LLVM  14.0.0git
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"
36 #include "llvm/IR/Attributes.h"
37 #include "llvm/IR/BasicBlock.h"
38 #include "llvm/IR/CFG.h"
39 #include "llvm/IR/Constant.h"
40 #include "llvm/IR/ConstantRange.h"
41 #include "llvm/IR/Constants.h"
42 #include "llvm/IR/DataLayout.h"
43 #include "llvm/IR/DerivedTypes.h"
44 #include "llvm/IR/Function.h"
45 #include "llvm/IR/GlobalValue.h"
46 #include "llvm/IR/GlobalVariable.h"
47 #include "llvm/IR/IRBuilder.h"
48 #include "llvm/IR/InstrTypes.h"
49 #include "llvm/IR/Instruction.h"
50 #include "llvm/IR/Instructions.h"
51 #include "llvm/IR/IntrinsicInst.h"
52 #include "llvm/IR/Intrinsics.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/PseudoProbe.h"
61 #include "llvm/IR/Type.h"
62 #include "llvm/IR/Use.h"
63 #include "llvm/IR/User.h"
64 #include "llvm/IR/Value.h"
65 #include "llvm/IR/ValueHandle.h"
67 #include "llvm/Support/Casting.h"
69 #include "llvm/Support/Debug.h"
71 #include "llvm/Support/KnownBits.h"
78 #include <algorithm>
79 #include <cassert>
80 #include <climits>
81 #include <cstddef>
82 #include <cstdint>
83 #include <iterator>
84 #include <map>
85 #include <set>
86 #include <tuple>
87 #include <utility>
88 #include <vector>
89 
90 using namespace llvm;
91 using namespace PatternMatch;
92 
93 #define DEBUG_TYPE "simplifycfg"
94 
96  "simplifycfg-require-and-preserve-domtree", cl::Hidden, cl::ZeroOrMore,
97  cl::init(false),
98  cl::desc("Temorary development switch used to gradually uplift SimplifyCFG "
99  "into preserving DomTree,"));
100 
101 // Chosen as 2 so as to be cheap, but still to have enough power to fold
102 // a select, so the "clamp" idiom (of a min followed by a max) will be caught.
103 // To catch this, we need to fold a compare and a select, hence '2' being the
104 // minimum reasonable default.
106  "phi-node-folding-threshold", cl::Hidden, cl::init(2),
107  cl::desc(
108  "Control the amount of phi node folding to perform (default = 2)"));
109 
111  "two-entry-phi-node-folding-threshold", cl::Hidden, cl::init(4),
112  cl::desc("Control the maximal total instruction cost that we are willing "
113  "to speculatively execute to fold a 2-entry PHI node into a "
114  "select (default = 4)"));
115 
116 static cl::opt<bool>
117  HoistCommon("simplifycfg-hoist-common", cl::Hidden, cl::init(true),
118  cl::desc("Hoist common instructions up to the parent block"));
119 
120 static cl::opt<bool>
121  SinkCommon("simplifycfg-sink-common", cl::Hidden, cl::init(true),
122  cl::desc("Sink common instructions down to the end block"));
123 
125  "simplifycfg-hoist-cond-stores", cl::Hidden, cl::init(true),
126  cl::desc("Hoist conditional stores if an unconditional store precedes"));
127 
129  "simplifycfg-merge-cond-stores", cl::Hidden, cl::init(true),
130  cl::desc("Hoist conditional stores even if an unconditional store does not "
131  "precede - hoist multiple conditional stores into a single "
132  "predicated store"));
133 
135  "simplifycfg-merge-cond-stores-aggressively", cl::Hidden, cl::init(false),
136  cl::desc("When merging conditional stores, do so even if the resultant "
137  "basic blocks are unlikely to be if-converted as a result"));
138 
140  "speculate-one-expensive-inst", cl::Hidden, cl::init(true),
141  cl::desc("Allow exactly one expensive instruction to be speculatively "
142  "executed"));
143 
145  "max-speculation-depth", cl::Hidden, cl::init(10),
146  cl::desc("Limit maximum recursion depth when calculating costs of "
147  "speculatively executed instructions"));
148 
149 static cl::opt<int>
150  MaxSmallBlockSize("simplifycfg-max-small-block-size", cl::Hidden,
151  cl::init(10),
152  cl::desc("Max size of a block which is still considered "
153  "small enough to thread through"));
154 
155 // Two is chosen to allow one negation and a logical combine.
156 static cl::opt<unsigned>
157  BranchFoldThreshold("simplifycfg-branch-fold-threshold", cl::Hidden,
158  cl::init(2),
159  cl::desc("Maximum cost of combining conditions when "
160  "folding branches"));
161 
162 STATISTIC(NumBitMaps, "Number of switch instructions turned into bitmaps");
163 STATISTIC(NumLinearMaps,
164  "Number of switch instructions turned into linear mapping");
165 STATISTIC(NumLookupTables,
166  "Number of switch instructions turned into lookup tables");
167 STATISTIC(
168  NumLookupTablesHoles,
169  "Number of switch instructions turned into lookup tables (holes checked)");
170 STATISTIC(NumTableCmpReuses, "Number of reused switch table lookup compares");
171 STATISTIC(NumFoldValueComparisonIntoPredecessors,
172  "Number of value comparisons folded into predecessor basic blocks");
173 STATISTIC(NumFoldBranchToCommonDest,
174  "Number of branches folded into predecessor basic block");
175 STATISTIC(
176  NumHoistCommonCode,
177  "Number of common instruction 'blocks' hoisted up to the begin block");
178 STATISTIC(NumHoistCommonInstrs,
179  "Number of common instructions hoisted up to the begin block");
180 STATISTIC(NumSinkCommonCode,
181  "Number of common instruction 'blocks' sunk down to the end block");
182 STATISTIC(NumSinkCommonInstrs,
183  "Number of common instructions sunk down to the end block");
184 STATISTIC(NumSpeculations, "Number of speculative executed instructions");
185 STATISTIC(NumInvokes,
186  "Number of invokes with empty resume blocks simplified into calls");
187 
188 namespace {
189 
190 // The first field contains the value that the switch produces when a certain
191 // case group is selected, and the second field is a vector containing the
192 // cases composing the case group.
193 using SwitchCaseResultVectorTy =
195 
196 // The first field contains the phi node that generates a result of the switch
197 // and the second field contains the value generated for a certain case in the
198 // switch for that PHI.
199 using SwitchCaseResultsTy = SmallVector<std::pair<PHINode *, Constant *>, 4>;
200 
201 /// ValueEqualityComparisonCase - Represents a case of a switch.
202 struct ValueEqualityComparisonCase {
204  BasicBlock *Dest;
205 
206  ValueEqualityComparisonCase(ConstantInt *Value, BasicBlock *Dest)
207  : Value(Value), Dest(Dest) {}
208 
209  bool operator<(ValueEqualityComparisonCase RHS) const {
210  // Comparing pointers is ok as we only rely on the order for uniquing.
211  return Value < RHS.Value;
212  }
213 
214  bool operator==(BasicBlock *RHSDest) const { return Dest == RHSDest; }
215 };
216 
217 class SimplifyCFGOpt {
218  const TargetTransformInfo &TTI;
219  DomTreeUpdater *DTU;
220  const DataLayout &DL;
221  ArrayRef<WeakVH> LoopHeaders;
223  bool Resimplify;
224 
225  Value *isValueEqualityComparison(Instruction *TI);
226  BasicBlock *GetValueEqualityComparisonCases(
227  Instruction *TI, std::vector<ValueEqualityComparisonCase> &Cases);
228  bool SimplifyEqualityComparisonWithOnlyPredecessor(Instruction *TI,
229  BasicBlock *Pred,
231  bool PerformValueComparisonIntoPredecessorFolding(Instruction *TI, Value *&CV,
232  Instruction *PTI,
234  bool FoldValueComparisonIntoPredecessors(Instruction *TI,
236 
237  bool simplifyResume(ResumeInst *RI, IRBuilder<> &Builder);
238  bool simplifySingleResume(ResumeInst *RI);
239  bool simplifyCommonResume(ResumeInst *RI);
240  bool simplifyCleanupReturn(CleanupReturnInst *RI);
241  bool simplifyUnreachable(UnreachableInst *UI);
242  bool simplifySwitch(SwitchInst *SI, IRBuilder<> &Builder);
243  bool simplifyIndirectBr(IndirectBrInst *IBI);
244  bool simplifyBranch(BranchInst *Branch, IRBuilder<> &Builder);
245  bool simplifyUncondBranch(BranchInst *BI, IRBuilder<> &Builder);
246  bool simplifyCondBranch(BranchInst *BI, IRBuilder<> &Builder);
247 
248  bool tryToSimplifyUncondBranchWithICmpInIt(ICmpInst *ICI,
250 
251  bool HoistThenElseCodeToIf(BranchInst *BI, const TargetTransformInfo &TTI,
252  bool EqTermsOnly);
253  bool SpeculativelyExecuteBB(BranchInst *BI, BasicBlock *ThenBB,
254  const TargetTransformInfo &TTI);
255  bool SimplifyTerminatorOnSelect(Instruction *OldTerm, Value *Cond,
256  BasicBlock *TrueBB, BasicBlock *FalseBB,
257  uint32_t TrueWeight, uint32_t FalseWeight);
258  bool SimplifyBranchOnICmpChain(BranchInst *BI, IRBuilder<> &Builder,
259  const DataLayout &DL);
260  bool SimplifySwitchOnSelect(SwitchInst *SI, SelectInst *Select);
261  bool SimplifyIndirectBrOnSelect(IndirectBrInst *IBI, SelectInst *SI);
262  bool TurnSwitchRangeIntoICmp(SwitchInst *SI, IRBuilder<> &Builder);
263 
264 public:
265  SimplifyCFGOpt(const TargetTransformInfo &TTI, DomTreeUpdater *DTU,
266  const DataLayout &DL, ArrayRef<WeakVH> LoopHeaders,
267  const SimplifyCFGOptions &Opts)
268  : TTI(TTI), DTU(DTU), DL(DL), LoopHeaders(LoopHeaders), Options(Opts) {
269  assert((!DTU || !DTU->hasPostDomTree()) &&
270  "SimplifyCFG is not yet capable of maintaining validity of a "
271  "PostDomTree, so don't ask for it.");
272  }
273 
274  bool simplifyOnce(BasicBlock *BB);
275  bool simplifyOnceImpl(BasicBlock *BB);
276  bool run(BasicBlock *BB);
277 
278  // Helper to set Resimplify and return change indication.
279  bool requestResimplify() {
280  Resimplify = true;
281  return true;
282  }
283 };
284 
285 } // end anonymous namespace
286 
287 /// Return true if it is safe to merge these two
288 /// terminator instructions together.
289 static bool
291  SmallSetVector<BasicBlock *, 4> *FailBlocks = nullptr) {
292  if (SI1 == SI2)
293  return false; // Can't merge with self!
294 
295  // It is not safe to merge these two switch instructions if they have a common
296  // successor, and if that successor has a PHI node, and if *that* PHI node has
297  // conflicting incoming values from the two switch blocks.
298  BasicBlock *SI1BB = SI1->getParent();
299  BasicBlock *SI2BB = SI2->getParent();
300 
301  SmallPtrSet<BasicBlock *, 16> SI1Succs(succ_begin(SI1BB), succ_end(SI1BB));
302  bool Fail = false;
303  for (BasicBlock *Succ : successors(SI2BB))
304  if (SI1Succs.count(Succ))
305  for (BasicBlock::iterator BBI = Succ->begin(); isa<PHINode>(BBI); ++BBI) {
306  PHINode *PN = cast<PHINode>(BBI);
307  if (PN->getIncomingValueForBlock(SI1BB) !=
308  PN->getIncomingValueForBlock(SI2BB)) {
309  if (FailBlocks)
310  FailBlocks->insert(Succ);
311  Fail = true;
312  }
313  }
314 
315  return !Fail;
316 }
317 
318 /// Update PHI nodes in Succ to indicate that there will now be entries in it
319 /// from the 'NewPred' block. The values that will be flowing into the PHI nodes
320 /// will be the same as those coming in from ExistPred, an existing predecessor
321 /// of Succ.
322 static void AddPredecessorToBlock(BasicBlock *Succ, BasicBlock *NewPred,
323  BasicBlock *ExistPred,
324  MemorySSAUpdater *MSSAU = nullptr) {
325  for (PHINode &PN : Succ->phis())
326  PN.addIncoming(PN.getIncomingValueForBlock(ExistPred), NewPred);
327  if (MSSAU)
328  if (auto *MPhi = MSSAU->getMemorySSA()->getMemoryAccess(Succ))
329  MPhi->addIncoming(MPhi->getIncomingValueForBlock(ExistPred), NewPred);
330 }
331 
332 /// Compute an abstract "cost" of speculating the given instruction,
333 /// which is assumed to be safe to speculate. TCC_Free means cheap,
334 /// TCC_Basic means less cheap, and TCC_Expensive means prohibitively
335 /// expensive.
337  const TargetTransformInfo &TTI) {
339  "Instruction is not safe to speculatively execute!");
341 }
342 
343 /// If we have a merge point of an "if condition" as accepted above,
344 /// return true if the specified value dominates the block. We
345 /// don't handle the true generality of domination here, just a special case
346 /// which works well enough for us.
347 ///
348 /// If AggressiveInsts is non-null, and if V does not dominate BB, we check to
349 /// see if V (which must be an instruction) and its recursive operands
350 /// that do not dominate BB have a combined cost lower than Budget and
351 /// are non-trapping. If both are true, the instruction is inserted into the
352 /// set and true is returned.
353 ///
354 /// The cost for most non-trapping instructions is defined as 1 except for
355 /// Select whose cost is 2.
356 ///
357 /// After this function returns, Cost is increased by the cost of
358 /// V plus its non-dominating operands. If that cost is greater than
359 /// Budget, false is returned and Cost is undefined.
361  SmallPtrSetImpl<Instruction *> &AggressiveInsts,
362  InstructionCost &Cost,
363  InstructionCost Budget,
364  const TargetTransformInfo &TTI,
365  unsigned Depth = 0) {
366  // It is possible to hit a zero-cost cycle (phi/gep instructions for example),
367  // so limit the recursion depth.
368  // TODO: While this recursion limit does prevent pathological behavior, it
369  // would be better to track visited instructions to avoid cycles.
370  if (Depth == MaxSpeculationDepth)
371  return false;
372 
373  Instruction *I = dyn_cast<Instruction>(V);
374  if (!I) {
375  // Non-instructions all dominate instructions, but not all constantexprs
376  // can be executed unconditionally.
377  if (ConstantExpr *C = dyn_cast<ConstantExpr>(V))
378  if (C->canTrap())
379  return false;
380  return true;
381  }
382  BasicBlock *PBB = I->getParent();
383 
384  // We don't want to allow weird loops that might have the "if condition" in
385  // the bottom of this block.
386  if (PBB == BB)
387  return false;
388 
389  // If this instruction is defined in a block that contains an unconditional
390  // branch to BB, then it must be in the 'conditional' part of the "if
391  // statement". If not, it definitely dominates the region.
392  BranchInst *BI = dyn_cast<BranchInst>(PBB->getTerminator());
393  if (!BI || BI->isConditional() || BI->getSuccessor(0) != BB)
394  return true;
395 
396  // If we have seen this instruction before, don't count it again.
397  if (AggressiveInsts.count(I))
398  return true;
399 
400  // Okay, it looks like the instruction IS in the "condition". Check to
401  // see if it's a cheap instruction to unconditionally compute, and if it
402  // only uses stuff defined outside of the condition. If so, hoist it out.
404  return false;
405 
406  Cost += computeSpeculationCost(I, TTI);
407 
408  // Allow exactly one instruction to be speculated regardless of its cost
409  // (as long as it is safe to do so).
410  // This is intended to flatten the CFG even if the instruction is a division
411  // or other expensive operation. The speculation of an expensive instruction
412  // is expected to be undone in CodeGenPrepare if the speculation has not
413  // enabled further IR optimizations.
414  if (Cost > Budget &&
415  (!SpeculateOneExpensiveInst || !AggressiveInsts.empty() || Depth > 0 ||
416  !Cost.isValid()))
417  return false;
418 
419  // Okay, we can only really hoist these out if their operands do
420  // not take us over the cost threshold.
421  for (Use &Op : I->operands())
422  if (!dominatesMergePoint(Op, BB, AggressiveInsts, Cost, Budget, TTI,
423  Depth + 1))
424  return false;
425  // Okay, it's safe to do this! Remember this instruction.
426  AggressiveInsts.insert(I);
427  return true;
428 }
429 
430 /// Extract ConstantInt from value, looking through IntToPtr
431 /// and PointerNullValue. Return NULL if value is not a constant int.
433  // Normal constant int.
434  ConstantInt *CI = dyn_cast<ConstantInt>(V);
435  if (CI || !isa<Constant>(V) || !V->getType()->isPointerTy())
436  return CI;
437 
438  // This is some kind of pointer constant. Turn it into a pointer-sized
439  // ConstantInt if possible.
440  IntegerType *PtrTy = cast<IntegerType>(DL.getIntPtrType(V->getType()));
441 
442  // Null pointer means 0, see SelectionDAGBuilder::getValue(const Value*).
443  if (isa<ConstantPointerNull>(V))
444  return ConstantInt::get(PtrTy, 0);
445 
446  // IntToPtr const int.
447  if (ConstantExpr *CE = dyn_cast<ConstantExpr>(V))
448  if (CE->getOpcode() == Instruction::IntToPtr)
449  if (ConstantInt *CI = dyn_cast<ConstantInt>(CE->getOperand(0))) {
450  // The constant is very likely to have the right type already.
451  if (CI->getType() == PtrTy)
452  return CI;
453  else
454  return cast<ConstantInt>(
455  ConstantExpr::getIntegerCast(CI, PtrTy, /*isSigned=*/false));
456  }
457  return nullptr;
458 }
459 
460 namespace {
461 
462 /// Given a chain of or (||) or and (&&) comparison of a value against a
463 /// constant, this will try to recover the information required for a switch
464 /// structure.
465 /// It will depth-first traverse the chain of comparison, seeking for patterns
466 /// like %a == 12 or %a < 4 and combine them to produce a set of integer
467 /// representing the different cases for the switch.
468 /// Note that if the chain is composed of '||' it will build the set of elements
469 /// that matches the comparisons (i.e. any of this value validate the chain)
470 /// while for a chain of '&&' it will build the set elements that make the test
471 /// fail.
472 struct ConstantComparesGatherer {
473  const DataLayout &DL;
474 
475  /// Value found for the switch comparison
476  Value *CompValue = nullptr;
477 
478  /// Extra clause to be checked before the switch
479  Value *Extra = nullptr;
480 
481  /// Set of integers to match in switch
483 
484  /// Number of comparisons matched in the and/or chain
485  unsigned UsedICmps = 0;
486 
487  /// Construct and compute the result for the comparison instruction Cond
488  ConstantComparesGatherer(Instruction *Cond, const DataLayout &DL) : DL(DL) {
489  gather(Cond);
490  }
491 
492  ConstantComparesGatherer(const ConstantComparesGatherer &) = delete;
493  ConstantComparesGatherer &
494  operator=(const ConstantComparesGatherer &) = delete;
495 
496 private:
497  /// Try to set the current value used for the comparison, it succeeds only if
498  /// it wasn't set before or if the new value is the same as the old one
499  bool setValueOnce(Value *NewVal) {
500  if (CompValue && CompValue != NewVal)
501  return false;
502  CompValue = NewVal;
503  return (CompValue != nullptr);
504  }
505 
506  /// Try to match Instruction "I" as a comparison against a constant and
507  /// populates the array Vals with the set of values that match (or do not
508  /// match depending on isEQ).
509  /// Return false on failure. On success, the Value the comparison matched
510  /// against is placed in CompValue.
511  /// If CompValue is already set, the function is expected to fail if a match
512  /// is found but the value compared to is different.
513  bool matchInstruction(Instruction *I, bool isEQ) {
514  // If this is an icmp against a constant, handle this as one of the cases.
515  ICmpInst *ICI;
516  ConstantInt *C;
517  if (!((ICI = dyn_cast<ICmpInst>(I)) &&
518  (C = GetConstantInt(I->getOperand(1), DL)))) {
519  return false;
520  }
521 
522  Value *RHSVal;
523  const APInt *RHSC;
524 
525  // Pattern match a special case
526  // (x & ~2^z) == y --> x == y || x == y|2^z
527  // This undoes a transformation done by instcombine to fuse 2 compares.
528  if (ICI->getPredicate() == (isEQ ? ICmpInst::ICMP_EQ : ICmpInst::ICMP_NE)) {
529  // It's a little bit hard to see why the following transformations are
530  // correct. Here is a CVC3 program to verify them for 64-bit values:
531 
532  /*
533  ONE : BITVECTOR(64) = BVZEROEXTEND(0bin1, 63);
534  x : BITVECTOR(64);
535  y : BITVECTOR(64);
536  z : BITVECTOR(64);
537  mask : BITVECTOR(64) = BVSHL(ONE, z);
538  QUERY( (y & ~mask = y) =>
539  ((x & ~mask = y) <=> (x = y OR x = (y | mask)))
540  );
541  QUERY( (y | mask = y) =>
542  ((x | mask = y) <=> (x = y OR x = (y & ~mask)))
543  );
544  */
545 
546  // Please note that each pattern must be a dual implication (<--> or
547  // iff). One directional implication can create spurious matches. If the
548  // implication is only one-way, an unsatisfiable condition on the left
549  // side can imply a satisfiable condition on the right side. Dual
550  // implication ensures that satisfiable conditions are transformed to
551  // other satisfiable conditions and unsatisfiable conditions are
552  // transformed to other unsatisfiable conditions.
553 
554  // Here is a concrete example of a unsatisfiable condition on the left
555  // implying a satisfiable condition on the right:
556  //
557  // mask = (1 << z)
558  // (x & ~mask) == y --> (x == y || x == (y | mask))
559  //
560  // Substituting y = 3, z = 0 yields:
561  // (x & -2) == 3 --> (x == 3 || x == 2)
562 
563  // Pattern match a special case:
564  /*
565  QUERY( (y & ~mask = y) =>
566  ((x & ~mask = y) <=> (x = y OR x = (y | mask)))
567  );
568  */
569  if (match(ICI->getOperand(0),
570  m_And(m_Value(RHSVal), m_APInt(RHSC)))) {
571  APInt Mask = ~*RHSC;
572  if (Mask.isPowerOf2() && (C->getValue() & ~Mask) == C->getValue()) {
573  // If we already have a value for the switch, it has to match!
574  if (!setValueOnce(RHSVal))
575  return false;
576 
577  Vals.push_back(C);
578  Vals.push_back(
579  ConstantInt::get(C->getContext(),
580  C->getValue() | Mask));
581  UsedICmps++;
582  return true;
583  }
584  }
585 
586  // Pattern match a special case:
587  /*
588  QUERY( (y | mask = y) =>
589  ((x | mask = y) <=> (x = y OR x = (y & ~mask)))
590  );
591  */
592  if (match(ICI->getOperand(0),
593  m_Or(m_Value(RHSVal), m_APInt(RHSC)))) {
594  APInt Mask = *RHSC;
595  if (Mask.isPowerOf2() && (C->getValue() | Mask) == C->getValue()) {
596  // If we already have a value for the switch, it has to match!
597  if (!setValueOnce(RHSVal))
598  return false;
599 
600  Vals.push_back(C);
601  Vals.push_back(ConstantInt::get(C->getContext(),
602  C->getValue() & ~Mask));
603  UsedICmps++;
604  return true;
605  }
606  }
607 
608  // If we already have a value for the switch, it has to match!
609  if (!setValueOnce(ICI->getOperand(0)))
610  return false;
611 
612  UsedICmps++;
613  Vals.push_back(C);
614  return ICI->getOperand(0);
615  }
616 
617  // If we have "x ult 3", for example, then we can add 0,1,2 to the set.
618  ConstantRange Span =
620 
621  // Shift the range if the compare is fed by an add. This is the range
622  // compare idiom as emitted by instcombine.
623  Value *CandidateVal = I->getOperand(0);
624  if (match(I->getOperand(0), m_Add(m_Value(RHSVal), m_APInt(RHSC)))) {
625  Span = Span.subtract(*RHSC);
626  CandidateVal = RHSVal;
627  }
628 
629  // If this is an and/!= check, then we are looking to build the set of
630  // value that *don't* pass the and chain. I.e. to turn "x ugt 2" into
631  // x != 0 && x != 1.
632  if (!isEQ)
633  Span = Span.inverse();
634 
635  // If there are a ton of values, we don't want to make a ginormous switch.
636  if (Span.isSizeLargerThan(8) || Span.isEmptySet()) {
637  return false;
638  }
639 
640  // If we already have a value for the switch, it has to match!
641  if (!setValueOnce(CandidateVal))
642  return false;
643 
644  // Add all values from the range to the set
645  for (APInt Tmp = Span.getLower(); Tmp != Span.getUpper(); ++Tmp)
646  Vals.push_back(ConstantInt::get(I->getContext(), Tmp));
647 
648  UsedICmps++;
649  return true;
650  }
651 
652  /// Given a potentially 'or'd or 'and'd together collection of icmp
653  /// eq/ne/lt/gt instructions that compare a value against a constant, extract
654  /// the value being compared, and stick the list constants into the Vals
655  /// vector.
656  /// One "Extra" case is allowed to differ from the other.
657  void gather(Value *V) {
658  bool isEQ = match(V, m_LogicalOr(m_Value(), m_Value()));
659 
660  // Keep a stack (SmallVector for efficiency) for depth-first traversal
662  SmallPtrSet<Value *, 8> Visited;
663 
664  // Initialize
665  Visited.insert(V);
666  DFT.push_back(V);
667 
668  while (!DFT.empty()) {
669  V = DFT.pop_back_val();
670 
671  if (Instruction *I = dyn_cast<Instruction>(V)) {
672  // If it is a || (or && depending on isEQ), process the operands.
673  Value *Op0, *Op1;
674  if (isEQ ? match(I, m_LogicalOr(m_Value(Op0), m_Value(Op1)))
675  : match(I, m_LogicalAnd(m_Value(Op0), m_Value(Op1)))) {
676  if (Visited.insert(Op1).second)
677  DFT.push_back(Op1);
678  if (Visited.insert(Op0).second)
679  DFT.push_back(Op0);
680 
681  continue;
682  }
683 
684  // Try to match the current instruction
685  if (matchInstruction(I, isEQ))
686  // Match succeed, continue the loop
687  continue;
688  }
689 
690  // One element of the sequence of || (or &&) could not be match as a
691  // comparison against the same value as the others.
692  // We allow only one "Extra" case to be checked before the switch
693  if (!Extra) {
694  Extra = V;
695  continue;
696  }
697  // Failed to parse a proper sequence, abort now
698  CompValue = nullptr;
699  break;
700  }
701  }
702 };
703 
704 } // end anonymous namespace
705 
707  MemorySSAUpdater *MSSAU = nullptr) {
708  Instruction *Cond = nullptr;
709  if (SwitchInst *SI = dyn_cast<SwitchInst>(TI)) {
710  Cond = dyn_cast<Instruction>(SI->getCondition());
711  } else if (BranchInst *BI = dyn_cast<BranchInst>(TI)) {
712  if (BI->isConditional())
713  Cond = dyn_cast<Instruction>(BI->getCondition());
714  } else if (IndirectBrInst *IBI = dyn_cast<IndirectBrInst>(TI)) {
715  Cond = dyn_cast<Instruction>(IBI->getAddress());
716  }
717 
718  TI->eraseFromParent();
719  if (Cond)
721 }
722 
723 /// Return true if the specified terminator checks
724 /// to see if a value is equal to constant integer value.
725 Value *SimplifyCFGOpt::isValueEqualityComparison(Instruction *TI) {
726  Value *CV = nullptr;
727  if (SwitchInst *SI = dyn_cast<SwitchInst>(TI)) {
728  // Do not permit merging of large switch instructions into their
729  // predecessors unless there is only one predecessor.
730  if (!SI->getParent()->hasNPredecessorsOrMore(128 / SI->getNumSuccessors()))
731  CV = SI->getCondition();
732  } else if (BranchInst *BI = dyn_cast<BranchInst>(TI))
733  if (BI->isConditional() && BI->getCondition()->hasOneUse())
734  if (ICmpInst *ICI = dyn_cast<ICmpInst>(BI->getCondition())) {
735  if (ICI->isEquality() && GetConstantInt(ICI->getOperand(1), DL))
736  CV = ICI->getOperand(0);
737  }
738 
739  // Unwrap any lossless ptrtoint cast.
740  if (CV) {
741  if (PtrToIntInst *PTII = dyn_cast<PtrToIntInst>(CV)) {
742  Value *Ptr = PTII->getPointerOperand();
743  if (PTII->getType() == DL.getIntPtrType(Ptr->getType()))
744  CV = Ptr;
745  }
746  }
747  return CV;
748 }
749 
750 /// Given a value comparison instruction,
751 /// decode all of the 'cases' that it represents and return the 'default' block.
752 BasicBlock *SimplifyCFGOpt::GetValueEqualityComparisonCases(
753  Instruction *TI, std::vector<ValueEqualityComparisonCase> &Cases) {
754  if (SwitchInst *SI = dyn_cast<SwitchInst>(TI)) {
755  Cases.reserve(SI->getNumCases());
756  for (auto Case : SI->cases())
757  Cases.push_back(ValueEqualityComparisonCase(Case.getCaseValue(),
758  Case.getCaseSuccessor()));
759  return SI->getDefaultDest();
760  }
761 
762  BranchInst *BI = cast<BranchInst>(TI);
763  ICmpInst *ICI = cast<ICmpInst>(BI->getCondition());
764  BasicBlock *Succ = BI->getSuccessor(ICI->getPredicate() == ICmpInst::ICMP_NE);
765  Cases.push_back(ValueEqualityComparisonCase(
766  GetConstantInt(ICI->getOperand(1), DL), Succ));
767  return BI->getSuccessor(ICI->getPredicate() == ICmpInst::ICMP_EQ);
768 }
769 
770 /// Given a vector of bb/value pairs, remove any entries
771 /// in the list that match the specified block.
772 static void
774  std::vector<ValueEqualityComparisonCase> &Cases) {
775  llvm::erase_value(Cases, BB);
776 }
777 
778 /// Return true if there are any keys in C1 that exist in C2 as well.
779 static bool ValuesOverlap(std::vector<ValueEqualityComparisonCase> &C1,
780  std::vector<ValueEqualityComparisonCase> &C2) {
781  std::vector<ValueEqualityComparisonCase> *V1 = &C1, *V2 = &C2;
782 
783  // Make V1 be smaller than V2.
784  if (V1->size() > V2->size())
785  std::swap(V1, V2);
786 
787  if (V1->empty())
788  return false;
789  if (V1->size() == 1) {
790  // Just scan V2.
791  ConstantInt *TheVal = (*V1)[0].Value;
792  for (unsigned i = 0, e = V2->size(); i != e; ++i)
793  if (TheVal == (*V2)[i].Value)
794  return true;
795  }
796 
797  // Otherwise, just sort both lists and compare element by element.
798  array_pod_sort(V1->begin(), V1->end());
799  array_pod_sort(V2->begin(), V2->end());
800  unsigned i1 = 0, i2 = 0, e1 = V1->size(), e2 = V2->size();
801  while (i1 != e1 && i2 != e2) {
802  if ((*V1)[i1].Value == (*V2)[i2].Value)
803  return true;
804  if ((*V1)[i1].Value < (*V2)[i2].Value)
805  ++i1;
806  else
807  ++i2;
808  }
809  return false;
810 }
811 
812 // Set branch weights on SwitchInst. This sets the metadata if there is at
813 // least one non-zero weight.
815  // Check that there is at least one non-zero weight. Otherwise, pass
816  // nullptr to setMetadata which will erase the existing metadata.
817  MDNode *N = nullptr;
818  if (llvm::any_of(Weights, [](uint32_t W) { return W != 0; }))
819  N = MDBuilder(SI->getParent()->getContext()).createBranchWeights(Weights);
820  SI->setMetadata(LLVMContext::MD_prof, N);
821 }
822 
823 // Similar to the above, but for branch and select instructions that take
824 // exactly 2 weights.
825 static void setBranchWeights(Instruction *I, uint32_t TrueWeight,
826  uint32_t FalseWeight) {
827  assert(isa<BranchInst>(I) || isa<SelectInst>(I));
828  // Check that there is at least one non-zero weight. Otherwise, pass
829  // nullptr to setMetadata which will erase the existing metadata.
830  MDNode *N = nullptr;
831  if (TrueWeight || FalseWeight)
832  N = MDBuilder(I->getParent()->getContext())
833  .createBranchWeights(TrueWeight, FalseWeight);
834  I->setMetadata(LLVMContext::MD_prof, N);
835 }
836 
837 /// If TI is known to be a terminator instruction and its block is known to
838 /// only have a single predecessor block, check to see if that predecessor is
839 /// also a value comparison with the same value, and if that comparison
840 /// determines the outcome of this comparison. If so, simplify TI. This does a
841 /// very limited form of jump threading.
842 bool SimplifyCFGOpt::SimplifyEqualityComparisonWithOnlyPredecessor(
843  Instruction *TI, BasicBlock *Pred, IRBuilder<> &Builder) {
844  Value *PredVal = isValueEqualityComparison(Pred->getTerminator());
845  if (!PredVal)
846  return false; // Not a value comparison in predecessor.
847 
848  Value *ThisVal = isValueEqualityComparison(TI);
849  assert(ThisVal && "This isn't a value comparison!!");
850  if (ThisVal != PredVal)
851  return false; // Different predicates.
852 
853  // TODO: Preserve branch weight metadata, similarly to how
854  // FoldValueComparisonIntoPredecessors preserves it.
855 
856  // Find out information about when control will move from Pred to TI's block.
857  std::vector<ValueEqualityComparisonCase> PredCases;
858  BasicBlock *PredDef =
859  GetValueEqualityComparisonCases(Pred->getTerminator(), PredCases);
860  EliminateBlockCases(PredDef, PredCases); // Remove default from cases.
861 
862  // Find information about how control leaves this block.
863  std::vector<ValueEqualityComparisonCase> ThisCases;
864  BasicBlock *ThisDef = GetValueEqualityComparisonCases(TI, ThisCases);
865  EliminateBlockCases(ThisDef, ThisCases); // Remove default from cases.
866 
867  // If TI's block is the default block from Pred's comparison, potentially
868  // simplify TI based on this knowledge.
869  if (PredDef == TI->getParent()) {
870  // If we are here, we know that the value is none of those cases listed in
871  // PredCases. If there are any cases in ThisCases that are in PredCases, we
872  // can simplify TI.
873  if (!ValuesOverlap(PredCases, ThisCases))
874  return false;
875 
876  if (isa<BranchInst>(TI)) {
877  // Okay, one of the successors of this condbr is dead. Convert it to a
878  // uncond br.
879  assert(ThisCases.size() == 1 && "Branch can only have one case!");
880  // Insert the new branch.
881  Instruction *NI = Builder.CreateBr(ThisDef);
882  (void)NI;
883 
884  // Remove PHI node entries for the dead edge.
885  ThisCases[0].Dest->removePredecessor(PredDef);
886 
887  LLVM_DEBUG(dbgs() << "Threading pred instr: " << *Pred->getTerminator()
888  << "Through successor TI: " << *TI << "Leaving: " << *NI
889  << "\n");
890 
892 
893  if (DTU)
894  DTU->applyUpdates(
895  {{DominatorTree::Delete, PredDef, ThisCases[0].Dest}});
896 
897  return true;
898  }
899 
900  SwitchInstProfUpdateWrapper SI = *cast<SwitchInst>(TI);
901  // Okay, TI has cases that are statically dead, prune them away.
902  SmallPtrSet<Constant *, 16> DeadCases;
903  for (unsigned i = 0, e = PredCases.size(); i != e; ++i)
904  DeadCases.insert(PredCases[i].Value);
905 
906  LLVM_DEBUG(dbgs() << "Threading pred instr: " << *Pred->getTerminator()
907  << "Through successor TI: " << *TI);
908 
909  SmallDenseMap<BasicBlock *, int, 8> NumPerSuccessorCases;
910  for (SwitchInst::CaseIt i = SI->case_end(), e = SI->case_begin(); i != e;) {
911  --i;
912  auto *Successor = i->getCaseSuccessor();
913  if (DTU)
914  ++NumPerSuccessorCases[Successor];
915  if (DeadCases.count(i->getCaseValue())) {
916  Successor->removePredecessor(PredDef);
917  SI.removeCase(i);
918  if (DTU)
919  --NumPerSuccessorCases[Successor];
920  }
921  }
922 
923  if (DTU) {
924  std::vector<DominatorTree::UpdateType> Updates;
925  for (const std::pair<BasicBlock *, int> &I : NumPerSuccessorCases)
926  if (I.second == 0)
927  Updates.push_back({DominatorTree::Delete, PredDef, I.first});
928  DTU->applyUpdates(Updates);
929  }
930 
931  LLVM_DEBUG(dbgs() << "Leaving: " << *TI << "\n");
932  return true;
933  }
934 
935  // Otherwise, TI's block must correspond to some matched value. Find out
936  // which value (or set of values) this is.
937  ConstantInt *TIV = nullptr;
938  BasicBlock *TIBB = TI->getParent();
939  for (unsigned i = 0, e = PredCases.size(); i != e; ++i)
940  if (PredCases[i].Dest == TIBB) {
941  if (TIV)
942  return false; // Cannot handle multiple values coming to this block.
943  TIV = PredCases[i].Value;
944  }
945  assert(TIV && "No edge from pred to succ?");
946 
947  // Okay, we found the one constant that our value can be if we get into TI's
948  // BB. Find out which successor will unconditionally be branched to.
949  BasicBlock *TheRealDest = nullptr;
950  for (unsigned i = 0, e = ThisCases.size(); i != e; ++i)
951  if (ThisCases[i].Value == TIV) {
952  TheRealDest = ThisCases[i].Dest;
953  break;
954  }
955 
956  // If not handled by any explicit cases, it is handled by the default case.
957  if (!TheRealDest)
958  TheRealDest = ThisDef;
959 
960  SmallPtrSet<BasicBlock *, 2> RemovedSuccs;
961 
962  // Remove PHI node entries for dead edges.
963  BasicBlock *CheckEdge = TheRealDest;
964  for (BasicBlock *Succ : successors(TIBB))
965  if (Succ != CheckEdge) {
966  if (Succ != TheRealDest)
967  RemovedSuccs.insert(Succ);
968  Succ->removePredecessor(TIBB);
969  } else
970  CheckEdge = nullptr;
971 
972  // Insert the new branch.
973  Instruction *NI = Builder.CreateBr(TheRealDest);
974  (void)NI;
975 
976  LLVM_DEBUG(dbgs() << "Threading pred instr: " << *Pred->getTerminator()
977  << "Through successor TI: " << *TI << "Leaving: " << *NI
978  << "\n");
979 
981  if (DTU) {
983  Updates.reserve(RemovedSuccs.size());
984  for (auto *RemovedSucc : RemovedSuccs)
985  Updates.push_back({DominatorTree::Delete, TIBB, RemovedSucc});
986  DTU->applyUpdates(Updates);
987  }
988  return true;
989 }
990 
991 namespace {
992 
993 /// This class implements a stable ordering of constant
994 /// integers that does not depend on their address. This is important for
995 /// applications that sort ConstantInt's to ensure uniqueness.
996 struct ConstantIntOrdering {
997  bool operator()(const ConstantInt *LHS, const ConstantInt *RHS) const {
998  return LHS->getValue().ult(RHS->getValue());
999  }
1000 };
1001 
1002 } // end anonymous namespace
1003 
1005  ConstantInt *const *P2) {
1006  const ConstantInt *LHS = *P1;
1007  const ConstantInt *RHS = *P2;
1008  if (LHS == RHS)
1009  return 0;
1010  return LHS->getValue().ult(RHS->getValue()) ? 1 : -1;
1011 }
1012 
1013 static inline bool HasBranchWeights(const Instruction *I) {
1014  MDNode *ProfMD = I->getMetadata(LLVMContext::MD_prof);
1015  if (ProfMD && ProfMD->getOperand(0))
1016  if (MDString *MDS = dyn_cast<MDString>(ProfMD->getOperand(0)))
1017  return MDS->getString().equals("branch_weights");
1018 
1019  return false;
1020 }
1021 
1022 /// Get Weights of a given terminator, the default weight is at the front
1023 /// of the vector. If TI is a conditional eq, we need to swap the branch-weight
1024 /// metadata.
1026  SmallVectorImpl<uint64_t> &Weights) {
1027  MDNode *MD = TI->getMetadata(LLVMContext::MD_prof);
1028  assert(MD);
1029  for (unsigned i = 1, e = MD->getNumOperands(); i < e; ++i) {
1030  ConstantInt *CI = mdconst::extract<ConstantInt>(MD->getOperand(i));
1031  Weights.push_back(CI->getValue().getZExtValue());
1032  }
1033 
1034  // If TI is a conditional eq, the default case is the false case,
1035  // and the corresponding branch-weight data is at index 2. We swap the
1036  // default weight to be the first entry.
1037  if (BranchInst *BI = dyn_cast<BranchInst>(TI)) {
1038  assert(Weights.size() == 2);
1039  ICmpInst *ICI = cast<ICmpInst>(BI->getCondition());
1040  if (ICI->getPredicate() == ICmpInst::ICMP_EQ)
1041  std::swap(Weights.front(), Weights.back());
1042  }
1043 }
1044 
1045 /// Keep halving the weights until all can fit in uint32_t.
1047  uint64_t Max = *std::max_element(Weights.begin(), Weights.end());
1048  if (Max > UINT_MAX) {
1049  unsigned Offset = 32 - countLeadingZeros(Max);
1050  for (uint64_t &I : Weights)
1051  I >>= Offset;
1052  }
1053 }
1054 
1056  BasicBlock *BB, BasicBlock *PredBlock, ValueToValueMapTy &VMap) {
1057  Instruction *PTI = PredBlock->getTerminator();
1058 
1059  // If we have bonus instructions, clone them into the predecessor block.
1060  // Note that there may be multiple predecessor blocks, so we cannot move
1061  // bonus instructions to a predecessor block.
1062  for (Instruction &BonusInst : *BB) {
1063  if (isa<DbgInfoIntrinsic>(BonusInst) || BonusInst.isTerminator())
1064  continue;
1065 
1066  Instruction *NewBonusInst = BonusInst.clone();
1067 
1068  if (PTI->getDebugLoc() != NewBonusInst->getDebugLoc()) {
1069  // Unless the instruction has the same !dbg location as the original
1070  // branch, drop it. When we fold the bonus instructions we want to make
1071  // sure we reset their debug locations in order to avoid stepping on
1072  // dead code caused by folding dead branches.
1073  NewBonusInst->setDebugLoc(DebugLoc());
1074  }
1075 
1076  RemapInstruction(NewBonusInst, VMap,
1078  VMap[&BonusInst] = NewBonusInst;
1079 
1080  // If we moved a load, we cannot any longer claim any knowledge about
1081  // its potential value. The previous information might have been valid
1082  // only given the branch precondition.
1083  // For an analogous reason, we must also drop all the metadata whose
1084  // semantics we don't understand. We *can* preserve !annotation, because
1085  // it is tied to the instruction itself, not the value or position.
1086  // Similarly strip attributes on call parameters that may cause UB in
1087  // location the call is moved to.
1089  LLVMContext::MD_annotation);
1090 
1091  PredBlock->getInstList().insert(PTI->getIterator(), NewBonusInst);
1092  NewBonusInst->takeName(&BonusInst);
1093  BonusInst.setName(NewBonusInst->getName() + ".old");
1094 
1095  // Update (liveout) uses of bonus instructions,
1096  // now that the bonus instruction has been cloned into predecessor.
1097  SSAUpdater SSAUpdate;
1098  SSAUpdate.Initialize(BonusInst.getType(),
1099  (NewBonusInst->getName() + ".merge").str());
1100  SSAUpdate.AddAvailableValue(BB, &BonusInst);
1101  SSAUpdate.AddAvailableValue(PredBlock, NewBonusInst);
1102  for (Use &U : make_early_inc_range(BonusInst.uses())) {
1103  auto *UI = cast<Instruction>(U.getUser());
1104  if (UI->getParent() != PredBlock)
1105  SSAUpdate.RewriteUseAfterInsertions(U);
1106  else // Use is in the same block as, and comes before, NewBonusInst.
1107  SSAUpdate.RewriteUse(U);
1108  }
1109  }
1110 }
1111 
1112 bool SimplifyCFGOpt::PerformValueComparisonIntoPredecessorFolding(
1113  Instruction *TI, Value *&CV, Instruction *PTI, IRBuilder<> &Builder) {
1114  BasicBlock *BB = TI->getParent();
1115  BasicBlock *Pred = PTI->getParent();
1116 
1118 
1119  // Figure out which 'cases' to copy from SI to PSI.
1120  std::vector<ValueEqualityComparisonCase> BBCases;
1121  BasicBlock *BBDefault = GetValueEqualityComparisonCases(TI, BBCases);
1122 
1123  std::vector<ValueEqualityComparisonCase> PredCases;
1124  BasicBlock *PredDefault = GetValueEqualityComparisonCases(PTI, PredCases);
1125 
1126  // Based on whether the default edge from PTI goes to BB or not, fill in
1127  // PredCases and PredDefault with the new switch cases we would like to
1128  // build.
1130 
1131  // Update the branch weight metadata along the way
1132  SmallVector<uint64_t, 8> Weights;
1133  bool PredHasWeights = HasBranchWeights(PTI);
1134  bool SuccHasWeights = HasBranchWeights(TI);
1135 
1136  if (PredHasWeights) {
1137  GetBranchWeights(PTI, Weights);
1138  // branch-weight metadata is inconsistent here.
1139  if (Weights.size() != 1 + PredCases.size())
1140  PredHasWeights = SuccHasWeights = false;
1141  } else if (SuccHasWeights)
1142  // If there are no predecessor weights but there are successor weights,
1143  // populate Weights with 1, which will later be scaled to the sum of
1144  // successor's weights
1145  Weights.assign(1 + PredCases.size(), 1);
1146 
1147  SmallVector<uint64_t, 8> SuccWeights;
1148  if (SuccHasWeights) {
1149  GetBranchWeights(TI, SuccWeights);
1150  // branch-weight metadata is inconsistent here.
1151  if (SuccWeights.size() != 1 + BBCases.size())
1152  PredHasWeights = SuccHasWeights = false;
1153  } else if (PredHasWeights)
1154  SuccWeights.assign(1 + BBCases.size(), 1);
1155 
1156  if (PredDefault == BB) {
1157  // If this is the default destination from PTI, only the edges in TI
1158  // that don't occur in PTI, or that branch to BB will be activated.
1159  std::set<ConstantInt *, ConstantIntOrdering> PTIHandled;
1160  for (unsigned i = 0, e = PredCases.size(); i != e; ++i)
1161  if (PredCases[i].Dest != BB)
1162  PTIHandled.insert(PredCases[i].Value);
1163  else {
1164  // The default destination is BB, we don't need explicit targets.
1165  std::swap(PredCases[i], PredCases.back());
1166 
1167  if (PredHasWeights || SuccHasWeights) {
1168  // Increase weight for the default case.
1169  Weights[0] += Weights[i + 1];
1170  std::swap(Weights[i + 1], Weights.back());
1171  Weights.pop_back();
1172  }
1173 
1174  PredCases.pop_back();
1175  --i;
1176  --e;
1177  }
1178 
1179  // Reconstruct the new switch statement we will be building.
1180  if (PredDefault != BBDefault) {
1181  PredDefault->removePredecessor(Pred);
1182  if (DTU && PredDefault != BB)
1183  Updates.push_back({DominatorTree::Delete, Pred, PredDefault});
1184  PredDefault = BBDefault;
1185  ++NewSuccessors[BBDefault];
1186  }
1187 
1188  unsigned CasesFromPred = Weights.size();
1189  uint64_t ValidTotalSuccWeight = 0;
1190  for (unsigned i = 0, e = BBCases.size(); i != e; ++i)
1191  if (!PTIHandled.count(BBCases[i].Value) && BBCases[i].Dest != BBDefault) {
1192  PredCases.push_back(BBCases[i]);
1193  ++NewSuccessors[BBCases[i].Dest];
1194  if (SuccHasWeights || PredHasWeights) {
1195  // The default weight is at index 0, so weight for the ith case
1196  // should be at index i+1. Scale the cases from successor by
1197  // PredDefaultWeight (Weights[0]).
1198  Weights.push_back(Weights[0] * SuccWeights[i + 1]);
1199  ValidTotalSuccWeight += SuccWeights[i + 1];
1200  }
1201  }
1202 
1203  if (SuccHasWeights || PredHasWeights) {
1204  ValidTotalSuccWeight += SuccWeights[0];
1205  // Scale the cases from predecessor by ValidTotalSuccWeight.
1206  for (unsigned i = 1; i < CasesFromPred; ++i)
1207  Weights[i] *= ValidTotalSuccWeight;
1208  // Scale the default weight by SuccDefaultWeight (SuccWeights[0]).
1209  Weights[0] *= SuccWeights[0];
1210  }
1211  } else {
1212  // If this is not the default destination from PSI, only the edges
1213  // in SI that occur in PSI with a destination of BB will be
1214  // activated.
1215  std::set<ConstantInt *, ConstantIntOrdering> PTIHandled;
1216  std::map<ConstantInt *, uint64_t> WeightsForHandled;
1217  for (unsigned i = 0, e = PredCases.size(); i != e; ++i)
1218  if (PredCases[i].Dest == BB) {
1219  PTIHandled.insert(PredCases[i].Value);
1220 
1221  if (PredHasWeights || SuccHasWeights) {
1222  WeightsForHandled[PredCases[i].Value] = Weights[i + 1];
1223  std::swap(Weights[i + 1], Weights.back());
1224  Weights.pop_back();
1225  }
1226 
1227  std::swap(PredCases[i], PredCases.back());
1228  PredCases.pop_back();
1229  --i;
1230  --e;
1231  }
1232 
1233  // Okay, now we know which constants were sent to BB from the
1234  // predecessor. Figure out where they will all go now.
1235  for (unsigned i = 0, e = BBCases.size(); i != e; ++i)
1236  if (PTIHandled.count(BBCases[i].Value)) {
1237  // If this is one we are capable of getting...
1238  if (PredHasWeights || SuccHasWeights)
1239  Weights.push_back(WeightsForHandled[BBCases[i].Value]);
1240  PredCases.push_back(BBCases[i]);
1241  ++NewSuccessors[BBCases[i].Dest];
1242  PTIHandled.erase(BBCases[i].Value); // This constant is taken care of
1243  }
1244 
1245  // If there are any constants vectored to BB that TI doesn't handle,
1246  // they must go to the default destination of TI.
1247  for (ConstantInt *I : PTIHandled) {
1248  if (PredHasWeights || SuccHasWeights)
1249  Weights.push_back(WeightsForHandled[I]);
1250  PredCases.push_back(ValueEqualityComparisonCase(I, BBDefault));
1251  ++NewSuccessors[BBDefault];
1252  }
1253  }
1254 
1255  // Okay, at this point, we know which new successor Pred will get. Make
1256  // sure we update the number of entries in the PHI nodes for these
1257  // successors.
1258  SmallPtrSet<BasicBlock *, 2> SuccsOfPred;
1259  if (DTU) {
1260  SuccsOfPred = {succ_begin(Pred), succ_end(Pred)};
1261  Updates.reserve(Updates.size() + NewSuccessors.size());
1262  }
1263  for (const std::pair<BasicBlock *, int /*Num*/> &NewSuccessor :
1264  NewSuccessors) {
1265  for (auto I : seq(0, NewSuccessor.second)) {
1266  (void)I;
1267  AddPredecessorToBlock(NewSuccessor.first, Pred, BB);
1268  }
1269  if (DTU && !SuccsOfPred.contains(NewSuccessor.first))
1270  Updates.push_back({DominatorTree::Insert, Pred, NewSuccessor.first});
1271  }
1272 
1273  Builder.SetInsertPoint(PTI);
1274  // Convert pointer to int before we switch.
1275  if (CV->getType()->isPointerTy()) {
1276  CV =
1277  Builder.CreatePtrToInt(CV, DL.getIntPtrType(CV->getType()), "magicptr");
1278  }
1279 
1280  // Now that the successors are updated, create the new Switch instruction.
1281  SwitchInst *NewSI = Builder.CreateSwitch(CV, PredDefault, PredCases.size());
1282  NewSI->setDebugLoc(PTI->getDebugLoc());
1283  for (ValueEqualityComparisonCase &V : PredCases)
1284  NewSI->addCase(V.Value, V.Dest);
1285 
1286  if (PredHasWeights || SuccHasWeights) {
1287  // Halve the weights if any of them cannot fit in an uint32_t
1288  FitWeights(Weights);
1289 
1290  SmallVector<uint32_t, 8> MDWeights(Weights.begin(), Weights.end());
1291 
1292  setBranchWeights(NewSI, MDWeights);
1293  }
1294 
1296 
1297  // Okay, last check. If BB is still a successor of PSI, then we must
1298  // have an infinite loop case. If so, add an infinitely looping block
1299  // to handle the case to preserve the behavior of the code.
1300  BasicBlock *InfLoopBlock = nullptr;
1301  for (unsigned i = 0, e = NewSI->getNumSuccessors(); i != e; ++i)
1302  if (NewSI->getSuccessor(i) == BB) {
1303  if (!InfLoopBlock) {
1304  // Insert it at the end of the function, because it's either code,
1305  // or it won't matter if it's hot. :)
1306  InfLoopBlock =
1307  BasicBlock::Create(BB->getContext(), "infloop", BB->getParent());
1308  BranchInst::Create(InfLoopBlock, InfLoopBlock);
1309  if (DTU)
1310  Updates.push_back(
1311  {DominatorTree::Insert, InfLoopBlock, InfLoopBlock});
1312  }
1313  NewSI->setSuccessor(i, InfLoopBlock);
1314  }
1315 
1316  if (DTU) {
1317  if (InfLoopBlock)
1318  Updates.push_back({DominatorTree::Insert, Pred, InfLoopBlock});
1319 
1320  Updates.push_back({DominatorTree::Delete, Pred, BB});
1321 
1322  DTU->applyUpdates(Updates);
1323  }
1324 
1325  ++NumFoldValueComparisonIntoPredecessors;
1326  return true;
1327 }
1328 
1329 /// The specified terminator is a value equality comparison instruction
1330 /// (either a switch or a branch on "X == c").
1331 /// See if any of the predecessors of the terminator block are value comparisons
1332 /// on the same value. If so, and if safe to do so, fold them together.
1333 bool SimplifyCFGOpt::FoldValueComparisonIntoPredecessors(Instruction *TI,
1334  IRBuilder<> &Builder) {
1335  BasicBlock *BB = TI->getParent();
1336  Value *CV = isValueEqualityComparison(TI); // CondVal
1337  assert(CV && "Not a comparison?");
1338 
1339  bool Changed = false;
1340 
1342  while (!Preds.empty()) {
1343  BasicBlock *Pred = Preds.pop_back_val();
1344  Instruction *PTI = Pred->getTerminator();
1345 
1346  // Don't try to fold into itself.
1347  if (Pred == BB)
1348  continue;
1349 
1350  // See if the predecessor is a comparison with the same value.
1351  Value *PCV = isValueEqualityComparison(PTI); // PredCondVal
1352  if (PCV != CV)
1353  continue;
1354 
1356  if (!SafeToMergeTerminators(TI, PTI, &FailBlocks)) {
1357  for (auto *Succ : FailBlocks) {
1358  if (!SplitBlockPredecessors(Succ, TI->getParent(), ".fold.split", DTU))
1359  return false;
1360  }
1361  }
1362 
1363  PerformValueComparisonIntoPredecessorFolding(TI, CV, PTI, Builder);
1364  Changed = true;
1365  }
1366  return Changed;
1367 }
1368 
1369 // If we would need to insert a select that uses the value of this invoke
1370 // (comments in HoistThenElseCodeToIf explain why we would need to do this), we
1371 // can't hoist the invoke, as there is nowhere to put the select in this case.
1373  Instruction *I1, Instruction *I2) {
1374  for (BasicBlock *Succ : successors(BB1)) {
1375  for (const PHINode &PN : Succ->phis()) {
1376  Value *BB1V = PN.getIncomingValueForBlock(BB1);
1377  Value *BB2V = PN.getIncomingValueForBlock(BB2);
1378  if (BB1V != BB2V && (BB1V == I1 || BB2V == I2)) {
1379  return false;
1380  }
1381  }
1382  }
1383  return true;
1384 }
1385 
1386 static bool passingValueIsAlwaysUndefined(Value *V, Instruction *I, bool PtrValueMayBeModified = false);
1387 
1388 /// Given a conditional branch that goes to BB1 and BB2, hoist any common code
1389 /// in the two blocks up into the branch block. The caller of this function
1390 /// guarantees that BI's block dominates BB1 and BB2. If EqTermsOnly is given,
1391 /// only perform hoisting in case both blocks only contain a terminator. In that
1392 /// case, only the original BI will be replaced and selects for PHIs are added.
1393 bool SimplifyCFGOpt::HoistThenElseCodeToIf(BranchInst *BI,
1394  const TargetTransformInfo &TTI,
1395  bool EqTermsOnly) {
1396  // This does very trivial matching, with limited scanning, to find identical
1397  // instructions in the two blocks. In particular, we don't want to get into
1398  // O(M*N) situations here where M and N are the sizes of BB1 and BB2. As
1399  // such, we currently just scan for obviously identical instructions in an
1400  // identical order.
1401  BasicBlock *BB1 = BI->getSuccessor(0); // The true destination.
1402  BasicBlock *BB2 = BI->getSuccessor(1); // The false destination
1403 
1404  // If either of the blocks has it's address taken, then we can't do this fold,
1405  // because the code we'd hoist would no longer run when we jump into the block
1406  // by it's address.
1407  if (BB1->hasAddressTaken() || BB2->hasAddressTaken())
1408  return false;
1409 
1410  BasicBlock::iterator BB1_Itr = BB1->begin();
1411  BasicBlock::iterator BB2_Itr = BB2->begin();
1412 
1413  Instruction *I1 = &*BB1_Itr++, *I2 = &*BB2_Itr++;
1414  // Skip debug info if it is not identical.
1415  DbgInfoIntrinsic *DBI1 = dyn_cast<DbgInfoIntrinsic>(I1);
1416  DbgInfoIntrinsic *DBI2 = dyn_cast<DbgInfoIntrinsic>(I2);
1417  if (!DBI1 || !DBI2 || !DBI1->isIdenticalToWhenDefined(DBI2)) {
1418  while (isa<DbgInfoIntrinsic>(I1))
1419  I1 = &*BB1_Itr++;
1420  while (isa<DbgInfoIntrinsic>(I2))
1421  I2 = &*BB2_Itr++;
1422  }
1423  // FIXME: Can we define a safety predicate for CallBr?
1424  if (isa<PHINode>(I1) || !I1->isIdenticalToWhenDefined(I2) ||
1425  (isa<InvokeInst>(I1) && !isSafeToHoistInvoke(BB1, BB2, I1, I2)) ||
1426  isa<CallBrInst>(I1))
1427  return false;
1428 
1429  BasicBlock *BIParent = BI->getParent();
1430 
1431  bool Changed = false;
1432 
1433  auto _ = make_scope_exit([&]() {
1434  if (Changed)
1435  ++NumHoistCommonCode;
1436  });
1437 
1438  // Check if only hoisting terminators is allowed. This does not add new
1439  // instructions to the hoist location.
1440  if (EqTermsOnly) {
1441  // Skip any debug intrinsics, as they are free to hoist.
1442  auto *I1NonDbg = &*skipDebugIntrinsics(I1->getIterator());
1443  auto *I2NonDbg = &*skipDebugIntrinsics(I2->getIterator());
1444  if (!I1NonDbg->isIdenticalToWhenDefined(I2NonDbg))
1445  return false;
1446  if (!I1NonDbg->isTerminator())
1447  return false;
1448  // Now we know that we only need to hoist debug instrinsics and the
1449  // terminator. Let the loop below handle those 2 cases.
1450  }
1451 
1452  do {
1453  // If we are hoisting the terminator instruction, don't move one (making a
1454  // broken BB), instead clone it, and remove BI.
1455  if (I1->isTerminator())
1456  goto HoistTerminator;
1457 
1458  // If we're going to hoist a call, make sure that the two instructions we're
1459  // commoning/hoisting are both marked with musttail, or neither of them is
1460  // marked as such. Otherwise, we might end up in a situation where we hoist
1461  // from a block where the terminator is a `ret` to a block where the terminator
1462  // is a `br`, and `musttail` calls expect to be followed by a return.
1463  auto *C1 = dyn_cast<CallInst>(I1);
1464  auto *C2 = dyn_cast<CallInst>(I2);
1465  if (C1 && C2)
1466  if (C1->isMustTailCall() != C2->isMustTailCall())
1467  return Changed;
1468 
1470  return Changed;
1471 
1472  // If any of the two call sites has nomerge attribute, stop hoisting.
1473  if (const auto *CB1 = dyn_cast<CallBase>(I1))
1474  if (CB1->cannotMerge())
1475  return Changed;
1476  if (const auto *CB2 = dyn_cast<CallBase>(I2))
1477  if (CB2->cannotMerge())
1478  return Changed;
1479 
1480  if (isa<DbgInfoIntrinsic>(I1) || isa<DbgInfoIntrinsic>(I2)) {
1481  assert (isa<DbgInfoIntrinsic>(I1) && isa<DbgInfoIntrinsic>(I2));
1482  // The debug location is an integral part of a debug info intrinsic
1483  // and can't be separated from it or replaced. Instead of attempting
1484  // to merge locations, simply hoist both copies of the intrinsic.
1485  BIParent->getInstList().splice(BI->getIterator(),
1486  BB1->getInstList(), I1);
1487  BIParent->getInstList().splice(BI->getIterator(),
1488  BB2->getInstList(), I2);
1489  Changed = true;
1490  } else {
1491  // For a normal instruction, we just move one to right before the branch,
1492  // then replace all uses of the other with the first. Finally, we remove
1493  // the now redundant second instruction.
1494  BIParent->getInstList().splice(BI->getIterator(),
1495  BB1->getInstList(), I1);
1496  if (!I2->use_empty())
1497  I2->replaceAllUsesWith(I1);
1498  I1->andIRFlags(I2);
1499  unsigned KnownIDs[] = {LLVMContext::MD_tbaa,
1500  LLVMContext::MD_range,
1501  LLVMContext::MD_fpmath,
1502  LLVMContext::MD_invariant_load,
1503  LLVMContext::MD_nonnull,
1504  LLVMContext::MD_invariant_group,
1505  LLVMContext::MD_align,
1506  LLVMContext::MD_dereferenceable,
1507  LLVMContext::MD_dereferenceable_or_null,
1508  LLVMContext::MD_mem_parallel_loop_access,
1509  LLVMContext::MD_access_group,
1510  LLVMContext::MD_preserve_access_index};
1511  combineMetadata(I1, I2, KnownIDs, true);
1512 
1513  // I1 and I2 are being combined into a single instruction. Its debug
1514  // location is the merged locations of the original instructions.
1515  I1->applyMergedLocation(I1->getDebugLoc(), I2->getDebugLoc());
1516 
1517  I2->eraseFromParent();
1518  Changed = true;
1519  }
1520  ++NumHoistCommonInstrs;
1521 
1522  I1 = &*BB1_Itr++;
1523  I2 = &*BB2_Itr++;
1524  // Skip debug info if it is not identical.
1525  DbgInfoIntrinsic *DBI1 = dyn_cast<DbgInfoIntrinsic>(I1);
1526  DbgInfoIntrinsic *DBI2 = dyn_cast<DbgInfoIntrinsic>(I2);
1527  if (!DBI1 || !DBI2 || !DBI1->isIdenticalToWhenDefined(DBI2)) {
1528  while (isa<DbgInfoIntrinsic>(I1))
1529  I1 = &*BB1_Itr++;
1530  while (isa<DbgInfoIntrinsic>(I2))
1531  I2 = &*BB2_Itr++;
1532  }
1533  } while (I1->isIdenticalToWhenDefined(I2));
1534 
1535  return true;
1536 
1537 HoistTerminator:
1538  // It may not be possible to hoist an invoke.
1539  // FIXME: Can we define a safety predicate for CallBr?
1540  if (isa<InvokeInst>(I1) && !isSafeToHoistInvoke(BB1, BB2, I1, I2))
1541  return Changed;
1542 
1543  // TODO: callbr hoisting currently disabled pending further study.
1544  if (isa<CallBrInst>(I1))
1545  return Changed;
1546 
1547  for (BasicBlock *Succ : successors(BB1)) {
1548  for (PHINode &PN : Succ->phis()) {
1549  Value *BB1V = PN.getIncomingValueForBlock(BB1);
1550  Value *BB2V = PN.getIncomingValueForBlock(BB2);
1551  if (BB1V == BB2V)
1552  continue;
1553 
1554  // Check for passingValueIsAlwaysUndefined here because we would rather
1555  // eliminate undefined control flow then converting it to a select.
1556  if (passingValueIsAlwaysUndefined(BB1V, &PN) ||
1557  passingValueIsAlwaysUndefined(BB2V, &PN))
1558  return Changed;
1559 
1560  if (isa<ConstantExpr>(BB1V) && !isSafeToSpeculativelyExecute(BB1V))
1561  return Changed;
1562  if (isa<ConstantExpr>(BB2V) && !isSafeToSpeculativelyExecute(BB2V))
1563  return Changed;
1564  }
1565  }
1566 
1567  // Okay, it is safe to hoist the terminator.
1568  Instruction *NT = I1->clone();
1569  BIParent->getInstList().insert(BI->getIterator(), NT);
1570  if (!NT->getType()->isVoidTy()) {
1571  I1->replaceAllUsesWith(NT);
1572  I2->replaceAllUsesWith(NT);
1573  NT->takeName(I1);
1574  }
1575  Changed = true;
1576  ++NumHoistCommonInstrs;
1577 
1578  // Ensure terminator gets a debug location, even an unknown one, in case
1579  // it involves inlinable calls.
1580  NT->applyMergedLocation(I1->getDebugLoc(), I2->getDebugLoc());
1581 
1582  // PHIs created below will adopt NT's merged DebugLoc.
1584 
1585  // Hoisting one of the terminators from our successor is a great thing.
1586  // Unfortunately, the successors of the if/else blocks may have PHI nodes in
1587  // them. If they do, all PHI entries for BB1/BB2 must agree for all PHI
1588  // nodes, so we insert select instruction to compute the final result.
1589  std::map<std::pair<Value *, Value *>, SelectInst *> InsertedSelects;
1590  for (BasicBlock *Succ : successors(BB1)) {
1591  for (PHINode &PN : Succ->phis()) {
1592  Value *BB1V = PN.getIncomingValueForBlock(BB1);
1593  Value *BB2V = PN.getIncomingValueForBlock(BB2);
1594  if (BB1V == BB2V)
1595  continue;
1596 
1597  // These values do not agree. Insert a select instruction before NT
1598  // that determines the right value.
1599  SelectInst *&SI = InsertedSelects[std::make_pair(BB1V, BB2V)];
1600  if (!SI) {
1601  // Propagate fast-math-flags from phi node to its replacement select.
1603  if (isa<FPMathOperator>(PN))
1604  Builder.setFastMathFlags(PN.getFastMathFlags());
1605 
1606  SI = cast<SelectInst>(
1607  Builder.CreateSelect(BI->getCondition(), BB1V, BB2V,
1608  BB1V->getName() + "." + BB2V->getName(), BI));
1609  }
1610 
1611  // Make the PHI node use the select for all incoming values for BB1/BB2
1612  for (unsigned i = 0, e = PN.getNumIncomingValues(); i != e; ++i)
1613  if (PN.getIncomingBlock(i) == BB1 || PN.getIncomingBlock(i) == BB2)
1614  PN.setIncomingValue(i, SI);
1615  }
1616  }
1617 
1619 
1620  // Update any PHI nodes in our new successors.
1621  for (BasicBlock *Succ : successors(BB1)) {
1622  AddPredecessorToBlock(Succ, BIParent, BB1);
1623  if (DTU)
1624  Updates.push_back({DominatorTree::Insert, BIParent, Succ});
1625  }
1626 
1627  if (DTU)
1628  for (BasicBlock *Succ : successors(BI))
1629  Updates.push_back({DominatorTree::Delete, BIParent, Succ});
1630 
1632  if (DTU)
1633  DTU->applyUpdates(Updates);
1634  return Changed;
1635 }
1636 
1637 // Check lifetime markers.
1638 static bool isLifeTimeMarker(const Instruction *I) {
1639  if (auto II = dyn_cast<IntrinsicInst>(I)) {
1640  switch (II->getIntrinsicID()) {
1641  default:
1642  break;
1643  case Intrinsic::lifetime_start:
1644  case Intrinsic::lifetime_end:
1645  return true;
1646  }
1647  }
1648  return false;
1649 }
1650 
1651 // TODO: Refine this. This should avoid cases like turning constant memcpy sizes
1652 // into variables.
1654  int OpIdx) {
1655  return !isa<IntrinsicInst>(I);
1656 }
1657 
1658 // All instructions in Insts belong to different blocks that all unconditionally
1659 // branch to a common successor. Analyze each instruction and return true if it
1660 // would be possible to sink them into their successor, creating one common
1661 // instruction instead. For every value that would be required to be provided by
1662 // PHI node (because an operand varies in each input block), add to PHIOperands.
1665  DenseMap<Instruction *, SmallVector<Value *, 4>> &PHIOperands) {
1666  // Prune out obviously bad instructions to move. Each instruction must have
1667  // exactly zero or one use, and we check later that use is by a single, common
1668  // PHI instruction in the successor.
1669  bool HasUse = !Insts.front()->user_empty();
1670  for (auto *I : Insts) {
1671  // These instructions may change or break semantics if moved.
1672  if (isa<PHINode>(I) || I->isEHPad() || isa<AllocaInst>(I) ||
1673  I->getType()->isTokenTy())
1674  return false;
1675 
1676  // Do not try to sink an instruction in an infinite loop - it can cause
1677  // this algorithm to infinite loop.
1678  if (I->getParent()->getSingleSuccessor() == I->getParent())
1679  return false;
1680 
1681  // Conservatively return false if I is an inline-asm instruction. Sinking
1682  // and merging inline-asm instructions can potentially create arguments
1683  // that cannot satisfy the inline-asm constraints.
1684  // If the instruction has nomerge attribute, return false.
1685  if (const auto *C = dyn_cast<CallBase>(I))
1686  if (C->isInlineAsm() || C->cannotMerge())
1687  return false;
1688 
1689  // Each instruction must have zero or one use.
1690  if (HasUse && !I->hasOneUse())
1691  return false;
1692  if (!HasUse && !I->user_empty())
1693  return false;
1694  }
1695 
1696  const Instruction *I0 = Insts.front();
1697  for (auto *I : Insts)
1698  if (!I->isSameOperationAs(I0))
1699  return false;
1700 
1701  // All instructions in Insts are known to be the same opcode. If they have a
1702  // use, check that the only user is a PHI or in the same block as the
1703  // instruction, because if a user is in the same block as an instruction we're
1704  // contemplating sinking, it must already be determined to be sinkable.
1705  if (HasUse) {
1706  auto *PNUse = dyn_cast<PHINode>(*I0->user_begin());
1707  auto *Succ = I0->getParent()->getTerminator()->getSuccessor(0);
1708  if (!all_of(Insts, [&PNUse,&Succ](const Instruction *I) -> bool {
1709  auto *U = cast<Instruction>(*I->user_begin());
1710  return (PNUse &&
1711  PNUse->getParent() == Succ &&
1712  PNUse->getIncomingValueForBlock(I->getParent()) == I) ||
1713  U->getParent() == I->getParent();
1714  }))
1715  return false;
1716  }
1717 
1718  // Because SROA can't handle speculating stores of selects, try not to sink
1719  // loads, stores or lifetime markers of allocas when we'd have to create a
1720  // PHI for the address operand. Also, because it is likely that loads or
1721  // stores of allocas will disappear when Mem2Reg/SROA is run, don't sink
1722  // them.
1723  // This can cause code churn which can have unintended consequences down
1724  // the line - see https://llvm.org/bugs/show_bug.cgi?id=30244.
1725  // FIXME: This is a workaround for a deficiency in SROA - see
1726  // https://llvm.org/bugs/show_bug.cgi?id=30188
1727  if (isa<StoreInst>(I0) && any_of(Insts, [](const Instruction *I) {
1728  return isa<AllocaInst>(I->getOperand(1)->stripPointerCasts());
1729  }))
1730  return false;
1731  if (isa<LoadInst>(I0) && any_of(Insts, [](const Instruction *I) {
1732  return isa<AllocaInst>(I->getOperand(0)->stripPointerCasts());
1733  }))
1734  return false;
1735  if (isLifeTimeMarker(I0) && any_of(Insts, [](const Instruction *I) {
1736  return isa<AllocaInst>(I->getOperand(1)->stripPointerCasts());
1737  }))
1738  return false;
1739 
1740  // For calls to be sinkable, they must all be indirect, or have same callee.
1741  // I.e. if we have two direct calls to different callees, we don't want to
1742  // turn that into an indirect call. Likewise, if we have an indirect call,
1743  // and a direct call, we don't actually want to have a single indirect call.
1744  if (isa<CallBase>(I0)) {
1745  auto IsIndirectCall = [](const Instruction *I) {
1746  return cast<CallBase>(I)->isIndirectCall();
1747  };
1748  bool HaveIndirectCalls = any_of(Insts, IsIndirectCall);
1749  bool AllCallsAreIndirect = all_of(Insts, IsIndirectCall);
1750  if (HaveIndirectCalls) {
1751  if (!AllCallsAreIndirect)
1752  return false;
1753  } else {
1754  // All callees must be identical.
1755  Value *Callee = nullptr;
1756  for (const Instruction *I : Insts) {
1757  Value *CurrCallee = cast<CallBase>(I)->getCalledOperand();
1758  if (!Callee)
1759  Callee = CurrCallee;
1760  else if (Callee != CurrCallee)
1761  return false;
1762  }
1763  }
1764  }
1765 
1766  for (unsigned OI = 0, OE = I0->getNumOperands(); OI != OE; ++OI) {
1767  Value *Op = I0->getOperand(OI);
1768  if (Op->getType()->isTokenTy())
1769  // Don't touch any operand of token type.
1770  return false;
1771 
1772  auto SameAsI0 = [&I0, OI](const Instruction *I) {
1773  assert(I->getNumOperands() == I0->getNumOperands());
1774  return I->getOperand(OI) == I0->getOperand(OI);
1775  };
1776  if (!all_of(Insts, SameAsI0)) {
1777  if ((isa<Constant>(Op) && !replacingOperandWithVariableIsCheap(I0, OI)) ||
1779  // We can't create a PHI from this GEP.
1780  return false;
1781  for (auto *I : Insts)
1782  PHIOperands[I].push_back(I->getOperand(OI));
1783  }
1784  }
1785  return true;
1786 }
1787 
1788 // Assuming canSinkInstructions(Blocks) has returned true, sink the last
1789 // instruction of every block in Blocks to their common successor, commoning
1790 // into one instruction.
1792  auto *BBEnd = Blocks[0]->getTerminator()->getSuccessor(0);
1793 
1794  // canSinkInstructions returning true guarantees that every block has at
1795  // least one non-terminator instruction.
1797  for (auto *BB : Blocks) {
1798  Instruction *I = BB->getTerminator();
1799  do {
1800  I = I->getPrevNode();
1801  } while (isa<DbgInfoIntrinsic>(I) && I != &BB->front());
1802  if (!isa<DbgInfoIntrinsic>(I))
1803  Insts.push_back(I);
1804  }
1805 
1806  // The only checking we need to do now is that all users of all instructions
1807  // are the same PHI node. canSinkInstructions should have checked this but
1808  // it is slightly over-aggressive - it gets confused by commutative
1809  // instructions so double-check it here.
1810  Instruction *I0 = Insts.front();
1811  if (!I0->user_empty()) {
1812  auto *PNUse = dyn_cast<PHINode>(*I0->user_begin());
1813  if (!all_of(Insts, [&PNUse](const Instruction *I) -> bool {
1814  auto *U = cast<Instruction>(*I->user_begin());
1815  return U == PNUse;
1816  }))
1817  return false;
1818  }
1819 
1820  // We don't need to do any more checking here; canSinkInstructions should
1821  // have done it all for us.
1822  SmallVector<Value*, 4> NewOperands;
1823  for (unsigned O = 0, E = I0->getNumOperands(); O != E; ++O) {
1824  // This check is different to that in canSinkInstructions. There, we
1825  // cared about the global view once simplifycfg (and instcombine) have
1826  // completed - it takes into account PHIs that become trivially
1827  // simplifiable. However here we need a more local view; if an operand
1828  // differs we create a PHI and rely on instcombine to clean up the very
1829  // small mess we may make.
1830  bool NeedPHI = any_of(Insts, [&I0, O](const Instruction *I) {
1831  return I->getOperand(O) != I0->getOperand(O);
1832  });
1833  if (!NeedPHI) {
1834  NewOperands.push_back(I0->getOperand(O));
1835  continue;
1836  }
1837 
1838  // Create a new PHI in the successor block and populate it.
1839  auto *Op = I0->getOperand(O);
1840  assert(!Op->getType()->isTokenTy() && "Can't PHI tokens!");
1841  auto *PN = PHINode::Create(Op->getType(), Insts.size(),
1842  Op->getName() + ".sink", &BBEnd->front());
1843  for (auto *I : Insts)
1844  PN->addIncoming(I->getOperand(O), I->getParent());
1845  NewOperands.push_back(PN);
1846  }
1847 
1848  // Arbitrarily use I0 as the new "common" instruction; remap its operands
1849  // and move it to the start of the successor block.
1850  for (unsigned O = 0, E = I0->getNumOperands(); O != E; ++O)
1851  I0->getOperandUse(O).set(NewOperands[O]);
1852  I0->moveBefore(&*BBEnd->getFirstInsertionPt());
1853 
1854  // Update metadata and IR flags, and merge debug locations.
1855  for (auto *I : Insts)
1856  if (I != I0) {
1857  // The debug location for the "common" instruction is the merged locations
1858  // of all the commoned instructions. We start with the original location
1859  // of the "common" instruction and iteratively merge each location in the
1860  // loop below.
1861  // This is an N-way merge, which will be inefficient if I0 is a CallInst.
1862  // However, as N-way merge for CallInst is rare, so we use simplified API
1863  // instead of using complex API for N-way merge.
1864  I0->applyMergedLocation(I0->getDebugLoc(), I->getDebugLoc());
1865  combineMetadataForCSE(I0, I, true);
1866  I0->andIRFlags(I);
1867  }
1868 
1869  if (!I0->user_empty()) {
1870  // canSinkLastInstruction checked that all instructions were used by
1871  // one and only one PHI node. Find that now, RAUW it to our common
1872  // instruction and nuke it.
1873  auto *PN = cast<PHINode>(*I0->user_begin());
1874  PN->replaceAllUsesWith(I0);
1875  PN->eraseFromParent();
1876  }
1877 
1878  // Finally nuke all instructions apart from the common instruction.
1879  for (auto *I : Insts)
1880  if (I != I0)
1881  I->eraseFromParent();
1882 
1883  return true;
1884 }
1885 
1886 namespace {
1887 
1888  // LockstepReverseIterator - Iterates through instructions
1889  // in a set of blocks in reverse order from the first non-terminator.
1890  // For example (assume all blocks have size n):
1891  // LockstepReverseIterator I([B1, B2, B3]);
1892  // *I-- = [B1[n], B2[n], B3[n]];
1893  // *I-- = [B1[n-1], B2[n-1], B3[n-1]];
1894  // *I-- = [B1[n-2], B2[n-2], B3[n-2]];
1895  // ...
1896  class LockstepReverseIterator {
1897  ArrayRef<BasicBlock*> Blocks;
1899  bool Fail;
1900 
1901  public:
1902  LockstepReverseIterator(ArrayRef<BasicBlock*> Blocks) : Blocks(Blocks) {
1903  reset();
1904  }
1905 
1906  void reset() {
1907  Fail = false;
1908  Insts.clear();
1909  for (auto *BB : Blocks) {
1910  Instruction *Inst = BB->getTerminator();
1911  for (Inst = Inst->getPrevNode(); Inst && isa<DbgInfoIntrinsic>(Inst);)
1912  Inst = Inst->getPrevNode();
1913  if (!Inst) {
1914  // Block wasn't big enough.
1915  Fail = true;
1916  return;
1917  }
1918  Insts.push_back(Inst);
1919  }
1920  }
1921 
1922  bool isValid() const {
1923  return !Fail;
1924  }
1925 
1926  void operator--() {
1927  if (Fail)
1928  return;
1929  for (auto *&Inst : Insts) {
1930  for (Inst = Inst->getPrevNode(); Inst && isa<DbgInfoIntrinsic>(Inst);)
1931  Inst = Inst->getPrevNode();
1932  // Already at beginning of block.
1933  if (!Inst) {
1934  Fail = true;
1935  return;
1936  }
1937  }
1938  }
1939 
1940  void operator++() {
1941  if (Fail)
1942  return;
1943  for (auto *&Inst : Insts) {
1944  for (Inst = Inst->getNextNode(); Inst && isa<DbgInfoIntrinsic>(Inst);)
1945  Inst = Inst->getNextNode();
1946  // Already at end of block.
1947  if (!Inst) {
1948  Fail = true;
1949  return;
1950  }
1951  }
1952  }
1953 
1955  return Insts;
1956  }
1957  };
1958 
1959 } // end anonymous namespace
1960 
1961 /// Check whether BB's predecessors end with unconditional branches. If it is
1962 /// true, sink any common code from the predecessors to BB.
1964  DomTreeUpdater *DTU) {
1965  // We support two situations:
1966  // (1) all incoming arcs are unconditional
1967  // (2) there are non-unconditional incoming arcs
1968  //
1969  // (2) is very common in switch defaults and
1970  // else-if patterns;
1971  //
1972  // if (a) f(1);
1973  // else if (b) f(2);
1974  //
1975  // produces:
1976  //
1977  // [if]
1978  // / \
1979  // [f(1)] [if]
1980  // | | \
1981  // | | |
1982  // | [f(2)]|
1983  // \ | /
1984  // [ end ]
1985  //
1986  // [end] has two unconditional predecessor arcs and one conditional. The
1987  // conditional refers to the implicit empty 'else' arc. This conditional
1988  // arc can also be caused by an empty default block in a switch.
1989  //
1990  // In this case, we attempt to sink code from all *unconditional* arcs.
1991  // If we can sink instructions from these arcs (determined during the scan
1992  // phase below) we insert a common successor for all unconditional arcs and
1993  // connect that to [end], to enable sinking:
1994  //
1995  // [if]
1996  // / \
1997  // [x(1)] [if]
1998  // | | \
1999  // | | \
2000  // | [x(2)] |
2001  // \ / |
2002  // [sink.split] |
2003  // \ /
2004  // [ end ]
2005  //
2006  SmallVector<BasicBlock*,4> UnconditionalPreds;
2007  bool HaveNonUnconditionalPredecessors = false;
2008  for (auto *PredBB : predecessors(BB)) {
2009  auto *PredBr = dyn_cast<BranchInst>(PredBB->getTerminator());
2010  if (PredBr && PredBr->isUnconditional())
2011  UnconditionalPreds.push_back(PredBB);
2012  else
2013  HaveNonUnconditionalPredecessors = true;
2014  }
2015  if (UnconditionalPreds.size() < 2)
2016  return false;
2017 
2018  // We take a two-step approach to tail sinking. First we scan from the end of
2019  // each block upwards in lockstep. If the n'th instruction from the end of each
2020  // block can be sunk, those instructions are added to ValuesToSink and we
2021  // carry on. If we can sink an instruction but need to PHI-merge some operands
2022  // (because they're not identical in each instruction) we add these to
2023  // PHIOperands.
2024  int ScanIdx = 0;
2025  SmallPtrSet<Value*,4> InstructionsToSink;
2027  LockstepReverseIterator LRI(UnconditionalPreds);
2028  while (LRI.isValid() &&
2029  canSinkInstructions(*LRI, PHIOperands)) {
2030  LLVM_DEBUG(dbgs() << "SINK: instruction can be sunk: " << *(*LRI)[0]
2031  << "\n");
2032  InstructionsToSink.insert((*LRI).begin(), (*LRI).end());
2033  ++ScanIdx;
2034  --LRI;
2035  }
2036 
2037  // If no instructions can be sunk, early-return.
2038  if (ScanIdx == 0)
2039  return false;
2040 
2041  // Okay, we *could* sink last ScanIdx instructions. But how many can we
2042  // actually sink before encountering instruction that is unprofitable to sink?
2043  auto ProfitableToSinkInstruction = [&](LockstepReverseIterator &LRI) {
2044  unsigned NumPHIdValues = 0;
2045  for (auto *I : *LRI)
2046  for (auto *V : PHIOperands[I]) {
2047  if (InstructionsToSink.count(V) == 0)
2048  ++NumPHIdValues;
2049  // FIXME: this check is overly optimistic. We may end up not sinking
2050  // said instruction, due to the very same profitability check.
2051  // See @creating_too_many_phis in sink-common-code.ll.
2052  }
2053  LLVM_DEBUG(dbgs() << "SINK: #phid values: " << NumPHIdValues << "\n");
2054  unsigned NumPHIInsts = NumPHIdValues / UnconditionalPreds.size();
2055  if ((NumPHIdValues % UnconditionalPreds.size()) != 0)
2056  NumPHIInsts++;
2057 
2058  return NumPHIInsts <= 1;
2059  };
2060 
2061  // We've determined that we are going to sink last ScanIdx instructions,
2062  // and recorded them in InstructionsToSink. Now, some instructions may be
2063  // unprofitable to sink. But that determination depends on the instructions
2064  // that we are going to sink.
2065 
2066  // First, forward scan: find the first instruction unprofitable to sink,
2067  // recording all the ones that are profitable to sink.
2068  // FIXME: would it be better, after we detect that not all are profitable.
2069  // to either record the profitable ones, or erase the unprofitable ones?
2070  // Maybe we need to choose (at runtime) the one that will touch least instrs?
2071  LRI.reset();
2072  int Idx = 0;
2073  SmallPtrSet<Value *, 4> InstructionsProfitableToSink;
2074  while (Idx < ScanIdx) {
2075  if (!ProfitableToSinkInstruction(LRI)) {
2076  // Too many PHIs would be created.
2077  LLVM_DEBUG(
2078  dbgs() << "SINK: stopping here, too many PHIs would be created!\n");
2079  break;
2080  }
2081  InstructionsProfitableToSink.insert((*LRI).begin(), (*LRI).end());
2082  --LRI;
2083  ++Idx;
2084  }
2085 
2086  // If no instructions can be sunk, early-return.
2087  if (Idx == 0)
2088  return false;
2089 
2090  // Did we determine that (only) some instructions are unprofitable to sink?
2091  if (Idx < ScanIdx) {
2092  // Okay, some instructions are unprofitable.
2093  ScanIdx = Idx;
2094  InstructionsToSink = InstructionsProfitableToSink;
2095 
2096  // But, that may make other instructions unprofitable, too.
2097  // So, do a backward scan, do any earlier instructions become unprofitable?
2098  assert(!ProfitableToSinkInstruction(LRI) &&
2099  "We already know that the last instruction is unprofitable to sink");
2100  ++LRI;
2101  --Idx;
2102  while (Idx >= 0) {
2103  // If we detect that an instruction becomes unprofitable to sink,
2104  // all earlier instructions won't be sunk either,
2105  // so preemptively keep InstructionsProfitableToSink in sync.
2106  // FIXME: is this the most performant approach?
2107  for (auto *I : *LRI)
2108  InstructionsProfitableToSink.erase(I);
2109  if (!ProfitableToSinkInstruction(LRI)) {
2110  // Everything starting with this instruction won't be sunk.
2111  ScanIdx = Idx;
2112  InstructionsToSink = InstructionsProfitableToSink;
2113  }
2114  ++LRI;
2115  --Idx;
2116  }
2117  }
2118 
2119  // If no instructions can be sunk, early-return.
2120  if (ScanIdx == 0)
2121  return false;
2122 
2123  bool Changed = false;
2124 
2125  if (HaveNonUnconditionalPredecessors) {
2126  // It is always legal to sink common instructions from unconditional
2127  // predecessors. However, if not all predecessors are unconditional,
2128  // this transformation might be pessimizing. So as a rule of thumb,
2129  // don't do it unless we'd sink at least one non-speculatable instruction.
2130  // See https://bugs.llvm.org/show_bug.cgi?id=30244
2131  LRI.reset();
2132  int Idx = 0;
2133  bool Profitable = false;
2134  while (Idx < ScanIdx) {
2135  if (!isSafeToSpeculativelyExecute((*LRI)[0])) {
2136  Profitable = true;
2137  break;
2138  }
2139  --LRI;
2140  ++Idx;
2141  }
2142  if (!Profitable)
2143  return false;
2144 
2145  LLVM_DEBUG(dbgs() << "SINK: Splitting edge\n");
2146  // We have a conditional edge and we're going to sink some instructions.
2147  // Insert a new block postdominating all blocks we're going to sink from.
2148  if (!SplitBlockPredecessors(BB, UnconditionalPreds, ".sink.split", DTU))
2149  // Edges couldn't be split.
2150  return false;
2151  Changed = true;
2152  }
2153 
2154  // Now that we've analyzed all potential sinking candidates, perform the
2155  // actual sink. We iteratively sink the last non-terminator of the source
2156  // blocks into their common successor unless doing so would require too
2157  // many PHI instructions to be generated (currently only one PHI is allowed
2158  // per sunk instruction).
2159  //
2160  // We can use InstructionsToSink to discount values needing PHI-merging that will
2161  // actually be sunk in a later iteration. This allows us to be more
2162  // aggressive in what we sink. This does allow a false positive where we
2163  // sink presuming a later value will also be sunk, but stop half way through
2164  // and never actually sink it which means we produce more PHIs than intended.
2165  // This is unlikely in practice though.
2166  int SinkIdx = 0;
2167  for (; SinkIdx != ScanIdx; ++SinkIdx) {
2168  LLVM_DEBUG(dbgs() << "SINK: Sink: "
2169  << *UnconditionalPreds[0]->getTerminator()->getPrevNode()
2170  << "\n");
2171 
2172  // Because we've sunk every instruction in turn, the current instruction to
2173  // sink is always at index 0.
2174  LRI.reset();
2175 
2176  if (!sinkLastInstruction(UnconditionalPreds)) {
2177  LLVM_DEBUG(
2178  dbgs()
2179  << "SINK: stopping here, failed to actually sink instruction!\n");
2180  break;
2181  }
2182 
2183  NumSinkCommonInstrs++;
2184  Changed = true;
2185  }
2186  if (SinkIdx != 0)
2187  ++NumSinkCommonCode;
2188  return Changed;
2189 }
2190 
2191 /// Determine if we can hoist sink a sole store instruction out of a
2192 /// conditional block.
2193 ///
2194 /// We are looking for code like the following:
2195 /// BrBB:
2196 /// store i32 %add, i32* %arrayidx2
2197 /// ... // No other stores or function calls (we could be calling a memory
2198 /// ... // function).
2199 /// %cmp = icmp ult %x, %y
2200 /// br i1 %cmp, label %EndBB, label %ThenBB
2201 /// ThenBB:
2202 /// store i32 %add5, i32* %arrayidx2
2203 /// br label EndBB
2204 /// EndBB:
2205 /// ...
2206 /// We are going to transform this into:
2207 /// BrBB:
2208 /// store i32 %add, i32* %arrayidx2
2209 /// ... //
2210 /// %cmp = icmp ult %x, %y
2211 /// %add.add5 = select i1 %cmp, i32 %add, %add5
2212 /// store i32 %add.add5, i32* %arrayidx2
2213 /// ...
2214 ///
2215 /// \return The pointer to the value of the previous store if the store can be
2216 /// hoisted into the predecessor block. 0 otherwise.
2218  BasicBlock *StoreBB, BasicBlock *EndBB) {
2219  StoreInst *StoreToHoist = dyn_cast<StoreInst>(I);
2220  if (!StoreToHoist)
2221  return nullptr;
2222 
2223  // Volatile or atomic.
2224  if (!StoreToHoist->isSimple())
2225  return nullptr;
2226 
2227  Value *StorePtr = StoreToHoist->getPointerOperand();
2228  Type *StoreTy = StoreToHoist->getValueOperand()->getType();
2229 
2230  // Look for a store to the same pointer in BrBB.
2231  unsigned MaxNumInstToLookAt = 9;
2232  // Skip pseudo probe intrinsic calls which are not really killing any memory
2233  // accesses.
2234  for (Instruction &CurI : reverse(BrBB->instructionsWithoutDebug(true))) {
2235  if (!MaxNumInstToLookAt)
2236  break;
2237  --MaxNumInstToLookAt;
2238 
2239  // Could be calling an instruction that affects memory like free().
2240  if (CurI.mayWriteToMemory() && !isa<StoreInst>(CurI))
2241  return nullptr;
2242 
2243  if (auto *SI = dyn_cast<StoreInst>(&CurI)) {
2244  // Found the previous store to same location and type. Make sure it is
2245  // simple, to avoid introducing a spurious non-atomic write after an
2246  // atomic write.
2247  if (SI->getPointerOperand() == StorePtr &&
2248  SI->getValueOperand()->getType() == StoreTy && SI->isSimple())
2249  // Found the previous store, return its value operand.
2250  return SI->getValueOperand();
2251  return nullptr; // Unknown store.
2252  }
2253  }
2254 
2255  return nullptr;
2256 }
2257 
2258 /// Estimate the cost of the insertion(s) and check that the PHI nodes can be
2259 /// converted to selects.
2261  BasicBlock *EndBB,
2262  unsigned &SpeculatedInstructions,
2263  InstructionCost &Cost,
2264  const TargetTransformInfo &TTI) {
2266  BB->getParent()->hasMinSize()
2269 
2270  bool HaveRewritablePHIs = false;
2271  for (PHINode &PN : EndBB->phis()) {
2272  Value *OrigV = PN.getIncomingValueForBlock(BB);
2273  Value *ThenV = PN.getIncomingValueForBlock(ThenBB);
2274 
2275  // FIXME: Try to remove some of the duplication with HoistThenElseCodeToIf.
2276  // Skip PHIs which are trivial.
2277  if (ThenV == OrigV)
2278  continue;
2279 
2280  Cost += TTI.getCmpSelInstrCost(Instruction::Select, PN.getType(), nullptr,
2282 
2283  // Don't convert to selects if we could remove undefined behavior instead.
2284  if (passingValueIsAlwaysUndefined(OrigV, &PN) ||
2285  passingValueIsAlwaysUndefined(ThenV, &PN))
2286  return false;
2287 
2288  HaveRewritablePHIs = true;
2289  ConstantExpr *OrigCE = dyn_cast<ConstantExpr>(OrigV);
2290  ConstantExpr *ThenCE = dyn_cast<ConstantExpr>(ThenV);
2291  if (!OrigCE && !ThenCE)
2292  continue; // Known safe and cheap.
2293 
2294  if ((ThenCE && !isSafeToSpeculativelyExecute(ThenCE)) ||
2295  (OrigCE && !isSafeToSpeculativelyExecute(OrigCE)))
2296  return false;
2297  InstructionCost OrigCost = OrigCE ? computeSpeculationCost(OrigCE, TTI) : 0;
2298  InstructionCost ThenCost = ThenCE ? computeSpeculationCost(ThenCE, TTI) : 0;
2299  InstructionCost MaxCost =
2301  if (OrigCost + ThenCost > MaxCost)
2302  return false;
2303 
2304  // Account for the cost of an unfolded ConstantExpr which could end up
2305  // getting expanded into Instructions.
2306  // FIXME: This doesn't account for how many operations are combined in the
2307  // constant expression.
2308  ++SpeculatedInstructions;
2309  if (SpeculatedInstructions > 1)
2310  return false;
2311  }
2312 
2313  return HaveRewritablePHIs;
2314 }
2315 
2316 /// Speculate a conditional basic block flattening the CFG.
2317 ///
2318 /// Note that this is a very risky transform currently. Speculating
2319 /// instructions like this is most often not desirable. Instead, there is an MI
2320 /// pass which can do it with full awareness of the resource constraints.
2321 /// However, some cases are "obvious" and we should do directly. An example of
2322 /// this is speculating a single, reasonably cheap instruction.
2323 ///
2324 /// There is only one distinct advantage to flattening the CFG at the IR level:
2325 /// it makes very common but simplistic optimizations such as are common in
2326 /// instcombine and the DAG combiner more powerful by removing CFG edges and
2327 /// modeling their effects with easier to reason about SSA value graphs.
2328 ///
2329 ///
2330 /// An illustration of this transform is turning this IR:
2331 /// \code
2332 /// BB:
2333 /// %cmp = icmp ult %x, %y
2334 /// br i1 %cmp, label %EndBB, label %ThenBB
2335 /// ThenBB:
2336 /// %sub = sub %x, %y
2337 /// br label BB2
2338 /// EndBB:
2339 /// %phi = phi [ %sub, %ThenBB ], [ 0, %EndBB ]
2340 /// ...
2341 /// \endcode
2342 ///
2343 /// Into this IR:
2344 /// \code
2345 /// BB:
2346 /// %cmp = icmp ult %x, %y
2347 /// %sub = sub %x, %y
2348 /// %cond = select i1 %cmp, 0, %sub
2349 /// ...
2350 /// \endcode
2351 ///
2352 /// \returns true if the conditional block is removed.
2353 bool SimplifyCFGOpt::SpeculativelyExecuteBB(BranchInst *BI, BasicBlock *ThenBB,
2354  const TargetTransformInfo &TTI) {
2355  // Be conservative for now. FP select instruction can often be expensive.
2356  Value *BrCond = BI->getCondition();
2357  if (isa<FCmpInst>(BrCond))
2358  return false;
2359 
2360  BasicBlock *BB = BI->getParent();
2361  BasicBlock *EndBB = ThenBB->getTerminator()->getSuccessor(0);
2362  InstructionCost Budget =
2364 
2365  // If ThenBB is actually on the false edge of the conditional branch, remember
2366  // to swap the select operands later.
2367  bool Invert = false;
2368  if (ThenBB != BI->getSuccessor(0)) {
2369  assert(ThenBB == BI->getSuccessor(1) && "No edge from 'if' block?");
2370  Invert = true;
2371  }
2372  assert(EndBB == BI->getSuccessor(!Invert) && "No edge from to end block");
2373 
2374  // If the branch is non-unpredictable, and is predicted to *not* branch to
2375  // the `then` block, then avoid speculating it.
2376  if (!BI->getMetadata(LLVMContext::MD_unpredictable)) {
2377  uint64_t TWeight, FWeight;
2378  if (BI->extractProfMetadata(TWeight, FWeight) && (TWeight + FWeight) != 0) {
2379  uint64_t EndWeight = Invert ? TWeight : FWeight;
2380  BranchProbability BIEndProb =
2381  BranchProbability::getBranchProbability(EndWeight, TWeight + FWeight);
2383  if (BIEndProb >= Likely)
2384  return false;
2385  }
2386  }
2387 
2388  // Keep a count of how many times instructions are used within ThenBB when
2389  // they are candidates for sinking into ThenBB. Specifically:
2390  // - They are defined in BB, and
2391  // - They have no side effects, and
2392  // - All of their uses are in ThenBB.
2393  SmallDenseMap<Instruction *, unsigned, 4> SinkCandidateUseCounts;
2394 
2395  SmallVector<Instruction *, 4> SpeculatedDbgIntrinsics;
2396 
2397  unsigned SpeculatedInstructions = 0;
2398  Value *SpeculatedStoreValue = nullptr;
2399  StoreInst *SpeculatedStore = nullptr;
2400  for (BasicBlock::iterator BBI = ThenBB->begin(),
2401  BBE = std::prev(ThenBB->end());
2402  BBI != BBE; ++BBI) {
2403  Instruction *I = &*BBI;
2404  // Skip debug info.
2405  if (isa<DbgInfoIntrinsic>(I)) {
2406  SpeculatedDbgIntrinsics.push_back(I);
2407  continue;
2408  }
2409 
2410  // Skip pseudo probes. The consequence is we lose track of the branch
2411  // probability for ThenBB, which is fine since the optimization here takes
2412  // place regardless of the branch probability.
2413  if (isa<PseudoProbeInst>(I)) {
2414  // The probe should be deleted so that it will not be over-counted when
2415  // the samples collected on the non-conditional path are counted towards
2416  // the conditional path. We leave it for the counts inference algorithm to
2417  // figure out a proper count for an unknown probe.
2418  SpeculatedDbgIntrinsics.push_back(I);
2419  continue;
2420  }
2421 
2422  // Only speculatively execute a single instruction (not counting the
2423  // terminator) for now.
2424  ++SpeculatedInstructions;
2425  if (SpeculatedInstructions > 1)
2426  return false;
2427 
2428  // Don't hoist the instruction if it's unsafe or expensive.
2430  !(HoistCondStores && (SpeculatedStoreValue = isSafeToSpeculateStore(
2431  I, BB, ThenBB, EndBB))))
2432  return false;
2433  if (!SpeculatedStoreValue &&
2436  return false;
2437 
2438  // Store the store speculation candidate.
2439  if (SpeculatedStoreValue)
2440  SpeculatedStore = cast<StoreInst>(I);
2441 
2442  // Do not hoist the instruction if any of its operands are defined but not
2443  // used in BB. The transformation will prevent the operand from
2444  // being sunk into the use block.
2445  for (Use &Op : I->operands()) {
2446  Instruction *OpI = dyn_cast<Instruction>(Op);
2447  if (!OpI || OpI->getParent() != BB || OpI->mayHaveSideEffects())
2448  continue; // Not a candidate for sinking.
2449 
2450  ++SinkCandidateUseCounts[OpI];
2451  }
2452  }
2453 
2454  // Consider any sink candidates which are only used in ThenBB as costs for
2455  // speculation. Note, while we iterate over a DenseMap here, we are summing
2456  // and so iteration order isn't significant.
2458  I = SinkCandidateUseCounts.begin(),
2459  E = SinkCandidateUseCounts.end();
2460  I != E; ++I)
2461  if (I->first->hasNUses(I->second)) {
2462  ++SpeculatedInstructions;
2463  if (SpeculatedInstructions > 1)
2464  return false;
2465  }
2466 
2467  // Check that we can insert the selects and that it's not too expensive to do
2468  // so.
2469  bool Convert = SpeculatedStore != nullptr;
2470  InstructionCost Cost = 0;
2471  Convert |= validateAndCostRequiredSelects(BB, ThenBB, EndBB,
2472  SpeculatedInstructions,
2473  Cost, TTI);
2474  if (!Convert || Cost > Budget)
2475  return false;
2476 
2477  // If we get here, we can hoist the instruction and if-convert.
2478  LLVM_DEBUG(dbgs() << "SPECULATIVELY EXECUTING BB" << *ThenBB << "\n";);
2479 
2480  // Insert a select of the value of the speculated store.
2481  if (SpeculatedStoreValue) {
2483  Value *TrueV = SpeculatedStore->getValueOperand();
2484  Value *FalseV = SpeculatedStoreValue;
2485  if (Invert)
2486  std::swap(TrueV, FalseV);
2487  Value *S = Builder.CreateSelect(
2488  BrCond, TrueV, FalseV, "spec.store.select", BI);
2489  SpeculatedStore->setOperand(0, S);
2490  SpeculatedStore->applyMergedLocation(BI->getDebugLoc(),
2491  SpeculatedStore->getDebugLoc());
2492  }
2493 
2494  // Metadata can be dependent on the condition we are hoisting above.
2495  // Conservatively strip all metadata on the instruction. Drop the debug loc
2496  // to avoid making it appear as if the condition is a constant, which would
2497  // be misleading while debugging.
2498  // Similarly strip attributes that maybe dependent on condition we are
2499  // hoisting above.
2500  for (auto &I : *ThenBB) {
2501  if (!SpeculatedStoreValue || &I != SpeculatedStore)
2502  I.setDebugLoc(DebugLoc());
2503  I.dropUndefImplyingAttrsAndUnknownMetadata();
2504  }
2505 
2506  // Hoist the instructions.
2507  BB->getInstList().splice(BI->getIterator(), ThenBB->getInstList(),
2508  ThenBB->begin(), std::prev(ThenBB->end()));
2509 
2510  // Insert selects and rewrite the PHI operands.
2512  for (PHINode &PN : EndBB->phis()) {
2513  unsigned OrigI = PN.getBasicBlockIndex(BB);
2514  unsigned ThenI = PN.getBasicBlockIndex(ThenBB);
2515  Value *OrigV = PN.getIncomingValue(OrigI);
2516  Value *ThenV = PN.getIncomingValue(ThenI);
2517 
2518  // Skip PHIs which are trivial.
2519  if (OrigV == ThenV)
2520  continue;
2521 
2522  // Create a select whose true value is the speculatively executed value and
2523  // false value is the pre-existing value. Swap them if the branch
2524  // destinations were inverted.
2525  Value *TrueV = ThenV, *FalseV = OrigV;
2526  if (Invert)
2527  std::swap(TrueV, FalseV);
2528  Value *V = Builder.CreateSelect(BrCond, TrueV, FalseV, "spec.select", BI);
2529  PN.setIncomingValue(OrigI, V);
2530  PN.setIncomingValue(ThenI, V);
2531  }
2532 
2533  // Remove speculated dbg intrinsics.
2534  // FIXME: Is it possible to do this in a more elegant way? Moving/merging the
2535  // dbg value for the different flows and inserting it after the select.
2536  for (Instruction *I : SpeculatedDbgIntrinsics)
2537  I->eraseFromParent();
2538 
2539  ++NumSpeculations;
2540  return true;
2541 }
2542 
2543 /// Return true if we can thread a branch across this block.
2545  int Size = 0;
2546 
2548  auto IsEphemeral = [&](const Value *V) {
2549  if (isa<AssumeInst>(V))
2550  return true;
2551  return isSafeToSpeculativelyExecute(V) &&
2552  all_of(V->users(),
2553  [&](const User *U) { return EphValues.count(U); });
2554  };
2555 
2556  // Walk the loop in reverse so that we can identify ephemeral values properly
2557  // (values only feeding assumes).
2558  for (Instruction &I : reverse(BB->instructionsWithoutDebug())) {
2559  // Can't fold blocks that contain noduplicate or convergent calls.
2560  if (CallInst *CI = dyn_cast<CallInst>(&I))
2561  if (CI->cannotDuplicate() || CI->isConvergent())
2562  return false;
2563 
2564  // Ignore ephemeral values which are deleted during codegen.
2565  if (IsEphemeral(&I))
2566  EphValues.insert(&I);
2567  // We will delete Phis while threading, so Phis should not be accounted in
2568  // block's size.
2569  else if (!isa<PHINode>(I)) {
2570  if (Size++ > MaxSmallBlockSize)
2571  return false; // Don't clone large BB's.
2572  }
2573 
2574  // We can only support instructions that do not define values that are
2575  // live outside of the current basic block.
2576  for (User *U : I.users()) {
2577  Instruction *UI = cast<Instruction>(U);
2578  if (UI->getParent() != BB || isa<PHINode>(UI))
2579  return false;
2580  }
2581 
2582  // Looks ok, continue checking.
2583  }
2584 
2585  return true;
2586 }
2587 
2588 /// If we have a conditional branch on a PHI node value that is defined in the
2589 /// same block as the branch and if any PHI entries are constants, thread edges
2590 /// corresponding to that entry to be branches to their ultimate destination.
2592  const DataLayout &DL, AssumptionCache *AC) {
2593  BasicBlock *BB = BI->getParent();
2594  PHINode *PN = dyn_cast<PHINode>(BI->getCondition());
2595  // NOTE: we currently cannot transform this case if the PHI node is used
2596  // outside of the block.
2597  if (!PN || PN->getParent() != BB || !PN->hasOneUse())
2598  return false;
2599 
2600  // Degenerate case of a single entry PHI.
2601  if (PN->getNumIncomingValues() == 1) {
2603  return true;
2604  }
2605 
2606  // Now we know that this block has multiple preds and two succs.
2608  return false;
2609 
2610  // Okay, this is a simple enough basic block. See if any phi values are
2611  // constants.
2612  for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
2613  ConstantInt *CB = dyn_cast<ConstantInt>(PN->getIncomingValue(i));
2614  if (!CB || !CB->getType()->isIntegerTy(1))
2615  continue;
2616 
2617  // Okay, we now know that all edges from PredBB should be revectored to
2618  // branch to RealDest.
2619  BasicBlock *PredBB = PN->getIncomingBlock(i);
2620  BasicBlock *RealDest = BI->getSuccessor(!CB->getZExtValue());
2621 
2622  if (RealDest == BB)
2623  continue; // Skip self loops.
2624  // Skip if the predecessor's terminator is an indirect branch.
2625  if (isa<IndirectBrInst>(PredBB->getTerminator()))
2626  continue;
2627 
2629 
2630  // The dest block might have PHI nodes, other predecessors and other
2631  // difficult cases. Instead of being smart about this, just insert a new
2632  // block that jumps to the destination block, effectively splitting
2633  // the edge we are about to create.
2634  BasicBlock *EdgeBB =
2635  BasicBlock::Create(BB->getContext(), RealDest->getName() + ".critedge",
2636  RealDest->getParent(), RealDest);
2637  BranchInst *CritEdgeBranch = BranchInst::Create(RealDest, EdgeBB);
2638  if (DTU)
2639  Updates.push_back({DominatorTree::Insert, EdgeBB, RealDest});
2640  CritEdgeBranch->setDebugLoc(BI->getDebugLoc());
2641 
2642  // Update PHI nodes.
2643  AddPredecessorToBlock(RealDest, EdgeBB, BB);
2644 
2645  // BB may have instructions that are being threaded over. Clone these
2646  // instructions into EdgeBB. We know that there will be no uses of the
2647  // cloned instructions outside of EdgeBB.
2648  BasicBlock::iterator InsertPt = EdgeBB->begin();
2649  DenseMap<Value *, Value *> TranslateMap; // Track translated values.
2650  for (BasicBlock::iterator BBI = BB->begin(); &*BBI != BI; ++BBI) {
2651  if (PHINode *PN = dyn_cast<PHINode>(BBI)) {
2652  TranslateMap[PN] = PN->getIncomingValueForBlock(PredBB);
2653  continue;
2654  }
2655  // Clone the instruction.
2656  Instruction *N = BBI->clone();
2657  if (BBI->hasName())
2658  N->setName(BBI->getName() + ".c");
2659 
2660  // Update operands due to translation.
2661  for (Use &Op : N->operands()) {
2662  DenseMap<Value *, Value *>::iterator PI = TranslateMap.find(Op);
2663  if (PI != TranslateMap.end())
2664  Op = PI->second;
2665  }
2666 
2667  // Check for trivial simplification.
2668  if (Value *V = SimplifyInstruction(N, {DL, nullptr, nullptr, AC})) {
2669  if (!BBI->use_empty())
2670  TranslateMap[&*BBI] = V;
2671  if (!N->mayHaveSideEffects()) {
2672  N->deleteValue(); // Instruction folded away, don't need actual inst
2673  N = nullptr;
2674  }
2675  } else {
2676  if (!BBI->use_empty())
2677  TranslateMap[&*BBI] = N;
2678  }
2679  if (N) {
2680  // Insert the new instruction into its new home.
2681  EdgeBB->getInstList().insert(InsertPt, N);
2682 
2683  // Register the new instruction with the assumption cache if necessary.
2684  if (auto *Assume = dyn_cast<AssumeInst>(N))
2685  if (AC)
2686  AC->registerAssumption(Assume);
2687  }
2688  }
2689 
2690  // Loop over all of the edges from PredBB to BB, changing them to branch
2691  // to EdgeBB instead.
2692  Instruction *PredBBTI = PredBB->getTerminator();
2693  for (unsigned i = 0, e = PredBBTI->getNumSuccessors(); i != e; ++i)
2694  if (PredBBTI->getSuccessor(i) == BB) {
2695  BB->removePredecessor(PredBB);
2696  PredBBTI->setSuccessor(i, EdgeBB);
2697  }
2698 
2699  if (DTU) {
2700  Updates.push_back({DominatorTree::Insert, PredBB, EdgeBB});
2701  Updates.push_back({DominatorTree::Delete, PredBB, BB});
2702 
2703  DTU->applyUpdates(Updates);
2704  }
2705 
2706  // Recurse, simplifying any other constants.
2707  return FoldCondBranchOnPHI(BI, DTU, DL, AC) || true;
2708  }
2709 
2710  return false;
2711 }
2712 
2713 /// Given a BB that starts with the specified two-entry PHI node,
2714 /// see if we can eliminate it.
2716  DomTreeUpdater *DTU, const DataLayout &DL) {
2717  // Ok, this is a two entry PHI node. Check to see if this is a simple "if
2718  // statement", which has a very simple dominance structure. Basically, we
2719  // are trying to find the condition that is being branched on, which
2720  // subsequently causes this merge to happen. We really want control
2721  // dependence information for this check, but simplifycfg can't keep it up
2722  // to date, and this catches most of the cases we care about anyway.
2723  BasicBlock *BB = PN->getParent();
2724 
2725  BasicBlock *IfTrue, *IfFalse;
2726  BranchInst *DomBI = GetIfCondition(BB, IfTrue, IfFalse);
2727  if (!DomBI)
2728  return false;
2729  Value *IfCond = DomBI->getCondition();
2730  // Don't bother if the branch will be constant folded trivially.
2731  if (isa<ConstantInt>(IfCond))
2732  return false;
2733 
2734  BasicBlock *DomBlock = DomBI->getParent();
2736  llvm::copy_if(
2737  PN->blocks(), std::back_inserter(IfBlocks), [](BasicBlock *IfBlock) {
2738  return cast<BranchInst>(IfBlock->getTerminator())->isUnconditional();
2739  });
2740  assert((IfBlocks.size() == 1 || IfBlocks.size() == 2) &&
2741  "Will have either one or two blocks to speculate.");
2742 
2743  // If the branch is non-unpredictable, see if we either predictably jump to
2744  // the merge bb (if we have only a single 'then' block), or if we predictably
2745  // jump to one specific 'then' block (if we have two of them).
2746  // It isn't beneficial to speculatively execute the code
2747  // from the block that we know is predictably not entered.
2748  if (!DomBI->getMetadata(LLVMContext::MD_unpredictable)) {
2749  uint64_t TWeight, FWeight;
2750  if (DomBI->extractProfMetadata(TWeight, FWeight) &&
2751  (TWeight + FWeight) != 0) {
2752  BranchProbability BITrueProb =
2753  BranchProbability::getBranchProbability(TWeight, TWeight + FWeight);
2755  BranchProbability BIFalseProb = BITrueProb.getCompl();
2756  if (IfBlocks.size() == 1) {
2757  BranchProbability BIBBProb =
2758  DomBI->getSuccessor(0) == BB ? BITrueProb : BIFalseProb;
2759  if (BIBBProb >= Likely)
2760  return false;
2761  } else {
2762  if (BITrueProb >= Likely || BIFalseProb >= Likely)
2763  return false;
2764  }
2765  }
2766  }
2767 
2768  // Don't try to fold an unreachable block. For example, the phi node itself
2769  // can't be the candidate if-condition for a select that we want to form.
2770  if (auto *IfCondPhiInst = dyn_cast<PHINode>(IfCond))
2771  if (IfCondPhiInst->getParent() == BB)
2772  return false;
2773 
2774  // Okay, we found that we can merge this two-entry phi node into a select.
2775  // Doing so would require us to fold *all* two entry phi nodes in this block.
2776  // At some point this becomes non-profitable (particularly if the target
2777  // doesn't support cmov's). Only do this transformation if there are two or
2778  // fewer PHI nodes in this block.
2779  unsigned NumPhis = 0;
2780  for (BasicBlock::iterator I = BB->begin(); isa<PHINode>(I); ++NumPhis, ++I)
2781  if (NumPhis > 2)
2782  return false;
2783 
2784  // Loop over the PHI's seeing if we can promote them all to select
2785  // instructions. While we are at it, keep track of the instructions
2786  // that need to be moved to the dominating block.
2787  SmallPtrSet<Instruction *, 4> AggressiveInsts;
2788  InstructionCost Cost = 0;
2789  InstructionCost Budget =
2791 
2792  bool Changed = false;
2793  for (BasicBlock::iterator II = BB->begin(); isa<PHINode>(II);) {
2794  PHINode *PN = cast<PHINode>(II++);
2795  if (Value *V = SimplifyInstruction(PN, {DL, PN})) {
2796  PN->replaceAllUsesWith(V);
2797  PN->eraseFromParent();
2798  Changed = true;
2799  continue;
2800  }
2801 
2802  if (!dominatesMergePoint(PN->getIncomingValue(0), BB, AggressiveInsts,
2803  Cost, Budget, TTI) ||
2804  !dominatesMergePoint(PN->getIncomingValue(1), BB, AggressiveInsts,
2805  Cost, Budget, TTI))
2806  return Changed;
2807  }
2808 
2809  // If we folded the first phi, PN dangles at this point. Refresh it. If
2810  // we ran out of PHIs then we simplified them all.
2811  PN = dyn_cast<PHINode>(BB->begin());
2812  if (!PN)
2813  return true;
2814 
2815  // Return true if at least one of these is a 'not', and another is either
2816  // a 'not' too, or a constant.
2817  auto CanHoistNotFromBothValues = [](Value *V0, Value *V1) {
2818  if (!match(V0, m_Not(m_Value())))
2819  std::swap(V0, V1);
2820  auto Invertible = m_CombineOr(m_Not(m_Value()), m_AnyIntegralConstant());
2821  return match(V0, m_Not(m_Value())) && match(V1, Invertible);
2822  };
2823 
2824  // Don't fold i1 branches on PHIs which contain binary operators or
2825  // (possibly inverted) select form of or/ands, unless one of
2826  // the incoming values is an 'not' and another one is freely invertible.
2827  // These can often be turned into switches and other things.
2828  auto IsBinOpOrAnd = [](Value *V) {
2829  return match(
2830  V, m_CombineOr(
2831  m_BinOp(),
2833  m_Select(m_Value(), m_Value(), m_ImmConstant()))));
2834  };
2835  if (PN->getType()->isIntegerTy(1) &&
2836  (IsBinOpOrAnd(PN->getIncomingValue(0)) ||
2837  IsBinOpOrAnd(PN->getIncomingValue(1)) || IsBinOpOrAnd(IfCond)) &&
2838  !CanHoistNotFromBothValues(PN->getIncomingValue(0),
2839  PN->getIncomingValue(1)))
2840  return Changed;
2841 
2842  // If all PHI nodes are promotable, check to make sure that all instructions
2843  // in the predecessor blocks can be promoted as well. If not, we won't be able
2844  // to get rid of the control flow, so it's not worth promoting to select
2845  // instructions.
2846  for (BasicBlock *IfBlock : IfBlocks)
2847  for (BasicBlock::iterator I = IfBlock->begin(); !I->isTerminator(); ++I)
2848  if (!AggressiveInsts.count(&*I) && !isa<DbgInfoIntrinsic>(I) &&
2849  !isa<PseudoProbeInst>(I)) {
2850  // This is not an aggressive instruction that we can promote.
2851  // Because of this, we won't be able to get rid of the control flow, so
2852  // the xform is not worth it.
2853  return Changed;
2854  }
2855 
2856  // If either of the blocks has it's address taken, we can't do this fold.
2857  if (any_of(IfBlocks,
2858  [](BasicBlock *IfBlock) { return IfBlock->hasAddressTaken(); }))
2859  return Changed;
2860 
2861  LLVM_DEBUG(dbgs() << "FOUND IF CONDITION! " << *IfCond
2862  << " T: " << IfTrue->getName()
2863  << " F: " << IfFalse->getName() << "\n");
2864 
2865  // If we can still promote the PHI nodes after this gauntlet of tests,
2866  // do all of the PHI's now.
2867 
2868  // Move all 'aggressive' instructions, which are defined in the
2869  // conditional parts of the if's up to the dominating block.
2870  for (BasicBlock *IfBlock : IfBlocks)
2871  hoistAllInstructionsInto(DomBlock, DomBI, IfBlock);
2872 
2874  // Propagate fast-math-flags from phi nodes to replacement selects.
2876  while (PHINode *PN = dyn_cast<PHINode>(BB->begin())) {
2877  if (isa<FPMathOperator>(PN))
2878  Builder.setFastMathFlags(PN->getFastMathFlags());
2879 
2880  // Change the PHI node into a select instruction.
2881  Value *TrueVal = PN->getIncomingValueForBlock(IfTrue);
2882  Value *FalseVal = PN->getIncomingValueForBlock(IfFalse);
2883 
2884  Value *Sel = Builder.CreateSelect(IfCond, TrueVal, FalseVal, "", DomBI);
2885  PN->replaceAllUsesWith(Sel);
2886  Sel->takeName(PN);
2887  PN->eraseFromParent();
2888  }
2889 
2890  // At this point, all IfBlocks are empty, so our if statement
2891  // has been flattened. Change DomBlock to jump directly to our new block to
2892  // avoid other simplifycfg's kicking in on the diamond.
2893  Builder.CreateBr(BB);
2894 
2896  if (DTU) {
2897  Updates.push_back({DominatorTree::Insert, DomBlock, BB});
2898  for (auto *Successor : successors(DomBlock))
2899  Updates.push_back({DominatorTree::Delete, DomBlock, Successor});
2900  }
2901 
2902  DomBI->eraseFromParent();
2903  if (DTU)
2904  DTU->applyUpdates(Updates);
2905 
2906  return true;
2907 }
2908 
2910  Instruction::BinaryOps Opc, Value *LHS,
2911  Value *RHS, const Twine &Name = "") {
2912  // Try to relax logical op to binary op.
2913  if (impliesPoison(RHS, LHS))
2914  return Builder.CreateBinOp(Opc, LHS, RHS, Name);
2915  if (Opc == Instruction::And)
2916  return Builder.CreateLogicalAnd(LHS, RHS, Name);
2917  if (Opc == Instruction::Or)
2918  return Builder.CreateLogicalOr(LHS, RHS, Name);
2919  llvm_unreachable("Invalid logical opcode");
2920 }
2921 
2922 /// Return true if either PBI or BI has branch weight available, and store
2923 /// the weights in {Pred|Succ}{True|False}Weight. If one of PBI and BI does
2924 /// not have branch weight, use 1:1 as its weight.
2926  uint64_t &PredTrueWeight,
2927  uint64_t &PredFalseWeight,
2928  uint64_t &SuccTrueWeight,
2929  uint64_t &SuccFalseWeight) {
2930  bool PredHasWeights =
2931  PBI->extractProfMetadata(PredTrueWeight, PredFalseWeight);
2932  bool SuccHasWeights =
2933  BI->extractProfMetadata(SuccTrueWeight, SuccFalseWeight);
2934  if (PredHasWeights || SuccHasWeights) {
2935  if (!PredHasWeights)
2936  PredTrueWeight = PredFalseWeight = 1;
2937  if (!SuccHasWeights)
2938  SuccTrueWeight = SuccFalseWeight = 1;
2939  return true;
2940  } else {
2941  return false;
2942  }
2943 }
2944 
2945 /// Determine if the two branches share a common destination and deduce a glue
2946 /// that joins the branches' conditions to arrive at the common destination if
2947 /// that would be profitable.
2950  const TargetTransformInfo *TTI) {
2951  assert(BI && PBI && BI->isConditional() && PBI->isConditional() &&
2952  "Both blocks must end with a conditional branches.");
2954  "PredBB must be a predecessor of BB.");
2955 
2956  // We have the potential to fold the conditions together, but if the
2957  // predecessor branch is predictable, we may not want to merge them.
2958  uint64_t PTWeight, PFWeight;
2959  BranchProbability PBITrueProb, Likely;
2960  if (TTI && !PBI->getMetadata(LLVMContext::MD_unpredictable) &&
2961  PBI->extractProfMetadata(PTWeight, PFWeight) &&
2962  (PTWeight + PFWeight) != 0) {
2963  PBITrueProb =
2964  BranchProbability::getBranchProbability(PTWeight, PTWeight + PFWeight);
2965  Likely = TTI->getPredictableBranchThreshold();
2966  }
2967 
2968  if (PBI->getSuccessor(0) == BI->getSuccessor(0)) {
2969  // Speculate the 2nd condition unless the 1st is probably true.
2970  if (PBITrueProb.isUnknown() || PBITrueProb < Likely)
2971  return {{Instruction::Or, false}};
2972  } else if (PBI->getSuccessor(1) == BI->getSuccessor(1)) {
2973  // Speculate the 2nd condition unless the 1st is probably false.
2974  if (PBITrueProb.isUnknown() || PBITrueProb.getCompl() < Likely)
2975  return {{Instruction::And, false}};
2976  } else if (PBI->getSuccessor(0) == BI->getSuccessor(1)) {
2977  // Speculate the 2nd condition unless the 1st is probably true.
2978  if (PBITrueProb.isUnknown() || PBITrueProb < Likely)
2979  return {{Instruction::And, true}};
2980  } else if (PBI->getSuccessor(1) == BI->getSuccessor(0)) {
2981  // Speculate the 2nd condition unless the 1st is probably false.
2982  if (PBITrueProb.isUnknown() || PBITrueProb.getCompl() < Likely)
2983  return {{Instruction::Or, true}};
2984  }
2985  return None;
2986 }
2987 
2989  DomTreeUpdater *DTU,
2990  MemorySSAUpdater *MSSAU,
2991  const TargetTransformInfo *TTI) {
2992  BasicBlock *BB = BI->getParent();
2993  BasicBlock *PredBlock = PBI->getParent();
2994 
2995  // Determine if the two branches share a common destination.
2997  bool InvertPredCond;
2998  std::tie(Opc, InvertPredCond) =
3000 
3001  LLVM_DEBUG(dbgs() << "FOLDING BRANCH TO COMMON DEST:\n" << *PBI << *BB);
3002 
3003  IRBuilder<> Builder(PBI);
3004  // The builder is used to create instructions to eliminate the branch in BB.
3005  // If BB's terminator has !annotation metadata, add it to the new
3006  // instructions.
3007  Builder.CollectMetadataToCopy(BB->getTerminator(),
3008  {LLVMContext::MD_annotation});
3009 
3010  // If we need to invert the condition in the pred block to match, do so now.
3011  if (InvertPredCond) {
3012  Value *NewCond = PBI->getCondition();
3013  if (NewCond->hasOneUse() && isa<CmpInst>(NewCond)) {
3014  CmpInst *CI = cast<CmpInst>(NewCond);
3015  CI->setPredicate(CI->getInversePredicate());
3016  } else {
3017  NewCond =
3018  Builder.CreateNot(NewCond, PBI->getCondition()->getName() + ".not");
3019  }
3020 
3021  PBI->setCondition(NewCond);
3022  PBI->swapSuccessors();
3023  }
3024 
3025  BasicBlock *UniqueSucc =
3026  PBI->getSuccessor(0) == BB ? BI->getSuccessor(0) : BI->getSuccessor(1);
3027 
3028  // Before cloning instructions, notify the successor basic block that it
3029  // is about to have a new predecessor. This will update PHI nodes,
3030  // which will allow us to update live-out uses of bonus instructions.
3031  AddPredecessorToBlock(UniqueSucc, PredBlock, BB, MSSAU);
3032 
3033  // Try to update branch weights.
3034  uint64_t PredTrueWeight, PredFalseWeight, SuccTrueWeight, SuccFalseWeight;
3035  if (extractPredSuccWeights(PBI, BI, PredTrueWeight, PredFalseWeight,
3036  SuccTrueWeight, SuccFalseWeight)) {
3037  SmallVector<uint64_t, 8> NewWeights;
3038 
3039  if (PBI->getSuccessor(0) == BB) {
3040  // PBI: br i1 %x, BB, FalseDest
3041  // BI: br i1 %y, UniqueSucc, FalseDest
3042  // TrueWeight is TrueWeight for PBI * TrueWeight for BI.
3043  NewWeights.push_back(PredTrueWeight * SuccTrueWeight);
3044  // FalseWeight is FalseWeight for PBI * TotalWeight for BI +
3045  // TrueWeight for PBI * FalseWeight for BI.
3046  // We assume that total weights of a BranchInst can fit into 32 bits.
3047  // Therefore, we will not have overflow using 64-bit arithmetic.
3048  NewWeights.push_back(PredFalseWeight *
3049  (SuccFalseWeight + SuccTrueWeight) +
3050  PredTrueWeight * SuccFalseWeight);
3051  } else {
3052  // PBI: br i1 %x, TrueDest, BB
3053  // BI: br i1 %y, TrueDest, UniqueSucc
3054  // TrueWeight is TrueWeight for PBI * TotalWeight for BI +
3055  // FalseWeight for PBI * TrueWeight for BI.
3056  NewWeights.push_back(PredTrueWeight * (SuccFalseWeight + SuccTrueWeight) +
3057  PredFalseWeight * SuccTrueWeight);
3058  // FalseWeight is FalseWeight for PBI * FalseWeight for BI.
3059  NewWeights.push_back(PredFalseWeight * SuccFalseWeight);
3060  }
3061 
3062  // Halve the weights if any of them cannot fit in an uint32_t
3063  FitWeights(NewWeights);
3064 
3065  SmallVector<uint32_t, 8> MDWeights(NewWeights.begin(), NewWeights.end());
3066  setBranchWeights(PBI, MDWeights[0], MDWeights[1]);
3067 
3068  // TODO: If BB is reachable from all paths through PredBlock, then we
3069  // could replace PBI's branch probabilities with BI's.
3070  } else
3071  PBI->setMetadata(LLVMContext::MD_prof, nullptr);
3072 
3073  // Now, update the CFG.
3074  PBI->setSuccessor(PBI->getSuccessor(0) != BB, UniqueSucc);
3075 
3076  if (DTU)
3077  DTU->applyUpdates({{DominatorTree::Insert, PredBlock, UniqueSucc},
3078  {DominatorTree::Delete, PredBlock, BB}});
3079 
3080  // If BI was a loop latch, it may have had associated loop metadata.
3081  // We need to copy it to the new latch, that is, PBI.
3082  if (MDNode *LoopMD = BI->getMetadata(LLVMContext::MD_loop))
3083  PBI->setMetadata(LLVMContext::MD_loop, LoopMD);
3084 
3085  ValueToValueMapTy VMap; // maps original values to cloned values
3087 
3088  // Now that the Cond was cloned into the predecessor basic block,
3089  // or/and the two conditions together.
3090  Value *BICond = VMap[BI->getCondition()];
3091  PBI->setCondition(
3092  createLogicalOp(Builder, Opc, PBI->getCondition(), BICond, "or.cond"));
3093 
3094  // Copy any debug value intrinsics into the end of PredBlock.
3095  for (Instruction &I : *BB) {
3096  if (isa<DbgInfoIntrinsic>(I)) {
3097  Instruction *NewI = I.clone();
3098  RemapInstruction(NewI, VMap,
3100  NewI->insertBefore(PBI);
3101  }
3102  }
3103 
3104  ++NumFoldBranchToCommonDest;
3105  return true;
3106 }
3107 
3108 /// If this basic block is simple enough, and if a predecessor branches to us
3109 /// and one of our successors, fold the block into the predecessor and use
3110 /// logical operations to pick the right destination.
3112  MemorySSAUpdater *MSSAU,
3113  const TargetTransformInfo *TTI,
3114  unsigned BonusInstThreshold) {
3115  // If this block ends with an unconditional branch,
3116  // let SpeculativelyExecuteBB() deal with it.
3117  if (!BI->isConditional())
3118  return false;
3119 
3120  BasicBlock *BB = BI->getParent();
3122  BB->getParent()->hasMinSize() ? TargetTransformInfo::TCK_CodeSize
3124 
3125  Instruction *Cond = dyn_cast<Instruction>(BI->getCondition());
3126 
3127  if (!Cond || (!isa<CmpInst>(Cond) && !isa<BinaryOperator>(Cond)) ||
3128  Cond->getParent() != BB || !Cond->hasOneUse())
3129  return false;
3130 
3131  // Cond is known to be a compare or binary operator. Check to make sure that
3132  // neither operand is a potentially-trapping constant expression.
3133  if (ConstantExpr *CE = dyn_cast<ConstantExpr>(Cond->getOperand(0)))
3134  if (CE->canTrap())
3135  return false;
3136  if (ConstantExpr *CE = dyn_cast<ConstantExpr>(Cond->getOperand(1)))
3137  if (CE->canTrap())
3138  return false;
3139 
3140  // Finally, don't infinitely unroll conditional loops.
3141  if (is_contained(successors(BB), BB))
3142  return false;
3143 
3144  // With which predecessors will we want to deal with?
3146  for (BasicBlock *PredBlock : predecessors(BB)) {
3147  BranchInst *PBI = dyn_cast<BranchInst>(PredBlock->getTerminator());
3148 
3149  // Check that we have two conditional branches. If there is a PHI node in
3150  // the common successor, verify that the same value flows in from both
3151  // blocks.
3152  if (!PBI || PBI->isUnconditional() || !SafeToMergeTerminators(BI, PBI))
3153  continue;
3154 
3155  // Determine if the two branches share a common destination.
3157  bool InvertPredCond;
3158  if (auto Recipe = shouldFoldCondBranchesToCommonDestination(BI, PBI, TTI))
3159  std::tie(Opc, InvertPredCond) = *Recipe;
3160  else
3161  continue;
3162 
3163  // Check the cost of inserting the necessary logic before performing the
3164  // transformation.
3165  if (TTI) {
3166  Type *Ty = BI->getCondition()->getType();
3168  if (InvertPredCond && (!PBI->getCondition()->hasOneUse() ||
3169  !isa<CmpInst>(PBI->getCondition())))
3170  Cost += TTI->getArithmeticInstrCost(Instruction::Xor, Ty, CostKind);
3171 
3172  if (Cost > BranchFoldThreshold)
3173  continue;
3174  }
3175 
3176  // Ok, we do want to deal with this predecessor. Record it.
3177  Preds.emplace_back(PredBlock);
3178  }
3179 
3180  // If there aren't any predecessors into which we can fold,
3181  // don't bother checking the cost.
3182  if (Preds.empty())
3183  return false;
3184 
3185  // Only allow this transformation if computing the condition doesn't involve
3186  // too many instructions and these involved instructions can be executed
3187  // unconditionally. We denote all involved instructions except the condition
3188  // as "bonus instructions", and only allow this transformation when the
3189  // number of the bonus instructions we'll need to create when cloning into
3190  // each predecessor does not exceed a certain threshold.
3191  unsigned NumBonusInsts = 0;
3192  const unsigned PredCount = Preds.size();
3193  for (Instruction &I : *BB) {
3194  // Don't check the branch condition comparison itself.
3195  if (&I == Cond)
3196  continue;
3197  // Ignore dbg intrinsics, and the terminator.
3198  if (isa<DbgInfoIntrinsic>(I) || isa<BranchInst>(I))
3199  continue;
3200  // I must be safe to execute unconditionally.
3202  return false;
3203 
3204  // Account for the cost of duplicating this instruction into each
3205  // predecessor.
3206  NumBonusInsts += PredCount;
3207  // Early exits once we reach the limit.
3208  if (NumBonusInsts > BonusInstThreshold)
3209  return false;
3210  }
3211 
3212  // Ok, we have the budget. Perform the transformation.
3213  for (BasicBlock *PredBlock : Preds) {
3214  auto *PBI = cast<BranchInst>(PredBlock->getTerminator());
3215  return performBranchToCommonDestFolding(BI, PBI, DTU, MSSAU, TTI);
3216  }
3217  return false;
3218 }
3219 
3220 // If there is only one store in BB1 and BB2, return it, otherwise return
3221 // nullptr.
3223  StoreInst *S = nullptr;
3224  for (auto *BB : {BB1, BB2}) {
3225  if (!BB)
3226  continue;
3227  for (auto &I : *BB)
3228  if (auto *SI = dyn_cast<StoreInst>(&I)) {
3229  if (S)
3230  // Multiple stores seen.
3231  return nullptr;
3232  else
3233  S = SI;
3234  }
3235  }
3236  return S;
3237 }
3238 
3240  Value *AlternativeV = nullptr) {
3241  // PHI is going to be a PHI node that allows the value V that is defined in
3242  // BB to be referenced in BB's only successor.
3243  //
3244  // If AlternativeV is nullptr, the only value we care about in PHI is V. It
3245  // doesn't matter to us what the other operand is (it'll never get used). We
3246  // could just create a new PHI with an undef incoming value, but that could
3247  // increase register pressure if EarlyCSE/InstCombine can't fold it with some
3248  // other PHI. So here we directly look for some PHI in BB's successor with V
3249  // as an incoming operand. If we find one, we use it, else we create a new
3250  // one.
3251  //
3252  // If AlternativeV is not nullptr, we care about both incoming values in PHI.
3253  // PHI must be exactly: phi <ty> [ %BB, %V ], [ %OtherBB, %AlternativeV]
3254  // where OtherBB is the single other predecessor of BB's only successor.
3255  PHINode *PHI = nullptr;
3256  BasicBlock *Succ = BB->getSingleSuccessor();
3257 
3258  for (auto I = Succ->begin(); isa<PHINode>(I); ++I)
3259  if (cast<PHINode>(I)->getIncomingValueForBlock(BB) == V) {
3260  PHI = cast<PHINode>(I);
3261  if (!AlternativeV)
3262  break;
3263 
3264  assert(Succ->hasNPredecessors(2));
3265  auto PredI = pred_begin(Succ);
3266  BasicBlock *OtherPredBB = *PredI == BB ? *++PredI : *PredI;
3267  if (PHI->getIncomingValueForBlock(OtherPredBB) == AlternativeV)
3268  break;
3269  PHI = nullptr;
3270  }
3271  if (PHI)
3272  return PHI;
3273 
3274  // If V is not an instruction defined in BB, just return it.
3275  if (!AlternativeV &&
3276  (!isa<Instruction>(V) || cast<Instruction>(V)->getParent() != BB))
3277  return V;
3278 
3279  PHI = PHINode::Create(V->getType(), 2, "simplifycfg.merge", &Succ->front());
3280  PHI->addIncoming(V, BB);
3281  for (BasicBlock *PredBB : predecessors(Succ))
3282  if (PredBB != BB)
3283  PHI->addIncoming(
3284  AlternativeV ? AlternativeV : UndefValue::get(V->getType()), PredBB);
3285  return PHI;
3286 }
3287 
3289  BasicBlock *PTB, BasicBlock *PFB, BasicBlock *QTB, BasicBlock *QFB,
3290  BasicBlock *PostBB, Value *Address, bool InvertPCond, bool InvertQCond,
3291  DomTreeUpdater *DTU, const DataLayout &DL, const TargetTransformInfo &TTI) {
3292  // For every pointer, there must be exactly two stores, one coming from
3293  // PTB or PFB, and the other from QTB or QFB. We don't support more than one
3294  // store (to any address) in PTB,PFB or QTB,QFB.
3295  // FIXME: We could relax this restriction with a bit more work and performance
3296  // testing.
3297  StoreInst *PStore = findUniqueStoreInBlocks(PTB, PFB);
3298  StoreInst *QStore = findUniqueStoreInBlocks(QTB, QFB);
3299  if (!PStore || !QStore)
3300  return false;
3301 
3302  // Now check the stores are compatible.
3303  if (!QStore->isUnordered() || !PStore->isUnordered())
3304  return false;
3305 
3306  // Check that sinking the store won't cause program behavior changes. Sinking
3307  // the store out of the Q blocks won't change any behavior as we're sinking
3308  // from a block to its unconditional successor. But we're moving a store from
3309  // the P blocks down through the middle block (QBI) and past both QFB and QTB.
3310  // So we need to check that there are no aliasing loads or stores in
3311  // QBI, QTB and QFB. We also need to check there are no conflicting memory
3312  // operations between PStore and the end of its parent block.
3313  //
3314  // The ideal way to do this is to query AliasAnalysis, but we don't
3315  // preserve AA currently so that is dangerous. Be super safe and just
3316  // check there are no other memory operations at all.
3317  for (auto &I : *QFB->getSinglePredecessor())
3318  if (I.mayReadOrWriteMemory())
3319  return false;
3320  for (auto &I : *QFB)
3321  if (&I != QStore && I.mayReadOrWriteMemory())
3322  return false;
3323  if (QTB)
3324  for (auto &I : *QTB)
3325  if (&I != QStore && I.mayReadOrWriteMemory())
3326  return false;
3327  for (auto I = BasicBlock::iterator(PStore), E = PStore->getParent()->end();
3328  I != E; ++I)
3329  if (&*I != PStore && I->mayReadOrWriteMemory())
3330  return false;
3331 
3332  // If we're not in aggressive mode, we only optimize if we have some
3333  // confidence that by optimizing we'll allow P and/or Q to be if-converted.
3334  auto IsWorthwhile = [&](BasicBlock *BB, ArrayRef<StoreInst *> FreeStores) {
3335  if (!BB)
3336  return true;
3337  // Heuristic: if the block can be if-converted/phi-folded and the
3338  // instructions inside are all cheap (arithmetic/GEPs), it's worthwhile to
3339  // thread this store.
3340  InstructionCost Cost = 0;
3341  InstructionCost Budget =
3343  for (auto &I : BB->instructionsWithoutDebug()) {
3344  // Consider terminator instruction to be free.
3345  if (I.isTerminator())
3346  continue;
3347  // If this is one the stores that we want to speculate out of this BB,
3348  // then don't count it's cost, consider it to be free.
3349  if (auto *S = dyn_cast<StoreInst>(&I))
3350  if (llvm::find(FreeStores, S))
3351  continue;
3352  // Else, we have a white-list of instructions that we are ak speculating.
3353  if (!isa<BinaryOperator>(I) && !isa<GetElementPtrInst>(I))
3354  return false; // Not in white-list - not worthwhile folding.
3355  // And finally, if this is a non-free instruction that we are okay
3356  // speculating, ensure that we consider the speculation budget.
3358  if (Cost > Budget)
3359  return false; // Eagerly refuse to fold as soon as we're out of budget.
3360  }
3361  assert(Cost <= Budget &&
3362  "When we run out of budget we will eagerly return from within the "
3363  "per-instruction loop.");
3364  return true;
3365  };
3366 
3367  const std::array<StoreInst *, 2> FreeStores = {PStore, QStore};
3369  (!IsWorthwhile(PTB, FreeStores) || !IsWorthwhile(PFB, FreeStores) ||
3370  !IsWorthwhile(QTB, FreeStores) || !IsWorthwhile(QFB, FreeStores)))
3371  return false;
3372 
3373  // If PostBB has more than two predecessors, we need to split it so we can
3374  // sink the store.
3375  if (std::next(pred_begin(PostBB), 2) != pred_end(PostBB)) {
3376  // We know that QFB's only successor is PostBB. And QFB has a single
3377  // predecessor. If QTB exists, then its only successor is also PostBB.
3378  // If QTB does not exist, then QFB's only predecessor has a conditional
3379  // branch to QFB and PostBB.
3380  BasicBlock *TruePred = QTB ? QTB : QFB->getSinglePredecessor();
3381  BasicBlock *NewBB =
3382  SplitBlockPredecessors(PostBB, {QFB, TruePred}, "condstore.split", DTU);
3383  if (!NewBB)
3384  return false;
3385  PostBB = NewBB;
3386  }
3387 
3388  // OK, we're going to sink the stores to PostBB. The store has to be
3389  // conditional though, so first create the predicate.
3390  Value *PCond = cast<BranchInst>(PFB->getSinglePredecessor()->getTerminator())
3391  ->getCondition();
3392  Value *QCond = cast<BranchInst>(QFB->getSinglePredecessor()->getTerminator())
3393  ->getCondition();
3394 
3396  PStore->getParent());
3398  QStore->getParent(), PPHI);
3399 
3400  IRBuilder<> QB(&*PostBB->getFirstInsertionPt());
3401 
3402  Value *PPred = PStore->getParent() == PTB ? PCond : QB.CreateNot(PCond);
3403  Value *QPred = QStore->getParent() == QTB ? QCond : QB.CreateNot(QCond);
3404 
3405  if (InvertPCond)
3406  PPred = QB.CreateNot(PPred);
3407  if (InvertQCond)
3408  QPred = QB.CreateNot(QPred);
3409  Value *CombinedPred = QB.CreateOr(PPred, QPred);
3410 
3411  auto *T = SplitBlockAndInsertIfThen(CombinedPred, &*QB.GetInsertPoint(),
3412  /*Unreachable=*/false,
3413  /*BranchWeights=*/nullptr, DTU);
3414  QB.SetInsertPoint(T);
3415  StoreInst *SI = cast<StoreInst>(QB.CreateStore(QPHI, Address));
3416  AAMDNodes AAMD;
3417  PStore->getAAMetadata(AAMD, /*Merge=*/false);
3418  PStore->getAAMetadata(AAMD, /*Merge=*/true);
3419  SI->setAAMetadata(AAMD);
3420  // Choose the minimum alignment. If we could prove both stores execute, we
3421  // could use biggest one. In this case, though, we only know that one of the
3422  // stores executes. And we don't know it's safe to take the alignment from a
3423  // store that doesn't execute.
3424  SI->setAlignment(std::min(PStore->getAlign(), QStore->getAlign()));
3425 
3426  QStore->eraseFromParent();
3427  PStore->eraseFromParent();
3428 
3429  return true;
3430 }
3431 
3433  DomTreeUpdater *DTU, const DataLayout &DL,
3434  const TargetTransformInfo &TTI) {
3435  // The intention here is to find diamonds or triangles (see below) where each
3436  // conditional block contains a store to the same address. Both of these
3437  // stores are conditional, so they can't be unconditionally sunk. But it may
3438  // be profitable to speculatively sink the stores into one merged store at the
3439  // end, and predicate the merged store on the union of the two conditions of
3440  // PBI and QBI.
3441  //
3442  // This can reduce the number of stores executed if both of the conditions are
3443  // true, and can allow the blocks to become small enough to be if-converted.
3444  // This optimization will also chain, so that ladders of test-and-set
3445  // sequences can be if-converted away.
3446  //
3447  // We only deal with simple diamonds or triangles:
3448  //
3449  // PBI or PBI or a combination of the two
3450  // / \ | \
3451  // PTB PFB | PFB
3452  // \ / | /
3453  // QBI QBI
3454  // / \ | \
3455  // QTB QFB | QFB
3456  // \ / | /
3457  // PostBB PostBB
3458  //
3459  // We model triangles as a type of diamond with a nullptr "true" block.
3460  // Triangles are canonicalized so that the fallthrough edge is represented by
3461  // a true condition, as in the diagram above.
3462  BasicBlock *PTB = PBI->getSuccessor(0);
3463  BasicBlock *PFB = PBI->getSuccessor(1);
3464  BasicBlock *QTB = QBI->getSuccessor(0);
3465  BasicBlock *QFB = QBI->getSuccessor(1);
3466  BasicBlock *PostBB = QFB->getSingleSuccessor();
3467 
3468  // Make sure we have a good guess for PostBB. If QTB's only successor is
3469  // QFB, then QFB is a better PostBB.
3470  if (QTB->getSingleSuccessor() == QFB)
3471  PostBB = QFB;
3472 
3473  // If we couldn't find a good PostBB, stop.
3474  if (!PostBB)
3475  return false;
3476 
3477  bool InvertPCond = false, InvertQCond = false;
3478  // Canonicalize fallthroughs to the true branches.
3479  if (PFB == QBI->getParent()) {
3480  std::swap(PFB, PTB);
3481  InvertPCond = true;
3482  }
3483  if (QFB == PostBB) {
3484  std::swap(QFB, QTB);
3485  InvertQCond = true;
3486  }
3487 
3488  // From this point on we can assume PTB or QTB may be fallthroughs but PFB
3489  // and QFB may not. Model fallthroughs as a nullptr block.
3490  if (PTB == QBI->getParent())
3491  PTB = nullptr;
3492  if (QTB == PostBB)
3493  QTB = nullptr;
3494 
3495  // Legality bailouts. We must have at least the non-fallthrough blocks and
3496  // the post-dominating block, and the non-fallthroughs must only have one
3497  // predecessor.
3498  auto HasOnePredAndOneSucc = [](BasicBlock *BB, BasicBlock *P, BasicBlock *S) {
3499  return BB->getSinglePredecessor() == P && BB->getSingleSuccessor() == S;
3500  };
3501  if (!HasOnePredAndOneSucc(PFB, PBI->getParent(), QBI->getParent()) ||
3502  !HasOnePredAndOneSucc(QFB, QBI->getParent(), PostBB))
3503  return false;
3504  if ((PTB && !HasOnePredAndOneSucc(PTB, PBI->getParent(), QBI->getParent())) ||
3505  (QTB && !HasOnePredAndOneSucc(QTB, QBI->getParent(), PostBB)))
3506  return false;
3507  if (!QBI->getParent()->hasNUses(2))
3508  return false;
3509 
3510  // OK, this is a sequence of two diamonds or triangles.
3511  // Check if there are stores in PTB or PFB that are repeated in QTB or QFB.
3512  SmallPtrSet<Value *, 4> PStoreAddresses, QStoreAddresses;
3513  for (auto *BB : {PTB, PFB}) {
3514  if (!BB)
3515  continue;
3516  for (auto &I : *BB)
3517  if (StoreInst *SI = dyn_cast<StoreInst>(&I))
3518  PStoreAddresses.insert(SI->getPointerOperand());
3519  }
3520  for (auto *BB : {QTB, QFB}) {
3521  if (!BB)
3522  continue;
3523  for (auto &I : *BB)
3524  if (StoreInst *SI = dyn_cast<StoreInst>(&I))
3525  QStoreAddresses.insert(SI->getPointerOperand());
3526  }
3527 
3528  set_intersect(PStoreAddresses, QStoreAddresses);
3529  // set_intersect mutates PStoreAddresses in place. Rename it here to make it
3530  // clear what it contains.
3531  auto &CommonAddresses = PStoreAddresses;
3532 
3533  bool Changed = false;
3534  for (auto *Address : CommonAddresses)
3535  Changed |=
3536  mergeConditionalStoreToAddress(PTB, PFB, QTB, QFB, PostBB, Address,
3537  InvertPCond, InvertQCond, DTU, DL, TTI);
3538  return Changed;
3539 }
3540 
3541 /// If the previous block ended with a widenable branch, determine if reusing
3542 /// the target block is profitable and legal. This will have the effect of
3543 /// "widening" PBI, but doesn't require us to reason about hosting safety.
3545  DomTreeUpdater *DTU) {
3546  // TODO: This can be generalized in two important ways:
3547  // 1) We can allow phi nodes in IfFalseBB and simply reuse all the input
3548  // values from the PBI edge.
3549  // 2) We can sink side effecting instructions into BI's fallthrough
3550  // successor provided they doesn't contribute to computation of
3551  // BI's condition.
3552  Value *CondWB, *WC;
3553  BasicBlock *IfTrueBB, *IfFalseBB;
3554  if (!parseWidenableBranch(PBI, CondWB, WC, IfTrueBB, IfFalseBB) ||
3555  IfTrueBB != BI->getParent() || !BI->getParent()->getSinglePredecessor())
3556  return false;
3557  if (!IfFalseBB->phis().empty())
3558  return false; // TODO
3559  // Use lambda to lazily compute expensive condition after cheap ones.
3560  auto NoSideEffects = [](BasicBlock &BB) {
3561  return !llvm::any_of(BB, [](const Instruction &I) {
3562  return I.mayWriteToMemory() || I.mayHaveSideEffects();
3563  });
3564  };
3565  if (BI->getSuccessor(1) != IfFalseBB && // no inf looping
3566  BI->getSuccessor(1)->getTerminatingDeoptimizeCall() && // profitability
3567  NoSideEffects(*BI->getParent())) {
3568  auto *OldSuccessor = BI->getSuccessor(1);
3569  OldSuccessor->removePredecessor(BI->getParent());
3570  BI->setSuccessor(1, IfFalseBB);
3571  if (DTU)
3572  DTU->applyUpdates(
3573  {{DominatorTree::Insert, BI->getParent(), IfFalseBB},
3574  {DominatorTree::Delete, BI->getParent(), OldSuccessor}});
3575  return true;
3576  }
3577  if (BI->getSuccessor(0) != IfFalseBB && // no inf looping
3578  BI->getSuccessor(0)->getTerminatingDeoptimizeCall() && // profitability
3579  NoSideEffects(*BI->getParent())) {
3580  auto *OldSuccessor = BI->getSuccessor(0);
3581  OldSuccessor->removePredecessor(BI->getParent());
3582  BI->setSuccessor(0, IfFalseBB);
3583  if (DTU)
3584  DTU->applyUpdates(
3585  {{DominatorTree::Insert, BI->getParent(), IfFalseBB},
3586  {DominatorTree::Delete, BI->getParent(), OldSuccessor}});
3587  return true;
3588  }
3589  return false;
3590 }
3591 
3592 /// If we have a conditional branch as a predecessor of another block,
3593 /// this function tries to simplify it. We know
3594 /// that PBI and BI are both conditional branches, and BI is in one of the
3595 /// successor blocks of PBI - PBI branches to BI.
3597  DomTreeUpdater *DTU,
3598  const DataLayout &DL,
3599  const TargetTransformInfo &TTI) {
3600  assert(PBI->isConditional() && BI->isConditional());
3601  BasicBlock *BB = BI->getParent();
3602 
3603  // If this block ends with a branch instruction, and if there is a
3604  // predecessor that ends on a branch of the same condition, make
3605  // this conditional branch redundant.
3606  if (PBI->getCondition() == BI->getCondition() &&
3607  PBI->getSuccessor(0) != PBI->getSuccessor(1)) {
3608  // Okay, the outcome of this conditional branch is statically
3609  // knowable. If this block had a single pred, handle specially.
3610  if (BB->getSinglePredecessor()) {
3611  // Turn this into a branch on constant.
3612  bool CondIsTrue = PBI->getSuccessor(0) == BB;
3613  BI->setCondition(
3614  ConstantInt::get(Type::getInt1Ty(BB->getContext()), CondIsTrue));
3615  return true; // Nuke the branch on constant.
3616  }
3617 
3618  // Otherwise, if there are multiple predecessors, insert a PHI that merges
3619  // in the constant and simplify the block result. Subsequent passes of
3620  // simplifycfg will thread the block.
3623  PHINode *NewPN = PHINode::Create(
3624  Type::getInt1Ty(BB->getContext()), std::distance(PB, PE),
3625  BI->getCondition()->getName() + ".pr", &BB->front());
3626  // Okay, we're going to insert the PHI node. Since PBI is not the only
3627  // predecessor, compute the PHI'd conditional value for all of the preds.
3628  // Any predecessor where the condition is not computable we keep symbolic.
3629  for (pred_iterator PI = PB; PI != PE; ++PI) {
3630  BasicBlock *P = *PI;
3631  if ((PBI = dyn_cast<BranchInst>(P->getTerminator())) && PBI != BI &&
3632  PBI->isConditional() && PBI->getCondition() == BI->getCondition() &&
3633  PBI->getSuccessor(0) != PBI->getSuccessor(1)) {
3634  bool CondIsTrue = PBI->getSuccessor(0) == BB;
3635  NewPN->addIncoming(
3636  ConstantInt::get(Type::getInt1Ty(BB->getContext()), CondIsTrue),
3637  P);
3638  } else {
3639  NewPN->addIncoming(BI->getCondition(), P);
3640  }
3641  }
3642 
3643  BI->setCondition(NewPN);
3644  return true;
3645  }
3646  }
3647 
3648  // If the previous block ended with a widenable branch, determine if reusing
3649  // the target block is profitable and legal. This will have the effect of
3650  // "widening" PBI, but doesn't require us to reason about hosting safety.
3651  if (tryWidenCondBranchToCondBranch(PBI, BI, DTU))
3652  return true;
3653 
3654  if (auto *CE = dyn_cast<ConstantExpr>(BI->getCondition()))
3655  if (CE->canTrap())
3656  return false;
3657 
3658  // If both branches are conditional and both contain stores to the same
3659  // address, remove the stores from the conditionals and create a conditional
3660  // merged store at the end.
3661  if (MergeCondStores && mergeConditionalStores(PBI, BI, DTU, DL, TTI))
3662  return true;
3663 
3664  // If this is a conditional branch in an empty block, and if any
3665  // predecessors are a conditional branch to one of our destinations,
3666  // fold the conditions into logical ops and one cond br.
3667 
3668  // Ignore dbg intrinsics.
3669  if (&*BB->instructionsWithoutDebug().begin() != BI)
3670  return false;
3671 
3672  int PBIOp, BIOp;
3673  if (PBI->getSuccessor(0) == BI->getSuccessor(0)) {
3674  PBIOp = 0;
3675  BIOp = 0;
3676  } else if (PBI->getSuccessor(0) == BI->getSuccessor(1)) {
3677  PBIOp = 0;
3678  BIOp = 1;
3679  } else if (PBI->getSuccessor(1) == BI->getSuccessor(0)) {
3680  PBIOp = 1;
3681  BIOp = 0;
3682  } else if (PBI->getSuccessor(1) == BI->getSuccessor(1)) {
3683  PBIOp = 1;
3684  BIOp = 1;
3685  } else {
3686  return false;
3687  }
3688 
3689  // Check to make sure that the other destination of this branch
3690  // isn't BB itself. If so, this is an infinite loop that will
3691  // keep getting unwound.
3692  if (PBI->getSuccessor(PBIOp) == BB)
3693  return false;
3694 
3695  // Do not perform this transformation if it would require
3696  // insertion of a large number of select instructions. For targets
3697  // without predication/cmovs, this is a big pessimization.
3698 
3699  // Also do not perform this transformation if any phi node in the common
3700  // destination block can trap when reached by BB or PBB (PR17073). In that
3701  // case, it would be unsafe to hoist the operation into a select instruction.
3702 
3703  BasicBlock *CommonDest = PBI->getSuccessor(PBIOp);
3704  BasicBlock *RemovedDest = PBI->getSuccessor(PBIOp ^ 1);
3705  unsigned NumPhis = 0;
3706  for (BasicBlock::iterator II = CommonDest->begin(); isa<PHINode>(II);
3707  ++II, ++NumPhis) {
3708  if (NumPhis > 2) // Disable this xform.
3709  return false;
3710 
3711  PHINode *PN = cast<PHINode>(II);
3712  Value *BIV = PN->getIncomingValueForBlock(BB);
3713  if (ConstantExpr *CE = dyn_cast<ConstantExpr>(BIV))
3714  if (CE->canTrap())
3715  return false;
3716 
3717  unsigned PBBIdx = PN->getBasicBlockIndex(PBI->getParent());
3718  Value *PBIV = PN->getIncomingValue(PBBIdx);
3719  if (ConstantExpr *CE = dyn_cast<ConstantExpr>(PBIV))
3720  if (CE->canTrap())
3721  return false;
3722  }
3723 
3724  // Finally, if everything is ok, fold the branches to logical ops.
3725  BasicBlock *OtherDest = BI->getSuccessor(BIOp ^ 1);
3726 
3727  LLVM_DEBUG(dbgs() << "FOLDING BRs:" << *PBI->getParent()
3728  << "AND: " << *BI->getParent());
3729 
3731 
3732  // If OtherDest *is* BB, then BB is a basic block with a single conditional
3733  // branch in it, where one edge (OtherDest) goes back to itself but the other
3734  // exits. We don't *know* that the program avoids the infinite loop
3735  // (even though that seems likely). If we do this xform naively, we'll end up
3736  // recursively unpeeling the loop. Since we know that (after the xform is
3737  // done) that the block *is* infinite if reached, we just make it an obviously
3738  // infinite loop with no cond branch.
3739  if (OtherDest == BB) {
3740  // Insert it at the end of the function, because it's either code,
3741  // or it won't matter if it's hot. :)
3742  BasicBlock *InfLoopBlock =
3743  BasicBlock::Create(BB->getContext(), "infloop", BB->getParent());
3744  BranchInst::Create(InfLoopBlock, InfLoopBlock);
3745  if (DTU)
3746  Updates.push_back({DominatorTree::Insert, InfLoopBlock, InfLoopBlock});
3747  OtherDest = InfLoopBlock;
3748  }
3749 
3750  LLVM_DEBUG(dbgs() << *PBI->getParent()->getParent());
3751 
3752  // BI may have other predecessors. Because of this, we leave
3753  // it alone, but modify PBI.
3754 
3755  // Make sure we get to CommonDest on True&True directions.
3756  Value *PBICond = PBI->getCondition();
3758  if (PBIOp)
3759  PBICond = Builder.CreateNot(PBICond, PBICond->getName() + ".not");
3760 
3761  Value *BICond = BI->getCondition();
3762  if (BIOp)
3763  BICond = Builder.CreateNot(BICond, BICond->getName() + ".not");
3764 
3765  // Merge the conditions.
3766  Value *Cond =
3767  createLogicalOp(Builder, Instruction::Or, PBICond, BICond, "brmerge");
3768 
3769  // Modify PBI to branch on the new condition to the new dests.
3770  PBI->setCondition(Cond);
3771  PBI->setSuccessor(0, CommonDest);
3772  PBI->setSuccessor(1, OtherDest);
3773 
3774  if (DTU) {
3775  Updates.push_back({DominatorTree::Insert, PBI->getParent(), OtherDest});
3776  Updates.push_back({DominatorTree::Delete, PBI->getParent(), RemovedDest});
3777 
3778  DTU->applyUpdates(Updates);
3779  }
3780 
3781  // Update branch weight for PBI.
3782  uint64_t PredTrueWeight, PredFalseWeight, SuccTrueWeight, SuccFalseWeight;
3783  uint64_t PredCommon, PredOther, SuccCommon, SuccOther;
3784  bool HasWeights =
3785  extractPredSuccWeights(PBI, BI, PredTrueWeight, PredFalseWeight,
3786  SuccTrueWeight, SuccFalseWeight);
3787  if (HasWeights) {
3788  PredCommon = PBIOp ? PredFalseWeight : PredTrueWeight;
3789  PredOther = PBIOp ? PredTrueWeight : PredFalseWeight;
3790  SuccCommon = BIOp ? SuccFalseWeight : SuccTrueWeight;
3791  SuccOther = BIOp ? SuccTrueWeight : SuccFalseWeight;
3792  // The weight to CommonDest should be PredCommon * SuccTotal +
3793  // PredOther * SuccCommon.
3794  // The weight to OtherDest should be PredOther * SuccOther.
3795  uint64_t NewWeights[2] = {PredCommon * (SuccCommon + SuccOther) +
3796  PredOther * SuccCommon,
3797  PredOther * SuccOther};
3798  // Halve the weights if any of them cannot fit in an uint32_t
3799  FitWeights(NewWeights);
3800 
3801  setBranchWeights(PBI, NewWeights[0], NewWeights[1]);
3802  }
3803 
3804  // OtherDest may have phi nodes. If so, add an entry from PBI's
3805  // block that are identical to the entries for BI's block.
3806  AddPredecessorToBlock(OtherDest, PBI->getParent(), BB);
3807 
3808  // We know that the CommonDest already had an edge from PBI to
3809  // it. If it has PHIs though, the PHIs may have different
3810  // entries for BB and PBI's BB. If so, insert a select to make
3811  // them agree.
3812  for (PHINode &PN : CommonDest->phis()) {
3813  Value *BIV = PN.getIncomingValueForBlock(BB);
3814  unsigned PBBIdx = PN.getBasicBlockIndex(PBI->getParent());
3815  Value *PBIV = PN.getIncomingValue(PBBIdx);
3816  if (BIV != PBIV) {
3817  // Insert a select in PBI to pick the right value.
3818  SelectInst *NV = cast<SelectInst>(
3819  Builder.CreateSelect(PBICond, PBIV, BIV, PBIV->getName() + ".mux"));
3820  PN.setIncomingValue(PBBIdx, NV);
3821  // Although the select has the same condition as PBI, the original branch
3822  // weights for PBI do not apply to the new select because the select's
3823  // 'logical' edges are incoming edges of the phi that is eliminated, not
3824  // the outgoing edges of PBI.
3825  if (HasWeights) {
3826  uint64_t PredCommon = PBIOp ? PredFalseWeight : PredTrueWeight;
3827  uint64_t PredOther = PBIOp ? PredTrueWeight : PredFalseWeight;
3828  uint64_t SuccCommon = BIOp ? SuccFalseWeight : SuccTrueWeight;
3829  uint64_t SuccOther = BIOp ? SuccTrueWeight : SuccFalseWeight;
3830  // The weight to PredCommonDest should be PredCommon * SuccTotal.
3831  // The weight to PredOtherDest should be PredOther * SuccCommon.
3832  uint64_t NewWeights[2] = {PredCommon * (SuccCommon + SuccOther),
3833  PredOther * SuccCommon};
3834 
3835  FitWeights(NewWeights);
3836 
3837  setBranchWeights(NV, NewWeights[0], NewWeights[1]);
3838  }
3839  }
3840  }
3841 
3842  LLVM_DEBUG(dbgs() << "INTO: " << *PBI->getParent());
3843  LLVM_DEBUG(dbgs() << *PBI->getParent()->getParent());
3844 
3845  // This basic block is probably dead. We know it has at least
3846  // one fewer predecessor.
3847  return true;
3848 }
3849 
3850 // Simplifies a terminator by replacing it with a branch to TrueBB if Cond is
3851 // true or to FalseBB if Cond is false.
3852 // Takes care of updating the successors and removing the old terminator.
3853 // Also makes sure not to introduce new successors by assuming that edges to
3854 // non-successor TrueBBs and FalseBBs aren't reachable.
3855 bool SimplifyCFGOpt::SimplifyTerminatorOnSelect(Instruction *OldTerm,
3856  Value *Cond, BasicBlock *TrueBB,
3857  BasicBlock *FalseBB,
3858  uint32_t TrueWeight,
3859  uint32_t FalseWeight) {
3860  auto *BB = OldTerm->getParent();
3861  // Remove any superfluous successor edges from the CFG.
3862  // First, figure out which successors to preserve.
3863  // If TrueBB and FalseBB are equal, only try to preserve one copy of that
3864  // successor.
3865  BasicBlock *KeepEdge1 = TrueBB;
3866  BasicBlock *KeepEdge2 = TrueBB != FalseBB ? FalseBB : nullptr;
3867 
3868  SmallPtrSet<BasicBlock *, 2> RemovedSuccessors;
3869 
3870  // Then remove the rest.
3871  for (BasicBlock *Succ : successors(OldTerm)) {
3872  // Make sure only to keep exactly one copy of each edge.
3873  if (Succ == KeepEdge1)
3874  KeepEdge1 = nullptr;
3875  else if (Succ == KeepEdge2)
3876  KeepEdge2 = nullptr;
3877  else {
3878  Succ->removePredecessor(BB,
3879  /*KeepOneInputPHIs=*/true);
3880 
3881  if (Succ != TrueBB && Succ != FalseBB)
3882  RemovedSuccessors.insert(Succ);
3883  }
3884  }
3885 
3886  IRBuilder<> Builder(OldTerm);
3887  Builder.SetCurrentDebugLocation(OldTerm->getDebugLoc());
3888 
3889  // Insert an appropriate new terminator.
3890  if (!KeepEdge1 && !KeepEdge2) {
3891  if (TrueBB == FalseBB) {
3892  // We were only looking for one successor, and it was present.
3893  // Create an unconditional branch to it.
3894  Builder.CreateBr(TrueBB);
3895  } else {
3896  // We found both of the successors we were looking for.
3897  // Create a conditional branch sharing the condition of the select.
3898  BranchInst *NewBI = Builder.CreateCondBr(Cond, TrueBB, FalseBB);
3899  if (TrueWeight != FalseWeight)
3900  setBranchWeights(NewBI, TrueWeight, FalseWeight);
3901  }
3902  } else if (KeepEdge1 && (KeepEdge2 || TrueBB == FalseBB)) {
3903  // Neither of the selected blocks were successors, so this
3904  // terminator must be unreachable.
3905  new UnreachableInst(OldTerm->getContext(), OldTerm);
3906  } else {
3907  // One of the selected values was a successor, but the other wasn't.
3908  // Insert an unconditional branch to the one that was found;
3909  // the edge to the one that wasn't must be unreachable.
3910  if (!KeepEdge1) {
3911  // Only TrueBB was found.
3912  Builder.CreateBr(TrueBB);
3913  } else {
3914  // Only FalseBB was found.
3915  Builder.CreateBr(FalseBB);
3916  }
3917  }
3918 
3919  EraseTerminatorAndDCECond(OldTerm);
3920 
3921  if (DTU) {
3923  Updates.reserve(RemovedSuccessors.size());
3924  for (auto *RemovedSuccessor : RemovedSuccessors)
3925  Updates.push_back({DominatorTree::Delete, BB, RemovedSuccessor});
3926  DTU->applyUpdates(Updates);
3927  }
3928 
3929  return true;
3930 }
3931 
3932 // Replaces
3933 // (switch (select cond, X, Y)) on constant X, Y
3934 // with a branch - conditional if X and Y lead to distinct BBs,
3935 // unconditional otherwise.
3936 bool SimplifyCFGOpt::SimplifySwitchOnSelect(SwitchInst *SI,
3937  SelectInst *Select) {
3938  // Check for constant integer values in the select.
3939  ConstantInt *TrueVal = dyn_cast<ConstantInt>(Select->getTrueValue());
3940  ConstantInt *FalseVal = dyn_cast<ConstantInt>(Select->getFalseValue());
3941  if (!TrueVal || !FalseVal)
3942  return false;
3943 
3944  // Find the relevant condition and destinations.
3945  Value *Condition = Select->getCondition();
3946  BasicBlock *TrueBB = SI->findCaseValue(TrueVal)->getCaseSuccessor();
3947  BasicBlock *FalseBB = SI->findCaseValue(FalseVal)->getCaseSuccessor();
3948 
3949  // Get weight for TrueBB and FalseBB.
3950  uint32_t TrueWeight = 0, FalseWeight = 0;
3951  SmallVector<uint64_t, 8> Weights;
3952  bool HasWeights = HasBranchWeights(SI);
3953  if (HasWeights) {
3954  GetBranchWeights(SI, Weights);
3955  if (Weights.size() == 1 + SI->getNumCases()) {
3956  TrueWeight =
3957  (uint32_t)Weights[SI->findCaseValue(TrueVal)->getSuccessorIndex()];
3958  FalseWeight =
3959  (uint32_t)Weights[SI->findCaseValue(FalseVal)->getSuccessorIndex()];
3960  }
3961  }
3962 
3963  // Perform the actual simplification.
3964  return SimplifyTerminatorOnSelect(SI, Condition, TrueBB, FalseBB, TrueWeight,
3965  FalseWeight);
3966 }
3967 
3968 // Replaces
3969 // (indirectbr (select cond, blockaddress(@fn, BlockA),
3970 // blockaddress(@fn, BlockB)))
3971 // with
3972 // (br cond, BlockA, BlockB).
3973 bool SimplifyCFGOpt::SimplifyIndirectBrOnSelect(IndirectBrInst *IBI,
3974  SelectInst *SI) {
3975  // Check that both operands of the select are block addresses.
3976  BlockAddress *TBA = dyn_cast<BlockAddress>(SI->getTrueValue());
3977  BlockAddress *FBA = dyn_cast<BlockAddress>(SI->getFalseValue());
3978  if (!TBA || !FBA)
3979  return false;
3980 
3981  // Extract the actual blocks.
3982  BasicBlock *TrueBB = TBA->getBasicBlock();
3983  BasicBlock *FalseBB = FBA->getBasicBlock();
3984 
3985  // Perform the actual simplification.
3986  return SimplifyTerminatorOnSelect(IBI, SI->getCondition(), TrueBB, FalseBB, 0,
3987  0);
3988 }
3989 
3990 /// This is called when we find an icmp instruction
3991 /// (a seteq/setne with a constant) as the only instruction in a
3992 /// block that ends with an uncond branch. We are looking for a very specific
3993 /// pattern that occurs when "A == 1 || A == 2 || A == 3" gets simplified. In
3994 /// this case, we merge the first two "or's of icmp" into a switch, but then the
3995 /// default value goes to an uncond block with a seteq in it, we get something
3996 /// like:
3997 ///
3998 /// switch i8 %A, label %DEFAULT [ i8 1, label %end i8 2, label %end ]
3999 /// DEFAULT:
4000 /// %tmp = icmp eq i8 %A, 92
4001 /// br label %end
4002 /// end:
4003 /// ... = phi i1 [ true, %entry ], [ %tmp, %DEFAULT ], [ true, %entry ]
4004 ///
4005 /// We prefer to split the edge to 'end' so that there is a true/false entry to
4006 /// the PHI, merging the third icmp into the switch.
4007 bool SimplifyCFGOpt::tryToSimplifyUncondBranchWithICmpInIt(
4008  ICmpInst *ICI, IRBuilder<> &Builder) {
4009  BasicBlock *BB = ICI->getParent();
4010 
4011  // If the block has any PHIs in it or the icmp has multiple uses, it is too
4012  // complex.
4013  if (isa<PHINode>(BB->begin()) || !ICI->hasOneUse())
4014  return false;
4015 
4016  Value *V = ICI->getOperand(0);
4017  ConstantInt *Cst = cast<ConstantInt>(ICI->getOperand(1));
4018 
4019  // The pattern we're looking for is where our only predecessor is a switch on
4020  // 'V' and this block is the default case for the switch. In this case we can
4021  // fold the compared value into the switch to simplify things.
4022  BasicBlock *Pred = BB->getSinglePredecessor();
4023  if (!Pred || !isa<SwitchInst>(Pred->getTerminator()))
4024  return false;
4025 
4026  SwitchInst *SI = cast<SwitchInst>(Pred->getTerminator());
4027  if (SI->getCondition() != V)
4028  return false;
4029 
4030  // If BB is reachable on a non-default case, then we simply know the value of
4031  // V in this block. Substitute it and constant fold the icmp instruction
4032  // away.
4033  if (SI->getDefaultDest() != BB) {
4034  ConstantInt *VVal = SI->findCaseDest(BB);
4035  assert(VVal && "Should have a unique destination value");
4036  ICI->setOperand(0, VVal);
4037 
4038  if (Value *V = SimplifyInstruction(ICI, {DL, ICI})) {
4039  ICI->replaceAllUsesWith(V);
4040  ICI->eraseFromParent();
4041  }
4042  // BB is now empty, so it is likely to simplify away.
4043  return requestResimplify();
4044  }
4045 
4046  // Ok, the block is reachable from the default dest. If the constant we're
4047  // comparing exists in one of the other edges, then we can constant fold ICI
4048  // and zap it.
4049  if (SI->findCaseValue(Cst) != SI->case_default()) {
4050  Value *V;
4051  if (ICI->getPredicate() == ICmpInst::ICMP_EQ)
4052  V = ConstantInt::getFalse(BB->getContext());
4053  else
4054  V = ConstantInt::getTrue(BB->getContext());
4055 
4056  ICI->replaceAllUsesWith(V);
4057  ICI->eraseFromParent();
4058  // BB is now empty, so it is likely to simplify away.
4059  return requestResimplify();
4060  }
4061 
4062  // The use of the icmp has to be in the 'end' block, by the only PHI node in
4063  // the block.
4064  BasicBlock *SuccBlock = BB->getTerminator()->getSuccessor(0);
4065  PHINode *PHIUse = dyn_cast<PHINode>(ICI->user_back());
4066  if (PHIUse == nullptr || PHIUse != &SuccBlock->front() ||
4067  isa<PHINode>(++BasicBlock::iterator(PHIUse)))
4068  return false;
4069 
4070  // If the icmp is a SETEQ, then the default dest gets false, the new edge gets
4071  // true in the PHI.
4072  Constant *DefaultCst = ConstantInt::getTrue(BB->getContext());
4073  Constant *NewCst = ConstantInt::getFalse(BB->getContext());
4074 
4075  if (ICI->getPredicate() == ICmpInst::ICMP_EQ)
4076  std::swap(DefaultCst, NewCst);
4077 
4078  // Replace ICI (which is used by the PHI for the default value) with true or
4079  // false depending on if it is EQ or NE.
4080  ICI->replaceAllUsesWith(DefaultCst);
4081  ICI->eraseFromParent();
4082 
4084 
4085  // Okay, the switch goes to this block on a default value. Add an edge from
4086  // the switch to the merge point on the compared value.
4087  BasicBlock *NewBB =
4088  BasicBlock::Create(BB->getContext(), "switch.edge", BB->getParent(), BB);
4089  {
4091  auto W0 = SIW.getSuccessorWeight(0);
4093  if (W0) {
4094  NewW = ((uint64_t(*W0) + 1) >> 1);
4095  SIW.setSuccessorWeight(0, *NewW);
4096  }
4097  SIW.addCase(Cst, NewBB, NewW);
4098  if (DTU)
4099  Updates.push_back({DominatorTree::Insert, Pred, NewBB});
4100  }
4101 
4102  // NewBB branches to the phi block, add the uncond branch and the phi entry.
4103  Builder.SetInsertPoint(NewBB);
4104  Builder.SetCurrentDebugLocation(SI->getDebugLoc());
4105  Builder.CreateBr(SuccBlock);
4106  PHIUse->addIncoming(NewCst, NewBB);
4107  if (DTU) {
4108  Updates.push_back({DominatorTree::Insert, NewBB, SuccBlock});
4109  DTU->applyUpdates(Updates);
4110  }
4111  return true;
4112 }
4113 
4114 /// The specified branch is a conditional branch.
4115 /// Check to see if it is branching on an or/and chain of icmp instructions, and
4116 /// fold it into a switch instruction if so.
4117 bool SimplifyCFGOpt::SimplifyBranchOnICmpChain(BranchInst *BI,
4119  const DataLayout &DL) {
4120  Instruction *Cond = dyn_cast<Instruction>(BI->getCondition());
4121  if (!Cond)
4122  return false;
4123 
4124  // Change br (X == 0 | X == 1), T, F into a switch instruction.
4125  // If this is a bunch of seteq's or'd together, or if it's a bunch of
4126  // 'setne's and'ed together, collect them.
4127 
4128  // Try to gather values from a chain of and/or to be turned into a switch
4129  ConstantComparesGatherer ConstantCompare(Cond, DL);
4130  // Unpack the result
4131  SmallVectorImpl<ConstantInt *> &Values = ConstantCompare.Vals;
4132  Value *CompVal = ConstantCompare.CompValue;
4133  unsigned UsedICmps = ConstantCompare.UsedICmps;
4134  Value *ExtraCase = ConstantCompare.Extra;
4135 
4136  // If we didn't have a multiply compared value, fail.
4137  if (!CompVal)
4138  return false;
4139 
4140  // Avoid turning single icmps into a switch.
4141  if (UsedICmps <= 1)
4142  return false;
4143 
4144  bool TrueWhenEqual = match(Cond, m_LogicalOr(m_Value(), m_Value()));
4145 
4146  // There might be duplicate constants in the list, which the switch
4147  // instruction can't handle, remove them now.
4148  array_pod_sort(Values.begin(), Values.end(), ConstantIntSortPredicate);
4149  Values.erase(std::unique(Values.begin(), Values.end()), Values.end());
4150 
4151  // If Extra was used, we require at least two switch values to do the
4152  // transformation. A switch with one value is just a conditional branch.
4153  if (ExtraCase && Values.size() < 2)
4154  return false;
4155 
4156  // TODO: Preserve branch weight metadata, similarly to how
4157  // FoldValueComparisonIntoPredecessors preserves it.
4158 
4159  // Figure out which block is which destination.
4160  BasicBlock *DefaultBB = BI->getSuccessor(1);
4161  BasicBlock *EdgeBB = BI->getSuccessor(0);
4162  if (!TrueWhenEqual)
4163  std::swap(DefaultBB, EdgeBB);
4164 
4165  BasicBlock *BB = BI->getParent();
4166 
4167  LLVM_DEBUG(dbgs() << "Converting 'icmp' chain with " << Values.size()
4168  << " cases into SWITCH. BB is:\n"
4169  << *BB);
4170 
4172 
4173  // If there are any extra values that couldn't be folded into the switch
4174  // then we evaluate them with an explicit branch first. Split the block
4175  // right before the condbr to handle it.
4176  if (ExtraCase) {
4177  BasicBlock *NewBB = SplitBlock(BB, BI, DTU, /*LI=*/nullptr,
4178  /*MSSAU=*/nullptr, "switch.early.test");
4179 
4180  // Remove the uncond branch added to the old block.
4181  Instruction *OldTI = BB->getTerminator();
4182  Builder.SetInsertPoint(OldTI);
4183 
4184  // There can be an unintended UB if extra values are Poison. Before the
4185  // transformation, extra values may not be evaluated according to the
4186  // condition, and it will not raise UB. But after transformation, we are
4187  // evaluating extra values before checking the condition, and it will raise
4188  // UB. It can be solved by adding freeze instruction to extra values.
4189  AssumptionCache *AC = Options.AC;
4190 
4191  if (!isGuaranteedNotToBeUndefOrPoison(ExtraCase, AC, BI, nullptr))
4192  ExtraCase = Builder.CreateFreeze(ExtraCase);
4193 
4194  if (TrueWhenEqual)
4195  Builder.CreateCondBr(ExtraCase, EdgeBB, NewBB);
4196  else
4197  Builder.CreateCondBr(ExtraCase, NewBB, EdgeBB);
4198 
4199  OldTI->eraseFromParent();
4200 
4201  if (DTU)
4202  Updates.push_back({DominatorTree::Insert, BB, EdgeBB});
4203 
4204  // If there are PHI nodes in EdgeBB, then we need to add a new entry to them
4205  // for the edge we just added.
4206  AddPredecessorToBlock(EdgeBB, BB, NewBB);
4207 
4208  LLVM_DEBUG(dbgs() << " ** 'icmp' chain unhandled condition: " << *ExtraCase
4209  << "\nEXTRABB = " << *BB);
4210  BB = NewBB;
4211  }
4212 
4213  Builder.SetInsertPoint(BI);
4214  // Convert pointer to int before we switch.
4215  if (CompVal->getType()->isPointerTy()) {
4216  CompVal = Builder.CreatePtrToInt(
4217  CompVal, DL.getIntPtrType(CompVal->getType()), "magicptr");
4218  }
4219 
4220  // Create the new switch instruction now.
4221  SwitchInst *New = Builder.CreateSwitch(CompVal, DefaultBB, Values.size());
4222 
4223  // Add all of the 'cases' to the switch instruction.
4224  for (unsigned i = 0, e = Values.size(); i != e; ++i)
4225  New->addCase(Values[i], EdgeBB);
4226 
4227  // We added edges from PI to the EdgeBB. As such, if there were any
4228  // PHI nodes in EdgeBB, they need entries to be added corresponding to
4229  // the number of edges added.
4230  for (BasicBlock::iterator BBI = EdgeBB->begin(); isa<PHINode>(BBI); ++BBI) {
4231  PHINode *PN = cast<PHINode>(BBI);
4232  Value *InVal = PN->getIncomingValueForBlock(BB);
4233  for (unsigned i = 0, e = Values.size() - 1; i != e; ++i)
4234  PN->addIncoming(InVal, BB);
4235  }
4236 
4237  // Erase the old branch instruction.
4239  if (DTU)
4240  DTU->applyUpdates(Updates);
4241 
4242  LLVM_DEBUG(dbgs() << " ** 'icmp' chain result is:\n" << *BB << '\n');
4243  return true;
4244 }
4245 
4246 bool SimplifyCFGOpt::simplifyResume(ResumeInst *RI, IRBuilder<> &Builder) {
4247  if (isa<PHINode>(RI->getValue()))
4248  return simplifyCommonResume(RI);
4249  else if (isa<LandingPadInst>(RI->getParent()->getFirstNonPHI()) &&
4250  RI->getValue() == RI->getParent()->getFirstNonPHI())
4251  // The resume must unwind the exception that caused control to branch here.
4252  return simplifySingleResume(RI);
4253 
4254  return false;
4255 }
4256 
4257 // Check if cleanup block is empty
4259  for (Instruction &I : R) {
4260  auto *II = dyn_cast<IntrinsicInst>(&I);
4261  if (!II)
4262  return false;
4263 
4264  Intrinsic::ID IntrinsicID = II->getIntrinsicID();
4265  switch (IntrinsicID) {
4266  case Intrinsic::dbg_declare:
4267  case Intrinsic::dbg_value:
4268  case Intrinsic::dbg_label:
4269  case Intrinsic::lifetime_end:
4270  break;
4271  default:
4272  return false;
4273  }
4274  }
4275  return true;
4276 }
4277 
4278 // Simplify resume that is shared by several landing pads (phi of landing pad).
4279 bool SimplifyCFGOpt::simplifyCommonResume(ResumeInst *RI) {
4280  BasicBlock *BB = RI->getParent();
4281 
4282  // Check that there are no other instructions except for debug and lifetime
4283  // intrinsics between the phi's and resume instruction.
4284  if (!isCleanupBlockEmpty(
4285  make_range(RI->getParent()->getFirstNonPHI(), BB->getTerminator())))
4286  return false;
4287 
4288  SmallSetVector<BasicBlock *, 4> TrivialUnwindBlocks;
4289  auto *PhiLPInst = cast<PHINode>(RI->getValue());
4290 
4291  // Check incoming blocks to see if any of them are trivial.
4292  for (unsigned Idx = 0, End = PhiLPInst->getNumIncomingValues(); Idx != End;
4293  Idx++) {
4294  auto *IncomingBB = PhiLPInst->getIncomingBlock(Idx);
4295  auto *IncomingValue = PhiLPInst->getIncomingValue(Idx);
4296 
4297  // If the block has other successors, we can not delete it because
4298  // it has other dependents.
4299  if (IncomingBB->getUniqueSuccessor() != BB)
4300  continue;
4301 
4302  auto *LandingPad = dyn_cast<LandingPadInst>(IncomingBB->getFirstNonPHI());
4303  // Not the landing pad that caused the control to branch here.
4304  if (IncomingValue != LandingPad)
4305  continue;
4306 
4307  if (isCleanupBlockEmpty(
4308  make_range(LandingPad->getNextNode(), IncomingBB->getTerminator())))
4309  TrivialUnwindBlocks.insert(IncomingBB);
4310  }
4311 
4312  // If no trivial unwind blocks, don't do any simplifications.
4313  if (TrivialUnwindBlocks.empty())
4314  return false;
4315 
4316  // Turn all invokes that unwind here into calls.
4317  for (auto *TrivialBB : TrivialUnwindBlocks) {
4318  // Blocks that will be simplified should be removed from the phi node.
4319  // Note there could be multiple edges to the resume block, and we need
4320  // to remove them all.
4321  while (PhiLPInst->getBasicBlockIndex(TrivialBB) != -1)
4322  BB->removePredecessor(TrivialBB, true);
4323 
4324  for (BasicBlock *Pred :
4326  removeUnwindEdge(Pred, DTU);
4327  ++NumInvokes;
4328  }
4329 
4330  // In each SimplifyCFG run, only the current processed block can be erased.
4331  // Otherwise, it will break the iteration of SimplifyCFG pass. So instead
4332  // of erasing TrivialBB, we only remove the branch to the common resume
4333  // block so that we can later erase the resume block since it has no
4334  // predecessors.
4335  TrivialBB->getTerminator()->eraseFromParent();
4336  new UnreachableInst(RI->getContext(), TrivialBB);
4337  if (DTU)
4338  DTU->applyUpdates({{DominatorTree::Delete, TrivialBB, BB}});
4339  }
4340 
4341  // Delete the resume block if all its predecessors have been removed.
4342  if (pred_empty(BB))
4343  DeleteDeadBlock(BB, DTU);
4344 
4345  return !TrivialUnwindBlocks.empty();
4346 }
4347 
4348 // Simplify resume that is only used by a single (non-phi) landing pad.
4349 bool SimplifyCFGOpt::simplifySingleResume(ResumeInst *RI) {
4350  BasicBlock *BB = RI->getParent();
4351  auto *LPInst = cast<LandingPadInst>(BB->getFirstNonPHI());
4352  assert(RI->getValue() == LPInst &&
4353  "Resume must unwind the exception that caused control to here");
4354 
4355  // Check that there are no other instructions except for debug intrinsics.
4356  if (!isCleanupBlockEmpty(
4357  make_range<Instruction *>(LPInst->getNextNode(), RI)))
4358  return false;
4359 
4360  // Turn all invokes that unwind here into calls and delete the basic block.
4362  removeUnwindEdge(Pred, DTU);
4363  ++NumInvokes;
4364  }
4365 
4366  // The landingpad is now unreachable. Zap it.
4367  DeleteDeadBlock(BB, DTU);
4368  return true;
4369 }
4370 
4372  // If this is a trivial cleanup pad that executes no instructions, it can be
4373  // eliminated. If the cleanup pad continues to the caller, any predecessor
4374  // that is an EH pad will be updated to continue to the caller and any
4375  // predecessor that terminates with an invoke instruction will have its invoke
4376  // instruction converted to a call instruction. If the cleanup pad being
4377  // simplified does not continue to the caller, each predecessor will be
4378  // updated to continue to the unwind destination of the cleanup pad being
4379  // simplified.
4380  BasicBlock *BB = RI->getParent();
4381  CleanupPadInst *CPInst = RI->getCleanupPad();
4382  if (CPInst->getParent() != BB)
4383  // This isn't an empty cleanup.
4384  return false;
4385 
4386  // We cannot kill the pad if it has multiple uses. This typically arises
4387  // from unreachable basic blocks.
4388  if (!CPInst->hasOneUse())
4389  return false;
4390 
4391  // Check that there are no other instructions except for benign intrinsics.
4392  if (!isCleanupBlockEmpty(
4393  make_range<Instruction *>(CPInst->getNextNode(), RI)))
4394  return false;
4395 
4396  // If the cleanup return we are simplifying unwinds to the caller, this will
4397  // set UnwindDest to nullptr.
4398  BasicBlock *UnwindDest = RI->getUnwindDest();
4399  Instruction *DestEHPad = UnwindDest ? UnwindDest->getFirstNonPHI() : nullptr;
4400 
4401  // We're about to remove BB from the control flow. Before we do, sink any
4402  // PHINodes into the unwind destination. Doing this before changing the
4403  // control flow avoids some potentially slow checks, since we can currently
4404  // be certain that UnwindDest and BB have no common predecessors (since they
4405  // are both EH pads).
4406  if (UnwindDest) {
4407  // First, go through the PHI nodes in UnwindDest and update any nodes that
4408  // reference the block we are removing
4409  for (PHINode &DestPN : UnwindDest->phis()) {
4410  int Idx = DestPN.getBasicBlockIndex(BB);
4411  // Since BB unwinds to UnwindDest, it has to be in the PHI node.
4412  assert(Idx != -1);
4413  // This PHI node has an incoming value that corresponds to a control
4414  // path through the cleanup pad we are removing. If the incoming
4415  // value is in the cleanup pad, it must be a PHINode (because we
4416  // verified above that the block is otherwise empty). Otherwise, the
4417  // value is either a constant or a value that dominates the cleanup
4418  // pad being removed.
4419  //
4420  // Because BB and UnwindDest are both EH pads, all of their
4421  // predecessors must unwind to these blocks, and since no instruction
4422  // can have multiple unwind destinations, there will be no overlap in
4423  // incoming blocks between SrcPN and DestPN.
4424  Value *SrcVal = DestPN.getIncomingValue(Idx);
4425  PHINode *SrcPN = dyn_cast<PHINode>(SrcVal);
4426 
4427  bool NeedPHITranslation = SrcPN && SrcPN->getParent() == BB;
4428  for (auto *Pred : predecessors(BB)) {
4429  Value *Incoming =
4430  NeedPHITranslation ? SrcPN->getIncomingValueForBlock(Pred) : SrcVal;
4431  DestPN.addIncoming(Incoming, Pred);
4432  }
4433  }
4434 
4435  // Sink any remaining PHI nodes directly into UnwindDest.
4436  Instruction *InsertPt = DestEHPad;
4437  for (PHINode &PN : make_early_inc_range(BB->phis())) {
4438  if (PN.use_empty() || !PN.isUsedOutsideOfBlock(BB))
4439  // If the PHI node has no uses or all of its uses are in this basic
4440  // block (meaning they are debug or lifetime intrinsics), just leave
4441  // it. It will be erased when we erase BB below.
4442  continue;
4443 
4444  // Otherwise, sink this PHI node into UnwindDest.
4445  // Any predecessors to UnwindDest which are not already represented
4446  // must be back edges which inherit the value from the path through
4447  // BB. In this case, the PHI value must reference itself.
4448  for (auto *pred : predecessors(UnwindDest))
4449  if (pred != BB)
4450  PN.addIncoming(&PN, pred);
4451  PN.moveBefore(InsertPt);
4452  // Also, add a dummy incoming value for the original BB itself,
4453  // so that the PHI is well-formed until we drop said predecessor.
4455  }
4456  }
4457 
4458  std::vector<DominatorTree::UpdateType> Updates;
4459 
4460  // We use make_early_inc_range here because we will remove all predecessors.
4462  if (UnwindDest == nullptr) {
4463  if (DTU) {
4464  DTU->applyUpdates(Updates);
4465  Updates.clear();
4466  }
4467  removeUnwindEdge(PredBB, DTU);
4468  ++NumInvokes;
4469  } else {
4470  BB->removePredecessor(PredBB);
4471  Instruction *TI = PredBB->getTerminator();
4472  TI->replaceUsesOfWith(BB, UnwindDest);
4473  if (DTU) {
4474  Updates.push_back({DominatorTree::Insert, PredBB, UnwindDest});
4475  Updates.push_back({DominatorTree::Delete, PredBB, BB});
4476  }
4477  }
4478  }
4479 
4480  if (DTU)
4481  DTU->applyUpdates(Updates);
4482 
4483  DeleteDeadBlock(BB, DTU);
4484 
4485  return true;
4486 }
4487 
4488 // Try to merge two cleanuppads together.
4490  // Skip any cleanuprets which unwind to caller, there is nothing to merge
4491  // with.
4492  BasicBlock *UnwindDest = RI->getUnwindDest();
4493  if (!UnwindDest)
4494  return false;
4495 
4496  // This cleanupret isn't the only predecessor of this cleanuppad, it wouldn't
4497  // be safe to merge without code duplication.
4498  if (UnwindDest->getSinglePredecessor() != RI->getParent())
4499  return false;
4500 
4501  // Verify that our cleanuppad's unwind destination is another cleanuppad.
4502  auto *SuccessorCleanupPad = dyn_cast<CleanupPadInst>(&UnwindDest->front());
4503  if (!SuccessorCleanupPad)
4504  return false;
4505 
4506  CleanupPadInst *PredecessorCleanupPad = RI->getCleanupPad();
4507  // Replace any uses of the successor cleanupad with the predecessor pad
4508  // The only cleanuppad uses should be this cleanupret, it's cleanupret and
4509  // funclet bundle operands.
4510  SuccessorCleanupPad->replaceAllUsesWith(PredecessorCleanupPad);
4511  // Remove the old cleanuppad.
4512  SuccessorCleanupPad->eraseFromParent();
4513  // Now, we simply replace the cleanupret with a branch to the unwind
4514  // destination.
4515  BranchInst::Create(UnwindDest, RI->getParent());
4516  RI->eraseFromParent();
4517 
4518  return true;
4519 }
4520 
4521 bool SimplifyCFGOpt::simplifyCleanupReturn(CleanupReturnInst *RI) {
4522  // It is possible to transiantly have an undef cleanuppad operand because we
4523  // have deleted some, but not all, dead blocks.
4524  // Eventually, this block will be deleted.
4525  if (isa<UndefValue>(RI->getOperand(0)))
4526  return false;
4527 
4528  if (mergeCleanupPad(RI))
4529  return true;
4530 
4531  if (removeEmptyCleanup(RI, DTU))
4532  return true;
4533 
4534  return false;
4535 }
4536 
4537 // WARNING: keep in sync with InstCombinerImpl::visitUnreachableInst()!
4538 bool SimplifyCFGOpt::simplifyUnreachable(UnreachableInst *UI) {
4539  BasicBlock *BB = UI->getParent();
4540 
4541  bool Changed = false;
4542 
4543  // If there are any instructions immediately before the unreachable that can
4544  // be removed, do so.
4545  while (UI->getIterator() != BB->begin()) {
4546  BasicBlock::iterator BBI = UI->getIterator();
4547  --BBI;
4548 
4550  break; // Can not drop any more instructions. We're done here.
4551  // Otherwise, this instruction can be freely erased,
4552  // even if it is not side-effect free.
4553 
4554  // Note that deleting EH's here is in fact okay, although it involves a bit
4555  // of subtle reasoning. If this inst is an EH, all the predecessors of this
4556  // block will be the unwind edges of Invoke/CatchSwitch/CleanupReturn,
4557  // and we can therefore guarantee this block will be erased.
4558 
4559  // Delete this instruction (any uses are guaranteed to be dead)
4560  BBI->replaceAllUsesWith(PoisonValue::get(BBI->getType()));
4561  BBI->eraseFromParent();
4562  Changed = true;
4563  }
4564 
4565  // If the unreachable instruction is the first in the block, take a gander
4566  // at all of the predecessors of this instruction, and simplify them.
4567  if (&BB->front() != UI)
4568  return Changed;
4569 
4570  std::vector<DominatorTree::UpdateType> Updates;
4571 
4573  for (unsigned i = 0, e = Preds.size(); i != e; ++i) {
4574  auto *Predecessor = Preds[i];
4575  Instruction *TI = Predecessor->getTerminator();
4576  IRBuilder<> Builder(TI);
4577  if (auto *BI = dyn_cast<BranchInst>(TI)) {
4578  // We could either have a proper unconditional branch,
4579  // or a degenerate conditional branch with matching destinations.
4580  if (all_of(BI->successors(),
4581  [BB](auto *Successor) { return Successor == BB; })) {
4582  new UnreachableInst(TI->getContext(), TI);
4583  TI->eraseFromParent();
4584  Changed = true;
4585  } else {
4586  assert(BI->isConditional() && "Can't get here with an uncond branch.");
4587  Value* Cond = BI->getCondition();
4588  assert(BI->getSuccessor(0) != BI->getSuccessor(1) &&
4589  "The destinations are guaranteed to be different here.");
4590  if (BI->getSuccessor(0) == BB) {
4591  Builder.CreateAssumption(Builder.CreateNot(Cond));
4592  Builder.CreateBr(BI->getSuccessor(1));
4593  } else {
4594  assert(BI->getSuccessor(1) == BB && "Incorrect CFG");
4595  Builder.CreateAssumption(Cond);
4596  Builder.CreateBr(BI->getSuccessor(0));
4597  }
4599  Changed = true;
4600  }
4601  if (DTU)
4602  Updates.push_back({DominatorTree::Delete, Predecessor, BB});
4603  } else if (auto *SI = dyn_cast<SwitchInst>(TI)) {
4605  for (auto i = SU->case_begin(), e = SU->case_end(); i != e;) {
4606  if (i->getCaseSuccessor() != BB) {
4607  ++i;
4608  continue;
4609  }
4610  BB->removePredecessor(SU->getParent());
4611  i = SU.removeCase(i);
4612  e = SU->case_end();
4613  Changed = true;
4614  }
4615  // Note that the default destination can't be removed!
4616  if (DTU && SI->getDefaultDest() != BB)
4617  Updates.push_back({DominatorTree::Delete, Predecessor, BB});
4618  } else if (auto *II = dyn_cast<InvokeInst>(TI)) {
4619  if (II->getUnwindDest() == BB) {
4620  if (DTU) {
4621  DTU->applyUpdates(Updates);
4622  Updates.clear();
4623  }
4624  removeUnwindEdge(TI->getParent(), DTU);
4625  Changed = true;
4626  }
4627  } else if (auto *CSI = dyn_cast<CatchSwitchInst>(TI)) {
4628  if (CSI->getUnwindDest() == BB) {
4629  if (DTU) {
4630  DTU->applyUpdates(Updates);
4631  Updates.clear();
4632  }
4633  removeUnwindEdge(TI->getParent(), DTU);
4634  Changed = true;
4635  continue;
4636  }
4637 
4638  for (CatchSwitchInst::handler_iterator I = CSI->handler_begin(),
4639  E = CSI->handler_end();
4640  I != E; ++I) {
4641  if (*I == BB) {
4642  CSI->removeHandler(I);
4643  --I;
4644  --E;
4645  Changed = true;
4646  }
4647  }
4648  if (DTU)
4649  Updates.push_back({DominatorTree::Delete, Predecessor, BB});
4650  if (CSI->getNumHandlers() == 0) {
4651  if (CSI->hasUnwindDest()) {
4652  // Redirect all predecessors of the block containing CatchSwitchInst
4653  // to instead branch to the CatchSwitchInst's unwind destination.
4654  if (DTU) {
4655  for (auto *PredecessorOfPredecessor : predecessors(Predecessor)) {
4656  Updates.push_back({DominatorTree::Insert,
4657  PredecessorOfPredecessor,
4658  CSI->getUnwindDest()});
4659  Updates.push_back({DominatorTree::Delete,
4660  PredecessorOfPredecessor, Predecessor});
4661  }
4662  }
4663  Predecessor->replaceAllUsesWith(CSI->getUnwindDest());
4664  } else {
4665  // Rewrite all preds to unwind to caller (or from invoke to call).
4666  if (DTU) {
4667  DTU->applyUpdates(Updates);
4668  Updates.clear();
4669  }
4670  SmallVector<BasicBlock *, 8> EHPreds(predecessors(Predecessor));
4671  for (BasicBlock *EHPred : EHPreds)
4672  removeUnwindEdge(EHPred, DTU);
4673  }
4674  // The catchswitch is no longer reachable.
4675  new UnreachableInst(CSI->getContext(), CSI);
4676  CSI->eraseFromParent();
4677  Changed = true;
4678  }
4679  } else if (auto *CRI = dyn_cast<CleanupReturnInst>(TI)) {
4680  (void)CRI;
4681  assert(CRI->hasUnwindDest() && CRI->getUnwindDest() == BB &&
4682  "Expected to always have an unwind to BB.");
4683  if (DTU)
4684  Updates.push_back({DominatorTree::Delete, Predecessor, BB});
4685  new UnreachableInst(TI->getContext(), TI);
4686  TI->eraseFromParent();
4687  Changed = true;
4688  }
4689  }
4690 
4691  if (DTU)
4692  DTU->applyUpdates(Updates);
4693 
4694  // If this block is now dead, remove it.
4695  if (pred_empty(BB) && BB != &BB->getParent()->getEntryBlock()) {
4696  DeleteDeadBlock(BB, DTU);
4697  return true;
4698  }
4699 
4700  return Changed;
4701 }
4702 
4704  assert(Cases.size() >= 1);
4705 
4706  array_pod_sort(Cases.begin(), Cases.end(), ConstantIntSortPredicate);
4707  for (size_t I = 1, E = Cases.size(); I != E; ++I) {
4708  if (Cases[I - 1]->getValue() != Cases[I]->getValue() + 1)
4709  return false;
4710  }
4711  return true;
4712 }
4713 
4715  DomTreeUpdater *DTU) {
4716  LLVM_DEBUG(dbgs() << "SimplifyCFG: switch default is dead.\n");
4717  auto *BB = Switch->getParent();
4718  BasicBlock *NewDefaultBlock = SplitBlockPredecessors(
4719  Switch->getDefaultDest(), Switch->getParent(), "", DTU);
4720  auto *OrigDefaultBlock = Switch->getDefaultDest();
4721  Switch->setDefaultDest(&*NewDefaultBlock);
4722  if (DTU)
4723  DTU->applyUpdates({{DominatorTree::Insert, BB, &*NewDefaultBlock},
4724  {DominatorTree::Delete, BB, OrigDefaultBlock}});
4725  SplitBlock(&*NewDefaultBlock, &NewDefaultBlock->front(), DTU);
4727  if (DTU)
4728  for (auto *Successor : successors(NewDefaultBlock))
4729  Updates.push_back({DominatorTree::Delete, NewDefaultBlock, Successor});
4730  auto *NewTerminator = NewDefaultBlock->getTerminator();
4731  new UnreachableInst(Switch->getContext(), NewTerminator);
4732  EraseTerminatorAndDCECond(NewTerminator);
4733  if (DTU)
4734  DTU->applyUpdates(Updates);
4735 }
4736 
4737 /// Turn a switch with two reachable destinations into an integer range
4738 /// comparison and branch.
4739 bool SimplifyCFGOpt::TurnSwitchRangeIntoICmp(SwitchInst *SI,
4740  IRBuilder<> &Builder) {
4741  assert(SI->getNumCases() > 1 && "Degenerate switch?");
4742 
4743  bool HasDefault =
4744  !isa<UnreachableInst>(SI->getDefaultDest()->getFirstNonPHIOrDbg());
4745 
4746  auto *BB = SI->getParent();
4747 
4748  // Partition the cases into two sets with different destinations.
4749  BasicBlock *DestA = HasDefault ? SI->getDefaultDest() : nullptr;
4750  BasicBlock *DestB = nullptr;
4753 
4754  for (auto Case : SI->cases()) {
4755  BasicBlock *Dest = Case.getCaseSuccessor();
4756  if (!DestA)
4757  DestA = Dest;
4758  if (Dest == DestA) {
4759  CasesA.push_back(Case.getCaseValue());
4760  continue;
4761  }
4762  if (!DestB)
4763  DestB = Dest;
4764  if (Dest == DestB) {
4765  CasesB.push_back(Case.getCaseValue());
4766  continue;
4767  }
4768  return false; // More than two destinations.
4769  }
4770 
4771  assert(DestA && DestB &&
4772  "Single-destination switch should have been folded.");
4773  assert(DestA != DestB);
4774  assert(DestB != SI->getDefaultDest());
4775  assert(!CasesB.empty() && "There must be non-default cases.");
4776  assert(!CasesA.empty() || HasDefault);
4777 
4778  // Figure out if one of the sets of cases form a contiguous range.
4779  SmallVectorImpl<ConstantInt *> *ContiguousCases = nullptr;
4780  BasicBlock *ContiguousDest = nullptr;
4781  BasicBlock *OtherDest = nullptr;
4782  if (!CasesA.empty() && CasesAreContiguous(CasesA)) {
4783  ContiguousCases = &CasesA;
4784  ContiguousDest = DestA;
4785  OtherDest = DestB;
4786  } else if (CasesAreContiguous(CasesB)) {
4787  ContiguousCases = &CasesB;
4788  ContiguousDest = DestB;
4789  OtherDest = DestA;
4790  } else
4791  return false;
4792 
4793  // Start building the compare and branch.
4794 
4795  Constant *Offset = ConstantExpr::getNeg(ContiguousCases->back());
4796  Constant *NumCases =
4797  ConstantInt::get(Offset->getType(), ContiguousCases->size());
4798 
4799  Value *Sub = SI->getCondition();
4800  if (!Offset->isNullValue())
4801  Sub = Builder.CreateAdd(Sub, Offset, Sub->getName() + ".off");
4802 
4803  Value *Cmp;
4804  // If NumCases overflowed, then all possible values jump to the successor.
4805  if (NumCases->isNullValue() && !ContiguousCases->empty())
4806  Cmp = ConstantInt::getTrue(SI->getContext());
4807  else
4808  Cmp = Builder.CreateICmpULT(Sub, NumCases, "switch");
4809  BranchInst *NewBI = Builder.CreateCondBr(Cmp, ContiguousDest, OtherDest);
4810 
4811  // Update weight for the newly-created conditional branch.
4812  if (HasBranchWeights(SI)) {
4813  SmallVector<uint64_t, 8> Weights;
4814  GetBranchWeights(SI, Weights);
4815  if (Weights.size() == 1 + SI->getNumCases()) {
4816  uint64_t TrueWeight = 0;
4817  uint64_t FalseWeight = 0;
4818  for (size_t I = 0, E = Weights.size(); I != E; ++I) {
4819  if (SI->getSuccessor(I) == ContiguousDest)
4820  TrueWeight += Weights[I];
4821  else
4822  FalseWeight += Weights[I];
4823  }
4824  while (TrueWeight > UINT32_MAX || FalseWeight > UINT32_MAX) {
4825  TrueWeight /= 2;
4826  FalseWeight /= 2;
4827  }
4828  setBranchWeights(NewBI, TrueWeight, FalseWeight);
4829  }
4830  }
4831 
4832  // Prune obsolete incoming values off the successors' PHI nodes.
4833  for (auto BBI = ContiguousDest->begin(); isa<PHINode>(BBI); ++BBI) {
4834  unsigned PreviousEdges = ContiguousCases->size();
4835  if (ContiguousDest == SI->getDefaultDest())
4836  ++PreviousEdges;
4837  for (unsigned I = 0, E = PreviousEdges - 1; I != E; ++I)
4838  cast<PHINode>(BBI)->removeIncomingValue(SI->getParent());
4839  }
4840  for (auto BBI = OtherDest->begin(); isa<PHINode>(BBI); ++BBI) {
4841  unsigned PreviousEdges = SI->getNumCases() - ContiguousCases->size();
4842  if (OtherDest == SI->getDefaultDest())
4843  ++PreviousEdges;
4844  for (unsigned I = 0, E = PreviousEdges - 1; I != E; ++I)
4845  cast<PHINode>(BBI)->removeIncomingValue(SI->getParent());
4846  }
4847 
4848  // Clean up the default block - it may have phis or other instructions before
4849  // the unreachable terminator.
4850  if (!HasDefault)
4852 
4853  auto *UnreachableDefault = SI->getDefaultDest();
4854 
4855  // Drop the switch.
4856  SI->eraseFromParent();
4857 
4858  if (!HasDefault && DTU)
4859  DTU->applyUpdates({{DominatorTree::Delete, BB, UnreachableDefault}});
4860 
4861  return true;
4862 }
4863 
4864 /// Compute masked bits for the condition of a switch
4865 /// and use it to remove dead cases.
4867  AssumptionCache *AC,
4868  const DataLayout &DL) {
4869  Value *Cond = SI->getCondition();
4870  unsigned Bits = Cond->getType()->getIntegerBitWidth();
4871  KnownBits Known = computeKnownBits(Cond, DL, 0, AC, SI);
4872 
4873  // We can also eliminate cases by determining that their values are outside of
4874  // the limited range of the condition based on how many significant (non-sign)
4875  // bits are in the condition value.
4876  unsigned ExtraSignBits = ComputeNumSignBits(Cond, DL, 0, AC, SI) - 1;
4877  unsigned MaxSignificantBitsInCond = Bits - ExtraSignBits;
4878 
4879  // Gather dead cases.
4881  SmallDenseMap<BasicBlock *, int, 8> NumPerSuccessorCases;
4882  for (auto &Case : SI->cases()) {
4883  auto *Successor = Case.getCaseSuccessor();
4884  if (DTU)
4885  ++NumPerSuccessorCases[Successor];
4886  const APInt &CaseVal = Case.getCaseValue()->getValue();
4887  if (Known.Zero.intersects(CaseVal) || !Known.One.isSubsetOf(CaseVal) ||
4888  (CaseVal.getMinSignedBits() > MaxSignificantBitsInCond)) {
4889  DeadCases.push_back(Case.getCaseValue());
4890  if (DTU)
4891  --NumPerSuccessorCases[Successor];
4892  LLVM_DEBUG(dbgs() << "SimplifyCFG: switch case " << CaseVal
4893  << " is dead.\n");
4894  }
4895  }
4896 
4897  // If we can prove that the cases must cover all possible values, the
4898  // default destination becomes dead and we can remove it. If we know some
4899  // of the bits in the value, we can use that to more precisely compute the
4900  // number of possible unique case values.
4901  bool HasDefault =
4902  !isa<UnreachableInst>(SI->getDefaultDest()->getFirstNonPHIOrDbg());
4903  const unsigned NumUnknownBits =
4904  Bits - (Known.Zero | Known.One).countPopulation();
4905  assert(NumUnknownBits <= Bits);
4906  if (HasDefault && DeadCases.empty() &&
4907  NumUnknownBits < 64 /* avoid overflow */ &&
4908  SI->getNumCases() == (1ULL << NumUnknownBits)) {
4910  return true;
4911  }
4912 
4913  if (DeadCases.empty())
4914  return false;
4915 
4917  for (ConstantInt *DeadCase : DeadCases) {
4918  SwitchInst::CaseIt CaseI = SI->findCaseValue(DeadCase);
4919  assert(CaseI != SI->case_default() &&
4920  "Case was not found. Probably mistake in DeadCases forming.");
4921  // Prune unused values from PHI nodes.
4922  CaseI->getCaseSuccessor()->removePredecessor(SI->getParent());
4923  SIW.removeCase(CaseI);
4924  }
4925 
4926  if (DTU) {
4927  std::vector<DominatorTree::UpdateType> Updates;
4928  for (const std::pair<BasicBlock *, int> &I : NumPerSuccessorCases)
4929  if (I.second == 0)
4930  Updates.push_back({DominatorTree::Delete, SI->getParent(), I.first});
4931  DTU->applyUpdates(Updates);
4932  }
4933 
4934  return true;
4935 }
4936 
4937 /// If BB would be eligible for simplification by
4938 /// TryToSimplifyUncondBranchFromEmptyBlock (i.e. it is empty and terminated
4939 /// by an unconditional branch), look at the phi node for BB in the successor
4940 /// block and see if the incoming value is equal to CaseValue. If so, return
4941 /// the phi node, and set PhiIndex to BB's index in the phi node.
4943  BasicBlock *BB, int *PhiIndex) {
4944  if (BB->getFirstNonPHIOrDbg() != BB->getTerminator())
4945  return nullptr; // BB must be empty to be a candidate for simplification.
4946  if (!BB->getSinglePredecessor())
4947  return nullptr; // BB must be dominated by the switch.
4948 
4949  BranchInst *Branch = dyn_cast<BranchInst>(BB->getTerminator());
4950  if (!Branch || !Branch->isUnconditional())
4951  return nullptr; // Terminator must be unconditional branch.
4952 
4953  BasicBlock *Succ = Branch->getSuccessor(0);
4954 
4955  for (PHINode &PHI : Succ->phis()) {
4956  int Idx = PHI.getBasicBlockIndex(BB);
4957  assert(Idx >= 0 && "PHI has no entry for predecessor?");
4958 
4959  Value *InValue = PHI.getIncomingValue(Idx);
4960  if (InValue != CaseValue)
4961  continue;
4962 
4963  *PhiIndex = Idx;
4964  return &PHI;
4965  }
4966 
4967  return nullptr;
4968 }
4969 
4970 /// Try to forward the condition of a switch instruction to a phi node
4971 /// dominated by the switch, if that would mean that some of the destination
4972 /// blocks of the switch can be folded away. Return true if a change is made.
4974  using ForwardingNodesMap = DenseMap<PHINode *, SmallVector<int, 4>>;
4975 
4976  ForwardingNodesMap ForwardingNodes;
4977  BasicBlock *SwitchBlock = SI->getParent();
4978  bool Changed = false;
4979  for (auto &Case : SI->cases()) {
4980  ConstantInt *CaseValue = Case.getCaseValue();
4981  BasicBlock *CaseDest = Case.getCaseSuccessor();
4982 
4983  // Replace phi operands in successor blocks that are using the constant case
4984  // value rather than the switch condition variable:
4985  // switchbb:
4986  // switch i32 %x, label %default [
4987  // i32 17, label %succ
4988  // ...
4989  // succ:
4990  // %r = phi i32 ... [ 17, %switchbb ] ...
4991  // -->
4992  // %r = phi i32 ... [ %x, %switchbb ] ...
4993 
4994  for (PHINode &Phi : CaseDest->phis()) {
4995  // This only works if there is exactly 1 incoming edge from the switch to
4996  // a phi. If there is >1, that means multiple cases of the switch map to 1
4997  // value in the phi, and that phi value is not the switch condition. Thus,
4998  // this transform would not make sense (the phi would be invalid because
4999  // a phi can't have different incoming values from the same block).
5000  int SwitchBBIdx = Phi.getBasicBlockIndex(SwitchBlock);
5001  if (Phi.getIncomingValue(SwitchBBIdx) == CaseValue &&
5002  count(Phi.blocks(), SwitchBlock) == 1) {
5003  Phi.setIncomingValue(SwitchBBIdx, SI->getCondition());
5004  Changed = true;
5005  }
5006  }
5007 
5008  // Collect phi nodes that are indirectly using this switch's case constants.
5009  int PhiIdx;
5010  if (auto *Phi = FindPHIForConditionForwarding(CaseValue, CaseDest, &PhiIdx))
5011  ForwardingNodes[Phi].push_back(PhiIdx);
5012  }
5013 
5014  for (auto &ForwardingNode : ForwardingNodes) {
5015  PHINode *Phi = ForwardingNode.first;
5016  SmallVectorImpl<int> &Indexes = ForwardingNode.second;
5017  if (Indexes.size() < 2)
5018  continue;
5019 
5020  for (int Index : Indexes)
5021  Phi->setIncomingValue(Index, SI->getCondition());
5022  Changed = true;
5023  }
5024 
5025  return Changed;
5026 }
5027 
5028 /// Return true if the backend will be able to handle
5029 /// initializing an array of constants like C.
5031  if (C->isThreadDependent())
5032  return false;
5033  if (C->isDLLImportDependent())
5034  return false;
5035 
5036  if (!isa<ConstantFP>(C) && !isa<ConstantInt>(C) &&
5037  !isa<ConstantPointerNull>(C) && !isa<GlobalValue>(C) &&
5038  !isa<UndefValue>(C) && !isa<ConstantExpr>(C))
5039  return false;
5040 
5041  if (ConstantExpr *CE = dyn_cast<ConstantExpr>(C)) {
5042  if (!CE->isGEPWithNoNotionalOverIndexing())
5043  return false;
5044  if (!ValidLookupTableConstant(CE->getOperand(0), TTI))
5045  return false;
5046  }
5047 
5049  return false;
5050 
5051  return true;
5052 }
5053 
5054 /// If V is a Constant, return it. Otherwise, try to look up
5055 /// its constant value in ConstantPool, returning 0 if it's not there.
5056 static Constant *
5059  if (Constant *C = dyn_cast<Constant>(V))
5060  return C;
5061  return ConstantPool.lookup(V);
5062 }
5063 
5064 /// Try to fold instruction I into a constant. This works for
5065 /// simple instructions such as binary operations where both operands are
5066 /// constant or can be replaced by constants from the ConstantPool. Returns the
5067 /// resulting constant on success, 0 otherwise.
5068 static Constant *
5071  if (SelectInst *Select = dyn_cast<SelectInst>(I)) {
5072  Constant *A = LookupConstant(Select->getCondition(), ConstantPool);
5073  if (!A)
5074  return nullptr;
5075  if (A->isAllOnesValue())
5076  return LookupConstant(Select->getTrueValue(), ConstantPool);
5077  if (A->isNullValue())
5078  return LookupConstant(Select->getFalseValue(), ConstantPool);
5079  return nullptr;
5080  }
5081 
5083  for (unsigned N = 0, E = I->getNumOperands(); N != E; ++N) {
5084  if (Constant *A = LookupConstant(I->getOperand(N), ConstantPool))
5085  COps.push_back(A);
5086  else
5087  return nullptr;
5088  }
5089 
5090  if (CmpInst *Cmp = dyn_cast<CmpInst>(I)) {
5091  return ConstantFoldCompareInstOperands(Cmp->getPredicate(), COps[0],
5092  COps[1], DL);
5093  }
5094 
5095  return ConstantFoldInstOperands(I, COps, DL);
5096 }
5097 
5098 /// Try to determine the resulting constant values in phi nodes
5099 /// at the common destination basic block, *CommonDest, for one of the case
5100 /// destionations CaseDest corresponding to value CaseVal (0 for the default
5101 /// case), of a switch instruction SI.
5102 static bool
5104  BasicBlock **CommonDest,
5105  SmallVectorImpl<std::pair<PHINode *, Constant *>> &Res,
5106  const DataLayout &DL, const TargetTransformInfo &TTI) {
5107  // The block from which we enter the common destination.
5108  BasicBlock *Pred = SI->getParent();
5109 
5110  // If CaseDest is empty except for some side-effect free instructions through
5111  // which we can constant-propagate the CaseVal, continue to its successor.
5113  ConstantPool.insert(std::make_pair(SI->getCondition(), CaseVal));
5114  for (Instruction &I :CaseDest->instructionsWithoutDebug()) {
5115  if (I.isTerminator()) {
5116  // If the terminator is a simple branch, continue to the next block.
5117  if (I.getNumSuccessors() != 1 || I.isExceptionalTerminator())
5118  return false;
5119  Pred = CaseDest;
5120  CaseDest = I.getSuccessor(0);
5121  } else if (Constant *C = ConstantFold(&I, DL, ConstantPool)) {
5122  // Instruction is side-effect free and constant.
5123 
5124  // If the instruction has uses outside this block or a phi node slot for
5125  // the block, it is not safe to bypass the instruction since it would then
5126  // no longer dominate all its uses.
5127  for (auto &Use : I.uses()) {
5128  User *User = Use.getUser();
5129  if (Instruction *I = dyn_cast<Instruction>(User))
5130  if (I->getParent() == CaseDest)
5131  continue;
5132  if (PHINode *Phi = dyn_cast<PHINode>(User))
5133  if (Phi->getIncomingBlock(Use) == CaseDest)
5134  continue;
5135  return false;
5136  }
5137 
5138  ConstantPool.insert(std::make_pair(&I, C));
5139  } else {
5140  break;
5141  }
5142  }
5143 
5144  // If we did not have a CommonDest before, use the current one.
5145  if (!*CommonDest)
5146  *CommonDest = CaseDest;
5147  // If the destination isn't the common one, abort.
5148  if (CaseDest != *CommonDest)
5149  return false;
5150 
5151  // Get the values for this case from phi nodes in the destination block.
5152  for (PHINode &PHI : (*CommonDest)->phis()) {
5153  int Idx = PHI.getBasicBlockIndex(Pred);
5154  if (Idx == -1)
5155  continue;
5156 
5157  Constant *ConstVal =
5158  LookupConstant(PHI.getIncomingValue(Idx), ConstantPool);
5159  if (!ConstVal)
5160  return false;
5161 
5162  // Be conservative about which kinds of constants we support.
5163  if (!ValidLookupTableConstant(ConstVal, TTI))
5164  return false;
5165 
5166  Res.push_back(std::make_pair(&PHI, ConstVal));
5167  }
5168 
5169  return Res.size() > 0;
5170 }
5171 
5172 // Helper function used to add CaseVal to the list of cases that generate
5173 // Result. Returns the updated number of cases that generate this result.
5174 static uintptr_t MapCaseToResult(ConstantInt *CaseVal,
5175  SwitchCaseResultVectorTy &UniqueResults,
5176  Constant *Result) {
5177  for (auto &I : UniqueResults) {
5178  if (I.first == Result) {
5179  I.second.push_back(CaseVal);
5180  return I.second.size();
5181  }
5182  }
5183  UniqueResults.push_back(
5184  std::make_pair(Result, SmallVector<ConstantInt *, 4>(1, CaseVal)));
5185  return 1;
5186 }
5187 
5188 // Helper function that initializes a map containing
5189 // results for the PHI node of the common destination block for a switch
5190 // instruction. Returns false if multiple PHI nodes have been found or if
5191 // there is not a common destination block for the switch.
5192 static bool
5194  SwitchCaseResultVectorTy &UniqueResults,
5195  Constant *&DefaultResult, const DataLayout &DL,
5196  const TargetTransformInfo &TTI,
5197  uintptr_t MaxUniqueResults, uintptr_t MaxCasesPerResult) {
5198  for (auto &I : SI->cases()) {
5199  ConstantInt *CaseVal = I.getCaseValue();
5200 
5201  // Resulting value at phi nodes for this case value.
5202  SwitchCaseResultsTy Results;
5203  if (!GetCaseResults(SI, CaseVal, I.getCaseSuccessor(), &CommonDest, Results,
5204  DL, TTI))
5205  return false;
5206 
5207  // Only one value per case is permitted.
5208  if (Results.size() > 1)
5209  return false;
5210 
5211  // Add the case->result mapping to UniqueResults.
5212  const uintptr_t NumCasesForResult =
5213  MapCaseToResult(CaseVal, UniqueResults, Results.begin()->second);
5214 
5215  // Early out if there are too many cases for this result.
5216  if (NumCasesForResult > MaxCasesPerResult)
5217  return false;
5218 
5219  // Early out if there are too many unique results.
5220  if (UniqueResults.size() > MaxUniqueResults)
5221  return false;
5222 
5223  // Check the PHI consistency.
5224  if (!PHI)
5225  PHI = Results[0].first;
5226  else if (PHI != Results[0].first)
5227  return false;
5228  }
5229  // Find the default result value.
5231  BasicBlock *DefaultDest = SI->getDefaultDest();
5232  GetCaseResults(SI, nullptr, SI->getDefaultDest(), &CommonDest, DefaultResults,
5233  DL, TTI);
5234  // If the default value is not found abort unless the default destination
5235  // is unreachable.
5236  DefaultResult =
5237  DefaultResults.size() == 1 ? DefaultResults.begin()->second : nullptr;
5238  if ((!DefaultResult &&
5239  !isa<UnreachableInst>(DefaultDest->getFirstNonPHIOrDbg())))
5240  return false;
5241 
5242  return true;
5243 }
5244 
5245 // Helper function that checks if it is possible to transform a switch with only
5246 // two cases (or two cases + default) that produces a result into a select.
5247 // Example:
5248 // switch (a) {
5249 // case 10: %0 = icmp eq i32 %a, 10
5250 // return 10; %1 = select i1 %0, i32 10, i32 4
5251 // case 20: ----> %2 = icmp eq i32 %a, 20
5252 // return 2; %3 = select i1 %2, i32 2, i32 %1
5253 // default:
5254 // return 4;
5255 // }
5256 static Value *ConvertTwoCaseSwitch(const SwitchCaseResultVectorTy &ResultVector,
5257  Constant *DefaultResult, Value *Condition,
5258  IRBuilder<> &Builder) {
5259  // If we are selecting between only two cases transform into a simple
5260  // select or a two-way select if default is possible.
5261  if (ResultVector.size() == 2 && ResultVector[0].second.size() == 1 &&
5262  ResultVector[1].second.size() == 1) {
5263  ConstantInt *const FirstCase = ResultVector[0].second[0];
5264  ConstantInt *const SecondCase = ResultVector[1].second[0];
5265 
5266  bool DefaultCanTrigger = DefaultResult;
5267  Value *SelectValue = ResultVector[1].first;
5268  if (DefaultCanTrigger) {
5269  Value *const ValueCompare =
5270  Builder.CreateICmpEQ(Condition, SecondCase, "switch.selectcmp");
5271  SelectValue = Builder.CreateSelect(ValueCompare, ResultVector[1].first,
5272  DefaultResult, "switch.select");
5273  }
5274  Value *const ValueCompare =
5275  Builder.CreateICmpEQ(Condition, FirstCase, "switch.selectcmp");
5276  return Builder.CreateSelect(ValueCompare, ResultVector[0].first,
5277  SelectValue, "switch.select");
5278  }
5279 
5280  // Handle the degenerate case where two cases have the same value.
5281  if (ResultVector.size() == 1 && ResultVector[0].second.size() == 2 &&
5282  DefaultResult) {
5283  Value *Cmp1 = Builder.CreateICmpEQ(
5284  Condition, ResultVector[0].second[0], "switch.selectcmp.case1");
5285  Value *Cmp2 = Builder.CreateICmpEQ(
5286  Condition, ResultVector[0].second[1], "switch.selectcmp.case2");
5287  Value *Cmp = Builder.CreateOr(Cmp1, Cmp2, "switch.selectcmp");
5288  return Builder.CreateSelect(Cmp, ResultVector[0].first, DefaultResult);
5289  }
5290 
5291  return nullptr;
5292 }
5293 
5294 // Helper function to cleanup a switch instruction that has been converted into
5295 // a select, fixing up PHI nodes and basic blocks.
5297  Value *SelectValue,
5299  DomTreeUpdater *DTU) {
5300  std::vector<DominatorTree::UpdateType> Updates;
5301 
5302  BasicBlock *SelectBB = SI->getParent();
5303  BasicBlock *DestBB = PHI->getParent();
5304 
5305  if (DTU && !