LLVM  7.0.0svn
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 using SetOfInstrs = SmallPtrSet<Instruction *, 16>;
227 using TypeIsSExt = PointerIntPair<Type *, 1, bool>;
228 using InstrToOrigTy = DenseMap<Instruction *, TypeIsSExt>;
229 using SExts = SmallVector<Instruction *, 16>;
230 using ValueToSExts = DenseMap<Value *, SExts>;
231 
232 class TypePromotionTransaction;
233 
234  class CodeGenPrepare : public FunctionPass {
235  const TargetMachine *TM = nullptr;
236  const TargetSubtargetInfo *SubtargetInfo;
237  const TargetLowering *TLI = nullptr;
238  const TargetRegisterInfo *TRI;
239  const TargetTransformInfo *TTI = nullptr;
240  const TargetLibraryInfo *TLInfo;
241  const LoopInfo *LI;
242  std::unique_ptr<BlockFrequencyInfo> BFI;
243  std::unique_ptr<BranchProbabilityInfo> BPI;
244 
245  /// As we scan instructions optimizing them, this is the next instruction
246  /// to optimize. Transforms that can invalidate this should update it.
247  BasicBlock::iterator CurInstIterator;
248 
249  /// Keeps track of non-local addresses that have been sunk into a block.
250  /// This allows us to avoid inserting duplicate code for blocks with
251  /// multiple load/stores of the same address. The usage of WeakTrackingVH
252  /// enables SunkAddrs to be treated as a cache whose entries can be
253  /// invalidated if a sunken address computation has been erased.
255 
256  /// Keeps track of all instructions inserted for the current function.
257  SetOfInstrs InsertedInsts;
258 
259  /// Keeps track of the type of the related instruction before their
260  /// promotion for the current function.
261  InstrToOrigTy PromotedInsts;
262 
263  /// Keep track of instructions removed during promotion.
264  SetOfInstrs RemovedInsts;
265 
266  /// Keep track of sext chains based on their initial value.
267  DenseMap<Value *, Instruction *> SeenChainsForSExt;
268 
269  /// Keep track of GEPs accessing the same data structures such as structs or
270  /// arrays that are candidates to be split later because of their large
271  /// size.
272  DenseMap<
275  LargeOffsetGEPMap;
276 
277  /// Keep track of new GEP base after splitting the GEPs having large offset.
278  SmallSet<AssertingVH<Value>, 2> NewGEPBases;
279 
280  /// Map serial numbers to Large offset GEPs.
281  DenseMap<AssertingVH<GetElementPtrInst>, int> LargeOffsetGEPID;
282 
283  /// Keep track of SExt promoted.
284  ValueToSExts ValToSExtendedUses;
285 
286  /// True if CFG is modified in any way.
287  bool ModifiedDT;
288 
289  /// True if optimizing for size.
290  bool OptSize;
291 
292  /// DataLayout for the Function being processed.
293  const DataLayout *DL = nullptr;
294 
295  public:
296  static char ID; // Pass identification, replacement for typeid
297 
298  CodeGenPrepare() : FunctionPass(ID) {
300  }
301 
302  bool runOnFunction(Function &F) override;
303 
304  StringRef getPassName() const override { return "CodeGen Prepare"; }
305 
306  void getAnalysisUsage(AnalysisUsage &AU) const override {
307  // FIXME: When we can selectively preserve passes, preserve the domtree.
312  }
313 
314  private:
315  bool eliminateFallThrough(Function &F);
316  bool eliminateMostlyEmptyBlocks(Function &F);
317  BasicBlock *findDestBlockOfMergeableEmptyBlock(BasicBlock *BB);
318  bool canMergeBlocks(const BasicBlock *BB, const BasicBlock *DestBB) const;
319  void eliminateMostlyEmptyBlock(BasicBlock *BB);
320  bool isMergingEmptyBlockProfitable(BasicBlock *BB, BasicBlock *DestBB,
321  bool isPreheader);
322  bool optimizeBlock(BasicBlock &BB, bool &ModifiedDT);
323  bool optimizeInst(Instruction *I, bool &ModifiedDT);
324  bool optimizeMemoryInst(Instruction *MemoryInst, Value *Addr,
325  Type *AccessTy, unsigned AddrSpace);
326  bool optimizeInlineAsmInst(CallInst *CS);
327  bool optimizeCallInst(CallInst *CI, bool &ModifiedDT);
328  bool optimizeExt(Instruction *&I);
329  bool optimizeExtUses(Instruction *I);
330  bool optimizeLoadExt(LoadInst *Load);
331  bool optimizeSelectInst(SelectInst *SI);
332  bool optimizeShuffleVectorInst(ShuffleVectorInst *SVI);
333  bool optimizeSwitchInst(SwitchInst *SI);
334  bool optimizeExtractElementInst(Instruction *Inst);
335  bool dupRetToEnableTailCallOpts(BasicBlock *BB);
336  bool placeDbgValues(Function &F);
337  bool canFormExtLd(const SmallVectorImpl<Instruction *> &MovedExts,
338  LoadInst *&LI, Instruction *&Inst, bool HasPromoted);
339  bool tryToPromoteExts(TypePromotionTransaction &TPT,
340  const SmallVectorImpl<Instruction *> &Exts,
341  SmallVectorImpl<Instruction *> &ProfitablyMovedExts,
342  unsigned CreatedInstsCost = 0);
343  bool mergeSExts(Function &F);
344  bool splitLargeGEPOffsets();
345  bool performAddressTypePromotion(
346  Instruction *&Inst,
347  bool AllowPromotionWithoutCommonHeader,
348  bool HasPromoted, TypePromotionTransaction &TPT,
349  SmallVectorImpl<Instruction *> &SpeculativelyMovedExts);
350  bool splitBranchCondition(Function &F);
351  bool simplifyOffsetableRelocate(Instruction &I);
352  };
353 
354 } // end anonymous namespace
355 
356 char CodeGenPrepare::ID = 0;
357 
358 INITIALIZE_PASS_BEGIN(CodeGenPrepare, DEBUG_TYPE,
359  "Optimize for code generation", false, false)
362  "Optimize for code generation", false, false)
363 
364 FunctionPass *llvm::createCodeGenPreparePass() { return new CodeGenPrepare(); }
365 
367  if (skipFunction(F))
368  return false;
369 
370  DL = &F.getParent()->getDataLayout();
371 
372  bool EverMadeChange = false;
373  // Clear per function information.
374  InsertedInsts.clear();
375  PromotedInsts.clear();
376 
377  ModifiedDT = false;
378  if (auto *TPC = getAnalysisIfAvailable<TargetPassConfig>()) {
379  TM = &TPC->getTM<TargetMachine>();
380  SubtargetInfo = TM->getSubtargetImpl(F);
381  TLI = SubtargetInfo->getTargetLowering();
382  TRI = SubtargetInfo->getRegisterInfo();
383  }
384  TLInfo = &getAnalysis<TargetLibraryInfoWrapperPass>().getTLI();
385  TTI = &getAnalysis<TargetTransformInfoWrapperPass>().getTTI(F);
386  LI = &getAnalysis<LoopInfoWrapperPass>().getLoopInfo();
387  BPI.reset(new BranchProbabilityInfo(F, *LI));
388  BFI.reset(new BlockFrequencyInfo(F, *BPI, *LI));
389  OptSize = F.optForSize();
390 
391  ProfileSummaryInfo *PSI =
392  getAnalysis<ProfileSummaryInfoWrapperPass>().getPSI();
394  if (PSI->isFunctionHotInCallGraph(&F, *BFI))
395  F.setSectionPrefix(".hot");
396  else if (PSI->isFunctionColdInCallGraph(&F, *BFI))
397  F.setSectionPrefix(".unlikely");
398  }
399 
400  /// This optimization identifies DIV instructions that can be
401  /// profitably bypassed and carried out with a shorter, faster divide.
402  if (!OptSize && !PSI->hasHugeWorkingSetSize() && TLI &&
403  TLI->isSlowDivBypassed()) {
404  const DenseMap<unsigned int, unsigned int> &BypassWidths =
405  TLI->getBypassSlowDivWidths();
406  BasicBlock* BB = &*F.begin();
407  while (BB != nullptr) {
408  // bypassSlowDivision may create new BBs, but we don't want to reapply the
409  // optimization to those blocks.
410  BasicBlock* Next = BB->getNextNode();
411  EverMadeChange |= bypassSlowDivision(BB, BypassWidths);
412  BB = Next;
413  }
414  }
415 
416  // Eliminate blocks that contain only PHI nodes and an
417  // unconditional branch.
418  EverMadeChange |= eliminateMostlyEmptyBlocks(F);
419 
420  // llvm.dbg.value is far away from the value then iSel may not be able
421  // handle it properly. iSel will drop llvm.dbg.value if it can not
422  // find a node corresponding to the value.
423  EverMadeChange |= placeDbgValues(F);
424 
425  if (!DisableBranchOpts)
426  EverMadeChange |= splitBranchCondition(F);
427 
428  // Split some critical edges where one of the sources is an indirect branch,
429  // to help generate sane code for PHIs involving such edges.
430  EverMadeChange |= SplitIndirectBrCriticalEdges(F);
431 
432  bool MadeChange = true;
433  while (MadeChange) {
434  MadeChange = false;
435  SeenChainsForSExt.clear();
436  ValToSExtendedUses.clear();
437  RemovedInsts.clear();
438  LargeOffsetGEPMap.clear();
439  LargeOffsetGEPID.clear();
440  for (Function::iterator I = F.begin(); I != F.end(); ) {
441  BasicBlock *BB = &*I++;
442  bool ModifiedDTOnIteration = false;
443  MadeChange |= optimizeBlock(*BB, ModifiedDTOnIteration);
444 
445  // Restart BB iteration if the dominator tree of the Function was changed
446  if (ModifiedDTOnIteration)
447  break;
448  }
449  if (EnableTypePromotionMerge && !ValToSExtendedUses.empty())
450  MadeChange |= mergeSExts(F);
451  if (!LargeOffsetGEPMap.empty())
452  MadeChange |= splitLargeGEPOffsets();
453 
454  // Really free removed instructions during promotion.
455  for (Instruction *I : RemovedInsts)
456  I->deleteValue();
457 
458  EverMadeChange |= MadeChange;
459  }
460 
461  SunkAddrs.clear();
462 
463  if (!DisableBranchOpts) {
464  MadeChange = false;
465  // Use a set vector to get deterministic iteration order. The order the
466  // blocks are removed may affect whether or not PHI nodes in successors
467  // are removed.
469  for (BasicBlock &BB : F) {
470  SmallVector<BasicBlock *, 2> Successors(succ_begin(&BB), succ_end(&BB));
471  MadeChange |= ConstantFoldTerminator(&BB, true);
472  if (!MadeChange) continue;
473 
475  II = Successors.begin(), IE = Successors.end(); II != IE; ++II)
476  if (pred_begin(*II) == pred_end(*II))
477  WorkList.insert(*II);
478  }
479 
480  // Delete the dead blocks and any of their dead successors.
481  MadeChange |= !WorkList.empty();
482  while (!WorkList.empty()) {
483  BasicBlock *BB = WorkList.pop_back_val();
484  SmallVector<BasicBlock*, 2> Successors(succ_begin(BB), succ_end(BB));
485 
486  DeleteDeadBlock(BB);
487 
489  II = Successors.begin(), IE = Successors.end(); II != IE; ++II)
490  if (pred_begin(*II) == pred_end(*II))
491  WorkList.insert(*II);
492  }
493 
494  // Merge pairs of basic blocks with unconditional branches, connected by
495  // a single edge.
496  if (EverMadeChange || MadeChange)
497  MadeChange |= eliminateFallThrough(F);
498 
499  EverMadeChange |= MadeChange;
500  }
501 
502  if (!DisableGCOpts) {
503  SmallVector<Instruction *, 2> Statepoints;
504  for (BasicBlock &BB : F)
505  for (Instruction &I : BB)
506  if (isStatepoint(I))
507  Statepoints.push_back(&I);
508  for (auto &I : Statepoints)
509  EverMadeChange |= simplifyOffsetableRelocate(*I);
510  }
511 
512  return EverMadeChange;
513 }
514 
515 /// Merge basic blocks which are connected by a single edge, where one of the
516 /// basic blocks has a single successor pointing to the other basic block,
517 /// which has a single predecessor.
518 bool CodeGenPrepare::eliminateFallThrough(Function &F) {
519  bool Changed = false;
520  // Scan all of the blocks in the function, except for the entry block.
521  // Use a temporary array to avoid iterator being invalidated when
522  // deleting blocks.
524  for (auto &Block : llvm::make_range(std::next(F.begin()), F.end()))
525  Blocks.push_back(&Block);
526 
527  for (auto &Block : Blocks) {
528  auto *BB = cast_or_null<BasicBlock>(Block);
529  if (!BB)
530  continue;
531  // If the destination block has a single pred, then this is a trivial
532  // edge, just collapse it.
533  BasicBlock *SinglePred = BB->getSinglePredecessor();
534 
535  // Don't merge if BB's address is taken.
536  if (!SinglePred || SinglePred == BB || BB->hasAddressTaken()) continue;
537 
538  BranchInst *Term = dyn_cast<BranchInst>(SinglePred->getTerminator());
539  if (Term && !Term->isConditional()) {
540  Changed = true;
541  LLVM_DEBUG(dbgs() << "To merge:\n" << *BB << "\n\n\n");
542 
543  // Merge BB into SinglePred and delete it.
545  }
546  }
547  return Changed;
548 }
549 
550 /// Find a destination block from BB if BB is mergeable empty block.
551 BasicBlock *CodeGenPrepare::findDestBlockOfMergeableEmptyBlock(BasicBlock *BB) {
552  // If this block doesn't end with an uncond branch, ignore it.
554  if (!BI || !BI->isUnconditional())
555  return nullptr;
556 
557  // If the instruction before the branch (skipping debug info) isn't a phi
558  // node, then other stuff is happening here.
559  BasicBlock::iterator BBI = BI->getIterator();
560  if (BBI != BB->begin()) {
561  --BBI;
562  while (isa<DbgInfoIntrinsic>(BBI)) {
563  if (BBI == BB->begin())
564  break;
565  --BBI;
566  }
567  if (!isa<DbgInfoIntrinsic>(BBI) && !isa<PHINode>(BBI))
568  return nullptr;
569  }
570 
571  // Do not break infinite loops.
572  BasicBlock *DestBB = BI->getSuccessor(0);
573  if (DestBB == BB)
574  return nullptr;
575 
576  if (!canMergeBlocks(BB, DestBB))
577  DestBB = nullptr;
578 
579  return DestBB;
580 }
581 
582 /// Eliminate blocks that contain only PHI nodes, debug info directives, and an
583 /// unconditional branch. Passes before isel (e.g. LSR/loopsimplify) often split
584 /// edges in ways that are non-optimal for isel. Start by eliminating these
585 /// blocks so we can split them the way we want them.
586 bool CodeGenPrepare::eliminateMostlyEmptyBlocks(Function &F) {
588  SmallVector<Loop *, 16> LoopList(LI->begin(), LI->end());
589  while (!LoopList.empty()) {
590  Loop *L = LoopList.pop_back_val();
591  LoopList.insert(LoopList.end(), L->begin(), L->end());
592  if (BasicBlock *Preheader = L->getLoopPreheader())
593  Preheaders.insert(Preheader);
594  }
595 
596  bool MadeChange = false;
597  // Copy blocks into a temporary array to avoid iterator invalidation issues
598  // as we remove them.
599  // Note that this intentionally skips the entry block.
601  for (auto &Block : llvm::make_range(std::next(F.begin()), F.end()))
602  Blocks.push_back(&Block);
603 
604  for (auto &Block : Blocks) {
605  BasicBlock *BB = cast_or_null<BasicBlock>(Block);
606  if (!BB)
607  continue;
608  BasicBlock *DestBB = findDestBlockOfMergeableEmptyBlock(BB);
609  if (!DestBB ||
610  !isMergingEmptyBlockProfitable(BB, DestBB, Preheaders.count(BB)))
611  continue;
612 
613  eliminateMostlyEmptyBlock(BB);
614  MadeChange = true;
615  }
616  return MadeChange;
617 }
618 
619 bool CodeGenPrepare::isMergingEmptyBlockProfitable(BasicBlock *BB,
620  BasicBlock *DestBB,
621  bool isPreheader) {
622  // Do not delete loop preheaders if doing so would create a critical edge.
623  // Loop preheaders can be good locations to spill registers. If the
624  // preheader is deleted and we create a critical edge, registers may be
625  // spilled in the loop body instead.
626  if (!DisablePreheaderProtect && isPreheader &&
627  !(BB->getSinglePredecessor() &&
629  return false;
630 
631  // Try to skip merging if the unique predecessor of BB is terminated by a
632  // switch or indirect branch instruction, and BB is used as an incoming block
633  // of PHIs in DestBB. In such case, merging BB and DestBB would cause ISel to
634  // add COPY instructions in the predecessor of BB instead of BB (if it is not
635  // merged). Note that the critical edge created by merging such blocks wont be
636  // split in MachineSink because the jump table is not analyzable. By keeping
637  // such empty block (BB), ISel will place COPY instructions in BB, not in the
638  // predecessor of BB.
639  BasicBlock *Pred = BB->getUniquePredecessor();
640  if (!Pred ||
641  !(isa<SwitchInst>(Pred->getTerminator()) ||
642  isa<IndirectBrInst>(Pred->getTerminator())))
643  return true;
644 
645  if (BB->getTerminator() != BB->getFirstNonPHI())
646  return true;
647 
648  // We use a simple cost heuristic which determine skipping merging is
649  // profitable if the cost of skipping merging is less than the cost of
650  // merging : Cost(skipping merging) < Cost(merging BB), where the
651  // Cost(skipping merging) is Freq(BB) * (Cost(Copy) + Cost(Branch)), and
652  // the Cost(merging BB) is Freq(Pred) * Cost(Copy).
653  // Assuming Cost(Copy) == Cost(Branch), we could simplify it to :
654  // Freq(Pred) / Freq(BB) > 2.
655  // Note that if there are multiple empty blocks sharing the same incoming
656  // value for the PHIs in the DestBB, we consider them together. In such
657  // case, Cost(merging BB) will be the sum of their frequencies.
658 
659  if (!isa<PHINode>(DestBB->begin()))
660  return true;
661 
662  SmallPtrSet<BasicBlock *, 16> SameIncomingValueBBs;
663 
664  // Find all other incoming blocks from which incoming values of all PHIs in
665  // DestBB are the same as the ones from BB.
666  for (pred_iterator PI = pred_begin(DestBB), E = pred_end(DestBB); PI != E;
667  ++PI) {
668  BasicBlock *DestBBPred = *PI;
669  if (DestBBPred == BB)
670  continue;
671 
672  if (llvm::all_of(DestBB->phis(), [&](const PHINode &DestPN) {
673  return DestPN.getIncomingValueForBlock(BB) ==
674  DestPN.getIncomingValueForBlock(DestBBPred);
675  }))
676  SameIncomingValueBBs.insert(DestBBPred);
677  }
678 
679  // See if all BB's incoming values are same as the value from Pred. In this
680  // case, no reason to skip merging because COPYs are expected to be place in
681  // Pred already.
682  if (SameIncomingValueBBs.count(Pred))
683  return true;
684 
685  BlockFrequency PredFreq = BFI->getBlockFreq(Pred);
686  BlockFrequency BBFreq = BFI->getBlockFreq(BB);
687 
688  for (auto SameValueBB : SameIncomingValueBBs)
689  if (SameValueBB->getUniquePredecessor() == Pred &&
690  DestBB == findDestBlockOfMergeableEmptyBlock(SameValueBB))
691  BBFreq += BFI->getBlockFreq(SameValueBB);
692 
693  return PredFreq.getFrequency() <=
695 }
696 
697 /// Return true if we can merge BB into DestBB if there is a single
698 /// unconditional branch between them, and BB contains no other non-phi
699 /// instructions.
700 bool CodeGenPrepare::canMergeBlocks(const BasicBlock *BB,
701  const BasicBlock *DestBB) const {
702  // We only want to eliminate blocks whose phi nodes are used by phi nodes in
703  // the successor. If there are more complex condition (e.g. preheaders),
704  // don't mess around with them.
705  for (const PHINode &PN : BB->phis()) {
706  for (const User *U : PN.users()) {
707  const Instruction *UI = cast<Instruction>(U);
708  if (UI->getParent() != DestBB || !isa<PHINode>(UI))
709  return false;
710  // If User is inside DestBB block and it is a PHINode then check
711  // incoming value. If incoming value is not from BB then this is
712  // a complex condition (e.g. preheaders) we want to avoid here.
713  if (UI->getParent() == DestBB) {
714  if (const PHINode *UPN = dyn_cast<PHINode>(UI))
715  for (unsigned I = 0, E = UPN->getNumIncomingValues(); I != E; ++I) {
716  Instruction *Insn = dyn_cast<Instruction>(UPN->getIncomingValue(I));
717  if (Insn && Insn->getParent() == BB &&
718  Insn->getParent() != UPN->getIncomingBlock(I))
719  return false;
720  }
721  }
722  }
723  }
724 
725  // If BB and DestBB contain any common predecessors, then the phi nodes in BB
726  // and DestBB may have conflicting incoming values for the block. If so, we
727  // can't merge the block.
728  const PHINode *DestBBPN = dyn_cast<PHINode>(DestBB->begin());
729  if (!DestBBPN) return true; // no conflict.
730 
731  // Collect the preds of BB.
733  if (const PHINode *BBPN = dyn_cast<PHINode>(BB->begin())) {
734  // It is faster to get preds from a PHI than with pred_iterator.
735  for (unsigned i = 0, e = BBPN->getNumIncomingValues(); i != e; ++i)
736  BBPreds.insert(BBPN->getIncomingBlock(i));
737  } else {
738  BBPreds.insert(pred_begin(BB), pred_end(BB));
739  }
740 
741  // Walk the preds of DestBB.
742  for (unsigned i = 0, e = DestBBPN->getNumIncomingValues(); i != e; ++i) {
743  BasicBlock *Pred = DestBBPN->getIncomingBlock(i);
744  if (BBPreds.count(Pred)) { // Common predecessor?
745  for (const PHINode &PN : DestBB->phis()) {
746  const Value *V1 = PN.getIncomingValueForBlock(Pred);
747  const Value *V2 = PN.getIncomingValueForBlock(BB);
748 
749  // If V2 is a phi node in BB, look up what the mapped value will be.
750  if (const PHINode *V2PN = dyn_cast<PHINode>(V2))
751  if (V2PN->getParent() == BB)
752  V2 = V2PN->getIncomingValueForBlock(Pred);
753 
754  // If there is a conflict, bail out.
755  if (V1 != V2) return false;
756  }
757  }
758  }
759 
760  return true;
761 }
762 
763 /// Eliminate a basic block that has only phi's and an unconditional branch in
764 /// it.
765 void CodeGenPrepare::eliminateMostlyEmptyBlock(BasicBlock *BB) {
766  BranchInst *BI = cast<BranchInst>(BB->getTerminator());
767  BasicBlock *DestBB = BI->getSuccessor(0);
768 
769  LLVM_DEBUG(dbgs() << "MERGING MOSTLY EMPTY BLOCKS - BEFORE:\n"
770  << *BB << *DestBB);
771 
772  // If the destination block has a single pred, then this is a trivial edge,
773  // just collapse it.
774  if (BasicBlock *SinglePred = DestBB->getSinglePredecessor()) {
775  if (SinglePred != DestBB) {
776  assert(SinglePred == BB &&
777  "Single predecessor not the same as predecessor");
778  // Merge DestBB into SinglePred/BB and delete it.
780  // Note: BB(=SinglePred) will not be deleted on this path.
781  // DestBB(=its single successor) is the one that was deleted.
782  LLVM_DEBUG(dbgs() << "AFTER:\n" << *SinglePred << "\n\n\n");
783  return;
784  }
785  }
786 
787  // Otherwise, we have multiple predecessors of BB. Update the PHIs in DestBB
788  // to handle the new incoming edges it is about to have.
789  for (PHINode &PN : DestBB->phis()) {
790  // Remove the incoming value for BB, and remember it.
791  Value *InVal = PN.removeIncomingValue(BB, false);
792 
793  // Two options: either the InVal is a phi node defined in BB or it is some
794  // value that dominates BB.
795  PHINode *InValPhi = dyn_cast<PHINode>(InVal);
796  if (InValPhi && InValPhi->getParent() == BB) {
797  // Add all of the input values of the input PHI as inputs of this phi.
798  for (unsigned i = 0, e = InValPhi->getNumIncomingValues(); i != e; ++i)
799  PN.addIncoming(InValPhi->getIncomingValue(i),
800  InValPhi->getIncomingBlock(i));
801  } else {
802  // Otherwise, add one instance of the dominating value for each edge that
803  // we will be adding.
804  if (PHINode *BBPN = dyn_cast<PHINode>(BB->begin())) {
805  for (unsigned i = 0, e = BBPN->getNumIncomingValues(); i != e; ++i)
806  PN.addIncoming(InVal, BBPN->getIncomingBlock(i));
807  } else {
808  for (pred_iterator PI = pred_begin(BB), E = pred_end(BB); PI != E; ++PI)
809  PN.addIncoming(InVal, *PI);
810  }
811  }
812  }
813 
814  // The PHIs are now updated, change everything that refers to BB to use
815  // DestBB and remove BB.
816  BB->replaceAllUsesWith(DestBB);
817  BB->eraseFromParent();
818  ++NumBlocksElim;
819 
820  LLVM_DEBUG(dbgs() << "AFTER:\n" << *DestBB << "\n\n\n");
821 }
822 
823 // Computes a map of base pointer relocation instructions to corresponding
824 // derived pointer relocation instructions given a vector of all relocate calls
826  const SmallVectorImpl<GCRelocateInst *> &AllRelocateCalls,
828  &RelocateInstMap) {
829  // Collect information in two maps: one primarily for locating the base object
830  // while filling the second map; the second map is the final structure holding
831  // a mapping between Base and corresponding Derived relocate calls
833  for (auto *ThisRelocate : AllRelocateCalls) {
834  auto K = std::make_pair(ThisRelocate->getBasePtrIndex(),
835  ThisRelocate->getDerivedPtrIndex());
836  RelocateIdxMap.insert(std::make_pair(K, ThisRelocate));
837  }
838  for (auto &Item : RelocateIdxMap) {
839  std::pair<unsigned, unsigned> Key = Item.first;
840  if (Key.first == Key.second)
841  // Base relocation: nothing to insert
842  continue;
843 
844  GCRelocateInst *I = Item.second;
845  auto BaseKey = std::make_pair(Key.first, Key.first);
846 
847  // We're iterating over RelocateIdxMap so we cannot modify it.
848  auto MaybeBase = RelocateIdxMap.find(BaseKey);
849  if (MaybeBase == RelocateIdxMap.end())
850  // TODO: We might want to insert a new base object relocate and gep off
851  // that, if there are enough derived object relocates.
852  continue;
853 
854  RelocateInstMap[MaybeBase->second].push_back(I);
855  }
856 }
857 
858 // Accepts a GEP and extracts the operands into a vector provided they're all
859 // small integer constants
861  SmallVectorImpl<Value *> &OffsetV) {
862  for (unsigned i = 1; i < GEP->getNumOperands(); i++) {
863  // Only accept small constant integer operands
864  auto Op = dyn_cast<ConstantInt>(GEP->getOperand(i));
865  if (!Op || Op->getZExtValue() > 20)
866  return false;
867  }
868 
869  for (unsigned i = 1; i < GEP->getNumOperands(); i++)
870  OffsetV.push_back(GEP->getOperand(i));
871  return true;
872 }
873 
874 // Takes a RelocatedBase (base pointer relocation instruction) and Targets to
875 // replace, computes a replacement, and affects it.
876 static bool
878  const SmallVectorImpl<GCRelocateInst *> &Targets) {
879  bool MadeChange = false;
880  // We must ensure the relocation of derived pointer is defined after
881  // relocation of base pointer. If we find a relocation corresponding to base
882  // defined earlier than relocation of base then we move relocation of base
883  // right before found relocation. We consider only relocation in the same
884  // basic block as relocation of base. Relocations from other basic block will
885  // be skipped by optimization and we do not care about them.
886  for (auto R = RelocatedBase->getParent()->getFirstInsertionPt();
887  &*R != RelocatedBase; ++R)
888  if (auto RI = dyn_cast<GCRelocateInst>(R))
889  if (RI->getStatepoint() == RelocatedBase->getStatepoint())
890  if (RI->getBasePtrIndex() == RelocatedBase->getBasePtrIndex()) {
891  RelocatedBase->moveBefore(RI);
892  break;
893  }
894 
895  for (GCRelocateInst *ToReplace : Targets) {
896  assert(ToReplace->getBasePtrIndex() == RelocatedBase->getBasePtrIndex() &&
897  "Not relocating a derived object of the original base object");
898  if (ToReplace->getBasePtrIndex() == ToReplace->getDerivedPtrIndex()) {
899  // A duplicate relocate call. TODO: coalesce duplicates.
900  continue;
901  }
902 
903  if (RelocatedBase->getParent() != ToReplace->getParent()) {
904  // Base and derived relocates are in different basic blocks.
905  // In this case transform is only valid when base dominates derived
906  // relocate. However it would be too expensive to check dominance
907  // for each such relocate, so we skip the whole transformation.
908  continue;
909  }
910 
911  Value *Base = ToReplace->getBasePtr();
912  auto Derived = dyn_cast<GetElementPtrInst>(ToReplace->getDerivedPtr());
913  if (!Derived || Derived->getPointerOperand() != Base)
914  continue;
915 
916  SmallVector<Value *, 2> OffsetV;
917  if (!getGEPSmallConstantIntOffsetV(Derived, OffsetV))
918  continue;
919 
920  // Create a Builder and replace the target callsite with a gep
921  assert(RelocatedBase->getNextNode() &&
922  "Should always have one since it's not a terminator");
923 
924  // Insert after RelocatedBase
925  IRBuilder<> Builder(RelocatedBase->getNextNode());
926  Builder.SetCurrentDebugLocation(ToReplace->getDebugLoc());
927 
928  // If gc_relocate does not match the actual type, cast it to the right type.
929  // In theory, there must be a bitcast after gc_relocate if the type does not
930  // match, and we should reuse it to get the derived pointer. But it could be
931  // cases like this:
932  // bb1:
933  // ...
934  // %g1 = call coldcc i8 addrspace(1)* @llvm.experimental.gc.relocate.p1i8(...)
935  // br label %merge
936  //
937  // bb2:
938  // ...
939  // %g2 = call coldcc i8 addrspace(1)* @llvm.experimental.gc.relocate.p1i8(...)
940  // br label %merge
941  //
942  // merge:
943  // %p1 = phi i8 addrspace(1)* [ %g1, %bb1 ], [ %g2, %bb2 ]
944  // %cast = bitcast i8 addrspace(1)* %p1 in to i32 addrspace(1)*
945  //
946  // In this case, we can not find the bitcast any more. So we insert a new bitcast
947  // no matter there is already one or not. In this way, we can handle all cases, and
948  // the extra bitcast should be optimized away in later passes.
949  Value *ActualRelocatedBase = RelocatedBase;
950  if (RelocatedBase->getType() != Base->getType()) {
951  ActualRelocatedBase =
952  Builder.CreateBitCast(RelocatedBase, Base->getType());
953  }
954  Value *Replacement = Builder.CreateGEP(
955  Derived->getSourceElementType(), ActualRelocatedBase, makeArrayRef(OffsetV));
956  Replacement->takeName(ToReplace);
957  // If the newly generated derived pointer's type does not match the original derived
958  // pointer's type, cast the new derived pointer to match it. Same reasoning as above.
959  Value *ActualReplacement = Replacement;
960  if (Replacement->getType() != ToReplace->getType()) {
961  ActualReplacement =
962  Builder.CreateBitCast(Replacement, ToReplace->getType());
963  }
964  ToReplace->replaceAllUsesWith(ActualReplacement);
965  ToReplace->eraseFromParent();
966 
967  MadeChange = true;
968  }
969  return MadeChange;
970 }
971 
972 // Turns this:
973 //
974 // %base = ...
975 // %ptr = gep %base + 15
976 // %tok = statepoint (%fun, i32 0, i32 0, i32 0, %base, %ptr)
977 // %base' = relocate(%tok, i32 4, i32 4)
978 // %ptr' = relocate(%tok, i32 4, i32 5)
979 // %val = load %ptr'
980 //
981 // into this:
982 //
983 // %base = ...
984 // %ptr = gep %base + 15
985 // %tok = statepoint (%fun, i32 0, i32 0, i32 0, %base, %ptr)
986 // %base' = gc.relocate(%tok, i32 4, i32 4)
987 // %ptr' = gep %base' + 15
988 // %val = load %ptr'
989 bool CodeGenPrepare::simplifyOffsetableRelocate(Instruction &I) {
990  bool MadeChange = false;
991  SmallVector<GCRelocateInst *, 2> AllRelocateCalls;
992 
993  for (auto *U : I.users())
994  if (GCRelocateInst *Relocate = dyn_cast<GCRelocateInst>(U))
995  // Collect all the relocate calls associated with a statepoint
996  AllRelocateCalls.push_back(Relocate);
997 
998  // We need atleast one base pointer relocation + one derived pointer
999  // relocation to mangle
1000  if (AllRelocateCalls.size() < 2)
1001  return false;
1002 
1003  // RelocateInstMap is a mapping from the base relocate instruction to the
1004  // corresponding derived relocate instructions
1006  computeBaseDerivedRelocateMap(AllRelocateCalls, RelocateInstMap);
1007  if (RelocateInstMap.empty())
1008  return false;
1009 
1010  for (auto &Item : RelocateInstMap)
1011  // Item.first is the RelocatedBase to offset against
1012  // Item.second is the vector of Targets to replace
1013  MadeChange = simplifyRelocatesOffABase(Item.first, Item.second);
1014  return MadeChange;
1015 }
1016 
1017 /// SinkCast - Sink the specified cast instruction into its user blocks
1018 static bool SinkCast(CastInst *CI) {
1019  BasicBlock *DefBB = CI->getParent();
1020 
1021  /// InsertedCasts - Only insert a cast in each block once.
1022  DenseMap<BasicBlock*, CastInst*> InsertedCasts;
1023 
1024  bool MadeChange = false;
1025  for (Value::user_iterator UI = CI->user_begin(), E = CI->user_end();
1026  UI != E; ) {
1027  Use &TheUse = UI.getUse();
1028  Instruction *User = cast<Instruction>(*UI);
1029 
1030  // Figure out which BB this cast is used in. For PHI's this is the
1031  // appropriate predecessor block.
1032  BasicBlock *UserBB = User->getParent();
1033  if (PHINode *PN = dyn_cast<PHINode>(User)) {
1034  UserBB = PN->getIncomingBlock(TheUse);
1035  }
1036 
1037  // Preincrement use iterator so we don't invalidate it.
1038  ++UI;
1039 
1040  // The first insertion point of a block containing an EH pad is after the
1041  // pad. If the pad is the user, we cannot sink the cast past the pad.
1042  if (User->isEHPad())
1043  continue;
1044 
1045  // If the block selected to receive the cast is an EH pad that does not
1046  // allow non-PHI instructions before the terminator, we can't sink the
1047  // cast.
1048  if (UserBB->getTerminator()->isEHPad())
1049  continue;
1050 
1051  // If this user is in the same block as the cast, don't change the cast.
1052  if (UserBB == DefBB) continue;
1053 
1054  // If we have already inserted a cast into this block, use it.
1055  CastInst *&InsertedCast = InsertedCasts[UserBB];
1056 
1057  if (!InsertedCast) {
1058  BasicBlock::iterator InsertPt = UserBB->getFirstInsertionPt();
1059  assert(InsertPt != UserBB->end());
1060  InsertedCast = CastInst::Create(CI->getOpcode(), CI->getOperand(0),
1061  CI->getType(), "", &*InsertPt);
1062  InsertedCast->setDebugLoc(CI->getDebugLoc());
1063  }
1064 
1065  // Replace a use of the cast with a use of the new cast.
1066  TheUse = InsertedCast;
1067  MadeChange = true;
1068  ++NumCastUses;
1069  }
1070 
1071  // If we removed all uses, nuke the cast.
1072  if (CI->use_empty()) {
1073  salvageDebugInfo(*CI);
1074  CI->eraseFromParent();
1075  MadeChange = true;
1076  }
1077 
1078  return MadeChange;
1079 }
1080 
1081 /// If the specified cast instruction is a noop copy (e.g. it's casting from
1082 /// one pointer type to another, i32->i8 on PPC), sink it into user blocks to
1083 /// reduce the number of virtual registers that must be created and coalesced.
1084 ///
1085 /// Return true if any changes are made.
1087  const DataLayout &DL) {
1088  // Sink only "cheap" (or nop) address-space casts. This is a weaker condition
1089  // than sinking only nop casts, but is helpful on some platforms.
1090  if (auto *ASC = dyn_cast<AddrSpaceCastInst>(CI)) {
1091  if (!TLI.isCheapAddrSpaceCast(ASC->getSrcAddressSpace(),
1092  ASC->getDestAddressSpace()))
1093  return false;
1094  }
1095 
1096  // If this is a noop copy,
1097  EVT SrcVT = TLI.getValueType(DL, CI->getOperand(0)->getType());
1098  EVT DstVT = TLI.getValueType(DL, CI->getType());
1099 
1100  // This is an fp<->int conversion?
1101  if (SrcVT.isInteger() != DstVT.isInteger())
1102  return false;
1103 
1104  // If this is an extension, it will be a zero or sign extension, which
1105  // isn't a noop.
1106  if (SrcVT.bitsLT(DstVT)) return false;
1107 
1108  // If these values will be promoted, find out what they will be promoted
1109  // to. This helps us consider truncates on PPC as noop copies when they
1110  // are.
1111  if (TLI.getTypeAction(CI->getContext(), SrcVT) ==
1113  SrcVT = TLI.getTypeToTransformTo(CI->getContext(), SrcVT);
1114  if (TLI.getTypeAction(CI->getContext(), DstVT) ==
1116  DstVT = TLI.getTypeToTransformTo(CI->getContext(), DstVT);
1117 
1118  // If, after promotion, these are the same types, this is a noop copy.
1119  if (SrcVT != DstVT)
1120  return false;
1121 
1122  return SinkCast(CI);
1123 }
1124 
1125 /// Try to combine CI into a call to the llvm.uadd.with.overflow intrinsic if
1126 /// possible.
1127 ///
1128 /// Return true if any changes were made.
1130  Value *A, *B;
1131  Instruction *AddI;
1132  if (!match(CI,
1134  return false;
1135 
1136  Type *Ty = AddI->getType();
1137  if (!isa<IntegerType>(Ty))
1138  return false;
1139 
1140  // We don't want to move around uses of condition values this late, so we we
1141  // check if it is legal to create the call to the intrinsic in the basic
1142  // block containing the icmp:
1143 
1144  if (AddI->getParent() != CI->getParent() && !AddI->hasOneUse())
1145  return false;
1146 
1147 #ifndef NDEBUG
1148  // Someday m_UAddWithOverflow may get smarter, but this is a safe assumption
1149  // for now:
1150  if (AddI->hasOneUse())
1151  assert(*AddI->user_begin() == CI && "expected!");
1152 #endif
1153 
1154  Module *M = CI->getModule();
1155  Value *F = Intrinsic::getDeclaration(M, Intrinsic::uadd_with_overflow, Ty);
1156 
1157  auto *InsertPt = AddI->hasOneUse() ? CI : AddI;
1158 
1159  auto *UAddWithOverflow =
1160  CallInst::Create(F, {A, B}, "uadd.overflow", InsertPt);
1161  auto *UAdd = ExtractValueInst::Create(UAddWithOverflow, 0, "uadd", InsertPt);
1162  auto *Overflow =
1163  ExtractValueInst::Create(UAddWithOverflow, 1, "overflow", InsertPt);
1164 
1165  CI->replaceAllUsesWith(Overflow);
1166  AddI->replaceAllUsesWith(UAdd);
1167  CI->eraseFromParent();
1168  AddI->eraseFromParent();
1169  return true;
1170 }
1171 
1172 /// Sink the given CmpInst into user blocks to reduce the number of virtual
1173 /// registers that must be created and coalesced. This is a clear win except on
1174 /// targets with multiple condition code registers (PowerPC), where it might
1175 /// lose; some adjustment may be wanted there.
1176 ///
1177 /// Return true if any changes are made.
1178 static bool SinkCmpExpression(CmpInst *CI, const TargetLowering *TLI) {
1179  BasicBlock *DefBB = CI->getParent();
1180 
1181  // Avoid sinking soft-FP comparisons, since this can move them into a loop.
1182  if (TLI && TLI->useSoftFloat() && isa<FCmpInst>(CI))
1183  return false;
1184 
1185  // Only insert a cmp in each block once.
1186  DenseMap<BasicBlock*, CmpInst*> InsertedCmps;
1187 
1188  bool MadeChange = false;
1189  for (Value::user_iterator UI = CI->user_begin(), E = CI->user_end();
1190  UI != E; ) {
1191  Use &TheUse = UI.getUse();
1192  Instruction *User = cast<Instruction>(*UI);
1193 
1194  // Preincrement use iterator so we don't invalidate it.
1195  ++UI;
1196 
1197  // Don't bother for PHI nodes.
1198  if (isa<PHINode>(User))
1199  continue;
1200 
1201  // Figure out which BB this cmp is used in.
1202  BasicBlock *UserBB = User->getParent();
1203 
1204  // If this user is in the same block as the cmp, don't change the cmp.
1205  if (UserBB == DefBB) continue;
1206 
1207  // If we have already inserted a cmp into this block, use it.
1208  CmpInst *&InsertedCmp = InsertedCmps[UserBB];
1209 
1210  if (!InsertedCmp) {
1211  BasicBlock::iterator InsertPt = UserBB->getFirstInsertionPt();
1212  assert(InsertPt != UserBB->end());
1213  InsertedCmp =
1214  CmpInst::Create(CI->getOpcode(), CI->getPredicate(),
1215  CI->getOperand(0), CI->getOperand(1), "", &*InsertPt);
1216  // Propagate the debug info.
1217  InsertedCmp->setDebugLoc(CI->getDebugLoc());
1218  }
1219 
1220  // Replace a use of the cmp with a use of the new cmp.
1221  TheUse = InsertedCmp;
1222  MadeChange = true;
1223  ++NumCmpUses;
1224  }
1225 
1226  // If we removed all uses, nuke the cmp.
1227  if (CI->use_empty()) {
1228  CI->eraseFromParent();
1229  MadeChange = true;
1230  }
1231 
1232  return MadeChange;
1233 }
1234 
1235 static bool OptimizeCmpExpression(CmpInst *CI, const TargetLowering *TLI) {
1236  if (SinkCmpExpression(CI, TLI))
1237  return true;
1238 
1239  if (CombineUAddWithOverflow(CI))
1240  return true;
1241 
1242  return false;
1243 }
1244 
1245 /// Duplicate and sink the given 'and' instruction into user blocks where it is
1246 /// used in a compare to allow isel to generate better code for targets where
1247 /// this operation can be combined.
1248 ///
1249 /// Return true if any changes are made.
1251  const TargetLowering &TLI,
1252  SetOfInstrs &InsertedInsts) {
1253  // Double-check that we're not trying to optimize an instruction that was
1254  // already optimized by some other part of this pass.
1255  assert(!InsertedInsts.count(AndI) &&
1256  "Attempting to optimize already optimized and instruction");
1257  (void) InsertedInsts;
1258 
1259  // Nothing to do for single use in same basic block.
1260  if (AndI->hasOneUse() &&
1261  AndI->getParent() == cast<Instruction>(*AndI->user_begin())->getParent())
1262  return false;
1263 
1264  // Try to avoid cases where sinking/duplicating is likely to increase register
1265  // pressure.
1266  if (!isa<ConstantInt>(AndI->getOperand(0)) &&
1267  !isa<ConstantInt>(AndI->getOperand(1)) &&
1268  AndI->getOperand(0)->hasOneUse() && AndI->getOperand(1)->hasOneUse())
1269  return false;
1270 
1271  for (auto *U : AndI->users()) {
1272  Instruction *User = cast<Instruction>(U);
1273 
1274  // Only sink for and mask feeding icmp with 0.
1275  if (!isa<ICmpInst>(User))
1276  return false;
1277 
1278  auto *CmpC = dyn_cast<ConstantInt>(User->getOperand(1));
1279  if (!CmpC || !CmpC->isZero())
1280  return false;
1281  }
1282 
1283  if (!TLI.isMaskAndCmp0FoldingBeneficial(*AndI))
1284  return false;
1285 
1286  LLVM_DEBUG(dbgs() << "found 'and' feeding only icmp 0;\n");
1287  LLVM_DEBUG(AndI->getParent()->dump());
1288 
1289  // Push the 'and' into the same block as the icmp 0. There should only be
1290  // one (icmp (and, 0)) in each block, since CSE/GVN should have removed any
1291  // others, so we don't need to keep track of which BBs we insert into.
1292  for (Value::user_iterator UI = AndI->user_begin(), E = AndI->user_end();
1293  UI != E; ) {
1294  Use &TheUse = UI.getUse();
1295  Instruction *User = cast<Instruction>(*UI);
1296 
1297  // Preincrement use iterator so we don't invalidate it.
1298  ++UI;
1299 
1300  LLVM_DEBUG(dbgs() << "sinking 'and' use: " << *User << "\n");
1301 
1302  // Keep the 'and' in the same place if the use is already in the same block.
1303  Instruction *InsertPt =
1304  User->getParent() == AndI->getParent() ? AndI : User;
1305  Instruction *InsertedAnd =
1306  BinaryOperator::Create(Instruction::And, AndI->getOperand(0),
1307  AndI->getOperand(1), "", InsertPt);
1308  // Propagate the debug info.
1309  InsertedAnd->setDebugLoc(AndI->getDebugLoc());
1310 
1311  // Replace a use of the 'and' with a use of the new 'and'.
1312  TheUse = InsertedAnd;
1313  ++NumAndUses;
1314  LLVM_DEBUG(User->getParent()->dump());
1315  }
1316 
1317  // We removed all uses, nuke the and.
1318  AndI->eraseFromParent();
1319  return true;
1320 }
1321 
1322 /// Check if the candidates could be combined with a shift instruction, which
1323 /// includes:
1324 /// 1. Truncate instruction
1325 /// 2. And instruction and the imm is a mask of the low bits:
1326 /// imm & (imm+1) == 0
1328  if (!isa<TruncInst>(User)) {
1329  if (User->getOpcode() != Instruction::And ||
1330  !isa<ConstantInt>(User->getOperand(1)))
1331  return false;
1332 
1333  const APInt &Cimm = cast<ConstantInt>(User->getOperand(1))->getValue();
1334 
1335  if ((Cimm & (Cimm + 1)).getBoolValue())
1336  return false;
1337  }
1338  return true;
1339 }
1340 
1341 /// Sink both shift and truncate instruction to the use of truncate's BB.
1342 static bool
1345  const TargetLowering &TLI, const DataLayout &DL) {
1346  BasicBlock *UserBB = User->getParent();
1347  DenseMap<BasicBlock *, CastInst *> InsertedTruncs;
1348  TruncInst *TruncI = dyn_cast<TruncInst>(User);
1349  bool MadeChange = false;
1350 
1351  for (Value::user_iterator TruncUI = TruncI->user_begin(),
1352  TruncE = TruncI->user_end();
1353  TruncUI != TruncE;) {
1354 
1355  Use &TruncTheUse = TruncUI.getUse();
1356  Instruction *TruncUser = cast<Instruction>(*TruncUI);
1357  // Preincrement use iterator so we don't invalidate it.
1358 
1359  ++TruncUI;
1360 
1361  int ISDOpcode = TLI.InstructionOpcodeToISD(TruncUser->getOpcode());
1362  if (!ISDOpcode)
1363  continue;
1364 
1365  // If the use is actually a legal node, there will not be an
1366  // implicit truncate.
1367  // FIXME: always querying the result type is just an
1368  // approximation; some nodes' legality is determined by the
1369  // operand or other means. There's no good way to find out though.
1370  if (TLI.isOperationLegalOrCustom(
1371  ISDOpcode, TLI.getValueType(DL, TruncUser->getType(), true)))
1372  continue;
1373 
1374  // Don't bother for PHI nodes.
1375  if (isa<PHINode>(TruncUser))
1376  continue;
1377 
1378  BasicBlock *TruncUserBB = TruncUser->getParent();
1379 
1380  if (UserBB == TruncUserBB)
1381  continue;
1382 
1383  BinaryOperator *&InsertedShift = InsertedShifts[TruncUserBB];
1384  CastInst *&InsertedTrunc = InsertedTruncs[TruncUserBB];
1385 
1386  if (!InsertedShift && !InsertedTrunc) {
1387  BasicBlock::iterator InsertPt = TruncUserBB->getFirstInsertionPt();
1388  assert(InsertPt != TruncUserBB->end());
1389  // Sink the shift
1390  if (ShiftI->getOpcode() == Instruction::AShr)
1391  InsertedShift = BinaryOperator::CreateAShr(ShiftI->getOperand(0), CI,
1392  "", &*InsertPt);
1393  else
1394  InsertedShift = BinaryOperator::CreateLShr(ShiftI->getOperand(0), CI,
1395  "", &*InsertPt);
1396 
1397  // Sink the trunc
1398  BasicBlock::iterator TruncInsertPt = TruncUserBB->getFirstInsertionPt();
1399  TruncInsertPt++;
1400  assert(TruncInsertPt != TruncUserBB->end());
1401 
1402  InsertedTrunc = CastInst::Create(TruncI->getOpcode(), InsertedShift,
1403  TruncI->getType(), "", &*TruncInsertPt);
1404 
1405  MadeChange = true;
1406 
1407  TruncTheUse = InsertedTrunc;
1408  }
1409  }
1410  return MadeChange;
1411 }
1412 
1413 /// Sink the shift *right* instruction into user blocks if the uses could
1414 /// potentially be combined with this shift instruction and generate BitExtract
1415 /// instruction. It will only be applied if the architecture supports BitExtract
1416 /// instruction. Here is an example:
1417 /// BB1:
1418 /// %x.extract.shift = lshr i64 %arg1, 32
1419 /// BB2:
1420 /// %x.extract.trunc = trunc i64 %x.extract.shift to i16
1421 /// ==>
1422 ///
1423 /// BB2:
1424 /// %x.extract.shift.1 = lshr i64 %arg1, 32
1425 /// %x.extract.trunc = trunc i64 %x.extract.shift.1 to i16
1426 ///
1427 /// CodeGen will recognize the pattern in BB2 and generate BitExtract
1428 /// instruction.
1429 /// Return true if any changes are made.
1431  const TargetLowering &TLI,
1432  const DataLayout &DL) {
1433  BasicBlock *DefBB = ShiftI->getParent();
1434 
1435  /// Only insert instructions in each block once.
1437 
1438  bool shiftIsLegal = TLI.isTypeLegal(TLI.getValueType(DL, ShiftI->getType()));
1439 
1440  bool MadeChange = false;
1441  for (Value::user_iterator UI = ShiftI->user_begin(), E = ShiftI->user_end();
1442  UI != E;) {
1443  Use &TheUse = UI.getUse();
1444  Instruction *User = cast<Instruction>(*UI);
1445  // Preincrement use iterator so we don't invalidate it.
1446  ++UI;
1447 
1448  // Don't bother for PHI nodes.
1449  if (isa<PHINode>(User))
1450  continue;
1451 
1452  if (!isExtractBitsCandidateUse(User))
1453  continue;
1454 
1455  BasicBlock *UserBB = User->getParent();
1456 
1457  if (UserBB == DefBB) {
1458  // If the shift and truncate instruction are in the same BB. The use of
1459  // the truncate(TruncUse) may still introduce another truncate if not
1460  // legal. In this case, we would like to sink both shift and truncate
1461  // instruction to the BB of TruncUse.
1462  // for example:
1463  // BB1:
1464  // i64 shift.result = lshr i64 opnd, imm
1465  // trunc.result = trunc shift.result to i16
1466  //
1467  // BB2:
1468  // ----> We will have an implicit truncate here if the architecture does
1469  // not have i16 compare.
1470  // cmp i16 trunc.result, opnd2
1471  //
1472  if (isa<TruncInst>(User) && shiftIsLegal
1473  // If the type of the truncate is legal, no truncate will be
1474  // introduced in other basic blocks.
1475  &&
1476  (!TLI.isTypeLegal(TLI.getValueType(DL, User->getType()))))
1477  MadeChange =
1478  SinkShiftAndTruncate(ShiftI, User, CI, InsertedShifts, TLI, DL);
1479 
1480  continue;
1481  }
1482  // If we have already inserted a shift into this block, use it.
1483  BinaryOperator *&InsertedShift = InsertedShifts[UserBB];
1484 
1485  if (!InsertedShift) {
1486  BasicBlock::iterator InsertPt = UserBB->getFirstInsertionPt();
1487  assert(InsertPt != UserBB->end());
1488 
1489  if (ShiftI->getOpcode() == Instruction::AShr)
1490  InsertedShift = BinaryOperator::CreateAShr(ShiftI->getOperand(0), CI,
1491  "", &*InsertPt);
1492  else
1493  InsertedShift = BinaryOperator::CreateLShr(ShiftI->getOperand(0), CI,
1494  "", &*InsertPt);
1495 
1496  MadeChange = true;
1497  }
1498 
1499  // Replace a use of the shift with a use of the new shift.
1500  TheUse = InsertedShift;
1501  }
1502 
1503  // If we removed all uses, nuke the shift.
1504  if (ShiftI->use_empty())
1505  ShiftI->eraseFromParent();
1506 
1507  return MadeChange;
1508 }
1509 
1510 /// If counting leading or trailing zeros is an expensive operation and a zero
1511 /// input is defined, add a check for zero to avoid calling the intrinsic.
1512 ///
1513 /// We want to transform:
1514 /// %z = call i64 @llvm.cttz.i64(i64 %A, i1 false)
1515 ///
1516 /// into:
1517 /// entry:
1518 /// %cmpz = icmp eq i64 %A, 0
1519 /// br i1 %cmpz, label %cond.end, label %cond.false
1520 /// cond.false:
1521 /// %z = call i64 @llvm.cttz.i64(i64 %A, i1 true)
1522 /// br label %cond.end
1523 /// cond.end:
1524 /// %ctz = phi i64 [ 64, %entry ], [ %z, %cond.false ]
1525 ///
1526 /// If the transform is performed, return true and set ModifiedDT to true.
1527 static bool despeculateCountZeros(IntrinsicInst *CountZeros,
1528  const TargetLowering *TLI,
1529  const DataLayout *DL,
1530  bool &ModifiedDT) {
1531  if (!TLI || !DL)
1532  return false;
1533 
1534  // If a zero input is undefined, it doesn't make sense to despeculate that.
1535  if (match(CountZeros->getOperand(1), m_One()))
1536  return false;
1537 
1538  // If it's cheap to speculate, there's nothing to do.
1539  auto IntrinsicID = CountZeros->getIntrinsicID();
1540  if ((IntrinsicID == Intrinsic::cttz && TLI->isCheapToSpeculateCttz()) ||
1541  (IntrinsicID == Intrinsic::ctlz && TLI->isCheapToSpeculateCtlz()))
1542  return false;
1543 
1544  // Only handle legal scalar cases. Anything else requires too much work.
1545  Type *Ty = CountZeros->getType();
1546  unsigned SizeInBits = Ty->getPrimitiveSizeInBits();
1547  if (Ty->isVectorTy() || SizeInBits > DL->getLargestLegalIntTypeSizeInBits())
1548  return false;
1549 
1550  // The intrinsic will be sunk behind a compare against zero and branch.
1551  BasicBlock *StartBlock = CountZeros->getParent();
1552  BasicBlock *CallBlock = StartBlock->splitBasicBlock(CountZeros, "cond.false");
1553 
1554  // Create another block after the count zero intrinsic. A PHI will be added
1555  // in this block to select the result of the intrinsic or the bit-width
1556  // constant if the input to the intrinsic is zero.
1557  BasicBlock::iterator SplitPt = ++(BasicBlock::iterator(CountZeros));
1558  BasicBlock *EndBlock = CallBlock->splitBasicBlock(SplitPt, "cond.end");
1559 
1560  // Set up a builder to create a compare, conditional branch, and PHI.
1561  IRBuilder<> Builder(CountZeros->getContext());
1562  Builder.SetInsertPoint(StartBlock->getTerminator());
1563  Builder.SetCurrentDebugLocation(CountZeros->getDebugLoc());
1564 
1565  // Replace the unconditional branch that was created by the first split with
1566  // a compare against zero and a conditional branch.
1567  Value *Zero = Constant::getNullValue(Ty);
1568  Value *Cmp = Builder.CreateICmpEQ(CountZeros->getOperand(0), Zero, "cmpz");
1569  Builder.CreateCondBr(Cmp, EndBlock, CallBlock);
1570  StartBlock->getTerminator()->eraseFromParent();
1571 
1572  // Create a PHI in the end block to select either the output of the intrinsic
1573  // or the bit width of the operand.
1574  Builder.SetInsertPoint(&EndBlock->front());
1575  PHINode *PN = Builder.CreatePHI(Ty, 2, "ctz");
1576  CountZeros->replaceAllUsesWith(PN);
1577  Value *BitWidth = Builder.getInt(APInt(SizeInBits, SizeInBits));
1578  PN->addIncoming(BitWidth, StartBlock);
1579  PN->addIncoming(CountZeros, CallBlock);
1580 
1581  // We are explicitly handling the zero case, so we can set the intrinsic's
1582  // undefined zero argument to 'true'. This will also prevent reprocessing the
1583  // intrinsic; we only despeculate when a zero input is defined.
1584  CountZeros->setArgOperand(1, Builder.getTrue());
1585  ModifiedDT = true;
1586  return true;
1587 }
1588 
1589 bool CodeGenPrepare::optimizeCallInst(CallInst *CI, bool &ModifiedDT) {
1590  BasicBlock *BB = CI->getParent();
1591 
1592  // Lower inline assembly if we can.
1593  // If we found an inline asm expession, and if the target knows how to
1594  // lower it to normal LLVM code, do so now.
1595  if (TLI && isa<InlineAsm>(CI->getCalledValue())) {
1596  if (TLI->ExpandInlineAsm(CI)) {
1597  // Avoid invalidating the iterator.
1598  CurInstIterator = BB->begin();
1599  // Avoid processing instructions out of order, which could cause
1600  // reuse before a value is defined.
1601  SunkAddrs.clear();
1602  return true;
1603  }
1604  // Sink address computing for memory operands into the block.
1605  if (optimizeInlineAsmInst(CI))
1606  return true;
1607  }
1608 
1609  // Align the pointer arguments to this call if the target thinks it's a good
1610  // idea
1611  unsigned MinSize, PrefAlign;
1612  if (TLI && TLI->shouldAlignPointerArgs(CI, MinSize, PrefAlign)) {
1613  for (auto &Arg : CI->arg_operands()) {
1614  // We want to align both objects whose address is used directly and
1615  // objects whose address is used in casts and GEPs, though it only makes
1616  // sense for GEPs if the offset is a multiple of the desired alignment and
1617  // if size - offset meets the size threshold.
1618  if (!Arg->getType()->isPointerTy())
1619  continue;
1620  APInt Offset(DL->getIndexSizeInBits(
1621  cast<PointerType>(Arg->getType())->getAddressSpace()),
1622  0);
1624  uint64_t Offset2 = Offset.getLimitedValue();
1625  if ((Offset2 & (PrefAlign-1)) != 0)
1626  continue;
1627  AllocaInst *AI;
1628  if ((AI = dyn_cast<AllocaInst>(Val)) && AI->getAlignment() < PrefAlign &&
1629  DL->getTypeAllocSize(AI->getAllocatedType()) >= MinSize + Offset2)
1630  AI->setAlignment(PrefAlign);
1631  // Global variables can only be aligned if they are defined in this
1632  // object (i.e. they are uniquely initialized in this object), and
1633  // over-aligning global variables that have an explicit section is
1634  // forbidden.
1635  GlobalVariable *GV;
1636  if ((GV = dyn_cast<GlobalVariable>(Val)) && GV->canIncreaseAlignment() &&
1637  GV->getPointerAlignment(*DL) < PrefAlign &&
1638  DL->getTypeAllocSize(GV->getValueType()) >=
1639  MinSize + Offset2)
1640  GV->setAlignment(PrefAlign);
1641  }
1642  // If this is a memcpy (or similar) then we may be able to improve the
1643  // alignment
1644  if (MemIntrinsic *MI = dyn_cast<MemIntrinsic>(CI)) {
1645  unsigned DestAlign = getKnownAlignment(MI->getDest(), *DL);
1646  if (DestAlign > MI->getDestAlignment())
1647  MI->setDestAlignment(DestAlign);
1648  if (MemTransferInst *MTI = dyn_cast<MemTransferInst>(MI)) {
1649  unsigned SrcAlign = getKnownAlignment(MTI->getSource(), *DL);
1650  if (SrcAlign > MTI->getSourceAlignment())
1651  MTI->setSourceAlignment(SrcAlign);
1652  }
1653  }
1654  }
1655 
1656  // If we have a cold call site, try to sink addressing computation into the
1657  // cold block. This interacts with our handling for loads and stores to
1658  // ensure that we can fold all uses of a potential addressing computation
1659  // into their uses. TODO: generalize this to work over profiling data
1660  if (!OptSize && CI->hasFnAttr(Attribute::Cold))
1661  for (auto &Arg : CI->arg_operands()) {
1662  if (!Arg->getType()->isPointerTy())
1663  continue;
1664  unsigned AS = Arg->getType()->getPointerAddressSpace();
1665  return optimizeMemoryInst(CI, Arg, Arg->getType(), AS);
1666  }
1667 
1669  if (II) {
1670  switch (II->getIntrinsicID()) {
1671  default: break;
1672  case Intrinsic::objectsize: {
1673  // Lower all uses of llvm.objectsize.*
1674  ConstantInt *RetVal =
1675  lowerObjectSizeCall(II, *DL, TLInfo, /*MustSucceed=*/true);
1676  // Substituting this can cause recursive simplifications, which can
1677  // invalidate our iterator. Use a WeakTrackingVH to hold onto it in case
1678  // this
1679  // happens.
1680  Value *CurValue = &*CurInstIterator;
1681  WeakTrackingVH IterHandle(CurValue);
1682 
1683  replaceAndRecursivelySimplify(CI, RetVal, TLInfo, nullptr);
1684 
1685  // If the iterator instruction was recursively deleted, start over at the
1686  // start of the block.
1687  if (IterHandle != CurValue) {
1688  CurInstIterator = BB->begin();
1689  SunkAddrs.clear();
1690  }
1691  return true;
1692  }
1693  case Intrinsic::aarch64_stlxr:
1694  case Intrinsic::aarch64_stxr: {
1695  ZExtInst *ExtVal = dyn_cast<ZExtInst>(CI->getArgOperand(0));
1696  if (!ExtVal || !ExtVal->hasOneUse() ||
1697  ExtVal->getParent() == CI->getParent())
1698  return false;
1699  // Sink a zext feeding stlxr/stxr before it, so it can be folded into it.
1700  ExtVal->moveBefore(CI);
1701  // Mark this instruction as "inserted by CGP", so that other
1702  // optimizations don't touch it.
1703  InsertedInsts.insert(ExtVal);
1704  return true;
1705  }
1706  case Intrinsic::launder_invariant_group:
1707  case Intrinsic::strip_invariant_group:
1708  II->replaceAllUsesWith(II->getArgOperand(0));
1709  II->eraseFromParent();
1710  return true;
1711 
1712  case Intrinsic::cttz:
1713  case Intrinsic::ctlz:
1714  // If counting zeros is expensive, try to avoid it.
1715  return despeculateCountZeros(II, TLI, DL, ModifiedDT);
1716  }
1717 
1718  if (TLI) {
1719  SmallVector<Value*, 2> PtrOps;
1720  Type *AccessTy;
1721  if (TLI->getAddrModeArguments(II, PtrOps, AccessTy))
1722  while (!PtrOps.empty()) {
1723  Value *PtrVal = PtrOps.pop_back_val();
1724  unsigned AS = PtrVal->getType()->getPointerAddressSpace();
1725  if (optimizeMemoryInst(II, PtrVal, AccessTy, AS))
1726  return true;
1727  }
1728  }
1729  }
1730 
1731  // From here on out we're working with named functions.
1732  if (!CI->getCalledFunction()) return false;
1733 
1734  // Lower all default uses of _chk calls. This is very similar
1735  // to what InstCombineCalls does, but here we are only lowering calls
1736  // to fortified library functions (e.g. __memcpy_chk) that have the default
1737  // "don't know" as the objectsize. Anything else should be left alone.
1738  FortifiedLibCallSimplifier Simplifier(TLInfo, true);
1739  if (Value *V = Simplifier.optimizeCall(CI)) {
1740  CI->replaceAllUsesWith(V);
1741  CI->eraseFromParent();
1742  return true;
1743  }
1744 
1745  return false;
1746 }
1747 
1748 /// Look for opportunities to duplicate return instructions to the predecessor
1749 /// to enable tail call optimizations. The case it is currently looking for is:
1750 /// @code
1751 /// bb0:
1752 /// %tmp0 = tail call i32 @f0()
1753 /// br label %return
1754 /// bb1:
1755 /// %tmp1 = tail call i32 @f1()
1756 /// br label %return
1757 /// bb2:
1758 /// %tmp2 = tail call i32 @f2()
1759 /// br label %return
1760 /// return:
1761 /// %retval = phi i32 [ %tmp0, %bb0 ], [ %tmp1, %bb1 ], [ %tmp2, %bb2 ]
1762 /// ret i32 %retval
1763 /// @endcode
1764 ///
1765 /// =>
1766 ///
1767 /// @code
1768 /// bb0:
1769 /// %tmp0 = tail call i32 @f0()
1770 /// ret i32 %tmp0
1771 /// bb1:
1772 /// %tmp1 = tail call i32 @f1()
1773 /// ret i32 %tmp1
1774 /// bb2:
1775 /// %tmp2 = tail call i32 @f2()
1776 /// ret i32 %tmp2
1777 /// @endcode
1778 bool CodeGenPrepare::dupRetToEnableTailCallOpts(BasicBlock *BB) {
1779  if (!TLI)
1780  return false;
1781 
1782  ReturnInst *RetI = dyn_cast<ReturnInst>(BB->getTerminator());
1783  if (!RetI)
1784  return false;
1785 
1786  PHINode *PN = nullptr;
1787  BitCastInst *BCI = nullptr;
1788  Value *V = RetI->getReturnValue();
1789  if (V) {
1790  BCI = dyn_cast<BitCastInst>(V);
1791  if (BCI)
1792  V = BCI->getOperand(0);
1793 
1794  PN = dyn_cast<PHINode>(V);
1795  if (!PN)
1796  return false;
1797  }
1798 
1799  if (PN && PN->getParent() != BB)
1800  return false;
1801 
1802  // Make sure there are no instructions between the PHI and return, or that the
1803  // return is the first instruction in the block.
1804  if (PN) {
1805  BasicBlock::iterator BI = BB->begin();
1806  do { ++BI; } while (isa<DbgInfoIntrinsic>(BI));
1807  if (&*BI == BCI)
1808  // Also skip over the bitcast.
1809  ++BI;
1810  if (&*BI != RetI)
1811  return false;
1812  } else {
1813  BasicBlock::iterator BI = BB->begin();
1814  while (isa<DbgInfoIntrinsic>(BI)) ++BI;
1815  if (&*BI != RetI)
1816  return false;
1817  }
1818 
1819  /// Only dup the ReturnInst if the CallInst is likely to be emitted as a tail
1820  /// call.
1821  const Function *F = BB->getParent();
1822  SmallVector<CallInst*, 4> TailCalls;
1823  if (PN) {
1824  for (unsigned I = 0, E = PN->getNumIncomingValues(); I != E; ++I) {
1825  CallInst *CI = dyn_cast<CallInst>(PN->getIncomingValue(I));
1826  // Make sure the phi value is indeed produced by the tail call.
1827  if (CI && CI->hasOneUse() && CI->getParent() == PN->getIncomingBlock(I) &&
1828  TLI->mayBeEmittedAsTailCall(CI) &&
1829  attributesPermitTailCall(F, CI, RetI, *TLI))
1830  TailCalls.push_back(CI);
1831  }
1832  } else {
1833  SmallPtrSet<BasicBlock*, 4> VisitedBBs;
1834  for (pred_iterator PI = pred_begin(BB), PE = pred_end(BB); PI != PE; ++PI) {
1835  if (!VisitedBBs.insert(*PI).second)
1836  continue;
1837 
1838  BasicBlock::InstListType &InstList = (*PI)->getInstList();
1839  BasicBlock::InstListType::reverse_iterator RI = InstList.rbegin();
1840  BasicBlock::InstListType::reverse_iterator RE = InstList.rend();
1841  do { ++RI; } while (RI != RE && isa<DbgInfoIntrinsic>(&*RI));
1842  if (RI == RE)
1843  continue;
1844 
1845  CallInst *CI = dyn_cast<CallInst>(&*RI);
1846  if (CI && CI->use_empty() && TLI->mayBeEmittedAsTailCall(CI) &&
1847  attributesPermitTailCall(F, CI, RetI, *TLI))
1848  TailCalls.push_back(CI);
1849  }
1850  }
1851 
1852  bool Changed = false;
1853  for (unsigned i = 0, e = TailCalls.size(); i != e; ++i) {
1854  CallInst *CI = TailCalls[i];
1855  CallSite CS(CI);
1856 
1857  // Conservatively require the attributes of the call to match those of the
1858  // return. Ignore noalias because it doesn't affect the call sequence.
1859  AttributeList CalleeAttrs = CS.getAttributes();
1860  if (AttrBuilder(CalleeAttrs, AttributeList::ReturnIndex)
1861  .removeAttribute(Attribute::NoAlias) !=
1863  .removeAttribute(Attribute::NoAlias))
1864  continue;
1865 
1866  // Make sure the call instruction is followed by an unconditional branch to
1867  // the return block.
1868  BasicBlock *CallBB = CI->getParent();
1869  BranchInst *BI = dyn_cast<BranchInst>(CallBB->getTerminator());
1870  if (!BI || !BI->isUnconditional() || BI->getSuccessor(0) != BB)
1871  continue;
1872 
1873  // Duplicate the return into CallBB.
1874  (void)FoldReturnIntoUncondBranch(RetI, BB, CallBB);
1875  ModifiedDT = Changed = true;
1876  ++NumRetsDup;
1877  }
1878 
1879  // If we eliminated all predecessors of the block, delete the block now.
1880  if (Changed && !BB->hasAddressTaken() && pred_begin(BB) == pred_end(BB))
1881  BB->eraseFromParent();
1882 
1883  return Changed;
1884 }
1885 
1886 //===----------------------------------------------------------------------===//
1887 // Memory Optimization
1888 //===----------------------------------------------------------------------===//
1889 
1890 namespace {
1891 
1892 /// This is an extended version of TargetLowering::AddrMode
1893 /// which holds actual Value*'s for register values.
1894 struct ExtAddrMode : public TargetLowering::AddrMode {
1895  Value *BaseReg = nullptr;
1896  Value *ScaledReg = nullptr;
1897  Value *OriginalValue = nullptr;
1898 
1899  enum FieldName {
1900  NoField = 0x00,
1901  BaseRegField = 0x01,
1902  BaseGVField = 0x02,
1903  BaseOffsField = 0x04,
1904  ScaledRegField = 0x08,
1905  ScaleField = 0x10,
1906  MultipleFields = 0xff
1907  };
1908 
1909  ExtAddrMode() = default;
1910 
1911  void print(raw_ostream &OS) const;
1912  void dump() const;
1913 
1914  FieldName compare(const ExtAddrMode &other) {
1915  // First check that the types are the same on each field, as differing types
1916  // is something we can't cope with later on.
1917  if (BaseReg && other.BaseReg &&
1918  BaseReg->getType() != other.BaseReg->getType())
1919  return MultipleFields;
1920  if (BaseGV && other.BaseGV &&
1921  BaseGV->getType() != other.BaseGV->getType())
1922  return MultipleFields;
1923  if (ScaledReg && other.ScaledReg &&
1924  ScaledReg->getType() != other.ScaledReg->getType())
1925  return MultipleFields;
1926 
1927  // Check each field to see if it differs.
1928  unsigned Result = NoField;
1929  if (BaseReg != other.BaseReg)
1930  Result |= BaseRegField;
1931  if (BaseGV != other.BaseGV)
1932  Result |= BaseGVField;
1933  if (BaseOffs != other.BaseOffs)
1934  Result |= BaseOffsField;
1935  if (ScaledReg != other.ScaledReg)
1936  Result |= ScaledRegField;
1937  // Don't count 0 as being a different scale, because that actually means
1938  // unscaled (which will already be counted by having no ScaledReg).
1939  if (Scale && other.Scale && Scale != other.Scale)
1940  Result |= ScaleField;
1941 
1942  if (countPopulation(Result) > 1)
1943  return MultipleFields;
1944  else
1945  return static_cast<FieldName>(Result);
1946  }
1947 
1948  // An AddrMode is trivial if it involves no calculation i.e. it is just a base
1949  // with no offset.
1950  bool isTrivial() {
1951  // An AddrMode is (BaseGV + BaseReg + BaseOffs + ScaleReg * Scale) so it is
1952  // trivial if at most one of these terms is nonzero, except that BaseGV and
1953  // BaseReg both being zero actually means a null pointer value, which we
1954  // consider to be 'non-zero' here.
1955  return !BaseOffs && !Scale && !(BaseGV && BaseReg);
1956  }
1957 
1958  Value *GetFieldAsValue(FieldName Field, Type *IntPtrTy) {
1959  switch (Field) {
1960  default:
1961  return nullptr;
1962  case BaseRegField:
1963  return BaseReg;
1964  case BaseGVField:
1965  return BaseGV;
1966  case ScaledRegField:
1967  return ScaledReg;
1968  case BaseOffsField:
1969  return ConstantInt::get(IntPtrTy, BaseOffs);
1970  }
1971  }
1972 
1973  void SetCombinedField(FieldName Field, Value *V,
1974  const SmallVectorImpl<ExtAddrMode> &AddrModes) {
1975  switch (Field) {
1976  default:
1977  llvm_unreachable("Unhandled fields are expected to be rejected earlier");
1978  break;
1979  case ExtAddrMode::BaseRegField:
1980  BaseReg = V;
1981  break;
1982  case ExtAddrMode::BaseGVField:
1983  // A combined BaseGV is an Instruction, not a GlobalValue, so it goes
1984  // in the BaseReg field.
1985  assert(BaseReg == nullptr);
1986  BaseReg = V;
1987  BaseGV = nullptr;
1988  break;
1989  case ExtAddrMode::ScaledRegField:
1990  ScaledReg = V;
1991  // If we have a mix of scaled and unscaled addrmodes then we want scale
1992  // to be the scale and not zero.
1993  if (!Scale)
1994  for (const ExtAddrMode &AM : AddrModes)
1995  if (AM.Scale) {
1996  Scale = AM.Scale;
1997  break;
1998  }
1999  break;
2000  case ExtAddrMode::BaseOffsField:
2001  // The offset is no longer a constant, so it goes in ScaledReg with a
2002  // scale of 1.
2003  assert(ScaledReg == nullptr);
2004  ScaledReg = V;
2005  Scale = 1;
2006  BaseOffs = 0;
2007  break;
2008  }
2009  }
2010 };
2011 
2012 } // end anonymous namespace
2013 
2014 #ifndef NDEBUG
2015 static inline raw_ostream &operator<<(raw_ostream &OS, const ExtAddrMode &AM) {
2016  AM.print(OS);
2017  return OS;
2018 }
2019 #endif
2020 
2021 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
2022 void ExtAddrMode::print(raw_ostream &OS) const {
2023  bool NeedPlus = false;
2024  OS << "[";
2025  if (BaseGV) {
2026  OS << (NeedPlus ? " + " : "")
2027  << "GV:";
2028  BaseGV->printAsOperand(OS, /*PrintType=*/false);
2029  NeedPlus = true;
2030  }
2031 
2032  if (BaseOffs) {
2033  OS << (NeedPlus ? " + " : "")
2034  << BaseOffs;
2035  NeedPlus = true;
2036  }
2037 
2038  if (BaseReg) {
2039  OS << (NeedPlus ? " + " : "")
2040  << "Base:";
2041  BaseReg->printAsOperand(OS, /*PrintType=*/false);
2042  NeedPlus = true;
2043  }
2044  if (Scale) {
2045  OS << (NeedPlus ? " + " : "")
2046  << Scale << "*";
2047  ScaledReg->printAsOperand(OS, /*PrintType=*/false);
2048  }
2049 
2050  OS << ']';
2051 }
2052 
2053 LLVM_DUMP_METHOD void ExtAddrMode::dump() const {
2054  print(dbgs());
2055  dbgs() << '\n';
2056 }
2057 #endif
2058 
2059 namespace {
2060 
2061 /// This class provides transaction based operation on the IR.
2062 /// Every change made through this class is recorded in the internal state and
2063 /// can be undone (rollback) until commit is called.
2064 class TypePromotionTransaction {
2065  /// This represents the common interface of the individual transaction.
2066  /// Each class implements the logic for doing one specific modification on
2067  /// the IR via the TypePromotionTransaction.
2068  class TypePromotionAction {
2069  protected:
2070  /// The Instruction modified.
2071  Instruction *Inst;
2072 
2073  public:
2074  /// Constructor of the action.
2075  /// The constructor performs the related action on the IR.
2076  TypePromotionAction(Instruction *Inst) : Inst(Inst) {}
2077 
2078  virtual ~TypePromotionAction() = default;
2079 
2080  /// Undo the modification done by this action.
2081  /// When this method is called, the IR must be in the same state as it was
2082  /// before this action was applied.
2083  /// \pre Undoing the action works if and only if the IR is in the exact same
2084  /// state as it was directly after this action was applied.
2085  virtual void undo() = 0;
2086 
2087  /// Advocate every change made by this action.
2088  /// When the results on the IR of the action are to be kept, it is important
2089  /// to call this function, otherwise hidden information may be kept forever.
2090  virtual void commit() {
2091  // Nothing to be done, this action is not doing anything.
2092  }
2093  };
2094 
2095  /// Utility to remember the position of an instruction.
2096  class InsertionHandler {
2097  /// Position of an instruction.
2098  /// Either an instruction:
2099  /// - Is the first in a basic block: BB is used.
2100  /// - Has a previous instruction: PrevInst is used.
2101  union {
2102  Instruction *PrevInst;
2103  BasicBlock *BB;
2104  } Point;
2105 
2106  /// Remember whether or not the instruction had a previous instruction.
2107  bool HasPrevInstruction;
2108 
2109  public:
2110  /// Record the position of \p Inst.
2111  InsertionHandler(Instruction *Inst) {
2112  BasicBlock::iterator It = Inst->getIterator();
2113  HasPrevInstruction = (It != (Inst->getParent()->begin()));
2114  if (HasPrevInstruction)
2115  Point.PrevInst = &*--It;
2116  else
2117  Point.BB = Inst->getParent();
2118  }
2119 
2120  /// Insert \p Inst at the recorded position.
2121  void insert(Instruction *Inst) {
2122  if (HasPrevInstruction) {
2123  if (Inst->getParent())
2124  Inst->removeFromParent();
2125  Inst->insertAfter(Point.PrevInst);
2126  } else {
2127  Instruction *Position = &*Point.BB->getFirstInsertionPt();
2128  if (Inst->getParent())
2129  Inst->moveBefore(Position);
2130  else
2131  Inst->insertBefore(Position);
2132  }
2133  }
2134  };
2135 
2136  /// Move an instruction before another.
2137  class InstructionMoveBefore : public TypePromotionAction {
2138  /// Original position of the instruction.
2139  InsertionHandler Position;
2140 
2141  public:
2142  /// Move \p Inst before \p Before.
2143  InstructionMoveBefore(Instruction *Inst, Instruction *Before)
2144  : TypePromotionAction(Inst), Position(Inst) {
2145  LLVM_DEBUG(dbgs() << "Do: move: " << *Inst << "\nbefore: " << *Before
2146  << "\n");
2147  Inst->moveBefore(Before);
2148  }
2149 
2150  /// Move the instruction back to its original position.
2151  void undo() override {
2152  LLVM_DEBUG(dbgs() << "Undo: moveBefore: " << *Inst << "\n");
2153  Position.insert(Inst);
2154  }
2155  };
2156 
2157  /// Set the operand of an instruction with a new value.
2158  class OperandSetter : public TypePromotionAction {
2159  /// Original operand of the instruction.
2160  Value *Origin;
2161 
2162  /// Index of the modified instruction.
2163  unsigned Idx;
2164 
2165  public:
2166  /// Set \p Idx operand of \p Inst with \p NewVal.
2167  OperandSetter(Instruction *Inst, unsigned Idx, Value *NewVal)
2168  : TypePromotionAction(Inst), Idx(Idx) {
2169  LLVM_DEBUG(dbgs() << "Do: setOperand: " << Idx << "\n"
2170  << "for:" << *Inst << "\n"
2171  << "with:" << *NewVal << "\n");
2172  Origin = Inst->getOperand(Idx);
2173  Inst->setOperand(Idx, NewVal);
2174  }
2175 
2176  /// Restore the original value of the instruction.
2177  void undo() override {
2178  LLVM_DEBUG(dbgs() << "Undo: setOperand:" << Idx << "\n"
2179  << "for: " << *Inst << "\n"
2180  << "with: " << *Origin << "\n");
2181  Inst->setOperand(Idx, Origin);
2182  }
2183  };
2184 
2185  /// Hide the operands of an instruction.
2186  /// Do as if this instruction was not using any of its operands.
2187  class OperandsHider : public TypePromotionAction {
2188  /// The list of original operands.
2189  SmallVector<Value *, 4> OriginalValues;
2190 
2191  public:
2192  /// Remove \p Inst from the uses of the operands of \p Inst.
2193  OperandsHider(Instruction *Inst) : TypePromotionAction(Inst) {
2194  LLVM_DEBUG(dbgs() << "Do: OperandsHider: " << *Inst << "\n");
2195  unsigned NumOpnds = Inst->getNumOperands();
2196  OriginalValues.reserve(NumOpnds);
2197  for (unsigned It = 0; It < NumOpnds; ++It) {
2198  // Save the current operand.
2199  Value *Val = Inst->getOperand(It);
2200  OriginalValues.push_back(Val);
2201  // Set a dummy one.
2202  // We could use OperandSetter here, but that would imply an overhead
2203  // that we are not willing to pay.
2204  Inst->setOperand(It, UndefValue::get(Val->getType()));
2205  }
2206  }
2207 
2208  /// Restore the original list of uses.
2209  void undo() override {
2210  LLVM_DEBUG(dbgs() << "Undo: OperandsHider: " << *Inst << "\n");
2211  for (unsigned It = 0, EndIt = OriginalValues.size(); It != EndIt; ++It)
2212  Inst->setOperand(It, OriginalValues[It]);
2213  }
2214  };
2215 
2216  /// Build a truncate instruction.
2217  class TruncBuilder : public TypePromotionAction {
2218  Value *Val;
2219 
2220  public:
2221  /// Build a truncate instruction of \p Opnd producing a \p Ty
2222  /// result.
2223  /// trunc Opnd to Ty.
2224  TruncBuilder(Instruction *Opnd, Type *Ty) : TypePromotionAction(Opnd) {
2225  IRBuilder<> Builder(Opnd);
2226  Val = Builder.CreateTrunc(Opnd, Ty, "promoted");
2227  LLVM_DEBUG(dbgs() << "Do: TruncBuilder: " << *Val << "\n");
2228  }
2229 
2230  /// Get the built value.
2231  Value *getBuiltValue() { return Val; }
2232 
2233  /// Remove the built instruction.
2234  void undo() override {
2235  LLVM_DEBUG(dbgs() << "Undo: TruncBuilder: " << *Val << "\n");
2236  if (Instruction *IVal = dyn_cast<Instruction>(Val))
2237  IVal->eraseFromParent();
2238  }
2239  };
2240 
2241  /// Build a sign extension instruction.
2242  class SExtBuilder : public TypePromotionAction {
2243  Value *Val;
2244 
2245  public:
2246  /// Build a sign extension instruction of \p Opnd producing a \p Ty
2247  /// result.
2248  /// sext Opnd to Ty.
2249  SExtBuilder(Instruction *InsertPt, Value *Opnd, Type *Ty)
2250  : TypePromotionAction(InsertPt) {
2251  IRBuilder<> Builder(InsertPt);
2252  Val = Builder.CreateSExt(Opnd, Ty, "promoted");
2253  LLVM_DEBUG(dbgs() << "Do: SExtBuilder: " << *Val << "\n");
2254  }
2255 
2256  /// Get the built value.
2257  Value *getBuiltValue() { return Val; }
2258 
2259  /// Remove the built instruction.
2260  void undo() override {
2261  LLVM_DEBUG(dbgs() << "Undo: SExtBuilder: " << *Val << "\n");
2262  if (Instruction *IVal = dyn_cast<Instruction>(Val))
2263  IVal->eraseFromParent();
2264  }
2265  };
2266 
2267  /// Build a zero extension instruction.
2268  class ZExtBuilder : public TypePromotionAction {
2269  Value *Val;
2270 
2271  public:
2272  /// Build a zero extension instruction of \p Opnd producing a \p Ty
2273  /// result.
2274  /// zext Opnd to Ty.
2275  ZExtBuilder(Instruction *InsertPt, Value *Opnd, Type *Ty)
2276  : TypePromotionAction(InsertPt) {
2277  IRBuilder<> Builder(InsertPt);
2278  Val = Builder.CreateZExt(Opnd, Ty, "promoted");
2279  LLVM_DEBUG(dbgs() << "Do: ZExtBuilder: " << *Val << "\n");
2280  }
2281 
2282  /// Get the built value.
2283  Value *getBuiltValue() { return Val; }
2284 
2285  /// Remove the built instruction.
2286  void undo() override {
2287  LLVM_DEBUG(dbgs() << "Undo: ZExtBuilder: " << *Val << "\n");
2288  if (Instruction *IVal = dyn_cast<Instruction>(Val))
2289  IVal->eraseFromParent();
2290  }
2291  };
2292 
2293  /// Mutate an instruction to another type.
2294  class TypeMutator : public TypePromotionAction {
2295  /// Record the original type.
2296  Type *OrigTy;
2297 
2298  public:
2299  /// Mutate the type of \p Inst into \p NewTy.
2300  TypeMutator(Instruction *Inst, Type *NewTy)
2301  : TypePromotionAction(Inst), OrigTy(Inst->getType()) {
2302  LLVM_DEBUG(dbgs() << "Do: MutateType: " << *Inst << " with " << *NewTy
2303  << "\n");
2304  Inst->mutateType(NewTy);
2305  }
2306 
2307  /// Mutate the instruction back to its original type.
2308  void undo() override {
2309  LLVM_DEBUG(dbgs() << "Undo: MutateType: " << *Inst << " with " << *OrigTy
2310  << "\n");
2311  Inst->mutateType(OrigTy);
2312  }
2313  };
2314 
2315  /// Replace the uses of an instruction by another instruction.
2316  class UsesReplacer : public TypePromotionAction {
2317  /// Helper structure to keep track of the replaced uses.
2318  struct InstructionAndIdx {
2319  /// The instruction using the instruction.
2320  Instruction *Inst;
2321 
2322  /// The index where this instruction is used for Inst.
2323  unsigned Idx;
2324 
2325  InstructionAndIdx(Instruction *Inst, unsigned Idx)
2326  : Inst(Inst), Idx(Idx) {}
2327  };
2328 
2329  /// Keep track of the original uses (pair Instruction, Index).
2330  SmallVector<InstructionAndIdx, 4> OriginalUses;
2331 
2332  using use_iterator = SmallVectorImpl<InstructionAndIdx>::iterator;
2333 
2334  public:
2335  /// Replace all the use of \p Inst by \p New.
2336  UsesReplacer(Instruction *Inst, Value *New) : TypePromotionAction(Inst) {
2337  LLVM_DEBUG(dbgs() << "Do: UsersReplacer: " << *Inst << " with " << *New
2338  << "\n");
2339  // Record the original uses.
2340  for (Use &U : Inst->uses()) {
2341  Instruction *UserI = cast<Instruction>(U.getUser());
2342  OriginalUses.push_back(InstructionAndIdx(UserI, U.getOperandNo()));
2343  }
2344  // Now, we can replace the uses.
2345  Inst->replaceAllUsesWith(New);
2346  }
2347 
2348  /// Reassign the original uses of Inst to Inst.
2349  void undo() override {
2350  LLVM_DEBUG(dbgs() << "Undo: UsersReplacer: " << *Inst << "\n");
2351  for (use_iterator UseIt = OriginalUses.begin(),
2352  EndIt = OriginalUses.end();
2353  UseIt != EndIt; ++UseIt) {
2354  UseIt->Inst->setOperand(UseIt->Idx, Inst);
2355  }
2356  }
2357  };
2358 
2359  /// Remove an instruction from the IR.
2360  class InstructionRemover : public TypePromotionAction {
2361  /// Original position of the instruction.
2362  InsertionHandler Inserter;
2363 
2364  /// Helper structure to hide all the link to the instruction. In other
2365  /// words, this helps to do as if the instruction was removed.
2366  OperandsHider Hider;
2367 
2368  /// Keep track of the uses replaced, if any.
2369  UsesReplacer *Replacer = nullptr;
2370 
2371  /// Keep track of instructions removed.
2372  SetOfInstrs &RemovedInsts;
2373 
2374  public:
2375  /// Remove all reference of \p Inst and optionally replace all its
2376  /// uses with New.
2377  /// \p RemovedInsts Keep track of the instructions removed by this Action.
2378  /// \pre If !Inst->use_empty(), then New != nullptr
2379  InstructionRemover(Instruction *Inst, SetOfInstrs &RemovedInsts,
2380  Value *New = nullptr)
2381  : TypePromotionAction(Inst), Inserter(Inst), Hider(Inst),
2382  RemovedInsts(RemovedInsts) {
2383  if (New)
2384  Replacer = new UsesReplacer(Inst, New);
2385  LLVM_DEBUG(dbgs() << "Do: InstructionRemover: " << *Inst << "\n");
2386  RemovedInsts.insert(Inst);
2387  /// The instructions removed here will be freed after completing
2388  /// optimizeBlock() for all blocks as we need to keep track of the
2389  /// removed instructions during promotion.
2390  Inst->removeFromParent();
2391  }
2392 
2393  ~InstructionRemover() override { delete Replacer; }
2394 
2395  /// Resurrect the instruction and reassign it to the proper uses if
2396  /// new value was provided when build this action.
2397  void undo() override {
2398  LLVM_DEBUG(dbgs() << "Undo: InstructionRemover: " << *Inst << "\n");
2399  Inserter.insert(Inst);
2400  if (Replacer)
2401  Replacer->undo();
2402  Hider.undo();
2403  RemovedInsts.erase(Inst);
2404  }
2405  };
2406 
2407 public:
2408  /// Restoration point.
2409  /// The restoration point is a pointer to an action instead of an iterator
2410  /// because the iterator may be invalidated but not the pointer.
2411  using ConstRestorationPt = const TypePromotionAction *;
2412 
2413  TypePromotionTransaction(SetOfInstrs &RemovedInsts)
2414  : RemovedInsts(RemovedInsts) {}
2415 
2416  /// Advocate every changes made in that transaction.
2417  void commit();
2418 
2419  /// Undo all the changes made after the given point.
2420  void rollback(ConstRestorationPt Point);
2421 
2422  /// Get the current restoration point.
2423  ConstRestorationPt getRestorationPoint() const;
2424 
2425  /// \name API for IR modification with state keeping to support rollback.
2426  /// @{
2427  /// Same as Instruction::setOperand.
2428  void setOperand(Instruction *Inst, unsigned Idx, Value *NewVal);
2429 
2430  /// Same as Instruction::eraseFromParent.
2431  void eraseInstruction(Instruction *Inst, Value *NewVal = nullptr);
2432 
2433  /// Same as Value::replaceAllUsesWith.
2434  void replaceAllUsesWith(Instruction *Inst, Value *New);
2435 
2436  /// Same as Value::mutateType.
2437  void mutateType(Instruction *Inst, Type *NewTy);
2438 
2439  /// Same as IRBuilder::createTrunc.
2440  Value *createTrunc(Instruction *Opnd, Type *Ty);
2441 
2442  /// Same as IRBuilder::createSExt.
2443  Value *createSExt(Instruction *Inst, Value *Opnd, Type *Ty);
2444 
2445  /// Same as IRBuilder::createZExt.
2446  Value *createZExt(Instruction *Inst, Value *Opnd, Type *Ty);
2447 
2448  /// Same as Instruction::moveBefore.
2449  void moveBefore(Instruction *Inst, Instruction *Before);
2450  /// @}
2451 
2452 private:
2453  /// The ordered list of actions made so far.
2455 
2456  using CommitPt = SmallVectorImpl<std::unique_ptr<TypePromotionAction>>::iterator;
2457 
2458  SetOfInstrs &RemovedInsts;
2459 };
2460 
2461 } // end anonymous namespace
2462 
2463 void TypePromotionTransaction::setOperand(Instruction *Inst, unsigned Idx,
2464  Value *NewVal) {
2465  Actions.push_back(llvm::make_unique<TypePromotionTransaction::OperandSetter>(
2466  Inst, Idx, NewVal));
2467 }
2468 
2469 void TypePromotionTransaction::eraseInstruction(Instruction *Inst,
2470  Value *NewVal) {
2471  Actions.push_back(
2472  llvm::make_unique<TypePromotionTransaction::InstructionRemover>(
2473  Inst, RemovedInsts, NewVal));
2474 }
2475 
2476 void TypePromotionTransaction::replaceAllUsesWith(Instruction *Inst,
2477  Value *New) {
2478  Actions.push_back(
2479  llvm::make_unique<TypePromotionTransaction::UsesReplacer>(Inst, New));
2480 }
2481 
2482 void TypePromotionTransaction::mutateType(Instruction *Inst, Type *NewTy) {
2483  Actions.push_back(
2484  llvm::make_unique<TypePromotionTransaction::TypeMutator>(Inst, NewTy));
2485 }
2486 
2487 Value *TypePromotionTransaction::createTrunc(Instruction *Opnd,
2488  Type *Ty) {
2489  std::unique_ptr<TruncBuilder> Ptr(new TruncBuilder(Opnd, Ty));
2490  Value *Val = Ptr->getBuiltValue();
2491  Actions.push_back(std::move(Ptr));
2492  return Val;
2493 }
2494 
2495 Value *TypePromotionTransaction::createSExt(Instruction *Inst,
2496  Value *Opnd, Type *Ty) {
2497  std::unique_ptr<SExtBuilder> Ptr(new SExtBuilder(Inst, Opnd, Ty));
2498  Value *Val = Ptr->getBuiltValue();
2499  Actions.push_back(std::move(Ptr));
2500  return Val;
2501 }
2502 
2503 Value *TypePromotionTransaction::createZExt(Instruction *Inst,
2504  Value *Opnd, Type *Ty) {
2505  std::unique_ptr<ZExtBuilder> Ptr(new ZExtBuilder(Inst, Opnd, Ty));
2506  Value *Val = Ptr->getBuiltValue();
2507  Actions.push_back(std::move(Ptr));
2508  return Val;
2509 }
2510 
2511 void TypePromotionTransaction::moveBefore(Instruction *Inst,
2512  Instruction *Before) {
2513  Actions.push_back(
2514  llvm::make_unique<TypePromotionTransaction::InstructionMoveBefore>(
2515  Inst, Before));
2516 }
2517 
2518 TypePromotionTransaction::ConstRestorationPt
2519 TypePromotionTransaction::getRestorationPoint() const {
2520  return !Actions.empty() ? Actions.back().get() : nullptr;
2521 }
2522 
2523 void TypePromotionTransaction::commit() {
2524  for (CommitPt It = Actions.begin(), EndIt = Actions.end(); It != EndIt;
2525  ++It)
2526  (*It)->commit();
2527  Actions.clear();
2528 }
2529 
2530 void TypePromotionTransaction::rollback(
2531  TypePromotionTransaction::ConstRestorationPt Point) {
2532  while (!Actions.empty() && Point != Actions.back().get()) {
2533  std::unique_ptr<TypePromotionAction> Curr = Actions.pop_back_val();
2534  Curr->undo();
2535  }
2536 }
2537 
2538 namespace {
2539 
2540 /// A helper class for matching addressing modes.
2541 ///
2542 /// This encapsulates the logic for matching the target-legal addressing modes.
2543 class AddressingModeMatcher {
2544  SmallVectorImpl<Instruction*> &AddrModeInsts;
2545  const TargetLowering &TLI;
2546  const TargetRegisterInfo &TRI;
2547  const DataLayout &DL;
2548 
2549  /// AccessTy/MemoryInst - This is the type for the access (e.g. double) and
2550  /// the memory instruction that we're computing this address for.
2551  Type *AccessTy;
2552  unsigned AddrSpace;
2553  Instruction *MemoryInst;
2554 
2555  /// This is the addressing mode that we're building up. This is
2556  /// part of the return value of this addressing mode matching stuff.
2557  ExtAddrMode &AddrMode;
2558 
2559  /// The instructions inserted by other CodeGenPrepare optimizations.
2560  const SetOfInstrs &InsertedInsts;
2561 
2562  /// A map from the instructions to their type before promotion.
2563  InstrToOrigTy &PromotedInsts;
2564 
2565  /// The ongoing transaction where every action should be registered.
2566  TypePromotionTransaction &TPT;
2567 
2568  // A GEP which has too large offset to be folded into the addressing mode.
2569  std::pair<AssertingVH<GetElementPtrInst>, int64_t> &LargeOffsetGEP;
2570 
2571  /// This is set to true when we should not do profitability checks.
2572  /// When true, IsProfitableToFoldIntoAddressingMode always returns true.
2573  bool IgnoreProfitability;
2574 
2575  AddressingModeMatcher(
2577  const TargetRegisterInfo &TRI, Type *AT, unsigned AS, Instruction *MI,
2578  ExtAddrMode &AM, const SetOfInstrs &InsertedInsts,
2579  InstrToOrigTy &PromotedInsts, TypePromotionTransaction &TPT,
2580  std::pair<AssertingVH<GetElementPtrInst>, int64_t> &LargeOffsetGEP)
2581  : AddrModeInsts(AMI), TLI(TLI), TRI(TRI),
2582  DL(MI->getModule()->getDataLayout()), AccessTy(AT), AddrSpace(AS),
2583  MemoryInst(MI), AddrMode(AM), InsertedInsts(InsertedInsts),
2584  PromotedInsts(PromotedInsts), TPT(TPT), LargeOffsetGEP(LargeOffsetGEP) {
2585  IgnoreProfitability = false;
2586  }
2587 
2588 public:
2589  /// Find the maximal addressing mode that a load/store of V can fold,
2590  /// give an access type of AccessTy. This returns a list of involved
2591  /// instructions in AddrModeInsts.
2592  /// \p InsertedInsts The instructions inserted by other CodeGenPrepare
2593  /// optimizations.
2594  /// \p PromotedInsts maps the instructions to their type before promotion.
2595  /// \p The ongoing transaction where every action should be registered.
2596  static ExtAddrMode
2597  Match(Value *V, Type *AccessTy, unsigned AS, Instruction *MemoryInst,
2598  SmallVectorImpl<Instruction *> &AddrModeInsts,
2599  const TargetLowering &TLI, const TargetRegisterInfo &TRI,
2600  const SetOfInstrs &InsertedInsts, InstrToOrigTy &PromotedInsts,
2601  TypePromotionTransaction &TPT,
2602  std::pair<AssertingVH<GetElementPtrInst>, int64_t> &LargeOffsetGEP) {
2603  ExtAddrMode Result;
2604 
2605  bool Success = AddressingModeMatcher(AddrModeInsts, TLI, TRI, AccessTy, AS,
2606  MemoryInst, Result, InsertedInsts,
2607  PromotedInsts, TPT, LargeOffsetGEP)
2608  .matchAddr(V, 0);
2609  (void)Success; assert(Success && "Couldn't select *anything*?");
2610  return Result;
2611  }
2612 
2613 private:
2614  bool matchScaledValue(Value *ScaleReg, int64_t Scale, unsigned Depth);
2615  bool matchAddr(Value *Addr, unsigned Depth);
2616  bool matchOperationAddr(User *AddrInst, unsigned Opcode, unsigned Depth,
2617  bool *MovedAway = nullptr);
2618  bool isProfitableToFoldIntoAddressingMode(Instruction *I,
2619  ExtAddrMode &AMBefore,
2620  ExtAddrMode &AMAfter);
2621  bool valueAlreadyLiveAtInst(Value *Val, Value *KnownLive1, Value *KnownLive2);
2622  bool isPromotionProfitable(unsigned NewCost, unsigned OldCost,
2623  Value *PromotedOperand) const;
2624 };
2625 
2626 /// Keep track of simplification of Phi nodes.
2627 /// Accept the set of all phi nodes and erase phi node from this set
2628 /// if it is simplified.
2629 class SimplificationTracker {
2631  const SimplifyQuery &SQ;
2632  // Tracks newly created Phi nodes. We use a SetVector to get deterministic
2633  // order when iterating over the set in MatchPhiSet.
2634  SmallSetVector<PHINode *, 32> AllPhiNodes;
2635  // Tracks newly created Select nodes.
2636  SmallPtrSet<SelectInst *, 32> AllSelectNodes;
2637 
2638 public:
2639  SimplificationTracker(const SimplifyQuery &sq)
2640  : SQ(sq) {}
2641 
2642  Value *Get(Value *V) {
2643  do {
2644  auto SV = Storage.find(V);
2645  if (SV == Storage.end())
2646  return V;
2647  V = SV->second;
2648  } while (true);
2649  }
2650 
2651  Value *Simplify(Value *Val) {
2652  SmallVector<Value *, 32> WorkList;
2653  SmallPtrSet<Value *, 32> Visited;
2654  WorkList.push_back(Val);
2655  while (!WorkList.empty()) {
2656  auto P = WorkList.pop_back_val();
2657  if (!Visited.insert(P).second)
2658  continue;
2659  if (auto *PI = dyn_cast<Instruction>(P))
2660  if (Value *V = SimplifyInstruction(cast<Instruction>(PI), SQ)) {
2661  for (auto *U : PI->users())
2662  WorkList.push_back(cast<Value>(U));
2663  Put(PI, V);
2664  PI->replaceAllUsesWith(V);
2665  if (auto *PHI = dyn_cast<PHINode>(PI))
2666  AllPhiNodes.remove(PHI);
2667  if (auto *Select = dyn_cast<SelectInst>(PI))
2668  AllSelectNodes.erase(Select);
2669  PI->eraseFromParent();
2670  }
2671  }
2672  return Get(Val);
2673  }
2674 
2675  void Put(Value *From, Value *To) {
2676  Storage.insert({ From, To });
2677  }
2678 
2679  void ReplacePhi(PHINode *From, PHINode *To) {
2680  Value* OldReplacement = Get(From);
2681  while (OldReplacement != From) {
2682  From = To;
2683  To = dyn_cast<PHINode>(OldReplacement);
2684  OldReplacement = Get(From);
2685  }
2686  assert(Get(To) == To && "Replacement PHI node is already replaced.");
2687  Put(From, To);
2688  From->replaceAllUsesWith(To);
2689  AllPhiNodes.remove(From);
2690  From->eraseFromParent();
2691  }
2692 
2693  SmallSetVector<PHINode *, 32>& newPhiNodes() { return AllPhiNodes; }
2694 
2695  void insertNewPhi(PHINode *PN) { AllPhiNodes.insert(PN); }
2696 
2697  void insertNewSelect(SelectInst *SI) { AllSelectNodes.insert(SI); }
2698 
2699  unsigned countNewPhiNodes() const { return AllPhiNodes.size(); }
2700 
2701  unsigned countNewSelectNodes() const { return AllSelectNodes.size(); }
2702 
2703  void destroyNewNodes(Type *CommonType) {
2704  // For safe erasing, replace the uses with dummy value first.
2705  auto Dummy = UndefValue::get(CommonType);
2706  for (auto I : AllPhiNodes) {
2708  I->eraseFromParent();
2709  }
2710  AllPhiNodes.clear();
2711  for (auto I : AllSelectNodes) {
2713  I->eraseFromParent();
2714  }
2715  AllSelectNodes.clear();
2716  }
2717 };
2718 
2719 /// A helper class for combining addressing modes.
2720 class AddressingModeCombiner {
2721  typedef std::pair<Value *, BasicBlock *> ValueInBB;
2722  typedef DenseMap<ValueInBB, Value *> FoldAddrToValueMapping;
2723  typedef std::pair<PHINode *, PHINode *> PHIPair;
2724 
2725 private:
2726  /// The addressing modes we've collected.
2727  SmallVector<ExtAddrMode, 16> AddrModes;
2728 
2729  /// The field in which the AddrModes differ, when we have more than one.
2730  ExtAddrMode::FieldName DifferentField = ExtAddrMode::NoField;
2731 
2732  /// Are the AddrModes that we have all just equal to their original values?
2733  bool AllAddrModesTrivial = true;
2734 
2735  /// Common Type for all different fields in addressing modes.
2736  Type *CommonType;
2737 
2738  /// SimplifyQuery for simplifyInstruction utility.
2739  const SimplifyQuery &SQ;
2740 
2741  /// Original Address.
2742  ValueInBB Original;
2743 
2744 public:
2745  AddressingModeCombiner(const SimplifyQuery &_SQ, ValueInBB OriginalValue)
2746  : CommonType(nullptr), SQ(_SQ), Original(OriginalValue) {}
2747 
2748  /// Get the combined AddrMode
2749  const ExtAddrMode &getAddrMode() const {
2750  return AddrModes[0];
2751  }
2752 
2753  /// Add a new AddrMode if it's compatible with the AddrModes we already
2754  /// have.
2755  /// \return True iff we succeeded in doing so.
2756  bool addNewAddrMode(ExtAddrMode &NewAddrMode) {
2757  // Take note of if we have any non-trivial AddrModes, as we need to detect
2758  // when all AddrModes are trivial as then we would introduce a phi or select
2759  // which just duplicates what's already there.
2760  AllAddrModesTrivial = AllAddrModesTrivial && NewAddrMode.isTrivial();
2761 
2762  // If this is the first addrmode then everything is fine.
2763  if (AddrModes.empty()) {
2764  AddrModes.emplace_back(NewAddrMode);
2765  return true;
2766  }
2767 
2768  // Figure out how different this is from the other address modes, which we
2769  // can do just by comparing against the first one given that we only care
2770  // about the cumulative difference.
2771  ExtAddrMode::FieldName ThisDifferentField =
2772  AddrModes[0].compare(NewAddrMode);
2773  if (DifferentField == ExtAddrMode::NoField)
2774  DifferentField = ThisDifferentField;
2775  else if (DifferentField != ThisDifferentField)
2776  DifferentField = ExtAddrMode::MultipleFields;
2777 
2778  // If NewAddrMode differs in more than one dimension we cannot handle it.
2779  bool CanHandle = DifferentField != ExtAddrMode::MultipleFields;
2780 
2781  // If Scale Field is different then we reject.
2782  CanHandle = CanHandle && DifferentField != ExtAddrMode::ScaleField;
2783 
2784  // We also must reject the case when base offset is different and
2785  // scale reg is not null, we cannot handle this case due to merge of
2786  // different offsets will be used as ScaleReg.
2787  CanHandle = CanHandle && (DifferentField != ExtAddrMode::BaseOffsField ||
2788  !NewAddrMode.ScaledReg);
2789 
2790  // We also must reject the case when GV is different and BaseReg installed
2791  // due to we want to use base reg as a merge of GV values.
2792  CanHandle = CanHandle && (DifferentField != ExtAddrMode::BaseGVField ||
2793  !NewAddrMode.HasBaseReg);
2794 
2795  // Even if NewAddMode is the same we still need to collect it due to
2796  // original value is different. And later we will need all original values
2797  // as anchors during finding the common Phi node.
2798  if (CanHandle)
2799  AddrModes.emplace_back(NewAddrMode);
2800  else
2801  AddrModes.clear();
2802 
2803  return CanHandle;
2804  }
2805 
2806  /// Combine the addressing modes we've collected into a single
2807  /// addressing mode.
2808  /// \return True iff we successfully combined them or we only had one so
2809  /// didn't need to combine them anyway.
2810  bool combineAddrModes() {
2811  // If we have no AddrModes then they can't be combined.
2812  if (AddrModes.size() == 0)
2813  return false;
2814 
2815  // A single AddrMode can trivially be combined.
2816  if (AddrModes.size() == 1 || DifferentField == ExtAddrMode::NoField)
2817  return true;
2818 
2819  // If the AddrModes we collected are all just equal to the value they are
2820  // derived from then combining them wouldn't do anything useful.
2821  if (AllAddrModesTrivial)
2822  return false;
2823 
2824  if (!addrModeCombiningAllowed())
2825  return false;
2826 
2827  // Build a map between <original value, basic block where we saw it> to
2828  // value of base register.
2829  // Bail out if there is no common type.
2830  FoldAddrToValueMapping Map;
2831  if (!initializeMap(Map))
2832  return false;
2833 
2834  Value *CommonValue = findCommon(Map);
2835  if (CommonValue)
2836  AddrModes[0].SetCombinedField(DifferentField, CommonValue, AddrModes);
2837  return CommonValue != nullptr;
2838  }
2839 
2840 private:
2841  /// Initialize Map with anchor values. For address seen in some BB
2842  /// we set the value of different field saw in this address.
2843  /// If address is not an instruction than basic block is set to null.
2844  /// At the same time we find a common type for different field we will
2845  /// use to create new Phi/Select nodes. Keep it in CommonType field.
2846  /// Return false if there is no common type found.
2847  bool initializeMap(FoldAddrToValueMapping &Map) {
2848  // Keep track of keys where the value is null. We will need to replace it
2849  // with constant null when we know the common type.
2850  SmallVector<ValueInBB, 2> NullValue;
2851  Type *IntPtrTy = SQ.DL.getIntPtrType(AddrModes[0].OriginalValue->getType());
2852  for (auto &AM : AddrModes) {
2853  BasicBlock *BB = nullptr;
2854  if (Instruction *I = dyn_cast<Instruction>(AM.OriginalValue))
2855  BB = I->getParent();
2856 
2857  Value *DV = AM.GetFieldAsValue(DifferentField, IntPtrTy);
2858  if (DV) {
2859  auto *Type = DV->getType();
2860  if (CommonType && CommonType != Type)
2861  return false;
2862  CommonType = Type;
2863  Map[{ AM.OriginalValue, BB }] = DV;
2864  } else {
2865  NullValue.push_back({ AM.OriginalValue, BB });
2866  }
2867  }
2868  assert(CommonType && "At least one non-null value must be!");
2869  for (auto VIBB : NullValue)
2870  Map[VIBB] = Constant::getNullValue(CommonType);
2871  return true;
2872  }
2873 
2874  /// We have mapping between value A and basic block where value A
2875  /// seen to other value B where B was a field in addressing mode represented
2876  /// by A. Also we have an original value C representing an address in some
2877  /// basic block. Traversing from C through phi and selects we ended up with
2878  /// A's in a map. This utility function tries to find a value V which is a
2879  /// field in addressing mode C and traversing through phi nodes and selects
2880  /// we will end up in corresponded values B in a map.
2881  /// The utility will create a new Phi/Selects if needed.
2882  // The simple example looks as follows:
2883  // BB1:
2884  // p1 = b1 + 40
2885  // br cond BB2, BB3
2886  // BB2:
2887  // p2 = b2 + 40
2888  // br BB3
2889  // BB3:
2890  // p = phi [p1, BB1], [p2, BB2]
2891  // v = load p
2892  // Map is
2893  // <p1, BB1> -> b1
2894  // <p2, BB2> -> b2
2895  // Request is
2896  // <p, BB3> -> ?
2897  // The function tries to find or build phi [b1, BB1], [b2, BB2] in BB3
2898  Value *findCommon(FoldAddrToValueMapping &Map) {
2899  // Tracks the simplification of newly created phi nodes. The reason we use
2900  // this mapping is because we will add new created Phi nodes in AddrToBase.
2901  // Simplification of Phi nodes is recursive, so some Phi node may
2902  // be simplified after we added it to AddrToBase.
2903  // Using this mapping we can find the current value in AddrToBase.
2904  SimplificationTracker ST(SQ);
2905 
2906  // First step, DFS to create PHI nodes for all intermediate blocks.
2907  // Also fill traverse order for the second step.
2908  SmallVector<ValueInBB, 32> TraverseOrder;
2909  InsertPlaceholders(Map, TraverseOrder, ST);
2910 
2911  // Second Step, fill new nodes by merged values and simplify if possible.
2912  FillPlaceholders(Map, TraverseOrder, ST);
2913 
2914  if (!AddrSinkNewSelects && ST.countNewSelectNodes() > 0) {
2915  ST.destroyNewNodes(CommonType);
2916  return nullptr;
2917  }
2918 
2919  // Now we'd like to match New Phi nodes to existed ones.
2920  unsigned PhiNotMatchedCount = 0;
2921  if (!MatchPhiSet(ST, AddrSinkNewPhis, PhiNotMatchedCount)) {
2922  ST.destroyNewNodes(CommonType);
2923  return nullptr;
2924  }
2925 
2926  auto *Result = ST.Get(Map.find(Original)->second);
2927  if (Result) {
2928  NumMemoryInstsPhiCreated += ST.countNewPhiNodes() + PhiNotMatchedCount;
2929  NumMemoryInstsSelectCreated += ST.countNewSelectNodes();
2930  }
2931  return Result;
2932  }
2933 
2934  /// Try to match PHI node to Candidate.
2935  /// Matcher tracks the matched Phi nodes.
2936  bool MatchPhiNode(PHINode *PHI, PHINode *Candidate,
2937  SmallSetVector<PHIPair, 8> &Matcher,
2938  SmallSetVector<PHINode *, 32> &PhiNodesToMatch) {
2939  SmallVector<PHIPair, 8> WorkList;
2940  Matcher.insert({ PHI, Candidate });
2941  WorkList.push_back({ PHI, Candidate });
2942  SmallSet<PHIPair, 8> Visited;
2943  while (!WorkList.empty()) {
2944  auto Item = WorkList.pop_back_val();
2945  if (!Visited.insert(Item).second)
2946  continue;
2947  // We iterate over all incoming values to Phi to compare them.
2948  // If values are different and both of them Phi and the first one is a
2949  // Phi we added (subject to match) and both of them is in the same basic
2950  // block then we can match our pair if values match. So we state that
2951  // these values match and add it to work list to verify that.
2952  for (auto B : Item.first->blocks()) {
2953  Value *FirstValue = Item.first->getIncomingValueForBlock(B);
2954  Value *SecondValue = Item.second->getIncomingValueForBlock(B);
2955  if (FirstValue == SecondValue)
2956  continue;
2957 
2958  PHINode *FirstPhi = dyn_cast<PHINode>(FirstValue);
2959  PHINode *SecondPhi = dyn_cast<PHINode>(SecondValue);
2960 
2961  // One of them is not Phi or
2962  // The first one is not Phi node from the set we'd like to match or
2963  // Phi nodes from different basic blocks then
2964  // we will not be able to match.
2965  if (!FirstPhi || !SecondPhi || !PhiNodesToMatch.count(FirstPhi) ||
2966  FirstPhi->getParent() != SecondPhi->getParent())
2967  return false;
2968 
2969  // If we already matched them then continue.
2970  if (Matcher.count({ FirstPhi, SecondPhi }))
2971  continue;
2972  // So the values are different and does not match. So we need them to
2973  // match.
2974  Matcher.insert({ FirstPhi, SecondPhi });
2975  // But me must check it.
2976  WorkList.push_back({ FirstPhi, SecondPhi });
2977  }
2978  }
2979  return true;
2980  }
2981 
2982  /// For the given set of PHI nodes (in the SimplificationTracker) try
2983  /// to find their equivalents.
2984  /// Returns false if this matching fails and creation of new Phi is disabled.
2985  bool MatchPhiSet(SimplificationTracker &ST, bool AllowNewPhiNodes,
2986  unsigned &PhiNotMatchedCount) {
2987  // Use a SetVector for Matched to make sure we do replacements (ReplacePhi)
2988  // in a deterministic order below.
2990  SmallPtrSet<PHINode *, 8> WillNotMatch;
2991  SmallSetVector<PHINode *, 32> &PhiNodesToMatch = ST.newPhiNodes();
2992  while (PhiNodesToMatch.size()) {
2993  PHINode *PHI = *PhiNodesToMatch.begin();
2994 
2995  // Add us, if no Phi nodes in the basic block we do not match.
2996  WillNotMatch.clear();
2997  WillNotMatch.insert(PHI);
2998 
2999  // Traverse all Phis until we found equivalent or fail to do that.
3000  bool IsMatched = false;
3001  for (auto &P : PHI->getParent()->phis()) {
3002  if (&P == PHI)
3003  continue;
3004  if ((IsMatched = MatchPhiNode(PHI, &P, Matched, PhiNodesToMatch)))
3005  break;
3006  // If it does not match, collect all Phi nodes from matcher.
3007  // if we end up with no match, them all these Phi nodes will not match
3008  // later.
3009  for (auto M : Matched)
3010  WillNotMatch.insert(M.first);
3011  Matched.clear();
3012  }
3013  if (IsMatched) {
3014  // Replace all matched values and erase them.
3015  for (auto MV : Matched)
3016  ST.ReplacePhi(MV.first, MV.second);
3017  Matched.clear();
3018  continue;
3019  }
3020  // If we are not allowed to create new nodes then bail out.
3021  if (!AllowNewPhiNodes)
3022  return false;
3023  // Just remove all seen values in matcher. They will not match anything.
3024  PhiNotMatchedCount += WillNotMatch.size();
3025  for (auto *P : WillNotMatch)
3026  PhiNodesToMatch.remove(P);
3027  }
3028  return true;
3029  }
3030  /// Fill the placeholder with values from predecessors and simplify it.
3031  void FillPlaceholders(FoldAddrToValueMapping &Map,
3032  SmallVectorImpl<ValueInBB> &TraverseOrder,
3033  SimplificationTracker &ST) {
3034  while (!TraverseOrder.empty()) {
3035  auto Current = TraverseOrder.pop_back_val();
3036  assert(Map.find(Current) != Map.end() && "No node to fill!!!");
3037  Value *CurrentValue = Current.first;
3038  BasicBlock *CurrentBlock = Current.second;
3039  Value *V = Map[Current];
3040 
3041  if (SelectInst *Select = dyn_cast<SelectInst>(V)) {
3042  // CurrentValue also must be Select.
3043  auto *CurrentSelect = cast<SelectInst>(CurrentValue);
3044  auto *TrueValue = CurrentSelect->getTrueValue();
3045  ValueInBB TrueItem = { TrueValue, isa<Instruction>(TrueValue)
3046  ? CurrentBlock
3047  : nullptr };
3048  assert(Map.find(TrueItem) != Map.end() && "No True Value!");
3049  Select->setTrueValue(ST.Get(Map[TrueItem]));
3050  auto *FalseValue = CurrentSelect->getFalseValue();
3051  ValueInBB FalseItem = { FalseValue, isa<Instruction>(FalseValue)
3052  ? CurrentBlock
3053  : nullptr };
3054  assert(Map.find(FalseItem) != Map.end() && "No False Value!");
3055  Select->setFalseValue(ST.Get(Map[FalseItem]));
3056  } else {
3057  // Must be a Phi node then.
3058  PHINode *PHI = cast<PHINode>(V);
3059  // Fill the Phi node with values from predecessors.
3060  bool IsDefinedInThisBB =
3061  cast<Instruction>(CurrentValue)->getParent() == CurrentBlock;
3062  auto *CurrentPhi = dyn_cast<PHINode>(CurrentValue);
3063  for (auto B : predecessors(CurrentBlock)) {
3064  Value *PV = IsDefinedInThisBB
3065  ? CurrentPhi->getIncomingValueForBlock(B)
3066  : CurrentValue;
3067  ValueInBB item = { PV, isa<Instruction>(PV) ? B : nullptr };
3068  assert(Map.find(item) != Map.end() && "No predecessor Value!");
3069  PHI->addIncoming(ST.Get(Map[item]), B);
3070  }
3071  }
3072  // Simplify if possible.
3073  Map[Current] = ST.Simplify(V);
3074  }
3075  }
3076 
3077  /// Starting from value recursively iterates over predecessors up to known
3078  /// ending values represented in a map. For each traversed block inserts
3079  /// a placeholder Phi or Select.
3080  /// Reports all new created Phi/Select nodes by adding them to set.
3081  /// Also reports and order in what basic blocks have been traversed.
3082  void InsertPlaceholders(FoldAddrToValueMapping &Map,
3083  SmallVectorImpl<ValueInBB> &TraverseOrder,
3084  SimplificationTracker &ST) {
3085  SmallVector<ValueInBB, 32> Worklist;
3086  assert((isa<PHINode>(Original.first) || isa<SelectInst>(Original.first)) &&
3087  "Address must be a Phi or Select node");
3088  auto *Dummy = UndefValue::get(CommonType);
3089  Worklist.push_back(Original);
3090  while (!Worklist.empty()) {
3091  auto Current = Worklist.pop_back_val();
3092  // If value is not an instruction it is something global, constant,
3093  // parameter and we can say that this value is observable in any block.
3094  // Set block to null to denote it.
3095  // Also please take into account that it is how we build anchors.
3096  if (!isa<Instruction>(Current.first))
3097  Current.second = nullptr;
3098  // if it is already visited or it is an ending value then skip it.
3099  if (Map.find(Current) != Map.end())
3100  continue;
3101  TraverseOrder.push_back(Current);
3102 
3103  Value *CurrentValue = Current.first;
3104  BasicBlock *CurrentBlock = Current.second;
3105  // CurrentValue must be a Phi node or select. All others must be covered
3106  // by anchors.
3107  Instruction *CurrentI = cast<Instruction>(CurrentValue);
3108  bool IsDefinedInThisBB = CurrentI->getParent() == CurrentBlock;
3109 
3110  unsigned PredCount = pred_size(CurrentBlock);
3111  // if Current Value is not defined in this basic block we are interested
3112  // in values in predecessors.
3113  if (!IsDefinedInThisBB) {
3114  assert(PredCount && "Unreachable block?!");
3115  PHINode *PHI = PHINode::Create(CommonType, PredCount, "sunk_phi",
3116  &CurrentBlock->front());
3117  Map[Current] = PHI;
3118  ST.insertNewPhi(PHI);
3119  // Add all predecessors in work list.
3120  for (auto B : predecessors(CurrentBlock))
3121  Worklist.push_back({ CurrentValue, B });
3122  continue;
3123  }
3124  // Value is defined in this basic block.
3125  if (SelectInst *OrigSelect = dyn_cast<SelectInst>(CurrentI)) {
3126  // Is it OK to get metadata from OrigSelect?!
3127  // Create a Select placeholder with dummy value.
3128  SelectInst *Select =
3129  SelectInst::Create(OrigSelect->getCondition(), Dummy, Dummy,
3130  OrigSelect->getName(), OrigSelect, OrigSelect);
3131  Map[Current] = Select;
3132  ST.insertNewSelect(Select);
3133  // We are interested in True and False value in this basic block.
3134  Worklist.push_back({ OrigSelect->getTrueValue(), CurrentBlock });
3135  Worklist.push_back({ OrigSelect->getFalseValue(), CurrentBlock });
3136  } else {
3137  // It must be a Phi node then.
3138  auto *CurrentPhi = cast<PHINode>(CurrentI);
3139  // Create new Phi node for merge of bases.
3140  assert(PredCount && "Unreachable block?!");
3141  PHINode *PHI = PHINode::Create(CommonType, PredCount, "sunk_phi",
3142  &CurrentBlock->front());
3143  Map[Current] = PHI;
3144  ST.insertNewPhi(PHI);
3145 
3146  // Add all predecessors in work list.
3147  for (auto B : predecessors(CurrentBlock))
3148  Worklist.push_back({ CurrentPhi->getIncomingValueForBlock(B), B });
3149  }
3150  }
3151  }
3152 
3153  bool addrModeCombiningAllowed() {
3155  return false;
3156  switch (DifferentField) {
3157  default:
3158  return false;
3159  case ExtAddrMode::BaseRegField:
3160  return AddrSinkCombineBaseReg;
3161  case ExtAddrMode::BaseGVField:
3162  return AddrSinkCombineBaseGV;
3163  case ExtAddrMode::BaseOffsField:
3164  return AddrSinkCombineBaseOffs;
3165  case ExtAddrMode::ScaledRegField:
3166  return AddrSinkCombineScaledReg;
3167  }
3168  }
3169 };
3170 } // end anonymous namespace
3171 
3172 /// Try adding ScaleReg*Scale to the current addressing mode.
3173 /// Return true and update AddrMode if this addr mode is legal for the target,
3174 /// false if not.
3175 bool AddressingModeMatcher::matchScaledValue(Value *ScaleReg, int64_t Scale,
3176  unsigned Depth) {
3177  // If Scale is 1, then this is the same as adding ScaleReg to the addressing
3178  // mode. Just process that directly.
3179  if (Scale == 1)
3180  return matchAddr(ScaleReg, Depth);
3181 
3182  // If the scale is 0, it takes nothing to add this.
3183  if (Scale == 0)
3184  return true;
3185 
3186  // If we already have a scale of this value, we can add to it, otherwise, we
3187  // need an available scale field.
3188  if (AddrMode.Scale != 0 && AddrMode.ScaledReg != ScaleReg)
3189  return false;
3190 
3191  ExtAddrMode TestAddrMode = AddrMode;
3192 
3193  // Add scale to turn X*4+X*3 -> X*7. This could also do things like
3194  // [A+B + A*7] -> [B+A*8].
3195  TestAddrMode.Scale += Scale;
3196  TestAddrMode.ScaledReg = ScaleReg;
3197 
3198  // If the new address isn't legal, bail out.
3199  if (!TLI.isLegalAddressingMode(DL, TestAddrMode, AccessTy, AddrSpace))
3200  return false;
3201 
3202  // It was legal, so commit it.
3203  AddrMode = TestAddrMode;
3204 
3205  // Okay, we decided that we can add ScaleReg+Scale to AddrMode. Check now
3206  // to see if ScaleReg is actually X+C. If so, we can turn this into adding
3207  // X*Scale + C*Scale to addr mode.
3208  ConstantInt *CI = nullptr; Value *AddLHS = nullptr;
3209  if (isa<Instruction>(ScaleReg) && // not a constant expr.
3210  match(ScaleReg, m_Add(m_Value(AddLHS), m_ConstantInt(CI)))) {
3211  TestAddrMode.ScaledReg = AddLHS;
3212  TestAddrMode.BaseOffs += CI->getSExtValue()*TestAddrMode.Scale;
3213 
3214  // If this addressing mode is legal, commit it and remember that we folded
3215  // this instruction.
3216  if (TLI.isLegalAddressingMode(DL, TestAddrMode, AccessTy, AddrSpace)) {
3217  AddrModeInsts.push_back(cast<Instruction>(ScaleReg));
3218  AddrMode = TestAddrMode;
3219  return true;
3220  }
3221  }
3222 
3223  // Otherwise, not (x+c)*scale, just return what we have.
3224  return true;
3225 }
3226 
3227 /// This is a little filter, which returns true if an addressing computation
3228 /// involving I might be folded into a load/store accessing it.
3229 /// This doesn't need to be perfect, but needs to accept at least
3230 /// the set of instructions that MatchOperationAddr can.
3232  switch (I->getOpcode()) {
3233  case Instruction::BitCast:
3234  case Instruction::AddrSpaceCast:
3235  // Don't touch identity bitcasts.
3236  if (I->getType() == I->getOperand(0)->getType())
3237  return false;
3238  return I->getType()->isIntOrPtrTy();
3239  case Instruction::PtrToInt:
3240  // PtrToInt is always a noop, as we know that the int type is pointer sized.
3241  return true;
3242  case Instruction::IntToPtr:
3243  // We know the input is intptr_t, so this is foldable.
3244  return true;
3245  case Instruction::Add:
3246  return true;
3247  case Instruction::Mul:
3248  case Instruction::Shl:
3249  // Can only handle X*C and X << C.
3250  return isa<ConstantInt>(I->getOperand(1));
3251  case Instruction::GetElementPtr:
3252  return true;
3253  default:
3254  return false;
3255  }
3256 }
3257 
3258 /// Check whether or not \p Val is a legal instruction for \p TLI.
3259 /// \note \p Val is assumed to be the product of some type promotion.
3260 /// Therefore if \p Val has an undefined state in \p TLI, this is assumed
3261 /// to be legal, as the non-promoted value would have had the same state.
3263  const DataLayout &DL, Value *Val) {
3264  Instruction *PromotedInst = dyn_cast<Instruction>(Val);
3265  if (!PromotedInst)
3266  return false;
3267  int ISDOpcode = TLI.InstructionOpcodeToISD(PromotedInst->getOpcode());
3268  // If the ISDOpcode is undefined, it was undefined before the promotion.
3269  if (!ISDOpcode)
3270  return true;
3271  // Otherwise, check if the promoted instruction is legal or not.
3272  return TLI.isOperationLegalOrCustom(
3273  ISDOpcode, TLI.getValueType(DL, PromotedInst->getType()));
3274 }
3275 
3276 namespace {
3277 
3278 /// Hepler class to perform type promotion.
3279 class TypePromotionHelper {
3280  /// Utility function to check whether or not a sign or zero extension
3281  /// of \p Inst with \p ConsideredExtType can be moved through \p Inst by
3282  /// either using the operands of \p Inst or promoting \p Inst.
3283  /// The type of the extension is defined by \p IsSExt.
3284  /// In other words, check if:
3285  /// ext (Ty Inst opnd1 opnd2 ... opndN) to ConsideredExtType.
3286  /// #1 Promotion applies:
3287  /// ConsideredExtType Inst (ext opnd1 to ConsideredExtType, ...).
3288  /// #2 Operand reuses:
3289  /// ext opnd1 to ConsideredExtType.
3290  /// \p PromotedInsts maps the instructions to their type before promotion.
3291  static bool canGetThrough(const Instruction *Inst, Type *ConsideredExtType,
3292  const InstrToOrigTy &PromotedInsts, bool IsSExt);
3293 
3294  /// Utility function to determine if \p OpIdx should be promoted when
3295  /// promoting \p Inst.
3296  static bool shouldExtOperand(const Instruction *Inst, int OpIdx) {
3297  return !(isa<SelectInst>(Inst) && OpIdx == 0);
3298  }
3299 
3300  /// Utility function to promote the operand of \p Ext when this
3301  /// operand is a promotable trunc or sext or zext.
3302  /// \p PromotedInsts maps the instructions to their type before promotion.
3303  /// \p CreatedInstsCost[out] contains the cost of all instructions
3304  /// created to promote the operand of Ext.
3305  /// Newly added extensions are inserted in \p Exts.
3306  /// Newly added truncates are inserted in \p Truncs.
3307  /// Should never be called directly.
3308  /// \return The promoted value which is used instead of Ext.
3309  static Value *promoteOperandForTruncAndAnyExt(
3310  Instruction *Ext, TypePromotionTransaction &TPT,
3311  InstrToOrigTy &PromotedInsts, unsigned &CreatedInstsCost,
3313  SmallVectorImpl<Instruction *> *Truncs, const TargetLowering &TLI);
3314 
3315  /// Utility function to promote the operand of \p Ext when this
3316  /// operand is promotable and is not a supported trunc or sext.
3317  /// \p PromotedInsts maps the instructions to their type before promotion.
3318  /// \p CreatedInstsCost[out] contains the cost of all the instructions
3319  /// created to promote the operand of Ext.
3320  /// Newly added extensions are inserted in \p Exts.
3321  /// Newly added truncates are inserted in \p Truncs.
3322  /// Should never be called directly.
3323  /// \return The promoted value which is used instead of Ext.
3324  static Value *promoteOperandForOther(Instruction *Ext,
3325  TypePromotionTransaction &TPT,
3326  InstrToOrigTy &PromotedInsts,
3327  unsigned &CreatedInstsCost,
3330  const TargetLowering &TLI, bool IsSExt);
3331 
3332  /// \see promoteOperandForOther.
3333  static Value *signExtendOperandForOther(
3334  Instruction *Ext, TypePromotionTransaction &TPT,
3335  InstrToOrigTy &PromotedInsts, unsigned &CreatedInstsCost,
3337  SmallVectorImpl<Instruction *> *Truncs, const TargetLowering &TLI) {
3338  return promoteOperandForOther(Ext, TPT, PromotedInsts, CreatedInstsCost,
3339  Exts, Truncs, TLI, true);
3340  }
3341 
3342  /// \see promoteOperandForOther.
3343  static Value *zeroExtendOperandForOther(
3344  Instruction *Ext, TypePromotionTransaction &TPT,
3345  InstrToOrigTy &PromotedInsts, unsigned &CreatedInstsCost,
3347  SmallVectorImpl<Instruction *> *Truncs, const TargetLowering &TLI) {
3348  return promoteOperandForOther(Ext, TPT, PromotedInsts, CreatedInstsCost,
3349  Exts, Truncs, TLI, false);
3350  }
3351 
3352 public:
3353  /// Type for the utility function that promotes the operand of Ext.
3354  using Action = Value *(*)(Instruction *Ext, TypePromotionTransaction &TPT,
3355  InstrToOrigTy &PromotedInsts,
3356  unsigned &CreatedInstsCost,
3359  const TargetLowering &TLI);
3360 
3361  /// Given a sign/zero extend instruction \p Ext, return the appropriate
3362  /// action to promote the operand of \p Ext instead of using Ext.
3363  /// \return NULL if no promotable action is possible with the current
3364  /// sign extension.
3365  /// \p InsertedInsts keeps track of all the instructions inserted by the
3366  /// other CodeGenPrepare optimizations. This information is important
3367  /// because we do not want to promote these instructions as CodeGenPrepare
3368  /// will reinsert them later. Thus creating an infinite loop: create/remove.
3369  /// \p PromotedInsts maps the instructions to their type before promotion.
3370  static Action getAction(Instruction *Ext, const SetOfInstrs &InsertedInsts,
3371  const TargetLowering &TLI,
3372  const InstrToOrigTy &PromotedInsts);
3373 };
3374 
3375 } // end anonymous namespace
3376 
3377 bool TypePromotionHelper::canGetThrough(const Instruction *Inst,
3378  Type *ConsideredExtType,
3379  const InstrToOrigTy &PromotedInsts,
3380  bool IsSExt) {
3381  // The promotion helper does not know how to deal with vector types yet.
3382  // To be able to fix that, we would need to fix the places where we
3383  // statically extend, e.g., constants and such.
3384  if (Inst->getType()->isVectorTy())
3385  return false;
3386 
3387  // We can always get through zext.
3388  if (isa<ZExtInst>(Inst))
3389  return true;
3390 
3391  // sext(sext) is ok too.
3392  if (IsSExt && isa<SExtInst>(Inst))
3393  return true;
3394 
3395  // We can get through binary operator, if it is legal. In other words, the
3396  // binary operator must have a nuw or nsw flag.
3397  const BinaryOperator *BinOp = dyn_cast<BinaryOperator>(Inst);
3398  if (BinOp && isa<OverflowingBinaryOperator>(BinOp) &&
3399  ((!IsSExt && BinOp->hasNoUnsignedWrap()) ||
3400  (IsSExt && BinOp->hasNoSignedWrap())))
3401  return true;
3402 
3403  // ext(and(opnd, cst)) --> and(ext(opnd), ext(cst))
3404  if ((Inst->getOpcode() == Instruction::And ||
3405  Inst->getOpcode() == Instruction::Or))
3406  return true;
3407 
3408  // ext(xor(opnd, cst)) --> xor(ext(opnd), ext(cst))
3409  if (Inst->getOpcode() == Instruction::Xor) {
3410  const ConstantInt *Cst = dyn_cast<ConstantInt>(Inst->getOperand(1));
3411  // Make sure it is not a NOT.
3412  if (Cst && !Cst->getValue().isAllOnesValue())
3413  return true;
3414  }
3415 
3416  // zext(shrl(opnd, cst)) --> shrl(zext(opnd), zext(cst))
3417  // It may change a poisoned value into a regular value, like
3418  // zext i32 (shrl i8 %val, 12) --> shrl i32 (zext i8 %val), 12
3419  // poisoned value regular value
3420  // It should be OK since undef covers valid value.
3421  if (Inst->getOpcode() == Instruction::LShr && !IsSExt)
3422  return true;
3423 
3424  // and(ext(shl(opnd, cst)), cst) --> and(shl(ext(opnd), ext(cst)), cst)
3425  // It may change a poisoned value into a regular value, like
3426  // zext i32 (shl i8 %val, 12) --> shl i32 (zext i8 %val), 12
3427  // poisoned value regular value
3428  // It should be OK since undef covers valid value.
3429  if (Inst->getOpcode() == Instruction::Shl && Inst->hasOneUse()) {
3430  const Instruction *ExtInst =
3431  dyn_cast<const Instruction>(*Inst->user_begin());
3432  if (ExtInst->hasOneUse()) {
3433  const Instruction *AndInst =
3434  dyn_cast<const Instruction>(*ExtInst->user_begin());
3435  if (AndInst && AndInst->getOpcode() == Instruction::And) {
3436  const ConstantInt *Cst = dyn_cast<ConstantInt>(AndInst->getOperand(1));
3437  if (Cst &&
3438  Cst->getValue().isIntN(Inst->getType()->getIntegerBitWidth()))
3439  return true;
3440  }
3441  }
3442  }
3443 
3444  // Check if we can do the following simplification.
3445  // ext(trunc(opnd)) --> ext(opnd)
3446  if (!isa<TruncInst>(Inst))
3447  return false;
3448 
3449  Value *OpndVal = Inst->getOperand(0);
3450  // Check if we can use this operand in the extension.
3451  // If the type is larger than the result type of the extension, we cannot.
3452  if (!OpndVal->getType()->isIntegerTy() ||
3453  OpndVal->getType()->getIntegerBitWidth() >
3454  ConsideredExtType->getIntegerBitWidth())
3455  return false;
3456 
3457  // If the operand of the truncate is not an instruction, we will not have
3458  // any information on the dropped bits.
3459  // (Actually we could for constant but it is not worth the extra logic).
3460  Instruction *Opnd = dyn_cast<Instruction>(OpndVal);
3461  if (!Opnd)
3462  return false;
3463 
3464  // Check if the source of the type is narrow enough.
3465  // I.e., check that trunc just drops extended bits of the same kind of
3466  // the extension.
3467  // #1 get the type of the operand and check the kind of the extended bits.
3468  const Type *OpndType;
3469  InstrToOrigTy::const_iterator It = PromotedInsts.find(Opnd);
3470  if (It != PromotedInsts.end() && It->second.getInt() == IsSExt)
3471  OpndType = It->second.getPointer();
3472  else if ((IsSExt && isa<SExtInst>(Opnd)) || (!IsSExt && isa<ZExtInst>(Opnd)))
3473  OpndType = Opnd->getOperand(0)->getType();
3474  else
3475  return false;
3476 
3477  // #2 check that the truncate just drops extended bits.
3478  return Inst->getType()->getIntegerBitWidth() >=
3479  OpndType->getIntegerBitWidth();
3480 }
3481 
3482 TypePromotionHelper::Action TypePromotionHelper::getAction(
3483  Instruction *Ext, const SetOfInstrs &InsertedInsts,
3484  const TargetLowering &TLI, const InstrToOrigTy &PromotedInsts) {
3485  assert((isa<SExtInst>(Ext) || isa<ZExtInst>(Ext)) &&
3486  "Unexpected instruction type");
3487  Instruction *ExtOpnd = dyn_cast<Instruction>(Ext->getOperand(0));
3488  Type *ExtTy = Ext->getType();
3489  bool IsSExt = isa<SExtInst>(Ext);
3490  // If the operand of the extension is not an instruction, we cannot
3491  // get through.
3492  // If it, check we can get through.
3493  if (!ExtOpnd || !canGetThrough(ExtOpnd, ExtTy, PromotedInsts, IsSExt))
3494  return nullptr;
3495 
3496  // Do not promote if the operand has been added by codegenprepare.
3497  // Otherwise, it means we are undoing an optimization that is likely to be
3498  // redone, thus causing potential infinite loop.
3499  if (isa<TruncInst>(ExtOpnd) && InsertedInsts.count(ExtOpnd))
3500  return nullptr;
3501 
3502  // SExt or Trunc instructions.
3503  // Return the related handler.
3504  if (isa<SExtInst>(ExtOpnd) || isa<TruncInst>(ExtOpnd) ||
3505  isa<ZExtInst>(ExtOpnd))
3506  return promoteOperandForTruncAndAnyExt;
3507 
3508  // Regular instruction.
3509  // Abort early if we will have to insert non-free instructions.
3510  if (!ExtOpnd->hasOneUse() && !TLI.isTruncateFree(ExtTy, ExtOpnd->getType()))
3511  return nullptr;
3512  return IsSExt ? signExtendOperandForOther : zeroExtendOperandForOther;
3513 }
3514 
3515 Value *TypePromotionHelper::promoteOperandForTruncAndAnyExt(
3516  Instruction *SExt, TypePromotionTransaction &TPT,
3517  InstrToOrigTy &PromotedInsts, unsigned &CreatedInstsCost,
3519  SmallVectorImpl<Instruction *> *Truncs, const TargetLowering &TLI) {
3520  // By construction, the operand of SExt is an instruction. Otherwise we cannot
3521  // get through it and this method should not be called.
3522  Instruction *SExtOpnd = cast<Instruction>(SExt->getOperand(0));
3523  Value *ExtVal = SExt;
3524  bool HasMergedNonFreeExt = false;
3525  if (isa<ZExtInst>(SExtOpnd)) {
3526  // Replace s|zext(zext(opnd))
3527  // => zext(opnd).
3528  HasMergedNonFreeExt = !TLI.isExtFree(SExtOpnd);
3529  Value *ZExt =
3530  TPT.createZExt(SExt, SExtOpnd->getOperand(0), SExt->getType());
3531  TPT.replaceAllUsesWith(SExt, ZExt);
3532  TPT.eraseInstruction(SExt);
3533  ExtVal = ZExt;
3534  } else {
3535  // Replace z|sext(trunc(opnd)) or sext(sext(opnd))
3536  // => z|sext(opnd).
3537  TPT.setOperand(SExt, 0, SExtOpnd->getOperand(0));
3538  }
3539  CreatedInstsCost = 0;
3540 
3541  // Remove dead code.
3542  if (SExtOpnd->use_empty())
3543  TPT.eraseInstruction(SExtOpnd);
3544 
3545  // Check if the extension is still needed.
3546  Instruction *ExtInst = dyn_cast<Instruction>(ExtVal);
3547  if (!ExtInst || ExtInst->getType() != ExtInst->getOperand(0)->getType()) {
3548  if (ExtInst) {
3549  if (Exts)
3550  Exts->push_back(ExtInst);
3551  CreatedInstsCost = !TLI.isExtFree(ExtInst) && !HasMergedNonFreeExt;
3552  }
3553  return ExtVal;
3554  }
3555 
3556  // At this point we have: ext ty opnd to ty.
3557  // Reassign the uses of ExtInst to the opnd and remove ExtInst.
3558  Value *NextVal = ExtInst->getOperand(0);
3559  TPT.eraseInstruction(ExtInst, NextVal);
3560  return NextVal;
3561 }
3562 
3563 Value *TypePromotionHelper::promoteOperandForOther(
3564  Instruction *Ext, TypePromotionTransaction &TPT,
3565  InstrToOrigTy &PromotedInsts, unsigned &CreatedInstsCost,
3567  SmallVectorImpl<Instruction *> *Truncs, const TargetLowering &TLI,
3568  bool IsSExt) {
3569  // By construction, the operand of Ext is an instruction. Otherwise we cannot
3570  // get through it and this method should not be called.
3571  Instruction *ExtOpnd = cast<Instruction>(Ext->getOperand(0));
3572  CreatedInstsCost = 0;
3573  if (!ExtOpnd->hasOneUse()) {
3574  // ExtOpnd will be promoted.
3575  // All its uses, but Ext, will need to use a truncated value of the
3576  // promoted version.
3577  // Create the truncate now.
3578  Value *Trunc = TPT.createTrunc(Ext, ExtOpnd->getType());
3579  if (Instruction *ITrunc = dyn_cast<Instruction>(Trunc)) {
3580  // Insert it just after the definition.
3581  ITrunc->moveAfter(ExtOpnd);
3582  if (Truncs)
3583  Truncs->push_back(ITrunc);
3584  }
3585 
3586  TPT.replaceAllUsesWith(ExtOpnd, Trunc);
3587  // Restore the operand of Ext (which has been replaced by the previous call
3588  // to replaceAllUsesWith) to avoid creating a cycle trunc <-> sext.
3589  TPT.setOperand(Ext, 0, ExtOpnd);
3590  }
3591 
3592  // Get through the Instruction:
3593  // 1. Update its type.
3594  // 2. Replace the uses of Ext by Inst.
3595  // 3. Extend each operand that needs to be extended.
3596 
3597  // Remember the original type of the instruction before promotion.
3598  // This is useful to know that the high bits are sign extended bits.
3599  PromotedInsts.insert(std::pair<Instruction *, TypeIsSExt>(
3600  ExtOpnd, TypeIsSExt(ExtOpnd->getType(), IsSExt)));
3601  // Step #1.
3602  TPT.mutateType(ExtOpnd, Ext->getType());
3603  // Step #2.
3604  TPT.replaceAllUsesWith(Ext, ExtOpnd);
3605  // Step #3.
3606  Instruction *ExtForOpnd = Ext;
3607 
3608  LLVM_DEBUG(dbgs() << "Propagate Ext to operands\n");
3609  for (int OpIdx = 0, EndOpIdx = ExtOpnd->getNumOperands(); OpIdx != EndOpIdx;
3610  ++OpIdx) {
3611  LLVM_DEBUG(dbgs() << "Operand:\n" << *(ExtOpnd->getOperand(OpIdx)) << '\n');
3612  if (ExtOpnd->getOperand(OpIdx)->getType() == Ext->getType() ||
3613  !shouldExtOperand(ExtOpnd, OpIdx)) {
3614  LLVM_DEBUG(dbgs() << "No need to propagate\n");
3615  continue;
3616  }
3617  // Check if we can statically extend the operand.
3618  Value *Opnd = ExtOpnd->getOperand(OpIdx);
3619  if (const ConstantInt *Cst = dyn_cast<ConstantInt>(Opnd)) {
3620  LLVM_DEBUG(dbgs() << "Statically extend\n");
3621  unsigned BitWidth = Ext->getType()->getIntegerBitWidth();
3622  APInt CstVal = IsSExt ? Cst->getValue().sext(BitWidth)
3623  : Cst->getValue().zext(BitWidth);
3624  TPT.setOperand(ExtOpnd, OpIdx, ConstantInt::get(Ext->getType(), CstVal));
3625  continue;
3626  }
3627  // UndefValue are typed, so we have to statically sign extend them.
3628  if (isa<UndefValue>(Opnd)) {
3629  LLVM_DEBUG(dbgs() << "Statically extend\n");
3630  TPT.setOperand(ExtOpnd, OpIdx, UndefValue::get(Ext->getType()));
3631  continue;
3632  }
3633 
3634  // Otherwise we have to explicitly sign extend the operand.
3635  // Check if Ext was reused to extend an operand.
3636  if (!ExtForOpnd) {
3637  // If yes, create a new one.
3638  LLVM_DEBUG(dbgs() << "More operands to ext\n");
3639  Value *ValForExtOpnd = IsSExt ? TPT.createSExt(Ext, Opnd, Ext->getType())
3640  : TPT.createZExt(Ext, Opnd, Ext->getType());
3641  if (!isa<Instruction>(ValForExtOpnd)) {
3642  TPT.setOperand(ExtOpnd, OpIdx, ValForExtOpnd);
3643  continue;
3644  }
3645  ExtForOpnd = cast<Instruction>(ValForExtOpnd);
3646  }
3647  if (Exts)
3648  Exts->push_back(ExtForOpnd);
3649  TPT.setOperand(ExtForOpnd, 0, Opnd);
3650 
3651  // Move the sign extension before the insertion point.
3652  TPT.moveBefore(ExtForOpnd, ExtOpnd);
3653  TPT.setOperand(ExtOpnd, OpIdx, ExtForOpnd);
3654  CreatedInstsCost += !TLI.isExtFree(ExtForOpnd);
3655  // If more sext are required, new instructions will have to be created.
3656  ExtForOpnd = nullptr;
3657  }
3658  if (ExtForOpnd == Ext) {
3659  LLVM_DEBUG(dbgs() << "Extension is useless now\n");
3660  TPT.eraseInstruction(Ext);
3661  }
3662  return ExtOpnd;
3663 }
3664 
3665 /// Check whether or not promoting an instruction to a wider type is profitable.
3666 /// \p NewCost gives the cost of extension instructions created by the
3667 /// promotion.
3668 /// \p OldCost gives the cost of extension instructions before the promotion
3669 /// plus the number of instructions that have been
3670 /// matched in the addressing mode the promotion.
3671 /// \p PromotedOperand is the value that has been promoted.
3672 /// \return True if the promotion is profitable, false otherwise.
3673 bool AddressingModeMatcher::isPromotionProfitable(
3674  unsigned NewCost, unsigned OldCost, Value *PromotedOperand) const {
3675  LLVM_DEBUG(dbgs() << "OldCost: " << OldCost << "\tNewCost: " << NewCost
3676  << '\n');
3677  // The cost of the new extensions is greater than the cost of the
3678  // old extension plus what we folded.
3679  // This is not profitable.
3680  if (NewCost > OldCost)
3681  return false;
3682  if (NewCost < OldCost)
3683  return true;
3684  // The promotion is neutral but it may help folding the sign extension in
3685  // loads for instance.
3686  // Check that we did not create an illegal instruction.
3687  return isPromotedInstructionLegal(TLI, DL, PromotedOperand);
3688 }
3689 
3690 /// Given an instruction or constant expr, see if we can fold the operation
3691 /// into the addressing mode. If so, update the addressing mode and return
3692 /// true, otherwise return false without modifying AddrMode.
3693 /// If \p MovedAway is not NULL, it contains the information of whether or
3694 /// not AddrInst has to be folded into the addressing mode on success.
3695 /// If \p MovedAway == true, \p AddrInst will not be part of the addressing
3696 /// because it has been moved away.
3697 /// Thus AddrInst must not be added in the matched instructions.
3698 /// This state can happen when AddrInst is a sext, since it may be moved away.
3699 /// Therefore, AddrInst may not be valid when MovedAway is true and it must
3700 /// not be referenced anymore.
3701 bool AddressingModeMatcher::matchOperationAddr(User *AddrInst, unsigned Opcode,
3702  unsigned Depth,
3703  bool *MovedAway) {
3704  // Avoid exponential behavior on extremely deep expression trees.
3705  if (Depth >= 5) return false;
3706 
3707  // By default, all matched instructions stay in place.
3708  if (MovedAway)
3709  *MovedAway = false;
3710 
3711  switch (Opcode) {
3712  case Instruction::PtrToInt:
3713  // PtrToInt is always a noop, as we know that the int type is pointer sized.
3714  return matchAddr(AddrInst->getOperand(0), Depth);
3715  case Instruction::IntToPtr: {
3716  auto AS = AddrInst->getType()->getPointerAddressSpace();
3717  auto PtrTy = MVT::getIntegerVT(DL.getPointerSizeInBits(AS));
3718  // This inttoptr is a no-op if the integer type is pointer sized.
3719  if (TLI.getValueType(DL, AddrInst->getOperand(0)->getType()) == PtrTy)
3720  return matchAddr(AddrInst->getOperand(0), Depth);
3721  return false;
3722  }
3723  case Instruction::BitCast:
3724  // BitCast is always a noop, and we can handle it as long as it is
3725  // int->int or pointer->pointer (we don't want int<->fp or something).
3726  if (AddrInst->getOperand(0)->getType()->isIntOrPtrTy() &&
3727  // Don't touch identity bitcasts. These were probably put here by LSR,
3728  // and we don't want to mess around with them. Assume it knows what it
3729  // is doing.
3730  AddrInst->getOperand(0)->getType() != AddrInst->getType())
3731  return matchAddr(AddrInst->getOperand(0), Depth);
3732  return false;
3733  case Instruction::AddrSpaceCast: {
3734  unsigned SrcAS
3735  = AddrInst->getOperand(0)->getType()->getPointerAddressSpace();
3736  unsigned DestAS = AddrInst->getType()->getPointerAddressSpace();
3737  if (TLI.isNoopAddrSpaceCast(SrcAS, DestAS))
3738  return matchAddr(AddrInst->getOperand(0), Depth);
3739  return false;
3740  }
3741  case Instruction::Add: {
3742  // Check to see if we can merge in the RHS then the LHS. If so, we win.
3743  ExtAddrMode BackupAddrMode = AddrMode;
3744  unsigned OldSize = AddrModeInsts.size();
3745  // Start a transaction at this point.
3746  // The LHS may match but not the RHS.
3747  // Therefore, we need a higher level restoration point to undo partially
3748  // matched operation.
3749  TypePromotionTransaction::ConstRestorationPt LastKnownGood =
3750  TPT.getRestorationPoint();
3751 
3752  if (matchAddr(AddrInst->getOperand(1), Depth+1) &&
3753  matchAddr(AddrInst->getOperand(0), Depth+1))
3754  return true;
3755 
3756  // Restore the old addr mode info.
3757  AddrMode = BackupAddrMode;
3758  AddrModeInsts.resize(OldSize);
3759  TPT.rollback(LastKnownGood);
3760 
3761  // Otherwise this was over-aggressive. Try merging in the LHS then the RHS.
3762  if (matchAddr(AddrInst->getOperand(0), Depth+1) &&
3763  matchAddr(AddrInst->getOperand(1), Depth+1))
3764  return true;
3765 
3766  // Otherwise we definitely can't merge the ADD in.
3767  AddrMode = BackupAddrMode;
3768  AddrModeInsts.resize(OldSize);
3769  TPT.rollback(LastKnownGood);
3770  break;
3771  }
3772  //case Instruction::Or:
3773  // TODO: We can handle "Or Val, Imm" iff this OR is equivalent to an ADD.
3774  //break;
3775  case Instruction::Mul:
3776  case Instruction::Shl: {
3777  // Can only handle X*C and X << C.
3778  ConstantInt *RHS = dyn_cast<ConstantInt>(AddrInst->getOperand(1));
3779  if (!RHS || RHS->getBitWidth() > 64)
3780  return false;
3781  int64_t Scale = RHS->getSExtValue();
3782  if (Opcode == Instruction::Shl)
3783  Scale = 1LL << Scale;
3784 
3785  return matchScaledValue(AddrInst->getOperand(0), Scale, Depth);
3786  }
3787  case Instruction::GetElementPtr: {
3788  // Scan the GEP. We check it if it contains constant offsets and at most
3789  // one variable offset.
3790  int VariableOperand = -1;
3791  unsigned VariableScale = 0;
3792 
3793  int64_t ConstantOffset = 0;
3794  gep_type_iterator GTI = gep_type_begin(AddrInst);
3795  for (unsigned i = 1, e = AddrInst->getNumOperands(); i != e; ++i, ++GTI) {
3796  if (StructType *STy = GTI.getStructTypeOrNull()) {
3797  const StructLayout *SL = DL.getStructLayout(STy);
3798  unsigned Idx =
3799  cast<ConstantInt>(AddrInst->getOperand(i))->getZExtValue();
3800  ConstantOffset += SL->getElementOffset(Idx);
3801  } else {
3802  uint64_t TypeSize = DL.getTypeAllocSize(GTI.getIndexedType());
3803  if (ConstantInt *CI = dyn_cast<ConstantInt>(AddrInst->getOperand(i))) {
3804  ConstantOffset += CI->getSExtValue() * TypeSize;
3805  } else if (TypeSize) { // Scales of zero don't do anything.
3806  // We only allow one variable index at the moment.
3807  if (VariableOperand != -1)
3808  return false;
3809 
3810  // Remember the variable index.
3811  VariableOperand = i;
3812  VariableScale = TypeSize;
3813  }
3814  }
3815  }
3816 
3817  // A common case is for the GEP to only do a constant offset. In this case,
3818  // just add it to the disp field and check validity.
3819  if (VariableOperand == -1) {
3820  AddrMode.BaseOffs += ConstantOffset;
3821  if (ConstantOffset == 0 ||
3822  TLI.isLegalAddressingMode(DL, AddrMode, AccessTy, AddrSpace)) {
3823  // Check to see if we can fold the base pointer in too.
3824  if (matchAddr(AddrInst->getOperand(0), Depth+1))
3825  return true;
3826  } else if (EnableGEPOffsetSplit && isa<GetElementPtrInst>(AddrInst) &&
3827  TLI.shouldConsiderGEPOffsetSplit() && Depth == 0 &&
3828  ConstantOffset > 0) {
3829  // Record GEPs with non-zero offsets as candidates for splitting in the
3830  // event that the offset cannot fit into the r+i addressing mode.
3831  // Simple and common case that only one GEP is used in calculating the
3832  // address for the memory access.
3833  Value *Base = AddrInst->getOperand(0);
3834  auto *BaseI = dyn_cast<Instruction>(Base);
3835  auto *GEP = cast<GetElementPtrInst>(AddrInst);
3836  if (isa<Argument>(Base) || isa<GlobalValue>(Base) ||
3837  (BaseI && !isa<CastInst>(BaseI) &&
3838  !isa<GetElementPtrInst>(BaseI))) {
3839  // If the base is an instruction, make sure the GEP is not in the same
3840  // basic block as the base. If the base is an argument or global
3841  // value, make sure the GEP is not in the entry block. Otherwise,
3842  // instruction selection can undo the split. Also make sure the
3843  // parent block allows inserting non-PHI instructions before the
3844  // terminator.
3845  BasicBlock *Parent =
3846  BaseI ? BaseI->getParent() : &GEP->getFunction()->getEntryBlock();
3847  if (GEP->getParent() != Parent && !Parent->getTerminator()->isEHPad())
3848  LargeOffsetGEP = std::make_pair(GEP, ConstantOffset);
3849  }
3850  }
3851  AddrMode.BaseOffs -= ConstantOffset;
3852  return false;
3853  }
3854 
3855  // Save the valid addressing mode in case we can't match.
3856  ExtAddrMode BackupAddrMode = AddrMode;
3857  unsigned OldSize = AddrModeInsts.size();
3858 
3859  // See if the scale and offset amount is valid for this target.
3860  AddrMode.BaseOffs += ConstantOffset;
3861 
3862  // Match the base operand of the GEP.
3863  if (!matchAddr(AddrInst->getOperand(0), Depth+1)) {
3864  // If it couldn't be matched, just stuff the value in a register.
3865  if (AddrMode.HasBaseReg) {
3866  AddrMode = BackupAddrMode;
3867  AddrModeInsts.resize(OldSize);
3868  return false;
3869  }
3870  AddrMode.HasBaseReg = true;
3871  AddrMode.BaseReg = AddrInst->getOperand(0);
3872  }
3873 
3874  // Match the remaining variable portion of the GEP.
3875  if (!matchScaledValue(AddrInst->getOperand(VariableOperand), VariableScale,
3876  Depth)) {
3877  // If it couldn't be matched, try stuffing the base into a register
3878  // instead of matching it, and retrying the match of the scale.
3879  AddrMode = BackupAddrMode;
3880  AddrModeInsts.resize(OldSize);
3881  if (AddrMode.HasBaseReg)
3882  return false;
3883  AddrMode.HasBaseReg = true;
3884  AddrMode.BaseReg = AddrInst->getOperand(0);
3885  AddrMode.BaseOffs += ConstantOffset;
3886  if (!matchScaledValue(AddrInst->getOperand(VariableOperand),
3887  VariableScale, Depth)) {
3888  // If even that didn't work, bail.
3889  AddrMode = BackupAddrMode;
3890  AddrModeInsts.resize(OldSize);
3891  return false;
3892  }
3893  }
3894 
3895  return true;
3896  }
3897  case Instruction::SExt:
3898  case Instruction::ZExt: {
3899  Instruction *Ext = dyn_cast<Instruction>(AddrInst);
3900  if (!Ext)
3901  return false;
3902 
3903  // Try to move this ext out of the way of the addressing mode.
3904  // Ask for a method for doing so.
3905  TypePromotionHelper::Action TPH =
3906  TypePromotionHelper::getAction(Ext, InsertedInsts, TLI, PromotedInsts);
3907  if (!TPH)
3908  return false;
3909 
3910  TypePromotionTransaction::ConstRestorationPt LastKnownGood =
3911  TPT.getRestorationPoint();
3912  unsigned CreatedInstsCost = 0;
3913  unsigned ExtCost = !TLI.isExtFree(Ext);
3914  Value *PromotedOperand =
3915  TPH(Ext, TPT, PromotedInsts, CreatedInstsCost, nullptr, nullptr, TLI);
3916  // SExt has been moved away.
3917  // Thus either it will be rematched later in the recursive calls or it is
3918  // gone. Anyway, we must not fold it into the addressing mode at this point.
3919  // E.g.,
3920  // op = add opnd, 1
3921  // idx = ext op
3922  // addr = gep base, idx
3923  // is now:
3924  // promotedOpnd = ext opnd <- no match here
3925  // op = promoted_add promotedOpnd, 1 <- match (later in recursive calls)
3926  // addr = gep base, op <- match
3927  if (MovedAway)
3928  *MovedAway = true;
3929 
3930  assert(PromotedOperand &&
3931  "TypePromotionHelper should have filtered out those cases");
3932 
3933  ExtAddrMode BackupAddrMode = AddrMode;
3934  unsigned OldSize = AddrModeInsts.size();
3935 
3936  if (!matchAddr(PromotedOperand, Depth) ||
3937  // The total of the new cost is equal to the cost of the created
3938  // instructions.
3939  // The total of the old cost is equal to the cost of the extension plus
3940  // what we have saved in the addressing mode.
3941  !isPromotionProfitable(CreatedInstsCost,
3942  ExtCost + (AddrModeInsts.size() - OldSize),
3943  PromotedOperand)) {
3944  AddrMode = BackupAddrMode;
3945  AddrModeInsts.resize(OldSize);
3946  LLVM_DEBUG(dbgs() << "Sign extension does not pay off: rollback\n");
3947  TPT.rollback(LastKnownGood);
3948  return false;
3949  }
3950  return true;
3951  }
3952  }
3953  return false;
3954 }
3955 
3956 /// If we can, try to add the value of 'Addr' into the current addressing mode.
3957 /// If Addr can't be added to AddrMode this returns false and leaves AddrMode
3958 /// unmodified. This assumes that Addr is either a pointer type or intptr_t
3959 /// for the target.
3960 ///
3961 bool AddressingModeMatcher::matchAddr(Value *Addr, unsigned Depth) {
3962  // Start a transaction at this point that we will rollback if the matching
3963  // fails.
3964  TypePromotionTransaction::ConstRestorationPt LastKnownGood =
3965  TPT.getRestorationPoint();
3966  if (ConstantInt *CI = dyn_cast<ConstantInt>(Addr)) {
3967  // Fold in immediates if legal for the target.
3968  AddrMode.BaseOffs += CI->getSExtValue();
3969  if (TLI.isLegalAddressingMode(DL, AddrMode, AccessTy, AddrSpace))
3970  return true;
3971  AddrMode.BaseOffs -= CI->getSExtValue();
3972  } else if (GlobalValue *GV = dyn_cast<GlobalValue>(Addr)) {
3973  // If this is a global variable, try to fold it into the addressing mode.
3974  if (!AddrMode.BaseGV) {
3975  AddrMode.BaseGV = GV;
3976  if (TLI.isLegalAddressingMode(DL, AddrMode, AccessTy, AddrSpace))
3977  return true;
3978  AddrMode.BaseGV = nullptr;
3979  }
3980  } else if (Instruction *I = dyn_cast<Instruction>(Addr)) {
3981  ExtAddrMode BackupAddrMode = AddrMode;
3982  unsigned OldSize = AddrModeInsts.size();
3983 
3984  // Check to see if it is possible to fold this operation.
3985  bool MovedAway = false;
3986  if (matchOperationAddr(I, I->getOpcode(), Depth, &MovedAway)) {
3987  // This instruction may have been moved away. If so, there is nothing
3988  // to check here.
3989  if (MovedAway)
3990  return true;
3991  // Okay, it's possible to fold this. Check to see if it is actually
3992  // *profitable* to do so. We use a simple cost model to avoid increasing
3993  // register pressure too much.
3994  if (I->hasOneUse() ||
3995  isProfitableToFoldIntoAddressingMode(I, BackupAddrMode, AddrMode)) {
3996  AddrModeInsts.push_back(I);
3997  return true;
3998  }
3999 
4000  // It isn't profitable to do this, roll back.
4001  //cerr << "NOT FOLDING: " << *I;
4002  AddrMode = BackupAddrMode;
4003  AddrModeInsts.resize(OldSize);
4004  TPT.rollback(LastKnownGood);
4005  }
4006  } else if (ConstantExpr *CE = dyn_cast<ConstantExpr>(Addr)) {
4007  if (matchOperationAddr(CE, CE->getOpcode(), Depth))
4008  return true;
4009  TPT.rollback(LastKnownGood);
4010  } else if (isa<ConstantPointerNull>(Addr)) {
4011  // Null pointer gets folded without affecting the addressing mode.
4012  return true;
4013  }
4014 
4015  // Worse case, the target should support [reg] addressing modes. :)
4016  if (!AddrMode.HasBaseReg) {
4017  AddrMode.HasBaseReg = true;
4018  AddrMode.BaseReg = Addr;
4019  // Still check for legality in case the target supports [imm] but not [i+r].
4020  if (TLI.isLegalAddressingMode(DL, AddrMode, AccessTy, AddrSpace))
4021  return true;
4022  AddrMode.HasBaseReg = false;
4023  AddrMode.BaseReg = nullptr;
4024  }
4025 
4026  // If the base register is already taken, see if we can do [r+r].
4027  if (AddrMode.Scale == 0) {
4028  AddrMode.Scale = 1;
4029  AddrMode.ScaledReg = Addr;
4030  if (TLI.isLegalAddressingMode(DL, AddrMode, AccessTy, AddrSpace))
4031  return true;
4032  AddrMode.Scale = 0;
4033  AddrMode.ScaledReg = nullptr;
4034  }
4035  // Couldn't match.
4036  TPT.rollback(LastKnownGood);
4037  return false;
4038 }
4039 
4040 /// Check to see if all uses of OpVal by the specified inline asm call are due
4041 /// to memory operands. If so, return true, otherwise return false.
4042 static bool IsOperandAMemoryOperand(CallInst *CI, InlineAsm *IA, Value *OpVal,
4043  const TargetLowering &TLI,
4044  const TargetRegisterInfo &TRI) {
4045  const Function *F = CI->getFunction();
4046  TargetLowering::AsmOperandInfoVector TargetConstraints =
4048  ImmutableCallSite(CI));
4049 
4050  for (unsigned i = 0, e = TargetConstraints.size(); i != e; ++i) {
4051  TargetLowering::AsmOperandInfo &OpInfo = TargetConstraints[i];
4052 
4053  // Compute the constraint code and ConstraintType to use.
4054  TLI.ComputeConstraintToUse(OpInfo, SDValue());
4055 
4056  // If this asm operand is our Value*, and if it isn't an indirect memory
4057  // operand, we can't fold it!
4058  if (OpInfo.CallOperandVal == OpVal &&
4060  !OpInfo.isIndirect))
4061  return false;
4062  }
4063 
4064  return true;
4065 }
4066 
4067 // Max number of memory uses to look at before aborting the search to conserve
4068 // compile time.
4069 static constexpr int MaxMemoryUsesToScan = 20;
4070 
4071 /// Recursively walk all the uses of I until we find a memory use.
4072 /// If we find an obviously non-foldable instruction, return true.
4073 /// Add the ultimately found memory instructions to MemoryUses.
4074 static bool FindAllMemoryUses(
4075  Instruction *I,
4076  SmallVectorImpl<std::pair<Instruction *, unsigned>> &MemoryUses,
4077  SmallPtrSetImpl<Instruction *> &ConsideredInsts, const TargetLowering &TLI,
4078  const TargetRegisterInfo &TRI, int SeenInsts = 0) {
4079  // If we already considered this instruction, we're done.
4080  if (!ConsideredInsts.insert(I).second)
4081  return false;
4082 
4083  // If this is an obviously unfoldable instruction, bail out.
4084  if (!MightBeFoldableInst(I))
4085  return true;
4086 
4087  const bool OptSize = I->getFunction()->optForSize();
4088 
4089  // Loop over all the uses, recursively processing them.
4090  for (Use &U : I->uses()) {
4091  // Conservatively return true if we're seeing a large number or a deep chain
4092  // of users. This avoids excessive compilation times in pathological cases.
4093  if (SeenInsts++ >= MaxMemoryUsesToScan)
4094  return true;
4095 
4096  Instruction *UserI = cast<Instruction>(U.getUser());
4097  if (LoadInst *LI = dyn_cast<LoadInst>(UserI)) {
4098  MemoryUses.push_back(std::make_pair(LI, U.getOperandNo()));
4099  continue;
4100  }
4101 
4102  if (StoreInst *SI = dyn_cast<StoreInst>(UserI)) {
4103  unsigned opNo = U.getOperandNo();
4104  if (opNo != StoreInst::getPointerOperandIndex())
4105  return true; // Storing addr, not into addr.
4106  MemoryUses.push_back(std::make_pair(SI, opNo));
4107  continue;
4108  }
4109 
4110  if (AtomicRMWInst *RMW = dyn_cast<AtomicRMWInst>(UserI)) {
4111  unsigned opNo = U.getOperandNo();
4113  return true; // Storing addr, not into addr.
4114  MemoryUses.push_back(std::make_pair(RMW, opNo));
4115  continue;
4116  }
4117 
4118  if (AtomicCmpXchgInst *CmpX = dyn_cast<AtomicCmpXchgInst>(UserI)) {
4119  unsigned opNo = U.getOperandNo();
4121  return true; // Storing addr, not into addr.
4122  MemoryUses.push_back(std::make_pair(CmpX, opNo));
4123  continue;
4124  }
4125 
4126  if (CallInst *CI = dyn_cast<CallInst>(UserI)) {
4127  // If this is a cold call, we can sink the addressing calculation into
4128  // the cold path. See optimizeCallInst
4129  if (!OptSize && CI->hasFnAttr(Attribute::Cold))
4130  continue;
4131 
4133  if (!IA) return true;
4134 
4135  // If this is a memory operand, we're cool, otherwise bail out.
4136  if (!IsOperandAMemoryOperand(CI, IA, I, TLI, TRI))
4137  return true;
4138  continue;
4139  }
4140 
4141  if (FindAllMemoryUses(UserI, MemoryUses, ConsideredInsts, TLI, TRI,
4142  SeenInsts))
4143  return true;
4144  }
4145 
4146  return false;
4147 }
4148 
4149 /// Return true if Val is already known to be live at the use site that we're
4150 /// folding it into. If so, there is no cost to include it in the addressing
4151 /// mode. KnownLive1 and KnownLive2 are two values that we know are live at the
4152 /// instruction already.
4153 bool AddressingModeMatcher::valueAlreadyLiveAtInst(Value *Val,Value *KnownLive1,
4154  Value *KnownLive2) {
4155  // If Val is either of the known-live values, we know it is live!
4156  if (Val == nullptr || Val == KnownLive1 || Val == KnownLive2)
4157  return true;
4158 
4159  // All values other than instructions and arguments (e.g. constants) are live.
4160  if (!isa<Instruction>(Val) && !isa<Argument>(Val)) return true;
4161 
4162  // If Val is a constant sized alloca in the entry block, it is live, this is
4163  // true because it is just a reference to the stack/frame pointer, which is
4164  // live for the whole function.
4165  if (AllocaInst *AI = dyn_cast<AllocaInst>(Val))
4166  if (AI->isStaticAlloca())
4167  return true;
4168 
4169  // Check to see if this value is already used in the memory instruction's
4170  // block. If so, it's already live into the block at the very least, so we
4171  // can reasonably fold it.
4172  return Val->isUsedInBasicBlock(MemoryInst->getParent());
4173 }
4174 
4175 /// It is possible for the addressing mode of the machine to fold the specified
4176 /// instruction into a load or store that ultimately uses it.
4177 /// However, the specified instruction has multiple uses.
4178 /// Given this, it may actually increase register pressure to fold it
4179 /// into the load. For example, consider this code:
4180 ///
4181 /// X = ...
4182 /// Y = X+1
4183 /// use(Y) -> nonload/store
4184 /// Z = Y+1
4185 /// load Z
4186 ///
4187 /// In this case, Y has multiple uses, and can be folded into the load of Z
4188 /// (yielding load [X+2]). However, doing this will cause both "X" and "X+1" to
4189 /// be live at the use(Y) line. If we don't fold Y into load Z, we use one
4190 /// fewer register. Since Y can't be folded into "use(Y)" we don't increase the
4191 /// number of computations either.
4192 ///
4193 /// Note that this (like most of CodeGenPrepare) is just a rough heuristic. If
4194 /// X was live across 'load Z' for other reasons, we actually *would* want to
4195 /// fold the addressing mode in the Z case. This would make Y die earlier.
4196 bool AddressingModeMatcher::
4197 isProfitableToFoldIntoAddressingMode(Instruction *I, ExtAddrMode &AMBefore,
4198  ExtAddrMode &AMAfter) {
4199  if (IgnoreProfitability) return true;
4200 
4201  // AMBefore is the addressing mode before this instruction was folded into it,
4202  // and AMAfter is the addressing mode after the instruction was folded. Get
4203  // the set of registers referenced by AMAfter and subtract out those
4204  // referenced by AMBefore: this is the set of values which folding in this
4205  // address extends the lifetime of.
4206  //
4207  // Note that there are only two potential values being referenced here,
4208  // BaseReg and ScaleReg (global addresses are always available, as are any
4209  // folded immediates).
4210  Value *BaseReg = AMAfter.BaseReg, *ScaledReg = AMAfter.ScaledReg;
4211 
4212  // If the BaseReg or ScaledReg was referenced by the previous addrmode, their
4213  // lifetime wasn't extended by adding this instruction.
4214  if (valueAlreadyLiveAtInst(BaseReg, AMBefore.BaseReg, AMBefore.ScaledReg))
4215  BaseReg = nullptr;
4216  if (valueAlreadyLiveAtInst(ScaledReg, AMBefore.BaseReg, AMBefore.ScaledReg))
4217  ScaledReg = nullptr;
4218 
4219  // If folding this instruction (and it's subexprs) didn't extend any live
4220  // ranges, we're ok with it.
4221  if (!BaseReg && !ScaledReg)
4222  return true;
4223 
4224  // If all uses of this instruction can have the address mode sunk into them,
4225  // we can remove the addressing mode and effectively trade one live register
4226  // for another (at worst.) In this context, folding an addressing mode into
4227  // the use is just a particularly nice way of sinking it.
4229  SmallPtrSet<Instruction*, 16> ConsideredInsts;
4230  if (FindAllMemoryUses(I, MemoryUses, ConsideredInsts, TLI, TRI))
4231  return false; // Has a non-memory, non-foldable use!
4232 
4233  // Now that we know that all uses of this instruction are part of a chain of
4234  // computation involving only operations that could theoretically be folded
4235  // into a memory use, loop over each of these memory operation uses and see
4236  // if they could *actually* fold the instruction. The assumption is that
4237  // addressing modes are cheap and that duplicating the computation involved
4238  // many times is worthwhile, even on a fastpath. For sinking candidates
4239  // (i.e. cold call sites), this serves as a way to prevent excessive code
4240  // growth since most architectures have some reasonable small and fast way to
4241  // compute an effective address. (i.e LEA on x86)
4242  SmallVector<Instruction*, 32> MatchedAddrModeInsts;
4243  for (unsigned i = 0, e = MemoryUses.size(); i != e; ++i) {
4244  Instruction *User = MemoryUses[i].first;
4245  unsigned OpNo = MemoryUses[i].second;
4246 
4247  // Get the access type of this use. If the use isn't a pointer, we don't
4248  // know what it accesses.
4249  Value *Address = User->getOperand(OpNo);
4250  PointerType *AddrTy = dyn_cast<PointerType>(Address->getType());
4251  if (!AddrTy)
4252  return false;
4253  Type *AddressAccessTy = AddrTy->getElementType();
4254  unsigned AS = AddrTy->getAddressSpace();
4255 
4256  // Do a match against the root of this address, ignoring profitability. This
4257  // will tell us if the addressing mode for the memory operation will
4258  // *actually* cover the shared instruction.
4259  ExtAddrMode Result;
4260  std::pair<AssertingVH<GetElementPtrInst>, int64_t> LargeOffsetGEP(nullptr,
4261  0);
4262  TypePromotionTransaction::ConstRestorationPt LastKnownGood =
4263  TPT.getRestorationPoint();
4264  AddressingModeMatcher Matcher(
4265  MatchedAddrModeInsts, TLI, TRI, AddressAccessTy, AS, MemoryInst, Result,
4266  InsertedInsts, PromotedInsts, TPT, LargeOffsetGEP);
4267  Matcher.IgnoreProfitability = true;
4268  bool Success = Matcher.matchAddr(Address, 0);
4269  (void)Success; assert(Success && "Couldn't select *anything*?");
4270 
4271  // The match was to check the profitability, the changes made are not
4272  // part of the original matcher. Therefore, they should be dropped
4273  // otherwise the original matcher will not present the right state.
4274  TPT.rollback(LastKnownGood);
4275 
4276  // If the match didn't cover I, then it won't be shared by it.
4277  if (!is_contained(MatchedAddrModeInsts, I))
4278  return false;
4279 
4280  MatchedAddrModeInsts.clear();
4281  }
4282 
4283  return true;
4284 }
4285 
4286 /// Return true if the specified values are defined in a
4287 /// different basic block than BB.
4288 static bool IsNonLocalValue(Value *V, BasicBlock *BB) {
4289  if (Instruction *I = dyn_cast<Instruction>(V))
4290  return I->getParent() != BB;
4291  return false;
4292 }
4293 
4294 /// Sink addressing mode computation immediate before MemoryInst if doing so
4295 /// can be done without increasing register pressure. The need for the
4296 /// register pressure constraint means this can end up being an all or nothing
4297 /// decision for all uses of the same addressing computation.
4298 ///
4299 /// Load and Store Instructions often have addressing modes that can do
4300 /// significant amounts of computation. As such, instruction selection will try
4301 /// to get the load or store to do as much computation as possible for the
4302 /// program. The problem is that isel can only see within a single block. As
4303 /// such, we sink as much legal addressing mode work into the block as possible.
4304 ///
4305 /// This method is used to optimize both load/store and inline asms with memory
4306 /// operands. It's also used to sink addressing computations feeding into cold
4307 /// call sites into their (cold) basic block.
4308 ///
4309 /// The motivation for handling sinking into cold blocks is that doing so can
4310 /// both enable other address mode sinking (by satisfying the register pressure
4311 /// constraint above), and reduce register pressure globally (by removing the
4312 /// addressing mode computation from the fast path entirely.).
4313 bool CodeGenPrepare::optimizeMemoryInst(Instruction *MemoryInst, Value *Addr,
4314  Type *AccessTy, unsigned AddrSpace) {
4315  Value *Repl = Addr;
4316 
4317  // Try to collapse single-value PHI nodes. This is necessary to undo
4318  // unprofitable PRE transformations.
4319  SmallVector<Value*, 8> worklist;
4320  SmallPtrSet<Value*, 16> Visited;
4321  worklist.push_back(Addr);
4322 
4323  // Use a worklist to iteratively look through PHI and select nodes, and
4324  // ensure that the addressing mode obtained from the non-PHI/select roots of
4325  // the graph are compatible.
4326  bool PhiOrSelectSeen = false;
4327  SmallVector<Instruction*, 16> AddrModeInsts;
4328  const SimplifyQuery SQ(*DL, TLInfo);
4329  AddressingModeCombiner AddrModes(SQ, { Addr, MemoryInst->getParent() });
4330  TypePromotionTransaction TPT(RemovedInsts);
4331  TypePromotionTransaction::ConstRestorationPt LastKnownGood =
4332  TPT.getRestorationPoint();
4333  while (!worklist.empty()) {
4334  Value *V = worklist.back();
4335  worklist.pop_back();
4336 
4337  // We allow traversing cyclic Phi nodes.
4338  // In case of success after this loop we ensure that traversing through
4339  // Phi nodes ends up with all cases to compute address of the form
4340  // BaseGV + Base + Scale * Index + Offset
4341  // where Scale and Offset are constans and BaseGV, Base and Index
4342  // are exactly the same Values in all cases.
4343  // It means that BaseGV, Scale and Offset dominate our memory instruction
4344  // and have the same value as they had in address computation represented
4345  // as Phi. So we can safely sink address computation to memory instruction.
4346  if (!Visited.insert(V).second)
4347  continue;
4348 
4349  // For a PHI node, push all of its incoming values.
4350  if (PHINode *P = dyn_cast<PHINode>(V)) {
4351  for (Value *IncValue : P->incoming_values())
4352  worklist.push_back(IncValue);
4353  PhiOrSelectSeen = true;
4354  continue;
4355  }
4356  // Similar for select.
4357  if (SelectInst *SI = dyn_cast<SelectInst>(V)) {
4358  worklist.push_back(SI->getFalseValue());
4359  worklist.push_back(SI->getTrueValue());
4360  PhiOrSelectSeen = true;
4361  continue;
4362  }
4363 
4364  // For non-PHIs, determine the addressing mode being computed. Note that
4365  // the result may differ depending on what other uses our candidate
4366  // addressing instructions might have.
4367  AddrModeInsts.clear();
4368  std::pair<AssertingVH<GetElementPtrInst>, int64_t> LargeOffsetGEP(nullptr,
4369  0);
4370  ExtAddrMode NewAddrMode = AddressingModeMatcher::Match(
4371  V, AccessTy, AddrSpace, MemoryInst, AddrModeInsts, *TLI, *TRI,
4372  InsertedInsts, PromotedInsts, TPT, LargeOffsetGEP);
4373 
4374  GetElementPtrInst *GEP = LargeOffsetGEP.first;
4375  if (GEP && GEP->getParent() != MemoryInst->getParent() &&
4376  !NewGEPBases.count(GEP)) {
4377  // If splitting the underlying data structure can reduce the offset of a
4378  // GEP, collect the GEP. Skip the GEPs that are the new bases of
4379  // previously split data structures.
4380  LargeOffsetGEPMap[GEP->getPointerOperand()].push_back(LargeOffsetGEP);
4381  if (LargeOffsetGEPID.find(GEP) == LargeOffsetGEPID.end())
4382  LargeOffsetGEPID[GEP] = LargeOffsetGEPID.size();
4383  }
4384 
4385  NewAddrMode.OriginalValue = V;
4386  if (!AddrModes.addNewAddrMode(NewAddrMode))
4387  break;
4388  }
4389 
4390  // Try to combine the AddrModes we've collected. If we couldn't collect any,
4391  // or we have multiple but either couldn't combine them or combining them
4392  // wouldn't do anything useful, bail out now.
4393  if (!AddrModes.combineAddrModes()) {
4394  TPT.rollback(LastKnownGood);
4395  return false;
4396  }
4397  TPT.commit();
4398 
4399  // Get the combined AddrMode (or the only AddrMode, if we only had one).
4400  ExtAddrMode AddrMode = AddrModes.getAddrMode();
4401 
4402  // If all the instructions matched are already in this BB, don't do anything.
4403  // If we saw a Phi node then it is not local definitely, and if we saw a select
4404  // then we want to push the address calculation past it even if it's already
4405  // in this BB.
4406  if (!PhiOrSelectSeen && none_of(AddrModeInsts, [&](Value *V) {
4407  return IsNonLocalValue(V, MemoryInst->getParent());
4408  })) {
4409  LLVM_DEBUG(dbgs() << "CGP: Found local addrmode: " << AddrMode
4410  << "\n");
4411  return false;
4412  }
4413 
4414  // Insert this computation right after this user. Since our caller is
4415  // scanning from the top of the BB to the bottom, reuse of the expr are
4416  // guaranteed to happen later.
4417  IRBuilder<> Builder(MemoryInst);
4418 
4419  // Now that we determined the addressing expression we want to use and know
4420  // that we have to sink it into this block. Check to see if we have already
4421  // done this for some other load/store instr in this block. If so, reuse
4422  // the computation. Before attempting reuse, check if the address is valid
4423  // as it may have been erased.
4424 
4425  WeakTrackingVH SunkAddrVH = SunkAddrs[Addr];
4426 
4427  Value * SunkAddr = SunkAddrVH.pointsToAliveValue() ? SunkAddrVH : nullptr;
4428  if (SunkAddr) {
4429  LLVM_DEBUG(dbgs() << "CGP: Reusing nonlocal addrmode: " << AddrMode
4430  << " for " << *MemoryInst << "\n");
4431  if (SunkAddr->getType() != Addr->getType())
4432  SunkAddr = Builder.CreatePointerCast(SunkAddr, Addr->getType());
4433  } else if (AddrSinkUsingGEPs ||
4434  (!AddrSinkUsingGEPs.getNumOccurrences() && TM && TTI->useAA())) {
4435  // By default, we use the GEP-based method when AA is used later. This
4436  // prevents new inttoptr/ptrtoint pairs from degrading AA capabilities.
4437  LLVM_DEBUG(dbgs() << "CGP: SINKING nonlocal addrmode: " << AddrMode
4438  << " for " << *MemoryInst << "\n");
4439  Type *IntPtrTy = DL->getIntPtrType(Addr->getType());
4440  Value *ResultPtr = nullptr, *ResultIndex = nullptr;
4441 
4442  // First, find the pointer.
4443  if (AddrMode.BaseReg && AddrMode.BaseReg->getType()->isPointerTy()) {
4444  ResultPtr = AddrMode.BaseReg;
4445  AddrMode.BaseReg = nullptr;
4446  }
4447 
4448  if (AddrMode.Scale && AddrMode.ScaledReg->getType()->isPointerTy()) {
4449  // We can't add more than one pointer together, nor can we scale a
4450  // pointer (both of which seem meaningless).
4451  if (ResultPtr || AddrMode.Scale != 1)
4452  return false;
4453 
4454  ResultPtr = AddrMode.ScaledReg;
4455  AddrMode.Scale = 0;
4456  }
4457 
4458  // It is only safe to sign extend the BaseReg if we know that the math
4459  // required to create it did not overflow before we extend it. Since
4460  // the original IR value was tossed in favor of a constant back when
4461  // the AddrMode was created we need to bail out gracefully if widths
4462  // do not match instead of extending it.
4463  //
4464  // (See below for code to add the scale.)
4465  if (AddrMode.Scale) {
4466  Type *ScaledRegTy = AddrMode.ScaledReg->getType();
4467  if (cast<IntegerType>(IntPtrTy)->getBitWidth() >
4468  cast<IntegerType>(ScaledRegTy)->getBitWidth())
4469  return false;
4470  }
4471 
4472  if (AddrMode.BaseGV) {
4473  if (ResultPtr)
4474  return false;
4475 
4476  ResultPtr = AddrMode.BaseGV;
4477  }
4478 
4479  // If the real base value actually came from an inttoptr, then the matcher
4480  // will look through it and provide only the integer value. In that case,
4481  // use it here.
4482  if (!DL->isNonIntegralPointerType(Addr->getType())) {
4483  if (!ResultPtr && AddrMode.BaseReg) {
4484  ResultPtr = Builder.CreateIntToPtr(AddrMode.BaseReg, Addr->getType(),
4485  "sunkaddr");
4486  AddrMode.BaseReg = nullptr;
4487  } else if (!ResultPtr && AddrMode.Scale == 1) {
4488  ResultPtr = Builder.CreateIntToPtr(AddrMode.ScaledReg, Addr->getType(),
4489  "sunkaddr");
4490  AddrMode.Scale = 0;
4491  }
4492  }
4493 
4494  if (!ResultPtr &&
4495  !AddrMode.BaseReg && !AddrMode.Scale && !AddrMode.BaseOffs) {
4496  SunkAddr = Constant::getNullValue(Addr->getType());
4497  } else if (!ResultPtr) {
4498  return false;
4499  } else {
4500  Type *I8PtrTy =
4501  Builder.getInt8PtrTy(Addr->getType()->getPointerAddressSpace());
4502  Type *I8Ty = Builder.getInt8Ty();
4503 
4504  // Start with the base register. Do this first so that subsequent address
4505  // matching finds it last, which will prevent it from trying to match it
4506  // as the scaled value in case it happens to be a mul. That would be
4507  // problematic if we've sunk a different mul for the scale, because then
4508  // we'd end up sinking both muls.
4509  if (AddrMode.BaseReg) {
4510  Value *V = AddrMode.BaseReg;
4511  if (V->getType() != IntPtrTy)
4512  V = Builder.CreateIntCast(V, IntPtrTy, /*isSigned=*/true, "sunkaddr");
4513 
4514  ResultIndex = V;
4515  }
4516 
4517  // Add the scale value.
4518  if (AddrMode.Scale) {
4519  Value *V = AddrMode.ScaledReg;
4520  if (V->getType() == IntPtrTy) {
4521  // done.
4522  } else {
4523  assert(cast<IntegerType>(IntPtrTy)->getBitWidth() <
4524  cast<IntegerType>(V->getType())->getBitWidth() &&
4525  "We can't transform if ScaledReg is too narrow");
4526  V = Builder.CreateTrunc(V, IntPtrTy, "sunkaddr");
4527  }
4528 
4529  if (AddrMode.Scale != 1)
4530  V = Builder.CreateMul(V, ConstantInt::get(IntPtrTy, AddrMode.Scale),
4531  "sunkaddr");
4532  if (ResultIndex)
4533  ResultIndex = Builder.CreateAdd(ResultIndex, V, "sunkaddr");
4534  else
4535  ResultIndex = V;
4536  }
4537 
4538  // Add in the Base Offset if present.
4539  if (AddrMode.BaseOffs) {
4540  Value *V = ConstantInt::get(IntPtrTy, AddrMode.BaseOffs);
4541  if (ResultIndex) {
4542  // We need to add this separately from the scale above to help with
4543  // SDAG consecutive load/store merging.
4544  if (ResultPtr->getType() != I8PtrTy)
4545  ResultPtr = Builder.CreatePointerCast(ResultPtr, I8PtrTy);
4546  ResultPtr = Builder.CreateGEP(I8Ty, ResultPtr, ResultIndex, "sunkaddr");
4547  }
4548 
4549  ResultIndex = V;
4550  }
4551 
4552  if (!ResultIndex) {
4553  SunkAddr = ResultPtr;
4554  } else {
4555  if (ResultPtr->getType() != I8PtrTy)
4556  ResultPtr = Builder.CreatePointerCast(ResultPtr, I8PtrTy);
4557  SunkAddr = Builder.CreateGEP(I8Ty, ResultPtr, ResultIndex, "sunkaddr");
4558  }
4559 
4560  if (SunkAddr->getType() != Addr->getType())
4561  SunkAddr = Builder.CreatePointerCast(SunkAddr, Addr->getType());
4562  }
4563  } else {
4564  // We'd require a ptrtoint/inttoptr down the line, which we can't do for
4565  // non-integral pointers, so in that case bail out now.
4566  Type *BaseTy = AddrMode.BaseReg ? AddrMode.BaseReg->getType() : nullptr;
4567  Type *ScaleTy = AddrMode.Scale ? AddrMode.ScaledReg->getType() : nullptr;
4568  PointerType *BasePtrTy = dyn_cast_or_null<PointerType>(BaseTy);
4569  PointerType *ScalePtrTy = dyn_cast_or_null<PointerType>(ScaleTy);
4570  if (DL->isNonIntegralPointerType(Addr->getType()) ||
4571  (BasePtrTy && DL->isNonIntegralPointerType(BasePtrTy)) ||
4572  (ScalePtrTy && DL->isNonIntegralPointerType(ScalePtrTy)) ||
4573  (AddrMode.BaseGV &&
4574  DL->isNonIntegralPointerType(AddrMode.BaseGV->getType())))
4575  return false;
4576 
4577  LLVM_DEBUG(dbgs() << "CGP: SINKING nonlocal addrmode: " << AddrMode
4578  << " for " << *MemoryInst << "\n");
4579  Type *IntPtrTy = DL->getIntPtrType(Addr->getType());
4580  Value *Result = nullptr;
4581 
4582  // Start with the base register. Do this first so that subsequent address
4583  // matching finds it last, which will prevent it from trying to match it
4584  // as the scaled value in case it happens to be a mul. That would be
4585  // problematic if we've sunk a different mul for the scale, because then
4586  // we'd end up sinking both muls.
4587  if (AddrMode.BaseReg) {
4588  Value *V = AddrMode.BaseReg;
4589  if (V->getType()->isPointerTy())
4590  V = Builder.CreatePtrToInt(V, IntPtrTy, "sunkaddr");
4591  if (V->getType() != IntPtrTy)
4592  V = Builder.CreateIntCast(V, IntPtrTy, /*isSigned=*/true, "sunkaddr");
4593  Result = V;
4594  }
4595 
4596  // Add the scale value.
4597  if (AddrMode.Scale) {
4598  Value *V = AddrMode.ScaledReg;
4599  if (V->getType() == IntPtrTy) {
4600  // done.
4601  } else if (V->getType()->isPointerTy()) {
4602  V = Builder.CreatePtrToInt(V, IntPtrTy, "sunkaddr");
4603  } else if (cast<IntegerType>(IntPtrTy)->getBitWidth() <
4604  cast<IntegerType>(V->getType())->getBitWidth()) {
4605  V = Builder.CreateTrunc(V, IntPtrTy, "sunkaddr");
4606  } else {
4607  // It is only safe to sign extend the BaseReg if we know that the math
4608  // required to create it did not overflow before we extend it. Since
4609  // the original IR value was tossed in favor of a constant back when
4610  // the AddrMode was created we need to bail out gracefully if widths
4611  // do not match instead of extending it.
4612  Instruction *I = dyn_cast_or_null<Instruction>(Result);
4613  if (I && (Result != AddrMode.BaseReg))
4614  I->eraseFromParent();
4615  return false;
4616  }
4617  if (AddrMode.Scale != 1)
4618  V = Builder.CreateMul(V, ConstantInt::get(IntPtrTy, AddrMode.Scale),
4619  "sunkaddr");
4620  if (Result)
4621  Result = Builder.CreateAdd(Result, V, "sunkaddr");
4622  else
4623  Result = V;
4624  }
4625 
4626  // Add in the BaseGV if present.
4627  if (AddrMode.BaseGV) {
4628  Value *V = Builder.CreatePtrToInt(AddrMode.BaseGV, IntPtrTy, "sunkaddr");
4629  if (Result)
4630  Result = Builder.CreateAdd(Result, V, "sunkaddr");
4631  else
4632  Result = V;
4633  }
4634 
4635  // Add in the Base Offset if present.
4636  if (AddrMode.BaseOffs) {
4637  Value *V = ConstantInt::get(IntPtrTy, AddrMode.BaseOffs);
4638  if (Result)
4639  Result = Builder.CreateAdd(Result, V, "sunkaddr");
4640  else
4641  Result = V;
4642  }
4643 
4644  if (!Result)
4645  SunkAddr = Constant::getNullValue(Addr->getType());
4646  else
4647  SunkAddr = Builder.CreateIntToPtr(Result, Addr->getType(), "sunkaddr");
4648  }
4649 
4650  MemoryInst->replaceUsesOfWith(Repl, SunkAddr);
4651  // Store the newly computed address into the cache. In the case we reused a
4652  // value, this should be idempotent.
4653  SunkAddrs[Addr] = WeakTrackingVH(SunkAddr);
4654 
4655  // If we have no uses, recursively delete the value and all dead instructions
4656  // using it.
4657  if (Repl->use_empty()) {
4658  // This can cause recursive deletion, which can invalidate our iterator.
4659  // Use a WeakTrackingVH to hold onto it in case this happens.
4660  Value *CurValue = &*CurInstIterator;
4661  WeakTrackingVH IterHandle(CurValue);
4662  BasicBlock *BB = CurInstIterator->getParent();
4663 
4665 
4666  if (IterHandle != CurValue) {
4667  // If the iterator instruction was recursively deleted, start over at the
4668  // start of the block.
4669  CurInstIterator = BB->begin();
4670  SunkAddrs.clear();
4671  }
4672  }
4673  ++NumMemoryInsts;
4674  return true;
4675 }
4676 
4677 /// If there are any memory operands, use OptimizeMemoryInst to sink their
4678 /// address computing into the block when possible / profitable.
4679 bool CodeGenPrepare::optimizeInlineAsmInst(CallInst *CS) {
4680  bool MadeChange = false;
4681 
4682  const TargetRegisterInfo *TRI =
4683  TM->getSubtargetImpl(*CS->getFunction())->getRegisterInfo();
4684  TargetLowering::AsmOperandInfoVector TargetConstraints =
4685  TLI->ParseConstraints(*DL, TRI, CS);
4686  unsigned ArgNo = 0;
4687  for (unsigned i = 0, e = TargetConstraints.size(); i != e; ++i) {
4688  TargetLowering::AsmOperandInfo &OpInfo = TargetConstraints[i];
4689 
4690  // Compute the constraint code and ConstraintType to use.
4691  TLI->ComputeConstraintToUse(OpInfo, SDValue());
4692 
4693  if (OpInfo.ConstraintType == TargetLowering::C_Memory &&
4694  OpInfo.isIndirect) {
4695  Value *OpVal = CS->getArgOperand(ArgNo++);
4696  MadeChange |= optimizeMemoryInst(CS, OpVal, OpVal->getType(), ~0u);
4697  } else if (OpInfo.Type == InlineAsm::isInput)
4698  ArgNo++;
4699  }
4700 
4701  return MadeChange;
4702 }
4703 
4704 /// Check if all the uses of \p Val are equivalent (or free) zero or
4705 /// sign extensions.
4706 static bool hasSameExtUse(Value *Val, const TargetLowering &TLI) {
4707  assert(!Val->use_empty() && "Input must have at least one use");
4708  const Instruction *FirstUser = cast<Instruction>(*Val->user_begin());
4709  bool IsSExt = isa<SExtInst>(FirstUser);
4710  Type *ExtTy = FirstUser->getType();
4711  for (const User *U : Val->users()) {
4712  const Instruction *UI = cast<Instruction>(U);
4713  if ((IsSExt && !isa<SExtInst>(UI)) || (!IsSExt && !isa<ZExtInst>(UI)))
4714  return false;
4715  Type *CurTy = UI->getType();
4716  // Same input and output types: Same instruction after CSE.
4717  if (CurTy == ExtTy)
4718  continue;
4719 
4720  // If IsSExt is true, we are in this situation:
4721  // a = Val
4722  // b = sext ty1 a to ty2
4723  // c = sext ty1 a to ty3
4724  // Assuming ty2 is shorter than ty3, this could be turned into:
4725  // a = Val
4726  // b = sext ty1 a to ty2
4727  // c = sext ty2 b to ty3
4728  // However, the last sext is not free.
4729  if (IsSExt)
4730  return false;
4731 
4732  // This is a ZExt, maybe this is free to extend from one type to another.
4733  // In that case, we would not account for a different use.
4734  Type *NarrowTy;
4735  Type *LargeTy;
4736  if (ExtTy->getScalarType()->getIntegerBitWidth() >
4737  CurTy->getScalarType()->getIntegerBitWidth()) {
4738  NarrowTy = CurTy;
4739  LargeTy = ExtTy;
4740  } else {
4741  NarrowTy = ExtTy;
4742  LargeTy = CurTy;
4743  }
4744 
4745  if (!TLI.isZExtFree(NarrowTy, LargeTy))
4746  return false;
4747  }
4748  // All uses are the same or can be derived from one another for free.
4749  return true;
4750 }
4751 
4752 /// Try to speculatively promote extensions in \p Exts and continue
4753 /// promoting through newly promoted operands recursively as far as doing so is
4754 /// profitable. Save extensions profitably moved up, in \p ProfitablyMovedExts.
4755 /// When some promotion happened, \p TPT contains the proper state to revert
4756 /// them.
4757 ///
4758 /// \return true if some promotion happened, false otherwise.
4759 bool CodeGenPrepare::tryToPromoteExts(
4760  TypePromotionTransaction &TPT, const SmallVectorImpl<Instruction *> &Exts,
4761  SmallVectorImpl<Instruction *> &ProfitablyMovedExts,
4762  unsigned CreatedInstsCost) {
4763  bool Promoted = false;
4764 
4765  // Iterate over all the extensions to try to promote them.
4766  for (auto I : Exts) {
4767  // Early check if we directly have ext(load).
4768  if (isa<LoadInst>(I->getOperand(0))) {
4769  ProfitablyMovedExts.push_back(I);
4770  continue;
4771  }
4772 
4773  // Check whether or not we want to do any promotion. The reason we have
4774  // this check inside the for loop is to catch the case where an extension
4775  // is directly fed by a load because in such case the extension can be moved
4776  // up without any promotion on its operands.
4777  if (!TLI || !TLI->enableExtLdPromotion() || DisableExtLdPromotion)
4778  return false;
4779 
4780  // Get the action to perform the promotion.
4781  TypePromotionHelper::Action TPH =
4782  TypePromotionHelper::getAction(I, InsertedInsts, *TLI, PromotedInsts);
4783  // Check if we can promote.
4784  if (!TPH) {
4785  // Save the current extension as we cannot move up through its operand.
4786  ProfitablyMovedExts.push_back(I);
4787  continue;
4788  }
4789 
4790  // Save the current state.
4791  TypePromotionTransaction::ConstRestorationPt LastKnownGood =
4792  TPT.getRestorationPoint();
4794  unsigned NewCreatedInstsCost = 0;
4795  unsigned ExtCost = !TLI->isExtFree(I);
4796  // Promote.
4797  Value *PromotedVal = TPH(I, TPT, PromotedInsts, NewCreatedInstsCost,
4798  &NewExts, nullptr, *TLI);
4799  assert(PromotedVal &&
4800  "TypePromotionHelper should have filtered out those cases");
4801 
4802  // We would be able to merge only one extension in a load.
4803  // Therefore, if we have more than 1 new extension we heuristically
4804  // cut this search path, because it means we degrade the code quality.
4805  // With exactly 2, the transformation is neutral, because we will merge
4806  // one extension but leave one. However, we optimistically keep going,
4807  // because the new extension may be removed too.
4808  long long TotalCreatedInstsCost = CreatedInstsCost + NewCreatedInstsCost;
4809  // FIXME: It would be possible to propagate a negative value instead of
4810  // conservatively ceiling it to 0.
4811  TotalCreatedInstsCost =
4812  std::max((long long)0, (TotalCreatedInstsCost - ExtCost));
4813  if (!StressExtLdPromotion &&
4814  (TotalCreatedInstsCost > 1 ||
4815  !isPromotedInstructionLegal(*TLI, *DL, PromotedVal))) {
4816  // This promotion is not profitable, rollback to the previous state, and
4817  // save the current extension in ProfitablyMovedExts as the latest
4818  // speculative promotion turned out to be unprofitable.
4819  TPT.rollback(LastKnownGood);
4820  ProfitablyMovedExts.push_back(I);
4821  continue;
4822  }
4823  // Continue promoting NewExts as far as doing so is profitable.
4824  SmallVector<Instruction *, 2> NewlyMovedExts;
4825  (void)tryToPromoteExts(TPT, NewExts, NewlyMovedExts, TotalCreatedInstsCost);
4826  bool NewPromoted = false;
4827  for (auto ExtInst : NewlyMovedExts) {
4828  Instruction *MovedExt = cast<Instruction>(ExtInst);
4829  Value *ExtOperand = MovedExt->getOperand(0);
4830  // If we have reached to a load, we need this extra profitability check
4831  // as it could potentially be merged into an ext(load).
4832  if (isa<LoadInst>(ExtOperand) &&
4833  !(StressExtLdPromotion || NewCreatedInstsCost <= ExtCost ||
4834  (ExtOperand->hasOneUse() || hasSameExtUse(ExtOperand, *TLI))))
4835  continue;
4836 
4837  ProfitablyMovedExts.push_back(MovedExt);
4838  NewPromoted = true;
4839  }
4840 
4841  // If none of speculative promotions for NewExts is profitable, rollback
4842  // and save the current extension (I) as the last profitable extension.
4843  if (!NewPromoted) {
4844  TPT.rollback(LastKnownGood);
4845  ProfitablyMovedExts.push_back(I);
4846  continue;
4847  }
4848  // The promotion is profitable.
4849  Promoted = true;
4850  }
4851  return Promoted;
4852 }
4853 
4854 /// Merging redundant sexts when one is dominating the other.
4855 bool CodeGenPrepare::mergeSExts(Function &F) {
4856  DominatorTree DT(F);
4857  bool Changed = false;
4858  for (auto &Entry : ValToSExtendedUses) {
4859  SExts &Insts = Entry.second;
4860  SExts CurPts;
4861  for (Instruction *Inst : Insts) {
4862  if (RemovedInsts.count(Inst) || !isa<SExtInst>(Inst) ||
4863  Inst->getOperand(0) != Entry.first)
4864  continue;
4865  bool inserted = false;
4866  for (auto &Pt : CurPts) {
4867  if (DT.dominates(Inst, Pt)) {
4868  Pt->replaceAllUsesWith(Inst);
4869  RemovedInsts.insert(Pt);
4870  Pt->removeFromParent();
4871  Pt = Inst;
4872  inserted = true;
4873  Changed = true;
4874  break;
4875  }
4876  if (!DT.dominates(Pt, Inst))
4877  // Give up if we need to merge in a common dominator as the
4878  // experiments show it is not profitable.
4879  continue;
4880  Inst->replaceAllUsesWith(Pt);
4881  RemovedInsts.insert(Inst);
4882  Inst->removeFromParent();
4883  inserted = true;
4884  Changed = true;
4885  break;
4886  }
4887  if (!inserted)
4888  CurPts.push_back(Inst);
4889  }
4890  }
4891  return Changed;
4892 }
4893 
4894 // Spliting large data structures so that the GEPs accessing them can have
4895 // smaller offsets so that they can be sunk to the same blocks as their users.
4896 // For example, a large struct starting from %base is splitted into two parts
4897 // where the second part starts from %new_base.
4898 //
4899 // Before:
4900 // BB0:
4901 // %base =
4902 //
4903 // BB1:
4904 // %gep0 = gep %base, off0
4905 // %gep1 = gep %base, off1
4906 // %gep2 = gep %base, off2
4907 //
4908 // BB2:
4909 // %load1 = load %gep0
4910 // %load2 = load %gep1
4911 // %load3 = load %gep2
4912 //
4913 // After:
4914 // BB0:
4915 // %base =
4916 // %new_base = gep %base, off0
4917 //
4918 // BB1:
4919 // %new_gep0 = %new_base
4920 // %new_gep1 = gep %new_base, off1 - off0
4921 // %new_gep2 = gep %new_base, off2 - off0
4922 //
4923 // BB2:
4924 // %load1 = load i32, i32* %new_gep0
4925 // %load2 = load i32, i32* %new_gep1
4926 // %load3 = load i32, i32* %new_gep2
4927 //
4928 // %new_gep1 and %new_gep2 can be sunk to BB2 now after the splitting because
4929 // their offsets are smaller enough to fit into the addressing mode.
4930 bool CodeGenPrepare::splitLargeGEPOffsets() {
4931  bool Changed = false;
4932  for (auto &Entry : LargeOffsetGEPMap) {
4933  Value *OldBase = Entry.first;
4935  &LargeOffsetGEPs = Entry.second;
4936  auto compareGEPOffset =
4937  [&](const std::pair<GetElementPtrInst *, int64_t> &LHS,
4938  const std::pair<GetElementPtrInst *, int64_t> &RHS) {
4939  if (LHS.first == RHS.first)
4940  return false;
4941  if (LHS.second != RHS.second)
4942  return LHS.second < RHS.second;
4943  return LargeOffsetGEPID[LHS.first] < LargeOffsetGEPID[RHS.first];
4944  };
4945  // Sorting all the GEPs of the same data structures based on the offsets.
4946  llvm::sort(LargeOffsetGEPs.begin(), LargeOffsetGEPs.end(),
4947  compareGEPOffset);
4948  LargeOffsetGEPs.erase(
4949  std::unique(LargeOffsetGEPs.begin(), LargeOffsetGEPs.end()),
4950  LargeOffsetGEPs.end());
4951  // Skip if all the GEPs have the same offsets.
4952  if (LargeOffsetGEPs.front().second == LargeOffsetGEPs.back().second)
4953  continue;
4954  GetElementPtrInst *BaseGEP = LargeOffsetGEPs.begin()->first;
4955  int64_t BaseOffset = LargeOffsetGEPs.begin()->second;
4956  Value *NewBaseGEP = nullptr;
4957 
4958  auto LargeOffsetGEP = LargeOffsetGEPs.begin();
4959  while (LargeOffsetGEP != LargeOffsetGEPs.end()) {
4960  GetElementPtrInst *GEP = LargeOffsetGEP->first;
4961  int64_t Offset = LargeOffsetGEP->second;
4962  if (Offset != BaseOffset) {
4964  AddrMode.BaseOffs = Offset - BaseOffset;
4965  // The result type of the GEP might not be the type of the memory
4966  // access.
4967  if (!TLI->isLegalAddressingMode(*DL, AddrMode,
4968  GEP->getResultElementType(),
4969  GEP->getAddressSpace())) {
4970  // We need to create a new base if the offset to the current base is
4971  // too large to fit into the addressing mode. So, a very large struct
4972  // may be splitted into several parts.
4973  BaseGEP = GEP;
4974  BaseOffset = Offset;
4975  NewBaseGEP = nullptr;
4976  }
4977  }
4978 
4979  // Generate a new GEP to replace the current one.
4980  IRBuilder<> Builder(GEP);
4981  Type *IntPtrTy = DL->getIntPtrType(GEP->getType());
4982  Type *I8PtrTy =
4983  Builder.getInt8PtrTy(GEP->getType()->getPointerAddressSpace());
4984  Type *I8Ty = Builder.getInt8Ty();
4985 
4986  if (!NewBaseGEP) {
4987  // Create a new base if we don't have one yet. Find the insertion
4988  // pointer for the new base first.
4989  BasicBlock::iterator NewBaseInsertPt;
4990  BasicBlock *NewBaseInsertBB;
4991  if (auto *BaseI = dyn_cast<Instruction>(OldBase)) {
4992  // If the base of the struct is an instruction, the new base will be
4993  // inserted close to it.
4994  NewBaseInsertBB = BaseI->getParent();
4995  if (isa<PHINode>(BaseI))
4996  NewBaseInsertPt = NewBaseInsertBB->getFirstInsertionPt();
4997  else if (InvokeInst *Invoke = dyn_cast<InvokeInst>(BaseI)) {
4998  NewBaseInsertBB =
4999  SplitEdge(NewBaseInsertBB, Invoke->getNormalDest());
5000  NewBaseInsertPt = NewBaseInsertBB->getFirstInsertionPt();
5001  } else
5002  NewBaseInsertPt = std::next(BaseI->getIterator());
5003  } else {
5004  // If the current base is an argument or global value, the new base
5005  // will be inserted to the entry block.
5006  NewBaseInsertBB = &BaseGEP->getFunction()->getEntryBlock();
5007  NewBaseInsertPt = NewBaseInsertBB->getFirstInsertionPt();
5008  }
5009  IRBuilder<> NewBaseBuilder(NewBaseInsertBB, NewBaseInsertPt);
5010  // Create a new base.
5011  Value *BaseIndex = ConstantInt::get(IntPtrTy, BaseOffset);
5012  NewBaseGEP = OldBase;
5013  if (NewBaseGEP->getType() != I8PtrTy)
5014  NewBaseGEP = NewBaseBuilder.CreatePointerCast(NewBaseGEP, I8PtrTy);
5015  NewBaseGEP =
5016  NewBaseBuilder.CreateGEP(I8Ty, NewBaseGEP, BaseIndex, "splitgep");
5017  NewGEPBases.insert(NewBaseGEP);
5018  }
5019 
5020  Value *NewGEP = NewBaseGEP;
5021  if (Offset == BaseOffset) {
5022  if (GEP->getType() != I8PtrTy)
5023  NewGEP = Builder.CreatePointerCast(NewGEP, GEP->getType());
5024  } else {
5025  // Calculate the new offset for the new GEP.
5026  Value *Index = ConstantInt::get(IntPtrTy, Offset - BaseOffset);
5027  NewGEP = Builder.CreateGEP(I8Ty, NewBaseGEP, Index);
5028 
5029  if (GEP->getType() != I8PtrTy)
5030  NewGEP = Builder.CreatePointerCast(NewGEP, GEP->getType());
5031  }
5032  GEP->replaceAllUsesWith(NewGEP);
5033  LargeOffsetGEPID.erase(GEP);
5034  LargeOffsetGEP = LargeOffsetGEPs.erase(LargeOffsetGEP);
5035  GEP->eraseFromParent();
5036  Changed = true;
5037  }
5038  }
5039  return Changed;
5040 }
5041 
5042 /// Return true, if an ext(load) can be formed from an extension in
5043 /// \p MovedExts.
5044 bool CodeGenPrepare::canFormExtLd(
5045  const SmallVectorImpl<Instruction *> &MovedExts, LoadInst *&LI,
5046  Instruction *&Inst, bool HasPromoted) {
5047  for (auto *MovedExtInst : MovedExts) {
5048  if (isa<LoadInst>(MovedExtInst->getOperand(0))) {
5049  LI = cast<LoadInst>(MovedExtInst->getOperand(0));
5050  Inst = MovedExtInst;
5051  break;
5052  }
5053  }
5054  if (!LI)
5055  return false;
5056 
5057  // If they're already in the same block, there's nothing to do.
5058  // Make the cheap checks first if we did not promote.
5059  // If we promoted, we need to check if it is indeed profitable.
5060  if (!HasPromoted && LI->getParent() == Inst->getParent())
5061  return false;
5062 
5063  return TLI->isExtLoad(LI, Inst, *DL);
5064 }
5065 
5066 /// Move a zext or sext fed by a load into the same basic block as the load,
5067 /// unless conditions are unfavorable. This allows SelectionDAG to fold the
5068 /// extend into the load.
5069 ///
5070 /// E.g.,
5071 /// \code
5072 /// %ld = load i32* %addr
5073 /// %add = add nuw i32 %ld, 4
5074 /// %zext = zext i32 %add to i64
5075 // \endcode
5076 /// =>
5077 /// \code
5078 /// %ld = load i32* %addr
5079 /// %zext = zext i32 %ld to i64
5080 /// %add = add nuw i64 %zext, 4
5081 /// \encode
5082 /// Note that the promotion in %add to i64 is done in tryToPromoteExts(), which
5083 /// allow us to match zext(load i32*) to i64.
5084 ///
5085 /// Also, try to promote the computations used to obtain a sign extended
5086 /// value used into memory accesses.
5087 /// E.g.,
5088 /// \code
5089 /// a = add nsw i32 b, 3
5090 /// d = sext i32 a to i64
5091 /// e = getelementptr ..., i64 d
5092 /// \endcode
5093 /// =>
5094 /// \code
5095 /// f = sext i32 b to i64
5096 /// a = add nsw i64 f, 3
5097 /// e = getelementptr ..., i64 a
5098 /// \endcode
5099 ///
5100 /// \p Inst[in/out] the extension may be modified during the process if some
5101 /// promotions apply.
5102 bool CodeGenPrepare::optimizeExt(Instruction *&Inst) {
5103  // ExtLoad formation and address type promotion infrastructure requires TLI to
5104  // be effective.
5105  if (!TLI)
5106  return false;
5107 
5108  bool AllowPromotionWithoutCommonHeader = false;
5109  /// See if it is an interesting sext operations for the address type
5110  /// promotion before trying to promote it, e.g., the ones with the right
5111  /// type and used in memory accesses.
5112  bool ATPConsiderable = TTI->shouldConsiderAddressTypePromotion(
5113  *Inst, AllowPromotionWithoutCommonHeader);
5114  TypePromotionTransaction TPT(RemovedInsts);
5115  TypePromotionTransaction::ConstRestorationPt LastKnownGood =
5116  TPT.getRestorationPoint();
5118  SmallVector<Instruction *, 2> SpeculativelyMovedExts;
5119  Exts.push_back(Inst);
5120 
5121  bool HasPromoted = tryToPromoteExts(TPT, Exts, SpeculativelyMovedExts);
5122 
5123  // Look for a load being extended.
5124  LoadInst *LI = nullptr;
5125  Instruction *ExtFedByLoad;
5126 
5127  // Try to promote a chain of computation if it allows to form an extended
5128  // load.
5129  if (canFormExtLd(SpeculativelyMovedExts, LI, ExtFedByLoad, HasPromoted)) {
5130  assert(LI && ExtFedByLoad && "Expect a valid load and extension");
5131  TPT.commit();
5132  // Move the extend into the same block as the load
5133  ExtFedByLoad->moveAfter(LI);
5134  // CGP does not check if the zext would be speculatively executed when moved
5135  // to the same basic block as the load. Preserving its original location
5136  // would pessimize the debugging experience, as well as negatively impact
5137  // the quality of sample pgo. We don't want to use "line 0" as that has a
5138  // size cost in the line-table section and logically the zext can be seen as
5139  // part of the load. Therefore we conservatively reuse the same debug
5140  // location for the load and the zext.
5141  ExtFedByLoad->setDebugLoc(LI->getDebugLoc());
5142  ++NumExtsMoved;
5143  Inst = ExtFedByLoad;
5144  return true;
5145  }
5146 
5147  // Continue promoting SExts if known as considerable depending on targets.
5148  if (ATPConsiderable &&
5149  performAddressTypePromotion(Inst, AllowPromotionWithoutCommonHeader,
5150  HasPromoted, TPT, SpeculativelyMovedExts))
5151  return true;
5152 
5153  TPT.rollback(LastKnownGood);
5154  return false;
5155 }
5156 
5157 // Perform address type promotion if doing so is profitable.
5158 // If AllowPromotionWithoutCommonHeader == false, we should find other sext
5159 // instructions that sign extended the same initial value. However, if
5160 // AllowPromotionWithoutCommonHeader == true, we expect promoting the
5161 // extension is just profitable.
5162 bool CodeGenPrepare::performAddressTypePromotion(
5163  Instruction *&Inst, bool AllowPromotionWithoutCommonHeader,
5164  bool HasPromoted, TypePromotionTransaction &TPT,
5165  SmallVectorImpl<Instruction *> &SpeculativelyMovedExts) {
5166  bool Promoted = false;
5167  SmallPtrSet<Instruction *, 1> UnhandledExts;
5168  bool AllSeenFirst = true;
5169  for (auto I : SpeculativelyMovedExts) {
5170  Value *HeadOfChain = I->getOperand(0);
5172  SeenChainsForSExt.find(HeadOfChain);
5173  // If there is an unhandled SExt which has the same header, try to promote
5174  // it as well.
5175  if (AlreadySeen != SeenChainsForSExt.end()) {
5176  if (AlreadySeen->second != nullptr)
5177  UnhandledExts.insert(AlreadySeen->second);
5178  AllSeenFirst = false;
5179  }
5180  }
5181 
5182  if (!AllSeenFirst || (AllowPromotionWithoutCommonHeader &&
5183  SpeculativelyMovedExts.size() == 1)) {
5184  TPT.commit();
5185  if (HasPromoted)
5186  Promoted = true;
5187  for (auto I : SpeculativelyMovedExts) {
5188  Value *HeadOfChain = I->getOperand(0);
5189  SeenChainsForSExt[HeadOfChain] = nullptr;
5190  ValToSExtendedUses[HeadOfChain].push_back(I);
5191  }
5192  // Update Inst as promotion happen.
5193  Inst = SpeculativelyMovedExts.pop_back_val();
5194  } else {
5195  // This is the first chain visited from the header, keep the current chain
5196  // as unhandled. Defer to promote this until we encounter another SExt
5197  // chain derived from the same header.
5198  for (auto I : SpeculativelyMovedExts) {
5199  Value *HeadOfChain = I->getOperand(0);
5200  SeenChainsForSExt[HeadOfChain] = Inst;
5201  }
5202  return false;
5203  }
5204 
5205  if (!AllSeenFirst && !UnhandledExts.empty())
5206  for (auto VisitedSExt : UnhandledExts) {
5207  if (RemovedInsts.count(VisitedSExt))
5208  continue;
5209  TypePromotionTransaction TPT(RemovedInsts);
5212  Exts.push_back(VisitedSExt);
5213  bool HasPromoted = tryToPromoteExts(TPT, Exts, Chains);
5214  TPT.commit();
5215  if (HasPromoted)
5216  Promoted = true;
5217  for (auto I : Chains) {
5218  Value *HeadOfChain = I->getOperand(0);
5219  // Mark this as handled.
5220  SeenChainsForSExt[HeadOfChain] = nullptr;
5221  ValToSExtendedUses[HeadOfChain].push_back(I);
5222  }
5223  }
5224  return Promoted;
5225 }
5226 
5227 bool CodeGenPrepare::optimizeExtUses(Instruction *I) {
5228  BasicBlock *DefBB = I->getParent();
5229 
5230  // If the result of a {s|z}ext and its source are both live out, rewrite all
5231  // other uses of the source with result of extension.
5232  Value *Src = I->getOperand(0);
5233  if (Src->hasOneUse())
5234  return false;
5235 
5236  // Only do this xform if truncating is free.
5237  if (TLI && !TLI->isTruncateFree(I->getType(), Src->getType()))
5238  return false;
5239 
5240  // Only safe to perform the optimization if the source is also defined in
5241  // this block.
5242  if (!isa<Instruction>(Src) || DefBB != cast<Instruction>(Src)->getParent())
5243  return false;
5244 
5245  bool DefIsLiveOut = false;
5246  for (User *U : I->users()) {
5247  Instruction *UI = cast<Instruction>(U);
5248 
5249  // Figure out which BB this ext is used in.
5250  BasicBlock *UserBB = UI->getParent();
5251  if (UserBB == DefBB) continue;
5252  DefIsLiveOut = true;
5253  break;
5254  }
5255  if (!DefIsLiveOut)
5256  return false;
5257 
5258  // Make sure none of the uses are PHI nodes.
5259  for (User *U : Src->users()) {
5260  Instruction *UI = cast<Instruction>(U);
5261  BasicBlock *UserBB = UI->getParent();
5262  if (UserBB == DefBB) continue;
5263  // Be conservative. We don't want this xform to end up introducing
5264  // reloads just before load / store instructions.
5265  if (isa<PHINode>(UI) || isa<LoadInst>(UI) || isa<StoreInst>(UI))
5266  return false;
5267  }
5268 
5269  // InsertedTruncs - Only insert one trunc in each block once.
5270  DenseMap<BasicBlock*, Instruction*> InsertedTruncs;
5271 
5272  bool MadeChange = false;
5273  for (Use &U : Src->uses()) {
5274  Instruction *User = cast<Instruction>(U.getUser());
5275 
5276  // Figure out which BB this ext is used in.
5277  BasicBlock *UserBB = User->getParent();
5278  if (UserBB == DefBB) continue;
5279 
5280  // Both src and def are live in this block. Rewrite the use.
5281  Instruction *&InsertedTrunc = InsertedTruncs[UserBB];
5282 
5283  if (!InsertedTrunc) {
5284  BasicBlock::iterator InsertPt = UserBB->getFirstInsertionPt();
5285  assert(InsertPt != UserBB->end());
5286  InsertedTrunc = new TruncInst(I, Src->getType(), "", &*InsertPt);
5287  InsertedInsts.insert(InsertedTrunc);
5288  }
5289 
5290  // Replace a use of the {s|z}ext source with a use of the result.
5291  U = InsertedTrunc;
5292  ++NumExtUses;
5293  MadeChange = true;
5294  }
5295 
5296  return MadeChange;
5297 }
5298 
5299 // Find loads whose uses only use some of the loaded value's bits. Add an "and"
5300 // just after the load if the target can fold this into one extload instruction,
5301 // with the hope of eliminating some of the other later "and" instructions using
5302 // the loaded value. "and"s that are made trivially redundant by the insertion
5303 // of the new "and" are removed by this function, while others (e.g. those whose
5304 // path from the load goes through a phi) are left for isel to potentially
5305 // remove.
5306 //
5307 // For example:
5308 //
5309 // b0:
5310 // x = load i32
5311 // ...
5312 // b1:
5313 // y = and x, 0xff
5314 // z = use y
5315 //
5316 // becomes:
5317 //
5318 // b0:
5319 // x = load i32
5320 // x' = and x, 0xff
5321 // ...
5322 // b1:
5323 // z = use x'
5324 //
5325 // whereas:
5326 //
5327 // b0:
5328 // x1 = load i32
5329 // ...
5330 // b1:
5331 // x2 = load i32
5332 // ...
5333 // b2:
5334 // x = phi x1, x2
5335 // y = and x, 0xff
5336 //
5337 // becomes (after a call to optimizeLoadExt for each load):
5338 //
5339 // b0:
5340 // x1 = load i32
5341 // x1' = and x1, 0xff
5342 // ...
5343 // b1:
5344 // x2 = load i32
5345 // x2' = and x2, 0xff
5346 // ...
5347 // b2:
5348 // x = phi x1', x2'
5349 // y = and x, 0xff
5350 bool CodeGenPrepare::optimizeLoadExt(LoadInst *Load) {
5351  if (!Load->isSimple() || !Load->getType()->isIntOrPtrTy())
5352  return false;
5353 
5354  // Skip loads we've already transformed.
5355  if (Load->hasOneUse() &&
5356  InsertedInsts.count(cast<Instruction>(*Load->user_begin())))
5357  return false;
5358 
5359  // Look at all uses of Load, looking through phis, to determine how many bits
5360  // of the loaded value are needed.
5363  SmallVector<Instruction *, 8> AndsToMaybeRemove;
5364  for (auto *U : Load->users())
5365  WorkList.push_back(cast<Instruction>(U));
5366 
5367  EVT LoadResultVT = TLI->getValueType(*DL, Load->getType());
5368  unsigned BitWidth = LoadResultVT.getSizeInBits();
5369  APInt DemandBits(BitWidth, 0);
5370  APInt WidestAndBits(BitWidth, 0);
5371 
5372  while (!WorkList.empty()) {
5373  Instruction *I = WorkList.back();
5374  WorkList.pop_back();
5375 
5376  // Break use-def graph loops.
5377  if (!Visited.insert(I).second)
5378  continue;
5379 
5380  // For a PHI node, push all of its users.
5381  if (auto *Phi = dyn_cast<PHINode>(I)) {
5382  for (auto *U : Phi->users())
5383  WorkList.push_back(cast<Instruction>(U));
5384  continue;
5385  }
5386 
5387  switch (I->getOpcode()) {
5388  case Instruction::And: {
5389  auto *AndC = dyn_cast<ConstantInt>(I->getOperand(1));
5390  if (!AndC)
5391  return false;
5392  APInt AndBits = AndC->getValue();
5393  DemandBits |= AndBits;
5394  // Keep track of the widest and mask we see.
5395  if (AndBits.ugt(WidestAndBits))
5396  WidestAndBits = AndBits;
5397  if (AndBits == WidestAndBits && I->getOperand(0) == Load)
5398  AndsToMaybeRemove.push_back(I);
5399  break;
5400  }
5401 
5402  case Instruction::Shl: {
5403  auto *ShlC = dyn_cast<ConstantInt>(I->getOperand(1));
5404  if (!ShlC)
5405  return false;
5406  uint64_t ShiftAmt = ShlC->getLimitedValue(BitWidth - 1);
5407  DemandBits.setLowBits(BitWidth - ShiftAmt);
5408  break;
5409  }
5410 
5411  case Instruction::Trunc: {
5412  EVT TruncVT = TLI->getValueType(*DL, I->getType());
5413  unsigned TruncBitWidth = TruncVT.getSizeInBits();
5414  DemandBits.setLowBits(TruncBitWidth);
5415  break;
5416  }
5417 
5418  default:
5419  return false;
5420  }
5421  }
5422 
5423  uint32_t ActiveBits = DemandBits.getActiveBits();
5424  // Avoid hoisting (and (load x) 1) since it is unlikely to be folded by the
5425  // target even if isLoadExtLegal says an i1 EXTLOAD is valid. For example,
5426  // for the AArch64 target isLoadExtLegal(ZEXTLOAD, i32, i1) returns true, but
5427  // (and (load x) 1) is not matched as a single instruction, rather as a LDR
5428  // followed by an AND.
5429  // TODO: Look into removing this restriction by fixing backends to either
5430  // return false for isLoadExtLegal for i1 or have them select this pattern to
5431  // a single instruction.
5432  //
5433  // Also avoid hoisting if we didn't see any ands with the exact DemandBits
5434  // mask, since these are the only ands that will be removed by isel.
5435  if (ActiveBits <= 1 || !DemandBits.isMask(ActiveBits) ||
5436  WidestAndBits != DemandBits)
5437  return false;
5438 
5439  LLVMContext &Ctx = Load->getType()->getContext();
5440  Type *TruncTy = Type::getIntNTy(Ctx, ActiveBits);
5441  EVT TruncVT = TLI->getValueType(*DL, TruncTy);
5442 
5443  // Reject cases that won't be matched as extloads.
5444  if (!LoadResultVT.bitsGT(TruncVT) || !TruncVT.isRound() ||
5445  !TLI->isLoadExtLegal(ISD::ZEXTLOAD, LoadResultVT, TruncVT))
5446  return false;
5447 
5448  IRBuilder<> Builder(Load->getNextNode());
5449  auto *NewAnd = dyn_cast<Instruction>(
5450  Builder.CreateAnd(Load, ConstantInt::get(Ctx, DemandBits)));
5451  // Mark this instruction as "inserted by CGP", so that other
5452  // optimizations don't touch it.
5453  InsertedInsts.insert(NewAnd);
5454 
5455  // Replace all uses of load with new and (except for the use of load in the
5456  // new and itself).
5457  Load->replaceAllUsesWith(NewAnd);
5458  NewAnd->setOperand(0, Load);
5459 
5460  // Remove any and instructions that are now redundant.
5461  for (auto *And : AndsToMaybeRemove)
5462  // Check that the and mask is the same as the one we decided to put on the
5463  // new and.
5464  if (cast<ConstantInt>(And->getOperand(1))->getValue() == DemandBits) {
5465  And->replaceAllUsesWith(NewAnd);
5466  if (&*CurInstIterator == And)
5467  CurInstIterator = std::next(And->getIterator());
5468  And->eraseFromParent();
5469  ++NumAndUses;
5470  }
5471 
5472  ++NumAndsAdded;
5473  return true;
5474 }
5475 
5476 /// Check if V (an operand of a select instruction) is an expensive instruction
5477 /// that is only used once.
5478 static bool sinkSelectOperand(const TargetTransformInfo *TTI, Value *V) {
5479  auto *I = dyn_cast<Instruction>(V);
5480  // If it's safe to speculatively execute, then it should not have side
5481  // effects; therefore, it's safe to sink and possibly *not* execute.
5482  return I && I->hasOneUse() && isSafeToSpeculativelyExecute(I) &&
5484 }
5485 
5486 /// Returns true if a SelectInst should be turned into an explicit branch.
5488  const TargetLowering *TLI,
5489  SelectInst *SI) {
5490  // If even a predictable select is cheap, then a branch can't be cheaper.
5491  if (!TLI->isPredictableSelectExpensive())
5492  return false;
5493 
5494  // FIXME: This should use the same heuristics as IfConversion to determine
5495  // whether a select is better represented as a branch.
5496 
5497  // If metadata tells us that the select condition is obviously predictable,
5498  // then we want to replace the select with a branch.
5499  uint64_t TrueWeight, FalseWeight;
5500  if (SI->extractProfMetadata(TrueWeight, FalseWeight)) {
5501  uint64_t Max = std::max(TrueWeight, FalseWeight);
5502  uint64_t Sum = TrueWeight + FalseWeight;
5503  if (Sum != 0) {
5504  auto Probability = BranchProbability::getBranchProbability(Max, Sum);
5505  if (Probability > TLI->getPredictableBranchThreshold())
5506  return true;
5507  }
5508  }
5509 
5510  CmpInst *Cmp = dyn_cast<CmpInst>(SI->getCondition());
5511 
5512  // If a branch is predictable, an out-of-order CPU can avoid blocking on its
5513  // comparison condition. If the compare has more than one use, there's
5514  // probably another cmov or setcc around, so it's not worth emitting a branch.
5515  if (!Cmp || !Cmp->hasOneUse())
5516  return false;
5517 
5518  // If either operand of the select is expensive and only needed on one side
5519  // of the select, we should form a branch.
5520  if (sinkSelectOperand(TTI, SI->getTrueValue()) ||
5521  sinkSelectOperand(TTI, SI->getFalseValue()))
5522  return true;
5523 
5524  return false;
5525 }
5526 
5527 /// If \p isTrue is true, return the true value of \p SI, otherwise return
5528 /// false value of \p SI. If the true/false value of \p SI is defined by any
5529 /// select instructions in \p Selects, look through the defining select
5530 /// instruction until the true/false value is not defined in \p Selects.
5532  SelectInst *SI, bool isTrue,
5533  const SmallPtrSet<const Instruction *, 2> &Selects) {
5534  Value *V;
5535 
5536  for (SelectInst *DefSI = SI; DefSI != nullptr && Selects.count(DefSI);
5537  DefSI = dyn_cast<SelectInst>(V)) {
5538  assert(DefSI->getCondition() == SI->getCondition() &&
5539  "The condition of DefSI does not match with SI");
5540  V = (isTrue ? DefSI->getTrueValue() : DefSI->getFalseValue());
5541  }
5542  return V;
5543 }
5544 
5545 /// If we have a SelectInst that will likely profit from branch prediction,
5546 /// turn it into a branch.
5547 bool CodeGenPrepare::optimizeSelectInst(SelectInst *SI) {
5548  // Find all consecutive select instructions that share the same condition.
5550  ASI.push_back(SI);
5552  It != SI->getParent()->end(); ++It) {
5553  SelectInst *I = dyn_cast<SelectInst>(&*It);
5554  if (I && SI->getCondition() == I->getCondition()) {
5555  ASI.push_back(I);
5556  } else {
5557  break;
5558  }
5559  }
5560 
5561  SelectInst *LastSI = ASI.back();
5562  // Increment the current iterator to skip all the rest of select instructions
5563  // because they will be either "not lowered" or "all lowered" to branch.
5564  CurInstIterator = std::next(LastSI->getIterator());
5565 
5566  bool VectorCond = !SI->getCondition()->getType()->isIntegerTy(1);
5567 
5568  // Can we convert the 'select' to CF ?
5569  if (DisableSelectToBranch || OptSize || !TLI || VectorCond ||
5571  return false;
5572 
5574  if (VectorCond)
5575  SelectKind = TargetLowering::VectorMaskSelect;
5576  else if (SI->getType()->isVectorTy())
5578  else
5579  SelectKind = TargetLowering::ScalarValSelect;
5580 
5581  if (TLI->isSelectSupported(SelectKind) &&
5582  !isFormingBranchFromSelectProfitable(TTI, TLI, SI))
5583  return false;
5584 
5585  ModifiedDT = true;
5586 
5587  // Transform a sequence like this:
5588  // start:
5589  // %cmp = cmp uge i32 %a, %b
5590  // %sel = select i1 %cmp, i32 %c, i32 %d
5591  //
5592  // Into:
5593  // start:
5594  // %cmp = cmp uge i32 %a, %b
5595  // br i1 %cmp, label %select.true, label %select.false
5596  // select.true:
5597  // br label %select.end
5598  // select.false:
5599  // br label %select.end
5600  // select.end:
5601  // %sel = phi i32 [ %c, %select.true ], [ %d, %select.false ]
5602  //
5603  // In addition, we may sink instructions that produce %c or %d from
5604  // the entry block into the destination(s) of the new branch.
5605  // If the true or false blocks do not contain a sunken instruction, that
5606  // block and its branch may be optimized away. In that case, one side of the
5607  // first branch will point directly to select.end, and the corresponding PHI
5608  // predecessor block will be the start block.
5609 
5610  // First, we split the block containing the select into 2 blocks.
5611  BasicBlock *StartBlock = SI->getParent();
5612  BasicBlock::iterator SplitPt = ++(BasicBlock::iterator(LastSI));
5613  BasicBlock *EndBlock = StartBlock->splitBasicBlock(SplitPt, "select.end");
5614 
5615  // Delete the unconditional branch that was just created by the split.
5616  StartBlock->getTerminator()->eraseFromParent();
5617 
5618  // These are the new basic blocks for the conditional branch.
5619  // At least one will become an actual new basic block.
5620  BasicBlock *TrueBlock = nullptr;
5621  BasicBlock *FalseBlock = nullptr;
5622  BranchInst *TrueBranch = nullptr;
5623  BranchInst *FalseBranch = nullptr;
5624 
5625  // Sink expensive instructions into the conditional blocks to avoid executing
5626  // them speculatively.
5627  for (SelectInst *SI : ASI) {
5628  if (sinkSelectOperand(TTI, SI->getTrueValue())) {
5629  if (TrueBlock == nullptr) {
5630  TrueBlock = BasicBlock::Create(SI->getContext(), "select.true.sink",
5631  EndBlock->getParent(), EndBlock);
5632  TrueBranch = BranchInst::Create(EndBlock, TrueBlock);
5633  }
5634  auto *TrueInst = cast<Instruction>(SI->getTrueValue());
5635  TrueInst->moveBefore(TrueBranch);
5636  }
5637  if (sinkSelectOperand(TTI, SI->getFalseValue())) {
5638  if (FalseBlock == nullptr) {
5639  FalseBlock = BasicBlock::Create(SI->getContext(), "select.false.sink",
5640  EndBlock->getParent(), EndBlock);
5641  FalseBranch = BranchInst::Create(EndBlock, FalseBlock);
5642  }
5643  auto *FalseInst = cast<Instruction>(SI->getFalseValue());
5644  FalseInst->moveBefore(FalseBranch);
5645  }
5646  }
5647 
5648  // If there was nothing to sink, then arbitrarily choose the 'false' side
5649  // for a new input value to the PHI.
5650  if (TrueBlock == FalseBlock) {
5651  assert(TrueBlock == nullptr &&
5652  "Unexpected basic block transform while optimizing select");
5653 
5654  FalseBlock = BasicBlock::Create(SI->getContext(), "select.false",
5655  EndBlock->getParent(), EndBlock);
5656  BranchInst::Create(EndBlock, FalseBlock);
5657  }
5658 
5659  // Insert the real conditional branch based on the original condition.
5660  // If we did not create a new block for one of the 'true' or 'false' paths
5661  // of the condition, it means that side of the branch goes to the end block
5662  // directly and the path originates from the start block from the point of
5663  // view of the new PHI.
5664  BasicBlock *TT, *FT;
5665  if (TrueBlock == nullptr) {
5666  TT = EndBlock;
5667  FT = FalseBlock;
5668  TrueBlock = StartBlock;
5669  } else if (FalseBlock == nullptr) {
5670  TT = TrueBlock;
5671  FT = EndBlock;
5672  FalseBlock = StartBlock;
5673  } else {
5674  TT = TrueBlock;
5675  FT = FalseBlock;
5676  }
5677  IRBuilder<>(SI).CreateCondBr(SI->getCondition(), TT, FT, SI);
5678 
5680  INS.insert(ASI.begin(), ASI.end());
5681  // Use reverse iterator because later select may use the value of the
5682  // earlier select, and we need to propagate value through earlier select
5683  // to get the PHI operand.
5684  for (auto It = ASI.rbegin(); It != ASI.rend(); ++It) {
5685  SelectInst *SI = *It;
5686  // The select itself is replaced with a PHI Node.
5687  PHINode *PN = PHINode::Create(SI->getType(), 2, "", &EndBlock->front());
5688  PN->takeName(SI);
5689  PN->addIncoming(getTrueOrFalseValue(SI, true, INS), TrueBlock);
5690  PN->addIncoming(getTrueOrFalseValue(SI, false, INS), FalseBlock);
5691 
5692  SI->replaceAllUsesWith(PN);
5693  SI->eraseFromParent();
5694  INS.erase(SI);
5695  ++NumSelectsExpanded;
5696  }
5697 
5698  // Instruct OptimizeBlock to skip to the next block.
5699  CurInstIterator = StartBlock->end();
5700  return true;
5701 }
5702 
5705  int SplatElem = -1;
5706  for (unsigned i = 0; i < Mask.size(); ++i) {
5707  if (SplatElem != -1 && Mask[i] != -1 && Mask[i] != SplatElem)
5708  return false;
5709  SplatElem = Mask[i];
5710  }
5711 
5712  return true;
5713 }
5714 
5715 /// Some targets have expensive vector shifts if the lanes aren't all the same
5716 /// (e.g. x86 only introduced "vpsllvd" and friends with AVX2). In these cases
5717 /// it's often worth sinking a shufflevector splat down to its use so that
5718 /// codegen can spot all lanes are identical.
5719 bool CodeGenPrepare::optimizeShuffleVectorInst(ShuffleVectorInst *SVI) {
5720  BasicBlock *DefBB = SVI->getParent();
5721 
5722  // Only do this xform if variable vector shifts are particularly expensive.
5723  if (!TLI || !TLI->isVectorShiftByScalarCheap(SVI->getType()))
5724  return false;
5725 
5726  // We only expect better codegen by sinking a shuffle if we can recognise a
5727  // constant splat.
5728  if (!isBroadcastShuffle(SVI))
5729  return false;
5730 
5731  // InsertedShuffles - Only insert a shuffle in each block once.
5732  DenseMap<BasicBlock*, Instruction*> InsertedShuffles;
5733 
5734  bool MadeChange = false;
5735  for (User *U : SVI->users()) {
5736  Instruction *UI = cast<Instruction>(U);
5737 
5738  // Figure out which BB this ext is used in.
5739  BasicBlock *UserBB = UI->getParent();
5740  if (UserBB == DefBB) continue;
5741 
5742  // For now only apply this when the splat is used by a shift instruction.
5743  if (!UI->isShift()) continue;
5744 
5745  // Everything checks out, sink the shuffle if the user's block doesn't
5746  // already have a copy.
5747  Instruction *&InsertedShuffle = InsertedShuffles[UserBB];
5748 
5749  if (!InsertedShuffle) {
5750  BasicBlock::iterator InsertPt = UserBB->getFirstInsertionPt();
5751  assert(InsertPt != UserBB->end());
5752  InsertedShuffle =
5753  new ShuffleVectorInst(SVI->getOperand(0), SVI->getOperand(1),
5754  SVI->getOperand(2), "", &*InsertPt);
5755  }
5756 
5757  UI->replaceUsesOfWith(SVI, InsertedShuffle);
5758  MadeChange = true;
5759  }
5760 
5761  // If we removed all uses, nuke the shuffle.
5762  if (SVI->use_empty()) {
5763  SVI->eraseFromParent();
5764  MadeChange = true;
5765  }
5766 
5767  return MadeChange;
5768 }
5769 
5770 bool CodeGenPrepare::optimizeSwitchInst(SwitchInst *SI) {
5771  if (!TLI || !DL)
5772  return false;
5773 
5774  Value *Cond = SI->getCondition();
5775  Type *OldType = Cond->getType();
5776  LLVMContext &Context = Cond->getContext();
5777  MVT RegType = TLI->getRegisterType(Context, TLI->getValueType(*DL, OldType));
5778  unsigned RegWidth = RegType.getSizeInBits();
5779 
5780  if (RegWidth <= cast<IntegerType>(OldType)->getBitWidth())
5781  return false;
5782 
5783  // If the register width is greater than the type width, expand the condition
5784  // of the switch instruction and each case constant to the width of the
5785  // register. By widening the type of the switch condition, subsequent
5786  // comparisons (for case comparisons) will not need to be extended to the