LLVM  15.0.0git
SimplifyCFG.cpp
Go to the documentation of this file.
1 //===- SimplifyCFG.cpp - Code to perform CFG simplification ---------------===//
2 //
3 // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
4 // See https://llvm.org/LICENSE.txt for license information.
5 // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
6 //
7 //===----------------------------------------------------------------------===//
8 //
9 // Peephole optimize the CFG.
10 //
11 //===----------------------------------------------------------------------===//
12 
13 #include "llvm/ADT/APInt.h"
14 #include "llvm/ADT/ArrayRef.h"
15 #include "llvm/ADT/DenseMap.h"
16 #include "llvm/ADT/MapVector.h"
17 #include "llvm/ADT/Optional.h"
18 #include "llvm/ADT/STLExtras.h"
19 #include "llvm/ADT/ScopeExit.h"
20 #include "llvm/ADT/Sequence.h"
21 #include "llvm/ADT/SetOperations.h"
22 #include "llvm/ADT/SetVector.h"
23 #include "llvm/ADT/SmallPtrSet.h"
24 #include "llvm/ADT/SmallVector.h"
25 #include "llvm/ADT/Statistic.h"
26 #include "llvm/ADT/StringRef.h"
37 #include "llvm/IR/Attributes.h"
38 #include "llvm/IR/BasicBlock.h"
39 #include "llvm/IR/CFG.h"
40 #include "llvm/IR/Constant.h"
41 #include "llvm/IR/ConstantRange.h"
42 #include "llvm/IR/Constants.h"
43 #include "llvm/IR/DataLayout.h"
44 #include "llvm/IR/DerivedTypes.h"
45 #include "llvm/IR/Function.h"
46 #include "llvm/IR/GlobalValue.h"
47 #include "llvm/IR/GlobalVariable.h"
48 #include "llvm/IR/IRBuilder.h"
49 #include "llvm/IR/InstrTypes.h"
50 #include "llvm/IR/Instruction.h"
51 #include "llvm/IR/Instructions.h"
52 #include "llvm/IR/IntrinsicInst.h"
53 #include "llvm/IR/LLVMContext.h"
54 #include "llvm/IR/MDBuilder.h"
55 #include "llvm/IR/Metadata.h"
56 #include "llvm/IR/Module.h"
57 #include "llvm/IR/NoFolder.h"
58 #include "llvm/IR/Operator.h"
59 #include "llvm/IR/PatternMatch.h"
60 #include "llvm/IR/Type.h"
61 #include "llvm/IR/Use.h"
62 #include "llvm/IR/User.h"
63 #include "llvm/IR/Value.h"
64 #include "llvm/IR/ValueHandle.h"
66 #include "llvm/Support/Casting.h"
68 #include "llvm/Support/Debug.h"
70 #include "llvm/Support/KnownBits.h"
76 #include <algorithm>
77 #include <cassert>
78 #include <climits>
79 #include <cstddef>
80 #include <cstdint>
81 #include <iterator>
82 #include <map>
83 #include <set>
84 #include <tuple>
85 #include <utility>
86 #include <vector>
87 
88 using namespace llvm;
89 using namespace PatternMatch;
90 
91 #define DEBUG_TYPE "simplifycfg"
92 
94  "simplifycfg-require-and-preserve-domtree", cl::Hidden, cl::ZeroOrMore,
95  cl::init(false),
96  cl::desc("Temorary development switch used to gradually uplift SimplifyCFG "
97  "into preserving DomTree,"));
98 
99 // Chosen as 2 so as to be cheap, but still to have enough power to fold
100 // a select, so the "clamp" idiom (of a min followed by a max) will be caught.
101 // To catch this, we need to fold a compare and a select, hence '2' being the
102 // minimum reasonable default.
104  "phi-node-folding-threshold", cl::Hidden, cl::init(2),
105  cl::desc(
106  "Control the amount of phi node folding to perform (default = 2)"));
107 
109  "two-entry-phi-node-folding-threshold", cl::Hidden, cl::init(4),
110  cl::desc("Control the maximal total instruction cost that we are willing "
111  "to speculatively execute to fold a 2-entry PHI node into a "
112  "select (default = 4)"));
113 
114 static cl::opt<bool>
115  HoistCommon("simplifycfg-hoist-common", cl::Hidden, cl::init(true),
116  cl::desc("Hoist common instructions up to the parent block"));
117 
118 static cl::opt<bool>
119  SinkCommon("simplifycfg-sink-common", cl::Hidden, cl::init(true),
120  cl::desc("Sink common instructions down to the end block"));
121 
123  "simplifycfg-hoist-cond-stores", cl::Hidden, cl::init(true),
124  cl::desc("Hoist conditional stores if an unconditional store precedes"));
125 
127  "simplifycfg-merge-cond-stores", cl::Hidden, cl::init(true),
128  cl::desc("Hoist conditional stores even if an unconditional store does not "
129  "precede - hoist multiple conditional stores into a single "
130  "predicated store"));
131 
133  "simplifycfg-merge-cond-stores-aggressively", cl::Hidden, cl::init(false),
134  cl::desc("When merging conditional stores, do so even if the resultant "
135  "basic blocks are unlikely to be if-converted as a result"));
136 
138  "speculate-one-expensive-inst", cl::Hidden, cl::init(true),
139  cl::desc("Allow exactly one expensive instruction to be speculatively "
140  "executed"));
141 
143  "max-speculation-depth", cl::Hidden, cl::init(10),
144  cl::desc("Limit maximum recursion depth when calculating costs of "
145  "speculatively executed instructions"));
146 
147 static cl::opt<int>
148  MaxSmallBlockSize("simplifycfg-max-small-block-size", cl::Hidden,
149  cl::init(10),
150  cl::desc("Max size of a block which is still considered "
151  "small enough to thread through"));
152 
153 // Two is chosen to allow one negation and a logical combine.
154 static cl::opt<unsigned>
155  BranchFoldThreshold("simplifycfg-branch-fold-threshold", cl::Hidden,
156  cl::init(2),
157  cl::desc("Maximum cost of combining conditions when "
158  "folding branches"));
159 
161  "simplifycfg-branch-fold-common-dest-vector-multiplier", cl::Hidden,
162  cl::init(2),
163  cl::desc("Multiplier to apply to threshold when determining whether or not "
164  "to fold branch to common destination when vector operations are "
165  "present"));
166 
168  "simplifycfg-merge-compatible-invokes", cl::Hidden, cl::init(true),
169  cl::desc("Allow SimplifyCFG to merge invokes together when appropriate"));
170 
172  "max-switch-cases-per-result", cl::Hidden, cl::init(16),
173  cl::desc("Limit cases to analyze when converting a switch to select"));
174 
175 STATISTIC(NumBitMaps, "Number of switch instructions turned into bitmaps");
176 STATISTIC(NumLinearMaps,
177  "Number of switch instructions turned into linear mapping");
178 STATISTIC(NumLookupTables,
179  "Number of switch instructions turned into lookup tables");
180 STATISTIC(
181  NumLookupTablesHoles,
182  "Number of switch instructions turned into lookup tables (holes checked)");
183 STATISTIC(NumTableCmpReuses, "Number of reused switch table lookup compares");
184 STATISTIC(NumFoldValueComparisonIntoPredecessors,
185  "Number of value comparisons folded into predecessor basic blocks");
186 STATISTIC(NumFoldBranchToCommonDest,
187  "Number of branches folded into predecessor basic block");
188 STATISTIC(
189  NumHoistCommonCode,
190  "Number of common instruction 'blocks' hoisted up to the begin block");
191 STATISTIC(NumHoistCommonInstrs,
192  "Number of common instructions hoisted up to the begin block");
193 STATISTIC(NumSinkCommonCode,
194  "Number of common instruction 'blocks' sunk down to the end block");
195 STATISTIC(NumSinkCommonInstrs,
196  "Number of common instructions sunk down to the end block");
197 STATISTIC(NumSpeculations, "Number of speculative executed instructions");
198 STATISTIC(NumInvokes,
199  "Number of invokes with empty resume blocks simplified into calls");
200 STATISTIC(NumInvokesMerged, "Number of invokes that were merged together");
201 STATISTIC(NumInvokeSetsFormed, "Number of invoke sets that were formed");
202 
203 namespace {
204 
205 // The first field contains the value that the switch produces when a certain
206 // case group is selected, and the second field is a vector containing the
207 // cases composing the case group.
208 using SwitchCaseResultVectorTy =
210 
211 // The first field contains the phi node that generates a result of the switch
212 // and the second field contains the value generated for a certain case in the
213 // switch for that PHI.
214 using SwitchCaseResultsTy = SmallVector<std::pair<PHINode *, Constant *>, 4>;
215 
216 /// ValueEqualityComparisonCase - Represents a case of a switch.
217 struct ValueEqualityComparisonCase {
219  BasicBlock *Dest;
220 
221  ValueEqualityComparisonCase(ConstantInt *Value, BasicBlock *Dest)
222  : Value(Value), Dest(Dest) {}
223 
224  bool operator<(ValueEqualityComparisonCase RHS) const {
225  // Comparing pointers is ok as we only rely on the order for uniquing.
226  return Value < RHS.Value;
227  }
228 
229  bool operator==(BasicBlock *RHSDest) const { return Dest == RHSDest; }
230 };
231 
232 class SimplifyCFGOpt {
233  const TargetTransformInfo &TTI;
234  DomTreeUpdater *DTU;
235  const DataLayout &DL;
236  ArrayRef<WeakVH> LoopHeaders;
238  bool Resimplify;
239 
240  Value *isValueEqualityComparison(Instruction *TI);
241  BasicBlock *GetValueEqualityComparisonCases(
242  Instruction *TI, std::vector<ValueEqualityComparisonCase> &Cases);
243  bool SimplifyEqualityComparisonWithOnlyPredecessor(Instruction *TI,
244  BasicBlock *Pred,
246  bool PerformValueComparisonIntoPredecessorFolding(Instruction *TI, Value *&CV,
247  Instruction *PTI,
249  bool FoldValueComparisonIntoPredecessors(Instruction *TI,
251 
252  bool simplifyResume(ResumeInst *RI, IRBuilder<> &Builder);
253  bool simplifySingleResume(ResumeInst *RI);
254  bool simplifyCommonResume(ResumeInst *RI);
255  bool simplifyCleanupReturn(CleanupReturnInst *RI);
256  bool simplifyUnreachable(UnreachableInst *UI);
257  bool simplifySwitch(SwitchInst *SI, IRBuilder<> &Builder);
258  bool simplifyIndirectBr(IndirectBrInst *IBI);
259  bool simplifyBranch(BranchInst *Branch, IRBuilder<> &Builder);
260  bool simplifyUncondBranch(BranchInst *BI, IRBuilder<> &Builder);
261  bool simplifyCondBranch(BranchInst *BI, IRBuilder<> &Builder);
262 
263  bool tryToSimplifyUncondBranchWithICmpInIt(ICmpInst *ICI,
265 
266  bool HoistThenElseCodeToIf(BranchInst *BI, const TargetTransformInfo &TTI,
267  bool EqTermsOnly);
268  bool SpeculativelyExecuteBB(BranchInst *BI, BasicBlock *ThenBB,
269  const TargetTransformInfo &TTI);
270  bool SimplifyTerminatorOnSelect(Instruction *OldTerm, Value *Cond,
271  BasicBlock *TrueBB, BasicBlock *FalseBB,
272  uint32_t TrueWeight, uint32_t FalseWeight);
273  bool SimplifyBranchOnICmpChain(BranchInst *BI, IRBuilder<> &Builder,
274  const DataLayout &DL);
275  bool SimplifySwitchOnSelect(SwitchInst *SI, SelectInst *Select);
276  bool SimplifyIndirectBrOnSelect(IndirectBrInst *IBI, SelectInst *SI);
277  bool TurnSwitchRangeIntoICmp(SwitchInst *SI, IRBuilder<> &Builder);
278 
279 public:
280  SimplifyCFGOpt(const TargetTransformInfo &TTI, DomTreeUpdater *DTU,
281  const DataLayout &DL, ArrayRef<WeakVH> LoopHeaders,
282  const SimplifyCFGOptions &Opts)
283  : TTI(TTI), DTU(DTU), DL(DL), LoopHeaders(LoopHeaders), Options(Opts) {
284  assert((!DTU || !DTU->hasPostDomTree()) &&
285  "SimplifyCFG is not yet capable of maintaining validity of a "
286  "PostDomTree, so don't ask for it.");
287  }
288 
289  bool simplifyOnce(BasicBlock *BB);
290  bool run(BasicBlock *BB);
291 
292  // Helper to set Resimplify and return change indication.
293  bool requestResimplify() {
294  Resimplify = true;
295  return true;
296  }
297 };
298 
299 } // end anonymous namespace
300 
301 /// Return true if all the PHI nodes in the basic block \p BB
302 /// receive compatible (identical) incoming values when coming from
303 /// all of the predecessor blocks that are specified in \p IncomingBlocks.
304 ///
305 /// Note that if the values aren't exactly identical, but \p EquivalenceSet
306 /// is provided, and *both* of the values are present in the set,
307 /// then they are considered equal.
309  BasicBlock *BB, ArrayRef<BasicBlock *> IncomingBlocks,
310  SmallPtrSetImpl<Value *> *EquivalenceSet = nullptr) {
311  assert(IncomingBlocks.size() == 2 &&
312  "Only for a pair of incoming blocks at the time!");
313 
314  // FIXME: it is okay if one of the incoming values is an `undef` value,
315  // iff the other incoming value is guaranteed to be a non-poison value.
316  // FIXME: it is okay if one of the incoming values is a `poison` value.
317  return all_of(BB->phis(), [IncomingBlocks, EquivalenceSet](PHINode &PN) {
318  Value *IV0 = PN.getIncomingValueForBlock(IncomingBlocks[0]);
319  Value *IV1 = PN.getIncomingValueForBlock(IncomingBlocks[1]);
320  if (IV0 == IV1)
321  return true;
322  if (EquivalenceSet && EquivalenceSet->contains(IV0) &&
323  EquivalenceSet->contains(IV1))
324  return true;
325  return false;
326  });
327 }
328 
329 /// Return true if it is safe to merge these two
330 /// terminator instructions together.
331 static bool
333  SmallSetVector<BasicBlock *, 4> *FailBlocks = nullptr) {
334  if (SI1 == SI2)
335  return false; // Can't merge with self!
336 
337  // It is not safe to merge these two switch instructions if they have a common
338  // successor, and if that successor has a PHI node, and if *that* PHI node has
339  // conflicting incoming values from the two switch blocks.
340  BasicBlock *SI1BB = SI1->getParent();
341  BasicBlock *SI2BB = SI2->getParent();
342 
343  SmallPtrSet<BasicBlock *, 16> SI1Succs(succ_begin(SI1BB), succ_end(SI1BB));
344  bool Fail = false;
345  for (BasicBlock *Succ : successors(SI2BB)) {
346  if (!SI1Succs.count(Succ))
347  continue;
348  if (IncomingValuesAreCompatible(Succ, {SI1BB, SI2BB}))
349  continue;
350  Fail = true;
351  if (FailBlocks)
352  FailBlocks->insert(Succ);
353  else
354  break;
355  }
356 
357  return !Fail;
358 }
359 
360 /// Update PHI nodes in Succ to indicate that there will now be entries in it
361 /// from the 'NewPred' block. The values that will be flowing into the PHI nodes
362 /// will be the same as those coming in from ExistPred, an existing predecessor
363 /// of Succ.
364 static void AddPredecessorToBlock(BasicBlock *Succ, BasicBlock *NewPred,
365  BasicBlock *ExistPred,
366  MemorySSAUpdater *MSSAU = nullptr) {
367  for (PHINode &PN : Succ->phis())
368  PN.addIncoming(PN.getIncomingValueForBlock(ExistPred), NewPred);
369  if (MSSAU)
370  if (auto *MPhi = MSSAU->getMemorySSA()->getMemoryAccess(Succ))
371  MPhi->addIncoming(MPhi->getIncomingValueForBlock(ExistPred), NewPred);
372 }
373 
374 /// Compute an abstract "cost" of speculating the given instruction,
375 /// which is assumed to be safe to speculate. TCC_Free means cheap,
376 /// TCC_Basic means less cheap, and TCC_Expensive means prohibitively
377 /// expensive.
379  const TargetTransformInfo &TTI) {
381  "Instruction is not safe to speculatively execute!");
383 }
384 
385 /// If we have a merge point of an "if condition" as accepted above,
386 /// return true if the specified value dominates the block. We
387 /// don't handle the true generality of domination here, just a special case
388 /// which works well enough for us.
389 ///
390 /// If AggressiveInsts is non-null, and if V does not dominate BB, we check to
391 /// see if V (which must be an instruction) and its recursive operands
392 /// that do not dominate BB have a combined cost lower than Budget and
393 /// are non-trapping. If both are true, the instruction is inserted into the
394 /// set and true is returned.
395 ///
396 /// The cost for most non-trapping instructions is defined as 1 except for
397 /// Select whose cost is 2.
398 ///
399 /// After this function returns, Cost is increased by the cost of
400 /// V plus its non-dominating operands. If that cost is greater than
401 /// Budget, false is returned and Cost is undefined.
403  SmallPtrSetImpl<Instruction *> &AggressiveInsts,
404  InstructionCost &Cost,
405  InstructionCost Budget,
406  const TargetTransformInfo &TTI,
407  unsigned Depth = 0) {
408  // It is possible to hit a zero-cost cycle (phi/gep instructions for example),
409  // so limit the recursion depth.
410  // TODO: While this recursion limit does prevent pathological behavior, it
411  // would be better to track visited instructions to avoid cycles.
412  if (Depth == MaxSpeculationDepth)
413  return false;
414 
415  Instruction *I = dyn_cast<Instruction>(V);
416  if (!I) {
417  // Non-instructions all dominate instructions, but not all constantexprs
418  // can be executed unconditionally.
419  if (ConstantExpr *C = dyn_cast<ConstantExpr>(V))
420  if (C->canTrap())
421  return false;
422  return true;
423  }
424  BasicBlock *PBB = I->getParent();
425 
426  // We don't want to allow weird loops that might have the "if condition" in
427  // the bottom of this block.
428  if (PBB == BB)
429  return false;
430 
431  // If this instruction is defined in a block that contains an unconditional
432  // branch to BB, then it must be in the 'conditional' part of the "if
433  // statement". If not, it definitely dominates the region.
434  BranchInst *BI = dyn_cast<BranchInst>(PBB->getTerminator());
435  if (!BI || BI->isConditional() || BI->getSuccessor(0) != BB)
436  return true;
437 
438  // If we have seen this instruction before, don't count it again.
439  if (AggressiveInsts.count(I))
440  return true;
441 
442  // Okay, it looks like the instruction IS in the "condition". Check to
443  // see if it's a cheap instruction to unconditionally compute, and if it
444  // only uses stuff defined outside of the condition. If so, hoist it out.
446  return false;
447 
448  Cost += computeSpeculationCost(I, TTI);
449 
450  // Allow exactly one instruction to be speculated regardless of its cost
451  // (as long as it is safe to do so).
452  // This is intended to flatten the CFG even if the instruction is a division
453  // or other expensive operation. The speculation of an expensive instruction
454  // is expected to be undone in CodeGenPrepare if the speculation has not
455  // enabled further IR optimizations.
456  if (Cost > Budget &&
457  (!SpeculateOneExpensiveInst || !AggressiveInsts.empty() || Depth > 0 ||
458  !Cost.isValid()))
459  return false;
460 
461  // Okay, we can only really hoist these out if their operands do
462  // not take us over the cost threshold.
463  for (Use &Op : I->operands())
464  if (!dominatesMergePoint(Op, BB, AggressiveInsts, Cost, Budget, TTI,
465  Depth + 1))
466  return false;
467  // Okay, it's safe to do this! Remember this instruction.
468  AggressiveInsts.insert(I);
469  return true;
470 }
471 
472 /// Extract ConstantInt from value, looking through IntToPtr
473 /// and PointerNullValue. Return NULL if value is not a constant int.
475  // Normal constant int.
476  ConstantInt *CI = dyn_cast<ConstantInt>(V);
477  if (CI || !isa<Constant>(V) || !V->getType()->isPointerTy())
478  return CI;
479 
480  // This is some kind of pointer constant. Turn it into a pointer-sized
481  // ConstantInt if possible.
482  IntegerType *PtrTy = cast<IntegerType>(DL.getIntPtrType(V->getType()));
483 
484  // Null pointer means 0, see SelectionDAGBuilder::getValue(const Value*).
485  if (isa<ConstantPointerNull>(V))
486  return ConstantInt::get(PtrTy, 0);
487 
488  // IntToPtr const int.
489  if (ConstantExpr *CE = dyn_cast<ConstantExpr>(V))
490  if (CE->getOpcode() == Instruction::IntToPtr)
491  if (ConstantInt *CI = dyn_cast<ConstantInt>(CE->getOperand(0))) {
492  // The constant is very likely to have the right type already.
493  if (CI->getType() == PtrTy)
494  return CI;
495  else
496  return cast<ConstantInt>(
497  ConstantExpr::getIntegerCast(CI, PtrTy, /*isSigned=*/false));
498  }
499  return nullptr;
500 }
501 
502 namespace {
503 
504 /// Given a chain of or (||) or and (&&) comparison of a value against a
505 /// constant, this will try to recover the information required for a switch
506 /// structure.
507 /// It will depth-first traverse the chain of comparison, seeking for patterns
508 /// like %a == 12 or %a < 4 and combine them to produce a set of integer
509 /// representing the different cases for the switch.
510 /// Note that if the chain is composed of '||' it will build the set of elements
511 /// that matches the comparisons (i.e. any of this value validate the chain)
512 /// while for a chain of '&&' it will build the set elements that make the test
513 /// fail.
514 struct ConstantComparesGatherer {
515  const DataLayout &DL;
516 
517  /// Value found for the switch comparison
518  Value *CompValue = nullptr;
519 
520  /// Extra clause to be checked before the switch
521  Value *Extra = nullptr;
522 
523  /// Set of integers to match in switch
525 
526  /// Number of comparisons matched in the and/or chain
527  unsigned UsedICmps = 0;
528 
529  /// Construct and compute the result for the comparison instruction Cond
530  ConstantComparesGatherer(Instruction *Cond, const DataLayout &DL) : DL(DL) {
531  gather(Cond);
532  }
533 
534  ConstantComparesGatherer(const ConstantComparesGatherer &) = delete;
535  ConstantComparesGatherer &
536  operator=(const ConstantComparesGatherer &) = delete;
537 
538 private:
539  /// Try to set the current value used for the comparison, it succeeds only if
540  /// it wasn't set before or if the new value is the same as the old one
541  bool setValueOnce(Value *NewVal) {
542  if (CompValue && CompValue != NewVal)
543  return false;
544  CompValue = NewVal;
545  return (CompValue != nullptr);
546  }
547 
548  /// Try to match Instruction "I" as a comparison against a constant and
549  /// populates the array Vals with the set of values that match (or do not
550  /// match depending on isEQ).
551  /// Return false on failure. On success, the Value the comparison matched
552  /// against is placed in CompValue.
553  /// If CompValue is already set, the function is expected to fail if a match
554  /// is found but the value compared to is different.
555  bool matchInstruction(Instruction *I, bool isEQ) {
556  // If this is an icmp against a constant, handle this as one of the cases.
557  ICmpInst *ICI;
558  ConstantInt *C;
559  if (!((ICI = dyn_cast<ICmpInst>(I)) &&
560  (C = GetConstantInt(I->getOperand(1), DL)))) {
561  return false;
562  }
563 
564  Value *RHSVal;
565  const APInt *RHSC;
566 
567  // Pattern match a special case
568  // (x & ~2^z) == y --> x == y || x == y|2^z
569  // This undoes a transformation done by instcombine to fuse 2 compares.
570  if (ICI->getPredicate() == (isEQ ? ICmpInst::ICMP_EQ : ICmpInst::ICMP_NE)) {
571  // It's a little bit hard to see why the following transformations are
572  // correct. Here is a CVC3 program to verify them for 64-bit values:
573 
574  /*
575  ONE : BITVECTOR(64) = BVZEROEXTEND(0bin1, 63);
576  x : BITVECTOR(64);
577  y : BITVECTOR(64);
578  z : BITVECTOR(64);
579  mask : BITVECTOR(64) = BVSHL(ONE, z);
580  QUERY( (y & ~mask = y) =>
581  ((x & ~mask = y) <=> (x = y OR x = (y | mask)))
582  );
583  QUERY( (y | mask = y) =>
584  ((x | mask = y) <=> (x = y OR x = (y & ~mask)))
585  );
586  */
587 
588  // Please note that each pattern must be a dual implication (<--> or
589  // iff). One directional implication can create spurious matches. If the
590  // implication is only one-way, an unsatisfiable condition on the left
591  // side can imply a satisfiable condition on the right side. Dual
592  // implication ensures that satisfiable conditions are transformed to
593  // other satisfiable conditions and unsatisfiable conditions are
594  // transformed to other unsatisfiable conditions.
595 
596  // Here is a concrete example of a unsatisfiable condition on the left
597  // implying a satisfiable condition on the right:
598  //
599  // mask = (1 << z)
600  // (x & ~mask) == y --> (x == y || x == (y | mask))
601  //
602  // Substituting y = 3, z = 0 yields:
603  // (x & -2) == 3 --> (x == 3 || x == 2)
604 
605  // Pattern match a special case:
606  /*
607  QUERY( (y & ~mask = y) =>
608  ((x & ~mask = y) <=> (x = y OR x = (y | mask)))
609  );
610  */
611  if (match(ICI->getOperand(0),
612  m_And(m_Value(RHSVal), m_APInt(RHSC)))) {
613  APInt Mask = ~*RHSC;
614  if (Mask.isPowerOf2() && (C->getValue() & ~Mask) == C->getValue()) {
615  // If we already have a value for the switch, it has to match!
616  if (!setValueOnce(RHSVal))
617  return false;
618 
619  Vals.push_back(C);
620  Vals.push_back(
621  ConstantInt::get(C->getContext(),
622  C->getValue() | Mask));
623  UsedICmps++;
624  return true;
625  }
626  }
627 
628  // Pattern match a special case:
629  /*
630  QUERY( (y | mask = y) =>
631  ((x | mask = y) <=> (x = y OR x = (y & ~mask)))
632  );
633  */
634  if (match(ICI->getOperand(0),
635  m_Or(m_Value(RHSVal), m_APInt(RHSC)))) {
636  APInt Mask = *RHSC;
637  if (Mask.isPowerOf2() && (C->getValue() | Mask) == C->getValue()) {
638  // If we already have a value for the switch, it has to match!
639  if (!setValueOnce(RHSVal))
640  return false;
641 
642  Vals.push_back(C);
643  Vals.push_back(ConstantInt::get(C->getContext(),
644  C->getValue() & ~Mask));
645  UsedICmps++;
646  return true;
647  }
648  }
649 
650  // If we already have a value for the switch, it has to match!
651  if (!setValueOnce(ICI->getOperand(0)))
652  return false;
653 
654  UsedICmps++;
655  Vals.push_back(C);
656  return ICI->getOperand(0);
657  }
658 
659  // If we have "x ult 3", for example, then we can add 0,1,2 to the set.
660  ConstantRange Span =
662 
663  // Shift the range if the compare is fed by an add. This is the range
664  // compare idiom as emitted by instcombine.
665  Value *CandidateVal = I->getOperand(0);
666  if (match(I->getOperand(0), m_Add(m_Value(RHSVal), m_APInt(RHSC)))) {
667  Span = Span.subtract(*RHSC);
668  CandidateVal = RHSVal;
669  }
670 
671  // If this is an and/!= check, then we are looking to build the set of
672  // value that *don't* pass the and chain. I.e. to turn "x ugt 2" into
673  // x != 0 && x != 1.
674  if (!isEQ)
675  Span = Span.inverse();
676 
677  // If there are a ton of values, we don't want to make a ginormous switch.
678  if (Span.isSizeLargerThan(8) || Span.isEmptySet()) {
679  return false;
680  }
681 
682  // If we already have a value for the switch, it has to match!
683  if (!setValueOnce(CandidateVal))
684  return false;
685 
686  // Add all values from the range to the set
687  for (APInt Tmp = Span.getLower(); Tmp != Span.getUpper(); ++Tmp)
688  Vals.push_back(ConstantInt::get(I->getContext(), Tmp));
689 
690  UsedICmps++;
691  return true;
692  }
693 
694  /// Given a potentially 'or'd or 'and'd together collection of icmp
695  /// eq/ne/lt/gt instructions that compare a value against a constant, extract
696  /// the value being compared, and stick the list constants into the Vals
697  /// vector.
698  /// One "Extra" case is allowed to differ from the other.
699  void gather(Value *V) {
700  bool isEQ = match(V, m_LogicalOr(m_Value(), m_Value()));
701 
702  // Keep a stack (SmallVector for efficiency) for depth-first traversal
704  SmallPtrSet<Value *, 8> Visited;
705 
706  // Initialize
707  Visited.insert(V);
708  DFT.push_back(V);
709 
710  while (!DFT.empty()) {
711  V = DFT.pop_back_val();
712 
713  if (Instruction *I = dyn_cast<Instruction>(V)) {
714  // If it is a || (or && depending on isEQ), process the operands.
715  Value *Op0, *Op1;
716  if (isEQ ? match(I, m_LogicalOr(m_Value(Op0), m_Value(Op1)))
717  : match(I, m_LogicalAnd(m_Value(Op0), m_Value(Op1)))) {
718  if (Visited.insert(Op1).second)
719  DFT.push_back(Op1);
720  if (Visited.insert(Op0).second)
721  DFT.push_back(Op0);
722 
723  continue;
724  }
725 
726  // Try to match the current instruction
727  if (matchInstruction(I, isEQ))
728  // Match succeed, continue the loop
729  continue;
730  }
731 
732  // One element of the sequence of || (or &&) could not be match as a
733  // comparison against the same value as the others.
734  // We allow only one "Extra" case to be checked before the switch
735  if (!Extra) {
736  Extra = V;
737  continue;
738  }
739  // Failed to parse a proper sequence, abort now
740  CompValue = nullptr;
741  break;
742  }
743  }
744 };
745 
746 } // end anonymous namespace
747 
749  MemorySSAUpdater *MSSAU = nullptr) {
750  Instruction *Cond = nullptr;
751  if (SwitchInst *SI = dyn_cast<SwitchInst>(TI)) {
752  Cond = dyn_cast<Instruction>(SI->getCondition());
753  } else if (BranchInst *BI = dyn_cast<BranchInst>(TI)) {
754  if (BI->isConditional())
755  Cond = dyn_cast<Instruction>(BI->getCondition());
756  } else if (IndirectBrInst *IBI = dyn_cast<IndirectBrInst>(TI)) {
757  Cond = dyn_cast<Instruction>(IBI->getAddress());
758  }
759 
760  TI->eraseFromParent();
761  if (Cond)
763 }
764 
765 /// Return true if the specified terminator checks
766 /// to see if a value is equal to constant integer value.
767 Value *SimplifyCFGOpt::isValueEqualityComparison(Instruction *TI) {
768  Value *CV = nullptr;
769  if (SwitchInst *SI = dyn_cast<SwitchInst>(TI)) {
770  // Do not permit merging of large switch instructions into their
771  // predecessors unless there is only one predecessor.
772  if (!SI->getParent()->hasNPredecessorsOrMore(128 / SI->getNumSuccessors()))
773  CV = SI->getCondition();
774  } else if (BranchInst *BI = dyn_cast<BranchInst>(TI))
775  if (BI->isConditional() && BI->getCondition()->hasOneUse())
776  if (ICmpInst *ICI = dyn_cast<ICmpInst>(BI->getCondition())) {
777  if (ICI->isEquality() && GetConstantInt(ICI->getOperand(1), DL))
778  CV = ICI->getOperand(0);
779  }
780 
781  // Unwrap any lossless ptrtoint cast.
782  if (CV) {
783  if (PtrToIntInst *PTII = dyn_cast<PtrToIntInst>(CV)) {
784  Value *Ptr = PTII->getPointerOperand();
785  if (PTII->getType() == DL.getIntPtrType(Ptr->getType()))
786  CV = Ptr;
787  }
788  }
789  return CV;
790 }
791 
792 /// Given a value comparison instruction,
793 /// decode all of the 'cases' that it represents and return the 'default' block.
794 BasicBlock *SimplifyCFGOpt::GetValueEqualityComparisonCases(
795  Instruction *TI, std::vector<ValueEqualityComparisonCase> &Cases) {
796  if (SwitchInst *SI = dyn_cast<SwitchInst>(TI)) {
797  Cases.reserve(SI->getNumCases());
798  for (auto Case : SI->cases())
799  Cases.push_back(ValueEqualityComparisonCase(Case.getCaseValue(),
800  Case.getCaseSuccessor()));
801  return SI->getDefaultDest();
802  }
803 
804  BranchInst *BI = cast<BranchInst>(TI);
805  ICmpInst *ICI = cast<ICmpInst>(BI->getCondition());
806  BasicBlock *Succ = BI->getSuccessor(ICI->getPredicate() == ICmpInst::ICMP_NE);
807  Cases.push_back(ValueEqualityComparisonCase(
808  GetConstantInt(ICI->getOperand(1), DL), Succ));
809  return BI->getSuccessor(ICI->getPredicate() == ICmpInst::ICMP_EQ);
810 }
811 
812 /// Given a vector of bb/value pairs, remove any entries
813 /// in the list that match the specified block.
814 static void
816  std::vector<ValueEqualityComparisonCase> &Cases) {
817  llvm::erase_value(Cases, BB);
818 }
819 
820 /// Return true if there are any keys in C1 that exist in C2 as well.
821 static bool ValuesOverlap(std::vector<ValueEqualityComparisonCase> &C1,
822  std::vector<ValueEqualityComparisonCase> &C2) {
823  std::vector<ValueEqualityComparisonCase> *V1 = &C1, *V2 = &C2;
824 
825  // Make V1 be smaller than V2.
826  if (V1->size() > V2->size())
827  std::swap(V1, V2);
828 
829  if (V1->empty())
830  return false;
831  if (V1->size() == 1) {
832  // Just scan V2.
833  ConstantInt *TheVal = (*V1)[0].Value;
834  for (unsigned i = 0, e = V2->size(); i != e; ++i)
835  if (TheVal == (*V2)[i].Value)
836  return true;
837  }
838 
839  // Otherwise, just sort both lists and compare element by element.
840  array_pod_sort(V1->begin(), V1->end());
841  array_pod_sort(V2->begin(), V2->end());
842  unsigned i1 = 0, i2 = 0, e1 = V1->size(), e2 = V2->size();
843  while (i1 != e1 && i2 != e2) {
844  if ((*V1)[i1].Value == (*V2)[i2].Value)
845  return true;
846  if ((*V1)[i1].Value < (*V2)[i2].Value)
847  ++i1;
848  else
849  ++i2;
850  }
851  return false;
852 }
853 
854 // Set branch weights on SwitchInst. This sets the metadata if there is at
855 // least one non-zero weight.
857  // Check that there is at least one non-zero weight. Otherwise, pass
858  // nullptr to setMetadata which will erase the existing metadata.
859  MDNode *N = nullptr;
860  if (llvm::any_of(Weights, [](uint32_t W) { return W != 0; }))
861  N = MDBuilder(SI->getParent()->getContext()).createBranchWeights(Weights);
862  SI->setMetadata(LLVMContext::MD_prof, N);
863 }
864 
865 // Similar to the above, but for branch and select instructions that take
866 // exactly 2 weights.
867 static void setBranchWeights(Instruction *I, uint32_t TrueWeight,
868  uint32_t FalseWeight) {
869  assert(isa<BranchInst>(I) || isa<SelectInst>(I));
870  // Check that there is at least one non-zero weight. Otherwise, pass
871  // nullptr to setMetadata which will erase the existing metadata.
872  MDNode *N = nullptr;
873  if (TrueWeight || FalseWeight)
874  N = MDBuilder(I->getParent()->getContext())
875  .createBranchWeights(TrueWeight, FalseWeight);
876  I->setMetadata(LLVMContext::MD_prof, N);
877 }
878 
879 /// If TI is known to be a terminator instruction and its block is known to
880 /// only have a single predecessor block, check to see if that predecessor is
881 /// also a value comparison with the same value, and if that comparison
882 /// determines the outcome of this comparison. If so, simplify TI. This does a
883 /// very limited form of jump threading.
884 bool SimplifyCFGOpt::SimplifyEqualityComparisonWithOnlyPredecessor(
885  Instruction *TI, BasicBlock *Pred, IRBuilder<> &Builder) {
886  Value *PredVal = isValueEqualityComparison(Pred->getTerminator());
887  if (!PredVal)
888  return false; // Not a value comparison in predecessor.
889 
890  Value *ThisVal = isValueEqualityComparison(TI);
891  assert(ThisVal && "This isn't a value comparison!!");
892  if (ThisVal != PredVal)
893  return false; // Different predicates.
894 
895  // TODO: Preserve branch weight metadata, similarly to how
896  // FoldValueComparisonIntoPredecessors preserves it.
897 
898  // Find out information about when control will move from Pred to TI's block.
899  std::vector<ValueEqualityComparisonCase> PredCases;
900  BasicBlock *PredDef =
901  GetValueEqualityComparisonCases(Pred->getTerminator(), PredCases);
902  EliminateBlockCases(PredDef, PredCases); // Remove default from cases.
903 
904  // Find information about how control leaves this block.
905  std::vector<ValueEqualityComparisonCase> ThisCases;
906  BasicBlock *ThisDef = GetValueEqualityComparisonCases(TI, ThisCases);
907  EliminateBlockCases(ThisDef, ThisCases); // Remove default from cases.
908 
909  // If TI's block is the default block from Pred's comparison, potentially
910  // simplify TI based on this knowledge.
911  if (PredDef == TI->getParent()) {
912  // If we are here, we know that the value is none of those cases listed in
913  // PredCases. If there are any cases in ThisCases that are in PredCases, we
914  // can simplify TI.
915  if (!ValuesOverlap(PredCases, ThisCases))
916  return false;
917 
918  if (isa<BranchInst>(TI)) {
919  // Okay, one of the successors of this condbr is dead. Convert it to a
920  // uncond br.
921  assert(ThisCases.size() == 1 && "Branch can only have one case!");
922  // Insert the new branch.
923  Instruction *NI = Builder.CreateBr(ThisDef);
924  (void)NI;
925 
926  // Remove PHI node entries for the dead edge.
927  ThisCases[0].Dest->removePredecessor(PredDef);
928 
929  LLVM_DEBUG(dbgs() << "Threading pred instr: " << *Pred->getTerminator()
930  << "Through successor TI: " << *TI << "Leaving: " << *NI
931  << "\n");
932 
934 
935  if (DTU)
936  DTU->applyUpdates(
937  {{DominatorTree::Delete, PredDef, ThisCases[0].Dest}});
938 
939  return true;
940  }
941 
942  SwitchInstProfUpdateWrapper SI = *cast<SwitchInst>(TI);
943  // Okay, TI has cases that are statically dead, prune them away.
944  SmallPtrSet<Constant *, 16> DeadCases;
945  for (unsigned i = 0, e = PredCases.size(); i != e; ++i)
946  DeadCases.insert(PredCases[i].Value);
947 
948  LLVM_DEBUG(dbgs() << "Threading pred instr: " << *Pred->getTerminator()
949  << "Through successor TI: " << *TI);
950 
951  SmallDenseMap<BasicBlock *, int, 8> NumPerSuccessorCases;
952  for (SwitchInst::CaseIt i = SI->case_end(), e = SI->case_begin(); i != e;) {
953  --i;
954  auto *Successor = i->getCaseSuccessor();
955  if (DTU)
956  ++NumPerSuccessorCases[Successor];
957  if (DeadCases.count(i->getCaseValue())) {
958  Successor->removePredecessor(PredDef);
959  SI.removeCase(i);
960  if (DTU)
961  --NumPerSuccessorCases[Successor];
962  }
963  }
964 
965  if (DTU) {
966  std::vector<DominatorTree::UpdateType> Updates;
967  for (const std::pair<BasicBlock *, int> &I : NumPerSuccessorCases)
968  if (I.second == 0)
969  Updates.push_back({DominatorTree::Delete, PredDef, I.first});
970  DTU->applyUpdates(Updates);
971  }
972 
973  LLVM_DEBUG(dbgs() << "Leaving: " << *TI << "\n");
974  return true;
975  }
976 
977  // Otherwise, TI's block must correspond to some matched value. Find out
978  // which value (or set of values) this is.
979  ConstantInt *TIV = nullptr;
980  BasicBlock *TIBB = TI->getParent();
981  for (unsigned i = 0, e = PredCases.size(); i != e; ++i)
982  if (PredCases[i].Dest == TIBB) {
983  if (TIV)
984  return false; // Cannot handle multiple values coming to this block.
985  TIV = PredCases[i].Value;
986  }
987  assert(TIV && "No edge from pred to succ?");
988 
989  // Okay, we found the one constant that our value can be if we get into TI's
990  // BB. Find out which successor will unconditionally be branched to.
991  BasicBlock *TheRealDest = nullptr;
992  for (unsigned i = 0, e = ThisCases.size(); i != e; ++i)
993  if (ThisCases[i].Value == TIV) {
994  TheRealDest = ThisCases[i].Dest;
995  break;
996  }
997 
998  // If not handled by any explicit cases, it is handled by the default case.
999  if (!TheRealDest)
1000  TheRealDest = ThisDef;
1001 
1002  SmallPtrSet<BasicBlock *, 2> RemovedSuccs;
1003 
1004  // Remove PHI node entries for dead edges.
1005  BasicBlock *CheckEdge = TheRealDest;
1006  for (BasicBlock *Succ : successors(TIBB))
1007  if (Succ != CheckEdge) {
1008  if (Succ != TheRealDest)
1009  RemovedSuccs.insert(Succ);
1010  Succ->removePredecessor(TIBB);
1011  } else
1012  CheckEdge = nullptr;
1013 
1014  // Insert the new branch.
1015  Instruction *NI = Builder.CreateBr(TheRealDest);
1016  (void)NI;
1017 
1018  LLVM_DEBUG(dbgs() << "Threading pred instr: " << *Pred->getTerminator()
1019  << "Through successor TI: " << *TI << "Leaving: " << *NI
1020  << "\n");
1021 
1023  if (DTU) {
1025  Updates.reserve(RemovedSuccs.size());
1026  for (auto *RemovedSucc : RemovedSuccs)
1027  Updates.push_back({DominatorTree::Delete, TIBB, RemovedSucc});
1028  DTU->applyUpdates(Updates);
1029  }
1030  return true;
1031 }
1032 
1033 namespace {
1034 
1035 /// This class implements a stable ordering of constant
1036 /// integers that does not depend on their address. This is important for
1037 /// applications that sort ConstantInt's to ensure uniqueness.
1038 struct ConstantIntOrdering {
1039  bool operator()(const ConstantInt *LHS, const ConstantInt *RHS) const {
1040  return LHS->getValue().ult(RHS->getValue());
1041  }
1042 };
1043 
1044 } // end anonymous namespace
1045 
1047  ConstantInt *const *P2) {
1048  const ConstantInt *LHS = *P1;
1049  const ConstantInt *RHS = *P2;
1050  if (LHS == RHS)
1051  return 0;
1052  return LHS->getValue().ult(RHS->getValue()) ? 1 : -1;
1053 }
1054 
1055 static inline bool HasBranchWeights(const Instruction *I) {
1056  MDNode *ProfMD = I->getMetadata(LLVMContext::MD_prof);
1057  if (ProfMD && ProfMD->getOperand(0))
1058  if (MDString *MDS = dyn_cast<MDString>(ProfMD->getOperand(0)))
1059  return MDS->getString().equals("branch_weights");
1060 
1061  return false;
1062 }
1063 
1064 /// Get Weights of a given terminator, the default weight is at the front
1065 /// of the vector. If TI is a conditional eq, we need to swap the branch-weight
1066 /// metadata.
1068  SmallVectorImpl<uint64_t> &Weights) {
1069  MDNode *MD = TI->getMetadata(LLVMContext::MD_prof);
1070  assert(MD);
1071  for (unsigned i = 1, e = MD->getNumOperands(); i < e; ++i) {
1072  ConstantInt *CI = mdconst::extract<ConstantInt>(MD->getOperand(i));
1073  Weights.push_back(CI->getValue().getZExtValue());
1074  }
1075 
1076  // If TI is a conditional eq, the default case is the false case,
1077  // and the corresponding branch-weight data is at index 2. We swap the
1078  // default weight to be the first entry.
1079  if (BranchInst *BI = dyn_cast<BranchInst>(TI)) {
1080  assert(Weights.size() == 2);
1081  ICmpInst *ICI = cast<ICmpInst>(BI->getCondition());
1082  if (ICI->getPredicate() == ICmpInst::ICMP_EQ)
1083  std::swap(Weights.front(), Weights.back());
1084  }
1085 }
1086 
1087 /// Keep halving the weights until all can fit in uint32_t.
1089  uint64_t Max = *std::max_element(Weights.begin(), Weights.end());
1090  if (Max > UINT_MAX) {
1091  unsigned Offset = 32 - countLeadingZeros(Max);
1092  for (uint64_t &I : Weights)
1093  I >>= Offset;
1094  }
1095 }
1096 
1098  BasicBlock *BB, BasicBlock *PredBlock, ValueToValueMapTy &VMap) {
1099  Instruction *PTI = PredBlock->getTerminator();
1100 
1101  // If we have bonus instructions, clone them into the predecessor block.
1102  // Note that there may be multiple predecessor blocks, so we cannot move
1103  // bonus instructions to a predecessor block.
1104  for (Instruction &BonusInst : *BB) {
1105  if (isa<DbgInfoIntrinsic>(BonusInst) || BonusInst.isTerminator())
1106  continue;
1107 
1108  Instruction *NewBonusInst = BonusInst.clone();
1109 
1110  if (PTI->getDebugLoc() != NewBonusInst->getDebugLoc()) {
1111  // Unless the instruction has the same !dbg location as the original
1112  // branch, drop it. When we fold the bonus instructions we want to make
1113  // sure we reset their debug locations in order to avoid stepping on
1114  // dead code caused by folding dead branches.
1115  NewBonusInst->setDebugLoc(DebugLoc());
1116  }
1117 
1118  RemapInstruction(NewBonusInst, VMap,
1120  VMap[&BonusInst] = NewBonusInst;
1121 
1122  // If we moved a load, we cannot any longer claim any knowledge about
1123  // its potential value. The previous information might have been valid
1124  // only given the branch precondition.
1125  // For an analogous reason, we must also drop all the metadata whose
1126  // semantics we don't understand. We *can* preserve !annotation, because
1127  // it is tied to the instruction itself, not the value or position.
1128  // Similarly strip attributes on call parameters that may cause UB in
1129  // location the call is moved to.
1131  LLVMContext::MD_annotation);
1132 
1133  PredBlock->getInstList().insert(PTI->getIterator(), NewBonusInst);
1134  NewBonusInst->takeName(&BonusInst);
1135  BonusInst.setName(NewBonusInst->getName() + ".old");
1136 
1137  // Update (liveout) uses of bonus instructions,
1138  // now that the bonus instruction has been cloned into predecessor.
1139  // Note that we expect to be in a block-closed SSA form for this to work!
1140  for (Use &U : make_early_inc_range(BonusInst.uses())) {
1141  auto *UI = cast<Instruction>(U.getUser());
1142  auto *PN = dyn_cast<PHINode>(UI);
1143  if (!PN) {
1144  assert(UI->getParent() == BB && BonusInst.comesBefore(UI) &&
1145  "If the user is not a PHI node, then it should be in the same "
1146  "block as, and come after, the original bonus instruction.");
1147  continue; // Keep using the original bonus instruction.
1148  }
1149  // Is this the block-closed SSA form PHI node?
1150  if (PN->getIncomingBlock(U) == BB)
1151  continue; // Great, keep using the original bonus instruction.
1152  // The only other alternative is an "use" when coming from
1153  // the predecessor block - here we should refer to the cloned bonus instr.
1154  assert(PN->getIncomingBlock(U) == PredBlock &&
1155  "Not in block-closed SSA form?");
1156  U.set(NewBonusInst);
1157  }
1158  }
1159 }
1160 
1161 bool SimplifyCFGOpt::PerformValueComparisonIntoPredecessorFolding(
1162  Instruction *TI, Value *&CV, Instruction *PTI, IRBuilder<> &Builder) {
1163  BasicBlock *BB = TI->getParent();
1164  BasicBlock *Pred = PTI->getParent();
1165 
1167 
1168  // Figure out which 'cases' to copy from SI to PSI.
1169  std::vector<ValueEqualityComparisonCase> BBCases;
1170  BasicBlock *BBDefault = GetValueEqualityComparisonCases(TI, BBCases);
1171 
1172  std::vector<ValueEqualityComparisonCase> PredCases;
1173  BasicBlock *PredDefault = GetValueEqualityComparisonCases(PTI, PredCases);
1174 
1175  // Based on whether the default edge from PTI goes to BB or not, fill in
1176  // PredCases and PredDefault with the new switch cases we would like to
1177  // build.
1179 
1180  // Update the branch weight metadata along the way
1181  SmallVector<uint64_t, 8> Weights;
1182  bool PredHasWeights = HasBranchWeights(PTI);
1183  bool SuccHasWeights = HasBranchWeights(TI);
1184 
1185  if (PredHasWeights) {
1186  GetBranchWeights(PTI, Weights);
1187  // branch-weight metadata is inconsistent here.
1188  if (Weights.size() != 1 + PredCases.size())
1189  PredHasWeights = SuccHasWeights = false;
1190  } else if (SuccHasWeights)
1191  // If there are no predecessor weights but there are successor weights,
1192  // populate Weights with 1, which will later be scaled to the sum of
1193  // successor's weights
1194  Weights.assign(1 + PredCases.size(), 1);
1195 
1196  SmallVector<uint64_t, 8> SuccWeights;
1197  if (SuccHasWeights) {
1198  GetBranchWeights(TI, SuccWeights);
1199  // branch-weight metadata is inconsistent here.
1200  if (SuccWeights.size() != 1 + BBCases.size())
1201  PredHasWeights = SuccHasWeights = false;
1202  } else if (PredHasWeights)
1203  SuccWeights.assign(1 + BBCases.size(), 1);
1204 
1205  if (PredDefault == BB) {
1206  // If this is the default destination from PTI, only the edges in TI
1207  // that don't occur in PTI, or that branch to BB will be activated.
1208  std::set<ConstantInt *, ConstantIntOrdering> PTIHandled;
1209  for (unsigned i = 0, e = PredCases.size(); i != e; ++i)
1210  if (PredCases[i].Dest != BB)
1211  PTIHandled.insert(PredCases[i].Value);
1212  else {
1213  // The default destination is BB, we don't need explicit targets.
1214  std::swap(PredCases[i], PredCases.back());
1215 
1216  if (PredHasWeights || SuccHasWeights) {
1217  // Increase weight for the default case.
1218  Weights[0] += Weights[i + 1];
1219  std::swap(Weights[i + 1], Weights.back());
1220  Weights.pop_back();
1221  }
1222 
1223  PredCases.pop_back();
1224  --i;
1225  --e;
1226  }
1227 
1228  // Reconstruct the new switch statement we will be building.
1229  if (PredDefault != BBDefault) {
1230  PredDefault->removePredecessor(Pred);
1231  if (DTU && PredDefault != BB)
1232  Updates.push_back({DominatorTree::Delete, Pred, PredDefault});
1233  PredDefault = BBDefault;
1234  ++NewSuccessors[BBDefault];
1235  }
1236 
1237  unsigned CasesFromPred = Weights.size();
1238  uint64_t ValidTotalSuccWeight = 0;
1239  for (unsigned i = 0, e = BBCases.size(); i != e; ++i)
1240  if (!PTIHandled.count(BBCases[i].Value) && BBCases[i].Dest != BBDefault) {
1241  PredCases.push_back(BBCases[i]);
1242  ++NewSuccessors[BBCases[i].Dest];
1243  if (SuccHasWeights || PredHasWeights) {
1244  // The default weight is at index 0, so weight for the ith case
1245  // should be at index i+1. Scale the cases from successor by
1246  // PredDefaultWeight (Weights[0]).
1247  Weights.push_back(Weights[0] * SuccWeights[i + 1]);
1248  ValidTotalSuccWeight += SuccWeights[i + 1];
1249  }
1250  }
1251 
1252  if (SuccHasWeights || PredHasWeights) {
1253  ValidTotalSuccWeight += SuccWeights[0];
1254  // Scale the cases from predecessor by ValidTotalSuccWeight.
1255  for (unsigned i = 1; i < CasesFromPred; ++i)
1256  Weights[i] *= ValidTotalSuccWeight;
1257  // Scale the default weight by SuccDefaultWeight (SuccWeights[0]).
1258  Weights[0] *= SuccWeights[0];
1259  }
1260  } else {
1261  // If this is not the default destination from PSI, only the edges
1262  // in SI that occur in PSI with a destination of BB will be
1263  // activated.
1264  std::set<ConstantInt *, ConstantIntOrdering> PTIHandled;
1265  std::map<ConstantInt *, uint64_t> WeightsForHandled;
1266  for (unsigned i = 0, e = PredCases.size(); i != e; ++i)
1267  if (PredCases[i].Dest == BB) {
1268  PTIHandled.insert(PredCases[i].Value);
1269 
1270  if (PredHasWeights || SuccHasWeights) {
1271  WeightsForHandled[PredCases[i].Value] = Weights[i + 1];
1272  std::swap(Weights[i + 1], Weights.back());
1273  Weights.pop_back();
1274  }
1275 
1276  std::swap(PredCases[i], PredCases.back());
1277  PredCases.pop_back();
1278  --i;
1279  --e;
1280  }
1281 
1282  // Okay, now we know which constants were sent to BB from the
1283  // predecessor. Figure out where they will all go now.
1284  for (unsigned i = 0, e = BBCases.size(); i != e; ++i)
1285  if (PTIHandled.count(BBCases[i].Value)) {
1286  // If this is one we are capable of getting...
1287  if (PredHasWeights || SuccHasWeights)
1288  Weights.push_back(WeightsForHandled[BBCases[i].Value]);
1289  PredCases.push_back(BBCases[i]);
1290  ++NewSuccessors[BBCases[i].Dest];
1291  PTIHandled.erase(BBCases[i].Value); // This constant is taken care of
1292  }
1293 
1294  // If there are any constants vectored to BB that TI doesn't handle,
1295  // they must go to the default destination of TI.
1296  for (ConstantInt *I : PTIHandled) {
1297  if (PredHasWeights || SuccHasWeights)
1298  Weights.push_back(WeightsForHandled[I]);
1299  PredCases.push_back(ValueEqualityComparisonCase(I, BBDefault));
1300  ++NewSuccessors[BBDefault];
1301  }
1302  }
1303 
1304  // Okay, at this point, we know which new successor Pred will get. Make
1305  // sure we update the number of entries in the PHI nodes for these
1306  // successors.
1307  SmallPtrSet<BasicBlock *, 2> SuccsOfPred;
1308  if (DTU) {
1309  SuccsOfPred = {succ_begin(Pred), succ_end(Pred)};
1310  Updates.reserve(Updates.size() + NewSuccessors.size());
1311  }
1312  for (const std::pair<BasicBlock *, int /*Num*/> &NewSuccessor :
1313  NewSuccessors) {
1314  for (auto I : seq(0, NewSuccessor.second)) {
1315  (void)I;
1316  AddPredecessorToBlock(NewSuccessor.first, Pred, BB);
1317  }
1318  if (DTU && !SuccsOfPred.contains(NewSuccessor.first))
1319  Updates.push_back({DominatorTree::Insert, Pred, NewSuccessor.first});
1320  }
1321 
1322  Builder.SetInsertPoint(PTI);
1323  // Convert pointer to int before we switch.
1324  if (CV->getType()->isPointerTy()) {
1325  CV =
1326  Builder.CreatePtrToInt(CV, DL.getIntPtrType(CV->getType()), "magicptr");
1327  }
1328 
1329  // Now that the successors are updated, create the new Switch instruction.
1330  SwitchInst *NewSI = Builder.CreateSwitch(CV, PredDefault, PredCases.size());
1331  NewSI->setDebugLoc(PTI->getDebugLoc());
1332  for (ValueEqualityComparisonCase &V : PredCases)
1333  NewSI->addCase(V.Value, V.Dest);
1334 
1335  if (PredHasWeights || SuccHasWeights) {
1336  // Halve the weights if any of them cannot fit in an uint32_t
1337  FitWeights(Weights);
1338 
1339  SmallVector<uint32_t, 8> MDWeights(Weights.begin(), Weights.end());
1340 
1341  setBranchWeights(NewSI, MDWeights);
1342  }
1343 
1345 
1346  // Okay, last check. If BB is still a successor of PSI, then we must
1347  // have an infinite loop case. If so, add an infinitely looping block
1348  // to handle the case to preserve the behavior of the code.
1349  BasicBlock *InfLoopBlock = nullptr;
1350  for (unsigned i = 0, e = NewSI->getNumSuccessors(); i != e; ++i)
1351  if (NewSI->getSuccessor(i) == BB) {
1352  if (!InfLoopBlock) {
1353  // Insert it at the end of the function, because it's either code,
1354  // or it won't matter if it's hot. :)
1355  InfLoopBlock =
1356  BasicBlock::Create(BB->getContext(), "infloop", BB->getParent());
1357  BranchInst::Create(InfLoopBlock, InfLoopBlock);
1358  if (DTU)
1359  Updates.push_back(
1360  {DominatorTree::Insert, InfLoopBlock, InfLoopBlock});
1361  }
1362  NewSI->setSuccessor(i, InfLoopBlock);
1363  }
1364 
1365  if (DTU) {
1366  if (InfLoopBlock)
1367  Updates.push_back({DominatorTree::Insert, Pred, InfLoopBlock});
1368 
1369  Updates.push_back({DominatorTree::Delete, Pred, BB});
1370 
1371  DTU->applyUpdates(Updates);
1372  }
1373 
1374  ++NumFoldValueComparisonIntoPredecessors;
1375  return true;
1376 }
1377 
1378 /// The specified terminator is a value equality comparison instruction
1379 /// (either a switch or a branch on "X == c").
1380 /// See if any of the predecessors of the terminator block are value comparisons
1381 /// on the same value. If so, and if safe to do so, fold them together.
1382 bool SimplifyCFGOpt::FoldValueComparisonIntoPredecessors(Instruction *TI,
1383  IRBuilder<> &Builder) {
1384  BasicBlock *BB = TI->getParent();
1385  Value *CV = isValueEqualityComparison(TI); // CondVal
1386  assert(CV && "Not a comparison?");
1387 
1388  bool Changed = false;
1389 
1391  while (!Preds.empty()) {
1392  BasicBlock *Pred = Preds.pop_back_val();
1393  Instruction *PTI = Pred->getTerminator();
1394 
1395  // Don't try to fold into itself.
1396  if (Pred == BB)
1397  continue;
1398 
1399  // See if the predecessor is a comparison with the same value.
1400  Value *PCV = isValueEqualityComparison(PTI); // PredCondVal
1401  if (PCV != CV)
1402  continue;
1403 
1405  if (!SafeToMergeTerminators(TI, PTI, &FailBlocks)) {
1406  for (auto *Succ : FailBlocks) {
1407  if (!SplitBlockPredecessors(Succ, TI->getParent(), ".fold.split", DTU))
1408  return false;
1409  }
1410  }
1411 
1412  PerformValueComparisonIntoPredecessorFolding(TI, CV, PTI, Builder);
1413  Changed = true;
1414  }
1415  return Changed;
1416 }
1417 
1418 // If we would need to insert a select that uses the value of this invoke
1419 // (comments in HoistThenElseCodeToIf explain why we would need to do this), we
1420 // can't hoist the invoke, as there is nowhere to put the select in this case.
1422  Instruction *I1, Instruction *I2) {
1423  for (BasicBlock *Succ : successors(BB1)) {
1424  for (const PHINode &PN : Succ->phis()) {
1425  Value *BB1V = PN.getIncomingValueForBlock(BB1);
1426  Value *BB2V = PN.getIncomingValueForBlock(BB2);
1427  if (BB1V != BB2V && (BB1V == I1 || BB2V == I2)) {
1428  return false;
1429  }
1430  }
1431  }
1432  return true;
1433 }
1434 
1435 static bool passingValueIsAlwaysUndefined(Value *V, Instruction *I, bool PtrValueMayBeModified = false);
1436 
1437 /// Given a conditional branch that goes to BB1 and BB2, hoist any common code
1438 /// in the two blocks up into the branch block. The caller of this function
1439 /// guarantees that BI's block dominates BB1 and BB2. If EqTermsOnly is given,
1440 /// only perform hoisting in case both blocks only contain a terminator. In that
1441 /// case, only the original BI will be replaced and selects for PHIs are added.
1442 bool SimplifyCFGOpt::HoistThenElseCodeToIf(BranchInst *BI,
1443  const TargetTransformInfo &TTI,
1444  bool EqTermsOnly) {
1445  // This does very trivial matching, with limited scanning, to find identical
1446  // instructions in the two blocks. In particular, we don't want to get into
1447  // O(M*N) situations here where M and N are the sizes of BB1 and BB2. As
1448  // such, we currently just scan for obviously identical instructions in an
1449  // identical order.
1450  BasicBlock *BB1 = BI->getSuccessor(0); // The true destination.
1451  BasicBlock *BB2 = BI->getSuccessor(1); // The false destination
1452 
1453  // If either of the blocks has it's address taken, then we can't do this fold,
1454  // because the code we'd hoist would no longer run when we jump into the block
1455  // by it's address.
1456  if (BB1->hasAddressTaken() || BB2->hasAddressTaken())
1457  return false;
1458 
1459  BasicBlock::iterator BB1_Itr = BB1->begin();
1460  BasicBlock::iterator BB2_Itr = BB2->begin();
1461 
1462  Instruction *I1 = &*BB1_Itr++, *I2 = &*BB2_Itr++;
1463  // Skip debug info if it is not identical.
1464  DbgInfoIntrinsic *DBI1 = dyn_cast<DbgInfoIntrinsic>(I1);
1465  DbgInfoIntrinsic *DBI2 = dyn_cast<DbgInfoIntrinsic>(I2);
1466  if (!DBI1 || !DBI2 || !DBI1->isIdenticalToWhenDefined(DBI2)) {
1467  while (isa<DbgInfoIntrinsic>(I1))
1468  I1 = &*BB1_Itr++;
1469  while (isa<DbgInfoIntrinsic>(I2))
1470  I2 = &*BB2_Itr++;
1471  }
1472  // FIXME: Can we define a safety predicate for CallBr?
1473  if (isa<PHINode>(I1) || !I1->isIdenticalToWhenDefined(I2) ||
1474  (isa<InvokeInst>(I1) && !isSafeToHoistInvoke(BB1, BB2, I1, I2)) ||
1475  isa<CallBrInst>(I1))
1476  return false;
1477 
1478  BasicBlock *BIParent = BI->getParent();
1479 
1480  bool Changed = false;
1481 
1482  auto _ = make_scope_exit([&]() {
1483  if (Changed)
1484  ++NumHoistCommonCode;
1485  });
1486 
1487  // Check if only hoisting terminators is allowed. This does not add new
1488  // instructions to the hoist location.
1489  if (EqTermsOnly) {
1490  // Skip any debug intrinsics, as they are free to hoist.
1491  auto *I1NonDbg = &*skipDebugIntrinsics(I1->getIterator());
1492  auto *I2NonDbg = &*skipDebugIntrinsics(I2->getIterator());
1493  if (!I1NonDbg->isIdenticalToWhenDefined(I2NonDbg))
1494  return false;
1495  if (!I1NonDbg->isTerminator())
1496  return false;
1497  // Now we know that we only need to hoist debug intrinsics and the
1498  // terminator. Let the loop below handle those 2 cases.
1499  }
1500 
1501  do {
1502  // If we are hoisting the terminator instruction, don't move one (making a
1503  // broken BB), instead clone it, and remove BI.
1504  if (I1->isTerminator())
1505  goto HoistTerminator;
1506 
1507  // If we're going to hoist a call, make sure that the two instructions we're
1508  // commoning/hoisting are both marked with musttail, or neither of them is
1509  // marked as such. Otherwise, we might end up in a situation where we hoist
1510  // from a block where the terminator is a `ret` to a block where the terminator
1511  // is a `br`, and `musttail` calls expect to be followed by a return.
1512  auto *C1 = dyn_cast<CallInst>(I1);
1513  auto *C2 = dyn_cast<CallInst>(I2);
1514  if (C1 && C2)
1515  if (C1->isMustTailCall() != C2->isMustTailCall())
1516  return Changed;
1517 
1519  return Changed;
1520 
1521  // If any of the two call sites has nomerge attribute, stop hoisting.
1522  if (const auto *CB1 = dyn_cast<CallBase>(I1))
1523  if (CB1->cannotMerge())
1524  return Changed;
1525  if (const auto *CB2 = dyn_cast<CallBase>(I2))
1526  if (CB2->cannotMerge())
1527  return Changed;
1528 
1529  if (isa<DbgInfoIntrinsic>(I1) || isa<DbgInfoIntrinsic>(I2)) {
1530  assert (isa<DbgInfoIntrinsic>(I1) && isa<DbgInfoIntrinsic>(I2));
1531  // The debug location is an integral part of a debug info intrinsic
1532  // and can't be separated from it or replaced. Instead of attempting
1533  // to merge locations, simply hoist both copies of the intrinsic.
1534  BIParent->getInstList().splice(BI->getIterator(),
1535  BB1->getInstList(), I1);
1536  BIParent->getInstList().splice(BI->getIterator(),
1537  BB2->getInstList(), I2);
1538  Changed = true;
1539  } else {
1540  // For a normal instruction, we just move one to right before the branch,
1541  // then replace all uses of the other with the first. Finally, we remove
1542  // the now redundant second instruction.
1543  BIParent->getInstList().splice(BI->getIterator(),
1544  BB1->getInstList(), I1);
1545  if (!I2->use_empty())
1546  I2->replaceAllUsesWith(I1);
1547  I1->andIRFlags(I2);
1548  unsigned KnownIDs[] = {LLVMContext::MD_tbaa,
1549  LLVMContext::MD_range,
1550  LLVMContext::MD_fpmath,
1551  LLVMContext::MD_invariant_load,
1552  LLVMContext::MD_nonnull,
1553  LLVMContext::MD_invariant_group,
1554  LLVMContext::MD_align,
1555  LLVMContext::MD_dereferenceable,
1556  LLVMContext::MD_dereferenceable_or_null,
1557  LLVMContext::MD_mem_parallel_loop_access,
1558  LLVMContext::MD_access_group,
1559  LLVMContext::MD_preserve_access_index};
1560  combineMetadata(I1, I2, KnownIDs, true);
1561 
1562  // I1 and I2 are being combined into a single instruction. Its debug
1563  // location is the merged locations of the original instructions.
1564  I1->applyMergedLocation(I1->getDebugLoc(), I2->getDebugLoc());
1565 
1566  I2->eraseFromParent();
1567  Changed = true;
1568  }
1569  ++NumHoistCommonInstrs;
1570 
1571  I1 = &*BB1_Itr++;
1572  I2 = &*BB2_Itr++;
1573  // Skip debug info if it is not identical.
1574  DbgInfoIntrinsic *DBI1 = dyn_cast<DbgInfoIntrinsic>(I1);
1575  DbgInfoIntrinsic *DBI2 = dyn_cast<DbgInfoIntrinsic>(I2);
1576  if (!DBI1 || !DBI2 || !DBI1->isIdenticalToWhenDefined(DBI2)) {
1577  while (isa<DbgInfoIntrinsic>(I1))
1578  I1 = &*BB1_Itr++;
1579  while (isa<DbgInfoIntrinsic>(I2))
1580  I2 = &*BB2_Itr++;
1581  }
1582  } while (I1->isIdenticalToWhenDefined(I2));
1583 
1584  return true;
1585 
1586 HoistTerminator:
1587  // It may not be possible to hoist an invoke.
1588  // FIXME: Can we define a safety predicate for CallBr?
1589  if (isa<InvokeInst>(I1) && !isSafeToHoistInvoke(BB1, BB2, I1, I2))
1590  return Changed;
1591 
1592  // TODO: callbr hoisting currently disabled pending further study.
1593  if (isa<CallBrInst>(I1))
1594  return Changed;
1595 
1596  for (BasicBlock *Succ : successors(BB1)) {
1597  for (PHINode &PN : Succ->phis()) {
1598  Value *BB1V = PN.getIncomingValueForBlock(BB1);
1599  Value *BB2V = PN.getIncomingValueForBlock(BB2);
1600  if (BB1V == BB2V)
1601  continue;
1602 
1603  // Check for passingValueIsAlwaysUndefined here because we would rather
1604  // eliminate undefined control flow then converting it to a select.
1605  if (passingValueIsAlwaysUndefined(BB1V, &PN) ||
1606  passingValueIsAlwaysUndefined(BB2V, &PN))
1607  return Changed;
1608 
1609  if (isa<ConstantExpr>(BB1V) && !isSafeToSpeculativelyExecute(BB1V))
1610  return Changed;
1611  if (isa<ConstantExpr>(BB2V) && !isSafeToSpeculativelyExecute(BB2V))
1612  return Changed;
1613  }
1614  }
1615 
1616  // Okay, it is safe to hoist the terminator.
1617  Instruction *NT = I1->clone();
1618  BIParent->getInstList().insert(BI->getIterator(), NT);
1619  if (!NT->getType()->isVoidTy()) {
1620  I1->replaceAllUsesWith(NT);
1621  I2->replaceAllUsesWith(NT);
1622  NT->takeName(I1);
1623  }
1624  Changed = true;
1625  ++NumHoistCommonInstrs;
1626 
1627  // Ensure terminator gets a debug location, even an unknown one, in case
1628  // it involves inlinable calls.
1629  NT->applyMergedLocation(I1->getDebugLoc(), I2->getDebugLoc());
1630 
1631  // PHIs created below will adopt NT's merged DebugLoc.
1633 
1634  // Hoisting one of the terminators from our successor is a great thing.
1635  // Unfortunately, the successors of the if/else blocks may have PHI nodes in
1636  // them. If they do, all PHI entries for BB1/BB2 must agree for all PHI
1637  // nodes, so we insert select instruction to compute the final result.
1638  std::map<std::pair<Value *, Value *>, SelectInst *> InsertedSelects;
1639  for (BasicBlock *Succ : successors(BB1)) {
1640  for (PHINode &PN : Succ->phis()) {
1641  Value *BB1V = PN.getIncomingValueForBlock(BB1);
1642  Value *BB2V = PN.getIncomingValueForBlock(BB2);
1643  if (BB1V == BB2V)
1644  continue;
1645 
1646  // These values do not agree. Insert a select instruction before NT
1647  // that determines the right value.
1648  SelectInst *&SI = InsertedSelects[std::make_pair(BB1V, BB2V)];
1649  if (!SI) {
1650  // Propagate fast-math-flags from phi node to its replacement select.
1652  if (isa<FPMathOperator>(PN))
1653  Builder.setFastMathFlags(PN.getFastMathFlags());
1654 
1655  SI = cast<SelectInst>(
1656  Builder.CreateSelect(BI->getCondition(), BB1V, BB2V,
1657  BB1V->getName() + "." + BB2V->getName(), BI));
1658  }
1659 
1660  // Make the PHI node use the select for all incoming values for BB1/BB2
1661  for (unsigned i = 0, e = PN.getNumIncomingValues(); i != e; ++i)
1662  if (PN.getIncomingBlock(i) == BB1 || PN.getIncomingBlock(i) == BB2)
1663  PN.setIncomingValue(i, SI);
1664  }
1665  }
1666 
1668 
1669  // Update any PHI nodes in our new successors.
1670  for (BasicBlock *Succ : successors(BB1)) {
1671  AddPredecessorToBlock(Succ, BIParent, BB1);
1672  if (DTU)
1673  Updates.push_back({DominatorTree::Insert, BIParent, Succ});
1674  }
1675 
1676  if (DTU)
1677  for (BasicBlock *Succ : successors(BI))
1678  Updates.push_back({DominatorTree::Delete, BIParent, Succ});
1679 
1681  if (DTU)
1682  DTU->applyUpdates(Updates);
1683  return Changed;
1684 }
1685 
1686 // Check lifetime markers.
1687 static bool isLifeTimeMarker(const Instruction *I) {
1688  if (auto II = dyn_cast<IntrinsicInst>(I)) {
1689  switch (II->getIntrinsicID()) {
1690  default:
1691  break;
1692  case Intrinsic::lifetime_start:
1693  case Intrinsic::lifetime_end:
1694  return true;
1695  }
1696  }
1697  return false;
1698 }
1699 
1700 // TODO: Refine this. This should avoid cases like turning constant memcpy sizes
1701 // into variables.
1703  int OpIdx) {
1704  return !isa<IntrinsicInst>(I);
1705 }
1706 
1707 // All instructions in Insts belong to different blocks that all unconditionally
1708 // branch to a common successor. Analyze each instruction and return true if it
1709 // would be possible to sink them into their successor, creating one common
1710 // instruction instead. For every value that would be required to be provided by
1711 // PHI node (because an operand varies in each input block), add to PHIOperands.
1714  DenseMap<Instruction *, SmallVector<Value *, 4>> &PHIOperands) {
1715  // Prune out obviously bad instructions to move. Each instruction must have
1716  // exactly zero or one use, and we check later that use is by a single, common
1717  // PHI instruction in the successor.
1718  bool HasUse = !Insts.front()->user_empty();
1719  for (auto *I : Insts) {
1720  // These instructions may change or break semantics if moved.
1721  if (isa<PHINode>(I) || I->isEHPad() || isa<AllocaInst>(I) ||
1722  I->getType()->isTokenTy())
1723  return false;
1724 
1725  // Do not try to sink an instruction in an infinite loop - it can cause
1726  // this algorithm to infinite loop.
1727  if (I->getParent()->getSingleSuccessor() == I->getParent())
1728  return false;
1729 
1730  // Conservatively return false if I is an inline-asm instruction. Sinking
1731  // and merging inline-asm instructions can potentially create arguments
1732  // that cannot satisfy the inline-asm constraints.
1733  // If the instruction has nomerge attribute, return false.
1734  if (const auto *C = dyn_cast<CallBase>(I))
1735  if (C->isInlineAsm() || C->cannotMerge())
1736  return false;
1737 
1738  // Each instruction must have zero or one use.
1739  if (HasUse && !I->hasOneUse())
1740  return false;
1741  if (!HasUse && !I->user_empty())
1742  return false;
1743  }
1744 
1745  const Instruction *I0 = Insts.front();
1746  for (auto *I : Insts)
1747  if (!I->isSameOperationAs(I0))
1748  return false;
1749 
1750  // All instructions in Insts are known to be the same opcode. If they have a
1751  // use, check that the only user is a PHI or in the same block as the
1752  // instruction, because if a user is in the same block as an instruction we're
1753  // contemplating sinking, it must already be determined to be sinkable.
1754  if (HasUse) {
1755  auto *PNUse = dyn_cast<PHINode>(*I0->user_begin());
1756  auto *Succ = I0->getParent()->getTerminator()->getSuccessor(0);
1757  if (!all_of(Insts, [&PNUse,&Succ](const Instruction *I) -> bool {
1758  auto *U = cast<Instruction>(*I->user_begin());
1759  return (PNUse &&
1760  PNUse->getParent() == Succ &&
1761  PNUse->getIncomingValueForBlock(I->getParent()) == I) ||
1762  U->getParent() == I->getParent();
1763  }))
1764  return false;
1765  }
1766 
1767  // Because SROA can't handle speculating stores of selects, try not to sink
1768  // loads, stores or lifetime markers of allocas when we'd have to create a
1769  // PHI for the address operand. Also, because it is likely that loads or
1770  // stores of allocas will disappear when Mem2Reg/SROA is run, don't sink
1771  // them.
1772  // This can cause code churn which can have unintended consequences down
1773  // the line - see https://llvm.org/bugs/show_bug.cgi?id=30244.
1774  // FIXME: This is a workaround for a deficiency in SROA - see
1775  // https://llvm.org/bugs/show_bug.cgi?id=30188
1776  if (isa<StoreInst>(I0) && any_of(Insts, [](const Instruction *I) {
1777  return isa<AllocaInst>(I->getOperand(1)->stripPointerCasts());
1778  }))
1779  return false;
1780  if (isa<LoadInst>(I0) && any_of(Insts, [](const Instruction *I) {
1781  return isa<AllocaInst>(I->getOperand(0)->stripPointerCasts());
1782  }))
1783  return false;
1784  if (isLifeTimeMarker(I0) && any_of(Insts, [](const Instruction *I) {
1785  return isa<AllocaInst>(I->getOperand(1)->stripPointerCasts());
1786  }))
1787  return false;
1788 
1789  // For calls to be sinkable, they must all be indirect, or have same callee.
1790  // I.e. if we have two direct calls to different callees, we don't want to
1791  // turn that into an indirect call. Likewise, if we have an indirect call,
1792  // and a direct call, we don't actually want to have a single indirect call.
1793  if (isa<CallBase>(I0)) {
1794  auto IsIndirectCall = [](const Instruction *I) {
1795  return cast<CallBase>(I)->isIndirectCall();
1796  };
1797  bool HaveIndirectCalls = any_of(Insts, IsIndirectCall);
1798  bool AllCallsAreIndirect = all_of(Insts, IsIndirectCall);
1799  if (HaveIndirectCalls) {
1800  if (!AllCallsAreIndirect)
1801  return false;
1802  } else {
1803  // All callees must be identical.
1804  Value *Callee = nullptr;
1805  for (const Instruction *I : Insts) {
1806  Value *CurrCallee = cast<CallBase>(I)->getCalledOperand();
1807  if (!Callee)
1808  Callee = CurrCallee;
1809  else if (Callee != CurrCallee)
1810  return false;
1811  }
1812  }
1813  }
1814 
1815  for (unsigned OI = 0, OE = I0->getNumOperands(); OI != OE; ++OI) {
1816  Value *Op = I0->getOperand(OI);
1817  if (Op->getType()->isTokenTy())
1818  // Don't touch any operand of token type.
1819  return false;
1820 
1821  auto SameAsI0 = [&I0, OI](const Instruction *I) {
1822  assert(I->getNumOperands() == I0->getNumOperands());
1823  return I->getOperand(OI) == I0->getOperand(OI);
1824  };
1825  if (!all_of(Insts, SameAsI0)) {
1826  if ((isa<Constant>(Op) && !replacingOperandWithVariableIsCheap(I0, OI)) ||
1828  // We can't create a PHI from this GEP.
1829  return false;
1830  for (auto *I : Insts)
1831  PHIOperands[I].push_back(I->getOperand(OI));
1832  }
1833  }
1834  return true;
1835 }
1836 
1837 // Assuming canSinkInstructions(Blocks) has returned true, sink the last
1838 // instruction of every block in Blocks to their common successor, commoning
1839 // into one instruction.
1841  auto *BBEnd = Blocks[0]->getTerminator()->getSuccessor(0);
1842 
1843  // canSinkInstructions returning true guarantees that every block has at
1844  // least one non-terminator instruction.
1846  for (auto *BB : Blocks) {
1847  Instruction *I = BB->getTerminator();
1848  do {
1849  I = I->getPrevNode();
1850  } while (isa<DbgInfoIntrinsic>(I) && I != &BB->front());
1851  if (!isa<DbgInfoIntrinsic>(I))
1852  Insts.push_back(I);
1853  }
1854 
1855  // The only checking we need to do now is that all users of all instructions
1856  // are the same PHI node. canSinkInstructions should have checked this but
1857  // it is slightly over-aggressive - it gets confused by commutative
1858  // instructions so double-check it here.
1859  Instruction *I0 = Insts.front();
1860  if (!I0->user_empty()) {
1861  auto *PNUse = dyn_cast<PHINode>(*I0->user_begin());
1862  if (!all_of(Insts, [&PNUse](const Instruction *I) -> bool {
1863  auto *U = cast<Instruction>(*I->user_begin());
1864  return U == PNUse;
1865  }))
1866  return false;
1867  }
1868 
1869  // We don't need to do any more checking here; canSinkInstructions should
1870  // have done it all for us.
1871  SmallVector<Value*, 4> NewOperands;
1872  for (unsigned O = 0, E = I0->getNumOperands(); O != E; ++O) {
1873  // This check is different to that in canSinkInstructions. There, we
1874  // cared about the global view once simplifycfg (and instcombine) have
1875  // completed - it takes into account PHIs that become trivially
1876  // simplifiable. However here we need a more local view; if an operand
1877  // differs we create a PHI and rely on instcombine to clean up the very
1878  // small mess we may make.
1879  bool NeedPHI = any_of(Insts, [&I0, O](const Instruction *I) {
1880  return I->getOperand(O) != I0->getOperand(O);
1881  });
1882  if (!NeedPHI) {
1883  NewOperands.push_back(I0->getOperand(O));
1884  continue;
1885  }
1886 
1887  // Create a new PHI in the successor block and populate it.
1888  auto *Op = I0->getOperand(O);
1889  assert(!Op->getType()->isTokenTy() && "Can't PHI tokens!");
1890  auto *PN = PHINode::Create(Op->getType(), Insts.size(),
1891  Op->getName() + ".sink", &BBEnd->front());
1892  for (auto *I : Insts)
1893  PN->addIncoming(I->getOperand(O), I->getParent());
1894  NewOperands.push_back(PN);
1895  }
1896 
1897  // Arbitrarily use I0 as the new "common" instruction; remap its operands
1898  // and move it to the start of the successor block.
1899  for (unsigned O = 0, E = I0->getNumOperands(); O != E; ++O)
1900  I0->getOperandUse(O).set(NewOperands[O]);
1901  I0->moveBefore(&*BBEnd->getFirstInsertionPt());
1902 
1903  // Update metadata and IR flags, and merge debug locations.
1904  for (auto *I : Insts)
1905  if (I != I0) {
1906  // The debug location for the "common" instruction is the merged locations
1907  // of all the commoned instructions. We start with the original location
1908  // of the "common" instruction and iteratively merge each location in the
1909  // loop below.
1910  // This is an N-way merge, which will be inefficient if I0 is a CallInst.
1911  // However, as N-way merge for CallInst is rare, so we use simplified API
1912  // instead of using complex API for N-way merge.
1913  I0->applyMergedLocation(I0->getDebugLoc(), I->getDebugLoc());
1914  combineMetadataForCSE(I0, I, true);
1915  I0->andIRFlags(I);
1916  }
1917 
1918  if (!I0->user_empty()) {
1919  // canSinkLastInstruction checked that all instructions were used by
1920  // one and only one PHI node. Find that now, RAUW it to our common
1921  // instruction and nuke it.
1922  auto *PN = cast<PHINode>(*I0->user_begin());
1923  PN->replaceAllUsesWith(I0);
1924  PN->eraseFromParent();
1925  }
1926 
1927  // Finally nuke all instructions apart from the common instruction.
1928  for (auto *I : Insts)
1929  if (I != I0)
1930  I->eraseFromParent();
1931 
1932  return true;
1933 }
1934 
1935 namespace {
1936 
1937  // LockstepReverseIterator - Iterates through instructions
1938  // in a set of blocks in reverse order from the first non-terminator.
1939  // For example (assume all blocks have size n):
1940  // LockstepReverseIterator I([B1, B2, B3]);
1941  // *I-- = [B1[n], B2[n], B3[n]];
1942  // *I-- = [B1[n-1], B2[n-1], B3[n-1]];
1943  // *I-- = [B1[n-2], B2[n-2], B3[n-2]];
1944  // ...
1945  class LockstepReverseIterator {
1946  ArrayRef<BasicBlock*> Blocks;
1948  bool Fail;
1949 
1950  public:
1951  LockstepReverseIterator(ArrayRef<BasicBlock*> Blocks) : Blocks(Blocks) {
1952  reset();
1953  }
1954 
1955  void reset() {
1956  Fail = false;
1957  Insts.clear();
1958  for (auto *BB : Blocks) {
1959  Instruction *Inst = BB->getTerminator();
1960  for (Inst = Inst->getPrevNode(); Inst && isa<DbgInfoIntrinsic>(Inst);)
1961  Inst = Inst->getPrevNode();
1962  if (!Inst) {
1963  // Block wasn't big enough.
1964  Fail = true;
1965  return;
1966  }
1967  Insts.push_back(Inst);
1968  }
1969  }
1970 
1971  bool isValid() const {
1972  return !Fail;
1973  }
1974 
1975  void operator--() {
1976  if (Fail)
1977  return;
1978  for (auto *&Inst : Insts) {
1979  for (Inst = Inst->getPrevNode(); Inst && isa<DbgInfoIntrinsic>(Inst);)
1980  Inst = Inst->getPrevNode();
1981  // Already at beginning of block.
1982  if (!Inst) {
1983  Fail = true;
1984  return;
1985  }
1986  }
1987  }
1988 
1989  void operator++() {
1990  if (Fail)
1991  return;
1992  for (auto *&Inst : Insts) {
1993  for (Inst = Inst->getNextNode(); Inst && isa<DbgInfoIntrinsic>(Inst);)
1994  Inst = Inst->getNextNode();
1995  // Already at end of block.
1996  if (!Inst) {
1997  Fail = true;
1998  return;
1999  }
2000  }
2001  }
2002 
2004  return Insts;
2005  }
2006  };
2007 
2008 } // end anonymous namespace
2009 
2010 /// Check whether BB's predecessors end with unconditional branches. If it is
2011 /// true, sink any common code from the predecessors to BB.
2013  DomTreeUpdater *DTU) {
2014  // We support two situations:
2015  // (1) all incoming arcs are unconditional
2016  // (2) there are non-unconditional incoming arcs
2017  //
2018  // (2) is very common in switch defaults and
2019  // else-if patterns;
2020  //
2021  // if (a) f(1);
2022  // else if (b) f(2);
2023  //
2024  // produces:
2025  //
2026  // [if]
2027  // / \
2028  // [f(1)] [if]
2029  // | | \
2030  // | | |
2031  // | [f(2)]|
2032  // \ | /
2033  // [ end ]
2034  //
2035  // [end] has two unconditional predecessor arcs and one conditional. The
2036  // conditional refers to the implicit empty 'else' arc. This conditional
2037  // arc can also be caused by an empty default block in a switch.
2038  //
2039  // In this case, we attempt to sink code from all *unconditional* arcs.
2040  // If we can sink instructions from these arcs (determined during the scan
2041  // phase below) we insert a common successor for all unconditional arcs and
2042  // connect that to [end], to enable sinking:
2043  //
2044  // [if]
2045  // / \
2046  // [x(1)] [if]
2047  // | | \
2048  // | | \
2049  // | [x(2)] |
2050  // \ / |
2051  // [sink.split] |
2052  // \ /
2053  // [ end ]
2054  //
2055  SmallVector<BasicBlock*,4> UnconditionalPreds;
2056  bool HaveNonUnconditionalPredecessors = false;
2057  for (auto *PredBB : predecessors(BB)) {
2058  auto *PredBr = dyn_cast<BranchInst>(PredBB->getTerminator());
2059  if (PredBr && PredBr->isUnconditional())
2060  UnconditionalPreds.push_back(PredBB);
2061  else
2062  HaveNonUnconditionalPredecessors = true;
2063  }
2064  if (UnconditionalPreds.size() < 2)
2065  return false;
2066 
2067  // We take a two-step approach to tail sinking. First we scan from the end of
2068  // each block upwards in lockstep. If the n'th instruction from the end of each
2069  // block can be sunk, those instructions are added to ValuesToSink and we
2070  // carry on. If we can sink an instruction but need to PHI-merge some operands
2071  // (because they're not identical in each instruction) we add these to
2072  // PHIOperands.
2073  int ScanIdx = 0;
2074  SmallPtrSet<Value*,4> InstructionsToSink;
2076  LockstepReverseIterator LRI(UnconditionalPreds);
2077  while (LRI.isValid() &&
2078  canSinkInstructions(*LRI, PHIOperands)) {
2079  LLVM_DEBUG(dbgs() << "SINK: instruction can be sunk: " << *(*LRI)[0]
2080  << "\n");
2081  InstructionsToSink.insert((*LRI).begin(), (*LRI).end());
2082  ++ScanIdx;
2083  --LRI;
2084  }
2085 
2086  // If no instructions can be sunk, early-return.
2087  if (ScanIdx == 0)
2088  return false;
2089 
2090  bool followedByDeoptOrUnreachable = IsBlockFollowedByDeoptOrUnreachable(BB);
2091 
2092  if (!followedByDeoptOrUnreachable) {
2093  // Okay, we *could* sink last ScanIdx instructions. But how many can we
2094  // actually sink before encountering instruction that is unprofitable to
2095  // sink?
2096  auto ProfitableToSinkInstruction = [&](LockstepReverseIterator &LRI) {
2097  unsigned NumPHIdValues = 0;
2098  for (auto *I : *LRI)
2099  for (auto *V : PHIOperands[I]) {
2100  if (!InstructionsToSink.contains(V))
2101  ++NumPHIdValues;
2102  // FIXME: this check is overly optimistic. We may end up not sinking
2103  // said instruction, due to the very same profitability check.
2104  // See @creating_too_many_phis in sink-common-code.ll.
2105  }
2106  LLVM_DEBUG(dbgs() << "SINK: #phid values: " << NumPHIdValues << "\n");
2107  unsigned NumPHIInsts = NumPHIdValues / UnconditionalPreds.size();
2108  if ((NumPHIdValues % UnconditionalPreds.size()) != 0)
2109  NumPHIInsts++;
2110 
2111  return NumPHIInsts <= 1;
2112  };
2113 
2114  // We've determined that we are going to sink last ScanIdx instructions,
2115  // and recorded them in InstructionsToSink. Now, some instructions may be
2116  // unprofitable to sink. But that determination depends on the instructions
2117  // that we are going to sink.
2118 
2119  // First, forward scan: find the first instruction unprofitable to sink,
2120  // recording all the ones that are profitable to sink.
2121  // FIXME: would it be better, after we detect that not all are profitable.
2122  // to either record the profitable ones, or erase the unprofitable ones?
2123  // Maybe we need to choose (at runtime) the one that will touch least
2124  // instrs?
2125  LRI.reset();
2126  int Idx = 0;
2127  SmallPtrSet<Value *, 4> InstructionsProfitableToSink;
2128  while (Idx < ScanIdx) {
2129  if (!ProfitableToSinkInstruction(LRI)) {
2130  // Too many PHIs would be created.
2131  LLVM_DEBUG(
2132  dbgs() << "SINK: stopping here, too many PHIs would be created!\n");
2133  break;
2134  }
2135  InstructionsProfitableToSink.insert((*LRI).begin(), (*LRI).end());
2136  --LRI;
2137  ++Idx;
2138  }
2139 
2140  // If no instructions can be sunk, early-return.
2141  if (Idx == 0)
2142  return false;
2143 
2144  // Did we determine that (only) some instructions are unprofitable to sink?
2145  if (Idx < ScanIdx) {
2146  // Okay, some instructions are unprofitable.
2147  ScanIdx = Idx;
2148  InstructionsToSink = InstructionsProfitableToSink;
2149 
2150  // But, that may make other instructions unprofitable, too.
2151  // So, do a backward scan, do any earlier instructions become
2152  // unprofitable?
2153  assert(
2154  !ProfitableToSinkInstruction(LRI) &&
2155  "We already know that the last instruction is unprofitable to sink");
2156  ++LRI;
2157  --Idx;
2158  while (Idx >= 0) {
2159  // If we detect that an instruction becomes unprofitable to sink,
2160  // all earlier instructions won't be sunk either,
2161  // so preemptively keep InstructionsProfitableToSink in sync.
2162  // FIXME: is this the most performant approach?
2163  for (auto *I : *LRI)
2164  InstructionsProfitableToSink.erase(I);
2165  if (!ProfitableToSinkInstruction(LRI)) {
2166  // Everything starting with this instruction won't be sunk.
2167  ScanIdx = Idx;
2168  InstructionsToSink = InstructionsProfitableToSink;
2169  }
2170  ++LRI;
2171  --Idx;
2172  }
2173  }
2174 
2175  // If no instructions can be sunk, early-return.
2176  if (ScanIdx == 0)
2177  return false;
2178  }
2179 
2180  bool Changed = false;
2181 
2182  if (HaveNonUnconditionalPredecessors) {
2183  if (!followedByDeoptOrUnreachable) {
2184  // It is always legal to sink common instructions from unconditional
2185  // predecessors. However, if not all predecessors are unconditional,
2186  // this transformation might be pessimizing. So as a rule of thumb,
2187  // don't do it unless we'd sink at least one non-speculatable instruction.
2188  // See https://bugs.llvm.org/show_bug.cgi?id=30244
2189  LRI.reset();
2190  int Idx = 0;
2191  bool Profitable = false;
2192  while (Idx < ScanIdx) {
2193  if (!isSafeToSpeculativelyExecute((*LRI)[0])) {
2194  Profitable = true;
2195  break;
2196  }
2197  --LRI;
2198  ++Idx;
2199  }
2200  if (!Profitable)
2201  return false;
2202  }
2203 
2204  LLVM_DEBUG(dbgs() << "SINK: Splitting edge\n");
2205  // We have a conditional edge and we're going to sink some instructions.
2206  // Insert a new block postdominating all blocks we're going to sink from.
2207  if (!SplitBlockPredecessors(BB, UnconditionalPreds, ".sink.split", DTU))
2208  // Edges couldn't be split.
2209  return false;
2210  Changed = true;
2211  }
2212 
2213  // Now that we've analyzed all potential sinking candidates, perform the
2214  // actual sink. We iteratively sink the last non-terminator of the source
2215  // blocks into their common successor unless doing so would require too
2216  // many PHI instructions to be generated (currently only one PHI is allowed
2217  // per sunk instruction).
2218  //
2219  // We can use InstructionsToSink to discount values needing PHI-merging that will
2220  // actually be sunk in a later iteration. This allows us to be more
2221  // aggressive in what we sink. This does allow a false positive where we
2222  // sink presuming a later value will also be sunk, but stop half way through
2223  // and never actually sink it which means we produce more PHIs than intended.
2224  // This is unlikely in practice though.
2225  int SinkIdx = 0;
2226  for (; SinkIdx != ScanIdx; ++SinkIdx) {
2227  LLVM_DEBUG(dbgs() << "SINK: Sink: "
2228  << *UnconditionalPreds[0]->getTerminator()->getPrevNode()
2229  << "\n");
2230 
2231  // Because we've sunk every instruction in turn, the current instruction to
2232  // sink is always at index 0.
2233  LRI.reset();
2234 
2235  if (!sinkLastInstruction(UnconditionalPreds)) {
2236  LLVM_DEBUG(
2237  dbgs()
2238  << "SINK: stopping here, failed to actually sink instruction!\n");
2239  break;
2240  }
2241 
2242  NumSinkCommonInstrs++;
2243  Changed = true;
2244  }
2245  if (SinkIdx != 0)
2246  ++NumSinkCommonCode;
2247  return Changed;
2248 }
2249 
2250 namespace {
2251 
2252 struct CompatibleSets {
2253  using SetTy = SmallVector<InvokeInst *, 2>;
2254 
2255  SmallVector<SetTy, 1> Sets;
2256 
2257  static bool shouldBelongToSameSet(ArrayRef<InvokeInst *> Invokes);
2258 
2259  SetTy &getCompatibleSet(InvokeInst *II);
2260 
2261  void insert(InvokeInst *II);
2262 };
2263 
2264 CompatibleSets::SetTy &CompatibleSets::getCompatibleSet(InvokeInst *II) {
2265  // Perform a linear scan over all the existing sets, see if the new `invoke`
2266  // is compatible with any particular set. Since we know that all the `invokes`
2267  // within a set are compatible, only check the first `invoke` in each set.
2268  // WARNING: at worst, this has quadratic complexity.
2269  for (CompatibleSets::SetTy &Set : Sets) {
2270  if (CompatibleSets::shouldBelongToSameSet({Set.front(), II}))
2271  return Set;
2272  }
2273 
2274  // Otherwise, we either had no sets yet, or this invoke forms a new set.
2275  return Sets.emplace_back();
2276 }
2277 
2278 void CompatibleSets::insert(InvokeInst *II) {
2279  getCompatibleSet(II).emplace_back(II);
2280 }
2281 
2282 bool CompatibleSets::shouldBelongToSameSet(ArrayRef<InvokeInst *> Invokes) {
2283  assert(Invokes.size() == 2 && "Always called with exactly two candidates.");
2284 
2285  // Can we theoretically merge these `invoke`s?
2286  auto IsIllegalToMerge = [](InvokeInst *II) {
2287  return II->cannotMerge() || II->isInlineAsm();
2288  };
2289  if (any_of(Invokes, IsIllegalToMerge))
2290  return false;
2291 
2292  // Either both `invoke`s must be direct,
2293  // or both `invoke`s must be indirect.
2294  auto IsIndirectCall = [](InvokeInst *II) { return II->isIndirectCall(); };
2295  bool HaveIndirectCalls = any_of(Invokes, IsIndirectCall);
2296  bool AllCallsAreIndirect = all_of(Invokes, IsIndirectCall);
2297  if (HaveIndirectCalls) {
2298  if (!AllCallsAreIndirect)
2299  return false;
2300  } else {
2301  // All callees must be identical.
2302  Value *Callee = nullptr;
2303  for (InvokeInst *II : Invokes) {
2304  Value *CurrCallee = II->getCalledOperand();
2305  assert(CurrCallee && "There is always a called operand.");
2306  if (!Callee)
2307  Callee = CurrCallee;
2308  else if (Callee != CurrCallee)
2309  return false;
2310  }
2311  }
2312 
2313  // Either both `invoke`s must not have a normal destination,
2314  // or both `invoke`s must have a normal destination,
2315  auto HasNormalDest = [](InvokeInst *II) {
2316  return !isa<UnreachableInst>(II->getNormalDest()->getFirstNonPHIOrDbg());
2317  };
2318  if (any_of(Invokes, HasNormalDest)) {
2319  // Do not merge `invoke` that does not have a normal destination with one
2320  // that does have a normal destination, even though doing so would be legal.
2321  if (!all_of(Invokes, HasNormalDest))
2322  return false;
2323 
2324  // All normal destinations must be identical.
2325  BasicBlock *NormalBB = nullptr;
2326  for (InvokeInst *II : Invokes) {
2327  BasicBlock *CurrNormalBB = II->getNormalDest();
2328  assert(CurrNormalBB && "There is always a 'continue to' basic block.");
2329  if (!NormalBB)
2330  NormalBB = CurrNormalBB;
2331  else if (NormalBB != CurrNormalBB)
2332  return false;
2333  }
2334 
2335  // In the normal destination, the incoming values for these two `invoke`s
2336  // must be compatible.
2337  SmallPtrSet<Value *, 16> EquivalenceSet(Invokes.begin(), Invokes.end());
2339  NormalBB, {Invokes[0]->getParent(), Invokes[1]->getParent()},
2340  &EquivalenceSet))
2341  return false;
2342  }
2343 
2344 #ifndef NDEBUG
2345  // All unwind destinations must be identical.
2346  // We know that because we have started from said unwind destination.
2347  BasicBlock *UnwindBB = nullptr;
2348  for (InvokeInst *II : Invokes) {
2349  BasicBlock *CurrUnwindBB = II->getUnwindDest();
2350  assert(CurrUnwindBB && "There is always an 'unwind to' basic block.");
2351  if (!UnwindBB)
2352  UnwindBB = CurrUnwindBB;
2353  else
2354  assert(UnwindBB == CurrUnwindBB && "Unexpected unwind destination.");
2355  }
2356 #endif
2357 
2358  // In the unwind destination, the incoming values for these two `invoke`s
2359  // must be compatible.
2361  Invokes.front()->getUnwindDest(),
2362  {Invokes[0]->getParent(), Invokes[1]->getParent()}))
2363  return false;
2364 
2365  // Ignoring arguments, these `invoke`s must be identical,
2366  // including operand bundles.
2367  const InvokeInst *II0 = Invokes.front();
2368  for (auto *II : Invokes.drop_front())
2369  if (!II->isSameOperationAs(II0))
2370  return false;
2371 
2372  // Can we theoretically form the data operands for the merged `invoke`?
2373  auto IsIllegalToMergeArguments = [](auto Ops) {
2374  Type *Ty = std::get<0>(Ops)->getType();
2375  assert(Ty == std::get<1>(Ops)->getType() && "Incompatible types?");
2376  return Ty->isTokenTy() && std::get<0>(Ops) != std::get<1>(Ops);
2377  };
2378  assert(Invokes.size() == 2 && "Always called with exactly two candidates.");
2379  if (any_of(zip(Invokes[0]->data_ops(), Invokes[1]->data_ops()),
2380  IsIllegalToMergeArguments))
2381  return false;
2382 
2383  return true;
2384 }
2385 
2386 } // namespace
2387 
2388 // Merge all invokes in the provided set, all of which are compatible
2389 // as per the `CompatibleSets::shouldBelongToSameSet()`.
2391  DomTreeUpdater *DTU) {
2392  assert(Invokes.size() >= 2 && "Must have at least two invokes to merge.");
2393 
2395  if (DTU)
2396  Updates.reserve(2 + 3 * Invokes.size());
2397 
2398  bool HasNormalDest =
2399  !isa<UnreachableInst>(Invokes[0]->getNormalDest()->getFirstNonPHIOrDbg());
2400 
2401  // Clone one of the invokes into a new basic block.
2402  // Since they are all compatible, it doesn't matter which invoke is cloned.
2403  InvokeInst *MergedInvoke = [&Invokes, HasNormalDest]() {
2404  InvokeInst *II0 = Invokes.front();
2405  BasicBlock *II0BB = II0->getParent();
2406  BasicBlock *InsertBeforeBlock =
2407  II0->getParent()->getIterator()->getNextNode();
2408  Function *Func = II0BB->getParent();
2409  LLVMContext &Ctx = II0->getContext();
2410 
2411  BasicBlock *MergedInvokeBB = BasicBlock::Create(
2412  Ctx, II0BB->getName() + ".invoke", Func, InsertBeforeBlock);
2413 
2414  auto *MergedInvoke = cast<InvokeInst>(II0->clone());
2415  // NOTE: all invokes have the same attributes, so no handling needed.
2416  MergedInvokeBB->getInstList().push_back(MergedInvoke);
2417 
2418  if (!HasNormalDest) {
2419  // This set does not have a normal destination,
2420  // so just form a new block with unreachable terminator.
2421  BasicBlock *MergedNormalDest = BasicBlock::Create(
2422  Ctx, II0BB->getName() + ".cont", Func, InsertBeforeBlock);
2423  new UnreachableInst(Ctx, MergedNormalDest);
2424  MergedInvoke->setNormalDest(MergedNormalDest);
2425  }
2426 
2427  // The unwind destination, however, remainds identical for all invokes here.
2428 
2429  return MergedInvoke;
2430  }();
2431 
2432  if (DTU) {
2433  // Predecessor blocks that contained these invokes will now branch to
2434  // the new block that contains the merged invoke, ...
2435  for (InvokeInst *II : Invokes)
2436  Updates.push_back(
2437  {DominatorTree::Insert, II->getParent(), MergedInvoke->getParent()});
2438 
2439  // ... which has the new `unreachable` block as normal destination,
2440  // or unwinds to the (same for all `invoke`s in this set) `landingpad`,
2441  for (BasicBlock *SuccBBOfMergedInvoke : successors(MergedInvoke))
2442  Updates.push_back({DominatorTree::Insert, MergedInvoke->getParent(),
2443  SuccBBOfMergedInvoke});
2444 
2445  // Since predecessor blocks now unconditionally branch to a new block,
2446  // they no longer branch to their original successors.
2447  for (InvokeInst *II : Invokes)
2448  for (BasicBlock *SuccOfPredBB : successors(II->getParent()))
2449  Updates.push_back(
2450  {DominatorTree::Delete, II->getParent(), SuccOfPredBB});
2451  }
2452 
2453  bool IsIndirectCall = Invokes[0]->isIndirectCall();
2454 
2455  // Form the merged operands for the merged invoke.
2456  for (Use &U : MergedInvoke->operands()) {
2457  // Only PHI together the indirect callees and data operands.
2458  if (MergedInvoke->isCallee(&U)) {
2459  if (!IsIndirectCall)
2460  continue;
2461  } else if (!MergedInvoke->isDataOperand(&U))
2462  continue;
2463 
2464  // Don't create trivial PHI's with all-identical incoming values.
2465  bool NeedPHI = any_of(Invokes, [&U](InvokeInst *II) {
2466  return II->getOperand(U.getOperandNo()) != U.get();
2467  });
2468  if (!NeedPHI)
2469  continue;
2470 
2471  // Form a PHI out of all the data ops under this index.
2472  PHINode *PN = PHINode::Create(
2473  U->getType(), /*NumReservedValues=*/Invokes.size(), "", MergedInvoke);
2474  for (InvokeInst *II : Invokes)
2475  PN->addIncoming(II->getOperand(U.getOperandNo()), II->getParent());
2476 
2477  U.set(PN);
2478  }
2479 
2480  // We've ensured that each PHI node has compatible (identical) incoming values
2481  // when coming from each of the `invoke`s in the current merge set,
2482  // so update the PHI nodes accordingly.
2483  for (BasicBlock *Succ : successors(MergedInvoke))
2484  AddPredecessorToBlock(Succ, /*NewPred=*/MergedInvoke->getParent(),
2485  /*ExistPred=*/Invokes.front()->getParent());
2486 
2487  // And finally, replace the original `invoke`s with an unconditional branch
2488  // to the block with the merged `invoke`. Also, give that merged `invoke`
2489  // the merged debugloc of all the original `invoke`s.
2490  const DILocation *MergedDebugLoc = nullptr;
2491  for (InvokeInst *II : Invokes) {
2492  // Compute the debug location common to all the original `invoke`s.
2493  if (!MergedDebugLoc)
2494  MergedDebugLoc = II->getDebugLoc();
2495  else
2496  MergedDebugLoc =
2497  DILocation::getMergedLocation(MergedDebugLoc, II->getDebugLoc());
2498 
2499  // And replace the old `invoke` with an unconditionally branch
2500  // to the block with the merged `invoke`.
2501  for (BasicBlock *OrigSuccBB : successors(II->getParent()))
2502  OrigSuccBB->removePredecessor(II->getParent());
2503  BranchInst::Create(MergedInvoke->getParent(), II->getParent());
2504  II->replaceAllUsesWith(MergedInvoke);
2505  II->eraseFromParent();
2506  ++NumInvokesMerged;
2507  }
2508  MergedInvoke->setDebugLoc(MergedDebugLoc);
2509  ++NumInvokeSetsFormed;
2510 
2511  if (DTU)
2512  DTU->applyUpdates(Updates);
2513 }
2514 
2515 /// If this block is a `landingpad` exception handling block, categorize all
2516 /// the predecessor `invoke`s into sets, with all `invoke`s in each set
2517 /// being "mergeable" together, and then merge invokes in each set together.
2518 ///
2519 /// This is a weird mix of hoisting and sinking. Visually, it goes from:
2520 /// [...] [...]
2521 /// | |
2522 /// [invoke0] [invoke1]
2523 /// / \ / \
2524 /// [cont0] [landingpad] [cont1]
2525 /// to:
2526 /// [...] [...]
2527 /// \ /
2528 /// [invoke]
2529 /// / \
2530 /// [cont] [landingpad]
2531 ///
2532 /// But of course we can only do that if the invokes share the `landingpad`,
2533 /// edges invoke0->cont0 and invoke1->cont1 are "compatible",
2534 /// and the invoked functions are "compatible".
2537  return false;
2538 
2539  bool Changed = false;
2540 
2541  // FIXME: generalize to all exception handling blocks?
2542  if (!BB->isLandingPad())
2543  return Changed;
2544 
2545  CompatibleSets Grouper;
2546 
2547  // Record all the predecessors of this `landingpad`. As per verifier,
2548  // the only allowed predecessor is the unwind edge of an `invoke`.
2549  // We want to group "compatible" `invokes` into the same set to be merged.
2550  for (BasicBlock *PredBB : predecessors(BB))
2551  Grouper.insert(cast<InvokeInst>(PredBB->getTerminator()));
2552 
2553  // And now, merge `invoke`s that were grouped togeter.
2554  for (ArrayRef<InvokeInst *> Invokes : Grouper.Sets) {
2555  if (Invokes.size() < 2)
2556  continue;
2557  Changed = true;
2558  MergeCompatibleInvokesImpl(Invokes, DTU);
2559  }
2560 
2561  return Changed;
2562 }
2563 
2564 /// Determine if we can hoist sink a sole store instruction out of a
2565 /// conditional block.
2566 ///
2567 /// We are looking for code like the following:
2568 /// BrBB:
2569 /// store i32 %add, i32* %arrayidx2
2570 /// ... // No other stores or function calls (we could be calling a memory
2571 /// ... // function).
2572 /// %cmp = icmp ult %x, %y
2573 /// br i1 %cmp, label %EndBB, label %ThenBB
2574 /// ThenBB:
2575 /// store i32 %add5, i32* %arrayidx2
2576 /// br label EndBB
2577 /// EndBB:
2578 /// ...
2579 /// We are going to transform this into:
2580 /// BrBB:
2581 /// store i32 %add, i32* %arrayidx2
2582 /// ... //
2583 /// %cmp = icmp ult %x, %y
2584 /// %add.add5 = select i1 %cmp, i32 %add, %add5
2585 /// store i32 %add.add5, i32* %arrayidx2
2586 /// ...
2587 ///
2588 /// \return The pointer to the value of the previous store if the store can be
2589 /// hoisted into the predecessor block. 0 otherwise.
2591  BasicBlock *StoreBB, BasicBlock *EndBB) {
2592  StoreInst *StoreToHoist = dyn_cast<StoreInst>(I);
2593  if (!StoreToHoist)
2594  return nullptr;
2595 
2596  // Volatile or atomic.
2597  if (!StoreToHoist->isSimple())
2598  return nullptr;
2599 
2600  Value *StorePtr = StoreToHoist->getPointerOperand();
2601  Type *StoreTy = StoreToHoist->getValueOperand()->getType();
2602 
2603  // Look for a store to the same pointer in BrBB.
2604  unsigned MaxNumInstToLookAt = 9;
2605  // Skip pseudo probe intrinsic calls which are not really killing any memory
2606  // accesses.
2607  for (Instruction &CurI : reverse(BrBB->instructionsWithoutDebug(true))) {
2608  if (!MaxNumInstToLookAt)
2609  break;
2610  --MaxNumInstToLookAt;
2611 
2612  // Could be calling an instruction that affects memory like free().
2613  if (CurI.mayWriteToMemory() && !isa<StoreInst>(CurI))
2614  return nullptr;
2615 
2616  if (auto *SI = dyn_cast<StoreInst>(&CurI)) {
2617  // Found the previous store to same location and type. Make sure it is
2618  // simple, to avoid introducing a spurious non-atomic write after an
2619  // atomic write.
2620  if (SI->getPointerOperand() == StorePtr &&
2621  SI->getValueOperand()->getType() == StoreTy && SI->isSimple())
2622  // Found the previous store, return its value operand.
2623  return SI->getValueOperand();
2624  return nullptr; // Unknown store.
2625  }
2626 
2627  if (auto *LI = dyn_cast<LoadInst>(&CurI)) {
2628  if (LI->getPointerOperand() == StorePtr && LI->getType() == StoreTy &&
2629  LI->isSimple()) {
2630  // Local objects (created by an `alloca` instruction) are always
2631  // writable, so once we are past a read from a location it is valid to
2632  // also write to that same location.
2633  // If the address of the local object never escapes the function, that
2634  // means it's never concurrently read or written, hence moving the store
2635  // from under the condition will not introduce a data race.
2636  auto *AI = dyn_cast<AllocaInst>(getUnderlyingObject(StorePtr));
2637  if (AI && !PointerMayBeCaptured(AI, false, true))
2638  // Found a previous load, return it.
2639  return LI;
2640  }
2641  // The load didn't work out, but we may still find a store.
2642  }
2643  }
2644 
2645  return nullptr;
2646 }
2647 
2648 /// Estimate the cost of the insertion(s) and check that the PHI nodes can be
2649 /// converted to selects.
2651  BasicBlock *EndBB,
2652  unsigned &SpeculatedInstructions,
2653  InstructionCost &Cost,
2654  const TargetTransformInfo &TTI) {
2656  BB->getParent()->hasMinSize()
2657  ? TargetTransformInfo::TCK_CodeSize
2658  : TargetTransformInfo::TCK_SizeAndLatency;
2659 
2660  bool HaveRewritablePHIs = false;
2661  for (PHINode &PN : EndBB->phis()) {
2662  Value *OrigV = PN.getIncomingValueForBlock(BB);
2663  Value *ThenV = PN.getIncomingValueForBlock(ThenBB);
2664 
2665  // FIXME: Try to remove some of the duplication with HoistThenElseCodeToIf.
2666  // Skip PHIs which are trivial.
2667  if (ThenV == OrigV)
2668  continue;
2669 
2670  Cost += TTI.getCmpSelInstrCost(Instruction::Select, PN.getType(), nullptr,
2671  CmpInst::BAD_ICMP_PREDICATE, CostKind);
2672 
2673  // Don't convert to selects if we could remove undefined behavior instead.
2674  if (passingValueIsAlwaysUndefined(OrigV, &PN) ||
2675  passingValueIsAlwaysUndefined(ThenV, &PN))
2676  return false;
2677 
2678  HaveRewritablePHIs = true;
2679  ConstantExpr *OrigCE = dyn_cast<ConstantExpr>(OrigV);
2680  ConstantExpr *ThenCE = dyn_cast<ConstantExpr>(ThenV);
2681  if (!OrigCE && !ThenCE)
2682  continue; // Known safe and cheap.
2683 
2684  if ((ThenCE && !isSafeToSpeculativelyExecute(ThenCE)) ||
2685  (OrigCE && !isSafeToSpeculativelyExecute(OrigCE)))
2686  return false;
2687  InstructionCost OrigCost = OrigCE ? computeSpeculationCost(OrigCE, TTI) : 0;
2688  InstructionCost ThenCost = ThenCE ? computeSpeculationCost(ThenCE, TTI) : 0;
2689  InstructionCost MaxCost =
2690  2 * PHINodeFoldingThreshold * TargetTransformInfo::TCC_Basic;
2691  if (OrigCost + ThenCost > MaxCost)
2692  return false;
2693 
2694  // Account for the cost of an unfolded ConstantExpr which could end up
2695  // getting expanded into Instructions.
2696  // FIXME: This doesn't account for how many operations are combined in the
2697  // constant expression.
2698  ++SpeculatedInstructions;
2699  if (SpeculatedInstructions > 1)
2700  return false;
2701  }
2702 
2703  return HaveRewritablePHIs;
2704 }
2705 
2706 /// Speculate a conditional basic block flattening the CFG.
2707 ///
2708 /// Note that this is a very risky transform currently. Speculating
2709 /// instructions like this is most often not desirable. Instead, there is an MI
2710 /// pass which can do it with full awareness of the resource constraints.
2711 /// However, some cases are "obvious" and we should do directly. An example of
2712 /// this is speculating a single, reasonably cheap instruction.
2713 ///
2714 /// There is only one distinct advantage to flattening the CFG at the IR level:
2715 /// it makes very common but simplistic optimizations such as are common in
2716 /// instcombine and the DAG combiner more powerful by removing CFG edges and
2717 /// modeling their effects with easier to reason about SSA value graphs.
2718 ///
2719 ///
2720 /// An illustration of this transform is turning this IR:
2721 /// \code
2722 /// BB:
2723 /// %cmp = icmp ult %x, %y
2724 /// br i1 %cmp, label %EndBB, label %ThenBB
2725 /// ThenBB:
2726 /// %sub = sub %x, %y
2727 /// br label BB2
2728 /// EndBB:
2729 /// %phi = phi [ %sub, %ThenBB ], [ 0, %EndBB ]
2730 /// ...
2731 /// \endcode
2732 ///
2733 /// Into this IR:
2734 /// \code
2735 /// BB:
2736 /// %cmp = icmp ult %x, %y
2737 /// %sub = sub %x, %y
2738 /// %cond = select i1 %cmp, 0, %sub
2739 /// ...
2740 /// \endcode
2741 ///
2742 /// \returns true if the conditional block is removed.
2743 bool SimplifyCFGOpt::SpeculativelyExecuteBB(BranchInst *BI, BasicBlock *ThenBB,
2744  const TargetTransformInfo &TTI) {
2745  // Be conservative for now. FP select instruction can often be expensive.
2746  Value *BrCond = BI->getCondition();
2747  if (isa<FCmpInst>(BrCond))
2748  return false;
2749 
2750  BasicBlock *BB = BI->getParent();
2751  BasicBlock *EndBB = ThenBB->getTerminator()->getSuccessor(0);
2752  InstructionCost Budget =
2753  PHINodeFoldingThreshold * TargetTransformInfo::TCC_Basic;
2754 
2755  // If ThenBB is actually on the false edge of the conditional branch, remember
2756  // to swap the select operands later.
2757  bool Invert = false;
2758  if (ThenBB != BI->getSuccessor(0)) {
2759  assert(ThenBB == BI->getSuccessor(1) && "No edge from 'if' block?");
2760  Invert = true;
2761  }
2762  assert(EndBB == BI->getSuccessor(!Invert) && "No edge from to end block");
2763 
2764  // If the branch is non-unpredictable, and is predicted to *not* branch to
2765  // the `then` block, then avoid speculating it.
2766  if (!BI->getMetadata(LLVMContext::MD_unpredictable)) {
2767  uint64_t TWeight, FWeight;
2768  if (BI->extractProfMetadata(TWeight, FWeight) && (TWeight + FWeight) != 0) {
2769  uint64_t EndWeight = Invert ? TWeight : FWeight;
2770  BranchProbability BIEndProb =
2771  BranchProbability::getBranchProbability(EndWeight, TWeight + FWeight);
2773  if (BIEndProb >= Likely)
2774  return false;
2775  }
2776  }
2777 
2778  // Keep a count of how many times instructions are used within ThenBB when
2779  // they are candidates for sinking into ThenBB. Specifically:
2780  // - They are defined in BB, and
2781  // - They have no side effects, and
2782  // - All of their uses are in ThenBB.
2783  SmallDenseMap<Instruction *, unsigned, 4> SinkCandidateUseCounts;
2784 
2785  SmallVector<Instruction *, 4> SpeculatedDbgIntrinsics;
2786 
2787  unsigned SpeculatedInstructions = 0;
2788  Value *SpeculatedStoreValue = nullptr;
2789  StoreInst *SpeculatedStore = nullptr;
2790  for (BasicBlock::iterator BBI = ThenBB->begin(),
2791  BBE = std::prev(ThenBB->end());
2792  BBI != BBE; ++BBI) {
2793  Instruction *I = &*BBI;
2794  // Skip debug info.
2795  if (isa<DbgInfoIntrinsic>(I)) {
2796  SpeculatedDbgIntrinsics.push_back(I);
2797  continue;
2798  }
2799 
2800  // Skip pseudo probes. The consequence is we lose track of the branch
2801  // probability for ThenBB, which is fine since the optimization here takes
2802  // place regardless of the branch probability.
2803  if (isa<PseudoProbeInst>(I)) {
2804  // The probe should be deleted so that it will not be over-counted when
2805  // the samples collected on the non-conditional path are counted towards
2806  // the conditional path. We leave it for the counts inference algorithm to
2807  // figure out a proper count for an unknown probe.
2808  SpeculatedDbgIntrinsics.push_back(I);
2809  continue;
2810  }
2811 
2812  // Only speculatively execute a single instruction (not counting the
2813  // terminator) for now.
2814  ++SpeculatedInstructions;
2815  if (SpeculatedInstructions > 1)
2816  return false;
2817 
2818  // Don't hoist the instruction if it's unsafe or expensive.
2820  !(HoistCondStores && (SpeculatedStoreValue = isSafeToSpeculateStore(
2821  I, BB, ThenBB, EndBB))))
2822  return false;
2823  if (!SpeculatedStoreValue &&
2825  PHINodeFoldingThreshold * TargetTransformInfo::TCC_Basic)
2826  return false;
2827 
2828  // Store the store speculation candidate.
2829  if (SpeculatedStoreValue)
2830  SpeculatedStore = cast<StoreInst>(I);
2831 
2832  // Do not hoist the instruction if any of its operands are defined but not
2833  // used in BB. The transformation will prevent the operand from
2834  // being sunk into the use block.
2835  for (Use &Op : I->operands()) {
2836  Instruction *OpI = dyn_cast<Instruction>(Op);
2837  if (!OpI || OpI->getParent() != BB || OpI->mayHaveSideEffects())
2838  continue; // Not a candidate for sinking.
2839 
2840  ++SinkCandidateUseCounts[OpI];
2841  }
2842  }
2843 
2844  // Consider any sink candidates which are only used in ThenBB as costs for
2845  // speculation. Note, while we iterate over a DenseMap here, we are summing
2846  // and so iteration order isn't significant.
2848  I = SinkCandidateUseCounts.begin(),
2849  E = SinkCandidateUseCounts.end();
2850  I != E; ++I)
2851  if (I->first->hasNUses(I->second)) {
2852  ++SpeculatedInstructions;
2853  if (SpeculatedInstructions > 1)
2854  return false;
2855  }
2856 
2857  // Check that we can insert the selects and that it's not too expensive to do
2858  // so.
2859  bool Convert = SpeculatedStore != nullptr;
2860  InstructionCost Cost = 0;
2861  Convert |= validateAndCostRequiredSelects(BB, ThenBB, EndBB,
2862  SpeculatedInstructions,
2863  Cost, TTI);
2864  if (!Convert || Cost > Budget)
2865  return false;
2866 
2867  // If we get here, we can hoist the instruction and if-convert.
2868  LLVM_DEBUG(dbgs() << "SPECULATIVELY EXECUTING BB" << *ThenBB << "\n";);
2869 
2870  // Insert a select of the value of the speculated store.
2871  if (SpeculatedStoreValue) {
2873  Value *TrueV = SpeculatedStore->getValueOperand();
2874  Value *FalseV = SpeculatedStoreValue;
2875  if (Invert)
2876  std::swap(TrueV, FalseV);
2877  Value *S = Builder.CreateSelect(
2878  BrCond, TrueV, FalseV, "spec.store.select", BI);
2879  SpeculatedStore->setOperand(0, S);
2880  SpeculatedStore->applyMergedLocation(BI->getDebugLoc(),
2881  SpeculatedStore->getDebugLoc());
2882  }
2883 
2884  // Metadata can be dependent on the condition we are hoisting above.
2885  // Conservatively strip all metadata on the instruction. Drop the debug loc
2886  // to avoid making it appear as if the condition is a constant, which would
2887  // be misleading while debugging.
2888  // Similarly strip attributes that maybe dependent on condition we are
2889  // hoisting above.
2890  for (auto &I : *ThenBB) {
2891  if (!SpeculatedStoreValue || &I != SpeculatedStore)
2892  I.setDebugLoc(DebugLoc());
2893  I.dropUndefImplyingAttrsAndUnknownMetadata();
2894  }
2895 
2896  // Hoist the instructions.
2897  BB->getInstList().splice(BI->getIterator(), ThenBB->getInstList(),
2898  ThenBB->begin(), std::prev(ThenBB->end()));
2899 
2900  // Insert selects and rewrite the PHI operands.
2902  for (PHINode &PN : EndBB->phis()) {
2903  unsigned OrigI = PN.getBasicBlockIndex(BB);
2904  unsigned ThenI = PN.getBasicBlockIndex(ThenBB);
2905  Value *OrigV = PN.getIncomingValue(OrigI);
2906  Value *ThenV = PN.getIncomingValue(ThenI);
2907 
2908  // Skip PHIs which are trivial.
2909  if (OrigV == ThenV)
2910  continue;
2911 
2912  // Create a select whose true value is the speculatively executed value and
2913  // false value is the pre-existing value. Swap them if the branch
2914  // destinations were inverted.
2915  Value *TrueV = ThenV, *FalseV = OrigV;
2916  if (Invert)
2917  std::swap(TrueV, FalseV);
2918  Value *V = Builder.CreateSelect(BrCond, TrueV, FalseV, "spec.select", BI);
2919  PN.setIncomingValue(OrigI, V);
2920  PN.setIncomingValue(ThenI, V);
2921  }
2922 
2923  // Remove speculated dbg intrinsics.
2924  // FIXME: Is it possible to do this in a more elegant way? Moving/merging the
2925  // dbg value for the different flows and inserting it after the select.
2926  for (Instruction *I : SpeculatedDbgIntrinsics)
2927  I->eraseFromParent();
2928 
2929  ++NumSpeculations;
2930  return true;
2931 }
2932 
2933 /// Return true if we can thread a branch across this block.
2935  int Size = 0;
2936 
2938  auto IsEphemeral = [&](const Instruction *I) {
2939  if (isa<AssumeInst>(I))
2940  return true;
2941  return !I->mayHaveSideEffects() && !I->isTerminator() &&
2942  all_of(I->users(),
2943  [&](const User *U) { return EphValues.count(U); });
2944  };
2945 
2946  // Walk the loop in reverse so that we can identify ephemeral values properly
2947  // (values only feeding assumes).
2948  for (Instruction &I : reverse(BB->instructionsWithoutDebug(false))) {
2949  // Can't fold blocks that contain noduplicate or convergent calls.
2950  if (CallInst *CI = dyn_cast<CallInst>(&I))
2951  if (CI->cannotDuplicate() || CI->isConvergent())
2952  return false;
2953 
2954  // Ignore ephemeral values which are deleted during codegen.
2955  if (IsEphemeral(&I))
2956  EphValues.insert(&I);
2957  // We will delete Phis while threading, so Phis should not be accounted in
2958  // block's size.
2959  else if (!isa<PHINode>(I)) {
2960  if (Size++ > MaxSmallBlockSize)
2961  return false; // Don't clone large BB's.
2962  }
2963 
2964  // We can only support instructions that do not define values that are
2965  // live outside of the current basic block.
2966  for (User *U : I.users()) {
2967  Instruction *UI = cast<Instruction>(U);
2968  if (UI->getParent() != BB || isa<PHINode>(UI))
2969  return false;
2970  }
2971 
2972  // Looks ok, continue checking.
2973  }
2974 
2975  return true;
2976 }
2977 
2978 static ConstantInt *
2980  SmallDenseMap<std::pair<BasicBlock *, BasicBlock *>,
2981  ConstantInt *> &Visited) {
2982  // Don't look past the block defining the value, we might get the value from
2983  // a previous loop iteration.
2984  auto *I = dyn_cast<Instruction>(V);
2985  if (I && I->getParent() == To)
2986  return nullptr;
2987 
2988  // We know the value if the From block branches on it.
2989  auto *BI = dyn_cast<BranchInst>(From->getTerminator());
2990  if (BI && BI->isConditional() && BI->getCondition() == V &&
2991  BI->getSuccessor(0) != BI->getSuccessor(1))
2992  return BI->getSuccessor(0) == To ? ConstantInt::getTrue(BI->getContext())
2994 
2995  // Limit the amount of blocks we inspect.
2996  if (Visited.size() >= 8)
2997  return nullptr;
2998 
2999  auto Pair = Visited.try_emplace({From, To}, nullptr);
3000  if (!Pair.second)
3001  return Pair.first->second;
3002 
3003  // Check whether the known value is the same for all predecessors.
3004  ConstantInt *Common = nullptr;
3005  for (BasicBlock *Pred : predecessors(From)) {
3006  ConstantInt *C = getKnownValueOnEdge(V, Pred, From, Visited);
3007  if (!C || (Common && Common != C))
3008  return nullptr;
3009  Common = C;
3010  }
3011  return Visited[{From, To}] = Common;
3012 }
3013 
3014 /// If we have a conditional branch on something for which we know the constant
3015 /// value in predecessors (e.g. a phi node in the current block), thread edges
3016 /// from the predecessor to their ultimate destination.
3017 static Optional<bool>
3019  const DataLayout &DL,
3020  AssumptionCache *AC) {
3022  BasicBlock *BB = BI->getParent();
3023  Value *Cond = BI->getCondition();
3024  PHINode *PN = dyn_cast<PHINode>(Cond);
3025  if (PN && PN->getParent() == BB) {
3026  // Degenerate case of a single entry PHI.
3027  if (PN->getNumIncomingValues() == 1) {
3029  return true;
3030  }
3031 
3032  for (Use &U : PN->incoming_values())
3033  if (auto *CB = dyn_cast<ConstantInt>(U))
3034  KnownValues.insert({PN->getIncomingBlock(U), CB});
3035  } else {
3037  for (BasicBlock *Pred : predecessors(BB)) {
3038  if (ConstantInt *CB = getKnownValueOnEdge(Cond, Pred, BB, Visited))
3039  KnownValues.insert({Pred, CB});
3040  }
3041  }
3042 
3043  if (KnownValues.empty())
3044  return false;
3045 
3046  // Now we know that this block has multiple preds and two succs.
3047  // Check that the block is small enough and values defined in the block are
3048  // not used outside of it.
3050  return false;
3051 
3052  for (const auto &Pair : KnownValues) {
3053  // Okay, we now know that all edges from PredBB should be revectored to
3054  // branch to RealDest.
3055  ConstantInt *CB = Pair.second;
3056  BasicBlock *PredBB = Pair.first;
3057  BasicBlock *RealDest = BI->getSuccessor(!CB->getZExtValue());
3058 
3059  if (RealDest == BB)
3060  continue; // Skip self loops.
3061  // Skip if the predecessor's terminator is an indirect branch.
3062  if (isa<IndirectBrInst>(PredBB->getTerminator()))
3063  continue;
3064 
3066 
3067  // The dest block might have PHI nodes, other predecessors and other
3068  // difficult cases. Instead of being smart about this, just insert a new
3069  // block that jumps to the destination block, effectively splitting
3070  // the edge we are about to create.
3071  BasicBlock *EdgeBB =
3072  BasicBlock::Create(BB->getContext(), RealDest->getName() + ".critedge",
3073  RealDest->getParent(), RealDest);
3074  BranchInst *CritEdgeBranch = BranchInst::Create(RealDest, EdgeBB);
3075  if (DTU)
3076  Updates.push_back({DominatorTree::Insert, EdgeBB, RealDest});
3077  CritEdgeBranch->setDebugLoc(BI->getDebugLoc());
3078 
3079  // Update PHI nodes.
3080  AddPredecessorToBlock(RealDest, EdgeBB, BB);
3081 
3082  // BB may have instructions that are being threaded over. Clone these
3083  // instructions into EdgeBB. We know that there will be no uses of the
3084  // cloned instructions outside of EdgeBB.
3085  BasicBlock::iterator InsertPt = EdgeBB->begin();
3086  DenseMap<Value *, Value *> TranslateMap; // Track translated values.
3087  TranslateMap[Cond] = Pair.second;
3088  for (BasicBlock::iterator BBI = BB->begin(); &*BBI != BI; ++BBI) {
3089  if (PHINode *PN = dyn_cast<PHINode>(BBI)) {
3090  TranslateMap[PN] = PN->getIncomingValueForBlock(PredBB);
3091  continue;
3092  }
3093  // Clone the instruction.
3094  Instruction *N = BBI->clone();
3095  if (BBI->hasName())
3096  N->setName(BBI->getName() + ".c");
3097 
3098  // Update operands due to translation.
3099  for (Use &Op : N->operands()) {
3100  DenseMap<Value *, Value *>::iterator PI = TranslateMap.find(Op);
3101  if (PI != TranslateMap.end())
3102  Op = PI->second;
3103  }
3104 
3105  // Check for trivial simplification.
3106  if (Value *V = SimplifyInstruction(N, {DL, nullptr, nullptr, AC})) {
3107  if (!BBI->use_empty())
3108  TranslateMap[&*BBI] = V;
3109  if (!N->mayHaveSideEffects()) {
3110  N->deleteValue(); // Instruction folded away, don't need actual inst
3111  N = nullptr;
3112  }
3113  } else {
3114  if (!BBI->use_empty())
3115  TranslateMap[&*BBI] = N;
3116  }
3117  if (N) {
3118  // Insert the new instruction into its new home.
3119  EdgeBB->getInstList().insert(InsertPt, N);
3120 
3121  // Register the new instruction with the assumption cache if necessary.
3122  if (auto *Assume = dyn_cast<AssumeInst>(N))
3123  if (AC)
3124  AC->registerAssumption(Assume);
3125  }
3126  }
3127 
3128  // Loop over all of the edges from PredBB to BB, changing them to branch
3129  // to EdgeBB instead.
3130  Instruction *PredBBTI = PredBB->getTerminator();
3131  for (unsigned i = 0, e = PredBBTI->getNumSuccessors(); i != e; ++i)
3132  if (PredBBTI->getSuccessor(i) == BB) {
3133  BB->removePredecessor(PredBB);
3134  PredBBTI->setSuccessor(i, EdgeBB);
3135  }
3136 
3137  if (DTU) {
3138  Updates.push_back({DominatorTree::Insert, PredBB, EdgeBB});
3139  Updates.push_back({DominatorTree::Delete, PredBB, BB});
3140 
3141  DTU->applyUpdates(Updates);
3142  }
3143 
3144  // Signal repeat, simplifying any other constants.
3145  return None;
3146  }
3147 
3148  return false;
3149 }
3150 
3152  DomTreeUpdater *DTU,
3153  const DataLayout &DL,
3154  AssumptionCache *AC) {
3155  Optional<bool> Result;
3156  bool EverChanged = false;
3157  do {
3158  // Note that None means "we changed things, but recurse further."
3159  Result = FoldCondBranchOnValueKnownInPredecessorImpl(BI, DTU, DL, AC);
3160  EverChanged |= Result == None || *Result;
3161  } while (Result == None);
3162  return EverChanged;
3163 }
3164 
3165 /// Given a BB that starts with the specified two-entry PHI node,
3166 /// see if we can eliminate it.
3168  DomTreeUpdater *DTU, const DataLayout &DL) {
3169  // Ok, this is a two entry PHI node. Check to see if this is a simple "if
3170  // statement", which has a very simple dominance structure. Basically, we
3171  // are trying to find the condition that is being branched on, which
3172  // subsequently causes this merge to happen. We really want control
3173  // dependence information for this check, but simplifycfg can't keep it up
3174  // to date, and this catches most of the cases we care about anyway.
3175  BasicBlock *BB = PN->getParent();
3176 
3177  BasicBlock *IfTrue, *IfFalse;
3178  BranchInst *DomBI = GetIfCondition(BB, IfTrue, IfFalse);
3179  if (!DomBI)
3180  return false;
3181  Value *IfCond = DomBI->getCondition();
3182  // Don't bother if the branch will be constant folded trivially.
3183  if (isa<ConstantInt>(IfCond))
3184  return false;
3185 
3186  BasicBlock *DomBlock = DomBI->getParent();
3188  llvm::copy_if(
3189  PN->blocks(), std::back_inserter(IfBlocks), [](BasicBlock *IfBlock) {
3190  return cast<BranchInst>(IfBlock->getTerminator())->isUnconditional();
3191  });
3192  assert((IfBlocks.size() == 1 || IfBlocks.size() == 2) &&
3193  "Will have either one or two blocks to speculate.");
3194 
3195  // If the branch is non-unpredictable, see if we either predictably jump to
3196  // the merge bb (if we have only a single 'then' block), or if we predictably
3197  // jump to one specific 'then' block (if we have two of them).
3198  // It isn't beneficial to speculatively execute the code
3199  // from the block that we know is predictably not entered.
3200  if (!DomBI->getMetadata(LLVMContext::MD_unpredictable)) {
3201  uint64_t TWeight, FWeight;
3202  if (DomBI->extractProfMetadata(TWeight, FWeight) &&
3203  (TWeight + FWeight) != 0) {
3204  BranchProbability BITrueProb =
3205  BranchProbability::getBranchProbability(TWeight, TWeight + FWeight);
3207  BranchProbability BIFalseProb = BITrueProb.getCompl();
3208  if (IfBlocks.size() == 1) {
3209  BranchProbability BIBBProb =
3210  DomBI->getSuccessor(0) == BB ? BITrueProb : BIFalseProb;
3211  if (BIBBProb >= Likely)
3212  return false;
3213  } else {
3214  if (BITrueProb >= Likely || BIFalseProb >= Likely)
3215  return false;
3216  }
3217  }
3218  }
3219 
3220  // Don't try to fold an unreachable block. For example, the phi node itself
3221  // can't be the candidate if-condition for a select that we want to form.
3222  if (auto *IfCondPhiInst = dyn_cast<PHINode>(IfCond))
3223  if (IfCondPhiInst->getParent() == BB)
3224  return false;
3225 
3226  // Okay, we found that we can merge this two-entry phi node into a select.
3227  // Doing so would require us to fold *all* two entry phi nodes in this block.
3228  // At some point this becomes non-profitable (particularly if the target
3229  // doesn't support cmov's). Only do this transformation if there are two or
3230  // fewer PHI nodes in this block.
3231  unsigned NumPhis = 0;
3232  for (BasicBlock::iterator I = BB->begin(); isa<PHINode>(I); ++NumPhis, ++I)
3233  if (NumPhis > 2)
3234  return false;
3235 
3236  // Loop over the PHI's seeing if we can promote them all to select
3237  // instructions. While we are at it, keep track of the instructions
3238  // that need to be moved to the dominating block.
3239  SmallPtrSet<Instruction *, 4> AggressiveInsts;
3240  InstructionCost Cost = 0;
3241  InstructionCost Budget =
3242  TwoEntryPHINodeFoldingThreshold * TargetTransformInfo::TCC_Basic;
3243 
3244  bool Changed = false;
3245  for (BasicBlock::iterator II = BB->begin(); isa<PHINode>(II);) {
3246  PHINode *PN = cast<PHINode>(II++);
3247  if (Value *V = SimplifyInstruction(PN, {DL, PN})) {
3248  PN->replaceAllUsesWith(V);
3249  PN->eraseFromParent();
3250  Changed = true;
3251  continue;
3252  }
3253 
3254  if (!dominatesMergePoint(PN->getIncomingValue(0), BB, AggressiveInsts,
3255  Cost, Budget, TTI) ||
3256  !dominatesMergePoint(PN->getIncomingValue(1), BB, AggressiveInsts,
3257  Cost, Budget, TTI))
3258  return Changed;
3259  }
3260 
3261  // If we folded the first phi, PN dangles at this point. Refresh it. If
3262  // we ran out of PHIs then we simplified them all.
3263  PN = dyn_cast<PHINode>(BB->begin());
3264  if (!PN)
3265  return true;
3266 
3267  // Return true if at least one of these is a 'not', and another is either
3268  // a 'not' too, or a constant.
3269  auto CanHoistNotFromBothValues = [](Value *V0, Value *V1) {
3270  if (!match(V0, m_Not(m_Value())))
3271  std::swap(V0, V1);
3272  auto Invertible = m_CombineOr(m_Not(m_Value()), m_AnyIntegralConstant());
3273  return match(V0, m_Not(m_Value())) && match(V1, Invertible);
3274  };
3275 
3276  // Don't fold i1 branches on PHIs which contain binary operators or
3277  // (possibly inverted) select form of or/ands, unless one of
3278  // the incoming values is an 'not' and another one is freely invertible.
3279  // These can often be turned into switches and other things.
3280  auto IsBinOpOrAnd = [](Value *V) {
3281  return match(
3282  V, m_CombineOr(
3283  m_BinOp(),
3285  m_Select(m_Value(), m_Value(), m_ImmConstant()))));
3286  };
3287  if (PN->getType()->isIntegerTy(1) &&
3288  (IsBinOpOrAnd(PN->getIncomingValue(0)) ||
3289  IsBinOpOrAnd(PN->getIncomingValue(1)) || IsBinOpOrAnd(IfCond)) &&
3290  !CanHoistNotFromBothValues(PN->getIncomingValue(0),
3291  PN->getIncomingValue(1)))
3292  return Changed;
3293 
3294  // If all PHI nodes are promotable, check to make sure that all instructions
3295  // in the predecessor blocks can be promoted as well. If not, we won't be able
3296  // to get rid of the control flow, so it's not worth promoting to select
3297  // instructions.
3298  for (BasicBlock *IfBlock : IfBlocks)
3299  for (BasicBlock::iterator I = IfBlock->begin(); !I->isTerminator(); ++I)
3300  if (!AggressiveInsts.count(&*I) && !I->isDebugOrPseudoInst()) {
3301  // This is not an aggressive instruction that we can promote.
3302  // Because of this, we won't be able to get rid of the control flow, so
3303  // the xform is not worth it.
3304  return Changed;
3305  }
3306 
3307  // If either of the blocks has it's address taken, we can't do this fold.
3308  if (any_of(IfBlocks,
3309  [](BasicBlock *IfBlock) { return IfBlock->hasAddressTaken(); }))
3310  return Changed;
3311 
3312  LLVM_DEBUG(dbgs() << "FOUND IF CONDITION! " << *IfCond
3313  << " T: " << IfTrue->getName()
3314  << " F: " << IfFalse->getName() << "\n");
3315 
3316  // If we can still promote the PHI nodes after this gauntlet of tests,
3317  // do all of the PHI's now.
3318 
3319  // Move all 'aggressive' instructions, which are defined in the
3320  // conditional parts of the if's up to the dominating block.
3321  for (BasicBlock *IfBlock : IfBlocks)
3322  hoistAllInstructionsInto(DomBlock, DomBI, IfBlock);
3323 
3325  // Propagate fast-math-flags from phi nodes to replacement selects.
3327  while (PHINode *PN = dyn_cast<PHINode>(BB->begin())) {
3328  if (isa<FPMathOperator>(PN))
3329  Builder.setFastMathFlags(PN->getFastMathFlags());
3330 
3331  // Change the PHI node into a select instruction.
3332  Value *TrueVal = PN->getIncomingValueForBlock(IfTrue);
3333  Value *FalseVal = PN->getIncomingValueForBlock(IfFalse);
3334 
3335  Value *Sel = Builder.CreateSelect(IfCond, TrueVal, FalseVal, "", DomBI);
3336  PN->replaceAllUsesWith(Sel);
3337  Sel->takeName(PN);
3338  PN->eraseFromParent();
3339  }
3340 
3341  // At this point, all IfBlocks are empty, so our if statement
3342  // has been flattened. Change DomBlock to jump directly to our new block to
3343  // avoid other simplifycfg's kicking in on the diamond.
3344  Builder.CreateBr(BB);
3345 
3347  if (DTU) {
3348  Updates.push_back({DominatorTree::Insert, DomBlock, BB});
3349  for (auto *Successor : successors(DomBlock))
3350  Updates.push_back({DominatorTree::Delete, DomBlock, Successor});
3351  }
3352 
3353  DomBI->eraseFromParent();
3354  if (DTU)
3355  DTU->applyUpdates(Updates);
3356 
3357  return true;
3358 }
3359 
3362  Value *RHS, const Twine &Name = "") {
3363  // Try to relax logical op to binary op.
3364  if (impliesPoison(RHS, LHS))
3365  return Builder.CreateBinOp(Opc, LHS, RHS, Name);
3366  if (Opc == Instruction::And)
3367  return Builder.CreateLogicalAnd(LHS, RHS, Name);
3368  if (Opc == Instruction::Or)
3369  return Builder.CreateLogicalOr(LHS, RHS, Name);
3370  llvm_unreachable("Invalid logical opcode");
3371 }
3372 
3373 /// Return true if either PBI or BI has branch weight available, and store
3374 /// the weights in {Pred|Succ}{True|False}Weight. If one of PBI and BI does
3375 /// not have branch weight, use 1:1 as its weight.
3377  uint64_t &PredTrueWeight,
3378  uint64_t &PredFalseWeight,
3379  uint64_t &SuccTrueWeight,
3380  uint64_t &SuccFalseWeight) {
3381  bool PredHasWeights =
3382  PBI->extractProfMetadata(PredTrueWeight, PredFalseWeight);
3383  bool SuccHasWeights =
3384  BI->extractProfMetadata(SuccTrueWeight, SuccFalseWeight);
3385  if (PredHasWeights || SuccHasWeights) {
3386  if (!PredHasWeights)
3387  PredTrueWeight = PredFalseWeight = 1;
3388  if (!SuccHasWeights)
3389  SuccTrueWeight = SuccFalseWeight = 1;
3390  return true;
3391  } else {
3392  return false;
3393  }
3394 }
3395 
3396 /// Determine if the two branches share a common destination and deduce a glue
3397 /// that joins the branches' conditions to arrive at the common destination if
3398 /// that would be profitable.
3401  const TargetTransformInfo *TTI) {
3402  assert(BI && PBI && BI->isConditional() && PBI->isConditional() &&
3403  "Both blocks must end with a conditional branches.");
3405  "PredBB must be a predecessor of BB.");
3406 
3407  // We have the potential to fold the conditions together, but if the
3408  // predecessor branch is predictable, we may not want to merge them.
3409  uint64_t PTWeight, PFWeight;
3410  BranchProbability PBITrueProb, Likely;
3411  if (TTI && !PBI->getMetadata(LLVMContext::MD_unpredictable) &&
3412  PBI->extractProfMetadata(PTWeight, PFWeight) &&
3413  (PTWeight + PFWeight) != 0) {
3414  PBITrueProb =
3415  BranchProbability::getBranchProbability(PTWeight, PTWeight + PFWeight);
3416  Likely = TTI->getPredictableBranchThreshold();
3417  }
3418 
3419  if (PBI->getSuccessor(0) == BI->getSuccessor(0)) {
3420  // Speculate the 2nd condition unless the 1st is probably true.
3421  if (PBITrueProb.isUnknown() || PBITrueProb < Likely)
3422  return {{Instruction::Or, false}};
3423  } else if (PBI->getSuccessor(1) == BI->getSuccessor(1)) {
3424  // Speculate the 2nd condition unless the 1st is probably false.
3425  if (PBITrueProb.isUnknown() || PBITrueProb.getCompl() < Likely)
3426  return {{Instruction::And, false}};
3427  } else if (PBI->getSuccessor(0) == BI->getSuccessor(1)) {
3428  // Speculate the 2nd condition unless the 1st is probably true.
3429  if (PBITrueProb.isUnknown() || PBITrueProb < Likely)
3430  return {{Instruction::And, true}};
3431  } else if (PBI->getSuccessor(1) == BI->getSuccessor(0)) {
3432  // Speculate the 2nd condition unless the 1st is probably false.
3433  if (PBITrueProb.isUnknown() || PBITrueProb.getCompl() < Likely)
3434  return {{Instruction::Or, true}};
3435  }
3436  return None;
3437 }
3438 
3440  DomTreeUpdater *DTU,
3441  MemorySSAUpdater *MSSAU,
3442  const TargetTransformInfo *TTI) {
3443  BasicBlock *BB = BI->getParent();
3444  BasicBlock *PredBlock = PBI->getParent();
3445 
3446  // Determine if the two branches share a common destination.
3448  bool InvertPredCond;
3449  std::tie(Opc, InvertPredCond) =
3451 
3452  LLVM_DEBUG(dbgs() << "FOLDING BRANCH TO COMMON DEST:\n" << *PBI << *BB);
3453 
3454  IRBuilder<> Builder(PBI);
3455  // The builder is used to create instructions to eliminate the branch in BB.
3456  // If BB's terminator has !annotation metadata, add it to the new
3457  // instructions.
3458  Builder.CollectMetadataToCopy(BB->getTerminator(),
3459  {LLVMContext::MD_annotation});
3460 
3461  // If we need to invert the condition in the pred block to match, do so now.
3462  if (InvertPredCond) {
3463  Value *NewCond = PBI->getCondition();
3464  if (NewCond->hasOneUse() && isa<CmpInst>(NewCond)) {
3465  CmpInst *CI = cast<CmpInst>(NewCond);
3466  CI->setPredicate(CI->getInversePredicate());
3467  } else {
3468  NewCond =
3469  Builder.CreateNot(NewCond, PBI->getCondition()->getName() + ".not");
3470  }
3471 
3472  PBI->setCondition(NewCond);
3473  PBI->swapSuccessors();
3474  }
3475 
3476  BasicBlock *UniqueSucc =
3477  PBI->getSuccessor(0) == BB ? BI->getSuccessor(0) : BI->getSuccessor(1);
3478 
3479  // Before cloning instructions, notify the successor basic block that it
3480  // is about to have a new predecessor. This will update PHI nodes,
3481  // which will allow us to update live-out uses of bonus instructions.
3482  AddPredecessorToBlock(UniqueSucc, PredBlock, BB, MSSAU);
3483 
3484  // Try to update branch weights.
3485  uint64_t PredTrueWeight, PredFalseWeight, SuccTrueWeight, SuccFalseWeight;
3486  if (extractPredSuccWeights(PBI, BI, PredTrueWeight, PredFalseWeight,
3487  SuccTrueWeight, SuccFalseWeight)) {
3488  SmallVector<uint64_t, 8> NewWeights;
3489 
3490  if (PBI->getSuccessor(0) == BB) {
3491  // PBI: br i1 %x, BB, FalseDest
3492  // BI: br i1 %y, UniqueSucc, FalseDest
3493  // TrueWeight is TrueWeight for PBI * TrueWeight for BI.
3494  NewWeights.push_back(PredTrueWeight * SuccTrueWeight);
3495  // FalseWeight is FalseWeight for PBI * TotalWeight for BI +
3496  // TrueWeight for PBI * FalseWeight for BI.
3497  // We assume that total weights of a BranchInst can fit into 32 bits.
3498  // Therefore, we will not have overflow using 64-bit arithmetic.
3499  NewWeights.push_back(PredFalseWeight *
3500  (SuccFalseWeight + SuccTrueWeight) +
3501  PredTrueWeight * SuccFalseWeight);
3502  } else {
3503  // PBI: br i1 %x, TrueDest, BB
3504  // BI: br i1 %y, TrueDest, UniqueSucc
3505  // TrueWeight is TrueWeight for PBI * TotalWeight for BI +
3506  // FalseWeight for PBI * TrueWeight for BI.
3507  NewWeights.push_back(PredTrueWeight * (SuccFalseWeight + SuccTrueWeight) +
3508  PredFalseWeight * SuccTrueWeight);
3509  // FalseWeight is FalseWeight for PBI * FalseWeight for BI.
3510  NewWeights.push_back(PredFalseWeight * SuccFalseWeight);
3511  }
3512 
3513  // Halve the weights if any of them cannot fit in an uint32_t
3514  FitWeights(NewWeights);
3515 
3516  SmallVector<uint32_t, 8> MDWeights(NewWeights.begin(), NewWeights.end());
3517  setBranchWeights(PBI, MDWeights[0], MDWeights[1]);
3518 
3519  // TODO: If BB is reachable from all paths through PredBlock, then we
3520  // could replace PBI's branch probabilities with BI's.
3521  } else
3522  PBI->setMetadata(LLVMContext::MD_prof, nullptr);
3523 
3524  // Now, update the CFG.
3525  PBI->setSuccessor(PBI->getSuccessor(0) != BB, UniqueSucc);
3526 
3527  if (DTU)
3528  DTU->applyUpdates({{DominatorTree::Insert, PredBlock, UniqueSucc},
3529  {DominatorTree::Delete, PredBlock, BB}});
3530 
3531  // If BI was a loop latch, it may have had associated loop metadata.
3532  // We need to copy it to the new latch, that is, PBI.
3533  if (MDNode *LoopMD = BI->getMetadata(LLVMContext::MD_loop))
3534  PBI->setMetadata(LLVMContext::MD_loop, LoopMD);
3535 
3536  ValueToValueMapTy VMap; // maps original values to cloned values
3538 
3539  // Now that the Cond was cloned into the predecessor basic block,
3540  // or/and the two conditions together.
3541  Value *BICond = VMap[BI->getCondition()];
3542  PBI->setCondition(
3543  createLogicalOp(Builder, Opc, PBI->getCondition(), BICond, "or.cond"));
3544 
3545  // Copy any debug value intrinsics into the end of PredBlock.
3546  for (Instruction &I : *BB) {
3547  if (isa<DbgInfoIntrinsic>(I)) {
3548  Instruction *NewI = I.clone();
3549  RemapInstruction(NewI, VMap,
3551  NewI->insertBefore(PBI);
3552  }
3553  }
3554 
3555  ++NumFoldBranchToCommonDest;
3556  return true;
3557 }
3558 
3559 /// Return if an instruction's type or any of its operands' types are a vector
3560 /// type.
3561 static bool isVectorOp(Instruction &I) {
3562  return I.getType()->isVectorTy() || any_of(I.operands(), [](Use &U) {
3563  return U->getType()->isVectorTy();
3564  });
3565 }
3566 
3567 /// If this basic block is simple enough, and if a predecessor branches to us
3568 /// and one of our successors, fold the block into the predecessor and use
3569 /// logical operations to pick the right destination.
3571  MemorySSAUpdater *MSSAU,
3572  const TargetTransformInfo *TTI,
3573  unsigned BonusInstThreshold) {
3574  // If this block ends with an unconditional branch,
3575  // let SpeculativelyExecuteBB() deal with it.
3576  if (!BI->isConditional())
3577  return false;
3578 
3579  BasicBlock *BB = BI->getParent();
3581  BB->getParent()->hasMinSize() ? TargetTransformInfo::TCK_CodeSize
3582  : TargetTransformInfo::TCK_SizeAndLatency;
3583 
3584  Instruction *Cond = dyn_cast<Instruction>(BI->getCondition());
3585 
3586  if (!Cond ||
3587  (!isa<CmpInst>(Cond) && !isa<BinaryOperator>(Cond) &&
3588  !isa<SelectInst>(Cond)) ||
3589  Cond->getParent() != BB || !Cond->hasOneUse())
3590  return false;
3591 
3592  // Cond is known to be a compare or binary operator. Check to make sure that
3593  // neither operand is a potentially-trapping constant expression.
3594  if (ConstantExpr *CE = dyn_cast<ConstantExpr>(Cond->getOperand(0)))
3595  if (CE->canTrap())
3596  return false;
3597  if (ConstantExpr *CE = dyn_cast<ConstantExpr>(Cond->getOperand(1)))
3598  if (CE->canTrap())
3599  return false;
3600 
3601  // Finally, don't infinitely unroll conditional loops.
3602  if (is_contained(successors(BB), BB))
3603  return false;
3604 
3605  // With which predecessors will we want to deal with?
3607  for (BasicBlock *PredBlock : predecessors(BB)) {
3608  BranchInst *PBI = dyn_cast<BranchInst>(PredBlock->getTerminator());
3609 
3610  // Check that we have two conditional branches. If there is a PHI node in
3611  // the common successor, verify that the same value flows in from both
3612  // blocks.
3613  if (!PBI || PBI->isUnconditional() || !SafeToMergeTerminators(BI, PBI))
3614  continue;
3615 
3616  // Determine if the two branches share a common destination.
3618  bool InvertPredCond;
3619  if (auto Recipe = shouldFoldCondBranchesToCommonDestination(BI, PBI, TTI))
3620  std::tie(Opc, InvertPredCond) = *Recipe;
3621  else
3622  continue;
3623 
3624  // Check the cost of inserting the necessary logic before performing the
3625  // transformation.
3626  if (TTI) {
3627  Type *Ty = BI->getCondition()->getType();
3629  if (InvertPredCond && (!PBI->getCondition()->hasOneUse() ||
3630  !isa<CmpInst>(PBI->getCondition())))
3631  Cost += TTI->getArithmeticInstrCost(Instruction::Xor, Ty, CostKind);
3632 
3633  if (Cost > BranchFoldThreshold)
3634  continue;
3635  }
3636 
3637  // Ok, we do want to deal with this predecessor. Record it.
3638  Preds.emplace_back(PredBlock);
3639  }
3640 
3641  // If there aren't any predecessors into which we can fold,
3642  // don't bother checking the cost.
3643  if (Preds.empty())
3644  return false;
3645 
3646  // Only allow this transformation if computing the condition doesn't involve
3647  // too many instructions and these involved instructions can be executed
3648  // unconditionally. We denote all involved instructions except the condition
3649  // as "bonus instructions", and only allow this transformation when the
3650  // number of the bonus instructions we'll need to create when cloning into
3651  // each predecessor does not exceed a certain threshold.
3652  unsigned NumBonusInsts = 0;
3653  bool SawVectorOp = false;
3654  const unsigned PredCount = Preds.size();
3655  for (Instruction &I : *BB) {
3656  // Don't check the branch condition comparison itself.
3657  if (&I == Cond)
3658  continue;
3659  // Ignore dbg intrinsics, and the terminator.
3660  if (isa<DbgInfoIntrinsic>(I) || isa<BranchInst>(I))
3661  continue;
3662  // I must be safe to execute unconditionally.
3664  return false;
3665  SawVectorOp |= isVectorOp(I);
3666 
3667  // Account for the cost of duplicating this instruction into each
3668  // predecessor. Ignore free instructions.
3669  if (!TTI ||
3670  TTI->getUserCost(&I, CostKind) != TargetTransformInfo::TCC_Free) {
3671  NumBonusInsts += PredCount;
3672 
3673  // Early exits once we reach the limit.
3674  if (NumBonusInsts >
3675  BonusInstThreshold * BranchFoldToCommonDestVectorMultiplier)
3676  return false;
3677  }
3678 
3679  auto IsBCSSAUse = [BB, &I](Use &U) {
3680  auto *UI = cast<Instruction>(U.getUser());
3681  if (auto *PN = dyn_cast<PHINode>(UI))
3682  return PN->getIncomingBlock(U) == BB;
3683  return UI->getParent() == BB && I.comesBefore(UI);
3684  };
3685 
3686  // Does this instruction require rewriting of uses?
3687  if (!all_of(I.uses(), IsBCSSAUse))
3688  return false;
3689  }
3690  if (NumBonusInsts >
3691  BonusInstThreshold *
3692  (SawVectorOp ? BranchFoldToCommonDestVectorMultiplier : 1))
3693  return false;
3694 
3695  // Ok, we have the budget. Perform the transformation.
3696  for (BasicBlock *PredBlock : Preds) {
3697  auto *PBI = cast<BranchInst>(PredBlock->getTerminator());
3698  return performBranchToCommonDestFolding(BI, PBI, DTU, MSSAU, TTI);
3699  }
3700  return false;
3701 }
3702 
3703 // If there is only one store in BB1 and BB2, return it, otherwise return
3704 // nullptr.
3706  StoreInst *S = nullptr;
3707  for (auto *BB : {BB1, BB2}) {
3708  if (!BB)
3709  continue;
3710  for (auto &I : *BB)
3711  if (auto *SI = dyn_cast<StoreInst>(&I)) {
3712  if (S)
3713  // Multiple stores seen.
3714  return nullptr;
3715  else
3716  S = SI;
3717  }
3718  }
3719  return S;
3720 }
3721 
3723  Value *AlternativeV = nullptr) {
3724  // PHI is going to be a PHI node that allows the value V that is defined in
3725  // BB to be referenced in BB's only successor.
3726  //
3727  // If AlternativeV is nullptr, the only value we care about in PHI is V. It
3728  // doesn't matter to us what the other operand is (it'll never get used). We
3729  // could just create a new PHI with an undef incoming value, but that could
3730  // increase register pressure if EarlyCSE/InstCombine can't fold it with some
3731  // other PHI. So here we directly look for some PHI in BB's successor with V
3732  // as an incoming operand. If we find one, we use it, else we create a new
3733  // one.
3734  //
3735  // If AlternativeV is not nullptr, we care about both incoming values in PHI.
3736  // PHI must be exactly: phi <ty> [ %BB, %V ], [ %OtherBB, %AlternativeV]
3737  // where OtherBB is the single other predecessor of BB's only successor.
3738  PHINode *PHI = nullptr;
3739  BasicBlock *Succ = BB->getSingleSuccessor();
3740 
3741  for (auto I = Succ->begin(); isa<PHINode>(I); ++I)
3742  if (cast<PHINode>(I)->getIncomingValueForBlock(BB) == V) {
3743  PHI = cast<PHINode>(I);
3744  if (!AlternativeV)
3745  break;
3746 
3747  assert(Succ->hasNPredecessors(2));
3748  auto PredI = pred_begin(Succ);
3749  BasicBlock *OtherPredBB = *PredI == BB ? *++PredI : *PredI;
3750  if (PHI->getIncomingValueForBlock(OtherPredBB) == AlternativeV)
3751  break;
3752  PHI = nullptr;
3753  }
3754  if (PHI)
3755  return PHI;
3756 
3757  // If V is not an instruction defined in BB, just return it.
3758  if (!AlternativeV &&
3759  (!isa<Instruction>(V) || cast<Instruction>(V)->getParent() != BB))
3760  return V;
3761 
3762  PHI = PHINode::Create(V->getType(), 2, "simplifycfg.merge", &Succ->front());
3763  PHI->addIncoming(V, BB);
3764  for (BasicBlock *PredBB : predecessors(Succ))
3765  if (PredBB != BB)
3766  PHI->addIncoming(
3767  AlternativeV ? AlternativeV : UndefValue::get(V->getType()), PredBB);
3768  return PHI;
3769 }
3770 
3772  BasicBlock *PTB, BasicBlock *PFB, BasicBlock *QTB, BasicBlock *QFB,
3773  BasicBlock *PostBB, Value *Address, bool InvertPCond, bool InvertQCond,
3774  DomTreeUpdater *DTU, const DataLayout &DL, const TargetTransformInfo &TTI) {
3775  // For every pointer, there must be exactly two stores, one coming from
3776  // PTB or PFB, and the other from QTB or QFB. We don't support more than one
3777  // store (to any address) in PTB,PFB or QTB,QFB.
3778  // FIXME: We could relax this restriction with a bit more work and performance
3779  // testing.
3780  StoreInst *PStore = findUniqueStoreInBlocks(PTB, PFB);
3781  StoreInst *QStore = findUniqueStoreInBlocks(QTB, QFB);
3782  if (!PStore || !QStore)
3783  return false;
3784 
3785  // Now check the stores are compatible.
3786  if (!QStore->isUnordered() || !PStore->isUnordered() ||
3787  PStore->getValueOperand()->getType() !=
3788  QStore->getValueOperand()->getType())
3789  return false;
3790 
3791  // Check that sinking the store won't cause program behavior changes. Sinking
3792  // the store out of the Q blocks won't change any behavior as we're sinking
3793  // from a block to its unconditional successor. But we're moving a store from
3794  // the P blocks down through the middle block (QBI) and past both QFB and QTB.
3795  // So we need to check that there are no aliasing loads or stores in
3796  // QBI, QTB and QFB. We also need to check there are no conflicting memory
3797  // operations between PStore and the end of its parent block.
3798  //
3799  // The ideal way to do this is to query AliasAnalysis, but we don't
3800  // preserve AA currently so that is dangerous. Be super safe and just
3801  // check there are no other memory operations at all.
3802  for (auto &I : *QFB->getSinglePredecessor())
3803  if (I.mayReadOrWriteMemory())
3804  return false;
3805  for (auto &I : *QFB)
3806  if (&I != QStore && I.mayReadOrWriteMemory())
3807  return false;
3808  if (QTB)
3809  for (auto &I : *QTB)
3810  if (&I != QStore && I.mayReadOrWriteMemory())
3811  return false;
3812  for (auto I = BasicBlock::iterator(PStore), E = PStore->getParent()->end();
3813  I != E; ++I)
3814  if (&*I != PStore && I->mayReadOrWriteMemory())
3815  return false;
3816 
3817  // If we're not in aggressive mode, we only optimize if we have some
3818  // confidence that by optimizing we'll allow P and/or Q to be if-converted.
3819  auto IsWorthwhile = [&](BasicBlock *BB, ArrayRef<StoreInst *> FreeStores) {
3820  if (!BB)
3821  return true;
3822  // Heuristic: if the block can be if-converted/phi-folded and the
3823  // instructions inside are all cheap (arithmetic/GEPs), it's worthwhile to
3824  // thread this store.
3825  InstructionCost Cost = 0;
3826  InstructionCost Budget =
3827  PHINodeFoldingThreshold * TargetTransformInfo::TCC_Basic;
3828  for (auto &I : BB->instructionsWithoutDebug(false)) {
3829  // Consider terminator instruction to be free.
3830  if (I.isTerminator())
3831  continue;
3832  // If this is one the stores that we want to speculate out of this BB,
3833  // then don't count it's cost, consider it to be free.
3834  if (auto *S = dyn_cast<StoreInst>(&I))
3835  if (llvm::find(FreeStores, S))
3836  continue;
3837  // Else, we have a white-list of instructions that we are ak speculating.
3838  if (!isa<BinaryOperator>(I) && !isa<GetElementPtrInst>(I))
3839  return false; // Not in white-list - not worthwhile folding.
3840  // And finally, if this is a non-free instruction that we are okay
3841  // speculating, ensure that we consider the speculation budget.
3842  Cost += TTI.getUserCost(&I, TargetTransformInfo::TCK_SizeAndLatency);
3843  if (Cost > Budget)
3844  return false; // Eagerly refuse to fold as soon as we're out of budget.
3845  }
3846  assert(Cost <= Budget &&
3847  "When we run out of budget we will eagerly return from within the "
3848  "per-instruction loop.");
3849  return true;
3850  };
3851 
3852  const std::array<StoreInst *, 2> FreeStores = {PStore, QStore};
3854  (!IsWorthwhile(PTB, FreeStores) || !IsWorthwhile(PFB, FreeStores) ||
3855  !IsWorthwhile(QTB, FreeStores) || !IsWorthwhile(QFB, FreeStores)))
3856  return false;
3857 
3858  // If PostBB has more than two predecessors, we need to split it so we can
3859  // sink the store.
3860  if (std::next(pred_begin(PostBB), 2) != pred_end(PostBB)) {
3861  // We know that QFB's only successor is PostBB. And QFB has a single
3862  // predecessor. If QTB exists, then its only successor is also PostBB.
3863  // If QTB does not exist, then QFB's only predecessor has a conditional
3864  // branch to QFB and PostBB.
3865  BasicBlock *TruePred = QTB ? QTB : QFB->getSinglePredecessor();
3866  BasicBlock *NewBB =
3867  SplitBlockPredecessors(PostBB, {QFB, TruePred}, "condstore.split", DTU);
3868  if (!NewBB)
3869  return false;
3870  PostBB = NewBB;
3871  }
3872 
3873  // OK, we're going to sink the stores to PostBB. The store has to be
3874  // conditional though, so first create the predicate.
3875  Value *PCond = cast<BranchInst>(PFB->getSinglePredecessor()->getTerminator())
3876  ->getCondition();
3877  Value *QCond = cast<BranchInst>(QFB->getSinglePredecessor()->getTerminator())
3878  ->getCondition();
3879 
3881  PStore->getParent());
3883  QStore->getParent(), PPHI);
3884 
3885  IRBuilder<> QB(&*PostBB->getFirstInsertionPt());
3886 
3887  Value *PPred = PStore->getParent() == PTB ? PCond : QB.CreateNot(PCond);
3888  Value *QPred = QStore->getParent() == QTB ? QCond : QB.CreateNot(QCond);
3889 
3890  if (InvertPCond)
3891  PPred = QB.CreateNot(PPred);
3892  if (InvertQCond)
3893  QPred = QB.CreateNot(QPred);
3894  Value *CombinedPred = QB.CreateOr(PPred, QPred);
3895 
3896  auto *T = SplitBlockAndInsertIfThen(CombinedPred, &*QB.GetInsertPoint(),
3897  /*Unreachable=*/false,
3898  /*BranchWeights=*/nullptr, DTU);
3899  QB.SetInsertPoint(T);
3900  StoreInst *SI = cast<StoreInst>(QB.CreateStore(QPHI, Address));
3901  SI->setAAMetadata(PStore->getAAMetadata().merge(QStore->getAAMetadata()));
3902  // Choose the minimum alignment. If we could prove both stores execute, we
3903  // could use biggest one. In this case, though, we only know that one of the
3904  // stores executes. And we don't know it's safe to take the alignment from a
3905  // store that doesn't execute.
3906  SI->setAlignment(std::min(PStore->getAlign(), QStore->getAlign()));
3907 
3908  QStore->eraseFromParent();
3909  PStore->eraseFromParent();
3910 
3911  return true;
3912 }
3913 
3915  DomTreeUpdater *DTU, const DataLayout &DL,
3916  const TargetTransformInfo &TTI) {
3917  // The intention here is to find diamonds or triangles (see below) where each
3918  // conditional block contains a store to the same address. Both of these
3919  // stores are conditional, so they can't be unconditionally sunk. But it may
3920  // be profitable to speculatively sink the stores into one merged store at the
3921  // end, and predicate the merged store on the union of the two conditions of
3922  // PBI and QBI.
3923  //
3924  // This can reduce the number of stores executed if both of the conditions are
3925  // true, and can allow the blocks to become small enough to be if-converted.
3926  // This optimization will also chain, so that ladders of test-and-set
3927  // sequences can be if-converted away.
3928  //
3929  // We only deal with simple diamonds or triangles:
3930  //
3931  // PBI or PBI or a combination of the two
3932  // / \ | \
3933  // PTB PFB | PFB
3934  // \ / | /
3935  // QBI QBI
3936  // / \ | \
3937  // QTB QFB | QFB
3938  // \ / | /
3939  // PostBB PostBB
3940  //
3941  // We model triangles as a type of diamond with a nullptr "true" block.
3942  // Triangles are canonicalized so that the fallthrough edge is represented by
3943  // a true condition, as in the diagram above.
3944  BasicBlock *PTB = PBI->getSuccessor(0);
3945  BasicBlock *PFB = PBI->getSuccessor(1);
3946  BasicBlock *QTB = QBI->getSuccessor(0);
3947  BasicBlock *QFB = QBI->getSuccessor(1);
3948  BasicBlock *PostBB = QFB->getSingleSuccessor();
3949 
3950  // Make sure we have a good guess for PostBB. If QTB's only successor is
3951  // QFB, then QFB is a better PostBB.
3952  if (QTB->getSingleSuccessor() == QFB)
3953  PostBB = QFB;
3954 
3955  // If we couldn't find a good PostBB, stop.
3956  if (!PostBB)
3957  return false;
3958 
3959  bool InvertPCond = false, InvertQCond = false;
3960  // Canonicalize fallthroughs to the true branches.
3961  if (PFB == QBI->getParent()) {
3962  std::swap(PFB, PTB);
3963  InvertPCond = true;
3964  }
3965  if (QFB == PostBB) {
3966  std::swap(QFB, QTB);
3967  InvertQCond = true;
3968  }
3969 
3970  // From this point on we can assume PTB or QTB may be fallthroughs but PFB
3971  // and QFB may not. Model fallthroughs as a nullptr block.
3972  if (PTB == QBI->getParent())
3973  PTB = nullptr;
3974  if (QTB == PostBB)
3975  QTB = nullptr;
3976 
3977  // Legality bailouts. We must have at least the non-fallthrough blocks and
3978  // the post-dominating block, and the non-fallthroughs must only have one
3979  // predecessor.
3980  auto HasOnePredAndOneSucc = [](BasicBlock *BB, BasicBlock *P, BasicBlock *S) {
3981  return BB->getSinglePredecessor() == P && BB->getSingleSuccessor() == S;
3982  };
3983  if (!HasOnePredAndOneSucc(PFB, PBI->getParent(), QBI->getParent()) ||
3984  !HasOnePredAndOneSucc(QFB, QBI->getParent(), PostBB))
3985  return false;
3986  if ((PTB && !HasOnePredAndOneSucc(PTB, PBI->getParent(), QBI->getParent())) ||
3987  (QTB && !HasOnePredAndOneSucc(QTB, QBI->getParent(), PostBB)))
3988  return false;
3989  if (!QBI->getParent()->hasNUses(2))
3990  return false;
3991 
3992  // OK, this is a sequence of two diamonds or triangles.
3993  // Check if there are stores in PTB or PFB that are repeated in QTB or QFB.
3994  SmallPtrSet<Value *, 4> PStoreAddresses, QStoreAddresses;
3995  for (auto *BB : {PTB, PFB}) {
3996  if (!BB)
3997  continue;
3998  for (auto &I : *BB)
3999  if (StoreInst *SI = dyn_cast<StoreInst>(&I))
4000  PStoreAddresses.insert(SI->getPointerOperand());
4001  }
4002  for (auto *BB : {QTB, QFB}) {
4003  if (!BB)
4004  continue;
4005  for (auto &I : *BB)
4006  if (StoreInst *SI = dyn_cast<StoreInst>(&I))
4007  QStoreAddresses.insert(SI->getPointerOperand());
4008  }
4009 
4010  set_intersect(PStoreAddresses, QStoreAddresses);
4011  // set_intersect mutates PStoreAddresses in place. Rename it here to make it
4012  // clear what it contains.
4013  auto &CommonAddresses = PStoreAddresses;
4014 
4015  bool Changed = false;
4016  for (auto *Address : CommonAddresses)
4017  Changed |=
4018  mergeConditionalStoreToAddress(PTB, PFB, QTB, QFB, PostBB, Address,
4019  InvertPCond, InvertQCond, DTU, DL, TTI);
4020  return Changed;
4021 }
4022 
4023 /// If the previous block ended with a widenable branch, determine if reusing
4024 /// the target block is profitable and legal. This will have the effect of
4025 /// "widening" PBI, but doesn't require us to reason about hosting safety.
4027  DomTreeUpdater *DTU) {
4028  // TODO: This can be generalized in two important ways:
4029  // 1) We can allow phi nodes in IfFalseBB and simply reuse all the input
4030  // values from the PBI edge.
4031  // 2) We can sink side effecting instructions into BI's fallthrough
4032  // successor provided they doesn't contribute to computation of
4033  // BI's condition.
4034  Value *CondWB, *WC;
4035  BasicBlock *IfTrueBB, *IfFalseBB;
4036  if (!parseWidenableBranch(PBI, CondWB, WC, IfTrueBB, IfFalseBB) ||
4037  IfTrueBB != BI->getParent() || !BI->getParent()->getSinglePredecessor())
4038  return false;
4039  if (!IfFalseBB->phis().empty())
4040  return false; // TODO
4041  // Use lambda to lazily compute expensive condition after cheap ones.
4042  auto NoSideEffects = [](BasicBlock &BB) {
4043  return llvm::none_of(BB, [](const Instruction &I) {
4044  return I.mayWriteToMemory() || I.mayHaveSideEffects();
4045  });
4046  };
4047  if (BI->getSuccessor(1) != IfFalseBB && // no inf looping
4048  BI->getSuccessor(1)->getTerminatingDeoptimizeCall() && // profitability
4049  NoSideEffects(*BI->getParent())) {
4050  auto *OldSuccessor = BI->getSuccessor(1);
4051  OldSuccessor->removePredecessor(BI->getParent());
4052  BI->setSuccessor(1, IfFalseBB);
4053  if (DTU)
4054  DTU->applyUpdates(
4055  {{DominatorTree::Insert, BI->getParent(), IfFalseBB},
4056  {DominatorTree::Delete, BI->getParent(), OldSuccessor}});
4057  return true;
4058  }
4059  if (BI->getSuccessor(0) != IfFalseBB && // no inf looping
4060  BI->getSuccessor(0)->getTerminatingDeoptimizeCall() && // profitability
4061  NoSideEffects(*BI->getParent())) {
4062  auto *OldSuccessor = BI->getSuccessor(0);
4063  OldSuccessor->removePredecessor(BI->getParent());
4064  BI->setSuccessor(0, IfFalseBB);
4065  if (DTU)
4066  DTU->applyUpdates(
4067  {{DominatorTree::Insert, BI->getParent(), IfFalseBB},
4068  {DominatorTree::Delete, BI->getParent(), OldSuccessor}});
4069  return true;
4070  }
4071  return false;
4072 }
4073 
4074 /// If we have a conditional branch as a predecessor of another block,
4075 /// this function tries to simplify it. We know
4076 /// that PBI and BI are both conditional branches, and BI is in one of the
4077 /// successor blocks of PBI - PBI branches to BI.
4079  DomTreeUpdater *DTU,
4080  const DataLayout &DL,
4081  const TargetTransformInfo &TTI) {
4082  assert(PBI->isConditional() && BI->isConditional());
4083  BasicBlock *BB = BI->getParent();
4084 
4085  // If this block ends with a branch instruction, and if there is a
4086  // predecessor that ends on a branch of the same condition, make
4087  // this conditional branch redundant.
4088  if (PBI->getCondition() == BI->getCondition() &&
4089  PBI->getSuccessor(0) != PBI->getSuccessor(1)) {
4090  // Okay, the outcome of this conditional branch is statically
4091  // knowable. If this block had a single pred, handle specially, otherwise
4092  // FoldCondBranchOnValueKnownInPredecessor() will handle it.
4093  if (BB->getSinglePredecessor()) {
4094  // Turn this into a branch on constant.
4095  bool CondIsTrue = PBI->getSuccessor(0) == BB;
4096  BI->setCondition(
4097  ConstantInt::get(Type::getInt1Ty(BB->getContext()), CondIsTrue));
4098  return true; // Nuke the branch on constant.
4099  }
4100  }
4101 
4102  // If the previous block ended with a widenable branch, determine if reusing
4103  // the target block is profitable and legal. This will have the effect of
4104  // "widening" PBI, but doesn't require us to reason about hosting safety.
4105  if (tryWidenCondBranchToCondBranch(PBI, BI, DTU))
4106  return true;
4107 
4108  if (auto *CE = dyn_cast<ConstantExpr>(BI->getCondition()))
4109  if (CE->canTrap())
4110  return false;
4111 
4112  // If both branches are conditional and both contain stores to the same
4113  // address, remove the stores from the conditionals and create a conditional
4114  // merged store at the end.
4115  if (MergeCondStores && mergeConditionalStores(PBI, BI, DTU, DL, TTI))
4116  return true;
4117 
4118  // If this is a conditional branch in an empty block, and if any
4119  // predecessors are a conditional branch to one of our destinations,
4120  // fold the conditions into logical ops and one cond br.
4121 
4122  // Ignore dbg intrinsics.
4123  if (&*BB->instructionsWithoutDebug(false).begin() != BI)
4124  return false;
4125 
4126  int PBIOp, BIOp;
4127  if (PBI->getSuccessor(0) == BI->getSuccessor(0)) {
4128  PBIOp = 0;
4129  BIOp = 0;
4130  } else if (PBI->getSuccessor(0) == BI->getSuccessor(1)) {
4131  PBIOp = 0;
4132  BIOp = 1;
4133  } else if (PBI->getSuccessor(1) == BI->getSuccessor(0)) {
4134  PBIOp = 1;
4135  BIOp = 0;
4136  } else if (PBI->getSuccessor(1) == BI->getSuccessor(1)) {
4137  PBIOp = 1;
4138  BIOp = 1;
4139  } else {
4140  return false;
4141  }
4142 
4143  // Check to make sure that the other destination of this branch
4144  // isn't BB itself. If so, this is an infinite loop that will
4145  // keep getting unwound.
4146  if (PBI->getSuccessor(PBIOp) == BB)
4147  return false;
4148 
4149  // Do not perform this transformation if it would require
4150  // insertion of a large number of select instructions. For targets
4151  // without predication/cmovs, this is a big pessimization.
4152 
4153  // Also do not perform this transformation if any phi node in the common
4154  // destination block can trap when reached by BB or PBB (PR17073). In that
4155  // case, it would be unsafe to hoist the operation into a select instruction.
4156 
4157  BasicBlock *CommonDest = PBI->getSuccessor(PBIOp);
4158  BasicBlock *RemovedDest = PBI->getSuccessor(PBIOp ^ 1);
4159  unsigned NumPhis = 0;
4160  for (BasicBlock::iterator II = CommonDest->begin(); isa<PHINode>(II);
4161  ++II, ++NumPhis) {
4162  if (NumPhis > 2) // Disable this xform.
4163  return false;
4164 
4165  PHINode *PN = cast<PHINode>(II);
4166  Value *BIV = PN->getIncomingValueForBlock(BB);
4167  if (ConstantExpr *CE = dyn_cast<ConstantExpr>(BIV))
4168  if (CE->canTrap())
4169  return false;
4170 
4171  unsigned PBBIdx = PN->getBasicBlockIndex(PBI->getParent());
4172  Value *PBIV = PN->getIncomingValue(PBBIdx);
4173  if (ConstantExpr *CE = dyn_cast<ConstantExpr>(PBIV))
4174  if (CE->canTrap())
4175  return false;
4176  }
4177 
4178  // Finally, if everything is ok, fold the branches to logical ops.
4179  BasicBlock *OtherDest = BI->getSuccessor(BIOp ^ 1);
4180 
4181  LLVM_DEBUG(dbgs() << "FOLDING BRs:" << *PBI->getParent()
4182  << "AND: " << *BI->getParent());
4183 
4185 
4186  // If OtherDest *is* BB, then BB is a basic block with a single conditional
4187  // branch in it, where one edge (OtherDest) goes back to itself but the other
4188  // exits. We don't *know* that the program avoids the infinite loop
4189  // (even though that seems likely). If we do this xform naively, we'll end up
4190  // recursively unpeeling the loop. Since we know that (after the xform is
4191  // done) that the block *is* infinite if reached, we just make it an obviously
4192  // infinite loop with no cond branch.
4193  if (OtherDest == BB) {
4194  // Insert it at the end of the function, because it's either code,
4195  // or it won't matter if it's hot. :)
4196  BasicBlock *InfLoopBlock =
4197  BasicBlock::Create(BB->getContext(), "infloop", BB->getParent());
4198  BranchInst::Create(InfLoopBlock, InfLoopBlock);
4199  if (DTU)
4200  Updates.push_back({DominatorTree::Insert, InfLoopBlock, InfLoopBlock});
4201  OtherDest = InfLoopBlock;
4202  }
4203 
4204  LLVM_DEBUG(dbgs() << *PBI->getParent()->getParent());
4205 
4206  // BI may have other predecessors. Because of this, we leave
4207  // it alone, but modify PBI.
4208 
4209  // Make sure we get to CommonDest on True&True directions.
4210  Value *PBICond = PBI->getCondition();
4212  if (PBIOp)
4213  PBICond = Builder.CreateNot(PBICond, PBICond->getName() + ".not");
4214 
4215  Value *BICond = BI->getCondition();
4216  if (BIOp)
4217  BICond = Builder.CreateNot(BICond, BICond->getName() + ".not");
4218 
4219  // Merge the conditions.
4220  Value *Cond =
4221  createLogicalOp(Builder, Instruction::Or, PBICond, BICond, "brmerge");
4222 
4223  // Modify PBI to branch on the new condition to the new dests.
4224  PBI->setCondition(Cond);
4225  PBI->setSuccessor(0, CommonDest);
4226  PBI->setSuccessor(1, OtherDest);
4227 
4228  if (DTU) {
4229  Updates.push_back({DominatorTree::Insert, PBI->getParent(), OtherDest});
4230  Updates.push_back({DominatorTree::Delete, PBI->getParent(), RemovedDest});
4231 
4232  DTU->applyUpdates(Updates);
4233  }
4234 
4235  // Update branch weight for PBI.
4236  uint64_t PredTrueWeight, PredFalseWeight, SuccTrueWeight, SuccFalseWeight;
4237  uint64_t PredCommon, PredOther, SuccCommon, SuccOther;
4238  bool HasWeights =
4239  extractPredSuccWeights(PBI, BI, PredTrueWeight, PredFalseWeight,
4240  SuccTrueWeight, SuccFalseWeight);
4241  if (HasWeights) {
4242  PredCommon = PBIOp ? PredFalseWeight : PredTrueWeight;
4243  PredOther = PBIOp ? PredTrueWeight : PredFalseWeight;
4244  SuccCommon = BIOp ? SuccFalseWeight : SuccTrueWeight;
4245  SuccOther = BIOp ? SuccTrueWeight : SuccFalseWeight;
4246  // The weight to CommonDest should be PredCommon * SuccTotal +
4247  // PredOther * SuccCommon.
4248  // The weight to OtherDest should be PredOther * SuccOther.
4249  uint64_t NewWeights[2] = {PredCommon * (SuccCommon + SuccOther) +
4250  PredOther * SuccCommon,
4251  PredOther * SuccOther};
4252  // Halve the weights if any of them cannot fit in an uint32_t
4253  FitWeights(NewWeights);
4254 
4255  setBranchWeights(PBI, NewWeights[0], NewWeights[1]);
4256  }
4257 
4258  // OtherDest may have phi nodes. If so, add an entry from PBI's
4259  // block that are identical to the entries for BI's block.
4260  AddPredecessorToBlock(OtherDest, PBI->getParent(), BB);
4261 
4262  // We know that the CommonDest already had an edge from PBI to
4263  // it. If it has PHIs though, the PHIs may have different
4264  // entries for BB and PBI's BB. If so, insert a select to make
4265  // them agree.
4266  for (PHINode &PN : CommonDest->phis()) {
4267  Value *BIV = PN.getIncomingValueForBlock(BB);
4268  unsigned PBBIdx = PN.getBasicBlockIndex(PBI->getParent());
4269  Value *PBIV = PN.getIncomingValue(PBBIdx);
4270  if (BIV != PBIV) {
4271  // Insert a select in PBI to pick the right value.
4272  SelectInst *NV = cast<SelectInst>(
4273  Builder.CreateSelect(PBICond, PBIV, BIV, PBIV->getName() + ".mux"));
4274  PN.setIncomingValue(PBBIdx, NV);
4275  // Although the select has the same condition as PBI, the original branch
4276  // weights for PBI do not apply to the new select because the select's
4277  // 'logical' edges are incoming edges of the phi that is eliminated, not
4278  // the outgoing edges of PBI.
4279  if (HasWeights) {
4280  uint64_t PredCommon = PBIOp ? PredFalseWeight : PredTrueWeight;
4281  uint64_t PredOther = PBIOp ? PredTrueWeight : PredFalseWeight;
4282  uint64_t SuccCommon = BIOp ? SuccFalseWeight : SuccTrueWeight;
4283  uint64_t SuccOther = BIOp ? SuccTrueWeight : SuccFalseWeight;
4284  // The weight to PredCommonDest should be PredCommon * SuccTotal.
4285  // The weight to PredOtherDest should be PredOther * SuccCommon.
4286  uint64_t NewWeights[2] = {PredCommon * (SuccCommon + SuccOther),
4287  PredOther * SuccCommon};
4288 
4289  FitWeights(NewWeights);
4290 
4291  setBranchWeights(NV, NewWeights[0], NewWeights[1]);
4292  }
4293  }
4294  }
4295 
4296  LLVM_DEBUG(dbgs() << "INTO: " << *PBI->getParent());
4297  LLVM_DEBUG(dbgs() << *PBI->getParent()->getParent());
4298 
4299  // This basic block is probably dead. We know it has at least
4300  // one fewer predecessor.
4301  return true;
4302 }
4303 
4304 // Simplifies a terminator by replacing it with a branch to TrueBB if Cond is
4305 // true or to FalseBB if Cond is false.
4306 // Takes care of updating the successors and removing the old terminator.
4307 // Also makes sure not to introduce new successors by assuming that edges to
4308 // non-successor TrueBBs and FalseBBs aren't reachable.
4309 bool SimplifyCFGOpt::SimplifyTerminatorOnSelect(Instruction *OldTerm,
4310  Value *Cond, BasicBlock *TrueBB,
4311  BasicBlock *FalseBB,
4312  uint32_t TrueWeight,
4313  uint32_t FalseWeight) {
4314  auto *BB = OldTerm->getParent();
4315  // Remove any superfluous successor edges from the CFG.
4316  // First, figure out which successors to preserve.
4317  // If TrueBB and FalseBB are equal, only try to preserve one copy of that
4318  // successor.
4319  BasicBlock *KeepEdge1 = TrueBB;
4320  BasicBlock *KeepEdge2 = TrueBB != FalseBB ? FalseBB : nullptr;
4321 
4322  SmallSetVector<BasicBlock *, 2> RemovedSuccessors;
4323 
4324  // Then remove the rest.
4325  for (BasicBlock *Succ : successors(OldTerm)) {
4326  // Make sure only to keep exactly one copy of each edge.
4327  if (Succ == KeepEdge1)
4328  KeepEdge1 = nullptr;
4329  else if (Succ == KeepEdge2)
4330  KeepEdge2 = nullptr;
4331  else {
4332  Succ->removePredecessor(BB,
4333  /*KeepOneInputPHIs=*/true);
4334 
4335  if (Succ != TrueBB && Succ != FalseBB)
4336  RemovedSuccessors.insert(Succ);
4337  }
4338  }
4339 
4340  IRBuilder<> Builder(OldTerm);
4341  Builder.SetCurrentDebugLocation(OldTerm->getDebugLoc());
4342 
4343  // Insert an appropriate new terminator.
4344  if (!KeepEdge1 && !KeepEdge2) {
4345  if (TrueBB == FalseBB) {
4346  // We were only looking for one successor, and it was present.
4347  // Create an unconditional branch to it.
4348  Builder.CreateBr(TrueBB);
4349  } else {
4350  // We found both of the successors we were looking for.
4351  // Create a conditional branch sharing the condition of the select.
4352  BranchInst *NewBI = Builder.CreateCondBr(Cond, TrueBB, FalseBB);
4353  if (TrueWeight != FalseWeight)
4354  setBranchWeights(NewBI, TrueWeight, FalseWeight);
4355  }
4356  } else if (KeepEdge1 && (KeepEdge2 || TrueBB == FalseBB)) {
4357  // Neither of the selected blocks were successors, so this
4358  // terminator must be unreachable.
4359  new UnreachableInst(OldTerm->getContext(), OldTerm);
4360  } else {
4361  // One of the selected values was a successor, but the other wasn't.
4362  // Insert an unconditional branch to the one that was found;
4363  // the edge to the one that wasn't must be unreachable.
4364  if (!KeepEdge1) {
4365  // Only TrueBB was found.
4366  Builder.CreateBr(TrueBB);
4367  } else {
4368  // Only FalseBB was found.
4369  Builder.CreateBr(FalseBB);
4370  }
4371  }
4372 
4373  EraseTerminatorAndDCECond(OldTerm);
4374 
4375  if (DTU) {
4377  Updates.reserve(RemovedSuccessors.size());
4378  for (auto *RemovedSuccessor : RemovedSuccessors)
4379  Updates.push_back({DominatorTree::Delete, BB, RemovedSuccessor});
4380  DTU->applyUpdates(Updates);
4381  }
4382 
4383  return true;
4384 }
4385 
4386 // Replaces
4387 // (switch (select cond, X, Y)) on constant X, Y
4388 // with a branch - conditional if X and Y lead to distinct BBs,
4389 // unconditional otherwise.
4390 bool SimplifyCFGOpt::SimplifySwitchOnSelect(SwitchInst *SI,
4391  SelectInst *Select) {
4392  // Check for constant integer values in the select.
4393  ConstantInt *TrueVal = dyn_cast<ConstantInt>(Select->getTrueValue());
4394  ConstantInt *FalseVal = dyn_cast<ConstantInt>(Select->getFalseValue());
4395  if (!TrueVal || !FalseVal)
4396  return false;
4397 
4398  // Find the relevant condition and destinations.
4399  Value *Condition = Select->getCondition();
4400  BasicBlock *TrueBB = SI->findCaseValue(TrueVal)->getCaseSuccessor();
4401  BasicBlock *FalseBB = SI->findCaseValue(FalseVal)->getCaseSuccessor();
4402 
4403  // Get weight for TrueBB and FalseBB.
4404  uint32_t TrueWeight = 0, FalseWeight = 0;
4405  SmallVector<uint64_t, 8> Weights;
4406  bool HasWeights = HasBranchWeights(SI);
4407  if (HasWeights) {
4408  GetBranchWeights(SI, Weights);
4409  if (Weights.size() == 1 + SI->getNumCases()) {
4410  TrueWeight =
4411  (uint32_t)Weights[SI->findCaseValue(TrueVal)->getSuccessorIndex()];
4412  FalseWeight =
4413  (uint32_t)Weights[SI->findCaseValue(FalseVal)->getSuccessorIndex()];
4414  }
4415  }
4416 
4417  // Perform the actual simplification.
4418  return SimplifyTerminatorOnSelect(SI, Condition, TrueBB, FalseBB, TrueWeight,
4419  FalseWeight);
4420 }
4421 
4422 // Replaces
4423 // (indirectbr (select cond, blockaddress(@fn, BlockA),
4424 // blockaddress(@fn, BlockB)))
4425 // with
4426 // (br cond, BlockA, BlockB).
4427 bool SimplifyCFGOpt::SimplifyIndirectBrOnSelect(IndirectBrInst *IBI,
4428  SelectInst *SI) {
4429  // Check that both operands of the select are block addresses.
4430  BlockAddress *TBA = dyn_cast<BlockAddress>(SI->getTrueValue());
4431  BlockAddress *FBA = dyn_cast<BlockAddress>(SI->getFalseValue());
4432  if (!TBA || !FBA)
4433  return false;
4434 
4435  // Extract the actual blocks.
4436  BasicBlock *TrueBB = TBA->getBasicBlock();
4437  BasicBlock *FalseBB = FBA->getBasicBlock();
4438 
4439  // Perform the actual simplification.
4440  return SimplifyTerminatorOnSelect(IBI, SI->getCondition(), TrueBB, FalseBB, 0,
4441  0);
4442 }
4443 
4444 /// This is called when we find an icmp instruction
4445 /// (a seteq/setne with a constant) as the only instruction in a
4446 /// block that ends with an uncond branch. We are looking for a very specific
4447 /// pattern that occurs when "A == 1 || A == 2 || A == 3" gets simplified. In
4448 /// this case, we merge the first two "or's of icmp" into a switch, but then the
4449 /// default value goes to an uncond block with a seteq in it, we get something
4450 /// like:
4451 ///
4452 /// switch i8 %A, label %DEFAULT [ i8 1, label %end i8 2, label %end ]
4453 /// DEFAULT:
4454 /// %tmp = icmp eq i8 %A, 92
4455 /// br label %end
4456 /// end:
4457 /// ... = phi i1 [ true, %entry ], [ %tmp, %DEFAULT ], [ true, %entry ]
4458 ///
4459 /// We prefer to split the edge to 'end' so that there is a true/false entry to
4460 /// the PHI, merging the third icmp into the switch.
4461 bool SimplifyCFGOpt::tryToSimplifyUncondBranchWithICmpInIt(
4462  ICmpInst *ICI, IRBuilder<> &Builder) {
4463  BasicBlock *BB = ICI->getParent();
4464 
4465  // If the block has any PHIs in it or the icmp has multiple uses, it is too
4466  // complex.
4467  if (isa<PHINode>(BB->begin()) || !ICI->hasOneUse())
4468  return false;
4469 
4470  Value *V = ICI->getOperand(0);
4471  ConstantInt *Cst = cast<ConstantInt>(ICI->getOperand(1));
4472 
4473  // The pattern we're looking for is where our only predecessor is a switch on
4474  // 'V' and this block is the default case for the switch. In this case we can
4475  // fold the compared value into the switch to simplify things.
4476  BasicBlock *Pred = BB->getSinglePredecessor();
4477  if (!Pred || !isa<SwitchInst>(Pred->getTerminator()))
4478  return false;
4479 
4480  SwitchInst *SI = cast<SwitchInst>(Pred->getTerminator());
4481  if (SI->getCondition() != V)
4482  return false;
4483 
4484  // If BB is reachable on a non-default case, then we simply know the value of
4485  // V in this block. Substitute it and constant fold the icmp instruction
4486  // away.
4487  if (SI->getDefaultDest() != BB) {
4488  ConstantInt *VVal = SI->findCaseDest(BB);
4489  assert(VVal && "Should have a unique destination value");
4490  ICI->setOperand(0, VVal);
4491 
4492  if (Value *V = SimplifyInstruction(ICI, {DL, ICI})) {
4493  ICI->replaceAllUsesWith(V);
4494  ICI->eraseFromParent();
4495  }
4496  // BB is now empty, so it is likely to simplify away.
4497  return requestResimplify();
4498  }
4499 
4500  // Ok, the block is reachable from the default dest. If the constant we're
4501  // comparing exists in one of the other edges, then we can constant fold ICI
4502  // and zap it.
4503  if (SI->findCaseValue(Cst) != SI->case_default()) {
4504  Value *V;
4505  if (ICI->getPredicate() == ICmpInst::ICMP_EQ)
4506  V = ConstantInt::getFalse(BB->getContext());
4507  else
4508  V = ConstantInt::getTrue(BB->getContext());
4509 
4510  ICI->replaceAllUsesWith(V);
4511  ICI->eraseFromParent();
4512  // BB is now empty, so it is likely to simplify away.
4513  return requestResimplify();
4514  }
4515 
4516  // The use of the icmp has to be in the 'end' block, by the only PHI node in
4517  // the block.
4518  BasicBlock *SuccBlock = BB->getTerminator()->getSuccessor(0);
4519  PHINode *PHIUse = dyn_cast<PHINode>(ICI->user_back());
4520  if (PHIUse == nullptr || PHIUse != &SuccBlock->front() ||
4521  isa<PHINode>(++BasicBlock::iterator(PHIUse)))
4522  return false;
4523 
4524  // If the icmp is a SETEQ, then the default dest gets false, the new edge gets
4525  // true in the PHI.
4526  Constant *DefaultCst = ConstantInt::getTrue(BB->getContext());
4527  Constant *NewCst = ConstantInt::getFalse(BB->getContext());
4528 
4529  if (ICI->getPredicate() == ICmpInst::ICMP_EQ)
4530  std::swap(DefaultCst, NewCst);
4531 
4532  // Replace ICI (which is used by the PHI for the default value) with true or
4533  // false depending on if it is EQ or NE.
4534  ICI->replaceAllUsesWith(DefaultCst);
4535  ICI->eraseFromParent();
4536 
4538 
4539  // Okay, the switch goes to this block on a default value. Add an edge from
4540  // the switch to the merge point on the compared value.
4541  BasicBlock *NewBB =
4542  BasicBlock::Create(BB->getContext(), "switch.edge", BB->getParent(), BB);
4543  {
4545  auto W0 = SIW.getSuccessorWeight(0);
4547  if (W0) {
4548  NewW = ((uint64_t(*W0) + 1) >> 1);
4549  SIW.setSuccessorWeight(0, *NewW);
4550  }
4551  SIW.addCase(Cst, NewBB, NewW);
4552  if (DTU)
4553  Updates.push_back({DominatorTree::Insert, Pred, NewBB});
4554  }
4555 
4556  // NewBB branches to the phi block, add the uncond branch and the phi entry.
4557  Builder.SetInsertPoint(NewBB);
4558  Builder.SetCurrentDebugLocation(SI->getDebugLoc());
4559  Builder.CreateBr(SuccBlock);
4560  PHIUse->addIncoming(NewCst, NewBB);
4561  if (DTU) {
4562  Updates.push_back({DominatorTree::Insert, NewBB, SuccBlock});
4563  DTU->applyUpdates(Updates);
4564  }
4565  return true;
4566 }
4567 
4568 /// The specified branch is a conditional branch.
4569 /// Check to see if it is branching on an or/and chain of icmp instructions, and
4570 /// fold it into a switch instruction if so.
4571 bool SimplifyCFGOpt::SimplifyBranchOnICmpChain(BranchInst *BI,
4573  const DataLayout &DL) {
4574  Instruction *Cond = dyn_cast<Instruction>(BI->getCondition());
4575  if (!Cond)
4576  return false;
4577 
4578  // Change br (X == 0 | X == 1), T, F into a switch instruction.
4579  // If this is a bunch of seteq's or'd together, or if it's a bunch of
4580  // 'setne's and'ed together, collect them.
4581 
4582  // Try to gather values from a chain of and/or to be turned into a switch
4583  ConstantComparesGatherer ConstantCompare(Cond, DL);
4584  // Unpack the result
4585  SmallVectorImpl<ConstantInt *> &Values = ConstantCompare.Vals;
4586  Value *CompVal = ConstantCompare.CompValue;
4587  unsigned UsedICmps = ConstantCompare.UsedICmps;
4588  Value *ExtraCase = ConstantCompare.Extra;
4589 
4590  // If we didn't have a multiply compared value, fail.
4591  if (!CompVal)
4592  return false;
4593 
4594  // Avoid turning single icmps into a switch.
4595  if (UsedICmps <= 1)
4596  return false;
4597 
4598  bool TrueWhenEqual = match(Cond, m_LogicalOr(m_Value(), m_Value()));
4599 
4600  // There might be duplicate constants in the list, which the switch
4601  // instruction can't handle, remove them now.
4602  array_pod_sort(Values.begin(), Values.end(), ConstantIntSortPredicate);
4603  Values.erase(std::unique(Values.begin(), Values.end()), Values.end());
4604 
4605  // If Extra was used, we require at least two switch values to do the
4606  // transformation. A switch with one value is just a conditional branch.
4607  if (ExtraCase && Values.size() < 2)
4608  return false;
4609 
4610  // TODO: Preserve branch weight metadata, similarly to how
4611  // FoldValueComparisonIntoPredecessors preserves it.
4612 
4613  // Figure out which block is which destination.
4614  BasicBlock *DefaultBB = BI->getSuccessor(1);
4615  BasicBlock *EdgeBB = BI->getSuccessor(0);
4616  if (!TrueWhenEqual)
4617  std::swap(DefaultBB, EdgeBB);
4618 
4619  BasicBlock *BB = BI->getParent();
4620 
4621  LLVM_DEBUG(dbgs() << "Converting 'icmp' chain with " << Values.size()
4622  << " cases into SWITCH. BB is:\n"
4623  << *BB);
4624 
4626 
4627  // If there are any extra values that couldn't be folded into the switch
4628  // then we evaluate them with an explicit branch first. Split the block
4629  // right before the condbr to handle it.
4630  if (ExtraCase) {
4631  BasicBlock *NewBB = SplitBlock(BB, BI, DTU, /*LI=*/nullptr,
4632  /*MSSAU=*/nullptr, "switch.early.test");
4633 
4634  // Remove the uncond branch added to the old block.
4635  Instruction *OldTI = BB->getTerminator();
4636  Builder.SetInsertPoint(OldTI);
4637 
4638  // There can be an unintended UB if extra values are Poison. Before the
4639  // transformation, extra values may not be evaluated according to the
4640  // condition, and it will not raise UB. But after transformation, we are
4641  // evaluating extra values before checking the condition, and it will raise
4642  // UB. It can be solved by adding freeze instruction to extra values.
4643  AssumptionCache *AC = Options.AC;
4644 
4645  if (!isGuaranteedNotToBeUndefOrPoison(ExtraCase, AC, BI, nullptr))
4646  ExtraCase = Builder.CreateFreeze(ExtraCase);
4647 
4648  if (TrueWhenEqual)
4649  Builder.CreateCondBr(ExtraCase, EdgeBB, NewBB);
4650  else
4651  Builder.CreateCondBr(ExtraCase, NewBB, EdgeBB);
4652 
4653  OldTI->eraseFromParent();
4654 
4655  if (DTU)
4656  Updates.push_back({DominatorTree::Insert, BB, EdgeBB});
4657 
4658  // If there are PHI nodes in EdgeBB, then we need to add a new entry to them
4659  // for the edge we just added.
4660  AddPredecessorToBlock(EdgeBB, BB, NewBB);
4661 
4662  LLVM_DEBUG(dbgs() << " ** 'icmp' chain unhandled condition: " << *ExtraCase
4663  << "\nEXTRABB = " << *BB);
4664  BB = NewBB;
4665  }
4666 
4667  Builder.SetInsertPoint(BI);
4668  // Convert pointer to int before we switch.
4669  if (CompVal->getType()->isPointerTy()) {
4670  CompVal = Builder.CreatePtrToInt(
4671  CompVal, DL.getIntPtrType(CompVal->getType()), "magicptr");
4672  }
4673 
4674  // Create the new switch instruction now.
4675  SwitchInst *New = Builder.CreateSwitch(CompVal, DefaultBB, Values.size());
4676 
4677  // Add all of the 'cases' to the switch instruction.
4678  for (unsigned i = 0, e = Values.size(); i != e; ++i)
4679  New->addCase(Values[i], EdgeBB);
4680 
4681  // We added edges from PI to the EdgeBB. As such, if there were any
4682  // PHI nodes in EdgeBB, they need entries to be added corresponding to
4683  // the number of edges added.
4684  for (BasicBlock::iterator BBI = EdgeBB->begin(); isa<PHINode>(BBI); ++BBI) {
4685  PHINode *PN = cast<PHINode>(BBI);
4686  Value *InVal = PN->getIncomingValueForBlock(BB);
4687  for (unsigned i = 0, e = Values.size() - 1; i != e; ++i)
4688  PN->addIncoming(InVal, BB);
4689  }
4690 
4691  // Erase the old branch instruction.
4693  if (DTU)
4694  DTU->applyUpdates(Updates);
4695 
4696  LLVM_DEBUG(dbgs() << " ** 'icmp' chain result is:\n" << *BB << '\n');
4697  return true;
4698 }
4699 
4700 bool SimplifyCFGOpt::simplifyResume(ResumeInst *RI, IRBuilder<> &Builder) {
4701  if (isa<PHINode>(RI->getValue()))
4702  return simplifyCommonResume(RI);
4703  else if (isa<LandingPadInst>(RI->getParent()->getFirstNonPHI()) &&
4704  RI->getValue() == RI->getParent()->getFirstNonPHI())
4705  // The resume must unwind the exception that caused control to branch here.
4706  return simplifySingleResume(RI);
4707 
4708  return false;
4709 }
4710 
4711 // Check if cleanup block is empty
4713  for (Instruction &I : R) {
4714  auto *II = dyn_cast<IntrinsicInst>(&I);
4715  if (!II)
4716  return false;
4717 
4718  Intrinsic::ID IntrinsicID = II->getIntrinsicID();
4719  switch (IntrinsicID) {
4720  case Intrinsic::dbg_declare:
4721  case Intrinsic::dbg_value:
4722  case Intrinsic::dbg_label:
4723  case Intrinsic::lifetime_end:
4724  break;
4725  default:
4726  return false;
4727  }
4728  }
4729  return true;
4730 }
4731 
4732 // Simplify resume that is shared by several landing pads (phi of landing pad).
4733 bool SimplifyCFGOpt::simplifyCommonResume(ResumeInst *RI) {
4734  BasicBlock *BB = RI->getParent();
4735 
4736  // Check that there are no other instructions except for debug and lifetime
4737  // intrinsics between the phi's and resume instruction.
4738  if (!isCleanupBlockEmpty(
4739  make_range(RI->getParent()->getFirstNonPHI(), BB->getTerminator())))
4740  return false;
4741 
4742  SmallSetVector<BasicBlock *, 4> TrivialUnwindBlocks;
4743  auto *PhiLPInst = cast<PHINode>(RI->getValue());
4744 
4745  // Check incoming blocks to see if any of them are trivial.
4746  for (unsigned Idx = 0, End = PhiLPInst->getNumIncomingValues(); Idx != End;
4747  Idx++) {
4748  auto *IncomingBB = PhiLPInst->getIncomingBlock(Idx);
4749  auto *IncomingValue = PhiLPInst->getIncomingValue(Idx);
4750 
4751  // If the block has other successors, we can not delete it because
4752  // it has other dependents.
4753  if (IncomingBB->getUniqueSuccessor() != BB)
4754  continue;
4755 
4756  auto *LandingPad = dyn_cast<LandingPadInst>(IncomingBB->getFirstNonPHI());
4757  // Not the landing pad that caused the control to branch here.
4758  if (IncomingValue != LandingPad)
4759  continue;
4760 
4761  if (isCleanupBlockEmpty(
4762  make_range(LandingPad->getNextNode(), IncomingBB->getTerminator())))
4763  TrivialUnwindBlocks.insert(IncomingBB);
4764  }
4765 
4766  // If no trivial unwind blocks, don't do any simplifications.
4767  if (TrivialUnwindBlocks.empty())
4768  return false;
4769 
4770  // Turn all invokes that unwind here into calls.
4771  for (auto *TrivialBB : TrivialUnwindBlocks) {
4772  // Blocks that will be simplified should be removed from the phi node.
4773  // Note there could be multiple edges to the resume block, and we need
4774  // to remove them all.
4775  while (PhiLPInst->getBasicBlockIndex(TrivialBB) != -1)
4776  BB->removePredecessor(TrivialBB, true);
4777 
4778  for (BasicBlock *Pred :
4780  removeUnwindEdge(Pred, DTU);
4781  ++NumInvokes;
4782  }
4783 
4784  // In each SimplifyCFG run, only the current processed block can be erased.
4785  // Otherwise, it will break the iteration of SimplifyCFG pass. So instead
4786  // of erasing TrivialBB, we only remove the branch to the common resume
4787  // block so that we can later erase the resume block since it has no
4788  // predecessors.
4789  TrivialBB->getTerminator()->eraseFromParent();
4790  new UnreachableInst(RI->getContext(), TrivialBB);
4791  if (DTU)
4792  DTU->applyUpdates({{DominatorTree::Delete, TrivialBB, BB}});
4793  }
4794 
4795  // Delete the resume block if all its predecessors have been removed.
4796  if (pred_empty(BB))
4797  DeleteDeadBlock(BB, DTU);
4798 
4799  return !TrivialUnwindBlocks.empty();
4800 }
4801 
4802 // Simplify resume that is only used by a single (non-phi) landing pad.
4803 bool SimplifyCFGOpt::simplifySingleResume(ResumeInst *RI) {
4804  BasicBlock *BB = RI->getParent();
4805  auto *LPInst = cast<LandingPadInst>(BB->getFirstNonPHI());
4806  assert(RI->getValue() == LPInst &&
4807  "Resume must unwind the exception that caused control to here");
4808 
4809  // Check that there are no other instructions except for debug intrinsics.
4810  if (!isCleanupBlockEmpty(
4811  make_range<Instruction *>(LPInst->getNextNode(), RI)))
4812  return false;
4813 
4814  // Turn all invokes that unwind here into calls and delete the basic block.
4816  removeUnwindEdge(Pred, DTU);
4817  ++NumInvokes;
4818  }
4819 
4820  // The landingpad is now unreachable. Zap it.
4821  DeleteDeadBlock(BB, DTU);
4822  return true;
4823 }
4824 
4826  // If this is a trivial cleanup pad that executes no instructions, it can be
4827  // eliminated. If the cleanup pad continues to the caller, any predecessor
4828  // that is an EH pad will be updated to continue to the caller and any
4829  // predecessor that terminates with an invoke instruction will have its invoke
4830  // instruction converted to a call instruction. If the cleanup pad being
4831  // simplified does not continue to the caller, each predecessor will be
4832  // updated to continue to the unwind destination of the cleanup pad being
4833  // simplified.
4834  BasicBlock *BB = RI->getParent();
4835  CleanupPadInst *CPInst = RI->getCleanupPad();
4836  if (CPInst->getParent() != BB)
4837  // This isn't an empty cleanup.
4838  return false;
4839 
4840  // We cannot kill the pad if it has multiple uses. This typically arises
4841  // from unreachable basic blocks.
4842  if (!CPInst->hasOneUse())
4843  return false;
4844 
4845  // Check that there are no other instructions except for benign intrinsics.
4846  if (!isCleanupBlockEmpty(
4847  make_range<Instruction *>(CPInst->getNextNode(), RI)))
4848  return false;
4849 
4850  // If the cleanup return we are simplifying unwinds to the caller, this will
4851  // set UnwindDest to nullptr.
4852  BasicBlock *UnwindDest = RI->getUnwindDest();
4853  Instruction *DestEHPad = UnwindDest ? UnwindDest->getFirstNonPHI() : nullptr;
4854 
4855  // We're about to remove BB from the control flow. Before we do, sink any
4856  // PHINodes into the unwind destination. Doing this before changing the
4857  // control flow avoids some potentially slow checks, since we can currently
4858  // be certain that UnwindDest and BB have no common predecessors (since they
4859  // are both EH pads).
4860  if (UnwindDest) {
4861  // First, go through the PHI nodes in UnwindDest and update any nodes that
4862  // reference the block we are removing
4863  for (PHINode &DestPN : UnwindDest->phis()) {
4864  int Idx = DestPN.getBasicBlockIndex(BB);
4865  // Since BB unwinds to UnwindDest, it has to be in the PHI node.
4866  assert(Idx != -1);
4867  // This PHI node has an incoming value that corresponds to a control
4868  // path through the cleanup pad we are removing. If the incoming
4869  // value is in the cleanup pad, it must be a PHINode (because we
4870  // verified above that the block is otherwise empty). Otherwise, the
4871  // value is either a constant or a value that dominates the cleanup
4872  // pad being removed.
4873  //
4874  // Because BB and UnwindDest are both EH pads, all of their
4875  // predecessors must unwind to these blocks, and since no instruction
4876  // can have multiple unwind destinations, there will be no overlap in
4877  // incoming blocks between SrcPN and DestPN.
4878  Value *SrcVal = DestPN.getIncomingValue(Idx);
4879  PHINode *SrcPN = dyn_cast<PHINode>(SrcVal);
4880 
4881  bool NeedPHITranslation = SrcPN && SrcPN->getParent() == BB;
4882  for (auto *Pred : predecessors(BB)) {
4883  Value *Incoming =
4884  NeedPHITranslation ? SrcPN->getIncomingValueForBlock(Pred) : SrcVal;
4885  DestPN.addIncoming(Incoming, Pred);
4886  }
4887  }
4888 
4889  // Sink any remaining PHI nodes directly into UnwindDest.
4890  Instruction *InsertPt = DestEHPad;
4891  for (PHINode &PN : make_early_inc_range(BB->phis())) {
4892  if (PN.use_empty() || !PN.isUsedOutsideOfBlock(BB))
4893  // If the PHI node has no uses or all of its uses are in this basic
4894  // block (meaning they are debug or lifetime intrinsics), just leave
4895  // it. It will be erased when we erase BB below.
4896  continue;
4897 
4898  // Otherwise, sink this PHI node into UnwindDest.
4899  // Any predecessors to UnwindDest which are not already represented
4900  // must be back edges which inherit the value from the path through
4901  // BB. In this case, the PHI value must reference itself.
4902  for (auto *pred : predecessors(UnwindDest))
4903  if (pred != BB)
4904  PN.addIncoming(&PN, pred);
4905  PN.moveBefore(InsertPt);
4906  // Also, add a dummy incoming value for the original BB itself,
4907  // so that the PHI is well-formed until we drop said predecessor.
4909  }
4910  }
4911 
4912  std::vector<DominatorTree::UpdateType> Updates;
4913 
4914  // We use make_early_inc_range here because we will remove all predecessors.
4916  if (UnwindDest == nullptr) {
4917  if (DTU) {
4918  DTU->applyUpdates(Updates);
4919  Updates.clear();
4920  }
4921  removeUnwindEdge(PredBB, DTU);
4922  ++NumInvokes;
4923  } else {
4924  BB->removePredecessor(PredBB);
4925  Instruction *TI = PredBB->getTerminator();
4926  TI->replaceUsesOfWith(BB, UnwindDest);
4927  if (DTU) {
4928  Updates.push_back({DominatorTree::Insert, PredBB, UnwindDest});
4929  Updates.push_back({DominatorTree::Delete, PredBB, BB});
4930  }
4931  }
4932  }
4933 
4934  if (DTU)
4935  DTU->applyUpdates(Updates);
4936 
4937  DeleteDeadBlock(BB, DTU);
4938 
4939  return true;
4940 }
4941 
4942 // Try to merge two cleanuppads together.
4944  // Skip any cleanuprets which unwind to caller, there is nothing to merge
4945  // with.
4946  BasicBlock *UnwindDest = RI->getUnwindDest();
4947  if (!UnwindDest)
4948  return false;
4949 
4950  // This cleanupret isn't the only predecessor of this cleanuppad, it wouldn't
4951  // be safe to merge without code duplication.
4952  if (UnwindDest->getSinglePredecessor() != RI->getParent())
4953  return false;
4954 
4955  // Verify that our cleanuppad's unwind destination is another cleanuppad.
4956  auto *SuccessorCleanupPad = dyn_cast<CleanupPadInst>(&UnwindDest->front());
4957  if (!SuccessorCleanupPad)
4958  return false;
4959 
4960  CleanupPadInst *PredecessorCleanupPad = RI->getCleanupPad();
4961  // Replace any uses of the successor cleanupad with the predecessor pad
4962  // The only cleanuppad uses should be this cleanupret, it's cleanupret and
4963  // funclet bundle operands.
4964  SuccessorCleanupPad->replaceAllUsesWith(PredecessorCleanupPad);
4965  // Remove the old cleanuppad.
4966  SuccessorCleanupPad->eraseFromParent();
4967  // Now, we simply replace the cleanupret with a branch to the unwind
4968  // destination.
4969  BranchInst::Create(UnwindDest, RI->getParent());
4970  RI->eraseFromParent();
4971 
4972  return true;
4973 }
4974 
4975 bool SimplifyCFGOpt::simplifyCleanupReturn(CleanupReturnInst *RI) {
4976  // It is possible to transiantly have an undef cleanuppad operand because we
4977  // have deleted some, but not all, dead blocks.
4978  // Eventually, this block will be deleted.
4979  if (isa<UndefValue>(RI->getOperand(0)))
4980  return false;
4981 
4982  if (mergeCleanupPad(RI))
4983  return true;
4984 
4985  if (removeEmptyCleanup(RI, DTU))
4986  return true;
4987 
4988  return false;
4989 }
4990 
4991 // WARNING: keep in sync with InstCombinerImpl::visitUnreachableInst()!
4992 bool SimplifyCFGOpt::simplifyUnreachable(UnreachableInst *UI) {
4993  BasicBlock *BB = UI->getParent();
4994 
4995  bool Changed = false;
4996 
4997  // If there are any instructions immediately before the unreachable that can
4998  // be removed, do so.
4999  while (UI->getIterator() != BB->begin()) {
5000  BasicBlock::iterator BBI = UI->getIterator();
5001  --BBI;
5002 
5004  break; // Can not drop any more instructions. We're done here.
5005  // Otherwise, this instruction can be freely erased,
5006  // even if it is not side-effect free.
5007 
5008  // Note that deleting EH's here is in fact okay, although it involves a bit
5009  // of subtle reasoning. If this inst is an EH, all the predecessors of this
5010  // block will be the unwind edges of Invoke/CatchSwitch/CleanupReturn,
5011  // and we can therefore guarantee this block will be erased.
5012 
5013  // Delete this instruction (any uses are guaranteed to be dead)
5014  BBI->replaceAllUsesWith(PoisonValue::get(BBI->getType()));
5015  BBI->eraseFromParent();
5016  Changed = true;
5017  }
5018 
5019  // If the unreachable instruction is the first in the block, take a gander
5020  // at all of the predecessors of this instruction, and simplify them.
5021  if (&BB->front() != UI)
5022  return Changed;
5023 
5024  std::vector<DominatorTree::UpdateType> Updates;
5025 
5027  for (unsigned i = 0, e = Preds.size(); i != e; ++i) {
5028  auto *Predecessor = Preds[i];
5029  Instruction *TI = Predecessor->getTerminator();
5030  IRBuilder<> Builder(TI);
5031  if (auto *BI = dyn_cast<BranchInst>(TI)) {
5032  // We could either have a proper unconditional branch,
5033  // or a degenerate conditional branch with matching destinations.
5034  if (all_of(BI->successors(),
5035  [BB](auto *Successor) { return Successor == BB; })) {
5036  new UnreachableInst(TI->getContext(), TI);
5037  TI->eraseFromParent();
5038  Changed = true;
5039  } else {
5040  assert(BI->isConditional() && "Can't get here with an uncond branch.");
5041  Value* Cond = BI->getCondition();
5042  assert(BI->getSuccessor(0) != BI->getSuccessor(1) &&
5043  "The destinations are guaranteed to be different here.");
5044  if (BI->getSuccessor(0) == BB) {
5045  Builder.CreateAssumption(Builder.CreateNot(Cond));
5046  Builder.CreateBr(BI->getSuccessor(1));
5047  } else {
5048  assert(BI->getSuccessor(1) == BB && "Incorrect CFG");
5049  Builder.CreateAssumption(Cond);
5050  Builder.CreateBr(BI->getSuccessor(0));
5051  }
5053  Changed = true;
5054  }
5055  if (DTU)
5056  Updates.push_back({DominatorTree::Delete, Predecessor, BB});
5057  } else if (auto *SI = dyn_cast<SwitchInst>(TI)) {
5059  for (auto i = SU->case_begin(), e = SU->case_end(); i != e;) {
5060  if (i->getCaseSuccessor() != BB) {
5061  ++i;
5062  continue;
5063  }
5064  BB->removePredecessor(SU->getParent());
5065  i = SU.removeCase(i);
5066  e = SU->case_end();
5067  Changed = true;
5068  }
5069  // Note that the default destination can't be removed!
5070  if (DTU && SI->getDefaultDest() != BB)
5071  Updates.push_back({DominatorTree::Delete, Predecessor, BB});
5072  } else if (auto *II = dyn_cast<InvokeInst>(TI)) {
5073  if (II->getUnwindDest() == BB) {
5074  if (DTU) {
5075  DTU->applyUpdates(Updates);
5076  Updates.clear();
5077  }
5078  removeUnwindEdge(TI->getParent(), DTU);
5079  Changed = true;
5080  }
5081  } else if (auto *CSI = dyn_cast<CatchSwitchInst>(TI)) {
5082  if (CSI->getUnwindDest() == BB) {
5083  if (DTU) {
5084  DTU->applyUpdates(Updates);
5085  Updates.clear();
5086  }
5087  removeUnwindEdge(TI->getParent(), DTU);
5088  Changed = true;
5089  continue;
5090  }
5091 
5092  for (CatchSwitchInst::handler_iterator I = CSI->handler_begin(),
5093  E = CSI->handler_end();
5094  I != E; ++I) {
5095  if (*I == BB) {
5096  CSI->removeHandler(I);
5097  --I;
5098  --E;
5099  Changed = true;
5100  }
5101  }
5102  if (DTU)
5103  Updates.push_back({DominatorTree::Delete, Predecessor, BB});
5104  if (CSI->getNumHandlers() == 0) {
5105  if (CSI->hasUnwindDest()) {
5106  // Redirect all predecessors of the block containing CatchSwitchInst
5107  // to instead branch to the CatchSwitchInst's unwind destination.
5108  if (DTU) {
5109  for (auto *PredecessorOfPredecessor : predecessors(Predecessor)) {
5110  Updates.push_back({DominatorTree::Insert,
5111  PredecessorOfPredecessor,
5112  CSI->getUnwindDest()});
5113  Updates.push_back({DominatorTree::Delete,
5114  PredecessorOfPredecessor, Predecessor});
5115  }
5116  }
5117  Predecessor->replaceAllUsesWith(CSI->getUnwindDest());
5118  } else {
5119  // Rewrite all preds to unwind to caller (or from invoke to call).
5120  if (DTU) {
5121  DTU->applyUpdates(Updates);
5122  Updates.clear();
5123  }
5124  SmallVector<BasicBlock *, 8> EHPreds(predecessors(Predecessor));
5125  for (BasicBlock *EHPred : EHPreds)
5126  removeUnwindEdge(EHPred, DTU);
5127  }
5128  // The catchswitch is no longer reachable.
5129  new UnreachableInst(CSI->getContext(), CSI);
5130  CSI->eraseFromParent();
5131  Changed = true;
5132  }
5133  } else if (auto *CRI = dyn_cast<CleanupReturnInst>(TI)) {
5134  (void)CRI;
5135  assert(CRI->hasUnwindDest() && CRI->getUnwindDest() == BB &&
5136  "Expected to always have an unwind to BB.");
5137  if (DTU)
5138  Updates.push_back({DominatorTree::Delete, Predecessor, BB});
5139  new UnreachableInst(TI->getContext(), TI);
5140  TI->eraseFromParent();
5141  Changed = true;
5142  }
5143  }
5144 
5145  if (DTU)
5146  DTU->applyUpdates(Updates);
5147 
5148  // If this block is now dead, remove it.
5149  if (pred_empty(BB) && BB != &BB->getParent()->getEntryBlock()) {
5150  DeleteDeadBlock(BB, DTU);
5151  return true;
5152  }
5153 
5154  return Changed;
5155 }
5156 
5158  assert(Cases.size() >= 1);
5159 
5160  array_pod_sort(Cases.begin(), Cases.end(), ConstantIntSortPredicate);
5161  for (size_t I = 1, E = Cases.size(); I != E; ++I) {
5162  if (Cases[I - 1]->getValue() != Cases[I]->getValue() + 1)
5163  return false;
5164  }
5165  return true;
5166 }
5167 
5169  DomTreeUpdater *DTU) {
5170  LLVM_DEBUG(dbgs() << "SimplifyCFG: switch default is dead.\n");
5171  auto *BB = Switch->getParent();
5172  auto *OrigDefaultBlock = Switch->getDefaultDest();
5173  OrigDefaultBlock->removePredecessor(BB);
5174  BasicBlock *NewDefaultBlock = BasicBlock::Create(
5175  BB->getContext(), BB->getName() + ".unreachabledefault", BB->getParent(),
5176  OrigDefaultBlock);
5177  new UnreachableInst(Switch->getContext(), NewDefaultBlock);
5178  Switch->setDefaultDest(&*NewDefaultBlock);
5179  if (DTU) {
5181  Updates.push_back({DominatorTree::Insert, BB, &*NewDefaultBlock});
5182  if (!is_contained(successors(BB), OrigDefaultBlock))
5183  Updates.push_back({DominatorTree::Delete, BB, &*OrigDefaultBlock});
5184  DTU->applyUpdates(Updates);
5185  }
5186 }
5187 
5188 /// Turn a switch with two reachable destinations into an integer range
5189 /// comparison and branch.
5190 bool SimplifyCFGOpt::TurnSwitchRangeIntoICmp(SwitchInst *SI,
5191  IRBuilder<> &Builder) {
5192  assert(SI->getNumCases() > 1 && "Degenerate switch?");
5193 
5194  bool HasDefault =
5195  !isa<UnreachableInst>(SI->getDefaultDest()->getFirstNonPHIOrDbg());
5196 
5197  auto *BB = SI->getParent();
5198 
5199  // Partition the cases into two sets with different destinations.
5200  BasicBlock *DestA = HasDefault ? SI->getDefaultDest() : nullptr;
5201  BasicBlock *DestB = nullptr;
5204 
5205  for (auto Case : SI->cases()) {
5206  BasicBlock *Dest = Case.getCaseSuccessor();
5207  if (!DestA)
5208  DestA = Dest;
5209  if (Dest == DestA) {
5210  CasesA.push_back(Case.getCaseValue());
5211  continue;
5212  }
5213  if (!DestB)
5214  DestB = Dest;
5215  if (Dest == DestB) {
5216  CasesB.push_back(Case.getCaseValue());
5217  continue;
5218  }
5219  return false; // More than two destinations.
5220  }
5221 
5222  assert(DestA && DestB &&
5223  "Single-destination switch should have been folded.");
5224  assert(DestA != DestB);
5225  assert(DestB != SI->getDefaultDest());
5226  assert(!CasesB.empty() && "There must be non-default cases.");
5227  assert(!CasesA.empty() || HasDefault);
5228 
5229  // Figure out if one of the sets of cases form a contiguous range.
5230  SmallVectorImpl<ConstantInt *> *ContiguousCases = nullptr;
5231  BasicBlock *ContiguousDest = nullptr;
5232  BasicBlock *OtherDest = nullptr;
5233  if (!CasesA.empty() && CasesAreContiguous(CasesA)) {
5234  ContiguousCases = &CasesA;
5235  ContiguousDest = DestA;
5236  OtherDest = DestB;
5237  } else if (CasesAreContiguous(CasesB)) {
5238  ContiguousCases = &CasesB;
5239  ContiguousDest = DestB;
5240  OtherDest = DestA;
5241  } else
5242  return false;
5243 
5244  // Start building the compare and branch.
5245 
5246  Constant *Offset = ConstantExpr::getNeg(ContiguousCases->back());
5247  Constant *NumCases =
5248  ConstantInt::get(Offset->getType(), ContiguousCases->size());
5249 
5250  Value *Sub = SI->getCondition();
5251  if (!Offset->isNullValue())
5252  Sub = Builder.CreateAdd(Sub, Offset, Sub->getName() + ".off");
5253 
5254  Value *Cmp;
5255  // If NumCases overflowed, then all possible values jump to the successor.
5256  if (NumCases->isNullValue() && !ContiguousCases->empty())
5257  Cmp = ConstantInt::getTrue(SI->getContext());
5258  else
5259  Cmp = Builder.CreateICmpULT(Sub, NumCases, "switch");
5260  BranchInst *NewBI = Builder.CreateCondBr(Cmp, ContiguousDest, OtherDest);
5261 
5262  // Update weight for the newly-created conditional branch.
5263  if (HasBranchWeights(SI)) {
5264  SmallVector<uint64_t, 8> Weights;
5265  GetBranchWeights(SI, Weights);
5266  if (Weights.size() == 1 + SI->getNumCases()) {
5267  uint64_t TrueWeight = 0;
5268  uint64_t FalseWeight = 0;
5269  for (size_t I = 0, E = Weights.size(); I != E; ++I) {
5270  if (SI->getSuccessor(I) == ContiguousDest)
5271  TrueWeight += Weights[I];
5272  else
5273  FalseWeight += Weights[I];
5274  }
5275  while (TrueWeight > UINT32_MAX || FalseWeight > UINT32_MAX) {
5276  TrueWeight /= 2;
5277  FalseWeight /= 2;
5278  }
5279  setBranchWeights(NewBI, TrueWeight, FalseWeight);
5280  }
5281  }
5282 
5283  // Prune obsolete incoming values off the successors' PHI nodes.
5284  for (auto BBI = ContiguousDest->begin(); isa<PHINode>(BBI); ++BBI) {
5285  unsigned PreviousEdges = ContiguousCases->size();
5286  if (ContiguousDest == SI->getDefaultDest())
5287  ++PreviousEdges;
5288  for (unsigned I = 0, E = PreviousEdges - 1; I != E; ++I)
5289  cast<PHINode>(BBI)->removeIncomingValue(SI->getParent());
5290  }
5291  for (auto BBI = OtherDest->begin(); isa<PHINode>(BBI); ++BBI) {
5292  unsigned PreviousEdges = SI->getNumCases() - ContiguousCases->size();
5293  if (OtherDest == SI->getDefaultDest())
5294  ++PreviousEdges;
5295  for (unsigned I = 0, E = PreviousEdges - 1; I != E; ++I)
5296  cast<PHINode>(BBI)->removeIncomingValue(SI->getParent());
5297  }
5298 
5299  // Clean up the default block - it may have phis or other instructions before
5300  // the unreachable terminator.
5301  if (!HasDefault)
5303 
5304  auto *UnreachableDefault = SI->getDefaultDest();
5305 
5306  // Drop the switch.
5307  SI->eraseFromParent();
5308 
5309  if (!HasDefault && DTU)
5310  DTU->applyUpdates({{DominatorTree::Delete, BB, UnreachableDefault}});
5311 
5312  return true;
5313 }
5314 
5315 /// Compute masked bits for the condition of a switch
5316 /// and use it to remove dead cases.
5318  AssumptionCache *AC,
5319  const DataLayout &DL) {
5320  Value *Cond = SI->getCondition();
5321  KnownBits Known = computeKnownBits(Cond, DL, 0, AC, SI);
5322 
5323  // We can also eliminate cases by determining that their values are outside of
5324  // the limited range of the condition based on how many significant (non-sign)
5325  // bits are in the condition value.
5326  unsigned MaxSignificantBitsInCond =
5327  ComputeMaxSignificantBits(Cond, DL, 0, AC, SI);
5328 
5329  // Gather dead cases.
5331  SmallDenseMap<BasicBlock *, int, 8> NumPerSuccessorCases;
5332  SmallVector<BasicBlock *, 8> UniqueSuccessors;
5333  for (auto &Case : SI->cases()) {
5334  auto *Successor = Case.getCaseSuccessor();
5335  if (DTU) {
5336  if (!NumPerSuccessorCases.