115  /// Original position of the instruction.
2116  InsertionHandler Position;
2117 
2118  public:
2119  /// \brief Move \p Inst before \p Before.
2120  InstructionMoveBefore(Instruction *Inst, Instruction *Before)
2121  : TypePromotionAction(Inst), Position(Inst) {
2122  DEBUG(dbgs() << "Do: move: " << *Inst << "\nbefore: " << *Before << "\n");
2123  Inst->moveBefore(Before);
2124  }
2125 
2126  /// \brief Move the instruction back to its original position.
2127  void undo() override {
2128  DEBUG(dbgs() << "Undo: moveBefore: " << *Inst << "\n");
2129  Position.insert(Inst);
2130  }
2131  };
2132 
2133  /// \brief Set the operand of an instruction with a new value.
2134  class OperandSetter : public TypePromotionAction {
2135  /// Original operand of the instruction.
2136  Value *Origin;
2137 
2138  /// Index of the modified instruction.
2139  unsigned Idx;
2140 
2141  public:
2142  /// \brief Set \p Idx operand of \p Inst with \p NewVal.
2143  OperandSetter(Instruction *Inst, unsigned Idx, Value *NewVal)
2144  : TypePromotionAction(Inst), Idx(Idx) {
2145  DEBUG(dbgs() << "Do: setOperand: " << Idx << "\n"
2146  << "for:" << *Inst << "\n"
2147  << "with:" << *NewVal << "\n");
2148  Origin = Inst->getOperand(Idx);
2149  Inst->setOperand(Idx, NewVal);
2150  }
2151 
2152  /// \brief Restore the original value of the instruction.
2153  void undo() override {
2154  DEBUG(dbgs() << "Undo: setOperand:" << Idx << "\n"
2155  << "for: " << *Inst << "\n"
2156  << "with: " << *Origin << "\n");
2157  Inst->setOperand(Idx, Origin);
2158  }
2159  };
2160 
2161  /// \brief Hide the operands of an instruction.
2162  /// Do as if this instruction was not using any of its operands.
2163  class OperandsHider : public TypePromotionAction {
2164  /// The list of original operands.
2165  SmallVector<Value *, 4> OriginalValues;
2166 
2167  public:
2168  /// \brief Remove \p Inst from the uses of the operands of \p Inst.
2169  OperandsHider(Instruction *Inst) : TypePromotionAction(Inst) {
2170  DEBUG(dbgs() << "Do: OperandsHider: " << *Inst << "\n");
2171  unsigned NumOpnds = Inst->getNumOperands();
2172  OriginalValues.reserve(NumOpnds);
2173  for (unsigned It = 0; It < NumOpnds; ++It) {
2174  // Save the current operand.
2175  Value *Val = Inst->getOperand(It);
2176  OriginalValues.push_back(Val);
2177  // Set a dummy one.
2178  // We could use OperandSetter here, but that would imply an overhead
2179  // that we are not willing to pay.
2180  Inst->setOperand(It, UndefValue::get(Val->getType()));
2181  }
2182  }
2183 
2184  /// \brief Restore the original list of uses.
2185  void undo() override {
2186  DEBUG(dbgs() << "Undo: OperandsHider: " << *Inst << "\n");
2187  for (unsigned It = 0, EndIt = OriginalValues.size(); It != EndIt; ++It)
2188  Inst->setOperand(It, OriginalValues[It]);
2189  }
2190  };
2191 
2192  /// \brief Build a truncate instruction.
2193  class TruncBuilder : public TypePromotionAction {
2194  Value *Val;
2195 
2196  public:
2197  /// \brief Build a truncate instruction of \p Opnd producing a \p Ty
2198  /// result.
2199  /// trunc Opnd to Ty.
2200  TruncBuilder(Instruction *Opnd, Type *Ty) : TypePromotionAction(Opnd) {
2201  IRBuilder<> Builder(Opnd);
2202  Val = Builder.CreateTrunc(Opnd, Ty, "promoted");
2203  DEBUG(dbgs() << "Do: TruncBuilder: " << *Val << "\n");
2204  }
2205 
2206  /// \brief Get the built value.
2207  Value *getBuiltValue() { return Val; }
2208 
2209  /// \brief Remove the built instruction.
2210  void undo() override {
2211  DEBUG(dbgs() << "Undo: TruncBuilder: " << *Val << "\n");
2212  if (Instruction *IVal = dyn_cast<Instruction>(Val))
2213  IVal->eraseFromParent();
2214  }
2215  };
2216 
2217  /// \brief Build a sign extension instruction.
2218  class SExtBuilder : public TypePromotionAction {
2219  Value *Val;
2220 
2221  public:
2222  /// \brief Build a sign extension instruction of \p Opnd producing a \p Ty
2223  /// result.
2224  /// sext Opnd to Ty.
2225  SExtBuilder(Instruction *InsertPt, Value *Opnd, Type *Ty)
2226  : TypePromotionAction(InsertPt) {
2227  IRBuilder<> Builder(InsertPt);
2228  Val = Builder.CreateSExt(Opnd, Ty, "promoted");
2229  DEBUG(dbgs() << "Do: SExtBuilder: " << *Val << "\n");
2230  }
2231 
2232  /// \brief Get the built value.
2233  Value *getBuiltValue() { return Val; }
2234 
2235  /// \brief Remove the built instruction.
2236  void undo() override {
2237  DEBUG(dbgs() << "Undo: SExtBuilder: " << *Val << "\n");
2238  if (Instruction *IVal = dyn_cast<Instruction>(Val))
2239  IVal->eraseFromParent();
2240  }
2241  };
2242 
2243  /// \brief Build a zero extension instruction.
2244  class ZExtBuilder : public TypePromotionAction {
2245  Value *Val;
2246 
2247  public:
2248  /// \brief Build a zero extension instruction of \p Opnd producing a \p Ty
2249  /// result.
2250  /// zext Opnd to Ty.
2251  ZExtBuilder(Instruction *InsertPt, Value *Opnd, Type *Ty)
2252  : TypePromotionAction(InsertPt) {
2253  IRBuilder<> Builder(InsertPt);
2254  Val = Builder.CreateZExt(Opnd, Ty, "promoted");
2255  DEBUG(dbgs() << "Do: ZExtBuilder: " << *Val << "\n");
2256  }
2257 
2258  /// \brief Get the built value.
2259  Value *getBuiltValue() { return Val; }
2260 
2261  /// \brief Remove the built instruction.
2262  void undo() override {
2263  DEBUG(dbgs() << "Undo: ZExtBuilder: " << *Val << "\n");
2264  if (Instruction *IVal = dyn_cast<Instruction>(Val))
2265  IVal->eraseFromParent();
2266  }
2267  };
2268 
2269  /// \brief Mutate an instruction to another type.
2270  class TypeMutator : public TypePromotionAction {
2271  /// Record the original type.
2272  Type *OrigTy;
2273 
2274  public:
2275  /// \brief Mutate the type of \p Inst into \p NewTy.
2276  TypeMutator(Instruction *Inst, Type *NewTy)
2277  : TypePromotionAction(Inst), OrigTy(Inst->getType()) {
2278  DEBUG(dbgs() << "Do: MutateType: " << *Inst << " with " << *NewTy
2279  << "\n");
2280  Inst->mutateType(NewTy);
2281  }
2282 
2283  /// \brief Mutate the instruction back to its original type.
2284  void undo() override {
2285  DEBUG(dbgs() << "Undo: MutateType: " << *Inst << " with " << *OrigTy
2286  << "\n");
2287  Inst->mutateType(OrigTy);
2288  }
2289  };
2290 
2291  /// \brief Replace the uses of an instruction by another instruction.
2292  class UsesReplacer : public TypePromotionAction {
2293  /// Helper structure to keep track of the replaced uses.
2294  struct InstructionAndIdx {
2295  /// The instruction using the instruction.
2296  Instruction *Inst;
2297 
2298  /// The index where this instruction is used for Inst.
2299  unsigned Idx;
2300 
2301  InstructionAndIdx(Instruction *Inst, unsigned Idx)
2302  : Inst(Inst), Idx(Idx) {}
2303  };
2304 
2305  /// Keep track of the original uses (pair Instruction, Index).
2306  SmallVector<InstructionAndIdx, 4> OriginalUses;
2307 
2308  using use_iterator = SmallVectorImpl<InstructionAndIdx>::iterator;
2309 
2310  public:
2311  /// \brief Replace all the use of \p Inst by \p New.
2312  UsesReplacer(Instruction *Inst, Value *New) : TypePromotionAction(Inst) {
2313  DEBUG(dbgs() << "Do: UsersReplacer: " << *Inst << " with " << *New
2314  << "\n");
2315  // Record the original uses.
2316  for (Use &U : Inst->uses()) {
2317  Instruction *UserI = cast<Instruction>(U.getUser());
2318  OriginalUses.push_back(InstructionAndIdx(UserI, U.getOperandNo()));
2319  }
2320  // Now, we can replace the uses.
2321  Inst->replaceAllUsesWith(New);
2322  }
2323 
2324  /// \brief Reassign the original uses of Inst to Inst.
2325  void undo() override {
2326  DEBUG(dbgs() << "Undo: UsersReplacer: " << *Inst << "\n");
2327  for (use_iterator UseIt = OriginalUses.begin(),
2328  EndIt = OriginalUses.end();
2329  UseIt != EndIt; ++UseIt) {
2330  UseIt->Inst->setOperand(UseIt->Idx, Inst);
2331  }
2332  }
2333  };
2334 
2335  /// \brief Remove an instruction from the IR.
2336  class InstructionRemover : public TypePromotionAction {
2337  /// Original position of the instruction.
2338  InsertionHandler Inserter;
2339 
2340  /// Helper structure to hide all the link to the instruction. In other
2341  /// words, this helps to do as if the instruction was removed.
2342  OperandsHider Hider;
2343 
2344  /// Keep track of the uses replaced, if any.
2345  UsesReplacer *Replacer = nullptr;
2346 
2347  /// Keep track of instructions removed.
2348  SetOfInstrs &RemovedInsts;
2349 
2350  public:
2351  /// \brief Remove all reference of \p Inst and optinally replace all its
2352  /// uses with New.
2353  /// \p RemovedInsts Keep track of the instructions removed by this Action.
2354  /// \pre If !Inst->use_empty(), then New != nullptr
2355  InstructionRemover(Instruction *Inst, SetOfInstrs &RemovedInsts,
2356  Value *New = nullptr)
2357  : TypePromotionAction(Inst), Inserter(Inst), Hider(Inst),
2358  RemovedInsts(RemovedInsts) {
2359  if (New)
2360  Replacer = new UsesReplacer(Inst, New);
2361  DEBUG(dbgs() << "Do: InstructionRemover: " << *Inst << "\n");
2362  RemovedInsts.insert(Inst);
2363  /// The instructions removed here will be freed after completing
2364  /// optimizeBlock() for all blocks as we need to keep track of the
2365  /// removed instructions during promotion.
2366  Inst->removeFromParent();
2367  }
2368 
2369  ~InstructionRemover() override { delete Replacer; }
2370 
2371  /// \brief Resurrect the instruction and reassign it to the proper uses if
2372  /// new value was provided when build this action.
2373  void undo() override {
2374  DEBUG(dbgs() << "Undo: InstructionRemover: " << *Inst << "\n");
2375  Inserter.insert(Inst);
2376  if (Replacer)
2377  Replacer->undo();
2378  Hider.undo();
2379  RemovedInsts.erase(Inst);
2380  }
2381  };
2382 
2383 public:
2384  /// Restoration point.
2385  /// The restoration point is a pointer to an action instead of an iterator
2386  /// because the iterator may be invalidated but not the pointer.
2387  using ConstRestorationPt = const TypePromotionAction *;
2388 
2389  TypePromotionTransaction(SetOfInstrs &RemovedInsts)
2390  : RemovedInsts(RemovedInsts) {}
2391 
2392  /// Advocate every changes made in that transaction.
2393  void commit();
2394 
2395  /// Undo all the changes made after the given point.
2396  void rollback(ConstRestorationPt Point);
2397 
2398  /// Get the current restoration point.
2399  ConstRestorationPt getRestorationPoint() const;
2400 
2401  /// \name API for IR modification with state keeping to support rollback.
2402  /// @{
2403  /// Same as Instruction::setOperand.
2404  void setOperand(Instruction *Inst, unsigned Idx, Value *NewVal);
2405 
2406  /// Same as Instruction::eraseFromParent.
2407  void eraseInstruction(Instruction *Inst, Value *NewVal = nullptr);
2408 
2409  /// Same as Value::replaceAllUsesWith.
2410  void replaceAllUsesWith(Instruction *Inst, Value *New);
2411 
2412  /// Same as Value::mutateType.
2413  void mutateType(Instruction *Inst, Type *NewTy);
2414 
2415  /// Same as IRBuilder::createTrunc.
2416  Value *createTrunc(Instruction *Opnd, Type *Ty);
2417 
2418  /// Same as IRBuilder::createSExt.
2419  Value *createSExt(Instruction *Inst, Value *Opnd, Type *Ty);
2420 
2421  /// Same as IRBuilder::createZExt.
2422  Value *createZExt(Instruction *Inst, Value *Opnd, Type *Ty);
2423 
2424  /// Same as Instruction::moveBefore.
2425  void moveBefore(Instruction *Inst, Instruction *Before);
2426  /// @}
2427 
2428 private:
2429  /// The ordered list of actions made so far.
2431 
2432  using CommitPt = SmallVectorImpl<std::unique_ptr<TypePromotionAction>>::iterator;
2433 
2434  SetOfInstrs &RemovedInsts;
2435 };
2436 
2437 } // end anonymous namespace
2438 
2439 void TypePromotionTransaction::setOperand(Instruction *Inst, unsigned Idx,
2440  Value *NewVal) {
2441  Actions.push_back(llvm::make_unique<TypePromotionTransaction::OperandSetter>(
2442  Inst, Idx, NewVal));
2443 }
2444 
2445 void TypePromotionTransaction::eraseInstruction(Instruction *Inst,
2446  Value *NewVal) {
2447  Actions.push_back(
2448  llvm::make_unique<TypePromotionTransaction::InstructionRemover>(
2449  Inst, RemovedInsts, NewVal));
2450 }
2451 
2452 void TypePromotionTransaction::replaceAllUsesWith(Instruction *Inst,
2453  Value *New) {
2454  Actions.push_back(
2455  llvm::make_unique<TypePromotionTransaction::UsesReplacer>(Inst, New));
2456 }
2457 
2458 void TypePromotionTransaction::mutateType(Instruction *Inst, Type *NewTy) {
2459  Actions.push_back(
2460  llvm::make_unique<TypePromotionTransaction::TypeMutator>(Inst, NewTy));
2461 }
2462 
2463 Value *TypePromotionTransaction::createTrunc(Instruction *Opnd,
2464  Type *Ty) {
2465  std::unique_ptr<TruncBuilder> Ptr(new TruncBuilder(Opnd, Ty));
2466  Value *Val = Ptr->getBuiltValue();
2467  Actions.push_back(std::move(Ptr));
2468  return Val;
2469 }
2470 
2471 Value *TypePromotionTransaction::createSExt(Instruction *Inst,
2472  Value *Opnd, Type *Ty) {
2473  std::unique_ptr<SExtBuilder> Ptr(new SExtBuilder(Inst, Opnd, Ty));
2474  Value *Val = Ptr->getBuiltValue();
2475  Actions.push_back(std::move(Ptr));
2476  return Val;
2477 }
2478 
2479 Value *TypePromotionTransaction::createZExt(Instruction *Inst,
2480  Value *Opnd, Type *Ty) {
2481  std::unique_ptr<ZExtBuilder> Ptr(new ZExtBuilder(Inst, Opnd, Ty));
2482  Value *Val = Ptr->getBuiltValue();
2483  Actions.push_back(std::move(Ptr));
2484  return Val;
2485 }
2486 
2487 void TypePromotionTransaction::moveBefore(Instruction *Inst,
2488  Instruction *Before) {
2489  Actions.push_back(
2490  llvm::make_unique<TypePromotionTransaction::InstructionMoveBefore>(
2491  Inst, Before));
2492 }
2493 
2494 TypePromotionTransaction::ConstRestorationPt
2495 TypePromotionTransaction::getRestorationPoint() const {
2496  return !Actions.empty() ? Actions.back().get() : nullptr;
2497 }
2498 
2499 void TypePromotionTransaction::commit() {
2500  for (CommitPt It = Actions.begin(), EndIt = Actions.end(); It != EndIt;
2501  ++It)
2502  (*It)->commit();
2503  Actions.clear();
2504 }
2505 
2506 void TypePromotionTransaction::rollback(
2507  TypePromotionTransaction::ConstRestorationPt Point) {
2508  while (!Actions.empty() && Point != Actions.back().get()) {
2509  std::unique_ptr<TypePromotionAction> Curr = Actions.pop_back_val();
2510  Curr->undo();
2511  }
2512 }
2513 
2514 namespace {
2515 
2516 /// \brief A helper class for matching addressing modes.
2517 ///
2518 /// This encapsulates the logic for matching the target-legal addressing modes.
2519 class AddressingModeMatcher {
2520  SmallVectorImpl<Instruction*> &AddrModeInsts;
2521  const TargetLowering &TLI;
2522  const TargetRegisterInfo &TRI;
2523  const DataLayout &DL;
2524 
2525  /// AccessTy/MemoryInst - This is the type for the access (e.g. double) and
2526  /// the memory instruction that we're computing this address for.
2527  Type *AccessTy;
2528  unsigned AddrSpace;
2529  Instruction *MemoryInst;
2530 
2531  /// This is the addressing mode that we're building up. This is
2532  /// part of the return value of this addressing mode matching stuff.
2533  ExtAddrMode &AddrMode;
2534 
2535  /// The instructions inserted by other CodeGenPrepare optimizations.
2536  const SetOfInstrs &InsertedInsts;
2537 
2538  /// A map from the instructions to their type before promotion.
2539  InstrToOrigTy &PromotedInsts;
2540 
2541  /// The ongoing transaction where every action should be registered.
2542  TypePromotionTransaction &TPT;
2543 
2544  /// This is set to true when we should not do profitability checks.
2545  /// When true, IsProfitableToFoldIntoAddressingMode always returns true.
2546  bool IgnoreProfitability;
2547 
2548  AddressingModeMatcher(SmallVectorImpl<Instruction *> &AMI,
2549  const TargetLowering &TLI,
2550  const TargetRegisterInfo &TRI,
2551  Type *AT, unsigned AS,
2552  Instruction *MI, ExtAddrMode &AM,
2553  const SetOfInstrs &InsertedInsts,
2554  InstrToOrigTy &PromotedInsts,
2555  TypePromotionTransaction &TPT)
2556  : AddrModeInsts(AMI), TLI(TLI), TRI(TRI),
2557  DL(MI->getModule()->getDataLayout()), AccessTy(AT), AddrSpace(AS),
2558  MemoryInst(MI), AddrMode(AM), InsertedInsts(InsertedInsts),
2559  PromotedInsts(PromotedInsts), TPT(TPT) {
2560  IgnoreProfitability = false;
2561  }
2562 
2563 public:
2564  /// Find the maximal addressing mode that a load/store of V can fold,
2565  /// give an access type of AccessTy. This returns a list of involved
2566  /// instructions in AddrModeInsts.
2567  /// \p InsertedInsts The instructions inserted by other CodeGenPrepare
2568  /// optimizations.
2569  /// \p PromotedInsts maps the instructions to their type before promotion.
2570  /// \p The ongoing transaction where every action should be registered.
2571  static ExtAddrMode Match(Value *V, Type *AccessTy, unsigned AS,
2572  Instruction *MemoryInst,
2573  SmallVectorImpl<Instruction*> &AddrModeInsts,
2574  const TargetLowering &TLI,
2575  const TargetRegisterInfo &TRI,
2576  const SetOfInstrs &InsertedInsts,
2577  InstrToOrigTy &PromotedInsts,
2578  TypePromotionTransaction &TPT) {
2579  ExtAddrMode Result;
2580 
2581  bool Success = AddressingModeMatcher(AddrModeInsts, TLI, TRI,
2582  AccessTy, AS,
2583  MemoryInst, Result, InsertedInsts,
2584  PromotedInsts, TPT).matchAddr(V, 0);
2585  (void)Success; assert(Success && "Couldn't select *anything*?");
2586  return Result;
2587  }
2588 
2589 private:
2590  bool matchScaledValue(Value *ScaleReg, int64_t Scale, unsigned Depth);
2591  bool matchAddr(Value *V, unsigned Depth);
2592  bool matchOperationAddr(User *Operation, unsigned Opcode, unsigned Depth,
2593  bool *MovedAway = nullptr);
2594  bool isProfitableToFoldIntoAddressingMode(Instruction *I,
2595  ExtAddrMode &AMBefore,
2596  ExtAddrMode &AMAfter);
2597  bool valueAlreadyLiveAtInst(Value *Val, Value *KnownLive1, Value *KnownLive2);
2598  bool isPromotionProfitable(unsigned NewCost, unsigned OldCost,
2599  Value *PromotedOperand) const;
2600 };
2601 
2602 /// \brief Keep track of simplification of Phi nodes.
2603 /// Accept the set of all phi nodes and erase phi node from this set
2604 /// if it is simplified.
2605 class SimplificationTracker {
2607  const SimplifyQuery &SQ;
2608  SmallPtrSetImpl<PHINode *> &AllPhiNodes;
2609  SmallPtrSetImpl<SelectInst *> &AllSelectNodes;
2610 
2611 public:
2612  SimplificationTracker(const SimplifyQuery &sq,
2615  : SQ(sq), AllPhiNodes(APN), AllSelectNodes(ASN) {}
2616 
2617  Value *Get(Value *V) {
2618  do {
2619  auto SV = Storage.find(V);
2620  if (SV == Storage.end())
2621  return V;
2622  V = SV->second;
2623  } while (true);
2624  }
2625 
2626  Value *Simplify(Value *Val) {
2627  SmallVector<Value *, 32> WorkList;
2628  SmallPtrSet<Value *, 32> Visited;
2629  WorkList.push_back(Val);
2630  while (!WorkList.empty()) {
2631  auto P = WorkList.pop_back_val();
2632  if (!Visited.insert(P).second)
2633  continue;
2634  if (auto *PI = dyn_cast<Instruction>(P))
2635  if (Value *V = SimplifyInstruction(cast<Instruction>(PI), SQ)) {
2636  for (auto *U : PI->users())
2637  WorkList.push_back(cast<Value>(U));
2638  Put(PI, V);
2639  PI->replaceAllUsesWith(V);
2640  if (auto *PHI = dyn_cast<PHINode>(PI))
2641  AllPhiNodes.erase(PHI);
2642  if (auto *Select = dyn_cast<SelectInst>(PI))
2643  AllSelectNodes.erase(Select);
2644  PI->eraseFromParent();
2645  }
2646  }
2647  return Get(Val);
2648  }
2649 
2650  void Put(Value *From, Value *To) {
2651  Storage.insert({ From, To });
2652  }
2653 };
2654 
2655 /// \brief A helper class for combining addressing modes.
2656 class AddressingModeCombiner {
2657  typedef std::pair<Value *, BasicBlock *> ValueInBB;
2658  typedef DenseMap<ValueInBB, Value *> FoldAddrToValueMapping;
2659  typedef std::pair<PHINode *, PHINode *> PHIPair;
2660 
2661 private:
2662  /// The addressing modes we've collected.
2663  SmallVector<ExtAddrMode, 16> AddrModes;
2664 
2665  /// The field in which the AddrModes differ, when we have more than one.
2666  ExtAddrMode::FieldName DifferentField = ExtAddrMode::NoField;
2667 
2668  /// Are the AddrModes that we have all just equal to their original values?
2669  bool AllAddrModesTrivial = true;
2670 
2671  /// Common Type for all different fields in addressing modes.
2672  Type *CommonType;
2673 
2674  /// SimplifyQuery for simplifyInstruction utility.
2675  const SimplifyQuery &SQ;
2676 
2677  /// Original Address.
2678  ValueInBB Original;
2679 
2680 public:
2681  AddressingModeCombiner(const SimplifyQuery &_SQ, ValueInBB OriginalValue)
2682  : CommonType(nullptr), SQ(_SQ), Original(OriginalValue) {}
2683 
2684  /// \brief Get the combined AddrMode
2685  const ExtAddrMode &getAddrMode() const {
2686  return AddrModes[0];
2687  }
2688 
2689  /// \brief Add a new AddrMode if it's compatible with the AddrModes we already
2690  /// have.
2691  /// \return True iff we succeeded in doing so.
2692  bool addNewAddrMode(ExtAddrMode &NewAddrMode) {
2693  // Take note of if we have any non-trivial AddrModes, as we need to detect
2694  // when all AddrModes are trivial as then we would introduce a phi or select
2695  // which just duplicates what's already there.
2696  AllAddrModesTrivial = AllAddrModesTrivial && NewAddrMode.isTrivial();
2697 
2698  // If this is the first addrmode then everything is fine.
2699  if (AddrModes.empty()) {
2700  AddrModes.emplace_back(NewAddrMode);
2701  return true;
2702  }
2703 
2704  // Figure out how different this is from the other address modes, which we
2705  // can do just by comparing against the first one given that we only care
2706  // about the cumulative difference.
2707  ExtAddrMode::FieldName ThisDifferentField =
2708  AddrModes[0].compare(NewAddrMode);
2709  if (DifferentField == ExtAddrMode::NoField)
2710  DifferentField = ThisDifferentField;
2711  else if (DifferentField != ThisDifferentField)
2712  DifferentField = ExtAddrMode::MultipleFields;
2713 
2714  // If NewAddrMode differs in only one dimension, and that dimension isn't
2715  // the amount that ScaledReg is scaled by, then we can handle it by
2716  // inserting a phi/select later on. Even if NewAddMode is the same
2717  // we still need to collect it due to original value is different.
2718  // And later we will need all original values as anchors during
2719  // finding the common Phi node.
2720  if (DifferentField != ExtAddrMode::MultipleFields &&
2721  DifferentField != ExtAddrMode::ScaleField) {
2722  AddrModes.emplace_back(NewAddrMode);
2723  return true;
2724  }
2725 
2726  // We couldn't combine NewAddrMode with the rest, so return failure.
2727  AddrModes.clear();
2728  return false;
2729  }
2730 
2731  /// \brief Combine the addressing modes we've collected into a single
2732  /// addressing mode.
2733  /// \return True iff we successfully combined them or we only had one so
2734  /// didn't need to combine them anyway.
2735  bool combineAddrModes() {
2736  // If we have no AddrModes then they can't be combined.
2737  if (AddrModes.size() == 0)
2738  return false;
2739 
2740  // A single AddrMode can trivially be combined.
2741  if (AddrModes.size() == 1 || DifferentField == ExtAddrMode::NoField)
2742  return true;
2743 
2744  // If the AddrModes we collected are all just equal to the value they are
2745  // derived from then combining them wouldn't do anything useful.
2746  if (AllAddrModesTrivial)
2747  return false;
2748 
2749  if (!addrModeCombiningAllowed())
2750  return false;
2751 
2752  // Build a map between <original value, basic block where we saw it> to
2753  // value of base register.
2754  // Bail out if there is no common type.
2755  FoldAddrToValueMapping Map;
2756  if (!initializeMap(Map))
2757  return false;
2758 
2759  Value *CommonValue = findCommon(Map);
2760  if (CommonValue)
2761  AddrModes[0].SetCombinedField(DifferentField, CommonValue, AddrModes);
2762  return CommonValue != nullptr;
2763  }
2764 
2765 private:
2766  /// \brief Initialize Map with anchor values. For address seen in some BB
2767  /// we set the value of different field saw in this address.
2768  /// If address is not an instruction than basic block is set to null.
2769  /// At the same time we find a common type for different field we will
2770  /// use to create new Phi/Select nodes. Keep it in CommonType field.
2771  /// Return false if there is no common type found.
2772  bool initializeMap(FoldAddrToValueMapping &Map) {
2773  // Keep track of keys where the value is null. We will need to replace it
2774  // with constant null when we know the common type.
2775  SmallVector<ValueInBB, 2> NullValue;
2776  Type *IntPtrTy = SQ.DL.getIntPtrType(AddrModes[0].OriginalValue->getType());
2777  for (auto &AM : AddrModes) {
2778  BasicBlock *BB = nullptr;
2779  if (Instruction *I = dyn_cast<Instruction>(AM.OriginalValue))
2780  BB = I->getParent();
2781 
2782  Value *DV = AM.GetFieldAsValue(DifferentField, IntPtrTy);
2783  if (DV) {
2784  auto *Type = DV->getType();
2785  if (CommonType && CommonType != Type)
2786  return false;
2787  CommonType = Type;
2788  Map[{ AM.OriginalValue, BB }] = DV;
2789  } else {
2790  NullValue.push_back({ AM.OriginalValue, BB });
2791  }
2792  }
2793  assert(CommonType && "At least one non-null value must be!");
2794  for (auto VIBB : NullValue)
2795  Map[VIBB] = Constant::getNullValue(CommonType);
2796  return true;
2797  }
2798 
2799  /// \brief We have mapping between value A and basic block where value A
2800  /// seen to other value B where B was a field in addressing mode represented
2801  /// by A. Also we have an original value C representin an address in some
2802  /// basic block. Traversing from C through phi and selects we ended up with
2803  /// A's in a map. This utility function tries to find a value V which is a
2804  /// field in addressing mode C and traversing through phi nodes and selects
2805  /// we will end up in corresponded values B in a map.
2806  /// The utility will create a new Phi/Selects if needed.
2807  // The simple example looks as follows:
2808  // BB1:
2809  // p1 = b1 + 40
2810  // br cond BB2, BB3
2811  // BB2:
2812  // p2 = b2 + 40
2813  // br BB3
2814  // BB3:
2815  // p = phi [p1, BB1], [p2, BB2]
2816  // v = load p
2817  // Map is
2818  // <p1, BB1> -> b1
2819  // <p2, BB2> -> b2
2820  // Request is
2821  // <p, BB3> -> ?
2822  // The function tries to find or build phi [b1, BB1], [b2, BB2] in BB3
2823  Value *findCommon(FoldAddrToValueMapping &Map) {
2824  // Tracks of new created Phi nodes.
2825  SmallPtrSet<PHINode *, 32> NewPhiNodes;
2826  // Tracks of new created Select nodes.
2827  SmallPtrSet<SelectInst *, 32> NewSelectNodes;
2828  // Tracks the simplification of new created phi nodes. The reason we use
2829  // this mapping is because we will add new created Phi nodes in AddrToBase.
2830  // Simplification of Phi nodes is recursive, so some Phi node may
2831  // be simplified after we added it to AddrToBase.
2832  // Using this mapping we can find the current value in AddrToBase.
2833  SimplificationTracker ST(SQ, NewPhiNodes, NewSelectNodes);
2834 
2835  // First step, DFS to create PHI nodes for all intermediate blocks.
2836  // Also fill traverse order for the second step.
2837  SmallVector<ValueInBB, 32> TraverseOrder;
2838  InsertPlaceholders(Map, TraverseOrder, NewPhiNodes, NewSelectNodes);
2839 
2840  // Second Step, fill new nodes by merged values and simplify if possible.
2841  FillPlaceholders(Map, TraverseOrder, ST);
2842 
2843  if (!AddrSinkNewSelects && NewSelectNodes.size() > 0) {
2844  DestroyNodes(NewPhiNodes);
2845  DestroyNodes(NewSelectNodes);
2846  return nullptr;
2847  }
2848 
2849  // Now we'd like to match New Phi nodes to existed ones.
2850  unsigned PhiNotMatchedCount = 0;
2851  if (!MatchPhiSet(NewPhiNodes, ST, AddrSinkNewPhis, PhiNotMatchedCount)) {
2852  DestroyNodes(NewPhiNodes);
2853  DestroyNodes(NewSelectNodes);
2854  return nullptr;
2855  }
2856 
2857  auto *Result = ST.Get(Map.find(Original)->second);
2858  if (Result) {
2859  NumMemoryInstsPhiCreated += NewPhiNodes.size() + PhiNotMatchedCount;
2860  NumMemoryInstsSelectCreated += NewSelectNodes.size();
2861  }
2862  return Result;
2863  }
2864 
2865  /// \brief Destroy nodes from a set.
2866  template <typename T> void DestroyNodes(SmallPtrSetImpl<T *> &Instructions) {
2867  // For safe erasing, replace the Phi with dummy value first.
2868  auto Dummy = UndefValue::get(CommonType);
2869  for (auto I : Instructions) {
2871  I->eraseFromParent();
2872  }
2873  }
2874 
2875  /// \brief Try to match PHI node to Candidate.
2876  /// Matcher tracks the matched Phi nodes.
2877  bool MatchPhiNode(PHINode *PHI, PHINode *Candidate,
2878  DenseSet<PHIPair> &Matcher,
2879  SmallPtrSetImpl<PHINode *> &PhiNodesToMatch) {
2880  SmallVector<PHIPair, 8> WorkList;
2881  Matcher.insert({ PHI, Candidate });
2882  WorkList.push_back({ PHI, Candidate });
2883  SmallSet<PHIPair, 8> Visited;
2884  while (!WorkList.empty()) {
2885  auto Item = WorkList.pop_back_val();
2886  if (!Visited.insert(Item).second)
2887  continue;
2888  // We iterate over all incoming values to Phi to compare them.
2889  // If values are different and both of them Phi and the first one is a
2890  // Phi we added (subject to match) and both of them is in the same basic
2891  // block then we can match our pair if values match. So we state that
2892  // these values match and add it to work list to verify that.
2893  for (auto B : Item.first->blocks()) {
2894  Value *FirstValue = Item.first->getIncomingValueForBlock(B);
2895  Value *SecondValue = Item.second->getIncomingValueForBlock(B);
2896  if (FirstValue == SecondValue)
2897  continue;
2898 
2899  PHINode *FirstPhi = dyn_cast<PHINode>(FirstValue);
2900  PHINode *SecondPhi = dyn_cast<PHINode>(SecondValue);
2901 
2902  // One of them is not Phi or
2903  // The first one is not Phi node from the set we'd like to match or
2904  // Phi nodes from different basic blocks then
2905  // we will not be able to match.
2906  if (!FirstPhi || !SecondPhi || !PhiNodesToMatch.count(FirstPhi) ||
2907  FirstPhi->getParent() != SecondPhi->getParent())
2908  return false;
2909 
2910  // If we already matched them then continue.
2911  if (Matcher.count({ FirstPhi, SecondPhi }))
2912  continue;
2913  // So the values are different and does not match. So we need them to
2914  // match.
2915  Matcher.insert({ FirstPhi, SecondPhi });
2916  // But me must check it.
2917  WorkList.push_back({ FirstPhi, SecondPhi });
2918  }
2919  }
2920  return true;
2921  }
2922 
2923  /// \brief For the given set of PHI nodes try to find their equivalents.
2924  /// Returns false if this matching fails and creation of new Phi is disabled.
2925  bool MatchPhiSet(SmallPtrSetImpl<PHINode *> &PhiNodesToMatch,
2926  SimplificationTracker &ST, bool AllowNewPhiNodes,
2927  unsigned &PhiNotMatchedCount) {
2928  DenseSet<PHIPair> Matched;
2929  SmallPtrSet<PHINode *, 8> WillNotMatch;
2930  while (PhiNodesToMatch.size()) {
2931  PHINode *PHI = *PhiNodesToMatch.begin();
2932 
2933  // Add us, if no Phi nodes in the basic block we do not match.
2934  WillNotMatch.clear();
2935  WillNotMatch.insert(PHI);
2936 
2937  // Traverse all Phis until we found equivalent or fail to do that.
2938  bool IsMatched = false;
2939  for (auto &P : PHI->getParent()->phis()) {
2940  if (&P == PHI)
2941  continue;
2942  if ((IsMatched = MatchPhiNode(PHI, &P, Matched, PhiNodesToMatch)))
2943  break;
2944  // If it does not match, collect all Phi nodes from matcher.
2945  // if we end up with no match, them all these Phi nodes will not match
2946  // later.
2947  for (auto M : Matched)
2948  WillNotMatch.insert(M.first);
2949  Matched.clear();
2950  }
2951  if (IsMatched) {
2952  // Replace all matched values and erase them.
2953  for (auto MV : Matched) {
2954  MV.first->replaceAllUsesWith(MV.second);
2955  PhiNodesToMatch.erase(MV.first);
2956  ST.Put(MV.first, MV.second);
2957  MV.first->eraseFromParent();
2958  }
2959  Matched.clear();
2960  continue;
2961  }
2962  // If we are not allowed to create new nodes then bail out.
2963  if (!AllowNewPhiNodes)
2964  return false;
2965  // Just remove all seen values in matcher. They will not match anything.
2966  PhiNotMatchedCount += WillNotMatch.size();
2967  for (auto *P : WillNotMatch)
2968  PhiNodesToMatch.erase(P);
2969  }
2970  return true;
2971  }
2972  /// \brief Fill the placeholder with values from predecessors and simplify it.
2973  void FillPlaceholders(FoldAddrToValueMapping &Map,
2974  SmallVectorImpl<ValueInBB> &TraverseOrder,
2975  SimplificationTracker &ST) {
2976  while (!TraverseOrder.empty()) {
2977  auto Current = TraverseOrder.pop_back_val();
2978  assert(Map.find(Current) != Map.end() && "No node to fill!!!");
2979  Value *CurrentValue = Current.first;
2980  BasicBlock *CurrentBlock = Current.second;
2981  Value *V = Map[Current];
2982 
2983  if (SelectInst *Select = dyn_cast<SelectInst>(V)) {
2984  // CurrentValue also must be Select.
2985  auto *CurrentSelect = cast<SelectInst>(CurrentValue);
2986  auto *TrueValue = CurrentSelect->getTrueValue();
2987  ValueInBB TrueItem = { TrueValue, isa<Instruction>(TrueValue)
2988  ? CurrentBlock
2989  : nullptr };
2990  assert(Map.find(TrueItem) != Map.end() && "No True Value!");
2991  Select->setTrueValue(Map[TrueItem]);
2992  auto *FalseValue = CurrentSelect->getFalseValue();
2993  ValueInBB FalseItem = { FalseValue, isa<Instruction>(FalseValue)
2994  ? CurrentBlock
2995  : nullptr };
2996  assert(Map.find(FalseItem) != Map.end() && "No False Value!");
2997  Select->setFalseValue(Map[FalseItem]);
2998  } else {
2999  // Must be a Phi node then.
3000  PHINode *PHI = cast<PHINode>(V);
3001  // Fill the Phi node with values from predecessors.
3002  bool IsDefinedInThisBB =
3003  cast<Instruction>(CurrentValue)->getParent() == CurrentBlock;
3004  auto *CurrentPhi = dyn_cast<PHINode>(CurrentValue);
3005  for (auto B : predecessors(CurrentBlock)) {
3006  Value *PV = IsDefinedInThisBB
3007  ? CurrentPhi->getIncomingValueForBlock(B)
3008  : CurrentValue;
3009  ValueInBB item = { PV, isa<Instruction>(PV) ? B : nullptr };
3010  assert(Map.find(item) != Map.end() && "No predecessor Value!");
3011  PHI->addIncoming(ST.Get(Map[item]), B);
3012  }
3013  }
3014  // Simplify if possible.
3015  Map[Current] = ST.Simplify(V);
3016  }
3017  }
3018 
3019  /// Starting from value recursively iterates over predecessors up to known
3020  /// ending values represented in a map. For each traversed block inserts
3021  /// a placeholder Phi or Select.
3022  /// Reports all new created Phi/Select nodes by adding them to set.
3023  /// Also reports and order in what basic blocks have been traversed.
3024  void InsertPlaceholders(FoldAddrToValueMapping &Map,
3025  SmallVectorImpl<ValueInBB> &TraverseOrder,
3026  SmallPtrSetImpl<PHINode *> &NewPhiNodes,
3027  SmallPtrSetImpl<SelectInst *> &NewSelectNodes) {
3028  SmallVector<ValueInBB, 32> Worklist;
3029  assert((isa<PHINode>(Original.first) || isa<SelectInst>(Original.first)) &&
3030  "Address must be a Phi or Select node");
3031  auto *Dummy = UndefValue::get(CommonType);
3032  Worklist.push_back(Original);
3033  while (!Worklist.empty()) {
3034  auto Current = Worklist.pop_back_val();
3035  // If value is not an instruction it is something global, constant,
3036  // parameter and we can say that this value is observable in any block.
3037  // Set block to null to denote it.
3038  // Also please take into account that it is how we build anchors.
3039  if (!isa<Instruction>(Current.first))
3040  Current.second = nullptr;
3041  // if it is already visited or it is an ending value then skip it.
3042  if (Map.find(Current) != Map.end())
3043  continue;
3044  TraverseOrder.push_back(Current);
3045 
3046  Value *CurrentValue = Current.first;
3047  BasicBlock *CurrentBlock = Current.second;
3048  // CurrentValue must be a Phi node or select. All others must be covered
3049  // by anchors.
3050  Instruction *CurrentI = cast<Instruction>(CurrentValue);
3051  bool IsDefinedInThisBB = CurrentI->getParent() == CurrentBlock;
3052 
3053  unsigned PredCount =
3054  std::distance(pred_begin(CurrentBlock), pred_end(CurrentBlock));
3055  // if Current Value is not defined in this basic block we are interested
3056  // in values in predecessors.
3057  if (!IsDefinedInThisBB) {
3058  assert(PredCount && "Unreachable block?!");
3059  PHINode *PHI = PHINode::Create(CommonType, PredCount, "sunk_phi",
3060  &CurrentBlock->front());
3061  Map[Current] = PHI;
3062  NewPhiNodes.insert(PHI);
3063  // Add all predecessors in work list.
3064  for (auto B : predecessors(CurrentBlock))
3065  Worklist.push_back({ CurrentValue, B });
3066  continue;
3067  }
3068  // Value is defined in this basic block.
3069  if (SelectInst *OrigSelect = dyn_cast<SelectInst>(CurrentI)) {
3070  // Is it OK to get metadata from OrigSelect?!
3071  // Create a Select placeholder with dummy value.
3072  SelectInst *Select =
3073  SelectInst::Create(OrigSelect->getCondition(), Dummy, Dummy,
3074  OrigSelect->getName(), OrigSelect, OrigSelect);
3075  Map[Current] = Select;
3076  NewSelectNodes.insert(Select);
3077  // We are interested in True and False value in this basic block.
3078  Worklist.push_back({ OrigSelect->getTrueValue(), CurrentBlock });
3079  Worklist.push_back({ OrigSelect->getFalseValue(), CurrentBlock });
3080  } else {
3081  // It must be a Phi node then.
3082  auto *CurrentPhi = cast<PHINode>(CurrentI);
3083  // Create new Phi node for merge of bases.
3084  assert(PredCount && "Unreachable block?!");
3085  PHINode *PHI = PHINode::Create(CommonType, PredCount, "sunk_phi",
3086  &CurrentBlock->front());
3087  Map[Current] = PHI;
3088  NewPhiNodes.insert(PHI);
3089 
3090  // Add all predecessors in work list.
3091  for (auto B : predecessors(CurrentBlock))
3092  Worklist.push_back({ CurrentPhi->getIncomingValueForBlock(B), B });
3093  }
3094  }
3095  }
3096 
3097  bool addrModeCombiningAllowed() {
3099  return false;
3100  switch (DifferentField) {
3101  default:
3102  return false;
3103  case ExtAddrMode::BaseRegField:
3104  return AddrSinkCombineBaseReg;
3105  case ExtAddrMode::BaseGVField:
3106  return AddrSinkCombineBaseGV;
3107  case ExtAddrMode::BaseOffsField:
3108  return AddrSinkCombineBaseOffs;
3109  case ExtAddrMode::ScaledRegField:
3110  return AddrSinkCombineScaledReg;
3111  }
3112  }
3113 };
3114 } // end anonymous namespace
3115 
3116 /// Try adding ScaleReg*Scale to the current addressing mode.
3117 /// Return true and update AddrMode if this addr mode is legal for the target,
3118 /// false if not.
3119 bool AddressingModeMatcher::matchScaledValue(Value *ScaleReg, int64_t Scale,
3120  unsigned Depth) {
3121  // If Scale is 1, then this is the same as adding ScaleReg to the addressing
3122  // mode. Just process that directly.
3123  if (Scale == 1)
3124  return matchAddr(ScaleReg, Depth);
3125 
3126  // If the scale is 0, it takes nothing to add this.
3127  if (Scale == 0)
3128  return true;
3129 
3130  // If we already have a scale of this value, we can add to it, otherwise, we
3131  // need an available scale field.
3132  if (AddrMode.Scale != 0 && AddrMode.ScaledReg != ScaleReg)
3133  return false;
3134 
3135  ExtAddrMode TestAddrMode = AddrMode;
3136 
3137  // Add scale to turn X*4+X*3 -> X*7. This could also do things like
3138  // [A+B + A*7] -> [B+A*8].
3139  TestAddrMode.Scale += Scale;
3140  TestAddrMode.ScaledReg = ScaleReg;
3141 
3142  // If the new address isn't legal, bail out.
3143  if (!TLI.isLegalAddressingMode(DL, TestAddrMode, AccessTy, AddrSpace))
3144  return false;
3145 
3146  // It was legal, so commit it.
3147  AddrMode = TestAddrMode;
3148 
3149  // Okay, we decided that we can add ScaleReg+Scale to AddrMode. Check now
3150  // to see if ScaleReg is actually X+C. If so, we can turn this into adding
3151  // X*Scale + C*Scale to addr mode.
3152  ConstantInt *CI = nullptr; Value *AddLHS = nullptr;
3153  if (isa<Instruction>(ScaleReg) && // not a constant expr.
3154  match(ScaleReg, m_Add(m_Value(AddLHS), m_ConstantInt(CI)))) {
3155  TestAddrMode.ScaledReg = AddLHS;
3156  TestAddrMode.BaseOffs += CI->getSExtValue()*TestAddrMode.Scale;
3157 
3158  // If this addressing mode is legal, commit it and remember that we folded
3159  // this instruction.
3160  if (TLI.isLegalAddressingMode(DL, TestAddrMode, AccessTy, AddrSpace)) {
3161  AddrModeInsts.push_back(cast<Instruction>(ScaleReg));
3162  AddrMode = TestAddrMode;
3163  return true;
3164  }
3165  }
3166 
3167  // Otherwise, not (x+c)*scale, just return what we have.
3168  return true;
3169 }
3170 
3171 /// This is a little filter, which returns true if an addressing computation
3172 /// involving I might be folded into a load/store accessing it.
3173 /// This doesn't need to be perfect, but needs to accept at least
3174 /// the set of instructions that MatchOperationAddr can.
3176  switch (I->getOpcode()) {
3177  case Instruction::BitCast:
3178  case Instruction::AddrSpaceCast:
3179  // Don't touch identity bitcasts.
3180  if (I->getType() == I->getOperand(0)->getType())
3181  return false;
3182  return I->getType()->isPointerTy() || I->getType()->isIntegerTy();
3183  case Instruction::PtrToInt:
3184  // PtrToInt is always a noop, as we know that the int type is pointer sized.
3185  return true;
3186  case Instruction::IntToPtr:
3187  // We know the input is intptr_t, so this is foldable.
3188  return true;
3189  case Instruction::Add:
3190  return true;
3191  case Instruction::Mul:
3192  case Instruction::Shl:
3193  // Can only handle X*C and X << C.
3194  return isa<ConstantInt>(I->getOperand(1));
3195  case Instruction::GetElementPtr:
3196  return true;
3197  default:
3198  return false;
3199  }
3200 }
3201 
3202 /// \brief Check whether or not \p Val is a legal instruction for \p TLI.
3203 /// \note \p Val is assumed to be the product of some type promotion.
3204 /// Therefore if \p Val has an undefined state in \p TLI, this is assumed
3205 /// to be legal, as the non-promoted value would have had the same state.
3207  const DataLayout &DL, Value *Val) {
3208  Instruction *PromotedInst = dyn_cast<Instruction>(Val);
3209  if (!PromotedInst)
3210  return false;
3211  int ISDOpcode = TLI.InstructionOpcodeToISD(PromotedInst->getOpcode());
3212  // If the ISDOpcode is undefined, it was undefined before the promotion.
3213  if (!ISDOpcode)
3214  return true;
3215  // Otherwise, check if the promoted instruction is legal or not.
3216  return TLI.isOperationLegalOrCustom(
3217  ISDOpcode, TLI.getValueType(DL, PromotedInst->getType()));
3218 }
3219 
3220 namespace {
3221 
3222 /// \brief Hepler class to perform type promotion.
3223 class TypePromotionHelper {
3224  /// \brief Utility function to check whether or not a sign or zero extension
3225  /// of \p Inst with \p ConsideredExtType can be moved through \p Inst by
3226  /// either using the operands of \p Inst or promoting \p Inst.
3227  /// The type of the extension is defined by \p IsSExt.
3228  /// In other words, check if:
3229  /// ext (Ty Inst opnd1 opnd2 ... opndN) to ConsideredExtType.
3230  /// #1 Promotion applies:
3231  /// ConsideredExtType Inst (ext opnd1 to ConsideredExtType, ...).
3232  /// #2 Operand reuses:
3233  /// ext opnd1 to ConsideredExtType.
3234  /// \p PromotedInsts maps the instructions to their type before promotion.
3235  static bool canGetThrough(const Instruction *Inst, Type *ConsideredExtType,
3236  const InstrToOrigTy &PromotedInsts, bool IsSExt);
3237 
3238  /// \brief Utility function to determine if \p OpIdx should be promoted when
3239  /// promoting \p Inst.
3240  static bool shouldExtOperand(const Instruction *Inst, int OpIdx) {
3241  return !(isa<SelectInst>(Inst) && OpIdx == 0);
3242  }
3243 
3244  /// \brief Utility function to promote the operand of \p Ext when this
3245  /// operand is a promotable trunc or sext or zext.
3246  /// \p PromotedInsts maps the instructions to their type before promotion.
3247  /// \p CreatedInstsCost[out] contains the cost of all instructions
3248  /// created to promote the operand of Ext.
3249  /// Newly added extensions are inserted in \p Exts.
3250  /// Newly added truncates are inserted in \p Truncs.
3251  /// Should never be called directly.
3252  /// \return The promoted value which is used instead of Ext.
3253  static Value *promoteOperandForTruncAndAnyExt(
3254  Instruction *Ext, TypePromotionTransaction &TPT,
3255  InstrToOrigTy &PromotedInsts, unsigned &CreatedInstsCost,
3257  SmallVectorImpl<Instruction *> *Truncs, const TargetLowering &TLI);
3258 
3259  /// \brief Utility function to promote the operand of \p Ext when this
3260  /// operand is promotable and is not a supported trunc or sext.
3261  /// \p PromotedInsts maps the instructions to their type before promotion.
3262  /// \p CreatedInstsCost[out] contains the cost of all the instructions
3263  /// created to promote the operand of Ext.
3264  /// Newly added extensions are inserted in \p Exts.
3265  /// Newly added truncates are inserted in \p Truncs.
3266  /// Should never be called directly.
3267  /// \return The promoted value which is used instead of Ext.
3268  static Value *promoteOperandForOther(Instruction *Ext,
3269  TypePromotionTransaction &TPT,
3270  InstrToOrigTy &PromotedInsts,
3271  unsigned &CreatedInstsCost,
3274  const TargetLowering &TLI, bool IsSExt);
3275 
3276  /// \see promoteOperandForOther.
3277  static Value *signExtendOperandForOther(
3278  Instruction *Ext, TypePromotionTransaction &TPT,
3279  InstrToOrigTy &PromotedInsts, unsigned &CreatedInstsCost,
3281  SmallVectorImpl<Instruction *> *Truncs, const TargetLowering &TLI) {
3282  return promoteOperandForOther(Ext, TPT, PromotedInsts, CreatedInstsCost,
3283  Exts, Truncs, TLI, true);
3284  }
3285 
3286  /// \see promoteOperandForOther.
3287  static Value *zeroExtendOperandForOther(
3288  Instruction *Ext, TypePromotionTransaction &TPT,
3289  InstrToOrigTy &PromotedInsts, unsigned &CreatedInstsCost,
3291  SmallVectorImpl<Instruction *> *Truncs, const TargetLowering &TLI) {
3292  return promoteOperandForOther(Ext, TPT, PromotedInsts, CreatedInstsCost,
3293  Exts, Truncs, TLI, false);
3294  }
3295 
3296 public:
3297  /// Type for the utility function that promotes the operand of Ext.
3298  using Action = Value *(*)(Instruction *Ext, TypePromotionTransaction &TPT,
3299  InstrToOrigTy &PromotedInsts,
3300  unsigned &CreatedInstsCost,
3303  const TargetLowering &TLI);
3304 
3305  /// \brief Given a sign/zero extend instruction \p Ext, return the approriate
3306  /// action to promote the operand of \p Ext instead of using Ext.
3307  /// \return NULL if no promotable action is possible with the current
3308  /// sign extension.
3309  /// \p InsertedInsts keeps track of all the instructions inserted by the
3310  /// other CodeGenPrepare optimizations. This information is important
3311  /// because we do not want to promote these instructions as CodeGenPrepare
3312  /// will reinsert them later. Thus creating an infinite loop: create/remove.
3313  /// \p PromotedInsts maps the instructions to their type before promotion.
3314  static Action getAction(Instruction *Ext, const SetOfInstrs &InsertedInsts,
3315  const TargetLowering &TLI,
3316  const InstrToOrigTy &PromotedInsts);
3317 };
3318 
3319 } // end anonymous namespace
3320 
3321 bool TypePromotionHelper::canGetThrough(const Instruction *Inst,
3322  Type *ConsideredExtType,
3323  const InstrToOrigTy &PromotedInsts,
3324  bool IsSExt) {
3325  // The promotion helper does not know how to deal with vector types yet.
3326  // To be able to fix that, we would need to fix the places where we
3327  // statically extend, e.g., constants and such.
3328  if (Inst->getType()->isVectorTy())
3329  return false;
3330 
3331  // We can always get through zext.
3332  if (isa<ZExtInst>(Inst))
3333  return true;
3334 
3335  // sext(sext) is ok too.
3336  if (IsSExt && isa<SExtInst>(Inst))
3337  return true;
3338 
3339  // We can get through binary operator, if it is legal. In other words, the
3340  // binary operator must have a nuw or nsw flag.
3341  const BinaryOperator *BinOp = dyn_cast<BinaryOperator>(Inst);
3342  if (BinOp && isa<OverflowingBinaryOperator>(BinOp) &&
3343  ((!IsSExt && BinOp->hasNoUnsignedWrap()) ||
3344  (IsSExt && BinOp->hasNoSignedWrap())))
3345  return true;
3346 
3347  // Check if we can do the following simplification.
3348  // ext(trunc(opnd)) --> ext(opnd)
3349  if (!isa<TruncInst>(Inst))
3350  return false;
3351 
3352  Value *OpndVal = Inst->getOperand(0);
3353  // Check if we can use this operand in the extension.
3354  // If the type is larger than the result type of the extension, we cannot.
3355  if (!OpndVal->getType()->isIntegerTy() ||
3356  OpndVal->getType()->getIntegerBitWidth() >
3357  ConsideredExtType->getIntegerBitWidth())
3358  return false;
3359 
3360  // If the operand of the truncate is not an instruction, we will not have
3361  // any information on the dropped bits.
3362  // (Actually we could for constant but it is not worth the extra logic).
3363  Instruction *Opnd = dyn_cast<Instruction>(OpndVal);
3364  if (!Opnd)
3365  return false;
3366 
3367  // Check if the source of the type is narrow enough.
3368  // I.e., check that trunc just drops extended bits of the same kind of
3369  // the extension.
3370  // #1 get the type of the operand and check the kind of the extended bits.
3371  const Type *OpndType;
3372  InstrToOrigTy::const_iterator It = PromotedInsts.find(Opnd);
3373  if (It != PromotedInsts.end() && It->second.getInt() == IsSExt)
3374  OpndType = It->second.getPointer();
3375  else if ((IsSExt && isa<SExtInst>(Opnd)) || (!IsSExt && isa<ZExtInst>(Opnd)))
3376  OpndType = Opnd->getOperand(0)->getType();
3377  else
3378  return false;
3379 
3380  // #2 check that the truncate just drops extended bits.
3381  return Inst->getType()->getIntegerBitWidth() >=
3382  OpndType->getIntegerBitWidth();
3383 }
3384 
3385 TypePromotionHelper::Action TypePromotionHelper::getAction(
3386  Instruction *Ext, const SetOfInstrs &InsertedInsts,
3387  const TargetLowering &TLI, const InstrToOrigTy &PromotedInsts) {
3388  assert((isa<SExtInst>(Ext) || isa<ZExtInst>(Ext)) &&
3389  "Unexpected instruction type");
3390  Instruction *ExtOpnd = dyn_cast<Instruction>(Ext->getOperand(0));
3391  Type *ExtTy = Ext->getType();
3392  bool IsSExt = isa<SExtInst>(Ext);
3393  // If the operand of the extension is not an instruction, we cannot
3394  // get through.
3395  // If it, check we can get through.
3396  if (!ExtOpnd || !canGetThrough(ExtOpnd, ExtTy, PromotedInsts, IsSExt))
3397  return nullptr;
3398 
3399  // Do not promote if the operand has been added by codegenprepare.
3400  // Otherwise, it means we are undoing an optimization that is likely to be
3401  // redone, thus causing potential infinite loop.
3402  if (isa<TruncInst>(ExtOpnd) && InsertedInsts.count(ExtOpnd))
3403  return nullptr;
3404 
3405  // SExt or Trunc instructions.
3406  // Return the related handler.
3407  if (isa<SExtInst>(ExtOpnd) || isa<TruncInst>(ExtOpnd) ||
3408  isa<ZExtInst>(ExtOpnd))
3409  return promoteOperandForTruncAndAnyExt;
3410 
3411  // Regular instruction.
3412  // Abort early if we will have to insert non-free instructions.
3413  if (!ExtOpnd->hasOneUse() && !TLI.isTruncateFree(ExtTy, ExtOpnd->getType()))
3414  return nullptr;
3415  return IsSExt ? signExtendOperandForOther : zeroExtendOperandForOther;
3416 }
3417 
3418 Value *TypePromotionHelper::promoteOperandForTruncAndAnyExt(
3419  Instruction *SExt, TypePromotionTransaction &TPT,
3420  InstrToOrigTy &PromotedInsts, unsigned &CreatedInstsCost,
3422  SmallVectorImpl<Instruction *> *Truncs, const TargetLowering &TLI) {
3423  // By construction, the operand of SExt is an instruction. Otherwise we cannot
3424  // get through it and this method should not be called.
3425  Instruction *SExtOpnd = cast<Instruction>(SExt->getOperand(0));
3426  Value *ExtVal = SExt;
3427  bool HasMergedNonFreeExt = false;
3428  if (isa<ZExtInst>(SExtOpnd)) {
3429  // Replace s|zext(zext(opnd))
3430  // => zext(opnd).
3431  HasMergedNonFreeExt = !TLI.isExtFree(SExtOpnd);
3432  Value *ZExt =
3433  TPT.createZExt(SExt, SExtOpnd->getOperand(0), SExt->getType());
3434  TPT.replaceAllUsesWith(SExt, ZExt);
3435  TPT.eraseInstruction(SExt);
3436  ExtVal = ZExt;
3437  } else {
3438  // Replace z|sext(trunc(opnd)) or sext(sext(opnd))
3439  // => z|sext(opnd).
3440  TPT.setOperand(SExt, 0, SExtOpnd->getOperand(0));
3441  }
3442  CreatedInstsCost = 0;
3443 
3444  // Remove dead code.
3445  if (SExtOpnd->use_empty())
3446  TPT.eraseInstruction(SExtOpnd);
3447 
3448  // Check if the extension is still needed.
3449  Instruction *ExtInst = dyn_cast<Instruction>(ExtVal);
3450  if (!ExtInst || ExtInst->getType() != ExtInst->getOperand(0)->getType()) {
3451  if (ExtInst) {
3452  if (Exts)
3453  Exts->push_back(ExtInst);
3454  CreatedInstsCost = !TLI.isExtFree(ExtInst) && !HasMergedNonFreeExt;
3455  }
3456  return ExtVal;
3457  }
3458 
3459  // At this point we have: ext ty opnd to ty.
3460  // Reassign the uses of ExtInst to the opnd and remove ExtInst.
3461  Value *NextVal = ExtInst->getOperand(0);
3462  TPT.eraseInstruction(ExtInst, NextVal);
3463  return NextVal;
3464 }
3465 
3466 Value *TypePromotionHelper::promoteOperandForOther(
3467  Instruction *Ext, TypePromotionTransaction &TPT,
3468  InstrToOrigTy &PromotedInsts, unsigned &CreatedInstsCost,
3470  SmallVectorImpl<Instruction *> *Truncs, const TargetLowering &TLI,
3471  bool IsSExt) {
3472  // By construction, the operand of Ext is an instruction. Otherwise we cannot
3473  // get through it and this method should not be called.
3474  Instruction *ExtOpnd = cast<Instruction>(Ext->getOperand(0));
3475  CreatedInstsCost = 0;
3476  if (!ExtOpnd->hasOneUse()) {
3477  // ExtOpnd will be promoted.
3478  // All its uses, but Ext, will need to use a truncated value of the
3479  // promoted version.
3480  // Create the truncate now.
3481  Value *Trunc = TPT.createTrunc(Ext, ExtOpnd->getType());
3482  if (Instruction *ITrunc = dyn_cast<Instruction>(Trunc)) {
3483  // Insert it just after the definition.
3484  ITrunc->moveAfter(ExtOpnd);
3485  if (Truncs)
3486  Truncs->push_back(ITrunc);
3487  }
3488 
3489  TPT.replaceAllUsesWith(ExtOpnd, Trunc);
3490  // Restore the operand of Ext (which has been replaced by the previous call
3491  // to replaceAllUsesWith) to avoid creating a cycle trunc <-> sext.
3492  TPT.setOperand(Ext, 0, ExtOpnd);
3493  }
3494 
3495  // Get through the Instruction:
3496  // 1. Update its type.
3497  // 2. Replace the uses of Ext by Inst.
3498  // 3. Extend each operand that needs to be extended.
3499 
3500  // Remember the original type of the instruction before promotion.
3501  // This is useful to know that the high bits are sign extended bits.
3502  PromotedInsts.insert(std::pair<Instruction *, TypeIsSExt>(
3503  ExtOpnd, TypeIsSExt(ExtOpnd->getType(), IsSExt)));
3504  // Step #1.
3505  TPT.mutateType(ExtOpnd, Ext->getType());
3506  // Step #2.
3507  TPT.replaceAllUsesWith(Ext, ExtOpnd);
3508  // Step #3.
3509  Instruction *ExtForOpnd = Ext;
3510 
3511  DEBUG(dbgs() << "Propagate Ext to operands\n");
3512  for (int OpIdx = 0, EndOpIdx = ExtOpnd->getNumOperands(); OpIdx != EndOpIdx;
3513  ++OpIdx) {
3514  DEBUG(dbgs() << "Operand:\n" << *(ExtOpnd->getOperand(OpIdx)) << '\n');
3515  if (ExtOpnd->getOperand(OpIdx)->getType() == Ext->getType() ||
3516  !shouldExtOperand(ExtOpnd, OpIdx)) {
3517  DEBUG(dbgs() << "No need to propagate\n");
3518  continue;
3519  }
3520  // Check if we can statically extend the operand.
3521  Value *Opnd = ExtOpnd->getOperand(OpIdx);
3522  if (const ConstantInt *Cst = dyn_cast<ConstantInt>(Opnd)) {
3523  DEBUG(dbgs() << "Statically extend\n");
3524  unsigned BitWidth = Ext->getType()->getIntegerBitWidth();
3525  APInt CstVal = IsSExt ? Cst->getValue().sext(BitWidth)
3526  : Cst->getValue().zext(BitWidth);
3527  TPT.setOperand(ExtOpnd, OpIdx, ConstantInt::get(Ext->getType(), CstVal));
3528  continue;
3529  }
3530  // UndefValue are typed, so we have to statically sign extend them.
3531  if (isa<UndefValue>(Opnd)) {
3532  DEBUG(dbgs() << "Statically extend\n");
3533  TPT.setOperand(ExtOpnd, OpIdx, UndefValue::get(Ext->getType()));
3534  continue;
3535  }
3536 
3537  // Otherwise we have to explicity sign extend the operand.
3538  // Check if Ext was reused to extend an operand.
3539  if (!ExtForOpnd) {
3540  // If yes, create a new one.
3541  DEBUG(dbgs() << "More operands to ext\n");
3542  Value *ValForExtOpnd = IsSExt ? TPT.createSExt(Ext, Opnd, Ext->getType())
3543  : TPT.createZExt(Ext, Opnd, Ext->getType());
3544  if (!isa<Instruction>(ValForExtOpnd)) {
3545  TPT.setOperand(ExtOpnd, OpIdx, ValForExtOpnd);
3546  continue;
3547  }
3548  ExtForOpnd = cast<Instruction>(ValForExtOpnd);
3549  }
3550  if (Exts)
3551  Exts->push_back(ExtForOpnd);
3552  TPT.setOperand(ExtForOpnd, 0, Opnd);
3553 
3554  // Move the sign extension before the insertion point.
3555  TPT.moveBefore(ExtForOpnd, ExtOpnd);
3556  TPT.setOperand(ExtOpnd, OpIdx, ExtForOpnd);
3557  CreatedInstsCost += !TLI.isExtFree(ExtForOpnd);
3558  // If more sext are required, new instructions will have to be created.
3559  ExtForOpnd = nullptr;
3560  }
3561  if (ExtForOpnd == Ext) {
3562  DEBUG(dbgs() << "Extension is useless now\n");
3563  TPT.eraseInstruction(Ext);
3564  }
3565  return ExtOpnd;
3566 }
3567 
3568 /// Check whether or not promoting an instruction to a wider type is profitable.
3569 /// \p NewCost gives the cost of extension instructions created by the
3570 /// promotion.
3571 /// \p OldCost gives the cost of extension instructions before the promotion
3572 /// plus the number of instructions that have been
3573 /// matched in the addressing mode the promotion.
3574 /// \p PromotedOperand is the value that has been promoted.
3575 /// \return True if the promotion is profitable, false otherwise.
3576 bool AddressingModeMatcher::isPromotionProfitable(
3577  unsigned NewCost, unsigned OldCost, Value *PromotedOperand) const {
3578  DEBUG(dbgs() << "OldCost: " << OldCost << "\tNewCost: " << NewCost << '\n');
3579  // The cost of the new extensions is greater than the cost of the
3580  // old extension plus what we folded.
3581  // This is not profitable.
3582  if (NewCost > OldCost)
3583  return false;
3584  if (NewCost < OldCost)
3585  return true;
3586  // The promotion is neutral but it may help folding the sign extension in
3587  // loads for instance.
3588  // Check that we did not create an illegal instruction.
3589  return isPromotedInstructionLegal(TLI, DL, PromotedOperand);
3590 }
3591 
3592 /// Given an instruction or constant expr, see if we can fold the operation
3593 /// into the addressing mode. If so, update the addressing mode and return
3594 /// true, otherwise return false without modifying AddrMode.
3595 /// If \p MovedAway is not NULL, it contains the information of whether or
3596 /// not AddrInst has to be folded into the addressing mode on success.
3597 /// If \p MovedAway == true, \p AddrInst will not be part of the addressing
3598 /// because it has been moved away.
3599 /// Thus AddrInst must not be added in the matched instructions.
3600 /// This state can happen when AddrInst is a sext, since it may be moved away.
3601 /// Therefore, AddrInst may not be valid when MovedAway is true and it must
3602 /// not be referenced anymore.
3603 bool AddressingModeMatcher::matchOperationAddr(User *AddrInst, unsigned Opcode,
3604  unsigned Depth,
3605  bool *MovedAway) {
3606  // Avoid exponential behavior on extremely deep expression trees.
3607  if (Depth >= 5) return false;
3608 
3609  // By default, all matched instructions stay in place.
3610  if (MovedAway)
3611  *MovedAway = false;
3612 
3613  switch (Opcode) {
3614  case Instruction::PtrToInt:
3615  // PtrToInt is always a noop, as we know that the int type is pointer sized.
3616  return matchAddr(AddrInst->getOperand(0), Depth);
3617  case Instruction::IntToPtr: {
3618  auto AS = AddrInst->getType()->getPointerAddressSpace();
3619  auto PtrTy = MVT::getIntegerVT(DL.getPointerSizeInBits(AS));
3620  // This inttoptr is a no-op if the integer type is pointer sized.
3621  if (TLI.getValueType(DL, AddrInst->getOperand(0)->getType()) == PtrTy)
3622  return matchAddr(AddrInst->getOperand(0), Depth);
3623  return false;
3624  }
3625  case Instruction::BitCast:
3626  // BitCast is always a noop, and we can handle it as long as it is
3627  // int->int or pointer->pointer (we don't want int<->fp or something).
3628  if ((AddrInst->getOperand(0)->getType()->isPointerTy() ||
3629  AddrInst->getOperand(0)->getType()->isIntegerTy()) &&
3630  // Don't touch identity bitcasts. These were probably put here by LSR,
3631  // and we don't want to mess around with them. Assume it knows what it
3632  // is doing.
3633  AddrInst->getOperand(0)->getType() != AddrInst->getType())
3634  return matchAddr(AddrInst->getOperand(0), Depth);
3635  return false;
3636  case Instruction::AddrSpaceCast: {
3637  unsigned SrcAS
3638  = AddrInst->getOperand(0)->getType()->getPointerAddressSpace();
3639  unsigned DestAS = AddrInst->getType()->getPointerAddressSpace();
3640  if (TLI.isNoopAddrSpaceCast(SrcAS, DestAS))
3641  return matchAddr(AddrInst->getOperand(0), Depth);
3642  return false;
3643  }
3644  case Instruction::Add: {
3645  // Check to see if we can merge in the RHS then the LHS. If so, we win.
3646  ExtAddrMode BackupAddrMode = AddrMode;
3647  unsigned OldSize = AddrModeInsts.size();
3648  // Start a transaction at this point.
3649  // The LHS may match but not the RHS.
3650  // Therefore, we need a higher level restoration point to undo partially
3651  // matched operation.
3652  TypePromotionTransaction::ConstRestorationPt LastKnownGood =
3653  TPT.getRestorationPoint();
3654 
3655  if (matchAddr(AddrInst->getOperand(1), Depth+1) &&
3656  matchAddr(AddrInst->getOperand(0), Depth+1))
3657  return true;
3658 
3659  // Restore the old addr mode info.
3660  AddrMode = BackupAddrMode;
3661  AddrModeInsts.resize(OldSize);
3662  TPT.rollback(LastKnownGood);
3663 
3664  // Otherwise this was over-aggressive. Try merging in the LHS then the RHS.
3665  if (matchAddr(AddrInst->getOperand(0), Depth+1) &&
3666  matchAddr(AddrInst->getOperand(1), Depth+1))
3667  return true;
3668 
3669  // Otherwise we definitely can't merge the ADD in.
3670  AddrMode = BackupAddrMode;
3671  AddrModeInsts.resize(OldSize);
3672  TPT.rollback(LastKnownGood);
3673  break;
3674  }
3675  //case Instruction::Or:
3676  // TODO: We can handle "Or Val, Imm" iff this OR is equivalent to an ADD.
3677  //break;
3678  case Instruction::Mul:
3679  case Instruction::Shl: {
3680  // Can only handle X*C and X << C.
3681  ConstantInt *RHS = dyn_cast<ConstantInt>(AddrInst->getOperand(1));
3682  if (!RHS || RHS->getBitWidth() > 64)
3683  return false;
3684  int64_t Scale = RHS->getSExtValue();
3685  if (Opcode == Instruction::Shl)
3686  Scale = 1LL << Scale;
3687 
3688  return matchScaledValue(AddrInst->getOperand(0), Scale, Depth);
3689  }
3690  case Instruction::GetElementPtr: {
3691  // Scan the GEP. We check it if it contains constant offsets and at most
3692  // one variable offset.
3693  int VariableOperand = -1;
3694  unsigned VariableScale = 0;
3695 
3696  int64_t ConstantOffset = 0;
3697  gep_type_iterator GTI = gep_type_begin(AddrInst);
3698  for (unsigned i = 1, e = AddrInst->getNumOperands(); i != e; ++i, ++GTI) {
3699  if (StructType *STy = GTI.getStructTypeOrNull()) {
3700  const StructLayout *SL = DL.getStructLayout(STy);
3701  unsigned Idx =
3702  cast<ConstantInt>(AddrInst->getOperand(i))->getZExtValue();
3703  ConstantOffset += SL->getElementOffset(Idx);
3704  } else {
3705  uint64_t TypeSize = DL.getTypeAllocSize(GTI.getIndexedType());
3706  if (ConstantInt *CI = dyn_cast<ConstantInt>(AddrInst->getOperand(i))) {
3707  ConstantOffset += CI->getSExtValue()*TypeSize;
3708  } else if (TypeSize) { // Scales of zero don't do anything.
3709  // We only allow one variable index at the moment.
3710  if (VariableOperand != -1)
3711  return false;
3712 
3713  // Remember the variable index.
3714  VariableOperand = i;
3715  VariableScale = TypeSize;
3716  }
3717  }
3718  }
3719 
3720  // A common case is for the GEP to only do a constant offset. In this case,
3721  // just add it to the disp field and check validity.
3722  if (VariableOperand == -1) {
3723  AddrMode.BaseOffs += ConstantOffset;
3724  if (ConstantOffset == 0 ||
3725  TLI.isLegalAddressingMode(DL, AddrMode, AccessTy, AddrSpace)) {
3726  // Check to see if we can fold the base pointer in too.
3727  if (matchAddr(AddrInst->getOperand(0), Depth+1))
3728  return true;
3729  }
3730  AddrMode.BaseOffs -= ConstantOffset;
3731  return false;
3732  }
3733 
3734  // Save the valid addressing mode in case we can't match.
3735  ExtAddrMode BackupAddrMode = AddrMode;
3736  unsigned OldSize = AddrModeInsts.size();
3737 
3738  // See if the scale and offset amount is valid for this target.
3739  AddrMode.BaseOffs += ConstantOffset;
3740 
3741  // Match the base operand of the GEP.
3742  if (!matchAddr(AddrInst->getOperand(0), Depth+1)) {
3743  // If it couldn't be matched, just stuff the value in a register.
3744  if (AddrMode.HasBaseReg) {
3745  AddrMode = BackupAddrMode;
3746  AddrModeInsts.resize(OldSize);
3747  return false;
3748  }
3749  AddrMode.HasBaseReg = true;
3750  AddrMode.BaseReg = AddrInst->getOperand(0);
3751  }
3752 
3753  // Match the remaining variable portion of the GEP.
3754  if (!matchScaledValue(AddrInst->getOperand(VariableOperand), VariableScale,
3755  Depth)) {
3756  // If it couldn't be matched, try stuffing the base into a register
3757  // instead of matching it, and retrying the match of the scale.
3758  AddrMode = BackupAddrMode;
3759  AddrModeInsts.resize(OldSize);
3760  if (AddrMode.HasBaseReg)
3761  return false;
3762  AddrMode.HasBaseReg = true;
3763  AddrMode.BaseReg = AddrInst->getOperand(0);
3764  AddrMode.BaseOffs += ConstantOffset;
3765  if (!matchScaledValue(AddrInst->getOperand(VariableOperand),
3766  VariableScale, Depth)) {
3767  // If even that didn't work, bail.
3768  AddrMode = BackupAddrMode;
3769  AddrModeInsts.resize(OldSize);
3770  return false;
3771  }
3772  }
3773 
3774  return true;
3775  }
3776  case Instruction::SExt:
3777  case Instruction::ZExt: {
3778  Instruction *Ext = dyn_cast<Instruction>(AddrInst);
3779  if (!Ext)
3780  return false;
3781 
3782  // Try to move this ext out of the way of the addressing mode.
3783  // Ask for a method for doing so.
3784  TypePromotionHelper::Action TPH =
3785  TypePromotionHelper::getAction(Ext, InsertedInsts, TLI, PromotedInsts);
3786  if (!TPH)
3787  return false;
3788 
3789  TypePromotionTransaction::ConstRestorationPt LastKnownGood =
3790  TPT.getRestorationPoint();
3791  unsigned CreatedInstsCost = 0;
3792  unsigned ExtCost = !TLI.isExtFree(Ext);
3793  Value *PromotedOperand =
3794  TPH(Ext, TPT, PromotedInsts, CreatedInstsCost, nullptr, nullptr, TLI);
3795  // SExt has been moved away.
3796  // Thus either it will be rematched later in the recursive calls or it is
3797  // gone. Anyway, we must not fold it into the addressing mode at this point.
3798  // E.g.,
3799  // op = add opnd, 1
3800  // idx = ext op
3801  // addr = gep base, idx
3802  // is now:
3803  // promotedOpnd = ext opnd <- no match here
3804  // op = promoted_add promotedOpnd, 1 <- match (later in recursive calls)
3805  // addr = gep base, op <- match
3806  if (MovedAway)
3807  *MovedAway = true;
3808 
3809  assert(PromotedOperand &&
3810  "TypePromotionHelper should have filtered out those cases");
3811 
3812  ExtAddrMode BackupAddrMode = AddrMode;
3813  unsigned OldSize = AddrModeInsts.size();
3814 
3815  if (!matchAddr(PromotedOperand, Depth) ||
3816  // The total of the new cost is equal to the cost of the created
3817  // instructions.
3818  // The total of the old cost is equal to the cost of the extension plus
3819  // what we have saved in the addressing mode.
3820  !isPromotionProfitable(CreatedInstsCost,
3821  ExtCost + (AddrModeInsts.size() - OldSize),
3822  PromotedOperand)) {
3823  AddrMode = BackupAddrMode;
3824  AddrModeInsts.resize(OldSize);
3825  DEBUG(dbgs() << "Sign extension does not pay off: rollback\n");
3826  TPT.rollback(LastKnownGood);
3827  return false;
3828  }
3829  return true;
3830  }
3831  }
3832  return false;
3833 }
3834 
3835 /// If we can, try to add the value of 'Addr' into the current addressing mode.
3836 /// If Addr can't be added to AddrMode this returns false and leaves AddrMode
3837 /// unmodified. This assumes that Addr is either a pointer type or intptr_t
3838 /// for the target.
3839 ///
3840 bool AddressingModeMatcher::matchAddr(Value *Addr, unsigned Depth) {
3841  // Start a transaction at this point that we will rollback if the matching
3842  // fails.
3843  TypePromotionTransaction::ConstRestorationPt LastKnownGood =
3844  TPT.getRestorationPoint();
3845  if (ConstantInt *CI = dyn_cast<ConstantInt>(Addr)) {
3846  // Fold in immediates if legal for the target.
3847  AddrMode.BaseOffs += CI->getSExtValue();
3848  if (TLI.isLegalAddressingMode(DL, AddrMode, AccessTy, AddrSpace))
3849  return true;
3850  AddrMode.BaseOffs -= CI->getSExtValue();
3851  } else if (GlobalValue *GV = dyn_cast<GlobalValue>(Addr)) {
3852  // If this is a global variable, try to fold it into the addressing mode.
3853  if (!AddrMode.BaseGV) {
3854  AddrMode.BaseGV = GV;
3855  if (TLI.isLegalAddressingMode(DL, AddrMode, AccessTy, AddrSpace))
3856  return true;
3857  AddrMode.BaseGV = nullptr;
3858  }
3859  } else if (Instruction *I = dyn_cast<Instruction>(Addr)) {
3860  ExtAddrMode BackupAddrMode = AddrMode;
3861  unsigned OldSize = AddrModeInsts.size();
3862 
3863  // Check to see if it is possible to fold this operation.
3864  bool MovedAway = false;
3865  if (matchOperationAddr(I, I->getOpcode(), Depth, &MovedAway)) {
3866  // This instruction may have been moved away. If so, there is nothing
3867  // to check here.
3868  if (MovedAway)
3869  return true;
3870  // Okay, it's possible to fold this. Check to see if it is actually
3871  // *profitable* to do so. We use a simple cost model to avoid increasing
3872  // register pressure too much.
3873  if (I->hasOneUse() ||
3874  isProfitableToFoldIntoAddressingMode(I, BackupAddrMode, AddrMode)) {
3875  AddrModeInsts.push_back(I);
3876  return true;
3877  }
3878 
3879  // It isn't profitable to do this, roll back.
3880  //cerr << "NOT FOLDING: " << *I;
3881  AddrMode = BackupAddrMode;
3882  AddrModeInsts.resize(OldSize);
3883  TPT.rollback(LastKnownGood);
3884  }
3885  } else if (ConstantExpr *CE = dyn_cast<ConstantExpr>(Addr)) {
3886  if (matchOperationAddr(CE, CE->getOpcode(), Depth))
3887  return true;
3888  TPT.rollback(LastKnownGood);
3889  } else if (isa<ConstantPointerNull>(Addr)) {
3890  // Null pointer gets folded without affecting the addressing mode.
3891  return true;
3892  }
3893 
3894  // Worse case, the target should support [reg] addressing modes. :)
3895  if (!AddrMode.HasBaseReg) {
3896  AddrMode.HasBaseReg = true;
3897  AddrMode.BaseReg = Addr;
3898  // Still check for legality in case the target supports [imm] but not [i+r].
3899  if (TLI.isLegalAddressingMode(DL, AddrMode, AccessTy, AddrSpace))
3900  return true;
3901  AddrMode.HasBaseReg = false;
3902  AddrMode.BaseReg = nullptr;
3903  }
3904 
3905  // If the base register is already taken, see if we can do [r+r].
3906  if (AddrMode.Scale == 0) {
3907  AddrMode.Scale = 1;
3908  AddrMode.ScaledReg = Addr;
3909  if (TLI.isLegalAddressingMode(DL, AddrMode, AccessTy, AddrSpace))
3910  return true;
3911  AddrMode.Scale = 0;
3912  AddrMode.ScaledReg = nullptr;
3913  }
3914  // Couldn't match.
3915  TPT.rollback(LastKnownGood);
3916  return false;
3917 }
3918 
3919 /// Check to see if all uses of OpVal by the specified inline asm call are due
3920 /// to memory operands. If so, return true, otherwise return false.
3921 static bool IsOperandAMemoryOperand(CallInst *CI, InlineAsm *IA, Value *OpVal,
3922  const TargetLowering &TLI,
3923  const TargetRegisterInfo &TRI) {
3924  const Function *F = CI->getFunction();
3925  TargetLowering::AsmOperandInfoVector TargetConstraints =
3926  TLI.ParseConstraints(F->getParent()->getDataLayout(), &TRI,
3927  ImmutableCallSite(CI));
3928 
3929  for (unsigned i = 0, e = TargetConstraints.size(); i != e; ++i) {
3930  TargetLowering::AsmOperandInfo &OpInfo = TargetConstraints[i];
3931 
3932  // Compute the constraint code and ConstraintType to use.
3933  TLI.ComputeConstraintToUse(OpInfo, SDValue());
3934 
3935  // If this asm operand is our Value*, and if it isn't an indirect memory
3936  // operand, we can't fold it!
3937  if (OpInfo.CallOperandVal == OpVal &&
3939  !OpInfo.isIndirect))
3940  return false;
3941  }
3942 
3943  return true;
3944 }
3945 
3946 // Max number of memory uses to look at before aborting the search to conserve
3947 // compile time.
3948 static constexpr int MaxMemoryUsesToScan = 20;
3949 
3950 /// Recursively walk all the uses of I until we find a memory use.
3951 /// If we find an obviously non-foldable instruction, return true.
3952 /// Add the ultimately found memory instructions to MemoryUses.
3953 static bool FindAllMemoryUses(
3954  Instruction *I,
3955  SmallVectorImpl<std::pair<Instruction *, unsigned>> &MemoryUses,
3956  SmallPtrSetImpl<Instruction *> &ConsideredInsts, const TargetLowering &TLI,
3957  const TargetRegisterInfo &TRI, int SeenInsts = 0) {
3958  // If we already considered this instruction, we're done.
3959  if (!ConsideredInsts.insert(I).second)
3960  return false;
3961 
3962  // If this is an obviously unfoldable instruction, bail out.
3963  if (!MightBeFoldableInst(I))
3964  return true;
3965 
3966  const bool OptSize = I->getFunction()->optForSize();
3967 
3968  // Loop over all the uses, recursively processing them.
3969  for (Use &U : I->uses()) {
3970  // Conservatively return true if we're seeing a large number or a deep chain
3971  // of users. This avoids excessive compilation times in pathological cases.
3972  if (SeenInsts++ >= MaxMemoryUsesToScan)
3973  return true;
3974 
3975  Instruction *UserI = cast<Instruction>(U.getUser());
3976  if (LoadInst *LI = dyn_cast<LoadInst>(UserI)) {
3977  MemoryUses.push_back(std::make_pair(LI, U.getOperandNo()));
3978  continue;
3979  }
3980 
3981  if (StoreInst *SI = dyn_cast<StoreInst>(UserI)) {
3982  unsigned opNo = U.getOperandNo();
3983  if (opNo != StoreInst::getPointerOperandIndex())
3984  return true; // Storing addr, not into addr.
3985  MemoryUses.push_back(std::make_pair(SI, opNo));
3986  continue;
3987  }
3988 
3989  if (AtomicRMWInst *RMW = dyn_cast<AtomicRMWInst>(UserI)) {
3990  unsigned opNo = U.getOperandNo();
3992  return true; // Storing addr, not into addr.
3993  MemoryUses.push_back(std::make_pair(RMW, opNo));
3994  continue;
3995  }
3996 
3997  if (AtomicCmpXchgInst *CmpX = dyn_cast<AtomicCmpXchgInst>(UserI)) {
3998  unsigned opNo = U.getOperandNo();
4000  return true; // Storing addr, not into addr.
4001  MemoryUses.push_back(std::make_pair(CmpX, opNo));
4002  continue;
4003  }
4004 
4005  if (CallInst *CI = dyn_cast<CallInst>(UserI)) {
4006  // If this is a cold call, we can sink the addressing calculation into
4007  // the cold path. See optimizeCallInst
4008  if (!OptSize && CI->hasFnAttr(Attribute::Cold))
4009  continue;
4010 
4012  if (!IA) return true;
4013 
4014  // If this is a memory operand, we're cool, otherwise bail out.
4015  if (!IsOperandAMemoryOperand(CI, IA, I, TLI, TRI))
4016  return true;
4017  continue;
4018  }
4019 
4020  if (FindAllMemoryUses(UserI, MemoryUses, ConsideredInsts, TLI, TRI,
4021  SeenInsts))
4022  return true;
4023  }
4024 
4025  return false;
4026 }
4027 
4028 /// Return true if Val is already known to be live at the use site that we're
4029 /// folding it into. If so, there is no cost to include it in the addressing
4030 /// mode. KnownLive1 and KnownLive2 are two values that we know are live at the
4031 /// instruction already.
4032 bool AddressingModeMatcher::valueAlreadyLiveAtInst(Value *Val,Value *KnownLive1,
4033  Value *KnownLive2) {
4034  // If Val is either of the known-live values, we know it is live!
4035  if (Val == nullptr || Val == KnownLive1 || Val == KnownLive2)
4036  return true;
4037 
4038  // All values other than instructions and arguments (e.g. constants) are live.
4039  if (!isa<Instruction>(Val) && !isa<Argument>(Val)) return true;
4040 
4041  // If Val is a constant sized alloca in the entry block, it is live, this is
4042  // true because it is just a reference to the stack/frame pointer, which is
4043  // live for the whole function.
4044  if (AllocaInst *AI = dyn_cast<AllocaInst>(Val))
4045  if (AI->isStaticAlloca())
4046  return true;
4047 
4048  // Check to see if this value is already used in the memory instruction's
4049  // block. If so, it's already live into the block at the very least, so we
4050  // can reasonably fold it.
4051  return Val->isUsedInBasicBlock(MemoryInst->getParent());
4052 }
4053 
4054 /// It is possible for the addressing mode of the machine to fold the specified
4055 /// instruction into a load or store that ultimately uses it.
4056 /// However, the specified instruction has multiple uses.
4057 /// Given this, it may actually increase register pressure to fold it
4058 /// into the load. For example, consider this code:
4059 ///
4060 /// X = ...
4061 /// Y = X+1
4062 /// use(Y) -> nonload/store
4063 /// Z = Y+1
4064 /// load Z
4065 ///
4066 /// In this case, Y has multiple uses, and can be folded into the load of Z
4067 /// (yielding load [X+2]). However, doing this will cause both "X" and "X+1" to
4068 /// be live at the use(Y) line. If we don't fold Y into load Z, we use one
4069 /// fewer register. Since Y can't be folded into "use(Y)" we don't increase the
4070 /// number of computations either.
4071 ///
4072 /// Note that this (like most of CodeGenPrepare) is just a rough heuristic. If
4073 /// X was live across 'load Z' for other reasons, we actually *would* want to
4074 /// fold the addressing mode in the Z case. This would make Y die earlier.
4075 bool AddressingModeMatcher::
4076 isProfitableToFoldIntoAddressingMode(Instruction *I, ExtAddrMode &AMBefore,
4077  ExtAddrMode &AMAfter) {
4078  if (IgnoreProfitability) return true;
4079 
4080  // AMBefore is the addressing mode before this instruction was folded into it,
4081  // and AMAfter is the addressing mode after the instruction was folded. Get
4082  // the set of registers referenced by AMAfter and subtract out those
4083  // referenced by AMBefore: this is the set of values which folding in this
4084  // address extends the lifetime of.
4085  //
4086  // Note that there are only two potential values being referenced here,
4087  // BaseReg and ScaleReg (global addresses are always available, as are any
4088  // folded immediates).
4089  Value *BaseReg = AMAfter.BaseReg, *ScaledReg = AMAfter.ScaledReg;
4090 
4091  // If the BaseReg or ScaledReg was referenced by the previous addrmode, their
4092  // lifetime wasn't extended by adding this instruction.
4093  if (valueAlreadyLiveAtInst(BaseReg, AMBefore.BaseReg, AMBefore.ScaledReg))
4094  BaseReg = nullptr;
4095  if (valueAlreadyLiveAtInst(ScaledReg, AMBefore.BaseReg, AMBefore.ScaledReg))
4096  ScaledReg = nullptr;
4097 
4098  // If folding this instruction (and it's subexprs) didn't extend any live
4099  // ranges, we're ok with it.
4100  if (!BaseReg && !ScaledReg)
4101  return true;
4102 
4103  // If all uses of this instruction can have the address mode sunk into them,
4104  // we can remove the addressing mode and effectively trade one live register
4105  // for another (at worst.) In this context, folding an addressing mode into
4106  // the use is just a particularly nice way of sinking it.
4108  SmallPtrSet<Instruction*, 16> ConsideredInsts;
4109  if (FindAllMemoryUses(I, MemoryUses, ConsideredInsts, TLI, TRI))
4110  return false; // Has a non-memory, non-foldable use!
4111 
4112  // Now that we know that all uses of this instruction are part of a chain of
4113  // computation involving only operations that could theoretically be folded
4114  // into a memory use, loop over each of these memory operation uses and see
4115  // if they could *actually* fold the instruction. The assumption is that
4116  // addressing modes are cheap and that duplicating the computation involved
4117  // many times is worthwhile, even on a fastpath. For sinking candidates
4118  // (i.e. cold call sites), this serves as a way to prevent excessive code
4119  // growth since most architectures have some reasonable small and fast way to
4120  // compute an effective address. (i.e LEA on x86)
4121  SmallVector<Instruction*, 32> MatchedAddrModeInsts;
4122  for (unsigned i = 0, e = MemoryUses.size(); i != e; ++i) {
4123  Instruction *User = MemoryUses[i].first;
4124  unsigned OpNo = MemoryUses[i].second;
4125 
4126  // Get the access type of this use. If the use isn't a pointer, we don't
4127  // know what it accesses.
4128  Value *Address = User->getOperand(OpNo);
4129  PointerType *AddrTy = dyn_cast<PointerType>(Address->getType());
4130  if (!AddrTy)
4131  return false;
4132  Type *AddressAccessTy = AddrTy->getElementType();
4133  unsigned AS = AddrTy->getAddressSpace();
4134 
4135  // Do a match against the root of this address, ignoring profitability. This
4136  // will tell us if the addressing mode for the memory operation will
4137  // *actually* cover the shared instruction.
4138  ExtAddrMode Result;
4139  TypePromotionTransaction::ConstRestorationPt LastKnownGood =
4140  TPT.getRestorationPoint();
4141  AddressingModeMatcher Matcher(MatchedAddrModeInsts, TLI, TRI,
4142  AddressAccessTy, AS,
4143  MemoryInst, Result, InsertedInsts,
4144  PromotedInsts, TPT);
4145  Matcher.IgnoreProfitability = true;
4146  bool Success = Matcher.matchAddr(Address, 0);
4147  (void)Success; assert(Success && "Couldn't select *anything*?");
4148 
4149  // The match was to check the profitability, the changes made are not
4150  // part of the original matcher. Therefore, they should be dropped
4151  // otherwise the original matcher will not present the right state.
4152  TPT.rollback(LastKnownGood);
4153 
4154  // If the match didn't cover I, then it won't be shared by it.
4155  if (!is_contained(MatchedAddrModeInsts, I))
4156  return false;
4157 
4158  MatchedAddrModeInsts.clear();
4159  }
4160 
4161  return true;
4162 }
4163 
4164 /// Return true if the specified values are defined in a
4165 /// different basic block than BB.
4166 static bool IsNonLocalValue(Value *V, BasicBlock *BB) {
4167  if (Instruction *I = dyn_cast<Instruction>(V))
4168  return I->getParent() != BB;
4169  return false;
4170 }
4171 
4172 /// Sink addressing mode computation immediate before MemoryInst if doing so
4173 /// can be done without increasing register pressure. The need for the
4174 /// register pressure constraint means this can end up being an all or nothing
4175 /// decision for all uses of the same addressing computation.
4176 ///
4177 /// Load and Store Instructions often have addressing modes that can do
4178 /// significant amounts of computation. As such, instruction selection will try
4179 /// to get the load or store to do as much computation as possible for the
4180 /// program. The problem is that isel can only see within a single block. As
4181 /// such, we sink as much legal addressing mode work into the block as possible.
4182 ///
4183 /// This method is used to optimize both load/store and inline asms with memory
4184 /// operands. It's also used to sink addressing computations feeding into cold
4185 /// call sites into their (cold) basic block.
4186 ///
4187 /// The motivation for handling sinking into cold blocks is that doing so can
4188 /// both enable other address mode sinking (by satisfying the register pressure
4189 /// constraint above), and reduce register pressure globally (by removing the
4190 /// addressing mode computation from the fast path entirely.).
4191 bool CodeGenPrepare::optimizeMemoryInst(Instruction *MemoryInst, Value *Addr,
4192  Type *AccessTy, unsigned AddrSpace) {
4193  Value *Repl = Addr;
4194 
4195  // Try to collapse single-value PHI nodes. This is necessary to undo
4196  // unprofitable PRE transformations.
4197  SmallVector<Value*, 8> worklist;
4198  SmallPtrSet<Value*, 16> Visited;
4199  worklist.push_back(Addr);
4200 
4201  // Use a worklist to iteratively look through PHI and select nodes, and
4202  // ensure that the addressing mode obtained from the non-PHI/select roots of
4203  // the graph are compatible.
4204  bool PhiOrSelectSeen = false;
4205  SmallVector<Instruction*, 16> AddrModeInsts;
4206  const SimplifyQuery SQ(*DL, TLInfo);
4207  AddressingModeCombiner AddrModes(SQ, { Addr, MemoryInst->getParent() });
4208  TypePromotionTransaction TPT(RemovedInsts);
4209  TypePromotionTransaction::ConstRestorationPt LastKnownGood =
4210  TPT.getRestorationPoint();
4211  while (!worklist.empty()) {
4212  Value *V = worklist.back();
4213  worklist.pop_back();
4214 
4215  // We allow traversing cyclic Phi nodes.
4216  // In case of success after this loop we ensure that traversing through
4217  // Phi nodes ends up with all cases to compute address of the form
4218  // BaseGV + Base + Scale * Index + Offset
4219  // where Scale and Offset are constans and BaseGV, Base and Index
4220  // are exactly the same Values in all cases.
4221  // It means that BaseGV, Scale and Offset dominate our memory instruction
4222  // and have the same value as they had in address computation represented
4223  // as Phi. So we can safely sink address computation to memory instruction.
4224  if (!Visited.insert(V).second)
4225  continue;
4226 
4227  // For a PHI node, push all of its incoming values.
4228  if (PHINode *P = dyn_cast<PHINode>(V)) {
4229  for (Value *IncValue : P->incoming_values())
4230  worklist.push_back(IncValue);
4231  PhiOrSelectSeen = true;
4232  continue;
4233  }
4234  // Similar for select.
4235  if (SelectInst *SI = dyn_cast<SelectInst>(V)) {
4236  worklist.push_back(SI->getFalseValue());
4237  worklist.push_back(SI->getTrueValue());
4238  PhiOrSelectSeen = true;
4239  continue;
4240  }
4241 
4242  // For non-PHIs, determine the addressing mode being computed. Note that
4243  // the result may differ depending on what other uses our candidate
4244  // addressing instructions might have.
4245  AddrModeInsts.clear();
4246  ExtAddrMode NewAddrMode = AddressingModeMatcher::Match(
4247  V, AccessTy, AddrSpace, MemoryInst, AddrModeInsts, *TLI, *TRI,
4248  InsertedInsts, PromotedInsts, TPT);
4249  NewAddrMode.OriginalValue = V;
4250 
4251  if (!AddrModes.addNewAddrMode(NewAddrMode))
4252  break;
4253  }
4254 
4255  // Try to combine the AddrModes we've collected. If we couldn't collect any,
4256  // or we have multiple but either couldn't combine them or combining them
4257  // wouldn't do anything useful, bail out now.
4258  if (!AddrModes.combineAddrModes()) {
4259  TPT.rollback(LastKnownGood);
4260  return false;
4261  }
4262  TPT.commit();
4263 
4264  // Get the combined AddrMode (or the only AddrMode, if we only had one).
4265  ExtAddrMode AddrMode = AddrModes.getAddrMode();
4266 
4267  // If all the instructions matched are already in this BB, don't do anything.
4268  // If we saw a Phi node then it is not local definitely, and if we saw a select
4269  // then we want to push the address calculation past it even if it's already
4270  // in this BB.
4271  if (!PhiOrSelectSeen && none_of(AddrModeInsts, [&](Value *V) {
4272  return IsNonLocalValue(V, MemoryInst->getParent());
4273  })) {
4274  DEBUG(dbgs() << "CGP: Found local addrmode: " << AddrMode << "\n");
4275  return false;
4276  }
4277 
4278  // Insert this computation right after this user. Since our caller is
4279  // scanning from the top of the BB to the bottom, reuse of the expr are
4280  // guaranteed to happen later.
4281  IRBuilder<> Builder(MemoryInst);
4282 
4283  // Now that we determined the addressing expression we want to use and know
4284  // that we have to sink it into this block. Check to see if we have already
4285  // done this for some other load/store instr in this block. If so, reuse
4286  // the computation. Before attempting reuse, check if the address is valid
4287  // as it may have been erased.
4288 
4289  WeakTrackingVH SunkAddrVH = SunkAddrs[Addr];
4290 
4291  Value * SunkAddr = SunkAddrVH.pointsToAliveValue() ? SunkAddrVH : nullptr;
4292  if (SunkAddr) {
4293  DEBUG(dbgs() << "CGP: Reusing nonlocal addrmode: " << AddrMode << " for "
4294  << *MemoryInst << "\n");
4295  if (SunkAddr->getType() != Addr->getType())
4296  SunkAddr = Builder.CreatePointerCast(SunkAddr, Addr->getType());
4297  } else if (AddrSinkUsingGEPs ||
4298  (!AddrSinkUsingGEPs.getNumOccurrences() && TM &&
4299  SubtargetInfo->useAA())) {
4300  // By default, we use the GEP-based method when AA is used later. This
4301  // prevents new inttoptr/ptrtoint pairs from degrading AA capabilities.
4302  DEBUG(dbgs() << "CGP: SINKING nonlocal addrmode: " << AddrMode << " for "
4303  << *MemoryInst << "\n");
4304  Type *IntPtrTy = DL->getIntPtrType(Addr->getType());
4305  Value *ResultPtr = nullptr, *ResultIndex = nullptr;
4306 
4307  // First, find the pointer.
4308  if (AddrMode.BaseReg && AddrMode.BaseReg->getType()->isPointerTy()) {
4309  ResultPtr = AddrMode.BaseReg;
4310  AddrMode.BaseReg = nullptr;
4311  }
4312 
4313  if (AddrMode.Scale && AddrMode.ScaledReg->getType()->isPointerTy()) {
4314  // We can't add more than one pointer together, nor can we scale a
4315  // pointer (both of which seem meaningless).
4316  if (ResultPtr || AddrMode.Scale != 1)
4317  return false;
4318 
4319  ResultPtr = AddrMode.ScaledReg;
4320  AddrMode.Scale = 0;
4321  }
4322 
4323  // It is only safe to sign extend the BaseReg if we know that the math
4324  // required to create it did not overflow before we extend it. Since
4325  // the original IR value was tossed in favor of a constant back when
4326  // the AddrMode was created we need to bail out gracefully if widths
4327  // do not match instead of extending it.
4328  //
4329  // (See below for code to add the scale.)
4330  if (AddrMode.Scale) {
4331  Type *ScaledRegTy = AddrMode.ScaledReg->getType();
4332  if (cast<IntegerType>(IntPtrTy)->getBitWidth() >
4333  cast<IntegerType>(ScaledRegTy)->getBitWidth())
4334  return false;
4335  }
4336 
4337  if (AddrMode.BaseGV) {
4338  if (ResultPtr)
4339  return false;
4340 
4341  ResultPtr = AddrMode.BaseGV;
4342  }
4343 
4344  // If the real base value actually came from an inttoptr, then the matcher
4345  // will look through it and provide only the integer value. In that case,
4346  // use it here.
4347  if (!DL->isNonIntegralPointerType(Addr->getType())) {
4348  if (!ResultPtr && AddrMode.BaseReg) {
4349  ResultPtr = Builder.CreateIntToPtr(AddrMode.BaseReg, Addr->getType(),
4350  "sunkaddr");
4351  AddrMode.BaseReg = nullptr;
4352  } else if (!ResultPtr && AddrMode.Scale == 1) {
4353  ResultPtr = Builder.CreateIntToPtr(AddrMode.ScaledReg, Addr->getType(),
4354  "sunkaddr");
4355  AddrMode.Scale = 0;
4356  }
4357  }
4358 
4359  if (!ResultPtr &&
4360  !AddrMode.BaseReg && !AddrMode.Scale && !AddrMode.BaseOffs) {
4361  SunkAddr = Constant::getNullValue(Addr->getType());
4362  } else if (!ResultPtr) {
4363  return false;
4364  } else {
4365  Type *I8PtrTy =
4366  Builder.getInt8PtrTy(Addr->getType()->getPointerAddressSpace());
4367  Type *I8Ty = Builder.getInt8Ty();
4368 
4369  // Start with the base register. Do this first so that subsequent address
4370  // matching finds it last, which will prevent it from trying to match it
4371  // as the scaled value in case it happens to be a mul. That would be
4372  // problematic if we've sunk a different mul for the scale, because then
4373  // we'd end up sinking both muls.
4374  if (AddrMode.BaseReg) {
4375  Value *V = AddrMode.BaseReg;
4376  if (V->getType() != IntPtrTy)
4377  V = Builder.CreateIntCast(V, IntPtrTy, /*isSigned=*/true, "sunkaddr");
4378 
4379  ResultIndex = V;
4380  }
4381 
4382  // Add the scale value.
4383  if (AddrMode.Scale) {
4384  Value *V = AddrMode.ScaledReg;
4385  if (V->getType() == IntPtrTy) {
4386  // done.
4387  } else {
4388  assert(cast<IntegerType>(IntPtrTy)->getBitWidth() <
4389  cast<IntegerType>(V->getType())->getBitWidth() &&
4390  "We can't transform if ScaledReg is too narrow");
4391  V = Builder.CreateTrunc(V, IntPtrTy, "sunkaddr");
4392  }
4393 
4394  if (AddrMode.Scale != 1)
4395  V = Builder.CreateMul(V, ConstantInt::get(IntPtrTy, AddrMode.Scale),
4396  "sunkaddr");
4397  if (ResultIndex)
4398  ResultIndex = Builder.CreateAdd(ResultIndex, V, "sunkaddr");
4399  else
4400  ResultIndex = V;
4401  }
4402 
4403  // Add in the Base Offset if present.
4404  if (AddrMode.BaseOffs) {
4405  Value *V = ConstantInt::get(IntPtrTy, AddrMode.BaseOffs);
4406  if (ResultIndex) {
4407  // We need to add this separately from the scale above to help with
4408  // SDAG consecutive load/store merging.
4409  if (ResultPtr->getType() != I8PtrTy)
4410  ResultPtr = Builder.CreatePointerCast(ResultPtr, I8PtrTy);
4411  ResultPtr = Builder.CreateGEP(I8Ty, ResultPtr, ResultIndex, "sunkaddr");
4412  }
4413 
4414  ResultIndex = V;
4415  }
4416 
4417  if (!ResultIndex) {
4418  SunkAddr = ResultPtr;
4419  } else {
4420  if (ResultPtr->getType() != I8PtrTy)
4421  ResultPtr = Builder.CreatePointerCast(ResultPtr, I8PtrTy);
4422  SunkAddr = Builder.CreateGEP(I8Ty, ResultPtr, ResultIndex, "sunkaddr");
4423  }
4424 
4425  if (SunkAddr->getType() != Addr->getType())
4426  SunkAddr = Builder.CreatePointerCast(SunkAddr, Addr->getType());
4427  }
4428  } else {
4429  // We'd require a ptrtoint/inttoptr down the line, which we can't do for
4430  // non-integral pointers, so in that case bail out now.
4431  Type *BaseTy = AddrMode.BaseReg ? AddrMode.BaseReg->getType() : nullptr;
4432  Type *ScaleTy = AddrMode.Scale ? AddrMode.ScaledReg->getType() : nullptr;
4433  PointerType *BasePtrTy = dyn_cast_or_null<PointerType>(BaseTy);
4434  PointerType *ScalePtrTy = dyn_cast_or_null<PointerType>(ScaleTy);
4435  if (DL->isNonIntegralPointerType(Addr->getType()) ||
4436  (BasePtrTy && DL->isNonIntegralPointerType(BasePtrTy)) ||
4437  (ScalePtrTy && DL->isNonIntegralPointerType(ScalePtrTy)) ||
4438  (AddrMode.BaseGV &&
4439  DL->isNonIntegralPointerType(AddrMode.BaseGV->getType())))
4440  return false;
4441 
4442  DEBUG(dbgs() << "CGP: SINKING nonlocal addrmode: " << AddrMode << " for "
4443  << *MemoryInst << "\n");
4444  Type *IntPtrTy = DL->getIntPtrType(Addr->getType());
4445  Value *Result = nullptr;
4446 
4447  // Start with the base register. Do this first so that subsequent address
4448  // matching finds it last, which will prevent it from trying to match it
4449  // as the scaled value in case it happens to be a mul. That would be
4450  // problematic if we've sunk a different mul for the scale, because then
4451  // we'd end up sinking both muls.
4452  if (AddrMode.BaseReg) {
4453  Value *V = AddrMode.BaseReg;
4454  if (V->getType()->isPointerTy())
4455  V = Builder.CreatePtrToInt(V, IntPtrTy, "sunkaddr");
4456  if (V->getType() != IntPtrTy)
4457  V = Builder.CreateIntCast(V, IntPtrTy, /*isSigned=*/true, "sunkaddr");
4458  Result = V;
4459  }
4460 
4461  // Add the scale value.
4462  if (AddrMode.Scale) {
4463  Value *V = AddrMode.ScaledReg;
4464  if (V->getType() == IntPtrTy) {
4465  // done.
4466  } else if (V->getType()->isPointerTy()) {
4467  V = Builder.CreatePtrToInt(V, IntPtrTy, "sunkaddr");
4468  } else if (cast<IntegerType>(IntPtrTy)->getBitWidth() <
4469  cast<IntegerType>(V->getType())->getBitWidth()) {
4470  V = Builder.CreateTrunc(V, IntPtrTy, "sunkaddr");
4471  } else {
4472  // It is only safe to sign extend the BaseReg if we know that the math
4473  // required to create it did not overflow before we extend it. Since
4474  // the original IR value was tossed in favor of a constant back when
4475  // the AddrMode was created we need to bail out gracefully if widths
4476  // do not match instead of extending it.
4477  Instruction *I = dyn_cast_or_null<Instruction>(Result);
4478  if (I && (Result != AddrMode.BaseReg))
4479  I->eraseFromParent();
4480  return false;
4481  }
4482  if (AddrMode.Scale != 1)
4483  V = Builder.CreateMul(V, ConstantInt::get(IntPtrTy, AddrMode.Scale),
4484  "sunkaddr");
4485  if (Result)
4486  Result = Builder.CreateAdd(Result, V, "sunkaddr");
4487  else
4488  Result = V;
4489  }
4490 
4491  // Add in the BaseGV if present.
4492  if (AddrMode.BaseGV) {
4493  Value *V = Builder.CreatePtrToInt(AddrMode.BaseGV, IntPtrTy, "sunkaddr");
4494  if (Result)
4495  Result = Builder.CreateAdd(Result, V, "sunkaddr");
4496  else
4497  Result = V;
4498  }
4499 
4500  // Add in the Base Offset if present.
4501  if (AddrMode.BaseOffs) {
4502  Value *V = ConstantInt::get(IntPtrTy, AddrMode.BaseOffs);
4503  if (Result)
4504  Result = Builder.CreateAdd(Result, V, "sunkaddr");
4505  else
4506  Result = V;
4507  }
4508 
4509  if (!Result)
4510  SunkAddr = Constant::getNullValue(Addr->getType());
4511  else
4512  SunkAddr = Builder.CreateIntToPtr(Result, Addr->getType(), "sunkaddr");
4513  }
4514 
4515  MemoryInst->replaceUsesOfWith(Repl, SunkAddr);
4516  // Store the newly computed address into the cache. In the case we reused a
4517  // value, this should be idempotent.
4518  SunkAddrs[Addr] = WeakTrackingVH(SunkAddr);
4519 
4520  // If we have no uses, recursively delete the value and all dead instructions
4521  // using it.
4522  if (Repl->use_empty()) {
4523  // This can cause recursive deletion, which can invalidate our iterator.
4524  // Use a WeakTrackingVH to hold onto it in case this happens.
4525  Value *CurValue = &*CurInstIterator;
4526  WeakTrackingVH IterHandle(CurValue);
4527  BasicBlock *BB = CurInstIterator->getParent();
4528 
4530 
4531  if (IterHandle != CurValue) {
4532  // If the iterator instruction was recursively deleted, start over at the
4533  // start of the block.
4534  CurInstIterator = BB->begin();
4535  SunkAddrs.clear();
4536  }
4537  }
4538  ++NumMemoryInsts;
4539  return true;
4540 }
4541 
4542 /// If there are any memory operands, use OptimizeMemoryInst to sink their
4543 /// address computing into the block when possible / profitable.
4544 bool CodeGenPrepare::optimizeInlineAsmInst(CallInst *CS) {
4545  bool MadeChange = false;
4546 
4547  const TargetRegisterInfo *TRI =
4548  TM->getSubtargetImpl(*CS->getFunction())->getRegisterInfo();
4549  TargetLowering::AsmOperandInfoVector TargetConstraints =
4550  TLI->ParseConstraints(*DL, TRI, CS);
4551  unsigned ArgNo = 0;
4552  for (unsigned i = 0, e = TargetConstraints.size(); i != e; ++i) {
4553  TargetLowering::AsmOperandInfo &OpInfo = TargetConstraints[i];
4554 
4555  // Compute the constraint code and ConstraintType to use.
4556  TLI->ComputeConstraintToUse(OpInfo, SDValue());
4557 
4558  if (OpInfo.ConstraintType == TargetLowering::C_Memory &&
4559  OpInfo.isIndirect) {
4560  Value *OpVal = CS->getArgOperand(ArgNo++);
4561  MadeChange |= optimizeMemoryInst(CS, OpVal, OpVal->getType(), ~0u);
4562  } else if (OpInfo.Type == InlineAsm::isInput)
4563  ArgNo++;
4564  }
4565 
4566  return MadeChange;
4567 }
4568 
4569 /// \brief Check if all the uses of \p Val are equivalent (or free) zero or
4570 /// sign extensions.
4571 static bool hasSameExtUse(Value *Val, const TargetLowering &TLI) {
4572  assert(!Val->use_empty() && "Input must have at least one use");
4573  const Instruction *FirstUser = cast<Instruction>(*Val->user_begin());
4574  bool IsSExt = isa<SExtInst>(FirstUser);
4575  Type *ExtTy = FirstUser->getType();
4576  for (const User *U : Val->users()) {
4577  const Instruction *UI = cast<Instruction>(U);
4578  if ((IsSExt && !isa<SExtInst>(UI)) || (!IsSExt && !isa<ZExtInst>(UI)))
4579  return false;
4580  Type *CurTy = UI->getType();
4581  // Same input and output types: Same instruction after CSE.
4582  if (CurTy == ExtTy)
4583  continue;
4584 
4585  // If IsSExt is true, we are in this situation:
4586  // a = Val
4587  // b = sext ty1 a to ty2
4588  // c = sext ty1 a to ty3
4589  // Assuming ty2 is shorter than ty3, this could be turned into:
4590  // a = Val
4591  // b = sext ty1 a to ty2
4592  // c = sext ty2 b to ty3
4593  // However, the last sext is not free.
4594  if (IsSExt)
4595  return false;
4596 
4597  // This is a ZExt, maybe this is free to extend from one type to another.
4598  // In that case, we would not account for a different use.
4599  Type *NarrowTy;
4600  Type *LargeTy;
4601  if (ExtTy->getScalarType()->getIntegerBitWidth() >
4602  CurTy->getScalarType()->getIntegerBitWidth()) {
4603  NarrowTy = CurTy;
4604  LargeTy = ExtTy;
4605  } else {
4606  NarrowTy = ExtTy;
4607  LargeTy = CurTy;
4608  }
4609 
4610  if (!TLI.isZExtFree(NarrowTy, LargeTy))
4611  return false;
4612  }
4613  // All uses are the same or can be derived from one another for free.
4614  return true;
4615 }
4616 
4617 /// \brief Try to speculatively promote extensions in \p Exts and continue
4618 /// promoting through newly promoted operands recursively as far as doing so is
4619 /// profitable. Save extensions profitably moved up, in \p ProfitablyMovedExts.
4620 /// When some promotion happened, \p TPT contains the proper state to revert
4621 /// them.
4622 ///
4623 /// \return true if some promotion happened, false otherwise.
4624 bool CodeGenPrepare::tryToPromoteExts(
4625  TypePromotionTransaction &TPT, const SmallVectorImpl<Instruction *> &Exts,
4626  SmallVectorImpl<Instruction *> &ProfitablyMovedExts,
4627  unsigned CreatedInstsCost) {
4628  bool Promoted = false;
4629 
4630  // Iterate over all the extensions to try to promote them.
4631  for (auto I : Exts) {
4632  // Early check if we directly have ext(load).
4633  if (isa<LoadInst>(I->getOperand(0))) {
4634  ProfitablyMovedExts.push_back(I);
4635  continue;
4636  }
4637 
4638  // Check whether or not we want to do any promotion. The reason we have
4639  // this check inside the for loop is to catch the case where an extension
4640  // is directly fed by a load because in such case the extension can be moved
4641  // up without any promotion on its operands.
4642  if (!TLI || !TLI->enableExtLdPromotion() || DisableExtLdPromotion)
4643  return false;
4644 
4645  // Get the action to perform the promotion.
4646  TypePromotionHelper::Action TPH =
4647  TypePromotionHelper::getAction(I, InsertedInsts, *TLI, PromotedInsts);
4648  // Check if we can promote.
4649  if (!TPH) {
4650  // Save the current extension as we cannot move up through its operand.
4651  ProfitablyMovedExts.push_back(I);
4652  continue;
4653  }
4654 
4655  // Save the current state.
4656  TypePromotionTransaction::ConstRestorationPt LastKnownGood =
4657  TPT.getRestorationPoint();
4659  unsigned NewCreatedInstsCost = 0;
4660  unsigned ExtCost = !TLI->isExtFree(I);
4661  // Promote.
4662  Value *PromotedVal = TPH(I, TPT, PromotedInsts, NewCreatedInstsCost,
4663  &NewExts, nullptr, *TLI);
4664  assert(PromotedVal &&
4665  "TypePromotionHelper should have filtered out those cases");
4666 
4667  // We would be able to merge only one extension in a load.
4668  // Therefore, if we have more than 1 new extension we heuristically
4669  // cut this search path, because it means we degrade the code quality.
4670  // With exactly 2, the transformation is neutral, because we will merge
4671  // one extension but leave one. However, we optimistically keep going,
4672  // because the new extension may be removed too.
4673  long long TotalCreatedInstsCost = CreatedInstsCost + NewCreatedInstsCost;
4674  // FIXME: It would be possible to propagate a negative value instead of
4675  // conservatively ceiling it to 0.
4676  TotalCreatedInstsCost =
4677  std::max((long long)0, (TotalCreatedInstsCost - ExtCost));
4678  if (!StressExtLdPromotion &&
4679  (TotalCreatedInstsCost > 1 ||
4680  !isPromotedInstructionLegal(*TLI, *DL, PromotedVal))) {
4681  // This promotion is not profitable, rollback to the previous state, and
4682  // save the current extension in ProfitablyMovedExts as the latest
4683  // speculative promotion turned out to be unprofitable.
4684  TPT.rollback(LastKnownGood);
4685  ProfitablyMovedExts.push_back(I);
4686  continue;
4687  }
4688  // Continue promoting NewExts as far as doing so is profitable.
4689  SmallVector<Instruction *, 2> NewlyMovedExts;
4690  (void)tryToPromoteExts(TPT, NewExts, NewlyMovedExts, TotalCreatedInstsCost);
4691  bool NewPromoted = false;
4692  for (auto ExtInst : NewlyMovedExts) {
4693  Instruction *MovedExt = cast<Instruction>(ExtInst);
4694  Value *ExtOperand = MovedExt->getOperand(0);
4695  // If we have reached to a load, we need this extra profitability check
4696  // as it could potentially be merged into an ext(load).
4697  if (isa<LoadInst>(ExtOperand) &&
4698  !(StressExtLdPromotion || NewCreatedInstsCost <= ExtCost ||
4699  (ExtOperand->hasOneUse() || hasSameExtUse(ExtOperand, *TLI))))
4700  continue;
4701 
4702  ProfitablyMovedExts.push_back(MovedExt);
4703  NewPromoted = true;
4704  }
4705 
4706  // If none of speculative promotions for NewExts is profitable, rollback
4707  // and save the current extension (I) as the last profitable extension.
4708  if (!NewPromoted) {
4709  TPT.rollback(LastKnownGood);
4710  ProfitablyMovedExts.push_back(I);
4711  continue;
4712  }
4713  // The promotion is profitable.
4714  Promoted = true;
4715  }
4716  return Promoted;
4717 }
4718 
4719 /// Merging redundant sexts when one is dominating the other.
4720 bool CodeGenPrepare::mergeSExts(Function &F) {
4721  DominatorTree DT(F);
4722  bool Changed = false;
4723  for (auto &Entry : ValToSExtendedUses) {
4724  SExts &Insts = Entry.second;
4725  SExts CurPts;
4726  for (Instruction *Inst : Insts) {
4727  if (RemovedInsts.count(Inst) || !isa<SExtInst>(Inst) ||
4728  Inst->getOperand(0) != Entry.first)
4729  continue;
4730  bool inserted = false;
4731  for (auto &Pt : CurPts) {
4732  if (DT.dominates(Inst, Pt)) {
4733  Pt->replaceAllUsesWith(Inst);
4734  RemovedInsts.insert(Pt);
4735  Pt->removeFromParent();
4736  Pt = Inst;
4737  inserted = true;
4738  Changed = true;
4739  break;
4740  }
4741  if (!DT.dominates(Pt, Inst))
4742  // Give up if we need to merge in a common dominator as the
4743  // expermients show it is not profitable.
4744  continue;
4745  Inst->replaceAllUsesWith(Pt);
4746  RemovedInsts.insert(Inst);
4747  Inst->removeFromParent();
4748  inserted = true;
4749  Changed = true;
4750  break;
4751  }
4752  if (!inserted)
4753  CurPts.push_back(Inst);
4754  }
4755  }
4756  return Changed;
4757 }
4758 
4759 /// Return true, if an ext(load) can be formed from an extension in
4760 /// \p MovedExts.
4761 bool CodeGenPrepare::canFormExtLd(
4762  const SmallVectorImpl<Instruction *> &MovedExts, LoadInst *&LI,
4763  Instruction *&Inst, bool HasPromoted) {
4764  for (auto *MovedExtInst : MovedExts) {
4765  if (isa<LoadInst>(MovedExtInst->getOperand(0))) {
4766  LI = cast<LoadInst>(MovedExtInst->getOperand(0));
4767  Inst = MovedExtInst;
4768  break;
4769  }
4770  }
4771  if (!LI)
4772  return false;
4773 
4774  // If they're already in the same block, there's nothing to do.
4775  // Make the cheap checks first if we did not promote.
4776  // If we promoted, we need to check if it is indeed profitable.
4777  if (!HasPromoted && LI->getParent() == Inst->getParent())
4778  return false;
4779 
4780  return TLI->isExtLoad(LI, Inst, *DL);
4781 }
4782 
4783 /// Move a zext or sext fed by a load into the same basic block as the load,
4784 /// unless conditions are unfavorable. This allows SelectionDAG to fold the
4785 /// extend into the load.
4786 ///
4787 /// E.g.,
4788 /// \code
4789 /// %ld = load i32* %addr
4790 /// %add = add nuw i32 %ld, 4
4791 /// %zext = zext i32 %add to i64
4792 // \endcode
4793 /// =>
4794 /// \code
4795 /// %ld = load i32* %addr
4796 /// %zext = zext i32 %ld to i64
4797 /// %add = add nuw i64 %zext, 4
4798 /// \encode
4799 /// Note that the promotion in %add to i64 is done in tryToPromoteExts(), which
4800 /// allow us to match zext(load i32*) to i64.
4801 ///
4802 /// Also, try to promote the computations used to obtain a sign extended
4803 /// value used into memory accesses.
4804 /// E.g.,
4805 /// \code
4806 /// a = add nsw i32 b, 3
4807 /// d = sext i32 a to i64
4808 /// e = getelementptr ..., i64 d
4809 /// \endcode
4810 /// =>
4811 /// \code
4812 /// f = sext i32 b to i64
4813 /// a = add nsw i64 f, 3
4814 /// e = getelementptr ..., i64 a
4815 /// \endcode
4816 ///
4817 /// \p Inst[in/out] the extension may be modified during the process if some
4818 /// promotions apply.
4819 bool CodeGenPrepare::optimizeExt(Instruction *&Inst) {
4820  // ExtLoad formation and address type promotion infrastructure requires TLI to
4821  // be effective.
4822  if (!TLI)
4823  return false;
4824 
4825  bool AllowPromotionWithoutCommonHeader = false;
4826  /// See if it is an interesting sext operations for the address type
4827  /// promotion before trying to promote it, e.g., the ones with the right
4828  /// type and used in memory accesses.
4829  bool ATPConsiderable = TTI->shouldConsiderAddressTypePromotion(
4830  *Inst, AllowPromotionWithoutCommonHeader);
4831  TypePromotionTransaction TPT(RemovedInsts);
4832  TypePromotionTransaction::ConstRestorationPt LastKnownGood =
4833  TPT.getRestorationPoint();
4835  SmallVector<Instruction *, 2> SpeculativelyMovedExts;
4836  Exts.push_back(Inst);
4837 
4838  bool HasPromoted = tryToPromoteExts(TPT, Exts, SpeculativelyMovedExts);
4839 
4840  // Look for a load being extended.
4841  LoadInst *LI = nullptr;
4842  Instruction *ExtFedByLoad;
4843 
4844  // Try to promote a chain of computation if it allows to form an extended
4845  // load.
4846  if (canFormExtLd(SpeculativelyMovedExts, LI, ExtFedByLoad, HasPromoted)) {
4847  assert(LI && ExtFedByLoad && "Expect a valid load and extension");
4848  TPT.commit();
4849  // Move the extend into the same block as the load
4850  ExtFedByLoad->moveAfter(LI);
4851  // CGP does not check if the zext would be speculatively executed when moved
4852  // to the same basic block as the load. Preserving its original location
4853  // would pessimize the debugging experience, as well as negatively impact
4854  // the quality of sample pgo. We don't want to use "line 0" as that has a
4855  // size cost in the line-table section and logically the zext can be seen as
4856  // part of the load. Therefore we conservatively reuse the same debug
4857  // location for the load and the zext.
4858  ExtFedByLoad->setDebugLoc(LI->getDebugLoc());
4859  ++NumExtsMoved;
4860  Inst = ExtFedByLoad;
4861  return true;
4862  }
4863 
4864  // Continue promoting SExts if known as considerable depending on targets.
4865  if (ATPConsiderable &&
4866  performAddressTypePromotion(Inst, AllowPromotionWithoutCommonHeader,
4867  HasPromoted, TPT, SpeculativelyMovedExts))
4868  return true;
4869 
4870  TPT.rollback(LastKnownGood);
4871  return false;
4872 }
4873 
4874 // Perform address type promotion if doing so is profitable.
4875 // If AllowPromotionWithoutCommonHeader == false, we should find other sext
4876 // instructions that sign extended the same initial value. However, if
4877 // AllowPromotionWithoutCommonHeader == true, we expect promoting the
4878 // extension is just profitable.
4879 bool CodeGenPrepare::performAddressTypePromotion(
4880  Instruction *&Inst, bool AllowPromotionWithoutCommonHeader,
4881  bool HasPromoted, TypePromotionTransaction &TPT,
4882  SmallVectorImpl<Instruction *> &SpeculativelyMovedExts) {
4883  bool Promoted = false;
4884  SmallPtrSet<Instruction *, 1> UnhandledExts;
4885  bool AllSeenFirst = true;
4886  for (auto I : SpeculativelyMovedExts) {
4887  Value *HeadOfChain = I->getOperand(0);
4889  SeenChainsForSExt.find(HeadOfChain);
4890  // If there is an unhandled SExt which has the same header, try to promote
4891  // it as well.
4892  if (AlreadySeen != SeenChainsForSExt.end()) {
4893  if (AlreadySeen->second != nullptr)
4894  UnhandledExts.insert(AlreadySeen->second);
4895  AllSeenFirst = false;
4896  }
4897  }
4898 
4899  if (!AllSeenFirst || (AllowPromotionWithoutCommonHeader &&
4900  SpeculativelyMovedExts.size() == 1)) {
4901  TPT.commit();
4902  if (HasPromoted)
4903  Promoted = true;
4904  for (auto I : SpeculativelyMovedExts) {
4905  Value *HeadOfChain = I->getOperand(0);
4906  SeenChainsForSExt[HeadOfChain] = nullptr;
4907  ValToSExtendedUses[HeadOfChain].push_back(I);
4908  }
4909  // Update Inst as promotion happen.
4910  Inst = SpeculativelyMovedExts.pop_back_val();
4911  } else {
4912  // This is the first chain visited from the header, keep the current chain
4913  // as unhandled. Defer to promote this until we encounter another SExt
4914  // chain derived from the same header.
4915  for (auto I : SpeculativelyMovedExts) {
4916  Value *HeadOfChain = I->getOperand(0);
4917  SeenChainsForSExt[HeadOfChain] = Inst;
4918  }
4919  return false;
4920  }
4921 
4922  if (!AllSeenFirst && !UnhandledExts.empty())
4923  for (auto VisitedSExt : UnhandledExts) {
4924  if (RemovedInsts.count(VisitedSExt))
4925  continue;
4926  TypePromotionTransaction TPT(RemovedInsts);
4929  Exts.push_back(VisitedSExt);
4930  bool HasPromoted = tryToPromoteExts(TPT, Exts, Chains);
4931  TPT.commit();
4932  if (HasPromoted)
4933  Promoted = true;
4934  for (auto I : Chains) {
4935  Value *HeadOfChain = I->getOperand(0);
4936  // Mark this as handled.
4937  SeenChainsForSExt[HeadOfChain] = nullptr;
4938  ValToSExtendedUses[HeadOfChain].push_back(I);
4939  }
4940  }
4941  return Promoted;
4942 }
4943 
4944 bool CodeGenPrepare::optimizeExtUses(Instruction *I) {
4945  BasicBlock *DefBB = I->getParent();
4946 
4947  // If the result of a {s|z}ext and its source are both live out, rewrite all
4948  // other uses of the source with result of extension.
4949  Value *Src = I->getOperand(0);
4950  if (Src->hasOneUse())
4951  return false;
4952 
4953  // Only do this xform if truncating is free.
4954  if (TLI && !TLI->isTruncateFree(I->getType(), Src->getType()))
4955  return false;
4956 
4957  // Only safe to perform the optimization if the source is also defined in
4958  // this block.
4959  if (!isa<Instruction>(Src) || DefBB != cast<Instruction>(Src)->getParent())
4960  return false;
4961 
4962  bool DefIsLiveOut = false;
4963  for (User *U : I->users()) {
4964  Instruction *UI = cast<Instruction>(U);
4965 
4966  // Figure out which BB this ext is used in.
4967  BasicBlock *UserBB = UI->getParent();
4968  if (UserBB == DefBB) continue;
4969  DefIsLiveOut = true;
4970  break;
4971  }
4972  if (!DefIsLiveOut)
4973  return false;
4974 
4975  // Make sure none of the uses are PHI nodes.
4976  for (User *U : Src->users()) {
4977  Instruction *UI = cast<Instruction>(U);
4978  BasicBlock *UserBB = UI->getParent();
4979  if (UserBB == DefBB) continue;
4980  // Be conservative. We don't want this xform to end up introducing
4981  // reloads just before load / store instructions.
4982  if (isa<PHINode>(UI) || isa<LoadInst>(UI) || isa<StoreInst>(UI))
4983  return false;
4984  }
4985 
4986  // InsertedTruncs - Only insert one trunc in each block once.
4987  DenseMap<BasicBlock*, Instruction*> InsertedTruncs;
4988 
4989  bool MadeChange = false;
4990  for (Use &U : Src->uses()) {
4991  Instruction *User = cast<Instruction>(U.getUser());
4992 
4993  // Figure out which BB this ext is used in.
4994  BasicBlock *UserBB = User->getParent();
4995  if (UserBB == DefBB) continue;
4996 
4997  // Both src and def are live in this block. Rewrite the use.
4998  Instruction *&InsertedTrunc = InsertedTruncs[UserBB];
4999 
5000  if (!InsertedTrunc) {
5001  BasicBlock::iterator InsertPt = UserBB->getFirstInsertionPt();
5002  assert(InsertPt != UserBB->end());
5003  InsertedTrunc = new TruncInst(I, Src->getType(), "", &*InsertPt);
5004  InsertedInsts.insert(InsertedTrunc);
5005  }
5006 
5007  // Replace a use of the {s|z}ext source with a use of the result.
5008  U = InsertedTrunc;
5009  ++NumExtUses;
5010  MadeChange = true;
5011  }
5012 
5013  return MadeChange;
5014 }
5015 
5016 // Find loads whose uses only use some of the loaded value's bits. Add an "and"
5017 // just after the load if the target can fold this into one extload instruction,
5018 // with the hope of eliminating some of the other later "and" instructions using
5019 // the loaded value. "and"s that are made trivially redundant by the insertion
5020 // of the new "and" are removed by this function, while others (e.g. those whose
5021 // path from the load goes through a phi) are left for isel to potentially
5022 // remove.
5023 //
5024 // For example:
5025 //
5026 // b0:
5027 // x = load i32
5028 // ...
5029 // b1:
5030 // y = and x, 0xff
5031 // z = use y
5032 //
5033 // becomes:
5034 //
5035 // b0:
5036 // x = load i32
5037 // x' = and x, 0xff
5038 // ...
5039 // b1:
5040 // z = use x'
5041 //
5042 // whereas:
5043 //
5044 // b0:
5045 // x1 = load i32
5046 // ...
5047 // b1:
5048 // x2 = load i32
5049 // ...
5050 // b2:
5051 // x = phi x1, x2
5052 // y = and x, 0xff
5053 //
5054 // becomes (after a call to optimizeLoadExt for each load):
5055 //
5056 // b0:
5057 // x1 = load i32
5058 // x1' = and x1, 0xff
5059 // ...
5060 // b1:
5061 // x2 = load i32
5062 // x2' = and x2, 0xff
5063 // ...
5064 // b2:
5065 // x = phi x1', x2'
5066 // y = and x, 0xff
5067 bool CodeGenPrepare::optimizeLoadExt(LoadInst *Load) {
5068  if (!Load->isSimple() ||
5069  !(Load->getType()->isIntegerTy() || Load->getType()->isPointerTy()))
5070  return false;
5071 
5072  // Skip loads we've already transformed.
5073  if (Load->hasOneUse() &&
5074  InsertedInsts.count(cast<Instruction>(*Load->user_begin())))
5075  return false;
5076 
5077  // Look at all uses of Load, looking through phis, to determine how many bits
5078  // of the loaded value are needed.
5081  SmallVector<Instruction *, 8> AndsToMaybeRemove;
5082  for (auto *U : Load->users())
5083  WorkList.push_back(cast<Instruction>(U));
5084 
5085  EVT LoadResultVT = TLI->getValueType(*DL, Load->getType());
5086  unsigned BitWidth = LoadResultVT.getSizeInBits();
5087  APInt DemandBits(BitWidth, 0);
5088  APInt WidestAndBits(BitWidth, 0);
5089 
5090  while (!WorkList.empty()) {
5091  Instruction *I = WorkList.back();
5092  WorkList.pop_back();
5093 
5094  // Break use-def graph loops.
5095  if (!Visited.insert(I).second)
5096  continue;
5097 
5098  // For a PHI node, push all of its users.
5099  if (auto *Phi = dyn_cast<PHINode>(I)) {
5100  for (auto *U : Phi->users())
5101  WorkList.push_back(cast<Instruction>(U));
5102  continue;
5103  }
5104 
5105  switch (I->getOpcode()) {
5106  case Instruction::And: {
5107  auto *AndC = dyn_cast<ConstantInt>(I->getOperand(1));
5108  if (!AndC)
5109  return false;
5110  APInt AndBits = AndC->getValue();
5111  DemandBits |= AndBits;
5112  // Keep track of the widest and mask we see.
5113  if (AndBits.ugt(WidestAndBits))
5114  WidestAndBits = AndBits;
5115  if (AndBits == WidestAndBits && I->getOperand(0) == Load)
5116  AndsToMaybeRemove.push_back(I);
5117  break;
5118  }
5119 
5120  case Instruction::Shl: {
5121  auto *ShlC = dyn_cast<ConstantInt>(I->getOperand(1));
5122  if (!ShlC)
5123  return false;
5124  uint64_t ShiftAmt = ShlC->getLimitedValue(BitWidth - 1);
5125  DemandBits.setLowBits(BitWidth - ShiftAmt);
5126  break;
5127  }
5128 
5129  case Instruction::Trunc: {
5130  EVT TruncVT = TLI->getValueType(*DL, I->getType());
5131  unsigned TruncBitWidth = TruncVT.getSizeInBits();
5132  DemandBits.setLowBits(TruncBitWidth);
5133  break;
5134  }
5135 
5136  default:
5137  return false;
5138  }
5139  }
5140 
5141  uint32_t ActiveBits = DemandBits.getActiveBits();
5142  // Avoid hoisting (and (load x) 1) since it is unlikely to be folded by the
5143  // target even if isLoadExtLegal says an i1 EXTLOAD is valid. For example,
5144  // for the AArch64 target isLoadExtLegal(ZEXTLOAD, i32, i1) returns true, but
5145  // (and (load x) 1) is not matched as a single instruction, rather as a LDR
5146  // followed by an AND.
5147  // TODO: Look into removing this restriction by fixing backends to either
5148  // return false for isLoadExtLegal for i1 or have them select this pattern to
5149  // a single instruction.
5150  //
5151  // Also avoid hoisting if we didn't see any ands with the exact DemandBits
5152  // mask, since these are the only ands that will be removed by isel.
5153  if (ActiveBits <= 1 || !DemandBits.isMask(ActiveBits) ||
5154  WidestAndBits != DemandBits)
5155  return false;
5156 
5157  LLVMContext &Ctx = Load->getType()->getContext();
5158  Type *TruncTy = Type::getIntNTy(Ctx, ActiveBits);
5159  EVT TruncVT = TLI->getValueType(*DL, TruncTy);
5160 
5161  // Reject cases that won't be matched as extloads.
5162  if (!LoadResultVT.bitsGT(TruncVT) || !TruncVT.isRound() ||
5163  !TLI->isLoadExtLegal(ISD::ZEXTLOAD, LoadResultVT, TruncVT))
5164  return false;
5165 
5166  IRBuilder<> Builder(Load->getNextNode());
5167  auto *NewAnd = dyn_cast<Instruction>(
5168  Builder.CreateAnd(Load, ConstantInt::get(Ctx, DemandBits)));
5169  // Mark this instruction as "inserted by CGP", so that other
5170  // optimizations don't touch it.
5171  InsertedInsts.insert(NewAnd);
5172 
5173  // Replace all uses of load with new and (except for the use of load in the
5174  // new and itself).
5175  Load->replaceAllUsesWith(NewAnd);
5176  NewAnd->setOperand(0, Load);
5177 
5178  // Remove any and instructions that are now redundant.
5179  for (auto *And : AndsToMaybeRemove)
5180  // Check that the and mask is the same as the one we decided to put on the
5181  // new and.
5182  if (cast<ConstantInt>(And->getOperand(1))->getValue() == DemandBits) {
5183  And->replaceAllUsesWith(NewAnd);
5184  if (&*CurInstIterator == And)
5185  CurInstIterator = std::next(And->getIterator());
5186  And->eraseFromParent();
5187  ++NumAndUses;
5188  }
5189 
5190  ++NumAndsAdded;
5191  return true;
5192 }
5193 
5194 /// Check if V (an operand of a select instruction) is an expensive instruction
5195 /// that is only used once.
5196 static bool sinkSelectOperand(const TargetTransformInfo *TTI, Value *V) {
5197  auto *I = dyn_cast<Instruction>(V);
5198  // If it's safe to speculatively execute, then it should not have side
5199  // effects; therefore, it's safe to sink and possibly *not* execute.
5200  return I && I->hasOneUse() && isSafeToSpeculativelyExecute(I) &&
5202 }
5203 
5204 /// Returns true if a SelectInst should be turned into an explicit branch.
5206  const TargetLowering *TLI,
5207  SelectInst *SI) {
5208  // If even a predictable select is cheap, then a branch can't be cheaper.
5209  if (!TLI->isPredictableSelectExpensive())
5210  return false;
5211 
5212  // FIXME: This should use the same heuristics as IfConversion to determine
5213  // whether a select is better represented as a branch.
5214 
5215  // If metadata tells us that the select condition is obviously predictable,
5216  // then we want to replace the select with a branch.
5217  uint64_t TrueWeight, FalseWeight;
5218  if (SI->extractProfMetadata(TrueWeight, FalseWeight)) {
5219  uint64_t Max = std::max(TrueWeight, FalseWeight);
5220  uint64_t Sum = TrueWeight + FalseWeight;
5221  if (Sum != 0) {
5222  auto Probability = BranchProbability::getBranchProbability(Max, Sum);
5223  if (Probability > TLI->getPredictableBranchThreshold())
5224  return true;
5225  }
5226  }
5227 
5228  CmpInst *Cmp = dyn_cast<CmpInst>(SI->getCondition());
5229 
5230  // If a branch is predictable, an out-of-order CPU can avoid blocking on its
5231  // comparison condition. If the compare has more than one use, there's
5232  // probably another cmov or setcc around, so it's not worth emitting a branch.
5233  if (!Cmp || !Cmp->hasOneUse())
5234  return false;
5235 
5236  // If either operand of the select is expensive and only needed on one side
5237  // of the select, we should form a branch.
5238  if (sinkSelectOperand(TTI, SI->getTrueValue()) ||
5239  sinkSelectOperand(TTI, SI->getFalseValue()))
5240  return true;
5241 
5242  return false;
5243 }
5244 
5245 /// If \p isTrue is true, return the true value of \p SI, otherwise return
5246 /// false value of \p SI. If the true/false value of \p SI is defined by any
5247 /// select instructions in \p Selects, look through the defining select
5248 /// instruction until the true/false value is not defined in \p Selects.
5250  SelectInst *SI, bool isTrue,
5251  const SmallPtrSet<const Instruction *, 2> &Selects) {
5252  Value *V;
5253 
5254  for (SelectInst *DefSI = SI; DefSI != nullptr && Selects.count(DefSI);
5255  DefSI = dyn_cast<SelectInst>(V)) {
5256  assert(DefSI->getCondition() == SI->getCondition() &&
5257  "The condition of DefSI does not match with SI");
5258  V = (isTrue ? DefSI->getTrueValue() : DefSI->getFalseValue());
5259  }
5260  return V;
5261 }
5262 
5263 /// If we have a SelectInst that will likely profit from branch prediction,
5264 /// turn it into a branch.
5265 bool CodeGenPrepare::optimizeSelectInst(SelectInst *SI) {
5266  // Find all consecutive select instructions that share the same condition.
5268  ASI.push_back(SI);
5270  It != SI->getParent()->end(); ++It) {
5271  SelectInst *I = dyn_cast<SelectInst>(&*It);
5272  if (I && SI->getCondition() == I->getCondition()) {
5273  ASI.push_back(I);
5274  } else {
5275  break;
5276  }
5277  }
5278 
5279  SelectInst *LastSI = ASI.back();
5280  // Increment the current iterator to skip all the rest of select instructions
5281  // because they will be either "not lowered" or "all lowered" to branch.
5282  CurInstIterator = std::next(LastSI->getIterator());
5283 
5284  bool VectorCond = !SI->getCondition()->getType()->isIntegerTy(1);
5285 
5286  // Can we convert the 'select' to CF ?
5287  if (DisableSelectToBranch || OptSize || !TLI || VectorCond ||
5289  return false;
5290 
5292  if (VectorCond)
5293  SelectKind = TargetLowering::VectorMaskSelect;
5294  else if (SI->getType()->isVectorTy())
5296  else
5297  SelectKind = TargetLowering::ScalarValSelect;
5298 
5299  if (TLI->isSelectSupported(SelectKind) &&
5300  !isFormingBranchFromSelectProfitable(TTI, TLI, SI))
5301  return false;
5302 
5303  ModifiedDT = true;
5304 
5305  // Transform a sequence like this:
5306  // start:
5307  // %cmp = cmp uge i32 %a, %b
5308  // %sel = select i1 %cmp, i32 %c, i32 %d
5309  //
5310  // Into:
5311  // start:
5312  // %cmp = cmp uge i32 %a, %b
5313  // br i1 %cmp, label %select.true, label %select.false
5314  // select.true:
5315  // br label %select.end
5316  // select.false:
5317  // br label %select.end
5318  // select.end:
5319  // %sel = phi i32 [ %c, %select.true ], [ %d, %select.false ]
5320  //
5321  // In addition, we may sink instructions that produce %c or %d from
5322  // the entry block into the destination(s) of the new branch.
5323  // If the true or false blocks do not contain a sunken instruction, that
5324  // block and its branch may be optimized away. In that case, one side of the
5325  // first branch will point directly to select.end, and the corresponding PHI
5326  // predecessor block will be the start block.
5327 
5328  // First, we split the block containing the select into 2 blocks.
5329  BasicBlock *StartBlock = SI->getParent();
5330  BasicBlock::iterator SplitPt = ++(BasicBlock::iterator(LastSI));
5331  BasicBlock *EndBlock = StartBlock->splitBasicBlock(SplitPt, "select.end");
5332 
5333  // Delete the unconditional branch that was just created by the split.
5334  StartBlock->getTerminator()->eraseFromParent();
5335 
5336  // These are the new basic blocks for the conditional branch.
5337  // At least one will become an actual new basic block.
5338  BasicBlock *TrueBlock = nullptr;
5339  BasicBlock *FalseBlock = nullptr;
5340  BranchInst *TrueBranch = nullptr;
5341  BranchInst *FalseBranch = nullptr;
5342 
5343  // Sink expensive instructions into the conditional blocks to avoid executing
5344  // them speculatively.
5345  for (SelectInst *SI : ASI) {
5346  if (sinkSelectOperand(TTI, SI->getTrueValue())) {
5347  if (TrueBlock == nullptr) {
5348  TrueBlock = BasicBlock::Create(SI->getContext(), "select.true.sink",
5349  EndBlock->getParent(), EndBlock);
5350  TrueBranch = BranchInst::Create(EndBlock, TrueBlock);
5351  }
5352  auto *TrueInst = cast<Instruction>(SI->getTrueValue());
5353  TrueInst->moveBefore(TrueBranch);
5354  }
5355  if (sinkSelectOperand(TTI, SI->getFalseValue())) {
5356  if (FalseBlock == nullptr) {
5357  FalseBlock = BasicBlock::Create(SI->getContext(), "select.false.sink",
5358  EndBlock->getParent(), EndBlock);
5359  FalseBranch = BranchInst::Create(EndBlock, FalseBlock);
5360  }
5361  auto *FalseInst = cast<Instruction>(SI->getFalseValue());
5362  FalseInst->moveBefore(FalseBranch);
5363  }
5364  }
5365 
5366  // If there was nothing to sink, then arbitrarily choose the 'false' side
5367  // for a new input value to the PHI.
5368  if (TrueBlock == FalseBlock) {
5369  assert(TrueBlock == nullptr &&
5370  "Unexpected basic block transform while optimizing select");
5371 
5372  FalseBlock = BasicBlock::Create(SI->getContext(), "select.false",
5373  EndBlock->getParent(), EndBlock);
5374  BranchInst::Create(EndBlock, FalseBlock);
5375  }
5376 
5377  // Insert the real conditional branch based on the original condition.
5378  // If we did not create a new block for one of the 'true' or 'false' paths
5379  // of the condition, it means that side of the branch goes to the end block
5380  // directly and the path originates from the start block from the point of
5381  // view of the new PHI.
5382  BasicBlock *TT, *FT;
5383  if (TrueBlock == nullptr) {
5384  TT = EndBlock;
5385  FT = FalseBlock;
5386  TrueBlock = StartBlock;
5387  } else if (FalseBlock == nullptr) {
5388  TT = TrueBlock;
5389  FT = EndBlock;
5390  FalseBlock = StartBlock;
5391  } else {
5392  TT = TrueBlock;
5393  FT = FalseBlock;
5394  }
5395  IRBuilder<>(SI).CreateCondBr(SI->getCondition(), TT, FT, SI);
5396 
5398  INS.insert(ASI.begin(), ASI.end());
5399  // Use reverse iterator because later select may use the value of the
5400  // earlier select, and we need to propagate value through earlier select
5401  // to get the PHI operand.
5402  for (auto It = ASI.rbegin(); It != ASI.rend(); ++It) {
5403  SelectInst *SI = *It;
5404  // The select itself is replaced with a PHI Node.
5405  PHINode *PN = PHINode::Create(SI->getType(), 2, "", &EndBlock->front());
5406  PN->takeName(SI);
5407  PN->addIncoming(getTrueOrFalseValue(SI, true, INS), TrueBlock);
5408  PN->addIncoming(getTrueOrFalseValue(SI, false, INS), FalseBlock);
5409 
5410  SI->replaceAllUsesWith(PN);
5411  SI->eraseFromParent();
5412  INS.erase(SI);
5413  ++NumSelectsExpanded;
5414  }
5415 
5416  // Instruct OptimizeBlock to skip to the next block.
5417  CurInstIterator = StartBlock->end();
5418  return true;
5419 }
5420 
5423  int SplatElem = -1;
5424  for (unsigned i = 0; i < Mask.size(); ++i) {
5425  if (SplatElem != -1 && Mask[i] != -1 && Mask[i] != SplatElem)
5426  return false;
5427  SplatElem = Mask[i];
5428  }
5429 
5430  return true;
5431 }
5432 
5433 /// Some targets have expensive vector shifts if the lanes aren't all the same
5434 /// (e.g. x86 only introduced "vpsllvd" and friends with AVX2). In these cases
5435 /// it's often worth sinking a shufflevector splat down to its use so that
5436 /// codegen can spot all lanes are identical.
5437 bool CodeGenPrepare::optimizeShuffleVectorInst(ShuffleVectorInst *SVI) {
5438  BasicBlock *DefBB = SVI->getParent();
5439 
5440  // Only do this xform if variable vector shifts are particularly expensive.
5441  if (!TLI || !TLI->isVectorShiftByScalarCheap(SVI->getType()))
5442  return false;
5443 
5444  // We only expect better codegen by sinking a shuffle if we can recognise a
5445  // constant splat.
5446  if (!isBroadcastShuffle(SVI))
5447  return false;
5448 
5449  // InsertedShuffles - Only insert a shuffle in each block once.
5450  DenseMap<BasicBlock*, Instruction*> InsertedShuffles;
5451 
5452  bool MadeChange = false;
5453  for (User *U : SVI->users()) {
5454  Instruction *UI = cast<Instruction>(U);
5455 
5456  // Figure out which BB this ext is used in.
5457  BasicBlock *UserBB = UI->getParent();
5458  if (UserBB == DefBB) continue;
5459 
5460  // For now only apply this when the splat is used by a shift instruction.
5461  if (!UI->isShift()) continue;
5462 
5463  // Everything checks out, sink the shuffle if the user's block doesn't
5464  // already have a copy.
5465  Instruction *&InsertedShuffle = InsertedShuffles[UserBB];
5466 
5467  if (!InsertedShuffle) {
5468  BasicBlock::iterator InsertPt = UserBB->getFirstInsertionPt();
5469  assert(InsertPt != UserBB->end());
5470  InsertedShuffle =
5471  new ShuffleVectorInst(SVI->getOperand(0), SVI->getOperand(1),
5472  SVI->getOperand(2), "", &*InsertPt);
5473  }
5474 
5475  UI->replaceUsesOfWith(SVI, InsertedShuffle);
5476  MadeChange = true;
5477  }
5478 
5479  // If we removed all uses, nuke the shuffle.
5480  if (SVI->use_empty()) {
5481  SVI->eraseFromParent();
5482  MadeChange = true;
5483  }
5484 
5485  return MadeChange;
5486 }
5487 
5488 bool CodeGenPrepare::optimizeSwitchInst(SwitchInst *SI) {
5489  if (!TLI || !DL)
5490  return false;
5491 
5492  Value *Cond = SI->getCondition();
5493  Type *OldType = Cond->getType();
5494  LLVMContext &Context = Cond->getContext();
5495  MVT RegType = TLI->getRegisterType(Context, TLI->getValueType(*DL, OldType));
5496  unsigned RegWidth = RegType.getSizeInBits();
5497 
5498  if (RegWidth <= cast<IntegerType>(OldType)->getBitWidth())
5499  return false;
5500 
5501  // If the register width is greater than the type width, expand the condition
5502  // of the switch instruction and each case constant to the width of the
5503  // register. By widening the type of the switch condition, subsequent
5504  // comparisons (for case comparisons) will not need to be extended to the
5505  // preferred register width, so we will potentially eliminate N-1 extends,
5506  // where N is the number of cases in the switch.
5507  auto *NewType = Type::getIntNTy(Context, RegWidth);
5508 
5509  // Zero-extend the switch condition and case constants unless the switch
5510  // condition is a function argument that is already being sign-extended.
5511  // In that case, we can avoid an unnecessary mask/extension by sign-extending
5512  // everything instead.
5513  Instruction::CastOps ExtType = Instruction::ZExt;
5514  if (auto *Arg = dyn_cast<Argument>(Cond))
5515  if (Arg->hasSExtAttr())
5516  ExtType = Instruction::SExt;
5517 
5518  auto *ExtInst = CastInst::Create(ExtType, Cond, NewType);
5519  ExtInst->insertBefore(SI);
5520  SI->setCondition(ExtInst);
5521  for (auto Case : SI->cases()) {
5522  APInt NarrowConst = Case.getCaseValue()->getValue();
5523  APInt WideConst = (ExtType == Instruction::ZExt) ?
5524  NarrowConst.zext(RegWidth) : NarrowConst.sext(RegWidth);
5525  Case.setValue(ConstantInt::get(Context, WideConst));
5526  }
5527 
5528  return true;
5529 }
5530 
5531 
5532 namespace {
5533 
5534 /// \brief Helper class to promote a scalar operation to a vector one.
5535 /// This class is used to move downward extractelement transition.
5536 /// E.g.,
5537 /// a = vector_op <2 x i32>
5538 /// b = extractelement <2 x i32> a, i32 0
5539 /// c = scalar_op b
5540 /// store c
5541 ///
5542 /// =>
5543 /// a = vector_op <2 x i32>
5544 /// c = vector_op a (equivalent to scalar_op on the related lane)
5545 /// * d = extractelement <2 x i32> c, i32 0
5546 /// * store d
5547 /// Assuming both extractelement and store can be combine, we get rid of the
5548 /// transition.
5549 class VectorPromoteHelper {
5550  /// DataLayout associated with the current module.
5551  const DataLayout &DL;
5552 
5553  /// Used to perform some checks on the legality of vector operations.
5554  const TargetLowering &TLI;
5555 
5556  /// Used to estimated the cost of the promoted chain.
5557  const TargetTransformInfo &TTI;
5558 
5559  /// The transition being moved downwards.
5560  Instruction *Transition;
5561 
5562  /// The sequence of instructions to be promoted.
5563  SmallVector<Instruction *, 4> InstsToBePromoted;
5564 
5565  /// Cost of combining a store and an extract.
5566  unsigned StoreExtractCombineCost;
5567 
5568  /// Instruction that will be combined with the transition.
5569  Instruction *CombineInst = nullptr;
5570 
5571  /// \brief The instruction that represents the current end of the transition.
5572  /// Since we are faking the promotion until we reach the end of the chain
5573  /// of computation, we need a way to get the current end of the transition.
5574  Instruction *getEndOfTransition() const {
5575  if (InstsToBePromoted.empty())
5576  return Transition;
5577  return InstsToBePromoted.back();
5578  }
5579 
5580  /// \brief Return the index of the original value in the transition.
5581  /// E.g., for "extractelement <2 x i32> c, i32 1" the original value,
5582  /// c, is at index 0.
5583  unsigned getTransitionOriginalValueIdx() const {
5584  assert(isa<ExtractElementInst>(Transition) &&
5585  "Other kind of transitions are not supported yet");
5586  return 0;
5587  }
5588 
5589  /// \brief Return the index of the index in the transition.
5590  /// E.g., for "extractelement <2 x i32> c, i32 0" the index
5591  /// is at index 1.
5592  unsigned getTransitionIdx() const {
5593  assert(isa<ExtractElementInst>(Transition) &&
5594  "Other kind of transitions are not supported yet");
5595  return 1;
5596  }
5597 
5598  /// \brief Get the type of the transition.
5599  /// This is the type of the original value.
5600  /// E.g., for "extractelement <2 x i32> c, i32 1" the type of the
5601  /// transition is <2 x i32>.
5602  Type *getTransitionType() const {
5603  return Transition->getOperand(getTransitionOriginalValueIdx())->getType();
5604  }
5605 
5606  /// \brief Promote \p ToBePromoted by moving \p Def downward through.
5607  /// I.e., we have the following sequence:
5608  /// Def = Transition <ty1> a to <ty2>
5609  /// b = ToBePromoted <ty2> Def, ...
5610  /// =>
5611  /// b = ToBePromoted <ty1> a, ...
5612  /// Def = Transition <ty1> ToBePromoted to <ty2>
5613  void promoteImpl(Instruction *ToBePromoted);
5614 
5615  /// \brief Check whether or not it is profitable to promote all the
5616  /// instructions enqueued to be promoted.
5617  bool isProfitableToPromote() {
5618  Value *ValIdx = Transition->getOperand(getTransitionOriginalValueIdx());
5619  unsigned Index = isa<ConstantInt>(ValIdx)
5620  ? cast<ConstantInt>(ValIdx)->getZExtValue()
5621  : -1;
5622  Type *PromotedType = getTransitionType();
5623 
5624  StoreInst *ST = cast<StoreInst>(CombineInst);
5625  unsigned AS = ST->getPointerAddressSpace();
5626  unsigned Align = ST->getAlignment();
5627  // Check if this store is supported.
5629  TLI.getValueType(DL, ST->getValueOperand()->getType()), AS,
5630  Align)) {
5631  // If this is not supported, there is no way we can combine
5632  // the extract with the store.
5633  return false;
5634  }
5635 
5636  // The scalar chain of computation has to pay for the transition
5637  // scalar to vector.
5638  // The vector chain has to account for the combining cost.
5639  uint64_t ScalarCost =
5640  TTI.getVectorInstrCost(Transition->getOpcode(), PromotedType, Index);
5641  uint64_t VectorCost = StoreExtractCombineCost;
5642  for (const auto &Inst : InstsToBePromoted) {
5643  // Compute the cost.
5644  // By construction, all instructions being promoted are arithmetic ones.
5645  // Moreover, one argument is a constant that can be viewed as a splat
5646  // constant.
5647  Value *Arg0 = Inst->getOperand(0);
5648  bool IsArg0Constant = isa<UndefValue>(Arg0) || isa<ConstantInt>(Arg0) ||
5649  isa<ConstantFP>(Arg0);
5656  ScalarCost += TTI.getArithmeticInstrCost(
5657  Inst->getOpcode(), Inst->getType(), Arg0OVK, Arg1OVK);
5658  VectorCost += TTI.getArithmeticInstrCost(Inst->getOpcode(), PromotedType,
5659  Arg0OVK, Arg1OVK);
5660  }
5661  DEBUG(dbgs() << "Estimated cost of computation to be promoted:\nScalar: "
5662  << ScalarCost << "\nVector: " << VectorCost << '\n');
5663  return ScalarCost > VectorCost;
5664  }
5665 
5666  /// \brief Generate a constant vector with \p Val with the same
5667  /// number of elements as the transition.
5668  /// \p UseSplat defines whether or not \p Val should be replicated
5669  /// across the whole vector.
5670  /// In other words, if UseSplat == true, we generate <Val, Val, ..., Val>,
5671  /// otherwise we generate a vector with as many undef as possible:
5672  /// <undef, ..., undef, Val, undef, ..., undef> where \p Val is only
5673  /// used at the index of the extract.
5674  Value *getConstantVector(Constant *Val, bool UseSplat) const {
5675  unsigned ExtractIdx = std::numeric_limits<unsigned>::max();
5676  if (!UseSplat) {
5677  // If we cannot determine where the constant must be, we have to
5678  // use a splat constant.
5679  Value *ValExtractIdx = Transition->getOperand(getTransitionIdx());
5680  if (ConstantInt *CstVal = dyn_cast<ConstantInt>(ValExtractIdx))
5681  ExtractIdx = CstVal->getSExtValue();
5682  else
5683  UseSplat = true;
5684  }
5685 
5686  unsigned End = getTransitionType()->getVectorNumElements();
5687  if (UseSplat)
5688  return ConstantVector::getSplat(End, Val);
5689 
5690  SmallVector<Constant *, 4> ConstVec;
5691  UndefValue *UndefVal = UndefValue::get(Val->getType());
5692  for (unsigned Idx = 0; Idx != End; ++Idx) {
5693  if (Idx == ExtractIdx)
5694  ConstVec.push_back(Val);
5695  else
5696  ConstVec.push_back(UndefVal);
5697  }
5698  return ConstantVector::get(ConstVec);
5699  }
5700 
5701  /// \brief Check if promoting to a vector type an operand at \p OperandIdx
5702  /// in \p Use can trigger undefined behavior.
5703  static bool canCauseUndefinedBehavior(const Instruction *Use,
5704  unsigned OperandIdx) {
5705  // This is not safe to introduce undef when the operand is on
5706  // the right hand side of a division-like instruction.
5707  if (OperandIdx != 1)
5708  return false;
5709  switch (Use->getOpcode()) {
5710  default:
5711  return false;
5712  case Instruction::SDiv:
5713  case Instruction::UDiv:
5714  case Instruction::SRem:
5715  case Instruction::URem:
5716  return true;
5717  case Instruction::FDiv:
5718  case Instruction::FRem:
5719  return !Use->hasNoNaNs();
5720  }
5721  llvm_unreachable(nullptr);
5722  }
5723 
5724 public:
5725  VectorPromoteHelper(const DataLayout &DL, const TargetLowering &TLI,
5726  const TargetTransformInfo &TTI, Instruction *Transition,
5727  unsigned CombineCost)
5728  : DL(DL), TLI(TLI), TTI(TTI), Transition(Transition),
5729  StoreExtractCombineCost(CombineCost) {
5730  assert(Transition && "Do not know how to promote null");
5731  }
5732 
5733  /// \brief Check if we can promote \p ToBePromoted to \p Type.
5734  bool canPromote(const Instruction *ToBePromoted) const {
5735  // We could support CastInst too.
5736  return isa<BinaryOperator>(ToBePromoted);
5737  }
5738 
5739  /// \brief Check if it is profitable to promote \p ToBePromoted
5740  /// by moving downward the transition through.
5741  bool shouldPromote(const Instruction *ToBePromoted) const {
5742  // Promote only if all the operands can be statically expanded.
5743  // Indeed, we do not want to introduce any new kind of transitions.
5744  for (const Use &U : ToBePromoted->operands()) {
5745  const Value *Val = U.get();
5746  if (Val == getEndOfTransition()) {
5747  // If the use is a division and the transition is on the rhs,
5748  // we cannot promote the operation, otherwise we may create a
5749  // division by zero.
5750  if (canCauseUndefinedBehavior(ToBePromoted, U.getOperandNo()))
5751  return false;
5752  continue;
5753  }
5754  if (!isa<ConstantInt>(Val) && !isa<UndefValue>(Val) &&
5755  !isa<ConstantFP>(Val))
5756  return false;
5757  }
5758  // Check that the resulting operation is legal.
5759  int ISDOpcode = TLI.InstructionOpcodeToISD(ToBePromoted->getOpcode());
5760  if (!ISDOpcode)
5761  return false;
5762  return StressStoreExtract ||
5764  ISDOpcode, TLI.getValueType(DL, getTransitionType(), true));
5765  }
5766 
5767  /// \brief Check whether or not \p Use can be combined
5768  /// with the transition.
5769  /// I.e., is it possible to do Use(Transition) => AnotherUse?
5770  bool canCombine(const Instruction *Use) { return isa<StoreInst>(Use); }
5771 
5772  /// \brief Record \p ToBePromoted as part of the chain to be promoted.
5773  void enqueueForPromotion(Instruction *ToBePromoted) {
5774  InstsToBePromoted.push_back(ToBePromoted);
5775  }
5776 
5777  /// \brief Set the instruction that will be combined with the transition.
5778  void recordCombineInstruction(Instruction *ToBeCombined) {
5779  assert(canCombine(ToBeCombined) && "Unsupported instruction to combine");
5780  CombineInst = ToBeCombined;
5781  }
5782 
5783  /// \brief Promote all the instructions enqueued for promotion if it is
5784  /// is profitable.
5785  /// \return True if the promotion happened, false otherwise.
5786  bool promote() {
5787  // Check if there is something to promote.