LLVM  8.0.0svn
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"
35 #include "llvm/CodeGen/Analysis.h"
42 #include "llvm/Config/llvm-config.h"
43 #include "llvm/IR/Argument.h"
44 #include "llvm/IR/Attributes.h"
45 #include "llvm/IR/BasicBlock.h"
46 #include "llvm/IR/CallSite.h"
47 #include "llvm/IR/Constant.h"
48 #include "llvm/IR/Constants.h"
49 #include "llvm/IR/DataLayout.h"
50 #include "llvm/IR/DerivedTypes.h"
51 #include "llvm/IR/Dominators.h"
52 #include "llvm/IR/Function.h"
54 #include "llvm/IR/GlobalValue.h"
55 #include "llvm/IR/GlobalVariable.h"
56 #include "llvm/IR/IRBuilder.h"
57 #include "llvm/IR/InlineAsm.h"
58 #include "llvm/IR/InstrTypes.h"
59 #include "llvm/IR/Instruction.h"
60 #include "llvm/IR/Instructions.h"
61 #include "llvm/IR/IntrinsicInst.h"
62 #include "llvm/IR/Intrinsics.h"
63 #include "llvm/IR/LLVMContext.h"
64 #include "llvm/IR/MDBuilder.h"
65 #include "llvm/IR/Module.h"
66 #include "llvm/IR/Operator.h"
67 #include "llvm/IR/PatternMatch.h"
68 #include "llvm/IR/Statepoint.h"
69 #include "llvm/IR/Type.h"
70 #include "llvm/IR/Use.h"
71 #include "llvm/IR/User.h"
72 #include "llvm/IR/Value.h"
73 #include "llvm/IR/ValueHandle.h"
74 #include "llvm/IR/ValueMap.h"
75 #include "llvm/Pass.h"
78 #include "llvm/Support/Casting.h"
80 #include "llvm/Support/Compiler.h"
81 #include "llvm/Support/Debug.h"
91 #include <algorithm>
92 #include <cassert>
93 #include <cstdint>
94 #include <iterator>
95 #include <limits>
96 #include <memory>
97 #include <utility>
98 #include <vector>
99 
100 using namespace llvm;
101 using namespace llvm::PatternMatch;
102 
103 #define DEBUG_TYPE "codegenprepare"
104 
105 STATISTIC(NumBlocksElim, "Number of blocks eliminated");
106 STATISTIC(NumPHIsElim, "Number of trivial PHIs eliminated");
107 STATISTIC(NumGEPsElim, "Number of GEPs converted to casts");
108 STATISTIC(NumCmpUses, "Number of uses of Cmp expressions replaced with uses of "
109  "sunken Cmps");
110 STATISTIC(NumCastUses, "Number of uses of Cast expressions replaced with uses "
111  "of sunken Casts");
112 STATISTIC(NumMemoryInsts, "Number of memory instructions whose address "
113  "computations were sunk");
114 STATISTIC(NumMemoryInstsPhiCreated,
115  "Number of phis created when address "
116  "computations were sunk to memory instructions");
117 STATISTIC(NumMemoryInstsSelectCreated,
118  "Number of select created when address "
119  "computations were sunk to memory instructions");
120 STATISTIC(NumExtsMoved, "Number of [s|z]ext instructions combined with loads");
121 STATISTIC(NumExtUses, "Number of uses of [s|z]ext instructions optimized");
122 STATISTIC(NumAndsAdded,
123  "Number of and mask instructions added to form ext loads");
124 STATISTIC(NumAndUses, "Number of uses of and mask instructions optimized");
125 STATISTIC(NumRetsDup, "Number of return instructions duplicated");
126 STATISTIC(NumDbgValueMoved, "Number of debug value instructions moved");
127 STATISTIC(NumSelectsExpanded, "Number of selects turned into branches");
128 STATISTIC(NumStoreExtractExposed, "Number of store(extractelement) exposed");
129 
131  "disable-cgp-branch-opts", cl::Hidden, cl::init(false),
132  cl::desc("Disable branch optimizations in CodeGenPrepare"));
133 
134 static cl::opt<bool>
135  DisableGCOpts("disable-cgp-gc-opts", cl::Hidden, cl::init(false),
136  cl::desc("Disable GC optimizations in CodeGenPrepare"));
137 
139  "disable-cgp-select2branch", cl::Hidden, cl::init(false),
140  cl::desc("Disable select to branch conversion."));
141 
143  "addr-sink-using-gep", cl::Hidden, cl::init(true),
144  cl::desc("Address sinking in CGP using GEPs."));
145 
147  "enable-andcmp-sinking", cl::Hidden, cl::init(true),
148  cl::desc("Enable sinkinig and/cmp into branches."));
149 
151  "disable-cgp-store-extract", cl::Hidden, cl::init(false),
152  cl::desc("Disable store(extract) optimizations in CodeGenPrepare"));
153 
155  "stress-cgp-store-extract", cl::Hidden, cl::init(false),
156  cl::desc("Stress test store(extract) optimizations in CodeGenPrepare"));
157 
159  "disable-cgp-ext-ld-promotion", cl::Hidden, cl::init(false),
160  cl::desc("Disable ext(promotable(ld)) -> promoted(ext(ld)) optimization in "
161  "CodeGenPrepare"));
162 
164  "stress-cgp-ext-ld-promotion", cl::Hidden, cl::init(false),
165  cl::desc("Stress test ext(promotable(ld)) -> promoted(ext(ld)) "
166  "optimization in CodeGenPrepare"));
167 
169  "disable-preheader-prot", cl::Hidden, cl::init(false),
170  cl::desc("Disable protection against removing loop preheaders"));
171 
173  "profile-guided-section-prefix", cl::Hidden, cl::init(true), cl::ZeroOrMore,
174  cl::desc("Use profile info to add section prefix for hot/cold functions"));
175 
177  "cgp-freq-ratio-to-skip-merge", cl::Hidden, cl::init(2),
178  cl::desc("Skip merging empty blocks if (frequency of empty block) / "
179  "(frequency of destination block) is greater than this ratio"));
180 
182  "force-split-store", cl::Hidden, cl::init(false),
183  cl::desc("Force store splitting no matter what the target query says."));
184 
185 static cl::opt<bool>
186 EnableTypePromotionMerge("cgp-type-promotion-merge", cl::Hidden,
187  cl::desc("Enable merging of redundant sexts when one is dominating"
188  " the other."), cl::init(true));
189 
191  "disable-complex-addr-modes", cl::Hidden, cl::init(false),
192  cl::desc("Disables combining addressing modes with different parts "
193  "in optimizeMemoryInst."));
194 
195 static cl::opt<bool>
196 AddrSinkNewPhis("addr-sink-new-phis", cl::Hidden, cl::init(false),
197  cl::desc("Allow creation of Phis in Address sinking."));
198 
199 static cl::opt<bool>
200 AddrSinkNewSelects("addr-sink-new-select", cl::Hidden, cl::init(true),
201  cl::desc("Allow creation of selects in Address sinking."));
202 
204  "addr-sink-combine-base-reg", cl::Hidden, cl::init(true),
205  cl::desc("Allow combining of BaseReg field in Address sinking."));
206 
208  "addr-sink-combine-base-gv", cl::Hidden, cl::init(true),
209  cl::desc("Allow combining of BaseGV field in Address sinking."));
210 
212  "addr-sink-combine-base-offs", cl::Hidden, cl::init(true),
213  cl::desc("Allow combining of BaseOffs field in Address sinking."));
214 
216  "addr-sink-combine-scaled-reg", cl::Hidden, cl::init(true),
217  cl::desc("Allow combining of ScaledReg field in Address sinking."));
218 
219 static cl::opt<bool>
220  EnableGEPOffsetSplit("cgp-split-large-offset-gep", cl::Hidden,
221  cl::init(true),
222  cl::desc("Enable splitting large offset of GEP."));
223 
224 namespace {
225 
226 enum ExtType {
227  ZeroExtension, // Zero extension has been seen.
228  SignExtension, // Sign extension has been seen.
229  BothExtension // This extension type is used if we saw sext after
230  // ZeroExtension had been set, or if we saw zext after
231  // SignExtension had been set. It makes the type
232  // information of a promoted instruction invalid.
233 };
234 
235 using SetOfInstrs = SmallPtrSet<Instruction *, 16>;
236 using TypeIsSExt = PointerIntPair<Type *, 2, ExtType>;
237 using InstrToOrigTy = DenseMap<Instruction *, TypeIsSExt>;
238 using SExts = SmallVector<Instruction *, 16>;
239 using ValueToSExts = DenseMap<Value *, SExts>;
240 
241 class TypePromotionTransaction;
242 
243  class CodeGenPrepare : public FunctionPass {
244  const TargetMachine *TM = nullptr;
245  const TargetSubtargetInfo *SubtargetInfo;
246  const TargetLowering *TLI = nullptr;
247  const TargetRegisterInfo *TRI;
248  const TargetTransformInfo *TTI = nullptr;
249  const TargetLibraryInfo *TLInfo;
250  const LoopInfo *LI;
251  std::unique_ptr<BlockFrequencyInfo> BFI;
252  std::unique_ptr<BranchProbabilityInfo> BPI;
253 
254  /// As we scan instructions optimizing them, this is the next instruction
255  /// to optimize. Transforms that can invalidate this should update it.
256  BasicBlock::iterator CurInstIterator;
257 
258  /// Keeps track of non-local addresses that have been sunk into a block.
259  /// This allows us to avoid inserting duplicate code for blocks with
260  /// multiple load/stores of the same address. The usage of WeakTrackingVH
261  /// enables SunkAddrs to be treated as a cache whose entries can be
262  /// invalidated if a sunken address computation has been erased.
264 
265  /// Keeps track of all instructions inserted for the current function.
266  SetOfInstrs InsertedInsts;
267 
268  /// Keeps track of the type of the related instruction before their
269  /// promotion for the current function.
270  InstrToOrigTy PromotedInsts;
271 
272  /// Keep track of instructions removed during promotion.
273  SetOfInstrs RemovedInsts;
274 
275  /// Keep track of sext chains based on their initial value.
276  DenseMap<Value *, Instruction *> SeenChainsForSExt;
277 
278  /// Keep track of GEPs accessing the same data structures such as structs or
279  /// arrays that are candidates to be split later because of their large
280  /// size.
281  MapVector<
284  LargeOffsetGEPMap;
285 
286  /// Keep track of new GEP base after splitting the GEPs having large offset.
287  SmallSet<AssertingVH<Value>, 2> NewGEPBases;
288 
289  /// Map serial numbers to Large offset GEPs.
290  DenseMap<AssertingVH<GetElementPtrInst>, int> LargeOffsetGEPID;
291 
292  /// Keep track of SExt promoted.
293  ValueToSExts ValToSExtendedUses;
294 
295  /// True if CFG is modified in any way.
296  bool ModifiedDT;
297 
298  /// True if optimizing for size.
299  bool OptSize;
300 
301  /// DataLayout for the Function being processed.
302  const DataLayout *DL = nullptr;
303 
304  public:
305  static char ID; // Pass identification, replacement for typeid
306 
307  CodeGenPrepare() : FunctionPass(ID) {
309  }
310 
311  bool runOnFunction(Function &F) override;
312 
313  StringRef getPassName() const override { return "CodeGen Prepare"; }
314 
315  void getAnalysisUsage(AnalysisUsage &AU) const override {
316  // FIXME: When we can selectively preserve passes, preserve the domtree.
321  }
322 
323  private:
324  bool eliminateFallThrough(Function &F);
325  bool eliminateMostlyEmptyBlocks(Function &F);
326  BasicBlock *findDestBlockOfMergeableEmptyBlock(BasicBlock *BB);
327  bool canMergeBlocks(const BasicBlock *BB, const BasicBlock *DestBB) const;
328  void eliminateMostlyEmptyBlock(BasicBlock *BB);
329  bool isMergingEmptyBlockProfitable(BasicBlock *BB, BasicBlock *DestBB,
330  bool isPreheader);
331  bool optimizeBlock(BasicBlock &BB, bool &ModifiedDT);
332  bool optimizeInst(Instruction *I, bool &ModifiedDT);
333  bool optimizeMemoryInst(Instruction *MemoryInst, Value *Addr,
334  Type *AccessTy, unsigned AddrSpace);
335  bool optimizeInlineAsmInst(CallInst *CS);
336  bool optimizeCallInst(CallInst *CI, bool &ModifiedDT);
337  bool optimizeExt(Instruction *&I);
338  bool optimizeExtUses(Instruction *I);
339  bool optimizeLoadExt(LoadInst *Load);
340  bool optimizeSelectInst(SelectInst *SI);
341  bool optimizeShuffleVectorInst(ShuffleVectorInst *SVI);
342  bool optimizeSwitchInst(SwitchInst *SI);
343  bool optimizeExtractElementInst(Instruction *Inst);
344  bool dupRetToEnableTailCallOpts(BasicBlock *BB);
345  bool placeDbgValues(Function &F);
346  bool canFormExtLd(const SmallVectorImpl<Instruction *> &MovedExts,
347  LoadInst *&LI, Instruction *&Inst, bool HasPromoted);
348  bool tryToPromoteExts(TypePromotionTransaction &TPT,
349  const SmallVectorImpl<Instruction *> &Exts,
350  SmallVectorImpl<Instruction *> &ProfitablyMovedExts,
351  unsigned CreatedInstsCost = 0);
352  bool mergeSExts(Function &F);
353  bool splitLargeGEPOffsets();
354  bool performAddressTypePromotion(
355  Instruction *&Inst,
356  bool AllowPromotionWithoutCommonHeader,
357  bool HasPromoted, TypePromotionTransaction &TPT,
358  SmallVectorImpl<Instruction *> &SpeculativelyMovedExts);
359  bool splitBranchCondition(Function &F);
360  bool simplifyOffsetableRelocate(Instruction &I);
361  };
362 
363 } // end anonymous namespace
364 
365 char CodeGenPrepare::ID = 0;
366 
367 INITIALIZE_PASS_BEGIN(CodeGenPrepare, DEBUG_TYPE,
368  "Optimize for code generation", false, false)
371  "Optimize for code generation", false, false)
372 
373 FunctionPass *llvm::createCodeGenPreparePass() { return new CodeGenPrepare(); }
374 
376  if (skipFunction(F))
377  return false;
378 
379  DL = &F.getParent()->getDataLayout();
380 
381  bool EverMadeChange = false;
382  // Clear per function information.
383  InsertedInsts.clear();
384  PromotedInsts.clear();
385 
386  ModifiedDT = false;
387  if (auto *TPC = getAnalysisIfAvailable<TargetPassConfig>()) {
388  TM = &TPC->getTM<TargetMachine>();
389  SubtargetInfo = TM->getSubtargetImpl(F);
390  TLI = SubtargetInfo->getTargetLowering();
391  TRI = SubtargetInfo->getRegisterInfo();
392  }
393  TLInfo = &getAnalysis<TargetLibraryInfoWrapperPass>().getTLI();
394  TTI = &getAnalysis<TargetTransformInfoWrapperPass>().getTTI(F);
395  LI = &getAnalysis<LoopInfoWrapperPass>().getLoopInfo();
396  BPI.reset(new BranchProbabilityInfo(F, *LI));
397  BFI.reset(new BlockFrequencyInfo(F, *BPI, *LI));
398  OptSize = F.optForSize();
399 
400  ProfileSummaryInfo *PSI =
401  getAnalysis<ProfileSummaryInfoWrapperPass>().getPSI();
403  if (PSI->isFunctionHotInCallGraph(&F, *BFI))
404  F.setSectionPrefix(".hot");
405  else if (PSI->isFunctionColdInCallGraph(&F, *BFI))
406  F.setSectionPrefix(".unlikely");
407  }
408 
409  /// This optimization identifies DIV instructions that can be
410  /// profitably bypassed and carried out with a shorter, faster divide.
411  if (!OptSize && !PSI->hasHugeWorkingSetSize() && TLI &&
412  TLI->isSlowDivBypassed()) {
413  const DenseMap<unsigned int, unsigned int> &BypassWidths =
414  TLI->getBypassSlowDivWidths();
415  BasicBlock* BB = &*F.begin();
416  while (BB != nullptr) {
417  // bypassSlowDivision may create new BBs, but we don't want to reapply the
418  // optimization to those blocks.
419  BasicBlock* Next = BB->getNextNode();
420  EverMadeChange |= bypassSlowDivision(BB, BypassWidths);
421  BB = Next;
422  }
423  }
424 
425  // Eliminate blocks that contain only PHI nodes and an
426  // unconditional branch.
427  EverMadeChange |= eliminateMostlyEmptyBlocks(F);
428 
429  if (!DisableBranchOpts)
430  EverMadeChange |= splitBranchCondition(F);
431 
432  // Split some critical edges where one of the sources is an indirect branch,
433  // to help generate sane code for PHIs involving such edges.
434  EverMadeChange |= SplitIndirectBrCriticalEdges(F);
435 
436  bool MadeChange = true;
437  while (MadeChange) {
438  MadeChange = false;
439  SeenChainsForSExt.clear();
440  ValToSExtendedUses.clear();
441  RemovedInsts.clear();
442  LargeOffsetGEPMap.clear();
443  LargeOffsetGEPID.clear();
444  for (Function::iterator I = F.begin(); I != F.end(); ) {
445  BasicBlock *BB = &*I++;
446  bool ModifiedDTOnIteration = false;
447  MadeChange |= optimizeBlock(*BB, ModifiedDTOnIteration);
448 
449  // Restart BB iteration if the dominator tree of the Function was changed
450  if (ModifiedDTOnIteration)
451  break;
452  }
453  if (EnableTypePromotionMerge && !ValToSExtendedUses.empty())
454  MadeChange |= mergeSExts(F);
455  if (!LargeOffsetGEPMap.empty())
456  MadeChange |= splitLargeGEPOffsets();
457 
458  // Really free removed instructions during promotion.
459  for (Instruction *I : RemovedInsts)
460  I->deleteValue();
461 
462  EverMadeChange |= MadeChange;
463  }
464 
465  SunkAddrs.clear();
466 
467  if (!DisableBranchOpts) {
468  MadeChange = false;
469  // Use a set vector to get deterministic iteration order. The order the
470  // blocks are removed may affect whether or not PHI nodes in successors
471  // are removed.
473  for (BasicBlock &BB : F) {
474  SmallVector<BasicBlock *, 2> Successors(succ_begin(&BB), succ_end(&BB));
475  MadeChange |= ConstantFoldTerminator(&BB, true);
476  if (!MadeChange) continue;
477 
479  II = Successors.begin(), IE = Successors.end(); II != IE; ++II)
480  if (pred_begin(*II) == pred_end(*II))
481  WorkList.insert(*II);
482  }
483 
484  // Delete the dead blocks and any of their dead successors.
485  MadeChange |= !WorkList.empty();
486  while (!WorkList.empty()) {
487  BasicBlock *BB = WorkList.pop_back_val();
488  SmallVector<BasicBlock*, 2> Successors(succ_begin(BB), succ_end(BB));
489 
490  DeleteDeadBlock(BB);
491 
493  II = Successors.begin(), IE = Successors.end(); II != IE; ++II)
494  if (pred_begin(*II) == pred_end(*II))
495  WorkList.insert(*II);
496  }
497 
498  // Merge pairs of basic blocks with unconditional branches, connected by
499  // a single edge.
500  if (EverMadeChange || MadeChange)
501  MadeChange |= eliminateFallThrough(F);
502 
503  EverMadeChange |= MadeChange;
504  }
505 
506  if (!DisableGCOpts) {
507  SmallVector<Instruction *, 2> Statepoints;
508  for (BasicBlock &BB : F)
509  for (Instruction &I : BB)
510  if (isStatepoint(I))
511  Statepoints.push_back(&I);
512  for (auto &I : Statepoints)
513  EverMadeChange |= simplifyOffsetableRelocate(*I);
514  }
515 
516  // Do this last to clean up use-before-def scenarios introduced by other
517  // preparatory transforms.
518  EverMadeChange |= placeDbgValues(F);
519 
520  return EverMadeChange;
521 }
522 
523 /// Merge basic blocks which are connected by a single edge, where one of the
524 /// basic blocks has a single successor pointing to the other basic block,
525 /// which has a single predecessor.
526 bool CodeGenPrepare::eliminateFallThrough(Function &F) {
527  bool Changed = false;
528  // Scan all of the blocks in the function, except for the entry block.
529  // Use a temporary array to avoid iterator being invalidated when
530  // deleting blocks.
532  for (auto &Block : llvm::make_range(std::next(F.begin()), F.end()))
533  Blocks.push_back(&Block);
534 
535  for (auto &Block : Blocks) {
536  auto *BB = cast_or_null<BasicBlock>(Block);
537  if (!BB)
538  continue;
539  // If the destination block has a single pred, then this is a trivial
540  // edge, just collapse it.
541  BasicBlock *SinglePred = BB->getSinglePredecessor();
542 
543  // Don't merge if BB's address is taken.
544  if (!SinglePred || SinglePred == BB || BB->hasAddressTaken()) continue;
545 
546  BranchInst *Term = dyn_cast<BranchInst>(SinglePred->getTerminator());
547  if (Term && !Term->isConditional()) {
548  Changed = true;
549  LLVM_DEBUG(dbgs() << "To merge:\n" << *BB << "\n\n\n");
550 
551  // Merge BB into SinglePred and delete it.
553  }
554  }
555  return Changed;
556 }
557 
558 /// Find a destination block from BB if BB is mergeable empty block.
559 BasicBlock *CodeGenPrepare::findDestBlockOfMergeableEmptyBlock(BasicBlock *BB) {
560  // If this block doesn't end with an uncond branch, ignore it.
562  if (!BI || !BI->isUnconditional())
563  return nullptr;
564 
565  // If the instruction before the branch (skipping debug info) isn't a phi
566  // node, then other stuff is happening here.
567  BasicBlock::iterator BBI = BI->getIterator();
568  if (BBI != BB->begin()) {
569  --BBI;
570  while (isa<DbgInfoIntrinsic>(BBI)) {
571  if (BBI == BB->begin())
572  break;
573  --BBI;
574  }
575  if (!isa<DbgInfoIntrinsic>(BBI) && !isa<PHINode>(BBI))
576  return nullptr;
577  }
578 
579  // Do not break infinite loops.
580  BasicBlock *DestBB = BI->getSuccessor(0);
581  if (DestBB == BB)
582  return nullptr;
583 
584  if (!canMergeBlocks(BB, DestBB))
585  DestBB = nullptr;
586 
587  return DestBB;
588 }
589 
590 /// Eliminate blocks that contain only PHI nodes, debug info directives, and an
591 /// unconditional branch. Passes before isel (e.g. LSR/loopsimplify) often split
592 /// edges in ways that are non-optimal for isel. Start by eliminating these
593 /// blocks so we can split them the way we want them.
594 bool CodeGenPrepare::eliminateMostlyEmptyBlocks(Function &F) {
596  SmallVector<Loop *, 16> LoopList(LI->begin(), LI->end());
597  while (!LoopList.empty()) {
598  Loop *L = LoopList.pop_back_val();
599  LoopList.insert(LoopList.end(), L->begin(), L->end());
600  if (BasicBlock *Preheader = L->getLoopPreheader())
601  Preheaders.insert(Preheader);
602  }
603 
604  bool MadeChange = false;
605  // Copy blocks into a temporary array to avoid iterator invalidation issues
606  // as we remove them.
607  // Note that this intentionally skips the entry block.
609  for (auto &Block : llvm::make_range(std::next(F.begin()), F.end()))
610  Blocks.push_back(&Block);
611 
612  for (auto &Block : Blocks) {
613  BasicBlock *BB = cast_or_null<BasicBlock>(Block);
614  if (!BB)
615  continue;
616  BasicBlock *DestBB = findDestBlockOfMergeableEmptyBlock(BB);
617  if (!DestBB ||
618  !isMergingEmptyBlockProfitable(BB, DestBB, Preheaders.count(BB)))
619  continue;
620 
621  eliminateMostlyEmptyBlock(BB);
622  MadeChange = true;
623  }
624  return MadeChange;
625 }
626 
627 bool CodeGenPrepare::isMergingEmptyBlockProfitable(BasicBlock *BB,
628  BasicBlock *DestBB,
629  bool isPreheader) {
630  // Do not delete loop preheaders if doing so would create a critical edge.
631  // Loop preheaders can be good locations to spill registers. If the
632  // preheader is deleted and we create a critical edge, registers may be
633  // spilled in the loop body instead.
634  if (!DisablePreheaderProtect && isPreheader &&
635  !(BB->getSinglePredecessor() &&
637  return false;
638 
639  // Try to skip merging if the unique predecessor of BB is terminated by a
640  // switch or indirect branch instruction, and BB is used as an incoming block
641  // of PHIs in DestBB. In such case, merging BB and DestBB would cause ISel to
642  // add COPY instructions in the predecessor of BB instead of BB (if it is not
643  // merged). Note that the critical edge created by merging such blocks wont be
644  // split in MachineSink because the jump table is not analyzable. By keeping
645  // such empty block (BB), ISel will place COPY instructions in BB, not in the
646  // predecessor of BB.
647  BasicBlock *Pred = BB->getUniquePredecessor();
648  if (!Pred ||
649  !(isa<SwitchInst>(Pred->getTerminator()) ||
650  isa<IndirectBrInst>(Pred->getTerminator())))
651  return true;
652 
653  if (BB->getTerminator() != BB->getFirstNonPHIOrDbg())
654  return true;
655 
656  // We use a simple cost heuristic which determine skipping merging is
657  // profitable if the cost of skipping merging is less than the cost of
658  // merging : Cost(skipping merging) < Cost(merging BB), where the
659  // Cost(skipping merging) is Freq(BB) * (Cost(Copy) + Cost(Branch)), and
660  // the Cost(merging BB) is Freq(Pred) * Cost(Copy).
661  // Assuming Cost(Copy) == Cost(Branch), we could simplify it to :
662  // Freq(Pred) / Freq(BB) > 2.
663  // Note that if there are multiple empty blocks sharing the same incoming
664  // value for the PHIs in the DestBB, we consider them together. In such
665  // case, Cost(merging BB) will be the sum of their frequencies.
666 
667  if (!isa<PHINode>(DestBB->begin()))
668  return true;
669 
670  SmallPtrSet<BasicBlock *, 16> SameIncomingValueBBs;
671 
672  // Find all other incoming blocks from which incoming values of all PHIs in
673  // DestBB are the same as the ones from BB.
674  for (pred_iterator PI = pred_begin(DestBB), E = pred_end(DestBB); PI != E;
675  ++PI) {
676  BasicBlock *DestBBPred = *PI;
677  if (DestBBPred == BB)
678  continue;
679 
680  if (llvm::all_of(DestBB->phis(), [&](const PHINode &DestPN) {
681  return DestPN.getIncomingValueForBlock(BB) ==
682  DestPN.getIncomingValueForBlock(DestBBPred);
683  }))
684  SameIncomingValueBBs.insert(DestBBPred);
685  }
686 
687  // See if all BB's incoming values are same as the value from Pred. In this
688  // case, no reason to skip merging because COPYs are expected to be place in
689  // Pred already.
690  if (SameIncomingValueBBs.count(Pred))
691  return true;
692 
693  BlockFrequency PredFreq = BFI->getBlockFreq(Pred);
694  BlockFrequency BBFreq = BFI->getBlockFreq(BB);
695 
696  for (auto SameValueBB : SameIncomingValueBBs)
697  if (SameValueBB->getUniquePredecessor() == Pred &&
698  DestBB == findDestBlockOfMergeableEmptyBlock(SameValueBB))
699  BBFreq += BFI->getBlockFreq(SameValueBB);
700 
701  return PredFreq.getFrequency() <=
703 }
704 
705 /// Return true if we can merge BB into DestBB if there is a single
706 /// unconditional branch between them, and BB contains no other non-phi
707 /// instructions.
708 bool CodeGenPrepare::canMergeBlocks(const BasicBlock *BB,
709  const BasicBlock *DestBB) const {
710  // We only want to eliminate blocks whose phi nodes are used by phi nodes in
711  // the successor. If there are more complex condition (e.g. preheaders),
712  // don't mess around with them.
713  for (const PHINode &PN : BB->phis()) {
714  for (const User *U : PN.users()) {
715  const Instruction *UI = cast<Instruction>(U);
716  if (UI->getParent() != DestBB || !isa<PHINode>(UI))
717  return false;
718  // If User is inside DestBB block and it is a PHINode then check
719  // incoming value. If incoming value is not from BB then this is
720  // a complex condition (e.g. preheaders) we want to avoid here.
721  if (UI->getParent() == DestBB) {
722  if (const PHINode *UPN = dyn_cast<PHINode>(UI))
723  for (unsigned I = 0, E = UPN->getNumIncomingValues(); I != E; ++I) {
724  Instruction *Insn = dyn_cast<Instruction>(UPN->getIncomingValue(I));
725  if (Insn && Insn->getParent() == BB &&
726  Insn->getParent() != UPN->getIncomingBlock(I))
727  return false;
728  }
729  }
730  }
731  }
732 
733  // If BB and DestBB contain any common predecessors, then the phi nodes in BB
734  // and DestBB may have conflicting incoming values for the block. If so, we
735  // can't merge the block.
736  const PHINode *DestBBPN = dyn_cast<PHINode>(DestBB->begin());
737  if (!DestBBPN) return true; // no conflict.
738 
739  // Collect the preds of BB.
741  if (const PHINode *BBPN = dyn_cast<PHINode>(BB->begin())) {
742  // It is faster to get preds from a PHI than with pred_iterator.
743  for (unsigned i = 0, e = BBPN->getNumIncomingValues(); i != e; ++i)
744  BBPreds.insert(BBPN->getIncomingBlock(i));
745  } else {
746  BBPreds.insert(pred_begin(BB), pred_end(BB));
747  }
748 
749  // Walk the preds of DestBB.
750  for (unsigned i = 0, e = DestBBPN->getNumIncomingValues(); i != e; ++i) {
751  BasicBlock *Pred = DestBBPN->getIncomingBlock(i);
752  if (BBPreds.count(Pred)) { // Common predecessor?
753  for (const PHINode &PN : DestBB->phis()) {
754  const Value *V1 = PN.getIncomingValueForBlock(Pred);
755  const Value *V2 = PN.getIncomingValueForBlock(BB);
756 
757  // If V2 is a phi node in BB, look up what the mapped value will be.
758  if (const PHINode *V2PN = dyn_cast<PHINode>(V2))
759  if (V2PN->getParent() == BB)
760  V2 = V2PN->getIncomingValueForBlock(Pred);
761 
762  // If there is a conflict, bail out.
763  if (V1 != V2) return false;
764  }
765  }
766  }
767 
768  return true;
769 }
770 
771 /// Eliminate a basic block that has only phi's and an unconditional branch in
772 /// it.
773 void CodeGenPrepare::eliminateMostlyEmptyBlock(BasicBlock *BB) {
774  BranchInst *BI = cast<BranchInst>(BB->getTerminator());
775  BasicBlock *DestBB = BI->getSuccessor(0);
776 
777  LLVM_DEBUG(dbgs() << "MERGING MOSTLY EMPTY BLOCKS - BEFORE:\n"
778  << *BB << *DestBB);
779 
780  // If the destination block has a single pred, then this is a trivial edge,
781  // just collapse it.
782  if (BasicBlock *SinglePred = DestBB->getSinglePredecessor()) {
783  if (SinglePred != DestBB) {
784  assert(SinglePred == BB &&
785  "Single predecessor not the same as predecessor");
786  // Merge DestBB into SinglePred/BB and delete it.
788  // Note: BB(=SinglePred) will not be deleted on this path.
789  // DestBB(=its single successor) is the one that was deleted.
790  LLVM_DEBUG(dbgs() << "AFTER:\n" << *SinglePred << "\n\n\n");
791  return;
792  }
793  }
794 
795  // Otherwise, we have multiple predecessors of BB. Update the PHIs in DestBB
796  // to handle the new incoming edges it is about to have.
797  for (PHINode &PN : DestBB->phis()) {
798  // Remove the incoming value for BB, and remember it.
799  Value *InVal = PN.removeIncomingValue(BB, false);
800 
801  // Two options: either the InVal is a phi node defined in BB or it is some
802  // value that dominates BB.
803  PHINode *InValPhi = dyn_cast<PHINode>(InVal);
804  if (InValPhi && InValPhi->getParent() == BB) {
805  // Add all of the input values of the input PHI as inputs of this phi.
806  for (unsigned i = 0, e = InValPhi->getNumIncomingValues(); i != e; ++i)
807  PN.addIncoming(InValPhi->getIncomingValue(i),
808  InValPhi->getIncomingBlock(i));
809  } else {
810  // Otherwise, add one instance of the dominating value for each edge that
811  // we will be adding.
812  if (PHINode *BBPN = dyn_cast<PHINode>(BB->begin())) {
813  for (unsigned i = 0, e = BBPN->getNumIncomingValues(); i != e; ++i)
814  PN.addIncoming(InVal, BBPN->getIncomingBlock(i));
815  } else {
816  for (pred_iterator PI = pred_begin(BB), E = pred_end(BB); PI != E; ++PI)
817  PN.addIncoming(InVal, *PI);
818  }
819  }
820  }
821 
822  // The PHIs are now updated, change everything that refers to BB to use
823  // DestBB and remove BB.
824  BB->replaceAllUsesWith(DestBB);
825  BB->eraseFromParent();
826  ++NumBlocksElim;
827 
828  LLVM_DEBUG(dbgs() << "AFTER:\n" << *DestBB << "\n\n\n");
829 }
830 
831 // Computes a map of base pointer relocation instructions to corresponding
832 // derived pointer relocation instructions given a vector of all relocate calls
834  const SmallVectorImpl<GCRelocateInst *> &AllRelocateCalls,
836  &RelocateInstMap) {
837  // Collect information in two maps: one primarily for locating the base object
838  // while filling the second map; the second map is the final structure holding
839  // a mapping between Base and corresponding Derived relocate calls
841  for (auto *ThisRelocate : AllRelocateCalls) {
842  auto K = std::make_pair(ThisRelocate->getBasePtrIndex(),
843  ThisRelocate->getDerivedPtrIndex());
844  RelocateIdxMap.insert(std::make_pair(K, ThisRelocate));
845  }
846  for (auto &Item : RelocateIdxMap) {
847  std::pair<unsigned, unsigned> Key = Item.first;
848  if (Key.first == Key.second)
849  // Base relocation: nothing to insert
850  continue;
851 
852  GCRelocateInst *I = Item.second;
853  auto BaseKey = std::make_pair(Key.first, Key.first);
854 
855  // We're iterating over RelocateIdxMap so we cannot modify it.
856  auto MaybeBase = RelocateIdxMap.find(BaseKey);
857  if (MaybeBase == RelocateIdxMap.end())
858  // TODO: We might want to insert a new base object relocate and gep off
859  // that, if there are enough derived object relocates.
860  continue;
861 
862  RelocateInstMap[MaybeBase->second].push_back(I);
863  }
864 }
865 
866 // Accepts a GEP and extracts the operands into a vector provided they're all
867 // small integer constants
869  SmallVectorImpl<Value *> &OffsetV) {
870  for (unsigned i = 1; i < GEP->getNumOperands(); i++) {
871  // Only accept small constant integer operands
872  auto Op = dyn_cast<ConstantInt>(GEP->getOperand(i));
873  if (!Op || Op->getZExtValue() > 20)
874  return false;
875  }
876 
877  for (unsigned i = 1; i < GEP->getNumOperands(); i++)
878  OffsetV.push_back(GEP->getOperand(i));
879  return true;
880 }
881 
882 // Takes a RelocatedBase (base pointer relocation instruction) and Targets to
883 // replace, computes a replacement, and affects it.
884 static bool
886  const SmallVectorImpl<GCRelocateInst *> &Targets) {
887  bool MadeChange = false;
888  // We must ensure the relocation of derived pointer is defined after
889  // relocation of base pointer. If we find a relocation corresponding to base
890  // defined earlier than relocation of base then we move relocation of base
891  // right before found relocation. We consider only relocation in the same
892  // basic block as relocation of base. Relocations from other basic block will
893  // be skipped by optimization and we do not care about them.
894  for (auto R = RelocatedBase->getParent()->getFirstInsertionPt();
895  &*R != RelocatedBase; ++R)
896  if (auto RI = dyn_cast<GCRelocateInst>(R))
897  if (RI->getStatepoint() == RelocatedBase->getStatepoint())
898  if (RI->getBasePtrIndex() == RelocatedBase->getBasePtrIndex()) {
899  RelocatedBase->moveBefore(RI);
900  break;
901  }
902 
903  for (GCRelocateInst *ToReplace : Targets) {
904  assert(ToReplace->getBasePtrIndex() == RelocatedBase->getBasePtrIndex() &&
905  "Not relocating a derived object of the original base object");
906  if (ToReplace->getBasePtrIndex() == ToReplace->getDerivedPtrIndex()) {
907  // A duplicate relocate call. TODO: coalesce duplicates.
908  continue;
909  }
910 
911  if (RelocatedBase->getParent() != ToReplace->getParent()) {
912  // Base and derived relocates are in different basic blocks.
913  // In this case transform is only valid when base dominates derived
914  // relocate. However it would be too expensive to check dominance
915  // for each such relocate, so we skip the whole transformation.
916  continue;
917  }
918 
919  Value *Base = ToReplace->getBasePtr();
920  auto Derived = dyn_cast<GetElementPtrInst>(ToReplace->getDerivedPtr());
921  if (!Derived || Derived->getPointerOperand() != Base)
922  continue;
923 
924  SmallVector<Value *, 2> OffsetV;
925  if (!getGEPSmallConstantIntOffsetV(Derived, OffsetV))
926  continue;
927 
928  // Create a Builder and replace the target callsite with a gep
929  assert(RelocatedBase->getNextNode() &&
930  "Should always have one since it's not a terminator");
931 
932  // Insert after RelocatedBase
933  IRBuilder<> Builder(RelocatedBase->getNextNode());
934  Builder.SetCurrentDebugLocation(ToReplace->getDebugLoc());
935 
936  // If gc_relocate does not match the actual type, cast it to the right type.
937  // In theory, there must be a bitcast after gc_relocate if the type does not
938  // match, and we should reuse it to get the derived pointer. But it could be
939  // cases like this:
940  // bb1:
941  // ...
942  // %g1 = call coldcc i8 addrspace(1)* @llvm.experimental.gc.relocate.p1i8(...)
943  // br label %merge
944  //
945  // bb2:
946  // ...
947  // %g2 = call coldcc i8 addrspace(1)* @llvm.experimental.gc.relocate.p1i8(...)
948  // br label %merge
949  //
950  // merge:
951  // %p1 = phi i8 addrspace(1)* [ %g1, %bb1 ], [ %g2, %bb2 ]
952  // %cast = bitcast i8 addrspace(1)* %p1 in to i32 addrspace(1)*
953  //
954  // In this case, we can not find the bitcast any more. So we insert a new bitcast
955  // no matter there is already one or not. In this way, we can handle all cases, and
956  // the extra bitcast should be optimized away in later passes.
957  Value *ActualRelocatedBase = RelocatedBase;
958  if (RelocatedBase->getType() != Base->getType()) {
959  ActualRelocatedBase =
960  Builder.CreateBitCast(RelocatedBase, Base->getType());
961  }
962  Value *Replacement = Builder.CreateGEP(
963  Derived->getSourceElementType(), ActualRelocatedBase, makeArrayRef(OffsetV));
964  Replacement->takeName(ToReplace);
965  // If the newly generated derived pointer's type does not match the original derived
966  // pointer's type, cast the new derived pointer to match it. Same reasoning as above.
967  Value *ActualReplacement = Replacement;
968  if (Replacement->getType() != ToReplace->getType()) {
969  ActualReplacement =
970  Builder.CreateBitCast(Replacement, ToReplace->getType());
971  }
972  ToReplace->replaceAllUsesWith(ActualReplacement);
973  ToReplace->eraseFromParent();
974 
975  MadeChange = true;
976  }
977  return MadeChange;
978 }
979 
980 // Turns this:
981 //
982 // %base = ...
983 // %ptr = gep %base + 15
984 // %tok = statepoint (%fun, i32 0, i32 0, i32 0, %base, %ptr)
985 // %base' = relocate(%tok, i32 4, i32 4)
986 // %ptr' = relocate(%tok, i32 4, i32 5)
987 // %val = load %ptr'
988 //
989 // into this:
990 //
991 // %base = ...
992 // %ptr = gep %base + 15
993 // %tok = statepoint (%fun, i32 0, i32 0, i32 0, %base, %ptr)
994 // %base' = gc.relocate(%tok, i32 4, i32 4)
995 // %ptr' = gep %base' + 15
996 // %val = load %ptr'
997 bool CodeGenPrepare::simplifyOffsetableRelocate(Instruction &I) {
998  bool MadeChange = false;
999  SmallVector<GCRelocateInst *, 2> AllRelocateCalls;
1000 
1001  for (auto *U : I.users())
1002  if (GCRelocateInst *Relocate = dyn_cast<GCRelocateInst>(U))
1003  // Collect all the relocate calls associated with a statepoint
1004  AllRelocateCalls.push_back(Relocate);
1005 
1006  // We need atleast one base pointer relocation + one derived pointer
1007  // relocation to mangle
1008  if (AllRelocateCalls.size() < 2)
1009  return false;
1010 
1011  // RelocateInstMap is a mapping from the base relocate instruction to the
1012  // corresponding derived relocate instructions
1014  computeBaseDerivedRelocateMap(AllRelocateCalls, RelocateInstMap);
1015  if (RelocateInstMap.empty())
1016  return false;
1017 
1018  for (auto &Item : RelocateInstMap)
1019  // Item.first is the RelocatedBase to offset against
1020  // Item.second is the vector of Targets to replace
1021  MadeChange = simplifyRelocatesOffABase(Item.first, Item.second);
1022  return MadeChange;
1023 }
1024 
1025 /// SinkCast - Sink the specified cast instruction into its user blocks
1026 static bool SinkCast(CastInst *CI) {
1027  BasicBlock *DefBB = CI->getParent();
1028 
1029  /// InsertedCasts - Only insert a cast in each block once.
1030  DenseMap<BasicBlock*, CastInst*> InsertedCasts;
1031 
1032  bool MadeChange = false;
1033  for (Value::user_iterator UI = CI->user_begin(), E = CI->user_end();
1034  UI != E; ) {
1035  Use &TheUse = UI.getUse();
1036  Instruction *User = cast<Instruction>(*UI);
1037 
1038  // Figure out which BB this cast is used in. For PHI's this is the
1039  // appropriate predecessor block.
1040  BasicBlock *UserBB = User->getParent();
1041  if (PHINode *PN = dyn_cast<PHINode>(User)) {
1042  UserBB = PN->getIncomingBlock(TheUse);
1043  }
1044 
1045  // Preincrement use iterator so we don't invalidate it.
1046  ++UI;
1047 
1048  // The first insertion point of a block containing an EH pad is after the
1049  // pad. If the pad is the user, we cannot sink the cast past the pad.
1050  if (User->isEHPad())
1051  continue;
1052 
1053  // If the block selected to receive the cast is an EH pad that does not
1054  // allow non-PHI instructions before the terminator, we can't sink the
1055  // cast.
1056  if (UserBB->getTerminator()->isEHPad())
1057  continue;
1058 
1059  // If this user is in the same block as the cast, don't change the cast.
1060  if (UserBB == DefBB) continue;
1061 
1062  // If we have already inserted a cast into this block, use it.
1063  CastInst *&InsertedCast = InsertedCasts[UserBB];
1064 
1065  if (!InsertedCast) {
1066  BasicBlock::iterator InsertPt = UserBB->getFirstInsertionPt();
1067  assert(InsertPt != UserBB->end());
1068  InsertedCast = CastInst::Create(CI->getOpcode(), CI->getOperand(0),
1069  CI->getType(), "", &*InsertPt);
1070  InsertedCast->setDebugLoc(CI->getDebugLoc());
1071  }
1072 
1073  // Replace a use of the cast with a use of the new cast.
1074  TheUse = InsertedCast;
1075  MadeChange = true;
1076  ++NumCastUses;
1077  }
1078 
1079  // If we removed all uses, nuke the cast.
1080  if (CI->use_empty()) {
1081  salvageDebugInfo(*CI);
1082  CI->eraseFromParent();
1083  MadeChange = true;
1084  }
1085 
1086  return MadeChange;
1087 }
1088 
1089 /// If the specified cast instruction is a noop copy (e.g. it's casting from
1090 /// one pointer type to another, i32->i8 on PPC), sink it into user blocks to
1091 /// reduce the number of virtual registers that must be created and coalesced.
1092 ///
1093 /// Return true if any changes are made.
1095  const DataLayout &DL) {
1096  // Sink only "cheap" (or nop) address-space casts. This is a weaker condition
1097  // than sinking only nop casts, but is helpful on some platforms.
1098  if (auto *ASC = dyn_cast<AddrSpaceCastInst>(CI)) {
1099  if (!TLI.isCheapAddrSpaceCast(ASC->getSrcAddressSpace(),
1100  ASC->getDestAddressSpace()))
1101  return false;
1102  }
1103 
1104  // If this is a noop copy,
1105  EVT SrcVT = TLI.getValueType(DL, CI->getOperand(0)->getType());
1106  EVT DstVT = TLI.getValueType(DL, CI->getType());
1107 
1108  // This is an fp<->int conversion?
1109  if (SrcVT.isInteger() != DstVT.isInteger())
1110  return false;
1111 
1112  // If this is an extension, it will be a zero or sign extension, which
1113  // isn't a noop.
1114  if (SrcVT.bitsLT(DstVT)) return false;
1115 
1116  // If these values will be promoted, find out what they will be promoted
1117  // to. This helps us consider truncates on PPC as noop copies when they
1118  // are.
1119  if (TLI.getTypeAction(CI->getContext(), SrcVT) ==
1121  SrcVT = TLI.getTypeToTransformTo(CI->getContext(), SrcVT);
1122  if (TLI.getTypeAction(CI->getContext(), DstVT) ==
1124  DstVT = TLI.getTypeToTransformTo(CI->getContext(), DstVT);
1125 
1126  // If, after promotion, these are the same types, this is a noop copy.
1127  if (SrcVT != DstVT)
1128  return false;
1129 
1130  return SinkCast(CI);
1131 }
1132 
1133 /// Try to combine CI into a call to the llvm.uadd.with.overflow intrinsic if
1134 /// possible.
1135 ///
1136 /// Return true if any changes were made.
1138  Value *A, *B;
1139  Instruction *AddI;
1140  if (!match(CI,
1142  return false;
1143 
1144  Type *Ty = AddI->getType();
1145  if (!isa<IntegerType>(Ty))
1146  return false;
1147 
1148  // We don't want to move around uses of condition values this late, so we we
1149  // check if it is legal to create the call to the intrinsic in the basic
1150  // block containing the icmp:
1151 
1152  if (AddI->getParent() != CI->getParent() && !AddI->hasOneUse())
1153  return false;
1154 
1155 #ifndef NDEBUG
1156  // Someday m_UAddWithOverflow may get smarter, but this is a safe assumption
1157  // for now:
1158  if (AddI->hasOneUse())
1159  assert(*AddI->user_begin() == CI && "expected!");
1160 #endif
1161 
1162  Module *M = CI->getModule();
1163  Value *F = Intrinsic::getDeclaration(M, Intrinsic::uadd_with_overflow, Ty);
1164 
1165  auto *InsertPt = AddI->hasOneUse() ? CI : AddI;
1166 
1167  DebugLoc Loc = CI->getDebugLoc();
1168  auto *UAddWithOverflow =
1169  CallInst::Create(F, {A, B}, "uadd.overflow", InsertPt);
1170  UAddWithOverflow->setDebugLoc(Loc);
1171  auto *UAdd = ExtractValueInst::Create(UAddWithOverflow, 0, "uadd", InsertPt);
1172  UAdd->setDebugLoc(Loc);
1173  auto *Overflow =
1174  ExtractValueInst::Create(UAddWithOverflow, 1, "overflow", InsertPt);
1175  Overflow->setDebugLoc(Loc);
1176 
1177  CI->replaceAllUsesWith(Overflow);
1178  AddI->replaceAllUsesWith(UAdd);
1179  CI->eraseFromParent();
1180  AddI->eraseFromParent();
1181  return true;
1182 }
1183 
1184 /// Sink the given CmpInst into user blocks to reduce the number of virtual
1185 /// registers that must be created and coalesced. This is a clear win except on
1186 /// targets with multiple condition code registers (PowerPC), where it might
1187 /// lose; some adjustment may be wanted there.
1188 ///
1189 /// Return true if any changes are made.
1190 static bool SinkCmpExpression(CmpInst *CI, const TargetLowering *TLI) {
1191  BasicBlock *DefBB = CI->getParent();
1192 
1193  // Avoid sinking soft-FP comparisons, since this can move them into a loop.
1194  if (TLI && TLI->useSoftFloat() && isa<FCmpInst>(CI))
1195  return false;
1196 
1197  // Only insert a cmp in each block once.
1198  DenseMap<BasicBlock*, CmpInst*> InsertedCmps;
1199 
1200  bool MadeChange = false;
1201  for (Value::user_iterator UI = CI->user_begin(), E = CI->user_end();
1202  UI != E; ) {
1203  Use &TheUse = UI.getUse();
1204  Instruction *User = cast<Instruction>(*UI);
1205 
1206  // Preincrement use iterator so we don't invalidate it.
1207  ++UI;
1208 
1209  // Don't bother for PHI nodes.
1210  if (isa<PHINode>(User))
1211  continue;
1212 
1213  // Figure out which BB this cmp is used in.
1214  BasicBlock *UserBB = User->getParent();
1215 
1216  // If this user is in the same block as the cmp, don't change the cmp.
1217  if (UserBB == DefBB) continue;
1218 
1219  // If we have already inserted a cmp into this block, use it.
1220  CmpInst *&InsertedCmp = InsertedCmps[UserBB];
1221 
1222  if (!InsertedCmp) {
1223  BasicBlock::iterator InsertPt = UserBB->getFirstInsertionPt();
1224  assert(InsertPt != UserBB->end());
1225  InsertedCmp =
1226  CmpInst::Create(CI->getOpcode(), CI->getPredicate(),
1227  CI->getOperand(0), CI->getOperand(1), "", &*InsertPt);
1228  // Propagate the debug info.
1229  InsertedCmp->setDebugLoc(CI->getDebugLoc());
1230  }
1231 
1232  // Replace a use of the cmp with a use of the new cmp.
1233  TheUse = InsertedCmp;
1234  MadeChange = true;
1235  ++NumCmpUses;
1236  }
1237 
1238  // If we removed all uses, nuke the cmp.
1239  if (CI->use_empty()) {
1240  CI->eraseFromParent();
1241  MadeChange = true;
1242  }
1243 
1244  return MadeChange;
1245 }
1246 
1247 static bool OptimizeCmpExpression(CmpInst *CI, const TargetLowering *TLI) {
1248  if (SinkCmpExpression(CI, TLI))
1249  return true;
1250 
1251  if (CombineUAddWithOverflow(CI))
1252  return true;
1253 
1254  return false;
1255 }
1256 
1257 /// Duplicate and sink the given 'and' instruction into user blocks where it is
1258 /// used in a compare to allow isel to generate better code for targets where
1259 /// this operation can be combined.
1260 ///
1261 /// Return true if any changes are made.
1263  const TargetLowering &TLI,
1264  SetOfInstrs &InsertedInsts) {
1265  // Double-check that we're not trying to optimize an instruction that was
1266  // already optimized by some other part of this pass.
1267  assert(!InsertedInsts.count(AndI) &&
1268  "Attempting to optimize already optimized and instruction");
1269  (void) InsertedInsts;
1270 
1271  // Nothing to do for single use in same basic block.
1272  if (AndI->hasOneUse() &&
1273  AndI->getParent() == cast<Instruction>(*AndI->user_begin())->getParent())
1274  return false;
1275 
1276  // Try to avoid cases where sinking/duplicating is likely to increase register
1277  // pressure.
1278  if (!isa<ConstantInt>(AndI->getOperand(0)) &&
1279  !isa<ConstantInt>(AndI->getOperand(1)) &&
1280  AndI->getOperand(0)->hasOneUse() && AndI->getOperand(1)->hasOneUse())
1281  return false;
1282 
1283  for (auto *U : AndI->users()) {
1284  Instruction *User = cast<Instruction>(U);
1285 
1286  // Only sink for and mask feeding icmp with 0.
1287  if (!isa<ICmpInst>(User))
1288  return false;
1289 
1290  auto *CmpC = dyn_cast<ConstantInt>(User->getOperand(1));
1291  if (!CmpC || !CmpC->isZero())
1292  return false;
1293  }
1294 
1295  if (!TLI.isMaskAndCmp0FoldingBeneficial(*AndI))
1296  return false;
1297 
1298  LLVM_DEBUG(dbgs() << "found 'and' feeding only icmp 0;\n");
1299  LLVM_DEBUG(AndI->getParent()->dump());
1300 
1301  // Push the 'and' into the same block as the icmp 0. There should only be
1302  // one (icmp (and, 0)) in each block, since CSE/GVN should have removed any
1303  // others, so we don't need to keep track of which BBs we insert into.
1304  for (Value::user_iterator UI = AndI->user_begin(), E = AndI->user_end();
1305  UI != E; ) {
1306  Use &TheUse = UI.getUse();
1307  Instruction *User = cast<Instruction>(*UI);
1308 
1309  // Preincrement use iterator so we don't invalidate it.
1310  ++UI;
1311 
1312  LLVM_DEBUG(dbgs() << "sinking 'and' use: " << *User << "\n");
1313 
1314  // Keep the 'and' in the same place if the use is already in the same block.
1315  Instruction *InsertPt =
1316  User->getParent() == AndI->getParent() ? AndI : User;
1317  Instruction *InsertedAnd =
1318  BinaryOperator::Create(Instruction::And, AndI->getOperand(0),
1319  AndI->getOperand(1), "", InsertPt);
1320  // Propagate the debug info.
1321  InsertedAnd->setDebugLoc(AndI->getDebugLoc());
1322 
1323  // Replace a use of the 'and' with a use of the new 'and'.
1324  TheUse = InsertedAnd;
1325  ++NumAndUses;
1326  LLVM_DEBUG(User->getParent()->dump());
1327  }
1328 
1329  // We removed all uses, nuke the and.
1330  AndI->eraseFromParent();
1331  return true;
1332 }
1333 
1334 /// Check if the candidates could be combined with a shift instruction, which
1335 /// includes:
1336 /// 1. Truncate instruction
1337 /// 2. And instruction and the imm is a mask of the low bits:
1338 /// imm & (imm+1) == 0
1340  if (!isa<TruncInst>(User)) {
1341  if (User->getOpcode() != Instruction::And ||
1342  !isa<ConstantInt>(User->getOperand(1)))
1343  return false;
1344 
1345  const APInt &Cimm = cast<ConstantInt>(User->getOperand(1))->getValue();
1346 
1347  if ((Cimm & (Cimm + 1)).getBoolValue())
1348  return false;
1349  }
1350  return true;
1351 }
1352 
1353 /// Sink both shift and truncate instruction to the use of truncate's BB.
1354 static bool
1357  const TargetLowering &TLI, const DataLayout &DL) {
1358  BasicBlock *UserBB = User->getParent();
1359  DenseMap<BasicBlock *, CastInst *> InsertedTruncs;
1360  TruncInst *TruncI = dyn_cast<TruncInst>(User);
1361  bool MadeChange = false;
1362 
1363  for (Value::user_iterator TruncUI = TruncI->user_begin(),
1364  TruncE = TruncI->user_end();
1365  TruncUI != TruncE;) {
1366 
1367  Use &TruncTheUse = TruncUI.getUse();
1368  Instruction *TruncUser = cast<Instruction>(*TruncUI);
1369  // Preincrement use iterator so we don't invalidate it.
1370 
1371  ++TruncUI;
1372 
1373  int ISDOpcode = TLI.InstructionOpcodeToISD(TruncUser->getOpcode());
1374  if (!ISDOpcode)
1375  continue;
1376 
1377  // If the use is actually a legal node, there will not be an
1378  // implicit truncate.
1379  // FIXME: always querying the result type is just an
1380  // approximation; some nodes' legality is determined by the
1381  // operand or other means. There's no good way to find out though.
1382  if (TLI.isOperationLegalOrCustom(
1383  ISDOpcode, TLI.getValueType(DL, TruncUser->getType(), true)))
1384  continue;
1385 
1386  // Don't bother for PHI nodes.
1387  if (isa<PHINode>(TruncUser))
1388  continue;
1389 
1390  BasicBlock *TruncUserBB = TruncUser->getParent();
1391 
1392  if (UserBB == TruncUserBB)
1393  continue;
1394 
1395  BinaryOperator *&InsertedShift = InsertedShifts[TruncUserBB];
1396  CastInst *&InsertedTrunc = InsertedTruncs[TruncUserBB];
1397 
1398  if (!InsertedShift && !InsertedTrunc) {
1399  BasicBlock::iterator InsertPt = TruncUserBB->getFirstInsertionPt();
1400  assert(InsertPt != TruncUserBB->end());
1401  // Sink the shift
1402  if (ShiftI->getOpcode() == Instruction::AShr)
1403  InsertedShift = BinaryOperator::CreateAShr(ShiftI->getOperand(0), CI,
1404  "", &*InsertPt);
1405  else
1406  InsertedShift = BinaryOperator::CreateLShr(ShiftI->getOperand(0), CI,
1407  "", &*InsertPt);
1408  InsertedShift->setDebugLoc(ShiftI->getDebugLoc());
1409 
1410  // Sink the trunc
1411  BasicBlock::iterator TruncInsertPt = TruncUserBB->getFirstInsertionPt();
1412  TruncInsertPt++;
1413  assert(TruncInsertPt != TruncUserBB->end());
1414 
1415  InsertedTrunc = CastInst::Create(TruncI->getOpcode(), InsertedShift,
1416  TruncI->getType(), "", &*TruncInsertPt);
1417  InsertedTrunc->setDebugLoc(TruncI->getDebugLoc());
1418 
1419  MadeChange = true;
1420 
1421  TruncTheUse = InsertedTrunc;
1422  }
1423  }
1424  return MadeChange;
1425 }
1426 
1427 /// Sink the shift *right* instruction into user blocks if the uses could
1428 /// potentially be combined with this shift instruction and generate BitExtract
1429 /// instruction. It will only be applied if the architecture supports BitExtract
1430 /// instruction. Here is an example:
1431 /// BB1:
1432 /// %x.extract.shift = lshr i64 %arg1, 32
1433 /// BB2:
1434 /// %x.extract.trunc = trunc i64 %x.extract.shift to i16
1435 /// ==>
1436 ///
1437 /// BB2:
1438 /// %x.extract.shift.1 = lshr i64 %arg1, 32
1439 /// %x.extract.trunc = trunc i64 %x.extract.shift.1 to i16
1440 ///
1441 /// CodeGen will recognize the pattern in BB2 and generate BitExtract
1442 /// instruction.
1443 /// Return true if any changes are made.
1445  const TargetLowering &TLI,
1446  const DataLayout &DL) {
1447  BasicBlock *DefBB = ShiftI->getParent();
1448 
1449  /// Only insert instructions in each block once.
1451 
1452  bool shiftIsLegal = TLI.isTypeLegal(TLI.getValueType(DL, ShiftI->getType()));
1453 
1454  bool MadeChange = false;
1455  for (Value::user_iterator UI = ShiftI->user_begin(), E = ShiftI->user_end();
1456  UI != E;) {
1457  Use &TheUse = UI.getUse();
1458  Instruction *User = cast<Instruction>(*UI);
1459  // Preincrement use iterator so we don't invalidate it.
1460  ++UI;
1461 
1462  // Don't bother for PHI nodes.
1463  if (isa<PHINode>(User))
1464  continue;
1465 
1466  if (!isExtractBitsCandidateUse(User))
1467  continue;
1468 
1469  BasicBlock *UserBB = User->getParent();
1470 
1471  if (UserBB == DefBB) {
1472  // If the shift and truncate instruction are in the same BB. The use of
1473  // the truncate(TruncUse) may still introduce another truncate if not
1474  // legal. In this case, we would like to sink both shift and truncate
1475  // instruction to the BB of TruncUse.
1476  // for example:
1477  // BB1:
1478  // i64 shift.result = lshr i64 opnd, imm
1479  // trunc.result = trunc shift.result to i16
1480  //
1481  // BB2:
1482  // ----> We will have an implicit truncate here if the architecture does
1483  // not have i16 compare.
1484  // cmp i16 trunc.result, opnd2
1485  //
1486  if (isa<TruncInst>(User) && shiftIsLegal
1487  // If the type of the truncate is legal, no truncate will be
1488  // introduced in other basic blocks.
1489  &&
1490  (!TLI.isTypeLegal(TLI.getValueType(DL, User->getType()))))
1491  MadeChange =
1492  SinkShiftAndTruncate(ShiftI, User, CI, InsertedShifts, TLI, DL);
1493 
1494  continue;
1495  }
1496  // If we have already inserted a shift into this block, use it.
1497  BinaryOperator *&InsertedShift = InsertedShifts[UserBB];
1498 
1499  if (!InsertedShift) {
1500  BasicBlock::iterator InsertPt = UserBB->getFirstInsertionPt();
1501  assert(InsertPt != UserBB->end());
1502 
1503  if (ShiftI->getOpcode() == Instruction::AShr)
1504  InsertedShift = BinaryOperator::CreateAShr(ShiftI->getOperand(0), CI,
1505  "", &*InsertPt);
1506  else
1507  InsertedShift = BinaryOperator::CreateLShr(ShiftI->getOperand(0), CI,
1508  "", &*InsertPt);
1509  InsertedShift->setDebugLoc(ShiftI->getDebugLoc());
1510 
1511  MadeChange = true;
1512  }
1513 
1514  // Replace a use of the shift with a use of the new shift.
1515  TheUse = InsertedShift;
1516  }
1517 
1518  // If we removed all uses, nuke the shift.
1519  if (ShiftI->use_empty()) {
1520  salvageDebugInfo(*ShiftI);
1521  ShiftI->eraseFromParent();
1522  }
1523 
1524  return MadeChange;
1525 }
1526 
1527 /// If counting leading or trailing zeros is an expensive operation and a zero
1528 /// input is defined, add a check for zero to avoid calling the intrinsic.
1529 ///
1530 /// We want to transform:
1531 /// %z = call i64 @llvm.cttz.i64(i64 %A, i1 false)
1532 ///
1533 /// into:
1534 /// entry:
1535 /// %cmpz = icmp eq i64 %A, 0
1536 /// br i1 %cmpz, label %cond.end, label %cond.false
1537 /// cond.false:
1538 /// %z = call i64 @llvm.cttz.i64(i64 %A, i1 true)
1539 /// br label %cond.end
1540 /// cond.end:
1541 /// %ctz = phi i64 [ 64, %entry ], [ %z, %cond.false ]
1542 ///
1543 /// If the transform is performed, return true and set ModifiedDT to true.
1544 static bool despeculateCountZeros(IntrinsicInst *CountZeros,
1545  const TargetLowering *TLI,
1546  const DataLayout *DL,
1547  bool &ModifiedDT) {
1548  if (!TLI || !DL)
1549  return false;
1550 
1551  // If a zero input is undefined, it doesn't make sense to despeculate that.
1552  if (match(CountZeros->getOperand(1), m_One()))
1553  return false;
1554 
1555  // If it's cheap to speculate, there's nothing to do.
1556  auto IntrinsicID = CountZeros->getIntrinsicID();
1557  if ((IntrinsicID == Intrinsic::cttz && TLI->isCheapToSpeculateCttz()) ||
1558  (IntrinsicID == Intrinsic::ctlz && TLI->isCheapToSpeculateCtlz()))
1559  return false;
1560 
1561  // Only handle legal scalar cases. Anything else requires too much work.
1562  Type *Ty = CountZeros->getType();
1563  unsigned SizeInBits = Ty->getPrimitiveSizeInBits();
1564  if (Ty->isVectorTy() || SizeInBits > DL->getLargestLegalIntTypeSizeInBits())
1565  return false;
1566 
1567  // The intrinsic will be sunk behind a compare against zero and branch.
1568  BasicBlock *StartBlock = CountZeros->getParent();
1569  BasicBlock *CallBlock = StartBlock->splitBasicBlock(CountZeros, "cond.false");
1570 
1571  // Create another block after the count zero intrinsic. A PHI will be added
1572  // in this block to select the result of the intrinsic or the bit-width
1573  // constant if the input to the intrinsic is zero.
1574  BasicBlock::iterator SplitPt = ++(BasicBlock::iterator(CountZeros));
1575  BasicBlock *EndBlock = CallBlock->splitBasicBlock(SplitPt, "cond.end");
1576 
1577  // Set up a builder to create a compare, conditional branch, and PHI.
1578  IRBuilder<> Builder(CountZeros->getContext());
1579  Builder.SetInsertPoint(StartBlock->getTerminator());
1580  Builder.SetCurrentDebugLocation(CountZeros->getDebugLoc());
1581 
1582  // Replace the unconditional branch that was created by the first split with
1583  // a compare against zero and a conditional branch.
1584  Value *Zero = Constant::getNullValue(Ty);
1585  Value *Cmp = Builder.CreateICmpEQ(CountZeros->getOperand(0), Zero, "cmpz");
1586  Builder.CreateCondBr(Cmp, EndBlock, CallBlock);
1587  StartBlock->getTerminator()->eraseFromParent();
1588 
1589  // Create a PHI in the end block to select either the output of the intrinsic
1590  // or the bit width of the operand.
1591  Builder.SetInsertPoint(&EndBlock->front());
1592  PHINode *PN = Builder.CreatePHI(Ty, 2, "ctz");
1593  CountZeros->replaceAllUsesWith(PN);
1594  Value *BitWidth = Builder.getInt(APInt(SizeInBits, SizeInBits));
1595  PN->addIncoming(BitWidth, StartBlock);
1596  PN->addIncoming(CountZeros, CallBlock);
1597 
1598  // We are explicitly handling the zero case, so we can set the intrinsic's
1599  // undefined zero argument to 'true'. This will also prevent reprocessing the
1600  // intrinsic; we only despeculate when a zero input is defined.
1601  CountZeros->setArgOperand(1, Builder.getTrue());
1602  ModifiedDT = true;
1603  return true;
1604 }
1605 
1606 bool CodeGenPrepare::optimizeCallInst(CallInst *CI, bool &ModifiedDT) {
1607  BasicBlock *BB = CI->getParent();
1608 
1609  // Lower inline assembly if we can.
1610  // If we found an inline asm expession, and if the target knows how to
1611  // lower it to normal LLVM code, do so now.
1612  if (TLI && isa<InlineAsm>(CI->getCalledValue())) {
1613  if (TLI->ExpandInlineAsm(CI)) {
1614  // Avoid invalidating the iterator.
1615  CurInstIterator = BB->begin();
1616  // Avoid processing instructions out of order, which could cause
1617  // reuse before a value is defined.
1618  SunkAddrs.clear();
1619  return true;
1620  }
1621  // Sink address computing for memory operands into the block.
1622  if (optimizeInlineAsmInst(CI))
1623  return true;
1624  }
1625 
1626  // Align the pointer arguments to this call if the target thinks it's a good
1627  // idea
1628  unsigned MinSize, PrefAlign;
1629  if (TLI && TLI->shouldAlignPointerArgs(CI, MinSize, PrefAlign)) {
1630  for (auto &Arg : CI->arg_operands()) {
1631  // We want to align both objects whose address is used directly and
1632  // objects whose address is used in casts and GEPs, though it only makes
1633  // sense for GEPs if the offset is a multiple of the desired alignment and
1634  // if size - offset meets the size threshold.
1635  if (!Arg->getType()->isPointerTy())
1636  continue;
1637  APInt Offset(DL->getIndexSizeInBits(
1638  cast<PointerType>(Arg->getType())->getAddressSpace()),
1639  0);
1641  uint64_t Offset2 = Offset.getLimitedValue();
1642  if ((Offset2 & (PrefAlign-1)) != 0)
1643  continue;
1644  AllocaInst *AI;
1645  if ((AI = dyn_cast<AllocaInst>(Val)) && AI->getAlignment() < PrefAlign &&
1646  DL->getTypeAllocSize(AI->getAllocatedType()) >= MinSize + Offset2)
1647  AI->setAlignment(PrefAlign);
1648  // Global variables can only be aligned if they are defined in this
1649  // object (i.e. they are uniquely initialized in this object), and
1650  // over-aligning global variables that have an explicit section is
1651  // forbidden.
1652  GlobalVariable *GV;
1653  if ((GV = dyn_cast<GlobalVariable>(Val)) && GV->canIncreaseAlignment() &&
1654  GV->getPointerAlignment(*DL) < PrefAlign &&
1655  DL->getTypeAllocSize(GV->getValueType()) >=
1656  MinSize + Offset2)
1657  GV->setAlignment(PrefAlign);
1658  }
1659  // If this is a memcpy (or similar) then we may be able to improve the
1660  // alignment
1661  if (MemIntrinsic *MI = dyn_cast<MemIntrinsic>(CI)) {
1662  unsigned DestAlign = getKnownAlignment(MI->getDest(), *DL);
1663  if (DestAlign > MI->getDestAlignment())
1664  MI->setDestAlignment(DestAlign);
1665  if (MemTransferInst *MTI = dyn_cast<MemTransferInst>(MI)) {
1666  unsigned SrcAlign = getKnownAlignment(MTI->getSource(), *DL);
1667  if (SrcAlign > MTI->getSourceAlignment())
1668  MTI->setSourceAlignment(SrcAlign);
1669  }
1670  }
1671  }
1672 
1673  // If we have a cold call site, try to sink addressing computation into the
1674  // cold block. This interacts with our handling for loads and stores to
1675  // ensure that we can fold all uses of a potential addressing computation
1676  // into their uses. TODO: generalize this to work over profiling data
1677  if (!OptSize && CI->hasFnAttr(Attribute::Cold))
1678  for (auto &Arg : CI->arg_operands()) {
1679  if (!Arg->getType()->isPointerTy())
1680  continue;
1681  unsigned AS = Arg->getType()->getPointerAddressSpace();
1682  return optimizeMemoryInst(CI, Arg, Arg->getType(), AS);
1683  }
1684 
1686  if (II) {
1687  switch (II->getIntrinsicID()) {
1688  default: break;
1689  case Intrinsic::objectsize: {
1690  // Lower all uses of llvm.objectsize.*
1691  ConstantInt *RetVal =
1692  lowerObjectSizeCall(II, *DL, TLInfo, /*MustSucceed=*/true);
1693  // Substituting this can cause recursive simplifications, which can
1694  // invalidate our iterator. Use a WeakTrackingVH to hold onto it in case
1695  // this
1696  // happens.
1697  Value *CurValue = &*CurInstIterator;
1698  WeakTrackingVH IterHandle(CurValue);
1699 
1700  replaceAndRecursivelySimplify(CI, RetVal, TLInfo, nullptr);
1701 
1702  // If the iterator instruction was recursively deleted, start over at the
1703  // start of the block.
1704  if (IterHandle != CurValue) {
1705  CurInstIterator = BB->begin();
1706  SunkAddrs.clear();
1707  }
1708  return true;
1709  }
1710  case Intrinsic::aarch64_stlxr:
1711  case Intrinsic::aarch64_stxr: {
1712  ZExtInst *ExtVal = dyn_cast<ZExtInst>(CI->getArgOperand(0));
1713  if (!ExtVal || !ExtVal->hasOneUse() ||
1714  ExtVal->getParent() == CI->getParent())
1715  return false;
1716  // Sink a zext feeding stlxr/stxr before it, so it can be folded into it.
1717  ExtVal->moveBefore(CI);
1718  // Mark this instruction as "inserted by CGP", so that other
1719  // optimizations don't touch it.
1720  InsertedInsts.insert(ExtVal);
1721  return true;
1722  }
1723  case Intrinsic::launder_invariant_group:
1724  case Intrinsic::strip_invariant_group:
1725  II->replaceAllUsesWith(II->getArgOperand(0));
1726  II->eraseFromParent();
1727  return true;
1728 
1729  case Intrinsic::cttz:
1730  case Intrinsic::ctlz:
1731  // If counting zeros is expensive, try to avoid it.
1732  return despeculateCountZeros(II, TLI, DL, ModifiedDT);
1733  }
1734 
1735  if (TLI) {
1736  SmallVector<Value*, 2> PtrOps;
1737  Type *AccessTy;
1738  if (TLI->getAddrModeArguments(II, PtrOps, AccessTy))
1739  while (!PtrOps.empty()) {
1740  Value *PtrVal = PtrOps.pop_back_val();
1741  unsigned AS = PtrVal->getType()->getPointerAddressSpace();
1742  if (optimizeMemoryInst(II, PtrVal, AccessTy, AS))
1743  return true;
1744  }
1745  }
1746  }
1747 
1748  // From here on out we're working with named functions.
1749  if (!CI->getCalledFunction()) return false;
1750 
1751  // Lower all default uses of _chk calls. This is very similar
1752  // to what InstCombineCalls does, but here we are only lowering calls
1753  // to fortified library functions (e.g. __memcpy_chk) that have the default
1754  // "don't know" as the objectsize. Anything else should be left alone.
1755  FortifiedLibCallSimplifier Simplifier(TLInfo, true);
1756  if (Value *V = Simplifier.optimizeCall(CI)) {
1757  CI->replaceAllUsesWith(V);
1758  CI->eraseFromParent();
1759  return true;
1760  }
1761 
1762  return false;
1763 }
1764 
1765 /// Look for opportunities to duplicate return instructions to the predecessor
1766 /// to enable tail call optimizations. The case it is currently looking for is:
1767 /// @code
1768 /// bb0:
1769 /// %tmp0 = tail call i32 @f0()
1770 /// br label %return
1771 /// bb1:
1772 /// %tmp1 = tail call i32 @f1()
1773 /// br label %return
1774 /// bb2:
1775 /// %tmp2 = tail call i32 @f2()
1776 /// br label %return
1777 /// return:
1778 /// %retval = phi i32 [ %tmp0, %bb0 ], [ %tmp1, %bb1 ], [ %tmp2, %bb2 ]
1779 /// ret i32 %retval
1780 /// @endcode
1781 ///
1782 /// =>
1783 ///
1784 /// @code
1785 /// bb0:
1786 /// %tmp0 = tail call i32 @f0()
1787 /// ret i32 %tmp0
1788 /// bb1:
1789 /// %tmp1 = tail call i32 @f1()
1790 /// ret i32 %tmp1
1791 /// bb2:
1792 /// %tmp2 = tail call i32 @f2()
1793 /// ret i32 %tmp2
1794 /// @endcode
1795 bool CodeGenPrepare::dupRetToEnableTailCallOpts(BasicBlock *BB) {
1796  if (!TLI)
1797  return false;
1798 
1799  ReturnInst *RetI = dyn_cast<ReturnInst>(BB->getTerminator());
1800  if (!RetI)
1801  return false;
1802 
1803  PHINode *PN = nullptr;
1804  BitCastInst *BCI = nullptr;
1805  Value *V = RetI->getReturnValue();
1806  if (V) {
1807  BCI = dyn_cast<BitCastInst>(V);
1808  if (BCI)
1809  V = BCI->getOperand(0);
1810 
1811  PN = dyn_cast<PHINode>(V);
1812  if (!PN)
1813  return false;
1814  }
1815 
1816  if (PN && PN->getParent() != BB)
1817  return false;
1818 
1819  // Make sure there are no instructions between the PHI and return, or that the
1820  // return is the first instruction in the block.
1821  if (PN) {
1822  BasicBlock::iterator BI = BB->begin();
1823  do { ++BI; } while (isa<DbgInfoIntrinsic>(BI));
1824  if (&*BI == BCI)
1825  // Also skip over the bitcast.
1826  ++BI;
1827  if (&*BI != RetI)
1828  return false;
1829  } else {
1830  BasicBlock::iterator BI = BB->begin();
1831  while (isa<DbgInfoIntrinsic>(BI)) ++BI;
1832  if (&*BI != RetI)
1833  return false;
1834  }
1835 
1836  /// Only dup the ReturnInst if the CallInst is likely to be emitted as a tail
1837  /// call.
1838  const Function *F = BB->getParent();
1839  SmallVector<CallInst*, 4> TailCalls;
1840  if (PN) {
1841  for (unsigned I = 0, E = PN->getNumIncomingValues(); I != E; ++I) {
1842  CallInst *CI = dyn_cast<CallInst>(PN->getIncomingValue(I));
1843  // Make sure the phi value is indeed produced by the tail call.
1844  if (CI && CI->hasOneUse() && CI->getParent() == PN->getIncomingBlock(I) &&
1845  TLI->mayBeEmittedAsTailCall(CI) &&
1846  attributesPermitTailCall(F, CI, RetI, *TLI))
1847  TailCalls.push_back(CI);
1848  }
1849  } else {
1850  SmallPtrSet<BasicBlock*, 4> VisitedBBs;
1851  for (pred_iterator PI = pred_begin(BB), PE = pred_end(BB); PI != PE; ++PI) {
1852  if (!VisitedBBs.insert(*PI).second)
1853  continue;
1854 
1855  BasicBlock::InstListType &InstList = (*PI)->getInstList();
1856  BasicBlock::InstListType::reverse_iterator RI = InstList.rbegin();
1857  BasicBlock::InstListType::reverse_iterator RE = InstList.rend();
1858  do { ++RI; } while (RI != RE && isa<DbgInfoIntrinsic>(&*RI));
1859  if (RI == RE)
1860  continue;
1861 
1862  CallInst *CI = dyn_cast<CallInst>(&*RI);
1863  if (CI && CI->use_empty() && TLI->mayBeEmittedAsTailCall(CI) &&
1864  attributesPermitTailCall(F, CI, RetI, *TLI))
1865  TailCalls.push_back(CI);
1866  }
1867  }
1868 
1869  bool Changed = false;
1870  for (unsigned i = 0, e = TailCalls.size(); i != e; ++i) {
1871  CallInst *CI = TailCalls[i];
1872  CallSite CS(CI);
1873 
1874  // Make sure the call instruction is followed by an unconditional branch to
1875  // the return block.
1876  BasicBlock *CallBB = CI->getParent();
1877  BranchInst *BI = dyn_cast<BranchInst>(CallBB->getTerminator());
1878  if (!BI || !BI->isUnconditional() || BI->getSuccessor(0) != BB)
1879  continue;
1880 
1881  // Duplicate the return into CallBB.
1882  (void)FoldReturnIntoUncondBranch(RetI, BB, CallBB);
1883  ModifiedDT = Changed = true;
1884  ++NumRetsDup;
1885  }
1886 
1887  // If we eliminated all predecessors of the block, delete the block now.
1888  if (Changed && !BB->hasAddressTaken() && pred_begin(BB) == pred_end(BB))
1889  BB->eraseFromParent();
1890 
1891  return Changed;
1892 }
1893 
1894 //===----------------------------------------------------------------------===//
1895 // Memory Optimization
1896 //===----------------------------------------------------------------------===//
1897 
1898 namespace {
1899 
1900 /// This is an extended version of TargetLowering::AddrMode
1901 /// which holds actual Value*'s for register values.
1902 struct ExtAddrMode : public TargetLowering::AddrMode {
1903  Value *BaseReg = nullptr;
1904  Value *ScaledReg = nullptr;
1905  Value *OriginalValue = nullptr;
1906 
1907  enum FieldName {
1908  NoField = 0x00,
1909  BaseRegField = 0x01,
1910  BaseGVField = 0x02,
1911  BaseOffsField = 0x04,
1912  ScaledRegField = 0x08,
1913  ScaleField = 0x10,
1914  MultipleFields = 0xff
1915  };
1916 
1917  ExtAddrMode() = default;
1918 
1919  void print(raw_ostream &OS) const;
1920  void dump() const;
1921 
1922  FieldName compare(const ExtAddrMode &other) {
1923  // First check that the types are the same on each field, as differing types
1924  // is something we can't cope with later on.
1925  if (BaseReg && other.BaseReg &&
1926  BaseReg->getType() != other.BaseReg->getType())
1927  return MultipleFields;
1928  if (BaseGV && other.BaseGV &&
1929  BaseGV->getType() != other.BaseGV->getType())
1930  return MultipleFields;
1931  if (ScaledReg && other.ScaledReg &&
1932  ScaledReg->getType() != other.ScaledReg->getType())
1933  return MultipleFields;
1934 
1935  // Check each field to see if it differs.
1936  unsigned Result = NoField;
1937  if (BaseReg != other.BaseReg)
1938  Result |= BaseRegField;
1939  if (BaseGV != other.BaseGV)
1940  Result |= BaseGVField;
1941  if (BaseOffs != other.BaseOffs)
1942  Result |= BaseOffsField;
1943  if (ScaledReg != other.ScaledReg)
1944  Result |= ScaledRegField;
1945  // Don't count 0 as being a different scale, because that actually means
1946  // unscaled (which will already be counted by having no ScaledReg).
1947  if (Scale && other.Scale && Scale != other.Scale)
1948  Result |= ScaleField;
1949 
1950  if (countPopulation(Result) > 1)
1951  return MultipleFields;
1952  else
1953  return static_cast<FieldName>(Result);
1954  }
1955 
1956  // An AddrMode is trivial if it involves no calculation i.e. it is just a base
1957  // with no offset.
1958  bool isTrivial() {
1959  // An AddrMode is (BaseGV + BaseReg + BaseOffs + ScaleReg * Scale) so it is
1960  // trivial if at most one of these terms is nonzero, except that BaseGV and
1961  // BaseReg both being zero actually means a null pointer value, which we
1962  // consider to be 'non-zero' here.
1963  return !BaseOffs && !Scale && !(BaseGV && BaseReg);
1964  }
1965 
1966  Value *GetFieldAsValue(FieldName Field, Type *IntPtrTy) {
1967  switch (Field) {
1968  default:
1969  return nullptr;
1970  case BaseRegField:
1971  return BaseReg;
1972  case BaseGVField:
1973  return BaseGV;
1974  case ScaledRegField:
1975  return ScaledReg;
1976  case BaseOffsField:
1977  return ConstantInt::get(IntPtrTy, BaseOffs);
1978  }
1979  }
1980 
1981  void SetCombinedField(FieldName Field, Value *V,
1982  const SmallVectorImpl<ExtAddrMode> &AddrModes) {
1983  switch (Field) {
1984  default:
1985  llvm_unreachable("Unhandled fields are expected to be rejected earlier");
1986  break;
1987  case ExtAddrMode::BaseRegField:
1988  BaseReg = V;
1989  break;
1990  case ExtAddrMode::BaseGVField:
1991  // A combined BaseGV is an Instruction, not a GlobalValue, so it goes
1992  // in the BaseReg field.
1993  assert(BaseReg == nullptr);
1994  BaseReg = V;
1995  BaseGV = nullptr;
1996  break;
1997  case ExtAddrMode::ScaledRegField:
1998  ScaledReg = V;
1999  // If we have a mix of scaled and unscaled addrmodes then we want scale
2000  // to be the scale and not zero.
2001  if (!Scale)
2002  for (const ExtAddrMode &AM : AddrModes)
2003  if (AM.Scale) {
2004  Scale = AM.Scale;
2005  break;
2006  }
2007  break;
2008  case ExtAddrMode::BaseOffsField:
2009  // The offset is no longer a constant, so it goes in ScaledReg with a
2010  // scale of 1.
2011  assert(ScaledReg == nullptr);
2012  ScaledReg = V;
2013  Scale = 1;
2014  BaseOffs = 0;
2015  break;
2016  }
2017  }
2018 };
2019 
2020 } // end anonymous namespace
2021 
2022 #ifndef NDEBUG
2023 static inline raw_ostream &operator<<(raw_ostream &OS, const ExtAddrMode &AM) {
2024  AM.print(OS);
2025  return OS;
2026 }
2027 #endif
2028 
2029 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
2030 void ExtAddrMode::print(raw_ostream &OS) const {
2031  bool NeedPlus = false;
2032  OS << "[";
2033  if (BaseGV) {
2034  OS << (NeedPlus ? " + " : "")
2035  << "GV:";
2036  BaseGV->printAsOperand(OS, /*PrintType=*/false);
2037  NeedPlus = true;
2038  }
2039 
2040  if (BaseOffs) {
2041  OS << (NeedPlus ? " + " : "")
2042  << BaseOffs;
2043  NeedPlus = true;
2044  }
2045 
2046  if (BaseReg) {
2047  OS << (NeedPlus ? " + " : "")
2048  << "Base:";
2049  BaseReg->printAsOperand(OS, /*PrintType=*/false);
2050  NeedPlus = true;
2051  }
2052  if (Scale) {
2053  OS << (NeedPlus ? " + " : "")
2054  << Scale << "*";
2055  ScaledReg->printAsOperand(OS, /*PrintType=*/false);
2056  }
2057 
2058  OS << ']';
2059 }
2060 
2061 LLVM_DUMP_METHOD void ExtAddrMode::dump() const {
2062  print(dbgs());
2063  dbgs() << '\n';
2064 }
2065 #endif
2066 
2067 namespace {
2068 
2069 /// This class provides transaction based operation on the IR.
2070 /// Every change made through this class is recorded in the internal state and
2071 /// can be undone (rollback) until commit is called.
2072 class TypePromotionTransaction {
2073  /// This represents the common interface of the individual transaction.
2074  /// Each class implements the logic for doing one specific modification on
2075  /// the IR via the TypePromotionTransaction.
2076  class TypePromotionAction {
2077  protected:
2078  /// The Instruction modified.
2079  Instruction *Inst;
2080 
2081  public:
2082  /// Constructor of the action.
2083  /// The constructor performs the related action on the IR.
2084  TypePromotionAction(Instruction *Inst) : Inst(Inst) {}
2085 
2086  virtual ~TypePromotionAction() = default;
2087 
2088  /// Undo the modification done by this action.
2089  /// When this method is called, the IR must be in the same state as it was
2090  /// before this action was applied.
2091  /// \pre Undoing the action works if and only if the IR is in the exact same
2092  /// state as it was directly after this action was applied.
2093  virtual void undo() = 0;
2094 
2095  /// Advocate every change made by this action.
2096  /// When the results on the IR of the action are to be kept, it is important
2097  /// to call this function, otherwise hidden information may be kept forever.
2098  virtual void commit() {
2099  // Nothing to be done, this action is not doing anything.
2100  }
2101  };
2102 
2103  /// Utility to remember the position of an instruction.
2104  class InsertionHandler {
2105  /// Position of an instruction.
2106  /// Either an instruction:
2107  /// - Is the first in a basic block: BB is used.
2108  /// - Has a previous instruction: PrevInst is used.
2109  union {
2110  Instruction *PrevInst;
2111  BasicBlock *BB;
2112  } Point;
2113 
2114  /// Remember whether or not the instruction had a previous instruction.
2115  bool HasPrevInstruction;
2116 
2117  public:
2118  /// Record the position of \p Inst.
2119  InsertionHandler(Instruction *Inst) {
2120  BasicBlock::iterator It = Inst->getIterator();
2121  HasPrevInstruction = (It != (Inst->getParent()->begin()));
2122  if (HasPrevInstruction)
2123  Point.PrevInst = &*--It;
2124  else
2125  Point.BB = Inst->getParent();
2126  }
2127 
2128  /// Insert \p Inst at the recorded position.
2129  void insert(Instruction *Inst) {
2130  if (HasPrevInstruction) {
2131  if (Inst->getParent())
2132  Inst->removeFromParent();
2133  Inst->insertAfter(Point.PrevInst);
2134  } else {
2135  Instruction *Position = &*Point.BB->getFirstInsertionPt();
2136  if (Inst->getParent())
2137  Inst->moveBefore(Position);
2138  else
2139  Inst->insertBefore(Position);
2140  }
2141  }
2142  };
2143 
2144  /// Move an instruction before another.
2145  class InstructionMoveBefore : public TypePromotionAction {
2146  /// Original position of the instruction.
2147  InsertionHandler Position;
2148 
2149  public:
2150  /// Move \p Inst before \p Before.
2151  InstructionMoveBefore(Instruction *Inst, Instruction *Before)
2152  : TypePromotionAction(Inst), Position(Inst) {
2153  LLVM_DEBUG(dbgs() << "Do: move: " << *Inst << "\nbefore: " << *Before
2154  << "\n");
2155  Inst->moveBefore(Before);
2156  }
2157 
2158  /// Move the instruction back to its original position.
2159  void undo() override {
2160  LLVM_DEBUG(dbgs() << "Undo: moveBefore: " << *Inst << "\n");
2161  Position.insert(Inst);
2162  }
2163  };
2164 
2165  /// Set the operand of an instruction with a new value.
2166  class OperandSetter : public TypePromotionAction {
2167  /// Original operand of the instruction.
2168  Value *Origin;
2169 
2170  /// Index of the modified instruction.
2171  unsigned Idx;
2172 
2173  public:
2174  /// Set \p Idx operand of \p Inst with \p NewVal.
2175  OperandSetter(Instruction *Inst, unsigned Idx, Value *NewVal)
2176  : TypePromotionAction(Inst), Idx(Idx) {
2177  LLVM_DEBUG(dbgs() << "Do: setOperand: " << Idx << "\n"
2178  << "for:" << *Inst << "\n"
2179  << "with:" << *NewVal << "\n");
2180  Origin = Inst->getOperand(Idx);
2181  Inst->setOperand(Idx, NewVal);
2182  }
2183 
2184  /// Restore the original value of the instruction.
2185  void undo() override {
2186  LLVM_DEBUG(dbgs() << "Undo: setOperand:" << Idx << "\n"
2187  << "for: " << *Inst << "\n"
2188  << "with: " << *Origin << "\n");
2189  Inst->setOperand(Idx, Origin);
2190  }
2191  };
2192 
2193  /// Hide the operands of an instruction.
2194  /// Do as if this instruction was not using any of its operands.
2195  class OperandsHider : public TypePromotionAction {
2196  /// The list of original operands.
2197  SmallVector<Value *, 4> OriginalValues;
2198 
2199  public:
2200  /// Remove \p Inst from the uses of the operands of \p Inst.
2201  OperandsHider(Instruction *Inst) : TypePromotionAction(Inst) {
2202  LLVM_DEBUG(dbgs() << "Do: OperandsHider: " << *Inst << "\n");
2203  unsigned NumOpnds = Inst->getNumOperands();
2204  OriginalValues.reserve(NumOpnds);
2205  for (unsigned It = 0; It < NumOpnds; ++It) {
2206  // Save the current operand.
2207  Value *Val = Inst->getOperand(It);
2208  OriginalValues.push_back(Val);
2209  // Set a dummy one.
2210  // We could use OperandSetter here, but that would imply an overhead
2211  // that we are not willing to pay.
2212  Inst->setOperand(It, UndefValue::get(Val->getType()));
2213  }
2214  }
2215 
2216  /// Restore the original list of uses.
2217  void undo() override {
2218  LLVM_DEBUG(dbgs() << "Undo: OperandsHider: " << *Inst << "\n");
2219  for (unsigned It = 0, EndIt = OriginalValues.size(); It != EndIt; ++It)
2220  Inst->setOperand(It, OriginalValues[It]);
2221  }
2222  };
2223 
2224  /// Build a truncate instruction.
2225  class TruncBuilder : public TypePromotionAction {
2226  Value *Val;
2227 
2228  public:
2229  /// Build a truncate instruction of \p Opnd producing a \p Ty
2230  /// result.
2231  /// trunc Opnd to Ty.
2232  TruncBuilder(Instruction *Opnd, Type *Ty) : TypePromotionAction(Opnd) {
2233  IRBuilder<> Builder(Opnd);
2234  Val = Builder.CreateTrunc(Opnd, Ty, "promoted");
2235  LLVM_DEBUG(dbgs() << "Do: TruncBuilder: " << *Val << "\n");
2236  }
2237 
2238  /// Get the built value.
2239  Value *getBuiltValue() { return Val; }
2240 
2241  /// Remove the built instruction.
2242  void undo() override {
2243  LLVM_DEBUG(dbgs() << "Undo: TruncBuilder: " << *Val << "\n");
2244  if (Instruction *IVal = dyn_cast<Instruction>(Val))
2245  IVal->eraseFromParent();
2246  }
2247  };
2248 
2249  /// Build a sign extension instruction.
2250  class SExtBuilder : public TypePromotionAction {
2251  Value *Val;
2252 
2253  public:
2254  /// Build a sign extension instruction of \p Opnd producing a \p Ty
2255  /// result.
2256  /// sext Opnd to Ty.
2257  SExtBuilder(Instruction *InsertPt, Value *Opnd, Type *Ty)
2258  : TypePromotionAction(InsertPt) {
2259  IRBuilder<> Builder(InsertPt);
2260  Val = Builder.CreateSExt(Opnd, Ty, "promoted");
2261  LLVM_DEBUG(dbgs() << "Do: SExtBuilder: " << *Val << "\n");
2262  }
2263 
2264  /// Get the built value.
2265  Value *getBuiltValue() { return Val; }
2266 
2267  /// Remove the built instruction.
2268  void undo() override {
2269  LLVM_DEBUG(dbgs() << "Undo: SExtBuilder: " << *Val << "\n");
2270  if (Instruction *IVal = dyn_cast<Instruction>(Val))
2271  IVal->eraseFromParent();
2272  }
2273  };
2274 
2275  /// Build a zero extension instruction.
2276  class ZExtBuilder : public TypePromotionAction {
2277  Value *Val;
2278 
2279  public:
2280  /// Build a zero extension instruction of \p Opnd producing a \p Ty
2281  /// result.
2282  /// zext Opnd to Ty.
2283  ZExtBuilder(Instruction *InsertPt, Value *Opnd, Type *Ty)
2284  : TypePromotionAction(InsertPt) {
2285  IRBuilder<> Builder(InsertPt);
2286  Val = Builder.CreateZExt(Opnd, Ty, "promoted");
2287  LLVM_DEBUG(dbgs() << "Do: ZExtBuilder: " << *Val << "\n");
2288  }
2289 
2290  /// Get the built value.
2291  Value *getBuiltValue() { return Val; }
2292 
2293  /// Remove the built instruction.
2294  void undo() override {
2295  LLVM_DEBUG(dbgs() << "Undo: ZExtBuilder: " << *Val << "\n");
2296  if (Instruction *IVal = dyn_cast<Instruction>(Val))
2297  IVal->eraseFromParent();
2298  }
2299  };
2300 
2301  /// Mutate an instruction to another type.
2302  class TypeMutator : public TypePromotionAction {
2303  /// Record the original type.
2304  Type *OrigTy;
2305 
2306  public:
2307  /// Mutate the type of \p Inst into \p NewTy.
2308  TypeMutator(Instruction *Inst, Type *NewTy)
2309  : TypePromotionAction(Inst), OrigTy(Inst->getType()) {
2310  LLVM_DEBUG(dbgs() << "Do: MutateType: " << *Inst << " with " << *NewTy
2311  << "\n");
2312  Inst->mutateType(NewTy);
2313  }
2314 
2315  /// Mutate the instruction back to its original type.
2316  void undo() override {
2317  LLVM_DEBUG(dbgs() << "Undo: MutateType: " << *Inst << " with " << *OrigTy
2318  << "\n");
2319  Inst->mutateType(OrigTy);
2320  }
2321  };
2322 
2323  /// Replace the uses of an instruction by another instruction.
2324  class UsesReplacer : public TypePromotionAction {
2325  /// Helper structure to keep track of the replaced uses.
2326  struct InstructionAndIdx {
2327  /// The instruction using the instruction.
2328  Instruction *Inst;
2329 
2330  /// The index where this instruction is used for Inst.
2331  unsigned Idx;
2332 
2333  InstructionAndIdx(Instruction *Inst, unsigned Idx)
2334  : Inst(Inst), Idx(Idx) {}
2335  };
2336 
2337  /// Keep track of the original uses (pair Instruction, Index).
2338  SmallVector<InstructionAndIdx, 4> OriginalUses;
2339 
2340  using use_iterator = SmallVectorImpl<InstructionAndIdx>::iterator;
2341 
2342  public:
2343  /// Replace all the use of \p Inst by \p New.
2344  UsesReplacer(Instruction *Inst, Value *New) : TypePromotionAction(Inst) {
2345  LLVM_DEBUG(dbgs() << "Do: UsersReplacer: " << *Inst << " with " << *New
2346  << "\n");
2347  // Record the original uses.
2348  for (Use &U : Inst->uses()) {
2349  Instruction *UserI = cast<Instruction>(U.getUser());
2350  OriginalUses.push_back(InstructionAndIdx(UserI, U.getOperandNo()));
2351  }
2352  // Now, we can replace the uses.
2353  Inst->replaceAllUsesWith(New);
2354  }
2355 
2356  /// Reassign the original uses of Inst to Inst.
2357  void undo() override {
2358  LLVM_DEBUG(dbgs() << "Undo: UsersReplacer: " << *Inst << "\n");
2359  for (use_iterator UseIt = OriginalUses.begin(),
2360  EndIt = OriginalUses.end();
2361  UseIt != EndIt; ++UseIt) {
2362  UseIt->Inst->setOperand(UseIt->Idx, Inst);
2363  }
2364  }
2365  };
2366 
2367  /// Remove an instruction from the IR.
2368  class InstructionRemover : public TypePromotionAction {
2369  /// Original position of the instruction.
2370  InsertionHandler Inserter;
2371 
2372  /// Helper structure to hide all the link to the instruction. In other
2373  /// words, this helps to do as if the instruction was removed.
2374  OperandsHider Hider;
2375 
2376  /// Keep track of the uses replaced, if any.
2377  UsesReplacer *Replacer = nullptr;
2378 
2379  /// Keep track of instructions removed.
2380  SetOfInstrs &RemovedInsts;
2381 
2382  public:
2383  /// Remove all reference of \p Inst and optionally replace all its
2384  /// uses with New.
2385  /// \p RemovedInsts Keep track of the instructions removed by this Action.
2386  /// \pre If !Inst->use_empty(), then New != nullptr
2387  InstructionRemover(Instruction *Inst, SetOfInstrs &RemovedInsts,
2388  Value *New = nullptr)
2389  : TypePromotionAction(Inst), Inserter(Inst), Hider(Inst),
2390  RemovedInsts(RemovedInsts) {
2391  if (New)
2392  Replacer = new UsesReplacer(Inst, New);
2393  LLVM_DEBUG(dbgs() << "Do: InstructionRemover: " << *Inst << "\n");
2394  RemovedInsts.insert(Inst);
2395  /// The instructions removed here will be freed after completing
2396  /// optimizeBlock() for all blocks as we need to keep track of the
2397  /// removed instructions during promotion.
2398  Inst->removeFromParent();
2399  }
2400 
2401  ~InstructionRemover() override { delete Replacer; }
2402 
2403  /// Resurrect the instruction and reassign it to the proper uses if
2404  /// new value was provided when build this action.
2405  void undo() override {
2406  LLVM_DEBUG(dbgs() << "Undo: InstructionRemover: " << *Inst << "\n");
2407  Inserter.insert(Inst);
2408  if (Replacer)
2409  Replacer->undo();
2410  Hider.undo();
2411  RemovedInsts.erase(Inst);
2412  }
2413  };
2414 
2415 public:
2416  /// Restoration point.
2417  /// The restoration point is a pointer to an action instead of an iterator
2418  /// because the iterator may be invalidated but not the pointer.
2419  using ConstRestorationPt = const TypePromotionAction *;
2420 
2421  TypePromotionTransaction(SetOfInstrs &RemovedInsts)
2422  : RemovedInsts(RemovedInsts) {}
2423 
2424  /// Advocate every changes made in that transaction.
2425  void commit();
2426 
2427  /// Undo all the changes made after the given point.
2428  void rollback(ConstRestorationPt Point);
2429 
2430  /// Get the current restoration point.
2431  ConstRestorationPt getRestorationPoint() const;
2432 
2433  /// \name API for IR modification with state keeping to support rollback.
2434  /// @{
2435  /// Same as Instruction::setOperand.
2436  void setOperand(Instruction *Inst, unsigned Idx, Value *NewVal);
2437 
2438  /// Same as Instruction::eraseFromParent.
2439  void eraseInstruction(Instruction *Inst, Value *NewVal = nullptr);
2440 
2441  /// Same as Value::replaceAllUsesWith.
2442  void replaceAllUsesWith(Instruction *Inst, Value *New);
2443 
2444  /// Same as Value::mutateType.
2445  void mutateType(Instruction *Inst, Type *NewTy);
2446 
2447  /// Same as IRBuilder::createTrunc.
2448  Value *createTrunc(Instruction *Opnd, Type *Ty);
2449 
2450  /// Same as IRBuilder::createSExt.
2451  Value *createSExt(Instruction *Inst, Value *Opnd, Type *Ty);
2452 
2453  /// Same as IRBuilder::createZExt.
2454  Value *createZExt(Instruction *Inst, Value *Opnd, Type *Ty);
2455 
2456  /// Same as Instruction::moveBefore.
2457  void moveBefore(Instruction *Inst, Instruction *Before);
2458  /// @}
2459 
2460 private:
2461  /// The ordered list of actions made so far.
2463 
2464  using CommitPt = SmallVectorImpl<std::unique_ptr<TypePromotionAction>>::iterator;
2465 
2466  SetOfInstrs &RemovedInsts;
2467 };
2468 
2469 } // end anonymous namespace
2470 
2471 void TypePromotionTransaction::setOperand(Instruction *Inst, unsigned Idx,
2472  Value *NewVal) {
2473  Actions.push_back(llvm::make_unique<TypePromotionTransaction::OperandSetter>(
2474  Inst, Idx, NewVal));
2475 }
2476 
2477 void TypePromotionTransaction::eraseInstruction(Instruction *Inst,
2478  Value *NewVal) {
2479  Actions.push_back(
2480  llvm::make_unique<TypePromotionTransaction::InstructionRemover>(
2481  Inst, RemovedInsts, NewVal));
2482 }
2483 
2484 void TypePromotionTransaction::replaceAllUsesWith(Instruction *Inst,
2485  Value *New) {
2486  Actions.push_back(
2487  llvm::make_unique<TypePromotionTransaction::UsesReplacer>(Inst, New));
2488 }
2489 
2490 void TypePromotionTransaction::mutateType(Instruction *Inst, Type *NewTy) {
2491  Actions.push_back(
2492  llvm::make_unique<TypePromotionTransaction::TypeMutator>(Inst, NewTy));
2493 }
2494 
2495 Value *TypePromotionTransaction::createTrunc(Instruction *Opnd,
2496  Type *Ty) {
2497  std::unique_ptr<TruncBuilder> Ptr(new TruncBuilder(Opnd, Ty));
2498  Value *Val = Ptr->getBuiltValue();
2499  Actions.push_back(std::move(Ptr));
2500  return Val;
2501 }
2502 
2503 Value *TypePromotionTransaction::createSExt(Instruction *Inst,
2504  Value *Opnd, Type *Ty) {
2505  std::unique_ptr<SExtBuilder> Ptr(new SExtBuilder(Inst, Opnd, Ty));
2506  Value *Val = Ptr->getBuiltValue();
2507  Actions.push_back(std::move(Ptr));
2508  return Val;
2509 }
2510 
2511 Value *TypePromotionTransaction::createZExt(Instruction *Inst,
2512  Value *Opnd, Type *Ty) {
2513  std::unique_ptr<ZExtBuilder> Ptr(new ZExtBuilder(Inst, Opnd, Ty));
2514  Value *Val = Ptr->getBuiltValue();
2515  Actions.push_back(std::move(Ptr));
2516  return Val;
2517 }
2518 
2519 void TypePromotionTransaction::moveBefore(Instruction *Inst,
2520  Instruction *Before) {
2521  Actions.push_back(
2522  llvm::make_unique<TypePromotionTransaction::InstructionMoveBefore>(
2523  Inst, Before));
2524 }
2525 
2526 TypePromotionTransaction::ConstRestorationPt
2527 TypePromotionTransaction::getRestorationPoint() const {
2528  return !Actions.empty() ? Actions.back().get() : nullptr;
2529 }
2530 
2531 void TypePromotionTransaction::commit() {
2532  for (CommitPt It = Actions.begin(), EndIt = Actions.end(); It != EndIt;
2533  ++It)
2534  (*It)->commit();
2535  Actions.clear();
2536 }
2537 
2538 void TypePromotionTransaction::rollback(
2539  TypePromotionTransaction::ConstRestorationPt Point) {
2540  while (!Actions.empty() && Point != Actions.back().get()) {
2541  std::unique_ptr<TypePromotionAction> Curr = Actions.pop_back_val();
2542  Curr->undo();
2543  }
2544 }
2545 
2546 namespace {
2547 
2548 /// A helper class for matching addressing modes.
2549 ///
2550 /// This encapsulates the logic for matching the target-legal addressing modes.
2551 class AddressingModeMatcher {
2552  SmallVectorImpl<Instruction*> &AddrModeInsts;
2553  const TargetLowering &TLI;
2554  const TargetRegisterInfo &TRI;
2555  const DataLayout &DL;
2556 
2557  /// AccessTy/MemoryInst - This is the type for the access (e.g. double) and
2558  /// the memory instruction that we're computing this address for.
2559  Type *AccessTy;
2560  unsigned AddrSpace;
2561  Instruction *MemoryInst;
2562 
2563  /// This is the addressing mode that we're building up. This is
2564  /// part of the return value of this addressing mode matching stuff.
2565  ExtAddrMode &AddrMode;
2566 
2567  /// The instructions inserted by other CodeGenPrepare optimizations.
2568  const SetOfInstrs &InsertedInsts;
2569 
2570  /// A map from the instructions to their type before promotion.
2571  InstrToOrigTy &PromotedInsts;
2572 
2573  /// The ongoing transaction where every action should be registered.
2574  TypePromotionTransaction &TPT;
2575 
2576  // A GEP which has too large offset to be folded into the addressing mode.
2577  std::pair<AssertingVH<GetElementPtrInst>, int64_t> &LargeOffsetGEP;
2578 
2579  /// This is set to true when we should not do profitability checks.
2580  /// When true, IsProfitableToFoldIntoAddressingMode always returns true.
2581  bool IgnoreProfitability;
2582 
2583  AddressingModeMatcher(
2585  const TargetRegisterInfo &TRI, Type *AT, unsigned AS, Instruction *MI,
2586  ExtAddrMode &AM, const SetOfInstrs &InsertedInsts,
2587  InstrToOrigTy &PromotedInsts, TypePromotionTransaction &TPT,
2588  std::pair<AssertingVH<GetElementPtrInst>, int64_t> &LargeOffsetGEP)
2589  : AddrModeInsts(AMI), TLI(TLI), TRI(TRI),
2590  DL(MI->getModule()->getDataLayout()), AccessTy(AT), AddrSpace(AS),
2591  MemoryInst(MI), AddrMode(AM), InsertedInsts(InsertedInsts),
2592  PromotedInsts(PromotedInsts), TPT(TPT), LargeOffsetGEP(LargeOffsetGEP) {
2593  IgnoreProfitability = false;
2594  }
2595 
2596 public:
2597  /// Find the maximal addressing mode that a load/store of V can fold,
2598  /// give an access type of AccessTy. This returns a list of involved
2599  /// instructions in AddrModeInsts.
2600  /// \p InsertedInsts The instructions inserted by other CodeGenPrepare
2601  /// optimizations.
2602  /// \p PromotedInsts maps the instructions to their type before promotion.
2603  /// \p The ongoing transaction where every action should be registered.
2604  static ExtAddrMode
2605  Match(Value *V, Type *AccessTy, unsigned AS, Instruction *MemoryInst,
2606  SmallVectorImpl<Instruction *> &AddrModeInsts,
2607  const TargetLowering &TLI, const TargetRegisterInfo &TRI,
2608  const SetOfInstrs &InsertedInsts, InstrToOrigTy &PromotedInsts,
2609  TypePromotionTransaction &TPT,
2610  std::pair<AssertingVH<GetElementPtrInst>, int64_t> &LargeOffsetGEP) {
2611  ExtAddrMode Result;
2612 
2613  bool Success = AddressingModeMatcher(AddrModeInsts, TLI, TRI, AccessTy, AS,
2614  MemoryInst, Result, InsertedInsts,
2615  PromotedInsts, TPT, LargeOffsetGEP)
2616  .matchAddr(V, 0);
2617  (void)Success; assert(Success && "Couldn't select *anything*?");
2618  return Result;
2619  }
2620 
2621 private:
2622  bool matchScaledValue(Value *ScaleReg, int64_t Scale, unsigned Depth);
2623  bool matchAddr(Value *Addr, unsigned Depth);
2624  bool matchOperationAddr(User *AddrInst, unsigned Opcode, unsigned Depth,
2625  bool *MovedAway = nullptr);
2626  bool isProfitableToFoldIntoAddressingMode(Instruction *I,
2627  ExtAddrMode &AMBefore,
2628  ExtAddrMode &AMAfter);
2629  bool valueAlreadyLiveAtInst(Value *Val, Value *KnownLive1, Value *KnownLive2);
2630  bool isPromotionProfitable(unsigned NewCost, unsigned OldCost,
2631  Value *PromotedOperand) const;
2632 };
2633 
2634 /// Keep track of simplification of Phi nodes.
2635 /// Accept the set of all phi nodes and erase phi node from this set
2636 /// if it is simplified.
2637 class SimplificationTracker {
2639  const SimplifyQuery &SQ;
2640  // Tracks newly created Phi nodes. We use a SetVector to get deterministic
2641  // order when iterating over the set in MatchPhiSet.
2642  SmallSetVector<PHINode *, 32> AllPhiNodes;
2643  // Tracks newly created Select nodes.
2644  SmallPtrSet<SelectInst *, 32> AllSelectNodes;
2645 
2646 public:
2647  SimplificationTracker(const SimplifyQuery &sq)
2648  : SQ(sq) {}
2649 
2650  Value *Get(Value *V) {
2651  do {
2652  auto SV = Storage.find(V);
2653  if (SV == Storage.end())
2654  return V;
2655  V = SV->second;
2656  } while (true);
2657  }
2658 
2659  Value *Simplify(Value *Val) {
2660  SmallVector<Value *, 32> WorkList;
2661  SmallPtrSet<Value *, 32> Visited;
2662  WorkList.push_back(Val);
2663  while (!WorkList.empty()) {
2664  auto P = WorkList.pop_back_val();
2665  if (!Visited.insert(P).second)
2666  continue;
2667  if (auto *PI = dyn_cast<Instruction>(P))
2668  if (Value *V = SimplifyInstruction(cast<Instruction>(PI), SQ)) {
2669  for (auto *U : PI->users())
2670  WorkList.push_back(cast<Value>(U));
2671  Put(PI, V);
2672  PI->replaceAllUsesWith(V);
2673  if (auto *PHI = dyn_cast<PHINode>(PI))
2674  AllPhiNodes.remove(PHI);
2675  if (auto *Select = dyn_cast<SelectInst>(PI))
2676  AllSelectNodes.erase(Select);
2677  PI->eraseFromParent();
2678  }
2679  }
2680  return Get(Val);
2681  }
2682 
2683  void Put(Value *From, Value *To) {
2684  Storage.insert({ From, To });
2685  }
2686 
2687  void ReplacePhi(PHINode *From, PHINode *To) {
2688  Value* OldReplacement = Get(From);
2689  while (OldReplacement != From) {
2690  From = To;
2691  To = dyn_cast<PHINode>(OldReplacement);
2692  OldReplacement = Get(From);
2693  }
2694  assert(Get(To) == To && "Replacement PHI node is already replaced.");
2695  Put(From, To);
2696  From->replaceAllUsesWith(To);
2697  AllPhiNodes.remove(From);
2698  From->eraseFromParent();
2699  }
2700 
2701  SmallSetVector<PHINode *, 32>& newPhiNodes() { return AllPhiNodes; }
2702 
2703  void insertNewPhi(PHINode *PN) { AllPhiNodes.insert(PN); }
2704 
2705  void insertNewSelect(SelectInst *SI) { AllSelectNodes.insert(SI); }
2706 
2707  unsigned countNewPhiNodes() const { return AllPhiNodes.size(); }
2708 
2709  unsigned countNewSelectNodes() const { return AllSelectNodes.size(); }
2710 
2711  void destroyNewNodes(Type *CommonType) {
2712  // For safe erasing, replace the uses with dummy value first.
2713  auto Dummy = UndefValue::get(CommonType);
2714  for (auto I : AllPhiNodes) {
2716  I->eraseFromParent();
2717  }
2718  AllPhiNodes.clear();
2719  for (auto I : AllSelectNodes) {
2721  I->eraseFromParent();
2722  }
2723  AllSelectNodes.clear();
2724  }
2725 };
2726 
2727 /// A helper class for combining addressing modes.
2728 class AddressingModeCombiner {
2729  typedef std::pair<Value *, BasicBlock *> ValueInBB;
2730  typedef DenseMap<ValueInBB, Value *> FoldAddrToValueMapping;
2731  typedef std::pair<PHINode *, PHINode *> PHIPair;
2732 
2733 private:
2734  /// The addressing modes we've collected.
2735  SmallVector<ExtAddrMode, 16> AddrModes;
2736 
2737  /// The field in which the AddrModes differ, when we have more than one.
2738  ExtAddrMode::FieldName DifferentField = ExtAddrMode::NoField;
2739 
2740  /// Are the AddrModes that we have all just equal to their original values?
2741  bool AllAddrModesTrivial = true;
2742 
2743  /// Common Type for all different fields in addressing modes.
2744  Type *CommonType;
2745 
2746  /// SimplifyQuery for simplifyInstruction utility.
2747  const SimplifyQuery &SQ;
2748 
2749  /// Original Address.
2750  ValueInBB Original;
2751 
2752 public:
2753  AddressingModeCombiner(const SimplifyQuery &_SQ, ValueInBB OriginalValue)
2754  : CommonType(nullptr), SQ(_SQ), Original(OriginalValue) {}
2755 
2756  /// Get the combined AddrMode
2757  const ExtAddrMode &getAddrMode() const {
2758  return AddrModes[0];
2759  }
2760 
2761  /// Add a new AddrMode if it's compatible with the AddrModes we already
2762  /// have.
2763  /// \return True iff we succeeded in doing so.
2764  bool addNewAddrMode(ExtAddrMode &NewAddrMode) {
2765  // Take note of if we have any non-trivial AddrModes, as we need to detect
2766  // when all AddrModes are trivial as then we would introduce a phi or select
2767  // which just duplicates what's already there.
2768  AllAddrModesTrivial = AllAddrModesTrivial && NewAddrMode.isTrivial();
2769 
2770  // If this is the first addrmode then everything is fine.
2771  if (AddrModes.empty()) {
2772  AddrModes.emplace_back(NewAddrMode);
2773  return true;
2774  }
2775 
2776  // Figure out how different this is from the other address modes, which we
2777  // can do just by comparing against the first one given that we only care
2778  // about the cumulative difference.
2779  ExtAddrMode::FieldName ThisDifferentField =
2780  AddrModes[0].compare(NewAddrMode);
2781  if (DifferentField == ExtAddrMode::NoField)
2782  DifferentField = ThisDifferentField;
2783  else if (DifferentField != ThisDifferentField)
2784  DifferentField = ExtAddrMode::MultipleFields;
2785 
2786  // If NewAddrMode differs in more than one dimension we cannot handle it.
2787  bool CanHandle = DifferentField != ExtAddrMode::MultipleFields;
2788 
2789  // If Scale Field is different then we reject.
2790  CanHandle = CanHandle && DifferentField != ExtAddrMode::ScaleField;
2791 
2792  // We also must reject the case when base offset is different and
2793  // scale reg is not null, we cannot handle this case due to merge of
2794  // different offsets will be used as ScaleReg.
2795  CanHandle = CanHandle && (DifferentField != ExtAddrMode::BaseOffsField ||
2796  !NewAddrMode.ScaledReg);
2797 
2798  // We also must reject the case when GV is different and BaseReg installed
2799  // due to we want to use base reg as a merge of GV values.
2800  CanHandle = CanHandle && (DifferentField != ExtAddrMode::BaseGVField ||
2801  !NewAddrMode.HasBaseReg);
2802 
2803  // Even if NewAddMode is the same we still need to collect it due to
2804  // original value is different. And later we will need all original values
2805  // as anchors during finding the common Phi node.
2806  if (CanHandle)
2807  AddrModes.emplace_back(NewAddrMode);
2808  else
2809  AddrModes.clear();
2810 
2811  return CanHandle;
2812  }
2813 
2814  /// Combine the addressing modes we've collected into a single
2815  /// addressing mode.
2816  /// \return True iff we successfully combined them or we only had one so
2817  /// didn't need to combine them anyway.
2818  bool combineAddrModes() {
2819  // If we have no AddrModes then they can't be combined.
2820  if (AddrModes.size() == 0)
2821  return false;
2822 
2823  // A single AddrMode can trivially be combined.
2824  if (AddrModes.size() == 1 || DifferentField == ExtAddrMode::NoField)
2825  return true;
2826 
2827  // If the AddrModes we collected are all just equal to the value they are
2828  // derived from then combining them wouldn't do anything useful.
2829  if (AllAddrModesTrivial)
2830  return false;
2831 
2832  if (!addrModeCombiningAllowed())
2833  return false;
2834 
2835  // Build a map between <original value, basic block where we saw it> to
2836  // value of base register.
2837  // Bail out if there is no common type.
2838  FoldAddrToValueMapping Map;
2839  if (!initializeMap(Map))
2840  return false;
2841 
2842  Value *CommonValue = findCommon(Map);
2843  if (CommonValue)
2844  AddrModes[0].SetCombinedField(DifferentField, CommonValue, AddrModes);
2845  return CommonValue != nullptr;
2846  }
2847 
2848 private:
2849  /// Initialize Map with anchor values. For address seen in some BB
2850  /// we set the value of different field saw in this address.
2851  /// If address is not an instruction than basic block is set to null.
2852  /// At the same time we find a common type for different field we will
2853  /// use to create new Phi/Select nodes. Keep it in CommonType field.
2854  /// Return false if there is no common type found.
2855  bool initializeMap(FoldAddrToValueMapping &Map) {
2856  // Keep track of keys where the value is null. We will need to replace it
2857  // with constant null when we know the common type.
2858  SmallVector<ValueInBB, 2> NullValue;
2859  Type *IntPtrTy = SQ.DL.getIntPtrType(AddrModes[0].OriginalValue->getType());
2860  for (auto &AM : AddrModes) {
2861  BasicBlock *BB = nullptr;
2862  if (Instruction *I = dyn_cast<Instruction>(AM.OriginalValue))
2863  BB = I->getParent();
2864 
2865  Value *DV = AM.GetFieldAsValue(DifferentField, IntPtrTy);
2866  if (DV) {
2867  auto *Type = DV->getType();
2868  if (CommonType && CommonType != Type)
2869  return false;
2870  CommonType = Type;
2871  Map[{ AM.OriginalValue, BB }] = DV;
2872  } else {
2873  NullValue.push_back({ AM.OriginalValue, BB });
2874  }
2875  }
2876  assert(CommonType && "At least one non-null value must be!");
2877  for (auto VIBB : NullValue)
2878  Map[VIBB] = Constant::getNullValue(CommonType);
2879  return true;
2880  }
2881 
2882  /// We have mapping between value A and basic block where value A
2883  /// seen to other value B where B was a field in addressing mode represented
2884  /// by A. Also we have an original value C representing an address in some
2885  /// basic block. Traversing from C through phi and selects we ended up with
2886  /// A's in a map. This utility function tries to find a value V which is a
2887  /// field in addressing mode C and traversing through phi nodes and selects
2888  /// we will end up in corresponded values B in a map.
2889  /// The utility will create a new Phi/Selects if needed.
2890  // The simple example looks as follows:
2891  // BB1:
2892  // p1 = b1 + 40
2893  // br cond BB2, BB3
2894  // BB2:
2895  // p2 = b2 + 40
2896  // br BB3
2897  // BB3:
2898  // p = phi [p1, BB1], [p2, BB2]
2899  // v = load p
2900  // Map is
2901  // <p1, BB1> -> b1
2902  // <p2, BB2> -> b2
2903  // Request is
2904  // <p, BB3> -> ?
2905  // The function tries to find or build phi [b1, BB1], [b2, BB2] in BB3
2906  Value *findCommon(FoldAddrToValueMapping &Map) {
2907  // Tracks the simplification of newly created phi nodes. The reason we use
2908  // this mapping is because we will add new created Phi nodes in AddrToBase.
2909  // Simplification of Phi nodes is recursive, so some Phi node may
2910  // be simplified after we added it to AddrToBase.
2911  // Using this mapping we can find the current value in AddrToBase.
2912  SimplificationTracker ST(SQ);
2913 
2914  // First step, DFS to create PHI nodes for all intermediate blocks.
2915  // Also fill traverse order for the second step.
2916  SmallVector<ValueInBB, 32> TraverseOrder;
2917  InsertPlaceholders(Map, TraverseOrder, ST);
2918 
2919  // Second Step, fill new nodes by merged values and simplify if possible.
2920  FillPlaceholders(Map, TraverseOrder, ST);
2921 
2922  if (!AddrSinkNewSelects && ST.countNewSelectNodes() > 0) {
2923  ST.destroyNewNodes(CommonType);
2924  return nullptr;
2925  }
2926 
2927  // Now we'd like to match New Phi nodes to existed ones.
2928  unsigned PhiNotMatchedCount = 0;
2929  if (!MatchPhiSet(ST, AddrSinkNewPhis, PhiNotMatchedCount)) {
2930  ST.destroyNewNodes(CommonType);
2931  return nullptr;
2932  }
2933 
2934  auto *Result = ST.Get(Map.find(Original)->second);
2935  if (Result) {
2936  NumMemoryInstsPhiCreated += ST.countNewPhiNodes() + PhiNotMatchedCount;
2937  NumMemoryInstsSelectCreated += ST.countNewSelectNodes();
2938  }
2939  return Result;
2940  }
2941 
2942  /// Try to match PHI node to Candidate.
2943  /// Matcher tracks the matched Phi nodes.
2944  bool MatchPhiNode(PHINode *PHI, PHINode *Candidate,
2945  SmallSetVector<PHIPair, 8> &Matcher,
2946  SmallSetVector<PHINode *, 32> &PhiNodesToMatch) {
2947  SmallVector<PHIPair, 8> WorkList;
2948  Matcher.insert({ PHI, Candidate });
2949  WorkList.push_back({ PHI, Candidate });
2950  SmallSet<PHIPair, 8> Visited;
2951  while (!WorkList.empty()) {
2952  auto Item = WorkList.pop_back_val();
2953  if (!Visited.insert(Item).second)
2954  continue;
2955  // We iterate over all incoming values to Phi to compare them.
2956  // If values are different and both of them Phi and the first one is a
2957  // Phi we added (subject to match) and both of them is in the same basic
2958  // block then we can match our pair if values match. So we state that
2959  // these values match and add it to work list to verify that.
2960  for (auto B : Item.first->blocks()) {
2961  Value *FirstValue = Item.first->getIncomingValueForBlock(B);
2962  Value *SecondValue = Item.second->getIncomingValueForBlock(B);
2963  if (FirstValue == SecondValue)
2964  continue;
2965 
2966  PHINode *FirstPhi = dyn_cast<PHINode>(FirstValue);
2967  PHINode *SecondPhi = dyn_cast<PHINode>(SecondValue);
2968 
2969  // One of them is not Phi or
2970  // The first one is not Phi node from the set we'd like to match or
2971  // Phi nodes from different basic blocks then
2972  // we will not be able to match.
2973  if (!FirstPhi || !SecondPhi || !PhiNodesToMatch.count(FirstPhi) ||
2974  FirstPhi->getParent() != SecondPhi->getParent())
2975  return false;
2976 
2977  // If we already matched them then continue.
2978  if (Matcher.count({ FirstPhi, SecondPhi }))
2979  continue;
2980  // So the values are different and does not match. So we need them to
2981  // match.
2982  Matcher.insert({ FirstPhi, SecondPhi });
2983  // But me must check it.
2984  WorkList.push_back({ FirstPhi, SecondPhi });
2985  }
2986  }
2987  return true;
2988  }
2989 
2990  /// For the given set of PHI nodes (in the SimplificationTracker) try
2991  /// to find their equivalents.
2992  /// Returns false if this matching fails and creation of new Phi is disabled.
2993  bool MatchPhiSet(SimplificationTracker &ST, bool AllowNewPhiNodes,
2994  unsigned &PhiNotMatchedCount) {
2995  // Use a SetVector for Matched to make sure we do replacements (ReplacePhi)
2996  // in a deterministic order below.
2998  SmallPtrSet<PHINode *, 8> WillNotMatch;
2999  SmallSetVector<PHINode *, 32> &PhiNodesToMatch = ST.newPhiNodes();
3000  while (PhiNodesToMatch.size()) {
3001  PHINode *PHI = *PhiNodesToMatch.begin();
3002 
3003  // Add us, if no Phi nodes in the basic block we do not match.
3004  WillNotMatch.clear();
3005  WillNotMatch.insert(PHI);
3006 
3007  // Traverse all Phis until we found equivalent or fail to do that.
3008  bool IsMatched = false;
3009  for (auto &P : PHI->getParent()->phis()) {
3010  if (&P == PHI)
3011  continue;
3012  if ((IsMatched = MatchPhiNode(PHI, &P, Matched, PhiNodesToMatch)))
3013  break;
3014  // If it does not match, collect all Phi nodes from matcher.
3015  // if we end up with no match, them all these Phi nodes will not match
3016  // later.
3017  for (auto M : Matched)
3018  WillNotMatch.insert(M.first);
3019  Matched.clear();
3020  }
3021  if (IsMatched) {
3022  // Replace all matched values and erase them.
3023  for (auto MV : Matched)
3024  ST.ReplacePhi(MV.first, MV.second);
3025  Matched.clear();
3026  continue;
3027  }
3028  // If we are not allowed to create new nodes then bail out.
3029  if (!AllowNewPhiNodes)
3030  return false;
3031  // Just remove all seen values in matcher. They will not match anything.
3032  PhiNotMatchedCount += WillNotMatch.size();
3033  for (auto *P : WillNotMatch)
3034  PhiNodesToMatch.remove(P);
3035  }
3036  return true;
3037  }
3038  /// Fill the placeholder with values from predecessors and simplify it.
3039  void FillPlaceholders(FoldAddrToValueMapping &Map,
3040  SmallVectorImpl<ValueInBB> &TraverseOrder,
3041  SimplificationTracker &ST) {
3042  while (!TraverseOrder.empty()) {
3043  auto Current = TraverseOrder.pop_back_val();
3044  assert(Map.find(Current) != Map.end() && "No node to fill!!!");
3045  Value *CurrentValue = Current.first;
3046  BasicBlock *CurrentBlock = Current.second;
3047  Value *V = Map[Current];
3048 
3049  if (SelectInst *Select = dyn_cast<SelectInst>(V)) {
3050  // CurrentValue also must be Select.
3051  auto *CurrentSelect = cast<SelectInst>(CurrentValue);
3052  auto *TrueValue = CurrentSelect->getTrueValue();
3053  ValueInBB TrueItem = { TrueValue, isa<Instruction>(TrueValue)
3054  ? CurrentBlock
3055  : nullptr };
3056  assert(Map.find(TrueItem) != Map.end() && "No True Value!");
3057  Select->setTrueValue(ST.Get(Map[TrueItem]));
3058  auto *FalseValue = CurrentSelect->getFalseValue();
3059  ValueInBB FalseItem = { FalseValue, isa<Instruction>(FalseValue)
3060  ? CurrentBlock
3061  : nullptr };
3062  assert(Map.find(FalseItem) != Map.end() && "No False Value!");
3063  Select->setFalseValue(ST.Get(Map[FalseItem]));
3064  } else {
3065  // Must be a Phi node then.
3066  PHINode *PHI = cast<PHINode>(V);
3067  // Fill the Phi node with values from predecessors.
3068  bool IsDefinedInThisBB =
3069  cast<Instruction>(CurrentValue)->getParent() == CurrentBlock;
3070  auto *CurrentPhi = dyn_cast<PHINode>(CurrentValue);
3071  for (auto B : predecessors(CurrentBlock)) {
3072  Value *PV = IsDefinedInThisBB
3073  ? CurrentPhi->getIncomingValueForBlock(B)
3074  : CurrentValue;
3075  ValueInBB item = { PV, isa<Instruction>(PV) ? B : nullptr };
3076  assert(Map.find(item) != Map.end() && "No predecessor Value!");
3077  PHI->addIncoming(ST.Get(Map[item]), B);
3078  }
3079  }
3080  // Simplify if possible.
3081  Map[Current] = ST.Simplify(V);
3082  }
3083  }
3084 
3085  /// Starting from value recursively iterates over predecessors up to known
3086  /// ending values represented in a map. For each traversed block inserts
3087  /// a placeholder Phi or Select.
3088  /// Reports all new created Phi/Select nodes by adding them to set.
3089  /// Also reports and order in what basic blocks have been traversed.
3090  void InsertPlaceholders(FoldAddrToValueMapping &Map,
3091  SmallVectorImpl<ValueInBB> &TraverseOrder,
3092  SimplificationTracker &ST) {
3093  SmallVector<ValueInBB, 32> Worklist;
3094  assert((isa<PHINode>(Original.first) || isa<SelectInst>(Original.first)) &&
3095  "Address must be a Phi or Select node");
3096  auto *Dummy = UndefValue::get(CommonType);
3097  Worklist.push_back(Original);
3098  while (!Worklist.empty()) {
3099  auto Current = Worklist.pop_back_val();
3100  // If value is not an instruction it is something global, constant,
3101  // parameter and we can say that this value is observable in any block.
3102  // Set block to null to denote it.
3103  // Also please take into account that it is how we build anchors.
3104  if (!isa<Instruction>(Current.first))
3105  Current.second = nullptr;
3106  // if it is already visited or it is an ending value then skip it.
3107  if (Map.find(Current) != Map.end())
3108  continue;
3109  TraverseOrder.push_back(Current);
3110 
3111  Value *CurrentValue = Current.first;
3112  BasicBlock *CurrentBlock = Current.second;
3113  // CurrentValue must be a Phi node or select. All others must be covered
3114  // by anchors.
3115  Instruction *CurrentI = cast<Instruction>(CurrentValue);
3116  bool IsDefinedInThisBB = CurrentI->getParent() == CurrentBlock;
3117 
3118  unsigned PredCount = pred_size(CurrentBlock);
3119  // if Current Value is not defined in this basic block we are interested
3120  // in values in predecessors.
3121  if (!IsDefinedInThisBB) {
3122  assert(PredCount && "Unreachable block?!");
3123  PHINode *PHI = PHINode::Create(CommonType, PredCount, "sunk_phi",
3124  &CurrentBlock->front());
3125  Map[Current] = PHI;
3126  ST.insertNewPhi(PHI);
3127  // Add all predecessors in work list.
3128  for (auto B : predecessors(CurrentBlock))
3129  Worklist.push_back({ CurrentValue, B });
3130  continue;
3131  }
3132  // Value is defined in this basic block.
3133  if (SelectInst *OrigSelect = dyn_cast<SelectInst>(CurrentI)) {
3134  // Is it OK to get metadata from OrigSelect?!
3135  // Create a Select placeholder with dummy value.
3136  SelectInst *Select =
3137  SelectInst::Create(OrigSelect->getCondition(), Dummy, Dummy,
3138  OrigSelect->getName(), OrigSelect, OrigSelect);
3139  Map[Current] = Select;
3140  ST.insertNewSelect(Select);
3141  // We are interested in True and False value in this basic block.
3142  Worklist.push_back({ OrigSelect->getTrueValue(), CurrentBlock });
3143  Worklist.push_back({ OrigSelect->getFalseValue(), CurrentBlock });
3144  } else {
3145  // It must be a Phi node then.
3146  auto *CurrentPhi = cast<PHINode>(CurrentI);
3147  // Create new Phi node for merge of bases.
3148  assert(PredCount && "Unreachable block?!");
3149  PHINode *PHI = PHINode::Create(CommonType, PredCount, "sunk_phi",
3150  &CurrentBlock->front());
3151  Map[Current] = PHI;
3152  ST.insertNewPhi(PHI);
3153 
3154  // Add all predecessors in work list.
3155  for (auto B : predecessors(CurrentBlock))
3156  Worklist.push_back({ CurrentPhi->getIncomingValueForBlock(B), B });
3157  }
3158  }
3159  }
3160 
3161  bool addrModeCombiningAllowed() {
3163  return false;
3164  switch (DifferentField) {
3165  default:
3166  return false;
3167  case ExtAddrMode::BaseRegField:
3168  return AddrSinkCombineBaseReg;
3169  case ExtAddrMode::BaseGVField:
3170  return AddrSinkCombineBaseGV;
3171  case ExtAddrMode::BaseOffsField:
3172  return AddrSinkCombineBaseOffs;
3173  case ExtAddrMode::ScaledRegField:
3174  return AddrSinkCombineScaledReg;
3175  }
3176  }
3177 };
3178 } // end anonymous namespace
3179 
3180 /// Try adding ScaleReg*Scale to the current addressing mode.
3181 /// Return true and update AddrMode if this addr mode is legal for the target,
3182 /// false if not.
3183 bool AddressingModeMatcher::matchScaledValue(Value *ScaleReg, int64_t Scale,
3184  unsigned Depth) {
3185  // If Scale is 1, then this is the same as adding ScaleReg to the addressing
3186  // mode. Just process that directly.
3187  if (Scale == 1)
3188  return matchAddr(ScaleReg, Depth);
3189 
3190  // If the scale is 0, it takes nothing to add this.
3191  if (Scale == 0)
3192  return true;
3193 
3194  // If we already have a scale of this value, we can add to it, otherwise, we
3195  // need an available scale field.
3196  if (AddrMode.Scale != 0 && AddrMode.ScaledReg != ScaleReg)
3197  return false;
3198 
3199  ExtAddrMode TestAddrMode = AddrMode;
3200 
3201  // Add scale to turn X*4+X*3 -> X*7. This could also do things like
3202  // [A+B + A*7] -> [B+A*8].
3203  TestAddrMode.Scale += Scale;
3204  TestAddrMode.ScaledReg = ScaleReg;
3205 
3206  // If the new address isn't legal, bail out.
3207  if (!TLI.isLegalAddressingMode(DL, TestAddrMode, AccessTy, AddrSpace))
3208  return false;
3209 
3210  // It was legal, so commit it.
3211  AddrMode = TestAddrMode;
3212 
3213  // Okay, we decided that we can add ScaleReg+Scale to AddrMode. Check now
3214  // to see if ScaleReg is actually X+C. If so, we can turn this into adding
3215  // X*Scale + C*Scale to addr mode.
3216  ConstantInt *CI = nullptr; Value *AddLHS = nullptr;
3217  if (isa<Instruction>(ScaleReg) && // not a constant expr.
3218  match(ScaleReg, m_Add(m_Value(AddLHS), m_ConstantInt(CI)))) {
3219  TestAddrMode.ScaledReg = AddLHS;
3220  TestAddrMode.BaseOffs += CI->getSExtValue()*TestAddrMode.Scale;
3221 
3222  // If this addressing mode is legal, commit it and remember that we folded
3223  // this instruction.
3224  if (TLI.isLegalAddressingMode(DL, TestAddrMode, AccessTy, AddrSpace)) {
3225  AddrModeInsts.push_back(cast<Instruction>(ScaleReg));
3226  AddrMode = TestAddrMode;
3227  return true;
3228  }
3229  }
3230 
3231  // Otherwise, not (x+c)*scale, just return what we have.
3232  return true;
3233 }
3234 
3235 /// This is a little filter, which returns true if an addressing computation
3236 /// involving I might be folded into a load/store accessing it.
3237 /// This doesn't need to be perfect, but needs to accept at least
3238 /// the set of instructions that MatchOperationAddr can.
3240  switch (I->getOpcode()) {
3241  case Instruction::BitCast:
3242  case Instruction::AddrSpaceCast:
3243  // Don't touch identity bitcasts.
3244  if (I->getType() == I->getOperand(0)->getType())
3245  return false;
3246  return I->getType()->isIntOrPtrTy();
3247  case Instruction::PtrToInt:
3248  // PtrToInt is always a noop, as we know that the int type is pointer sized.
3249  return true;
3250  case Instruction::IntToPtr:
3251  // We know the input is intptr_t, so this is foldable.
3252  return true;
3253  case Instruction::Add:
3254  return true;
3255  case Instruction::Mul:
3256  case Instruction::Shl:
3257  // Can only handle X*C and X << C.
3258  return isa<ConstantInt>(I->getOperand(1));
3259  case Instruction::GetElementPtr:
3260  return true;
3261  default:
3262  return false;
3263  }
3264 }
3265 
3266 /// Check whether or not \p Val is a legal instruction for \p TLI.
3267 /// \note \p Val is assumed to be the product of some type promotion.
3268 /// Therefore if \p Val has an undefined state in \p TLI, this is assumed
3269 /// to be legal, as the non-promoted value would have had the same state.
3271  const DataLayout &DL, Value *Val) {
3272  Instruction *PromotedInst = dyn_cast<Instruction>(Val);
3273  if (!PromotedInst)
3274  return false;
3275  int ISDOpcode = TLI.InstructionOpcodeToISD(PromotedInst->getOpcode());
3276  // If the ISDOpcode is undefined, it was undefined before the promotion.
3277  if (!ISDOpcode)
3278  return true;
3279  // Otherwise, check if the promoted instruction is legal or not.
3280  return TLI.isOperationLegalOrCustom(
3281  ISDOpcode, TLI.getValueType(DL, PromotedInst->getType()));
3282 }
3283 
3284 namespace {
3285 
3286 /// Hepler class to perform type promotion.
3287 class TypePromotionHelper {
3288  /// Utility function to add a promoted instruction \p ExtOpnd to
3289  /// \p PromotedInsts and record the type of extension we have seen.
3290  static void addPromotedInst(InstrToOrigTy &PromotedInsts,
3291  Instruction *ExtOpnd,
3292  bool IsSExt) {
3293  ExtType ExtTy = IsSExt ? SignExtension : ZeroExtension;
3294  InstrToOrigTy::iterator It = PromotedInsts.find(ExtOpnd);
3295  if (It != PromotedInsts.end()) {
3296  // If the new extension is same as original, the information in
3297  // PromotedInsts[ExtOpnd] is still correct.
3298  if (It->second.getInt() == ExtTy)
3299  return;
3300 
3301  // Now the new extension is different from old extension, we make
3302  // the type information invalid by setting extension type to
3303  // BothExtension.
3304  ExtTy = BothExtension;
3305  }
3306  PromotedInsts[ExtOpnd] = TypeIsSExt(ExtOpnd->getType(), ExtTy);
3307  }
3308 
3309  /// Utility function to query the original type of instruction \p Opnd
3310  /// with a matched extension type. If the extension doesn't match, we
3311  /// cannot use the information we had on the original type.
3312  /// BothExtension doesn't match any extension type.
3313  static const Type *getOrigType(const InstrToOrigTy &PromotedInsts,
3314  Instruction *Opnd,
3315  bool IsSExt) {
3316  ExtType ExtTy = IsSExt ? SignExtension : ZeroExtension;
3317  InstrToOrigTy::const_iterator It = PromotedInsts.find(Opnd);
3318  if (It != PromotedInsts.end() && It->second.getInt() == ExtTy)
3319  return It->second.getPointer();
3320  return nullptr;
3321  }
3322 
3323  /// Utility function to check whether or not a sign or zero extension
3324  /// of \p Inst with \p ConsideredExtType can be moved through \p Inst by
3325  /// either using the operands of \p Inst or promoting \p Inst.
3326  /// The type of the extension is defined by \p IsSExt.
3327  /// In other words, check if:
3328  /// ext (Ty Inst opnd1 opnd2 ... opndN) to ConsideredExtType.
3329  /// #1 Promotion applies:
3330  /// ConsideredExtType Inst (ext opnd1 to ConsideredExtType, ...).
3331  /// #2 Operand reuses:
3332  /// ext opnd1 to ConsideredExtType.
3333  /// \p PromotedInsts maps the instructions to their type before promotion.
3334  static bool canGetThrough(const Instruction *Inst, Type *ConsideredExtType,
3335  const InstrToOrigTy &PromotedInsts, bool IsSExt);
3336 
3337  /// Utility function to determine if \p OpIdx should be promoted when
3338  /// promoting \p Inst.
3339  static bool shouldExtOperand(const Instruction *Inst, int OpIdx) {
3340  return !(isa<SelectInst>(Inst) && OpIdx == 0);
3341  }
3342 
3343  /// Utility function to promote the operand of \p Ext when this
3344  /// operand is a promotable trunc or sext or zext.
3345  /// \p PromotedInsts maps the instructions to their type before promotion.
3346  /// \p CreatedInstsCost[out] contains the cost of all instructions
3347  /// created to promote the operand of Ext.
3348  /// Newly added extensions are inserted in \p Exts.
3349  /// Newly added truncates are inserted in \p Truncs.
3350  /// Should never be called directly.
3351  /// \return The promoted value which is used instead of Ext.
3352  static Value *promoteOperandForTruncAndAnyExt(
3353  Instruction *Ext, TypePromotionTransaction &TPT,
3354  InstrToOrigTy &PromotedInsts, unsigned &CreatedInstsCost,
3356  SmallVectorImpl<Instruction *> *Truncs, const TargetLowering &TLI);
3357 
3358  /// Utility function to promote the operand of \p Ext when this
3359  /// operand is promotable and is not a supported trunc or sext.
3360  /// \p PromotedInsts maps the instructions to their type before promotion.
3361  /// \p CreatedInstsCost[out] contains the cost of all the instructions
3362  /// created to promote the operand of Ext.
3363  /// Newly added extensions are inserted in \p Exts.
3364  /// Newly added truncates are inserted in \p Truncs.
3365  /// Should never be called directly.
3366  /// \return The promoted value which is used instead of Ext.
3367  static Value *promoteOperandForOther(Instruction *Ext,
3368  TypePromotionTransaction &TPT,
3369  InstrToOrigTy &PromotedInsts,
3370  unsigned &CreatedInstsCost,
3373  const TargetLowering &TLI, bool IsSExt);
3374 
3375  /// \see promoteOperandForOther.
3376  static Value *signExtendOperandForOther(
3377  Instruction *Ext, TypePromotionTransaction &TPT,
3378  InstrToOrigTy &PromotedInsts, unsigned &CreatedInstsCost,
3380  SmallVectorImpl<Instruction *> *Truncs, const TargetLowering &TLI) {
3381  return promoteOperandForOther(Ext, TPT, PromotedInsts, CreatedInstsCost,
3382  Exts, Truncs, TLI, true);
3383  }
3384 
3385  /// \see promoteOperandForOther.
3386  static Value *zeroExtendOperandForOther(
3387  Instruction *Ext, TypePromotionTransaction &TPT,
3388  InstrToOrigTy &PromotedInsts, unsigned &CreatedInstsCost,
3390  SmallVectorImpl<Instruction *> *Truncs, const TargetLowering &TLI) {
3391  return promoteOperandForOther(Ext, TPT, PromotedInsts, CreatedInstsCost,
3392  Exts, Truncs, TLI, false);
3393  }
3394 
3395 public:
3396  /// Type for the utility function that promotes the operand of Ext.
3397  using Action = Value *(*)(Instruction *Ext, TypePromotionTransaction &TPT,
3398  InstrToOrigTy &PromotedInsts,
3399  unsigned &CreatedInstsCost,
3402  const TargetLowering &TLI);
3403 
3404  /// Given a sign/zero extend instruction \p Ext, return the appropriate
3405  /// action to promote the operand of \p Ext instead of using Ext.
3406  /// \return NULL if no promotable action is possible with the current
3407  /// sign extension.
3408  /// \p InsertedInsts keeps track of all the instructions inserted by the
3409  /// other CodeGenPrepare optimizations. This information is important
3410  /// because we do not want to promote these instructions as CodeGenPrepare
3411  /// will reinsert them later. Thus creating an infinite loop: create/remove.
3412  /// \p PromotedInsts maps the instructions to their type before promotion.
3413  static Action getAction(Instruction *Ext, const SetOfInstrs &InsertedInsts,
3414  const TargetLowering &TLI,
3415  const InstrToOrigTy &PromotedInsts);
3416 };
3417 
3418 } // end anonymous namespace
3419 
3420 bool TypePromotionHelper::canGetThrough(const Instruction *Inst,
3421  Type *ConsideredExtType,
3422  const InstrToOrigTy &PromotedInsts,
3423  bool IsSExt) {
3424  // The promotion helper does not know how to deal with vector types yet.
3425  // To be able to fix that, we would need to fix the places where we
3426  // statically extend, e.g., constants and such.
3427  if (Inst->getType()->isVectorTy())
3428  return false;
3429 
3430  // We can always get through zext.
3431  if (isa<ZExtInst>(Inst))
3432  return true;
3433 
3434  // sext(sext) is ok too.
3435  if (IsSExt && isa<SExtInst>(Inst))
3436  return true;
3437 
3438  // We can get through binary operator, if it is legal. In other words, the
3439  // binary operator must have a nuw or nsw flag.
3440  const BinaryOperator *BinOp = dyn_cast<BinaryOperator>(Inst);
3441  if (BinOp && isa<OverflowingBinaryOperator>(BinOp) &&
3442  ((!IsSExt && BinOp->hasNoUnsignedWrap()) ||
3443  (IsSExt && BinOp->hasNoSignedWrap())))
3444  return true;
3445 
3446  // ext(and(opnd, cst)) --> and(ext(opnd), ext(cst))
3447  if ((Inst->getOpcode() == Instruction::And ||
3448  Inst->getOpcode() == Instruction::Or))
3449  return true;
3450 
3451  // ext(xor(opnd, cst)) --> xor(ext(opnd), ext(cst))
3452  if (Inst->getOpcode() == Instruction::Xor) {
3453  const ConstantInt *Cst = dyn_cast<ConstantInt>(Inst->getOperand(1));
3454  // Make sure it is not a NOT.
3455  if (Cst && !Cst->getValue().isAllOnesValue())
3456  return true;
3457  }
3458 
3459  // zext(shrl(opnd, cst)) --> shrl(zext(opnd), zext(cst))
3460  // It may change a poisoned value into a regular value, like
3461  // zext i32 (shrl i8 %val, 12) --> shrl i32 (zext i8 %val), 12
3462  // poisoned value regular value
3463  // It should be OK since undef covers valid value.
3464  if (Inst->getOpcode() == Instruction::LShr && !IsSExt)
3465  return true;
3466 
3467  // and(ext(shl(opnd, cst)), cst) --> and(shl(ext(opnd), ext(cst)), cst)
3468  // It may change a poisoned value into a regular value, like
3469  // zext i32 (shl i8 %val, 12) --> shl i32 (zext i8 %val), 12
3470  // poisoned value regular value
3471  // It should be OK since undef covers valid value.
3472  if (Inst->getOpcode() == Instruction::Shl && Inst->hasOneUse()) {
3473  const Instruction *ExtInst =
3474  dyn_cast<const Instruction>(*Inst->user_begin());
3475  if (ExtInst->hasOneUse()) {
3476  const Instruction *AndInst =
3477  dyn_cast<const Instruction>(*ExtInst->user_begin());
3478  if (AndInst && AndInst->getOpcode() == Instruction::And) {
3479  const ConstantInt *Cst = dyn_cast<ConstantInt>(AndInst->getOperand(1));
3480  if (Cst &&
3481  Cst->getValue().isIntN(Inst->getType()->getIntegerBitWidth()))
3482  return true;
3483  }
3484  }
3485  }
3486 
3487  // Check if we can do the following simplification.
3488  // ext(trunc(opnd)) --> ext(opnd)
3489  if (!isa<TruncInst>(Inst))
3490  return false;
3491 
3492  Value *OpndVal = Inst->getOperand(0);
3493  // Check if we can use this operand in the extension.
3494  // If the type is larger than the result type of the extension, we cannot.
3495  if (!OpndVal->getType()->isIntegerTy() ||
3496  OpndVal->getType()->getIntegerBitWidth() >
3497  ConsideredExtType->getIntegerBitWidth())
3498  return false;
3499 
3500  // If the operand of the truncate is not an instruction, we will not have
3501  // any information on the dropped bits.
3502  // (Actually we could for constant but it is not worth the extra logic).
3503  Instruction *Opnd = dyn_cast<Instruction>(OpndVal);
3504  if (!Opnd)
3505  return false;
3506 
3507  // Check if the source of the type is narrow enough.
3508  // I.e., check that trunc just drops extended bits of the same kind of
3509  // the extension.
3510  // #1 get the type of the operand and check the kind of the extended bits.
3511  const Type *OpndType = getOrigType(PromotedInsts, Opnd, IsSExt);
3512  if (OpndType)
3513  ;
3514  else if ((IsSExt && isa<SExtInst>(Opnd)) || (!IsSExt && isa<ZExtInst>(Opnd)))
3515  OpndType = Opnd->getOperand(0)->getType();
3516  else
3517  return false;
3518 
3519  // #2 check that the truncate just drops extended bits.
3520  return Inst->getType()->getIntegerBitWidth() >=
3521  OpndType->getIntegerBitWidth();
3522 }
3523 
3524 TypePromotionHelper::Action TypePromotionHelper::getAction(
3525  Instruction *Ext, const SetOfInstrs &InsertedInsts,
3526  const TargetLowering &TLI, const InstrToOrigTy &PromotedInsts) {
3527  assert((isa<SExtInst>(Ext) || isa<ZExtInst>(Ext)) &&
3528  "Unexpected instruction type");
3529  Instruction *ExtOpnd = dyn_cast<Instruction>(Ext->getOperand(0));
3530  Type *ExtTy = Ext->getType();
3531  bool IsSExt = isa<SExtInst>(Ext);
3532  // If the operand of the extension is not an instruction, we cannot
3533  // get through.
3534  // If it, check we can get through.
3535  if (!ExtOpnd || !canGetThrough(ExtOpnd, ExtTy, PromotedInsts, IsSExt))
3536  return nullptr;
3537 
3538  // Do not promote if the operand has been added by codegenprepare.
3539  // Otherwise, it means we are undoing an optimization that is likely to be
3540  // redone, thus causing potential infinite loop.
3541  if (isa<TruncInst>(ExtOpnd) && InsertedInsts.count(ExtOpnd))
3542  return nullptr;
3543 
3544  // SExt or Trunc instructions.
3545  // Return the related handler.
3546  if (isa<SExtInst>(ExtOpnd) || isa<TruncInst>(ExtOpnd) ||
3547  isa<ZExtInst>(ExtOpnd))
3548  return promoteOperandForTruncAndAnyExt;
3549 
3550  // Regular instruction.
3551  // Abort early if we will have to insert non-free instructions.
3552  if (!ExtOpnd->hasOneUse() && !TLI.isTruncateFree(ExtTy, ExtOpnd->getType()))
3553  return nullptr;
3554  return IsSExt ? signExtendOperandForOther : zeroExtendOperandForOther;
3555 }
3556 
3557 Value *TypePromotionHelper::promoteOperandForTruncAndAnyExt(
3558  Instruction *SExt, TypePromotionTransaction &TPT,
3559  InstrToOrigTy &PromotedInsts, unsigned &CreatedInstsCost,
3561  SmallVectorImpl<Instruction *> *Truncs, const TargetLowering &TLI) {
3562  // By construction, the operand of SExt is an instruction. Otherwise we cannot
3563  // get through it and this method should not be called.
3564  Instruction *SExtOpnd = cast<Instruction>(SExt->getOperand(0));
3565  Value *ExtVal = SExt;
3566  bool HasMergedNonFreeExt = false;
3567  if (isa<ZExtInst>(SExtOpnd)) {
3568  // Replace s|zext(zext(opnd))
3569  // => zext(opnd).
3570  HasMergedNonFreeExt = !TLI.isExtFree(SExtOpnd);
3571  Value *ZExt =
3572  TPT.createZExt(SExt, SExtOpnd->getOperand(0), SExt->getType());
3573  TPT.replaceAllUsesWith(SExt, ZExt);
3574  TPT.eraseInstruction(SExt);
3575  ExtVal = ZExt;
3576  } else {
3577  // Replace z|sext(trunc(opnd)) or sext(sext(opnd))
3578  // => z|sext(opnd).
3579  TPT.setOperand(SExt, 0, SExtOpnd->getOperand(0));
3580  }
3581  CreatedInstsCost = 0;
3582 
3583  // Remove dead code.
3584  if (SExtOpnd->use_empty())
3585  TPT.eraseInstruction(SExtOpnd);
3586 
3587  // Check if the extension is still needed.
3588  Instruction *ExtInst = dyn_cast<Instruction>(ExtVal);
3589  if (!ExtInst || ExtInst->getType() != ExtInst->getOperand(0)->getType()) {
3590  if (ExtInst) {
3591  if (Exts)
3592  Exts->push_back(ExtInst);
3593  CreatedInstsCost = !TLI.isExtFree(ExtInst) && !HasMergedNonFreeExt;
3594  }
3595  return ExtVal;
3596  }
3597 
3598  // At this point we have: ext ty opnd to ty.
3599  // Reassign the uses of ExtInst to the opnd and remove ExtInst.
3600  Value *NextVal = ExtInst->getOperand(0);
3601  TPT.eraseInstruction(ExtInst, NextVal);
3602  return NextVal;
3603 }
3604 
3605 Value *TypePromotionHelper::promoteOperandForOther(
3606  Instruction *Ext, TypePromotionTransaction &TPT,
3607  InstrToOrigTy &PromotedInsts, unsigned &CreatedInstsCost,
3609  SmallVectorImpl<Instruction *> *Truncs, const TargetLowering &TLI,
3610  bool IsSExt) {
3611  // By construction, the operand of Ext is an instruction. Otherwise we cannot
3612  // get through it and this method should not be called.
3613  Instruction *ExtOpnd = cast<Instruction>(Ext->getOperand(0));
3614  CreatedInstsCost = 0;
3615  if (!ExtOpnd->hasOneUse()) {
3616  // ExtOpnd will be promoted.
3617  // All its uses, but Ext, will need to use a truncated value of the
3618  // promoted version.
3619  // Create the truncate now.
3620  Value *Trunc = TPT.createTrunc(Ext, ExtOpnd->getType());
3621  if (Instruction *ITrunc = dyn_cast<Instruction>(Trunc)) {
3622  // Insert it just after the definition.
3623  ITrunc->moveAfter(ExtOpnd);
3624  if (Truncs)
3625  Truncs->push_back(ITrunc);
3626  }
3627 
3628  TPT.replaceAllUsesWith(ExtOpnd, Trunc);
3629  // Restore the operand of Ext (which has been replaced by the previous call
3630  // to replaceAllUsesWith) to avoid creating a cycle trunc <-> sext.
3631  TPT.setOperand(Ext, 0, ExtOpnd);
3632  }
3633 
3634  // Get through the Instruction:
3635  // 1. Update its type.
3636  // 2. Replace the uses of Ext by Inst.
3637  // 3. Extend each operand that needs to be extended.
3638 
3639  // Remember the original type of the instruction before promotion.
3640  // This is useful to know that the high bits are sign extended bits.
3641  addPromotedInst(PromotedInsts, ExtOpnd, IsSExt);
3642  // Step #1.
3643  TPT.mutateType(ExtOpnd, Ext->getType());
3644  // Step #2.
3645  TPT.replaceAllUsesWith(Ext, ExtOpnd);
3646  // Step #3.
3647  Instruction *ExtForOpnd = Ext;
3648 
3649  LLVM_DEBUG(dbgs() << "Propagate Ext to operands\n");
3650  for (int OpIdx = 0, EndOpIdx = ExtOpnd->getNumOperands(); OpIdx != EndOpIdx;
3651  ++OpIdx) {
3652  LLVM_DEBUG(dbgs() << "Operand:\n" << *(ExtOpnd->getOperand(OpIdx)) << '\n');
3653  if (ExtOpnd->getOperand(OpIdx)->getType() == Ext->getType() ||
3654  !shouldExtOperand(ExtOpnd, OpIdx)) {
3655  LLVM_DEBUG(dbgs() << "No need to propagate\n");
3656  continue;
3657  }
3658  // Check if we can statically extend the operand.
3659  Value *Opnd = ExtOpnd->getOperand(OpIdx);
3660  if (const ConstantInt *Cst = dyn_cast<ConstantInt>(Opnd)) {
3661  LLVM_DEBUG(dbgs() << "Statically extend\n");
3662  unsigned BitWidth = Ext->getType()->getIntegerBitWidth();
3663  APInt CstVal = IsSExt ? Cst->getValue().sext(BitWidth)
3664  : Cst->getValue().zext(BitWidth);
3665  TPT.setOperand(ExtOpnd, OpIdx, ConstantInt::get(Ext->getType(), CstVal));
3666  continue;
3667  }
3668  // UndefValue are typed, so we have to statically sign extend them.
3669  if (isa<UndefValue>(Opnd)) {
3670  LLVM_DEBUG(dbgs() << "Statically extend\n");
3671  TPT.setOperand(ExtOpnd, OpIdx, UndefValue::get(Ext->getType()));
3672  continue;
3673  }
3674 
3675  // Otherwise we have to explicitly sign extend the operand.
3676  // Check if Ext was reused to extend an operand.
3677  if (!ExtForOpnd) {
3678  // If yes, create a new one.
3679  LLVM_DEBUG(dbgs() << "More operands to ext\n");
3680  Value *ValForExtOpnd = IsSExt ? TPT.createSExt(Ext, Opnd, Ext->getType())
3681  : TPT.createZExt(Ext, Opnd, Ext->getType());
3682  if (!isa<Instruction>(ValForExtOpnd)) {
3683  TPT.setOperand(ExtOpnd, OpIdx, ValForExtOpnd);
3684  continue;
3685  }
3686  ExtForOpnd = cast<Instruction>(ValForExtOpnd);
3687  }
3688  if (Exts)
3689  Exts->push_back(ExtForOpnd);
3690  TPT.setOperand(ExtForOpnd, 0, Opnd);
3691 
3692  // Move the sign extension before the insertion point.
3693  TPT.moveBefore(ExtForOpnd, ExtOpnd);
3694  TPT.setOperand(ExtOpnd, OpIdx, ExtForOpnd);
3695  CreatedInstsCost += !TLI.isExtFree(ExtForOpnd);
3696  // If more sext are required, new instructions will have to be created.
3697  ExtForOpnd = nullptr;
3698  }
3699  if (ExtForOpnd == Ext) {
3700  LLVM_DEBUG(dbgs() << "Extension is useless now\n");
3701  TPT.eraseInstruction(Ext);
3702  }
3703  return ExtOpnd;
3704 }
3705 
3706 /// Check whether or not promoting an instruction to a wider type is profitable.
3707 /// \p NewCost gives the cost of extension instructions created by the
3708 /// promotion.
3709 /// \p OldCost gives the cost of extension instructions before the promotion
3710 /// plus the number of instructions that have been
3711 /// matched in the addressing mode the promotion.
3712 /// \p PromotedOperand is the value that has been promoted.
3713 /// \return True if the promotion is profitable, false otherwise.
3714 bool AddressingModeMatcher::isPromotionProfitable(
3715  unsigned NewCost, unsigned OldCost, Value *PromotedOperand) const {
3716  LLVM_DEBUG(dbgs() << "OldCost: " << OldCost << "\tNewCost: " << NewCost
3717  << '\n');
3718  // The cost of the new extensions is greater than the cost of the
3719  // old extension plus what we folded.
3720  // This is not profitable.
3721  if (NewCost > OldCost)
3722  return false;
3723  if (NewCost < OldCost)
3724  return true;
3725  // The promotion is neutral but it may help folding the sign extension in
3726  // loads for instance.
3727  // Check that we did not create an illegal instruction.
3728  return isPromotedInstructionLegal(TLI, DL, PromotedOperand);
3729 }
3730 
3731 /// Given an instruction or constant expr, see if we can fold the operation
3732 /// into the addressing mode. If so, update the addressing mode and return
3733 /// true, otherwise return false without modifying AddrMode.
3734 /// If \p MovedAway is not NULL, it contains the information of whether or
3735 /// not AddrInst has to be folded into the addressing mode on success.
3736 /// If \p MovedAway == true, \p AddrInst will not be part of the addressing
3737 /// because it has been moved away.
3738 /// Thus AddrInst must not be added in the matched instructions.
3739 /// This state can happen when AddrInst is a sext, since it may be moved away.
3740 /// Therefore, AddrInst may not be valid when MovedAway is true and it must
3741 /// not be referenced anymore.
3742 bool AddressingModeMatcher::matchOperationAddr(User *AddrInst, unsigned Opcode,
3743  unsigned Depth,
3744  bool *MovedAway) {
3745  // Avoid exponential behavior on extremely deep expression trees.
3746  if (Depth >= 5) return false;
3747 
3748  // By default, all matched instructions stay in place.
3749  if (MovedAway)
3750  *MovedAway = false;
3751 
3752  switch (Opcode) {
3753  case Instruction::PtrToInt:
3754  // PtrToInt is always a noop, as we know that the int type is pointer sized.
3755  return matchAddr(AddrInst->getOperand(0), Depth);
3756  case Instruction::IntToPtr: {
3757  auto AS = AddrInst->getType()->getPointerAddressSpace();
3758  auto PtrTy = MVT::getIntegerVT(DL.getPointerSizeInBits(AS));
3759  // This inttoptr is a no-op if the integer type is pointer sized.
3760  if (TLI.getValueType(DL, AddrInst->getOperand(0)->getType()) == PtrTy)
3761  return matchAddr(AddrInst->getOperand(0), Depth);
3762  return false;
3763  }
3764  case Instruction::BitCast:
3765  // BitCast is always a noop, and we can handle it as long as it is
3766  // int->int or pointer->pointer (we don't want int<->fp or something).
3767  if (AddrInst->getOperand(0)->getType()->isIntOrPtrTy() &&
3768  // Don't touch identity bitcasts. These were probably put here by LSR,
3769  // and we don't want to mess around with them. Assume it knows what it
3770  // is doing.
3771  AddrInst->getOperand(0)->getType() != AddrInst->getType())
3772  return matchAddr(AddrInst->getOperand(0), Depth);
3773  return false;
3774  case Instruction::AddrSpaceCast: {
3775  unsigned SrcAS
3776  = AddrInst->getOperand(0)->getType()->getPointerAddressSpace();
3777  unsigned DestAS = AddrInst->getType()->getPointerAddressSpace();
3778  if (TLI.isNoopAddrSpaceCast(SrcAS, DestAS))
3779  return matchAddr(AddrInst->getOperand(0), Depth);
3780  return false;
3781  }
3782  case Instruction::Add: {
3783  // Check to see if we can merge in the RHS then the LHS. If so, we win.
3784  ExtAddrMode BackupAddrMode = AddrMode;
3785  unsigned OldSize = AddrModeInsts.size();
3786  // Start a transaction at this point.
3787  // The LHS may match but not the RHS.
3788  // Therefore, we need a higher level restoration point to undo partially
3789  // matched operation.
3790  TypePromotionTransaction::ConstRestorationPt LastKnownGood =
3791  TPT.getRestorationPoint();
3792 
3793  if (matchAddr(AddrInst->getOperand(1), Depth+1) &&
3794  matchAddr(AddrInst->getOperand(0), Depth+1))
3795  return true;
3796 
3797  // Restore the old addr mode info.
3798  AddrMode = BackupAddrMode;
3799  AddrModeInsts.resize(OldSize);
3800  TPT.rollback(LastKnownGood);
3801 
3802  // Otherwise this was over-aggressive. Try merging in the LHS then the RHS.
3803  if (matchAddr(AddrInst->getOperand(0), Depth+1) &&
3804  matchAddr(AddrInst->getOperand(1), Depth+1))
3805  return true;
3806 
3807  // Otherwise we definitely can't merge the ADD in.
3808  AddrMode = BackupAddrMode;
3809  AddrModeInsts.resize(OldSize);
3810  TPT.rollback(LastKnownGood);
3811  break;
3812  }
3813  //case Instruction::Or:
3814  // TODO: We can handle "Or Val, Imm" iff this OR is equivalent to an ADD.
3815  //break;
3816  case Instruction::Mul:
3817  case Instruction::Shl: {
3818  // Can only handle X*C and X << C.
3819  ConstantInt *RHS = dyn_cast<ConstantInt>(AddrInst->getOperand(1));
3820  if (!RHS || RHS->getBitWidth() > 64)
3821  return false;
3822  int64_t Scale = RHS->getSExtValue();
3823  if (Opcode == Instruction::Shl)
3824  Scale = 1LL << Scale;
3825 
3826  return matchScaledValue(AddrInst->getOperand(0), Scale, Depth);
3827  }
3828  case Instruction::GetElementPtr: {
3829  // Scan the GEP. We check it if it contains constant offsets and at most
3830  // one variable offset.
3831  int VariableOperand = -1;
3832  unsigned VariableScale = 0;
3833 
3834  int64_t ConstantOffset = 0;
3835  gep_type_iterator GTI = gep_type_begin(AddrInst);
3836  for (unsigned i = 1, e = AddrInst->getNumOperands(); i != e; ++i, ++GTI) {
3837  if (StructType *STy = GTI.getStructTypeOrNull()) {
3838  const StructLayout *SL = DL.getStructLayout(STy);
3839  unsigned Idx =
3840  cast<ConstantInt>(AddrInst->getOperand(i))->getZExtValue();
3841  ConstantOffset += SL->getElementOffset(Idx);
3842  } else {
3843  uint64_t TypeSize = DL.getTypeAllocSize(GTI.getIndexedType());
3844  if (ConstantInt *CI = dyn_cast<ConstantInt>(AddrInst->getOperand(i))) {
3845  const APInt &CVal = CI->getValue();
3846  if (CVal.getMinSignedBits() <= 64) {
3847  ConstantOffset += CVal.getSExtValue() * TypeSize;
3848  continue;
3849  }
3850  }
3851  if (TypeSize) { // Scales of zero don't do anything.
3852  // We only allow one variable index at the moment.
3853  if (VariableOperand != -1)
3854  return false;
3855 
3856  // Remember the variable index.
3857  VariableOperand = i;
3858  VariableScale = TypeSize;
3859  }
3860  }
3861  }
3862 
3863  // A common case is for the GEP to only do a constant offset. In this case,
3864  // just add it to the disp field and check validity.
3865  if (VariableOperand == -1) {
3866  AddrMode.BaseOffs += ConstantOffset;
3867  if (ConstantOffset == 0 ||
3868  TLI.isLegalAddressingMode(DL, AddrMode, AccessTy, AddrSpace)) {
3869  // Check to see if we can fold the base pointer in too.
3870  if (matchAddr(AddrInst->getOperand(0), Depth+1))
3871  return true;
3872  } else if (EnableGEPOffsetSplit && isa<GetElementPtrInst>(AddrInst) &&
3873  TLI.shouldConsiderGEPOffsetSplit() && Depth == 0 &&
3874  ConstantOffset > 0) {
3875  // Record GEPs with non-zero offsets as candidates for splitting in the
3876  // event that the offset cannot fit into the r+i addressing mode.
3877  // Simple and common case that only one GEP is used in calculating the
3878  // address for the memory access.
3879  Value *Base = AddrInst->getOperand(0);
3880  auto *BaseI = dyn_cast<Instruction>(Base);
3881  auto *GEP = cast<GetElementPtrInst>(AddrInst);
3882  if (isa<Argument>(Base) || isa<GlobalValue>(Base) ||
3883  (BaseI && !isa<CastInst>(BaseI) &&
3884  !isa<GetElementPtrInst>(BaseI))) {
3885  // If the base is an instruction, make sure the GEP is not in the same
3886  // basic block as the base. If the base is an argument or global
3887  // value, make sure the GEP is not in the entry block. Otherwise,
3888  // instruction selection can undo the split. Also make sure the
3889  // parent block allows inserting non-PHI instructions before the
3890  // terminator.
3891  BasicBlock *Parent =
3892  BaseI ? BaseI->getParent() : &GEP->getFunction()->getEntryBlock();
3893  if (GEP->getParent() != Parent && !Parent->getTerminator()->isEHPad())
3894  LargeOffsetGEP = std::make_pair(GEP, ConstantOffset);
3895  }
3896  }
3897  AddrMode.BaseOffs -= ConstantOffset;
3898  return false;
3899  }
3900 
3901  // Save the valid addressing mode in case we can't match.
3902  ExtAddrMode BackupAddrMode = AddrMode;
3903  unsigned OldSize = AddrModeInsts.size();
3904 
3905  // See if the scale and offset amount is valid for this target.
3906  AddrMode.BaseOffs += ConstantOffset;
3907 
3908  // Match the base operand of the GEP.
3909  if (!matchAddr(AddrInst->getOperand(0), Depth+1)) {
3910  // If it couldn't be matched, just stuff the value in a register.
3911  if (AddrMode.HasBaseReg) {
3912  AddrMode = BackupAddrMode;
3913  AddrModeInsts.resize(OldSize);
3914  return false;
3915  }
3916  AddrMode.HasBaseReg = true;
3917  AddrMode.BaseReg = AddrInst->getOperand(0);
3918  }
3919 
3920  // Match the remaining variable portion of the GEP.
3921  if (!matchScaledValue(AddrInst->getOperand(VariableOperand), VariableScale,
3922  Depth)) {
3923  // If it couldn't be matched, try stuffing the base into a register
3924  // instead of matching it, and retrying the match of the scale.
3925  AddrMode = BackupAddrMode;
3926  AddrModeInsts.resize(OldSize);
3927  if (AddrMode.HasBaseReg)
3928  return false;
3929  AddrMode.HasBaseReg = true;
3930  AddrMode.BaseReg = AddrInst->getOperand(0);
3931  AddrMode.BaseOffs += ConstantOffset;
3932  if (!matchScaledValue(AddrInst->getOperand(VariableOperand),
3933  VariableScale, Depth)) {
3934  // If even that didn't work, bail.
3935  AddrMode = BackupAddrMode;
3936  AddrModeInsts.resize(OldSize);
3937  return false;
3938  }
3939  }
3940 
3941  return true;
3942  }
3943  case Instruction::SExt:
3944  case Instruction::ZExt: {
3945  Instruction *Ext = dyn_cast<Instruction>(AddrInst);
3946  if (!Ext)
3947  return false;
3948 
3949  // Try to move this ext out of the way of the addressing mode.
3950  // Ask for a method for doing so.
3951  TypePromotionHelper::Action TPH =
3952  TypePromotionHelper::getAction(Ext, InsertedInsts, TLI, PromotedInsts);
3953  if (!TPH)
3954  return false;
3955 
3956  TypePromotionTransaction::ConstRestorationPt LastKnownGood =
3957  TPT.getRestorationPoint();
3958  unsigned CreatedInstsCost = 0;
3959  unsigned ExtCost = !TLI.isExtFree(Ext);
3960  Value *PromotedOperand =
3961  TPH(Ext, TPT, PromotedInsts, CreatedInstsCost, nullptr, nullptr, TLI);
3962  // SExt has been moved away.
3963  // Thus either it will be rematched later in the recursive calls or it is
3964  // gone. Anyway, we must not fold it into the addressing mode at this point.
3965  // E.g.,
3966  // op = add opnd, 1
3967  // idx = ext op
3968  // addr = gep base, idx
3969  // is now:
3970  // promotedOpnd = ext opnd <- no match here
3971  // op = promoted_add promotedOpnd, 1 <- match (later in recursive calls)
3972  // addr = gep base, op <- match
3973  if (MovedAway)
3974  *MovedAway = true;
3975 
3976  assert(PromotedOperand &&
3977  "TypePromotionHelper should have filtered out those cases");
3978 
3979  ExtAddrMode BackupAddrMode = AddrMode;
3980  unsigned OldSize = AddrModeInsts.size();
3981 
3982  if (!matchAddr(PromotedOperand, Depth) ||
3983  // The total of the new cost is equal to the cost of the created
3984  // instructions.
3985  // The total of the old cost is equal to the cost of the extension plus
3986  // what we have saved in the addressing mode.
3987  !isPromotionProfitable(CreatedInstsCost,
3988  ExtCost + (AddrModeInsts.size() - OldSize),
3989  PromotedOperand)) {
3990  AddrMode = BackupAddrMode;
3991  AddrModeInsts.resize(OldSize);
3992  LLVM_DEBUG(dbgs() << "Sign extension does not pay off: rollback\n");
3993  TPT.rollback(LastKnownGood);
3994  return false;
3995  }
3996  return true;
3997  }
3998  }
3999  return false;
4000 }
4001 
4002 /// If we can, try to add the value of 'Addr' into the current addressing mode.
4003 /// If Addr can't be added to AddrMode this returns false and leaves AddrMode
4004 /// unmodified. This assumes that Addr is either a pointer type or intptr_t
4005 /// for the target.
4006 ///
4007 bool AddressingModeMatcher::matchAddr(Value *Addr, unsigned Depth) {
4008  // Start a transaction at this point that we will rollback if the matching
4009  // fails.
4010  TypePromotionTransaction::ConstRestorationPt LastKnownGood =
4011  TPT.getRestorationPoint();
4012  if (ConstantInt *CI = dyn_cast<ConstantInt>(Addr)) {
4013  // Fold in immediates if legal for the target.
4014  AddrMode.BaseOffs += CI->getSExtValue();
4015  if (TLI.isLegalAddressingMode(DL, AddrMode, AccessTy, AddrSpace))
4016  return true;
4017  AddrMode.BaseOffs -= CI->getSExtValue();
4018  } else if (GlobalValue *GV = dyn_cast<GlobalValue>(Addr)) {
4019  // If this is a global variable, try to fold it into the addressing mode.
4020  if (!AddrMode.BaseGV) {
4021  AddrMode.BaseGV = GV;
4022  if (TLI.isLegalAddressingMode(DL, AddrMode, AccessTy, AddrSpace))
4023  return true;
4024  AddrMode.BaseGV = nullptr;
4025  }
4026  } else if (Instruction *I = dyn_cast<Instruction>(Addr)) {
4027  ExtAddrMode BackupAddrMode = AddrMode;
4028  unsigned OldSize = AddrModeInsts.size();
4029 
4030  // Check to see if it is possible to fold this operation.
4031  bool MovedAway = false;
4032  if (matchOperationAddr(I, I->getOpcode(), Depth, &MovedAway)) {
4033  // This instruction may have been moved away. If so, there is nothing
4034  // to check here.
4035  if (MovedAway)
4036  return true;
4037  // Okay, it's possible to fold this. Check to see if it is actually
4038  // *profitable* to do so. We use a simple cost model to avoid increasing
4039  // register pressure too much.
4040  if (I->hasOneUse() ||
4041  isProfitableToFoldIntoAddressingMode(I, BackupAddrMode, AddrMode)) {
4042  AddrModeInsts.push_back(I);
4043  return true;
4044  }
4045 
4046  // It isn't profitable to do this, roll back.
4047  //cerr << "NOT FOLDING: " << *I;
4048  AddrMode = BackupAddrMode;
4049  AddrModeInsts.resize(OldSize);
4050  TPT.rollback(LastKnownGood);
4051  }
4052  } else if (ConstantExpr *CE = dyn_cast<ConstantExpr>(Addr)) {
4053  if (matchOperationAddr(CE, CE->getOpcode(), Depth))
4054  return true;
4055  TPT.rollback(LastKnownGood);
4056  } else if (isa<ConstantPointerNull>(Addr)) {
4057  // Null pointer gets folded without affecting the addressing mode.
4058  return true;
4059  }
4060 
4061  // Worse case, the target should support [reg] addressing modes. :)
4062  if (!AddrMode.HasBaseReg) {
4063  AddrMode.HasBaseReg = true;
4064  AddrMode.BaseReg = Addr;
4065  // Still check for legality in case the target supports [imm] but not [i+r].
4066  if (TLI.isLegalAddressingMode(DL, AddrMode, AccessTy, AddrSpace))
4067  return true;
4068  AddrMode.HasBaseReg = false;
4069  AddrMode.BaseReg = nullptr;
4070  }
4071 
4072  // If the base register is already taken, see if we can do [r+r].
4073  if (AddrMode.Scale == 0) {
4074  AddrMode.Scale = 1;
4075  AddrMode.ScaledReg = Addr;
4076  if (TLI.isLegalAddressingMode(DL, AddrMode, AccessTy, AddrSpace))
4077  return true;
4078  AddrMode.Scale = 0;
4079  AddrMode.ScaledReg = nullptr;
4080  }
4081  // Couldn't match.
4082  TPT.rollback(LastKnownGood);
4083  return false;
4084 }
4085 
4086 /// Check to see if all uses of OpVal by the specified inline asm call are due
4087 /// to memory operands. If so, return true, otherwise return false.
4088 static bool IsOperandAMemoryOperand(CallInst *CI, InlineAsm *IA, Value *OpVal,
4089  const TargetLowering &TLI,
4090  const TargetRegisterInfo &TRI) {
4091  const Function *F = CI->getFunction();
4092  TargetLowering::AsmOperandInfoVector TargetConstraints =
4094  ImmutableCallSite(CI));
4095 
4096  for (unsigned i = 0, e = TargetConstraints.size(); i != e; ++i) {
4097  TargetLowering::AsmOperandInfo &OpInfo = TargetConstraints[i];
4098 
4099  // Compute the constraint code and ConstraintType to use.
4100  TLI.ComputeConstraintToUse(OpInfo, SDValue());
4101 
4102  // If this asm operand is our Value*, and if it isn't an indirect memory
4103  // operand, we can't fold it!
4104  if (OpInfo.CallOperandVal == OpVal &&
4106  !OpInfo.isIndirect))
4107  return false;
4108  }
4109 
4110  return true;
4111 }
4112 
4113 // Max number of memory uses to look at before aborting the search to conserve
4114 // compile time.
4115 static constexpr int MaxMemoryUsesToScan = 20;
4116 
4117 /// Recursively walk all the uses of I until we find a memory use.
4118 /// If we find an obviously non-foldable instruction, return true.
4119 /// Add the ultimately found memory instructions to MemoryUses.
4120 static bool FindAllMemoryUses(
4121  Instruction *I,
4122  SmallVectorImpl<std::pair<Instruction *, unsigned>> &MemoryUses,
4123  SmallPtrSetImpl<Instruction *> &ConsideredInsts, const TargetLowering &TLI,
4124  const TargetRegisterInfo &TRI, int SeenInsts = 0) {
4125  // If we already considered this instruction, we're done.
4126  if (!ConsideredInsts.insert(I).second)
4127  return false;
4128 
4129  // If this is an obviously unfoldable instruction, bail out.
4130  if (!MightBeFoldableInst(I))
4131  return true;
4132 
4133  const bool OptSize = I->getFunction()->optForSize();
4134 
4135  // Loop over all the uses, recursively processing them.
4136  for (Use &U : I->uses()) {
4137  // Conservatively return true if we're seeing a large number or a deep chain
4138  // of users. This avoids excessive compilation times in pathological cases.
4139  if (SeenInsts++ >= MaxMemoryUsesToScan)
4140  return true;
4141 
4142  Instruction *UserI = cast<Instruction>(U.getUser());
4143  if (LoadInst *LI = dyn_cast<LoadInst>(UserI)) {
4144  MemoryUses.push_back(std::make_pair(LI, U.getOperandNo()));
4145  continue;
4146  }
4147 
4148  if (StoreInst *SI = dyn_cast<StoreInst>(UserI)) {
4149  unsigned opNo = U.getOperandNo();
4150  if (opNo != StoreInst::getPointerOperandIndex())
4151  return true; // Storing addr, not into addr.
4152  MemoryUses.push_back(std::make_pair(SI, opNo));
4153  continue;
4154  }
4155 
4156  if (AtomicRMWInst *RMW = dyn_cast<AtomicRMWInst>(UserI)) {
4157  unsigned opNo = U.getOperandNo();
4159  return true; // Storing addr, not into addr.
4160  MemoryUses.push_back(std::make_pair(RMW, opNo));
4161  continue;
4162  }
4163 
4164  if (AtomicCmpXchgInst *CmpX = dyn_cast<AtomicCmpXchgInst>(UserI)) {
4165  unsigned opNo = U.getOperandNo();
4167  return true; // Storing addr, not into addr.
4168  MemoryUses.push_back(std::make_pair(CmpX, opNo));
4169  continue;
4170  }
4171 
4172  if (CallInst *CI = dyn_cast<CallInst>(UserI)) {
4173  // If this is a cold call, we can sink the addressing calculation into
4174  // the cold path. See optimizeCallInst
4175  if (!OptSize && CI->hasFnAttr(Attribute::Cold))
4176  continue;
4177 
4179  if (!IA) return true;
4180 
4181  // If this is a memory operand, we're cool, otherwise bail out.
4182  if (!IsOperandAMemoryOperand(CI, IA, I, TLI, TRI))
4183  return true;
4184  continue;
4185  }
4186 
4187  if (FindAllMemoryUses(UserI, MemoryUses, ConsideredInsts, TLI, TRI,
4188  SeenInsts))
4189  return true;
4190  }
4191 
4192  return false;
4193 }
4194 
4195 /// Return true if Val is already known to be live at the use site that we're
4196 /// folding it into. If so, there is no cost to include it in the addressing
4197 /// mode. KnownLive1 and KnownLive2 are two values that we know are live at the
4198 /// instruction already.
4199 bool AddressingModeMatcher::valueAlreadyLiveAtInst(Value *Val,Value *KnownLive1,
4200  Value *KnownLive2) {
4201  // If Val is either of the known-live values, we know it is live!
4202  if (Val == nullptr || Val == KnownLive1 || Val == KnownLive2)
4203  return true;
4204 
4205  // All values other than instructions and arguments (e.g. constants) are live.
4206  if (!isa<Instruction>(Val) && !isa<Argument>(Val)) return true;
4207 
4208  // If Val is a constant sized alloca in the entry block, it is live, this is
4209  // true because it is just a reference to the stack/frame pointer, which is
4210  // live for the whole function.
4211  if (AllocaInst *AI = dyn_cast<AllocaInst>(Val))
4212  if (AI->isStaticAlloca())
4213  return true;
4214 
4215  // Check to see if this value is already used in the memory instruction's
4216  // block. If so, it's already live into the block at the very least, so we
4217  // can reasonably fold it.
4218  return Val->isUsedInBasicBlock(MemoryInst->getParent());
4219 }
4220 
4221 /// It is possible for the addressing mode of the machine to fold the specified
4222 /// instruction into a load or store that ultimately uses it.
4223 /// However, the specified instruction has multiple uses.
4224 /// Given this, it may actually increase register pressure to fold it
4225 /// into the load. For example, consider this code:
4226 ///
4227 /// X = ...
4228 /// Y = X+1
4229 /// use(Y) -> nonload/store
4230 /// Z = Y+1
4231 /// load Z
4232 ///
4233 /// In this case, Y has multiple uses, and can be folded into the load of Z
4234 /// (yielding load [X+2]). However, doing this will cause both "X" and "X+1" to
4235 /// be live at the use(Y) line. If we don't fold Y into load Z, we use one
4236 /// fewer register. Since Y can't be folded into "use(Y)" we don't increase the
4237 /// number of computations either.
4238 ///
4239 /// Note that this (like most of CodeGenPrepare) is just a rough heuristic. If
4240 /// X was live across 'load Z' for other reasons, we actually *would* want to
4241 /// fold the addressing mode in the Z case. This would make Y die earlier.
4242 bool AddressingModeMatcher::
4243 isProfitableToFoldIntoAddressingMode(Instruction *I, ExtAddrMode &AMBefore,
4244  ExtAddrMode &AMAfter) {
4245  if (IgnoreProfitability) return true;
4246 
4247  // AMBefore is the addressing mode before this instruction was folded into it,
4248  // and AMAfter is the addressing mode after the instruction was folded. Get
4249  // the set of registers referenced by AMAfter and subtract out those
4250  // referenced by AMBefore: this is the set of values which folding in this
4251  // address extends the lifetime of.
4252  //
4253  // Note that there are only two potential values being referenced here,
4254  // BaseReg and ScaleReg (global addresses are always available, as are any
4255  // folded immediates).
4256  Value *BaseReg = AMAfter.BaseReg, *ScaledReg = AMAfter.ScaledReg;
4257 
4258  // If the BaseReg or ScaledReg was referenced by the previous addrmode, their
4259  // lifetime wasn't extended by adding this instruction.
4260  if (valueAlreadyLiveAtInst(BaseReg, AMBefore.BaseReg, AMBefore.ScaledReg))
4261  BaseReg = nullptr;
4262  if (valueAlreadyLiveAtInst(ScaledReg, AMBefore.BaseReg, AMBefore.ScaledReg))
4263  ScaledReg = nullptr;
4264 
4265  // If folding this instruction (and it's subexprs) didn't extend any live
4266  // ranges, we're ok with it.
4267  if (!BaseReg && !ScaledReg)
4268  return true;
4269 
4270  // If all uses of this instruction can have the address mode sunk into them,
4271  // we can remove the addressing mode and effectively trade one live register
4272  // for another (at worst.) In this context, folding an addressing mode into
4273  // the use is just a particularly nice way of sinking it.
4275  SmallPtrSet<Instruction*, 16> ConsideredInsts;
4276  if (FindAllMemoryUses(I, MemoryUses, ConsideredInsts, TLI, TRI))
4277  return false; // Has a non-memory, non-foldable use!
4278 
4279  // Now that we know that all uses of this instruction are part of a chain of
4280  // computation involving only operations that could theoretically be folded
4281  // into a memory use, loop over each of these memory operation uses and see
4282  // if they could *actually* fold the instruction. The assumption is that
4283  // addressing modes are cheap and that duplicating the computation involved
4284  // many times is worthwhile, even on a fastpath. For sinking candidates
4285  // (i.e. cold call sites), this serves as a way to prevent excessive code
4286  // growth since most architectures have some reasonable small and fast way to
4287  // compute an effective address. (i.e LEA on x86)
4288  SmallVector<Instruction*, 32> MatchedAddrModeInsts;
4289  for (unsigned i = 0, e = MemoryUses.size(); i != e; ++i) {
4290  Instruction *User = MemoryUses[i].first;
4291  unsigned OpNo = MemoryUses[i].second;
4292 
4293  // Get the access type of this use. If the use isn't a pointer, we don't
4294  // know what it accesses.
4295  Value *Address = User->getOperand(OpNo);
4296  PointerType *AddrTy = dyn_cast<PointerType>(Address->getType());
4297  if (!AddrTy)
4298  return false;
4299  Type *AddressAccessTy = AddrTy->getElementType();
4300  unsigned AS = AddrTy->getAddressSpace();
4301 
4302  // Do a match against the root of this address, ignoring profitability. This
4303  // will tell us if the addressing mode for the memory operation will
4304  // *actually* cover the shared instruction.
4305  ExtAddrMode Result;
4306  std::pair<AssertingVH<GetElementPtrInst>, int64_t> LargeOffsetGEP(nullptr,
4307  0);
4308  TypePromotionTransaction::ConstRestorationPt LastKnownGood =
4309  TPT.getRestorationPoint();
4310  AddressingModeMatcher Matcher(
4311  MatchedAddrModeInsts, TLI, TRI, AddressAccessTy, AS, MemoryInst, Result,
4312  InsertedInsts, PromotedInsts, TPT, LargeOffsetGEP);
4313  Matcher.IgnoreProfitability = true;
4314  bool Success = Matcher.matchAddr(Address, 0);
4315  (void)Success; assert(Success && "Couldn't select *anything*?");
4316 
4317  // The match was to check the profitability, the changes made are not
4318  // part of the original matcher. Therefore, they should be dropped
4319  // otherwise the original matcher will not present the right state.
4320  TPT.rollback(LastKnownGood);
4321 
4322  // If the match didn't cover I, then it won't be shared by it.
4323  if (!is_contained(MatchedAddrModeInsts, I))
4324  return false;
4325 
4326  MatchedAddrModeInsts.clear();
4327  }
4328 
4329  return true;
4330 }
4331 
4332 /// Return true if the specified values are defined in a
4333 /// different basic block than BB.
4334 static bool IsNonLocalValue(Value *V, BasicBlock *BB) {
4335  if (Instruction *I = dyn_cast<Instruction>(V))
4336  return I->getParent() != BB;
4337  return false;
4338 }
4339 
4340 /// Sink addressing mode computation immediate before MemoryInst if doing so
4341 /// can be done without increasing register pressure. The need for the
4342 /// register pressure constraint means this can end up being an all or nothing
4343 /// decision for all uses of the same addressing computation.
4344 ///
4345 /// Load and Store Instructions often have addressing modes that can do
4346 /// significant amounts of computation. As such, instruction selection will try
4347 /// to get the load or store to do as much computation as possible for the
4348 /// program. The problem is that isel can only see within a single block. As
4349 /// such, we sink as much legal addressing mode work into the block as possible.
4350 ///
4351 /// This method is used to optimize both load/store and inline asms with memory
4352 /// operands. It's also used to sink addressing computations feeding into cold
4353 /// call sites into their (cold) basic block.
4354 ///
4355 /// The motivation for handling sinking into cold blocks is that doing so can
4356 /// both enable other address mode sinking (by satisfying the register pressure
4357 /// constraint above), and reduce register pressure globally (by removing the
4358 /// addressing mode computation from the fast path entirely.).
4359 bool CodeGenPrepare::optimizeMemoryInst(Instruction *MemoryInst, Value *Addr,
4360  Type *AccessTy, unsigned AddrSpace) {
4361  Value *Repl = Addr;
4362 
4363  // Try to collapse single-value PHI nodes. This is necessary to undo
4364  // unprofitable PRE transformations.
4365  SmallVector<Value*, 8> worklist;
4366  SmallPtrSet<Value*, 16> Visited;
4367  worklist.push_back(Addr);
4368 
4369  // Use a worklist to iteratively look through PHI and select nodes, and
4370  // ensure that the addressing mode obtained from the non-PHI/select roots of
4371  // the graph are compatible.
4372  bool PhiOrSelectSeen = false;
4373  SmallVector<Instruction*, 16> AddrModeInsts;
4374  const SimplifyQuery SQ(*DL, TLInfo);
4375  AddressingModeCombiner AddrModes(SQ, { Addr, MemoryInst->getParent() });
4376  TypePromotionTransaction TPT(RemovedInsts);
4377  TypePromotionTransaction::ConstRestorationPt LastKnownGood =
4378  TPT.getRestorationPoint();
4379  while (!worklist.empty()) {
4380  Value *V = worklist.back();
4381  worklist.pop_back();
4382 
4383  // We allow traversing cyclic Phi nodes.
4384  // In case of success after this loop we ensure that traversing through
4385  // Phi nodes ends up with all cases to compute address of the form
4386  // BaseGV + Base + Scale * Index + Offset
4387  // where Scale and Offset are constans and BaseGV, Base and Index
4388  // are exactly the same Values in all cases.
4389  // It means that BaseGV, Scale and Offset dominate our memory instruction
4390  // and have the same value as they had in address computation represented
4391  // as Phi. So we can safely sink address computation to memory instruction.
4392  if (!Visited.insert(V).second)
4393  continue;
4394 
4395  // For a PHI node, push all of its incoming values.
4396  if (PHINode *P = dyn_cast<PHINode>(V)) {
4397  for (Value *IncValue : P->incoming_values())
4398  worklist.push_back(IncValue);
4399  PhiOrSelectSeen = true;
4400  continue;
4401  }
4402  // Similar for select.
4403  if (SelectInst *SI = dyn_cast<SelectInst>(V)) {
4404  worklist.push_back(SI->getFalseValue());
4405  worklist.push_back(SI->getTrueValue());
4406  PhiOrSelectSeen = true;
4407  continue;
4408  }
4409 
4410  // For non-PHIs, determine the addressing mode being computed. Note that
4411  // the result may differ depending on what other uses our candidate
4412  // addressing instructions might have.
4413  AddrModeInsts.clear();
4414  std::pair<AssertingVH<GetElementPtrInst>, int64_t> LargeOffsetGEP(nullptr,
4415  0);
4416  ExtAddrMode NewAddrMode = AddressingModeMatcher::Match(
4417  V, AccessTy, AddrSpace, MemoryInst, AddrModeInsts, *TLI, *TRI,
4418  InsertedInsts, PromotedInsts, TPT, LargeOffsetGEP);
4419 
4420  GetElementPtrInst *GEP = LargeOffsetGEP.first;
4421  if (GEP && GEP->getParent() != MemoryInst->getParent() &&
4422  !NewGEPBases.count(GEP)) {
4423  // If splitting the underlying data structure can reduce the offset of a
4424  // GEP, collect the GEP. Skip the GEPs that are the new bases of
4425  // previously split data structures.
4426  LargeOffsetGEPMap[GEP->getPointerOperand()].push_back(LargeOffsetGEP);
4427  if (LargeOffsetGEPID.find(GEP) == LargeOffsetGEPID.end())
4428  LargeOffsetGEPID[GEP] = LargeOffsetGEPID.size();
4429  }
4430 
4431  NewAddrMode.OriginalValue = V;
4432  if (!AddrModes.addNewAddrMode(NewAddrMode))
4433  break;
4434  }
4435 
4436  // Try to combine the AddrModes we've collected. If we couldn't collect any,
4437  // or we have multiple but either couldn't combine them or combining them
4438  // wouldn't do anything useful, bail out now.
4439  if (!AddrModes.combineAddrModes()) {
4440  TPT.rollback(LastKnownGood);
4441  return false;
4442  }
4443  TPT.commit();
4444 
4445  // Get the combined AddrMode (or the only AddrMode, if we only had one).
4446  ExtAddrMode AddrMode = AddrModes.getAddrMode();
4447 
4448  // If all the instructions matched are already in this BB, don't do anything.
4449  // If we saw a Phi node then it is not local definitely, and if we saw a select
4450  // then we want to push the address calculation past it even if it's already
4451  // in this BB.
4452  if (!PhiOrSelectSeen && none_of(AddrModeInsts, [&](Value *V) {
4453  return IsNonLocalValue(V, MemoryInst->getParent());
4454  })) {
4455  LLVM_DEBUG(dbgs() << "CGP: Found local addrmode: " << AddrMode
4456  << "\n");
4457  return false;
4458  }
4459 
4460  // Insert this computation right after this user. Since our caller is
4461  // scanning from the top of the BB to the bottom, reuse of the expr are
4462  // guaranteed to happen later.
4463  IRBuilder<> Builder(MemoryInst);
4464 
4465  // Now that we determined the addressing expression we want to use and know
4466  // that we have to sink it into this block. Check to see if we have already
4467  // done this for some other load/store instr in this block. If so, reuse
4468  // the computation. Before attempting reuse, check if the address is valid
4469  // as it may have been erased.
4470 
4471  WeakTrackingVH SunkAddrVH = SunkAddrs[Addr];
4472 
4473  Value * SunkAddr = SunkAddrVH.pointsToAliveValue() ? SunkAddrVH : nullptr;
4474  if (SunkAddr) {
4475  LLVM_DEBUG(dbgs() << "CGP: Reusing nonlocal addrmode: " << AddrMode
4476  << " for " << *MemoryInst << "\n");
4477  if (SunkAddr->getType() != Addr->getType())
4478  SunkAddr = Builder.CreatePointerCast(SunkAddr, Addr->getType());
4479  } else if (AddrSinkUsingGEPs ||
4480  (!AddrSinkUsingGEPs.getNumOccurrences() && TM && TTI->useAA())) {
4481  // By default, we use the GEP-based method when AA is used later. This
4482  // prevents new inttoptr/ptrtoint pairs from degrading AA capabilities.
4483  LLVM_DEBUG(dbgs() << "CGP: SINKING nonlocal addrmode: " << AddrMode
4484  << " for " << *MemoryInst << "\n");
4485  Type *IntPtrTy = DL->getIntPtrType(Addr->getType());
4486  Value *ResultPtr = nullptr, *ResultIndex = nullptr;
4487 
4488  // First, find the pointer.
4489  if (AddrMode.BaseReg && AddrMode.BaseReg->getType()->isPointerTy()) {
4490  ResultPtr = AddrMode.BaseReg;
4491  AddrMode.BaseReg = nullptr;
4492  }
4493 
4494  if (AddrMode.Scale && AddrMode.ScaledReg->getType()->isPointerTy()) {
4495  // We can't add more than one pointer together, nor can we scale a
4496  // pointer (both of which seem meaningless).
4497  if (ResultPtr || AddrMode.Scale != 1)
4498  return false;
4499 
4500  ResultPtr = AddrMode.ScaledReg;
4501  AddrMode.Scale = 0;
4502  }
4503 
4504  // It is only safe to sign extend the BaseReg if we know that the math
4505  // required to create it did not overflow before we extend it. Since
4506  // the original IR value was tossed in favor of a constant back when
4507  // the AddrMode was created we need to bail out gracefully if widths
4508  // do not match instead of extending it.
4509  //
4510  // (See below for code to add the scale.)
4511  if (AddrMode.Scale) {
4512  Type *ScaledRegTy = AddrMode.ScaledReg->getType();
4513  if (cast<IntegerType>(IntPtrTy)->getBitWidth() >
4514  cast<IntegerType>(ScaledRegTy)->getBitWidth())
4515  return false;
4516  }
4517 
4518  if (AddrMode.BaseGV) {
4519  if (ResultPtr)
4520  return false;
4521 
4522  ResultPtr = AddrMode.BaseGV;
4523  }
4524 
4525  // If the real base value actually came from an inttoptr, then the matcher
4526  // will look through it and provide only the integer value. In that case,
4527  // use it here.
4528  if (!DL->isNonIntegralPointerType(Addr->getType())) {
4529  if (!ResultPtr && AddrMode.BaseReg) {
4530  ResultPtr = Builder.CreateIntToPtr(AddrMode.BaseReg, Addr->getType(),
4531  "sunkaddr");
4532  AddrMode.BaseReg = nullptr;
4533  } else if (!ResultPtr && AddrMode.Scale == 1) {
4534  ResultPtr = Builder.CreateIntToPtr(AddrMode.ScaledReg, Addr->getType(),
4535  "sunkaddr");
4536  AddrMode.Scale = 0;
4537  }
4538  }
4539 
4540  if (!ResultPtr &&
4541  !AddrMode.BaseReg && !AddrMode.Scale && !AddrMode.BaseOffs) {
4542  SunkAddr = Constant::getNullValue(Addr->getType());
4543  } else if (!ResultPtr) {
4544  return false;
4545  } else {
4546  Type *I8PtrTy =
4547  Builder.getInt8PtrTy(Addr->getType()->getPointerAddressSpace());
4548  Type *I8Ty = Builder.getInt8Ty();
4549 
4550  // Start with the base register. Do this first so that subsequent address
4551  // matching finds it last, which will prevent it from trying to match it
4552  // as the scaled value in case it happens to be a mul. That would be
4553  // problematic if we've sunk a different mul for the scale, because then
4554  // we'd end up sinking both muls.
4555  if (AddrMode.BaseReg) {
4556  Value *V = AddrMode.BaseReg;
4557  if (V->getType() != IntPtrTy)
4558  V = Builder.CreateIntCast(V, IntPtrTy, /*isSigned=*/true, "sunkaddr");
4559 
4560  ResultIndex = V;
4561  }
4562 
4563  // Add the scale value.
4564  if (AddrMode.Scale) {
4565  Value *V = AddrMode.ScaledReg;
4566  if (V->getType() == IntPtrTy) {
4567  // done.
4568  } else {
4569  assert(cast<IntegerType>(IntPtrTy)->getBitWidth() <
4570  cast<IntegerType>(V->getType())->getBitWidth() &&
4571  "We can't transform if ScaledReg is too narrow");
4572  V = Builder.CreateTrunc(V, IntPtrTy, "sunkaddr");
4573  }
4574 
4575  if (AddrMode.Scale != 1)
4576  V = Builder.CreateMul(V, ConstantInt::get(IntPtrTy, AddrMode.Scale),
4577  "sunkaddr");
4578  if (ResultIndex)
4579  ResultIndex = Builder.CreateAdd(ResultIndex, V, "sunkaddr");
4580  else
4581  ResultIndex = V;
4582  }
4583 
4584  // Add in the Base Offset if present.
4585  if (AddrMode.BaseOffs) {
4586  Value *V = ConstantInt::get(IntPtrTy, AddrMode.BaseOffs);
4587  if (ResultIndex) {
4588  // We need to add this separately from the scale above to help with
4589  // SDAG consecutive load/store merging.
4590  if (ResultPtr->getType() != I8PtrTy)
4591  ResultPtr = Builder.CreatePointerCast(ResultPtr, I8PtrTy);
4592  ResultPtr = Builder.CreateGEP(I8Ty, ResultPtr, ResultIndex, "sunkaddr");
4593  }
4594 
4595  ResultIndex = V;
4596  }
4597 
4598  if (!ResultIndex) {
4599  SunkAddr = ResultPtr;
4600  } else {
4601  if (ResultPtr->getType() != I8PtrTy)
4602  ResultPtr = Builder.CreatePointerCast(ResultPtr, I8PtrTy);
4603  SunkAddr = Builder.CreateGEP(I8Ty, ResultPtr, ResultIndex, "sunkaddr");
4604  }
4605 
4606  if (SunkAddr->getType() != Addr->getType())
4607  SunkAddr = Builder.CreatePointerCast(SunkAddr, Addr->getType());
4608  }
4609  } else {
4610  // We'd require a ptrtoint/inttoptr down the line, which we can't do for
4611  // non-integral pointers, so in that case bail out now.
4612  Type *BaseTy = AddrMode.BaseReg ? AddrMode.BaseReg->getType() : nullptr;
4613  Type *ScaleTy = AddrMode.Scale ? AddrMode.ScaledReg->getType() : nullptr;
4614  PointerType *BasePtrTy = dyn_cast_or_null<PointerType>(BaseTy);
4615  PointerType *ScalePtrTy = dyn_cast_or_null<PointerType>(ScaleTy);
4616  if (DL->isNonIntegralPointerType(Addr->getType()) ||
4617  (BasePtrTy && DL->isNonIntegralPointerType(BasePtrTy)) ||
4618  (ScalePtrTy && DL->isNonIntegralPointerType(ScalePtrTy)) ||
4619  (AddrMode.BaseGV &&
4620  DL->isNonIntegralPointerType(AddrMode.BaseGV->getType())))
4621  return false;
4622 
4623  LLVM_DEBUG(dbgs() << "CGP: SINKING nonlocal addrmode: " << AddrMode
4624  << " for " << *MemoryInst << "\n");
4625  Type *IntPtrTy = DL->getIntPtrType(Addr->getType());
4626  Value *Result = nullptr;
4627 
4628  // Start with the base register. Do this first so that subsequent address
4629  // matching finds it last, which will prevent it from trying to match it
4630  // as the scaled value in case it happens to be a mul. That would be
4631  // problematic if we've sunk a different mul for the scale, because then
4632  // we'd end up sinking both muls.
4633  if (AddrMode.BaseReg) {
4634  Value *V = AddrMode.BaseReg;
4635  if (V->getType()->isPointerTy())
4636  V = Builder.CreatePtrToInt(V, IntPtrTy, "sunkaddr");
4637  if (V->getType() != IntPtrTy)
4638  V = Builder.CreateIntCast(V, IntPtrTy, /*isSigned=*/true, "sunkaddr");
4639  Result = V;
4640  }
4641 
4642  // Add the scale value.
4643  if (AddrMode.Scale) {
4644  Value *V = AddrMode.ScaledReg;
4645  if (V->getType() == IntPtrTy) {
4646  // done.
4647  } else if (V->getType()->isPointerTy()) {
4648  V = Builder.CreatePtrToInt(V, IntPtrTy, "sunkaddr");
4649  } else if (cast<IntegerType>(IntPtrTy)->getBitWidth() <
4650  cast<IntegerType>(V->getType())->getBitWidth()) {
4651  V = Builder.CreateTrunc(V, IntPtrTy, "sunkaddr");
4652  } else {
4653  // It is only safe to sign extend the BaseReg if we know that the math
4654  // required to create it did not overflow before we extend it. Since
4655  // the original IR value was tossed in favor of a constant back when
4656  // the AddrMode was created we need to bail out gracefully if widths
4657  // do not match instead of extending it.
4658  Instruction *I = dyn_cast_or_null<Instruction>(Result);
4659  if (I && (Result != AddrMode.BaseReg))
4660  I->eraseFromParent();
4661  return false;
4662  }
4663  if (AddrMode.Scale != 1)
4664  V = Builder.CreateMul(V, ConstantInt::get(IntPtrTy, AddrMode.Scale),
4665  "sunkaddr");
4666  if (Result)
4667  Result = Builder.CreateAdd(Result, V, "sunkaddr");
4668  else
4669  Result = V;
4670  }
4671 
4672  // Add in the BaseGV if present.
4673  if (AddrMode.BaseGV) {
4674  Value *V = Builder.CreatePtrToInt(AddrMode.BaseGV, IntPtrTy, "sunkaddr");
4675  if (Result)
4676  Result = Builder.CreateAdd(Result, V, "sunkaddr");
4677  else
4678  Result = V;
4679  }
4680 
4681  // Add in the Base Offset if present.
4682  if (AddrMode.BaseOffs) {
4683  Value *V = ConstantInt::get(IntPtrTy, AddrMode.BaseOffs);
4684  if (Result)
4685  Result = Builder.CreateAdd(Result, V, "sunkaddr");
4686  else
4687  Result = V;
4688  }
4689 
4690  if (!Result)
4691  SunkAddr = Constant::getNullValue(Addr->getType());
4692  else
4693  SunkAddr = Builder.CreateIntToPtr(Result, Addr->getType(), "sunkaddr");
4694  }
4695 
4696  MemoryInst->replaceUsesOfWith(Repl, SunkAddr);
4697  // Store the newly computed address into the cache. In the case we reused a
4698  // value, this should be idempotent.
4699  SunkAddrs[Addr] = WeakTrackingVH(SunkAddr);
4700 
4701  // If we have no uses, recursively delete the value and all dead instructions
4702  // using it.
4703  if (Repl->use_empty()) {
4704  // This can cause recursive deletion, which can invalidate our iterator.
4705  // Use a WeakTrackingVH to hold onto it in case this happens.
4706  Value *CurValue = &*CurInstIterator;
4707  WeakTrackingVH IterHandle(CurValue);
4708  BasicBlock *BB = CurInstIterator->getParent();
4709 
4711 
4712  if (IterHandle != CurValue) {
4713  // If the iterator instruction was recursively deleted, start over at the
4714  // start of the block.
4715  CurInstIterator = BB->begin();
4716  SunkAddrs.clear();
4717  }
4718  }
4719  ++NumMemoryInsts;
4720  return true;
4721 }
4722 
4723 /// If there are any memory operands, use OptimizeMemoryInst to sink their
4724 /// address computing into the block when possible / profitable.
4725 bool CodeGenPrepare::optimizeInlineAsmInst(CallInst *CS) {
4726  bool MadeChange = false;
4727 
4728  const TargetRegisterInfo *TRI =
4729  TM->getSubtargetImpl(*CS->getFunction())->getRegisterInfo();
4730  TargetLowering::AsmOperandInfoVector TargetConstraints =
4731  TLI->ParseConstraints(*DL, TRI, CS);
4732  unsigned ArgNo = 0;
4733  for (unsigned i = 0, e = TargetConstraints.size(); i != e; ++i) {
4734  TargetLowering::AsmOperandInfo &OpInfo = TargetConstraints[i];
4735 
4736  // Compute the constraint code and ConstraintType to use.
4737  TLI->ComputeConstraintToUse(OpInfo, SDValue());
4738 
4739  if (OpInfo.ConstraintType == TargetLowering::C_Memory &&
4740  OpInfo.isIndirect) {
4741  Value *OpVal = CS->getArgOperand(ArgNo++);
4742  MadeChange |= optimizeMemoryInst(CS, OpVal, OpVal->getType(), ~0u);
4743  } else if (OpInfo.Type == InlineAsm::isInput)
4744  ArgNo++;
4745  }
4746 
4747  return MadeChange;
4748 }
4749 
4750 /// Check if all the uses of \p Val are equivalent (or free) zero or
4751 /// sign extensions.
4752 static bool hasSameExtUse(Value *Val, const TargetLowering &TLI) {
4753  assert(!Val->use_empty() && "Input must have at least one use");
4754  const Instruction *FirstUser = cast<Instruction>(*Val->user_begin());
4755  bool IsSExt = isa<SExtInst>(FirstUser);
4756  Type *ExtTy = FirstUser->getType();
4757  for (const User *U : Val->users()) {
4758  const Instruction *UI = cast<Instruction>(U);
4759  if ((IsSExt && !isa<SExtInst>(UI)) || (!IsSExt && !isa<ZExtInst>(UI)))
4760  return false;
4761  Type *CurTy = UI->getType();
4762  // Same input and output types: Same instruction after CSE.
4763  if (CurTy == ExtTy)
4764  continue;
4765 
4766  // If IsSExt is true, we are in this situation:
4767  // a = Val
4768  // b = sext ty1 a to ty2
4769  // c = sext ty1 a to ty3
4770  // Assuming ty2 is shorter than ty3, this could be turned into:
4771  // a = Val
4772  // b = sext ty1 a to ty2
4773  // c = sext ty2 b to ty3
4774  // However, the last sext is not free.
4775  if (IsSExt)
4776  return false;
4777 
4778  // This is a ZExt, maybe this is free to extend from one type to another.
4779  // In that case, we would not account for a different use.
4780  Type *NarrowTy;
4781  Type *LargeTy;
4782  if (ExtTy->getScalarType()->getIntegerBitWidth() >
4783  CurTy->getScalarType()->getIntegerBitWidth()) {
4784  NarrowTy = CurTy;
4785  LargeTy = ExtTy;
4786  } else {
4787  NarrowTy = ExtTy;
4788  LargeTy = CurTy;
4789  }
4790 
4791  if (!TLI.isZExtFree(NarrowTy, LargeTy))
4792  return false;
4793  }
4794  // All uses are the same or can be derived from one another for free.
4795  return true;
4796 }
4797 
4798 /// Try to speculatively promote extensions in \p Exts and continue
4799 /// promoting through newly promoted operands recursively as far as doing so is
4800 /// profitable. Save extensions profitably moved up, in \p ProfitablyMovedExts.
4801 /// When some promotion happened, \p TPT contains the proper state to revert
4802 /// them.
4803 ///
4804 /// \return true if some promotion happened, false otherwise.
4805 bool CodeGenPrepare::tryToPromoteExts(
4806  TypePromotionTransaction &TPT, const SmallVectorImpl<Instruction *> &Exts,
4807  SmallVectorImpl<Instruction *> &ProfitablyMovedExts,
4808  unsigned CreatedInstsCost) {
4809  bool Promoted = false;
4810 
4811  // Iterate over all the extensions to try to promote them.
4812  for (auto I : Exts) {
4813  // Early check if we directly have ext(load).
4814  if (isa<LoadInst>(I->getOperand(0))) {
4815  ProfitablyMovedExts.push_back(I);
4816  continue;
4817  }
4818 
4819  // Check whether or not we want to do any promotion. The reason we have
4820  // this check inside the for loop is to catch the case where an extension
4821  // is directly fed by a load because in such case the extension can be moved
4822  // up without any promotion on its operands.
4823  if (!TLI || !TLI->enableExtLdPromotion() || DisableExtLdPromotion)
4824  return false;
4825 
4826  // Get the action to perform the promotion.
4827  TypePromotionHelper::Action TPH =
4828  TypePromotionHelper::getAction(I, InsertedInsts, *TLI, PromotedInsts);
4829  // Check if we can promote.
4830  if (!TPH) {
4831  // Save the current extension as we cannot move up through its operand.
4832  ProfitablyMovedExts.push_back(I);
4833  continue;
4834  }
4835 
4836  // Save the current state.
4837  TypePromotionTransaction::ConstRestorationPt LastKnownGood =
4838  TPT.getRestorationPoint();
4840  unsigned NewCreatedInstsCost = 0;
4841  unsigned ExtCost = !TLI->isExtFree(I);
4842  // Promote.
4843  Value *PromotedVal = TPH(I, TPT, PromotedInsts, NewCreatedInstsCost,
4844  &NewExts, nullptr, *TLI);
4845  assert(PromotedVal &&
4846  "TypePromotionHelper should have filtered out those cases");
4847 
4848  // We would be able to merge only one extension in a load.
4849  // Therefore, if we have more than 1 new extension we heuristically
4850  // cut this search path, because it means we degrade the code quality.
4851  // With exactly 2, the transformation is neutral, because we will merge
4852  // one extension but leave one. However, we optimistically keep going,
4853  // because the new extension may be removed too.
4854  long long TotalCreatedInstsCost = CreatedInstsCost + NewCreatedInstsCost;
4855  // FIXME: It would be possible to propagate a negative value instead of
4856  // conservatively ceiling it to 0.
4857  TotalCreatedInstsCost =
4858  std::max((long long)0, (TotalCreatedInstsCost - ExtCost));
4859  if (!StressExtLdPromotion &&
4860  (TotalCreatedInstsCost > 1 ||
4861  !isPromotedInstructionLegal(*TLI, *DL, PromotedVal))) {
4862  // This promotion is not profitable, rollback to the previous state, and
4863  // save the current extension in ProfitablyMovedExts as the latest
4864  // speculative promotion turned out to be unprofitable.
4865  TPT.rollback(LastKnownGood);
4866  ProfitablyMovedExts.push_back(I);
4867  continue;
4868  }
4869  // Continue promoting NewExts as far as doing so is profitable.
4870  SmallVector<Instruction *, 2> NewlyMovedExts;
4871  (void)tryToPromoteExts(TPT, NewExts, NewlyMovedExts, TotalCreatedInstsCost);
4872  bool NewPromoted = false;
4873  for (auto ExtInst : NewlyMovedExts) {
4874  Instruction *MovedExt = cast<Instruction>(ExtInst);
4875  Value *ExtOperand = MovedExt->getOperand(0);
4876  // If we have reached to a load, we need this extra profitability check
4877  // as it could potentially be merged into an ext(load).
4878  if (isa<LoadInst>(ExtOperand) &&
4879  !(StressExtLdPromotion || NewCreatedInstsCost <= ExtCost ||
4880  (ExtOperand->hasOneUse() || hasSameExtUse(ExtOperand, *TLI))))
4881  continue;
4882 
4883  ProfitablyMovedExts.push_back(MovedExt);
4884  NewPromoted = true;
4885  }
4886 
4887  // If none of speculative promotions for NewExts is profitable, rollback
4888  // and save the current extension (I) as the last profitable extension.
4889  if (!NewPromoted) {
4890  TPT.rollback(LastKnownGood);
4891  ProfitablyMovedExts.push_back(I);
4892  continue;
4893  }
4894  // The promotion is profitable.
4895  Promoted = true;
4896  }
4897  return Promoted;
4898 }
4899 
4900 /// Merging redundant sexts when one is dominating the other.
4901 bool CodeGenPrepare::mergeSExts(Function &F) {
4902  DominatorTree DT(F);
4903  bool Changed = false;
4904  for (auto &Entry : ValToSExtendedUses) {
4905  SExts &Insts = Entry.second;
4906  SExts CurPts;
4907  for (Instruction *Inst : Insts) {
4908  if (RemovedInsts.count(Inst) || !isa<SExtInst>(Inst) ||
4909  Inst->getOperand(0) != Entry.first)
4910  continue;
4911  bool inserted = false;
4912  for (auto &Pt : CurPts) {
4913  if (DT.dominates(Inst, Pt)) {
4914  Pt->replaceAllUsesWith(Inst);
4915  RemovedInsts.insert(Pt);
4916  Pt->removeFromParent();
4917  Pt = Inst;
4918  inserted = true;
4919  Changed = true;
4920  break;
4921  }
4922  if (!DT.dominates(Pt, Inst))
4923  // Give up if we need to merge in a common dominator as the
4924  // experiments show it is not profitable.
4925  continue;
4926  Inst->replaceAllUsesWith(Pt);
4927  RemovedInsts.insert(Inst);
4928  Inst->removeFromParent();
4929  inserted = true;
4930  Changed = true;
4931  break;
4932  }
4933  if (!inserted)
4934  CurPts.push_back(Inst);
4935  }
4936  }
4937  return Changed;
4938 }
4939 
4940 // Spliting large data structures so that the GEPs accessing them can have
4941 // smaller offsets so that they can be sunk to the same blocks as their users.
4942 // For example, a large struct starting from %base is splitted into two parts
4943 // where the second part starts from %new_base.
4944 //
4945 // Before:
4946 // BB0:
4947 // %base =
4948 //
4949 // BB1:
4950 // %gep0 = gep %base, off0
4951 // %gep1 = gep %base, off1
4952 // %gep2 = gep %base, off2
4953 //
4954 // BB2:
4955 // %load1 = load %gep0
4956 // %load2 = load %gep1
4957 // %load3 = load %gep2
4958 //
4959 // After:
4960 // BB0:
4961 // %base =
4962 // %new_base = gep %base, off0
4963 //
4964 // BB1:
4965 // %new_gep0 = %new_base
4966 // %new_gep1 = gep %new_base, off1 - off0
4967 // %new_gep2 = gep %new_base, off2 - off0
4968 //
4969 // BB2:
4970 // %load1 = load i32, i32* %new_gep0
4971 // %load2 = load i32, i32* %new_gep1
4972 // %load3 = load i32, i32* %new_gep2
4973 //
4974 // %new_gep1 and %new_gep2 can be sunk to BB2 now after the splitting because
4975 // their offsets are smaller enough to fit into the addressing mode.
4976 bool CodeGenPrepare::splitLargeGEPOffsets() {
4977  bool Changed = false;
4978  for (auto &Entry : LargeOffsetGEPMap) {
4979  Value *OldBase = Entry.first;
4981  &LargeOffsetGEPs = Entry.second;
4982  auto compareGEPOffset =
4983  [&](const std::pair<GetElementPtrInst *, int64_t> &LHS,
4984  const std::pair<GetElementPtrInst *, int64_t> &RHS) {
4985  if (LHS.first == RHS.first)
4986  return false;
4987  if (LHS.second != RHS.second)
4988  return LHS.second < RHS.second;
4989  return LargeOffsetGEPID[LHS.first] < LargeOffsetGEPID[RHS.first];
4990  };
4991  // Sorting all the GEPs of the same data structures based on the offsets.
4992  llvm::sort(LargeOffsetGEPs, compareGEPOffset);
4993  LargeOffsetGEPs.erase(
4994  std::unique(LargeOffsetGEPs.begin(), LargeOffsetGEPs.end()),
4995  LargeOffsetGEPs.end());
4996  // Skip if all the GEPs have the same offsets.
4997  if (LargeOffsetGEPs.front().second == LargeOffsetGEPs.back().second)
4998  continue;
4999  GetElementPtrInst *BaseGEP = LargeOffsetGEPs.begin()->first;
5000  int64_t BaseOffset = LargeOffsetGEPs.begin()->second;
5001  Value *NewBaseGEP = nullptr;
5002 
5003  auto LargeOffsetGEP = LargeOffsetGEPs.begin();
5004  while (LargeOffsetGEP != LargeOffsetGEPs.end()) {
5005  GetElementPtrInst *GEP = LargeOffsetGEP->first;
5006  int64_t Offset = LargeOffsetGEP->second;
5007  if (Offset != BaseOffset) {
5009  AddrMode.BaseOffs = Offset - BaseOffset;
5010  // The result type of the GEP might not be the type of the memory
5011  // access.
5012  if (!TLI->isLegalAddressingMode(*DL, AddrMode,
5013  GEP->getResultElementType(),
5014  GEP->getAddressSpace())) {
5015  // We need to create a new base if the offset to the current base is
5016  // too large to fit into the addressing mode. So, a very large struct
5017  // may be splitted into several parts.
5018  BaseGEP = GEP;
5019  BaseOffset = Offset;
5020  NewBaseGEP = nullptr;
5021  }
5022  }
5023 
5024  // Generate a new GEP to replace the current one.
5025  IRBuilder<> Builder(GEP);
5026  Type *IntPtrTy = DL->getIntPtrType(GEP->getType());
5027  Type *I8PtrTy =
5028  Builder.getInt8PtrTy(GEP->getType()->getPointerAddressSpace());
5029  Type *I8Ty = Builder.getInt8Ty();
5030 
5031  if (!NewBaseGEP) {
5032  // Create a new base if we don't have one yet. Find the insertion
5033  // pointer for the new base first.
5034  BasicBlock::iterator NewBaseInsertPt;
5035  BasicBlock *NewBaseInsertBB;
5036  if (auto *BaseI = dyn_cast<Instruction>(OldBase)) {
5037  // If the base of the struct is an instruction, the new base will be
5038  // inserted close to it.
5039  NewBaseInsertBB = BaseI->getParent();
5040  if (isa<PHINode>(BaseI))
5041  NewBaseInsertPt = NewBaseInsertBB->getFirstInsertionPt();
5042  else if (InvokeInst *Invoke = dyn_cast<InvokeInst>(BaseI)) {
5043  NewBaseInsertBB =
5044  SplitEdge(NewBaseInsertBB, Invoke->getNormalDest());
5045  NewBaseInsertPt = NewBaseInsertBB->getFirstInsertionPt();
5046  } else
5047  NewBaseInsertPt = std::next(BaseI->getIterator());
5048  } else {
5049  // If the current base is an argument or global value, the new base
5050  // will be inserted to the entry block.
5051  NewBaseInsertBB = &BaseGEP->getFunction()->getEntryBlock();
5052  NewBaseInsertPt = NewBaseInsertBB->getFirstInsertionPt();
5053  }
5054  IRBuilder<> NewBaseBuilder(NewBaseInsertBB, NewBaseInsertPt);
5055  // Create a new base.
5056  Value *BaseIndex = ConstantInt::get(IntPtrTy, BaseOffset);
5057  NewBaseGEP = OldBase;
5058  if (NewBaseGEP->getType() != I8PtrTy)
5059  NewBaseGEP = NewBaseBuilder.CreatePointerCast(NewBaseGEP, I8PtrTy);
5060  NewBaseGEP =
5061  NewBaseBuilder.CreateGEP(I8Ty, NewBaseGEP, BaseIndex, "splitgep");
5062  NewGEPBases.insert(NewBaseGEP);
5063  }
5064 
5065  Value *NewGEP = NewBaseGEP;
5066  if (Offset == BaseOffset) {
5067  if (GEP->getType() != I8PtrTy)
5068  NewGEP = Builder.CreatePointerCast(NewGEP, GEP->getType());
5069  } else {
5070  // Calculate the new offset for the new GEP.
5071  Value *Index = ConstantInt::get(IntPtrTy, Offset - BaseOffset);
5072  NewGEP = Builder.CreateGEP(I8Ty, NewBaseGEP, Index);
5073 
5074  if (GEP->getType() != I8PtrTy)
5075  NewGEP = Builder.CreatePointerCast(NewGEP, GEP->getType());
5076  }
5077  GEP->replaceAllUsesWith(NewGEP);
5078  LargeOffsetGEPID.erase(GEP);
5079  LargeOffsetGEP = LargeOffsetGEPs.erase(LargeOffsetGEP);
5080  GEP->eraseFromParent();
5081  Changed = true;
5082  }
5083  }
5084  return Changed;
5085 }
5086 
5087 /// Return true, if an ext(load) can be formed from an extension in
5088 /// \p MovedExts.
5089 bool CodeGenPrepare::canFormExtLd(
5090  const SmallVectorImpl<Instruction *> &MovedExts, LoadInst *&LI,
5091  Instruction *&Inst, bool HasPromoted) {
5092  for (auto *MovedExtInst : MovedExts) {
5093  if (isa<LoadInst>(MovedExtInst->getOperand(0))) {
5094  LI = cast<LoadInst>(MovedExtInst->getOperand(0));
5095  Inst = MovedExtInst;
5096  break;
5097  }
5098  }
5099  if (!LI)
5100  return false;
5101 
5102  // If they're already in the same block, there's nothing to do.
5103  // Make the cheap checks first if we did not promote.
5104  // If we promoted, we need to check if it is indeed profitable.
5105  if (!HasPromoted && LI->getParent() == Inst->getParent())
5106  return false;
5107 
5108  return TLI->isExtLoad(LI, Inst, *DL);
5109 }
5110 
5111 /// Move a zext or sext fed by a load into the same basic block as the load,
5112 /// unless conditions are unfavorable. This allows SelectionDAG to fold the
5113 /// extend into the load.
5114 ///
5115 /// E.g.,
5116 /// \code
5117 /// %ld = load i32* %addr
5118 /// %add = add nuw i32 %ld, 4
5119 /// %zext = zext i32 %add to i64
5120 // \endcode
5121 /// =>
5122 /// \code
5123 /// %ld = load i32* %addr
5124 /// %zext = zext i32 %ld to i64
5125 /// %add = add nuw i64 %zext, 4
5126 /// \encode
5127 /// Note that the promotion in %add to i64 is done in tryToPromoteExts(), which
5128 /// allow us to match zext(load i32*) to i64.
5129 ///
5130 /// Also, try to promote the computations used to obtain a sign extended
5131 /// value used into memory accesses.
5132 /// E.g.,
5133 /// \code
5134 /// a = add nsw i32 b, 3
5135 /// d = sext i32 a to i64
5136 /// e = getelementptr ..., i64 d
5137 /// \endcode
5138 /// =>
5139 /// \code
5140 /// f = sext i32 b to i64
5141 /// a = add nsw i64 f, 3
5142 /// e = getelementptr ..., i64 a
5143 /// \endcode
5144 ///
5145 /// \p Inst[in/out] the extension may be modified during the process if some
5146 /// promotions apply.
5147 bool CodeGenPrepare::optimizeExt(Instruction *&Inst) {
5148  // ExtLoad formation and address type promotion infrastructure requires TLI to
5149  // be effective.
5150  if (!TLI)
5151  return false;
5152 
5153  bool AllowPromotionWithoutCommonHeader = false;
5154  /// See if it is an interesting sext operations for the address type
5155  /// promotion before trying to promote it, e.g., the ones with the right
5156  /// type and used in memory accesses.
5157  bool ATPConsiderable = TTI->shouldConsiderAddressTypePromotion(
5158  *Inst, AllowPromotionWithoutCommonHeader);
5159  TypePromotionTransaction TPT(RemovedInsts);
5160  TypePromotionTransaction::ConstRestorationPt LastKnownGood =
5161  TPT.getRestorationPoint();
5163  SmallVector<Instruction *, 2> SpeculativelyMovedExts;
5164  Exts.push_back(Inst);
5165 
5166  bool HasPromoted = tryToPromoteExts(TPT, Exts, SpeculativelyMovedExts);
5167 
5168  // Look for a load being extended.
5169  LoadInst *LI = nullptr;
5170  Instruction *ExtFedByLoad;
5171 
5172  // Try to promote a chain of computation if it allows to form an extended
5173  // load.
5174  if (canFormExtLd(SpeculativelyMovedExts, LI, ExtFedByLoad, HasPromoted)) {
5175  assert(LI && ExtFedByLoad && "Expect a valid load and extension");
5176  TPT.commit();
5177  // Move the extend into the same block as the load
5178  ExtFedByLoad->moveAfter(LI);
5179  // CGP does not check if the zext would be speculatively executed when moved
5180  // to the same basic block as the load. Preserving its original location
5181  // would pessimize the debugging experience, as well as negatively impact
5182  // the quality of sample pgo. We don't want to use "line 0" as that has a
5183  // size cost in the line-table section and logically the zext can be seen as
5184  // part of the load. Therefore we conservatively reuse the same debug
5185  // location for the load and the zext.
5186  ExtFedByLoad->setDebugLoc(LI->getDebugLoc());
5187  ++NumExtsMoved;
5188  Inst = ExtFedByLoad;
5189  return true;
5190  }
5191 
5192  // Continue promoting SExts if known as considerable depending on targets.
5193  if (ATPConsiderable &&
5194  performAddressTypePromotion(Inst, AllowPromotionWithoutCommonHeader,
5195  HasPromoted, TPT, SpeculativelyMovedExts))
5196  return true;
5197 
5198  TPT.rollback(LastKnownGood);
5199  return false;
5200 }
5201 
5202 // Perform address type promotion if doing so is profitable.
5203 // If AllowPromotionWithoutCommonHeader == false, we should find other sext
5204 // instructions that sign extended the same initial value. However, if
5205 // AllowPromotionWithoutCommonHeader == true, we expect promoting the
5206 // extension is just profitable.
5207 bool CodeGenPrepare::performAddressTypePromotion(
5208  Instruction *&Inst, bool AllowPromotionWithoutCommonHeader,
5209  bool HasPromoted, TypePromotionTransaction &TPT,
5210  SmallVectorImpl<Instruction *> &SpeculativelyMovedExts) {
5211  bool Promoted = false;
5212  SmallPtrSet<Instruction *, 1> UnhandledExts;
5213  bool AllSeenFirst = true;
5214  for (auto I : SpeculativelyMovedExts) {
5215  Value *HeadOfChain = I->getOperand(0);
5217  SeenChainsForSExt.find(HeadOfChain);
5218  // If there is an unhandled SExt which has the same header, try to promote
5219  // it as well.
5220  if (AlreadySeen != SeenChainsForSExt.end()) {
5221  if (AlreadySeen->second != nullptr)
5222  UnhandledExts.insert(AlreadySeen->second);
5223  AllSeenFirst = false;
5224  }
5225  }
5226 
5227  if (!AllSeenFirst || (AllowPromotionWithoutCommonHeader &&
5228  SpeculativelyMovedExts.size() == 1)) {
5229  TPT.commit();
5230  if (HasPromoted)
5231  Promoted = true;
5232  for (auto I : SpeculativelyMovedExts) {
5233  Value *HeadOfChain = I->getOperand(0);
5234  SeenChainsForSExt[HeadOfChain] = nullptr;
5235  ValToSExtendedUses[HeadOfChain].push_back(I);
5236  }
5237  // Update Inst as promotion happen.
5238  Inst = SpeculativelyMovedExts.pop_back_val();
5239  } else {
5240  // This is the first chain visited from the header, keep the current chain
5241  // as unhandled. Defer to promote this until we encounter another SExt
5242  // chain derived from the same header.
5243  for (auto I : SpeculativelyMovedExts) {
5244  Value *HeadOfChain = I->getOperand(0);
5245  SeenChainsForSExt[HeadOfChain] = Inst;
5246  }
5247  return false;
5248  }
5249 
5250  if (!AllSeenFirst && !UnhandledExts.empty())
5251  for (auto VisitedSExt : UnhandledExts) {
5252  if (RemovedInsts.count(VisitedSExt))
5253  continue;
5254  TypePromotionTransaction TPT(RemovedInsts);
5257  Exts.push_back(VisitedSExt);
5258  bool HasPromoted = tryToPromoteExts(TPT, Exts, Chains);
5259  TPT.commit();
5260  if (HasPromoted)
5261  Promoted = true;
5262  for (auto I : Chains) {
5263  Value *HeadOfChain = I->getOperand(0);
5264  // Mark this as handled.
5265  SeenChainsForSExt[HeadOfChain] = nullptr;
5266  ValToSExtendedUses[HeadOfChain].push_back(I);
5267  }
5268  }
5269  return Promoted;
5270 }
5271 
5272 bool CodeGenPrepare::optimizeExtUses(Instruction *I) {
5273  BasicBlock *DefBB = I->getParent();
5274 
5275  // If the result of a {s|z}ext and its source are both live out, rewrite all
5276  // other uses of the source with result of extension.
5277  Value *Src = I->getOperand(0);
5278  if (Src->hasOneUse())
5279  return false;
5280 
5281  // Only do this xform if truncating is free.
5282  if (TLI && !TLI->isTruncateFree(I->getType(), Src->getType()))
5283  return false;
5284 
5285  // Only safe to perform the optimization if the source is also defined in
5286  // this block.
5287  if (!isa<Instruction>(Src) || DefBB != cast<Instruction>(Src)->getParent())
5288  return false;
5289 
5290  bool DefIsLiveOut = false;
5291  for (User *U : I->users()) {
5292  Instruction *UI = cast<Instruction>(U);
5293 
5294  // Figure out which BB this ext is used in.
5295  BasicBlock *UserBB = UI->getParent();
5296  if (UserBB == DefBB) continue;
5297  DefIsLiveOut = true;
5298  break;
5299  }
5300  if (!DefIsLiveOut)
5301  return false;
5302 
5303  // Make sure none of the uses are PHI nodes.
5304  for (User *U : Src->users()) {
5305  Instruction *UI = cast<Instruction>(U);
5306  BasicBlock *UserBB = UI->getParent();
5307  if (UserBB == DefBB) continue;
5308  // Be conservative. We don't want this xform to end up introducing
5309  // reloads just before load / store instructions.
5310  if (isa<PHINode>(UI) || isa<LoadInst>(UI) || isa<StoreInst>(UI))
5311  return false;
5312  }
5313 
5314  // InsertedTruncs - Only insert one trunc in each block once.
5315  DenseMap<BasicBlock*, Instruction*> InsertedTruncs;
5316 
5317  bool MadeChange = false;
5318  for (Use &U : Src->uses()) {
5319  Instruction *User = cast<Instruction>(U.getUser());
5320 
5321  // Figure out which BB this ext is used in.
5322  BasicBlock *UserBB = User->getParent();
5323  if (UserBB == DefBB) continue;
5324 
5325  // Both src and def are live in this block. Rewrite the use.
5326  Instruction *&InsertedTrunc = InsertedTruncs[UserBB];
5327 
5328  if (!InsertedTrunc) {
5329  BasicBlock::iterator InsertPt = UserBB->getFirstInsertionPt();
5330  assert(InsertPt != UserBB->end());
5331  InsertedTrunc = new TruncInst(I, Src->getType(), "", &*InsertPt);
5332  InsertedInsts.insert(InsertedTrunc);
5333  }
5334 
5335  // Replace a use of the {s|z}ext source with a use of the result.
5336  U = InsertedTrunc;
5337  ++NumExtUses;
5338  MadeChange = true;
5339  }
5340 
5341  return MadeChange;
5342 }
5343 
5344 // Find loads whose uses only use some of the loaded value's bits. Add an "and"
5345 // just after the load if the target can fold this into one extload instruction,
5346 // with the hope of eliminating some of the other later "and" instructions using
5347 // the loaded value. "and"s that are made trivially redundant by the insertion
5348 // of the new "and" are removed by this function, while others (e.g. those whose
5349 // path from the load goes through a phi) are left for isel to potentially
5350 // remove.
5351 //
5352 // For example:
5353 //
5354 // b0:
5355 // x = load i32
5356 // ...
5357 // b1:
5358 // y = and x, 0xff
5359 // z = use y
5360 //
5361 // becomes:
5362 //
5363 // b0:
5364 // x = load i32
5365 // x' = and x, 0xff
5366 // ...
5367 // b1:
5368 // z = use x'
5369 //
5370 // whereas:
5371 //
5372 // b0:
5373 // x1 = load i32
5374 // ...
5375 // b1:
5376 // x2 = load i32
5377 // ...
5378 // b2:
5379 // x = phi x1, x2
5380 // y = and x, 0xff
5381 //
5382 // becomes (after a call to optimizeLoadExt for each load):
5383 //
5384 // b0:
5385 // x1 = load i32
5386 // x1' = and x1, 0xff
5387 // ...
5388 // b1:
5389 // x2 = load i32
5390 // x2' = and x2, 0xff
5391 // ...
5392 // b2:
5393 // x = phi x1', x2'
5394 // y = and x, 0xff
5395 bool CodeGenPrepare::optimizeLoadExt(LoadInst *Load) {
5396  if (!Load->isSimple() || !Load->getType()->isIntOrPtrTy())
5397  return false;
5398 
5399  // Skip loads we've already transformed.
5400  if (Load->hasOneUse() &&
5401  InsertedInsts.count(cast<Instruction>(*Load->user_begin())))
5402  return false;
5403 
5404  // Look at all uses of Load, looking through phis, to determine how many bits
5405  // of the loaded value are needed.
5408  SmallVector<Instruction *, 8> AndsToMaybeRemove;
5409  for (auto *U : Load->users())
5410  WorkList.push_back(cast<Instruction>(U));
5411 
5412  EVT LoadResultVT = TLI->getValueType(*DL, Load->getType());
5413  unsigned BitWidth = LoadResultVT.getSizeInBits();
5414  APInt DemandBits(BitWidth, 0);
5415  APInt WidestAndBits(BitWidth, 0);
5416 
5417  while (!WorkList.empty()) {
5418  Instruction *I = WorkList.back();
5419  WorkList.pop_back();
5420 
5421  // Break use-def graph loops.
5422  if (!Visited.insert(I).second)
5423  continue;
5424 
5425  // For a PHI node, push all of its users.
5426  if (auto *Phi = dyn_cast<PHINode>(I)) {
5427  for (auto *U : Phi->users())
5428  WorkList.push_back(cast<Instruction>(U));
5429  continue;
5430  }
5431 
5432  switch (I->getOpcode()) {
5433  case Instruction::And: {
5434  auto *AndC = dyn_cast<ConstantInt>(I->getOperand(1));
5435  if (!AndC)
5436  return false;
5437  APInt AndBits = AndC->getValue();
5438  DemandBits |= AndBits;
5439  // Keep track of the widest and mask we see.
5440  if (AndBits.ugt(WidestAndBits))
5441  WidestAndBits = AndBits;
5442  if (AndBits == WidestAndBits && I->getOperand(0) == Load)
5443  AndsToMaybeRemove.push_back(I);
5444  break;
5445  }
5446 
5447  case Instruction::Shl: {
5448  auto *ShlC = dyn_cast<ConstantInt>(I->getOperand(1));
5449  if (!ShlC)
5450  return false;
5451  uint64_t ShiftAmt = ShlC->getLimitedValue(BitWidth - 1);
5452  DemandBits.setLowBits(BitWidth - ShiftAmt);
5453  break;
5454  }
5455 
5456  case Instruction::Trunc: {
5457  EVT TruncVT = TLI->getValueType(*DL, I->getType());
5458  unsigned TruncBitWidth = TruncVT.getSizeInBits();
5459  DemandBits.setLowBits(TruncBitWidth);
5460  break;
5461  }
5462 
5463  default:
5464  return false;
5465  }
5466  }
5467 
5468  uint32_t ActiveBits = DemandBits.getActiveBits();
5469  // Avoid hoisting (and (load x) 1) since it is unlikely to be folded by the
5470  // target even if isLoadExtLegal says an i1 EXTLOAD is valid. For example,
5471  // for the AArch64 target isLoadExtLegal(ZEXTLOAD, i32, i1) returns true, but
5472  // (and (load x) 1) is not matched as a single instruction, rather as a LDR
5473  // followed by an AND.
5474  // TODO: Look into removing this restriction by fixing backends to either
5475  // return false for isLoadExtLegal for i1 or have them select this pattern to
5476  // a single instruction.
5477  //
5478  // Also avoid hoisting if we didn't see any ands with the exact DemandBits
5479  // mask, since these are the only ands that will be removed by isel.
5480  if (ActiveBits <= 1 || !DemandBits.isMask(ActiveBits) ||
5481  WidestAndBits != DemandBits)
5482  return false;
5483 
5484  LLVMContext &Ctx = Load->getType()->getContext();
5485  Type *TruncTy = Type::getIntNTy(Ctx, ActiveBits);
5486  EVT TruncVT = TLI->getValueType(*DL, TruncTy);
5487 
5488  // Reject cases that won't be matched as extloads.
5489  if (!LoadResultVT.bitsGT(TruncVT) || !TruncVT.isRound() ||
5490  !TLI->isLoadExtLegal(ISD::ZEXTLOAD, LoadResultVT, TruncVT))
5491  return false;
5492 
5493  IRBuilder<> Builder(Load->getNextNode());
5494  auto *NewAnd = dyn_cast<Instruction>(
5495  Builder.CreateAnd(Load, ConstantInt::get(Ctx, DemandBits)));
5496  // Mark this instruction as "inserted by CGP", so that other
5497  // optimizations don't touch it.
5498  InsertedInsts.insert(NewAnd);
5499 
5500  // Replace all uses of load with new and (except for the use of load in the
5501  // new and itself).
5502  Load->replaceAllUsesWith(NewAnd);
5503  NewAnd->setOperand(0, Load);
5504 
5505  // Remove any and instructions that are now redundant.
5506  for (auto *And : AndsToMaybeRemove)
5507  // Check that the and mask is the same as the one we decided to put on the
5508  // new and.
5509  if (cast<ConstantInt>(And->getOperand(1))->getValue() == DemandBits) {
5510  And->replaceAllUsesWith(NewAnd);
5511  if (&*CurInstIterator == And)
5512  CurInstIterator = std::next(And->getIterator());
5513  And->eraseFromParent();
5514  ++NumAndUses;
5515  }
5516 
5517  ++NumAndsAdded;
5518  return true;
5519 }
5520 
5521 /// Check if V (an operand of a select instruction) is an expensive instruction
5522 /// that is only used once.
5523 static bool sinkSelectOperand(const TargetTransformInfo *TTI, Value *V) {
5524  auto *I = dyn_cast<Instruction>(V);
5525  // If it's safe to speculatively execute, then it should not have side
5526  // effects; therefore, it's safe to sink and possibly *not* execute.
5527  return I && I->hasOneUse() && isSafeToSpeculativelyExecute(I) &&
5529 }
5530 
5531 /// Returns true if a SelectInst should be turned into an explicit branch.
5533  const TargetLowering *TLI,
5534  SelectInst *SI) {
5535  // If even a predictable select is cheap, then a branch can't be cheaper.
5536  if (!TLI->isPredictableSelectExpensive())
5537  return false;
5538 
5539  // FIXME: This should use the same heuristics as IfConversion to determine
5540  // whether a select is better represented as a branch.
5541 
5542  // If metadata tells us that the select condition is obviously predictable,
5543  // then we want to replace the select with a branch.
5544  uint64_t TrueWeight, FalseWeight;
5545  if (SI->extractProfMetadata(TrueWeight, FalseWeight)) {
5546  uint64_t Max = std::max(TrueWeight, FalseWeight);
5547  uint64_t Sum = TrueWeight + FalseWeight;
5548  if (Sum != 0) {
5549  auto Probability = BranchProbability::getBranchProbability(Max, Sum);
5550  if (Probability > TLI->getPredictableBranchThreshold())
5551  return true;
5552  }
5553  }
5554 
5555  CmpInst *Cmp = dyn_cast<CmpInst>(SI->getCondition());
5556 
5557  // If a branch is predictable, an out-of-order CPU can avoid blocking on its
5558  // comparison condition. If the compare has more than one use, there's
5559  // probably another cmov or setcc around, so it's not worth emitting a branch.
5560  if (!Cmp || !Cmp->hasOneUse())
5561  return false;
5562 
5563  // If either operand of the select is expensive and only needed on one side
5564  // of the select, we should form a branch.
5565  if (sinkSelectOperand(TTI, SI->getTrueValue()) ||
5566  sinkSelectOperand(TTI, SI->getFalseValue()))
5567  return true;
5568 
5569  return false;
5570 }
5571 
5572 /// If \p isTrue is true, return the true value of \p SI, otherwise return
5573 /// false value of \p SI. If the true/false value of \p SI is defined by any
5574 /// select instructions in \p Selects, look through the defining select
5575 /// instruction until the true/false value is not defined in \p Selects.
5577  SelectInst *SI, bool isTrue,
5578  const SmallPtrSet<const Instruction *, 2> &Selects) {
5579  Value *V;
5580 
5581  for (SelectInst *DefSI = SI; DefSI != nullptr && Selects.count(DefSI);
5582  DefSI = dyn_cast<SelectInst>(V)) {
5583  assert(DefSI->getCondition() == SI->getCondition() &&
5584  "The condition of DefSI does not match with SI");
5585  V = (isTrue ? DefSI->getTrueValue() : DefSI->getFalseValue());
5586  }
5587  return V;
5588 }
5589 
5590 /// If we have a SelectInst that will likely profit from branch prediction,
5591 /// turn it into a branch.
5592 bool CodeGenPrepare::optimizeSelectInst(SelectInst *SI) {
5593  // If branch conversion isn't desirable, exit early.
5594  if (DisableSelectToBranch || OptSize || !TLI)
5595  return false;
5596 
5597  // Find all consecutive select instructions that share the same condition.
5599  ASI.push_back(SI);
5601  It != SI->getParent()->end(); ++It) {
5602  SelectInst *I = dyn_cast<SelectInst>(&*It);
5603  if (I && SI->getCondition() == I->getCondition()) {
5604  ASI.push_back(I);
5605  } else {
5606  break;
5607  }
5608  }
5609 
5610  SelectInst *LastSI = ASI.back();
5611  // Increment the current iterator to skip all the rest of select instructions
5612  // because they will be either "not lowered" or "all lowered" to branch.
5613  CurInstIterator = std::next(LastSI->getIterator());
5614 
5615  bool VectorCond = !SI->getCondition()->getType()->isIntegerTy(1);
5616 
5617  // Can we convert the 'select' to CF ?
5618  if (VectorCond || SI->getMetadata(LLVMContext::MD_unpredictable))
5619  return false;
5620 
5622  if (VectorCond)
5623  SelectKind = TargetLowering::VectorMaskSelect;
5624  else if (SI->getType()->isVectorTy())
5626  else
5627  SelectKind = TargetLowering::ScalarValSelect;
5628 
5629  if (TLI->isSelectSupported(SelectKind) &&
5630  !isFormingBranchFromSelectProfitable(TTI, TLI, SI))
5631  return false;
5632 
5633  ModifiedDT = true;
5634 
5635  // Transform a sequence like this:
5636  // start:
5637  // %cmp = cmp uge i32 %a, %b
5638  // %sel = select i1 %cmp, i32 %c, i32 %d
5639  //
5640  // Into:
5641  // start:
5642  // %cmp = cmp uge i32 %a, %b
5643  // br i1 %cmp, label %select.true, label %select.false
5644  // select.true:
5645  // br label %select.end
5646  // select.false:
5647  // br label %select.end
5648  // select.end:
5649  // %sel = phi i32 [ %c, %select.true ], [ %d, %select.false ]
5650  //
5651  // In addition, we may sink instructions that produce %c or %d from
5652  // the entry block into the destination(s) of the new branch.
5653  // If the true or false blocks do not contain a sunken instruction, that
5654  // block and its branch may be optimized away. In that case, one side of the
5655  // first branch will point directly to select.end, and the corresponding PHI
5656  // predecessor block will be the start block.
5657 
5658  // First, we split the block containing the select into 2 blocks.
5659  BasicBlock *StartBlock = SI->getParent();
5660  BasicBlock::iterator SplitPt = ++(BasicBlock::iterator(LastSI));
5661  BasicBlock *EndBlock = StartBlock->splitBasicBlock(SplitPt, "select.end");
5662 
5663  // Delete the unconditional branch that was just created by the split.
5664  StartBlock->getTerminator()->eraseFromParent();
5665 
5666  // These are the new basic blocks for the conditional branch.
5667  // At least one will become an actual new basic block.
5668  BasicBlock *TrueBlock = nullptr;
5669  BasicBlock *FalseBlock = nullptr;
5670  BranchInst *TrueBranch = nullptr;
5671  BranchInst *FalseBranch = nullptr;
5672 
5673  // Sink expensive instructions into the conditional blocks to avoid executing
5674  // them speculatively.
5675  for (SelectInst *SI : ASI) {
5676  if (sinkSelectOperand(TTI, SI->getTrueValue())) {
5677  if (TrueBlock == nullptr) {
5678  TrueBlock = BasicBlock::Create(SI->getContext(), "select.true.sink",
5679  EndBlock->getParent(), EndBlock);
5680  TrueBranch = BranchInst::Create(EndBlock, TrueBlock);
5681  TrueBranch->setDebugLoc(SI->getDebugLoc());
5682  }
5683  auto *TrueInst = cast<Instruction>(SI->getTrueValue());
5684  TrueInst->moveBefore(TrueBranch);
5685  }
5686  if (sinkSelectOperand(TTI, SI->getFalseValue())) {
5687  if (FalseBlock == nullptr) {
5688  FalseBlock = BasicBlock::Create(SI->getContext(), "select.false.sink",
5689  EndBlock->getParent(), EndBlock);
5690  FalseBranch = BranchInst::Create(EndBlock, FalseBlock);
5691  FalseBranch->setDebugLoc(SI->getDebugLoc());
5692  }
5693  auto *FalseInst = cast<Instruction>(SI->getFalseValue());
5694  FalseInst->moveBefore(FalseBranch);
5695  }
5696  }
5697 
5698  // If there was nothing to sink, then arbitrarily choose the 'false' side
5699  // for a new input value to the PHI.
5700  if (TrueBlock == FalseBlock) {
5701  assert(TrueBlock == nullptr &&
5702  "Unexpected basic block transform while optimizing select");
5703 
5704  FalseBlock = BasicBlock::Create(SI->getContext(), "select.false",
5705  EndBlock->getParent(), EndBlock);
5706  auto *FalseBranch = BranchInst::Create(EndBlock, FalseBlock);
5707  FalseBranch->setDebugLoc(SI->getDebugLoc());
5708  }
5709 
5710  // Insert the real conditional branch based on the original condition.
5711  // If we did not create a new block for one of the 'true' or 'false' paths
5712  // of the condition, it means that side of the branch goes to the end block
5713  // directly and the path originates from the start block from the point of
5714  // view of the new PHI.
5715  BasicBlock *TT, *FT;
5716  if (TrueBlock == nullptr) {
5717  TT = EndBlock;
5718  FT = FalseBlock;
5719  TrueBlock = StartBlock;
5720  } else if (FalseBlock == nullptr) {
5721  TT = TrueBlock;
5722  FT = EndBlock;
5723  FalseBlock = StartBlock;
5724  } else {
5725  TT = TrueBlock;
5726  FT = FalseBlock;
5727  }
5728  IRBuilder<>(SI).CreateCondBr(SI->getCondition(), TT, FT, SI);
5729 
5731  INS.insert(ASI.begin(), ASI.end());
5732  // Use reverse iterator because later select may use the value of the
5733  // earlier select, and we need to propagate value through earlier select
5734  // to get the PHI operand.
5735  for (auto It = ASI.rbegin(); It != ASI.rend(); ++It) {
5736  SelectInst *SI = *It;
5737  // The select itself is replaced with a PHI Node.
5738  PHINode *PN = PHINode::Create(SI->getType(), 2, "", &EndBlock->front());
5739  PN->takeName(SI);
5740  PN->addIncoming(getTrueOrFalseValue(SI, true, INS), TrueBlock);
5741  PN->addIncoming(getTrueOrFalseValue(SI, false, INS), FalseBlock);
5742  PN->setDebugLoc(SI->getDebugLoc());
5743 
5744  SI->replaceAllUsesWith(PN);
5745  SI->eraseFromParent();
5746  INS.erase(SI);
5747  ++NumSelectsExpanded;
5748  }
5749 
5750  // Instruct OptimizeBlock to skip to the next block.
5751  CurInstIterator = StartBlock->end();
5752  return true;
5753 }
5754 
5757  int SplatElem = -1;
5758  for (unsigned i = 0; i < Mask.size(); ++i) {
5759  if (SplatElem != -1 && Mask[i] != -1 && Mask[i] != SplatElem)
5760  return false;
5761  SplatElem = Mask[i];
5762  }
5763 
5764  return true;
5765 }
5766 
5767 /// Some targets have expensive vector shifts if the lanes aren't all the same
5768 /// (e.g. x86 only introduced "vpsllvd" and friends with AVX2). In these cases
5769 /// it's often worth sinking a shufflevector splat down to its use so that
5770 /// codegen can spot all lanes are identical.
5771 bool CodeGenPrepare::optimizeShuffleVectorInst(ShuffleVectorInst *SVI) {
5772  BasicBlock *DefBB = SVI->getParent();
5773 
5774  // Only do this xform if variable vector shifts are particularly expensive.
5775  if (!TLI || !TLI->isVectorShiftByScalarCheap(SVI->getType()))
5776  return false;
5777 
5778  // We only expect better codegen by sinking a shuffle if we can recognise a
5779  // constant splat.
5780  if (!isBroadcastShuffle(SVI))
5781  return false;
5782 
5783  // InsertedShuffles - Only insert a shuffle in each block once.
5784  DenseMap<BasicBlock*, Instruction*> InsertedShuffles;
5785 
5786  bool MadeChange = false;
5787  for (User *U : SVI->users()) {
5788  Instruction *UI = cast<Instruction>(U);
5789 
5790  // Figure out which BB this ext is used in.
5791  BasicBlock *UserBB = UI->getParent();
5792  if (UserBB == DefBB) continue;
5793 
5794  // For now only apply this when the splat is used by a shift instruction.
5795  if (!UI->isShift()) continue;
5796 
5797  // Everything checks out, sink the shuffle if the user's block doesn't
5798  // already have a copy.
5799