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,
95 
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 /// Check whether this is a potentially trapping constant.
386 static bool canTrap(const Value *V) {
387  if (auto *C = dyn_cast<Constant>(V))
388  return C->canTrap();
389  return false;
390 }
391 
392 /// If we have a merge point of an "if condition" as accepted above,
393 /// return true if the specified value dominates the block. We
394 /// don't handle the true generality of domination here, just a special case
395 /// which works well enough for us.
396 ///
397 /// If AggressiveInsts is non-null, and if V does not dominate BB, we check to
398 /// see if V (which must be an instruction) and its recursive operands
399 /// that do not dominate BB have a combined cost lower than Budget and
400 /// are non-trapping. If both are true, the instruction is inserted into the
401 /// set and true is returned.
402 ///
403 /// The cost for most non-trapping instructions is defined as 1 except for
404 /// Select whose cost is 2.
405 ///
406 /// After this function returns, Cost is increased by the cost of
407 /// V plus its non-dominating operands. If that cost is greater than
408 /// Budget, false is returned and Cost is undefined.
410  SmallPtrSetImpl<Instruction *> &AggressiveInsts,
411  InstructionCost &Cost,
412  InstructionCost Budget,
413  const TargetTransformInfo &TTI,
414  unsigned Depth = 0) {
415  // It is possible to hit a zero-cost cycle (phi/gep instructions for example),
416  // so limit the recursion depth.
417  // TODO: While this recursion limit does prevent pathological behavior, it
418  // would be better to track visited instructions to avoid cycles.
419  if (Depth == MaxSpeculationDepth)
420  return false;
421 
422  Instruction *I = dyn_cast<Instruction>(V);
423  if (!I) {
424  // Non-instructions all dominate instructions, but not all constantexprs
425  // can be executed unconditionally.
426  return !canTrap(V);
427  }
428  BasicBlock *PBB = I->getParent();
429 
430  // We don't want to allow weird loops that might have the "if condition" in
431  // the bottom of this block.
432  if (PBB == BB)
433  return false;
434 
435  // If this instruction is defined in a block that contains an unconditional
436  // branch to BB, then it must be in the 'conditional' part of the "if
437  // statement". If not, it definitely dominates the region.
438  BranchInst *BI = dyn_cast<BranchInst>(PBB->getTerminator());
439  if (!BI || BI->isConditional() || BI->getSuccessor(0) != BB)
440  return true;
441 
442  // If we have seen this instruction before, don't count it again.
443  if (AggressiveInsts.count(I))
444  return true;
445 
446  // Okay, it looks like the instruction IS in the "condition". Check to
447  // see if it's a cheap instruction to unconditionally compute, and if it
448  // only uses stuff defined outside of the condition. If so, hoist it out.
450  return false;
451 
452  Cost += computeSpeculationCost(I, TTI);
453 
454  // Allow exactly one instruction to be speculated regardless of its cost
455  // (as long as it is safe to do so).
456  // This is intended to flatten the CFG even if the instruction is a division
457  // or other expensive operation. The speculation of an expensive instruction
458  // is expected to be undone in CodeGenPrepare if the speculation has not
459  // enabled further IR optimizations.
460  if (Cost > Budget &&
461  (!SpeculateOneExpensiveInst || !AggressiveInsts.empty() || Depth > 0 ||
462  !Cost.isValid()))
463  return false;
464 
465  // Okay, we can only really hoist these out if their operands do
466  // not take us over the cost threshold.
467  for (Use &Op : I->operands())
468  if (!dominatesMergePoint(Op, BB, AggressiveInsts, Cost, Budget, TTI,
469  Depth + 1))
470  return false;
471  // Okay, it's safe to do this! Remember this instruction.
472  AggressiveInsts.insert(I);
473  return true;
474 }
475 
476 /// Extract ConstantInt from value, looking through IntToPtr
477 /// and PointerNullValue. Return NULL if value is not a constant int.
479  // Normal constant int.
480  ConstantInt *CI = dyn_cast<ConstantInt>(V);
481  if (CI || !isa<Constant>(V) || !V->getType()->isPointerTy())
482  return CI;
483 
484  // This is some kind of pointer constant. Turn it into a pointer-sized
485  // ConstantInt if possible.
486  IntegerType *PtrTy = cast<IntegerType>(DL.getIntPtrType(V->getType()));
487 
488  // Null pointer means 0, see SelectionDAGBuilder::getValue(const Value*).
489  if (isa<ConstantPointerNull>(V))
490  return ConstantInt::get(PtrTy, 0);
491 
492  // IntToPtr const int.
493  if (ConstantExpr *CE = dyn_cast<ConstantExpr>(V))
494  if (CE->getOpcode() == Instruction::IntToPtr)
495  if (ConstantInt *CI = dyn_cast<ConstantInt>(CE->getOperand(0))) {
496  // The constant is very likely to have the right type already.
497  if (CI->getType() == PtrTy)
498  return CI;
499  else
500  return cast<ConstantInt>(
501  ConstantExpr::getIntegerCast(CI, PtrTy, /*isSigned=*/false));
502  }
503  return nullptr;
504 }
505 
506 namespace {
507 
508 /// Given a chain of or (||) or and (&&) comparison of a value against a
509 /// constant, this will try to recover the information required for a switch
510 /// structure.
511 /// It will depth-first traverse the chain of comparison, seeking for patterns
512 /// like %a == 12 or %a < 4 and combine them to produce a set of integer
513 /// representing the different cases for the switch.
514 /// Note that if the chain is composed of '||' it will build the set of elements
515 /// that matches the comparisons (i.e. any of this value validate the chain)
516 /// while for a chain of '&&' it will build the set elements that make the test
517 /// fail.
518 struct ConstantComparesGatherer {
519  const DataLayout &DL;
520 
521  /// Value found for the switch comparison
522  Value *CompValue = nullptr;
523 
524  /// Extra clause to be checked before the switch
525  Value *Extra = nullptr;
526 
527  /// Set of integers to match in switch
529 
530  /// Number of comparisons matched in the and/or chain
531  unsigned UsedICmps = 0;
532 
533  /// Construct and compute the result for the comparison instruction Cond
534  ConstantComparesGatherer(Instruction *Cond, const DataLayout &DL) : DL(DL) {
535  gather(Cond);
536  }
537 
538  ConstantComparesGatherer(const ConstantComparesGatherer &) = delete;
539  ConstantComparesGatherer &
540  operator=(const ConstantComparesGatherer &) = delete;
541 
542 private:
543  /// Try to set the current value used for the comparison, it succeeds only if
544  /// it wasn't set before or if the new value is the same as the old one
545  bool setValueOnce(Value *NewVal) {
546  if (CompValue && CompValue != NewVal)
547  return false;
548  CompValue = NewVal;
549  return (CompValue != nullptr);
550  }
551 
552  /// Try to match Instruction "I" as a comparison against a constant and
553  /// populates the array Vals with the set of values that match (or do not
554  /// match depending on isEQ).
555  /// Return false on failure. On success, the Value the comparison matched
556  /// against is placed in CompValue.
557  /// If CompValue is already set, the function is expected to fail if a match
558  /// is found but the value compared to is different.
559  bool matchInstruction(Instruction *I, bool isEQ) {
560  // If this is an icmp against a constant, handle this as one of the cases.
561  ICmpInst *ICI;
562  ConstantInt *C;
563  if (!((ICI = dyn_cast<ICmpInst>(I)) &&
564  (C = GetConstantInt(I->getOperand(1), DL)))) {
565  return false;
566  }
567 
568  Value *RHSVal;
569  const APInt *RHSC;
570 
571  // Pattern match a special case
572  // (x & ~2^z) == y --> x == y || x == y|2^z
573  // This undoes a transformation done by instcombine to fuse 2 compares.
574  if (ICI->getPredicate() == (isEQ ? ICmpInst::ICMP_EQ : ICmpInst::ICMP_NE)) {
575  // It's a little bit hard to see why the following transformations are
576  // correct. Here is a CVC3 program to verify them for 64-bit values:
577 
578  /*
579  ONE : BITVECTOR(64) = BVZEROEXTEND(0bin1, 63);
580  x : BITVECTOR(64);
581  y : BITVECTOR(64);
582  z : BITVECTOR(64);
583  mask : BITVECTOR(64) = BVSHL(ONE, z);
584  QUERY( (y & ~mask = y) =>
585  ((x & ~mask = y) <=> (x = y OR x = (y | mask)))
586  );
587  QUERY( (y | mask = y) =>
588  ((x | mask = y) <=> (x = y OR x = (y & ~mask)))
589  );
590  */
591 
592  // Please note that each pattern must be a dual implication (<--> or
593  // iff). One directional implication can create spurious matches. If the
594  // implication is only one-way, an unsatisfiable condition on the left
595  // side can imply a satisfiable condition on the right side. Dual
596  // implication ensures that satisfiable conditions are transformed to
597  // other satisfiable conditions and unsatisfiable conditions are
598  // transformed to other unsatisfiable conditions.
599 
600  // Here is a concrete example of a unsatisfiable condition on the left
601  // implying a satisfiable condition on the right:
602  //
603  // mask = (1 << z)
604  // (x & ~mask) == y --> (x == y || x == (y | mask))
605  //
606  // Substituting y = 3, z = 0 yields:
607  // (x & -2) == 3 --> (x == 3 || x == 2)
608 
609  // Pattern match a special case:
610  /*
611  QUERY( (y & ~mask = y) =>
612  ((x & ~mask = y) <=> (x = y OR x = (y | mask)))
613  );
614  */
615  if (match(ICI->getOperand(0),
616  m_And(m_Value(RHSVal), m_APInt(RHSC)))) {
617  APInt Mask = ~*RHSC;
618  if (Mask.isPowerOf2() && (C->getValue() & ~Mask) == C->getValue()) {
619  // If we already have a value for the switch, it has to match!
620  if (!setValueOnce(RHSVal))
621  return false;
622 
623  Vals.push_back(C);
624  Vals.push_back(
625  ConstantInt::get(C->getContext(),
626  C->getValue() | Mask));
627  UsedICmps++;
628  return true;
629  }
630  }
631 
632  // Pattern match a special case:
633  /*
634  QUERY( (y | mask = y) =>
635  ((x | mask = y) <=> (x = y OR x = (y & ~mask)))
636  );
637  */
638  if (match(ICI->getOperand(0),
639  m_Or(m_Value(RHSVal), m_APInt(RHSC)))) {
640  APInt Mask = *RHSC;
641  if (Mask.isPowerOf2() && (C->getValue() | Mask) == C->getValue()) {
642  // If we already have a value for the switch, it has to match!
643  if (!setValueOnce(RHSVal))
644  return false;
645 
646  Vals.push_back(C);
647  Vals.push_back(ConstantInt::get(C->getContext(),
648  C->getValue() & ~Mask));
649  UsedICmps++;
650  return true;
651  }
652  }
653 
654  // If we already have a value for the switch, it has to match!
655  if (!setValueOnce(ICI->getOperand(0)))
656  return false;
657 
658  UsedICmps++;
659  Vals.push_back(C);
660  return ICI->getOperand(0);
661  }
662 
663  // If we have "x ult 3", for example, then we can add 0,1,2 to the set.
664  ConstantRange Span =
666 
667  // Shift the range if the compare is fed by an add. This is the range
668  // compare idiom as emitted by instcombine.
669  Value *CandidateVal = I->getOperand(0);
670  if (match(I->getOperand(0), m_Add(m_Value(RHSVal), m_APInt(RHSC)))) {
671  Span = Span.subtract(*RHSC);
672  CandidateVal = RHSVal;
673  }
674 
675  // If this is an and/!= check, then we are looking to build the set of
676  // value that *don't* pass the and chain. I.e. to turn "x ugt 2" into
677  // x != 0 && x != 1.
678  if (!isEQ)
679  Span = Span.inverse();
680 
681  // If there are a ton of values, we don't want to make a ginormous switch.
682  if (Span.isSizeLargerThan(8) || Span.isEmptySet()) {
683  return false;
684  }
685 
686  // If we already have a value for the switch, it has to match!
687  if (!setValueOnce(CandidateVal))
688  return false;
689 
690  // Add all values from the range to the set
691  for (APInt Tmp = Span.getLower(); Tmp != Span.getUpper(); ++Tmp)
692  Vals.push_back(ConstantInt::get(I->getContext(), Tmp));
693 
694  UsedICmps++;
695  return true;
696  }
697 
698  /// Given a potentially 'or'd or 'and'd together collection of icmp
699  /// eq/ne/lt/gt instructions that compare a value against a constant, extract
700  /// the value being compared, and stick the list constants into the Vals
701  /// vector.
702  /// One "Extra" case is allowed to differ from the other.
703  void gather(Value *V) {
704  bool isEQ = match(V, m_LogicalOr(m_Value(), m_Value()));
705 
706  // Keep a stack (SmallVector for efficiency) for depth-first traversal
708  SmallPtrSet<Value *, 8> Visited;
709 
710  // Initialize
711  Visited.insert(V);
712  DFT.push_back(V);
713 
714  while (!DFT.empty()) {
715  V = DFT.pop_back_val();
716 
717  if (Instruction *I = dyn_cast<Instruction>(V)) {
718  // If it is a || (or && depending on isEQ), process the operands.
719  Value *Op0, *Op1;
720  if (isEQ ? match(I, m_LogicalOr(m_Value(Op0), m_Value(Op1)))
721  : match(I, m_LogicalAnd(m_Value(Op0), m_Value(Op1)))) {
722  if (Visited.insert(Op1).second)
723  DFT.push_back(Op1);
724  if (Visited.insert(Op0).second)
725  DFT.push_back(Op0);
726 
727  continue;
728  }
729 
730  // Try to match the current instruction
731  if (matchInstruction(I, isEQ))
732  // Match succeed, continue the loop
733  continue;
734  }
735 
736  // One element of the sequence of || (or &&) could not be match as a
737  // comparison against the same value as the others.
738  // We allow only one "Extra" case to be checked before the switch
739  if (!Extra) {
740  Extra = V;
741  continue;
742  }
743  // Failed to parse a proper sequence, abort now
744  CompValue = nullptr;
745  break;
746  }
747  }
748 };
749 
750 } // end anonymous namespace
751 
753  MemorySSAUpdater *MSSAU = nullptr) {
754  Instruction *Cond = nullptr;
755  if (SwitchInst *SI = dyn_cast<SwitchInst>(TI)) {
756  Cond = dyn_cast<Instruction>(SI->getCondition());
757  } else if (BranchInst *BI = dyn_cast<BranchInst>(TI)) {
758  if (BI->isConditional())
759  Cond = dyn_cast<Instruction>(BI->getCondition());
760  } else if (IndirectBrInst *IBI = dyn_cast<IndirectBrInst>(TI)) {
761  Cond = dyn_cast<Instruction>(IBI->getAddress());
762  }
763 
764  TI->eraseFromParent();
765  if (Cond)
767 }
768 
769 /// Return true if the specified terminator checks
770 /// to see if a value is equal to constant integer value.
771 Value *SimplifyCFGOpt::isValueEqualityComparison(Instruction *TI) {
772  Value *CV = nullptr;
773  if (SwitchInst *SI = dyn_cast<SwitchInst>(TI)) {
774  // Do not permit merging of large switch instructions into their
775  // predecessors unless there is only one predecessor.
776  if (!SI->getParent()->hasNPredecessorsOrMore(128 / SI->getNumSuccessors()))
777  CV = SI->getCondition();
778  } else if (BranchInst *BI = dyn_cast<BranchInst>(TI))
779  if (BI->isConditional() && BI->getCondition()->hasOneUse())
780  if (ICmpInst *ICI = dyn_cast<ICmpInst>(BI->getCondition())) {
781  if (ICI->isEquality() && GetConstantInt(ICI->getOperand(1), DL))
782  CV = ICI->getOperand(0);
783  }
784 
785  // Unwrap any lossless ptrtoint cast.
786  if (CV) {
787  if (PtrToIntInst *PTII = dyn_cast<PtrToIntInst>(CV)) {
788  Value *Ptr = PTII->getPointerOperand();
789  if (PTII->getType() == DL.getIntPtrType(Ptr->getType()))
790  CV = Ptr;
791  }
792  }
793  return CV;
794 }
795 
796 /// Given a value comparison instruction,
797 /// decode all of the 'cases' that it represents and return the 'default' block.
798 BasicBlock *SimplifyCFGOpt::GetValueEqualityComparisonCases(
799  Instruction *TI, std::vector<ValueEqualityComparisonCase> &Cases) {
800  if (SwitchInst *SI = dyn_cast<SwitchInst>(TI)) {
801  Cases.reserve(SI->getNumCases());
802  for (auto Case : SI->cases())
803  Cases.push_back(ValueEqualityComparisonCase(Case.getCaseValue(),
804  Case.getCaseSuccessor()));
805  return SI->getDefaultDest();
806  }
807 
808  BranchInst *BI = cast<BranchInst>(TI);
809  ICmpInst *ICI = cast<ICmpInst>(BI->getCondition());
810  BasicBlock *Succ = BI->getSuccessor(ICI->getPredicate() == ICmpInst::ICMP_NE);
811  Cases.push_back(ValueEqualityComparisonCase(
812  GetConstantInt(ICI->getOperand(1), DL), Succ));
813  return BI->getSuccessor(ICI->getPredicate() == ICmpInst::ICMP_EQ);
814 }
815 
816 /// Given a vector of bb/value pairs, remove any entries
817 /// in the list that match the specified block.
818 static void
820  std::vector<ValueEqualityComparisonCase> &Cases) {
821  llvm::erase_value(Cases, BB);
822 }
823 
824 /// Return true if there are any keys in C1 that exist in C2 as well.
825 static bool ValuesOverlap(std::vector<ValueEqualityComparisonCase> &C1,
826  std::vector<ValueEqualityComparisonCase> &C2) {
827  std::vector<ValueEqualityComparisonCase> *V1 = &C1, *V2 = &C2;
828 
829  // Make V1 be smaller than V2.
830  if (V1->size() > V2->size())
831  std::swap(V1, V2);
832 
833  if (V1->empty())
834  return false;
835  if (V1->size() == 1) {
836  // Just scan V2.
837  ConstantInt *TheVal = (*V1)[0].Value;
838  for (unsigned i = 0, e = V2->size(); i != e; ++i)
839  if (TheVal == (*V2)[i].Value)
840  return true;
841  }
842 
843  // Otherwise, just sort both lists and compare element by element.
844  array_pod_sort(V1->begin(), V1->end());
845  array_pod_sort(V2->begin(), V2->end());
846  unsigned i1 = 0, i2 = 0, e1 = V1->size(), e2 = V2->size();
847  while (i1 != e1 && i2 != e2) {
848  if ((*V1)[i1].Value == (*V2)[i2].Value)
849  return true;
850  if ((*V1)[i1].Value < (*V2)[i2].Value)
851  ++i1;
852  else
853  ++i2;
854  }
855  return false;
856 }
857 
858 // Set branch weights on SwitchInst. This sets the metadata if there is at
859 // least one non-zero weight.
861  // Check that there is at least one non-zero weight. Otherwise, pass
862  // nullptr to setMetadata which will erase the existing metadata.
863  MDNode *N = nullptr;
864  if (llvm::any_of(Weights, [](uint32_t W) { return W != 0; }))
865  N = MDBuilder(SI->getParent()->getContext()).createBranchWeights(Weights);
866  SI->setMetadata(LLVMContext::MD_prof, N);
867 }
868 
869 // Similar to the above, but for branch and select instructions that take
870 // exactly 2 weights.
871 static void setBranchWeights(Instruction *I, uint32_t TrueWeight,
872  uint32_t FalseWeight) {
873  assert(isa<BranchInst>(I) || isa<SelectInst>(I));
874  // Check that there is at least one non-zero weight. Otherwise, pass
875  // nullptr to setMetadata which will erase the existing metadata.
876  MDNode *N = nullptr;
877  if (TrueWeight || FalseWeight)
878  N = MDBuilder(I->getParent()->getContext())
879  .createBranchWeights(TrueWeight, FalseWeight);
880  I->setMetadata(LLVMContext::MD_prof, N);
881 }
882 
883 /// If TI is known to be a terminator instruction and its block is known to
884 /// only have a single predecessor block, check to see if that predecessor is
885 /// also a value comparison with the same value, and if that comparison
886 /// determines the outcome of this comparison. If so, simplify TI. This does a
887 /// very limited form of jump threading.
888 bool SimplifyCFGOpt::SimplifyEqualityComparisonWithOnlyPredecessor(
889  Instruction *TI, BasicBlock *Pred, IRBuilder<> &Builder) {
890  Value *PredVal = isValueEqualityComparison(Pred->getTerminator());
891  if (!PredVal)
892  return false; // Not a value comparison in predecessor.
893 
894  Value *ThisVal = isValueEqualityComparison(TI);
895  assert(ThisVal && "This isn't a value comparison!!");
896  if (ThisVal != PredVal)
897  return false; // Different predicates.
898 
899  // TODO: Preserve branch weight metadata, similarly to how
900  // FoldValueComparisonIntoPredecessors preserves it.
901 
902  // Find out information about when control will move from Pred to TI's block.
903  std::vector<ValueEqualityComparisonCase> PredCases;
904  BasicBlock *PredDef =
905  GetValueEqualityComparisonCases(Pred->getTerminator(), PredCases);
906  EliminateBlockCases(PredDef, PredCases); // Remove default from cases.
907 
908  // Find information about how control leaves this block.
909  std::vector<ValueEqualityComparisonCase> ThisCases;
910  BasicBlock *ThisDef = GetValueEqualityComparisonCases(TI, ThisCases);
911  EliminateBlockCases(ThisDef, ThisCases); // Remove default from cases.
912 
913  // If TI's block is the default block from Pred's comparison, potentially
914  // simplify TI based on this knowledge.
915  if (PredDef == TI->getParent()) {
916  // If we are here, we know that the value is none of those cases listed in
917  // PredCases. If there are any cases in ThisCases that are in PredCases, we
918  // can simplify TI.
919  if (!ValuesOverlap(PredCases, ThisCases))
920  return false;
921 
922  if (isa<BranchInst>(TI)) {
923  // Okay, one of the successors of this condbr is dead. Convert it to a
924  // uncond br.
925  assert(ThisCases.size() == 1 && "Branch can only have one case!");
926  // Insert the new branch.
927  Instruction *NI = Builder.CreateBr(ThisDef);
928  (void)NI;
929 
930  // Remove PHI node entries for the dead edge.
931  ThisCases[0].Dest->removePredecessor(PredDef);
932 
933  LLVM_DEBUG(dbgs() << "Threading pred instr: " << *Pred->getTerminator()
934  << "Through successor TI: " << *TI << "Leaving: " << *NI
935  << "\n");
936 
938 
939  if (DTU)
940  DTU->applyUpdates(
941  {{DominatorTree::Delete, PredDef, ThisCases[0].Dest}});
942 
943  return true;
944  }
945 
946  SwitchInstProfUpdateWrapper SI = *cast<SwitchInst>(TI);
947  // Okay, TI has cases that are statically dead, prune them away.
948  SmallPtrSet<Constant *, 16> DeadCases;
949  for (unsigned i = 0, e = PredCases.size(); i != e; ++i)
950  DeadCases.insert(PredCases[i].Value);
951 
952  LLVM_DEBUG(dbgs() << "Threading pred instr: " << *Pred->getTerminator()
953  << "Through successor TI: " << *TI);
954 
955  SmallDenseMap<BasicBlock *, int, 8> NumPerSuccessorCases;
956  for (SwitchInst::CaseIt i = SI->case_end(), e = SI->case_begin(); i != e;) {
957  --i;
958  auto *Successor = i->getCaseSuccessor();
959  if (DTU)
960  ++NumPerSuccessorCases[Successor];
961  if (DeadCases.count(i->getCaseValue())) {
962  Successor->removePredecessor(PredDef);
963  SI.removeCase(i);
964  if (DTU)
965  --NumPerSuccessorCases[Successor];
966  }
967  }
968 
969  if (DTU) {
970  std::vector<DominatorTree::UpdateType> Updates;
971  for (const std::pair<BasicBlock *, int> &I : NumPerSuccessorCases)
972  if (I.second == 0)
973  Updates.push_back({DominatorTree::Delete, PredDef, I.first});
974  DTU->applyUpdates(Updates);
975  }
976 
977  LLVM_DEBUG(dbgs() << "Leaving: " << *TI << "\n");
978  return true;
979  }
980 
981  // Otherwise, TI's block must correspond to some matched value. Find out
982  // which value (or set of values) this is.
983  ConstantInt *TIV = nullptr;
984  BasicBlock *TIBB = TI->getParent();
985  for (unsigned i = 0, e = PredCases.size(); i != e; ++i)
986  if (PredCases[i].Dest == TIBB) {
987  if (TIV)
988  return false; // Cannot handle multiple values coming to this block.
989  TIV = PredCases[i].Value;
990  }
991  assert(TIV && "No edge from pred to succ?");
992 
993  // Okay, we found the one constant that our value can be if we get into TI's
994  // BB. Find out which successor will unconditionally be branched to.
995  BasicBlock *TheRealDest = nullptr;
996  for (unsigned i = 0, e = ThisCases.size(); i != e; ++i)
997  if (ThisCases[i].Value == TIV) {
998  TheRealDest = ThisCases[i].Dest;
999  break;
1000  }
1001 
1002  // If not handled by any explicit cases, it is handled by the default case.
1003  if (!TheRealDest)
1004  TheRealDest = ThisDef;
1005 
1006  SmallPtrSet<BasicBlock *, 2> RemovedSuccs;
1007 
1008  // Remove PHI node entries for dead edges.
1009  BasicBlock *CheckEdge = TheRealDest;
1010  for (BasicBlock *Succ : successors(TIBB))
1011  if (Succ != CheckEdge) {
1012  if (Succ != TheRealDest)
1013  RemovedSuccs.insert(Succ);
1014  Succ->removePredecessor(TIBB);
1015  } else
1016  CheckEdge = nullptr;
1017 
1018  // Insert the new branch.
1019  Instruction *NI = Builder.CreateBr(TheRealDest);
1020  (void)NI;
1021 
1022  LLVM_DEBUG(dbgs() << "Threading pred instr: " << *Pred->getTerminator()
1023  << "Through successor TI: " << *TI << "Leaving: " << *NI
1024  << "\n");
1025 
1027  if (DTU) {
1029  Updates.reserve(RemovedSuccs.size());
1030  for (auto *RemovedSucc : RemovedSuccs)
1031  Updates.push_back({DominatorTree::Delete, TIBB, RemovedSucc});
1032  DTU->applyUpdates(Updates);
1033  }
1034  return true;
1035 }
1036 
1037 namespace {
1038 
1039 /// This class implements a stable ordering of constant
1040 /// integers that does not depend on their address. This is important for
1041 /// applications that sort ConstantInt's to ensure uniqueness.
1042 struct ConstantIntOrdering {
1043  bool operator()(const ConstantInt *LHS, const ConstantInt *RHS) const {
1044  return LHS->getValue().ult(RHS->getValue());
1045  }
1046 };
1047 
1048 } // end anonymous namespace
1049 
1051  ConstantInt *const *P2) {
1052  const ConstantInt *LHS = *P1;
1053  const ConstantInt *RHS = *P2;
1054  if (LHS == RHS)
1055  return 0;
1056  return LHS->getValue().ult(RHS->getValue()) ? 1 : -1;
1057 }
1058 
1059 static inline bool HasBranchWeights(const Instruction *I) {
1060  MDNode *ProfMD = I->getMetadata(LLVMContext::MD_prof);
1061  if (ProfMD && ProfMD->getOperand(0))
1062  if (MDString *MDS = dyn_cast<MDString>(ProfMD->getOperand(0)))
1063  return MDS->getString().equals("branch_weights");
1064 
1065  return false;
1066 }
1067 
1068 /// Get Weights of a given terminator, the default weight is at the front
1069 /// of the vector. If TI is a conditional eq, we need to swap the branch-weight
1070 /// metadata.
1072  SmallVectorImpl<uint64_t> &Weights) {
1073  MDNode *MD = TI->getMetadata(LLVMContext::MD_prof);
1074  assert(MD);
1075  for (unsigned i = 1, e = MD->getNumOperands(); i < e; ++i) {
1076  ConstantInt *CI = mdconst::extract<ConstantInt>(MD->getOperand(i));
1077  Weights.push_back(CI->getValue().getZExtValue());
1078  }
1079 
1080  // If TI is a conditional eq, the default case is the false case,
1081  // and the corresponding branch-weight data is at index 2. We swap the
1082  // default weight to be the first entry.
1083  if (BranchInst *BI = dyn_cast<BranchInst>(TI)) {
1084  assert(Weights.size() == 2);
1085  ICmpInst *ICI = cast<ICmpInst>(BI->getCondition());
1086  if (ICI->getPredicate() == ICmpInst::ICMP_EQ)
1087  std::swap(Weights.front(), Weights.back());
1088  }
1089 }
1090 
1091 /// Keep halving the weights until all can fit in uint32_t.
1093  uint64_t Max = *std::max_element(Weights.begin(), Weights.end());
1094  if (Max > UINT_MAX) {
1095  unsigned Offset = 32 - countLeadingZeros(Max);
1096  for (uint64_t &I : Weights)
1097  I >>= Offset;
1098  }
1099 }
1100 
1102  BasicBlock *BB, BasicBlock *PredBlock, ValueToValueMapTy &VMap) {
1103  Instruction *PTI = PredBlock->getTerminator();
1104 
1105  // If we have bonus instructions, clone them into the predecessor block.
1106  // Note that there may be multiple predecessor blocks, so we cannot move
1107  // bonus instructions to a predecessor block.
1108  for (Instruction &BonusInst : *BB) {
1109  if (isa<DbgInfoIntrinsic>(BonusInst) || BonusInst.isTerminator())
1110  continue;
1111 
1112  Instruction *NewBonusInst = BonusInst.clone();
1113 
1114  if (PTI->getDebugLoc() != NewBonusInst->getDebugLoc()) {
1115  // Unless the instruction has the same !dbg location as the original
1116  // branch, drop it. When we fold the bonus instructions we want to make
1117  // sure we reset their debug locations in order to avoid stepping on
1118  // dead code caused by folding dead branches.
1119  NewBonusInst->setDebugLoc(DebugLoc());
1120  }
1121 
1122  RemapInstruction(NewBonusInst, VMap,
1124  VMap[&BonusInst] = NewBonusInst;
1125 
1126  // If we moved a load, we cannot any longer claim any knowledge about
1127  // its potential value. The previous information might have been valid
1128  // only given the branch precondition.
1129  // For an analogous reason, we must also drop all the metadata whose
1130  // semantics we don't understand. We *can* preserve !annotation, because
1131  // it is tied to the instruction itself, not the value or position.
1132  // Similarly strip attributes on call parameters that may cause UB in
1133  // location the call is moved to.
1135  LLVMContext::MD_annotation);
1136 
1137  PredBlock->getInstList().insert(PTI->getIterator(), NewBonusInst);
1138  NewBonusInst->takeName(&BonusInst);
1139  BonusInst.setName(NewBonusInst->getName() + ".old");
1140 
1141  // Update (liveout) uses of bonus instructions,
1142  // now that the bonus instruction has been cloned into predecessor.
1143  // Note that we expect to be in a block-closed SSA form for this to work!
1144  for (Use &U : make_early_inc_range(BonusInst.uses())) {
1145  auto *UI = cast<Instruction>(U.getUser());
1146  auto *PN = dyn_cast<PHINode>(UI);
1147  if (!PN) {
1148  assert(UI->getParent() == BB && BonusInst.comesBefore(UI) &&
1149  "If the user is not a PHI node, then it should be in the same "
1150  "block as, and come after, the original bonus instruction.");
1151  continue; // Keep using the original bonus instruction.
1152  }
1153  // Is this the block-closed SSA form PHI node?
1154  if (PN->getIncomingBlock(U) == BB)
1155  continue; // Great, keep using the original bonus instruction.
1156  // The only other alternative is an "use" when coming from
1157  // the predecessor block - here we should refer to the cloned bonus instr.
1158  assert(PN->getIncomingBlock(U) == PredBlock &&
1159  "Not in block-closed SSA form?");
1160  U.set(NewBonusInst);
1161  }
1162  }
1163 }
1164 
1165 bool SimplifyCFGOpt::PerformValueComparisonIntoPredecessorFolding(
1166  Instruction *TI, Value *&CV, Instruction *PTI, IRBuilder<> &Builder) {
1167  BasicBlock *BB = TI->getParent();
1168  BasicBlock *Pred = PTI->getParent();
1169 
1171 
1172  // Figure out which 'cases' to copy from SI to PSI.
1173  std::vector<ValueEqualityComparisonCase> BBCases;
1174  BasicBlock *BBDefault = GetValueEqualityComparisonCases(TI, BBCases);
1175 
1176  std::vector<ValueEqualityComparisonCase> PredCases;
1177  BasicBlock *PredDefault = GetValueEqualityComparisonCases(PTI, PredCases);
1178 
1179  // Based on whether the default edge from PTI goes to BB or not, fill in
1180  // PredCases and PredDefault with the new switch cases we would like to
1181  // build.
1183 
1184  // Update the branch weight metadata along the way
1185  SmallVector<uint64_t, 8> Weights;
1186  bool PredHasWeights = HasBranchWeights(PTI);
1187  bool SuccHasWeights = HasBranchWeights(TI);
1188 
1189  if (PredHasWeights) {
1190  GetBranchWeights(PTI, Weights);
1191  // branch-weight metadata is inconsistent here.
1192  if (Weights.size() != 1 + PredCases.size())
1193  PredHasWeights = SuccHasWeights = false;
1194  } else if (SuccHasWeights)
1195  // If there are no predecessor weights but there are successor weights,
1196  // populate Weights with 1, which will later be scaled to the sum of
1197  // successor's weights
1198  Weights.assign(1 + PredCases.size(), 1);
1199 
1200  SmallVector<uint64_t, 8> SuccWeights;
1201  if (SuccHasWeights) {
1202  GetBranchWeights(TI, SuccWeights);
1203  // branch-weight metadata is inconsistent here.
1204  if (SuccWeights.size() != 1 + BBCases.size())
1205  PredHasWeights = SuccHasWeights = false;
1206  } else if (PredHasWeights)
1207  SuccWeights.assign(1 + BBCases.size(), 1);
1208 
1209  if (PredDefault == BB) {
1210  // If this is the default destination from PTI, only the edges in TI
1211  // that don't occur in PTI, or that branch to BB will be activated.
1212  std::set<ConstantInt *, ConstantIntOrdering> PTIHandled;
1213  for (unsigned i = 0, e = PredCases.size(); i != e; ++i)
1214  if (PredCases[i].Dest != BB)
1215  PTIHandled.insert(PredCases[i].Value);
1216  else {
1217  // The default destination is BB, we don't need explicit targets.
1218  std::swap(PredCases[i], PredCases.back());
1219 
1220  if (PredHasWeights || SuccHasWeights) {
1221  // Increase weight for the default case.
1222  Weights[0] += Weights[i + 1];
1223  std::swap(Weights[i + 1], Weights.back());
1224  Weights.pop_back();
1225  }
1226 
1227  PredCases.pop_back();
1228  --i;
1229  --e;
1230  }
1231 
1232  // Reconstruct the new switch statement we will be building.
1233  if (PredDefault != BBDefault) {
1234  PredDefault->removePredecessor(Pred);
1235  if (DTU && PredDefault != BB)
1236  Updates.push_back({DominatorTree::Delete, Pred, PredDefault});
1237  PredDefault = BBDefault;
1238  ++NewSuccessors[BBDefault];
1239  }
1240 
1241  unsigned CasesFromPred = Weights.size();
1242  uint64_t ValidTotalSuccWeight = 0;
1243  for (unsigned i = 0, e = BBCases.size(); i != e; ++i)
1244  if (!PTIHandled.count(BBCases[i].Value) && BBCases[i].Dest != BBDefault) {
1245  PredCases.push_back(BBCases[i]);
1246  ++NewSuccessors[BBCases[i].Dest];
1247  if (SuccHasWeights || PredHasWeights) {
1248  // The default weight is at index 0, so weight for the ith case
1249  // should be at index i+1. Scale the cases from successor by
1250  // PredDefaultWeight (Weights[0]).
1251  Weights.push_back(Weights[0] * SuccWeights[i + 1]);
1252  ValidTotalSuccWeight += SuccWeights[i + 1];
1253  }
1254  }
1255 
1256  if (SuccHasWeights || PredHasWeights) {
1257  ValidTotalSuccWeight += SuccWeights[0];
1258  // Scale the cases from predecessor by ValidTotalSuccWeight.
1259  for (unsigned i = 1; i < CasesFromPred; ++i)
1260  Weights[i] *= ValidTotalSuccWeight;
1261  // Scale the default weight by SuccDefaultWeight (SuccWeights[0]).
1262  Weights[0] *= SuccWeights[0];
1263  }
1264  } else {
1265  // If this is not the default destination from PSI, only the edges
1266  // in SI that occur in PSI with a destination of BB will be
1267  // activated.
1268  std::set<ConstantInt *, ConstantIntOrdering> PTIHandled;
1269  std::map<ConstantInt *, uint64_t> WeightsForHandled;
1270  for (unsigned i = 0, e = PredCases.size(); i != e; ++i)
1271  if (PredCases[i].Dest == BB) {
1272  PTIHandled.insert(PredCases[i].Value);
1273 
1274  if (PredHasWeights || SuccHasWeights) {
1275  WeightsForHandled[PredCases[i].Value] = Weights[i + 1];
1276  std::swap(Weights[i + 1], Weights.back());
1277  Weights.pop_back();
1278  }
1279 
1280  std::swap(PredCases[i], PredCases.back());
1281  PredCases.pop_back();
1282  --i;
1283  --e;
1284  }
1285 
1286  // Okay, now we know which constants were sent to BB from the
1287  // predecessor. Figure out where they will all go now.
1288  for (unsigned i = 0, e = BBCases.size(); i != e; ++i)
1289  if (PTIHandled.count(BBCases[i].Value)) {
1290  // If this is one we are capable of getting...
1291  if (PredHasWeights || SuccHasWeights)
1292  Weights.push_back(WeightsForHandled[BBCases[i].Value]);
1293  PredCases.push_back(BBCases[i]);
1294  ++NewSuccessors[BBCases[i].Dest];
1295  PTIHandled.erase(BBCases[i].Value); // This constant is taken care of
1296  }
1297 
1298  // If there are any constants vectored to BB that TI doesn't handle,
1299  // they must go to the default destination of TI.
1300  for (ConstantInt *I : PTIHandled) {
1301  if (PredHasWeights || SuccHasWeights)
1302  Weights.push_back(WeightsForHandled[I]);
1303  PredCases.push_back(ValueEqualityComparisonCase(I, BBDefault));
1304  ++NewSuccessors[BBDefault];
1305  }
1306  }
1307 
1308  // Okay, at this point, we know which new successor Pred will get. Make
1309  // sure we update the number of entries in the PHI nodes for these
1310  // successors.
1311  SmallPtrSet<BasicBlock *, 2> SuccsOfPred;
1312  if (DTU) {
1313  SuccsOfPred = {succ_begin(Pred), succ_end(Pred)};
1314  Updates.reserve(Updates.size() + NewSuccessors.size());
1315  }
1316  for (const std::pair<BasicBlock *, int /*Num*/> &NewSuccessor :
1317  NewSuccessors) {
1318  for (auto I : seq(0, NewSuccessor.second)) {
1319  (void)I;
1320  AddPredecessorToBlock(NewSuccessor.first, Pred, BB);
1321  }
1322  if (DTU && !SuccsOfPred.contains(NewSuccessor.first))
1323  Updates.push_back({DominatorTree::Insert, Pred, NewSuccessor.first});
1324  }
1325 
1326  Builder.SetInsertPoint(PTI);
1327  // Convert pointer to int before we switch.
1328  if (CV->getType()->isPointerTy()) {
1329  CV =
1330  Builder.CreatePtrToInt(CV, DL.getIntPtrType(CV->getType()), "magicptr");
1331  }
1332 
1333  // Now that the successors are updated, create the new Switch instruction.
1334  SwitchInst *NewSI = Builder.CreateSwitch(CV, PredDefault, PredCases.size());
1335  NewSI->setDebugLoc(PTI->getDebugLoc());
1336  for (ValueEqualityComparisonCase &V : PredCases)
1337  NewSI->addCase(V.Value, V.Dest);
1338 
1339  if (PredHasWeights || SuccHasWeights) {
1340  // Halve the weights if any of them cannot fit in an uint32_t
1341  FitWeights(Weights);
1342 
1343  SmallVector<uint32_t, 8> MDWeights(Weights.begin(), Weights.end());
1344 
1345  setBranchWeights(NewSI, MDWeights);
1346  }
1347 
1349 
1350  // Okay, last check. If BB is still a successor of PSI, then we must
1351  // have an infinite loop case. If so, add an infinitely looping block
1352  // to handle the case to preserve the behavior of the code.
1353  BasicBlock *InfLoopBlock = nullptr;
1354  for (unsigned i = 0, e = NewSI->getNumSuccessors(); i != e; ++i)
1355  if (NewSI->getSuccessor(i) == BB) {
1356  if (!InfLoopBlock) {
1357  // Insert it at the end of the function, because it's either code,
1358  // or it won't matter if it's hot. :)
1359  InfLoopBlock =
1360  BasicBlock::Create(BB->getContext(), "infloop", BB->getParent());
1361  BranchInst::Create(InfLoopBlock, InfLoopBlock);
1362  if (DTU)
1363  Updates.push_back(
1364  {DominatorTree::Insert, InfLoopBlock, InfLoopBlock});
1365  }
1366  NewSI->setSuccessor(i, InfLoopBlock);
1367  }
1368 
1369  if (DTU) {
1370  if (InfLoopBlock)
1371  Updates.push_back({DominatorTree::Insert, Pred, InfLoopBlock});
1372 
1373  Updates.push_back({DominatorTree::Delete, Pred, BB});
1374 
1375  DTU->applyUpdates(Updates);
1376  }
1377 
1378  ++NumFoldValueComparisonIntoPredecessors;
1379  return true;
1380 }
1381 
1382 /// The specified terminator is a value equality comparison instruction
1383 /// (either a switch or a branch on "X == c").
1384 /// See if any of the predecessors of the terminator block are value comparisons
1385 /// on the same value. If so, and if safe to do so, fold them together.
1386 bool SimplifyCFGOpt::FoldValueComparisonIntoPredecessors(Instruction *TI,
1387  IRBuilder<> &Builder) {
1388  BasicBlock *BB = TI->getParent();
1389  Value *CV = isValueEqualityComparison(TI); // CondVal
1390  assert(CV && "Not a comparison?");
1391 
1392  bool Changed = false;
1393 
1395  while (!Preds.empty()) {
1396  BasicBlock *Pred = Preds.pop_back_val();
1397  Instruction *PTI = Pred->getTerminator();
1398 
1399  // Don't try to fold into itself.
1400  if (Pred == BB)
1401  continue;
1402 
1403  // See if the predecessor is a comparison with the same value.
1404  Value *PCV = isValueEqualityComparison(PTI); // PredCondVal
1405  if (PCV != CV)
1406  continue;
1407 
1409  if (!SafeToMergeTerminators(TI, PTI, &FailBlocks)) {
1410  for (auto *Succ : FailBlocks) {
1411  if (!SplitBlockPredecessors(Succ, TI->getParent(), ".fold.split", DTU))
1412  return false;
1413  }
1414  }
1415 
1416  PerformValueComparisonIntoPredecessorFolding(TI, CV, PTI, Builder);
1417  Changed = true;
1418  }
1419  return Changed;
1420 }
1421 
1422 // If we would need to insert a select that uses the value of this invoke
1423 // (comments in HoistThenElseCodeToIf explain why we would need to do this), we
1424 // can't hoist the invoke, as there is nowhere to put the select in this case.
1426  Instruction *I1, Instruction *I2) {
1427  for (BasicBlock *Succ : successors(BB1)) {
1428  for (const PHINode &PN : Succ->phis()) {
1429  Value *BB1V = PN.getIncomingValueForBlock(BB1);
1430  Value *BB2V = PN.getIncomingValueForBlock(BB2);
1431  if (BB1V != BB2V && (BB1V == I1 || BB2V == I2)) {
1432  return false;
1433  }
1434  }
1435  }
1436  return true;
1437 }
1438 
1439 static bool passingValueIsAlwaysUndefined(Value *V, Instruction *I, bool PtrValueMayBeModified = false);
1440 
1441 /// Given a conditional branch that goes to BB1 and BB2, hoist any common code
1442 /// in the two blocks up into the branch block. The caller of this function
1443 /// guarantees that BI's block dominates BB1 and BB2. If EqTermsOnly is given,
1444 /// only perform hoisting in case both blocks only contain a terminator. In that
1445 /// case, only the original BI will be replaced and selects for PHIs are added.
1446 bool SimplifyCFGOpt::HoistThenElseCodeToIf(BranchInst *BI,
1447  const TargetTransformInfo &TTI,
1448  bool EqTermsOnly) {
1449  // This does very trivial matching, with limited scanning, to find identical
1450  // instructions in the two blocks. In particular, we don't want to get into
1451  // O(M*N) situations here where M and N are the sizes of BB1 and BB2. As
1452  // such, we currently just scan for obviously identical instructions in an
1453  // identical order.
1454  BasicBlock *BB1 = BI->getSuccessor(0); // The true destination.
1455  BasicBlock *BB2 = BI->getSuccessor(1); // The false destination
1456 
1457  // If either of the blocks has it's address taken, then we can't do this fold,
1458  // because the code we'd hoist would no longer run when we jump into the block
1459  // by it's address.
1460  if (BB1->hasAddressTaken() || BB2->hasAddressTaken())
1461  return false;
1462 
1463  BasicBlock::iterator BB1_Itr = BB1->begin();
1464  BasicBlock::iterator BB2_Itr = BB2->begin();
1465 
1466  Instruction *I1 = &*BB1_Itr++, *I2 = &*BB2_Itr++;
1467  // Skip debug info if it is not identical.
1468  DbgInfoIntrinsic *DBI1 = dyn_cast<DbgInfoIntrinsic>(I1);
1469  DbgInfoIntrinsic *DBI2 = dyn_cast<DbgInfoIntrinsic>(I2);
1470  if (!DBI1 || !DBI2 || !DBI1->isIdenticalToWhenDefined(DBI2)) {
1471  while (isa<DbgInfoIntrinsic>(I1))
1472  I1 = &*BB1_Itr++;
1473  while (isa<DbgInfoIntrinsic>(I2))
1474  I2 = &*BB2_Itr++;
1475  }
1476  // FIXME: Can we define a safety predicate for CallBr?
1477  if (isa<PHINode>(I1) || !I1->isIdenticalToWhenDefined(I2) ||
1478  (isa<InvokeInst>(I1) && !isSafeToHoistInvoke(BB1, BB2, I1, I2)) ||
1479  isa<CallBrInst>(I1))
1480  return false;
1481 
1482  BasicBlock *BIParent = BI->getParent();
1483 
1484  bool Changed = false;
1485 
1486  auto _ = make_scope_exit([&]() {
1487  if (Changed)
1488  ++NumHoistCommonCode;
1489  });
1490 
1491  // Check if only hoisting terminators is allowed. This does not add new
1492  // instructions to the hoist location.
1493  if (EqTermsOnly) {
1494  // Skip any debug intrinsics, as they are free to hoist.
1495  auto *I1NonDbg = &*skipDebugIntrinsics(I1->getIterator());
1496  auto *I2NonDbg = &*skipDebugIntrinsics(I2->getIterator());
1497  if (!I1NonDbg->isIdenticalToWhenDefined(I2NonDbg))
1498  return false;
1499  if (!I1NonDbg->isTerminator())
1500  return false;
1501  // Now we know that we only need to hoist debug intrinsics and the
1502  // terminator. Let the loop below handle those 2 cases.
1503  }
1504 
1505  do {
1506  // If we are hoisting the terminator instruction, don't move one (making a
1507  // broken BB), instead clone it, and remove BI.
1508  if (I1->isTerminator())
1509  goto HoistTerminator;
1510 
1511  // If we're going to hoist a call, make sure that the two instructions we're
1512  // commoning/hoisting are both marked with musttail, or neither of them is
1513  // marked as such. Otherwise, we might end up in a situation where we hoist
1514  // from a block where the terminator is a `ret` to a block where the terminator
1515  // is a `br`, and `musttail` calls expect to be followed by a return.
1516  auto *C1 = dyn_cast<CallInst>(I1);
1517  auto *C2 = dyn_cast<CallInst>(I2);
1518  if (C1 && C2)
1519  if (C1->isMustTailCall() != C2->isMustTailCall())
1520  return Changed;
1521 
1523  return Changed;
1524 
1525  // If any of the two call sites has nomerge attribute, stop hoisting.
1526  if (const auto *CB1 = dyn_cast<CallBase>(I1))
1527  if (CB1->cannotMerge())
1528  return Changed;
1529  if (const auto *CB2 = dyn_cast<CallBase>(I2))
1530  if (CB2->cannotMerge())
1531  return Changed;
1532 
1533  if (isa<DbgInfoIntrinsic>(I1) || isa<DbgInfoIntrinsic>(I2)) {
1534  assert (isa<DbgInfoIntrinsic>(I1) && isa<DbgInfoIntrinsic>(I2));
1535  // The debug location is an integral part of a debug info intrinsic
1536  // and can't be separated from it or replaced. Instead of attempting
1537  // to merge locations, simply hoist both copies of the intrinsic.
1538  BIParent->getInstList().splice(BI->getIterator(),
1539  BB1->getInstList(), I1);
1540  BIParent->getInstList().splice(BI->getIterator(),
1541  BB2->getInstList(), I2);
1542  Changed = true;
1543  } else {
1544  // For a normal instruction, we just move one to right before the branch,
1545  // then replace all uses of the other with the first. Finally, we remove
1546  // the now redundant second instruction.
1547  BIParent->getInstList().splice(BI->getIterator(),
1548  BB1->getInstList(), I1);
1549  if (!I2->use_empty())
1550  I2->replaceAllUsesWith(I1);
1551  I1->andIRFlags(I2);
1552  unsigned KnownIDs[] = {LLVMContext::MD_tbaa,
1553  LLVMContext::MD_range,
1554  LLVMContext::MD_fpmath,
1555  LLVMContext::MD_invariant_load,
1556  LLVMContext::MD_nonnull,
1557  LLVMContext::MD_invariant_group,
1558  LLVMContext::MD_align,
1559  LLVMContext::MD_dereferenceable,
1560  LLVMContext::MD_dereferenceable_or_null,
1561  LLVMContext::MD_mem_parallel_loop_access,
1562  LLVMContext::MD_access_group,
1563  LLVMContext::MD_preserve_access_index};
1564  combineMetadata(I1, I2, KnownIDs, true);
1565 
1566  // I1 and I2 are being combined into a single instruction. Its debug
1567  // location is the merged locations of the original instructions.
1568  I1->applyMergedLocation(I1->getDebugLoc(), I2->getDebugLoc());
1569 
1570  I2->eraseFromParent();
1571  Changed = true;
1572  }
1573  ++NumHoistCommonInstrs;
1574 
1575  I1 = &*BB1_Itr++;
1576  I2 = &*BB2_Itr++;
1577  // Skip debug info if it is not identical.
1578  DbgInfoIntrinsic *DBI1 = dyn_cast<DbgInfoIntrinsic>(I1);
1579  DbgInfoIntrinsic *DBI2 = dyn_cast<DbgInfoIntrinsic>(I2);
1580  if (!DBI1 || !DBI2 || !DBI1->isIdenticalToWhenDefined(DBI2)) {
1581  while (isa<DbgInfoIntrinsic>(I1))
1582  I1 = &*BB1_Itr++;
1583  while (isa<DbgInfoIntrinsic>(I2))
1584  I2 = &*BB2_Itr++;
1585  }
1586  } while (I1->isIdenticalToWhenDefined(I2));
1587 
1588  return true;
1589 
1590 HoistTerminator:
1591  // It may not be possible to hoist an invoke.
1592  // FIXME: Can we define a safety predicate for CallBr?
1593  if (isa<InvokeInst>(I1) && !isSafeToHoistInvoke(BB1, BB2, I1, I2))
1594  return Changed;
1595 
1596  // TODO: callbr hoisting currently disabled pending further study.
1597  if (isa<CallBrInst>(I1))
1598  return Changed;
1599 
1600  for (BasicBlock *Succ : successors(BB1)) {
1601  for (PHINode &PN : Succ->phis()) {
1602  Value *BB1V = PN.getIncomingValueForBlock(BB1);
1603  Value *BB2V = PN.getIncomingValueForBlock(BB2);
1604  if (BB1V == BB2V)
1605  continue;
1606 
1607  // Check for passingValueIsAlwaysUndefined here because we would rather
1608  // eliminate undefined control flow then converting it to a select.
1609  if (passingValueIsAlwaysUndefined(BB1V, &PN) ||
1610  passingValueIsAlwaysUndefined(BB2V, &PN))
1611  return Changed;
1612 
1613  if (isa<ConstantExpr>(BB1V) && !isSafeToSpeculativelyExecute(BB1V))
1614  return Changed;
1615  if (isa<ConstantExpr>(BB2V) && !isSafeToSpeculativelyExecute(BB2V))
1616  return Changed;
1617  }
1618  }
1619 
1620  // Okay, it is safe to hoist the terminator.
1621  Instruction *NT = I1->clone();
1622  BIParent->getInstList().insert(BI->getIterator(), NT);
1623  if (!NT->getType()->isVoidTy()) {
1624  I1->replaceAllUsesWith(NT);
1625  I2->replaceAllUsesWith(NT);
1626  NT->takeName(I1);
1627  }
1628  Changed = true;
1629  ++NumHoistCommonInstrs;
1630 
1631  // Ensure terminator gets a debug location, even an unknown one, in case
1632  // it involves inlinable calls.
1633  NT->applyMergedLocation(I1->getDebugLoc(), I2->getDebugLoc());
1634 
1635  // PHIs created below will adopt NT's merged DebugLoc.
1637 
1638  // Hoisting one of the terminators from our successor is a great thing.
1639  // Unfortunately, the successors of the if/else blocks may have PHI nodes in
1640  // them. If they do, all PHI entries for BB1/BB2 must agree for all PHI
1641  // nodes, so we insert select instruction to compute the final result.
1642  std::map<std::pair<Value *, Value *>, SelectInst *> InsertedSelects;
1643  for (BasicBlock *Succ : successors(BB1)) {
1644  for (PHINode &PN : Succ->phis()) {
1645  Value *BB1V = PN.getIncomingValueForBlock(BB1);
1646  Value *BB2V = PN.getIncomingValueForBlock(BB2);
1647  if (BB1V == BB2V)
1648  continue;
1649 
1650  // These values do not agree. Insert a select instruction before NT
1651  // that determines the right value.
1652  SelectInst *&SI = InsertedSelects[std::make_pair(BB1V, BB2V)];
1653  if (!SI) {
1654  // Propagate fast-math-flags from phi node to its replacement select.
1656  if (isa<FPMathOperator>(PN))
1657  Builder.setFastMathFlags(PN.getFastMathFlags());
1658 
1659  SI = cast<SelectInst>(
1660  Builder.CreateSelect(BI->getCondition(), BB1V, BB2V,
1661  BB1V->getName() + "." + BB2V->getName(), BI));
1662  }
1663 
1664  // Make the PHI node use the select for all incoming values for BB1/BB2
1665  for (unsigned i = 0, e = PN.getNumIncomingValues(); i != e; ++i)
1666  if (PN.getIncomingBlock(i) == BB1 || PN.getIncomingBlock(i) == BB2)
1667  PN.setIncomingValue(i, SI);
1668  }
1669  }
1670 
1672 
1673  // Update any PHI nodes in our new successors.
1674  for (BasicBlock *Succ : successors(BB1)) {
1675  AddPredecessorToBlock(Succ, BIParent, BB1);
1676  if (DTU)
1677  Updates.push_back({DominatorTree::Insert, BIParent, Succ});
1678  }
1679 
1680  if (DTU)
1681  for (BasicBlock *Succ : successors(BI))
1682  Updates.push_back({DominatorTree::Delete, BIParent, Succ});
1683 
1685  if (DTU)
1686  DTU->applyUpdates(Updates);
1687  return Changed;
1688 }
1689 
1690 // Check lifetime markers.
1691 static bool isLifeTimeMarker(const Instruction *I) {
1692  if (auto II = dyn_cast<IntrinsicInst>(I)) {
1693  switch (II->getIntrinsicID()) {
1694  default:
1695  break;
1696  case Intrinsic::lifetime_start:
1697  case Intrinsic::lifetime_end:
1698  return true;
1699  }
1700  }
1701  return false;
1702 }
1703 
1704 // TODO: Refine this. This should avoid cases like turning constant memcpy sizes
1705 // into variables.
1707  int OpIdx) {
1708  return !isa<IntrinsicInst>(I);
1709 }
1710 
1711 // All instructions in Insts belong to different blocks that all unconditionally
1712 // branch to a common successor. Analyze each instruction and return true if it
1713 // would be possible to sink them into their successor, creating one common
1714 // instruction instead. For every value that would be required to be provided by
1715 // PHI node (because an operand varies in each input block), add to PHIOperands.
1718  DenseMap<Instruction *, SmallVector<Value *, 4>> &PHIOperands) {
1719  // Prune out obviously bad instructions to move. Each instruction must have
1720  // exactly zero or one use, and we check later that use is by a single, common
1721  // PHI instruction in the successor.
1722  bool HasUse = !Insts.front()->user_empty();
1723  for (auto *I : Insts) {
1724  // These instructions may change or break semantics if moved.
1725  if (isa<PHINode>(I) || I->isEHPad() || isa<AllocaInst>(I) ||
1726  I->getType()->isTokenTy())
1727  return false;
1728 
1729  // Do not try to sink an instruction in an infinite loop - it can cause
1730  // this algorithm to infinite loop.
1731  if (I->getParent()->getSingleSuccessor() == I->getParent())
1732  return false;
1733 
1734  // Conservatively return false if I is an inline-asm instruction. Sinking
1735  // and merging inline-asm instructions can potentially create arguments
1736  // that cannot satisfy the inline-asm constraints.
1737  // If the instruction has nomerge attribute, return false.
1738  if (const auto *C = dyn_cast<CallBase>(I))
1739  if (C->isInlineAsm() || C->cannotMerge())
1740  return false;
1741 
1742  // Each instruction must have zero or one use.
1743  if (HasUse && !I->hasOneUse())
1744  return false;
1745  if (!HasUse && !I->user_empty())
1746  return false;
1747  }
1748 
1749  const Instruction *I0 = Insts.front();
1750  for (auto *I : Insts)
1751  if (!I->isSameOperationAs(I0))
1752  return false;
1753 
1754  // All instructions in Insts are known to be the same opcode. If they have a
1755  // use, check that the only user is a PHI or in the same block as the
1756  // instruction, because if a user is in the same block as an instruction we're
1757  // contemplating sinking, it must already be determined to be sinkable.
1758  if (HasUse) {
1759  auto *PNUse = dyn_cast<PHINode>(*I0->user_begin());
1760  auto *Succ = I0->getParent()->getTerminator()->getSuccessor(0);
1761  if (!all_of(Insts, [&PNUse,&Succ](const Instruction *I) -> bool {
1762  auto *U = cast<Instruction>(*I->user_begin());
1763  return (PNUse &&
1764  PNUse->getParent() == Succ &&
1765  PNUse->getIncomingValueForBlock(I->getParent()) == I) ||
1766  U->getParent() == I->getParent();
1767  }))
1768  return false;
1769  }
1770 
1771  // Because SROA can't handle speculating stores of selects, try not to sink
1772  // loads, stores or lifetime markers of allocas when we'd have to create a
1773  // PHI for the address operand. Also, because it is likely that loads or
1774  // stores of allocas will disappear when Mem2Reg/SROA is run, don't sink
1775  // them.
1776  // This can cause code churn which can have unintended consequences down
1777  // the line - see https://llvm.org/bugs/show_bug.cgi?id=30244.
1778  // FIXME: This is a workaround for a deficiency in SROA - see
1779  // https://llvm.org/bugs/show_bug.cgi?id=30188
1780  if (isa<StoreInst>(I0) && any_of(Insts, [](const Instruction *I) {
1781  return isa<AllocaInst>(I->getOperand(1)->stripPointerCasts());
1782  }))
1783  return false;
1784  if (isa<LoadInst>(I0) && any_of(Insts, [](const Instruction *I) {
1785  return isa<AllocaInst>(I->getOperand(0)->stripPointerCasts());
1786  }))
1787  return false;
1788  if (isLifeTimeMarker(I0) && any_of(Insts, [](const Instruction *I) {
1789  return isa<AllocaInst>(I->getOperand(1)->stripPointerCasts());
1790  }))
1791  return false;
1792 
1793  // For calls to be sinkable, they must all be indirect, or have same callee.
1794  // I.e. if we have two direct calls to different callees, we don't want to
1795  // turn that into an indirect call. Likewise, if we have an indirect call,
1796  // and a direct call, we don't actually want to have a single indirect call.
1797  if (isa<CallBase>(I0)) {
1798  auto IsIndirectCall = [](const Instruction *I) {
1799  return cast<CallBase>(I)->isIndirectCall();
1800  };
1801  bool HaveIndirectCalls = any_of(Insts, IsIndirectCall);
1802  bool AllCallsAreIndirect = all_of(Insts, IsIndirectCall);
1803  if (HaveIndirectCalls) {
1804  if (!AllCallsAreIndirect)
1805  return false;
1806  } else {
1807  // All callees must be identical.
1808  Value *Callee = nullptr;
1809  for (const Instruction *I : Insts) {
1810  Value *CurrCallee = cast<CallBase>(I)->getCalledOperand();
1811  if (!Callee)
1812  Callee = CurrCallee;
1813  else if (Callee != CurrCallee)
1814  return false;
1815  }
1816  }
1817  }
1818 
1819  for (unsigned OI = 0, OE = I0->getNumOperands(); OI != OE; ++OI) {
1820  Value *Op = I0->getOperand(OI);
1821  if (Op->getType()->isTokenTy())
1822  // Don't touch any operand of token type.
1823  return false;
1824 
1825  auto SameAsI0 = [&I0, OI](const Instruction *I) {
1826  assert(I->getNumOperands() == I0->getNumOperands());
1827  return I->getOperand(OI) == I0->getOperand(OI);
1828  };
1829  if (!all_of(Insts, SameAsI0)) {
1830  if ((isa<Constant>(Op) && !replacingOperandWithVariableIsCheap(I0, OI)) ||
1832  // We can't create a PHI from this GEP.
1833  return false;
1834  for (auto *I : Insts)
1835  PHIOperands[I].push_back(I->getOperand(OI));
1836  }
1837  }
1838  return true;
1839 }
1840 
1841 // Assuming canSinkInstructions(Blocks) has returned true, sink the last
1842 // instruction of every block in Blocks to their common successor, commoning
1843 // into one instruction.
1845  auto *BBEnd = Blocks[0]->getTerminator()->getSuccessor(0);
1846 
1847  // canSinkInstructions returning true guarantees that every block has at
1848  // least one non-terminator instruction.
1850  for (auto *BB : Blocks) {
1851  Instruction *I = BB->getTerminator();
1852  do {
1853  I = I->getPrevNode();
1854  } while (isa<DbgInfoIntrinsic>(I) && I != &BB->front());
1855  if (!isa<DbgInfoIntrinsic>(I))
1856  Insts.push_back(I);
1857  }
1858 
1859  // The only checking we need to do now is that all users of all instructions
1860  // are the same PHI node. canSinkInstructions should have checked this but
1861  // it is slightly over-aggressive - it gets confused by commutative
1862  // instructions so double-check it here.
1863  Instruction *I0 = Insts.front();
1864  if (!I0->user_empty()) {
1865  auto *PNUse = dyn_cast<PHINode>(*I0->user_begin());
1866  if (!all_of(Insts, [&PNUse](const Instruction *I) -> bool {
1867  auto *U = cast<Instruction>(*I->user_begin());
1868  return U == PNUse;
1869  }))
1870  return false;
1871  }
1872 
1873  // We don't need to do any more checking here; canSinkInstructions should
1874  // have done it all for us.
1875  SmallVector<Value*, 4> NewOperands;
1876  for (unsigned O = 0, E = I0->getNumOperands(); O != E; ++O) {
1877  // This check is different to that in canSinkInstructions. There, we
1878  // cared about the global view once simplifycfg (and instcombine) have
1879  // completed - it takes into account PHIs that become trivially
1880  // simplifiable. However here we need a more local view; if an operand
1881  // differs we create a PHI and rely on instcombine to clean up the very
1882  // small mess we may make.
1883  bool NeedPHI = any_of(Insts, [&I0, O](const Instruction *I) {
1884  return I->getOperand(O) != I0->getOperand(O);
1885  });
1886  if (!NeedPHI) {
1887  NewOperands.push_back(I0->getOperand(O));
1888  continue;
1889  }
1890 
1891  // Create a new PHI in the successor block and populate it.
1892  auto *Op = I0->getOperand(O);
1893  assert(!Op->getType()->isTokenTy() && "Can't PHI tokens!");
1894  auto *PN = PHINode::Create(Op->getType(), Insts.size(),
1895  Op->getName() + ".sink", &BBEnd->front());
1896  for (auto *I : Insts)
1897  PN->addIncoming(I->getOperand(O), I->getParent());
1898  NewOperands.push_back(PN);
1899  }
1900 
1901  // Arbitrarily use I0 as the new "common" instruction; remap its operands
1902  // and move it to the start of the successor block.
1903  for (unsigned O = 0, E = I0->getNumOperands(); O != E; ++O)
1904  I0->getOperandUse(O).set(NewOperands[O]);
1905  I0->moveBefore(&*BBEnd->getFirstInsertionPt());
1906 
1907  // Update metadata and IR flags, and merge debug locations.
1908  for (auto *I : Insts)
1909  if (I != I0) {
1910  // The debug location for the "common" instruction is the merged locations
1911  // of all the commoned instructions. We start with the original location
1912  // of the "common" instruction and iteratively merge each location in the
1913  // loop below.
1914  // This is an N-way merge, which will be inefficient if I0 is a CallInst.
1915  // However, as N-way merge for CallInst is rare, so we use simplified API
1916  // instead of using complex API for N-way merge.
1917  I0->applyMergedLocation(I0->getDebugLoc(), I->getDebugLoc());
1918  combineMetadataForCSE(I0, I, true);
1919  I0->andIRFlags(I);
1920  }
1921 
1922  if (!I0->user_empty()) {
1923  // canSinkLastInstruction checked that all instructions were used by
1924  // one and only one PHI node. Find that now, RAUW it to our common
1925  // instruction and nuke it.
1926  auto *PN = cast<PHINode>(*I0->user_begin());
1927  PN->replaceAllUsesWith(I0);
1928  PN->eraseFromParent();
1929  }
1930 
1931  // Finally nuke all instructions apart from the common instruction.
1932  for (auto *I : Insts)
1933  if (I != I0)
1934  I->eraseFromParent();
1935 
1936  return true;
1937 }
1938 
1939 namespace {
1940 
1941  // LockstepReverseIterator - Iterates through instructions
1942  // in a set of blocks in reverse order from the first non-terminator.
1943  // For example (assume all blocks have size n):
1944  // LockstepReverseIterator I([B1, B2, B3]);
1945  // *I-- = [B1[n], B2[n], B3[n]];
1946  // *I-- = [B1[n-1], B2[n-1], B3[n-1]];
1947  // *I-- = [B1[n-2], B2[n-2], B3[n-2]];
1948  // ...
1949  class LockstepReverseIterator {
1950  ArrayRef<BasicBlock*> Blocks;
1952  bool Fail;
1953 
1954  public:
1955  LockstepReverseIterator(ArrayRef<BasicBlock*> Blocks) : Blocks(Blocks) {
1956  reset();
1957  }
1958 
1959  void reset() {
1960  Fail = false;
1961  Insts.clear();
1962  for (auto *BB : Blocks) {
1963  Instruction *Inst = BB->getTerminator();
1964  for (Inst = Inst->getPrevNode(); Inst && isa<DbgInfoIntrinsic>(Inst);)
1965  Inst = Inst->getPrevNode();
1966  if (!Inst) {
1967  // Block wasn't big enough.
1968  Fail = true;
1969  return;
1970  }
1971  Insts.push_back(Inst);
1972  }
1973  }
1974 
1975  bool isValid() const {
1976  return !Fail;
1977  }
1978 
1979  void operator--() {
1980  if (Fail)
1981  return;
1982  for (auto *&Inst : Insts) {
1983  for (Inst = Inst->getPrevNode(); Inst && isa<DbgInfoIntrinsic>(Inst);)
1984  Inst = Inst->getPrevNode();
1985  // Already at beginning of block.
1986  if (!Inst) {
1987  Fail = true;
1988  return;
1989  }
1990  }
1991  }
1992 
1993  void operator++() {
1994  if (Fail)
1995  return;
1996  for (auto *&Inst : Insts) {
1997  for (Inst = Inst->getNextNode(); Inst && isa<DbgInfoIntrinsic>(Inst);)
1998  Inst = Inst->getNextNode();
1999  // Already at end of block.
2000  if (!Inst) {
2001  Fail = true;
2002  return;
2003  }
2004  }
2005  }
2006 
2008  return Insts;
2009  }
2010  };
2011 
2012 } // end anonymous namespace
2013 
2014 /// Check whether BB's predecessors end with unconditional branches. If it is
2015 /// true, sink any common code from the predecessors to BB.
2017  DomTreeUpdater *DTU) {
2018  // We support two situations:
2019  // (1) all incoming arcs are unconditional
2020  // (2) there are non-unconditional incoming arcs
2021  //
2022  // (2) is very common in switch defaults and
2023  // else-if patterns;
2024  //
2025  // if (a) f(1);
2026  // else if (b) f(2);
2027  //
2028  // produces:
2029  //
2030  // [if]
2031  // / \
2032  // [f(1)] [if]
2033  // | | \
2034  // | | |
2035  // | [f(2)]|
2036  // \ | /
2037  // [ end ]
2038  //
2039  // [end] has two unconditional predecessor arcs and one conditional. The
2040  // conditional refers to the implicit empty 'else' arc. This conditional
2041  // arc can also be caused by an empty default block in a switch.
2042  //
2043  // In this case, we attempt to sink code from all *unconditional* arcs.
2044  // If we can sink instructions from these arcs (determined during the scan
2045  // phase below) we insert a common successor for all unconditional arcs and
2046  // connect that to [end], to enable sinking:
2047  //
2048  // [if]
2049  // / \
2050  // [x(1)] [if]
2051  // | | \
2052  // | | \
2053  // | [x(2)] |
2054  // \ / |
2055  // [sink.split] |
2056  // \ /
2057  // [ end ]
2058  //
2059  SmallVector<BasicBlock*,4> UnconditionalPreds;
2060  bool HaveNonUnconditionalPredecessors = false;
2061  for (auto *PredBB : predecessors(BB)) {
2062  auto *PredBr = dyn_cast<BranchInst>(PredBB->getTerminator());
2063  if (PredBr && PredBr->isUnconditional())
2064  UnconditionalPreds.push_back(PredBB);
2065  else
2066  HaveNonUnconditionalPredecessors = true;
2067  }
2068  if (UnconditionalPreds.size() < 2)
2069  return false;
2070 
2071  // We take a two-step approach to tail sinking. First we scan from the end of
2072  // each block upwards in lockstep. If the n'th instruction from the end of each
2073  // block can be sunk, those instructions are added to ValuesToSink and we
2074  // carry on. If we can sink an instruction but need to PHI-merge some operands
2075  // (because they're not identical in each instruction) we add these to
2076  // PHIOperands.
2077  int ScanIdx = 0;
2078  SmallPtrSet<Value*,4> InstructionsToSink;
2080  LockstepReverseIterator LRI(UnconditionalPreds);
2081  while (LRI.isValid() &&
2082  canSinkInstructions(*LRI, PHIOperands)) {
2083  LLVM_DEBUG(dbgs() << "SINK: instruction can be sunk: " << *(*LRI)[0]
2084  << "\n");
2085  InstructionsToSink.insert((*LRI).begin(), (*LRI).end());
2086  ++ScanIdx;
2087  --LRI;
2088  }
2089 
2090  // If no instructions can be sunk, early-return.
2091  if (ScanIdx == 0)
2092  return false;
2093 
2094  bool followedByDeoptOrUnreachable = IsBlockFollowedByDeoptOrUnreachable(BB);
2095 
2096  if (!followedByDeoptOrUnreachable) {
2097  // Okay, we *could* sink last ScanIdx instructions. But how many can we
2098  // actually sink before encountering instruction that is unprofitable to
2099  // sink?
2100  auto ProfitableToSinkInstruction = [&](LockstepReverseIterator &LRI) {
2101  unsigned NumPHIdValues = 0;
2102  for (auto *I : *LRI)
2103  for (auto *V : PHIOperands[I]) {
2104  if (!InstructionsToSink.contains(V))
2105  ++NumPHIdValues;
2106  // FIXME: this check is overly optimistic. We may end up not sinking
2107  // said instruction, due to the very same profitability check.
2108  // See @creating_too_many_phis in sink-common-code.ll.
2109  }
2110  LLVM_DEBUG(dbgs() << "SINK: #phid values: " << NumPHIdValues << "\n");
2111  unsigned NumPHIInsts = NumPHIdValues / UnconditionalPreds.size();
2112  if ((NumPHIdValues % UnconditionalPreds.size()) != 0)
2113  NumPHIInsts++;
2114 
2115  return NumPHIInsts <= 1;
2116  };
2117 
2118  // We've determined that we are going to sink last ScanIdx instructions,
2119  // and recorded them in InstructionsToSink. Now, some instructions may be
2120  // unprofitable to sink. But that determination depends on the instructions
2121  // that we are going to sink.
2122 
2123  // First, forward scan: find the first instruction unprofitable to sink,
2124  // recording all the ones that are profitable to sink.
2125  // FIXME: would it be better, after we detect that not all are profitable.
2126  // to either record the profitable ones, or erase the unprofitable ones?
2127  // Maybe we need to choose (at runtime) the one that will touch least
2128  // instrs?
2129  LRI.reset();
2130  int Idx = 0;
2131  SmallPtrSet<Value *, 4> InstructionsProfitableToSink;
2132  while (Idx < ScanIdx) {
2133  if (!ProfitableToSinkInstruction(LRI)) {
2134  // Too many PHIs would be created.
2135  LLVM_DEBUG(
2136  dbgs() << "SINK: stopping here, too many PHIs would be created!\n");
2137  break;
2138  }
2139  InstructionsProfitableToSink.insert((*LRI).begin(), (*LRI).end());
2140  --LRI;
2141  ++Idx;
2142  }
2143 
2144  // If no instructions can be sunk, early-return.
2145  if (Idx == 0)
2146  return false;
2147 
2148  // Did we determine that (only) some instructions are unprofitable to sink?
2149  if (Idx < ScanIdx) {
2150  // Okay, some instructions are unprofitable.
2151  ScanIdx = Idx;
2152  InstructionsToSink = InstructionsProfitableToSink;
2153 
2154  // But, that may make other instructions unprofitable, too.
2155  // So, do a backward scan, do any earlier instructions become
2156  // unprofitable?
2157  assert(
2158  !ProfitableToSinkInstruction(LRI) &&
2159  "We already know that the last instruction is unprofitable to sink");
2160  ++LRI;
2161  --Idx;
2162  while (Idx >= 0) {
2163  // If we detect that an instruction becomes unprofitable to sink,
2164  // all earlier instructions won't be sunk either,
2165  // so preemptively keep InstructionsProfitableToSink in sync.
2166  // FIXME: is this the most performant approach?
2167  for (auto *I : *LRI)
2168  InstructionsProfitableToSink.erase(I);
2169  if (!ProfitableToSinkInstruction(LRI)) {
2170  // Everything starting with this instruction won't be sunk.
2171  ScanIdx = Idx;
2172  InstructionsToSink = InstructionsProfitableToSink;
2173  }
2174  ++LRI;
2175  --Idx;
2176  }
2177  }
2178 
2179  // If no instructions can be sunk, early-return.
2180  if (ScanIdx == 0)
2181  return false;
2182  }
2183 
2184  bool Changed = false;
2185 
2186  if (HaveNonUnconditionalPredecessors) {
2187  if (!followedByDeoptOrUnreachable) {
2188  // It is always legal to sink common instructions from unconditional
2189  // predecessors. However, if not all predecessors are unconditional,
2190  // this transformation might be pessimizing. So as a rule of thumb,
2191  // don't do it unless we'd sink at least one non-speculatable instruction.
2192  // See https://bugs.llvm.org/show_bug.cgi?id=30244
2193  LRI.reset();
2194  int Idx = 0;
2195  bool Profitable = false;
2196  while (Idx < ScanIdx) {
2197  if (!isSafeToSpeculativelyExecute((*LRI)[0])) {
2198  Profitable = true;
2199  break;
2200  }
2201  --LRI;
2202  ++Idx;
2203  }
2204  if (!Profitable)
2205  return false;
2206  }
2207 
2208  LLVM_DEBUG(dbgs() << "SINK: Splitting edge\n");
2209  // We have a conditional edge and we're going to sink some instructions.
2210  // Insert a new block postdominating all blocks we're going to sink from.
2211  if (!SplitBlockPredecessors(BB, UnconditionalPreds, ".sink.split", DTU))
2212  // Edges couldn't be split.
2213  return false;
2214  Changed = true;
2215  }
2216 
2217  // Now that we've analyzed all potential sinking candidates, perform the
2218  // actual sink. We iteratively sink the last non-terminator of the source
2219  // blocks into their common successor unless doing so would require too
2220  // many PHI instructions to be generated (currently only one PHI is allowed
2221  // per sunk instruction).
2222  //
2223  // We can use InstructionsToSink to discount values needing PHI-merging that will
2224  // actually be sunk in a later iteration. This allows us to be more
2225  // aggressive in what we sink. This does allow a false positive where we
2226  // sink presuming a later value will also be sunk, but stop half way through
2227  // and never actually sink it which means we produce more PHIs than intended.
2228  // This is unlikely in practice though.
2229  int SinkIdx = 0;
2230  for (; SinkIdx != ScanIdx; ++SinkIdx) {
2231  LLVM_DEBUG(dbgs() << "SINK: Sink: "
2232  << *UnconditionalPreds[0]->getTerminator()->getPrevNode()
2233  << "\n");
2234 
2235  // Because we've sunk every instruction in turn, the current instruction to
2236  // sink is always at index 0.
2237  LRI.reset();
2238 
2239  if (!sinkLastInstruction(UnconditionalPreds)) {
2240  LLVM_DEBUG(
2241  dbgs()
2242  << "SINK: stopping here, failed to actually sink instruction!\n");
2243  break;
2244  }
2245 
2246  NumSinkCommonInstrs++;
2247  Changed = true;
2248  }
2249  if (SinkIdx != 0)
2250  ++NumSinkCommonCode;
2251  return Changed;
2252 }
2253 
2254 namespace {
2255 
2256 struct CompatibleSets {
2257  using SetTy = SmallVector<InvokeInst *, 2>;
2258 
2259  SmallVector<SetTy, 1> Sets;
2260 
2261  static bool shouldBelongToSameSet(ArrayRef<InvokeInst *> Invokes);
2262 
2263  SetTy &getCompatibleSet(InvokeInst *II);
2264 
2265  void insert(InvokeInst *II);
2266 };
2267 
2268 CompatibleSets::SetTy &CompatibleSets::getCompatibleSet(InvokeInst *II) {
2269  // Perform a linear scan over all the existing sets, see if the new `invoke`
2270  // is compatible with any particular set. Since we know that all the `invokes`
2271  // within a set are compatible, only check the first `invoke` in each set.
2272  // WARNING: at worst, this has quadratic complexity.
2273  for (CompatibleSets::SetTy &Set : Sets) {
2274  if (CompatibleSets::shouldBelongToSameSet({Set.front(), II}))
2275  return Set;
2276  }
2277 
2278  // Otherwise, we either had no sets yet, or this invoke forms a new set.
2279  return Sets.emplace_back();
2280 }
2281 
2282 void CompatibleSets::insert(InvokeInst *II) {
2283  getCompatibleSet(II).emplace_back(II);
2284 }
2285 
2286 bool CompatibleSets::shouldBelongToSameSet(ArrayRef<InvokeInst *> Invokes) {
2287  assert(Invokes.size() == 2 && "Always called with exactly two candidates.");
2288 
2289  // Can we theoretically merge these `invoke`s?
2290  auto IsIllegalToMerge = [](InvokeInst *II) {
2291  return II->cannotMerge() || II->isInlineAsm();
2292  };
2293  if (any_of(Invokes, IsIllegalToMerge))
2294  return false;
2295 
2296  // Either both `invoke`s must be direct,
2297  // or both `invoke`s must be indirect.
2298  auto IsIndirectCall = [](InvokeInst *II) { return II->isIndirectCall(); };
2299  bool HaveIndirectCalls = any_of(Invokes, IsIndirectCall);
2300  bool AllCallsAreIndirect = all_of(Invokes, IsIndirectCall);
2301  if (HaveIndirectCalls) {
2302  if (!AllCallsAreIndirect)
2303  return false;
2304  } else {
2305  // All callees must be identical.
2306  Value *Callee = nullptr;
2307  for (InvokeInst *II : Invokes) {
2308  Value *CurrCallee = II->getCalledOperand();
2309  assert(CurrCallee && "There is always a called operand.");
2310  if (!Callee)
2311  Callee = CurrCallee;
2312  else if (Callee != CurrCallee)
2313  return false;
2314  }
2315  }
2316 
2317  // Either both `invoke`s must not have a normal destination,
2318  // or both `invoke`s must have a normal destination,
2319  auto HasNormalDest = [](InvokeInst *II) {
2320  return !isa<UnreachableInst>(II->getNormalDest()->getFirstNonPHIOrDbg());
2321  };
2322  if (any_of(Invokes, HasNormalDest)) {
2323  // Do not merge `invoke` that does not have a normal destination with one
2324  // that does have a normal destination, even though doing so would be legal.
2325  if (!all_of(Invokes, HasNormalDest))
2326  return false;
2327 
2328  // All normal destinations must be identical.
2329  BasicBlock *NormalBB = nullptr;
2330  for (InvokeInst *II : Invokes) {
2331  BasicBlock *CurrNormalBB = II->getNormalDest();
2332  assert(CurrNormalBB && "There is always a 'continue to' basic block.");
2333  if (!NormalBB)
2334  NormalBB = CurrNormalBB;
2335  else if (NormalBB != CurrNormalBB)
2336  return false;
2337  }
2338 
2339  // In the normal destination, the incoming values for these two `invoke`s
2340  // must be compatible.
2341  SmallPtrSet<Value *, 16> EquivalenceSet(Invokes.begin(), Invokes.end());
2343  NormalBB, {Invokes[0]->getParent(), Invokes[1]->getParent()},
2344  &EquivalenceSet))
2345  return false;
2346  }
2347 
2348 #ifndef NDEBUG
2349  // All unwind destinations must be identical.
2350  // We know that because we have started from said unwind destination.
2351  BasicBlock *UnwindBB = nullptr;
2352  for (InvokeInst *II : Invokes) {
2353  BasicBlock *CurrUnwindBB = II->getUnwindDest();
2354  assert(CurrUnwindBB && "There is always an 'unwind to' basic block.");
2355  if (!UnwindBB)
2356  UnwindBB = CurrUnwindBB;
2357  else
2358  assert(UnwindBB == CurrUnwindBB && "Unexpected unwind destination.");
2359  }
2360 #endif
2361 
2362  // In the unwind destination, the incoming values for these two `invoke`s
2363  // must be compatible.
2365  Invokes.front()->getUnwindDest(),
2366  {Invokes[0]->getParent(), Invokes[1]->getParent()}))
2367  return false;
2368 
2369  // Ignoring arguments, these `invoke`s must be identical,
2370  // including operand bundles.
2371  const InvokeInst *II0 = Invokes.front();
2372  for (auto *II : Invokes.drop_front())
2373  if (!II->isSameOperationAs(II0))
2374  return false;
2375 
2376  // Can we theoretically form the data operands for the merged `invoke`?
2377  auto IsIllegalToMergeArguments = [](auto Ops) {
2378  Type *Ty = std::get<0>(Ops)->getType();
2379  assert(Ty == std::get<1>(Ops)->getType() && "Incompatible types?");
2380  return Ty->isTokenTy() && std::get<0>(Ops) != std::get<1>(Ops);
2381  };
2382  assert(Invokes.size() == 2 && "Always called with exactly two candidates.");
2383  if (any_of(zip(Invokes[0]->data_ops(), Invokes[1]->data_ops()),
2384  IsIllegalToMergeArguments))
2385  return false;
2386 
2387  return true;
2388 }
2389 
2390 } // namespace
2391 
2392 // Merge all invokes in the provided set, all of which are compatible
2393 // as per the `CompatibleSets::shouldBelongToSameSet()`.
2395  DomTreeUpdater *DTU) {
2396  assert(Invokes.size() >= 2 && "Must have at least two invokes to merge.");
2397 
2399  if (DTU)
2400  Updates.reserve(2 + 3 * Invokes.size());
2401 
2402  bool HasNormalDest =
2403  !isa<UnreachableInst>(Invokes[0]->getNormalDest()->getFirstNonPHIOrDbg());
2404 
2405  // Clone one of the invokes into a new basic block.
2406  // Since they are all compatible, it doesn't matter which invoke is cloned.
2407  InvokeInst *MergedInvoke = [&Invokes, HasNormalDest]() {
2408  InvokeInst *II0 = Invokes.front();
2409  BasicBlock *II0BB = II0->getParent();
2410  BasicBlock *InsertBeforeBlock =
2411  II0->getParent()->getIterator()->getNextNode();
2412  Function *Func = II0BB->getParent();
2413  LLVMContext &Ctx = II0->getContext();
2414 
2415  BasicBlock *MergedInvokeBB = BasicBlock::Create(
2416  Ctx, II0BB->getName() + ".invoke", Func, InsertBeforeBlock);
2417 
2418  auto *MergedInvoke = cast<InvokeInst>(II0->clone());
2419  // NOTE: all invokes have the same attributes, so no handling needed.
2420  MergedInvokeBB->getInstList().push_back(MergedInvoke);
2421 
2422  if (!HasNormalDest) {
2423  // This set does not have a normal destination,
2424  // so just form a new block with unreachable terminator.
2425  BasicBlock *MergedNormalDest = BasicBlock::Create(
2426  Ctx, II0BB->getName() + ".cont", Func, InsertBeforeBlock);
2427  new UnreachableInst(Ctx, MergedNormalDest);
2428  MergedInvoke->setNormalDest(MergedNormalDest);
2429  }
2430 
2431  // The unwind destination, however, remainds identical for all invokes here.
2432 
2433  return MergedInvoke;
2434  }();
2435 
2436  if (DTU) {
2437  // Predecessor blocks that contained these invokes will now branch to
2438  // the new block that contains the merged invoke, ...
2439  for (InvokeInst *II : Invokes)
2440  Updates.push_back(
2441  {DominatorTree::Insert, II->getParent(), MergedInvoke->getParent()});
2442 
2443  // ... which has the new `unreachable` block as normal destination,
2444  // or unwinds to the (same for all `invoke`s in this set) `landingpad`,
2445  for (BasicBlock *SuccBBOfMergedInvoke : successors(MergedInvoke))
2446  Updates.push_back({DominatorTree::Insert, MergedInvoke->getParent(),
2447  SuccBBOfMergedInvoke});
2448 
2449  // Since predecessor blocks now unconditionally branch to a new block,
2450  // they no longer branch to their original successors.
2451  for (InvokeInst *II : Invokes)
2452  for (BasicBlock *SuccOfPredBB : successors(II->getParent()))
2453  Updates.push_back(
2454  {DominatorTree::Delete, II->getParent(), SuccOfPredBB});
2455  }
2456 
2457  bool IsIndirectCall = Invokes[0]->isIndirectCall();
2458 
2459  // Form the merged operands for the merged invoke.
2460  for (Use &U : MergedInvoke->operands()) {
2461  // Only PHI together the indirect callees and data operands.
2462  if (MergedInvoke->isCallee(&U)) {
2463  if (!IsIndirectCall)
2464  continue;
2465  } else if (!MergedInvoke->isDataOperand(&U))
2466  continue;
2467 
2468  // Don't create trivial PHI's with all-identical incoming values.
2469  bool NeedPHI = any_of(Invokes, [&U](InvokeInst *II) {
2470  return II->getOperand(U.getOperandNo()) != U.get();
2471  });
2472  if (!NeedPHI)
2473  continue;
2474 
2475  // Form a PHI out of all the data ops under this index.
2476  PHINode *PN = PHINode::Create(
2477  U->getType(), /*NumReservedValues=*/Invokes.size(), "", MergedInvoke);
2478  for (InvokeInst *II : Invokes)
2479  PN->addIncoming(II->getOperand(U.getOperandNo()), II->getParent());
2480 
2481  U.set(PN);
2482  }
2483 
2484  // We've ensured that each PHI node has compatible (identical) incoming values
2485  // when coming from each of the `invoke`s in the current merge set,
2486  // so update the PHI nodes accordingly.
2487  for (BasicBlock *Succ : successors(MergedInvoke))
2488  AddPredecessorToBlock(Succ, /*NewPred=*/MergedInvoke->getParent(),
2489  /*ExistPred=*/Invokes.front()->getParent());
2490 
2491  // And finally, replace the original `invoke`s with an unconditional branch
2492  // to the block with the merged `invoke`. Also, give that merged `invoke`
2493  // the merged debugloc of all the original `invoke`s.
2494  const DILocation *MergedDebugLoc = nullptr;
2495  for (InvokeInst *II : Invokes) {
2496  // Compute the debug location common to all the original `invoke`s.
2497  if (!MergedDebugLoc)
2498  MergedDebugLoc = II->getDebugLoc();
2499  else
2500  MergedDebugLoc =
2501  DILocation::getMergedLocation(MergedDebugLoc, II->getDebugLoc());
2502 
2503  // And replace the old `invoke` with an unconditionally branch
2504  // to the block with the merged `invoke`.
2505  for (BasicBlock *OrigSuccBB : successors(II->getParent()))
2506  OrigSuccBB->removePredecessor(II->getParent());
2507  BranchInst::Create(MergedInvoke->getParent(), II->getParent());
2508  II->replaceAllUsesWith(MergedInvoke);
2509  II->eraseFromParent();
2510  ++NumInvokesMerged;
2511  }
2512  MergedInvoke->setDebugLoc(MergedDebugLoc);
2513  ++NumInvokeSetsFormed;
2514 
2515  if (DTU)
2516  DTU->applyUpdates(Updates);
2517 }
2518 
2519 /// If this block is a `landingpad` exception handling block, categorize all
2520 /// the predecessor `invoke`s into sets, with all `invoke`s in each set
2521 /// being "mergeable" together, and then merge invokes in each set together.
2522 ///
2523 /// This is a weird mix of hoisting and sinking. Visually, it goes from:
2524 /// [...] [...]
2525 /// | |
2526 /// [invoke0] [invoke1]
2527 /// / \ / \
2528 /// [cont0] [landingpad] [cont1]
2529 /// to:
2530 /// [...] [...]
2531 /// \ /
2532 /// [invoke]
2533 /// / \
2534 /// [cont] [landingpad]
2535 ///
2536 /// But of course we can only do that if the invokes share the `landingpad`,
2537 /// edges invoke0->cont0 and invoke1->cont1 are "compatible",
2538 /// and the invoked functions are "compatible".
2541  return false;
2542 
2543  bool Changed = false;
2544 
2545  // FIXME: generalize to all exception handling blocks?
2546  if (!BB->isLandingPad())
2547  return Changed;
2548 
2549  CompatibleSets Grouper;
2550 
2551  // Record all the predecessors of this `landingpad`. As per verifier,
2552  // the only allowed predecessor is the unwind edge of an `invoke`.
2553  // We want to group "compatible" `invokes` into the same set to be merged.
2554  for (BasicBlock *PredBB : predecessors(BB))
2555  Grouper.insert(cast<InvokeInst>(PredBB->getTerminator()));
2556 
2557  // And now, merge `invoke`s that were grouped togeter.
2558  for (ArrayRef<InvokeInst *> Invokes : Grouper.Sets) {
2559  if (Invokes.size() < 2)
2560  continue;
2561  Changed = true;
2562  MergeCompatibleInvokesImpl(Invokes, DTU);
2563  }
2564 
2565  return Changed;
2566 }
2567 
2568 /// Determine if we can hoist sink a sole store instruction out of a
2569 /// conditional block.
2570 ///
2571 /// We are looking for code like the following:
2572 /// BrBB:
2573 /// store i32 %add, i32* %arrayidx2
2574 /// ... // No other stores or function calls (we could be calling a memory
2575 /// ... // function).
2576 /// %cmp = icmp ult %x, %y
2577 /// br i1 %cmp, label %EndBB, label %ThenBB
2578 /// ThenBB:
2579 /// store i32 %add5, i32* %arrayidx2
2580 /// br label EndBB
2581 /// EndBB:
2582 /// ...
2583 /// We are going to transform this into:
2584 /// BrBB:
2585 /// store i32 %add, i32* %arrayidx2
2586 /// ... //
2587 /// %cmp = icmp ult %x, %y
2588 /// %add.add5 = select i1 %cmp, i32 %add, %add5
2589 /// store i32 %add.add5, i32* %arrayidx2
2590 /// ...
2591 ///
2592 /// \return The pointer to the value of the previous store if the store can be
2593 /// hoisted into the predecessor block. 0 otherwise.
2595  BasicBlock *StoreBB, BasicBlock *EndBB) {
2596  StoreInst *StoreToHoist = dyn_cast<StoreInst>(I);
2597  if (!StoreToHoist)
2598  return nullptr;
2599 
2600  // Volatile or atomic.
2601  if (!StoreToHoist->isSimple())
2602  return nullptr;
2603 
2604  Value *StorePtr = StoreToHoist->getPointerOperand();
2605  Type *StoreTy = StoreToHoist->getValueOperand()->getType();
2606 
2607  // Look for a store to the same pointer in BrBB.
2608  unsigned MaxNumInstToLookAt = 9;
2609  // Skip pseudo probe intrinsic calls which are not really killing any memory
2610  // accesses.
2611  for (Instruction &CurI : reverse(BrBB->instructionsWithoutDebug(true))) {
2612  if (!MaxNumInstToLookAt)
2613  break;
2614  --MaxNumInstToLookAt;
2615 
2616  // Could be calling an instruction that affects memory like free().
2617  if (CurI.mayWriteToMemory() && !isa<StoreInst>(CurI))
2618  return nullptr;
2619 
2620  if (auto *SI = dyn_cast<StoreInst>(&CurI)) {
2621  // Found the previous store to same location and type. Make sure it is
2622  // simple, to avoid introducing a spurious non-atomic write after an
2623  // atomic write.
2624  if (SI->getPointerOperand() == StorePtr &&
2625  SI->getValueOperand()->getType() == StoreTy && SI->isSimple())
2626  // Found the previous store, return its value operand.
2627  return SI->getValueOperand();
2628  return nullptr; // Unknown store.
2629  }
2630 
2631  if (auto *LI = dyn_cast<LoadInst>(&CurI)) {
2632  if (LI->getPointerOperand() == StorePtr && LI->getType() == StoreTy &&
2633  LI->isSimple()) {
2634  // Local objects (created by an `alloca` instruction) are always
2635  // writable, so once we are past a read from a location it is valid to
2636  // also write to that same location.
2637  // If the address of the local object never escapes the function, that
2638  // means it's never concurrently read or written, hence moving the store
2639  // from under the condition will not introduce a data race.
2640  auto *AI = dyn_cast<AllocaInst>(getUnderlyingObject(StorePtr));
2641  if (AI && !PointerMayBeCaptured(AI, false, true))
2642  // Found a previous load, return it.
2643  return LI;
2644  }
2645  // The load didn't work out, but we may still find a store.
2646  }
2647  }
2648 
2649  return nullptr;
2650 }
2651 
2652 /// Estimate the cost of the insertion(s) and check that the PHI nodes can be
2653 /// converted to selects.
2655  BasicBlock *EndBB,
2656  unsigned &SpeculatedInstructions,
2657  InstructionCost &Cost,
2658  const TargetTransformInfo &TTI) {
2660  BB->getParent()->hasMinSize()
2661  ? TargetTransformInfo::TCK_CodeSize
2662  : TargetTransformInfo::TCK_SizeAndLatency;
2663 
2664  bool HaveRewritablePHIs = false;
2665  for (PHINode &PN : EndBB->phis()) {
2666  Value *OrigV = PN.getIncomingValueForBlock(BB);
2667  Value *ThenV = PN.getIncomingValueForBlock(ThenBB);
2668 
2669  // FIXME: Try to remove some of the duplication with HoistThenElseCodeToIf.
2670  // Skip PHIs which are trivial.
2671  if (ThenV == OrigV)
2672  continue;
2673 
2674  Cost += TTI.getCmpSelInstrCost(Instruction::Select, PN.getType(), nullptr,
2675  CmpInst::BAD_ICMP_PREDICATE, CostKind);
2676 
2677  // Don't convert to selects if we could remove undefined behavior instead.
2678  if (passingValueIsAlwaysUndefined(OrigV, &PN) ||
2679  passingValueIsAlwaysUndefined(ThenV, &PN))
2680  return false;
2681 
2682  if (canTrap(OrigV) || canTrap(ThenV))
2683  return false;
2684 
2685  HaveRewritablePHIs = true;
2686  ConstantExpr *OrigCE = dyn_cast<ConstantExpr>(OrigV);
2687  ConstantExpr *ThenCE = dyn_cast<ConstantExpr>(ThenV);
2688  if (!OrigCE && !ThenCE)
2689  continue; // Known cheap (FIXME: Maybe not true for aggregates).
2690 
2691  InstructionCost OrigCost = OrigCE ? computeSpeculationCost(OrigCE, TTI) : 0;
2692  InstructionCost ThenCost = ThenCE ? computeSpeculationCost(ThenCE, TTI) : 0;
2693  InstructionCost MaxCost =
2694  2 * PHINodeFoldingThreshold * TargetTransformInfo::TCC_Basic;
2695  if (OrigCost + ThenCost > MaxCost)
2696  return false;
2697 
2698  // Account for the cost of an unfolded ConstantExpr which could end up
2699  // getting expanded into Instructions.
2700  // FIXME: This doesn't account for how many operations are combined in the
2701  // constant expression.
2702  ++SpeculatedInstructions;
2703  if (SpeculatedInstructions > 1)
2704  return false;
2705  }
2706 
2707  return HaveRewritablePHIs;
2708 }
2709 
2710 /// Speculate a conditional basic block flattening the CFG.
2711 ///
2712 /// Note that this is a very risky transform currently. Speculating
2713 /// instructions like this is most often not desirable. Instead, there is an MI
2714 /// pass which can do it with full awareness of the resource constraints.
2715 /// However, some cases are "obvious" and we should do directly. An example of
2716 /// this is speculating a single, reasonably cheap instruction.
2717 ///
2718 /// There is only one distinct advantage to flattening the CFG at the IR level:
2719 /// it makes very common but simplistic optimizations such as are common in
2720 /// instcombine and the DAG combiner more powerful by removing CFG edges and
2721 /// modeling their effects with easier to reason about SSA value graphs.
2722 ///
2723 ///
2724 /// An illustration of this transform is turning this IR:
2725 /// \code
2726 /// BB:
2727 /// %cmp = icmp ult %x, %y
2728 /// br i1 %cmp, label %EndBB, label %ThenBB
2729 /// ThenBB:
2730 /// %sub = sub %x, %y
2731 /// br label BB2
2732 /// EndBB:
2733 /// %phi = phi [ %sub, %ThenBB ], [ 0, %EndBB ]
2734 /// ...
2735 /// \endcode
2736 ///
2737 /// Into this IR:
2738 /// \code
2739 /// BB:
2740 /// %cmp = icmp ult %x, %y
2741 /// %sub = sub %x, %y
2742 /// %cond = select i1 %cmp, 0, %sub
2743 /// ...
2744 /// \endcode
2745 ///
2746 /// \returns true if the conditional block is removed.
2747 bool SimplifyCFGOpt::SpeculativelyExecuteBB(BranchInst *BI, BasicBlock *ThenBB,
2748  const TargetTransformInfo &TTI) {
2749  // Be conservative for now. FP select instruction can often be expensive.
2750  Value *BrCond = BI->getCondition();
2751  if (isa<FCmpInst>(BrCond))
2752  return false;
2753 
2754  BasicBlock *BB = BI->getParent();
2755  BasicBlock *EndBB = ThenBB->getTerminator()->getSuccessor(0);
2756  InstructionCost Budget =
2757  PHINodeFoldingThreshold * TargetTransformInfo::TCC_Basic;
2758 
2759  // If ThenBB is actually on the false edge of the conditional branch, remember
2760  // to swap the select operands later.
2761  bool Invert = false;
2762  if (ThenBB != BI->getSuccessor(0)) {
2763  assert(ThenBB == BI->getSuccessor(1) && "No edge from 'if' block?");
2764  Invert = true;
2765  }
2766  assert(EndBB == BI->getSuccessor(!Invert) && "No edge from to end block");
2767 
2768  // If the branch is non-unpredictable, and is predicted to *not* branch to
2769  // the `then` block, then avoid speculating it.
2770  if (!BI->getMetadata(LLVMContext::MD_unpredictable)) {
2771  uint64_t TWeight, FWeight;
2772  if (BI->extractProfMetadata(TWeight, FWeight) && (TWeight + FWeight) != 0) {
2773  uint64_t EndWeight = Invert ? TWeight : FWeight;
2774  BranchProbability BIEndProb =
2775  BranchProbability::getBranchProbability(EndWeight, TWeight + FWeight);
2777  if (BIEndProb >= Likely)
2778  return false;
2779  }
2780  }
2781 
2782  // Keep a count of how many times instructions are used within ThenBB when
2783  // they are candidates for sinking into ThenBB. Specifically:
2784  // - They are defined in BB, and
2785  // - They have no side effects, and
2786  // - All of their uses are in ThenBB.
2787  SmallDenseMap<Instruction *, unsigned, 4> SinkCandidateUseCounts;
2788 
2789  SmallVector<Instruction *, 4> SpeculatedDbgIntrinsics;
2790 
2791  unsigned SpeculatedInstructions = 0;
2792  Value *SpeculatedStoreValue = nullptr;
2793  StoreInst *SpeculatedStore = nullptr;
2794  for (BasicBlock::iterator BBI = ThenBB->begin(),
2795  BBE = std::prev(ThenBB->end());
2796  BBI != BBE; ++BBI) {
2797  Instruction *I = &*BBI;
2798  // Skip debug info.
2799  if (isa<DbgInfoIntrinsic>(I)) {
2800  SpeculatedDbgIntrinsics.push_back(I);
2801  continue;
2802  }
2803 
2804  // Skip pseudo probes. The consequence is we lose track of the branch
2805  // probability for ThenBB, which is fine since the optimization here takes
2806  // place regardless of the branch probability.
2807  if (isa<PseudoProbeInst>(I)) {
2808  // The probe should be deleted so that it will not be over-counted when
2809  // the samples collected on the non-conditional path are counted towards
2810  // the conditional path. We leave it for the counts inference algorithm to
2811  // figure out a proper count for an unknown probe.
2812  SpeculatedDbgIntrinsics.push_back(I);
2813  continue;
2814  }
2815 
2816  // Only speculatively execute a single instruction (not counting the
2817  // terminator) for now.
2818  ++SpeculatedInstructions;
2819  if (SpeculatedInstructions > 1)
2820  return false;
2821 
2822  // Don't hoist the instruction if it's unsafe or expensive.
2824  !(HoistCondStores && (SpeculatedStoreValue = isSafeToSpeculateStore(
2825  I, BB, ThenBB, EndBB))))
2826  return false;
2827  if (!SpeculatedStoreValue &&
2829  PHINodeFoldingThreshold * TargetTransformInfo::TCC_Basic)
2830  return false;
2831 
2832  // Store the store speculation candidate.
2833  if (SpeculatedStoreValue)
2834  SpeculatedStore = cast<StoreInst>(I);
2835 
2836  // Do not hoist the instruction if any of its operands are defined but not
2837  // used in BB. The transformation will prevent the operand from
2838  // being sunk into the use block.
2839  for (Use &Op : I->operands()) {
2840  Instruction *OpI = dyn_cast<Instruction>(Op);
2841  if (!OpI || OpI->getParent() != BB || OpI->mayHaveSideEffects())
2842  continue; // Not a candidate for sinking.
2843 
2844  ++SinkCandidateUseCounts[OpI];
2845  }
2846  }
2847 
2848  // Consider any sink candidates which are only used in ThenBB as costs for
2849  // speculation. Note, while we iterate over a DenseMap here, we are summing
2850  // and so iteration order isn't significant.
2852  I = SinkCandidateUseCounts.begin(),
2853  E = SinkCandidateUseCounts.end();
2854  I != E; ++I)
2855  if (I->first->hasNUses(I->second)) {
2856  ++SpeculatedInstructions;
2857  if (SpeculatedInstructions > 1)
2858  return false;
2859  }
2860 
2861  // Check that we can insert the selects and that it's not too expensive to do
2862  // so.
2863  bool Convert = SpeculatedStore != nullptr;
2864  InstructionCost Cost = 0;
2865  Convert |= validateAndCostRequiredSelects(BB, ThenBB, EndBB,
2866  SpeculatedInstructions,
2867  Cost, TTI);
2868  if (!Convert || Cost > Budget)
2869  return false;
2870 
2871  // If we get here, we can hoist the instruction and if-convert.
2872  LLVM_DEBUG(dbgs() << "SPECULATIVELY EXECUTING BB" << *ThenBB << "\n";);
2873 
2874  // Insert a select of the value of the speculated store.
2875  if (SpeculatedStoreValue) {
2877  Value *TrueV = SpeculatedStore->getValueOperand();
2878  Value *FalseV = SpeculatedStoreValue;
2879  if (Invert)
2880  std::swap(TrueV, FalseV);
2881  Value *S = Builder.CreateSelect(
2882  BrCond, TrueV, FalseV, "spec.store.select", BI);
2883  SpeculatedStore->setOperand(0, S);
2884  SpeculatedStore->applyMergedLocation(BI->getDebugLoc(),
2885  SpeculatedStore->getDebugLoc());
2886  }
2887 
2888  // Metadata can be dependent on the condition we are hoisting above.
2889  // Conservatively strip all metadata on the instruction. Drop the debug loc
2890  // to avoid making it appear as if the condition is a constant, which would
2891  // be misleading while debugging.
2892  // Similarly strip attributes that maybe dependent on condition we are
2893  // hoisting above.
2894  for (auto &I : *ThenBB) {
2895  if (!SpeculatedStoreValue || &I != SpeculatedStore)
2896  I.setDebugLoc(DebugLoc());
2897  I.dropUndefImplyingAttrsAndUnknownMetadata();
2898  }
2899 
2900  // Hoist the instructions.
2901  BB->getInstList().splice(BI->getIterator(), ThenBB->getInstList(),
2902  ThenBB->begin(), std::prev(ThenBB->end()));
2903 
2904  // Insert selects and rewrite the PHI operands.
2906  for (PHINode &PN : EndBB->phis()) {
2907  unsigned OrigI = PN.getBasicBlockIndex(BB);
2908  unsigned ThenI = PN.getBasicBlockIndex(ThenBB);
2909  Value *OrigV = PN.getIncomingValue(OrigI);
2910  Value *ThenV = PN.getIncomingValue(ThenI);
2911 
2912  // Skip PHIs which are trivial.
2913  if (OrigV == ThenV)
2914  continue;
2915 
2916  // Create a select whose true value is the speculatively executed value and
2917  // false value is the pre-existing value. Swap them if the branch
2918  // destinations were inverted.
2919  Value *TrueV = ThenV, *FalseV = OrigV;
2920  if (Invert)
2921  std::swap(TrueV, FalseV);
2922  Value *V = Builder.CreateSelect(BrCond, TrueV, FalseV, "spec.select", BI);
2923  PN.setIncomingValue(OrigI, V);
2924  PN.setIncomingValue(ThenI, V);
2925  }
2926 
2927  // Remove speculated dbg intrinsics.
2928  // FIXME: Is it possible to do this in a more elegant way? Moving/merging the
2929  // dbg value for the different flows and inserting it after the select.
2930  for (Instruction *I : SpeculatedDbgIntrinsics)
2931  I->eraseFromParent();
2932 
2933  ++NumSpeculations;
2934  return true;
2935 }
2936 
2937 /// Return true if we can thread a branch across this block.
2939  int Size = 0;
2940 
2942  auto IsEphemeral = [&](const Instruction *I) {
2943  if (isa<AssumeInst>(I))
2944  return true;
2945  return !I->mayHaveSideEffects() && !I->isTerminator() &&
2946  all_of(I->users(),
2947  [&](const User *U) { return EphValues.count(U); });
2948  };
2949 
2950  // Walk the loop in reverse so that we can identify ephemeral values properly
2951  // (values only feeding assumes).
2952  for (Instruction &I : reverse(BB->instructionsWithoutDebug(false))) {
2953  // Can't fold blocks that contain noduplicate or convergent calls.
2954  if (CallInst *CI = dyn_cast<CallInst>(&I))
2955  if (CI->cannotDuplicate() || CI->isConvergent())
2956  return false;
2957 
2958  // Ignore ephemeral values which are deleted during codegen.
2959  if (IsEphemeral(&I))
2960  EphValues.insert(&I);
2961  // We will delete Phis while threading, so Phis should not be accounted in
2962  // block's size.
2963  else if (!isa<PHINode>(I)) {
2964  if (Size++ > MaxSmallBlockSize)
2965  return false; // Don't clone large BB's.
2966  }
2967 
2968  // We can only support instructions that do not define values that are
2969  // live outside of the current basic block.
2970  for (User *U : I.users()) {
2971  Instruction *UI = cast<Instruction>(U);
2972  if (UI->getParent() != BB || isa<PHINode>(UI))
2973  return false;
2974  }
2975 
2976  // Looks ok, continue checking.
2977  }
2978 
2979  return true;
2980 }
2981 
2982 static ConstantInt *
2984  SmallDenseMap<std::pair<BasicBlock *, BasicBlock *>,
2985  ConstantInt *> &Visited) {
2986  // Don't look past the block defining the value, we might get the value from
2987  // a previous loop iteration.
2988  auto *I = dyn_cast<Instruction>(V);
2989  if (I && I->getParent() == To)
2990  return nullptr;
2991 
2992  // We know the value if the From block branches on it.
2993  auto *BI = dyn_cast<BranchInst>(From->getTerminator());
2994  if (BI && BI->isConditional() && BI->getCondition() == V &&
2995  BI->getSuccessor(0) != BI->getSuccessor(1))
2996  return BI->getSuccessor(0) == To ? ConstantInt::getTrue(BI->getContext())
2998 
2999  // Limit the amount of blocks we inspect.
3000  if (Visited.size() >= 8)
3001  return nullptr;
3002 
3003  auto Pair = Visited.try_emplace({From, To}, nullptr);
3004  if (!Pair.second)
3005  return Pair.first->second;
3006 
3007  // Check whether the known value is the same for all predecessors.
3008  ConstantInt *Common = nullptr;
3009  for (BasicBlock *Pred : predecessors(From)) {
3010  ConstantInt *C = getKnownValueOnEdge(V, Pred, From, Visited);
3011  if (!C || (Common && Common != C))
3012  return nullptr;
3013  Common = C;
3014  }
3015  return Visited[{From, To}] = Common;
3016 }
3017 
3018 /// If we have a conditional branch on something for which we know the constant
3019 /// value in predecessors (e.g. a phi node in the current block), thread edges
3020 /// from the predecessor to their ultimate destination.
3021 static Optional<bool>
3023  const DataLayout &DL,
3024  AssumptionCache *AC) {
3026  BasicBlock *BB = BI->getParent();
3027  Value *Cond = BI->getCondition();
3028  PHINode *PN = dyn_cast<PHINode>(Cond);
3029  if (PN && PN->getParent() == BB) {
3030  // Degenerate case of a single entry PHI.
3031  if (PN->getNumIncomingValues() == 1) {
3033  return true;
3034  }
3035 
3036  for (Use &U : PN->incoming_values())
3037  if (auto *CB = dyn_cast<ConstantInt>(U))
3038  KnownValues.insert({PN->getIncomingBlock(U), CB});
3039  } else {
3041  for (BasicBlock *Pred : predecessors(BB)) {
3042  if (ConstantInt *CB = getKnownValueOnEdge(Cond, Pred, BB, Visited))
3043  KnownValues.insert({Pred, CB});
3044  }
3045  }
3046 
3047  if (KnownValues.empty())
3048  return false;
3049 
3050  // Now we know that this block has multiple preds and two succs.
3051  // Check that the block is small enough and values defined in the block are
3052  // not used outside of it.
3054  return false;
3055 
3056  for (const auto &Pair : KnownValues) {
3057  // Okay, we now know that all edges from PredBB should be revectored to
3058  // branch to RealDest.
3059  ConstantInt *CB = Pair.second;
3060  BasicBlock *PredBB = Pair.first;
3061  BasicBlock *RealDest = BI->getSuccessor(!CB->getZExtValue());
3062 
3063  if (RealDest == BB)
3064  continue; // Skip self loops.
3065  // Skip if the predecessor's terminator is an indirect branch.
3066  if (isa<IndirectBrInst>(PredBB->getTerminator()))
3067  continue;
3068 
3070 
3071  // The dest block might have PHI nodes, other predecessors and other
3072  // difficult cases. Instead of being smart about this, just insert a new
3073  // block that jumps to the destination block, effectively splitting
3074  // the edge we are about to create.
3075  BasicBlock *EdgeBB =
3076  BasicBlock::Create(BB->getContext(), RealDest->getName() + ".critedge",
3077  RealDest->getParent(), RealDest);
3078  BranchInst *CritEdgeBranch = BranchInst::Create(RealDest, EdgeBB);
3079  if (DTU)
3080  Updates.push_back({DominatorTree::Insert, EdgeBB, RealDest});
3081  CritEdgeBranch->setDebugLoc(BI->getDebugLoc());
3082 
3083  // Update PHI nodes.
3084  AddPredecessorToBlock(RealDest, EdgeBB, BB);
3085 
3086  // BB may have instructions that are being threaded over. Clone these
3087  // instructions into EdgeBB. We know that there will be no uses of the
3088  // cloned instructions outside of EdgeBB.
3089  BasicBlock::iterator InsertPt = EdgeBB->begin();
3090  DenseMap<Value *, Value *> TranslateMap; // Track translated values.
3091  TranslateMap[Cond] = Pair.second;
3092  for (BasicBlock::iterator BBI = BB->begin(); &*BBI != BI; ++BBI) {
3093  if (PHINode *PN = dyn_cast<PHINode>(BBI)) {
3094  TranslateMap[PN] = PN->getIncomingValueForBlock(PredBB);
3095  continue;
3096  }
3097  // Clone the instruction.
3098  Instruction *N = BBI->clone();
3099  if (BBI->hasName())
3100  N->setName(BBI->getName() + ".c");
3101 
3102  // Update operands due to translation.
3103  for (Use &Op : N->operands()) {
3104  DenseMap<Value *, Value *>::iterator PI = TranslateMap.find(Op);
3105  if (PI != TranslateMap.end())
3106  Op = PI->second;
3107  }
3108 
3109  // Check for trivial simplification.
3110  if (Value *V = simplifyInstruction(N, {DL, nullptr, nullptr, AC})) {
3111  if (!BBI->use_empty())
3112  TranslateMap[&*BBI] = V;
3113  if (!N->mayHaveSideEffects()) {
3114  N->deleteValue(); // Instruction folded away, don't need actual inst
3115  N = nullptr;
3116  }
3117  } else {
3118  if (!BBI->use_empty())
3119  TranslateMap[&*BBI] = N;
3120  }
3121  if (N) {
3122  // Insert the new instruction into its new home.
3123  EdgeBB->getInstList().insert(InsertPt, N);
3124 
3125  // Register the new instruction with the assumption cache if necessary.
3126  if (auto *Assume = dyn_cast<AssumeInst>(N))
3127  if (AC)
3128  AC->registerAssumption(Assume);
3129  }
3130  }
3131 
3132  // Loop over all of the edges from PredBB to BB, changing them to branch
3133  // to EdgeBB instead.
3134  Instruction *PredBBTI = PredBB->getTerminator();
3135  for (unsigned i = 0, e = PredBBTI->getNumSuccessors(); i != e; ++i)
3136  if (PredBBTI->getSuccessor(i) == BB) {
3137  BB->removePredecessor(PredBB);
3138  PredBBTI->setSuccessor(i, EdgeBB);
3139  }
3140 
3141  if (DTU) {
3142  Updates.push_back({DominatorTree::Insert, PredBB, EdgeBB});
3143  Updates.push_back({DominatorTree::Delete, PredBB, BB});
3144 
3145  DTU->applyUpdates(Updates);
3146  }
3147 
3148  // For simplicity, we created a separate basic block for the edge. Merge
3149  // it back into the predecessor if possible. This not only avoids
3150  // unnecessary SimplifyCFG iterations, but also makes sure that we don't
3151  // bypass the check for trivial cycles above.
3152  MergeBlockIntoPredecessor(EdgeBB, DTU);
3153 
3154  // Signal repeat, simplifying any other constants.
3155  return None;
3156  }
3157 
3158  return false;
3159 }
3160 
3162  DomTreeUpdater *DTU,
3163  const DataLayout &DL,
3164  AssumptionCache *AC) {
3165  Optional<bool> Result;
3166  bool EverChanged = false;
3167  do {
3168  // Note that None means "we changed things, but recurse further."
3169  Result = FoldCondBranchOnValueKnownInPredecessorImpl(BI, DTU, DL, AC);
3170  EverChanged |= Result == None || *Result;
3171  } while (Result == None);
3172  return EverChanged;
3173 }
3174 
3175 /// Given a BB that starts with the specified two-entry PHI node,
3176 /// see if we can eliminate it.
3178  DomTreeUpdater *DTU, const DataLayout &DL) {
3179  // Ok, this is a two entry PHI node. Check to see if this is a simple "if
3180  // statement", which has a very simple dominance structure. Basically, we
3181  // are trying to find the condition that is being branched on, which
3182  // subsequently causes this merge to happen. We really want control
3183  // dependence information for this check, but simplifycfg can't keep it up
3184  // to date, and this catches most of the cases we care about anyway.
3185  BasicBlock *BB = PN->getParent();
3186 
3187  BasicBlock *IfTrue, *IfFalse;
3188  BranchInst *DomBI = GetIfCondition(BB, IfTrue, IfFalse);
3189  if (!DomBI)
3190  return false;
3191  Value *IfCond = DomBI->getCondition();
3192  // Don't bother if the branch will be constant folded trivially.
3193  if (isa<ConstantInt>(IfCond))
3194  return false;
3195 
3196  BasicBlock *DomBlock = DomBI->getParent();
3198  llvm::copy_if(
3199  PN->blocks(), std::back_inserter(IfBlocks), [](BasicBlock *IfBlock) {
3200  return cast<BranchInst>(IfBlock->getTerminator())->isUnconditional();
3201  });
3202  assert((IfBlocks.size() == 1 || IfBlocks.size() == 2) &&
3203  "Will have either one or two blocks to speculate.");
3204 
3205  // If the branch is non-unpredictable, see if we either predictably jump to
3206  // the merge bb (if we have only a single 'then' block), or if we predictably
3207  // jump to one specific 'then' block (if we have two of them).
3208  // It isn't beneficial to speculatively execute the code
3209  // from the block that we know is predictably not entered.
3210  if (!DomBI->getMetadata(LLVMContext::MD_unpredictable)) {
3211  uint64_t TWeight, FWeight;
3212  if (DomBI->extractProfMetadata(TWeight, FWeight) &&
3213  (TWeight + FWeight) != 0) {
3214  BranchProbability BITrueProb =
3215  BranchProbability::getBranchProbability(TWeight, TWeight + FWeight);
3217  BranchProbability BIFalseProb = BITrueProb.getCompl();
3218  if (IfBlocks.size() == 1) {
3219  BranchProbability BIBBProb =
3220  DomBI->getSuccessor(0) == BB ? BITrueProb : BIFalseProb;
3221  if (BIBBProb >= Likely)
3222  return false;
3223  } else {
3224  if (BITrueProb >= Likely || BIFalseProb >= Likely)
3225  return false;
3226  }
3227  }
3228  }
3229 
3230  // Don't try to fold an unreachable block. For example, the phi node itself
3231  // can't be the candidate if-condition for a select that we want to form.
3232  if (auto *IfCondPhiInst = dyn_cast<PHINode>(IfCond))
3233  if (IfCondPhiInst->getParent() == BB)
3234  return false;
3235 
3236  // Okay, we found that we can merge this two-entry phi node into a select.
3237  // Doing so would require us to fold *all* two entry phi nodes in this block.
3238  // At some point this becomes non-profitable (particularly if the target
3239  // doesn't support cmov's). Only do this transformation if there are two or
3240  // fewer PHI nodes in this block.
3241  unsigned NumPhis = 0;
3242  for (BasicBlock::iterator I = BB->begin(); isa<PHINode>(I); ++NumPhis, ++I)
3243  if (NumPhis > 2)
3244  return false;
3245 
3246  // Loop over the PHI's seeing if we can promote them all to select
3247  // instructions. While we are at it, keep track of the instructions
3248  // that need to be moved to the dominating block.
3249  SmallPtrSet<Instruction *, 4> AggressiveInsts;
3250  InstructionCost Cost = 0;
3251  InstructionCost Budget =
3252  TwoEntryPHINodeFoldingThreshold * TargetTransformInfo::TCC_Basic;
3253 
3254  bool Changed = false;
3255  for (BasicBlock::iterator II = BB->begin(); isa<PHINode>(II);) {
3256  PHINode *PN = cast<PHINode>(II++);
3257  if (Value *V = simplifyInstruction(PN, {DL, PN})) {
3258  PN->replaceAllUsesWith(V);
3259  PN->eraseFromParent();
3260  Changed = true;
3261  continue;
3262  }
3263 
3264  if (!dominatesMergePoint(PN->getIncomingValue(0), BB, AggressiveInsts,
3265  Cost, Budget, TTI) ||
3266  !dominatesMergePoint(PN->getIncomingValue(1), BB, AggressiveInsts,
3267  Cost, Budget, TTI))
3268  return Changed;
3269  }
3270 
3271  // If we folded the first phi, PN dangles at this point. Refresh it. If
3272  // we ran out of PHIs then we simplified them all.
3273  PN = dyn_cast<PHINode>(BB->begin());
3274  if (!PN)
3275  return true;
3276 
3277  // Return true if at least one of these is a 'not', and another is either
3278  // a 'not' too, or a constant.
3279  auto CanHoistNotFromBothValues = [](Value *V0, Value *V1) {
3280  if (!match(V0, m_Not(m_Value())))
3281  std::swap(V0, V1);
3282  auto Invertible = m_CombineOr(m_Not(m_Value()), m_AnyIntegralConstant());
3283  return match(V0, m_Not(m_Value())) && match(V1, Invertible);
3284  };
3285 
3286  // Don't fold i1 branches on PHIs which contain binary operators or
3287  // (possibly inverted) select form of or/ands, unless one of
3288  // the incoming values is an 'not' and another one is freely invertible.
3289  // These can often be turned into switches and other things.
3290  auto IsBinOpOrAnd = [](Value *V) {
3291  return match(
3292  V, m_CombineOr(
3293  m_BinOp(),
3295  m_Select(m_Value(), m_Value(), m_ImmConstant()))));
3296  };
3297  if (PN->getType()->isIntegerTy(1) &&
3298  (IsBinOpOrAnd(PN->getIncomingValue(0)) ||
3299  IsBinOpOrAnd(PN->getIncomingValue(1)) || IsBinOpOrAnd(IfCond)) &&
3300  !CanHoistNotFromBothValues(PN->getIncomingValue(0),
3301  PN->getIncomingValue(1)))
3302  return Changed;
3303 
3304  // If all PHI nodes are promotable, check to make sure that all instructions
3305  // in the predecessor blocks can be promoted as well. If not, we won't be able
3306  // to get rid of the control flow, so it's not worth promoting to select
3307  // instructions.
3308  for (BasicBlock *IfBlock : IfBlocks)
3309  for (BasicBlock::iterator I = IfBlock->begin(); !I->isTerminator(); ++I)
3310  if (!AggressiveInsts.count(&*I) && !I->isDebugOrPseudoInst()) {
3311  // This is not an aggressive instruction that we can promote.
3312  // Because of this, we won't be able to get rid of the control flow, so
3313  // the xform is not worth it.
3314  return Changed;
3315  }
3316 
3317  // If either of the blocks has it's address taken, we can't do this fold.
3318  if (any_of(IfBlocks,
3319  [](BasicBlock *IfBlock) { return IfBlock->hasAddressTaken(); }))
3320  return Changed;
3321 
3322  LLVM_DEBUG(dbgs() << "FOUND IF CONDITION! " << *IfCond
3323  << " T: " << IfTrue->getName()
3324  << " F: " << IfFalse->getName() << "\n");
3325 
3326  // If we can still promote the PHI nodes after this gauntlet of tests,
3327  // do all of the PHI's now.
3328 
3329  // Move all 'aggressive' instructions, which are defined in the
3330  // conditional parts of the if's up to the dominating block.
3331  for (BasicBlock *IfBlock : IfBlocks)
3332  hoistAllInstructionsInto(DomBlock, DomBI, IfBlock);
3333 
3335  // Propagate fast-math-flags from phi nodes to replacement selects.
3337  while (PHINode *PN = dyn_cast<PHINode>(BB->begin())) {
3338  if (isa<FPMathOperator>(PN))
3339  Builder.setFastMathFlags(PN->getFastMathFlags());
3340 
3341  // Change the PHI node into a select instruction.
3342  Value *TrueVal = PN->getIncomingValueForBlock(IfTrue);
3343  Value *FalseVal = PN->getIncomingValueForBlock(IfFalse);
3344 
3345  Value *Sel = Builder.CreateSelect(IfCond, TrueVal, FalseVal, "", DomBI);
3346  PN->replaceAllUsesWith(Sel);
3347  Sel->takeName(PN);
3348  PN->eraseFromParent();
3349  }
3350 
3351  // At this point, all IfBlocks are empty, so our if statement
3352  // has been flattened. Change DomBlock to jump directly to our new block to
3353  // avoid other simplifycfg's kicking in on the diamond.
3354  Builder.CreateBr(BB);
3355 
3357  if (DTU) {
3358  Updates.push_back({DominatorTree::Insert, DomBlock, BB});
3359  for (auto *Successor : successors(DomBlock))
3360  Updates.push_back({DominatorTree::Delete, DomBlock, Successor});
3361  }
3362 
3363  DomBI->eraseFromParent();
3364  if (DTU)
3365  DTU->applyUpdates(Updates);
3366 
3367  return true;
3368 }
3369 
3372  Value *RHS, const Twine &Name = "") {
3373  // Try to relax logical op to binary op.
3374  if (impliesPoison(RHS, LHS))
3375  return Builder.CreateBinOp(Opc, LHS, RHS, Name);
3376  if (Opc == Instruction::And)
3377  return Builder.CreateLogicalAnd(LHS, RHS, Name);
3378  if (Opc == Instruction::Or)
3379  return Builder.CreateLogicalOr(LHS, RHS, Name);
3380  llvm_unreachable("Invalid logical opcode");
3381 }
3382 
3383 /// Return true if either PBI or BI has branch weight available, and store
3384 /// the weights in {Pred|Succ}{True|False}Weight. If one of PBI and BI does
3385 /// not have branch weight, use 1:1 as its weight.
3387  uint64_t &PredTrueWeight,
3388  uint64_t &PredFalseWeight,
3389  uint64_t &SuccTrueWeight,
3390  uint64_t &SuccFalseWeight) {
3391  bool PredHasWeights =
3392  PBI->extractProfMetadata(PredTrueWeight, PredFalseWeight);
3393  bool SuccHasWeights =
3394  BI->extractProfMetadata(SuccTrueWeight, SuccFalseWeight);
3395  if (PredHasWeights || SuccHasWeights) {
3396  if (!PredHasWeights)
3397  PredTrueWeight = PredFalseWeight = 1;
3398  if (!SuccHasWeights)
3399  SuccTrueWeight = SuccFalseWeight = 1;
3400  return true;
3401  } else {
3402  return false;
3403  }
3404 }
3405 
3406 /// Determine if the two branches share a common destination and deduce a glue
3407 /// that joins the branches' conditions to arrive at the common destination if
3408 /// that would be profitable.
3411  const TargetTransformInfo *TTI) {
3412  assert(BI && PBI && BI->isConditional() && PBI->isConditional() &&
3413  "Both blocks must end with a conditional branches.");
3415  "PredBB must be a predecessor of BB.");
3416 
3417  // We have the potential to fold the conditions together, but if the
3418  // predecessor branch is predictable, we may not want to merge them.
3419  uint64_t PTWeight, PFWeight;
3420  BranchProbability PBITrueProb, Likely;
3421  if (TTI && !PBI->getMetadata(LLVMContext::MD_unpredictable) &&
3422  PBI->extractProfMetadata(PTWeight, PFWeight) &&
3423  (PTWeight + PFWeight) != 0) {
3424  PBITrueProb =
3425  BranchProbability::getBranchProbability(PTWeight, PTWeight + PFWeight);
3426  Likely = TTI->getPredictableBranchThreshold();
3427  }
3428 
3429  if (PBI->getSuccessor(0) == BI->getSuccessor(0)) {
3430  // Speculate the 2nd condition unless the 1st is probably true.
3431  if (PBITrueProb.isUnknown() || PBITrueProb < Likely)
3432  return {{Instruction::Or, false}};
3433  } else if (PBI->getSuccessor(1) == BI->getSuccessor(1)) {
3434  // Speculate the 2nd condition unless the 1st is probably false.
3435  if (PBITrueProb.isUnknown() || PBITrueProb.getCompl() < Likely)
3436  return {{Instruction::And, false}};
3437  } else if (PBI->getSuccessor(0) == BI->getSuccessor(1)) {
3438  // Speculate the 2nd condition unless the 1st is probably true.
3439  if (PBITrueProb.isUnknown() || PBITrueProb < Likely)
3440  return {{Instruction::And, true}};
3441  } else if (PBI->getSuccessor(1) == BI->getSuccessor(0)) {
3442  // Speculate the 2nd condition unless the 1st is probably false.
3443  if (PBITrueProb.isUnknown() || PBITrueProb.getCompl() < Likely)
3444  return {{Instruction::Or, true}};
3445  }
3446  return None;
3447 }
3448 
3450  DomTreeUpdater *DTU,
3451  MemorySSAUpdater *MSSAU,
3452  const TargetTransformInfo *TTI) {
3453  BasicBlock *BB = BI->getParent();
3454  BasicBlock *PredBlock = PBI->getParent();
3455 
3456  // Determine if the two branches share a common destination.
3458  bool InvertPredCond;
3459  std::tie(Opc, InvertPredCond) =
3461 
3462  LLVM_DEBUG(dbgs() << "FOLDING BRANCH TO COMMON DEST:\n" << *PBI << *BB);
3463 
3464  IRBuilder<> Builder(PBI);
3465  // The builder is used to create instructions to eliminate the branch in BB.
3466  // If BB's terminator has !annotation metadata, add it to the new
3467  // instructions.
3468  Builder.CollectMetadataToCopy(BB->getTerminator(),
3469  {LLVMContext::MD_annotation});
3470 
3471  // If we need to invert the condition in the pred block to match, do so now.
3472  if (InvertPredCond) {
3473  Value *NewCond = PBI->getCondition();
3474  if (NewCond->hasOneUse() && isa<CmpInst>(NewCond)) {
3475  CmpInst *CI = cast<CmpInst>(NewCond);
3476  CI->setPredicate(CI->getInversePredicate());
3477  } else {
3478  NewCond =
3479  Builder.CreateNot(NewCond, PBI->getCondition()->getName() + ".not");
3480  }
3481 
3482  PBI->setCondition(NewCond);
3483  PBI->swapSuccessors();
3484  }
3485 
3486  BasicBlock *UniqueSucc =
3487  PBI->getSuccessor(0) == BB ? BI->getSuccessor(0) : BI->getSuccessor(1);
3488 
3489  // Before cloning instructions, notify the successor basic block that it
3490  // is about to have a new predecessor. This will update PHI nodes,
3491  // which will allow us to update live-out uses of bonus instructions.
3492  AddPredecessorToBlock(UniqueSucc, PredBlock, BB, MSSAU);
3493 
3494  // Try to update branch weights.
3495  uint64_t PredTrueWeight, PredFalseWeight, SuccTrueWeight, SuccFalseWeight;
3496  if (extractPredSuccWeights(PBI, BI, PredTrueWeight, PredFalseWeight,
3497  SuccTrueWeight, SuccFalseWeight)) {
3498  SmallVector<uint64_t, 8> NewWeights;
3499 
3500  if (PBI->getSuccessor(0) == BB) {
3501  // PBI: br i1 %x, BB, FalseDest
3502  // BI: br i1 %y, UniqueSucc, FalseDest
3503  // TrueWeight is TrueWeight for PBI * TrueWeight for BI.
3504  NewWeights.push_back(PredTrueWeight * SuccTrueWeight);
3505  // FalseWeight is FalseWeight for PBI * TotalWeight for BI +
3506  // TrueWeight for PBI * FalseWeight for BI.
3507  // We assume that total weights of a BranchInst can fit into 32 bits.
3508  // Therefore, we will not have overflow using 64-bit arithmetic.
3509  NewWeights.push_back(PredFalseWeight *
3510  (SuccFalseWeight + SuccTrueWeight) +
3511  PredTrueWeight * SuccFalseWeight);
3512  } else {
3513  // PBI: br i1 %x, TrueDest, BB
3514  // BI: br i1 %y, TrueDest, UniqueSucc
3515  // TrueWeight is TrueWeight for PBI * TotalWeight for BI +
3516  // FalseWeight for PBI * TrueWeight for BI.
3517  NewWeights.push_back(PredTrueWeight * (SuccFalseWeight + SuccTrueWeight) +
3518  PredFalseWeight * SuccTrueWeight);
3519  // FalseWeight is FalseWeight for PBI * FalseWeight for BI.
3520  NewWeights.push_back(PredFalseWeight * SuccFalseWeight);
3521  }
3522 
3523  // Halve the weights if any of them cannot fit in an uint32_t
3524  FitWeights(NewWeights);
3525 
3526  SmallVector<uint32_t, 8> MDWeights(NewWeights.begin(), NewWeights.end());
3527  setBranchWeights(PBI, MDWeights[0], MDWeights[1]);
3528 
3529  // TODO: If BB is reachable from all paths through PredBlock, then we
3530  // could replace PBI's branch probabilities with BI's.
3531  } else
3532  PBI->setMetadata(LLVMContext::MD_prof, nullptr);
3533 
3534  // Now, update the CFG.
3535  PBI->setSuccessor(PBI->getSuccessor(0) != BB, UniqueSucc);
3536 
3537  if (DTU)
3538  DTU->applyUpdates({{DominatorTree::Insert, PredBlock, UniqueSucc},
3539  {DominatorTree::Delete, PredBlock, BB}});
3540 
3541  // If BI was a loop latch, it may have had associated loop metadata.
3542  // We need to copy it to the new latch, that is, PBI.
3543  if (MDNode *LoopMD = BI->getMetadata(LLVMContext::MD_loop))
3544  PBI->setMetadata(LLVMContext::MD_loop, LoopMD);
3545 
3546  ValueToValueMapTy VMap; // maps original values to cloned values
3548 
3549  // Now that the Cond was cloned into the predecessor basic block,
3550  // or/and the two conditions together.
3551  Value *BICond = VMap[BI->getCondition()];
3552  PBI->setCondition(
3553  createLogicalOp(Builder, Opc, PBI->getCondition(), BICond, "or.cond"));
3554 
3555  // Copy any debug value intrinsics into the end of PredBlock.
3556  for (Instruction &I : *BB) {
3557  if (isa<DbgInfoIntrinsic>(I)) {
3558  Instruction *NewI = I.clone();
3559  RemapInstruction(NewI, VMap,
3561  NewI->insertBefore(PBI);
3562  }
3563  }
3564 
3565  ++NumFoldBranchToCommonDest;
3566  return true;
3567 }
3568 
3569 /// Return if an instruction's type or any of its operands' types are a vector
3570 /// type.
3571 static bool isVectorOp(Instruction &I) {
3572  return I.getType()->isVectorTy() || any_of(I.operands(), [](Use &U) {
3573  return U->getType()->isVectorTy();
3574  });
3575 }
3576 
3577 /// If this basic block is simple enough, and if a predecessor branches to us
3578 /// and one of our successors, fold the block into the predecessor and use
3579 /// logical operations to pick the right destination.
3581  MemorySSAUpdater *MSSAU,
3582  const TargetTransformInfo *TTI,
3583  unsigned BonusInstThreshold) {
3584  // If this block ends with an unconditional branch,
3585  // let SpeculativelyExecuteBB() deal with it.
3586  if (!BI->isConditional())
3587  return false;
3588 
3589  BasicBlock *BB = BI->getParent();
3591  BB->getParent()->hasMinSize() ? TargetTransformInfo::TCK_CodeSize
3592  : TargetTransformInfo::TCK_SizeAndLatency;
3593 
3594  Instruction *Cond = dyn_cast<Instruction>(BI->getCondition());
3595 
3596  if (!Cond ||
3597  (!isa<CmpInst>(Cond) && !isa<BinaryOperator>(Cond) &&
3598  !isa<SelectInst>(Cond)) ||
3599  Cond->getParent() != BB || !Cond->hasOneUse())
3600  return false;
3601 
3602  // Cond is known to be a compare or binary operator. Check to make sure that
3603  // neither operand is a potentially-trapping constant expression.
3604  if (canTrap(Cond->getOperand(0)))
3605  return false;
3606  if (canTrap(Cond->getOperand(1)))
3607  return false;
3608 
3609  // Finally, don't infinitely unroll conditional loops.
3610  if (is_contained(successors(BB), BB))
3611  return false;
3612 
3613  // With which predecessors will we want to deal with?
3615  for (BasicBlock *PredBlock : predecessors(BB)) {
3616  BranchInst *PBI = dyn_cast<BranchInst>(PredBlock->getTerminator());
3617 
3618  // Check that we have two conditional branches. If there is a PHI node in
3619  // the common successor, verify that the same value flows in from both
3620  // blocks.
3621  if (!PBI || PBI->isUnconditional() || !SafeToMergeTerminators(BI, PBI))
3622  continue;
3623 
3624  // Determine if the two branches share a common destination.
3626  bool InvertPredCond;
3627  if (auto Recipe = shouldFoldCondBranchesToCommonDestination(BI, PBI, TTI))
3628  std::tie(Opc, InvertPredCond) = *Recipe;
3629  else
3630  continue;
3631 
3632  // Check the cost of inserting the necessary logic before performing the
3633  // transformation.
3634  if (TTI) {
3635  Type *Ty = BI->getCondition()->getType();
3637  if (InvertPredCond && (!PBI->getCondition()->hasOneUse() ||
3638  !isa<CmpInst>(PBI->getCondition())))
3639  Cost += TTI->getArithmeticInstrCost(Instruction::Xor, Ty, CostKind);
3640 
3641  if (Cost > BranchFoldThreshold)
3642  continue;
3643  }
3644 
3645  // Ok, we do want to deal with this predecessor. Record it.
3646  Preds.emplace_back(PredBlock);
3647  }
3648 
3649  // If there aren't any predecessors into which we can fold,
3650  // don't bother checking the cost.
3651  if (Preds.empty())
3652  return false;
3653 
3654  // Only allow this transformation if computing the condition doesn't involve
3655  // too many instructions and these involved instructions can be executed
3656  // unconditionally. We denote all involved instructions except the condition
3657  // as "bonus instructions", and only allow this transformation when the
3658  // number of the bonus instructions we'll need to create when cloning into
3659  // each predecessor does not exceed a certain threshold.
3660  unsigned NumBonusInsts = 0;
3661  bool SawVectorOp = false;
3662  const unsigned PredCount = Preds.size();
3663  for (Instruction &I : *BB) {
3664  // Don't check the branch condition comparison itself.
3665  if (&I == Cond)
3666  continue;
3667  // Ignore dbg intrinsics, and the terminator.
3668  if (isa<DbgInfoIntrinsic>(I) || isa<BranchInst>(I))
3669  continue;
3670  // I must be safe to execute unconditionally.
3672  return false;
3673  SawVectorOp |= isVectorOp(I);
3674 
3675  // Account for the cost of duplicating this instruction into each
3676  // predecessor. Ignore free instructions.
3677  if (!TTI ||
3678  TTI->getUserCost(&I, CostKind) != TargetTransformInfo::TCC_Free) {
3679  NumBonusInsts += PredCount;
3680 
3681  // Early exits once we reach the limit.
3682  if (NumBonusInsts >
3683  BonusInstThreshold * BranchFoldToCommonDestVectorMultiplier)
3684  return false;
3685  }
3686 
3687  auto IsBCSSAUse = [BB, &I](Use &U) {
3688  auto *UI = cast<Instruction>(U.getUser());
3689  if (auto *PN = dyn_cast<PHINode>(UI))
3690  return PN->getIncomingBlock(U) == BB;
3691  return UI->getParent() == BB && I.comesBefore(UI);
3692  };
3693 
3694  // Does this instruction require rewriting of uses?
3695  if (!all_of(I.uses(), IsBCSSAUse))
3696  return false;
3697  }
3698  if (NumBonusInsts >
3699  BonusInstThreshold *
3700  (SawVectorOp ? BranchFoldToCommonDestVectorMultiplier : 1))
3701  return false;
3702 
3703  // Ok, we have the budget. Perform the transformation.
3704  for (BasicBlock *PredBlock : Preds) {
3705  auto *PBI = cast<BranchInst>(PredBlock->getTerminator());
3706  return performBranchToCommonDestFolding(BI, PBI, DTU, MSSAU, TTI);
3707  }
3708  return false;
3709 }
3710 
3711 // If there is only one store in BB1 and BB2, return it, otherwise return
3712 // nullptr.
3714  StoreInst *S = nullptr;
3715  for (auto *BB : {BB1, BB2}) {
3716  if (!BB)
3717  continue;
3718  for (auto &I : *BB)
3719  if (auto *SI = dyn_cast<StoreInst>(&I)) {
3720  if (S)
3721  // Multiple stores seen.
3722  return nullptr;
3723  else
3724  S = SI;
3725  }
3726  }
3727  return S;
3728 }
3729 
3731  Value *AlternativeV = nullptr) {
3732  // PHI is going to be a PHI node that allows the value V that is defined in
3733  // BB to be referenced in BB's only successor.
3734  //
3735  // If AlternativeV is nullptr, the only value we care about in PHI is V. It
3736  // doesn't matter to us what the other operand is (it'll never get used). We
3737  // could just create a new PHI with an undef incoming value, but that could
3738  // increase register pressure if EarlyCSE/InstCombine can't fold it with some
3739  // other PHI. So here we directly look for some PHI in BB's successor with V
3740  // as an incoming operand. If we find one, we use it, else we create a new
3741  // one.
3742  //
3743  // If AlternativeV is not nullptr, we care about both incoming values in PHI.
3744  // PHI must be exactly: phi <ty> [ %BB, %V ], [ %OtherBB, %AlternativeV]
3745  // where OtherBB is the single other predecessor of BB's only successor.
3746  PHINode *PHI = nullptr;
3747  BasicBlock *Succ = BB->getSingleSuccessor();
3748 
3749  for (auto I = Succ->begin(); isa<PHINode>(I); ++I)
3750  if (cast<PHINode>(I)->getIncomingValueForBlock(BB) == V) {
3751  PHI = cast<PHINode>(I);
3752  if (!AlternativeV)
3753  break;
3754 
3755  assert(Succ->hasNPredecessors(2));
3756  auto PredI = pred_begin(Succ);
3757  BasicBlock *OtherPredBB = *PredI == BB ? *++PredI : *PredI;
3758  if (PHI->getIncomingValueForBlock(OtherPredBB) == AlternativeV)
3759  break;
3760  PHI = nullptr;
3761  }
3762  if (PHI)
3763  return PHI;
3764 
3765  // If V is not an instruction defined in BB, just return it.
3766  if (!AlternativeV &&
3767  (!isa<Instruction>(V) || cast<Instruction>(V)->getParent() != BB))
3768  return V;
3769 
3770  PHI = PHINode::Create(V->getType(), 2, "simplifycfg.merge", &Succ->front());
3771  PHI->addIncoming(V, BB);
3772  for (BasicBlock *PredBB : predecessors(Succ))
3773  if (PredBB != BB)
3774  PHI->addIncoming(
3775  AlternativeV ? AlternativeV : UndefValue::get(V->getType()), PredBB);
3776  return PHI;
3777 }
3778 
3780  BasicBlock *PTB, BasicBlock *PFB, BasicBlock *QTB, BasicBlock *QFB,
3781  BasicBlock *PostBB, Value *Address, bool InvertPCond, bool InvertQCond,
3782  DomTreeUpdater *DTU, const DataLayout &DL, const TargetTransformInfo &TTI) {
3783  // For every pointer, there must be exactly two stores, one coming from
3784  // PTB or PFB, and the other from QTB or QFB. We don't support more than one
3785  // store (to any address) in PTB,PFB or QTB,QFB.
3786  // FIXME: We could relax this restriction with a bit more work and performance
3787  // testing.
3788  StoreInst *PStore = findUniqueStoreInBlocks(PTB, PFB);
3789  StoreInst *QStore = findUniqueStoreInBlocks(QTB, QFB);
3790  if (!PStore || !QStore)
3791  return false;
3792 
3793  // Now check the stores are compatible.
3794  if (!QStore->isUnordered() || !PStore->isUnordered() ||
3795  PStore->getValueOperand()->getType() !=
3796  QStore->getValueOperand()->getType())
3797  return false;
3798 
3799  // Check that sinking the store won't cause program behavior changes. Sinking
3800  // the store out of the Q blocks won't change any behavior as we're sinking
3801  // from a block to its unconditional successor. But we're moving a store from
3802  // the P blocks down through the middle block (QBI) and past both QFB and QTB.
3803  // So we need to check that there are no aliasing loads or stores in
3804  // QBI, QTB and QFB. We also need to check there are no conflicting memory
3805  // operations between PStore and the end of its parent block.
3806  //
3807  // The ideal way to do this is to query AliasAnalysis, but we don't
3808  // preserve AA currently so that is dangerous. Be super safe and just
3809  // check there are no other memory operations at all.
3810  for (auto &I : *QFB->getSinglePredecessor())
3811  if (I.mayReadOrWriteMemory())
3812  return false;
3813  for (auto &I : *QFB)
3814  if (&I != QStore && I.mayReadOrWriteMemory())
3815  return false;
3816  if (QTB)
3817  for (auto &I : *QTB)
3818  if (&I != QStore && I.mayReadOrWriteMemory())
3819  return false;
3820  for (auto I = BasicBlock::iterator(PStore), E = PStore->getParent()->end();
3821  I != E; ++I)
3822  if (&*I != PStore && I->mayReadOrWriteMemory())
3823  return false;
3824 
3825  // If we're not in aggressive mode, we only optimize if we have some
3826  // confidence that by optimizing we'll allow P and/or Q to be if-converted.
3827  auto IsWorthwhile = [&](BasicBlock *BB, ArrayRef<StoreInst *> FreeStores) {
3828  if (!BB)
3829  return true;
3830  // Heuristic: if the block can be if-converted/phi-folded and the
3831  // instructions inside are all cheap (arithmetic/GEPs), it's worthwhile to
3832  // thread this store.
3833  InstructionCost Cost = 0;
3834  InstructionCost Budget =
3835  PHINodeFoldingThreshold * TargetTransformInfo::TCC_Basic;
3836  for (auto &I : BB->instructionsWithoutDebug(false)) {
3837  // Consider terminator instruction to be free.
3838  if (I.isTerminator())
3839  continue;
3840  // If this is one the stores that we want to speculate out of this BB,
3841  // then don't count it's cost, consider it to be free.
3842  if (auto *S = dyn_cast<StoreInst>(&I))
3843  if (llvm::find(FreeStores, S))
3844  continue;
3845  // Else, we have a white-list of instructions that we are ak speculating.
3846  if (!isa<BinaryOperator>(I) && !isa<GetElementPtrInst>(I))
3847  return false; // Not in white-list - not worthwhile folding.
3848  // And finally, if this is a non-free instruction that we are okay
3849  // speculating, ensure that we consider the speculation budget.
3850  Cost += TTI.getUserCost(&I, TargetTransformInfo::TCK_SizeAndLatency);
3851  if (Cost > Budget)
3852  return false; // Eagerly refuse to fold as soon as we're out of budget.
3853  }
3854  assert(Cost <= Budget &&
3855  "When we run out of budget we will eagerly return from within the "
3856  "per-instruction loop.");
3857  return true;
3858  };
3859 
3860  const std::array<StoreInst *, 2> FreeStores = {PStore, QStore};
3862  (!IsWorthwhile(PTB, FreeStores) || !IsWorthwhile(PFB, FreeStores) ||
3863  !IsWorthwhile(QTB, FreeStores) || !IsWorthwhile(QFB, FreeStores)))
3864  return false;
3865 
3866  // If PostBB has more than two predecessors, we need to split it so we can
3867  // sink the store.
3868  if (std::next(pred_begin(PostBB), 2) != pred_end(PostBB)) {
3869  // We know that QFB's only successor is PostBB. And QFB has a single
3870  // predecessor. If QTB exists, then its only successor is also PostBB.
3871  // If QTB does not exist, then QFB's only predecessor has a conditional
3872  // branch to QFB and PostBB.
3873  BasicBlock *TruePred = QTB ? QTB : QFB->getSinglePredecessor();
3874  BasicBlock *NewBB =
3875  SplitBlockPredecessors(PostBB, {QFB, TruePred}, "condstore.split", DTU);
3876  if (!NewBB)
3877  return false;
3878  PostBB = NewBB;
3879  }
3880 
3881  // OK, we're going to sink the stores to PostBB. The store has to be
3882  // conditional though, so first create the predicate.
3883  Value *PCond = cast<BranchInst>(PFB->getSinglePredecessor()->getTerminator())
3884  ->getCondition();
3885  Value *QCond = cast<BranchInst>(QFB->getSinglePredecessor()->getTerminator())
3886  ->getCondition();
3887 
3889  PStore->getParent());
3891  QStore->getParent(), PPHI);
3892 
3893  IRBuilder<> QB(&*PostBB->getFirstInsertionPt());
3894 
3895  Value *PPred = PStore->getParent() == PTB ? PCond : QB.CreateNot(PCond);
3896  Value *QPred = QStore->getParent() == QTB ? QCond : QB.CreateNot(QCond);
3897 
3898  if (InvertPCond)
3899  PPred = QB.CreateNot(PPred);
3900  if (InvertQCond)
3901  QPred = QB.CreateNot(QPred);
3902  Value *CombinedPred = QB.CreateOr(PPred, QPred);
3903 
3904  auto *T = SplitBlockAndInsertIfThen(CombinedPred, &*QB.GetInsertPoint(),
3905  /*Unreachable=*/false,
3906  /*BranchWeights=*/nullptr, DTU);
3907  QB.SetInsertPoint(T);
3908  StoreInst *SI = cast<StoreInst>(QB.CreateStore(QPHI, Address));
3909  SI->setAAMetadata(PStore->getAAMetadata().merge(QStore->getAAMetadata()));
3910  // Choose the minimum alignment. If we could prove both stores execute, we
3911  // could use biggest one. In this case, though, we only know that one of the
3912  // stores executes. And we don't know it's safe to take the alignment from a
3913  // store that doesn't execute.
3914  SI->setAlignment(std::min(PStore->getAlign(), QStore->getAlign()));
3915 
3916  QStore->eraseFromParent();
3917  PStore->eraseFromParent();
3918 
3919  return true;
3920 }
3921 
3923  DomTreeUpdater *DTU, const DataLayout &DL,
3924  const TargetTransformInfo &TTI) {
3925  // The intention here is to find diamonds or triangles (see below) where each
3926  // conditional block contains a store to the same address. Both of these
3927  // stores are conditional, so they can't be unconditionally sunk. But it may
3928  // be profitable to speculatively sink the stores into one merged store at the
3929  // end, and predicate the merged store on the union of the two conditions of
3930  // PBI and QBI.
3931  //
3932  // This can reduce the number of stores executed if both of the conditions are
3933  // true, and can allow the blocks to become small enough to be if-converted.
3934  // This optimization will also chain, so that ladders of test-and-set
3935  // sequences can be if-converted away.
3936  //
3937  // We only deal with simple diamonds or triangles:
3938  //
3939  // PBI or PBI or a combination of the two
3940  // / \ | \
3941  // PTB PFB | PFB
3942  // \ / | /
3943  // QBI QBI
3944  // / \ | \
3945  // QTB QFB | QFB
3946  // \ / | /
3947  // PostBB PostBB
3948  //
3949  // We model triangles as a type of diamond with a nullptr "true" block.
3950  // Triangles are canonicalized so that the fallthrough edge is represented by
3951  // a true condition, as in the diagram above.
3952  BasicBlock *PTB = PBI->getSuccessor(0);
3953  BasicBlock *PFB = PBI->getSuccessor(1);
3954  BasicBlock *QTB = QBI->getSuccessor(0);
3955  BasicBlock *QFB = QBI->getSuccessor(1);
3956  BasicBlock *PostBB = QFB->getSingleSuccessor();
3957 
3958  // Make sure we have a good guess for PostBB. If QTB's only successor is
3959  // QFB, then QFB is a better PostBB.
3960  if (QTB->getSingleSuccessor() == QFB)
3961  PostBB = QFB;
3962 
3963  // If we couldn't find a good PostBB, stop.
3964  if (!PostBB)
3965  return false;
3966 
3967  bool InvertPCond = false, InvertQCond = false;
3968  // Canonicalize fallthroughs to the true branches.
3969  if (PFB == QBI->getParent()) {
3970  std::swap(PFB, PTB);
3971  InvertPCond = true;
3972  }
3973  if (QFB == PostBB) {
3974  std::swap(QFB, QTB);
3975  InvertQCond = true;
3976  }
3977 
3978  // From this point on we can assume PTB or QTB may be fallthroughs but PFB
3979  // and QFB may not. Model fallthroughs as a nullptr block.
3980  if (PTB == QBI->getParent())
3981  PTB = nullptr;
3982  if (QTB == PostBB)
3983  QTB = nullptr;
3984 
3985  // Legality bailouts. We must have at least the non-fallthrough blocks and
3986  // the post-dominating block, and the non-fallthroughs must only have one
3987  // predecessor.
3988  auto HasOnePredAndOneSucc = [](BasicBlock *BB, BasicBlock *P, BasicBlock *S) {
3989  return BB->getSinglePredecessor() == P && BB->getSingleSuccessor() == S;
3990  };
3991  if (!HasOnePredAndOneSucc(PFB, PBI->getParent(), QBI->getParent()) ||
3992  !HasOnePredAndOneSucc(QFB, QBI->getParent(), PostBB))
3993  return false;
3994  if ((PTB && !HasOnePredAndOneSucc(PTB, PBI->getParent(), QBI->getParent())) ||
3995  (QTB && !HasOnePredAndOneSucc(QTB, QBI->getParent(), PostBB)))
3996  return false;
3997  if (!QBI->getParent()->hasNUses(2))
3998  return false;
3999 
4000  // OK, this is a sequence of two diamonds or triangles.
4001  // Check if there are stores in PTB or PFB that are repeated in QTB or QFB.
4002  SmallPtrSet<Value *, 4> PStoreAddresses, QStoreAddresses;
4003  for (auto *BB : {PTB, PFB}) {
4004  if (!BB)
4005  continue;
4006  for (auto &I : *BB)
4007  if (StoreInst *SI = dyn_cast<StoreInst>(&I))
4008  PStoreAddresses.insert(SI->getPointerOperand());
4009  }
4010  for (auto *BB : {QTB, QFB}) {
4011  if (!BB)
4012  continue;
4013  for (auto &I : *BB)
4014  if (StoreInst *SI = dyn_cast<StoreInst>(&I))
4015  QStoreAddresses.insert(SI->getPointerOperand());
4016  }
4017 
4018  set_intersect(PStoreAddresses, QStoreAddresses);
4019  // set_intersect mutates PStoreAddresses in place. Rename it here to make it
4020  // clear what it contains.
4021  auto &CommonAddresses = PStoreAddresses;
4022 
4023  bool Changed = false;
4024  for (auto *Address : CommonAddresses)
4025  Changed |=
4026  mergeConditionalStoreToAddress(PTB, PFB, QTB, QFB, PostBB, Address,
4027  InvertPCond, InvertQCond, DTU, DL, TTI);
4028  return Changed;
4029 }
4030 
4031 /// If the previous block ended with a widenable branch, determine if reusing
4032 /// the target block is profitable and legal. This will have the effect of
4033 /// "widening" PBI, but doesn't require us to reason about hosting safety.
4035  DomTreeUpdater *DTU) {
4036  // TODO: This can be generalized in two important ways:
4037  // 1) We can allow phi nodes in IfFalseBB and simply reuse all the input
4038  // values from the PBI edge.
4039  // 2) We can sink side effecting instructions into BI's fallthrough
4040  // successor provided they doesn't contribute to computation of
4041  // BI's condition.
4042  Value *CondWB, *WC;
4043  BasicBlock *IfTrueBB, *IfFalseBB;
4044  if (!parseWidenableBranch(PBI, CondWB, WC, IfTrueBB, IfFalseBB) ||
4045  IfTrueBB != BI->getParent() || !BI->getParent()->getSinglePredecessor())
4046  return false;
4047  if (!IfFalseBB->phis().empty())
4048  return false; // TODO
4049  // Use lambda to lazily compute expensive condition after cheap ones.
4050  auto NoSideEffects = [](BasicBlock &BB) {
4051  return llvm::none_of(BB, [](const Instruction &I) {
4052  return I.mayWriteToMemory() || I.mayHaveSideEffects();
4053  });
4054  };
4055  if (BI->getSuccessor(1) != IfFalseBB && // no inf looping
4056  BI->getSuccessor(1)->getTerminatingDeoptimizeCall() && // profitability
4057  NoSideEffects(*BI->getParent())) {
4058  auto *OldSuccessor = BI->getSuccessor(1);
4059  OldSuccessor->removePredecessor(BI->getParent());
4060  BI->setSuccessor(1, IfFalseBB);
4061  if (DTU)
4062  DTU->applyUpdates(
4063  {{DominatorTree::Insert, BI->getParent(), IfFalseBB},
4064  {DominatorTree::Delete, BI->getParent(), OldSuccessor}});
4065  return true;
4066  }
4067  if (BI->getSuccessor(0) != IfFalseBB && // no inf looping
4068  BI->getSuccessor(0)->getTerminatingDeoptimizeCall() && // profitability
4069  NoSideEffects(*BI->getParent())) {
4070  auto *OldSuccessor = BI->getSuccessor(0);
4071  OldSuccessor->removePredecessor(BI->getParent());
4072  BI->setSuccessor(0, IfFalseBB);
4073  if (DTU)
4074  DTU->applyUpdates(
4075  {{DominatorTree::Insert, BI->getParent(), IfFalseBB},
4076  {DominatorTree::Delete, BI->getParent(), OldSuccessor}});
4077  return true;
4078  }
4079  return false;
4080 }
4081 
4082 /// If we have a conditional branch as a predecessor of another block,
4083 /// this function tries to simplify it. We know
4084 /// that PBI and BI are both conditional branches, and BI is in one of the
4085 /// successor blocks of PBI - PBI branches to BI.
4087  DomTreeUpdater *DTU,
4088  const DataLayout &DL,
4089  const TargetTransformInfo &TTI) {
4090  assert(PBI->isConditional() && BI->isConditional());
4091  BasicBlock *BB = BI->getParent();
4092 
4093  // If this block ends with a branch instruction, and if there is a
4094  // predecessor that ends on a branch of the same condition, make
4095  // this conditional branch redundant.
4096  if (PBI->getCondition() == BI->getCondition() &&
4097  PBI->getSuccessor(0) != PBI->getSuccessor(1)) {
4098  // Okay, the outcome of this conditional branch is statically
4099  // knowable. If this block had a single pred, handle specially, otherwise
4100  // FoldCondBranchOnValueKnownInPredecessor() will handle it.
4101  if (BB->getSinglePredecessor()) {
4102  // Turn this into a branch on constant.
4103  bool CondIsTrue = PBI->getSuccessor(0) == BB;
4104  BI->setCondition(
4105  ConstantInt::get(Type::getInt1Ty(BB->getContext()), CondIsTrue));
4106  return true; // Nuke the branch on constant.
4107  }
4108  }
4109 
4110  // If the previous block ended with a widenable branch, determine if reusing
4111  // the target block is profitable and legal. This will have the effect of
4112  // "widening" PBI, but doesn't require us to reason about hosting safety.
4113  if (tryWidenCondBranchToCondBranch(PBI, BI, DTU))
4114  return true;
4115 
4116  if (canTrap(BI->getCondition()))
4117  return false;
4118 
4119  // If both branches are conditional and both contain stores to the same
4120  // address, remove the stores from the conditionals and create a conditional
4121  // merged store at the end.
4122  if (MergeCondStores && mergeConditionalStores(PBI, BI, DTU, DL, TTI))
4123  return true;
4124 
4125  // If this is a conditional branch in an empty block, and if any
4126  // predecessors are a conditional branch to one of our destinations,
4127  // fold the conditions into logical ops and one cond br.
4128 
4129  // Ignore dbg intrinsics.
4130  if (&*BB->instructionsWithoutDebug(false).begin() != BI)
4131  return false;
4132 
4133  int PBIOp, BIOp;
4134  if (PBI->getSuccessor(0) == BI->getSuccessor(0)) {
4135  PBIOp = 0;
4136  BIOp = 0;
4137  } else if (PBI->getSuccessor(0) == BI->getSuccessor(1)) {
4138  PBIOp = 0;
4139  BIOp = 1;
4140  } else if (PBI->getSuccessor(1) == BI->getSuccessor(0)) {
4141  PBIOp = 1;
4142  BIOp = 0;
4143  } else if (PBI->getSuccessor(1) == BI->getSuccessor(1)) {
4144  PBIOp = 1;
4145  BIOp = 1;
4146  } else {
4147  return false;
4148  }
4149 
4150  // Check to make sure that the other destination of this branch
4151  // isn't BB itself. If so, this is an infinite loop that will
4152  // keep getting unwound.
4153  if (PBI->getSuccessor(PBIOp) == BB)
4154  return false;
4155 
4156  // Do not perform this transformation if it would require
4157  // insertion of a large number of select instructions. For targets
4158  // without predication/cmovs, this is a big pessimization.
4159 
4160  // Also do not perform this transformation if any phi node in the common
4161  // destination block can trap when reached by BB or PBB (PR17073). In that
4162  // case, it would be unsafe to hoist the operation into a select instruction.
4163 
4164  BasicBlock *CommonDest = PBI->getSuccessor(PBIOp);
4165  BasicBlock *RemovedDest = PBI->getSuccessor(PBIOp ^ 1);
4166  unsigned NumPhis = 0;
4167  for (BasicBlock::iterator II = CommonDest->begin(); isa<PHINode>(II);
4168  ++II, ++NumPhis) {
4169  if (NumPhis > 2) // Disable this xform.
4170  return false;
4171 
4172  PHINode *PN = cast<PHINode>(II);
4173  Value *BIV = PN->getIncomingValueForBlock(BB);
4174  if (canTrap(BIV))
4175  return false;
4176 
4177  unsigned PBBIdx = PN->getBasicBlockIndex(PBI->getParent());
4178  Value *PBIV = PN->getIncomingValue(PBBIdx);
4179  if (canTrap(PBIV))
4180  return false;
4181  }
4182 
4183  // Finally, if everything is ok, fold the branches to logical ops.
4184  BasicBlock *OtherDest = BI->getSuccessor(BIOp ^ 1);
4185 
4186  LLVM_DEBUG(dbgs() << "FOLDING BRs:" << *PBI->getParent()
4187  << "AND: " << *BI->getParent());
4188 
4190 
4191  // If OtherDest *is* BB, then BB is a basic block with a single conditional
4192  // branch in it, where one edge (OtherDest) goes back to itself but the other
4193  // exits. We don't *know* that the program avoids the infinite loop
4194  // (even though that seems likely). If we do this xform naively, we'll end up
4195  // recursively unpeeling the loop. Since we know that (after the xform is
4196  // done) that the block *is* infinite if reached, we just make it an obviously
4197  // infinite loop with no cond branch.
4198  if (OtherDest == BB) {
4199  // Insert it at the end of the function, because it's either code,
4200  // or it won't matter if it's hot. :)
4201  BasicBlock *InfLoopBlock =
4202  BasicBlock::Create(BB->getContext(), "infloop", BB->getParent());
4203  BranchInst::Create(InfLoopBlock, InfLoopBlock);
4204  if (DTU)
4205  Updates.push_back({DominatorTree::Insert, InfLoopBlock, InfLoopBlock});
4206  OtherDest = InfLoopBlock;
4207  }
4208 
4209  LLVM_DEBUG(dbgs() << *PBI->getParent()->getParent());
4210 
4211  // BI may have other predecessors. Because of this, we leave
4212  // it alone, but modify PBI.
4213 
4214  // Make sure we get to CommonDest on True&True directions.
4215  Value *PBICond = PBI->getCondition();
4217  if (PBIOp)
4218  PBICond = Builder.CreateNot(PBICond, PBICond->getName() + ".not");
4219 
4220  Value *BICond = BI->getCondition();
4221  if (BIOp)
4222  BICond = Builder.CreateNot(BICond, BICond->getName() + ".not");
4223 
4224  // Merge the conditions.
4225  Value *Cond =
4226  createLogicalOp(Builder, Instruction::Or, PBICond, BICond, "brmerge");
4227 
4228  // Modify PBI to branch on the new condition to the new dests.
4229  PBI->setCondition(Cond);
4230  PBI->setSuccessor(0, CommonDest);
4231  PBI->setSuccessor(1, OtherDest);
4232 
4233  if (DTU) {
4234  Updates.push_back({DominatorTree::Insert, PBI->getParent(), OtherDest});
4235  Updates.push_back({DominatorTree::Delete, PBI->getParent(), RemovedDest});
4236 
4237  DTU->applyUpdates(Updates);
4238  }
4239 
4240  // Update branch weight for PBI.
4241  uint64_t PredTrueWeight, PredFalseWeight, SuccTrueWeight, SuccFalseWeight;
4242  uint64_t PredCommon, PredOther, SuccCommon, SuccOther;
4243  bool HasWeights =
4244  extractPredSuccWeights(PBI, BI, PredTrueWeight, PredFalseWeight,
4245  SuccTrueWeight, SuccFalseWeight);
4246  if (HasWeights) {
4247  PredCommon = PBIOp ? PredFalseWeight : PredTrueWeight;
4248  PredOther = PBIOp ? PredTrueWeight : PredFalseWeight;
4249  SuccCommon = BIOp ? SuccFalseWeight : SuccTrueWeight;
4250  SuccOther = BIOp ? SuccTrueWeight : SuccFalseWeight;
4251  // The weight to CommonDest should be PredCommon * SuccTotal +
4252  // PredOther * SuccCommon.
4253  // The weight to OtherDest should be PredOther * SuccOther.
4254  uint64_t NewWeights[2] = {PredCommon * (SuccCommon + SuccOther) +
4255  PredOther * SuccCommon,
4256  PredOther * SuccOther};
4257  // Halve the weights if any of them cannot fit in an uint32_t
4258  FitWeights(NewWeights);
4259 
4260  setBranchWeights(PBI, NewWeights[0], NewWeights[1]);
4261  }
4262 
4263  // OtherDest may have phi nodes. If so, add an entry from PBI's
4264  // block that are identical to the entries for BI's block.
4265  AddPredecessorToBlock(OtherDest, PBI->getParent(), BB);
4266 
4267  // We know that the CommonDest already had an edge from PBI to
4268  // it. If it has PHIs though, the PHIs may have different
4269  // entries for BB and PBI's BB. If so, insert a select to make
4270  // them agree.
4271  for (PHINode &PN : CommonDest->phis()) {
4272  Value *BIV = PN.getIncomingValueForBlock(BB);
4273  unsigned PBBIdx = PN.getBasicBlockIndex(PBI->getParent());
4274  Value *PBIV = PN.getIncomingValue(PBBIdx);
4275  if (BIV != PBIV) {
4276  // Insert a select in PBI to pick the right value.
4277  SelectInst *NV = cast<SelectInst>(
4278  Builder.CreateSelect(PBICond, PBIV, BIV, PBIV->getName() + ".mux"));
4279  PN.setIncomingValue(PBBIdx, NV);
4280  // Although the select has the same condition as PBI, the original branch
4281  // weights for PBI do not apply to the new select because the select's
4282  // 'logical' edges are incoming edges of the phi that is eliminated, not
4283  // the outgoing edges of PBI.
4284  if (HasWeights) {
4285  uint64_t PredCommon = PBIOp ? PredFalseWeight : PredTrueWeight;
4286  uint64_t PredOther = PBIOp ? PredTrueWeight : PredFalseWeight;
4287  uint64_t SuccCommon = BIOp ? SuccFalseWeight : SuccTrueWeight;
4288  uint64_t SuccOther = BIOp ? SuccTrueWeight : SuccFalseWeight;
4289  // The weight to PredCommonDest should be PredCommon * SuccTotal.
4290  // The weight to PredOtherDest should be PredOther * SuccCommon.
4291  uint64_t NewWeights[2] = {PredCommon * (SuccCommon + SuccOther),
4292  PredOther * SuccCommon};
4293 
4294  FitWeights(NewWeights);
4295 
4296  setBranchWeights(NV, NewWeights[0], NewWeights[1]);
4297  }
4298  }
4299  }
4300 
4301  LLVM_DEBUG(dbgs() << "INTO: " << *PBI->getParent());
4302  LLVM_DEBUG(dbgs() << *PBI->getParent()->getParent());
4303 
4304  // This basic block is probably dead. We know it has at least
4305  // one fewer predecessor.
4306  return true;
4307 }
4308 
4309 // Simplifies a terminator by replacing it with a branch to TrueBB if Cond is
4310 // true or to FalseBB if Cond is false.
4311 // Takes care of updating the successors and removing the old terminator.
4312 // Also makes sure not to introduce new successors by assuming that edges to
4313 // non-successor TrueBBs and FalseBBs aren't reachable.
4314 bool SimplifyCFGOpt::SimplifyTerminatorOnSelect(Instruction *OldTerm,
4315  Value *Cond, BasicBlock *TrueBB,
4316  BasicBlock *FalseBB,
4317  uint32_t TrueWeight,
4318  uint32_t FalseWeight) {
4319  auto *BB = OldTerm->getParent();
4320  // Remove any superfluous successor edges from the CFG.
4321  // First, figure out which successors to preserve.
4322  // If TrueBB and FalseBB are equal, only try to preserve one copy of that
4323  // successor.
4324  BasicBlock *KeepEdge1 = TrueBB;
4325  BasicBlock *KeepEdge2 = TrueBB != FalseBB ? FalseBB : nullptr;
4326 
4327  SmallSetVector<BasicBlock *, 2> RemovedSuccessors;
4328 
4329  // Then remove the rest.
4330  for (BasicBlock *Succ : successors(OldTerm)) {
4331  // Make sure only to keep exactly one copy of each edge.
4332  if (Succ == KeepEdge1)
4333  KeepEdge1 = nullptr;
4334  else if (Succ == KeepEdge2)
4335  KeepEdge2 = nullptr;
4336  else {
4337  Succ->removePredecessor(BB,
4338  /*KeepOneInputPHIs=*/true);
4339 
4340  if (Succ != TrueBB && Succ != FalseBB)
4341  RemovedSuccessors.insert(Succ);
4342  }
4343  }
4344 
4345  IRBuilder<> Builder(OldTerm);
4346  Builder.SetCurrentDebugLocation(OldTerm->getDebugLoc());
4347 
4348  // Insert an appropriate new terminator.
4349  if (!KeepEdge1 && !KeepEdge2) {
4350  if (TrueBB == FalseBB) {
4351  // We were only looking for one successor, and it was present.
4352  // Create an unconditional branch to it.
4353  Builder.CreateBr(TrueBB);
4354  } else {
4355  // We found both of the successors we were looking for.
4356  // Create a conditional branch sharing the condition of the select.
4357  BranchInst *NewBI = Builder.CreateCondBr(Cond, TrueBB, FalseBB);
4358  if (TrueWeight != FalseWeight)
4359  setBranchWeights(NewBI, TrueWeight, FalseWeight);
4360  }
4361  } else if (KeepEdge1 && (KeepEdge2 || TrueBB == FalseBB)) {
4362  // Neither of the selected blocks were successors, so this
4363  // terminator must be unreachable.
4364  new UnreachableInst(OldTerm->getContext(), OldTerm);
4365  } else {
4366  // One of the selected values was a successor, but the other wasn't.
4367  // Insert an unconditional branch to the one that was found;
4368  // the edge to the one that wasn't must be unreachable.
4369  if (!KeepEdge1) {
4370  // Only TrueBB was found.
4371  Builder.CreateBr(TrueBB);
4372  } else {
4373  // Only FalseBB was found.
4374  Builder.CreateBr(FalseBB);
4375  }
4376  }
4377 
4378  EraseTerminatorAndDCECond(OldTerm);
4379 
4380  if (DTU) {
4382  Updates.reserve(RemovedSuccessors.size());
4383  for (auto *RemovedSuccessor : RemovedSuccessors)
4384  Updates.push_back({DominatorTree::Delete, BB, RemovedSuccessor});
4385  DTU->applyUpdates(Updates);
4386  }
4387 
4388  return true;
4389 }
4390 
4391 // Replaces
4392 // (switch (select cond, X, Y)) on constant X, Y
4393 // with a branch - conditional if X and Y lead to distinct BBs,
4394 // unconditional otherwise.
4395 bool SimplifyCFGOpt::SimplifySwitchOnSelect(SwitchInst *SI,
4396  SelectInst *Select) {
4397  // Check for constant integer values in the select.
4398  ConstantInt *TrueVal = dyn_cast<ConstantInt>(Select->getTrueValue());
4399  ConstantInt *FalseVal = dyn_cast<ConstantInt>(Select->getFalseValue());
4400  if (!TrueVal || !FalseVal)
4401  return false;
4402 
4403  // Find the relevant condition and destinations.
4404  Value *Condition = Select->getCondition();
4405  BasicBlock *TrueBB = SI->findCaseValue(TrueVal)->getCaseSuccessor();
4406  BasicBlock *FalseBB = SI->findCaseValue(FalseVal)->getCaseSuccessor();
4407 
4408  // Get weight for TrueBB and FalseBB.
4409  uint32_t TrueWeight = 0, FalseWeight = 0;
4410  SmallVector<uint64_t, 8> Weights;
4411  bool HasWeights = HasBranchWeights(SI);
4412  if (HasWeights) {
4413  GetBranchWeights(SI, Weights);
4414  if (Weights.size() == 1 + SI->getNumCases()) {
4415  TrueWeight =
4416  (uint32_t)Weights[SI->findCaseValue(TrueVal)->getSuccessorIndex()];
4417  FalseWeight =
4418  (uint32_t)Weights[SI->findCaseValue(FalseVal)->getSuccessorIndex()];
4419  }
4420  }
4421 
4422  // Perform the actual simplification.
4423  return SimplifyTerminatorOnSelect(SI, Condition, TrueBB, FalseBB, TrueWeight,
4424  FalseWeight);
4425 }
4426 
4427 // Replaces
4428 // (indirectbr (select cond, blockaddress(@fn, BlockA),
4429 // blockaddress(@fn, BlockB)))
4430 // with
4431 // (br cond, BlockA, BlockB).
4432 bool SimplifyCFGOpt::SimplifyIndirectBrOnSelect(IndirectBrInst *IBI,
4433  SelectInst *SI) {
4434  // Check that both operands of the select are block addresses.
4435  BlockAddress *TBA = dyn_cast<BlockAddress>(SI->getTrueValue());
4436  BlockAddress *FBA = dyn_cast<BlockAddress>(SI->getFalseValue());
4437  if (!TBA || !FBA)
4438  return false;
4439 
4440  // Extract the actual blocks.
4441  BasicBlock *TrueBB = TBA->getBasicBlock();
4442  BasicBlock *FalseBB = FBA->getBasicBlock();
4443 
4444  // Perform the actual simplification.
4445  return SimplifyTerminatorOnSelect(IBI, SI->getCondition(), TrueBB, FalseBB, 0,
4446  0);
4447 }
4448 
4449 /// This is called when we find an icmp instruction
4450 /// (a seteq/setne with a constant) as the only instruction in a
4451 /// block that ends with an uncond branch. We are looking for a very specific
4452 /// pattern that occurs when "A == 1 || A == 2 || A == 3" gets simplified. In
4453 /// this case, we merge the first two "or's of icmp" into a switch, but then the
4454 /// default value goes to an uncond block with a seteq in it, we get something
4455 /// like:
4456 ///
4457 /// switch i8 %A, label %DEFAULT [ i8 1, label %end i8 2, label %end ]
4458 /// DEFAULT:
4459 /// %tmp = icmp eq i8 %A, 92
4460 /// br label %end
4461 /// end:
4462 /// ... = phi i1 [ true, %entry ], [ %tmp, %DEFAULT ], [ true, %entry ]
4463 ///
4464 /// We prefer to split the edge to 'end' so that there is a true/false entry to
4465 /// the PHI, merging the third icmp into the switch.
4466 bool SimplifyCFGOpt::tryToSimplifyUncondBranchWithICmpInIt(
4467  ICmpInst *ICI, IRBuilder<> &Builder) {
4468  BasicBlock *BB = ICI->getParent();
4469 
4470  // If the block has any PHIs in it or the icmp has multiple uses, it is too
4471  // complex.
4472  if (isa<PHINode>(BB->begin()) || !ICI->hasOneUse())
4473  return false;
4474 
4475  Value *V = ICI->getOperand(0);
4476  ConstantInt *Cst = cast<ConstantInt>(ICI->getOperand(1));
4477 
4478  // The pattern we're looking for is where our only predecessor is a switch on
4479  // 'V' and this block is the default case for the switch. In this case we can
4480  // fold the compared value into the switch to simplify things.
4481  BasicBlock *Pred = BB->getSinglePredecessor();
4482  if (!Pred || !isa<SwitchInst>(Pred->getTerminator()))
4483  return false;
4484 
4485  SwitchInst *SI = cast<SwitchInst>(Pred->getTerminator());
4486  if (SI->getCondition() != V)
4487  return false;
4488 
4489  // If BB is reachable on a non-default case, then we simply know the value of
4490  // V in this block. Substitute it and constant fold the icmp instruction
4491  // away.
4492  if (SI->getDefaultDest() != BB) {
4493  ConstantInt *VVal = SI->findCaseDest(BB);
4494  assert(VVal && "Should have a unique destination value");
4495  ICI->setOperand(0, VVal);
4496 
4497  if (Value *V = simplifyInstruction(ICI, {DL, ICI})) {
4498  ICI->replaceAllUsesWith(V);
4499  ICI->eraseFromParent();
4500  }
4501  // BB is now empty, so it is likely to simplify away.
4502  return requestResimplify();
4503  }
4504 
4505  // Ok, the block is reachable from the default dest. If the constant we're
4506  // comparing exists in one of the other edges, then we can constant fold ICI
4507  // and zap it.
4508  if (SI->findCaseValue(Cst) != SI->case_default()) {
4509  Value *V;
4510  if (ICI->getPredicate() == ICmpInst::ICMP_EQ)
4511  V = ConstantInt::getFalse(BB->getContext());
4512  else
4513  V = ConstantInt::getTrue(BB->getContext());
4514 
4515  ICI->replaceAllUsesWith(V);
4516  ICI->eraseFromParent();
4517  // BB is now empty, so it is likely to simplify away.
4518  return requestResimplify();
4519  }
4520 
4521  // The use of the icmp has to be in the 'end' block, by the only PHI node in
4522  // the block.
4523  BasicBlock *SuccBlock = BB->getTerminator()->getSuccessor(0);
4524  PHINode *PHIUse = dyn_cast<PHINode>(ICI->user_back());
4525  if (PHIUse == nullptr || PHIUse != &SuccBlock->front() ||
4526  isa<PHINode>(++BasicBlock::iterator(PHIUse)))
4527  return false;
4528 
4529  // If the icmp is a SETEQ, then the default dest gets false, the new edge gets
4530  // true in the PHI.
4531  Constant *DefaultCst = ConstantInt::getTrue(BB->getContext());
4532  Constant *NewCst = ConstantInt::getFalse(BB->getContext());
4533 
4534  if (ICI->getPredicate() == ICmpInst::ICMP_EQ)
4535  std::swap(DefaultCst, NewCst);
4536 
4537  // Replace ICI (which is used by the PHI for the default value) with true or
4538  // false depending on if it is EQ or NE.
4539  ICI->replaceAllUsesWith(DefaultCst);
4540  ICI->eraseFromParent();
4541 
4543 
4544  // Okay, the switch goes to this block on a default value. Add an edge from
4545  // the switch to the merge point on the compared value.
4546  BasicBlock *NewBB =
4547  BasicBlock::Create(BB->getContext(), "switch.edge", BB->getParent(), BB);
4548  {
4550  auto W0 = SIW.getSuccessorWeight(0);
4552  if (W0) {
4553  NewW = ((uint64_t(*W0) + 1) >> 1);
4554  SIW.setSuccessorWeight(0, *NewW);
4555  }
4556  SIW.addCase(Cst, NewBB, NewW);
4557  if (DTU)
4558  Updates.push_back({DominatorTree::Insert, Pred, NewBB});
4559  }
4560 
4561  // NewBB branches to the phi block, add the uncond branch and the phi entry.
4562  Builder.SetInsertPoint(NewBB);
4563  Builder.SetCurrentDebugLocation(SI->getDebugLoc());
4564  Builder.CreateBr(SuccBlock);
4565  PHIUse->addIncoming(NewCst, NewBB);
4566  if (DTU) {
4567  Updates.push_back({DominatorTree::Insert, NewBB, SuccBlock});
4568  DTU->applyUpdates(Updates);
4569  }
4570  return true;
4571 }
4572 
4573 /// The specified branch is a conditional branch.
4574 /// Check to see if it is branching on an or/and chain of icmp instructions, and
4575 /// fold it into a switch instruction if so.
4576 bool SimplifyCFGOpt::SimplifyBranchOnICmpChain(BranchInst *BI,
4578  const DataLayout &DL) {
4579  Instruction *Cond = dyn_cast<Instruction>(BI->getCondition());
4580  if (!Cond)
4581  return false;
4582 
4583  // Change br (X == 0 | X == 1), T, F into a switch instruction.
4584  // If this is a bunch of seteq's or'd together, or if it's a bunch of
4585  // 'setne's and'ed together, collect them.
4586 
4587  // Try to gather values from a chain of and/or to be turned into a switch
4588  ConstantComparesGatherer ConstantCompare(Cond, DL);
4589  // Unpack the result
4590  SmallVectorImpl<ConstantInt *> &Values = ConstantCompare.Vals;
4591  Value *CompVal = ConstantCompare.CompValue;
4592  unsigned UsedICmps = ConstantCompare.UsedICmps;
4593  Value *ExtraCase = ConstantCompare.Extra;
4594 
4595  // If we didn't have a multiply compared value, fail.
4596  if (!CompVal)
4597  return false;
4598 
4599  // Avoid turning single icmps into a switch.
4600  if (UsedICmps <= 1)
4601  return false;
4602 
4603  bool TrueWhenEqual = match(Cond, m_LogicalOr(m_Value(), m_Value()));
4604 
4605  // There might be duplicate constants in the list, which the switch
4606  // instruction can't handle, remove them now.
4607  array_pod_sort(Values.begin(), Values.end(), ConstantIntSortPredicate);
4608  Values.erase(std::unique(Values.begin(), Values.end()), Values.end());
4609 
4610  // If Extra was used, we require at least two switch values to do the
4611  // transformation. A switch with one value is just a conditional branch.
4612  if (ExtraCase && Values.size() < 2)
4613  return false;
4614 
4615  // TODO: Preserve branch weight metadata, similarly to how
4616  // FoldValueComparisonIntoPredecessors preserves it.
4617 
4618  // Figure out which block is which destination.
4619  BasicBlock *DefaultBB = BI->getSuccessor(1);
4620  BasicBlock *EdgeBB = BI->getSuccessor(0);
4621  if (!TrueWhenEqual)
4622  std::swap(DefaultBB, EdgeBB);
4623 
4624  BasicBlock *BB = BI->getParent();
4625 
4626  LLVM_DEBUG(dbgs() << "Converting 'icmp' chain with " << Values.size()
4627  << " cases into SWITCH. BB is:\n"
4628  << *BB);
4629 
4631 
4632  // If there are any extra values that couldn't be folded into the switch
4633  // then we evaluate them with an explicit branch first. Split the block
4634  // right before the condbr to handle it.
4635  if (ExtraCase) {
4636  BasicBlock *NewBB = SplitBlock(BB, BI, DTU, /*LI=*/nullptr,
4637  /*MSSAU=*/nullptr, "switch.early.test");
4638 
4639  // Remove the uncond branch added to the old block.
4640  Instruction *OldTI = BB->getTerminator();
4641  Builder.SetInsertPoint(OldTI);
4642 
4643  // There can be an unintended UB if extra values are Poison. Before the
4644  // transformation, extra values may not be evaluated according to the
4645  // condition, and it will not raise UB. But after transformation, we are
4646  // evaluating extra values before checking the condition, and it will raise
4647  // UB. It can be solved by adding freeze instruction to extra values.
4648  AssumptionCache *AC = Options.AC;
4649 
4650  if (!isGuaranteedNotToBeUndefOrPoison(ExtraCase, AC, BI, nullptr))
4651  ExtraCase = Builder.CreateFreeze(ExtraCase);
4652 
4653  if (TrueWhenEqual)
4654  Builder.CreateCondBr(ExtraCase, EdgeBB, NewBB);
4655  else
4656  Builder.CreateCondBr(ExtraCase, NewBB, EdgeBB);
4657 
4658  OldTI->eraseFromParent();
4659 
4660  if (DTU)
4661  Updates.push_back({DominatorTree::Insert, BB, EdgeBB});
4662 
4663  // If there are PHI nodes in EdgeBB, then we need to add a new entry to them
4664  // for the edge we just added.
4665  AddPredecessorToBlock(EdgeBB, BB, NewBB);
4666 
4667  LLVM_DEBUG(dbgs() << " ** 'icmp' chain unhandled condition: " << *ExtraCase
4668  << "\nEXTRABB = " << *BB);
4669  BB = NewBB;
4670  }
4671 
4672  Builder.SetInsertPoint(BI);
4673  // Convert pointer to int before we switch.
4674  if (CompVal->getType()->isPointerTy()) {
4675  CompVal = Builder.CreatePtrToInt(
4676  CompVal, DL.getIntPtrType(CompVal->getType()), "magicptr");
4677  }
4678 
4679  // Create the new switch instruction now.
4680  SwitchInst *New = Builder.CreateSwitch(CompVal, DefaultBB, Values.size());
4681 
4682  // Add all of the 'cases' to the switch instruction.
4683  for (unsigned i = 0, e = Values.size(); i != e; ++i)
4684  New->addCase(Values[i], EdgeBB);
4685 
4686  // We added edges from PI to the EdgeBB. As such, if there were any
4687  // PHI nodes in EdgeBB, they need entries to be added corresponding to
4688  // the number of edges added.
4689  for (BasicBlock::iterator BBI = EdgeBB->begin(); isa<PHINode>(BBI); ++BBI) {
4690  PHINode *PN = cast<PHINode>(BBI);
4691  Value *InVal = PN->getIncomingValueForBlock(BB);
4692  for (unsigned i = 0, e = Values.size() - 1; i != e; ++i)
4693  PN->addIncoming(InVal, BB);
4694  }
4695 
4696  // Erase the old branch instruction.
4698  if (DTU)
4699  DTU->applyUpdates(Updates);
4700 
4701  LLVM_DEBUG(dbgs() << " ** 'icmp' chain result is:\n" << *BB << '\n');
4702  return true;
4703 }
4704 
4705 bool SimplifyCFGOpt::simplifyResume(ResumeInst *RI, IRBuilder<> &Builder) {
4706  if (isa<PHINode>(RI->getValue()))
4707  return simplifyCommonResume(RI);
4708  else if (isa<LandingPadInst>(RI->getParent()->getFirstNonPHI()) &&
4709  RI->getValue() == RI->getParent()->getFirstNonPHI())
4710  // The resume must unwind the exception that caused control to branch here.
4711  return simplifySingleResume(RI);
4712 
4713  return false;
4714 }
4715 
4716 // Check if cleanup block is empty
4718  for (Instruction &I : R) {
4719  auto *II = dyn_cast<IntrinsicInst>(&I);
4720  if (!II)
4721  return false;
4722 
4723  Intrinsic::ID IntrinsicID = II->getIntrinsicID();
4724  switch (IntrinsicID) {
4725  case Intrinsic::dbg_declare:
4726  case Intrinsic::dbg_value:
4727  case Intrinsic::dbg_label:
4728  case Intrinsic::lifetime_end:
4729  break;
4730  default:
4731  return false;
4732  }
4733  }
4734  return true;
4735 }
4736 
4737 // Simplify resume that is shared by several landing pads (phi of landing pad).
4738 bool SimplifyCFGOpt::simplifyCommonResume(ResumeInst *RI) {
4739  BasicBlock *BB = RI->getParent();
4740 
4741  // Check that there are no other instructions except for debug and lifetime
4742  // intrinsics between the phi's and resume instruction.
4743  if (!isCleanupBlockEmpty(
4744  make_range(RI->getParent()->getFirstNonPHI(), BB->getTerminator())))
4745  return false;
4746 
4747  SmallSetVector<BasicBlock *, 4> TrivialUnwindBlocks;
4748  auto *PhiLPInst = cast<PHINode>(RI->getValue());
4749 
4750  // Check incoming blocks to see if any of them are trivial.
4751  for (unsigned Idx = 0, End = PhiLPInst->getNumIncomingValues(); Idx != End;
4752  Idx++) {
4753  auto *IncomingBB = PhiLPInst->getIncomingBlock(Idx);
4754  auto *IncomingValue = PhiLPInst->getIncomingValue(Idx);
4755 
4756  // If the block has other successors, we can not delete it because
4757  // it has other dependents.
4758  if (IncomingBB->getUniqueSuccessor() != BB)
4759  continue;
4760 
4761  auto *LandingPad = dyn_cast<LandingPadInst>(IncomingBB->getFirstNonPHI());
4762  // Not the landing pad that caused the control to branch here.
4763  if (IncomingValue != LandingPad)
4764  continue;
4765 
4766  if (isCleanupBlockEmpty(
4767  make_range(LandingPad->getNextNode(), IncomingBB->getTerminator())))
4768  TrivialUnwindBlocks.insert(IncomingBB);
4769  }
4770 
4771  // If no trivial unwind blocks, don't do any simplifications.
4772  if (TrivialUnwindBlocks.empty())
4773  return false;
4774 
4775  // Turn all invokes that unwind here into calls.
4776  for (auto *TrivialBB : TrivialUnwindBlocks) {
4777  // Blocks that will be simplified should be removed from the phi node.
4778  // Note there could be multiple edges to the resume block, and we need
4779  // to remove them all.
4780  while (PhiLPInst->getBasicBlockIndex(TrivialBB) != -1)
4781  BB->removePredecessor(TrivialBB, true);
4782 
4783  for (BasicBlock *Pred :
4785  removeUnwindEdge(Pred, DTU);
4786  ++NumInvokes;
4787  }
4788 
4789  // In each SimplifyCFG run, only the current processed block can be erased.
4790  // Otherwise, it will break the iteration of SimplifyCFG pass. So instead
4791  // of erasing TrivialBB, we only remove the branch to the common resume
4792  // block so that we can later erase the resume block since it has no
4793  // predecessors.
4794  TrivialBB->getTerminator()->eraseFromParent();
4795  new UnreachableInst(RI->getContext(), TrivialBB);
4796  if (DTU)
4797  DTU->applyUpdates({{DominatorTree::Delete, TrivialBB, BB}});
4798  }
4799 
4800  // Delete the resume block if all its predecessors have been removed.
4801  if (pred_empty(BB))
4802  DeleteDeadBlock(BB, DTU);
4803 
4804  return !TrivialUnwindBlocks.empty();
4805 }
4806 
4807 // Simplify resume that is only used by a single (non-phi) landing pad.
4808 bool SimplifyCFGOpt::simplifySingleResume(ResumeInst *RI) {
4809  BasicBlock *BB = RI->getParent();
4810  auto *LPInst = cast<LandingPadInst>(BB->getFirstNonPHI());
4811  assert(RI->getValue() == LPInst &&
4812  "Resume must unwind the exception that caused control to here");
4813 
4814  // Check that there are no other instructions except for debug intrinsics.
4815  if (!isCleanupBlockEmpty(
4816  make_range<Instruction *>(LPInst->getNextNode(), RI)))
4817  return false;
4818 
4819  // Turn all invokes that unwind here into calls and delete the basic block.
4821  removeUnwindEdge(Pred, DTU);
4822  ++NumInvokes;
4823  }
4824 
4825  // The landingpad is now unreachable. Zap it.
4826  DeleteDeadBlock(BB, DTU);
4827  return true;
4828 }
4829 
4831  // If this is a trivial cleanup pad that executes no instructions, it can be
4832  // eliminated. If the cleanup pad continues to the caller, any predecessor
4833  // that is an EH pad will be updated to continue to the caller and any
4834  // predecessor that terminates with an invoke instruction will have its invoke
4835  // instruction converted to a call instruction. If the cleanup pad being
4836  // simplified does not continue to the caller, each predecessor will be
4837  // updated to continue to the unwind destination of the cleanup pad being
4838  // simplified.
4839  BasicBlock *BB = RI->getParent();
4840  CleanupPadInst *CPInst = RI->getCleanupPad();
4841  if (CPInst->getParent() != BB)
4842  // This isn't an empty cleanup.
4843  return false;
4844 
4845  // We cannot kill the pad if it has multiple uses. This typically arises
4846  // from unreachable basic blocks.
4847  if (!CPInst->hasOneUse())
4848  return false;
4849 
4850  // Check that there are no other instructions except for benign intrinsics.
4851  if (!isCleanupBlockEmpty(
4852  make_range<Instruction *>(CPInst->getNextNode(), RI)))
4853  return false;
4854 
4855  // If the cleanup return we are simplifying unwinds to the caller, this will
4856  // set UnwindDest to nullptr.
4857  BasicBlock *UnwindDest = RI->getUnwindDest();
4858  Instruction *DestEHPad = UnwindDest ? UnwindDest->getFirstNonPHI() : nullptr;
4859 
4860  // We're about to remove BB from the control flow. Before we do, sink any
4861  // PHINodes into the unwind destination. Doing this before changing the
4862  // control flow avoids some potentially slow checks, since we can currently
4863  // be certain that UnwindDest and BB have no common predecessors (since they
4864  // are both EH pads).
4865  if (UnwindDest) {
4866  // First, go through the PHI nodes in UnwindDest and update any nodes that
4867  // reference the block we are removing
4868  for (PHINode &DestPN : UnwindDest->phis()) {
4869  int Idx = DestPN.getBasicBlockIndex(BB);
4870  // Since BB unwinds to UnwindDest, it has to be in the PHI node.
4871  assert(Idx != -1);
4872  // This PHI node has an incoming value that corresponds to a control
4873  // path through the cleanup pad we are removing. If the incoming
4874  // value is in the cleanup pad, it must be a PHINode (because we
4875  // verified above that the block is otherwise empty). Otherwise, the
4876  // value is either a constant or a value that dominates the cleanup
4877  // pad being removed.
4878  //
4879  // Because BB and UnwindDest are both EH pads, all of their
4880  // predecessors must unwind to these blocks, and since no instruction
4881  // can have multiple unwind destinations, there will be no overlap in
4882  // incoming blocks between SrcPN and DestPN.
4883  Value *SrcVal = DestPN.getIncomingValue(Idx);
4884  PHINode *SrcPN = dyn_cast<PHINode>(SrcVal);
4885 
4886  bool NeedPHITranslation = SrcPN && SrcPN->getParent() == BB;
4887  for (auto *Pred : predecessors(BB)) {
4888  Value *Incoming =
4889  NeedPHITranslation ? SrcPN->getIncomingValueForBlock(Pred) : SrcVal;
4890  DestPN.addIncoming(Incoming, Pred);
4891  }
4892  }
4893 
4894  // Sink any remaining PHI nodes directly into UnwindDest.
4895  Instruction *InsertPt = DestEHPad;
4896  for (PHINode &PN : make_early_inc_range(BB->phis())) {
4897  if (PN.use_empty() || !PN.isUsedOutsideOfBlock(BB))
4898  // If the PHI node has no uses or all of its uses are in this basic
4899  // block (meaning they are debug or lifetime intrinsics), just leave
4900  // it. It will be erased when we erase BB below.
4901  continue;
4902 
4903  // Otherwise, sink this PHI node into UnwindDest.
4904  // Any predecessors to UnwindDest which are not already represented
4905  // must be back edges which inherit the value from the path through
4906  // BB. In this case, the PHI value must reference itself.
4907  for (auto *pred : predecessors(UnwindDest))
4908  if (pred != BB)
4909  PN.addIncoming(&PN, pred);
4910  PN.moveBefore(InsertPt);
4911  // Also, add a dummy incoming value for the original BB itself,
4912  // so that the PHI is well-formed until we drop said predecessor.
4914  }
4915  }
4916 
4917  std::vector<DominatorTree::UpdateType> Updates;
4918 
4919  // We use make_early_inc_range here because we will remove all predecessors.
4921  if (UnwindDest == nullptr) {
4922  if (DTU) {
4923  DTU->applyUpdates(Updates);
4924  Updates.clear();
4925  }
4926  removeUnwindEdge(PredBB, DTU);
4927  ++NumInvokes;
4928  } else {
4929  BB->removePredecessor(PredBB);
4930  Instruction *TI = PredBB->getTerminator();
4931  TI->replaceUsesOfWith(BB, UnwindDest);
4932  if (DTU) {
4933  Updates.push_back({DominatorTree::Insert, PredBB, UnwindDest});
4934  Updates.push_back({DominatorTree::Delete, PredBB, BB});
4935  }
4936  }
4937  }
4938 
4939  if (DTU)
4940  DTU->applyUpdates(Updates);
4941 
4942  DeleteDeadBlock(BB, DTU);
4943 
4944  return true;
4945 }
4946 
4947 // Try to merge two cleanuppads together.
4949  // Skip any cleanuprets which unwind to caller, there is nothing to merge
4950  // with.
4951  BasicBlock *UnwindDest = RI->getUnwindDest();
4952  if (!UnwindDest)
4953  return false;
4954 
4955  // This cleanupret isn't the only predecessor of this cleanuppad, it wouldn't
4956  // be safe to merge without code duplication.
4957  if (UnwindDest->getSinglePredecessor() != RI->getParent())
4958  return false;
4959 
4960  // Verify that our cleanuppad's unwind destination is another cleanuppad.
4961  auto *SuccessorCleanupPad = dyn_cast<CleanupPadInst>(&UnwindDest->front());
4962  if (!SuccessorCleanupPad)
4963  return false;
4964 
4965  CleanupPadInst *PredecessorCleanupPad = RI->getCleanupPad();
4966  // Replace any uses of the successor cleanupad with the predecessor pad
4967  // The only cleanuppad uses should be this cleanupret, it's cleanupret and
4968  // funclet bundle operands.
4969  SuccessorCleanupPad->replaceAllUsesWith(PredecessorCleanupPad);
4970  // Remove the old cleanuppad.
4971  SuccessorCleanupPad->eraseFromParent();
4972  // Now, we simply replace the cleanupret with a branch to the unwind
4973  // destination.
4974  BranchInst::Create(UnwindDest, RI->getParent());
4975  RI->eraseFromParent();
4976 
4977  return true;
4978 }
4979 
4980 bool SimplifyCFGOpt::simplifyCleanupReturn(CleanupReturnInst *RI) {
4981  // It is possible to transiantly have an undef cleanuppad operand because we
4982  // have deleted some, but not all, dead blocks.
4983  // Eventually, this block will be deleted.
4984  if (isa<UndefValue>(RI->getOperand(0)))
4985  return false;
4986 
4987  if (mergeCleanupPad(RI))
4988  return true;
4989 
4990  if (removeEmptyCleanup(RI, DTU))
4991  return true;
4992 
4993  return false;
4994 }
4995 
4996 // WARNING: keep in sync with InstCombinerImpl::visitUnreachableInst()!
4997 bool SimplifyCFGOpt::simplifyUnreachable(UnreachableInst *UI) {
4998  BasicBlock *BB = UI->getParent();
4999 
5000  bool Changed = false;
5001 
5002  // If there are any instructions immediately before the unreachable that can
5003  // be removed, do so.
5004  while (UI->getIterator() != BB->begin()) {
5005  BasicBlock::iterator BBI = UI->getIterator();
5006  --BBI;
5007 
5009  break; // Can not drop any more instructions. We're done here.
5010  // Otherwise, this instruction can be freely erased,
5011  // even if it is not side-effect free.
5012 
5013  // Note that deleting EH's here is in fact okay, although it involves a bit
5014  // of subtle reasoning. If this inst is an EH, all the predecessors of this
5015  // block will be the unwind edges of Invoke/CatchSwitch/CleanupReturn,
5016  // and we can therefore guarantee this block will be erased.
5017 
5018  // Delete this instruction (any uses are guaranteed to be dead)
5019  BBI->replaceAllUsesWith(PoisonValue::get(BBI->getType()));
5020  BBI->eraseFromParent();
5021  Changed = true;
5022  }
5023 
5024  // If the unreachable instruction is the first in the block, take a gander
5025  // at all of the predecessors of this instruction, and simplify them.
5026  if (&BB->front() != UI)
5027  return Changed;
5028 
5029  std::vector<DominatorTree::UpdateType> Updates;
5030 
5032  for (unsigned i = 0, e = Preds.size(); i != e; ++i) {
5033  auto *Predecessor = Preds[i];
5034  Instruction *TI = Predecessor->getTerminator();
5035  IRBuilder<> Builder(TI);
5036  if (auto *BI = dyn_cast<BranchInst>(TI)) {
5037  // We could either have a proper unconditional branch,
5038  // or a degenerate conditional branch with matching destinations.
5039  if (all_of(BI->successors(),
5040  [BB](auto *Successor) { return Successor == BB; })) {
5041  new UnreachableInst(TI->getContext(), TI);
5042  TI->eraseFromParent();
5043  Changed = true;
5044  } else {
5045  assert(BI->isConditional() && "Can't get here with an uncond branch.");
5046  Value* Cond = BI->getCondition();
5047  assert(BI->getSuccessor(0) != BI->getSuccessor(1) &&
5048  "The destinations are guaranteed to be different here.");
5049  if (BI->getSuccessor(0) == BB) {
5050  Builder.CreateAssumption(Builder.CreateNot(Cond));
5051  Builder.CreateBr(BI->getSuccessor(1));
5052  } else {
5053  assert(BI->getSuccessor(1) == BB && "Incorrect CFG");
5054  Builder.CreateAssumption(Cond);
5055  Builder.CreateBr(BI->getSuccessor(0));
5056  }
5058  Changed = true;
5059  }
5060  if (DTU)
5061  Updates.push_back({DominatorTree::Delete, Predecessor, BB});
5062  } else if (auto *SI = dyn_cast<SwitchInst>(TI)) {
5064  for (auto i = SU->case_begin(), e = SU->case_end(); i != e;) {
5065  if (i->getCaseSuccessor() != BB) {
5066  ++i;
5067  continue;
5068  }
5069  BB->removePredecessor(SU->getParent());
5070  i = SU.removeCase(i);
5071  e = SU->case_end();
5072  Changed = true;
5073  }
5074  // Note that the default destination can't be removed!
5075  if (DTU && SI->getDefaultDest() != BB)
5076  Updates.push_back({DominatorTree::Delete, Predecessor, BB});
5077  } else if (auto *II = dyn_cast<InvokeInst>(TI)) {
5078  if (II->getUnwindDest() == BB) {
5079  if (DTU) {
5080  DTU->applyUpdates(Updates);
5081  Updates.clear();
5082  }
5083  removeUnwindEdge(TI->getParent(), DTU);
5084  Changed = true;
5085  }
5086  } else if (auto *CSI = dyn_cast<CatchSwitchInst>(TI)) {
5087  if (CSI->getUnwindDest() == BB) {
5088  if (DTU) {
5089  DTU->applyUpdates(Updates);
5090  Updates.clear();
5091  }
5092  removeUnwindEdge(TI->getParent(), DTU);
5093  Changed = true;
5094  continue;
5095  }
5096 
5097  for (CatchSwitchInst::handler_iterator I = CSI->handler_begin(),
5098  E = CSI->handler_end();
5099  I != E; ++I) {
5100  if (*I == BB) {
5101  CSI->removeHandler(I);
5102  --I;
5103  --E;
5104  Changed = true;
5105  }
5106  }
5107  if (DTU)
5108  Updates.push_back({DominatorTree::Delete, Predecessor, BB});
5109  if (CSI->getNumHandlers() == 0) {
5110  if (CSI->hasUnwindDest()) {
5111  // Redirect all predecessors of the block containing CatchSwitchInst
5112  // to instead branch to the CatchSwitchInst's unwind destination.
5113  if (DTU) {
5114  for (auto *PredecessorOfPredecessor : predecessors(Predecessor)) {
5115  Updates.push_back({DominatorTree::Insert,
5116  PredecessorOfPredecessor,
5117  CSI->getUnwindDest()});
5118  Updates.push_back({DominatorTree::Delete,
5119  PredecessorOfPredecessor, Predecessor});
5120  }
5121  }
5122  Predecessor->replaceAllUsesWith(CSI->getUnwindDest());
5123  } else {
5124  // Rewrite all preds to unwind to caller (or from invoke to call).
5125  if (DTU) {
5126  DTU->applyUpdates(Updates);
5127  Updates.clear();
5128  }
5129  SmallVector<BasicBlock *, 8> EHPreds(predecessors(Predecessor));
5130  for (BasicBlock *EHPred : EHPreds)
5131  removeUnwindEdge(EHPred, DTU);
5132  }
5133  // The catchswitch is no longer reachable.
5134  new UnreachableInst(CSI->getContext(), CSI);
5135  CSI->eraseFromParent();
5136  Changed = true;
5137  }
5138  } else if (auto *CRI = dyn_cast<CleanupReturnInst>(TI)) {
5139  (void)CRI;
5140  assert(CRI->hasUnwindDest() && CRI->getUnwindDest() == BB &&
5141  "Expected to always have an unwind to BB.");
5142  if (DTU)
5143  Updates.push_back({DominatorTree::Delete, Predecessor, BB});
5144  new UnreachableInst(TI->getContext(), TI);
5145  TI->eraseFromParent();
5146  Changed = true;
5147  }
5148  }
5149 
5150  if (DTU)
5151  DTU->applyUpdates(Updates);
5152 
5153  // If this block is now dead, remove it.
5154  if (pred_empty(BB) && BB != &BB->getParent()->getEntryBlock()) {
5155  DeleteDeadBlock(BB, DTU);
5156  return true;
5157  }
5158 
5159  return Changed;
5160 }
5161 
5163  assert(Cases.size() >= 1);
5164 
5165  array_pod_sort(Cases.begin(), Cases.end(), ConstantIntSortPredicate);
5166  for (size_t I = 1, E = Cases.size(); I != E; ++I) {
5167  if (Cases[I - 1]->getValue() != Cases[I]->getValue() + 1)
5168  return false;
5169  }
5170  return true;
5171 }
5172 
5174  DomTreeUpdater *DTU) {
5175  LLVM_DEBUG(dbgs() << "SimplifyCFG: switch default is dead.\n");
5176  auto *BB = Switch->getParent();
5177  auto *OrigDefaultBlock = Switch->getDefaultDest();
5178  OrigDefaultBlock->removePredecessor(BB);
5179  BasicBlock *NewDefaultBlock = BasicBlock::Create(
5180  BB->getContext(), BB->getName() + ".unreachabledefault", BB->getParent(),
5181  OrigDefaultBlock);
5182  new UnreachableInst(Switch->getContext(), NewDefaultBlock);
5183  Switch->setDefaultDest(&*NewDefaultBlock);
5184  if (DTU) {
5186  Updates.push_back({DominatorTree::Insert, BB, &*NewDefaultBlock});
5187  if (!is_contained(successors(BB), OrigDefaultBlock))
5188  Updates.push_back({DominatorTree::Delete, BB, &*OrigDefaultBlock});
5189  DTU->applyUpdates(Updates);
5190  }
5191 }
5192 
5193 /// Turn a switch into an integer range comparison and branch.
5194 /// Switches with more than 2 destinations are ignored.
5195 /// Switches with 1 destination are also ignored.
5196 bool SimplifyCFGOpt::TurnSwitchRangeIntoICmp(SwitchInst *SI,
5197  IRBuilder<> &Builder) {
5198  assert(SI->getNumCases() > 1 && "Degenerate switch?");
5199 
5200  bool HasDefault =
5201  !isa<UnreachableInst>(SI->getDefaultDest()->getFirstNonPHIOrDbg());
5202 
5203  auto *BB = SI->getParent();
5204 
5205  // Partition the cases into two sets with different destinations.
5206  BasicBlock *DestA = HasDefault ? SI->getDefaultDest() : nullptr;
5207  BasicBlock *DestB = nullptr;
5210 
5211  for (auto Case : SI->cases()) {
5212  BasicBlock *Dest = Case.getCaseSuccessor();
5213  if (!DestA)
5214  DestA = Dest;
5215  if (Dest == DestA) {
5216  CasesA.push_back(Case.getCaseValue());
5217  continue;
5218  }
5219  if (!DestB)
5220  DestB = Dest;
5221  if (Dest == DestB) {
5222  CasesB.push_back(Case.getCaseValue());
5223  continue;
5224  }
5225  return false; // More than two destinations.
5226  }
5227  if (!DestB)
5228  return false; // All destinations are the same and the default is unreachable
5229 
5230  assert(DestA && DestB &&
5231  "Single-destination switch should have been folded.");
5232  assert(DestA != DestB);
5233  assert(DestB != SI->getDefaultDest());
5234  assert(!CasesB.empty() && "There must be non-default cases.");
5235  assert(!CasesA.empty() || HasDefault);
5236 
5237  // Figure out if one of the sets of cases form a contiguous range.
5238  SmallVectorImpl<ConstantInt *> *ContiguousCases = nullptr;
5239  BasicBlock *ContiguousDest = nullptr;
5240  BasicBlock *OtherDest = nullptr;
5241  if (!CasesA.empty() && CasesAreContiguous(CasesA)) {
5242  ContiguousCases = &CasesA;
5243  ContiguousDest = DestA;
5244  OtherDest = DestB;
5245  } else if (CasesAreContiguous(CasesB)) {
5246  ContiguousCases = &CasesB;
5247  ContiguousDest = DestB;
5248  OtherDest = DestA;
5249  } else
5250  return false;
5251 
5252  // Start building the compare and branch.
5253 
5254  Constant *Offset = ConstantExpr::getNeg(ContiguousCases->back());
5255  Constant *NumCases =
5256  ConstantInt::get(Offset->getType(), ContiguousCases->size());
5257 
5258  Value *Sub = SI->getCondition();
5259  if (!Offset->isNullValue())
5260  Sub = Builder.CreateAdd(Sub, Offset, Sub->getName() + ".off");
5261 
5262  Value *Cmp;
5263  // If NumCases overflowed, then all possible values jump to the successor.
5264  if (NumCases->isNullValue() && !ContiguousCases->empty())
5265  Cmp = ConstantInt::getTrue(SI->getContext());
5266  else
5267  Cmp = Builder.CreateICmpULT(Sub, NumCases, "switch");
5268  BranchInst *NewBI = Builder.CreateCondBr(Cmp, ContiguousDest, OtherDest);
5269 
5270  // Update weight for the newly-created conditional branch.
5271  if (HasBranchWeights(SI)) {
5272  SmallVector<uint64_t, 8> Weights;
5273  GetBranchWeights(SI, Weights);
5274  if (Weights.size() == 1 + SI->getNumCases()) {
5275  uint64_t TrueWeight = 0;
5276  uint64_t FalseWeight = 0;
5277  for (size_t I = 0, E = Weights.size(); I != E; ++I) {
5278  if (SI->getSuccessor(I) == ContiguousDest)
5279  TrueWeight += Weights[I];
5280  else
5281  FalseWeight += Weights[I];
5282  }
5283  while (TrueWeight > UINT32_MAX || FalseWeight > UINT32_MAX) {
5284  TrueWeight /= 2;
5285  FalseWeight /= 2;
5286  }
5287  setBranchWeights(NewBI, TrueWeight, FalseWeight);
5288  }
5289  }
5290 
5291  // Prune obsolete incoming values off the successors' PHI nodes.
5292  for (auto BBI = ContiguousDest->begin(); isa<PHINode>(BBI); ++BBI) {
5293  unsigned PreviousEdges = ContiguousCases->size();
5294  if (ContiguousDest == SI->getDefaultDest())
5295  ++PreviousEdges;
5296  for (unsigned I = 0, E = PreviousEdges - 1; I != E; ++I)
5297  cast<PHINode>(BBI)->removeIncomingValue(SI->getParent());
5298  }
5299  for (auto BBI = OtherDest->begin(); isa<PHINode>(BBI); ++BBI) {
5300  unsigned PreviousEdges = SI->getNumCases() - ContiguousCases->size();
5301  if (OtherDest == SI->getDefaultDest())
5302  ++PreviousEdges;
5303  for (unsigned I = 0, E = PreviousEdges - 1; I != E; ++I)
5304  cast<PHINode>(BBI)->removeIncomingValue(SI->getParent());
5305  }
5306 
5307  // Clean up the default block - it may have phis or other instructions before
5308  // the unreachable terminator.
5309  if (!HasDefault)
5311 
5312  auto *UnreachableDefault = SI->getDefaultDest();
5313 
5314  // Drop the switch.
5315  SI->eraseFromParent();
5316 
5317  if (!HasDefault && DTU)
5318  DTU->applyUpdates({{DominatorTree::Delete, BB, UnreachableDefault}});
5319 
5320  return true;
5321 }
5322 
5323 /// Compute masked bits for the condition of a switch
5324 /// and use it to remove dead cases.
5326  AssumptionCache *AC,
5327  const DataLayout &DL) {
5328  Value *Cond = SI->getCondition();
5329  KnownBits Known = computeKnownBits(Cond, DL, 0, AC, SI);
5330 
5331  // We can also eliminate cases by determining that their values are outside of
5332  // the limited range of the condition based on how many significant (non-sign)
5333  // bits are in the condition value.
5334  unsigned MaxSignificantBitsInCond =
5335  ComputeMaxSignificantBits(Cond, DL, 0, AC, SI);
5336 
5337  // Gather dead cases.
5338  SmallVe