LLVM  8.0.0svn
SimplifyCFG.cpp
Go to the documentation of this file.
1 //===- SimplifyCFG.cpp - Code to perform CFG simplification ---------------===//
2 //
3 // The LLVM Compiler Infrastructure
4 //
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
7 //
8 //===----------------------------------------------------------------------===//
9 //
10 // Peephole optimize the CFG.
11 //
12 //===----------------------------------------------------------------------===//
13 
14 #include "llvm/ADT/APInt.h"
15 #include "llvm/ADT/ArrayRef.h"
16 #include "llvm/ADT/DenseMap.h"
17 #include "llvm/ADT/Optional.h"
18 #include "llvm/ADT/STLExtras.h"
19 #include "llvm/ADT/SetOperations.h"
20 #include "llvm/ADT/SetVector.h"
21 #include "llvm/ADT/SmallPtrSet.h"
22 #include "llvm/ADT/SmallVector.h"
23 #include "llvm/ADT/Statistic.h"
24 #include "llvm/ADT/StringRef.h"
32 #include "llvm/IR/Attributes.h"
33 #include "llvm/IR/BasicBlock.h"
34 #include "llvm/IR/CFG.h"
35 #include "llvm/IR/CallSite.h"
36 #include "llvm/IR/Constant.h"
37 #include "llvm/IR/ConstantRange.h"
38 #include "llvm/IR/Constants.h"
39 #include "llvm/IR/DataLayout.h"
40 #include "llvm/IR/DerivedTypes.h"
41 #include "llvm/IR/Function.h"
42 #include "llvm/IR/GlobalValue.h"
43 #include "llvm/IR/GlobalVariable.h"
44 #include "llvm/IR/IRBuilder.h"
45 #include "llvm/IR/InstrTypes.h"
46 #include "llvm/IR/Instruction.h"
47 #include "llvm/IR/Instructions.h"
48 #include "llvm/IR/IntrinsicInst.h"
49 #include "llvm/IR/Intrinsics.h"
50 #include "llvm/IR/LLVMContext.h"
51 #include "llvm/IR/MDBuilder.h"
52 #include "llvm/IR/Metadata.h"
53 #include "llvm/IR/Module.h"
54 #include "llvm/IR/NoFolder.h"
55 #include "llvm/IR/Operator.h"
56 #include "llvm/IR/PatternMatch.h"
57 #include "llvm/IR/Type.h"
58 #include "llvm/IR/Use.h"
59 #include "llvm/IR/User.h"
60 #include "llvm/IR/Value.h"
61 #include "llvm/Support/Casting.h"
63 #include "llvm/Support/Debug.h"
65 #include "llvm/Support/KnownBits.h"
70 #include <algorithm>
71 #include <cassert>
72 #include <climits>
73 #include <cstddef>
74 #include <cstdint>
75 #include <iterator>
76 #include <map>
77 #include <set>
78 #include <tuple>
79 #include <utility>
80 #include <vector>
81 
82 using namespace llvm;
83 using namespace PatternMatch;
84 
85 #define DEBUG_TYPE "simplifycfg"
86 
87 // Chosen as 2 so as to be cheap, but still to have enough power to fold
88 // a select, so the "clamp" idiom (of a min followed by a max) will be caught.
89 // To catch this, we need to fold a compare and a select, hence '2' being the
90 // minimum reasonable default.
92  "phi-node-folding-threshold", cl::Hidden, cl::init(2),
93  cl::desc(
94  "Control the amount of phi node folding to perform (default = 2)"));
95 
96 static cl::opt<bool> DupRet(
97  "simplifycfg-dup-ret", cl::Hidden, cl::init(false),
98  cl::desc("Duplicate return instructions into unconditional branches"));
99 
100 static cl::opt<bool>
101  SinkCommon("simplifycfg-sink-common", cl::Hidden, cl::init(true),
102  cl::desc("Sink common instructions down to the end block"));
103 
105  "simplifycfg-hoist-cond-stores", cl::Hidden, cl::init(true),
106  cl::desc("Hoist conditional stores if an unconditional store precedes"));
107 
109  "simplifycfg-merge-cond-stores", cl::Hidden, cl::init(true),
110  cl::desc("Hoist conditional stores even if an unconditional store does not "
111  "precede - hoist multiple conditional stores into a single "
112  "predicated store"));
113 
115  "simplifycfg-merge-cond-stores-aggressively", cl::Hidden, cl::init(false),
116  cl::desc("When merging conditional stores, do so even if the resultant "
117  "basic blocks are unlikely to be if-converted as a result"));
118 
120  "speculate-one-expensive-inst", cl::Hidden, cl::init(true),
121  cl::desc("Allow exactly one expensive instruction to be speculatively "
122  "executed"));
123 
125  "max-speculation-depth", cl::Hidden, cl::init(10),
126  cl::desc("Limit maximum recursion depth when calculating costs of "
127  "speculatively executed instructions"));
128 
129 STATISTIC(NumBitMaps, "Number of switch instructions turned into bitmaps");
130 STATISTIC(NumLinearMaps,
131  "Number of switch instructions turned into linear mapping");
132 STATISTIC(NumLookupTables,
133  "Number of switch instructions turned into lookup tables");
134 STATISTIC(
135  NumLookupTablesHoles,
136  "Number of switch instructions turned into lookup tables (holes checked)");
137 STATISTIC(NumTableCmpReuses, "Number of reused switch table lookup compares");
138 STATISTIC(NumSinkCommons,
139  "Number of common instructions sunk down to the end block");
140 STATISTIC(NumSpeculations, "Number of speculative executed instructions");
141 
142 namespace {
143 
144 // The first field contains the value that the switch produces when a certain
145 // case group is selected, and the second field is a vector containing the
146 // cases composing the case group.
147 using SwitchCaseResultVectorTy =
149 
150 // The first field contains the phi node that generates a result of the switch
151 // and the second field contains the value generated for a certain case in the
152 // switch for that PHI.
153 using SwitchCaseResultsTy = SmallVector<std::pair<PHINode *, Constant *>, 4>;
154 
155 /// ValueEqualityComparisonCase - Represents a case of a switch.
156 struct ValueEqualityComparisonCase {
158  BasicBlock *Dest;
159 
160  ValueEqualityComparisonCase(ConstantInt *Value, BasicBlock *Dest)
161  : Value(Value), Dest(Dest) {}
162 
163  bool operator<(ValueEqualityComparisonCase RHS) const {
164  // Comparing pointers is ok as we only rely on the order for uniquing.
165  return Value < RHS.Value;
166  }
167 
168  bool operator==(BasicBlock *RHSDest) const { return Dest == RHSDest; }
169 };
170 
171 class SimplifyCFGOpt {
172  const TargetTransformInfo &TTI;
173  const DataLayout &DL;
174  SmallPtrSetImpl<BasicBlock *> *LoopHeaders;
175  const SimplifyCFGOptions &Options;
176  bool Resimplify;
177 
178  Value *isValueEqualityComparison(Instruction *TI);
179  BasicBlock *GetValueEqualityComparisonCases(
180  Instruction *TI, std::vector<ValueEqualityComparisonCase> &Cases);
181  bool SimplifyEqualityComparisonWithOnlyPredecessor(Instruction *TI,
182  BasicBlock *Pred,
183  IRBuilder<> &Builder);
184  bool FoldValueComparisonIntoPredecessors(Instruction *TI,
185  IRBuilder<> &Builder);
186 
187  bool SimplifyReturn(ReturnInst *RI, IRBuilder<> &Builder);
188  bool SimplifyResume(ResumeInst *RI, IRBuilder<> &Builder);
189  bool SimplifySingleResume(ResumeInst *RI);
190  bool SimplifyCommonResume(ResumeInst *RI);
191  bool SimplifyCleanupReturn(CleanupReturnInst *RI);
192  bool SimplifyUnreachable(UnreachableInst *UI);
193  bool SimplifySwitch(SwitchInst *SI, IRBuilder<> &Builder);
194  bool SimplifyIndirectBr(IndirectBrInst *IBI);
195  bool SimplifyUncondBranch(BranchInst *BI, IRBuilder<> &Builder);
196  bool SimplifyCondBranch(BranchInst *BI, IRBuilder<> &Builder);
197 
198  bool tryToSimplifyUncondBranchWithICmpInIt(ICmpInst *ICI,
199  IRBuilder<> &Builder);
200 
201 public:
202  SimplifyCFGOpt(const TargetTransformInfo &TTI, const DataLayout &DL,
203  SmallPtrSetImpl<BasicBlock *> *LoopHeaders,
204  const SimplifyCFGOptions &Opts)
205  : TTI(TTI), DL(DL), LoopHeaders(LoopHeaders), Options(Opts) {}
206 
207  bool run(BasicBlock *BB);
208  bool simplifyOnce(BasicBlock *BB);
209 
210  // Helper to set Resimplify and return change indication.
211  bool requestResimplify() {
212  Resimplify = true;
213  return true;
214  }
215 };
216 
217 } // end anonymous namespace
218 
219 /// Return true if it is safe to merge these two
220 /// terminator instructions together.
221 static bool
223  SmallSetVector<BasicBlock *, 4> *FailBlocks = nullptr) {
224  if (SI1 == SI2)
225  return false; // Can't merge with self!
226 
227  // It is not safe to merge these two switch instructions if they have a common
228  // successor, and if that successor has a PHI node, and if *that* PHI node has
229  // conflicting incoming values from the two switch blocks.
230  BasicBlock *SI1BB = SI1->getParent();
231  BasicBlock *SI2BB = SI2->getParent();
232 
233  SmallPtrSet<BasicBlock *, 16> SI1Succs(succ_begin(SI1BB), succ_end(SI1BB));
234  bool Fail = false;
235  for (BasicBlock *Succ : successors(SI2BB))
236  if (SI1Succs.count(Succ))
237  for (BasicBlock::iterator BBI = Succ->begin(); isa<PHINode>(BBI); ++BBI) {
238  PHINode *PN = cast<PHINode>(BBI);
239  if (PN->getIncomingValueForBlock(SI1BB) !=
240  PN->getIncomingValueForBlock(SI2BB)) {
241  if (FailBlocks)
242  FailBlocks->insert(Succ);
243  Fail = true;
244  }
245  }
246 
247  return !Fail;
248 }
249 
250 /// Return true if it is safe and profitable to merge these two terminator
251 /// instructions together, where SI1 is an unconditional branch. PhiNodes will
252 /// store all PHI nodes in common successors.
253 static bool
255  Instruction *Cond,
256  SmallVectorImpl<PHINode *> &PhiNodes) {
257  if (SI1 == SI2)
258  return false; // Can't merge with self!
259  assert(SI1->isUnconditional() && SI2->isConditional());
260 
261  // We fold the unconditional branch if we can easily update all PHI nodes in
262  // common successors:
263  // 1> We have a constant incoming value for the conditional branch;
264  // 2> We have "Cond" as the incoming value for the unconditional branch;
265  // 3> SI2->getCondition() and Cond have same operands.
266  CmpInst *Ci2 = dyn_cast<CmpInst>(SI2->getCondition());
267  if (!Ci2)
268  return false;
269  if (!(Cond->getOperand(0) == Ci2->getOperand(0) &&
270  Cond->getOperand(1) == Ci2->getOperand(1)) &&
271  !(Cond->getOperand(0) == Ci2->getOperand(1) &&
272  Cond->getOperand(1) == Ci2->getOperand(0)))
273  return false;
274 
275  BasicBlock *SI1BB = SI1->getParent();
276  BasicBlock *SI2BB = SI2->getParent();
277  SmallPtrSet<BasicBlock *, 16> SI1Succs(succ_begin(SI1BB), succ_end(SI1BB));
278  for (BasicBlock *Succ : successors(SI2BB))
279  if (SI1Succs.count(Succ))
280  for (BasicBlock::iterator BBI = Succ->begin(); isa<PHINode>(BBI); ++BBI) {
281  PHINode *PN = cast<PHINode>(BBI);
282  if (PN->getIncomingValueForBlock(SI1BB) != Cond ||
283  !isa<ConstantInt>(PN->getIncomingValueForBlock(SI2BB)))
284  return false;
285  PhiNodes.push_back(PN);
286  }
287  return true;
288 }
289 
290 /// Update PHI nodes in Succ to indicate that there will now be entries in it
291 /// from the 'NewPred' block. The values that will be flowing into the PHI nodes
292 /// will be the same as those coming in from ExistPred, an existing predecessor
293 /// of Succ.
294 static void AddPredecessorToBlock(BasicBlock *Succ, BasicBlock *NewPred,
295  BasicBlock *ExistPred) {
296  for (PHINode &PN : Succ->phis())
297  PN.addIncoming(PN.getIncomingValueForBlock(ExistPred), NewPred);
298 }
299 
300 /// Compute an abstract "cost" of speculating the given instruction,
301 /// which is assumed to be safe to speculate. TCC_Free means cheap,
302 /// TCC_Basic means less cheap, and TCC_Expensive means prohibitively
303 /// expensive.
304 static unsigned ComputeSpeculationCost(const User *I,
305  const TargetTransformInfo &TTI) {
307  "Instruction is not safe to speculatively execute!");
308  return TTI.getUserCost(I);
309 }
310 
311 /// If we have a merge point of an "if condition" as accepted above,
312 /// return true if the specified value dominates the block. We
313 /// don't handle the true generality of domination here, just a special case
314 /// which works well enough for us.
315 ///
316 /// If AggressiveInsts is non-null, and if V does not dominate BB, we check to
317 /// see if V (which must be an instruction) and its recursive operands
318 /// that do not dominate BB have a combined cost lower than CostRemaining and
319 /// are non-trapping. If both are true, the instruction is inserted into the
320 /// set and true is returned.
321 ///
322 /// The cost for most non-trapping instructions is defined as 1 except for
323 /// Select whose cost is 2.
324 ///
325 /// After this function returns, CostRemaining is decreased by the cost of
326 /// V plus its non-dominating operands. If that cost is greater than
327 /// CostRemaining, false is returned and CostRemaining is undefined.
329  SmallPtrSetImpl<Instruction *> &AggressiveInsts,
330  unsigned &CostRemaining,
331  const TargetTransformInfo &TTI,
332  unsigned Depth = 0) {
333  // It is possible to hit a zero-cost cycle (phi/gep instructions for example),
334  // so limit the recursion depth.
335  // TODO: While this recursion limit does prevent pathological behavior, it
336  // would be better to track visited instructions to avoid cycles.
337  if (Depth == MaxSpeculationDepth)
338  return false;
339 
341  if (!I) {
342  // Non-instructions all dominate instructions, but not all constantexprs
343  // can be executed unconditionally.
344  if (ConstantExpr *C = dyn_cast<ConstantExpr>(V))
345  if (C->canTrap())
346  return false;
347  return true;
348  }
349  BasicBlock *PBB = I->getParent();
350 
351  // We don't want to allow weird loops that might have the "if condition" in
352  // the bottom of this block.
353  if (PBB == BB)
354  return false;
355 
356  // If this instruction is defined in a block that contains an unconditional
357  // branch to BB, then it must be in the 'conditional' part of the "if
358  // statement". If not, it definitely dominates the region.
360  if (!BI || BI->isConditional() || BI->getSuccessor(0) != BB)
361  return true;
362 
363  // If we have seen this instruction before, don't count it again.
364  if (AggressiveInsts.count(I))
365  return true;
366 
367  // Okay, it looks like the instruction IS in the "condition". Check to
368  // see if it's a cheap instruction to unconditionally compute, and if it
369  // only uses stuff defined outside of the condition. If so, hoist it out.
371  return false;
372 
373  unsigned Cost = ComputeSpeculationCost(I, TTI);
374 
375  // Allow exactly one instruction to be speculated regardless of its cost
376  // (as long as it is safe to do so).
377  // This is intended to flatten the CFG even if the instruction is a division
378  // or other expensive operation. The speculation of an expensive instruction
379  // is expected to be undone in CodeGenPrepare if the speculation has not
380  // enabled further IR optimizations.
381  if (Cost > CostRemaining &&
382  (!SpeculateOneExpensiveInst || !AggressiveInsts.empty() || Depth > 0))
383  return false;
384 
385  // Avoid unsigned wrap.
386  CostRemaining = (Cost > CostRemaining) ? 0 : CostRemaining - Cost;
387 
388  // Okay, we can only really hoist these out if their operands do
389  // not take us over the cost threshold.
390  for (User::op_iterator i = I->op_begin(), e = I->op_end(); i != e; ++i)
391  if (!DominatesMergePoint(*i, BB, AggressiveInsts, CostRemaining, TTI,
392  Depth + 1))
393  return false;
394  // Okay, it's safe to do this! Remember this instruction.
395  AggressiveInsts.insert(I);
396  return true;
397 }
398 
399 /// Extract ConstantInt from value, looking through IntToPtr
400 /// and PointerNullValue. Return NULL if value is not a constant int.
401 static ConstantInt *GetConstantInt(Value *V, const DataLayout &DL) {
402  // Normal constant int.
403  ConstantInt *CI = dyn_cast<ConstantInt>(V);
404  if (CI || !isa<Constant>(V) || !V->getType()->isPointerTy())
405  return CI;
406 
407  // This is some kind of pointer constant. Turn it into a pointer-sized
408  // ConstantInt if possible.
409  IntegerType *PtrTy = cast<IntegerType>(DL.getIntPtrType(V->getType()));
410 
411  // Null pointer means 0, see SelectionDAGBuilder::getValue(const Value*).
412  if (isa<ConstantPointerNull>(V))
413  return ConstantInt::get(PtrTy, 0);
414 
415  // IntToPtr const int.
416  if (ConstantExpr *CE = dyn_cast<ConstantExpr>(V))
417  if (CE->getOpcode() == Instruction::IntToPtr)
418  if (ConstantInt *CI = dyn_cast<ConstantInt>(CE->getOperand(0))) {
419  // The constant is very likely to have the right type already.
420  if (CI->getType() == PtrTy)
421  return CI;
422  else
423  return cast<ConstantInt>(
424  ConstantExpr::getIntegerCast(CI, PtrTy, /*isSigned=*/false));
425  }
426  return nullptr;
427 }
428 
429 namespace {
430 
431 /// Given a chain of or (||) or and (&&) comparison of a value against a
432 /// constant, this will try to recover the information required for a switch
433 /// structure.
434 /// It will depth-first traverse the chain of comparison, seeking for patterns
435 /// like %a == 12 or %a < 4 and combine them to produce a set of integer
436 /// representing the different cases for the switch.
437 /// Note that if the chain is composed of '||' it will build the set of elements
438 /// that matches the comparisons (i.e. any of this value validate the chain)
439 /// while for a chain of '&&' it will build the set elements that make the test
440 /// fail.
441 struct ConstantComparesGatherer {
442  const DataLayout &DL;
443 
444  /// Value found for the switch comparison
445  Value *CompValue = nullptr;
446 
447  /// Extra clause to be checked before the switch
448  Value *Extra = nullptr;
449 
450  /// Set of integers to match in switch
452 
453  /// Number of comparisons matched in the and/or chain
454  unsigned UsedICmps = 0;
455 
456  /// Construct and compute the result for the comparison instruction Cond
457  ConstantComparesGatherer(Instruction *Cond, const DataLayout &DL) : DL(DL) {
458  gather(Cond);
459  }
460 
461  ConstantComparesGatherer(const ConstantComparesGatherer &) = delete;
462  ConstantComparesGatherer &
463  operator=(const ConstantComparesGatherer &) = delete;
464 
465 private:
466  /// Try to set the current value used for the comparison, it succeeds only if
467  /// it wasn't set before or if the new value is the same as the old one
468  bool setValueOnce(Value *NewVal) {
469  if (CompValue && CompValue != NewVal)
470  return false;
471  CompValue = NewVal;
472  return (CompValue != nullptr);
473  }
474 
475  /// Try to match Instruction "I" as a comparison against a constant and
476  /// populates the array Vals with the set of values that match (or do not
477  /// match depending on isEQ).
478  /// Return false on failure. On success, the Value the comparison matched
479  /// against is placed in CompValue.
480  /// If CompValue is already set, the function is expected to fail if a match
481  /// is found but the value compared to is different.
482  bool matchInstruction(Instruction *I, bool isEQ) {
483  // If this is an icmp against a constant, handle this as one of the cases.
484  ICmpInst *ICI;
485  ConstantInt *C;
486  if (!((ICI = dyn_cast<ICmpInst>(I)) &&
487  (C = GetConstantInt(I->getOperand(1), DL)))) {
488  return false;
489  }
490 
491  Value *RHSVal;
492  const APInt *RHSC;
493 
494  // Pattern match a special case
495  // (x & ~2^z) == y --> x == y || x == y|2^z
496  // This undoes a transformation done by instcombine to fuse 2 compares.
497  if (ICI->getPredicate() == (isEQ ? ICmpInst::ICMP_EQ : ICmpInst::ICMP_NE)) {
498  // It's a little bit hard to see why the following transformations are
499  // correct. Here is a CVC3 program to verify them for 64-bit values:
500 
501  /*
502  ONE : BITVECTOR(64) = BVZEROEXTEND(0bin1, 63);
503  x : BITVECTOR(64);
504  y : BITVECTOR(64);
505  z : BITVECTOR(64);
506  mask : BITVECTOR(64) = BVSHL(ONE, z);
507  QUERY( (y & ~mask = y) =>
508  ((x & ~mask = y) <=> (x = y OR x = (y | mask)))
509  );
510  QUERY( (y | mask = y) =>
511  ((x | mask = y) <=> (x = y OR x = (y & ~mask)))
512  );
513  */
514 
515  // Please note that each pattern must be a dual implication (<--> or
516  // iff). One directional implication can create spurious matches. If the
517  // implication is only one-way, an unsatisfiable condition on the left
518  // side can imply a satisfiable condition on the right side. Dual
519  // implication ensures that satisfiable conditions are transformed to
520  // other satisfiable conditions and unsatisfiable conditions are
521  // transformed to other unsatisfiable conditions.
522 
523  // Here is a concrete example of a unsatisfiable condition on the left
524  // implying a satisfiable condition on the right:
525  //
526  // mask = (1 << z)
527  // (x & ~mask) == y --> (x == y || x == (y | mask))
528  //
529  // Substituting y = 3, z = 0 yields:
530  // (x & -2) == 3 --> (x == 3 || x == 2)
531 
532  // Pattern match a special case:
533  /*
534  QUERY( (y & ~mask = y) =>
535  ((x & ~mask = y) <=> (x = y OR x = (y | mask)))
536  );
537  */
538  if (match(ICI->getOperand(0),
539  m_And(m_Value(RHSVal), m_APInt(RHSC)))) {
540  APInt Mask = ~*RHSC;
541  if (Mask.isPowerOf2() && (C->getValue() & ~Mask) == C->getValue()) {
542  // If we already have a value for the switch, it has to match!
543  if (!setValueOnce(RHSVal))
544  return false;
545 
546  Vals.push_back(C);
547  Vals.push_back(
549  C->getValue() | Mask));
550  UsedICmps++;
551  return true;
552  }
553  }
554 
555  // Pattern match a special case:
556  /*
557  QUERY( (y | mask = y) =>
558  ((x | mask = y) <=> (x = y OR x = (y & ~mask)))
559  );
560  */
561  if (match(ICI->getOperand(0),
562  m_Or(m_Value(RHSVal), m_APInt(RHSC)))) {
563  APInt Mask = *RHSC;
564  if (Mask.isPowerOf2() && (C->getValue() | Mask) == C->getValue()) {
565  // If we already have a value for the switch, it has to match!
566  if (!setValueOnce(RHSVal))
567  return false;
568 
569  Vals.push_back(C);
571  C->getValue() & ~Mask));
572  UsedICmps++;
573  return true;
574  }
575  }
576 
577  // If we already have a value for the switch, it has to match!
578  if (!setValueOnce(ICI->getOperand(0)))
579  return false;
580 
581  UsedICmps++;
582  Vals.push_back(C);
583  return ICI->getOperand(0);
584  }
585 
586  // If we have "x ult 3", for example, then we can add 0,1,2 to the set.
588  ICI->getPredicate(), C->getValue());
589 
590  // Shift the range if the compare is fed by an add. This is the range
591  // compare idiom as emitted by instcombine.
592  Value *CandidateVal = I->getOperand(0);
593  if (match(I->getOperand(0), m_Add(m_Value(RHSVal), m_APInt(RHSC)))) {
594  Span = Span.subtract(*RHSC);
595  CandidateVal = RHSVal;
596  }
597 
598  // If this is an and/!= check, then we are looking to build the set of
599  // value that *don't* pass the and chain. I.e. to turn "x ugt 2" into
600  // x != 0 && x != 1.
601  if (!isEQ)
602  Span = Span.inverse();
603 
604  // If there are a ton of values, we don't want to make a ginormous switch.
605  if (Span.isSizeLargerThan(8) || Span.isEmptySet()) {
606  return false;
607  }
608 
609  // If we already have a value for the switch, it has to match!
610  if (!setValueOnce(CandidateVal))
611  return false;
612 
613  // Add all values from the range to the set
614  for (APInt Tmp = Span.getLower(); Tmp != Span.getUpper(); ++Tmp)
615  Vals.push_back(ConstantInt::get(I->getContext(), Tmp));
616 
617  UsedICmps++;
618  return true;
619  }
620 
621  /// Given a potentially 'or'd or 'and'd together collection of icmp
622  /// eq/ne/lt/gt instructions that compare a value against a constant, extract
623  /// the value being compared, and stick the list constants into the Vals
624  /// vector.
625  /// One "Extra" case is allowed to differ from the other.
626  void gather(Value *V) {
628  bool isEQ = (I->getOpcode() == Instruction::Or);
629 
630  // Keep a stack (SmallVector for efficiency) for depth-first traversal
632  SmallPtrSet<Value *, 8> Visited;
633 
634  // Initialize
635  Visited.insert(V);
636  DFT.push_back(V);
637 
638  while (!DFT.empty()) {
639  V = DFT.pop_back_val();
640 
641  if (Instruction *I = dyn_cast<Instruction>(V)) {
642  // If it is a || (or && depending on isEQ), process the operands.
643  if (I->getOpcode() == (isEQ ? Instruction::Or : Instruction::And)) {
644  if (Visited.insert(I->getOperand(1)).second)
645  DFT.push_back(I->getOperand(1));
646  if (Visited.insert(I->getOperand(0)).second)
647  DFT.push_back(I->getOperand(0));
648  continue;
649  }
650 
651  // Try to match the current instruction
652  if (matchInstruction(I, isEQ))
653  // Match succeed, continue the loop
654  continue;
655  }
656 
657  // One element of the sequence of || (or &&) could not be match as a
658  // comparison against the same value as the others.
659  // We allow only one "Extra" case to be checked before the switch
660  if (!Extra) {
661  Extra = V;
662  continue;
663  }
664  // Failed to parse a proper sequence, abort now
665  CompValue = nullptr;
666  break;
667  }
668  }
669 };
670 
671 } // end anonymous namespace
672 
674  Instruction *Cond = nullptr;
675  if (SwitchInst *SI = dyn_cast<SwitchInst>(TI)) {
676  Cond = dyn_cast<Instruction>(SI->getCondition());
677  } else if (BranchInst *BI = dyn_cast<BranchInst>(TI)) {
678  if (BI->isConditional())
679  Cond = dyn_cast<Instruction>(BI->getCondition());
680  } else if (IndirectBrInst *IBI = dyn_cast<IndirectBrInst>(TI)) {
681  Cond = dyn_cast<Instruction>(IBI->getAddress());
682  }
683 
684  TI->eraseFromParent();
685  if (Cond)
687 }
688 
689 /// Return true if the specified terminator checks
690 /// to see if a value is equal to constant integer value.
691 Value *SimplifyCFGOpt::isValueEqualityComparison(Instruction *TI) {
692  Value *CV = nullptr;
693  if (SwitchInst *SI = dyn_cast<SwitchInst>(TI)) {
694  // Do not permit merging of large switch instructions into their
695  // predecessors unless there is only one predecessor.
696  if (!SI->getParent()->hasNPredecessorsOrMore(128 / SI->getNumSuccessors()))
697  CV = SI->getCondition();
698  } else if (BranchInst *BI = dyn_cast<BranchInst>(TI))
699  if (BI->isConditional() && BI->getCondition()->hasOneUse())
700  if (ICmpInst *ICI = dyn_cast<ICmpInst>(BI->getCondition())) {
701  if (ICI->isEquality() && GetConstantInt(ICI->getOperand(1), DL))
702  CV = ICI->getOperand(0);
703  }
704 
705  // Unwrap any lossless ptrtoint cast.
706  if (CV) {
707  if (PtrToIntInst *PTII = dyn_cast<PtrToIntInst>(CV)) {
708  Value *Ptr = PTII->getPointerOperand();
709  if (PTII->getType() == DL.getIntPtrType(Ptr->getType()))
710  CV = Ptr;
711  }
712  }
713  return CV;
714 }
715 
716 /// Given a value comparison instruction,
717 /// decode all of the 'cases' that it represents and return the 'default' block.
718 BasicBlock *SimplifyCFGOpt::GetValueEqualityComparisonCases(
719  Instruction *TI, std::vector<ValueEqualityComparisonCase> &Cases) {
720  if (SwitchInst *SI = dyn_cast<SwitchInst>(TI)) {
721  Cases.reserve(SI->getNumCases());
722  for (auto Case : SI->cases())
723  Cases.push_back(ValueEqualityComparisonCase(Case.getCaseValue(),
724  Case.getCaseSuccessor()));
725  return SI->getDefaultDest();
726  }
727 
728  BranchInst *BI = cast<BranchInst>(TI);
729  ICmpInst *ICI = cast<ICmpInst>(BI->getCondition());
730  BasicBlock *Succ = BI->getSuccessor(ICI->getPredicate() == ICmpInst::ICMP_NE);
731  Cases.push_back(ValueEqualityComparisonCase(
732  GetConstantInt(ICI->getOperand(1), DL), Succ));
733  return BI->getSuccessor(ICI->getPredicate() == ICmpInst::ICMP_EQ);
734 }
735 
736 /// Given a vector of bb/value pairs, remove any entries
737 /// in the list that match the specified block.
738 static void
740  std::vector<ValueEqualityComparisonCase> &Cases) {
741  Cases.erase(std::remove(Cases.begin(), Cases.end(), BB), Cases.end());
742 }
743 
744 /// Return true if there are any keys in C1 that exist in C2 as well.
745 static bool ValuesOverlap(std::vector<ValueEqualityComparisonCase> &C1,
746  std::vector<ValueEqualityComparisonCase> &C2) {
747  std::vector<ValueEqualityComparisonCase> *V1 = &C1, *V2 = &C2;
748 
749  // Make V1 be smaller than V2.
750  if (V1->size() > V2->size())
751  std::swap(V1, V2);
752 
753  if (V1->empty())
754  return false;
755  if (V1->size() == 1) {
756  // Just scan V2.
757  ConstantInt *TheVal = (*V1)[0].Value;
758  for (unsigned i = 0, e = V2->size(); i != e; ++i)
759  if (TheVal == (*V2)[i].Value)
760  return true;
761  }
762 
763  // Otherwise, just sort both lists and compare element by element.
764  array_pod_sort(V1->begin(), V1->end());
765  array_pod_sort(V2->begin(), V2->end());
766  unsigned i1 = 0, i2 = 0, e1 = V1->size(), e2 = V2->size();
767  while (i1 != e1 && i2 != e2) {
768  if ((*V1)[i1].Value == (*V2)[i2].Value)
769  return true;
770  if ((*V1)[i1].Value < (*V2)[i2].Value)
771  ++i1;
772  else
773  ++i2;
774  }
775  return false;
776 }
777 
778 // Set branch weights on SwitchInst. This sets the metadata if there is at
779 // least one non-zero weight.
781  // Check that there is at least one non-zero weight. Otherwise, pass
782  // nullptr to setMetadata which will erase the existing metadata.
783  MDNode *N = nullptr;
784  if (llvm::any_of(Weights, [](uint32_t W) { return W != 0; }))
785  N = MDBuilder(SI->getParent()->getContext()).createBranchWeights(Weights);
787 }
788 
789 // Similar to the above, but for branch and select instructions that take
790 // exactly 2 weights.
791 static void setBranchWeights(Instruction *I, uint32_t TrueWeight,
792  uint32_t FalseWeight) {
793  assert(isa<BranchInst>(I) || isa<SelectInst>(I));
794  // Check that there is at least one non-zero weight. Otherwise, pass
795  // nullptr to setMetadata which will erase the existing metadata.
796  MDNode *N = nullptr;
797  if (TrueWeight || FalseWeight)
798  N = MDBuilder(I->getParent()->getContext())
799  .createBranchWeights(TrueWeight, FalseWeight);
801 }
802 
803 /// If TI is known to be a terminator instruction and its block is known to
804 /// only have a single predecessor block, check to see if that predecessor is
805 /// also a value comparison with the same value, and if that comparison
806 /// determines the outcome of this comparison. If so, simplify TI. This does a
807 /// very limited form of jump threading.
808 bool SimplifyCFGOpt::SimplifyEqualityComparisonWithOnlyPredecessor(
809  Instruction *TI, BasicBlock *Pred, IRBuilder<> &Builder) {
810  Value *PredVal = isValueEqualityComparison(Pred->getTerminator());
811  if (!PredVal)
812  return false; // Not a value comparison in predecessor.
813 
814  Value *ThisVal = isValueEqualityComparison(TI);
815  assert(ThisVal && "This isn't a value comparison!!");
816  if (ThisVal != PredVal)
817  return false; // Different predicates.
818 
819  // TODO: Preserve branch weight metadata, similarly to how
820  // FoldValueComparisonIntoPredecessors preserves it.
821 
822  // Find out information about when control will move from Pred to TI's block.
823  std::vector<ValueEqualityComparisonCase> PredCases;
824  BasicBlock *PredDef =
825  GetValueEqualityComparisonCases(Pred->getTerminator(), PredCases);
826  EliminateBlockCases(PredDef, PredCases); // Remove default from cases.
827 
828  // Find information about how control leaves this block.
829  std::vector<ValueEqualityComparisonCase> ThisCases;
830  BasicBlock *ThisDef = GetValueEqualityComparisonCases(TI, ThisCases);
831  EliminateBlockCases(ThisDef, ThisCases); // Remove default from cases.
832 
833  // If TI's block is the default block from Pred's comparison, potentially
834  // simplify TI based on this knowledge.
835  if (PredDef == TI->getParent()) {
836  // If we are here, we know that the value is none of those cases listed in
837  // PredCases. If there are any cases in ThisCases that are in PredCases, we
838  // can simplify TI.
839  if (!ValuesOverlap(PredCases, ThisCases))
840  return false;
841 
842  if (isa<BranchInst>(TI)) {
843  // Okay, one of the successors of this condbr is dead. Convert it to a
844  // uncond br.
845  assert(ThisCases.size() == 1 && "Branch can only have one case!");
846  // Insert the new branch.
847  Instruction *NI = Builder.CreateBr(ThisDef);
848  (void)NI;
849 
850  // Remove PHI node entries for the dead edge.
851  ThisCases[0].Dest->removePredecessor(TI->getParent());
852 
853  LLVM_DEBUG(dbgs() << "Threading pred instr: " << *Pred->getTerminator()
854  << "Through successor TI: " << *TI << "Leaving: " << *NI
855  << "\n");
856 
858  return true;
859  }
860 
861  SwitchInst *SI = cast<SwitchInst>(TI);
862  // Okay, TI has cases that are statically dead, prune them away.
863  SmallPtrSet<Constant *, 16> DeadCases;
864  for (unsigned i = 0, e = PredCases.size(); i != e; ++i)
865  DeadCases.insert(PredCases[i].Value);
866 
867  LLVM_DEBUG(dbgs() << "Threading pred instr: " << *Pred->getTerminator()
868  << "Through successor TI: " << *TI);
869 
870  // Collect branch weights into a vector.
871  SmallVector<uint32_t, 8> Weights;
873  bool HasWeight = MD && (MD->getNumOperands() == 2 + SI->getNumCases());
874  if (HasWeight)
875  for (unsigned MD_i = 1, MD_e = MD->getNumOperands(); MD_i < MD_e;
876  ++MD_i) {
877  ConstantInt *CI = mdconst::extract<ConstantInt>(MD->getOperand(MD_i));
878  Weights.push_back(CI->getValue().getZExtValue());
879  }
880  for (SwitchInst::CaseIt i = SI->case_end(), e = SI->case_begin(); i != e;) {
881  --i;
882  if (DeadCases.count(i->getCaseValue())) {
883  if (HasWeight) {
884  std::swap(Weights[i->getCaseIndex() + 1], Weights.back());
885  Weights.pop_back();
886  }
887  i->getCaseSuccessor()->removePredecessor(TI->getParent());
888  SI->removeCase(i);
889  }
890  }
891  if (HasWeight && Weights.size() >= 2)
892  setBranchWeights(SI, Weights);
893 
894  LLVM_DEBUG(dbgs() << "Leaving: " << *TI << "\n");
895  return true;
896  }
897 
898  // Otherwise, TI's block must correspond to some matched value. Find out
899  // which value (or set of values) this is.
900  ConstantInt *TIV = nullptr;
901  BasicBlock *TIBB = TI->getParent();
902  for (unsigned i = 0, e = PredCases.size(); i != e; ++i)
903  if (PredCases[i].Dest == TIBB) {
904  if (TIV)
905  return false; // Cannot handle multiple values coming to this block.
906  TIV = PredCases[i].Value;
907  }
908  assert(TIV && "No edge from pred to succ?");
909 
910  // Okay, we found the one constant that our value can be if we get into TI's
911  // BB. Find out which successor will unconditionally be branched to.
912  BasicBlock *TheRealDest = nullptr;
913  for (unsigned i = 0, e = ThisCases.size(); i != e; ++i)
914  if (ThisCases[i].Value == TIV) {
915  TheRealDest = ThisCases[i].Dest;
916  break;
917  }
918 
919  // If not handled by any explicit cases, it is handled by the default case.
920  if (!TheRealDest)
921  TheRealDest = ThisDef;
922 
923  // Remove PHI node entries for dead edges.
924  BasicBlock *CheckEdge = TheRealDest;
925  for (BasicBlock *Succ : successors(TIBB))
926  if (Succ != CheckEdge)
927  Succ->removePredecessor(TIBB);
928  else
929  CheckEdge = nullptr;
930 
931  // Insert the new branch.
932  Instruction *NI = Builder.CreateBr(TheRealDest);
933  (void)NI;
934 
935  LLVM_DEBUG(dbgs() << "Threading pred instr: " << *Pred->getTerminator()
936  << "Through successor TI: " << *TI << "Leaving: " << *NI
937  << "\n");
938 
940  return true;
941 }
942 
943 namespace {
944 
945 /// This class implements a stable ordering of constant
946 /// integers that does not depend on their address. This is important for
947 /// applications that sort ConstantInt's to ensure uniqueness.
948 struct ConstantIntOrdering {
949  bool operator()(const ConstantInt *LHS, const ConstantInt *RHS) const {
950  return LHS->getValue().ult(RHS->getValue());
951  }
952 };
953 
954 } // end anonymous namespace
955 
956 static int ConstantIntSortPredicate(ConstantInt *const *P1,
957  ConstantInt *const *P2) {
958  const ConstantInt *LHS = *P1;
959  const ConstantInt *RHS = *P2;
960  if (LHS == RHS)
961  return 0;
962  return LHS->getValue().ult(RHS->getValue()) ? 1 : -1;
963 }
964 
965 static inline bool HasBranchWeights(const Instruction *I) {
967  if (ProfMD && ProfMD->getOperand(0))
968  if (MDString *MDS = dyn_cast<MDString>(ProfMD->getOperand(0)))
969  return MDS->getString().equals("branch_weights");
970 
971  return false;
972 }
973 
974 /// Get Weights of a given terminator, the default weight is at the front
975 /// of the vector. If TI is a conditional eq, we need to swap the branch-weight
976 /// metadata.
978  SmallVectorImpl<uint64_t> &Weights) {
980  assert(MD);
981  for (unsigned i = 1, e = MD->getNumOperands(); i < e; ++i) {
982  ConstantInt *CI = mdconst::extract<ConstantInt>(MD->getOperand(i));
983  Weights.push_back(CI->getValue().getZExtValue());
984  }
985 
986  // If TI is a conditional eq, the default case is the false case,
987  // and the corresponding branch-weight data is at index 2. We swap the
988  // default weight to be the first entry.
989  if (BranchInst *BI = dyn_cast<BranchInst>(TI)) {
990  assert(Weights.size() == 2);
991  ICmpInst *ICI = cast<ICmpInst>(BI->getCondition());
992  if (ICI->getPredicate() == ICmpInst::ICMP_EQ)
993  std::swap(Weights.front(), Weights.back());
994  }
995 }
996 
997 /// Keep halving the weights until all can fit in uint32_t.
998 static void FitWeights(MutableArrayRef<uint64_t> Weights) {
999  uint64_t Max = *std::max_element(Weights.begin(), Weights.end());
1000  if (Max > UINT_MAX) {
1001  unsigned Offset = 32 - countLeadingZeros(Max);
1002  for (uint64_t &I : Weights)
1003  I >>= Offset;
1004  }
1005 }
1006 
1007 /// The specified terminator is a value equality comparison instruction
1008 /// (either a switch or a branch on "X == c").
1009 /// See if any of the predecessors of the terminator block are value comparisons
1010 /// on the same value. If so, and if safe to do so, fold them together.
1011 bool SimplifyCFGOpt::FoldValueComparisonIntoPredecessors(Instruction *TI,
1012  IRBuilder<> &Builder) {
1013  BasicBlock *BB = TI->getParent();
1014  Value *CV = isValueEqualityComparison(TI); // CondVal
1015  assert(CV && "Not a comparison?");
1016  bool Changed = false;
1017 
1019  while (!Preds.empty()) {
1020  BasicBlock *Pred = Preds.pop_back_val();
1021 
1022  // See if the predecessor is a comparison with the same value.
1023  Instruction *PTI = Pred->getTerminator();
1024  Value *PCV = isValueEqualityComparison(PTI); // PredCondVal
1025 
1026  if (PCV == CV && TI != PTI) {
1027  SmallSetVector<BasicBlock*, 4> FailBlocks;
1028  if (!SafeToMergeTerminators(TI, PTI, &FailBlocks)) {
1029  for (auto *Succ : FailBlocks) {
1030  if (!SplitBlockPredecessors(Succ, TI->getParent(), ".fold.split"))
1031  return false;
1032  }
1033  }
1034 
1035  // Figure out which 'cases' to copy from SI to PSI.
1036  std::vector<ValueEqualityComparisonCase> BBCases;
1037  BasicBlock *BBDefault = GetValueEqualityComparisonCases(TI, BBCases);
1038 
1039  std::vector<ValueEqualityComparisonCase> PredCases;
1040  BasicBlock *PredDefault = GetValueEqualityComparisonCases(PTI, PredCases);
1041 
1042  // Based on whether the default edge from PTI goes to BB or not, fill in
1043  // PredCases and PredDefault with the new switch cases we would like to
1044  // build.
1045  SmallVector<BasicBlock *, 8> NewSuccessors;
1046 
1047  // Update the branch weight metadata along the way
1048  SmallVector<uint64_t, 8> Weights;
1049  bool PredHasWeights = HasBranchWeights(PTI);
1050  bool SuccHasWeights = HasBranchWeights(TI);
1051 
1052  if (PredHasWeights) {
1053  GetBranchWeights(PTI, Weights);
1054  // branch-weight metadata is inconsistent here.
1055  if (Weights.size() != 1 + PredCases.size())
1056  PredHasWeights = SuccHasWeights = false;
1057  } else if (SuccHasWeights)
1058  // If there are no predecessor weights but there are successor weights,
1059  // populate Weights with 1, which will later be scaled to the sum of
1060  // successor's weights
1061  Weights.assign(1 + PredCases.size(), 1);
1062 
1063  SmallVector<uint64_t, 8> SuccWeights;
1064  if (SuccHasWeights) {
1065  GetBranchWeights(TI, SuccWeights);
1066  // branch-weight metadata is inconsistent here.
1067  if (SuccWeights.size() != 1 + BBCases.size())
1068  PredHasWeights = SuccHasWeights = false;
1069  } else if (PredHasWeights)
1070  SuccWeights.assign(1 + BBCases.size(), 1);
1071 
1072  if (PredDefault == BB) {
1073  // If this is the default destination from PTI, only the edges in TI
1074  // that don't occur in PTI, or that branch to BB will be activated.
1075  std::set<ConstantInt *, ConstantIntOrdering> PTIHandled;
1076  for (unsigned i = 0, e = PredCases.size(); i != e; ++i)
1077  if (PredCases[i].Dest != BB)
1078  PTIHandled.insert(PredCases[i].Value);
1079  else {
1080  // The default destination is BB, we don't need explicit targets.
1081  std::swap(PredCases[i], PredCases.back());
1082 
1083  if (PredHasWeights || SuccHasWeights) {
1084  // Increase weight for the default case.
1085  Weights[0] += Weights[i + 1];
1086  std::swap(Weights[i + 1], Weights.back());
1087  Weights.pop_back();
1088  }
1089 
1090  PredCases.pop_back();
1091  --i;
1092  --e;
1093  }
1094 
1095  // Reconstruct the new switch statement we will be building.
1096  if (PredDefault != BBDefault) {
1097  PredDefault->removePredecessor(Pred);
1098  PredDefault = BBDefault;
1099  NewSuccessors.push_back(BBDefault);
1100  }
1101 
1102  unsigned CasesFromPred = Weights.size();
1103  uint64_t ValidTotalSuccWeight = 0;
1104  for (unsigned i = 0, e = BBCases.size(); i != e; ++i)
1105  if (!PTIHandled.count(BBCases[i].Value) &&
1106  BBCases[i].Dest != BBDefault) {
1107  PredCases.push_back(BBCases[i]);
1108  NewSuccessors.push_back(BBCases[i].Dest);
1109  if (SuccHasWeights || PredHasWeights) {
1110  // The default weight is at index 0, so weight for the ith case
1111  // should be at index i+1. Scale the cases from successor by
1112  // PredDefaultWeight (Weights[0]).
1113  Weights.push_back(Weights[0] * SuccWeights[i + 1]);
1114  ValidTotalSuccWeight += SuccWeights[i + 1];
1115  }
1116  }
1117 
1118  if (SuccHasWeights || PredHasWeights) {
1119  ValidTotalSuccWeight += SuccWeights[0];
1120  // Scale the cases from predecessor by ValidTotalSuccWeight.
1121  for (unsigned i = 1; i < CasesFromPred; ++i)
1122  Weights[i] *= ValidTotalSuccWeight;
1123  // Scale the default weight by SuccDefaultWeight (SuccWeights[0]).
1124  Weights[0] *= SuccWeights[0];
1125  }
1126  } else {
1127  // If this is not the default destination from PSI, only the edges
1128  // in SI that occur in PSI with a destination of BB will be
1129  // activated.
1130  std::set<ConstantInt *, ConstantIntOrdering> PTIHandled;
1131  std::map<ConstantInt *, uint64_t> WeightsForHandled;
1132  for (unsigned i = 0, e = PredCases.size(); i != e; ++i)
1133  if (PredCases[i].Dest == BB) {
1134  PTIHandled.insert(PredCases[i].Value);
1135 
1136  if (PredHasWeights || SuccHasWeights) {
1137  WeightsForHandled[PredCases[i].Value] = Weights[i + 1];
1138  std::swap(Weights[i + 1], Weights.back());
1139  Weights.pop_back();
1140  }
1141 
1142  std::swap(PredCases[i], PredCases.back());
1143  PredCases.pop_back();
1144  --i;
1145  --e;
1146  }
1147 
1148  // Okay, now we know which constants were sent to BB from the
1149  // predecessor. Figure out where they will all go now.
1150  for (unsigned i = 0, e = BBCases.size(); i != e; ++i)
1151  if (PTIHandled.count(BBCases[i].Value)) {
1152  // If this is one we are capable of getting...
1153  if (PredHasWeights || SuccHasWeights)
1154  Weights.push_back(WeightsForHandled[BBCases[i].Value]);
1155  PredCases.push_back(BBCases[i]);
1156  NewSuccessors.push_back(BBCases[i].Dest);
1157  PTIHandled.erase(
1158  BBCases[i].Value); // This constant is taken care of
1159  }
1160 
1161  // If there are any constants vectored to BB that TI doesn't handle,
1162  // they must go to the default destination of TI.
1163  for (ConstantInt *I : PTIHandled) {
1164  if (PredHasWeights || SuccHasWeights)
1165  Weights.push_back(WeightsForHandled[I]);
1166  PredCases.push_back(ValueEqualityComparisonCase(I, BBDefault));
1167  NewSuccessors.push_back(BBDefault);
1168  }
1169  }
1170 
1171  // Okay, at this point, we know which new successor Pred will get. Make
1172  // sure we update the number of entries in the PHI nodes for these
1173  // successors.
1174  for (BasicBlock *NewSuccessor : NewSuccessors)
1175  AddPredecessorToBlock(NewSuccessor, Pred, BB);
1176 
1177  Builder.SetInsertPoint(PTI);
1178  // Convert pointer to int before we switch.
1179  if (CV->getType()->isPointerTy()) {
1180  CV = Builder.CreatePtrToInt(CV, DL.getIntPtrType(CV->getType()),
1181  "magicptr");
1182  }
1183 
1184  // Now that the successors are updated, create the new Switch instruction.
1185  SwitchInst *NewSI =
1186  Builder.CreateSwitch(CV, PredDefault, PredCases.size());
1187  NewSI->setDebugLoc(PTI->getDebugLoc());
1188  for (ValueEqualityComparisonCase &V : PredCases)
1189  NewSI->addCase(V.Value, V.Dest);
1190 
1191  if (PredHasWeights || SuccHasWeights) {
1192  // Halve the weights if any of them cannot fit in an uint32_t
1193  FitWeights(Weights);
1194 
1195  SmallVector<uint32_t, 8> MDWeights(Weights.begin(), Weights.end());
1196 
1197  setBranchWeights(NewSI, MDWeights);
1198  }
1199 
1201 
1202  // Okay, last check. If BB is still a successor of PSI, then we must
1203  // have an infinite loop case. If so, add an infinitely looping block
1204  // to handle the case to preserve the behavior of the code.
1205  BasicBlock *InfLoopBlock = nullptr;
1206  for (unsigned i = 0, e = NewSI->getNumSuccessors(); i != e; ++i)
1207  if (NewSI->getSuccessor(i) == BB) {
1208  if (!InfLoopBlock) {
1209  // Insert it at the end of the function, because it's either code,
1210  // or it won't matter if it's hot. :)
1211  InfLoopBlock = BasicBlock::Create(BB->getContext(), "infloop",
1212  BB->getParent());
1213  BranchInst::Create(InfLoopBlock, InfLoopBlock);
1214  }
1215  NewSI->setSuccessor(i, InfLoopBlock);
1216  }
1217 
1218  Changed = true;
1219  }
1220  }
1221  return Changed;
1222 }
1223 
1224 // If we would need to insert a select that uses the value of this invoke
1225 // (comments in HoistThenElseCodeToIf explain why we would need to do this), we
1226 // can't hoist the invoke, as there is nowhere to put the select in this case.
1228  Instruction *I1, Instruction *I2) {
1229  for (BasicBlock *Succ : successors(BB1)) {
1230  for (const PHINode &PN : Succ->phis()) {
1231  Value *BB1V = PN.getIncomingValueForBlock(BB1);
1232  Value *BB2V = PN.getIncomingValueForBlock(BB2);
1233  if (BB1V != BB2V && (BB1V == I1 || BB2V == I2)) {
1234  return false;
1235  }
1236  }
1237  }
1238  return true;
1239 }
1240 
1242 
1243 /// Given a conditional branch that goes to BB1 and BB2, hoist any common code
1244 /// in the two blocks up into the branch block. The caller of this function
1245 /// guarantees that BI's block dominates BB1 and BB2.
1247  const TargetTransformInfo &TTI) {
1248  // This does very trivial matching, with limited scanning, to find identical
1249  // instructions in the two blocks. In particular, we don't want to get into
1250  // O(M*N) situations here where M and N are the sizes of BB1 and BB2. As
1251  // such, we currently just scan for obviously identical instructions in an
1252  // identical order.
1253  BasicBlock *BB1 = BI->getSuccessor(0); // The true destination.
1254  BasicBlock *BB2 = BI->getSuccessor(1); // The false destination
1255 
1256  BasicBlock::iterator BB1_Itr = BB1->begin();
1257  BasicBlock::iterator BB2_Itr = BB2->begin();
1258 
1259  Instruction *I1 = &*BB1_Itr++, *I2 = &*BB2_Itr++;
1260  // Skip debug info if it is not identical.
1263  if (!DBI1 || !DBI2 || !DBI1->isIdenticalToWhenDefined(DBI2)) {
1264  while (isa<DbgInfoIntrinsic>(I1))
1265  I1 = &*BB1_Itr++;
1266  while (isa<DbgInfoIntrinsic>(I2))
1267  I2 = &*BB2_Itr++;
1268  }
1269  if (isa<PHINode>(I1) || !I1->isIdenticalToWhenDefined(I2) ||
1270  (isa<InvokeInst>(I1) && !isSafeToHoistInvoke(BB1, BB2, I1, I2)))
1271  return false;
1272 
1273  BasicBlock *BIParent = BI->getParent();
1274 
1275  bool Changed = false;
1276  do {
1277  // If we are hoisting the terminator instruction, don't move one (making a
1278  // broken BB), instead clone it, and remove BI.
1279  if (I1->isTerminator())
1280  goto HoistTerminator;
1281 
1282  // If we're going to hoist a call, make sure that the two instructions we're
1283  // commoning/hoisting are both marked with musttail, or neither of them is
1284  // marked as such. Otherwise, we might end up in a situation where we hoist
1285  // from a block where the terminator is a `ret` to a block where the terminator
1286  // is a `br`, and `musttail` calls expect to be followed by a return.
1287  auto *C1 = dyn_cast<CallInst>(I1);
1288  auto *C2 = dyn_cast<CallInst>(I2);
1289  if (C1 && C2)
1290  if (C1->isMustTailCall() != C2->isMustTailCall())
1291  return Changed;
1292 
1293  if (!TTI.isProfitableToHoist(I1) || !TTI.isProfitableToHoist(I2))
1294  return Changed;
1295 
1296  if (isa<DbgInfoIntrinsic>(I1) || isa<DbgInfoIntrinsic>(I2)) {
1297  assert (isa<DbgInfoIntrinsic>(I1) && isa<DbgInfoIntrinsic>(I2));
1298  // The debug location is an integral part of a debug info intrinsic
1299  // and can't be separated from it or replaced. Instead of attempting
1300  // to merge locations, simply hoist both copies of the intrinsic.
1301  BIParent->getInstList().splice(BI->getIterator(),
1302  BB1->getInstList(), I1);
1303  BIParent->getInstList().splice(BI->getIterator(),
1304  BB2->getInstList(), I2);
1305  Changed = true;
1306  } else {
1307  // For a normal instruction, we just move one to right before the branch,
1308  // then replace all uses of the other with the first. Finally, we remove
1309  // the now redundant second instruction.
1310  BIParent->getInstList().splice(BI->getIterator(),
1311  BB1->getInstList(), I1);
1312  if (!I2->use_empty())
1313  I2->replaceAllUsesWith(I1);
1314  I1->andIRFlags(I2);
1315  unsigned KnownIDs[] = {LLVMContext::MD_tbaa,
1325  combineMetadata(I1, I2, KnownIDs, true);
1326 
1327  // I1 and I2 are being combined into a single instruction. Its debug
1328  // location is the merged locations of the original instructions.
1329  I1->applyMergedLocation(I1->getDebugLoc(), I2->getDebugLoc());
1330 
1331  I2->eraseFromParent();
1332  Changed = true;
1333  }
1334 
1335  I1 = &*BB1_Itr++;
1336  I2 = &*BB2_Itr++;
1337  // Skip debug info if it is not identical.
1340  if (!DBI1 || !DBI2 || !DBI1->isIdenticalToWhenDefined(DBI2)) {
1341  while (isa<DbgInfoIntrinsic>(I1))
1342  I1 = &*BB1_Itr++;
1343  while (isa<DbgInfoIntrinsic>(I2))
1344  I2 = &*BB2_Itr++;
1345  }
1346  } while (I1->isIdenticalToWhenDefined(I2));
1347 
1348  return true;
1349 
1350 HoistTerminator:
1351  // It may not be possible to hoist an invoke.
1352  if (isa<InvokeInst>(I1) && !isSafeToHoistInvoke(BB1, BB2, I1, I2))
1353  return Changed;
1354 
1355  for (BasicBlock *Succ : successors(BB1)) {
1356  for (PHINode &PN : Succ->phis()) {
1357  Value *BB1V = PN.getIncomingValueForBlock(BB1);
1358  Value *BB2V = PN.getIncomingValueForBlock(BB2);
1359  if (BB1V == BB2V)
1360  continue;
1361 
1362  // Check for passingValueIsAlwaysUndefined here because we would rather
1363  // eliminate undefined control flow then converting it to a select.
1364  if (passingValueIsAlwaysUndefined(BB1V, &PN) ||
1365  passingValueIsAlwaysUndefined(BB2V, &PN))
1366  return Changed;
1367 
1368  if (isa<ConstantExpr>(BB1V) && !isSafeToSpeculativelyExecute(BB1V))
1369  return Changed;
1370  if (isa<ConstantExpr>(BB2V) && !isSafeToSpeculativelyExecute(BB2V))
1371  return Changed;
1372  }
1373  }
1374 
1375  // As the parent basic block terminator is a branch instruction which is
1376  // removed at the end of the current transformation, use its previous
1377  // non-debug instruction, as the reference insertion point, which will
1378  // provide the debug location for generated select instructions. For BBs
1379  // with only debug instructions, use an empty debug location.
1380  Instruction *InsertPt =
1382 
1383  // Okay, it is safe to hoist the terminator.
1384  Instruction *NT = I1->clone();
1385  BIParent->getInstList().insert(BI->getIterator(), NT);
1386  if (!NT->getType()->isVoidTy()) {
1387  I1->replaceAllUsesWith(NT);
1388  I2->replaceAllUsesWith(NT);
1389  NT->takeName(I1);
1390  }
1391 
1392  // Ensure terminator gets a debug location, even an unknown one, in case
1393  // it involves inlinable calls.
1394  NT->applyMergedLocation(I1->getDebugLoc(), I2->getDebugLoc());
1395 
1396  IRBuilder<NoFolder> Builder(NT);
1397  // If an earlier instruction in this BB had a location, adopt it, otherwise
1398  // clear debug locations.
1399  Builder.SetCurrentDebugLocation(InsertPt ? InsertPt->getDebugLoc()
1400  : DebugLoc());
1401 
1402  // Hoisting one of the terminators from our successor is a great thing.
1403  // Unfortunately, the successors of the if/else blocks may have PHI nodes in
1404  // them. If they do, all PHI entries for BB1/BB2 must agree for all PHI
1405  // nodes, so we insert select instruction to compute the final result.
1406  std::map<std::pair<Value *, Value *>, SelectInst *> InsertedSelects;
1407  for (BasicBlock *Succ : successors(BB1)) {
1408  for (PHINode &PN : Succ->phis()) {
1409  Value *BB1V = PN.getIncomingValueForBlock(BB1);
1410  Value *BB2V = PN.getIncomingValueForBlock(BB2);
1411  if (BB1V == BB2V)
1412  continue;
1413 
1414  // These values do not agree. Insert a select instruction before NT
1415  // that determines the right value.
1416  SelectInst *&SI = InsertedSelects[std::make_pair(BB1V, BB2V)];
1417  if (!SI)
1418  SI = cast<SelectInst>(
1419  Builder.CreateSelect(BI->getCondition(), BB1V, BB2V,
1420  BB1V->getName() + "." + BB2V->getName(), BI));
1421 
1422  // Make the PHI node use the select for all incoming values for BB1/BB2
1423  for (unsigned i = 0, e = PN.getNumIncomingValues(); i != e; ++i)
1424  if (PN.getIncomingBlock(i) == BB1 || PN.getIncomingBlock(i) == BB2)
1425  PN.setIncomingValue(i, SI);
1426  }
1427  }
1428 
1429  // Update any PHI nodes in our new successors.
1430  for (BasicBlock *Succ : successors(BB1))
1431  AddPredecessorToBlock(Succ, BIParent, BB1);
1432 
1434  return true;
1435 }
1436 
1437 // All instructions in Insts belong to different blocks that all unconditionally
1438 // branch to a common successor. Analyze each instruction and return true if it
1439 // would be possible to sink them into their successor, creating one common
1440 // instruction instead. For every value that would be required to be provided by
1441 // PHI node (because an operand varies in each input block), add to PHIOperands.
1444  DenseMap<Instruction *, SmallVector<Value *, 4>> &PHIOperands) {
1445  // Prune out obviously bad instructions to move. Any non-store instruction
1446  // must have exactly one use, and we check later that use is by a single,
1447  // common PHI instruction in the successor.
1448  for (auto *I : Insts) {
1449  // These instructions may change or break semantics if moved.
1450  if (isa<PHINode>(I) || I->isEHPad() || isa<AllocaInst>(I) ||
1451  I->getType()->isTokenTy())
1452  return false;
1453 
1454  // Conservatively return false if I is an inline-asm instruction. Sinking
1455  // and merging inline-asm instructions can potentially create arguments
1456  // that cannot satisfy the inline-asm constraints.
1457  if (const auto *C = dyn_cast<CallInst>(I))
1458  if (C->isInlineAsm())
1459  return false;
1460 
1461  // Everything must have only one use too, apart from stores which
1462  // have no uses.
1463  if (!isa<StoreInst>(I) && !I->hasOneUse())
1464  return false;
1465  }
1466 
1467  const Instruction *I0 = Insts.front();
1468  for (auto *I : Insts)
1469  if (!I->isSameOperationAs(I0))
1470  return false;
1471 
1472  // All instructions in Insts are known to be the same opcode. If they aren't
1473  // stores, check the only user of each is a PHI or in the same block as the
1474  // instruction, because if a user is in the same block as an instruction
1475  // we're contemplating sinking, it must already be determined to be sinkable.
1476  if (!isa<StoreInst>(I0)) {
1477  auto *PNUse = dyn_cast<PHINode>(*I0->user_begin());
1478  auto *Succ = I0->getParent()->getTerminator()->getSuccessor(0);
1479  if (!all_of(Insts, [&PNUse,&Succ](const Instruction *I) -> bool {
1480  auto *U = cast<Instruction>(*I->user_begin());
1481  return (PNUse &&
1482  PNUse->getParent() == Succ &&
1483  PNUse->getIncomingValueForBlock(I->getParent()) == I) ||
1484  U->getParent() == I->getParent();
1485  }))
1486  return false;
1487  }
1488 
1489  // Because SROA can't handle speculating stores of selects, try not
1490  // to sink loads or stores of allocas when we'd have to create a PHI for
1491  // the address operand. Also, because it is likely that loads or stores
1492  // of allocas will disappear when Mem2Reg/SROA is run, don't sink them.
1493  // This can cause code churn which can have unintended consequences down
1494  // the line - see https://llvm.org/bugs/show_bug.cgi?id=30244.
1495  // FIXME: This is a workaround for a deficiency in SROA - see
1496  // https://llvm.org/bugs/show_bug.cgi?id=30188
1497  if (isa<StoreInst>(I0) && any_of(Insts, [](const Instruction *I) {
1498  return isa<AllocaInst>(I->getOperand(1));
1499  }))
1500  return false;
1501  if (isa<LoadInst>(I0) && any_of(Insts, [](const Instruction *I) {
1502  return isa<AllocaInst>(I->getOperand(0));
1503  }))
1504  return false;
1505 
1506  for (unsigned OI = 0, OE = I0->getNumOperands(); OI != OE; ++OI) {
1507  if (I0->getOperand(OI)->getType()->isTokenTy())
1508  // Don't touch any operand of token type.
1509  return false;
1510 
1511  auto SameAsI0 = [&I0, OI](const Instruction *I) {
1512  assert(I->getNumOperands() == I0->getNumOperands());
1513  return I->getOperand(OI) == I0->getOperand(OI);
1514  };
1515  if (!all_of(Insts, SameAsI0)) {
1516  if (!canReplaceOperandWithVariable(I0, OI))
1517  // We can't create a PHI from this GEP.
1518  return false;
1519  // Don't create indirect calls! The called value is the final operand.
1520  if ((isa<CallInst>(I0) || isa<InvokeInst>(I0)) && OI == OE - 1) {
1521  // FIXME: if the call was *already* indirect, we should do this.
1522  return false;
1523  }
1524  for (auto *I : Insts)
1525  PHIOperands[I].push_back(I->getOperand(OI));
1526  }
1527  }
1528  return true;
1529 }
1530 
1531 // Assuming canSinkLastInstruction(Blocks) has returned true, sink the last
1532 // instruction of every block in Blocks to their common successor, commoning
1533 // into one instruction.
1535  auto *BBEnd = Blocks[0]->getTerminator()->getSuccessor(0);
1536 
1537  // canSinkLastInstruction returning true guarantees that every block has at
1538  // least one non-terminator instruction.
1540  for (auto *BB : Blocks) {
1541  Instruction *I = BB->getTerminator();
1542  do {
1543  I = I->getPrevNode();
1544  } while (isa<DbgInfoIntrinsic>(I) && I != &BB->front());
1545  if (!isa<DbgInfoIntrinsic>(I))
1546  Insts.push_back(I);
1547  }
1548 
1549  // The only checking we need to do now is that all users of all instructions
1550  // are the same PHI node. canSinkLastInstruction should have checked this but
1551  // it is slightly over-aggressive - it gets confused by commutative instructions
1552  // so double-check it here.
1553  Instruction *I0 = Insts.front();
1554  if (!isa<StoreInst>(I0)) {
1555  auto *PNUse = dyn_cast<PHINode>(*I0->user_begin());
1556  if (!all_of(Insts, [&PNUse](const Instruction *I) -> bool {
1557  auto *U = cast<Instruction>(*I->user_begin());
1558  return U == PNUse;
1559  }))
1560  return false;
1561  }
1562 
1563  // We don't need to do any more checking here; canSinkLastInstruction should
1564  // have done it all for us.
1565  SmallVector<Value*, 4> NewOperands;
1566  for (unsigned O = 0, E = I0->getNumOperands(); O != E; ++O) {
1567  // This check is different to that in canSinkLastInstruction. There, we
1568  // cared about the global view once simplifycfg (and instcombine) have
1569  // completed - it takes into account PHIs that become trivially
1570  // simplifiable. However here we need a more local view; if an operand
1571  // differs we create a PHI and rely on instcombine to clean up the very
1572  // small mess we may make.
1573  bool NeedPHI = any_of(Insts, [&I0, O](const Instruction *I) {
1574  return I->getOperand(O) != I0->getOperand(O);
1575  });
1576  if (!NeedPHI) {
1577  NewOperands.push_back(I0->getOperand(O));
1578  continue;
1579  }
1580 
1581  // Create a new PHI in the successor block and populate it.
1582  auto *Op = I0->getOperand(O);
1583  assert(!Op->getType()->isTokenTy() && "Can't PHI tokens!");
1584  auto *PN = PHINode::Create(Op->getType(), Insts.size(),
1585  Op->getName() + ".sink", &BBEnd->front());
1586  for (auto *I : Insts)
1587  PN->addIncoming(I->getOperand(O), I->getParent());
1588  NewOperands.push_back(PN);
1589  }
1590 
1591  // Arbitrarily use I0 as the new "common" instruction; remap its operands
1592  // and move it to the start of the successor block.
1593  for (unsigned O = 0, E = I0->getNumOperands(); O != E; ++O)
1594  I0->getOperandUse(O).set(NewOperands[O]);
1595  I0->moveBefore(&*BBEnd->getFirstInsertionPt());
1596 
1597  // Update metadata and IR flags, and merge debug locations.
1598  for (auto *I : Insts)
1599  if (I != I0) {
1600  // The debug location for the "common" instruction is the merged locations
1601  // of all the commoned instructions. We start with the original location
1602  // of the "common" instruction and iteratively merge each location in the
1603  // loop below.
1604  // This is an N-way merge, which will be inefficient if I0 is a CallInst.
1605  // However, as N-way merge for CallInst is rare, so we use simplified API
1606  // instead of using complex API for N-way merge.
1607  I0->applyMergedLocation(I0->getDebugLoc(), I->getDebugLoc());
1608  combineMetadataForCSE(I0, I, true);
1609  I0->andIRFlags(I);
1610  }
1611 
1612  if (!isa<StoreInst>(I0)) {
1613  // canSinkLastInstruction checked that all instructions were used by
1614  // one and only one PHI node. Find that now, RAUW it to our common
1615  // instruction and nuke it.
1616  assert(I0->hasOneUse());
1617  auto *PN = cast<PHINode>(*I0->user_begin());
1618  PN->replaceAllUsesWith(I0);
1619  PN->eraseFromParent();
1620  }
1621 
1622  // Finally nuke all instructions apart from the common instruction.
1623  for (auto *I : Insts)
1624  if (I != I0)
1625  I->eraseFromParent();
1626 
1627  return true;
1628 }
1629 
1630 namespace {
1631 
1632  // LockstepReverseIterator - Iterates through instructions
1633  // in a set of blocks in reverse order from the first non-terminator.
1634  // For example (assume all blocks have size n):
1635  // LockstepReverseIterator I([B1, B2, B3]);
1636  // *I-- = [B1[n], B2[n], B3[n]];
1637  // *I-- = [B1[n-1], B2[n-1], B3[n-1]];
1638  // *I-- = [B1[n-2], B2[n-2], B3[n-2]];
1639  // ...
1640  class LockstepReverseIterator {
1641  ArrayRef<BasicBlock*> Blocks;
1643  bool Fail;
1644 
1645  public:
1646  LockstepReverseIterator(ArrayRef<BasicBlock*> Blocks) : Blocks(Blocks) {
1647  reset();
1648  }
1649 
1650  void reset() {
1651  Fail = false;
1652  Insts.clear();
1653  for (auto *BB : Blocks) {
1654  Instruction *Inst = BB->getTerminator();
1655  for (Inst = Inst->getPrevNode(); Inst && isa<DbgInfoIntrinsic>(Inst);)
1656  Inst = Inst->getPrevNode();
1657  if (!Inst) {
1658  // Block wasn't big enough.
1659  Fail = true;
1660  return;
1661  }
1662  Insts.push_back(Inst);
1663  }
1664  }
1665 
1666  bool isValid() const {
1667  return !Fail;
1668  }
1669 
1670  void operator--() {
1671  if (Fail)
1672  return;
1673  for (auto *&Inst : Insts) {
1674  for (Inst = Inst->getPrevNode(); Inst && isa<DbgInfoIntrinsic>(Inst);)
1675  Inst = Inst->getPrevNode();
1676  // Already at beginning of block.
1677  if (!Inst) {
1678  Fail = true;
1679  return;
1680  }
1681  }
1682  }
1683 
1685  return Insts;
1686  }
1687  };
1688 
1689 } // end anonymous namespace
1690 
1691 /// Check whether BB's predecessors end with unconditional branches. If it is
1692 /// true, sink any common code from the predecessors to BB.
1693 /// We also allow one predecessor to end with conditional branch (but no more
1694 /// than one).
1696  // We support two situations:
1697  // (1) all incoming arcs are unconditional
1698  // (2) one incoming arc is conditional
1699  //
1700  // (2) is very common in switch defaults and
1701  // else-if patterns;
1702  //
1703  // if (a) f(1);
1704  // else if (b) f(2);
1705  //
1706  // produces:
1707  //
1708  // [if]
1709  // / \
1710  // [f(1)] [if]
1711  // | | \
1712  // | | |
1713  // | [f(2)]|
1714  // \ | /
1715  // [ end ]
1716  //
1717  // [end] has two unconditional predecessor arcs and one conditional. The
1718  // conditional refers to the implicit empty 'else' arc. This conditional
1719  // arc can also be caused by an empty default block in a switch.
1720  //
1721  // In this case, we attempt to sink code from all *unconditional* arcs.
1722  // If we can sink instructions from these arcs (determined during the scan
1723  // phase below) we insert a common successor for all unconditional arcs and
1724  // connect that to [end], to enable sinking:
1725  //
1726  // [if]
1727  // / \
1728  // [x(1)] [if]
1729  // | | \
1730  // | | \
1731  // | [x(2)] |
1732  // \ / |
1733  // [sink.split] |
1734  // \ /
1735  // [ end ]
1736  //
1737  SmallVector<BasicBlock*,4> UnconditionalPreds;
1738  Instruction *Cond = nullptr;
1739  for (auto *B : predecessors(BB)) {
1740  auto *T = B->getTerminator();
1741  if (isa<BranchInst>(T) && cast<BranchInst>(T)->isUnconditional())
1742  UnconditionalPreds.push_back(B);
1743  else if ((isa<BranchInst>(T) || isa<SwitchInst>(T)) && !Cond)
1744  Cond = T;
1745  else
1746  return false;
1747  }
1748  if (UnconditionalPreds.size() < 2)
1749  return false;
1750 
1751  bool Changed = false;
1752  // We take a two-step approach to tail sinking. First we scan from the end of
1753  // each block upwards in lockstep. If the n'th instruction from the end of each
1754  // block can be sunk, those instructions are added to ValuesToSink and we
1755  // carry on. If we can sink an instruction but need to PHI-merge some operands
1756  // (because they're not identical in each instruction) we add these to
1757  // PHIOperands.
1758  unsigned ScanIdx = 0;
1759  SmallPtrSet<Value*,4> InstructionsToSink;
1761  LockstepReverseIterator LRI(UnconditionalPreds);
1762  while (LRI.isValid() &&
1763  canSinkInstructions(*LRI, PHIOperands)) {
1764  LLVM_DEBUG(dbgs() << "SINK: instruction can be sunk: " << *(*LRI)[0]
1765  << "\n");
1766  InstructionsToSink.insert((*LRI).begin(), (*LRI).end());
1767  ++ScanIdx;
1768  --LRI;
1769  }
1770 
1771  auto ProfitableToSinkInstruction = [&](LockstepReverseIterator &LRI) {
1772  unsigned NumPHIdValues = 0;
1773  for (auto *I : *LRI)
1774  for (auto *V : PHIOperands[I])
1775  if (InstructionsToSink.count(V) == 0)
1776  ++NumPHIdValues;
1777  LLVM_DEBUG(dbgs() << "SINK: #phid values: " << NumPHIdValues << "\n");
1778  unsigned NumPHIInsts = NumPHIdValues / UnconditionalPreds.size();
1779  if ((NumPHIdValues % UnconditionalPreds.size()) != 0)
1780  NumPHIInsts++;
1781 
1782  return NumPHIInsts <= 1;
1783  };
1784 
1785  if (ScanIdx > 0 && Cond) {
1786  // Check if we would actually sink anything first! This mutates the CFG and
1787  // adds an extra block. The goal in doing this is to allow instructions that
1788  // couldn't be sunk before to be sunk - obviously, speculatable instructions
1789  // (such as trunc, add) can be sunk and predicated already. So we check that
1790  // we're going to sink at least one non-speculatable instruction.
1791  LRI.reset();
1792  unsigned Idx = 0;
1793  bool Profitable = false;
1794  while (ProfitableToSinkInstruction(LRI) && Idx < ScanIdx) {
1795  if (!isSafeToSpeculativelyExecute((*LRI)[0])) {
1796  Profitable = true;
1797  break;
1798  }
1799  --LRI;
1800  ++Idx;
1801  }
1802  if (!Profitable)
1803  return false;
1804 
1805  LLVM_DEBUG(dbgs() << "SINK: Splitting edge\n");
1806  // We have a conditional edge and we're going to sink some instructions.
1807  // Insert a new block postdominating all blocks we're going to sink from.
1808  if (!SplitBlockPredecessors(BB, UnconditionalPreds, ".sink.split"))
1809  // Edges couldn't be split.
1810  return false;
1811  Changed = true;
1812  }
1813 
1814  // Now that we've analyzed all potential sinking candidates, perform the
1815  // actual sink. We iteratively sink the last non-terminator of the source
1816  // blocks into their common successor unless doing so would require too
1817  // many PHI instructions to be generated (currently only one PHI is allowed
1818  // per sunk instruction).
1819  //
1820  // We can use InstructionsToSink to discount values needing PHI-merging that will
1821  // actually be sunk in a later iteration. This allows us to be more
1822  // aggressive in what we sink. This does allow a false positive where we
1823  // sink presuming a later value will also be sunk, but stop half way through
1824  // and never actually sink it which means we produce more PHIs than intended.
1825  // This is unlikely in practice though.
1826  for (unsigned SinkIdx = 0; SinkIdx != ScanIdx; ++SinkIdx) {
1827  LLVM_DEBUG(dbgs() << "SINK: Sink: "
1828  << *UnconditionalPreds[0]->getTerminator()->getPrevNode()
1829  << "\n");
1830 
1831  // Because we've sunk every instruction in turn, the current instruction to
1832  // sink is always at index 0.
1833  LRI.reset();
1834  if (!ProfitableToSinkInstruction(LRI)) {
1835  // Too many PHIs would be created.
1836  LLVM_DEBUG(
1837  dbgs() << "SINK: stopping here, too many PHIs would be created!\n");
1838  break;
1839  }
1840 
1841  if (!sinkLastInstruction(UnconditionalPreds))
1842  return Changed;
1843  NumSinkCommons++;
1844  Changed = true;
1845  }
1846  return Changed;
1847 }
1848 
1849 /// Determine if we can hoist sink a sole store instruction out of a
1850 /// conditional block.
1851 ///
1852 /// We are looking for code like the following:
1853 /// BrBB:
1854 /// store i32 %add, i32* %arrayidx2
1855 /// ... // No other stores or function calls (we could be calling a memory
1856 /// ... // function).
1857 /// %cmp = icmp ult %x, %y
1858 /// br i1 %cmp, label %EndBB, label %ThenBB
1859 /// ThenBB:
1860 /// store i32 %add5, i32* %arrayidx2
1861 /// br label EndBB
1862 /// EndBB:
1863 /// ...
1864 /// We are going to transform this into:
1865 /// BrBB:
1866 /// store i32 %add, i32* %arrayidx2
1867 /// ... //
1868 /// %cmp = icmp ult %x, %y
1869 /// %add.add5 = select i1 %cmp, i32 %add, %add5
1870 /// store i32 %add.add5, i32* %arrayidx2
1871 /// ...
1872 ///
1873 /// \return The pointer to the value of the previous store if the store can be
1874 /// hoisted into the predecessor block. 0 otherwise.
1876  BasicBlock *StoreBB, BasicBlock *EndBB) {
1877  StoreInst *StoreToHoist = dyn_cast<StoreInst>(I);
1878  if (!StoreToHoist)
1879  return nullptr;
1880 
1881  // Volatile or atomic.
1882  if (!StoreToHoist->isSimple())
1883  return nullptr;
1884 
1885  Value *StorePtr = StoreToHoist->getPointerOperand();
1886 
1887  // Look for a store to the same pointer in BrBB.
1888  unsigned MaxNumInstToLookAt = 9;
1889  for (Instruction &CurI : reverse(BrBB->instructionsWithoutDebug())) {
1890  if (!MaxNumInstToLookAt)
1891  break;
1892  --MaxNumInstToLookAt;
1893 
1894  // Could be calling an instruction that affects memory like free().
1895  if (CurI.mayHaveSideEffects() && !isa<StoreInst>(CurI))
1896  return nullptr;
1897 
1898  if (auto *SI = dyn_cast<StoreInst>(&CurI)) {
1899  // Found the previous store make sure it stores to the same location.
1900  if (SI->getPointerOperand() == StorePtr)
1901  // Found the previous store, return its value operand.
1902  return SI->getValueOperand();
1903  return nullptr; // Unknown store.
1904  }
1905  }
1906 
1907  return nullptr;
1908 }
1909 
1910 /// Speculate a conditional basic block flattening the CFG.
1911 ///
1912 /// Note that this is a very risky transform currently. Speculating
1913 /// instructions like this is most often not desirable. Instead, there is an MI
1914 /// pass which can do it with full awareness of the resource constraints.
1915 /// However, some cases are "obvious" and we should do directly. An example of
1916 /// this is speculating a single, reasonably cheap instruction.
1917 ///
1918 /// There is only one distinct advantage to flattening the CFG at the IR level:
1919 /// it makes very common but simplistic optimizations such as are common in
1920 /// instcombine and the DAG combiner more powerful by removing CFG edges and
1921 /// modeling their effects with easier to reason about SSA value graphs.
1922 ///
1923 ///
1924 /// An illustration of this transform is turning this IR:
1925 /// \code
1926 /// BB:
1927 /// %cmp = icmp ult %x, %y
1928 /// br i1 %cmp, label %EndBB, label %ThenBB
1929 /// ThenBB:
1930 /// %sub = sub %x, %y
1931 /// br label BB2
1932 /// EndBB:
1933 /// %phi = phi [ %sub, %ThenBB ], [ 0, %EndBB ]
1934 /// ...
1935 /// \endcode
1936 ///
1937 /// Into this IR:
1938 /// \code
1939 /// BB:
1940 /// %cmp = icmp ult %x, %y
1941 /// %sub = sub %x, %y
1942 /// %cond = select i1 %cmp, 0, %sub
1943 /// ...
1944 /// \endcode
1945 ///
1946 /// \returns true if the conditional block is removed.
1948  const TargetTransformInfo &TTI) {
1949  // Be conservative for now. FP select instruction can often be expensive.
1950  Value *BrCond = BI->getCondition();
1951  if (isa<FCmpInst>(BrCond))
1952  return false;
1953 
1954  BasicBlock *BB = BI->getParent();
1955  BasicBlock *EndBB = ThenBB->getTerminator()->getSuccessor(0);
1956 
1957  // If ThenBB is actually on the false edge of the conditional branch, remember
1958  // to swap the select operands later.
1959  bool Invert = false;
1960  if (ThenBB != BI->getSuccessor(0)) {
1961  assert(ThenBB == BI->getSuccessor(1) && "No edge from 'if' block?");
1962  Invert = true;
1963  }
1964  assert(EndBB == BI->getSuccessor(!Invert) && "No edge from to end block");
1965 
1966  // Keep a count of how many times instructions are used within ThenBB when
1967  // they are candidates for sinking into ThenBB. Specifically:
1968  // - They are defined in BB, and
1969  // - They have no side effects, and
1970  // - All of their uses are in ThenBB.
1971  SmallDenseMap<Instruction *, unsigned, 4> SinkCandidateUseCounts;
1972 
1973  SmallVector<Instruction *, 4> SpeculatedDbgIntrinsics;
1974 
1975  unsigned SpeculationCost = 0;
1976  Value *SpeculatedStoreValue = nullptr;
1977  StoreInst *SpeculatedStore = nullptr;
1978  for (BasicBlock::iterator BBI = ThenBB->begin(),
1979  BBE = std::prev(ThenBB->end());
1980  BBI != BBE; ++BBI) {
1981  Instruction *I = &*BBI;
1982  // Skip debug info.
1983  if (isa<DbgInfoIntrinsic>(I)) {
1984  SpeculatedDbgIntrinsics.push_back(I);
1985  continue;
1986  }
1987 
1988  // Only speculatively execute a single instruction (not counting the
1989  // terminator) for now.
1990  ++SpeculationCost;
1991  if (SpeculationCost > 1)
1992  return false;
1993 
1994  // Don't hoist the instruction if it's unsafe or expensive.
1995  if (!isSafeToSpeculativelyExecute(I) &&
1996  !(HoistCondStores && (SpeculatedStoreValue = isSafeToSpeculateStore(
1997  I, BB, ThenBB, EndBB))))
1998  return false;
1999  if (!SpeculatedStoreValue &&
2000  ComputeSpeculationCost(I, TTI) >
2002  return false;
2003 
2004  // Store the store speculation candidate.
2005  if (SpeculatedStoreValue)
2006  SpeculatedStore = cast<StoreInst>(I);
2007 
2008  // Do not hoist the instruction if any of its operands are defined but not
2009  // used in BB. The transformation will prevent the operand from
2010  // being sunk into the use block.
2011  for (User::op_iterator i = I->op_begin(), e = I->op_end(); i != e; ++i) {
2012  Instruction *OpI = dyn_cast<Instruction>(*i);
2013  if (!OpI || OpI->getParent() != BB || OpI->mayHaveSideEffects())
2014  continue; // Not a candidate for sinking.
2015 
2016  ++SinkCandidateUseCounts[OpI];
2017  }
2018  }
2019 
2020  // Consider any sink candidates which are only used in ThenBB as costs for
2021  // speculation. Note, while we iterate over a DenseMap here, we are summing
2022  // and so iteration order isn't significant.
2024  I = SinkCandidateUseCounts.begin(),
2025  E = SinkCandidateUseCounts.end();
2026  I != E; ++I)
2027  if (I->first->hasNUses(I->second)) {
2028  ++SpeculationCost;
2029  if (SpeculationCost > 1)
2030  return false;
2031  }
2032 
2033  // Check that the PHI nodes can be converted to selects.
2034  bool HaveRewritablePHIs = false;
2035  for (PHINode &PN : EndBB->phis()) {
2036  Value *OrigV = PN.getIncomingValueForBlock(BB);
2037  Value *ThenV = PN.getIncomingValueForBlock(ThenBB);
2038 
2039  // FIXME: Try to remove some of the duplication with HoistThenElseCodeToIf.
2040  // Skip PHIs which are trivial.
2041  if (ThenV == OrigV)
2042  continue;
2043 
2044  // Don't convert to selects if we could remove undefined behavior instead.
2045  if (passingValueIsAlwaysUndefined(OrigV, &PN) ||
2046  passingValueIsAlwaysUndefined(ThenV, &PN))
2047  return false;
2048 
2049  HaveRewritablePHIs = true;
2050  ConstantExpr *OrigCE = dyn_cast<ConstantExpr>(OrigV);
2051  ConstantExpr *ThenCE = dyn_cast<ConstantExpr>(ThenV);
2052  if (!OrigCE && !ThenCE)
2053  continue; // Known safe and cheap.
2054 
2055  if ((ThenCE && !isSafeToSpeculativelyExecute(ThenCE)) ||
2056  (OrigCE && !isSafeToSpeculativelyExecute(OrigCE)))
2057  return false;
2058  unsigned OrigCost = OrigCE ? ComputeSpeculationCost(OrigCE, TTI) : 0;
2059  unsigned ThenCost = ThenCE ? ComputeSpeculationCost(ThenCE, TTI) : 0;
2060  unsigned MaxCost =
2062  if (OrigCost + ThenCost > MaxCost)
2063  return false;
2064 
2065  // Account for the cost of an unfolded ConstantExpr which could end up
2066  // getting expanded into Instructions.
2067  // FIXME: This doesn't account for how many operations are combined in the
2068  // constant expression.
2069  ++SpeculationCost;
2070  if (SpeculationCost > 1)
2071  return false;
2072  }
2073 
2074  // If there are no PHIs to process, bail early. This helps ensure idempotence
2075  // as well.
2076  if (!HaveRewritablePHIs && !(HoistCondStores && SpeculatedStoreValue))
2077  return false;
2078 
2079  // If we get here, we can hoist the instruction and if-convert.
2080  LLVM_DEBUG(dbgs() << "SPECULATIVELY EXECUTING BB" << *ThenBB << "\n";);
2081 
2082  // Insert a select of the value of the speculated store.
2083  if (SpeculatedStoreValue) {
2084  IRBuilder<NoFolder> Builder(BI);
2085  Value *TrueV = SpeculatedStore->getValueOperand();
2086  Value *FalseV = SpeculatedStoreValue;
2087  if (Invert)
2088  std::swap(TrueV, FalseV);
2089  Value *S = Builder.CreateSelect(
2090  BrCond, TrueV, FalseV, "spec.store.select", BI);
2091  SpeculatedStore->setOperand(0, S);
2092  SpeculatedStore->applyMergedLocation(BI->getDebugLoc(),
2093  SpeculatedStore->getDebugLoc());
2094  }
2095 
2096  // Metadata can be dependent on the condition we are hoisting above.
2097  // Conservatively strip all metadata on the instruction.
2098  for (auto &I : *ThenBB)
2100 
2101  // Hoist the instructions.
2102  BB->getInstList().splice(BI->getIterator(), ThenBB->getInstList(),
2103  ThenBB->begin(), std::prev(ThenBB->end()));
2104 
2105  // Insert selects and rewrite the PHI operands.
2106  IRBuilder<NoFolder> Builder(BI);
2107  for (PHINode &PN : EndBB->phis()) {
2108  unsigned OrigI = PN.getBasicBlockIndex(BB);
2109  unsigned ThenI = PN.getBasicBlockIndex(ThenBB);
2110  Value *OrigV = PN.getIncomingValue(OrigI);
2111  Value *ThenV = PN.getIncomingValue(ThenI);
2112 
2113  // Skip PHIs which are trivial.
2114  if (OrigV == ThenV)
2115  continue;
2116 
2117  // Create a select whose true value is the speculatively executed value and
2118  // false value is the preexisting value. Swap them if the branch
2119  // destinations were inverted.
2120  Value *TrueV = ThenV, *FalseV = OrigV;
2121  if (Invert)
2122  std::swap(TrueV, FalseV);
2123  Value *V = Builder.CreateSelect(
2124  BrCond, TrueV, FalseV, "spec.select", BI);
2125  PN.setIncomingValue(OrigI, V);
2126  PN.setIncomingValue(ThenI, V);
2127  }
2128 
2129  // Remove speculated dbg intrinsics.
2130  // FIXME: Is it possible to do this in a more elegant way? Moving/merging the
2131  // dbg value for the different flows and inserting it after the select.
2132  for (Instruction *I : SpeculatedDbgIntrinsics)
2133  I->eraseFromParent();
2134 
2135  ++NumSpeculations;
2136  return true;
2137 }
2138 
2139 /// Return true if we can thread a branch across this block.
2141  unsigned Size = 0;
2142 
2143  for (Instruction &I : BB->instructionsWithoutDebug()) {
2144  if (Size > 10)
2145  return false; // Don't clone large BB's.
2146  ++Size;
2147 
2148  // We can only support instructions that do not define values that are
2149  // live outside of the current basic block.
2150  for (User *U : I.users()) {
2151  Instruction *UI = cast<Instruction>(U);
2152  if (UI->getParent() != BB || isa<PHINode>(UI))
2153  return false;
2154  }
2155 
2156  // Looks ok, continue checking.
2157  }
2158 
2159  return true;
2160 }
2161 
2162 /// If we have a conditional branch on a PHI node value that is defined in the
2163 /// same block as the branch and if any PHI entries are constants, thread edges
2164 /// corresponding to that entry to be branches to their ultimate destination.
2165 static bool FoldCondBranchOnPHI(BranchInst *BI, const DataLayout &DL,
2166  AssumptionCache *AC) {
2167  BasicBlock *BB = BI->getParent();
2168  PHINode *PN = dyn_cast<PHINode>(BI->getCondition());
2169  // NOTE: we currently cannot transform this case if the PHI node is used
2170  // outside of the block.
2171  if (!PN || PN->getParent() != BB || !PN->hasOneUse())
2172  return false;
2173 
2174  // Degenerate case of a single entry PHI.
2175  if (PN->getNumIncomingValues() == 1) {
2177  return true;
2178  }
2179 
2180  // Now we know that this block has multiple preds and two succs.
2182  return false;
2183 
2184  // Can't fold blocks that contain noduplicate or convergent calls.
2185  if (any_of(*BB, [](const Instruction &I) {
2186  const CallInst *CI = dyn_cast<CallInst>(&I);
2187  return CI && (CI->cannotDuplicate() || CI->isConvergent());
2188  }))
2189  return false;
2190 
2191  // Okay, this is a simple enough basic block. See if any phi values are
2192  // constants.
2193  for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
2195  if (!CB || !CB->getType()->isIntegerTy(1))
2196  continue;
2197 
2198  // Okay, we now know that all edges from PredBB should be revectored to
2199  // branch to RealDest.
2200  BasicBlock *PredBB = PN->getIncomingBlock(i);
2201  BasicBlock *RealDest = BI->getSuccessor(!CB->getZExtValue());
2202 
2203  if (RealDest == BB)
2204  continue; // Skip self loops.
2205  // Skip if the predecessor's terminator is an indirect branch.
2206  if (isa<IndirectBrInst>(PredBB->getTerminator()))
2207  continue;
2208 
2209  // The dest block might have PHI nodes, other predecessors and other
2210  // difficult cases. Instead of being smart about this, just insert a new
2211  // block that jumps to the destination block, effectively splitting
2212  // the edge we are about to create.
2213  BasicBlock *EdgeBB =
2214  BasicBlock::Create(BB->getContext(), RealDest->getName() + ".critedge",
2215  RealDest->getParent(), RealDest);
2216  BranchInst::Create(RealDest, EdgeBB);
2217 
2218  // Update PHI nodes.
2219  AddPredecessorToBlock(RealDest, EdgeBB, BB);
2220 
2221  // BB may have instructions that are being threaded over. Clone these
2222  // instructions into EdgeBB. We know that there will be no uses of the
2223  // cloned instructions outside of EdgeBB.
2224  BasicBlock::iterator InsertPt = EdgeBB->begin();
2225  DenseMap<Value *, Value *> TranslateMap; // Track translated values.
2226  for (BasicBlock::iterator BBI = BB->begin(); &*BBI != BI; ++BBI) {
2227  if (PHINode *PN = dyn_cast<PHINode>(BBI)) {
2228  TranslateMap[PN] = PN->getIncomingValueForBlock(PredBB);
2229  continue;
2230  }
2231  // Clone the instruction.
2232  Instruction *N = BBI->clone();
2233  if (BBI->hasName())
2234  N->setName(BBI->getName() + ".c");
2235 
2236  // Update operands due to translation.
2237  for (User::op_iterator i = N->op_begin(), e = N->op_end(); i != e; ++i) {
2238  DenseMap<Value *, Value *>::iterator PI = TranslateMap.find(*i);
2239  if (PI != TranslateMap.end())
2240  *i = PI->second;
2241  }
2242 
2243  // Check for trivial simplification.
2244  if (Value *V = SimplifyInstruction(N, {DL, nullptr, nullptr, AC})) {
2245  if (!BBI->use_empty())
2246  TranslateMap[&*BBI] = V;
2247  if (!N->mayHaveSideEffects()) {
2248  N->deleteValue(); // Instruction folded away, don't need actual inst
2249  N = nullptr;
2250  }
2251  } else {
2252  if (!BBI->use_empty())
2253  TranslateMap[&*BBI] = N;
2254  }
2255  // Insert the new instruction into its new home.
2256  if (N)
2257  EdgeBB->getInstList().insert(InsertPt, N);
2258 
2259  // Register the new instruction with the assumption cache if necessary.
2260  if (auto *II = dyn_cast_or_null<IntrinsicInst>(N))
2261  if (II->getIntrinsicID() == Intrinsic::assume)
2262  AC->registerAssumption(II);
2263  }
2264 
2265  // Loop over all of the edges from PredBB to BB, changing them to branch
2266  // to EdgeBB instead.
2267  Instruction *PredBBTI = PredBB->getTerminator();
2268  for (unsigned i = 0, e = PredBBTI->getNumSuccessors(); i != e; ++i)
2269  if (PredBBTI->getSuccessor(i) == BB) {
2270  BB->removePredecessor(PredBB);
2271  PredBBTI->setSuccessor(i, EdgeBB);
2272  }
2273 
2274  // Recurse, simplifying any other constants.
2275  return FoldCondBranchOnPHI(BI, DL, AC) || true;
2276  }
2277 
2278  return false;
2279 }
2280 
2281 /// Given a BB that starts with the specified two-entry PHI node,
2282 /// see if we can eliminate it.
2284  const DataLayout &DL) {
2285  // Ok, this is a two entry PHI node. Check to see if this is a simple "if
2286  // statement", which has a very simple dominance structure. Basically, we
2287  // are trying to find the condition that is being branched on, which
2288  // subsequently causes this merge to happen. We really want control
2289  // dependence information for this check, but simplifycfg can't keep it up
2290  // to date, and this catches most of the cases we care about anyway.
2291  BasicBlock *BB = PN->getParent();
2292  const Function *Fn = BB->getParent();
2293  if (Fn && Fn->hasFnAttribute(Attribute::OptForFuzzing))
2294  return false;
2295 
2296  BasicBlock *IfTrue, *IfFalse;
2297  Value *IfCond = GetIfCondition(BB, IfTrue, IfFalse);
2298  if (!IfCond ||
2299  // Don't bother if the branch will be constant folded trivially.
2300  isa<ConstantInt>(IfCond))
2301  return false;
2302 
2303  // Okay, we found that we can merge this two-entry phi node into a select.
2304  // Doing so would require us to fold *all* two entry phi nodes in this block.
2305  // At some point this becomes non-profitable (particularly if the target
2306  // doesn't support cmov's). Only do this transformation if there are two or
2307  // fewer PHI nodes in this block.
2308  unsigned NumPhis = 0;
2309  for (BasicBlock::iterator I = BB->begin(); isa<PHINode>(I); ++NumPhis, ++I)
2310  if (NumPhis > 2)
2311  return false;
2312 
2313  // Loop over the PHI's seeing if we can promote them all to select
2314  // instructions. While we are at it, keep track of the instructions
2315  // that need to be moved to the dominating block.
2316  SmallPtrSet<Instruction *, 4> AggressiveInsts;
2317  unsigned MaxCostVal0 = PHINodeFoldingThreshold,
2318  MaxCostVal1 = PHINodeFoldingThreshold;
2319  MaxCostVal0 *= TargetTransformInfo::TCC_Basic;
2320  MaxCostVal1 *= TargetTransformInfo::TCC_Basic;
2321 
2322  for (BasicBlock::iterator II = BB->begin(); isa<PHINode>(II);) {
2323  PHINode *PN = cast<PHINode>(II++);
2324  if (Value *V = SimplifyInstruction(PN, {DL, PN})) {
2325  PN->replaceAllUsesWith(V);
2326  PN->eraseFromParent();
2327  continue;
2328  }
2329 
2330  if (!DominatesMergePoint(PN->getIncomingValue(0), BB, AggressiveInsts,
2331  MaxCostVal0, TTI) ||
2332  !DominatesMergePoint(PN->getIncomingValue(1), BB, AggressiveInsts,
2333  MaxCostVal1, TTI))
2334  return false;
2335  }
2336 
2337  // If we folded the first phi, PN dangles at this point. Refresh it. If
2338  // we ran out of PHIs then we simplified them all.
2339  PN = dyn_cast<PHINode>(BB->begin());
2340  if (!PN)
2341  return true;
2342 
2343  // Don't fold i1 branches on PHIs which contain binary operators. These can
2344  // often be turned into switches and other things.
2345  if (PN->getType()->isIntegerTy(1) &&
2346  (isa<BinaryOperator>(PN->getIncomingValue(0)) ||
2347  isa<BinaryOperator>(PN->getIncomingValue(1)) ||
2348  isa<BinaryOperator>(IfCond)))
2349  return false;
2350 
2351  // If all PHI nodes are promotable, check to make sure that all instructions
2352  // in the predecessor blocks can be promoted as well. If not, we won't be able
2353  // to get rid of the control flow, so it's not worth promoting to select
2354  // instructions.
2355  BasicBlock *DomBlock = nullptr;
2356  BasicBlock *IfBlock1 = PN->getIncomingBlock(0);
2357  BasicBlock *IfBlock2 = PN->getIncomingBlock(1);
2358  if (cast<BranchInst>(IfBlock1->getTerminator())->isConditional()) {
2359  IfBlock1 = nullptr;
2360  } else {
2361  DomBlock = *pred_begin(IfBlock1);
2362  for (BasicBlock::iterator I = IfBlock1->begin(); !I->isTerminator(); ++I)
2363  if (!AggressiveInsts.count(&*I) && !isa<DbgInfoIntrinsic>(I)) {
2364  // This is not an aggressive instruction that we can promote.
2365  // Because of this, we won't be able to get rid of the control flow, so
2366  // the xform is not worth it.
2367  return false;
2368  }
2369  }
2370 
2371  if (cast<BranchInst>(IfBlock2->getTerminator())->isConditional()) {
2372  IfBlock2 = nullptr;
2373  } else {
2374  DomBlock = *pred_begin(IfBlock2);
2375  for (BasicBlock::iterator I = IfBlock2->begin(); !I->isTerminator(); ++I)
2376  if (!AggressiveInsts.count(&*I) && !isa<DbgInfoIntrinsic>(I)) {
2377  // This is not an aggressive instruction that we can promote.
2378  // Because of this, we won't be able to get rid of the control flow, so
2379  // the xform is not worth it.
2380  return false;
2381  }
2382  }
2383 
2384  LLVM_DEBUG(dbgs() << "FOUND IF CONDITION! " << *IfCond
2385  << " T: " << IfTrue->getName()
2386  << " F: " << IfFalse->getName() << "\n");
2387 
2388  // If we can still promote the PHI nodes after this gauntlet of tests,
2389  // do all of the PHI's now.
2390  Instruction *InsertPt = DomBlock->getTerminator();
2391  IRBuilder<NoFolder> Builder(InsertPt);
2392 
2393  // Move all 'aggressive' instructions, which are defined in the
2394  // conditional parts of the if's up to the dominating block.
2395  if (IfBlock1)
2396  hoistAllInstructionsInto(DomBlock, InsertPt, IfBlock1);
2397  if (IfBlock2)
2398  hoistAllInstructionsInto(DomBlock, InsertPt, IfBlock2);
2399 
2400  while (PHINode *PN = dyn_cast<PHINode>(BB->begin())) {
2401  // Change the PHI node into a select instruction.
2402  Value *TrueVal = PN->getIncomingValue(PN->getIncomingBlock(0) == IfFalse);
2403  Value *FalseVal = PN->getIncomingValue(PN->getIncomingBlock(0) == IfTrue);
2404 
2405  Value *Sel = Builder.CreateSelect(IfCond, TrueVal, FalseVal, "", InsertPt);
2406  PN->replaceAllUsesWith(Sel);
2407  Sel->takeName(PN);
2408  PN->eraseFromParent();
2409  }
2410 
2411  // At this point, IfBlock1 and IfBlock2 are both empty, so our if statement
2412  // has been flattened. Change DomBlock to jump directly to our new block to
2413  // avoid other simplifycfg's kicking in on the diamond.
2414  Instruction *OldTI = DomBlock->getTerminator();
2415  Builder.SetInsertPoint(OldTI);
2416  Builder.CreateBr(BB);
2417  OldTI->eraseFromParent();
2418  return true;
2419 }
2420 
2421 /// If we found a conditional branch that goes to two returning blocks,
2422 /// try to merge them together into one return,
2423 /// introducing a select if the return values disagree.
2425  IRBuilder<> &Builder) {
2426  assert(BI->isConditional() && "Must be a conditional branch");
2427  BasicBlock *TrueSucc = BI->getSuccessor(0);
2428  BasicBlock *FalseSucc = BI->getSuccessor(1);
2429  ReturnInst *TrueRet = cast<ReturnInst>(TrueSucc->getTerminator());
2430  ReturnInst *FalseRet = cast<ReturnInst>(FalseSucc->getTerminator());
2431 
2432  // Check to ensure both blocks are empty (just a return) or optionally empty
2433  // with PHI nodes. If there are other instructions, merging would cause extra
2434  // computation on one path or the other.
2435  if (!TrueSucc->getFirstNonPHIOrDbg()->isTerminator())
2436  return false;
2437  if (!FalseSucc->getFirstNonPHIOrDbg()->isTerminator())
2438  return false;
2439 
2440  Builder.SetInsertPoint(BI);
2441  // Okay, we found a branch that is going to two return nodes. If
2442  // there is no return value for this function, just change the
2443  // branch into a return.
2444  if (FalseRet->getNumOperands() == 0) {
2445  TrueSucc->removePredecessor(BI->getParent());
2446  FalseSucc->removePredecessor(BI->getParent());
2447  Builder.CreateRetVoid();
2449  return true;
2450  }
2451 
2452  // Otherwise, figure out what the true and false return values are
2453  // so we can insert a new select instruction.
2454  Value *TrueValue = TrueRet->getReturnValue();
2455  Value *FalseValue = FalseRet->getReturnValue();
2456 
2457  // Unwrap any PHI nodes in the return blocks.
2458  if (PHINode *TVPN = dyn_cast_or_null<PHINode>(TrueValue))
2459  if (TVPN->getParent() == TrueSucc)
2460  TrueValue = TVPN->getIncomingValueForBlock(BI->getParent());
2461  if (PHINode *FVPN = dyn_cast_or_null<PHINode>(FalseValue))
2462  if (FVPN->getParent() == FalseSucc)
2463  FalseValue = FVPN->getIncomingValueForBlock(BI->getParent());
2464 
2465  // In order for this transformation to be safe, we must be able to
2466  // unconditionally execute both operands to the return. This is
2467  // normally the case, but we could have a potentially-trapping
2468  // constant expression that prevents this transformation from being
2469  // safe.
2470  if (ConstantExpr *TCV = dyn_cast_or_null<ConstantExpr>(TrueValue))
2471  if (TCV->canTrap())
2472  return false;
2473  if (ConstantExpr *FCV = dyn_cast_or_null<ConstantExpr>(FalseValue))
2474  if (FCV->canTrap())
2475  return false;
2476 
2477  // Okay, we collected all the mapped values and checked them for sanity, and
2478  // defined to really do this transformation. First, update the CFG.
2479  TrueSucc->removePredecessor(BI->getParent());
2480  FalseSucc->removePredecessor(BI->getParent());
2481 
2482  // Insert select instructions where needed.
2483  Value *BrCond = BI->getCondition();
2484  if (TrueValue) {
2485  // Insert a select if the results differ.
2486  if (TrueValue == FalseValue || isa<UndefValue>(FalseValue)) {
2487  } else if (isa<UndefValue>(TrueValue)) {
2488  TrueValue = FalseValue;
2489  } else {
2490  TrueValue =
2491  Builder.CreateSelect(BrCond, TrueValue, FalseValue, "retval", BI);
2492  }
2493  }
2494 
2495  Value *RI =
2496  !TrueValue ? Builder.CreateRetVoid() : Builder.CreateRet(TrueValue);
2497 
2498  (void)RI;
2499 
2500  LLVM_DEBUG(dbgs() << "\nCHANGING BRANCH TO TWO RETURNS INTO SELECT:"
2501  << "\n " << *BI << "NewRet = " << *RI << "TRUEBLOCK: "
2502  << *TrueSucc << "FALSEBLOCK: " << *FalseSucc);
2503 
2505 
2506  return true;
2507 }
2508 
2509 /// Return true if the given instruction is available
2510 /// in its predecessor block. If yes, the instruction will be removed.
2512  if (!isa<BinaryOperator>(Inst) && !isa<CmpInst>(Inst))
2513  return false;
2514  for (Instruction &I : *PB) {
2515  Instruction *PBI = &I;
2516  // Check whether Inst and PBI generate the same value.
2517  if (Inst->isIdenticalTo(PBI)) {
2518  Inst->replaceAllUsesWith(PBI);
2519  Inst->eraseFromParent();
2520  return true;
2521  }
2522  }
2523  return false;
2524 }
2525 
2526 /// Return true if either PBI or BI has branch weight available, and store
2527 /// the weights in {Pred|Succ}{True|False}Weight. If one of PBI and BI does
2528 /// not have branch weight, use 1:1 as its weight.
2530  uint64_t &PredTrueWeight,
2531  uint64_t &PredFalseWeight,
2532  uint64_t &SuccTrueWeight,
2533  uint64_t &SuccFalseWeight) {
2534  bool PredHasWeights =
2535  PBI->extractProfMetadata(PredTrueWeight, PredFalseWeight);
2536  bool SuccHasWeights =
2537  BI->extractProfMetadata(SuccTrueWeight, SuccFalseWeight);
2538  if (PredHasWeights || SuccHasWeights) {
2539  if (!PredHasWeights)
2540  PredTrueWeight = PredFalseWeight = 1;
2541  if (!SuccHasWeights)
2542  SuccTrueWeight = SuccFalseWeight = 1;
2543  return true;
2544  } else {
2545  return false;
2546  }
2547 }
2548 
2549 /// If this basic block is simple enough, and if a predecessor branches to us
2550 /// and one of our successors, fold the block into the predecessor and use
2551 /// logical operations to pick the right destination.
2552 bool llvm::FoldBranchToCommonDest(BranchInst *BI, unsigned BonusInstThreshold) {
2553  BasicBlock *BB = BI->getParent();
2554 
2555  const unsigned PredCount = pred_size(BB);
2556 
2557  Instruction *Cond = nullptr;
2558  if (BI->isConditional())
2559  Cond = dyn_cast<Instruction>(BI->getCondition());
2560  else {
2561  // For unconditional branch, check for a simple CFG pattern, where
2562  // BB has a single predecessor and BB's successor is also its predecessor's
2563  // successor. If such pattern exists, check for CSE between BB and its
2564  // predecessor.
2565  if (BasicBlock *PB = BB->getSinglePredecessor())
2566  if (BranchInst *PBI = dyn_cast<BranchInst>(PB->getTerminator()))
2567  if (PBI->isConditional() &&
2568  (BI->getSuccessor(0) == PBI->getSuccessor(0) ||
2569  BI->getSuccessor(0) == PBI->getSuccessor(1))) {
2570  for (auto I = BB->instructionsWithoutDebug().begin(),
2571  E = BB->instructionsWithoutDebug().end();
2572  I != E;) {
2573  Instruction *Curr = &*I++;
2574  if (isa<CmpInst>(Curr)) {
2575  Cond = Curr;
2576  break;
2577  }
2578  // Quit if we can't remove this instruction.
2579  if (!tryCSEWithPredecessor(Curr, PB))
2580  return false;
2581  }
2582  }
2583 
2584  if (!Cond)
2585  return false;
2586  }
2587 
2588  if (!Cond || (!isa<CmpInst>(Cond) && !isa<BinaryOperator>(Cond)) ||
2589  Cond->getParent() != BB || !Cond->hasOneUse())
2590  return false;
2591 
2592  // Make sure the instruction after the condition is the cond branch.
2593  BasicBlock::iterator CondIt = ++Cond->getIterator();
2594 
2595  // Ignore dbg intrinsics.
2596  while (isa<DbgInfoIntrinsic>(CondIt))
2597  ++CondIt;
2598 
2599  if (&*CondIt != BI)
2600  return false;
2601 
2602  // Only allow this transformation if computing the condition doesn't involve
2603  // too many instructions and these involved instructions can be executed
2604  // unconditionally. We denote all involved instructions except the condition
2605  // as "bonus instructions", and only allow this transformation when the
2606  // number of the bonus instructions we'll need to create when cloning into
2607  // each predecessor does not exceed a certain threshold.
2608  unsigned NumBonusInsts = 0;
2609  for (auto I = BB->begin(); Cond != &*I; ++I) {
2610  // Ignore dbg intrinsics.
2611  if (isa<DbgInfoIntrinsic>(I))
2612  continue;
2613  if (!I->hasOneUse() || !isSafeToSpeculativelyExecute(&*I))
2614  return false;
2615  // I has only one use and can be executed unconditionally.
2617  if (User == nullptr || User->getParent() != BB)
2618  return false;
2619  // I is used in the same BB. Since BI uses Cond and doesn't have more slots
2620  // to use any other instruction, User must be an instruction between next(I)
2621  // and Cond.
2622 
2623  // Account for the cost of duplicating this instruction into each
2624  // predecessor.
2625  NumBonusInsts += PredCount;
2626  // Early exits once we reach the limit.
2627  if (NumBonusInsts > BonusInstThreshold)
2628  return false;
2629  }
2630 
2631  // Cond is known to be a compare or binary operator. Check to make sure that
2632  // neither operand is a potentially-trapping constant expression.
2633  if (ConstantExpr *CE = dyn_cast<ConstantExpr>(Cond->getOperand(0)))
2634  if (CE->canTrap())
2635  return false;
2636  if (ConstantExpr *CE = dyn_cast<ConstantExpr>(Cond->getOperand(1)))
2637  if (CE->canTrap())
2638  return false;
2639 
2640  // Finally, don't infinitely unroll conditional loops.
2641  BasicBlock *TrueDest = BI->getSuccessor(0);
2642  BasicBlock *FalseDest = (BI->isConditional()) ? BI->getSuccessor(1) : nullptr;
2643  if (TrueDest == BB || FalseDest == BB)
2644  return false;
2645 
2646  for (pred_iterator PI = pred_begin(BB), E = pred_end(BB); PI != E; ++PI) {
2647  BasicBlock *PredBlock = *PI;
2648  BranchInst *PBI = dyn_cast<BranchInst>(PredBlock->getTerminator());
2649 
2650  // Check that we have two conditional branches. If there is a PHI node in
2651  // the common successor, verify that the same value flows in from both
2652  // blocks.
2654  if (!PBI || PBI->isUnconditional() ||
2655  (BI->isConditional() && !SafeToMergeTerminators(BI, PBI)) ||
2656  (!BI->isConditional() &&
2657  !isProfitableToFoldUnconditional(BI, PBI, Cond, PHIs)))
2658  continue;
2659 
2660  // Determine if the two branches share a common destination.
2661  Instruction::BinaryOps Opc = Instruction::BinaryOpsEnd;
2662  bool InvertPredCond = false;
2663 
2664  if (BI->isConditional()) {
2665  if (PBI->getSuccessor(0) == TrueDest) {
2666  Opc = Instruction::Or;
2667  } else if (PBI->getSuccessor(1) == FalseDest) {
2668  Opc = Instruction::And;
2669  } else if (PBI->getSuccessor(0) == FalseDest) {
2670  Opc = Instruction::And;
2671  InvertPredCond = true;
2672  } else if (PBI->getSuccessor(1) == TrueDest) {
2673  Opc = Instruction::Or;
2674  InvertPredCond = true;
2675  } else {
2676  continue;
2677  }
2678  } else {
2679  if (PBI->getSuccessor(0) != TrueDest && PBI->getSuccessor(1) != TrueDest)
2680  continue;
2681  }
2682 
2683  LLVM_DEBUG(dbgs() << "FOLDING BRANCH TO COMMON DEST:\n" << *PBI << *BB);
2684  IRBuilder<> Builder(PBI);
2685 
2686  // If we need to invert the condition in the pred block to match, do so now.
2687  if (InvertPredCond) {
2688  Value *NewCond = PBI->getCondition();
2689 
2690  if (NewCond->hasOneUse() && isa<CmpInst>(NewCond)) {
2691  CmpInst *CI = cast<CmpInst>(NewCond);
2692  CI->setPredicate(CI->getInversePredicate());
2693  } else {
2694  NewCond =
2695  Builder.CreateNot(NewCond, PBI->getCondition()->getName() + ".not");
2696  }
2697 
2698  PBI->setCondition(NewCond);
2699  PBI->swapSuccessors();
2700  }
2701 
2702  // If we have bonus instructions, clone them into the predecessor block.
2703  // Note that there may be multiple predecessor blocks, so we cannot move
2704  // bonus instructions to a predecessor block.
2705  ValueToValueMapTy VMap; // maps original values to cloned values
2706  // We already make sure Cond is the last instruction before BI. Therefore,
2707  // all instructions before Cond other than DbgInfoIntrinsic are bonus
2708  // instructions.
2709  for (auto BonusInst = BB->begin(); Cond != &*BonusInst; ++BonusInst) {
2710  if (isa<DbgInfoIntrinsic>(BonusInst))
2711  continue;
2712  Instruction *NewBonusInst = BonusInst->clone();
2713  RemapInstruction(NewBonusInst, VMap,
2715  VMap[&*BonusInst] = NewBonusInst;
2716 
2717  // If we moved a load, we cannot any longer claim any knowledge about
2718  // its potential value. The previous information might have been valid
2719  // only given the branch precondition.
2720  // For an analogous reason, we must also drop all the metadata whose
2721  // semantics we don't understand.
2722  NewBonusInst->dropUnknownNonDebugMetadata();
2723 
2724  PredBlock->getInstList().insert(PBI->getIterator(), NewBonusInst);
2725  NewBonusInst->takeName(&*BonusInst);
2726  BonusInst->setName(BonusInst->getName() + ".old");
2727  }
2728 
2729  // Clone Cond into the predecessor basic block, and or/and the
2730  // two conditions together.
2731  Instruction *CondInPred = Cond->clone();
2732  RemapInstruction(CondInPred, VMap,
2734  PredBlock->getInstList().insert(PBI->getIterator(), CondInPred);
2735  CondInPred->takeName(Cond);
2736  Cond->setName(CondInPred->getName() + ".old");
2737 
2738  if (BI->isConditional()) {
2739  Instruction *NewCond = cast<Instruction>(
2740  Builder.CreateBinOp(Opc, PBI->getCondition(), CondInPred, "or.cond"));
2741  PBI->setCondition(NewCond);
2742 
2743  uint64_t PredTrueWeight, PredFalseWeight, SuccTrueWeight, SuccFalseWeight;
2744  bool HasWeights =
2745  extractPredSuccWeights(PBI, BI, PredTrueWeight, PredFalseWeight,
2746  SuccTrueWeight, SuccFalseWeight);
2747  SmallVector<uint64_t, 8> NewWeights;
2748 
2749  if (PBI->getSuccessor(0) == BB) {
2750  if (HasWeights) {
2751  // PBI: br i1 %x, BB, FalseDest
2752  // BI: br i1 %y, TrueDest, FalseDest
2753  // TrueWeight is TrueWeight for PBI * TrueWeight for BI.
2754  NewWeights.push_back(PredTrueWeight * SuccTrueWeight);
2755  // FalseWeight is FalseWeight for PBI * TotalWeight for BI +
2756  // TrueWeight for PBI * FalseWeight for BI.
2757  // We assume that total weights of a BranchInst can fit into 32 bits.
2758  // Therefore, we will not have overflow using 64-bit arithmetic.
2759  NewWeights.push_back(PredFalseWeight *
2760  (SuccFalseWeight + SuccTrueWeight) +
2761  PredTrueWeight * SuccFalseWeight);
2762  }
2763  AddPredecessorToBlock(TrueDest, PredBlock, BB);
2764  PBI->setSuccessor(0, TrueDest);
2765  }
2766  if (PBI->getSuccessor(1) == BB) {
2767  if (HasWeights) {
2768  // PBI: br i1 %x, TrueDest, BB
2769  // BI: br i1 %y, TrueDest, FalseDest
2770  // TrueWeight is TrueWeight for PBI * TotalWeight for BI +
2771  // FalseWeight for PBI * TrueWeight for BI.
2772  NewWeights.push_back(PredTrueWeight *
2773  (SuccFalseWeight + SuccTrueWeight) +
2774  PredFalseWeight * SuccTrueWeight);
2775  // FalseWeight is FalseWeight for PBI * FalseWeight for BI.
2776  NewWeights.push_back(PredFalseWeight * SuccFalseWeight);
2777  }
2778  AddPredecessorToBlock(FalseDest, PredBlock, BB);
2779  PBI->setSuccessor(1, FalseDest);
2780  }
2781  if (NewWeights.size() == 2) {
2782  // Halve the weights if any of them cannot fit in an uint32_t
2783  FitWeights(NewWeights);
2784 
2785  SmallVector<uint32_t, 8> MDWeights(NewWeights.begin(),
2786  NewWeights.end());
2787  setBranchWeights(PBI, MDWeights[0], MDWeights[1]);
2788  } else
2789  PBI->setMetadata(LLVMContext::MD_prof, nullptr);
2790  } else {
2791  // Update PHI nodes in the common successors.
2792  for (unsigned i = 0, e = PHIs.size(); i != e; ++i) {
2793  ConstantInt *PBI_C = cast<ConstantInt>(
2794  PHIs[i]->getIncomingValueForBlock(PBI->getParent()));
2795  assert(PBI_C->getType()->isIntegerTy(1));
2796  Instruction *MergedCond = nullptr;
2797  if (PBI->getSuccessor(0) == TrueDest) {
2798  // Create (PBI_Cond and PBI_C) or (!PBI_Cond and BI_Value)
2799  // PBI_C is true: PBI_Cond or (!PBI_Cond and BI_Value)
2800  // is false: !PBI_Cond and BI_Value
2801  Instruction *NotCond = cast<Instruction>(
2802  Builder.CreateNot(PBI->getCondition(), "not.cond"));
2803  MergedCond = cast<Instruction>(
2804  Builder.CreateBinOp(Instruction::And, NotCond, CondInPred,
2805  "and.cond"));
2806  if (PBI_C->isOne())
2807  MergedCond = cast<Instruction>(Builder.CreateBinOp(
2808  Instruction::Or, PBI->getCondition(), MergedCond, "or.cond"));
2809  } else {
2810  // Create (PBI_Cond and BI_Value) or (!PBI_Cond and PBI_C)
2811  // PBI_C is true: (PBI_Cond and BI_Value) or (!PBI_Cond)
2812  // is false: PBI_Cond and BI_Value
2813  MergedCond = cast<Instruction>(Builder.CreateBinOp(
2814  Instruction::And, PBI->getCondition(), CondInPred, "and.cond"));
2815  if (PBI_C->isOne()) {
2816  Instruction *NotCond = cast<Instruction>(
2817  Builder.CreateNot(PBI->getCondition(), "not.cond"));
2818  MergedCond = cast<Instruction>(Builder.CreateBinOp(
2819  Instruction::Or, NotCond, MergedCond, "or.cond"));
2820  }
2821  }
2822  // Update PHI Node.
2823  PHIs[i]->setIncomingValue(PHIs[i]->getBasicBlockIndex(PBI->getParent()),
2824  MergedCond);
2825  }
2826  // Change PBI from Conditional to Unconditional.
2827  BranchInst *New_PBI = BranchInst::Create(TrueDest, PBI);
2829  PBI = New_PBI;
2830  }
2831 
2832  // If BI was a loop latch, it may have had associated loop metadata.
2833  // We need to copy it to the new latch, that is, PBI.
2834  if (MDNode *LoopMD = BI->getMetadata(LLVMContext::MD_loop))
2835  PBI->setMetadata(LLVMContext::MD_loop, LoopMD);
2836 
2837  // TODO: If BB is reachable from all paths through PredBlock, then we
2838  // could replace PBI's branch probabilities with BI's.
2839 
2840  // Copy any debug value intrinsics into the end of PredBlock.
2841  for (Instruction &I : *BB)
2842  if (isa<DbgInfoIntrinsic>(I))
2843  I.clone()->insertBefore(PBI);
2844 
2845  return true;
2846  }
2847  return false;
2848 }
2849 
2850 // If there is only one store in BB1 and BB2, return it, otherwise return
2851 // nullptr.
2853  StoreInst *S = nullptr;
2854  for (auto *BB : {BB1, BB2}) {
2855  if (!BB)
2856  continue;
2857  for (auto &I : *BB)
2858  if (auto *SI = dyn_cast<StoreInst>(&I)) {
2859  if (S)
2860  // Multiple stores seen.
2861  return nullptr;
2862  else
2863  S = SI;
2864  }
2865  }
2866  return S;
2867 }
2868 
2870  Value *AlternativeV = nullptr) {
2871  // PHI is going to be a PHI node that allows the value V that is defined in
2872  // BB to be referenced in BB's only successor.
2873  //
2874  // If AlternativeV is nullptr, the only value we care about in PHI is V. It
2875  // doesn't matter to us what the other operand is (it'll never get used). We
2876  // could just create a new PHI with an undef incoming value, but that could
2877  // increase register pressure if EarlyCSE/InstCombine can't fold it with some
2878  // other PHI. So here we directly look for some PHI in BB's successor with V
2879  // as an incoming operand. If we find one, we use it, else we create a new
2880  // one.
2881  //
2882  // If AlternativeV is not nullptr, we care about both incoming values in PHI.
2883  // PHI must be exactly: phi <ty> [ %BB, %V ], [ %OtherBB, %AlternativeV]
2884  // where OtherBB is the single other predecessor of BB's only successor.
2885  PHINode *PHI = nullptr;
2886  BasicBlock *Succ = BB->getSingleSuccessor();
2887 
2888  for (auto I = Succ->begin(); isa<PHINode>(I); ++I)
2889  if (cast<PHINode>(I)->getIncomingValueForBlock(BB) == V) {
2890  PHI = cast<PHINode>(I);
2891  if (!AlternativeV)
2892  break;
2893 
2894  assert(Succ->hasNPredecessors(2));
2895  auto PredI = pred_begin(Succ);
2896  BasicBlock *OtherPredBB = *PredI == BB ? *++PredI : *PredI;
2897  if (PHI->getIncomingValueForBlock(OtherPredBB) == AlternativeV)
2898  break;
2899  PHI = nullptr;
2900  }
2901  if (PHI)
2902  return PHI;
2903 
2904  // If V is not an instruction defined in BB, just return it.
2905  if (!AlternativeV &&
2906  (!isa<Instruction>(V) || cast<Instruction>(V)->getParent() != BB))
2907  return V;
2908 
2909  PHI = PHINode::Create(V->getType(), 2, "simplifycfg.merge", &Succ->front());
2910  PHI->addIncoming(V, BB);
2911  for (BasicBlock *PredBB : predecessors(Succ))
2912  if (PredBB != BB)
2913  PHI->addIncoming(
2914  AlternativeV ? AlternativeV : UndefValue::get(V->getType()), PredBB);
2915  return PHI;
2916 }
2917 
2919  BasicBlock *QTB, BasicBlock *QFB,
2920  BasicBlock *PostBB, Value *Address,
2921  bool InvertPCond, bool InvertQCond,
2922  const DataLayout &DL) {
2923  auto IsaBitcastOfPointerType = [](const Instruction &I) {
2924  return Operator::getOpcode(&I) == Instruction::BitCast &&
2925  I.getType()->isPointerTy();
2926  };
2927 
2928  // If we're not in aggressive mode, we only optimize if we have some
2929  // confidence that by optimizing we'll allow P and/or Q to be if-converted.
2930  auto IsWorthwhile = [&](BasicBlock *BB) {
2931  if (!BB)
2932  return true;
2933  // Heuristic: if the block can be if-converted/phi-folded and the
2934  // instructions inside are all cheap (arithmetic/GEPs), it's worthwhile to
2935  // thread this store.
2936  unsigned N = 0;
2937  for (auto &I : BB->instructionsWithoutDebug()) {
2938  // Cheap instructions viable for folding.
2939  if (isa<BinaryOperator>(I) || isa<GetElementPtrInst>(I) ||
2940  isa<StoreInst>(I))
2941  ++N;
2942  // Free instructions.
2943  else if (I.isTerminator() || IsaBitcastOfPointerType(I))
2944  continue;
2945  else
2946  return false;
2947  }
2948  // The store we want to merge is counted in N, so add 1 to make sure
2949  // we're counting the instructions that would be left.
2950  return N <= (PHINodeFoldingThreshold + 1);
2951  };
2952 
2954  (!IsWorthwhile(PTB) || !IsWorthwhile(PFB) || !IsWorthwhile(QTB) ||
2955  !IsWorthwhile(QFB)))
2956  return false;
2957 
2958  // For every pointer, there must be exactly two stores, one coming from
2959  // PTB or PFB, and the other from QTB or QFB. We don't support more than one
2960  // store (to any address) in PTB,PFB or QTB,QFB.
2961  // FIXME: We could relax this restriction with a bit more work and performance
2962  // testing.
2963  StoreInst *PStore = findUniqueStoreInBlocks(PTB, PFB);
2964  StoreInst *QStore = findUniqueStoreInBlocks(QTB, QFB);
2965  if (!PStore || !QStore)
2966  return false;
2967 
2968  // Now check the stores are compatible.
2969  if (!QStore->isUnordered() || !PStore->isUnordered())
2970  return false;
2971 
2972  // Check that sinking the store won't cause program behavior changes. Sinking
2973  // the store out of the Q blocks won't change any behavior as we're sinking
2974  // from a block to its unconditional successor. But we're moving a store from
2975  // the P blocks down through the middle block (QBI) and past both QFB and QTB.
2976  // So we need to check that there are no aliasing loads or stores in
2977  // QBI, QTB and QFB. We also need to check there are no conflicting memory
2978  // operations between PStore and the end of its parent block.
2979  //
2980  // The ideal way to do this is to query AliasAnalysis, but we don't
2981  // preserve AA currently so that is dangerous. Be super safe and just
2982  // check there are no other memory operations at all.
2983  for (auto &I : *QFB->getSinglePredecessor())
2984  if (I.mayReadOrWriteMemory())
2985  return false;
2986  for (auto &I : *QFB)
2987  if (&I != QStore && I.mayReadOrWriteMemory())
2988  return false;
2989  if (QTB)
2990  for (auto &I : *QTB)
2991  if (&I != QStore && I.mayReadOrWriteMemory())
2992  return false;
2993  for (auto I = BasicBlock::iterator(PStore), E = PStore->getParent()->end();
2994  I != E; ++I)
2995  if (&*I != PStore && I->mayReadOrWriteMemory())
2996  return false;
2997 
2998  // If PostBB has more than two predecessors, we need to split it so we can
2999  // sink the store.
3000  if (std::next(pred_begin(PostBB), 2) != pred_end(PostBB)) {
3001  // We know that QFB's only successor is PostBB. And QFB has a single
3002  // predecessor. If QTB exists, then its only successor is also PostBB.
3003  // If QTB does not exist, then QFB's only predecessor has a conditional
3004  // branch to QFB and PostBB.
3005  BasicBlock *TruePred = QTB ? QTB : QFB->getSinglePredecessor();
3006  BasicBlock *NewBB = SplitBlockPredecessors(PostBB, { QFB, TruePred},
3007  "condstore.split");
3008  if (!NewBB)
3009  return false;
3010  PostBB = NewBB;
3011  }
3012 
3013  // OK, we're going to sink the stores to PostBB. The store has to be
3014  // conditional though, so first create the predicate.
3015  Value *PCond = cast<BranchInst>(PFB->getSinglePredecessor()->getTerminator())
3016  ->getCondition();
3017  Value *QCond = cast<BranchInst>(QFB->getSinglePredecessor()->getTerminator())
3018  ->getCondition();
3019 
3021  PStore->getParent());
3023  QStore->getParent(), PPHI);
3024 
3025  IRBuilder<> QB(&*PostBB->getFirstInsertionPt());
3026 
3027  Value *PPred = PStore->getParent() == PTB ? PCond : QB.CreateNot(PCond);
3028  Value *QPred = QStore->getParent() == QTB ? QCond : QB.CreateNot(QCond);
3029 
3030  if (InvertPCond)
3031  PPred = QB.CreateNot(PPred);
3032  if (InvertQCond)
3033  QPred = QB.CreateNot(QPred);
3034  Value *CombinedPred = QB.CreateOr(PPred, QPred);
3035 
3036  auto *T =
3037  SplitBlockAndInsertIfThen(CombinedPred, &*QB.GetInsertPoint(), false);
3038  QB.SetInsertPoint(T);
3039  StoreInst *SI = cast<StoreInst>(QB.CreateStore(QPHI, Address));
3040  AAMDNodes AAMD;
3041  PStore->getAAMetadata(AAMD, /*Merge=*/false);
3042  PStore->getAAMetadata(AAMD, /*Merge=*/true);
3043  SI->setAAMetadata(AAMD);
3044  unsigned PAlignment = PStore->getAlignment();
3045  unsigned QAlignment = QStore->getAlignment();
3046  unsigned TypeAlignment =
3048  unsigned MinAlignment;
3049  unsigned MaxAlignment;
3050  std::tie(MinAlignment, MaxAlignment) = std::minmax(PAlignment, QAlignment);
3051  // Choose the minimum alignment. If we could prove both stores execute, we
3052  // could use biggest one. In this case, though, we only know that one of the
3053  // stores executes. And we don't know it's safe to take the alignment from a
3054  // store that doesn't execute.
3055  if (MinAlignment != 0) {
3056  // Choose the minimum of all non-zero alignments.
3057  SI->setAlignment(MinAlignment);
3058  } else if (MaxAlignment != 0) {
3059  // Choose the minimal alignment between the non-zero alignment and the ABI
3060  // default alignment for the type of the stored value.
3061  SI->setAlignment(std::min(MaxAlignment, TypeAlignment));
3062  } else {
3063  // If both alignments are zero, use ABI default alignment for the type of
3064  // the stored value.
3065  SI->setAlignment(TypeAlignment);
3066  }
3067 
3068  QStore->eraseFromParent();
3069  PStore->eraseFromParent();
3070 
3071  return true;
3072 }
3073 
3075  const DataLayout &DL) {
3076  // The intention here is to find diamonds or triangles (see below) where each
3077  // conditional block contains a store to the same address. Both of these
3078  // stores are conditional, so they can't be unconditionally sunk. But it may
3079  // be profitable to speculatively sink the stores into one merged store at the
3080  // end, and predicate the merged store on the union of the two conditions of
3081  // PBI and QBI.
3082  //
3083  // This can reduce the number of stores executed if both of the conditions are
3084  // true, and can allow the blocks to become small enough to be if-converted.
3085  // This optimization will also chain, so that ladders of test-and-set
3086  // sequences can be if-converted away.
3087  //
3088  // We only deal with simple diamonds or triangles:
3089  //
3090  // PBI or PBI or a combination of the two
3091  // / \ | \
3092  // PTB PFB | PFB
3093  // \ / | /
3094  // QBI QBI
3095  // / \ | \
3096  // QTB QFB | QFB
3097  // \ / | /
3098  // PostBB PostBB
3099  //
3100  // We model triangles as a type of diamond with a nullptr "true" block.
3101  // Triangles are canonicalized so that the fallthrough edge is represented by
3102  // a true condition, as in the diagram above.
3103  BasicBlock *PTB = PBI->getSuccessor(0);
3104  BasicBlock *PFB = PBI->getSuccessor(1);
3105  BasicBlock *QTB = QBI->getSuccessor(0);
3106  BasicBlock *QFB = QBI->getSuccessor(1);
3107  BasicBlock *PostBB = QFB->getSingleSuccessor();
3108 
3109  // Make sure we have a good guess for PostBB. If QTB's only successor is
3110  // QFB, then QFB is a better PostBB.
3111  if (QTB->getSingleSuccessor() == QFB)
3112  PostBB = QFB;
3113 
3114  // If we couldn't find a good PostBB, stop.
3115  if (!PostBB)
3116  return false;
3117 
3118  bool InvertPCond = false, InvertQCond = false;
3119  // Canonicalize fallthroughs to the true branches.
3120  if (PFB == QBI->getParent()) {
3121  std::swap(PFB, PTB);
3122  InvertPCond = true;
3123  }
3124  if (QFB == PostBB) {
3125  std::swap(QFB, QTB);
3126  InvertQCond = true;
3127  }
3128 
3129  // From this point on we can assume PTB or QTB may be fallthroughs but PFB
3130  // and QFB may not. Model fallthroughs as a nullptr block.
3131  if (PTB == QBI->getParent())
3132  PTB = nullptr;
3133  if (QTB == PostBB)
3134  QTB = nullptr;
3135 
3136  // Legality bailouts. We must have at least the non-fallthrough blocks and
3137  // the post-dominating block, and the non-fallthroughs must only have one
3138  // predecessor.
3139  auto HasOnePredAndOneSucc = [](BasicBlock *BB, BasicBlock *P, BasicBlock *S) {
3140  return BB->getSinglePredecessor() == P && BB->getSingleSuccessor() == S;
3141  };
3142  if (!HasOnePredAndOneSucc(PFB, PBI->getParent(), QBI->getParent()) ||
3143  !HasOnePredAndOneSucc(QFB, QBI->getParent(), PostBB))
3144  return false;
3145  if ((PTB && !HasOnePredAndOneSucc(PTB, PBI->getParent(), QBI->getParent())) ||
3146  (QTB && !HasOnePredAndOneSucc(QTB, QBI->getParent(), PostBB)))
3147  return false;
3148  if (!QBI->getParent()->hasNUses(2))
3149  return false;
3150 
3151  // OK, this is a sequence of two diamonds or triangles.
3152  // Check if there are stores in PTB or PFB that are repeated in QTB or QFB.
3153  SmallPtrSet<Value *, 4> PStoreAddresses, QStoreAddresses;
3154  for (auto *BB : {PTB, PFB}) {
3155  if (!BB)
3156  continue;
3157  for (auto &I : *BB)
3158  if (StoreInst *SI = dyn_cast<StoreInst>(&I))
3159  PStoreAddresses.insert(SI->getPointerOperand());
3160  }
3161  for (auto *BB : {QTB, QFB}) {
3162  if (!BB)
3163  continue;
3164  for (auto &I : *BB)
3165  if (StoreInst *SI = dyn_cast<StoreInst>(&I))
3166  QStoreAddresses.insert(SI->getPointerOperand());
3167  }
3168 
3169  set_intersect(PStoreAddresses, QStoreAddresses);
3170  // set_intersect mutates PStoreAddresses in place. Rename it here to make it
3171  // clear what it contains.
3172  auto &CommonAddresses = PStoreAddresses;
3173 
3174  bool Changed = false;
3175  for (auto *Address : CommonAddresses)
3176  Changed |= mergeConditionalStoreToAddress(
3177  PTB, PFB, QTB, QFB, PostBB, Address, InvertPCond, InvertQCond, DL);
3178  return Changed;
3179 }
3180 
3181 /// If we have a conditional branch as a predecessor of another block,
3182 /// this function tries to simplify it. We know
3183 /// that PBI and BI are both conditional branches, and BI is in one of the
3184 /// successor blocks of PBI - PBI branches to BI.
3186  const DataLayout &DL) {
3187  assert(PBI->isConditional() && BI->isConditional());
3188  BasicBlock *BB = BI->getParent();
3189 
3190  // If this block ends with a branch instruction, and if there is a
3191  // predecessor that ends on a branch of the same condition, make
3192  // this conditional branch redundant.
3193  if (PBI->getCondition() == BI->getCondition() &&
3194  PBI->getSuccessor(0) != PBI->getSuccessor(1)) {
3195  // Okay, the outcome of this conditional branch is statically
3196  // knowable. If this block had a single pred, handle specially.
3197  if (BB->getSinglePredecessor()) {
3198  // Turn this into a branch on constant.
3199  bool CondIsTrue = PBI->getSuccessor(0) == BB;
3200  BI->setCondition(
3201  ConstantInt::get(Type::getInt1Ty(BB->getContext()), CondIsTrue));
3202  return true; // Nuke the branch on constant.
3203  }
3204 
3205  // Otherwise, if there are multiple predecessors, insert a PHI that merges
3206  // in the constant and simplify the block result. Subsequent passes of
3207  // simplifycfg will thread the block.
3209  pred_iterator PB = pred_begin(BB), PE = pred_end(BB);
3210  PHINode *NewPN = PHINode::Create(
3211  Type::getInt1Ty(BB->getContext()), std::distance(PB, PE),
3212  BI->getCondition()->getName() + ".pr", &BB->front());
3213  // Okay, we're going to insert the PHI node. Since PBI is not the only
3214  // predecessor, compute the PHI'd conditional value for all of the preds.
3215  // Any predecessor where the condition is not computable we keep symbolic.
3216  for (pred_iterator PI = PB; PI != PE; ++PI) {
3217  BasicBlock *P = *PI;
3218  if ((PBI = dyn_cast<BranchInst>(P->getTerminator())) && PBI != BI &&
3219  PBI->isConditional() && PBI->getCondition() == BI->getCondition() &&
3220  PBI->getSuccessor(0) != PBI->getSuccessor(1)) {
3221  bool CondIsTrue = PBI->getSuccessor(0) == BB;
3222  NewPN->addIncoming(
3223  ConstantInt::get(Type::getInt1Ty(BB->getContext()), CondIsTrue),
3224  P);
3225  } else {
3226  NewPN->addIncoming(BI->getCondition(), P);
3227  }
3228  }
3229 
3230  BI->setCondition(NewPN);
3231  return true;
3232  }
3233  }
3234 
3235  if (auto *CE = dyn_cast<ConstantExpr>(BI->getCondition()))
3236  if (CE->canTrap())
3237  return false;
3238 
3239  // If both branches are conditional and both contain stores to the same
3240  // address, remove the stores from the conditionals and create a conditional
3241  // merged store at the end.
3242  if (MergeCondStores && mergeConditionalStores(PBI, BI, DL))
3243  return true;
3244 
3245  // If this is a conditional branch in an empty block, and if any
3246  // predecessors are a conditional branch to one of our destinations,
3247  // fold the conditions into logical ops and one cond br.
3248 
3249  // Ignore dbg intrinsics.
3250  if (&*BB->instructionsWithoutDebug().begin() != BI)
3251  return false;
3252 
3253  int PBIOp, BIOp;
3254  if (PBI->getSuccessor(0) == BI->getSuccessor(0)) {
3255  PBIOp = 0;
3256  BIOp = 0;
3257  } else if (PBI->getSuccessor(0) == BI->getSuccessor(1)) {
3258  PBIOp = 0;
3259  BIOp = 1;
3260  } else if (PBI->getSuccessor(1) == BI->getSuccessor(0)) {
3261  PBIOp = 1;
3262  BIOp = 0;
3263  } else if (PBI->getSuccessor(1) == BI->getSuccessor(1)) {
3264  PBIOp = 1;
3265  BIOp = 1;
3266  } else {
3267  return false;
3268  }
3269 
3270  // Check to make sure that the other destination of this branch
3271  // isn't BB itself. If so, this is an infinite loop that will
3272  // keep getting unwound.
3273  if (PBI->getSuccessor(PBIOp) == BB)
3274  return false;
3275 
3276  // Do not perform this transformation if it would require
3277  // insertion of a large number of select instructions. For targets
3278  // without predication/cmovs, this is a big pessimization.
3279 
3280  // Also do not perform this transformation if any phi node in the common
3281  // destination block can trap when reached by BB or PBB (PR17073). In that
3282  // case, it would be unsafe to hoist the operation into a select instruction.
3283 
3284  BasicBlock *CommonDest = PBI->getSuccessor(PBIOp);
3285  unsigned NumPhis = 0;
3286  for (BasicBlock::iterator II = CommonDest->begin(); isa<PHINode>(II);
3287  ++II, ++NumPhis) {
3288  if (NumPhis > 2) // Disable this xform.
3289  return false;
3290 
3291  PHINode *PN = cast<PHINode>(II);
3292  Value *BIV = PN->getIncomingValueForBlock(BB);
3293  if (ConstantExpr *CE = dyn_cast<ConstantExpr>(BIV))
3294  if (CE->canTrap())
3295  return false;
3296 
3297  unsigned PBBIdx = PN->getBasicBlockIndex(PBI->getParent());
3298  Value *PBIV = PN->getIncomingValue(PBBIdx);
3299  if (ConstantExpr *CE = dyn_cast<ConstantExpr>(PBIV))
3300  if (CE->canTrap())
3301  return false;
3302  }
3303 
3304  // Finally, if everything is ok, fold the branches to logical ops.
3305  BasicBlock *OtherDest = BI->getSuccessor(BIOp ^ 1);
3306 
3307  LLVM_DEBUG(dbgs() << "FOLDING BRs:" << *PBI->getParent()
3308  << "AND: " << *BI->getParent());
3309 
3310  // If OtherDest *is* BB, then BB is a basic block with a single conditional
3311  // branch in it, where one edge (OtherDest) goes back to itself but the other
3312  // exits. We don't *know* that the program avoids the infinite loop
3313  // (even though that seems likely). If we do this xform naively, we'll end up
3314  // recursively unpeeling the loop. Since we know that (after the xform is
3315  // done) that the block *is* infinite if reached, we just make it an obviously
3316  // infinite loop with no cond branch.
3317  if (OtherDest == BB) {
3318  // Insert it at the end of the function, because it's either code,
3319  // or it won't matter if it's hot. :)
3320  BasicBlock *InfLoopBlock =
3321  BasicBlock::Create(BB->getContext(), "infloop", BB->getParent());
3322  BranchInst::Create(InfLoopBlock, InfLoopBlock);
3323  OtherDest = InfLoopBlock;
3324  }
3325 
3326  LLVM_DEBUG(dbgs() << *PBI->getParent()->getParent());
3327 
3328  // BI may have other predecessors. Because of this, we leave
3329  // it alone, but modify PBI.
3330 
3331  // Make sure we get to CommonDest on True&True directions.
3332  Value *PBICond = PBI->getCondition();
3333  IRBuilder<NoFolder> Builder(PBI);
3334  if (PBIOp)
3335  PBICond = Builder.CreateNot(PBICond, PBICond->getName() + ".not");
3336 
3337  Value *BICond = BI->getCondition();
3338  if (BIOp)
3339  BICond = Builder.CreateNot(BICond, BICond->getName() + ".not");
3340 
3341  // Merge the conditions.
3342  Value *Cond = Builder.CreateOr(PBICond, BICond, "brmerge");
3343 
3344  // Modify PBI to branch on the new condition to the new dests.
3345  PBI->setCondition(Cond);
3346  PBI->setSuccessor(0, CommonDest);
3347  PBI->setSuccessor(1, OtherDest);
3348 
3349  // Update branch weight for PBI.
3350  uint64_t PredTrueWeight, PredFalseWeight, SuccTrueWeight, SuccFalseWeight;
3351  uint64_t PredCommon, PredOther, SuccCommon, SuccOther;
3352  bool HasWeights =
3353  extractPredSuccWeights(PBI, BI, PredTrueWeight, PredFalseWeight,
3354  SuccTrueWeight, SuccFalseWeight);
3355  if (HasWeights) {
3356  PredCommon = PBIOp ? PredFalseWeight : PredTrueWeight;
3357  PredOther = PBIOp ? PredTrueWeight : PredFalseWeight;
3358  SuccCommon = BIOp ? SuccFalseWeight : SuccTrueWeight;
3359  SuccOther = BIOp ? SuccTrueWeight : SuccFalseWeight;
3360  // The weight to CommonDest should be PredCommon * SuccTotal +
3361  // PredOther * SuccCommon.
3362  // The weight to OtherDest should be PredOther * SuccOther.
3363  uint64_t NewWeights[2] = {PredCommon * (SuccCommon + SuccOther) +
3364  PredOther * SuccCommon,
3365  PredOther * SuccOther};
3366  // Halve the weights if any of them cannot fit in an uint32_t
3367  FitWeights(NewWeights);
3368 
3369  setBranchWeights(PBI, NewWeights[0], NewWeights[1]);
3370  }
3371 
3372  // OtherDest may have phi nodes. If so, add an entry from PBI's
3373  // block that are identical to the entries for BI's block.
3374  AddPredecessorToBlock(OtherDest, PBI->getParent(), BB);
3375 
3376  // We know that the CommonDest already had an edge from PBI to
3377  // it. If it has PHIs though, the PHIs may have different
3378  // entries for BB and PBI's BB. If so, insert a select to make
3379  // them agree.
3380  for (PHINode &PN : CommonDest->phis()) {
3381  Value *BIV = PN.getIncomingValueForBlock(BB);
3382  unsigned PBBIdx = PN.getBasicBlockIndex(PBI->getParent());
3383  Value *PBIV = PN.getIncomingValue(PBBIdx);
3384  if (BIV != PBIV) {
3385  // Insert a select in PBI to pick the right value.
3386  SelectInst *NV = cast<SelectInst>(
3387  Builder.CreateSelect(PBICond, PBIV, BIV, PBIV->getName() + ".mux"));
3388  PN.setIncomingValue(PBBIdx, NV);
3389  // Although the select has the same condition as PBI, the original branch
3390  // weights for PBI do not apply to the new select because the select's
3391  // 'logical' edges are incoming edges of the phi that is eliminated, not
3392  // the outgoing edges of PBI.
3393  if (HasWeights) {
3394  uint64_t PredCommon = PBIOp ? PredFalseWeight : PredTrueWeight;
3395  uint64_t PredOther = PBIOp ? PredTrueWeight : PredFalseWeight;
3396  uint64_t SuccCommon = BIOp ? SuccFalseWeight : SuccTrueWeight;
3397  uint64_t SuccOther = BIOp ? SuccTrueWeight : SuccFalseWeight;
3398  // The weight to PredCommonDest should be PredCommon * SuccTotal.
3399  // The weight to PredOtherDest should be PredOther * SuccCommon.
3400  uint64_t NewWeights[2] = {PredCommon * (SuccCommon + SuccOther),
3401  PredOther * SuccCommon};
3402 
3403  FitWeights(NewWeights);
3404 
3405  setBranchWeights(NV, NewWeights[0], NewWeights[1]);
3406  }
3407  }
3408  }
3409 
3410  LLVM_DEBUG(dbgs() << "INTO: " << *PBI->getParent());
3411  LLVM_DEBUG(dbgs() << *PBI->getParent()->getParent());
3412 
3413  // This basic block is probably dead. We know it has at least
3414  // one fewer predecessor.
3415  return true;
3416 }
3417 
3418 // Simplifies a terminator by replacing it with a branch to TrueBB if Cond is
3419 // true or to FalseBB if Cond is false.
3420 // Takes care of updating the successors and removing the old terminator.
3421 // Also makes sure not to introduce new successors by assuming that edges to
3422 // non-successor TrueBBs and FalseBBs aren't reachable.
3423 static bool SimplifyTerminatorOnSelect(Instruction *OldTerm, Value *Cond,
3424  BasicBlock *TrueBB, BasicBlock *FalseBB,
3425  uint32_t TrueWeight,
3426  uint32_t FalseWeight) {
3427  // Remove any superfluous successor edges from the CFG.
3428  // First, figure out which successors to preserve.
3429  // If TrueBB and FalseBB are equal, only try to preserve one copy of that
3430  // successor.
3431  BasicBlock *KeepEdge1 = TrueBB;
3432  BasicBlock *KeepEdge2 = TrueBB != FalseBB ? FalseBB : nullptr;
3433 
3434  // Then remove the rest.
3435  for (BasicBlock *Succ : successors(OldTerm)) {
3436  // Make sure only to keep exactly one copy of each edge.
3437  if (Succ == KeepEdge1)
3438  KeepEdge1 = nullptr;
3439  else if (Succ == KeepEdge2)
3440  KeepEdge2 = nullptr;
3441  else
3442  Succ->removePredecessor(OldTerm->getParent(),
3443  /*DontDeleteUselessPHIs=*/true);
3444  }
3445 
3446  IRBuilder<> Builder(OldTerm);
3447  Builder.SetCurrentDebugLocation(OldTerm->getDebugLoc());
3448 
3449  // Insert an appropriate new terminator.
3450  if (!KeepEdge1 && !KeepEdge2) {
3451  if (TrueBB == FalseBB)
3452  // We were only looking for one successor, and it was present.
3453  // Create an unconditional branch to it.
3454  Builder.CreateBr(TrueBB);
3455  else {
3456  // We found both of the successors we were looking for.
3457  // Create a conditional branch sharing the condition of the select.
3458  BranchInst *NewBI = Builder.CreateCondBr(Cond, TrueBB, FalseBB);
3459  if (TrueWeight != FalseWeight)
3460  setBranchWeights(NewBI, TrueWeight, FalseWeight);
3461  }
3462  } else if (KeepEdge1 && (KeepEdge2 || TrueBB == FalseBB)) {
3463  // Neither of the selected blocks were successors, so this
3464  // terminator must be unreachable.
3465  new UnreachableInst(OldTerm->getContext(), OldTerm);
3466  } else {
3467  // One of the selected values was a successor, but the other wasn't.
3468  // Insert an unconditional branch to the one that was found;
3469  // the edge to the one that wasn't must be unreachable.
3470  if (!KeepEdge1)
3471  // Only TrueBB was found.
3472  Builder.CreateBr(TrueBB);
3473  else
3474  // Only FalseBB was found.
3475  Builder.CreateBr(FalseBB);
3476  }
3477 
3478  EraseTerminatorAndDCECond(OldTerm);
3479  return true;
3480 }
3481 
3482 // Replaces
3483 // (switch (select cond, X, Y)) on constant X, Y
3484 // with a branch - conditional if X and Y lead to distinct BBs,
3485 // unconditional otherwise.
3487  // Check for constant integer values in the select.
3488  ConstantInt *TrueVal = dyn_cast<ConstantInt>(Select->getTrueValue());
3489  ConstantInt *FalseVal = dyn_cast<ConstantInt>(Select->getFalseValue());
3490  if (!TrueVal || !FalseVal)
3491  return false;
3492 
3493  // Find the relevant condition and destinations.
3494  Value *Condition = Select->getCondition();
3495  BasicBlock *TrueBB = SI->findCaseValue(TrueVal)->getCaseSuccessor();
3496  BasicBlock *FalseBB = SI->findCaseValue(FalseVal)->getCaseSuccessor();
3497 
3498  // Get weight for TrueBB and FalseBB.
3499  uint32_t TrueWeight = 0, FalseWeight = 0;
3500  SmallVector<uint64_t, 8> Weights;
3501  bool HasWeights = HasBranchWeights(SI);
3502  if (HasWeights) {
3503  GetBranchWeights(SI, Weights);
3504  if (Weights.size() == 1 + SI->getNumCases()) {
3505  TrueWeight =
3506  (uint32_t)Weights[SI->findCaseValue(TrueVal)->getSuccessorIndex()];
3507  FalseWeight =
3508  (uint32_t)Weights[SI->findCaseValue(FalseVal)->getSuccessorIndex()];
3509  }
3510  }
3511 
3512  // Perform the actual simplification.
3513  return SimplifyTerminatorOnSelect(SI, Condition, TrueBB, FalseBB, TrueWeight,
3514  FalseWeight);
3515 }
3516 
3517 // Replaces
3518 // (indirectbr (select cond, blockaddress(@fn, BlockA),
3519 // blockaddress(@fn, BlockB)))
3520 // with
3521 // (br cond, BlockA, BlockB).
3523  // Check that both operands of the select are block addresses.
3526  if (!TBA || !FBA)
3527  return false;
3528 
3529  // Extract the actual blocks.
3530  BasicBlock *TrueBB = TBA->getBasicBlock();
3531  BasicBlock *FalseBB = FBA->getBasicBlock();
3532 
3533  // Perform the actual simplification.
3534  return SimplifyTerminatorOnSelect(IBI, SI->getCondition(), TrueBB, FalseBB, 0,
3535  0);
3536 }
3537 
3538 /// This is called when we find an icmp instruction
3539 /// (a seteq/setne with a constant) as the only instruction in a
3540 /// block that ends with an uncond branch. We are looking for a very specific
3541 /// pattern that occurs when "A == 1 || A == 2 || A == 3" gets simplified. In
3542 /// this case, we merge the first two "or's of icmp" into a switch, but then the
3543 /// default value goes to an uncond block with a seteq in it, we get something
3544 /// like:
3545 ///
3546 /// switch i8 %A, label %DEFAULT [ i8 1, label %end i8 2, label %end ]
3547 /// DEFAULT:
3548 /// %tmp = icmp eq i8 %A, 92
3549 /// br label %end
3550 /// end:
3551 /// ... = phi i1 [ true, %entry ], [ %tmp, %DEFAULT ], [ true, %entry ]
3552 ///
3553 /// We prefer to split the edge to 'end' so that there is a true/false entry to
3554 /// the PHI, merging the third icmp into the switch.
3555 bool SimplifyCFGOpt::tryToSimplifyUncondBranchWithICmpInIt(
3556  ICmpInst *ICI, IRBuilder<> &Builder) {
3557  BasicBlock *BB = ICI->getParent();
3558 
3559  // If the block has any PHIs in it or the icmp has multiple uses, it is too
3560  // complex.
3561  if (isa<PHINode>(BB->begin()) || !ICI->hasOneUse())
3562  return false;
3563 
3564  Value *V = ICI->getOperand(0);
3565  ConstantInt *Cst = cast<ConstantInt>(ICI->getOperand(1));
3566 
3567  // The pattern we're looking for is where our only predecessor is a switch on
3568  // 'V' and this block is the default case for the switch. In this case we can
3569  // fold the compared value into the switch to simplify things.
3570  BasicBlock *Pred = BB->getSinglePredecessor();
3571  if (!Pred || !isa<SwitchInst>(Pred->getTerminator()))
3572  return false;
3573 
3574  SwitchInst *SI = cast<SwitchInst>(Pred->getTerminator());
3575  if (SI->getCondition() != V)
3576  return false;
3577 
3578  // If BB is reachable on a non-default case, then we simply know the value of
3579  // V in this block. Substitute it and constant fold the icmp instruction
3580  // away.
3581  if (SI->getDefaultDest() != BB) {
3582  ConstantInt *VVal = SI->findCaseDest(BB);
3583  assert(VVal && "Should have a unique destination value");
3584  ICI->setOperand(0, VVal);
3585 
3586  if (Value *V = SimplifyInstruction(ICI, {DL, ICI})) {
3587  ICI->replaceAllUsesWith(V);
3588  ICI->eraseFromParent();
3589  }
3590  // BB is now empty, so it is likely to simplify away.
3591  return requestResimplify();
3592  }
3593 
3594  // Ok, the block is reachable from the default dest. If the constant we're
3595  // comparing exists in one of the other edges, then we can constant fold ICI
3596  // and zap it.
3597  if (SI->findCaseValue(Cst) != SI->case_default()) {
3598  Value *V;
3599  if (ICI->getPredicate() == ICmpInst::ICMP_EQ)
3600  V = ConstantInt::getFalse(BB->getContext());
3601  else
3602  V = ConstantInt::getTrue(BB->getContext());
3603 
3604  ICI->replaceAllUsesWith(V);
3605  ICI->eraseFromParent();
3606  // BB is now empty, so it is likely to simplify away.
3607  return requestResimplify();
3608  }
3609 
3610  // The use of the icmp has to be in the 'end' block, by the only PHI node in
3611  // the block.
3612  BasicBlock *SuccBlock = BB->getTerminator()->getSuccessor(0);
3613  PHINode *PHIUse = dyn_cast<PHINode>(ICI->user_back());
3614  if (PHIUse == nullptr || PHIUse != &SuccBlock->front() ||
3615  isa<PHINode>(++BasicBlock::iterator(PHIUse)))
3616  return false;
3617 
3618  // If the icmp is a SETEQ, then the default dest gets false, the new edge gets
3619  // true in the PHI.
3620  Constant *DefaultCst = ConstantInt::getTrue(BB->getContext());
3621  Constant *NewCst = ConstantInt::getFalse(BB->getContext());
3622 
3623  if (ICI->getPredicate() == ICmpInst::ICMP_EQ)
3624  std::swap(DefaultCst, NewCst);
3625 
3626  // Replace ICI (which is used by the PHI for the default value) with true or
3627  // false depending on if it is EQ or NE.
3628  ICI->replaceAllUsesWith(DefaultCst);
3629  ICI->eraseFromParent();
3630 
3631  // Okay, the switch goes to this block on a default value. Add an edge from
3632  // the switch to the merge point on the compared value.
3633  BasicBlock *NewBB =
3634  BasicBlock::Create(BB->getContext(), "switch.edge", BB->getParent(), BB);
3635  SmallVector<uint64_t, 8> Weights;
3636  bool HasWeights = HasBranchWeights(SI);
3637  if (HasWeights) {
3638  GetBranchWeights(SI, Weights);
3639  if (Weights.size() == 1 + SI->getNumCases()) {
3640  // Split weight for default case to case for "Cst".
3641  Weights[0] = (Weights[0] + 1) >> 1;
3642  Weights.push_back(Weights[0]);
3643 
3644  SmallVector<uint32_t, 8> MDWeights(Weights.begin(), Weights.end());
3645  setBranchWeights(SI, MDWeights);
3646  }
3647  }
3648  SI->addCase(Cst, NewBB);
3649 
3650  // NewBB branches to the phi block, add the uncond branch and the phi entry.
3651  Builder.SetInsertPoint(NewBB);
3652  Builder.SetCurrentDebugLocation(SI->getDebugLoc());
3653  Builder.CreateBr(SuccBlock);
3654  PHIUse->addIncoming(NewCst, NewBB);
3655  return true;
3656 }
3657 
3658 /// The specified branch is a conditional branch.
3659 /// Check to see if it is branching on an or/and chain of icmp instructions, and
3660 /// fold it into a switch instruction if so.
3662  const DataLayout &DL) {
3663  Instruction *Cond = dyn_cast<Instruction>(BI->getCondition());
3664  if (!Cond)
3665  return false;
3666 
3667  // Change br (X == 0 | X == 1), T, F into a switch instruction.
3668  // If this is a bunch of seteq's or'd together, or if it's a bunch of
3669  // 'setne's and'ed together, collect them.
3670 
3671  // Try to gather values from a chain of and/or to be turned into a switch
3672  ConstantComparesGatherer ConstantCompare(Cond, DL);
3673  // Unpack the result
3674  SmallVectorImpl<ConstantInt *> &Values = ConstantCompare.Vals;
3675  Value *CompVal = ConstantCompare.CompValue;
3676  unsigned UsedICmps = ConstantCompare.UsedICmps;
3677  Value *ExtraCase = ConstantCompare.Extra;
3678 
3679  // If we didn't have a multiply compared value, fail.
3680  if (!CompVal)
3681  return false;
3682 
3683  // Avoid turning single icmps into a switch.
3684  if (UsedICmps <= 1)
3685  return false;
3686 
3687  bool TrueWhenEqual = (Cond->getOpcode() == Instruction::Or);
3688 
3689  // There might be duplicate constants in the list, which the switch
3690  // instruction can't handle, remove them now.
3691  array_pod_sort(Values.begin(), Values.end(), ConstantIntSortPredicate);
3692  Values.erase(std::unique(Values.begin(), Values.end()), Values.end());
3693 
3694  // If Extra was used, we require at least two switch values to do the
3695  // transformation. A switch with one value is just a conditional branch.
3696  if (ExtraCase && Values.size() < 2)
3697  return false;
3698 
3699  // TODO: Preserve branch weight metadata, similarly to how
3700  // FoldValueComparisonIntoPredecessors preserves it.
3701 
3702  // Figure out which block is which destination.
3703  BasicBlock *DefaultBB = BI->getSuccessor(1);
3704  BasicBlock *EdgeBB = BI->getSuccessor(0);
3705  if (!TrueWhenEqual)
3706  std::swap(DefaultBB, EdgeBB);
3707 
3708  BasicBlock *BB = BI->getParent();
3709 
3710  LLVM_DEBUG(dbgs() << "Converting 'icmp' chain with " << Values.size()
3711  << " cases into SWITCH. BB is:\n"
3712  << *BB);
3713 
3714  // If there are any extra values that couldn't be folded into the switch
3715  // then we evaluate them with an explicit branch first. Split the block
3716  // right before the condbr to handle it.
3717  if (ExtraCase) {
3718  BasicBlock *NewBB =
3719  BB->splitBasicBlock(BI->getIterator(), "switch.early.test");
3720  // Remove the uncond branch added to the old block.
3721  Instruction *OldTI = BB->getTerminator();
3722  Builder.SetInsertPoint(OldTI);
3723 
3724  if (TrueWhenEqual)
3725  Builder.CreateCondBr(ExtraCase, EdgeBB, NewBB);
3726  else
3727  Builder.CreateCondBr(ExtraCase, NewBB, EdgeBB);
3728 
3729  OldTI->eraseFromParent();
3730 
3731  // If there are PHI nodes in EdgeBB, then we need to add a new entry to them
3732  // for the edge we just added.
3733  AddPredecessorToBlock(EdgeBB, BB, NewBB);
3734 
3735  LLVM_DEBUG(dbgs() << " ** 'icmp' chain unhandled condition: " << *ExtraCase
3736  << "\nEXTRABB = " << *BB);
3737  BB = NewBB;
3738  }
3739 
3740  Builder.SetInsertPoint(BI);
3741  // Convert pointer to int before we switch.
3742  if (CompVal->getType()->isPointerTy()) {
3743  CompVal = Builder.CreatePtrToInt(
3744  CompVal, DL.getIntPtrType(CompVal->getType()), "magicptr");
3745  }
3746 
3747  // Create the new switch instruction now.
3748  SwitchInst *New = Builder.CreateSwitch(CompVal, DefaultBB, Values.size());
3749 
3750  // Add all of the 'cases' to the switch instruction.
3751  for (unsigned i = 0, e = Values.size(); i != e; ++i)
3752  New->addCase(Values[i], EdgeBB);
3753 
3754  // We added edges from PI to the EdgeBB. As such, if there were any
3755  // PHI nodes in EdgeBB, they need entries to be added corresponding to
3756  // the number of edges added.
3757  for (BasicBlock::iterator BBI = EdgeBB->begin(); isa<PHINode>(BBI); ++BBI) {
3758  PHINode *PN = cast<PHINode>(BBI);
3759  Value *InVal = PN->getIncomingValueForBlock(BB);
3760  for (unsigned i = 0, e = Values.size() - 1; i != e; ++i)
3761  PN->addIncoming(InVal, BB);
3762  }
3763 
3764  // Erase the old branch instruction.
3766 
3767  LLVM_DEBUG(dbgs() << " ** 'icmp' chain result is:\n" << *BB << '\n');
3768  return true;
3769 }
3770 
3771 bool SimplifyCFGOpt::SimplifyResume(ResumeInst *RI, IRBuilder<> &Builder) {
3772  if (isa<PHINode>(RI->getValue()))
3773  return SimplifyCommonResume(RI);
3774  else if (isa<LandingPadInst>(RI->getParent()->getFirstNonPHI()) &&
3775  RI->getValue() == RI->getParent()->getFirstNonPHI())
3776  // The resume must unwind the exception that caused control to branch here.
3777  return SimplifySingleResume(RI);
3778 
3779  return false;
3780 }
3781 
3782 // Simplify resume that is shared by several landing pads (phi of landing pad).
3783 bool SimplifyCFGOpt::SimplifyCommonResume(ResumeInst *RI) {
3784  BasicBlock *BB = RI->getParent();
3785 
3786  // Check that there are no other instructions except for debug intrinsics
3787  // between the phi of landing pads (RI->getValue()) and resume instruction.
3788  BasicBlock::iterator I = cast<Instruction>(RI->getValue())->getIterator(),
3789  E = RI->getIterator();
3790  while (++I != E)
3791  if (!isa<DbgInfoIntrinsic>(I))
3792  return false;
3793 
3794  SmallSetVector<BasicBlock *, 4> TrivialUnwindBlocks;
3795  auto *PhiLPInst = cast<PHINode>(RI->getValue());
3796 
3797  // Check incoming blocks to see if any of them are trivial.
3798  for (unsigned Idx = 0, End = PhiLPInst->getNumIncomingValues(); Idx != End;
3799  Idx++) {
3800  auto *IncomingBB = PhiLPInst->getIncomingBlock(Idx);
3801  auto *IncomingValue = PhiLPInst->getIncomingValue(Idx);
3802 
3803  // If the block has other successors, we can not delete it because
3804  // it has other dependents.
3805  if (IncomingBB->getUniqueSuccessor() != BB)
3806  continue;
3807 
3808  auto *LandingPad = dyn_cast<LandingPadInst>(IncomingBB->getFirstNonPHI());
3809  // Not the landing pad that caused the control to branch here.
3810  if (IncomingValue != LandingPad)
3811  continue;
3812 
3813  bool isTrivial = true;
3814 
3815  I = IncomingBB->getFirstNonPHI()->getIterator();
3816  E = IncomingBB->getTerminator()->getIterator();
3817  while (++I != E)
3818  if (!isa<DbgInfoIntrinsic>(I)) {
3819  isTrivial = false;
3820  break;
3821  }
3822 
3823  if (isTrivial)
3824  TrivialUnwindBlocks.insert(IncomingBB);
3825  }
3826 
3827  // If no trivial unwind blocks, don't do any simplifications.
3828  if (TrivialUnwindBlocks.empty())
3829  return false;
3830 
3831  // Turn all invokes that unwind here into calls.
3832  for (auto *TrivialBB : TrivialUnwindBlocks) {
3833  // Blocks that will be simplified should be removed from the phi node.
3834  // Note there could be multiple edges to the resume block, and we need
3835  // to remove them all.
3836  while (PhiLPInst->getBasicBlockIndex(TrivialBB) != -1)
3837  BB->removePredecessor(TrivialBB, true);
3838 
3839  for (pred_iterator PI = pred_begin(TrivialBB), PE = pred_end(TrivialBB);
3840  PI != PE;) {
3841  BasicBlock *Pred = *PI++;
3842  removeUnwindEdge(Pred);
3843  }
3844 
3845  // In each SimplifyCFG run, only the current processed block can be erased.
3846  // Otherwise, it will break the iteration of SimplifyCFG pass. So instead
3847  // of erasing TrivialBB, we only remove the branch to the common resume
3848  // block so that we can later erase the resume block since it has no
3849  // predecessors.
3850  TrivialBB->getTerminator()->eraseFromParent();
3851  new UnreachableInst(RI->getContext(), TrivialBB);
3852  }
3853 
3854  // Delete the resume block if all its predecessors have been removed.
3855  if (pred_empty(BB))
3856  BB->eraseFromParent();
3857 
3858  return !TrivialUnwindBlocks.empty();
3859 }
3860 
3861 // Simplify resume that is only used by a single (non-phi) landing pad.
3862 bool SimplifyCFGOpt::SimplifySingleResume(ResumeInst *RI) {
3863  BasicBlock *BB = RI->getParent();
3865  assert(RI->getValue() == LPInst &&
3866  "Resume must unwind the exception that caused control to here");
3867 
3868  // Check that there are no other instructions except for debug intrinsics.
3869  BasicBlock::iterator I = LPInst->getIterator(), E = RI->getIterator();
3870  while (++I != E)
3871  if (!isa<DbgInfoIntrinsic>(I))
3872  return false;
3873 
3874  // Turn all invokes that unwind here into calls and delete the basic block.
3875  for (pred_iterator PI = pred_begin(BB), PE = pred_end(BB); PI != PE;) {
3876  BasicBlock *Pred = *PI++;
3877  removeUnwindEdge(Pred);
3878  }
3879 
3880  // The landingpad is now unreachable. Zap it.
3881  if (LoopHeaders)
3882  LoopHeaders->erase(BB);
3883  BB->eraseFromParent();
3884  return true;
3885 }
3886 
3888  // If this is a trivial cleanup pad that executes no instructions, it can be
3889  // eliminated. If the cleanup pad continues to the caller, any predecessor
3890  // that is an EH pad will be updated to continue to the caller and any
3891  // predecessor that terminates with an invoke instruction will have its invoke
3892  // instruction converted to a call instruction. If the cleanup pad being
3893  // simplified does not continue to the caller, each predecessor will be
3894  // updated to continue to the unwind destination of the cleanup pad being
3895  // simplified.
3896  BasicBlock *BB = RI->getParent();
3897  CleanupPadInst *CPInst = RI->getCleanupPad();
3898  if (CPInst->getParent() != BB)
3899  // This isn't an empty cleanup.
3900  return false;
3901 
3902  // We cannot kill the pad if it has multiple uses. This typically arises
3903  // from unreachable basic blocks.
3904  if (!CPInst->hasOneUse())
3905  return false;
3906 
3907  // Check that there are no other instructions except for benign intrinsics.
3908  BasicBlock::iterator I = CPInst->getIterator(), E = RI->getIterator();
3909  while (++I != E) {
3910  auto *II = dyn_cast<IntrinsicInst>(I);
3911  if (!II)
3912  return false;
3913 
3914  Intrinsic::ID IntrinsicID = II->getIntrinsicID();
3915  switch (IntrinsicID) {
3916  case Intrinsic::dbg_declare:
3917  case Intrinsic::dbg_value:
3918  case Intrinsic::dbg_label:
3919  case Intrinsic::lifetime_end:
3920  break;
3921  default:
3922  return false;
3923  }
3924  }
3925 
3926  // If the cleanup return we are simplifying unwinds to the caller, this will
3927  // set UnwindDest to nullptr.
3928  BasicBlock *UnwindDest = RI->getUnwindDest();
3929  Instruction *DestEHPad = UnwindDest ? UnwindDest->getFirstNonPHI() : nullptr;
3930 
3931  // We're about to remove BB from the control flow. Before we do, sink any
3932  // PHINodes into the unwind destination. Doing this before changing the
3933  // control flow avoids some potentially slow checks, since we can currently
3934  // be certain that UnwindDest and BB have no common predecessors (since they
3935  // are both EH pads).
3936  if (UnwindDest) {
3937  // First, go through the PHI nodes in UnwindDest and update any nodes that
3938  // reference the block we are removing
3939  for (BasicBlock::iterator I = UnwindDest->begin(),
3940  IE = DestEHPad->getIterator();
3941  I != IE; ++I) {
3942  PHINode *DestPN = cast<PHINode>(I);
3943 
3944  int Idx = DestPN->getBasicBlockIndex(BB);
3945  // Since BB unwinds to UnwindDest, it has to be in the PHI node.
3946  assert(Idx != -1);
3947  // This PHI node has an incoming value that corresponds to a control
3948  // path through the cleanup pad we are removing. If the incoming
3949  // value is in the cleanup pad, it must be a PHINode (because we
3950  // verified above that the block is otherwise empty). Otherwise, the
3951  // value is either a constant or a value that dominates the cleanup
3952  // pad being removed.
3953  //
3954  // Because BB and UnwindDest are both EH pads, all of their
3955  // predecessors must unwind to these blocks, and since no instruction
3956  // can have multiple unwind destinations, there will be no overlap in
3957  // incoming blocks between SrcPN and DestPN.
3958  Value *SrcVal = DestPN->getIncomingValue(Idx);
3959  PHINode *SrcPN = dyn_cast<PHINode>(SrcVal);
3960 
3961  // Remove the entry for the block we are deleting.
3962  DestPN->removeIncomingValue(Idx, false);
3963 
3964  if (SrcPN && SrcPN->getParent() == BB) {
3965  // If the incoming value was a PHI node in the cleanup pad we are
3966  // removing, we need to merge that PHI node's incoming values into
3967  // DestPN.
3968  for (unsigned SrcIdx = 0, SrcE = SrcPN->getNumIncomingValues();
3969  SrcIdx != SrcE; ++SrcIdx) {
3970  DestPN->addIncoming(SrcPN->getIncomingValue(SrcIdx),
3971  SrcPN->getIncomingBlock(SrcIdx));
3972  }
3973  } else {
3974  // Otherwise, the incoming value came from above BB and
3975  // so we can just reuse it. We must associate all of BB's
3976  // predecessors with this value.
3977  for (auto *pred : predecessors(BB)) {
3978  DestPN->addIncoming(SrcVal, pred);
3979  }
3980  }
3981  }
3982 
3983  // Sink any remaining PHI nodes directly into UnwindDest.
3984  Instruction *InsertPt = DestEHPad;
3985  for (BasicBlock::iterator I = BB->begin(),
3986  IE = BB->getFirstNonPHI()->getIterator();
3987  I != IE;) {
3988  // The iterator must be incremented here because the instructions are
3989  // being moved to another block.
3990  PHINode *PN = cast<PHINode>(I++);
3991  if (PN->use_empty())
3992  // If the PHI node has no uses, just leave it. It will be erased
3993  // when we erase BB below.
3994  continue;
3995 
3996  // Otherwise, sink this PHI node into UnwindDest.
3997  // Any predecessors to UnwindDest which are not already represented
3998  // must be back edges which inherit the value from the path through
3999  // BB. In this case, the PHI value must reference itself.
4000  for (auto *pred : predecessors(UnwindDest))
4001  if (pred != BB)
4002  PN->addIncoming(PN, pred);
4003  PN->moveBefore(InsertPt);
4004  }
4005  }
4006 
4007  for (pred_iterator PI = pred_begin(BB), PE = pred_end(BB); PI != PE;) {
4008  // The iterator must be updated here because we are removing this pred.
4009  BasicBlock *PredBB = *PI++;
4010  if (UnwindDest == nullptr) {
4011  removeUnwindEdge(PredBB);
4012  } else {
4013  Instruction *TI = PredBB->getTerminator();
4014  TI->replaceUsesOfWith(BB, UnwindDest);
4015  }
4016  }
4017 
4018  // The cleanup pad is now unreachable. Zap it.
4019  BB->eraseFromParent();
4020  return true;
4021 }
4022 
4023 // Try to merge two cleanuppads together.
4025  // Skip any cleanuprets which unwind to caller, there is nothing to merge
4026  // with.
4027  BasicBlock *UnwindDest = RI->getUnwindDest();
4028  if (!UnwindDest)
4029  return false;
4030 
4031  // This cleanupret isn't the only predecessor of this cleanuppad, it wouldn't
4032  // be safe to merge without code duplication.
4033  if (UnwindDest->getSinglePredecessor() != RI->getParent())
4034  return false;
4035 
4036  // Verify that our cleanuppad's unwind destination is another cleanuppad.
4037  auto *SuccessorCleanupPad = dyn_cast<CleanupPadInst>(&UnwindDest->front());
4038  if (!SuccessorCleanupPad)
4039  return false;
4040 
4041  CleanupPadInst *PredecessorCleanupPad = RI->getCleanupPad();
4042  // Replace any uses of the successor cleanupad with the predecessor pad
4043  // The only cleanuppad uses should be this cleanupret, it's cleanupret and
4044  // funclet bundle operands.
4045  SuccessorCleanupPad->replaceAllUsesWith(PredecessorCleanupPad);
4046  // Remove the old cleanuppad.
4047  SuccessorCleanupPad->eraseFromParent();
4048  // Now, we simply replace the cleanupret with a branch to the unwind
4049  // destination.
4050  BranchInst::Create(UnwindDest, RI->getParent());
4051  RI->eraseFromParent();
4052 
4053  return true;
4054 }
4055 
4056 bool SimplifyCFGOpt::SimplifyCleanupReturn(CleanupReturnInst *RI) {
4057  // It is possible to transiantly have an undef cleanuppad operand because we
4058  // have deleted some, but not all, dead blocks.
4059  // Eventually, this block will be deleted.
4060  if (isa<UndefValue>(RI->getOperand(0)))
4061  return false;
4062 
4063  if (mergeCleanupPad(RI))
4064  return true;
4065 
4066  if (removeEmptyCleanup(RI))
4067  return true;
4068 
4069  return false;
4070 }
4071 
4072 bool SimplifyCFGOpt::SimplifyReturn(ReturnInst *RI, IRBuilder<> &Builder) {
4073  BasicBlock *BB = RI->getParent();
4074  if (!BB->getFirstNonPHIOrDbg()->isTerminator())
4075  return false;
4076 
4077  // Find predecessors that end with branches.
4078  SmallVector<BasicBlock *, 8> UncondBranchPreds;
4079  SmallVector<BranchInst *, 8> CondBranchPreds;
4080  for (pred_iterator PI = pred_begin(BB), E = pred_end(BB); PI != E; ++PI) {
4081  BasicBlock *P = *PI;
4082  Instruction *PTI = P->getTerminator();
4083  if (BranchInst *BI = dyn_cast<BranchInst>(PTI)) {
4084  if (BI->isUnconditional())
4085  UncondBranchPreds.push_back(P);
4086  else
4087  CondBranchPreds.push_back(BI);
4088  }
4089  }
4090 
4091  // If we found some, do the transformation!
4092  if (!UncondBranchPreds.empty() && DupRet) {
4093  while (!UncondBranchPreds.empty()) {
4094  BasicBlock *Pred = UncondBranchPreds.pop_back_val();
4095  LLVM_DEBUG(dbgs() << "FOLDING: " << *BB
4096  << "INTO UNCOND BRANCH PRED: " << *Pred);
4097  (void)FoldReturnIntoUncondBranch(RI, BB, Pred);
4098  }
4099 
4100  // If we eliminated all predecessors of the block, delete the block now.
4101  if (pred_empty(BB)) {
4102  // We know there are no successors, so just nuke the block.
4103  if (LoopHeaders)
4104  LoopHeaders->erase(BB);
4105  BB->eraseFromParent();
4106  }
4107 
4108  return true;
4109  }
4110 
4111  // Check out all of the conditional branches going to this return
4112  // instruction. If any of them just select between returns, change the
4113  // branch itself into a select/return pair.
4114  while (!CondBranchPreds.empty()) {
4115  BranchInst *BI = CondBranchPreds.pop_back_val();
4116 
4117  // Check to see if the non-BB successor is also a return block.
4118  if (isa<ReturnInst>(BI->getSuccessor(0)->getTerminator()) &&
4119  isa<ReturnInst>(BI->getSuccessor(1)->getTerminator()) &&
4120  SimplifyCondBranchToTwoReturns(BI, Builder))
4121  return true;
4122  }
4123  return false;
4124 }
4125 
4126 bool SimplifyCFGOpt::SimplifyUnreachable(UnreachableInst *UI) {
4127  BasicBlock *BB = UI->getParent();
4128 
4129  bool Changed = false;
4130 
4131  // If there are any instructions immediately before the unreachable that can
4132  // be removed, do so.
4133  while (UI->getIterator() != BB->begin()) {
4134  BasicBlock::iterator BBI = UI->getIterator();
4135  --BBI;
4136  // Do not delete instructions that can have side effects which might cause
4137  // the unreachable to not be reachable; specifically, calls and volatile
4138  // operations may have this effect.
4139  if (isa<CallInst>(BBI) && !isa<DbgInfoIntrinsic>(BBI))
4140  break;
4141 
4142  if (BBI->mayHaveSideEffects()) {
4143  if (auto *SI = dyn_cast<StoreInst>(BBI)) {
4144  if (SI->isVolatile())
4145  break;
4146  } else if (auto *LI = dyn_cast<LoadInst>(BBI)) {
4147  if (LI->isVolatile())
4148  break;
4149  } else if (auto *RMWI = dyn_cast<AtomicRMWInst>(BBI)) {
4150  if (RMWI->isVolatile())
4151  break;
4152  } else if (auto *CXI = dyn_cast<AtomicCmpXchgInst>(BBI)) {
4153  if (CXI->isVolatile())
4154  break;
4155  } else if (isa<CatchPadInst>(BBI)) {
4156  // A catchpad may invoke exception object constructors and such, which
4157  // in some languages can be arbitrary code, so be conservative by
4158  // default.
4159  // For CoreCLR, it just involves a type test, so can be removed.
4162  break;
4163  } else if (!isa<FenceInst>(BBI) && !isa<VAArgInst>(BBI) &&
4164  !isa<LandingPadInst>(BBI)) {
4165  break;
4166  }
4167  // Note that deleting LandingPad's here is in fact okay, although it
4168  // involves a bit of subtle reasoning. If this inst is a LandingPad,
4169  // all the predecessors of this block will be the unwind edges of Invokes,
4170  // and we can therefore guarantee this block will be erased.
4171  }
4172 
4173  // Delete this instruction (any uses are guaranteed to be dead)
4174  if (!BBI->use_empty())
4175  BBI->replaceAllUsesWith(UndefValue::get(BBI->getType()));
4176  BBI->eraseFromParent();
4177  Changed = true;
4178  }
4179 
4180  // If the unreachable instruction is the first in the block, take a gander
4181  // at all of the predecessors of this instruction, and simplify them.
4182  if (&BB->front() != UI)
4183  return Changed;
4184 
4186  for (unsigned i = 0, e = Preds.size(); i != e; ++i) {
4187  Instruction *TI = Preds[i]->getTerminator();
4188  IRBuilder<> Builder(TI);
4189  if (auto *BI = dyn_cast<BranchInst>(TI)) {
4190  if (BI->isUnconditional()) {
4191  if (BI->getSuccessor(0) == BB) {
4192  new UnreachableInst(TI->getContext(), TI);
4193  TI->eraseFromParent();
4194  Changed = true;
4195  }
4196  } else {
4197  if (BI->getSuccessor(0) == BB) {
4198  Builder.CreateBr(BI->getSuccessor(1));
4200  } else if (BI->getSuccessor(1) == BB) {
4201  Builder.CreateBr(BI->getSuccessor(0));
4203  Changed = true;
4204  }
4205  }
4206  } else if (auto *SI = dyn_cast<SwitchInst>(TI)) {
4207  for (auto i = SI->case_begin(), e = SI->case_end(); i != e;) {
4208  if (i->getCaseSuccessor() != BB) {
4209  ++i;
4210  continue;
4211  }
4212  BB->removePredecessor(SI->getParent());
4213  i = SI->removeCase(i);
4214  e = SI->case_end();
4215  Changed = true;
4216  }
4217  } else if (auto *II = dyn_cast<InvokeInst>(TI)) {
4218  if (II->getUnwindDest() == BB) {
4219  removeUnwindEdge(TI->getParent());
4220  Changed = true;
4221  }
4222  } else if (auto *CSI = dyn_cast<CatchSwitchInst>(TI)) {
4223  if (CSI->getUnwindDest() == BB) {
4224  removeUnwindEdge(TI->getParent());
4225  Changed = true;
4226  continue;
4227  }
4228 
4229  for (CatchSwitchInst::handler_iterator I = CSI->handler_begin(),
4230  E = CSI->handler_end();
4231  I != E; ++I) {
4232  if (*I == BB) {
4233  CSI->removeHandler(I);
4234  --I;
4235  --E;
4236  Changed = true;
4237  }
4238  }
4239  if (CSI->getNumHandlers() == 0) {
4240  BasicBlock *CatchSwitchBB = CSI->getParent();
4241  if (CSI->hasUnwindDest()) {
4242  // Redirect preds to the unwind dest
4243  CatchSwitchBB->replaceAllUsesWith(CSI->getUnwindDest());
4244  } else {
4245  // Rewrite all preds to unwind to caller (or from invoke to call).
4246  SmallVector<BasicBlock *, 8> EHPreds(predecessors(CatchSwitchBB));
4247  for (BasicBlock *EHPred : EHPreds)
4248  removeUnwindEdge(EHPred);
4249  }
4250  // The catchswitch is no longer reachable.
4251  new UnreachableInst(CSI->getContext(), CSI);
4252  CSI->eraseFromParent();
4253  Changed = true;
4254  }
4255  } else if (isa<CleanupReturnInst>(TI)) {
4256  new UnreachableInst(TI->getContext(), TI);
4257  TI->eraseFromParent();
4258  Changed = true;
4259  }
4260  }
4261 
4262  // If this block is now dead, remove it.
4263  if (pred_empty(BB) && BB != &BB->getParent()->getEntryBlock()) {
4264  // We know there are no successors, so just nuke the block.
4265  if (LoopHeaders)
4266  LoopHeaders->erase(BB);
4267  BB->eraseFromParent();
4268  return true;
4269  }
4270 
4271  return Changed;
4272 }
4273 
4275  assert(Cases.size() >= 1);
4276 
4277  array_pod_sort(Cases.begin(), Cases.end(), ConstantIntSortPredicate);
4278  for (size_t I = 1, E = Cases.size(); I != E; ++I) {
4279  if (Cases[I - 1]->getValue() != Cases[I]->getValue() + 1)
4280  return false;
4281  }
4282  return true;
4283 }
4284 
4285 /// Turn a switch with two reachable destinations into an integer range
4286 /// comparison and branch.
4288  assert(SI->getNumCases() > 1 && "Degenerate switch?");
4289 
4290  bool HasDefault =
4291  !isa<UnreachableInst>(SI->getDefaultDest()->getFirstNonPHIOrDbg());
4292 
4293  // Partition the cases into two sets with different destinations.
4294  BasicBlock *DestA = HasDefault ? SI->getDefaultDest() : nullptr;
4295  BasicBlock *DestB = nullptr;
4298 
4299  for (auto Case : SI->cases()) {
4300  BasicBlock *Dest = Case.getCaseSuccessor();
4301  if (!DestA)
4302  DestA = Dest;
4303  if (Dest == DestA) {
4304  CasesA.push_back(Case.getCaseValue());
4305  continue;
4306  }
4307  if (!DestB)
4308  DestB = Dest;
4309  if (Dest == DestB) {
4310  CasesB.push_back(Case.getCaseValue());
4311  continue;
4312  }
4313  return false; // More than two destinations.
4314  }
4315 
4316  assert(DestA && DestB &&
4317  "Single-destination switch should have been folded.");
4318  assert(DestA != DestB);
4319  assert(DestB != SI->getDefaultDest());
4320  assert(!CasesB.empty() && "There must be non-default cases.");
4321  assert(!CasesA.empty() || HasDefault);
4322 
4323  // Figure out if one of the sets of cases form a contiguous range.
4324  SmallVectorImpl<ConstantInt *> *ContiguousCases = nullptr;
4325  BasicBlock *ContiguousDest = nullptr;
4326  BasicBlock *OtherDest = nullptr;
4327  if (!CasesA.empty() && CasesAreContiguous(CasesA)) {
4328  ContiguousCases = &CasesA;
4329  ContiguousDest = DestA;
4330  OtherDest = DestB;
4331  } else if (CasesAreContiguous(CasesB)) {
4332  ContiguousCases = &CasesB;
4333  ContiguousDest = DestB;
4334  OtherDest = DestA;
4335  } else
4336  return false;
4337 
4338  // Start building the compare and branch.
4339 
4340  Constant *Offset = ConstantExpr::getNeg(ContiguousCases->back());
4341  Constant *NumCases =
4342  ConstantInt::get(Offset->getType(), ContiguousCases->size());
4343 
4344  Value *Sub = SI->getCondition();
4345  if (!Offset->isNullValue())
4346  Sub = Builder.CreateAdd(Sub, Offset, Sub->getName() + ".off");
4347 
4348  Value *Cmp;
4349  // If NumCases overflowed, then all possible values jump to the successor.
4350  if (NumCases->isNullValue() && !ContiguousCases->empty())
4351  Cmp = ConstantInt::getTrue(SI->getContext());
4352  else
4353  Cmp = Builder.CreateICmpULT(Sub, NumCases, "switch");
4354  BranchInst *NewBI = Builder.CreateCondBr(Cmp, ContiguousDest, OtherDest);
4355 
4356  // Update weight for the newly-created conditional branch.
4357  if (HasBranchWeights(SI)) {
4358  SmallVector<uint64_t, 8> Weights;
4359  GetBranchWeights(SI, Weights);
4360  if (Weights.size() == 1 + SI->getNumCases()) {
4361  uint64_t TrueWeight = 0;
4362  uint64_t FalseWeight = 0;
4363  for (size_t I = 0, E = Weights.size(); I != E; ++I) {
4364  if (SI->getSuccessor(I) == ContiguousDest)
4365  TrueWeight += Weights[I];
4366  else
4367  FalseWeight += Weights[I];
4368  }
4369  while (TrueWeight > UINT32_MAX || FalseWeight > UINT32_MAX) {
4370  TrueWeight /= 2;
4371  FalseWeight /= 2;
4372  }
4373  setBranchWeights(NewBI, TrueWeight, FalseWeight);
4374  }
4375  }
4376 
4377  // Prune obsolete incoming values off the successors' PHI nodes.
4378  for (auto BBI = ContiguousDest->begin(); isa<PHINode>(BBI); ++BBI) {
4379  unsigned PreviousEdges = ContiguousCases->size();
4380  if (ContiguousDest == SI->getDefaultDest())
4381  ++PreviousEdges;
4382  for (unsigned I = 0, E = PreviousEdges - 1; I != E; ++I)
4383  cast<PHINode>(BBI)->removeIncomingValue(SI->getParent());
4384  }
4385  for (auto BBI = OtherDest->begin(); isa<PHINode>(BBI); ++BBI) {
4386  unsigned PreviousEdges = SI->getNumCases() - ContiguousCases->size();
4387  if (OtherDest == SI->getDefaultDest())
4388  ++PreviousEdges;
4389  for (unsigned I = 0, E = PreviousEdges - 1; I != E; ++I)
4390  cast<PHINode>(BBI)->removeIncomingValue(SI->getParent());
4391  }
4392 
4393  // Drop the switch.
4394  SI->eraseFromParent();
4395 
4396  return true;
4397 }
4398 
4399 /// Compute masked bits for the condition of a switch
4400 /// and use it to remove dead cases.
4402  const DataLayout &DL) {
4403  Value *Cond = SI->getCondition();
4404  unsigned Bits = Cond->getType()->getIntegerBitWidth();
4405  KnownBits Known = computeKnownBits(Cond, DL, 0, AC, SI);
4406 
4407  // We can also eliminate cases by determining that their values are outside of
4408  // the limited range of the condition based on how many significant (non-sign)
4409  // bits are in the condition value.
4410  unsigned ExtraSignBits = ComputeNumSignBits(Cond, DL, 0, AC, SI) - 1;
4411  unsigned MaxSignificantBitsInCond = Bits - ExtraSignBits;
4412 
4413  // Gather dead cases.
4415  for (auto &Case : SI->cases()) {
4416  const APInt &CaseVal = Case.getCaseValue()->getValue();
4417  if (Known.Zero.intersects(CaseVal) || !Known.One.isSubsetOf(CaseVal) ||
4418  (CaseVal.getMinSignedBits() > MaxSignificantBitsInCond)) {
4419  DeadCases.push_back(Case.getCaseValue());
4420  LLVM_DEBUG(dbgs() << "SimplifyCFG: switch case " << CaseVal
4421  << " is dead.\n");
4422  }
4423  }
4424 
4425  // If we can prove that the cases must cover all possible values, the
4426  // default destination becomes dead and we can remove it. If we know some
4427  // of the bits in the value, we can use that to more precisely compute the
4428  // number of possible unique case values.
4429  bool HasDefault =
4430  !isa<UnreachableInst>(SI->getDefaultDest()->getFirstNonPHIOrDbg());
4431  const unsigned NumUnknownBits =
4432  Bits - (Known.Zero | Known.One).countPopulation();
4433  assert(NumUnknownBits <= Bits);
4434  if (HasDefault && DeadCases.empty() &&
4435  NumUnknownBits < 64 /* avoid overflow */ &&
4436  SI->getNumCases() == (1ULL << NumUnknownBits)) {
4437  LLVM_DEBUG(dbgs() << "SimplifyCFG: switch default is dead.\n");
4438  BasicBlock *NewDefault =
4440  SI->setDefaultDest(&*NewDefault);
4441  SplitBlock(&*NewDefault, &NewDefault->front());
4442  auto *OldTI = NewDefault->getTerminator();
4443  new UnreachableInst(SI->getContext(), OldTI);
4445  return true;
4446  }
4447 
4448  SmallVector<uint64_t, 8> Weights;
4449  bool HasWeight = HasBranchWeights(SI);
4450  if (HasWeight) {
4451  GetBranchWeights(SI, Weights);
4452  HasWeight = (Weights.size() == 1 + SI->getNumCases());
4453  }
4454 
4455  // Remove dead cases from the switch.
4456  for (ConstantInt *DeadCase : DeadCases) {
4457  SwitchInst::CaseIt CaseI = SI->findCaseValue(DeadCase);
4458  assert(CaseI != SI->case_default() &&
4459  "Case was not found. Probably mistake in DeadCases forming.");
4460  if (HasWeight) {
4461  std::swap(Weights[CaseI->getCaseIndex() + 1], Weights.back());
4462  Weights.pop_back();
4463  }
4464 
4465  // Prune unused values from PHI nodes.
4466  CaseI->getCaseSuccessor()->removePredecessor(SI->getParent());
4467  SI->removeCase(CaseI);
4468  }
4469  if (HasWeight && Weights.size() >= 2) {
4470  SmallVector<uint32_t, 8> MDWeights(Weights.begin(), Weights.end());
4471  setBranchWeights(SI, MDWeights);
4472  }
4473 
4474  return !DeadCases.empty();
4475 }
4476 
4477 /// If BB would be eligible for simplification by
4478 /// TryToSimplifyUncondBranchFromEmptyBlock (i.e. it is empty and terminated
4479 /// by an unconditional branch), look at the phi node for BB in the successor
4480 /// block and see if the incoming value is equal to CaseValue. If so, return
4481 /// the phi node, and set PhiIndex to BB's index in the phi node.
4483  BasicBlock *BB, int *PhiIndex) {
4484  if (BB->getFirstNonPHIOrDbg() != BB->getTerminator())
4485  return nullptr; // BB must be empty to be a candidate for simplification.
4486  if (!BB->getSinglePredecessor())
4487  return nullptr; // BB must be dominated by the switch.
4488 
4490  if (!Branch || !Branch->isUnconditional())
4491  return nullptr; // Terminator must be unconditional branch.
4492 
4493  BasicBlock *Succ = Branch->getSuccessor(0);
4494 
4495  for (PHINode &PHI : Succ->phis()) {
4496  int Idx = PHI.getBasicBlockIndex(BB);
4497  assert(Idx >= 0 && "PHI has no entry for predecessor?");
4498 
4499  Value *InValue = PHI.getIncomingValue(Idx);
4500  if (InValue != CaseValue)
4501  continue;
4502 
4503  *PhiIndex = Idx;
4504  return &PHI;
4505  }
4506 
4507  return nullptr;
4508 }
4509 
4510 /// Try to forward the condition of a switch instruction to a phi node
4511 /// dominated by the switch, if that would mean that some of the destination
4512 /// blocks of the switch can be folded away. Return true if a change is made.
4514  using ForwardingNodesMap = DenseMap<PHINode *, SmallVector<int, 4>>;
4515 
4516  ForwardingNodesMap ForwardingNodes;
4517  BasicBlock *SwitchBlock = SI->getParent();
4518  bool Changed = false;
4519  for (auto &Case : SI->cases()) {
4520  ConstantInt *CaseValue = Case.getCaseValue();
4521  BasicBlock *CaseDest = Case.getCaseSuccessor();
4522 
4523  // Replace phi operands in successor blocks that are using the constant case
4524  // value rather than the switch condition variable:
4525  // switchbb:
4526  // switch i32 %x, label %default [
4527  // i32 17, label %succ
4528  // ...
4529  // succ:
4530  // %r = phi i32 ... [ 17, %switchbb ] ...
4531  // -->
4532  // %r = phi i32 ... [ %x, %switchbb ] ...
4533 
4534  for (PHINode &Phi : CaseDest->phis()) {
4535  // This only works if there is exactly 1 incoming edge from the switch to
4536  // a phi. If there is >1, that means multiple cases of the switch map to 1
4537  // value in the phi, and that phi value is not the switch condition. Thus,
4538  // this transform would not make sense (the phi would be invalid because
4539  // a phi can't have different incoming values from the same block).
4540  int SwitchBBIdx = Phi.getBasicBlockIndex(SwitchBlock);
4541  if (Phi.getIncomingValue(SwitchBBIdx) == CaseValue &&
4542  count(Phi.blocks(), SwitchBlock) == 1) {
4543  Phi.setIncomingValue(SwitchBBIdx, SI->getCondition());
4544  Changed = true;
4545  }
4546  }
4547 
4548  // Collect phi nodes that are indirectly using this switch's case constants.
4549  int PhiIdx;
4550  if (auto *Phi = FindPHIForConditionForwarding(CaseValue, CaseDest, &PhiIdx))
4551  ForwardingNodes[Phi].push_back(PhiIdx);
4552  }
4553 
4554  for (auto &ForwardingNode : ForwardingNodes) {
4555  PHINode *Phi = ForwardingNode.first;
4556  SmallVectorImpl<int> &Indexes = ForwardingNode.second;
4557  if (Indexes.size() < 2)
4558  continue;
4559 
4560  for (int Index : Indexes)
4561  Phi->setIncomingValue(Index, SI->getCondition());
4562  Changed = true;
4563  }
4564 
4565  return Changed;
4566 }
4567 
4568 /// Return true if the backend will be able to handle
4569 /// initializing an array of constants like C.
4571  if (C->isThreadDependent())
4572  return false;
4573  if (C->isDLLImportDependent())
4574  return false;
4575 
4576  if (!isa<ConstantFP>(C) && !isa<ConstantInt>(C) &&
4577  !isa<ConstantPointerNull>(C) && !isa<GlobalValue>(C) &&
4578  !isa<UndefValue>(C) && !isa<ConstantExpr>(C))
4579  return false;
4580 
4581  if (ConstantExpr *CE = dyn_cast<ConstantExpr>(C)) {
4582  if (!CE->isGEPWithNoNotionalOverIndexing())
4583  return false;
4584  if (!ValidLookupTableConstant(CE->getOperand(0), TTI))
4585  return false;
4586  }
4587 
4589  return false;
4590 
4591  return true;
4592 }
4593 
4594 /// If V is a Constant, return it. Otherwise, try to look up
4595 /// its constant value in ConstantPool, returning 0 if it's not there.
4596 static Constant *
4599  if (Constant *C = dyn_cast<Constant>(V))
4600  return C;
4601  return ConstantPool.lookup(V);
4602 }
4603 
4604 /// Try to fold instruction I into a constant. This works for
4605 /// simple instructions such as binary operations where both operands are
4606 /// constant or can be replaced by constants from the ConstantPool. Returns the
4607 /// resulting constant on success, 0 otherwise.
4608 static Constant *
4611  if (SelectInst *Select = dyn_cast<SelectInst>(I)) {
4612  Constant *A = LookupConstant(Select->getCondition(), ConstantPool);
4613  if (!A)
4614  return nullptr;
4615  if (A->isAllOnesValue())
4616  return LookupConstant(Select->getTrueValue(), ConstantPool);
4617  if (A->isNullValue())
4618  return LookupConstant(Select->getFalseValue(), ConstantPool);
4619  return nullptr;
4620  }
4621 
4623  for (unsigned N = 0, E = I->getNumOperands(); N != E; ++N) {
4625  COps.push_back(A);
4626  else
4627  return nullptr;
4628  }
4629 
4630  if (CmpInst *Cmp = dyn_cast<CmpInst>(I)) {
4631  return ConstantFoldCompareInstOperands(Cmp->getPredicate(), COps[0],
4632  COps[1], DL);
4633  }
4634 
4635  return ConstantFoldInstOperands(I, COps, DL);
4636 }
4637 
4638 /// Try to determine the resulting constant values in phi nodes
4639 /// at the common destination basic block, *CommonDest, for one of the case
4640 /// destionations CaseDest corresponding to value CaseVal (0 for the default
4641 /// case), of a switch instruction SI.
4642 static bool
4644  BasicBlock **CommonDest,
4645  SmallVectorImpl<std::pair<PHINode *, Constant *>> &Res,
4646  const DataLayout &DL, const TargetTransformInfo &TTI) {
4647  // The block from which we enter the common destination.
4648  BasicBlock *Pred = SI->getParent();
4649 
4650  // If CaseDest is empty except for some side-effect free instructions through
4651  // which we can constant-propagate the CaseVal, continue to its successor.
4653  ConstantPool.insert(std::make_pair(SI->getCondition(), CaseVal));
4654  for (Instruction &I :CaseDest->instructionsWithoutDebug()) {
4655  if (I.isTerminator()) {
4656  // If the terminator is a simple branch, continue to the next block.
4657  if (I.getNumSuccessors() != 1 || I.isExceptionalTerminator())
4658  return false;
4659  Pred = CaseDest;
4660  CaseDest = I.getSuccessor(0);
4661  } else if (Constant *C = ConstantFold(&I, DL, ConstantPool)) {
4662  // Instruction is side-effect free and constant.
4663 
4664  // If the instruction has uses outside this block or a phi node slot for
4665  // the block, it is not safe to bypass the instruction since it would then
4666  // no longer dominate all its uses.
4667  for (auto &Use : I.uses()) {
4668  User *User = Use.getUser();
4669  if (Instruction *I = dyn_cast<Instruction>(User))
4670  if (I->getParent() == CaseDest)
4671  continue;
4672  if (PHINode *Phi = dyn_cast<PHINode>(User))
4673  if (Phi->getIncomingBlock(Use) == CaseDest)
4674  continue;
4675  return false;
4676  }
4677 
4678  ConstantPool.insert(std::make_pair(&I, C));
4679  } else {
4680  break;
4681  }
4682  }
4683 
4684  // If we did not have a CommonDest before, use the current one.
4685  if (!*CommonDest)
4686  *CommonDest = CaseDest;
4687  // If the destination isn't the common one, abort.
4688  if (CaseDest != *CommonDest)
4689  return false;
4690 
4691  // Get the values for this case from phi nodes in the destination block.
4692  for (PHINode &PHI : (*CommonDest)->phis()) {
4693  int Idx = PHI.getBasicBlockIndex(Pred);
4694  if (Idx == -1)
4695  continue;
4696 
4697  Constant *ConstVal =
4698  LookupConstant(PHI.getIncomingValue(Idx), ConstantPool);
4699  if (!ConstVal)
4700  return false;
4701 
4702  // Be conservative about which kinds of constants we support.
4703  if (!ValidLookupTableConstant(ConstVal, TTI))
4704  return false;
4705 
4706  Res.push_back(std::make_pair(&PHI, ConstVal));
4707  }
4708 
4709  return Res.size() > 0;
4710 }
4711 
4712 // Helper function used to add CaseVal to the list of cases that generate
4713 // Result. Returns the updated number of cases that generate this result.
4714 static uintptr_t MapCaseToResult(ConstantInt *CaseVal,
4715  SwitchCaseResultVectorTy &UniqueResults,
4716  Constant *Result) {
4717  for (auto &I : UniqueResults) {
4718  if (I.first == Result) {
4719  I.second.push_back(CaseVal);
4720  return I.second.size();
4721  }
4722  }
4723  UniqueResults.push_back(
4724  std::make_pair(Result, SmallVector<ConstantInt *, 4>(1, CaseVal)));
4725  return 1;
4726 }
4727 
4728 // Helper function that initializes a map containing
4729 // results for the PHI node of the common destination block for a switch
4730 // instruction. Returns false if multiple PHI nodes have been found or if
4731 // there is not a common destination block for the switch.
4732 static bool
4734  SwitchCaseResultVectorTy &UniqueResults,
4735  Constant *&DefaultResult, const DataLayout &DL,
4736  const TargetTransformInfo &TTI,
4737  uintptr_t MaxUniqueResults, uintptr_t MaxCasesPerResult) {
4738  for (auto &I : SI->cases()) {
4739  ConstantInt *CaseVal = I.getCaseValue();
4740 
4741  // Resulting value at phi nodes for this case value.
4742  SwitchCaseResultsTy Results;
4743  if (!GetCaseResults(SI, CaseVal, I.getCaseSuccessor(), &CommonDest, Results,
4744  DL, TTI))
4745  return false;
4746 
4747  // Only one value per case is permitted.
4748  if (Results.size() > 1)
4749  return false;
4750 
4751  // Add the case->result mapping to UniqueResults.
4752  const uintptr_t NumCasesForResult =
4753  MapCaseToResult(CaseVal, UniqueResults, Results.begin()->second);
4754 
4755  // Early out if there are too many cases for this result.
4756  if (NumCasesForResult > MaxCasesPerResult)
4757  return false;
4758 
4759  // Early out if there are too many unique results.
4760  if (UniqueResults.size() > MaxUniqueResults)
4761  return false;
4762 
4763  // Check the PHI consistency.
4764  if (!PHI)
4765  PHI = Results[0].first;
4766  else if (PHI != Results[0].first)
4767  return false;
4768  }
4769  // Find the default result value.
4771  BasicBlock *DefaultDest = SI->getDefaultDest();
4772  GetCaseResults(SI, nullptr, SI->getDefaultDest(), &CommonDest, DefaultResults,
4773  DL, TTI);
4774  // If the default value is not found abort unless the default destination
4775  // is unreachable.
4776  DefaultResult =
4777  DefaultResults.size() == 1 ? DefaultResults.begin()->second : nullptr;
4778  if ((!DefaultResult &&
4779  !isa<UnreachableInst>(DefaultDest->getFirstNonPHIOrDbg())))
4780  return false;
4781 
4782  return true;
4783 }
4784 
4785 // Helper function that checks if it is possible to transform a switch with only
4786 // two cases (or two cases + default) that produces a result into a select.
4787 // Example:
4788 // switch (a) {
4789 // case 10: %0 = icmp eq i32 %a, 10
4790 // return 10; %1 = select i1 %0, i32 10, i32 4
4791 // case 20: ----> %2 = icmp eq i32 %a, 20
4792 // return 2; %3 = select i1 %2, i32 2, i32 %1
4793 // default:
4794 // return 4;
4795 // }
4796 static Value *ConvertTwoCaseSwitch(const SwitchCaseResultVectorTy &ResultVector,
4797  Constant *DefaultResult, Value *Condition,
4798  IRBuilder<> &Builder) {
4799  assert(ResultVector.size() == 2 &&
4800  "We should have exactly two unique results at this point");
4801  // If we are selecting between only two cases transform into a simple
4802  // select or a two-way select if default is possible.
4803  if (ResultVector[0].second.size() == 1 &&
4804  ResultVector[1].second.size() == 1) {
4805  ConstantInt *const FirstCase = ResultVector[0].second[0];
4806  ConstantInt *const SecondCase = ResultVector[1].second[0];
4807 
4808  bool DefaultCanTrigger = DefaultResult;
4809  Value *SelectValue = ResultVector[1].first;
4810  if (DefaultCanTrigger) {
4811  Value *const ValueCompare =
4812  Builder.CreateICmpEQ(Condition, SecondCase, "switch.selectcmp");
4813  SelectValue = Builder.CreateSelect(ValueCompare, ResultVector[1].first,
4814  DefaultResult, "switch.select");
4815  }
4816  Value *const ValueCompare =
4817  Builder.CreateICmpEQ(Condition, FirstCase, "switch.selectcmp");
4818  return Builder.CreateSelect(ValueCompare, ResultVector[0].first,
4819  SelectValue, "switch.select");
4820  }
4821 
4822  return nullptr;
4823 }
4824 
4825 // Helper function to cleanup a switch instruction that has been converted into
4826 // a select, fixing up PHI nodes and basic blocks.
4828  Value *SelectValue,
4829  IRBuilder<> &Builder) {
4830  BasicBlock *SelectBB = SI->getParent();
4831  while (PHI->getBasicBlockIndex(SelectBB) >= 0)
4832  PHI->removeIncomingValue(SelectBB);
4833  PHI->addIncoming(SelectValue, SelectBB);
4834 
4835  Builder.CreateBr(PHI->getParent());
4836 
4837  // Remove the switch.
4838  for (unsigned i = 0, e = SI->getNumSuccessors(); i < e; ++i) {
4839  BasicBlock *Succ = SI->getSuccessor(i);
4840 
4841  if (Succ == PHI->getParent())
4842  continue;
4843  Succ->removePredecessor(SelectBB);
4844  }
4845  SI->eraseFromParent();
4846 }
4847 
4848 /// If the switch is only used to initialize one or more
4849 /// phi nodes in a common successor block with only two different
4850 /// constant values, replace the switch with select.
4851 static bool switchToSelect(SwitchInst *SI, IRBuilder<> &Builder,
4852  const DataLayout &DL,
4853  const TargetTransformInfo &TTI) {
4854  Value *const Cond = SI->getCondition();
4855  PHINode *PHI = nullptr;
4856  BasicBlock *CommonDest = nullptr;
4857  Constant *DefaultResult;
4858  SwitchCaseResultVectorTy UniqueResults;
4859  // Collect all the cases that will deliver the same value from the switch.
4860  if (!InitializeUniqueCases(SI, PHI, CommonDest, UniqueResults, DefaultResult,
4861  DL, TTI, 2, 1))
4862  return false;
4863  // Selects choose between maximum two values.
4864  if (UniqueResults.size() != 2)
4865  return false;
4866  assert(PHI != nullptr && "PHI for value select not found");
4867 
4868  Builder.SetInsertPoint(SI);
4869  Value *SelectValue =
4870  ConvertTwoCaseSwitch(UniqueResults, DefaultResult, Cond, Builder);
4871  if (SelectValue) {
4872  RemoveSwitchAfterSelectConversion(SI, PHI, SelectValue, Builder);
4873  return true;
4874  }
4875  // The switch couldn't be converted into a select.
4876  return false;
4877 }
4878 
4879 namespace {
4880 
4881 /// This class represents a lookup table that can be used to replace a switch.
4882 class SwitchLookupTable {
4883 public:
4884  /// Create a lookup table to use as a switch replacement with the contents
4885  /// of Values, using DefaultValue to fill any holes in the table.
4886  SwitchLookupTable(
4887  Module &M, uint64_t TableSize, ConstantInt *Offset,
4888  const SmallVectorImpl<std::pair<ConstantInt *, Constant *>> &Values,
4889  Constant *DefaultValue, const DataLayout &DL, const StringRef &FuncName);
4890 
4891  /// Build instructions with Builder to retrieve the value at
4892  /// the position given by Index in the lookup table.
4893  Value *BuildLookup(Value *Index, IRBuilder<> &Builder);
4894 
4895  /// Return true if a table with TableSize elements of
4896  /// type ElementType would fit in a target-legal register.
4897  static bool WouldFitInRegister(const DataLayout &DL, uint64_t TableSize,
4898  Type *ElementType);
4899 
4900 private:
4901  // Depending on the contents of the table, it can be represented in
4902  // different ways.
4903  enum {
4904  // For tables where each element contains the same value, we just have to
4905  // store that single value and return it for each lookup.
4906  SingleValueKind,
4907 
4908  // For tables where there is a linear relationship between table index
4909  // and values. We calculate the result with a simple multiplication
4910  // and addition instead of a table lookup.
4911  LinearMapKind,
4912 
4913  // For small tables with integer elements, we can pack them into a bitmap
4914  // that fits into a target-legal register. Values are retrieved by
4915  // shift and mask operations.
4916  BitMapKind,
4917 
4918  // The table is stored as an array of values. Values are retrieved by load
4919  // instructions from the table.
4920  ArrayKind
4921  } Kind;
4922 
4923  // For SingleValueKind, this is the single value.
4924  Constant *SingleValue = nullptr;
4925 
4926  // For BitMapKind, this is the bitmap.
4927  ConstantInt *BitMap = nullptr;
4928  IntegerType *BitMapElementTy = nullptr;
4929 
4930  // For LinearMapKind, these are the constants used to derive the value.
4931  ConstantInt *LinearOffset = nullptr;
4932  ConstantInt *LinearMultiplier = nullptr;
4933 
4934  // For ArrayKind, this is the array.
4935  GlobalVariable *Array = nullptr;
4936 };
4937 
4938 } // end anonymous namespace
4939 
4940 SwitchLookupTable::SwitchLookupTable(
4941  Module &M, uint64_t TableSize, ConstantInt *Offset,
4942  const SmallVectorImpl<std::pair<ConstantInt *, Constant *>> &Values,
4943  Constant *DefaultValue, const DataLayout &DL, const StringRef &FuncName) {
4944  assert(Values.size() && "Can't build lookup table without values!");
4945  assert(TableSize >= Values.size() && "Can't fit values in table!");
4946 
4947  // If all values in the table are equal, this is that value.
4948  SingleValue = Values.begin()->second;
4949 
4950  Type *ValueType = Values.begin()->second->getType();
4951 
4952  // Build up the table contents.
4953  SmallVector<Constant *, 64> TableContents(TableSize);
4954  for (size_t I = 0, E = Values.size(); I != E; ++I) {
4955  ConstantInt *CaseVal = Values[I].first;
4956  Constant *CaseRes = Values[I].second;
4957  assert(CaseRes->getType() == ValueType);
4958 
4959  uint64_t Idx = (CaseVal->getValue() - Offset->getValue()).getLimitedValue();
4960  TableContents[Idx] = CaseRes;
4961 
4962  if (CaseRes != SingleValue)
4963  SingleValue = nullptr;
4964  }
4965 
4966  // Fill in any holes in the table with the default result.
4967  if (Values.size() < TableSize) {
4968  assert(DefaultValue &&
4969  "Need a default value to fill the lookup table holes.");
4970  assert(DefaultValue->getType() == ValueType);
4971  for (uint64_t I = 0; I < TableSize; ++I) {
4972  if (!TableContents[I])
4973  TableContents[I] = DefaultValue;
4974  }
4975 
4976  if (DefaultValue != SingleValue)
4977  SingleValue = nullptr;
4978  }
4979 
4980  // If each element in the table contains the same value, we only need to store
4981  // that single value.
4982  if (SingleValue) {
4983  Kind = SingleValueKind;
4984  return;
4985  }
4986 
4987  // Check if we can derive the value with a linear transformation from the
4988  // table index.
4989  if (isa<IntegerType>(ValueType)) {
4990  bool LinearMappingPossible = true;
4991  APInt PrevVal;
4992  APInt DistToPrev;
4993  assert(TableSize >= 2 && "Should be a SingleValue table.");
4994  // Check if there is the same distance between two consecutive values.
4995  for (uint64_t I = 0; I < TableSize; ++I) {
4996  ConstantInt *ConstVal = dyn_cast<ConstantInt>(TableContents[I]);
4997  if (!ConstVal) {
4998  // This is an undef. We could deal with it, but undefs in lookup tables
4999  // are very seldom. It's probably not worth the additional complexity.
5000  LinearMappingPossible = false;
5001  break;
5002  }
5003  const APInt &Val = ConstVal->getValue();
5004  if (I != 0) {
5005  APInt Dist = Val - PrevVal;
5006  if (I == 1) {
5007  DistToPrev = Dist;
5008  } else if (Dist != DistToPrev) {
5009  LinearMappingPossible = false;
5010  break;
5011  }
5012  }
5013  PrevVal = Val;
5014  }
5015  if (LinearMappingPossible) {
5016  LinearOffset = cast<ConstantInt>(TableContents[0]);
5017  LinearMultiplier = ConstantInt::get(M.getContext(), DistToPrev);
5018  Kind = LinearMapKind;
5019  ++NumLinearMaps;
5020  return;
5021  }
5022  }
5023 
5024  // If the type is integer and the table fits in a register, build a bitmap.
5025  if (WouldFitInRegister(DL, TableSize, ValueType)) {
5026  IntegerType *IT = cast<IntegerType>(ValueType);
5027  APInt TableInt(TableSize * IT->getBitWidth(), 0);
5028  for (uint64_t I = TableSize; I > 0; --I) {
5029  TableInt <<= IT->getBitWidth();
5030  // Insert values into the bitmap. Undef values are set to zero.
5031  if (!isa<UndefValue>(TableContents[I - 1])) {
5032  ConstantInt *Val = cast<ConstantInt>(TableContents[I - 1]);
5033  TableInt |= Val->getValue().zext(TableInt.getBitWidth());
5034  }
5035  }
5036  BitMap = ConstantInt::get(M.getContext(), TableInt);
5037  BitMapElementTy = IT;
5038  Kind = BitMapKind;
5039  ++NumBitMaps;
5040  return;
5041  }
5042 
5043  // Store the table in an array.
5044  ArrayType *ArrayTy = ArrayType::get(ValueType, TableSize);
5045  Constant *Initializer = ConstantArray::get(ArrayTy, TableContents);
5046 
5047  Array = new GlobalVariable(M, ArrayTy, /*constant=*/true,
5048  GlobalVariable::PrivateLinkage, Initializer,
5049  "switch.table." + FuncName);
5050  Array->setUnnamedAddr(GlobalValue::UnnamedAddr::Global);
5051  // Set the alignment to that of an array items. We will be only loading one
5052  // value out of it.
5053  Array->setAlignment(DL.getPrefTypeAlignment(ValueType));
5054  Kind = ArrayKind;
5055 }
5056 
5057 Value *SwitchLookupTable::BuildLookup(Value *Index, IRBuilder<> &Builder) {
5058  switch (Kind) {
5059  case SingleValueKind:
5060  return SingleValue;
5061  case LinearMapKind: {
5062  // Derive the result value from the input value.
5063  Value *Result = Builder.CreateIntCast(Index, LinearMultiplier->getType(),
5064  false, "switch.idx.cast");
5065  if (!LinearMultiplier->isOne())
5066  Result = Builder.CreateMul(Result, LinearMultiplier, "switch.idx.mult");
5067  if (!LinearOffset->isZero())
5068  Result = Builder.CreateAdd(Result, LinearOffset, "switch.offset");
5069  return Result;
5070  }
5071  case BitMapKind: {
5072  // Type of the bitmap (e.g. i59).
5073  IntegerType *MapTy = BitMap->getType();
5074 
5075  // Cast Index to the same type as the bitmap.
5076  // Note: The Index is <= the number of elements in the table, so
5077  // truncating it to the width of the bitmask is safe.
5078  Value *ShiftAmt = Builder.CreateZExtOrTrunc(Index, MapTy, "switch.cast");
5079 
5080  // Multiply the shift amount by the element width.
5081  ShiftAmt = Builder.CreateMul(
5082  ShiftAmt, ConstantInt::get(MapTy, BitMapElementTy->getBitWidth()),
5083  "switch.shiftamt");
5084 
5085  // Shift down.
5086  Value *DownShifted =
5087  Builder.CreateLShr(BitMap, ShiftAmt, "switch.downshift");
5088  // Mask off.
5089  return Builder.CreateTrunc(DownShifted, BitMapElementTy, "switch.masked");
5090  }
5091  case ArrayKind: {
5092  // Make sure the table index will not overflow when treated as signed.
5093  IntegerType *IT = cast<IntegerType>(Index->getType());
5094  uint64_t TableSize =
5095  Array->getInitializer()->getType()->getArrayNumElements();
5096  if (TableSize > (1ULL << (IT->getBitWidth() - 1)))
5097  Index = Builder.CreateZExt(
5098  Index, IntegerType::get(IT->getContext(), IT->getBitWidth() + 1),
5099  "switch.tableidx.zext");
5100 
5101  Value *GEPIndices[] = {Builder.getInt32(0), Index};
5102  Value *GEP = Builder.CreateInBoundsGEP(Array->getValueType(), Array,
5103  GEPIndices, "switch.gep");
5104  return Builder.CreateLoad(GEP, "switch.load");
5105  }
5106  }
5107  llvm_unreachable("Unknown lookup table kind!");
5108 }
5109 
5110 bool SwitchLookupTable::WouldFitInRegister(const DataLayout &DL,
5111  uint64_t TableSize,
5112  Type *ElementType) {
5113  auto *IT = dyn_cast<IntegerType>(ElementType);
5114  if (!IT)
5115  return false;
5116  // FIXME: If the type is wider than it needs to be, e.g. i8 but all values
5117  // are <= 15, we could try to narrow the type.
5118 
5119  // Avoid overflow, fitsInLegalInteger uses unsigned int for the width.
5120  if (TableSize >= UINT_MAX / IT->getBitWidth())
5121  return false;
5122  return DL.fitsInLegalInteger(TableSize * IT->getBitWidth());
5123 }
5124 
5125 /// Determine whether a lookup table should be built for this switch, based on
5126 /// the number of cases, size of the table, and the types of the results.
5127 static bool
5128 ShouldBuildLookupTable(SwitchInst *SI, uint64_t TableSize,
5129  const TargetTransformInfo &TTI, const DataLayout &DL,
5130  const SmallDenseMap<PHINode *, Type *> &ResultTypes) {
5131  if (SI->getNumCases() > TableSize || TableSize >= UINT64_MAX / 10)
5132  return false; // TableSize overflowed, or mul below might overflow.
5133 
5134  bool AllTablesFitInRegister = true;
5135  bool HasIllegalType = false;
5136  for (const auto &I : ResultTypes) {
5137  Type *Ty = I.second;
5138 
5139  // Saturate this flag to true.
5140  HasIllegalType = HasIllegalType || !TTI.isTypeLegal(Ty);
5141 
5142  // Saturate this flag to false.
5143  AllTablesFitInRegister =
5144  AllTablesFitInRegister &&
5145  SwitchLookupTable::WouldFitInRegister(DL, TableSize, Ty);
5146 
5147  // If both flags saturate, we're done. NOTE: This *only* works with
5148  // saturating flags, and all flags have to saturate first due to the
5149  // non-deterministic behavior of iterating over a dense map.
5150  if (HasIllegalType && !AllTablesFitInRegister)
5151  break;
5152  }
5153 
5154  // If each table would fit in a register, we should build it anyway.
5155  if (AllTablesFitInRegister)
5156  return true;
5157 
5158  // Don't build a table that doesn't fit in-register if it has illegal types.
5159  if (HasIllegalType)
5160  return false;
5161 
5162  // The table density should be at least 40%. This is the same criterion as for
5163  // jump tables, see SelectionDAGBuilder::handleJTSwitchCase.
5164  // FIXME: Find the best cut-off.
5165  return SI->getNumCases() * 10 >= TableSize * 4;
5166 }
5167 
5168 /// Try to reuse the switch table index compare. Following pattern:
5169 /// \code
5170 /// if (idx < tablesize)
5171 /// r = table[idx]; // table does not contain default_value
5172 /// else
5173 /// r = default_value;
5174 /// if (r != default_value)
5175 /// ...
5176 /// \endcode
5177 /// Is optimized to:
5178 /// \code
5179 /// cond = idx < tablesize;
5180 /// if (cond)
5181 /// r = table[idx];
5182 /// else
5183 /// r = default_value;
5184 /// if (cond)
5185 /// ...
5186 /// \endcode
5187 /// Jump threading will then eliminate the second if(cond).
5188 static void reuseTableCompare(
5189  User *PhiUser, BasicBlock *PhiBlock, BranchInst *RangeCheckBranch,
5190  Constant *DefaultValue,
5191  const SmallVectorImpl<std::pair<ConstantInt *, Constant *>> &Values) {
5192  ICmpInst *CmpInst = dyn_cast<ICmpInst>(PhiUser);
5193  if (!CmpInst)
5194  return;
5195 
5196  // We require that the compare is in the same block as the phi so that jump
5197  // threading can do its work afterwards.
5198  if (CmpInst->getParent() != PhiBlock)
5199  return;
5200 
5201  Constant *CmpOp1 = dyn_cast<Constant>(CmpInst->getOperand(1));
5202  if (!CmpOp1)
5203  return;
5204 
5205  Value *RangeCmp = RangeCheckBranch->getCondition();
5206  Constant *TrueConst = ConstantInt::getTrue(RangeCmp->getType());
5207  Constant *FalseConst = ConstantInt::getFalse(RangeCmp->getType());
5208 
5209  // Check if the compare with the default value is constant true or false.
5210  Constant *DefaultConst = ConstantExpr::getICmp(CmpInst->getPredicate(),
5211  DefaultValue, CmpOp1, true);
5212  if (DefaultConst != TrueConst && DefaultConst != FalseConst)
5213  return;
5214 
5215  // Check if the compare with the case values is distinct from the default
5216  // compare result.
5217  for (auto ValuePair : Values) {
5218  Constant *CaseConst = ConstantExpr::getICmp(CmpInst->getPredicate(),
5219  ValuePair.second, CmpOp1, true);
5220  if (!CaseConst || CaseConst == DefaultConst || isa<UndefValue>(CaseConst))
5221  return;
5222  assert((CaseConst == TrueConst || CaseConst == FalseConst) &&
5223  "Expect true or false as compare result.");
5224  }
5225 
5226  // Check if the branch instruction dominates the phi node. It's a simple
5227  // dominance check, but sufficient for our needs.
5228  // Although this check is invariant in the calling loops, it's better to do it
5229  // at this late stage. Practically we do it at most once for a switch.
5230  BasicBlock *BranchBlock = RangeCheckBranch->getParent();
5231  for (auto PI = pred_begin(PhiBlock), E = pred_end(PhiBlock); PI != E; ++PI) {
5232  BasicBlock *Pred = *PI;
5233  if (Pred != BranchBlock && Pred->getUniquePredecessor() != BranchBlock)
5234  return;
5235  }
5236 
5237  if (DefaultConst == FalseConst) {
5238  // The compare yields the same result. We can replace it.
5239  CmpInst->replaceAllUsesWith(RangeCmp);
5240  ++NumTableCmpReuses;
5241  } else {
5242  // The compare yields the same result, just inverted. We can replace it.
5243  Value *InvertedTableCmp = BinaryOperator::CreateXor(
5244  RangeCmp, ConstantInt::get(RangeCmp->getType(), 1), "inverted.cmp",
5245  RangeCheckBranch);
5246  CmpInst->replaceAllUsesWith(InvertedTableCmp);
5247  ++NumTableCmpReuses;
5248  }
5249 }
5250 
5251 /// If the switch is only used to initialize one or more phi nodes in a common
5252 /// successor block with different constant values, replace the switch with
5253 /// lookup tables.
5255  const DataLayout &DL,
5256  const TargetTransformInfo &TTI) {
5257  assert(SI->getNumCases() > 1 && "Degenerate switch?");
5258 
5259  Function *Fn = SI->getParent()->getParent();
5260  // Only build lookup table when we have a target that supports it or the
5261  // attribute is not set.
5262  if (!TTI.shouldBuildLookupTables() ||
5263  (Fn->getFnAttribute("no-jump-tables").getValueAsString() == "true"))
5264  return false;
5265 
5266  // FIXME: If the switch is too sparse for a lookup table, perhaps we could
5267  // split off a dense part and build a lookup table for that.
5268 
5269  // FIXME: This creates arrays of GEPs to constant strings, which means each
5270  // GEP needs a runtime relocation in PIC code. We should just build one big
5271  // string and lookup indices into that.
5272 
5273  // Ignore switches with less than three cases. Lookup tables will not make
5274  // them faster, so we don't analyze them.
5275  if (SI->getNumCases() < 3)
5276  return false;
5277 
5278  // Figure out the corresponding result for each case value and phi node in the
5279  // common destination, as well as the min and max case values.
5280  assert(!empty(SI->cases()));
5281  SwitchInst::CaseIt CI = SI->case_begin();
5282  ConstantInt *MinCaseVal = CI->getCaseValue();
5283  ConstantInt *MaxCaseVal = CI->getCaseValue();
5284 
5285  BasicBlock *CommonDest = nullptr;
5286 
5287  using ResultListTy = SmallVector<std::pair<ConstantInt *, Constant *>, 4>;
5289 
5290  SmallDenseMap<PHINode *, Constant *> DefaultResults;
5293 
5294  for (SwitchInst::CaseIt E = SI->case_end(); CI != E; ++CI) {
5295  ConstantInt *CaseVal = CI->getCaseValue();
5296  if (CaseVal->getValue().slt(MinCaseVal->getValue()))
5297  MinCaseVal = CaseVal;
5298  if (CaseVal->getValue().sgt(MaxCaseVal->getValue()))
5299  MaxCaseVal = CaseVal;
5300 
5301  // Resulting value at phi nodes for this case value.
5302  using ResultsTy = SmallVector<std::pair<PHINode *, Constant *>, 4>;
5303  ResultsTy Results;
5304  if (!GetCaseResults(SI, CaseVal, CI->getCaseSuccessor(), &CommonDest,
5305  Results, DL, TTI))
5306  return false;
5307 
5308  // Append the result from this case to the list for each phi.
5309  for (const auto &I : Results) {
5310  PHINode *PHI = I.first;
5311  Constant *Value = I.second;
5312  if (!ResultLists.count(PHI))
5313  PHIs.push_back(PHI);
5314  ResultLists[PHI].push_back(std::make_pair(CaseVal, Value));
5315  }
5316  }
5317 
5318  // Keep track of the result types.
5319  for (PHINode *PHI : PHIs) {
5320  ResultTypes[PHI] = ResultLists[PHI][0].second->getType();
5321  }
5322 
5323  uint64_t NumResults = ResultLists[PHIs[0]].size();
5324  APInt RangeSpread = MaxCaseVal->getValue() - MinCaseVal->getValue();
5325  uint64_t TableSize = RangeSpread.getLimitedValue() + 1;
5326  bool TableHasHoles = (NumResults < TableSize);
5327 
5328  // If the table has holes, we need a constant result for the default case
5329  // or a bitmask that fits in a register.
5330  SmallVector<std::pair<PHINode *, Constant *>, 4> DefaultResultsList;
5331  bool HasDefaultResults =
5332  GetCaseResults(SI, nullptr, SI->getDefaultDest(), &CommonDest,
5333  DefaultResultsList, DL, TTI);
5334 
5335  bool NeedMask = (TableHasHoles && !HasDefaultResults);
5336  if (NeedMask) {
5337  // As an extra penalty for the validity test we require more cases.
5338  if (SI->getNumCases() < 4) // FIXME: Find best threshold value (benchmark).
5339  return false;
5340  if (!DL.fitsInLegalInteger(TableSize))
5341  return false;
5342  }
5343 
5344  for (const auto &I : DefaultResultsList) {
5345  PHINode *PHI = I.first;
5346  Constant *Result = I.second;
5347  DefaultResults[PHI] = Result;
5348  }
5349 
5350  if (!ShouldBuildLookupTable(SI, TableSize, TTI, DL, ResultTypes))
5351  return false;
5352 
5353  // Create the BB that does the lookups.
5354  Module &Mod = *CommonDest->getParent()->getParent();
5355  BasicBlock *LookupBB = BasicBlock::Create(
5356  Mod.getContext(), "switch.lookup", CommonDest->getParent(), CommonDest);
5357 
5358  // Compute the table index value.
5359  Builder.SetInsertPoint(SI);
5360  Value *TableIndex;
5361  if (MinCaseVal->isNullValue())
5362  TableIndex = SI->getCondition();
5363  else
5364  TableIndex = Builder.CreateSub(SI->getCondition(), MinCaseVal,
5365  "switch.tableidx");
5366 
5367  // Compute the maximum table size representable by the integer type we are
5368  // switching upon.
5369  unsigned CaseSize = MinCaseVal->getType()->getPrimitiveSizeInBits();
5370  uint64_t MaxTableSize = CaseSize > 63 ? UINT64_MAX : 1ULL << CaseSize;
5371  assert(MaxTableSize >= TableSize &&
5372  "It is impossible for a switch to have more entries than the max "
5373  "representable value of its input integer type's size.");
5374 
5375  // If the default destination is unreachable, or if the lookup table covers
5376  // all values of the conditional variable, branch directly to the lookup table
5377  // BB. Otherwise, check that the condition is within the case range.
5378  const bool DefaultIsReachable =
5379  !isa<UnreachableInst>(SI->getDefaultDest()->getFirstNonPHIOrDbg());
5380  const bool GeneratingCoveredLookupTable = (MaxTableSize == TableSize);
5381  BranchInst *RangeCheckBranch = nullptr;
5382 
5383  if (!DefaultIsReachable || GeneratingCoveredLookupTable) {
5384  Builder.CreateBr(LookupBB);
5385  // Note: We call removeProdecessor later since we need to be able to get the
5386  // PHI value for the default case in case we're using a bit mask.
5387  } else {
5388  Value *Cmp = Builder.CreateICmpULT(
5389  TableIndex, ConstantInt::get(MinCaseVal->getType(), TableSize));
5390  RangeCheckBranch =
5391  Builder.CreateCondBr(Cmp, LookupBB, SI->getDefaultDest());
5392  }
5393 
5394  // Populate the BB that does the lookups.
5395  Builder.SetInsertPoint(LookupBB);
5396 
5397  if (NeedMask) {
5398  // Before doing the lookup, we do the hole check. The LookupBB is therefore
5399  // re-purposed to do the hole check, and we create a new LookupBB.
5400  BasicBlock *MaskBB = LookupBB;
5401  MaskBB->setName("switch.hole_check");
5402  LookupBB = BasicBlock::Create(Mod.getContext(), "switch.lookup",
5403  CommonDest->getParent(), CommonDest);
5404 
5405  // Make the mask's bitwidth at least 8-bit and a power-of-2 to avoid
5406  // unnecessary illegal types.
5407  uint64_t TableSizePowOf2 = NextPowerOf2(std::max(7ULL, TableSize - 1ULL));
5408  APInt MaskInt(TableSizePowOf2, 0);
5409  APInt One(TableSizePowOf2, 1);
5410  // Build bitmask; fill in a 1 bit for every case.
5411  const ResultListTy &ResultList = ResultLists[PHIs[0]];
5412  for (size_t I = 0, E = ResultList.size(); I != E; ++I) {
5413  uint64_t Idx = (ResultList[I].first->getValue() - MinCaseVal->getValue())
5414  .getLimitedValue();
5415  MaskInt |= One << Idx;
5416  }
5417  ConstantInt *TableMask = ConstantInt::get(Mod.getContext(), MaskInt);
5418 
5419  // Get the TableIndex'th bit of the bitmask.
5420  // If this bit is 0 (meaning hole) jump to the default destination,
5421  // else continue with table lookup.
5422  IntegerType *MapTy = TableMask->getType();
5423  Value *MaskIndex =
5424  Builder.CreateZExtOrTrunc(TableIndex, MapTy, "switch.maskindex");
5425  Value *Shifted = Builder.CreateLShr(TableMask, MaskIndex, "switch.shifted");
5426  Value *LoBit = Builder.CreateTrunc(
5427  Shifted, Type::getInt1Ty(Mod.getContext()), "switch.lobit");
5428  Builder.CreateCondBr(LoBit, LookupBB, SI->getDefaultDest());
5429 
5430  Builder.SetInsertPoint(LookupBB);
5431  AddPredecessorToBlock(SI->getDefaultDest(), MaskBB, SI->getParent());
5432  }
5433 
5434  if (!DefaultIsReachable || GeneratingCoveredLookupTable) {
5435  // We cached PHINodes in PHIs. To avoid accessing deleted PHINodes later,
5436  // do not delete PHINodes here.
5438  /*DontDeleteUselessPHIs=*/true);
5439  }
5440 
5441  bool ReturnedEarly = false;
5442  for (PHINode *PHI : PHIs) {
5443  const ResultListTy &ResultList = ResultLists[PHI];
5444 
5445  // If using a bitmask, use any value to fill the lookup table holes.
5446  Constant *DV = NeedMask ? ResultLists[PHI][0].second : DefaultResults[PHI];
5447  StringRef FuncName = Fn->getName();
5448  SwitchLookupTable Table(Mod, TableSize, MinCaseVal, ResultList, DV, DL,
5449  FuncName);
5450 
5451  Value *Result = Table.BuildLookup(TableIndex, Builder);
5452 
5453  // If the result is used to return immediately from the function, we want to
5454  // do that right here.
5455  if (PHI->hasOneUse() && isa<ReturnInst>(*PHI->user_begin()) &&
5456  PHI->user_back() == CommonDest->getFirstNonPHIOrDbg()) {
5457  Builder.CreateRet(Result);
5458  ReturnedEarly = true;
5459  break;
5460  }
5461 
5462  // Do a small peephole optimization: re-use the switch table compare if
5463  // possible.
5464  if (!TableHasHoles && HasDefaultResults && RangeCheckBranch) {
5465  BasicBlock *PhiBlock = PHI->getParent();
5466  // Search for compare instructions which use the phi.
5467  for (auto *User : PHI->users()) {
5468  reuseTableCompare(User, PhiBlock, RangeCheckBranch, DV, ResultList);
5469  }
5470  }
5471 
5472  PHI->addIncoming(Result, LookupBB);
5473  }
5474 
5475  if (!ReturnedEarly)
5476  Builder.CreateBr(CommonDest);
5477 
5478  // Remove the switch.
5479  for (unsigned i = 0, e = SI->getNumSuccessors(); i < e; ++i) {
5480  BasicBlock *Succ = SI->getSuccessor(i);
5481 
5482  if (Succ == SI->getDefaultDest())
5483  continue;
5484  Succ->removePredecessor(SI->getParent());
5485  }
5486  SI->eraseFromParent();
5487 
5488  ++NumLookupTables;
5489  if (NeedMask)
5490  ++NumLookupTablesHoles;
5491  return true;
5492 }
5493 
5494 static bool isSwitchDense(ArrayRef<int64_t> Values) {
5495  // See also SelectionDAGBuilder::isDense(), which this function was based on.
5496  uint64_t Diff = (uint64_t)Values.back() - (uint64_t)Values.front();
5497  uint64_t Range = Diff + 1;
5498  uint64_t NumCases = Values.size();
5499  // 40% is the default density for building a jump table in optsize/minsize mode.
5500  uint64_t MinDensity = 40;
5501 
5502  return NumCases * 100 >= Range * MinDensity;
5503 }
5504 
5505 /// Try to transform a switch that has "holes" in it to a contiguous sequence
5506 /// of cases.
5507 ///
5508 /// A switch such as: switch(i) {case 5: case 9: case 13: case 17:} can be
5509 /// range-reduced to: switch ((i-5) / 4) {case 0: case 1: case 2: case 3:}.
5510 ///
5511 /// This converts a sparse switch into a dense switch which allows better
5512 /// lowering and could also allow transforming into a lookup table.
5514  const DataLayout &DL,
5515  const TargetTransformInfo &TTI) {
5516  auto *CondTy = cast<IntegerType>(SI->getCondition()->getType());
5517  if (CondTy->getIntegerBitWidth() > 64 ||
5518  !DL.fitsInLegalInteger(CondTy->getIntegerBitWidth()))
5519  return false;
5520  // Only bother with this optimization if there are more than 3 switch cases;
5521  // SDAG will only bother creating jump tables for 4 or more cases.
5522  if (SI->getNumCases() < 4)
5523  return false;
5524 
5525  // This transform is agnostic to the signedness of the input or case values. We
5526  // can treat the case values as signed or unsigned. We can optimize more common
5527  // cases such as a sequence crossing zero {-4,0,4,8} if we interpret case values
5528  // as signed.
5529  SmallVector<int64_t,4> Values;
5530  for (auto &C : SI->cases())
5531  Values.push_back(C.getCaseValue()->getValue().getSExtValue());
5532  llvm::sort(Values);
5533 
5534  // If the switch is already dense, there's nothing useful to do here.
5535  if (isSwitchDense(Values))
5536  return false;
5537 
5538  // First, transform the values such that they start at zero and ascend.
5539  int64_t Base = Values[0];
5540  for (auto &V : Values)
5541  V -= (uint64_t)(Base);
5542 
5543  // Now we have signed numbers that have been shifted so that, given enough
5544  // precision, there are no negative values. Since the rest of the transform
5545  // is bitwise only, we switch now to an unsigned representation.
5546  uint64_t GCD = 0;
5547  for (auto &V : Values)
5548  GCD = GreatestCommonDivisor64(GCD, (uint64_t)V);
5549 
5550  // This transform can be done speculatively because it is so cheap - it results
5551  // in a single rotate operation being inserted. This can only happen if the
5552  // factor extracted is a power of 2.
5553  // FIXME: If the GCD is an odd number we can multiply by the multiplicative
5554  // inverse of GCD and then perform this transform.
5555  // FIXME: It's possible that optimizing a switch on powers of two might also
5556  // be beneficial - flag values are often powers of two and we could use a CLZ
5557  // as the key function.
5558  if (GCD <= 1 || !isPowerOf2_64(GCD))
5559  // No common divisor found or too expensive to compute key function.
5560  return false;
5561 
5562  unsigned Shift = Log2_64(GCD);
5563  for (auto &V : Values)
5564  V = (int64_t)((uint64_t)V >> Shift);
5565 
5566  if (!isSwitchDense(Values))
5567  // Transform didn't create a dense switch.
5568  return false;
5569 
5570  // The obvious transform is to shift the switch condition right and emit a
5571  // check that the condition actually cleanly divided by GCD, i.e.
5572  // C & (1 << Shift - 1) == 0
5573  // inserting a new CFG edge to handle the case where it didn't divide cleanly.
5574  //
5575  // A cheaper way of doing this is a simple ROTR(C, Shift). This performs the
5576  // shift and puts the shifted-off bits in the uppermost bits. If any of these
5577  // are nonzero then the switch condition will be very large and will hit the
5578  // default case.
5579 
5580  auto *Ty = cast<IntegerType>(SI->getCondition()->getType());
5581  Builder.SetInsertPoint(SI);
5582  auto *ShiftC = ConstantInt::get(Ty, Shift);
5583  auto *Sub = Builder.CreateSub(SI->getCondition(), ConstantInt::get(Ty, Base));
5584  auto *LShr = Builder.CreateLShr(Sub, ShiftC);
5585  auto *Shl = Builder.CreateShl(Sub, Ty->getBitWidth() - Shift);
5586  auto *Rot = Builder.CreateOr(LShr, Shl);
5587  SI->replaceUsesOfWith(SI->getCondition(), Rot);
5588 
5589  for (auto Case : SI->cases()) {
5590  auto *Orig = Case.getCaseValue();
5591  auto Sub = Orig->getValue() - APInt(Ty->getBitWidth(), Base);
5592  Case.setValue(
5593  cast<ConstantInt>(ConstantInt::get(Ty, Sub.lshr(ShiftC->getValue()))));
5594  }
5595  return true;
5596 }