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(TerminatorInst *TI);
179  BasicBlock *GetValueEqualityComparisonCases(
180  TerminatorInst *TI, std::vector<ValueEqualityComparisonCase> &Cases);
181  bool SimplifyEqualityComparisonWithOnlyPredecessor(TerminatorInst *TI,
182  BasicBlock *Pred,
183  IRBuilder<> &Builder);
184  bool FoldValueComparisonIntoPredecessors(TerminatorInst *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(TerminatorInst *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->getNumSuccessors() * pred_size(SI->getParent()) <= 128)
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  TerminatorInst *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  TerminatorInst *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 TerminatorInst, 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(TerminatorInst *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  TerminatorInst *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  // Okay, it is safe to hoist the terminator.
1376  Instruction *NT = I1->clone();
1377  BIParent->getInstList().insert(BI->getIterator(), NT);
1378  if (!NT->getType()->isVoidTy()) {
1379  I1->replaceAllUsesWith(NT);
1380  I2->replaceAllUsesWith(NT);
1381  NT->takeName(I1);
1382  }
1383 
1384  IRBuilder<NoFolder> Builder(NT);
1385  // Hoisting one of the terminators from our successor is a great thing.
1386  // Unfortunately, the successors of the if/else blocks may have PHI nodes in
1387  // them. If they do, all PHI entries for BB1/BB2 must agree for all PHI
1388  // nodes, so we insert select instruction to compute the final result.
1389  std::map<std::pair<Value *, Value *>, SelectInst *> InsertedSelects;
1390  for (BasicBlock *Succ : successors(BB1)) {
1391  for (PHINode &PN : Succ->phis()) {
1392  Value *BB1V = PN.getIncomingValueForBlock(BB1);
1393  Value *BB2V = PN.getIncomingValueForBlock(BB2);
1394  if (BB1V == BB2V)
1395  continue;
1396 
1397  // These values do not agree. Insert a select instruction before NT
1398  // that determines the right value.
1399  SelectInst *&SI = InsertedSelects[std::make_pair(BB1V, BB2V)];
1400  if (!SI)
1401  SI = cast<SelectInst>(
1402  Builder.CreateSelect(BI->getCondition(), BB1V, BB2V,
1403  BB1V->getName() + "." + BB2V->getName(), BI));
1404 
1405  // Make the PHI node use the select for all incoming values for BB1/BB2
1406  for (unsigned i = 0, e = PN.getNumIncomingValues(); i != e; ++i)
1407  if (PN.getIncomingBlock(i) == BB1 || PN.getIncomingBlock(i) == BB2)
1408  PN.setIncomingValue(i, SI);
1409  }
1410  }
1411 
1412  // Update any PHI nodes in our new successors.
1413  for (BasicBlock *Succ : successors(BB1))
1414  AddPredecessorToBlock(Succ, BIParent, BB1);
1415 
1417  return true;
1418 }
1419 
1420 // All instructions in Insts belong to different blocks that all unconditionally
1421 // branch to a common successor. Analyze each instruction and return true if it
1422 // would be possible to sink them into their successor, creating one common
1423 // instruction instead. For every value that would be required to be provided by
1424 // PHI node (because an operand varies in each input block), add to PHIOperands.
1427  DenseMap<Instruction *, SmallVector<Value *, 4>> &PHIOperands) {
1428  // Prune out obviously bad instructions to move. Any non-store instruction
1429  // must have exactly one use, and we check later that use is by a single,
1430  // common PHI instruction in the successor.
1431  for (auto *I : Insts) {
1432  // These instructions may change or break semantics if moved.
1433  if (isa<PHINode>(I) || I->isEHPad() || isa<AllocaInst>(I) ||
1434  I->getType()->isTokenTy())
1435  return false;
1436 
1437  // Conservatively return false if I is an inline-asm instruction. Sinking
1438  // and merging inline-asm instructions can potentially create arguments
1439  // that cannot satisfy the inline-asm constraints.
1440  if (const auto *C = dyn_cast<CallInst>(I))
1441  if (C->isInlineAsm())
1442  return false;
1443 
1444  // Everything must have only one use too, apart from stores which
1445  // have no uses.
1446  if (!isa<StoreInst>(I) && !I->hasOneUse())
1447  return false;
1448  }
1449 
1450  const Instruction *I0 = Insts.front();
1451  for (auto *I : Insts)
1452  if (!I->isSameOperationAs(I0))
1453  return false;
1454 
1455  // All instructions in Insts are known to be the same opcode. If they aren't
1456  // stores, check the only user of each is a PHI or in the same block as the
1457  // instruction, because if a user is in the same block as an instruction
1458  // we're contemplating sinking, it must already be determined to be sinkable.
1459  if (!isa<StoreInst>(I0)) {
1460  auto *PNUse = dyn_cast<PHINode>(*I0->user_begin());
1461  auto *Succ = I0->getParent()->getTerminator()->getSuccessor(0);
1462  if (!all_of(Insts, [&PNUse,&Succ](const Instruction *I) -> bool {
1463  auto *U = cast<Instruction>(*I->user_begin());
1464  return (PNUse &&
1465  PNUse->getParent() == Succ &&
1466  PNUse->getIncomingValueForBlock(I->getParent()) == I) ||
1467  U->getParent() == I->getParent();
1468  }))
1469  return false;
1470  }
1471 
1472  // Because SROA can't handle speculating stores of selects, try not
1473  // to sink loads or stores of allocas when we'd have to create a PHI for
1474  // the address operand. Also, because it is likely that loads or stores
1475  // of allocas will disappear when Mem2Reg/SROA is run, don't sink them.
1476  // This can cause code churn which can have unintended consequences down
1477  // the line - see https://llvm.org/bugs/show_bug.cgi?id=30244.
1478  // FIXME: This is a workaround for a deficiency in SROA - see
1479  // https://llvm.org/bugs/show_bug.cgi?id=30188
1480  if (isa<StoreInst>(I0) && any_of(Insts, [](const Instruction *I) {
1481  return isa<AllocaInst>(I->getOperand(1));
1482  }))
1483  return false;
1484  if (isa<LoadInst>(I0) && any_of(Insts, [](const Instruction *I) {
1485  return isa<AllocaInst>(I->getOperand(0));
1486  }))
1487  return false;
1488 
1489  for (unsigned OI = 0, OE = I0->getNumOperands(); OI != OE; ++OI) {
1490  if (I0->getOperand(OI)->getType()->isTokenTy())
1491  // Don't touch any operand of token type.
1492  return false;
1493 
1494  auto SameAsI0 = [&I0, OI](const Instruction *I) {
1495  assert(I->getNumOperands() == I0->getNumOperands());
1496  return I->getOperand(OI) == I0->getOperand(OI);
1497  };
1498  if (!all_of(Insts, SameAsI0)) {
1499  if (!canReplaceOperandWithVariable(I0, OI))
1500  // We can't create a PHI from this GEP.
1501  return false;
1502  // Don't create indirect calls! The called value is the final operand.
1503  if ((isa<CallInst>(I0) || isa<InvokeInst>(I0)) && OI == OE - 1) {
1504  // FIXME: if the call was *already* indirect, we should do this.
1505  return false;
1506  }
1507  for (auto *I : Insts)
1508  PHIOperands[I].push_back(I->getOperand(OI));
1509  }
1510  }
1511  return true;
1512 }
1513 
1514 // Assuming canSinkLastInstruction(Blocks) has returned true, sink the last
1515 // instruction of every block in Blocks to their common successor, commoning
1516 // into one instruction.
1518  auto *BBEnd = Blocks[0]->getTerminator()->getSuccessor(0);
1519 
1520  // canSinkLastInstruction returning true guarantees that every block has at
1521  // least one non-terminator instruction.
1523  for (auto *BB : Blocks) {
1524  Instruction *I = BB->getTerminator();
1525  do {
1526  I = I->getPrevNode();
1527  } while (isa<DbgInfoIntrinsic>(I) && I != &BB->front());
1528  if (!isa<DbgInfoIntrinsic>(I))
1529  Insts.push_back(I);
1530  }
1531 
1532  // The only checking we need to do now is that all users of all instructions
1533  // are the same PHI node. canSinkLastInstruction should have checked this but
1534  // it is slightly over-aggressive - it gets confused by commutative instructions
1535  // so double-check it here.
1536  Instruction *I0 = Insts.front();
1537  if (!isa<StoreInst>(I0)) {
1538  auto *PNUse = dyn_cast<PHINode>(*I0->user_begin());
1539  if (!all_of(Insts, [&PNUse](const Instruction *I) -> bool {
1540  auto *U = cast<Instruction>(*I->user_begin());
1541  return U == PNUse;
1542  }))
1543  return false;
1544  }
1545 
1546  // We don't need to do any more checking here; canSinkLastInstruction should
1547  // have done it all for us.
1548  SmallVector<Value*, 4> NewOperands;
1549  for (unsigned O = 0, E = I0->getNumOperands(); O != E; ++O) {
1550  // This check is different to that in canSinkLastInstruction. There, we
1551  // cared about the global view once simplifycfg (and instcombine) have
1552  // completed - it takes into account PHIs that become trivially
1553  // simplifiable. However here we need a more local view; if an operand
1554  // differs we create a PHI and rely on instcombine to clean up the very
1555  // small mess we may make.
1556  bool NeedPHI = any_of(Insts, [&I0, O](const Instruction *I) {
1557  return I->getOperand(O) != I0->getOperand(O);
1558  });
1559  if (!NeedPHI) {
1560  NewOperands.push_back(I0->getOperand(O));
1561  continue;
1562  }
1563 
1564  // Create a new PHI in the successor block and populate it.
1565  auto *Op = I0->getOperand(O);
1566  assert(!Op->getType()->isTokenTy() && "Can't PHI tokens!");
1567  auto *PN = PHINode::Create(Op->getType(), Insts.size(),
1568  Op->getName() + ".sink", &BBEnd->front());
1569  for (auto *I : Insts)
1570  PN->addIncoming(I->getOperand(O), I->getParent());
1571  NewOperands.push_back(PN);
1572  }
1573 
1574  // Arbitrarily use I0 as the new "common" instruction; remap its operands
1575  // and move it to the start of the successor block.
1576  for (unsigned O = 0, E = I0->getNumOperands(); O != E; ++O)
1577  I0->getOperandUse(O).set(NewOperands[O]);
1578  I0->moveBefore(&*BBEnd->getFirstInsertionPt());
1579 
1580  // Update metadata and IR flags, and merge debug locations.
1581  for (auto *I : Insts)
1582  if (I != I0) {
1583  // The debug location for the "common" instruction is the merged locations
1584  // of all the commoned instructions. We start with the original location
1585  // of the "common" instruction and iteratively merge each location in the
1586  // loop below.
1587  // This is an N-way merge, which will be inefficient if I0 is a CallInst.
1588  // However, as N-way merge for CallInst is rare, so we use simplified API
1589  // instead of using complex API for N-way merge.
1590  I0->applyMergedLocation(I0->getDebugLoc(), I->getDebugLoc());
1591  combineMetadataForCSE(I0, I, true);
1592  I0->andIRFlags(I);
1593  }
1594 
1595  if (!isa<StoreInst>(I0)) {
1596  // canSinkLastInstruction checked that all instructions were used by
1597  // one and only one PHI node. Find that now, RAUW it to our common
1598  // instruction and nuke it.
1599  assert(I0->hasOneUse());
1600  auto *PN = cast<PHINode>(*I0->user_begin());
1601  PN->replaceAllUsesWith(I0);
1602  PN->eraseFromParent();
1603  }
1604 
1605  // Finally nuke all instructions apart from the common instruction.
1606  for (auto *I : Insts)
1607  if (I != I0)
1608  I->eraseFromParent();
1609 
1610  return true;
1611 }
1612 
1613 namespace {
1614 
1615  // LockstepReverseIterator - Iterates through instructions
1616  // in a set of blocks in reverse order from the first non-terminator.
1617  // For example (assume all blocks have size n):
1618  // LockstepReverseIterator I([B1, B2, B3]);
1619  // *I-- = [B1[n], B2[n], B3[n]];
1620  // *I-- = [B1[n-1], B2[n-1], B3[n-1]];
1621  // *I-- = [B1[n-2], B2[n-2], B3[n-2]];
1622  // ...
1623  class LockstepReverseIterator {
1624  ArrayRef<BasicBlock*> Blocks;
1626  bool Fail;
1627 
1628  public:
1629  LockstepReverseIterator(ArrayRef<BasicBlock*> Blocks) : Blocks(Blocks) {
1630  reset();
1631  }
1632 
1633  void reset() {
1634  Fail = false;
1635  Insts.clear();
1636  for (auto *BB : Blocks) {
1637  Instruction *Inst = BB->getTerminator();
1638  for (Inst = Inst->getPrevNode(); Inst && isa<DbgInfoIntrinsic>(Inst);)
1639  Inst = Inst->getPrevNode();
1640  if (!Inst) {
1641  // Block wasn't big enough.
1642  Fail = true;
1643  return;
1644  }
1645  Insts.push_back(Inst);
1646  }
1647  }
1648 
1649  bool isValid() const {
1650  return !Fail;
1651  }
1652 
1653  void operator--() {
1654  if (Fail)
1655  return;
1656  for (auto *&Inst : Insts) {
1657  for (Inst = Inst->getPrevNode(); Inst && isa<DbgInfoIntrinsic>(Inst);)
1658  Inst = Inst->getPrevNode();
1659  // Already at beginning of block.
1660  if (!Inst) {
1661  Fail = true;
1662  return;
1663  }
1664  }
1665  }
1666 
1668  return Insts;
1669  }
1670  };
1671 
1672 } // end anonymous namespace
1673 
1674 /// Check whether BB's predecessors end with unconditional branches. If it is
1675 /// true, sink any common code from the predecessors to BB.
1676 /// We also allow one predecessor to end with conditional branch (but no more
1677 /// than one).
1679  // We support two situations:
1680  // (1) all incoming arcs are unconditional
1681  // (2) one incoming arc is conditional
1682  //
1683  // (2) is very common in switch defaults and
1684  // else-if patterns;
1685  //
1686  // if (a) f(1);
1687  // else if (b) f(2);
1688  //
1689  // produces:
1690  //
1691  // [if]
1692  // / \
1693  // [f(1)] [if]
1694  // | | \
1695  // | | |
1696  // | [f(2)]|
1697  // \ | /
1698  // [ end ]
1699  //
1700  // [end] has two unconditional predecessor arcs and one conditional. The
1701  // conditional refers to the implicit empty 'else' arc. This conditional
1702  // arc can also be caused by an empty default block in a switch.
1703  //
1704  // In this case, we attempt to sink code from all *unconditional* arcs.
1705  // If we can sink instructions from these arcs (determined during the scan
1706  // phase below) we insert a common successor for all unconditional arcs and
1707  // connect that to [end], to enable sinking:
1708  //
1709  // [if]
1710  // / \
1711  // [x(1)] [if]
1712  // | | \
1713  // | | \
1714  // | [x(2)] |
1715  // \ / |
1716  // [sink.split] |
1717  // \ /
1718  // [ end ]
1719  //
1720  SmallVector<BasicBlock*,4> UnconditionalPreds;
1721  Instruction *Cond = nullptr;
1722  for (auto *B : predecessors(BB)) {
1723  auto *T = B->getTerminator();
1724  if (isa<BranchInst>(T) && cast<BranchInst>(T)->isUnconditional())
1725  UnconditionalPreds.push_back(B);
1726  else if ((isa<BranchInst>(T) || isa<SwitchInst>(T)) && !Cond)
1727  Cond = T;
1728  else
1729  return false;
1730  }
1731  if (UnconditionalPreds.size() < 2)
1732  return false;
1733 
1734  bool Changed = false;
1735  // We take a two-step approach to tail sinking. First we scan from the end of
1736  // each block upwards in lockstep. If the n'th instruction from the end of each
1737  // block can be sunk, those instructions are added to ValuesToSink and we
1738  // carry on. If we can sink an instruction but need to PHI-merge some operands
1739  // (because they're not identical in each instruction) we add these to
1740  // PHIOperands.
1741  unsigned ScanIdx = 0;
1742  SmallPtrSet<Value*,4> InstructionsToSink;
1744  LockstepReverseIterator LRI(UnconditionalPreds);
1745  while (LRI.isValid() &&
1746  canSinkInstructions(*LRI, PHIOperands)) {
1747  LLVM_DEBUG(dbgs() << "SINK: instruction can be sunk: " << *(*LRI)[0]
1748  << "\n");
1749  InstructionsToSink.insert((*LRI).begin(), (*LRI).end());
1750  ++ScanIdx;
1751  --LRI;
1752  }
1753 
1754  auto ProfitableToSinkInstruction = [&](LockstepReverseIterator &LRI) {
1755  unsigned NumPHIdValues = 0;
1756  for (auto *I : *LRI)
1757  for (auto *V : PHIOperands[I])
1758  if (InstructionsToSink.count(V) == 0)
1759  ++NumPHIdValues;
1760  LLVM_DEBUG(dbgs() << "SINK: #phid values: " << NumPHIdValues << "\n");
1761  unsigned NumPHIInsts = NumPHIdValues / UnconditionalPreds.size();
1762  if ((NumPHIdValues % UnconditionalPreds.size()) != 0)
1763  NumPHIInsts++;
1764 
1765  return NumPHIInsts <= 1;
1766  };
1767 
1768  if (ScanIdx > 0 && Cond) {
1769  // Check if we would actually sink anything first! This mutates the CFG and
1770  // adds an extra block. The goal in doing this is to allow instructions that
1771  // couldn't be sunk before to be sunk - obviously, speculatable instructions
1772  // (such as trunc, add) can be sunk and predicated already. So we check that
1773  // we're going to sink at least one non-speculatable instruction.
1774  LRI.reset();
1775  unsigned Idx = 0;
1776  bool Profitable = false;
1777  while (ProfitableToSinkInstruction(LRI) && Idx < ScanIdx) {
1778  if (!isSafeToSpeculativelyExecute((*LRI)[0])) {
1779  Profitable = true;
1780  break;
1781  }
1782  --LRI;
1783  ++Idx;
1784  }
1785  if (!Profitable)
1786  return false;
1787 
1788  LLVM_DEBUG(dbgs() << "SINK: Splitting edge\n");
1789  // We have a conditional edge and we're going to sink some instructions.
1790  // Insert a new block postdominating all blocks we're going to sink from.
1791  if (!SplitBlockPredecessors(BB, UnconditionalPreds, ".sink.split"))
1792  // Edges couldn't be split.
1793  return false;
1794  Changed = true;
1795  }
1796 
1797  // Now that we've analyzed all potential sinking candidates, perform the
1798  // actual sink. We iteratively sink the last non-terminator of the source
1799  // blocks into their common successor unless doing so would require too
1800  // many PHI instructions to be generated (currently only one PHI is allowed
1801  // per sunk instruction).
1802  //
1803  // We can use InstructionsToSink to discount values needing PHI-merging that will
1804  // actually be sunk in a later iteration. This allows us to be more
1805  // aggressive in what we sink. This does allow a false positive where we
1806  // sink presuming a later value will also be sunk, but stop half way through
1807  // and never actually sink it which means we produce more PHIs than intended.
1808  // This is unlikely in practice though.
1809  for (unsigned SinkIdx = 0; SinkIdx != ScanIdx; ++SinkIdx) {
1810  LLVM_DEBUG(dbgs() << "SINK: Sink: "
1811  << *UnconditionalPreds[0]->getTerminator()->getPrevNode()
1812  << "\n");
1813 
1814  // Because we've sunk every instruction in turn, the current instruction to
1815  // sink is always at index 0.
1816  LRI.reset();
1817  if (!ProfitableToSinkInstruction(LRI)) {
1818  // Too many PHIs would be created.
1819  LLVM_DEBUG(
1820  dbgs() << "SINK: stopping here, too many PHIs would be created!\n");
1821  break;
1822  }
1823 
1824  if (!sinkLastInstruction(UnconditionalPreds))
1825  return Changed;
1826  NumSinkCommons++;
1827  Changed = true;
1828  }
1829  return Changed;
1830 }
1831 
1832 /// Determine if we can hoist sink a sole store instruction out of a
1833 /// conditional block.
1834 ///
1835 /// We are looking for code like the following:
1836 /// BrBB:
1837 /// store i32 %add, i32* %arrayidx2
1838 /// ... // No other stores or function calls (we could be calling a memory
1839 /// ... // function).
1840 /// %cmp = icmp ult %x, %y
1841 /// br i1 %cmp, label %EndBB, label %ThenBB
1842 /// ThenBB:
1843 /// store i32 %add5, i32* %arrayidx2
1844 /// br label EndBB
1845 /// EndBB:
1846 /// ...
1847 /// We are going to transform this into:
1848 /// BrBB:
1849 /// store i32 %add, i32* %arrayidx2
1850 /// ... //
1851 /// %cmp = icmp ult %x, %y
1852 /// %add.add5 = select i1 %cmp, i32 %add, %add5
1853 /// store i32 %add.add5, i32* %arrayidx2
1854 /// ...
1855 ///
1856 /// \return The pointer to the value of the previous store if the store can be
1857 /// hoisted into the predecessor block. 0 otherwise.
1859  BasicBlock *StoreBB, BasicBlock *EndBB) {
1860  StoreInst *StoreToHoist = dyn_cast<StoreInst>(I);
1861  if (!StoreToHoist)
1862  return nullptr;
1863 
1864  // Volatile or atomic.
1865  if (!StoreToHoist->isSimple())
1866  return nullptr;
1867 
1868  Value *StorePtr = StoreToHoist->getPointerOperand();
1869 
1870  // Look for a store to the same pointer in BrBB.
1871  unsigned MaxNumInstToLookAt = 9;
1872  for (Instruction &CurI : reverse(BrBB->instructionsWithoutDebug())) {
1873  if (!MaxNumInstToLookAt)
1874  break;
1875  --MaxNumInstToLookAt;
1876 
1877  // Could be calling an instruction that affects memory like free().
1878  if (CurI.mayHaveSideEffects() && !isa<StoreInst>(CurI))
1879  return nullptr;
1880 
1881  if (auto *SI = dyn_cast<StoreInst>(&CurI)) {
1882  // Found the previous store make sure it stores to the same location.
1883  if (SI->getPointerOperand() == StorePtr)
1884  // Found the previous store, return its value operand.
1885  return SI->getValueOperand();
1886  return nullptr; // Unknown store.
1887  }
1888  }
1889 
1890  return nullptr;
1891 }
1892 
1893 /// Speculate a conditional basic block flattening the CFG.
1894 ///
1895 /// Note that this is a very risky transform currently. Speculating
1896 /// instructions like this is most often not desirable. Instead, there is an MI
1897 /// pass which can do it with full awareness of the resource constraints.
1898 /// However, some cases are "obvious" and we should do directly. An example of
1899 /// this is speculating a single, reasonably cheap instruction.
1900 ///
1901 /// There is only one distinct advantage to flattening the CFG at the IR level:
1902 /// it makes very common but simplistic optimizations such as are common in
1903 /// instcombine and the DAG combiner more powerful by removing CFG edges and
1904 /// modeling their effects with easier to reason about SSA value graphs.
1905 ///
1906 ///
1907 /// An illustration of this transform is turning this IR:
1908 /// \code
1909 /// BB:
1910 /// %cmp = icmp ult %x, %y
1911 /// br i1 %cmp, label %EndBB, label %ThenBB
1912 /// ThenBB:
1913 /// %sub = sub %x, %y
1914 /// br label BB2
1915 /// EndBB:
1916 /// %phi = phi [ %sub, %ThenBB ], [ 0, %EndBB ]
1917 /// ...
1918 /// \endcode
1919 ///
1920 /// Into this IR:
1921 /// \code
1922 /// BB:
1923 /// %cmp = icmp ult %x, %y
1924 /// %sub = sub %x, %y
1925 /// %cond = select i1 %cmp, 0, %sub
1926 /// ...
1927 /// \endcode
1928 ///
1929 /// \returns true if the conditional block is removed.
1931  const TargetTransformInfo &TTI) {
1932  // Be conservative for now. FP select instruction can often be expensive.
1933  Value *BrCond = BI->getCondition();
1934  if (isa<FCmpInst>(BrCond))
1935  return false;
1936 
1937  BasicBlock *BB = BI->getParent();
1938  BasicBlock *EndBB = ThenBB->getTerminator()->getSuccessor(0);
1939 
1940  // If ThenBB is actually on the false edge of the conditional branch, remember
1941  // to swap the select operands later.
1942  bool Invert = false;
1943  if (ThenBB != BI->getSuccessor(0)) {
1944  assert(ThenBB == BI->getSuccessor(1) && "No edge from 'if' block?");
1945  Invert = true;
1946  }
1947  assert(EndBB == BI->getSuccessor(!Invert) && "No edge from to end block");
1948 
1949  // Keep a count of how many times instructions are used within ThenBB when
1950  // they are candidates for sinking into ThenBB. Specifically:
1951  // - They are defined in BB, and
1952  // - They have no side effects, and
1953  // - All of their uses are in ThenBB.
1954  SmallDenseMap<Instruction *, unsigned, 4> SinkCandidateUseCounts;
1955 
1956  SmallVector<Instruction *, 4> SpeculatedDbgIntrinsics;
1957 
1958  unsigned SpeculationCost = 0;
1959  Value *SpeculatedStoreValue = nullptr;
1960  StoreInst *SpeculatedStore = nullptr;
1961  for (BasicBlock::iterator BBI = ThenBB->begin(),
1962  BBE = std::prev(ThenBB->end());
1963  BBI != BBE; ++BBI) {
1964  Instruction *I = &*BBI;
1965  // Skip debug info.
1966  if (isa<DbgInfoIntrinsic>(I)) {
1967  SpeculatedDbgIntrinsics.push_back(I);
1968  continue;
1969  }
1970 
1971  // Only speculatively execute a single instruction (not counting the
1972  // terminator) for now.
1973  ++SpeculationCost;
1974  if (SpeculationCost > 1)
1975  return false;
1976 
1977  // Don't hoist the instruction if it's unsafe or expensive.
1978  if (!isSafeToSpeculativelyExecute(I) &&
1979  !(HoistCondStores && (SpeculatedStoreValue = isSafeToSpeculateStore(
1980  I, BB, ThenBB, EndBB))))
1981  return false;
1982  if (!SpeculatedStoreValue &&
1983  ComputeSpeculationCost(I, TTI) >
1985  return false;
1986 
1987  // Store the store speculation candidate.
1988  if (SpeculatedStoreValue)
1989  SpeculatedStore = cast<StoreInst>(I);
1990 
1991  // Do not hoist the instruction if any of its operands are defined but not
1992  // used in BB. The transformation will prevent the operand from
1993  // being sunk into the use block.
1994  for (User::op_iterator i = I->op_begin(), e = I->op_end(); i != e; ++i) {
1995  Instruction *OpI = dyn_cast<Instruction>(*i);
1996  if (!OpI || OpI->getParent() != BB || OpI->mayHaveSideEffects())
1997  continue; // Not a candidate for sinking.
1998 
1999  ++SinkCandidateUseCounts[OpI];
2000  }
2001  }
2002 
2003  // Consider any sink candidates which are only used in ThenBB as costs for
2004  // speculation. Note, while we iterate over a DenseMap here, we are summing
2005  // and so iteration order isn't significant.
2007  I = SinkCandidateUseCounts.begin(),
2008  E = SinkCandidateUseCounts.end();
2009  I != E; ++I)
2010  if (I->first->hasNUses(I->second)) {
2011  ++SpeculationCost;
2012  if (SpeculationCost > 1)
2013  return false;
2014  }
2015 
2016  // Check that the PHI nodes can be converted to selects.
2017  bool HaveRewritablePHIs = false;
2018  for (PHINode &PN : EndBB->phis()) {
2019  Value *OrigV = PN.getIncomingValueForBlock(BB);
2020  Value *ThenV = PN.getIncomingValueForBlock(ThenBB);
2021 
2022  // FIXME: Try to remove some of the duplication with HoistThenElseCodeToIf.
2023  // Skip PHIs which are trivial.
2024  if (ThenV == OrigV)
2025  continue;
2026 
2027  // Don't convert to selects if we could remove undefined behavior instead.
2028  if (passingValueIsAlwaysUndefined(OrigV, &PN) ||
2029  passingValueIsAlwaysUndefined(ThenV, &PN))
2030  return false;
2031 
2032  HaveRewritablePHIs = true;
2033  ConstantExpr *OrigCE = dyn_cast<ConstantExpr>(OrigV);
2034  ConstantExpr *ThenCE = dyn_cast<ConstantExpr>(ThenV);
2035  if (!OrigCE && !ThenCE)
2036  continue; // Known safe and cheap.
2037 
2038  if ((ThenCE && !isSafeToSpeculativelyExecute(ThenCE)) ||
2039  (OrigCE && !isSafeToSpeculativelyExecute(OrigCE)))
2040  return false;
2041  unsigned OrigCost = OrigCE ? ComputeSpeculationCost(OrigCE, TTI) : 0;
2042  unsigned ThenCost = ThenCE ? ComputeSpeculationCost(ThenCE, TTI) : 0;
2043  unsigned MaxCost =
2045  if (OrigCost + ThenCost > MaxCost)
2046  return false;
2047 
2048  // Account for the cost of an unfolded ConstantExpr which could end up
2049  // getting expanded into Instructions.
2050  // FIXME: This doesn't account for how many operations are combined in the
2051  // constant expression.
2052  ++SpeculationCost;
2053  if (SpeculationCost > 1)
2054  return false;
2055  }
2056 
2057  // If there are no PHIs to process, bail early. This helps ensure idempotence
2058  // as well.
2059  if (!HaveRewritablePHIs && !(HoistCondStores && SpeculatedStoreValue))
2060  return false;
2061 
2062  // If we get here, we can hoist the instruction and if-convert.
2063  LLVM_DEBUG(dbgs() << "SPECULATIVELY EXECUTING BB" << *ThenBB << "\n";);
2064 
2065  // Insert a select of the value of the speculated store.
2066  if (SpeculatedStoreValue) {
2067  IRBuilder<NoFolder> Builder(BI);
2068  Value *TrueV = SpeculatedStore->getValueOperand();
2069  Value *FalseV = SpeculatedStoreValue;
2070  if (Invert)
2071  std::swap(TrueV, FalseV);
2072  Value *S = Builder.CreateSelect(
2073  BrCond, TrueV, FalseV, "spec.store.select", BI);
2074  SpeculatedStore->setOperand(0, S);
2075  SpeculatedStore->applyMergedLocation(BI->getDebugLoc(),
2076  SpeculatedStore->getDebugLoc());
2077  }
2078 
2079  // Metadata can be dependent on the condition we are hoisting above.
2080  // Conservatively strip all metadata on the instruction.
2081  for (auto &I : *ThenBB)
2083 
2084  // Hoist the instructions.
2085  BB->getInstList().splice(BI->getIterator(), ThenBB->getInstList(),
2086  ThenBB->begin(), std::prev(ThenBB->end()));
2087 
2088  // Insert selects and rewrite the PHI operands.
2089  IRBuilder<NoFolder> Builder(BI);
2090  for (PHINode &PN : EndBB->phis()) {
2091  unsigned OrigI = PN.getBasicBlockIndex(BB);
2092  unsigned ThenI = PN.getBasicBlockIndex(ThenBB);
2093  Value *OrigV = PN.getIncomingValue(OrigI);
2094  Value *ThenV = PN.getIncomingValue(ThenI);
2095 
2096  // Skip PHIs which are trivial.
2097  if (OrigV == ThenV)
2098  continue;
2099 
2100  // Create a select whose true value is the speculatively executed value and
2101  // false value is the preexisting value. Swap them if the branch
2102  // destinations were inverted.
2103  Value *TrueV = ThenV, *FalseV = OrigV;
2104  if (Invert)
2105  std::swap(TrueV, FalseV);
2106  Value *V = Builder.CreateSelect(
2107  BrCond, TrueV, FalseV, "spec.select", BI);
2108  PN.setIncomingValue(OrigI, V);
2109  PN.setIncomingValue(ThenI, V);
2110  }
2111 
2112  // Remove speculated dbg intrinsics.
2113  // FIXME: Is it possible to do this in a more elegant way? Moving/merging the
2114  // dbg value for the different flows and inserting it after the select.
2115  for (Instruction *I : SpeculatedDbgIntrinsics)
2116  I->eraseFromParent();
2117 
2118  ++NumSpeculations;
2119  return true;
2120 }
2121 
2122 /// Return true if we can thread a branch across this block.
2124  unsigned Size = 0;
2125 
2126  for (Instruction &I : BB->instructionsWithoutDebug()) {
2127  if (Size > 10)
2128  return false; // Don't clone large BB's.
2129  ++Size;
2130 
2131  // We can only support instructions that do not define values that are
2132  // live outside of the current basic block.
2133  for (User *U : I.users()) {
2134  Instruction *UI = cast<Instruction>(U);
2135  if (UI->getParent() != BB || isa<PHINode>(UI))
2136  return false;
2137  }
2138 
2139  // Looks ok, continue checking.
2140  }
2141 
2142  return true;
2143 }
2144 
2145 /// If we have a conditional branch on a PHI node value that is defined in the
2146 /// same block as the branch and if any PHI entries are constants, thread edges
2147 /// corresponding to that entry to be branches to their ultimate destination.
2148 static bool FoldCondBranchOnPHI(BranchInst *BI, const DataLayout &DL,
2149  AssumptionCache *AC) {
2150  BasicBlock *BB = BI->getParent();
2151  PHINode *PN = dyn_cast<PHINode>(BI->getCondition());
2152  // NOTE: we currently cannot transform this case if the PHI node is used
2153  // outside of the block.
2154  if (!PN || PN->getParent() != BB || !PN->hasOneUse())
2155  return false;
2156 
2157  // Degenerate case of a single entry PHI.
2158  if (PN->getNumIncomingValues() == 1) {
2160  return true;
2161  }
2162 
2163  // Now we know that this block has multiple preds and two succs.
2165  return false;
2166 
2167  // Can't fold blocks that contain noduplicate or convergent calls.
2168  if (any_of(*BB, [](const Instruction &I) {
2169  const CallInst *CI = dyn_cast<CallInst>(&I);
2170  return CI && (CI->cannotDuplicate() || CI->isConvergent());
2171  }))
2172  return false;
2173 
2174  // Okay, this is a simple enough basic block. See if any phi values are
2175  // constants.
2176  for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
2178  if (!CB || !CB->getType()->isIntegerTy(1))
2179  continue;
2180 
2181  // Okay, we now know that all edges from PredBB should be revectored to
2182  // branch to RealDest.
2183  BasicBlock *PredBB = PN->getIncomingBlock(i);
2184  BasicBlock *RealDest = BI->getSuccessor(!CB->getZExtValue());
2185 
2186  if (RealDest == BB)
2187  continue; // Skip self loops.
2188  // Skip if the predecessor's terminator is an indirect branch.
2189  if (isa<IndirectBrInst>(PredBB->getTerminator()))
2190  continue;
2191 
2192  // The dest block might have PHI nodes, other predecessors and other
2193  // difficult cases. Instead of being smart about this, just insert a new
2194  // block that jumps to the destination block, effectively splitting
2195  // the edge we are about to create.
2196  BasicBlock *EdgeBB =
2197  BasicBlock::Create(BB->getContext(), RealDest->getName() + ".critedge",
2198  RealDest->getParent(), RealDest);
2199  BranchInst::Create(RealDest, EdgeBB);
2200 
2201  // Update PHI nodes.
2202  AddPredecessorToBlock(RealDest, EdgeBB, BB);
2203 
2204  // BB may have instructions that are being threaded over. Clone these
2205  // instructions into EdgeBB. We know that there will be no uses of the
2206  // cloned instructions outside of EdgeBB.
2207  BasicBlock::iterator InsertPt = EdgeBB->begin();
2208  DenseMap<Value *, Value *> TranslateMap; // Track translated values.
2209  for (BasicBlock::iterator BBI = BB->begin(); &*BBI != BI; ++BBI) {
2210  if (PHINode *PN = dyn_cast<PHINode>(BBI)) {
2211  TranslateMap[PN] = PN->getIncomingValueForBlock(PredBB);
2212  continue;
2213  }
2214  // Clone the instruction.
2215  Instruction *N = BBI->clone();
2216  if (BBI->hasName())
2217  N->setName(BBI->getName() + ".c");
2218 
2219  // Update operands due to translation.
2220  for (User::op_iterator i = N->op_begin(), e = N->op_end(); i != e; ++i) {
2221  DenseMap<Value *, Value *>::iterator PI = TranslateMap.find(*i);
2222  if (PI != TranslateMap.end())
2223  *i = PI->second;
2224  }
2225 
2226  // Check for trivial simplification.
2227  if (Value *V = SimplifyInstruction(N, {DL, nullptr, nullptr, AC})) {
2228  if (!BBI->use_empty())
2229  TranslateMap[&*BBI] = V;
2230  if (!N->mayHaveSideEffects()) {
2231  N->deleteValue(); // Instruction folded away, don't need actual inst
2232  N = nullptr;
2233  }
2234  } else {
2235  if (!BBI->use_empty())
2236  TranslateMap[&*BBI] = N;
2237  }
2238  // Insert the new instruction into its new home.
2239  if (N)
2240  EdgeBB->getInstList().insert(InsertPt, N);
2241 
2242  // Register the new instruction with the assumption cache if necessary.
2243  if (auto *II = dyn_cast_or_null<IntrinsicInst>(N))
2244  if (II->getIntrinsicID() == Intrinsic::assume)
2245  AC->registerAssumption(II);
2246  }
2247 
2248  // Loop over all of the edges from PredBB to BB, changing them to branch
2249  // to EdgeBB instead.
2250  TerminatorInst *PredBBTI = PredBB->getTerminator();
2251  for (unsigned i = 0, e = PredBBTI->getNumSuccessors(); i != e; ++i)
2252  if (PredBBTI->getSuccessor(i) == BB) {
2253  BB->removePredecessor(PredBB);
2254  PredBBTI->setSuccessor(i, EdgeBB);
2255  }
2256 
2257  // Recurse, simplifying any other constants.
2258  return FoldCondBranchOnPHI(BI, DL, AC) || true;
2259  }
2260 
2261  return false;
2262 }
2263 
2264 /// Given a BB that starts with the specified two-entry PHI node,
2265 /// see if we can eliminate it.
2267  const DataLayout &DL) {
2268  // Ok, this is a two entry PHI node. Check to see if this is a simple "if
2269  // statement", which has a very simple dominance structure. Basically, we
2270  // are trying to find the condition that is being branched on, which
2271  // subsequently causes this merge to happen. We really want control
2272  // dependence information for this check, but simplifycfg can't keep it up
2273  // to date, and this catches most of the cases we care about anyway.
2274  BasicBlock *BB = PN->getParent();
2275  const Function *Fn = BB->getParent();
2276  if (Fn && Fn->hasFnAttribute(Attribute::OptForFuzzing))
2277  return false;
2278 
2279  BasicBlock *IfTrue, *IfFalse;
2280  Value *IfCond = GetIfCondition(BB, IfTrue, IfFalse);
2281  if (!IfCond ||
2282  // Don't bother if the branch will be constant folded trivially.
2283  isa<ConstantInt>(IfCond))
2284  return false;
2285 
2286  // Okay, we found that we can merge this two-entry phi node into a select.
2287  // Doing so would require us to fold *all* two entry phi nodes in this block.
2288  // At some point this becomes non-profitable (particularly if the target
2289  // doesn't support cmov's). Only do this transformation if there are two or
2290  // fewer PHI nodes in this block.
2291  unsigned NumPhis = 0;
2292  for (BasicBlock::iterator I = BB->begin(); isa<PHINode>(I); ++NumPhis, ++I)
2293  if (NumPhis > 2)
2294  return false;
2295 
2296  // Loop over the PHI's seeing if we can promote them all to select
2297  // instructions. While we are at it, keep track of the instructions
2298  // that need to be moved to the dominating block.
2299  SmallPtrSet<Instruction *, 4> AggressiveInsts;
2300  unsigned MaxCostVal0 = PHINodeFoldingThreshold,
2301  MaxCostVal1 = PHINodeFoldingThreshold;
2302  MaxCostVal0 *= TargetTransformInfo::TCC_Basic;
2303  MaxCostVal1 *= TargetTransformInfo::TCC_Basic;
2304 
2305  for (BasicBlock::iterator II = BB->begin(); isa<PHINode>(II);) {
2306  PHINode *PN = cast<PHINode>(II++);
2307  if (Value *V = SimplifyInstruction(PN, {DL, PN})) {
2308  PN->replaceAllUsesWith(V);
2309  PN->eraseFromParent();
2310  continue;
2311  }
2312 
2313  if (!DominatesMergePoint(PN->getIncomingValue(0), BB, AggressiveInsts,
2314  MaxCostVal0, TTI) ||
2315  !DominatesMergePoint(PN->getIncomingValue(1), BB, AggressiveInsts,
2316  MaxCostVal1, TTI))
2317  return false;
2318  }
2319 
2320  // If we folded the first phi, PN dangles at this point. Refresh it. If
2321  // we ran out of PHIs then we simplified them all.
2322  PN = dyn_cast<PHINode>(BB->begin());
2323  if (!PN)
2324  return true;
2325 
2326  // Don't fold i1 branches on PHIs which contain binary operators. These can
2327  // often be turned into switches and other things.
2328  if (PN->getType()->isIntegerTy(1) &&
2329  (isa<BinaryOperator>(PN->getIncomingValue(0)) ||
2330  isa<BinaryOperator>(PN->getIncomingValue(1)) ||
2331  isa<BinaryOperator>(IfCond)))
2332  return false;
2333 
2334  // If all PHI nodes are promotable, check to make sure that all instructions
2335  // in the predecessor blocks can be promoted as well. If not, we won't be able
2336  // to get rid of the control flow, so it's not worth promoting to select
2337  // instructions.
2338  BasicBlock *DomBlock = nullptr;
2339  BasicBlock *IfBlock1 = PN->getIncomingBlock(0);
2340  BasicBlock *IfBlock2 = PN->getIncomingBlock(1);
2341  if (cast<BranchInst>(IfBlock1->getTerminator())->isConditional()) {
2342  IfBlock1 = nullptr;
2343  } else {
2344  DomBlock = *pred_begin(IfBlock1);
2345  for (BasicBlock::iterator I = IfBlock1->begin(); !I->isTerminator(); ++I)
2346  if (!AggressiveInsts.count(&*I) && !isa<DbgInfoIntrinsic>(I)) {
2347  // This is not an aggressive instruction that we can promote.
2348  // Because of this, we won't be able to get rid of the control flow, so
2349  // the xform is not worth it.
2350  return false;
2351  }
2352  }
2353 
2354  if (cast<BranchInst>(IfBlock2->getTerminator())->isConditional()) {
2355  IfBlock2 = nullptr;
2356  } else {
2357  DomBlock = *pred_begin(IfBlock2);
2358  for (BasicBlock::iterator I = IfBlock2->begin(); !I->isTerminator(); ++I)
2359  if (!AggressiveInsts.count(&*I) && !isa<DbgInfoIntrinsic>(I)) {
2360  // This is not an aggressive instruction that we can promote.
2361  // Because of this, we won't be able to get rid of the control flow, so
2362  // the xform is not worth it.
2363  return false;
2364  }
2365  }
2366 
2367  LLVM_DEBUG(dbgs() << "FOUND IF CONDITION! " << *IfCond
2368  << " T: " << IfTrue->getName()
2369  << " F: " << IfFalse->getName() << "\n");
2370 
2371  // If we can still promote the PHI nodes after this gauntlet of tests,
2372  // do all of the PHI's now.
2373  Instruction *InsertPt = DomBlock->getTerminator();
2374  IRBuilder<NoFolder> Builder(InsertPt);
2375 
2376  // Move all 'aggressive' instructions, which are defined in the
2377  // conditional parts of the if's up to the dominating block.
2378  if (IfBlock1) {
2379  for (auto &I : *IfBlock1) {
2381  dropDebugUsers(I);
2382  }
2383  DomBlock->getInstList().splice(InsertPt->getIterator(),
2384  IfBlock1->getInstList(), IfBlock1->begin(),
2385  IfBlock1->getTerminator()->getIterator());
2386  }
2387  if (IfBlock2) {
2388  for (auto &I : *IfBlock2) {
2390  dropDebugUsers(I);
2391  }
2392  DomBlock->getInstList().splice(InsertPt->getIterator(),
2393  IfBlock2->getInstList(), IfBlock2->begin(),
2394  IfBlock2->getTerminator()->getIterator());
2395  }
2396 
2397  while (PHINode *PN = dyn_cast<PHINode>(BB->begin())) {
2398  // Change the PHI node into a select instruction.
2399  Value *TrueVal = PN->getIncomingValue(PN->getIncomingBlock(0) == IfFalse);
2400  Value *FalseVal = PN->getIncomingValue(PN->getIncomingBlock(0) == IfTrue);
2401 
2402  Value *Sel = Builder.CreateSelect(IfCond, TrueVal, FalseVal, "", InsertPt);
2403  PN->replaceAllUsesWith(Sel);
2404  Sel->takeName(PN);
2405  PN->eraseFromParent();
2406  }
2407 
2408  // At this point, IfBlock1 and IfBlock2 are both empty, so our if statement
2409  // has been flattened. Change DomBlock to jump directly to our new block to
2410  // avoid other simplifycfg's kicking in on the diamond.
2411  TerminatorInst *OldTI = DomBlock->getTerminator();
2412  Builder.SetInsertPoint(OldTI);
2413  Builder.CreateBr(BB);
2414  OldTI->eraseFromParent();
2415  return true;
2416 }
2417 
2418 /// If we found a conditional branch that goes to two returning blocks,
2419 /// try to merge them together into one return,
2420 /// introducing a select if the return values disagree.
2422  IRBuilder<> &Builder) {
2423  assert(BI->isConditional() && "Must be a conditional branch");
2424  BasicBlock *TrueSucc = BI->getSuccessor(0);
2425  BasicBlock *FalseSucc = BI->getSuccessor(1);
2426  ReturnInst *TrueRet = cast<ReturnInst>(TrueSucc->getTerminator());
2427  ReturnInst *FalseRet = cast<ReturnInst>(FalseSucc->getTerminator());
2428 
2429  // Check to ensure both blocks are empty (just a return) or optionally empty
2430  // with PHI nodes. If there are other instructions, merging would cause extra
2431  // computation on one path or the other.
2432  if (!TrueSucc->getFirstNonPHIOrDbg()->isTerminator())
2433  return false;
2434  if (!FalseSucc->getFirstNonPHIOrDbg()->isTerminator())
2435  return false;
2436 
2437  Builder.SetInsertPoint(BI);
2438  // Okay, we found a branch that is going to two return nodes. If
2439  // there is no return value for this function, just change the
2440  // branch into a return.
2441  if (FalseRet->getNumOperands() == 0) {
2442  TrueSucc->removePredecessor(BI->getParent());
2443  FalseSucc->removePredecessor(BI->getParent());
2444  Builder.CreateRetVoid();
2446  return true;
2447  }
2448 
2449  // Otherwise, figure out what the true and false return values are
2450  // so we can insert a new select instruction.
2451  Value *TrueValue = TrueRet->getReturnValue();
2452  Value *FalseValue = FalseRet->getReturnValue();
2453 
2454  // Unwrap any PHI nodes in the return blocks.
2455  if (PHINode *TVPN = dyn_cast_or_null<PHINode>(TrueValue))
2456  if (TVPN->getParent() == TrueSucc)
2457  TrueValue = TVPN->getIncomingValueForBlock(BI->getParent());
2458  if (PHINode *FVPN = dyn_cast_or_null<PHINode>(FalseValue))
2459  if (FVPN->getParent() == FalseSucc)
2460  FalseValue = FVPN->getIncomingValueForBlock(BI->getParent());
2461 
2462  // In order for this transformation to be safe, we must be able to
2463  // unconditionally execute both operands to the return. This is
2464  // normally the case, but we could have a potentially-trapping
2465  // constant expression that prevents this transformation from being
2466  // safe.
2467  if (ConstantExpr *TCV = dyn_cast_or_null<ConstantExpr>(TrueValue))
2468  if (TCV->canTrap())
2469  return false;
2470  if (ConstantExpr *FCV = dyn_cast_or_null<ConstantExpr>(FalseValue))
2471  if (FCV->canTrap())
2472  return false;
2473 
2474  // Okay, we collected all the mapped values and checked them for sanity, and
2475  // defined to really do this transformation. First, update the CFG.
2476  TrueSucc->removePredecessor(BI->getParent());
2477  FalseSucc->removePredecessor(BI->getParent());
2478 
2479  // Insert select instructions where needed.
2480  Value *BrCond = BI->getCondition();
2481  if (TrueValue) {
2482  // Insert a select if the results differ.
2483  if (TrueValue == FalseValue || isa<UndefValue>(FalseValue)) {
2484  } else if (isa<UndefValue>(TrueValue)) {
2485  TrueValue = FalseValue;
2486  } else {
2487  TrueValue =
2488  Builder.CreateSelect(BrCond, TrueValue, FalseValue, "retval", BI);
2489  }
2490  }
2491 
2492  Value *RI =
2493  !TrueValue ? Builder.CreateRetVoid() : Builder.CreateRet(TrueValue);
2494 
2495  (void)RI;
2496 
2497  LLVM_DEBUG(dbgs() << "\nCHANGING BRANCH TO TWO RETURNS INTO SELECT:"
2498  << "\n " << *BI << "NewRet = " << *RI << "TRUEBLOCK: "
2499  << *TrueSucc << "FALSEBLOCK: " << *FalseSucc);
2500 
2502 
2503  return true;
2504 }
2505 
2506 /// Return true if the given instruction is available
2507 /// in its predecessor block. If yes, the instruction will be removed.
2509  if (!isa<BinaryOperator>(Inst) && !isa<CmpInst>(Inst))
2510  return false;
2511  for (Instruction &I : *PB) {
2512  Instruction *PBI = &I;
2513  // Check whether Inst and PBI generate the same value.
2514  if (Inst->isIdenticalTo(PBI)) {
2515  Inst->replaceAllUsesWith(PBI);
2516  Inst->eraseFromParent();
2517  return true;
2518  }
2519  }
2520  return false;
2521 }
2522 
2523 /// Return true if either PBI or BI has branch weight available, and store
2524 /// the weights in {Pred|Succ}{True|False}Weight. If one of PBI and BI does
2525 /// not have branch weight, use 1:1 as its weight.
2527  uint64_t &PredTrueWeight,
2528  uint64_t &PredFalseWeight,
2529  uint64_t &SuccTrueWeight,
2530  uint64_t &SuccFalseWeight) {
2531  bool PredHasWeights =
2532  PBI->extractProfMetadata(PredTrueWeight, PredFalseWeight);
2533  bool SuccHasWeights =
2534  BI->extractProfMetadata(SuccTrueWeight, SuccFalseWeight);
2535  if (PredHasWeights || SuccHasWeights) {
2536  if (!PredHasWeights)
2537  PredTrueWeight = PredFalseWeight = 1;
2538  if (!SuccHasWeights)
2539  SuccTrueWeight = SuccFalseWeight = 1;
2540  return true;
2541  } else {
2542  return false;
2543  }
2544 }
2545 
2546 /// If this basic block is simple enough, and if a predecessor branches to us
2547 /// and one of our successors, fold the block into the predecessor and use
2548 /// logical operations to pick the right destination.
2549 bool llvm::FoldBranchToCommonDest(BranchInst *BI, unsigned BonusInstThreshold) {
2550  BasicBlock *BB = BI->getParent();
2551 
2552  const unsigned PredCount = pred_size(BB);
2553 
2554  Instruction *Cond = nullptr;
2555  if (BI->isConditional())
2556  Cond = dyn_cast<Instruction>(BI->getCondition());
2557  else {
2558  // For unconditional branch, check for a simple CFG pattern, where
2559  // BB has a single predecessor and BB's successor is also its predecessor's
2560  // successor. If such pattern exists, check for CSE between BB and its
2561  // predecessor.
2562  if (BasicBlock *PB = BB->getSinglePredecessor())
2563  if (BranchInst *PBI = dyn_cast<BranchInst>(PB->getTerminator()))
2564  if (PBI->isConditional() &&
2565  (BI->getSuccessor(0) == PBI->getSuccessor(0) ||
2566  BI->getSuccessor(0) == PBI->getSuccessor(1))) {
2567  for (auto I = BB->instructionsWithoutDebug().begin(),
2568  E = BB->instructionsWithoutDebug().end();
2569  I != E;) {
2570  Instruction *Curr = &*I++;
2571  if (isa<CmpInst>(Curr)) {
2572  Cond = Curr;
2573  break;
2574  }
2575  // Quit if we can't remove this instruction.
2576  if (!tryCSEWithPredecessor(Curr, PB))
2577  return false;
2578  }
2579  }
2580 
2581  if (!Cond)
2582  return false;
2583  }
2584 
2585  if (!Cond || (!isa<CmpInst>(Cond) && !isa<BinaryOperator>(Cond)) ||
2586  Cond->getParent() != BB || !Cond->hasOneUse())
2587  return false;
2588 
2589  // Make sure the instruction after the condition is the cond branch.
2590  BasicBlock::iterator CondIt = ++Cond->getIterator();
2591 
2592  // Ignore dbg intrinsics.
2593  while (isa<DbgInfoIntrinsic>(CondIt))
2594  ++CondIt;
2595 
2596  if (&*CondIt != BI)
2597  return false;
2598 
2599  // Only allow this transformation if computing the condition doesn't involve
2600  // too many instructions and these involved instructions can be executed
2601  // unconditionally. We denote all involved instructions except the condition
2602  // as "bonus instructions", and only allow this transformation when the
2603  // number of the bonus instructions we'll need to create when cloning into
2604  // each predecessor does not exceed a certain threshold.
2605  unsigned NumBonusInsts = 0;
2606  for (auto I = BB->begin(); Cond != &*I; ++I) {
2607  // Ignore dbg intrinsics.
2608  if (isa<DbgInfoIntrinsic>(I))
2609  continue;
2610  if (!I->hasOneUse() || !isSafeToSpeculativelyExecute(&*I))
2611  return false;
2612  // I has only one use and can be executed unconditionally.
2614  if (User == nullptr || User->getParent() != BB)
2615  return false;
2616  // I is used in the same BB. Since BI uses Cond and doesn't have more slots
2617  // to use any other instruction, User must be an instruction between next(I)
2618  // and Cond.
2619 
2620  // Account for the cost of duplicating this instruction into each
2621  // predecessor.
2622  NumBonusInsts += PredCount;
2623  // Early exits once we reach the limit.
2624  if (NumBonusInsts > BonusInstThreshold)
2625  return false;
2626  }
2627 
2628  // Cond is known to be a compare or binary operator. Check to make sure that
2629  // neither operand is a potentially-trapping constant expression.
2630  if (ConstantExpr *CE = dyn_cast<ConstantExpr>(Cond->getOperand(0)))
2631  if (CE->canTrap())
2632  return false;
2633  if (ConstantExpr *CE = dyn_cast<ConstantExpr>(Cond->getOperand(1)))
2634  if (CE->canTrap())
2635  return false;
2636 
2637  // Finally, don't infinitely unroll conditional loops.
2638  BasicBlock *TrueDest = BI->getSuccessor(0);
2639  BasicBlock *FalseDest = (BI->isConditional()) ? BI->getSuccessor(1) : nullptr;
2640  if (TrueDest == BB || FalseDest == BB)
2641  return false;
2642 
2643  for (pred_iterator PI = pred_begin(BB), E = pred_end(BB); PI != E; ++PI) {
2644  BasicBlock *PredBlock = *PI;
2645  BranchInst *PBI = dyn_cast<BranchInst>(PredBlock->getTerminator());
2646 
2647  // Check that we have two conditional branches. If there is a PHI node in
2648  // the common successor, verify that the same value flows in from both
2649  // blocks.
2651  if (!PBI || PBI->isUnconditional() ||
2652  (BI->isConditional() && !SafeToMergeTerminators(BI, PBI)) ||
2653  (!BI->isConditional() &&
2654  !isProfitableToFoldUnconditional(BI, PBI, Cond, PHIs)))
2655  continue;
2656 
2657  // Determine if the two branches share a common destination.
2658  Instruction::BinaryOps Opc = Instruction::BinaryOpsEnd;
2659  bool InvertPredCond = false;
2660 
2661  if (BI->isConditional()) {
2662  if (PBI->getSuccessor(0) == TrueDest) {
2663  Opc = Instruction::Or;
2664  } else if (PBI->getSuccessor(1) == FalseDest) {
2665  Opc = Instruction::And;
2666  } else if (PBI->getSuccessor(0) == FalseDest) {
2667  Opc = Instruction::And;
2668  InvertPredCond = true;
2669  } else if (PBI->getSuccessor(1) == TrueDest) {
2670  Opc = Instruction::Or;
2671  InvertPredCond = true;
2672  } else {
2673  continue;
2674  }
2675  } else {
2676  if (PBI->getSuccessor(0) != TrueDest && PBI->getSuccessor(1) != TrueDest)
2677  continue;
2678  }
2679 
2680  LLVM_DEBUG(dbgs() << "FOLDING BRANCH TO COMMON DEST:\n" << *PBI << *BB);
2681  IRBuilder<> Builder(PBI);
2682 
2683  // If we need to invert the condition in the pred block to match, do so now.
2684  if (InvertPredCond) {
2685  Value *NewCond = PBI->getCondition();
2686 
2687  if (NewCond->hasOneUse() && isa<CmpInst>(NewCond)) {
2688  CmpInst *CI = cast<CmpInst>(NewCond);
2689  CI->setPredicate(CI->getInversePredicate());
2690  } else {
2691  NewCond =
2692  Builder.CreateNot(NewCond, PBI->getCondition()->getName() + ".not");
2693  }
2694 
2695  PBI->setCondition(NewCond);
2696  PBI->swapSuccessors();
2697  }
2698 
2699  // If we have bonus instructions, clone them into the predecessor block.
2700  // Note that there may be multiple predecessor blocks, so we cannot move
2701  // bonus instructions to a predecessor block.
2702  ValueToValueMapTy VMap; // maps original values to cloned values
2703  // We already make sure Cond is the last instruction before BI. Therefore,
2704  // all instructions before Cond other than DbgInfoIntrinsic are bonus
2705  // instructions.
2706  for (auto BonusInst = BB->begin(); Cond != &*BonusInst; ++BonusInst) {
2707  if (isa<DbgInfoIntrinsic>(BonusInst))
2708  continue;
2709  Instruction *NewBonusInst = BonusInst->clone();
2710  RemapInstruction(NewBonusInst, VMap,
2712  VMap[&*BonusInst] = NewBonusInst;
2713 
2714  // If we moved a load, we cannot any longer claim any knowledge about
2715  // its potential value. The previous information might have been valid
2716  // only given the branch precondition.
2717  // For an analogous reason, we must also drop all the metadata whose
2718  // semantics we don't understand.
2719  NewBonusInst->dropUnknownNonDebugMetadata();
2720 
2721  PredBlock->getInstList().insert(PBI->getIterator(), NewBonusInst);
2722  NewBonusInst->takeName(&*BonusInst);
2723  BonusInst->setName(BonusInst->getName() + ".old");
2724  }
2725 
2726  // Clone Cond into the predecessor basic block, and or/and the
2727  // two conditions together.
2728  Instruction *CondInPred = Cond->clone();
2729  RemapInstruction(CondInPred, VMap,
2731  PredBlock->getInstList().insert(PBI->getIterator(), CondInPred);
2732  CondInPred->takeName(Cond);
2733  Cond->setName(CondInPred->getName() + ".old");
2734 
2735  if (BI->isConditional()) {
2736  Instruction *NewCond = cast<Instruction>(
2737  Builder.CreateBinOp(Opc, PBI->getCondition(), CondInPred, "or.cond"));
2738  PBI->setCondition(NewCond);
2739 
2740  uint64_t PredTrueWeight, PredFalseWeight, SuccTrueWeight, SuccFalseWeight;
2741  bool HasWeights =
2742  extractPredSuccWeights(PBI, BI, PredTrueWeight, PredFalseWeight,
2743  SuccTrueWeight, SuccFalseWeight);
2744  SmallVector<uint64_t, 8> NewWeights;
2745 
2746  if (PBI->getSuccessor(0) == BB) {
2747  if (HasWeights) {
2748  // PBI: br i1 %x, BB, FalseDest
2749  // BI: br i1 %y, TrueDest, FalseDest
2750  // TrueWeight is TrueWeight for PBI * TrueWeight for BI.
2751  NewWeights.push_back(PredTrueWeight * SuccTrueWeight);
2752  // FalseWeight is FalseWeight for PBI * TotalWeight for BI +
2753  // TrueWeight for PBI * FalseWeight for BI.
2754  // We assume that total weights of a BranchInst can fit into 32 bits.
2755  // Therefore, we will not have overflow using 64-bit arithmetic.
2756  NewWeights.push_back(PredFalseWeight *
2757  (SuccFalseWeight + SuccTrueWeight) +
2758  PredTrueWeight * SuccFalseWeight);
2759  }
2760  AddPredecessorToBlock(TrueDest, PredBlock, BB);
2761  PBI->setSuccessor(0, TrueDest);
2762  }
2763  if (PBI->getSuccessor(1) == BB) {
2764  if (HasWeights) {
2765  // PBI: br i1 %x, TrueDest, BB
2766  // BI: br i1 %y, TrueDest, FalseDest
2767  // TrueWeight is TrueWeight for PBI * TotalWeight for BI +
2768  // FalseWeight for PBI * TrueWeight for BI.
2769  NewWeights.push_back(PredTrueWeight *
2770  (SuccFalseWeight + SuccTrueWeight) +
2771  PredFalseWeight * SuccTrueWeight);
2772  // FalseWeight is FalseWeight for PBI * FalseWeight for BI.
2773  NewWeights.push_back(PredFalseWeight * SuccFalseWeight);
2774  }
2775  AddPredecessorToBlock(FalseDest, PredBlock, BB);
2776  PBI->setSuccessor(1, FalseDest);
2777  }
2778  if (NewWeights.size() == 2) {
2779  // Halve the weights if any of them cannot fit in an uint32_t
2780  FitWeights(NewWeights);
2781 
2782  SmallVector<uint32_t, 8> MDWeights(NewWeights.begin(),
2783  NewWeights.end());
2784  setBranchWeights(PBI, MDWeights[0], MDWeights[1]);
2785  } else
2786  PBI->setMetadata(LLVMContext::MD_prof, nullptr);
2787  } else {
2788  // Update PHI nodes in the common successors.
2789  for (unsigned i = 0, e = PHIs.size(); i != e; ++i) {
2790  ConstantInt *PBI_C = cast<ConstantInt>(
2791  PHIs[i]->getIncomingValueForBlock(PBI->getParent()));
2792  assert(PBI_C->getType()->isIntegerTy(1));
2793  Instruction *MergedCond = nullptr;
2794  if (PBI->getSuccessor(0) == TrueDest) {
2795  // Create (PBI_Cond and PBI_C) or (!PBI_Cond and BI_Value)
2796  // PBI_C is true: PBI_Cond or (!PBI_Cond and BI_Value)
2797  // is false: !PBI_Cond and BI_Value
2798  Instruction *NotCond = cast<Instruction>(
2799  Builder.CreateNot(PBI->getCondition(), "not.cond"));
2800  MergedCond = cast<Instruction>(
2801  Builder.CreateBinOp(Instruction::And, NotCond, CondInPred,
2802  "and.cond"));
2803  if (PBI_C->isOne())
2804  MergedCond = cast<Instruction>(Builder.CreateBinOp(
2805  Instruction::Or, PBI->getCondition(), MergedCond, "or.cond"));
2806  } else {
2807  // Create (PBI_Cond and BI_Value) or (!PBI_Cond and PBI_C)
2808  // PBI_C is true: (PBI_Cond and BI_Value) or (!PBI_Cond)
2809  // is false: PBI_Cond and BI_Value
2810  MergedCond = cast<Instruction>(Builder.CreateBinOp(
2811  Instruction::And, PBI->getCondition(), CondInPred, "and.cond"));
2812  if (PBI_C->isOne()) {
2813  Instruction *NotCond = cast<Instruction>(
2814  Builder.CreateNot(PBI->getCondition(), "not.cond"));
2815  MergedCond = cast<Instruction>(Builder.CreateBinOp(
2816  Instruction::Or, NotCond, MergedCond, "or.cond"));
2817  }
2818  }
2819  // Update PHI Node.
2820  PHIs[i]->setIncomingValue(PHIs[i]->getBasicBlockIndex(PBI->getParent()),
2821  MergedCond);
2822  }
2823  // Change PBI from Conditional to Unconditional.
2824  BranchInst *New_PBI = BranchInst::Create(TrueDest, PBI);
2826  PBI = New_PBI;
2827  }
2828 
2829  // If BI was a loop latch, it may have had associated loop metadata.
2830  // We need to copy it to the new latch, that is, PBI.
2831  if (MDNode *LoopMD = BI->getMetadata(LLVMContext::MD_loop))
2832  PBI->setMetadata(LLVMContext::MD_loop, LoopMD);
2833 
2834  // TODO: If BB is reachable from all paths through PredBlock, then we
2835  // could replace PBI's branch probabilities with BI's.
2836 
2837  // Copy any debug value intrinsics into the end of PredBlock.
2838  for (Instruction &I : *BB)
2839  if (isa<DbgInfoIntrinsic>(I))
2840  I.clone()->insertBefore(PBI);
2841 
2842  return true;
2843  }
2844  return false;
2845 }
2846 
2847 // If there is only one store in BB1 and BB2, return it, otherwise return
2848 // nullptr.
2850  StoreInst *S = nullptr;
2851  for (auto *BB : {BB1, BB2}) {
2852  if (!BB)
2853  continue;
2854  for (auto &I : *BB)
2855  if (auto *SI = dyn_cast<StoreInst>(&I)) {
2856  if (S)
2857  // Multiple stores seen.
2858  return nullptr;
2859  else
2860  S = SI;
2861  }
2862  }
2863  return S;
2864 }
2865 
2867  Value *AlternativeV = nullptr) {
2868  // PHI is going to be a PHI node that allows the value V that is defined in
2869  // BB to be referenced in BB's only successor.
2870  //
2871  // If AlternativeV is nullptr, the only value we care about in PHI is V. It
2872  // doesn't matter to us what the other operand is (it'll never get used). We
2873  // could just create a new PHI with an undef incoming value, but that could
2874  // increase register pressure if EarlyCSE/InstCombine can't fold it with some
2875  // other PHI. So here we directly look for some PHI in BB's successor with V
2876  // as an incoming operand. If we find one, we use it, else we create a new
2877  // one.
2878  //
2879  // If AlternativeV is not nullptr, we care about both incoming values in PHI.
2880  // PHI must be exactly: phi <ty> [ %BB, %V ], [ %OtherBB, %AlternativeV]
2881  // where OtherBB is the single other predecessor of BB's only successor.
2882  PHINode *PHI = nullptr;
2883  BasicBlock *Succ = BB->getSingleSuccessor();
2884 
2885  for (auto I = Succ->begin(); isa<PHINode>(I); ++I)
2886  if (cast<PHINode>(I)->getIncomingValueForBlock(BB) == V) {
2887  PHI = cast<PHINode>(I);
2888  if (!AlternativeV)
2889  break;
2890 
2891  assert(pred_size(Succ) == 2);
2892  auto PredI = pred_begin(Succ);
2893  BasicBlock *OtherPredBB = *PredI == BB ? *++PredI : *PredI;
2894  if (PHI->getIncomingValueForBlock(OtherPredBB) == AlternativeV)
2895  break;
2896  PHI = nullptr;
2897  }
2898  if (PHI)
2899  return PHI;
2900 
2901  // If V is not an instruction defined in BB, just return it.
2902  if (!AlternativeV &&
2903  (!isa<Instruction>(V) || cast<Instruction>(V)->getParent() != BB))
2904  return V;
2905 
2906  PHI = PHINode::Create(V->getType(), 2, "simplifycfg.merge", &Succ->front());
2907  PHI->addIncoming(V, BB);
2908  for (BasicBlock *PredBB : predecessors(Succ))
2909  if (PredBB != BB)
2910  PHI->addIncoming(
2911  AlternativeV ? AlternativeV : UndefValue::get(V->getType()), PredBB);
2912  return PHI;
2913 }
2914 
2916  BasicBlock *QTB, BasicBlock *QFB,
2917  BasicBlock *PostBB, Value *Address,
2918  bool InvertPCond, bool InvertQCond,
2919  const DataLayout &DL) {
2920  auto IsaBitcastOfPointerType = [](const Instruction &I) {
2921  return Operator::getOpcode(&I) == Instruction::BitCast &&
2922  I.getType()->isPointerTy();
2923  };
2924 
2925  // If we're not in aggressive mode, we only optimize if we have some
2926  // confidence that by optimizing we'll allow P and/or Q to be if-converted.
2927  auto IsWorthwhile = [&](BasicBlock *BB) {
2928  if (!BB)
2929  return true;
2930  // Heuristic: if the block can be if-converted/phi-folded and the
2931  // instructions inside are all cheap (arithmetic/GEPs), it's worthwhile to
2932  // thread this store.
2933  unsigned N = 0;
2934  for (auto &I : BB->instructionsWithoutDebug()) {
2935  // Cheap instructions viable for folding.
2936  if (isa<BinaryOperator>(I) || isa<GetElementPtrInst>(I) ||
2937  isa<StoreInst>(I))
2938  ++N;
2939  // Free instructions.
2940  else if (I.isTerminator() || IsaBitcastOfPointerType(I))
2941  continue;
2942  else
2943  return false;
2944  }
2945  // The store we want to merge is counted in N, so add 1 to make sure
2946  // we're counting the instructions that would be left.
2947  return N <= (PHINodeFoldingThreshold + 1);
2948  };
2949 
2951  (!IsWorthwhile(PTB) || !IsWorthwhile(PFB) || !IsWorthwhile(QTB) ||
2952  !IsWorthwhile(QFB)))
2953  return false;
2954 
2955  // For every pointer, there must be exactly two stores, one coming from
2956  // PTB or PFB, and the other from QTB or QFB. We don't support more than one
2957  // store (to any address) in PTB,PFB or QTB,QFB.
2958  // FIXME: We could relax this restriction with a bit more work and performance
2959  // testing.
2960  StoreInst *PStore = findUniqueStoreInBlocks(PTB, PFB);
2961  StoreInst *QStore = findUniqueStoreInBlocks(QTB, QFB);
2962  if (!PStore || !QStore)
2963  return false;
2964 
2965  // Now check the stores are compatible.
2966  if (!QStore->isUnordered() || !PStore->isUnordered())
2967  return false;
2968 
2969  // Check that sinking the store won't cause program behavior changes. Sinking
2970  // the store out of the Q blocks won't change any behavior as we're sinking
2971  // from a block to its unconditional successor. But we're moving a store from
2972  // the P blocks down through the middle block (QBI) and past both QFB and QTB.
2973  // So we need to check that there are no aliasing loads or stores in
2974  // QBI, QTB and QFB. We also need to check there are no conflicting memory
2975  // operations between PStore and the end of its parent block.
2976  //
2977  // The ideal way to do this is to query AliasAnalysis, but we don't
2978  // preserve AA currently so that is dangerous. Be super safe and just
2979  // check there are no other memory operations at all.
2980  for (auto &I : *QFB->getSinglePredecessor())
2981  if (I.mayReadOrWriteMemory())
2982  return false;
2983  for (auto &I : *QFB)
2984  if (&I != QStore && I.mayReadOrWriteMemory())
2985  return false;
2986  if (QTB)
2987  for (auto &I : *QTB)
2988  if (&I != QStore && I.mayReadOrWriteMemory())
2989  return false;
2990  for (auto I = BasicBlock::iterator(PStore), E = PStore->getParent()->end();
2991  I != E; ++I)
2992  if (&*I != PStore && I->mayReadOrWriteMemory())
2993  return false;
2994 
2995  // If PostBB has more than two predecessors, we need to split it so we can
2996  // sink the store.
2997  if (std::next(pred_begin(PostBB), 2) != pred_end(PostBB)) {
2998  // We know that QFB's only successor is PostBB. And QFB has a single
2999  // predecessor. If QTB exists, then its only successor is also PostBB.
3000  // If QTB does not exist, then QFB's only predecessor has a conditional
3001  // branch to QFB and PostBB.
3002  BasicBlock *TruePred = QTB ? QTB : QFB->getSinglePredecessor();
3003  BasicBlock *NewBB = SplitBlockPredecessors(PostBB, { QFB, TruePred},
3004  "condstore.split");
3005  if (!NewBB)
3006  return false;
3007  PostBB = NewBB;
3008  }
3009 
3010  // OK, we're going to sink the stores to PostBB. The store has to be
3011  // conditional though, so first create the predicate.
3012  Value *PCond = cast<BranchInst>(PFB->getSinglePredecessor()->getTerminator())
3013  ->getCondition();
3014  Value *QCond = cast<BranchInst>(QFB->getSinglePredecessor()->getTerminator())
3015  ->getCondition();
3016 
3018  PStore->getParent());
3020  QStore->getParent(), PPHI);
3021 
3022  IRBuilder<> QB(&*PostBB->getFirstInsertionPt());
3023 
3024  Value *PPred = PStore->getParent() == PTB ? PCond : QB.CreateNot(PCond);
3025  Value *QPred = QStore->getParent() == QTB ? QCond : QB.CreateNot(QCond);
3026 
3027  if (InvertPCond)
3028  PPred = QB.CreateNot(PPred);
3029  if (InvertQCond)
3030  QPred = QB.CreateNot(QPred);
3031  Value *CombinedPred = QB.CreateOr(PPred, QPred);
3032 
3033  auto *T =
3034  SplitBlockAndInsertIfThen(CombinedPred, &*QB.GetInsertPoint(), false);
3035  QB.SetInsertPoint(T);
3036  StoreInst *SI = cast<StoreInst>(QB.CreateStore(QPHI, Address));
3037  AAMDNodes AAMD;
3038  PStore->getAAMetadata(AAMD, /*Merge=*/false);
3039  PStore->getAAMetadata(AAMD, /*Merge=*/true);
3040  SI->setAAMetadata(AAMD);
3041  unsigned PAlignment = PStore->getAlignment();
3042  unsigned QAlignment = QStore->getAlignment();
3043  unsigned TypeAlignment =
3045  unsigned MinAlignment;
3046  unsigned MaxAlignment;
3047  std::tie(MinAlignment, MaxAlignment) = std::minmax(PAlignment, QAlignment);
3048  // Choose the minimum alignment. If we could prove both stores execute, we
3049  // could use biggest one. In this case, though, we only know that one of the
3050  // stores executes. And we don't know it's safe to take the alignment from a
3051  // store that doesn't execute.
3052  if (MinAlignment != 0) {
3053  // Choose the minimum of all non-zero alignments.
3054  SI->setAlignment(MinAlignment);
3055  } else if (MaxAlignment != 0) {
3056  // Choose the minimal alignment between the non-zero alignment and the ABI
3057  // default alignment for the type of the stored value.
3058  SI->setAlignment(std::min(MaxAlignment, TypeAlignment));
3059  } else {
3060  // If both alignments are zero, use ABI default alignment for the type of
3061  // the stored value.
3062  SI->setAlignment(TypeAlignment);
3063  }
3064 
3065  QStore->eraseFromParent();
3066  PStore->eraseFromParent();
3067 
3068  return true;
3069 }
3070 
3072  const DataLayout &DL) {
3073  // The intention here is to find diamonds or triangles (see below) where each
3074  // conditional block contains a store to the same address. Both of these
3075  // stores are conditional, so they can't be unconditionally sunk. But it may
3076  // be profitable to speculatively sink the stores into one merged store at the
3077  // end, and predicate the merged store on the union of the two conditions of
3078  // PBI and QBI.
3079  //
3080  // This can reduce the number of stores executed if both of the conditions are
3081  // true, and can allow the blocks to become small enough to be if-converted.
3082  // This optimization will also chain, so that ladders of test-and-set
3083  // sequences can be if-converted away.
3084  //
3085  // We only deal with simple diamonds or triangles:
3086  //
3087  // PBI or PBI or a combination of the two
3088  // / \ | \
3089  // PTB PFB | PFB
3090  // \ / | /
3091  // QBI QBI
3092  // / \ | \
3093  // QTB QFB | QFB
3094  // \ / | /
3095  // PostBB PostBB
3096  //
3097  // We model triangles as a type of diamond with a nullptr "true" block.
3098  // Triangles are canonicalized so that the fallthrough edge is represented by
3099  // a true condition, as in the diagram above.
3100  BasicBlock *PTB = PBI->getSuccessor(0);
3101  BasicBlock *PFB = PBI->getSuccessor(1);
3102  BasicBlock *QTB = QBI->getSuccessor(0);
3103  BasicBlock *QFB = QBI->getSuccessor(1);
3104  BasicBlock *PostBB = QFB->getSingleSuccessor();
3105 
3106  // Make sure we have a good guess for PostBB. If QTB's only successor is
3107  // QFB, then QFB is a better PostBB.
3108  if (QTB->getSingleSuccessor() == QFB)
3109  PostBB = QFB;
3110 
3111  // If we couldn't find a good PostBB, stop.
3112  if (!PostBB)
3113  return false;
3114 
3115  bool InvertPCond = false, InvertQCond = false;
3116  // Canonicalize fallthroughs to the true branches.
3117  if (PFB == QBI->getParent()) {
3118  std::swap(PFB, PTB);
3119  InvertPCond = true;
3120  }
3121  if (QFB == PostBB) {
3122  std::swap(QFB, QTB);
3123  InvertQCond = true;
3124  }
3125 
3126  // From this point on we can assume PTB or QTB may be fallthroughs but PFB
3127  // and QFB may not. Model fallthroughs as a nullptr block.
3128  if (PTB == QBI->getParent())
3129  PTB = nullptr;
3130  if (QTB == PostBB)
3131  QTB = nullptr;
3132 
3133  // Legality bailouts. We must have at least the non-fallthrough blocks and
3134  // the post-dominating block, and the non-fallthroughs must only have one
3135  // predecessor.
3136  auto HasOnePredAndOneSucc = [](BasicBlock *BB, BasicBlock *P, BasicBlock *S) {
3137  return BB->getSinglePredecessor() == P && BB->getSingleSuccessor() == S;
3138  };
3139  if (!HasOnePredAndOneSucc(PFB, PBI->getParent(), QBI->getParent()) ||
3140  !HasOnePredAndOneSucc(QFB, QBI->getParent(), PostBB))
3141  return false;
3142  if ((PTB && !HasOnePredAndOneSucc(PTB, PBI->getParent(), QBI->getParent())) ||
3143  (QTB && !HasOnePredAndOneSucc(QTB, QBI->getParent(), PostBB)))
3144  return false;
3145  if (!QBI->getParent()->hasNUses(2))
3146  return false;
3147 
3148  // OK, this is a sequence of two diamonds or triangles.
3149  // Check if there are stores in PTB or PFB that are repeated in QTB or QFB.
3150  SmallPtrSet<Value *, 4> PStoreAddresses, QStoreAddresses;
3151  for (auto *BB : {PTB, PFB}) {
3152  if (!BB)
3153  continue;
3154  for (auto &I : *BB)
3155  if (StoreInst *SI = dyn_cast<StoreInst>(&I))
3156  PStoreAddresses.insert(SI->getPointerOperand());
3157  }
3158  for (auto *BB : {QTB, QFB}) {
3159  if (!BB)
3160  continue;
3161  for (auto &I : *BB)
3162  if (StoreInst *SI = dyn_cast<StoreInst>(&I))
3163  QStoreAddresses.insert(SI->getPointerOperand());
3164  }
3165 
3166  set_intersect(PStoreAddresses, QStoreAddresses);
3167  // set_intersect mutates PStoreAddresses in place. Rename it here to make it
3168  // clear what it contains.
3169  auto &CommonAddresses = PStoreAddresses;
3170 
3171  bool Changed = false;
3172  for (auto *Address : CommonAddresses)
3173  Changed |= mergeConditionalStoreToAddress(
3174  PTB, PFB, QTB, QFB, PostBB, Address, InvertPCond, InvertQCond, DL);
3175  return Changed;
3176 }
3177 
3178 /// If we have a conditional branch as a predecessor of another block,
3179 /// this function tries to simplify it. We know
3180 /// that PBI and BI are both conditional branches, and BI is in one of the
3181 /// successor blocks of PBI - PBI branches to BI.
3183  const DataLayout &DL) {
3184  assert(PBI->isConditional() && BI->isConditional());
3185  BasicBlock *BB = BI->getParent();
3186 
3187  // If this block ends with a branch instruction, and if there is a
3188  // predecessor that ends on a branch of the same condition, make
3189  // this conditional branch redundant.
3190  if (PBI->getCondition() == BI->getCondition() &&
3191  PBI->getSuccessor(0) != PBI->getSuccessor(1)) {
3192  // Okay, the outcome of this conditional branch is statically
3193  // knowable. If this block had a single pred, handle specially.
3194  if (BB->getSinglePredecessor()) {
3195  // Turn this into a branch on constant.
3196  bool CondIsTrue = PBI->getSuccessor(0) == BB;
3197  BI->setCondition(
3198  ConstantInt::get(Type::getInt1Ty(BB->getContext()), CondIsTrue));
3199  return true; // Nuke the branch on constant.
3200  }
3201 
3202  // Otherwise, if there are multiple predecessors, insert a PHI that merges
3203  // in the constant and simplify the block result. Subsequent passes of
3204  // simplifycfg will thread the block.
3206  pred_iterator PB = pred_begin(BB), PE = pred_end(BB);
3207  PHINode *NewPN = PHINode::Create(
3208  Type::getInt1Ty(BB->getContext()), std::distance(PB, PE),
3209  BI->getCondition()->getName() + ".pr", &BB->front());
3210  // Okay, we're going to insert the PHI node. Since PBI is not the only
3211  // predecessor, compute the PHI'd conditional value for all of the preds.
3212  // Any predecessor where the condition is not computable we keep symbolic.
3213  for (pred_iterator PI = PB; PI != PE; ++PI) {
3214  BasicBlock *P = *PI;
3215  if ((PBI = dyn_cast<BranchInst>(P->getTerminator())) && PBI != BI &&
3216  PBI->isConditional() && PBI->getCondition() == BI->getCondition() &&
3217  PBI->getSuccessor(0) != PBI->getSuccessor(1)) {
3218  bool CondIsTrue = PBI->getSuccessor(0) == BB;
3219  NewPN->addIncoming(
3220  ConstantInt::get(Type::getInt1Ty(BB->getContext()), CondIsTrue),
3221  P);
3222  } else {
3223  NewPN->addIncoming(BI->getCondition(), P);
3224  }
3225  }
3226 
3227  BI->setCondition(NewPN);
3228  return true;
3229  }
3230  }
3231 
3232  if (auto *CE = dyn_cast<ConstantExpr>(BI->getCondition()))
3233  if (CE->canTrap())
3234  return false;
3235 
3236  // If both branches are conditional and both contain stores to the same
3237  // address, remove the stores from the conditionals and create a conditional
3238  // merged store at the end.
3239  if (MergeCondStores && mergeConditionalStores(PBI, BI, DL))
3240  return true;
3241 
3242  // If this is a conditional branch in an empty block, and if any
3243  // predecessors are a conditional branch to one of our destinations,
3244  // fold the conditions into logical ops and one cond br.
3245 
3246  // Ignore dbg intrinsics.
3247  if (&*BB->instructionsWithoutDebug().begin() != BI)
3248  return false;
3249 
3250  int PBIOp, BIOp;
3251  if (PBI->getSuccessor(0) == BI->getSuccessor(0)) {
3252  PBIOp = 0;
3253  BIOp = 0;
3254  } else if (PBI->getSuccessor(0) == BI->getSuccessor(1)) {
3255  PBIOp = 0;
3256  BIOp = 1;
3257  } else if (PBI->getSuccessor(1) == BI->getSuccessor(0)) {
3258  PBIOp = 1;
3259  BIOp = 0;
3260  } else if (PBI->getSuccessor(1) == BI->getSuccessor(1)) {
3261  PBIOp = 1;
3262  BIOp = 1;
3263  } else {
3264  return false;
3265  }
3266 
3267  // Check to make sure that the other destination of this branch
3268  // isn't BB itself. If so, this is an infinite loop that will
3269  // keep getting unwound.
3270  if (PBI->getSuccessor(PBIOp) == BB)
3271  return false;
3272 
3273  // Do not perform this transformation if it would require
3274  // insertion of a large number of select instructions. For targets
3275  // without predication/cmovs, this is a big pessimization.
3276 
3277  // Also do not perform this transformation if any phi node in the common
3278  // destination block can trap when reached by BB or PBB (PR17073). In that
3279  // case, it would be unsafe to hoist the operation into a select instruction.
3280 
3281  BasicBlock *CommonDest = PBI->getSuccessor(PBIOp);
3282  unsigned NumPhis = 0;
3283  for (BasicBlock::iterator II = CommonDest->begin(); isa<PHINode>(II);
3284  ++II, ++NumPhis) {
3285  if (NumPhis > 2) // Disable this xform.
3286  return false;
3287 
3288  PHINode *PN = cast<PHINode>(II);
3289  Value *BIV = PN->getIncomingValueForBlock(BB);
3290  if (ConstantExpr *CE = dyn_cast<ConstantExpr>(BIV))
3291  if (CE->canTrap())
3292  return false;
3293 
3294  unsigned PBBIdx = PN->getBasicBlockIndex(PBI->getParent());
3295  Value *PBIV = PN->getIncomingValue(PBBIdx);
3296  if (ConstantExpr *CE = dyn_cast<ConstantExpr>(PBIV))
3297  if (CE->canTrap())
3298  return false;
3299  }
3300 
3301  // Finally, if everything is ok, fold the branches to logical ops.
3302  BasicBlock *OtherDest = BI->getSuccessor(BIOp ^ 1);
3303 
3304  LLVM_DEBUG(dbgs() << "FOLDING BRs:" << *PBI->getParent()
3305  << "AND: " << *BI->getParent());
3306 
3307  // If OtherDest *is* BB, then BB is a basic block with a single conditional
3308  // branch in it, where one edge (OtherDest) goes back to itself but the other
3309  // exits. We don't *know* that the program avoids the infinite loop
3310  // (even though that seems likely). If we do this xform naively, we'll end up
3311  // recursively unpeeling the loop. Since we know that (after the xform is
3312  // done) that the block *is* infinite if reached, we just make it an obviously
3313  // infinite loop with no cond branch.
3314  if (OtherDest == BB) {
3315  // Insert it at the end of the function, because it's either code,
3316  // or it won't matter if it's hot. :)
3317  BasicBlock *InfLoopBlock =
3318  BasicBlock::Create(BB->getContext(), "infloop", BB->getParent());
3319  BranchInst::Create(InfLoopBlock, InfLoopBlock);
3320  OtherDest = InfLoopBlock;
3321  }
3322 
3323  LLVM_DEBUG(dbgs() << *PBI->getParent()->getParent());
3324 
3325  // BI may have other predecessors. Because of this, we leave
3326  // it alone, but modify PBI.
3327 
3328  // Make sure we get to CommonDest on True&True directions.
3329  Value *PBICond = PBI->getCondition();
3330  IRBuilder<NoFolder> Builder(PBI);
3331  if (PBIOp)
3332  PBICond = Builder.CreateNot(PBICond, PBICond->getName() + ".not");
3333 
3334  Value *BICond = BI->getCondition();
3335  if (BIOp)
3336  BICond = Builder.CreateNot(BICond, BICond->getName() + ".not");
3337 
3338  // Merge the conditions.
3339  Value *Cond = Builder.CreateOr(PBICond, BICond, "brmerge");
3340 
3341  // Modify PBI to branch on the new condition to the new dests.
3342  PBI->setCondition(Cond);
3343  PBI->setSuccessor(0, CommonDest);
3344  PBI->setSuccessor(1, OtherDest);
3345 
3346  // Update branch weight for PBI.
3347  uint64_t PredTrueWeight, PredFalseWeight, SuccTrueWeight, SuccFalseWeight;
3348  uint64_t PredCommon, PredOther, SuccCommon, SuccOther;
3349  bool HasWeights =
3350  extractPredSuccWeights(PBI, BI, PredTrueWeight, PredFalseWeight,
3351  SuccTrueWeight, SuccFalseWeight);
3352  if (HasWeights) {
3353  PredCommon = PBIOp ? PredFalseWeight : PredTrueWeight;
3354  PredOther = PBIOp ? PredTrueWeight : PredFalseWeight;
3355  SuccCommon = BIOp ? SuccFalseWeight : SuccTrueWeight;
3356  SuccOther = BIOp ? SuccTrueWeight : SuccFalseWeight;
3357  // The weight to CommonDest should be PredCommon * SuccTotal +
3358  // PredOther * SuccCommon.
3359  // The weight to OtherDest should be PredOther * SuccOther.
3360  uint64_t NewWeights[2] = {PredCommon * (SuccCommon + SuccOther) +
3361  PredOther * SuccCommon,
3362  PredOther * SuccOther};
3363  // Halve the weights if any of them cannot fit in an uint32_t
3364  FitWeights(NewWeights);
3365 
3366  setBranchWeights(PBI, NewWeights[0], NewWeights[1]);
3367  }
3368 
3369  // OtherDest may have phi nodes. If so, add an entry from PBI's
3370  // block that are identical to the entries for BI's block.
3371  AddPredecessorToBlock(OtherDest, PBI->getParent(), BB);
3372 
3373  // We know that the CommonDest already had an edge from PBI to
3374  // it. If it has PHIs though, the PHIs may have different
3375  // entries for BB and PBI's BB. If so, insert a select to make
3376  // them agree.
3377  for (PHINode &PN : CommonDest->phis()) {
3378  Value *BIV = PN.getIncomingValueForBlock(BB);
3379  unsigned PBBIdx = PN.getBasicBlockIndex(PBI->getParent());
3380  Value *PBIV = PN.getIncomingValue(PBBIdx);
3381  if (BIV != PBIV) {
3382  // Insert a select in PBI to pick the right value.
3383  SelectInst *NV = cast<SelectInst>(
3384  Builder.CreateSelect(PBICond, PBIV, BIV, PBIV->getName() + ".mux"));
3385  PN.setIncomingValue(PBBIdx, NV);
3386  // Although the select has the same condition as PBI, the original branch
3387  // weights for PBI do not apply to the new select because the select's
3388  // 'logical' edges are incoming edges of the phi that is eliminated, not
3389  // the outgoing edges of PBI.
3390  if (HasWeights) {
3391  uint64_t PredCommon = PBIOp ? PredFalseWeight : PredTrueWeight;
3392  uint64_t PredOther = PBIOp ? PredTrueWeight : PredFalseWeight;
3393  uint64_t SuccCommon = BIOp ? SuccFalseWeight : SuccTrueWeight;
3394  uint64_t SuccOther = BIOp ? SuccTrueWeight : SuccFalseWeight;
3395  // The weight to PredCommonDest should be PredCommon * SuccTotal.
3396  // The weight to PredOtherDest should be PredOther * SuccCommon.
3397  uint64_t NewWeights[2] = {PredCommon * (SuccCommon + SuccOther),
3398  PredOther * SuccCommon};
3399 
3400  FitWeights(NewWeights);
3401 
3402  setBranchWeights(NV, NewWeights[0], NewWeights[1]);
3403  }
3404  }
3405  }
3406 
3407  LLVM_DEBUG(dbgs() << "INTO: " << *PBI->getParent());
3408  LLVM_DEBUG(dbgs() << *PBI->getParent()->getParent());
3409 
3410  // This basic block is probably dead. We know it has at least
3411  // one fewer predecessor.
3412  return true;
3413 }
3414 
3415 // Simplifies a terminator by replacing it with a branch to TrueBB if Cond is
3416 // true or to FalseBB if Cond is false.
3417 // Takes care of updating the successors and removing the old terminator.
3418 // Also makes sure not to introduce new successors by assuming that edges to
3419 // non-successor TrueBBs and FalseBBs aren't reachable.
3421  BasicBlock *TrueBB, BasicBlock *FalseBB,
3422  uint32_t TrueWeight,
3423  uint32_t FalseWeight) {
3424  // Remove any superfluous successor edges from the CFG.
3425  // First, figure out which successors to preserve.
3426  // If TrueBB and FalseBB are equal, only try to preserve one copy of that
3427  // successor.
3428  BasicBlock *KeepEdge1 = TrueBB;
3429  BasicBlock *KeepEdge2 = TrueBB != FalseBB ? FalseBB : nullptr;
3430 
3431  // Then remove the rest.
3432  for (BasicBlock *Succ : successors(OldTerm)) {
3433  // Make sure only to keep exactly one copy of each edge.
3434  if (Succ == KeepEdge1)
3435  KeepEdge1 = nullptr;
3436  else if (Succ == KeepEdge2)
3437  KeepEdge2 = nullptr;
3438  else
3439  Succ->removePredecessor(OldTerm->getParent(),
3440  /*DontDeleteUselessPHIs=*/true);
3441  }
3442 
3443  IRBuilder<> Builder(OldTerm);
3444  Builder.SetCurrentDebugLocation(OldTerm->getDebugLoc());
3445 
3446  // Insert an appropriate new terminator.
3447  if (!KeepEdge1 && !KeepEdge2) {
3448  if (TrueBB == FalseBB)
3449  // We were only looking for one successor, and it was present.
3450  // Create an unconditional branch to it.
3451  Builder.CreateBr(TrueBB);
3452  else {
3453  // We found both of the successors we were looking for.
3454  // Create a conditional branch sharing the condition of the select.
3455  BranchInst *NewBI = Builder.CreateCondBr(Cond, TrueBB, FalseBB);
3456  if (TrueWeight != FalseWeight)
3457  setBranchWeights(NewBI, TrueWeight, FalseWeight);
3458  }
3459  } else if (KeepEdge1 && (KeepEdge2 || TrueBB == FalseBB)) {
3460  // Neither of the selected blocks were successors, so this
3461  // terminator must be unreachable.
3462  new UnreachableInst(OldTerm->getContext(), OldTerm);
3463  } else {
3464  // One of the selected values was a successor, but the other wasn't.
3465  // Insert an unconditional branch to the one that was found;
3466  // the edge to the one that wasn't must be unreachable.
3467  if (!KeepEdge1)
3468  // Only TrueBB was found.
3469  Builder.CreateBr(TrueBB);
3470  else
3471  // Only FalseBB was found.
3472  Builder.CreateBr(FalseBB);
3473  }
3474 
3476  return true;
3477 }
3478 
3479 // Replaces
3480 // (switch (select cond, X, Y)) on constant X, Y
3481 // with a branch - conditional if X and Y lead to distinct BBs,
3482 // unconditional otherwise.
3484  // Check for constant integer values in the select.
3485  ConstantInt *TrueVal = dyn_cast<ConstantInt>(Select->getTrueValue());
3486  ConstantInt *FalseVal = dyn_cast<ConstantInt>(Select->getFalseValue());
3487  if (!TrueVal || !FalseVal)
3488  return false;
3489 
3490  // Find the relevant condition and destinations.
3491  Value *Condition = Select->getCondition();
3492  BasicBlock *TrueBB = SI->findCaseValue(TrueVal)->getCaseSuccessor();
3493  BasicBlock *FalseBB = SI->findCaseValue(FalseVal)->getCaseSuccessor();
3494 
3495  // Get weight for TrueBB and FalseBB.
3496  uint32_t TrueWeight = 0, FalseWeight = 0;
3497  SmallVector<uint64_t, 8> Weights;
3498  bool HasWeights = HasBranchWeights(SI);
3499  if (HasWeights) {
3500  GetBranchWeights(SI, Weights);
3501  if (Weights.size() == 1 + SI->getNumCases()) {
3502  TrueWeight =
3503  (uint32_t)Weights[SI->findCaseValue(TrueVal)->getSuccessorIndex()];
3504  FalseWeight =
3505  (uint32_t)Weights[SI->findCaseValue(FalseVal)->getSuccessorIndex()];
3506  }
3507  }
3508 
3509  // Perform the actual simplification.
3510  return SimplifyTerminatorOnSelect(SI, Condition, TrueBB, FalseBB, TrueWeight,
3511  FalseWeight);
3512 }
3513 
3514 // Replaces
3515 // (indirectbr (select cond, blockaddress(@fn, BlockA),
3516 // blockaddress(@fn, BlockB)))
3517 // with
3518 // (br cond, BlockA, BlockB).
3520  // Check that both operands of the select are block addresses.
3523  if (!TBA || !FBA)
3524  return false;
3525 
3526  // Extract the actual blocks.
3527  BasicBlock *TrueBB = TBA->getBasicBlock();
3528  BasicBlock *FalseBB = FBA->getBasicBlock();
3529 
3530  // Perform the actual simplification.
3531  return SimplifyTerminatorOnSelect(IBI, SI->getCondition(), TrueBB, FalseBB, 0,
3532  0);
3533 }
3534 
3535 /// This is called when we find an icmp instruction
3536 /// (a seteq/setne with a constant) as the only instruction in a
3537 /// block that ends with an uncond branch. We are looking for a very specific
3538 /// pattern that occurs when "A == 1 || A == 2 || A == 3" gets simplified. In
3539 /// this case, we merge the first two "or's of icmp" into a switch, but then the
3540 /// default value goes to an uncond block with a seteq in it, we get something
3541 /// like:
3542 ///
3543 /// switch i8 %A, label %DEFAULT [ i8 1, label %end i8 2, label %end ]
3544 /// DEFAULT:
3545 /// %tmp = icmp eq i8 %A, 92
3546 /// br label %end
3547 /// end:
3548 /// ... = phi i1 [ true, %entry ], [ %tmp, %DEFAULT ], [ true, %entry ]
3549 ///
3550 /// We prefer to split the edge to 'end' so that there is a true/false entry to
3551 /// the PHI, merging the third icmp into the switch.
3552 bool SimplifyCFGOpt::tryToSimplifyUncondBranchWithICmpInIt(
3553  ICmpInst *ICI, IRBuilder<> &Builder) {
3554  BasicBlock *BB = ICI->getParent();
3555 
3556  // If the block has any PHIs in it or the icmp has multiple uses, it is too
3557  // complex.
3558  if (isa<PHINode>(BB->begin()) || !ICI->hasOneUse())
3559  return false;
3560 
3561  Value *V = ICI->getOperand(0);
3562  ConstantInt *Cst = cast<ConstantInt>(ICI->getOperand(1));
3563 
3564  // The pattern we're looking for is where our only predecessor is a switch on
3565  // 'V' and this block is the default case for the switch. In this case we can
3566  // fold the compared value into the switch to simplify things.
3567  BasicBlock *Pred = BB->getSinglePredecessor();
3568  if (!Pred || !isa<SwitchInst>(Pred->getTerminator()))
3569  return false;
3570 
3571  SwitchInst *SI = cast<SwitchInst>(Pred->getTerminator());
3572  if (SI->getCondition() != V)
3573  return false;
3574 
3575  // If BB is reachable on a non-default case, then we simply know the value of
3576  // V in this block. Substitute it and constant fold the icmp instruction
3577  // away.
3578  if (SI->getDefaultDest() != BB) {
3579  ConstantInt *VVal = SI->findCaseDest(BB);
3580  assert(VVal && "Should have a unique destination value");
3581  ICI->setOperand(0, VVal);
3582 
3583  if (Value *V = SimplifyInstruction(ICI, {DL, ICI})) {
3584  ICI->replaceAllUsesWith(V);
3585  ICI->eraseFromParent();
3586  }
3587  // BB is now empty, so it is likely to simplify away.
3588  return requestResimplify();
3589  }
3590 
3591  // Ok, the block is reachable from the default dest. If the constant we're
3592  // comparing exists in one of the other edges, then we can constant fold ICI
3593  // and zap it.
3594  if (SI->findCaseValue(Cst) != SI->case_default()) {
3595  Value *V;
3596  if (ICI->getPredicate() == ICmpInst::ICMP_EQ)
3597  V = ConstantInt::getFalse(BB->getContext());
3598  else
3599  V = ConstantInt::getTrue(BB->getContext());
3600 
3601  ICI->replaceAllUsesWith(V);
3602  ICI->eraseFromParent();
3603  // BB is now empty, so it is likely to simplify away.
3604  return requestResimplify();
3605  }
3606 
3607  // The use of the icmp has to be in the 'end' block, by the only PHI node in
3608  // the block.
3609  BasicBlock *SuccBlock = BB->getTerminator()->getSuccessor(0);
3610  PHINode *PHIUse = dyn_cast<PHINode>(ICI->user_back());
3611  if (PHIUse == nullptr || PHIUse != &SuccBlock->front() ||
3612  isa<PHINode>(++BasicBlock::iterator(PHIUse)))
3613  return false;
3614 
3615  // If the icmp is a SETEQ, then the default dest gets false, the new edge gets
3616  // true in the PHI.
3617  Constant *DefaultCst = ConstantInt::getTrue(BB->getContext());
3618  Constant *NewCst = ConstantInt::getFalse(BB->getContext());
3619 
3620  if (ICI->getPredicate() == ICmpInst::ICMP_EQ)
3621  std::swap(DefaultCst, NewCst);
3622 
3623  // Replace ICI (which is used by the PHI for the default value) with true or
3624  // false depending on if it is EQ or NE.
3625  ICI->replaceAllUsesWith(DefaultCst);
3626  ICI->eraseFromParent();
3627 
3628  // Okay, the switch goes to this block on a default value. Add an edge from
3629  // the switch to the merge point on the compared value.
3630  BasicBlock *NewBB =
3631  BasicBlock::Create(BB->getContext(), "switch.edge", BB->getParent(), BB);
3632  SmallVector<uint64_t, 8> Weights;
3633  bool HasWeights = HasBranchWeights(SI);
3634  if (HasWeights) {
3635  GetBranchWeights(SI, Weights);
3636  if (Weights.size() == 1 + SI->getNumCases()) {
3637  // Split weight for default case to case for "Cst".
3638  Weights[0] = (Weights[0] + 1) >> 1;
3639  Weights.push_back(Weights[0]);
3640 
3641  SmallVector<uint32_t, 8> MDWeights(Weights.begin(), Weights.end());
3642  setBranchWeights(SI, MDWeights);
3643  }
3644  }
3645  SI->addCase(Cst, NewBB);
3646 
3647  // NewBB branches to the phi block, add the uncond branch and the phi entry.
3648  Builder.SetInsertPoint(NewBB);
3649  Builder.SetCurrentDebugLocation(SI->getDebugLoc());
3650  Builder.CreateBr(SuccBlock);
3651  PHIUse->addIncoming(NewCst, NewBB);
3652  return true;
3653 }
3654 
3655 /// The specified branch is a conditional branch.
3656 /// Check to see if it is branching on an or/and chain of icmp instructions, and
3657 /// fold it into a switch instruction if so.
3659  const DataLayout &DL) {
3660  Instruction *Cond = dyn_cast<Instruction>(BI->getCondition());
3661  if (!Cond)
3662  return false;
3663 
3664  // Change br (X == 0 | X == 1), T, F into a switch instruction.
3665  // If this is a bunch of seteq's or'd together, or if it's a bunch of
3666  // 'setne's and'ed together, collect them.
3667 
3668  // Try to gather values from a chain of and/or to be turned into a switch
3669  ConstantComparesGatherer ConstantCompare(Cond, DL);
3670  // Unpack the result
3671  SmallVectorImpl<ConstantInt *> &Values = ConstantCompare.Vals;
3672  Value *CompVal = ConstantCompare.CompValue;
3673  unsigned UsedICmps = ConstantCompare.UsedICmps;
3674  Value *ExtraCase = ConstantCompare.Extra;
3675 
3676  // If we didn't have a multiply compared value, fail.
3677  if (!CompVal)
3678  return false;
3679 
3680  // Avoid turning single icmps into a switch.
3681  if (UsedICmps <= 1)
3682  return false;
3683 
3684  bool TrueWhenEqual = (Cond->getOpcode() == Instruction::Or);
3685 
3686  // There might be duplicate constants in the list, which the switch
3687  // instruction can't handle, remove them now.
3688  array_pod_sort(Values.begin(), Values.end(), ConstantIntSortPredicate);
3689  Values.erase(std::unique(Values.begin(), Values.end()), Values.end());
3690 
3691  // If Extra was used, we require at least two switch values to do the
3692  // transformation. A switch with one value is just a conditional branch.
3693  if (ExtraCase && Values.size() < 2)
3694  return false;
3695 
3696  // TODO: Preserve branch weight metadata, similarly to how
3697  // FoldValueComparisonIntoPredecessors preserves it.
3698 
3699  // Figure out which block is which destination.
3700  BasicBlock *DefaultBB = BI->getSuccessor(1);
3701  BasicBlock *EdgeBB = BI->getSuccessor(0);
3702  if (!TrueWhenEqual)
3703  std::swap(DefaultBB, EdgeBB);
3704 
3705  BasicBlock *BB = BI->getParent();
3706 
3707  LLVM_DEBUG(dbgs() << "Converting 'icmp' chain with " << Values.size()
3708  << " cases into SWITCH. BB is:\n"
3709  << *BB);
3710 
3711  // If there are any extra values that couldn't be folded into the switch
3712  // then we evaluate them with an explicit branch first. Split the block
3713  // right before the condbr to handle it.
3714  if (ExtraCase) {
3715  BasicBlock *NewBB =
3716  BB->splitBasicBlock(BI->getIterator(), "switch.early.test");
3717  // Remove the uncond branch added to the old block.
3718  TerminatorInst *OldTI = BB->getTerminator();
3719  Builder.SetInsertPoint(OldTI);
3720 
3721  if (TrueWhenEqual)
3722  Builder.CreateCondBr(ExtraCase, EdgeBB, NewBB);
3723  else
3724  Builder.CreateCondBr(ExtraCase, NewBB, EdgeBB);
3725 
3726  OldTI->eraseFromParent();
3727 
3728  // If there are PHI nodes in EdgeBB, then we need to add a new entry to them
3729  // for the edge we just added.
3730  AddPredecessorToBlock(EdgeBB, BB, NewBB);
3731 
3732  LLVM_DEBUG(dbgs() << " ** 'icmp' chain unhandled condition: " << *ExtraCase
3733  << "\nEXTRABB = " << *BB);
3734  BB = NewBB;
3735  }
3736 
3737  Builder.SetInsertPoint(BI);
3738  // Convert pointer to int before we switch.
3739  if (CompVal->getType()->isPointerTy()) {
3740  CompVal = Builder.CreatePtrToInt(
3741  CompVal, DL.getIntPtrType(CompVal->getType()), "magicptr");
3742  }
3743 
3744  // Create the new switch instruction now.
3745  SwitchInst *New = Builder.CreateSwitch(CompVal, DefaultBB, Values.size());
3746 
3747  // Add all of the 'cases' to the switch instruction.
3748  for (unsigned i = 0, e = Values.size(); i != e; ++i)
3749  New->addCase(Values[i], EdgeBB);
3750 
3751  // We added edges from PI to the EdgeBB. As such, if there were any
3752  // PHI nodes in EdgeBB, they need entries to be added corresponding to
3753  // the number of edges added.
3754  for (BasicBlock::iterator BBI = EdgeBB->begin(); isa<PHINode>(BBI); ++BBI) {
3755  PHINode *PN = cast<PHINode>(BBI);
3756  Value *InVal = PN->getIncomingValueForBlock(BB);
3757  for (unsigned i = 0, e = Values.size() - 1; i != e; ++i)
3758  PN->addIncoming(InVal, BB);
3759  }
3760 
3761  // Erase the old branch instruction.
3763 
3764  LLVM_DEBUG(dbgs() << " ** 'icmp' chain result is:\n" << *BB << '\n');
3765  return true;
3766 }
3767 
3768 bool SimplifyCFGOpt::SimplifyResume(ResumeInst *RI, IRBuilder<> &Builder) {
3769  if (isa<PHINode>(RI->getValue()))
3770  return SimplifyCommonResume(RI);
3771  else if (isa<LandingPadInst>(RI->getParent()->getFirstNonPHI()) &&
3772  RI->getValue() == RI->getParent()->getFirstNonPHI())
3773  // The resume must unwind the exception that caused control to branch here.
3774  return SimplifySingleResume(RI);
3775 
3776  return false;
3777 }
3778 
3779 // Simplify resume that is shared by several landing pads (phi of landing pad).
3780 bool SimplifyCFGOpt::SimplifyCommonResume(ResumeInst *RI) {
3781  BasicBlock *BB = RI->getParent();
3782 
3783  // Check that there are no other instructions except for debug intrinsics
3784  // between the phi of landing pads (RI->getValue()) and resume instruction.
3785  BasicBlock::iterator I = cast<Instruction>(RI->getValue())->getIterator(),
3786  E = RI->getIterator();
3787  while (++I != E)
3788  if (!isa<DbgInfoIntrinsic>(I))
3789  return false;
3790 
3791  SmallSetVector<BasicBlock *, 4> TrivialUnwindBlocks;
3792  auto *PhiLPInst = cast<PHINode>(RI->getValue());
3793 
3794  // Check incoming blocks to see if any of them are trivial.
3795  for (unsigned Idx = 0, End = PhiLPInst->getNumIncomingValues(); Idx != End;
3796  Idx++) {
3797  auto *IncomingBB = PhiLPInst->getIncomingBlock(Idx);
3798  auto *IncomingValue = PhiLPInst->getIncomingValue(Idx);
3799 
3800  // If the block has other successors, we can not delete it because
3801  // it has other dependents.
3802  if (IncomingBB->getUniqueSuccessor() != BB)
3803  continue;
3804 
3805  auto *LandingPad = dyn_cast<LandingPadInst>(IncomingBB->getFirstNonPHI());
3806  // Not the landing pad that caused the control to branch here.
3807  if (IncomingValue != LandingPad)
3808  continue;
3809 
3810  bool isTrivial = true;
3811 
3812  I = IncomingBB->getFirstNonPHI()->getIterator();
3813  E = IncomingBB->getTerminator()->getIterator();
3814  while (++I != E)
3815  if (!isa<DbgInfoIntrinsic>(I)) {
3816  isTrivial = false;
3817  break;
3818  }
3819 
3820  if (isTrivial)
3821  TrivialUnwindBlocks.insert(IncomingBB);
3822  }
3823 
3824  // If no trivial unwind blocks, don't do any simplifications.
3825  if (TrivialUnwindBlocks.empty())
3826  return false;
3827 
3828  // Turn all invokes that unwind here into calls.
3829  for (auto *TrivialBB : TrivialUnwindBlocks) {
3830  // Blocks that will be simplified should be removed from the phi node.
3831  // Note there could be multiple edges to the resume block, and we need
3832  // to remove them all.
3833  while (PhiLPInst->getBasicBlockIndex(TrivialBB) != -1)
3834  BB->removePredecessor(TrivialBB, true);
3835 
3836  for (pred_iterator PI = pred_begin(TrivialBB), PE = pred_end(TrivialBB);
3837  PI != PE;) {
3838  BasicBlock *Pred = *PI++;
3839  removeUnwindEdge(Pred);
3840  }
3841 
3842  // In each SimplifyCFG run, only the current processed block can be erased.
3843  // Otherwise, it will break the iteration of SimplifyCFG pass. So instead
3844  // of erasing TrivialBB, we only remove the branch to the common resume
3845  // block so that we can later erase the resume block since it has no
3846  // predecessors.
3847  TrivialBB->getTerminator()->eraseFromParent();
3848  new UnreachableInst(RI->getContext(), TrivialBB);
3849  }
3850 
3851  // Delete the resume block if all its predecessors have been removed.
3852  if (pred_empty(BB))
3853  BB->eraseFromParent();
3854 
3855  return !TrivialUnwindBlocks.empty();
3856 }
3857 
3858 // Simplify resume that is only used by a single (non-phi) landing pad.
3859 bool SimplifyCFGOpt::SimplifySingleResume(ResumeInst *RI) {
3860  BasicBlock *BB = RI->getParent();
3862  assert(RI->getValue() == LPInst &&
3863  "Resume must unwind the exception that caused control to here");
3864 
3865  // Check that there are no other instructions except for debug intrinsics.
3866  BasicBlock::iterator I = LPInst->getIterator(), E = RI->getIterator();
3867  while (++I != E)
3868  if (!isa<DbgInfoIntrinsic>(I))
3869  return false;
3870 
3871  // Turn all invokes that unwind here into calls and delete the basic block.
3872  for (pred_iterator PI = pred_begin(BB), PE = pred_end(BB); PI != PE;) {
3873  BasicBlock *Pred = *PI++;
3874  removeUnwindEdge(Pred);
3875  }
3876 
3877  // The landingpad is now unreachable. Zap it.
3878  if (LoopHeaders)
3879  LoopHeaders->erase(BB);
3880  BB->eraseFromParent();
3881  return true;
3882 }
3883 
3885  // If this is a trivial cleanup pad that executes no instructions, it can be
3886  // eliminated. If the cleanup pad continues to the caller, any predecessor
3887  // that is an EH pad will be updated to continue to the caller and any
3888  // predecessor that terminates with an invoke instruction will have its invoke
3889  // instruction converted to a call instruction. If the cleanup pad being
3890  // simplified does not continue to the caller, each predecessor will be
3891  // updated to continue to the unwind destination of the cleanup pad being
3892  // simplified.
3893  BasicBlock *BB = RI->getParent();
3894  CleanupPadInst *CPInst = RI->getCleanupPad();
3895  if (CPInst->getParent() != BB)
3896  // This isn't an empty cleanup.
3897  return false;
3898 
3899  // We cannot kill the pad if it has multiple uses. This typically arises
3900  // from unreachable basic blocks.
3901  if (!CPInst->hasOneUse())
3902  return false;
3903 
3904  // Check that there are no other instructions except for benign intrinsics.
3905  BasicBlock::iterator I = CPInst->getIterator(), E = RI->getIterator();
3906  while (++I != E) {
3907  auto *II = dyn_cast<IntrinsicInst>(I);
3908  if (!II)
3909  return false;
3910 
3911  Intrinsic::ID IntrinsicID = II->getIntrinsicID();
3912  switch (IntrinsicID) {
3913  case Intrinsic::dbg_declare:
3914  case Intrinsic::dbg_value:
3915  case Intrinsic::dbg_label:
3916  case Intrinsic::lifetime_end:
3917  break;
3918  default:
3919  return false;
3920  }
3921  }
3922 
3923  // If the cleanup return we are simplifying unwinds to the caller, this will
3924  // set UnwindDest to nullptr.
3925  BasicBlock *UnwindDest = RI->getUnwindDest();
3926  Instruction *DestEHPad = UnwindDest ? UnwindDest->getFirstNonPHI() : nullptr;
3927 
3928  // We're about to remove BB from the control flow. Before we do, sink any
3929  // PHINodes into the unwind destination. Doing this before changing the
3930  // control flow avoids some potentially slow checks, since we can currently
3931  // be certain that UnwindDest and BB have no common predecessors (since they
3932  // are both EH pads).
3933  if (UnwindDest) {
3934  // First, go through the PHI nodes in UnwindDest and update any nodes that
3935  // reference the block we are removing
3936  for (BasicBlock::iterator I = UnwindDest->begin(),
3937  IE = DestEHPad->getIterator();
3938  I != IE; ++I) {
3939  PHINode *DestPN = cast<PHINode>(I);
3940 
3941  int Idx = DestPN->getBasicBlockIndex(BB);
3942  // Since BB unwinds to UnwindDest, it has to be in the PHI node.
3943  assert(Idx != -1);
3944  // This PHI node has an incoming value that corresponds to a control
3945  // path through the cleanup pad we are removing. If the incoming
3946  // value is in the cleanup pad, it must be a PHINode (because we
3947  // verified above that the block is otherwise empty). Otherwise, the
3948  // value is either a constant or a value that dominates the cleanup
3949  // pad being removed.
3950  //
3951  // Because BB and UnwindDest are both EH pads, all of their
3952  // predecessors must unwind to these blocks, and since no instruction
3953  // can have multiple unwind destinations, there will be no overlap in
3954  // incoming blocks between SrcPN and DestPN.
3955  Value *SrcVal = DestPN->getIncomingValue(Idx);
3956  PHINode *SrcPN = dyn_cast<PHINode>(SrcVal);
3957 
3958  // Remove the entry for the block we are deleting.
3959  DestPN->removeIncomingValue(Idx, false);
3960 
3961  if (SrcPN && SrcPN->getParent() == BB) {
3962  // If the incoming value was a PHI node in the cleanup pad we are
3963  // removing, we need to merge that PHI node's incoming values into
3964  // DestPN.
3965  for (unsigned SrcIdx = 0, SrcE = SrcPN->getNumIncomingValues();
3966  SrcIdx != SrcE; ++SrcIdx) {
3967  DestPN->addIncoming(SrcPN->getIncomingValue(SrcIdx),
3968  SrcPN->getIncomingBlock(SrcIdx));
3969  }
3970  } else {
3971  // Otherwise, the incoming value came from above BB and
3972  // so we can just reuse it. We must associate all of BB's
3973  // predecessors with this value.
3974  for (auto *pred : predecessors(BB)) {
3975  DestPN->addIncoming(SrcVal, pred);
3976  }
3977  }
3978  }
3979 
3980  // Sink any remaining PHI nodes directly into UnwindDest.
3981  Instruction *InsertPt = DestEHPad;
3982  for (BasicBlock::iterator I = BB->begin(),
3983  IE = BB->getFirstNonPHI()->getIterator();
3984  I != IE;) {
3985  // The iterator must be incremented here because the instructions are
3986  // being moved to another block.
3987  PHINode *PN = cast<PHINode>(I++);
3988  if (PN->use_empty())
3989  // If the PHI node has no uses, just leave it. It will be erased
3990  // when we erase BB below.
3991  continue;
3992 
3993  // Otherwise, sink this PHI node into UnwindDest.
3994  // Any predecessors to UnwindDest which are not already represented
3995  // must be back edges which inherit the value from the path through
3996  // BB. In this case, the PHI value must reference itself.
3997  for (auto *pred : predecessors(UnwindDest))
3998  if (pred != BB)
3999  PN->addIncoming(PN, pred);
4000  PN->moveBefore(InsertPt);
4001  }
4002  }
4003 
4004  for (pred_iterator PI = pred_begin(BB), PE = pred_end(BB); PI != PE;) {
4005  // The iterator must be updated here because we are removing this pred.
4006  BasicBlock *PredBB = *PI++;
4007  if (UnwindDest == nullptr) {
4008  removeUnwindEdge(PredBB);
4009  } else {
4010  TerminatorInst *TI = PredBB->getTerminator();
4011  TI->replaceUsesOfWith(BB, UnwindDest);
4012  }
4013  }
4014 
4015  // The cleanup pad is now unreachable. Zap it.
4016  BB->eraseFromParent();
4017  return true;
4018 }
4019 
4020 // Try to merge two cleanuppads together.
4022  // Skip any cleanuprets which unwind to caller, there is nothing to merge
4023  // with.
4024  BasicBlock *UnwindDest = RI->getUnwindDest();
4025  if (!UnwindDest)
4026  return false;
4027 
4028  // This cleanupret isn't the only predecessor of this cleanuppad, it wouldn't
4029  // be safe to merge without code duplication.
4030  if (UnwindDest->getSinglePredecessor() != RI->getParent())
4031  return false;
4032 
4033  // Verify that our cleanuppad's unwind destination is another cleanuppad.
4034  auto *SuccessorCleanupPad = dyn_cast<CleanupPadInst>(&UnwindDest->front());
4035  if (!SuccessorCleanupPad)
4036  return false;
4037 
4038  CleanupPadInst *PredecessorCleanupPad = RI->getCleanupPad();
4039  // Replace any uses of the successor cleanupad with the predecessor pad
4040  // The only cleanuppad uses should be this cleanupret, it's cleanupret and
4041  // funclet bundle operands.
4042  SuccessorCleanupPad->replaceAllUsesWith(PredecessorCleanupPad);
4043  // Remove the old cleanuppad.
4044  SuccessorCleanupPad->eraseFromParent();
4045  // Now, we simply replace the cleanupret with a branch to the unwind
4046  // destination.
4047  BranchInst::Create(UnwindDest, RI->getParent());
4048  RI->eraseFromParent();
4049 
4050  return true;
4051 }
4052 
4053 bool SimplifyCFGOpt::SimplifyCleanupReturn(CleanupReturnInst *RI) {
4054  // It is possible to transiantly have an undef cleanuppad operand because we
4055  // have deleted some, but not all, dead blocks.
4056  // Eventually, this block will be deleted.
4057  if (isa<UndefValue>(RI->getOperand(0)))
4058  return false;
4059 
4060  if (mergeCleanupPad(RI))
4061  return true;
4062 
4063  if (removeEmptyCleanup(RI))
4064  return true;
4065 
4066  return false;
4067 }
4068 
4069 bool SimplifyCFGOpt::SimplifyReturn(ReturnInst *RI, IRBuilder<> &Builder) {
4070  BasicBlock *BB = RI->getParent();
4071  if (!BB->getFirstNonPHIOrDbg()->isTerminator())
4072  return false;
4073 
4074  // Find predecessors that end with branches.
4075  SmallVector<BasicBlock *, 8> UncondBranchPreds;
4076  SmallVector<BranchInst *, 8> CondBranchPreds;
4077  for (pred_iterator PI = pred_begin(BB), E = pred_end(BB); PI != E; ++PI) {
4078  BasicBlock *P = *PI;
4079  TerminatorInst *PTI = P->getTerminator();
4080  if (BranchInst *BI = dyn_cast<BranchInst>(PTI)) {
4081  if (BI->isUnconditional())
4082  UncondBranchPreds.push_back(P);
4083  else
4084  CondBranchPreds.push_back(BI);
4085  }
4086  }
4087 
4088  // If we found some, do the transformation!
4089  if (!UncondBranchPreds.empty() && DupRet) {
4090  while (!UncondBranchPreds.empty()) {
4091  BasicBlock *Pred = UncondBranchPreds.pop_back_val();
4092  LLVM_DEBUG(dbgs() << "FOLDING: " << *BB
4093  << "INTO UNCOND BRANCH PRED: " << *Pred);
4094  (void)FoldReturnIntoUncondBranch(RI, BB, Pred);
4095  }
4096 
4097  // If we eliminated all predecessors of the block, delete the block now.
4098  if (pred_empty(BB)) {
4099  // We know there are no successors, so just nuke the block.
4100  if (LoopHeaders)
4101  LoopHeaders->erase(BB);
4102  BB->eraseFromParent();
4103  }
4104 
4105  return true;
4106  }
4107 
4108  // Check out all of the conditional branches going to this return
4109  // instruction. If any of them just select between returns, change the
4110  // branch itself into a select/return pair.
4111  while (!CondBranchPreds.empty()) {
4112  BranchInst *BI = CondBranchPreds.pop_back_val();
4113 
4114  // Check to see if the non-BB successor is also a return block.
4115  if (isa<ReturnInst>(BI->getSuccessor(0)->getTerminator()) &&
4116  isa<ReturnInst>(BI->getSuccessor(1)->getTerminator()) &&
4117  SimplifyCondBranchToTwoReturns(BI, Builder))
4118  return true;
4119  }
4120  return false;
4121 }
4122 
4123 bool SimplifyCFGOpt::SimplifyUnreachable(UnreachableInst *UI) {
4124  BasicBlock *BB = UI->getParent();
4125 
4126  bool Changed = false;
4127 
4128  // If there are any instructions immediately before the unreachable that can
4129  // be removed, do so.
4130  while (UI->getIterator() != BB->begin()) {
4131  BasicBlock::iterator BBI = UI->getIterator();
4132  --BBI;
4133  // Do not delete instructions that can have side effects which might cause
4134  // the unreachable to not be reachable; specifically, calls and volatile
4135  // operations may have this effect.
4136  if (isa<CallInst>(BBI) && !isa<DbgInfoIntrinsic>(BBI))
4137  break;
4138 
4139  if (BBI->mayHaveSideEffects()) {
4140  if (auto *SI = dyn_cast<StoreInst>(BBI)) {
4141  if (SI->isVolatile())
4142  break;
4143  } else if (auto *LI = dyn_cast<LoadInst>(BBI)) {
4144  if (LI->isVolatile())
4145  break;
4146  } else if (auto *RMWI = dyn_cast<AtomicRMWInst>(BBI)) {
4147  if (RMWI->isVolatile())
4148  break;
4149  } else if (auto *CXI = dyn_cast<AtomicCmpXchgInst>(BBI)) {
4150  if (CXI->isVolatile())
4151  break;
4152  } else if (isa<CatchPadInst>(BBI)) {
4153  // A catchpad may invoke exception object constructors and such, which
4154  // in some languages can be arbitrary code, so be conservative by
4155  // default.
4156  // For CoreCLR, it just involves a type test, so can be removed.
4159  break;
4160  } else if (!isa<FenceInst>(BBI) && !isa<VAArgInst>(BBI) &&
4161  !isa<LandingPadInst>(BBI)) {
4162  break;
4163  }
4164  // Note that deleting LandingPad's here is in fact okay, although it
4165  // involves a bit of subtle reasoning. If this inst is a LandingPad,
4166  // all the predecessors of this block will be the unwind edges of Invokes,
4167  // and we can therefore guarantee this block will be erased.
4168  }
4169 
4170  // Delete this instruction (any uses are guaranteed to be dead)
4171  if (!BBI->use_empty())
4172  BBI->replaceAllUsesWith(UndefValue::get(BBI->getType()));
4173  BBI->eraseFromParent();
4174  Changed = true;
4175  }
4176 
4177  // If the unreachable instruction is the first in the block, take a gander
4178  // at all of the predecessors of this instruction, and simplify them.
4179  if (&BB->front() != UI)
4180  return Changed;
4181 
4183  for (unsigned i = 0, e = Preds.size(); i != e; ++i) {
4184  TerminatorInst *TI = Preds[i]->getTerminator();
4185  IRBuilder<> Builder(TI);
4186  if (auto *BI = dyn_cast<BranchInst>(TI)) {
4187  if (BI->isUnconditional()) {
4188  if (BI->getSuccessor(0) == BB) {
4189  new UnreachableInst(TI->getContext(), TI);
4190  TI->eraseFromParent();
4191  Changed = true;
4192  }
4193  } else {
4194  if (BI->getSuccessor(0) == BB) {
4195  Builder.CreateBr(BI->getSuccessor(1));
4197  } else if (BI->getSuccessor(1) == BB) {
4198  Builder.CreateBr(BI->getSuccessor(0));
4200  Changed = true;
4201  }
4202  }
4203  } else if (auto *SI = dyn_cast<SwitchInst>(TI)) {
4204  for (auto i = SI->case_begin(), e = SI->case_end(); i != e;) {
4205  if (i->getCaseSuccessor() != BB) {
4206  ++i;
4207  continue;
4208  }
4209  BB->removePredecessor(SI->getParent());
4210  i = SI->removeCase(i);
4211  e = SI->case_end();
4212  Changed = true;
4213  }
4214  } else if (auto *II = dyn_cast<InvokeInst>(TI)) {
4215  if (II->getUnwindDest() == BB) {
4216  removeUnwindEdge(TI->getParent());
4217  Changed = true;
4218  }
4219  } else if (auto *CSI = dyn_cast<CatchSwitchInst>(TI)) {
4220  if (CSI->getUnwindDest() == BB) {
4221  removeUnwindEdge(TI->getParent());
4222  Changed = true;
4223  continue;
4224  }
4225 
4226  for (CatchSwitchInst::handler_iterator I = CSI->handler_begin(),
4227  E = CSI->handler_end();
4228  I != E; ++I) {
4229  if (*I == BB) {
4230  CSI->removeHandler(I);
4231  --I;
4232  --E;
4233  Changed = true;
4234  }
4235  }
4236  if (CSI->getNumHandlers() == 0) {
4237  BasicBlock *CatchSwitchBB = CSI->getParent();
4238  if (CSI->hasUnwindDest()) {
4239  // Redirect preds to the unwind dest
4240  CatchSwitchBB->replaceAllUsesWith(CSI->getUnwindDest());
4241  } else {
4242  // Rewrite all preds to unwind to caller (or from invoke to call).
4243  SmallVector<BasicBlock *, 8> EHPreds(predecessors(CatchSwitchBB));
4244  for (BasicBlock *EHPred : EHPreds)
4245  removeUnwindEdge(EHPred);
4246  }
4247  // The catchswitch is no longer reachable.
4248  new UnreachableInst(CSI->getContext(), CSI);
4249  CSI->eraseFromParent();
4250  Changed = true;
4251  }
4252  } else if (isa<CleanupReturnInst>(TI)) {
4253  new UnreachableInst(TI->getContext(), TI);
4254  TI->eraseFromParent();
4255  Changed = true;
4256  }
4257  }
4258 
4259  // If this block is now dead, remove it.
4260  if (pred_empty(BB) && BB != &BB->getParent()->getEntryBlock()) {
4261  // We know there are no successors, so just nuke the block.
4262  if (LoopHeaders)
4263  LoopHeaders->erase(BB);
4264  BB->eraseFromParent();
4265  return true;
4266  }
4267 
4268  return Changed;
4269 }
4270 
4272  assert(Cases.size() >= 1);
4273 
4274  array_pod_sort(Cases.begin(), Cases.end(), ConstantIntSortPredicate);
4275  for (size_t I = 1, E = Cases.size(); I != E; ++I) {
4276  if (Cases[I - 1]->getValue() != Cases[I]->getValue() + 1)
4277  return false;
4278  }
4279  return true;
4280 }
4281 
4282 /// Turn a switch with two reachable destinations into an integer range
4283 /// comparison and branch.
4285  assert(SI->getNumCases() > 1 && "Degenerate switch?");
4286 
4287  bool HasDefault =
4288  !isa<UnreachableInst>(SI->getDefaultDest()->getFirstNonPHIOrDbg());
4289 
4290  // Partition the cases into two sets with different destinations.
4291  BasicBlock *DestA = HasDefault ? SI->getDefaultDest() : nullptr;
4292  BasicBlock *DestB = nullptr;
4295 
4296  for (auto Case : SI->cases()) {
4297  BasicBlock *Dest = Case.getCaseSuccessor();
4298  if (!DestA)
4299  DestA = Dest;
4300  if (Dest == DestA) {
4301  CasesA.push_back(Case.getCaseValue());
4302  continue;
4303  }
4304  if (!DestB)
4305  DestB = Dest;
4306  if (Dest == DestB) {
4307  CasesB.push_back(Case.getCaseValue());
4308  continue;
4309  }
4310  return false; // More than two destinations.
4311  }
4312 
4313  assert(DestA && DestB &&
4314  "Single-destination switch should have been folded.");
4315  assert(DestA != DestB);
4316  assert(DestB != SI->getDefaultDest());
4317  assert(!CasesB.empty() && "There must be non-default cases.");
4318  assert(!CasesA.empty() || HasDefault);
4319 
4320  // Figure out if one of the sets of cases form a contiguous range.
4321  SmallVectorImpl<ConstantInt *> *ContiguousCases = nullptr;
4322  BasicBlock *ContiguousDest = nullptr;
4323  BasicBlock *OtherDest = nullptr;
4324  if (!CasesA.empty() && CasesAreContiguous(CasesA)) {
4325  ContiguousCases = &CasesA;
4326  ContiguousDest = DestA;
4327  OtherDest = DestB;
4328  } else if (CasesAreContiguous(CasesB)) {
4329  ContiguousCases = &CasesB;
4330  ContiguousDest = DestB;
4331  OtherDest = DestA;
4332  } else
4333  return false;
4334 
4335  // Start building the compare and branch.
4336 
4337  Constant *Offset = ConstantExpr::getNeg(ContiguousCases->back());
4338  Constant *NumCases =
4339  ConstantInt::get(Offset->getType(), ContiguousCases->size());
4340 
4341  Value *Sub = SI->getCondition();
4342  if (!Offset->isNullValue())
4343  Sub = Builder.CreateAdd(Sub, Offset, Sub->getName() + ".off");
4344 
4345  Value *Cmp;
4346  // If NumCases overflowed, then all possible values jump to the successor.
4347  if (NumCases->isNullValue() && !ContiguousCases->empty())
4348  Cmp = ConstantInt::getTrue(SI->getContext());
4349  else
4350  Cmp = Builder.CreateICmpULT(Sub, NumCases, "switch");
4351  BranchInst *NewBI = Builder.CreateCondBr(Cmp, ContiguousDest, OtherDest);
4352 
4353  // Update weight for the newly-created conditional branch.
4354  if (HasBranchWeights(SI)) {
4355  SmallVector<uint64_t, 8> Weights;
4356  GetBranchWeights(SI, Weights);
4357  if (Weights.size() == 1 + SI->getNumCases()) {
4358  uint64_t TrueWeight = 0;
4359  uint64_t FalseWeight = 0;
4360  for (size_t I = 0, E = Weights.size(); I != E; ++I) {
4361  if (SI->getSuccessor(I) == ContiguousDest)
4362  TrueWeight += Weights[I];
4363  else
4364  FalseWeight += Weights[I];
4365  }
4366  while (TrueWeight > UINT32_MAX || FalseWeight > UINT32_MAX) {
4367  TrueWeight /= 2;
4368  FalseWeight /= 2;
4369  }
4370  setBranchWeights(NewBI, TrueWeight, FalseWeight);
4371  }
4372  }
4373 
4374  // Prune obsolete incoming values off the successors' PHI nodes.
4375  for (auto BBI = ContiguousDest->begin(); isa<PHINode>(BBI); ++BBI) {
4376  unsigned PreviousEdges = ContiguousCases->size();
4377  if (ContiguousDest == SI->getDefaultDest())
4378  ++PreviousEdges;
4379  for (unsigned I = 0, E = PreviousEdges - 1; I != E; ++I)
4380  cast<PHINode>(BBI)->removeIncomingValue(SI->getParent());
4381  }
4382  for (auto BBI = OtherDest->begin(); isa<PHINode>(BBI); ++BBI) {
4383  unsigned PreviousEdges = SI->getNumCases() - ContiguousCases->size();
4384  if (OtherDest == SI->getDefaultDest())
4385  ++PreviousEdges;
4386  for (unsigned I = 0, E = PreviousEdges - 1; I != E; ++I)
4387  cast<PHINode>(BBI)->removeIncomingValue(SI->getParent());
4388  }
4389 
4390  // Drop the switch.
4391  SI->eraseFromParent();
4392 
4393  return true;
4394 }
4395 
4396 /// Compute masked bits for the condition of a switch
4397 /// and use it to remove dead cases.
4399  const DataLayout &DL) {
4400  Value *Cond = SI->getCondition();
4401  unsigned Bits = Cond->getType()->getIntegerBitWidth();
4402  KnownBits Known = computeKnownBits(Cond, DL, 0, AC, SI);
4403 
4404  // We can also eliminate cases by determining that their values are outside of
4405  // the limited range of the condition based on how many significant (non-sign)
4406  // bits are in the condition value.
4407  unsigned ExtraSignBits = ComputeNumSignBits(Cond, DL, 0, AC, SI) - 1;
4408  unsigned MaxSignificantBitsInCond = Bits - ExtraSignBits;
4409 
4410  // Gather dead cases.
4412  for (auto &Case : SI->cases()) {
4413  const APInt &CaseVal = Case.getCaseValue()->getValue();
4414  if (Known.Zero.intersects(CaseVal) || !Known.One.isSubsetOf(CaseVal) ||
4415  (CaseVal.getMinSignedBits() > MaxSignificantBitsInCond)) {
4416  DeadCases.push_back(Case.getCaseValue());
4417  LLVM_DEBUG(dbgs() << "SimplifyCFG: switch case " << CaseVal
4418  << " is dead.\n");
4419  }
4420  }
4421 
4422  // If we can prove that the cases must cover all possible values, the
4423  // default destination becomes dead and we can remove it. If we know some
4424  // of the bits in the value, we can use that to more precisely compute the
4425  // number of possible unique case values.
4426  bool HasDefault =
4427  !isa<UnreachableInst>(SI->getDefaultDest()->getFirstNonPHIOrDbg());
4428  const unsigned NumUnknownBits =
4429  Bits - (Known.Zero | Known.One).countPopulation();
4430  assert(NumUnknownBits <= Bits);
4431  if (HasDefault && DeadCases.empty() &&
4432  NumUnknownBits < 64 /* avoid overflow */ &&
4433  SI->getNumCases() == (1ULL << NumUnknownBits)) {
4434  LLVM_DEBUG(dbgs() << "SimplifyCFG: switch default is dead.\n");
4435  BasicBlock *NewDefault =
4437  SI->setDefaultDest(&*NewDefault);
4438  SplitBlock(&*NewDefault, &NewDefault->front());
4439  auto *OldTI = NewDefault->getTerminator();
4440  new UnreachableInst(SI->getContext(), OldTI);
4442  return true;
4443  }
4444 
4445  SmallVector<uint64_t, 8> Weights;
4446  bool HasWeight = HasBranchWeights(SI);
4447  if (HasWeight) {
4448  GetBranchWeights(SI, Weights);
4449  HasWeight = (Weights.size() == 1 + SI->getNumCases());
4450  }
4451 
4452  // Remove dead cases from the switch.
4453  for (ConstantInt *DeadCase : DeadCases) {
4454  SwitchInst::CaseIt CaseI = SI->findCaseValue(DeadCase);
4455  assert(CaseI != SI->case_default() &&
4456  "Case was not found. Probably mistake in DeadCases forming.");
4457  if (HasWeight) {
4458  std::swap(Weights[CaseI->getCaseIndex() + 1], Weights.back());
4459  Weights.pop_back();
4460  }
4461 
4462  // Prune unused values from PHI nodes.
4463  CaseI->getCaseSuccessor()->removePredecessor(SI->getParent());
4464  SI->removeCase(CaseI);
4465  }
4466  if (HasWeight && Weights.size() >= 2) {
4467  SmallVector<uint32_t, 8> MDWeights(Weights.begin(), Weights.end());
4468  setBranchWeights(SI, MDWeights);
4469  }
4470 
4471  return !DeadCases.empty();
4472 }
4473 
4474 /// If BB would be eligible for simplification by
4475 /// TryToSimplifyUncondBranchFromEmptyBlock (i.e. it is empty and terminated
4476 /// by an unconditional branch), look at the phi node for BB in the successor
4477 /// block and see if the incoming value is equal to CaseValue. If so, return
4478 /// the phi node, and set PhiIndex to BB's index in the phi node.
4480  BasicBlock *BB, int *PhiIndex) {
4481  if (BB->getFirstNonPHIOrDbg() != BB->getTerminator())
4482  return nullptr; // BB must be empty to be a candidate for simplification.
4483  if (!BB->getSinglePredecessor())
4484  return nullptr; // BB must be dominated by the switch.
4485 
4487  if (!Branch || !Branch->isUnconditional())
4488  return nullptr; // Terminator must be unconditional branch.
4489 
4490  BasicBlock *Succ = Branch->getSuccessor(0);
4491 
4492  for (PHINode &PHI : Succ->phis()) {
4493  int Idx = PHI.getBasicBlockIndex(BB);
4494  assert(Idx >= 0 && "PHI has no entry for predecessor?");
4495 
4496  Value *InValue = PHI.getIncomingValue(Idx);
4497  if (InValue != CaseValue)
4498  continue;
4499 
4500  *PhiIndex = Idx;
4501  return &PHI;
4502  }
4503 
4504  return nullptr;
4505 }
4506 
4507 /// Try to forward the condition of a switch instruction to a phi node
4508 /// dominated by the switch, if that would mean that some of the destination
4509 /// blocks of the switch can be folded away. Return true if a change is made.
4511  using ForwardingNodesMap = DenseMap<PHINode *, SmallVector<int, 4>>;
4512 
4513  ForwardingNodesMap ForwardingNodes;
4514  BasicBlock *SwitchBlock = SI->getParent();
4515  bool Changed = false;
4516  for (auto &Case : SI->cases()) {
4517  ConstantInt *CaseValue = Case.getCaseValue();
4518  BasicBlock *CaseDest = Case.getCaseSuccessor();
4519 
4520  // Replace phi operands in successor blocks that are using the constant case
4521  // value rather than the switch condition variable:
4522  // switchbb:
4523  // switch i32 %x, label %default [
4524  // i32 17, label %succ
4525  // ...
4526  // succ:
4527  // %r = phi i32 ... [ 17, %switchbb ] ...
4528  // -->
4529  // %r = phi i32 ... [ %x, %switchbb ] ...
4530 
4531  for (PHINode &Phi : CaseDest->phis()) {
4532  // This only works if there is exactly 1 incoming edge from the switch to
4533  // a phi. If there is >1, that means multiple cases of the switch map to 1
4534  // value in the phi, and that phi value is not the switch condition. Thus,
4535  // this transform would not make sense (the phi would be invalid because
4536  // a phi can't have different incoming values from the same block).
4537  int SwitchBBIdx = Phi.getBasicBlockIndex(SwitchBlock);
4538  if (Phi.getIncomingValue(SwitchBBIdx) == CaseValue &&
4539  count(Phi.blocks(), SwitchBlock) == 1) {
4540  Phi.setIncomingValue(SwitchBBIdx, SI->getCondition());
4541  Changed = true;
4542  }
4543  }
4544 
4545  // Collect phi nodes that are indirectly using this switch's case constants.
4546  int PhiIdx;
4547  if (auto *Phi = FindPHIForConditionForwarding(CaseValue, CaseDest, &PhiIdx))
4548  ForwardingNodes[Phi].push_back(PhiIdx);
4549  }
4550 
4551  for (auto &ForwardingNode : ForwardingNodes) {
4552  PHINode *Phi = ForwardingNode.first;
4553  SmallVectorImpl<int> &Indexes = ForwardingNode.second;
4554  if (Indexes.size() < 2)
4555  continue;
4556 
4557  for (int Index : Indexes)
4558  Phi->setIncomingValue(Index, SI->getCondition());
4559  Changed = true;
4560  }
4561 
4562  return Changed;
4563 }
4564 
4565 /// Return true if the backend will be able to handle
4566 /// initializing an array of constants like C.
4568  if (C->isThreadDependent())
4569  return false;
4570  if (C->isDLLImportDependent())
4571  return false;
4572 
4573  if (!isa<ConstantFP>(C) && !isa<ConstantInt>(C) &&
4574  !isa<ConstantPointerNull>(C) && !isa<GlobalValue>(C) &&
4575  !isa<UndefValue>(C) && !isa<ConstantExpr>(C))
4576  return false;
4577 
4578  if (ConstantExpr *CE = dyn_cast<ConstantExpr>(C)) {
4579  if (!CE->isGEPWithNoNotionalOverIndexing())
4580  return false;
4581  if (!ValidLookupTableConstant(CE->getOperand(0), TTI))
4582  return false;
4583  }
4584 
4586  return false;
4587 
4588  return true;
4589 }
4590 
4591 /// If V is a Constant, return it. Otherwise, try to look up
4592 /// its constant value in ConstantPool, returning 0 if it's not there.
4593 static Constant *
4596  if (Constant *C = dyn_cast<Constant>(V))
4597  return C;
4598  return ConstantPool.lookup(V);
4599 }
4600 
4601 /// Try to fold instruction I into a constant. This works for
4602 /// simple instructions such as binary operations where both operands are
4603 /// constant or can be replaced by constants from the ConstantPool. Returns the
4604 /// resulting constant on success, 0 otherwise.
4605 static Constant *
4608  if (SelectInst *Select = dyn_cast<SelectInst>(I)) {
4609  Constant *A = LookupConstant(Select->getCondition(), ConstantPool);
4610  if (!A)
4611  return nullptr;
4612  if (A->isAllOnesValue())
4613  return LookupConstant(Select->getTrueValue(), ConstantPool);
4614  if (A->isNullValue())
4615  return LookupConstant(Select->getFalseValue(), ConstantPool);
4616  return nullptr;
4617  }
4618 
4620  for (unsigned N = 0, E = I->getNumOperands(); N != E; ++N) {
4622  COps.push_back(A);
4623  else
4624  return nullptr;
4625  }
4626 
4627  if (CmpInst *Cmp = dyn_cast<CmpInst>(I)) {
4628  return ConstantFoldCompareInstOperands(Cmp->getPredicate(), COps[0],
4629  COps[1], DL);
4630  }
4631 
4632  return ConstantFoldInstOperands(I, COps, DL);
4633 }
4634 
4635 /// Try to determine the resulting constant values in phi nodes
4636 /// at the common destination basic block, *CommonDest, for one of the case
4637 /// destionations CaseDest corresponding to value CaseVal (0 for the default
4638 /// case), of a switch instruction SI.
4639 static bool
4641  BasicBlock **CommonDest,
4642  SmallVectorImpl<std::pair<PHINode *, Constant *>> &Res,
4643  const DataLayout &DL, const TargetTransformInfo &TTI) {
4644  // The block from which we enter the common destination.
4645  BasicBlock *Pred = SI->getParent();
4646 
4647  // If CaseDest is empty except for some side-effect free instructions through
4648  // which we can constant-propagate the CaseVal, continue to its successor.
4650  ConstantPool.insert(std::make_pair(SI->getCondition(), CaseVal));
4651  for (Instruction &I :CaseDest->instructionsWithoutDebug()) {
4652  if (TerminatorInst *T = dyn_cast<TerminatorInst>(&I)) {
4653  // If the terminator is a simple branch, continue to the next block.
4654  if (T->getNumSuccessors() != 1 || T->isExceptionalTerminator())
4655  return false;
4656  Pred = CaseDest;
4657  CaseDest = T->getSuccessor(0);
4658  } else if (Constant *C = ConstantFold(&I, DL, ConstantPool)) {
4659  // Instruction is side-effect free and constant.
4660 
4661  // If the instruction has uses outside this block or a phi node slot for
4662  // the block, it is not safe to bypass the instruction since it would then
4663  // no longer dominate all its uses.
4664  for (auto &Use : I.uses()) {
4665  User *User = Use.getUser();
4666  if (Instruction *I = dyn_cast<Instruction>(User))
4667  if (I->getParent() == CaseDest)
4668  continue;
4669  if (PHINode *Phi = dyn_cast<PHINode>(User))
4670  if (Phi->getIncomingBlock(Use) == CaseDest)
4671  continue;
4672  return false;
4673  }
4674 
4675  ConstantPool.insert(std::make_pair(&I, C));
4676  } else {
4677  break;
4678  }
4679  }
4680 
4681  // If we did not have a CommonDest before, use the current one.
4682  if (!*CommonDest)
4683  *CommonDest = CaseDest;
4684  // If the destination isn't the common one, abort.
4685  if (CaseDest != *CommonDest)
4686  return false;
4687 
4688  // Get the values for this case from phi nodes in the destination block.
4689  for (PHINode &PHI : (*CommonDest)->phis()) {
4690  int Idx = PHI.getBasicBlockIndex(Pred);
4691  if (Idx == -1)
4692  continue;
4693 
4694  Constant *ConstVal =
4695  LookupConstant(PHI.getIncomingValue(Idx), ConstantPool);
4696  if (!ConstVal)
4697  return false;
4698 
4699  // Be conservative about which kinds of constants we support.
4700  if (!ValidLookupTableConstant(ConstVal, TTI))
4701  return false;
4702 
4703  Res.push_back(std::make_pair(&PHI, ConstVal));
4704  }
4705 
4706  return Res.size() > 0;
4707 }
4708 
4709 // Helper function used to add CaseVal to the list of cases that generate
4710 // Result. Returns the updated number of cases that generate this result.
4711 static uintptr_t MapCaseToResult(ConstantInt *CaseVal,
4712  SwitchCaseResultVectorTy &UniqueResults,
4713  Constant *Result) {
4714  for (auto &I : UniqueResults) {
4715  if (I.first == Result) {
4716  I.second.push_back(CaseVal);
4717  return I.second.size();
4718  }
4719  }
4720  UniqueResults.push_back(
4721  std::make_pair(Result, SmallVector<ConstantInt *, 4>(1, CaseVal)));
4722  return 1;
4723 }
4724 
4725 // Helper function that initializes a map containing
4726 // results for the PHI node of the common destination block for a switch
4727 // instruction. Returns false if multiple PHI nodes have been found or if
4728 // there is not a common destination block for the switch.
4729 static bool
4731  SwitchCaseResultVectorTy &UniqueResults,
4732  Constant *&DefaultResult, const DataLayout &DL,
4733  const TargetTransformInfo &TTI,
4734  uintptr_t MaxUniqueResults, uintptr_t MaxCasesPerResult) {
4735  for (auto &I : SI->cases()) {
4736  ConstantInt *CaseVal = I.getCaseValue();
4737 
4738  // Resulting value at phi nodes for this case value.
4739  SwitchCaseResultsTy Results;
4740  if (!GetCaseResults(SI, CaseVal, I.getCaseSuccessor(), &CommonDest, Results,
4741  DL, TTI))
4742  return false;
4743 
4744  // Only one value per case is permitted.
4745  if (Results.size() > 1)
4746  return false;
4747 
4748  // Add the case->result mapping to UniqueResults.
4749  const uintptr_t NumCasesForResult =
4750  MapCaseToResult(CaseVal, UniqueResults, Results.begin()->second);
4751 
4752  // Early out if there are too many cases for this result.
4753  if (NumCasesForResult > MaxCasesPerResult)
4754  return false;
4755 
4756  // Early out if there are too many unique results.
4757  if (UniqueResults.size() > MaxUniqueResults)
4758  return false;
4759 
4760  // Check the PHI consistency.
4761  if (!PHI)
4762  PHI = Results[0].first;
4763  else if (PHI != Results[0].first)
4764  return false;
4765  }
4766  // Find the default result value.
4768  BasicBlock *DefaultDest = SI->getDefaultDest();
4769  GetCaseResults(SI, nullptr, SI->getDefaultDest(), &CommonDest, DefaultResults,
4770  DL, TTI);
4771  // If the default value is not found abort unless the default destination
4772  // is unreachable.
4773  DefaultResult =
4774  DefaultResults.size() == 1 ? DefaultResults.begin()->second : nullptr;
4775  if ((!DefaultResult &&
4776  !isa<UnreachableInst>(DefaultDest->getFirstNonPHIOrDbg())))
4777  return false;
4778 
4779  return true;
4780 }
4781 
4782 // Helper function that checks if it is possible to transform a switch with only
4783 // two cases (or two cases + default) that produces a result into a select.
4784 // Example:
4785 // switch (a) {
4786 // case 10: %0 = icmp eq i32 %a, 10
4787 // return 10; %1 = select i1 %0, i32 10, i32 4
4788 // case 20: ----> %2 = icmp eq i32 %a, 20
4789 // return 2; %3 = select i1 %2, i32 2, i32 %1
4790 // default:
4791 // return 4;
4792 // }
4793 static Value *ConvertTwoCaseSwitch(const SwitchCaseResultVectorTy &ResultVector,
4794  Constant *DefaultResult, Value *Condition,
4795  IRBuilder<> &Builder) {
4796  assert(ResultVector.size() == 2 &&
4797  "We should have exactly two unique results at this point");
4798  // If we are selecting between only two cases transform into a simple
4799  // select or a two-way select if default is possible.
4800  if (ResultVector[0].second.size() == 1 &&
4801  ResultVector[1].second.size() == 1) {
4802  ConstantInt *const FirstCase = ResultVector[0].second[0];
4803  ConstantInt *const SecondCase = ResultVector[1].second[0];
4804 
4805  bool DefaultCanTrigger = DefaultResult;
4806  Value *SelectValue = ResultVector[1].first;
4807  if (DefaultCanTrigger) {
4808  Value *const ValueCompare =
4809  Builder.CreateICmpEQ(Condition, SecondCase, "switch.selectcmp");
4810  SelectValue = Builder.CreateSelect(ValueCompare, ResultVector[1].first,
4811  DefaultResult, "switch.select");
4812  }
4813  Value *const ValueCompare =
4814  Builder.CreateICmpEQ(Condition, FirstCase, "switch.selectcmp");
4815  return Builder.CreateSelect(ValueCompare, ResultVector[0].first,
4816  SelectValue, "switch.select");
4817  }
4818 
4819  return nullptr;
4820 }
4821 
4822 // Helper function to cleanup a switch instruction that has been converted into
4823 // a select, fixing up PHI nodes and basic blocks.
4825  Value *SelectValue,
4826  IRBuilder<> &Builder) {
4827  BasicBlock *SelectBB = SI->getParent();
4828  while (PHI->getBasicBlockIndex(SelectBB) >= 0)
4829  PHI->removeIncomingValue(SelectBB);
4830  PHI->addIncoming(SelectValue, SelectBB);
4831 
4832  Builder.CreateBr(PHI->getParent());
4833 
4834  // Remove the switch.
4835  for (unsigned i = 0, e = SI->getNumSuccessors(); i < e; ++i) {
4836  BasicBlock *Succ = SI->getSuccessor(i);
4837 
4838  if (Succ == PHI->getParent())
4839  continue;
4840  Succ->removePredecessor(SelectBB);
4841  }
4842  SI->eraseFromParent();
4843 }
4844 
4845 /// If the switch is only used to initialize one or more
4846 /// phi nodes in a common successor block with only two different
4847 /// constant values, replace the switch with select.
4848 static bool switchToSelect(SwitchInst *SI, IRBuilder<> &Builder,
4849  const DataLayout &DL,
4850  const TargetTransformInfo &TTI) {
4851  Value *const Cond = SI->getCondition();
4852  PHINode *PHI = nullptr;
4853  BasicBlock *CommonDest = nullptr;
4854  Constant *DefaultResult;
4855  SwitchCaseResultVectorTy UniqueResults;
4856  // Collect all the cases that will deliver the same value from the switch.
4857  if (!InitializeUniqueCases(SI, PHI, CommonDest, UniqueResults, DefaultResult,
4858  DL, TTI, 2, 1))
4859  return false;
4860  // Selects choose between maximum two values.
4861  if (UniqueResults.size() != 2)
4862  return false;
4863  assert(PHI != nullptr && "PHI for value select not found");
4864 
4865  Builder.SetInsertPoint(SI);
4866  Value *SelectValue =
4867  ConvertTwoCaseSwitch(UniqueResults, DefaultResult, Cond, Builder);
4868  if (SelectValue) {
4869  RemoveSwitchAfterSelectConversion(SI, PHI, SelectValue, Builder);
4870  return true;
4871  }
4872  // The switch couldn't be converted into a select.
4873  return false;
4874 }
4875 
4876 namespace {
4877 
4878 /// This class represents a lookup table that can be used to replace a switch.
4879 class SwitchLookupTable {
4880 public:
4881  /// Create a lookup table to use as a switch replacement with the contents
4882  /// of Values, using DefaultValue to fill any holes in the table.
4883  SwitchLookupTable(
4884  Module &M, uint64_t TableSize, ConstantInt *Offset,
4885  const SmallVectorImpl<std::pair<ConstantInt *, Constant *>> &Values,
4886  Constant *DefaultValue, const DataLayout &DL, const StringRef &FuncName);
4887 
4888  /// Build instructions with Builder to retrieve the value at
4889  /// the position given by Index in the lookup table.
4890  Value *BuildLookup(Value *Index, IRBuilder<> &Builder);
4891 
4892  /// Return true if a table with TableSize elements of
4893  /// type ElementType would fit in a target-legal register.
4894  static bool WouldFitInRegister(const DataLayout &DL, uint64_t TableSize,
4895  Type *ElementType);
4896 
4897 private:
4898  // Depending on the contents of the table, it can be represented in
4899  // different ways.
4900  enum {
4901  // For tables where each element contains the same value, we just have to
4902  // store that single value and return it for each lookup.
4903  SingleValueKind,
4904 
4905  // For tables where there is a linear relationship between table index
4906  // and values. We calculate the result with a simple multiplication
4907  // and addition instead of a table lookup.
4908  LinearMapKind,
4909 
4910  // For small tables with integer elements, we can pack them into a bitmap
4911  // that fits into a target-legal register. Values are retrieved by
4912  // shift and mask operations.
4913  BitMapKind,
4914 
4915  // The table is stored as an array of values. Values are retrieved by load
4916  // instructions from the table.
4917  ArrayKind
4918  } Kind;
4919 
4920  // For SingleValueKind, this is the single value.
4921  Constant *SingleValue = nullptr;
4922 
4923  // For BitMapKind, this is the bitmap.
4924  ConstantInt *BitMap = nullptr;
4925  IntegerType *BitMapElementTy = nullptr;
4926 
4927  // For LinearMapKind, these are the constants used to derive the value.
4928  ConstantInt *LinearOffset = nullptr;
4929  ConstantInt *LinearMultiplier = nullptr;
4930 
4931  // For ArrayKind, this is the array.
4932  GlobalVariable *Array = nullptr;
4933 };
4934 
4935 } // end anonymous namespace
4936 
4937 SwitchLookupTable::SwitchLookupTable(
4938  Module &M, uint64_t TableSize, ConstantInt *Offset,
4939  const SmallVectorImpl<std::pair<ConstantInt *, Constant *>> &Values,
4940  Constant *DefaultValue, const DataLayout &DL, const StringRef &FuncName) {
4941  assert(Values.size() && "Can't build lookup table without values!");
4942  assert(TableSize >= Values.size() && "Can't fit values in table!");
4943 
4944  // If all values in the table are equal, this is that value.
4945  SingleValue = Values.begin()->second;
4946 
4947  Type *ValueType = Values.begin()->second->getType();
4948 
4949  // Build up the table contents.
4950  SmallVector<Constant *, 64> TableContents(TableSize);
4951  for (size_t I = 0, E = Values.size(); I != E; ++I) {
4952  ConstantInt *CaseVal = Values[I].first;
4953  Constant *CaseRes = Values[I].second;
4954  assert(CaseRes->getType() == ValueType);
4955 
4956  uint64_t Idx = (CaseVal->getValue() - Offset->getValue()).getLimitedValue();
4957  TableContents[Idx] = CaseRes;
4958 
4959  if (CaseRes != SingleValue)
4960  SingleValue = nullptr;
4961  }
4962 
4963  // Fill in any holes in the table with the default result.
4964  if (Values.size() < TableSize) {
4965  assert(DefaultValue &&
4966  "Need a default value to fill the lookup table holes.");
4967  assert(DefaultValue->getType() == ValueType);
4968  for (uint64_t I = 0; I < TableSize; ++I) {
4969  if (!TableContents[I])
4970  TableContents[I] = DefaultValue;
4971  }
4972 
4973  if (DefaultValue != SingleValue)
4974  SingleValue = nullptr;
4975  }
4976 
4977  // If each element in the table contains the same value, we only need to store
4978  // that single value.
4979  if (SingleValue) {
4980  Kind = SingleValueKind;
4981  return;
4982  }
4983 
4984  // Check if we can derive the value with a linear transformation from the
4985  // table index.
4986  if (isa<IntegerType>(ValueType)) {
4987  bool LinearMappingPossible = true;
4988  APInt PrevVal;
4989  APInt DistToPrev;
4990  assert(TableSize >= 2 && "Should be a SingleValue table.");
4991  // Check if there is the same distance between two consecutive values.
4992  for (uint64_t I = 0; I < TableSize; ++I) {
4993  ConstantInt *ConstVal = dyn_cast<ConstantInt>(TableContents[I]);
4994  if (!ConstVal) {
4995  // This is an undef. We could deal with it, but undefs in lookup tables
4996  // are very seldom. It's probably not worth the additional complexity.
4997  LinearMappingPossible = false;
4998  break;
4999  }
5000  const APInt &Val = ConstVal->getValue();
5001  if (I != 0) {
5002  APInt Dist = Val - PrevVal;
5003  if (I == 1) {
5004  DistToPrev = Dist;
5005  } else if (Dist != DistToPrev) {
5006  LinearMappingPossible = false;
5007  break;
5008  }
5009  }
5010  PrevVal = Val;
5011  }
5012  if (LinearMappingPossible) {
5013  LinearOffset = cast<ConstantInt>(TableContents[0]);
5014  LinearMultiplier = ConstantInt::get(M.getContext(), DistToPrev);
5015  Kind = LinearMapKind;
5016  ++NumLinearMaps;
5017  return;
5018  }
5019  }
5020 
5021  // If the type is integer and the table fits in a register, build a bitmap.
5022  if (WouldFitInRegister(DL, TableSize, ValueType)) {
5023  IntegerType *IT = cast<IntegerType>(ValueType);
5024  APInt TableInt(TableSize * IT->getBitWidth(), 0);
5025  for (uint64_t I = TableSize; I > 0; --I) {
5026  TableInt <<= IT->getBitWidth();
5027  // Insert values into the bitmap. Undef values are set to zero.
5028  if (!isa<UndefValue>(TableContents[I - 1])) {
5029  ConstantInt *Val = cast<ConstantInt>(TableContents[I - 1]);
5030  TableInt |= Val->getValue().zext(TableInt.getBitWidth());
5031  }
5032  }
5033  BitMap = ConstantInt::get(M.getContext(), TableInt);
5034  BitMapElementTy = IT;
5035  Kind = BitMapKind;
5036  ++NumBitMaps;
5037  return;
5038  }
5039 
5040  // Store the table in an array.
5041  ArrayType *ArrayTy = ArrayType::get(ValueType, TableSize);
5042  Constant *Initializer = ConstantArray::get(ArrayTy, TableContents);
5043 
5044  Array = new GlobalVariable(M, ArrayTy, /*constant=*/true,
5045  GlobalVariable::PrivateLinkage, Initializer,
5046  "switch.table." + FuncName);
5047  Array->setUnnamedAddr(GlobalValue::UnnamedAddr::Global);
5048  // Set the alignment to that of an array items. We will be only loading one
5049  // value out of it.
5050  Array->setAlignment(DL.getPrefTypeAlignment(ValueType));
5051  Kind = ArrayKind;
5052 }
5053 
5054 Value *SwitchLookupTable::BuildLookup(Value *Index, IRBuilder<> &Builder) {
5055  switch (Kind) {
5056  case SingleValueKind:
5057  return SingleValue;
5058  case LinearMapKind: {
5059  // Derive the result value from the input value.
5060  Value *Result = Builder.CreateIntCast(Index, LinearMultiplier->getType(),
5061  false, "switch.idx.cast");
5062  if (!LinearMultiplier->isOne())
5063  Result = Builder.CreateMul(Result, LinearMultiplier, "switch.idx.mult");
5064  if (!LinearOffset->isZero())
5065  Result = Builder.CreateAdd(Result, LinearOffset, "switch.offset");
5066  return Result;
5067  }
5068  case BitMapKind: {
5069  // Type of the bitmap (e.g. i59).
5070  IntegerType *MapTy = BitMap->getType();
5071 
5072  // Cast Index to the same type as the bitmap.
5073  // Note: The Index is <= the number of elements in the table, so
5074  // truncating it to the width of the bitmask is safe.
5075  Value *ShiftAmt = Builder.CreateZExtOrTrunc(Index, MapTy, "switch.cast");
5076 
5077  // Multiply the shift amount by the element width.
5078  ShiftAmt = Builder.CreateMul(
5079  ShiftAmt, ConstantInt::get(MapTy, BitMapElementTy->getBitWidth()),
5080  "switch.shiftamt");
5081 
5082  // Shift down.
5083  Value *DownShifted =
5084  Builder.CreateLShr(BitMap, ShiftAmt, "switch.downshift");
5085  // Mask off.
5086  return Builder.CreateTrunc(DownShifted, BitMapElementTy, "switch.masked");
5087  }
5088  case ArrayKind: {
5089  // Make sure the table index will not overflow when treated as signed.
5090  IntegerType *IT = cast<IntegerType>(Index->getType());
5091  uint64_t TableSize =
5092  Array->getInitializer()->getType()->getArrayNumElements();
5093  if (TableSize > (1ULL << (IT->getBitWidth() - 1)))
5094  Index = Builder.CreateZExt(
5095  Index, IntegerType::get(IT->getContext(), IT->getBitWidth() + 1),
5096  "switch.tableidx.zext");
5097 
5098  Value *GEPIndices[] = {Builder.getInt32(0), Index};
5099  Value *GEP = Builder.CreateInBoundsGEP(Array->getValueType(), Array,
5100  GEPIndices, "switch.gep");
5101  return Builder.CreateLoad(GEP, "switch.load");
5102  }
5103  }
5104  llvm_unreachable("Unknown lookup table kind!");
5105 }
5106 
5107 bool SwitchLookupTable::WouldFitInRegister(const DataLayout &DL,
5108  uint64_t TableSize,
5109  Type *ElementType) {
5110  auto *IT = dyn_cast<IntegerType>(ElementType);
5111  if (!IT)
5112  return false;
5113  // FIXME: If the type is wider than it needs to be, e.g. i8 but all values
5114  // are <= 15, we could try to narrow the type.
5115 
5116  // Avoid overflow, fitsInLegalInteger uses unsigned int for the width.
5117  if (TableSize >= UINT_MAX / IT->getBitWidth())
5118  return false;
5119  return DL.fitsInLegalInteger(TableSize * IT->getBitWidth());
5120 }
5121 
5122 /// Determine whether a lookup table should be built for this switch, based on
5123 /// the number of cases, size of the table, and the types of the results.
5124 static bool
5125 ShouldBuildLookupTable(SwitchInst *SI, uint64_t TableSize,
5126  const TargetTransformInfo &TTI, const DataLayout &DL,
5127  const SmallDenseMap<PHINode *, Type *> &ResultTypes) {
5128  if (SI->getNumCases() > TableSize || TableSize >= UINT64_MAX / 10)
5129  return false; // TableSize overflowed, or mul below might overflow.
5130 
5131  bool AllTablesFitInRegister = true;
5132  bool HasIllegalType = false;
5133  for (const auto &I : ResultTypes) {
5134  Type *Ty = I.second;
5135 
5136  // Saturate this flag to true.
5137  HasIllegalType = HasIllegalType || !TTI.isTypeLegal(Ty);
5138 
5139  // Saturate this flag to false.
5140  AllTablesFitInRegister =
5141  AllTablesFitInRegister &&
5142  SwitchLookupTable::WouldFitInRegister(DL, TableSize, Ty);
5143 
5144  // If both flags saturate, we're done. NOTE: This *only* works with
5145  // saturating flags, and all flags have to saturate first due to the
5146  // non-deterministic behavior of iterating over a dense map.
5147  if (HasIllegalType && !AllTablesFitInRegister)
5148  break;
5149  }
5150 
5151  // If each table would fit in a register, we should build it anyway.
5152  if (AllTablesFitInRegister)
5153  return true;
5154 
5155  // Don't build a table that doesn't fit in-register if it has illegal types.
5156  if (HasIllegalType)
5157  return false;
5158 
5159  // The table density should be at least 40%. This is the same criterion as for
5160  // jump tables, see SelectionDAGBuilder::handleJTSwitchCase.
5161  // FIXME: Find the best cut-off.
5162  return SI->getNumCases() * 10 >= TableSize * 4;
5163 }
5164 
5165 /// Try to reuse the switch table index compare. Following pattern:
5166 /// \code
5167 /// if (idx < tablesize)
5168 /// r = table[idx]; // table does not contain default_value
5169 /// else
5170 /// r = default_value;
5171 /// if (r != default_value)
5172 /// ...
5173 /// \endcode
5174 /// Is optimized to:
5175 /// \code
5176 /// cond = idx < tablesize;
5177 /// if (cond)
5178 /// r = table[idx];
5179 /// else
5180 /// r = default_value;
5181 /// if (cond)
5182 /// ...
5183 /// \endcode
5184 /// Jump threading will then eliminate the second if(cond).
5185 static void reuseTableCompare(
5186  User *PhiUser, BasicBlock *PhiBlock, BranchInst *RangeCheckBranch,
5187  Constant *DefaultValue,
5188  const SmallVectorImpl<std::pair<ConstantInt *, Constant *>> &Values) {
5189  ICmpInst *CmpInst = dyn_cast<ICmpInst>(PhiUser);
5190  if (!CmpInst)
5191  return;
5192 
5193  // We require that the compare is in the same block as the phi so that jump
5194  // threading can do its work afterwards.
5195  if (CmpInst->getParent() != PhiBlock)
5196  return;
5197 
5198  Constant *CmpOp1 = dyn_cast<Constant>(CmpInst->getOperand(1));
5199  if (!CmpOp1)
5200  return;
5201 
5202  Value *RangeCmp = RangeCheckBranch->getCondition();
5203  Constant *TrueConst = ConstantInt::getTrue(RangeCmp->getType());
5204  Constant *FalseConst = ConstantInt::getFalse(RangeCmp->getType());
5205 
5206  // Check if the compare with the default value is constant true or false.
5207  Constant *DefaultConst = ConstantExpr::getICmp(CmpInst->getPredicate(),
5208  DefaultValue, CmpOp1, true);
5209  if (DefaultConst != TrueConst && DefaultConst != FalseConst)
5210  return;
5211 
5212  // Check if the compare with the case values is distinct from the default
5213  // compare result.
5214  for (auto ValuePair : Values) {
5215  Constant *CaseConst = ConstantExpr::getICmp(CmpInst->getPredicate(),
5216  ValuePair.second, CmpOp1, true);
5217  if (!CaseConst || CaseConst == DefaultConst || isa<UndefValue>(CaseConst))
5218  return;
5219  assert((CaseConst == TrueConst || CaseConst == FalseConst) &&
5220  "Expect true or false as compare result.");
5221  }
5222 
5223  // Check if the branch instruction dominates the phi node. It's a simple
5224  // dominance check, but sufficient for our needs.
5225  // Although this check is invariant in the calling loops, it's better to do it
5226  // at this late stage. Practically we do it at most once for a switch.
5227  BasicBlock *BranchBlock = RangeCheckBranch->getParent();
5228  for (auto PI = pred_begin(PhiBlock), E = pred_end(PhiBlock); PI != E; ++PI) {
5229  BasicBlock *Pred = *PI;
5230  if (Pred != BranchBlock && Pred->getUniquePredecessor() != BranchBlock)
5231  return;
5232  }
5233 
5234  if (DefaultConst == FalseConst) {
5235  // The compare yields the same result. We can replace it.
5236  CmpInst->replaceAllUsesWith(RangeCmp);
5237  ++NumTableCmpReuses;
5238  } else {
5239  // The compare yields the same result, just inverted. We can replace it.
5240  Value *InvertedTableCmp = BinaryOperator::CreateXor(
5241  RangeCmp, ConstantInt::get(RangeCmp->getType(), 1), "inverted.cmp",
5242  RangeCheckBranch);
5243  CmpInst->replaceAllUsesWith(InvertedTableCmp);
5244  ++NumTableCmpReuses;
5245  }
5246 }
5247 
5248 /// If the switch is only used to initialize one or more phi nodes in a common
5249 /// successor block with different constant values, replace the switch with
5250 /// lookup tables.
5252  const DataLayout &DL,
5253  const TargetTransformInfo &TTI) {
5254  assert(SI->getNumCases() > 1 && "Degenerate switch?");
5255 
5256  Function *Fn = SI->getParent()->getParent();
5257  // Only build lookup table when we have a target that supports it or the
5258  // attribute is not set.
5259  if (!TTI.shouldBuildLookupTables() ||
5260  (Fn->getFnAttribute("no-jump-tables").getValueAsString() == "true"))
5261  return false;
5262 
5263  // FIXME: If the switch is too sparse for a lookup table, perhaps we could
5264  // split off a dense part and build a lookup table for that.
5265 
5266  // FIXME: This creates arrays of GEPs to constant strings, which means each
5267  // GEP needs a runtime relocation in PIC code. We should just build one big
5268  // string and lookup indices into that.
5269 
5270  // Ignore switches with less than three cases. Lookup tables will not make
5271  // them faster, so we don't analyze them.
5272  if (SI->getNumCases() < 3)
5273  return false;
5274 
5275  // Figure out the corresponding result for each case value and phi node in the
5276  // common destination, as well as the min and max case values.
5277  assert(SI->case_begin() != SI->case_end());
5278  SwitchInst::CaseIt CI = SI->case_begin();
5279  ConstantInt *MinCaseVal = CI->getCaseValue();
5280  ConstantInt *MaxCaseVal = CI->getCaseValue();
5281 
5282  BasicBlock *CommonDest = nullptr;
5283 
5284  using ResultListTy = SmallVector<std::pair<ConstantInt *, Constant *>, 4>;
5286 
5287  SmallDenseMap<PHINode *, Constant *> DefaultResults;
5290 
5291  for (SwitchInst::CaseIt E = SI->case_end(); CI != E; ++CI) {
5292  ConstantInt *CaseVal = CI->getCaseValue();
5293  if (CaseVal->getValue().slt(MinCaseVal->getValue()))
5294  MinCaseVal = CaseVal;
5295  if (CaseVal->getValue().sgt(MaxCaseVal->getValue()))
5296  MaxCaseVal = CaseVal;
5297 
5298  // Resulting value at phi nodes for this case value.
5299  using ResultsTy = SmallVector<std::pair<PHINode *, Constant *>, 4>;
5300  ResultsTy Results;
5301  if (!GetCaseResults(SI, CaseVal, CI->getCaseSuccessor(), &CommonDest,
5302  Results, DL, TTI))
5303  return false;
5304 
5305  // Append the result from this case to the list for each phi.
5306  for (const auto &I : Results) {
5307  PHINode *PHI = I.first;
5308  Constant *Value = I.second;
5309  if (!ResultLists.count(PHI))
5310  PHIs.push_back(PHI);
5311  ResultLists[PHI].push_back(std::make_pair(CaseVal, Value));
5312  }
5313  }
5314 
5315  // Keep track of the result types.
5316  for (PHINode *PHI : PHIs) {
5317  ResultTypes[PHI] = ResultLists[PHI][0].second->getType();
5318  }
5319 
5320  uint64_t NumResults = ResultLists[PHIs[0]].size();
5321  APInt RangeSpread = MaxCaseVal->getValue() - MinCaseVal->getValue();
5322  uint64_t TableSize = RangeSpread.getLimitedValue() + 1;
5323  bool TableHasHoles = (NumResults < TableSize);
5324 
5325  // If the table has holes, we need a constant result for the default case
5326  // or a bitmask that fits in a register.
5327  SmallVector<std::pair<PHINode *, Constant *>, 4> DefaultResultsList;
5328  bool HasDefaultResults =
5329  GetCaseResults(SI, nullptr, SI->getDefaultDest(), &CommonDest,
5330  DefaultResultsList, DL, TTI);
5331 
5332  bool NeedMask = (TableHasHoles && !HasDefaultResults);
5333  if (NeedMask) {
5334  // As an extra penalty for the validity test we require more cases.
5335  if (SI->getNumCases() < 4) // FIXME: Find best threshold value (benchmark).
5336  return false;
5337  if (!DL.fitsInLegalInteger(TableSize))
5338  return false;
5339  }
5340 
5341  for (const auto &I : DefaultResultsList) {
5342  PHINode *PHI = I.first;
5343  Constant *Result = I.second;
5344  DefaultResults[PHI] = Result;
5345  }
5346 
5347  if (!ShouldBuildLookupTable(SI, TableSize, TTI, DL, ResultTypes))
5348  return false;
5349 
5350  // Create the BB that does the lookups.
5351  Module &Mod = *CommonDest->getParent()->getParent();
5352  BasicBlock *LookupBB = BasicBlock::Create(
5353  Mod.getContext(), "switch.lookup", CommonDest->getParent(), CommonDest);
5354 
5355  // Compute the table index value.
5356  Builder.SetInsertPoint(SI);
5357  Value *TableIndex;
5358  if (MinCaseVal->isNullValue())
5359  TableIndex = SI->getCondition();
5360  else
5361  TableIndex = Builder.CreateSub(SI->getCondition(), MinCaseVal,
5362  "switch.tableidx");
5363 
5364  // Compute the maximum table size representable by the integer type we are
5365  // switching upon.
5366  unsigned CaseSize = MinCaseVal->getType()->getPrimitiveSizeInBits();
5367  uint64_t MaxTableSize = CaseSize > 63 ? UINT64_MAX : 1ULL << CaseSize;
5368  assert(MaxTableSize >= TableSize &&
5369  "It is impossible for a switch to have more entries than the max "
5370  "representable value of its input integer type's size.");
5371 
5372  // If the default destination is unreachable, or if the lookup table covers
5373  // all values of the conditional variable, branch directly to the lookup table
5374  // BB. Otherwise, check that the condition is within the case range.
5375  const bool DefaultIsReachable =
5376  !isa<UnreachableInst>(SI->getDefaultDest()->getFirstNonPHIOrDbg());
5377  const bool GeneratingCoveredLookupTable = (MaxTableSize == TableSize);
5378  BranchInst *RangeCheckBranch = nullptr;
5379 
5380  if (!DefaultIsReachable || GeneratingCoveredLookupTable) {
5381  Builder.CreateBr(LookupBB);
5382  // Note: We call removeProdecessor later since we need to be able to get the
5383  // PHI value for the default case in case we're using a bit mask.
5384  } else {
5385  Value *Cmp = Builder.CreateICmpULT(
5386  TableIndex, ConstantInt::get(MinCaseVal->getType(), TableSize));
5387  RangeCheckBranch =
5388  Builder.CreateCondBr(Cmp, LookupBB, SI->getDefaultDest());
5389  }
5390 
5391  // Populate the BB that does the lookups.
5392  Builder.SetInsertPoint(LookupBB);
5393 
5394  if (NeedMask) {
5395  // Before doing the lookup, we do the hole check. The LookupBB is therefore
5396  // re-purposed to do the hole check, and we create a new LookupBB.
5397  BasicBlock *MaskBB = LookupBB;
5398  MaskBB->setName("switch.hole_check");
5399  LookupBB = BasicBlock::Create(Mod.getContext(), "switch.lookup",
5400  CommonDest->getParent(), CommonDest);
5401 
5402  // Make the mask's bitwidth at least 8-bit and a power-of-2 to avoid
5403  // unnecessary illegal types.
5404  uint64_t TableSizePowOf2 = NextPowerOf2(std::max(7ULL, TableSize - 1ULL));
5405  APInt MaskInt(TableSizePowOf2, 0);
5406  APInt One(TableSizePowOf2, 1);
5407  // Build bitmask; fill in a 1 bit for every case.
5408  const ResultListTy &ResultList = ResultLists[PHIs[0]];
5409  for (size_t I = 0, E = ResultList.size(); I != E; ++I) {
5410  uint64_t Idx = (ResultList[I].first->getValue() - MinCaseVal->getValue())
5411  .getLimitedValue();
5412  MaskInt |= One << Idx;
5413  }
5414  ConstantInt *TableMask = ConstantInt::get(Mod.getContext(), MaskInt);
5415 
5416  // Get the TableIndex'th bit of the bitmask.
5417  // If this bit is 0 (meaning hole) jump to the default destination,
5418  // else continue with table lookup.
5419  IntegerType *MapTy = TableMask->getType();
5420  Value *MaskIndex =
5421  Builder.CreateZExtOrTrunc(TableIndex, MapTy, "switch.maskindex");
5422  Value *Shifted = Builder.CreateLShr(TableMask, MaskIndex, "switch.shifted");
5423  Value *LoBit = Builder.CreateTrunc(
5424  Shifted, Type::getInt1Ty(Mod.getContext()), "switch.lobit");
5425  Builder.CreateCondBr(LoBit, LookupBB, SI->getDefaultDest());
5426 
5427  Builder.SetInsertPoint(LookupBB);
5428  AddPredecessorToBlock(SI->getDefaultDest(), MaskBB, SI->getParent());
5429  }
5430 
5431  if (!DefaultIsReachable || GeneratingCoveredLookupTable) {
5432  // We cached PHINodes in PHIs. To avoid accessing deleted PHINodes later,
5433  // do not delete PHINodes here.
5435  /*DontDeleteUselessPHIs=*/true);
5436  }
5437 
5438  bool ReturnedEarly = false;
5439  for (PHINode *PHI : PHIs) {
5440  const ResultListTy &ResultList = ResultLists[PHI];
5441 
5442  // If using a bitmask, use any value to fill the lookup table holes.
5443  Constant *DV = NeedMask ? ResultLists[PHI][0].second : DefaultResults[PHI];
5444  StringRef FuncName = Fn->getName();
5445  SwitchLookupTable Table(Mod, TableSize, MinCaseVal, ResultList, DV, DL,
5446  FuncName);
5447 
5448  Value *Result = Table.BuildLookup(TableIndex, Builder);
5449 
5450  // If the result is used to return immediately from the function, we want to
5451  // do that right here.
5452  if (PHI->hasOneUse() && isa<ReturnInst>(*PHI->user_begin()) &&
5453  PHI->user_back() == CommonDest->getFirstNonPHIOrDbg()) {
5454  Builder.CreateRet(Result);
5455  ReturnedEarly = true;
5456  break;
5457  }
5458 
5459  // Do a small peephole optimization: re-use the switch table compare if
5460  // possible.
5461  if (!TableHasHoles && HasDefaultResults && RangeCheckBranch) {
5462  BasicBlock *PhiBlock = PHI->getParent();
5463  // Search for compare instructions which use the phi.
5464  for (auto *User : PHI->users()) {
5465  reuseTableCompare(User, PhiBlock, RangeCheckBranch, DV, ResultList);
5466  }
5467  }
5468 
5469  PHI->addIncoming(Result, LookupBB);
5470  }
5471 
5472  if (!ReturnedEarly)
5473  Builder.CreateBr(CommonDest);
5474 
5475  // Remove the switch.
5476  for (unsigned i = 0, e = SI->getNumSuccessors(); i < e; ++i) {
5477  BasicBlock *Succ = SI->getSuccessor(i);
5478 
5479  if (Succ == SI->getDefaultDest())
5480  continue;
5481  Succ->removePredecessor(SI->getParent());
5482  }
5483  SI->eraseFromParent();
5484 
5485  ++NumLookupTables;
5486  if (NeedMask)
5487  ++NumLookupTablesHoles;
5488  return true;
5489 }
5490 
5491 static bool isSwitchDense(ArrayRef<int64_t> Values) {
5492  // See also SelectionDAGBuilder::isDense(), which this function was based on.
5493  uint64_t Diff = (uint64_t)Values.back() - (uint64_t)Values.front();
5494  uint64_t Range = Diff + 1;
5495  uint64_t NumCases = Values.size();
5496  // 40% is the default density for building a jump table in optsize/minsize mode.
5497  uint64_t MinDensity = 40;
5498 
5499  return NumCases * 100 >= Range * MinDensity;
5500 }
5501 
5502 /// Try to transform a switch that has "holes" in it to a contiguous sequence
5503 /// of cases.
5504 ///
5505 /// A switch such as: switch(i) {case 5: case 9: case 13: case 17:} can be
5506 /// range-reduced to: switch ((i-5) / 4) {case 0: case 1: case 2: case 3:}.
5507 ///
5508 /// This converts a sparse switch into a dense switch which allows better
5509 /// lowering and could also allow transforming into a lookup table.
5511  const DataLayout &DL,
5512  const TargetTransformInfo &TTI) {
5513  auto *CondTy = cast<IntegerType>(SI->getCondition()->getType());
5514  if (CondTy->getIntegerBitWidth() > 64 ||
5515  !DL.fitsInLegalInteger(CondTy->getIntegerBitWidth()))
5516  return false;
5517  // Only bother with this optimization if there are more than 3 switch cases;
5518  // SDAG will only bother creating jump tables for 4 or more cases.
5519  if (SI->getNumCases() < 4)
5520  return false;
5521 
5522  // This transform is agnostic to the signedness of the input or case values. We
5523  // can treat the case values as signed or unsigned. We can optimize more common
5524  // cases such as a sequence crossing zero {-4,0,4,8} if we interpret case values
5525  // as signed.
5526  SmallVector<int64_t,4> Values;
5527  for (auto &C : SI->cases())
5528  Values.push_back(C.getCaseValue()->getValue().getSExtValue());
5529  llvm::sort(Values);
5530 
5531  // If the switch is already dense, there's nothing useful to do here.
5532  if (isSwitchDense(Values))
5533  return false;
5534 
5535  // First, transform the values such that they start at zero and ascend.
5536  int64_t Base = Values[0];
5537  for (auto &V : Values)
5538  V -= (uint64_t)(Base);
5539 
5540  // Now we have signed numbers that have been shifted so that, given enough
5541  // precision, there are no negative values. Since the rest of the transform
5542  // is bitwise only, we switch now to an unsigned representation.
5543  uint64_t GCD = 0;
5544  for (auto &V : Values)
5545  GCD = GreatestCommonDivisor64(GCD, (uint64_t)V);
5546 
5547  // This transform can be done speculatively because it is so cheap - it results
5548  // in a single rotate operation being inserted. This can only happen if the
5549  // factor extracted is a power of 2.
5550  // FIXME: If the GCD is an odd number we can multiply by the multiplicative
5551  // inverse of GCD and then perform this transform.
5552  // FIXME: It's possible that optimizing a switch on powers of two might also
5553  // be beneficial - flag values are often powers of two and we could use a CLZ
5554  // as the key function.
5555  if (GCD <= 1 || !isPowerOf2_64(GCD))
5556  // No common divisor found or too expensive to compute key function.
5557  return false;
5558 
5559  unsigned Shift = Log2_64(GCD);
5560  for (auto &V : Values)
5561  V = (int64_t)((uint64_t)V >> Shift);
5562 
5563  if (!isSwitchDense(Values))
5564  // Transform didn't create a dense switch.
5565  return false;
5566 
5567  // The obvious transform is to shift the switch condition right and emit a
5568  // check that the condition actually cleanly divided by GCD, i.e.
5569  // C & (1 << Shift - 1) == 0
5570  // inserting a new CFG edge to handle the case where it didn't divide cleanly.
5571  //
5572  // A cheaper way of doing this is a simple ROTR(C, Shift). This performs the
5573  // shift and puts the shifted-off bits in the uppermost bits. If any of these
5574  // are nonzero then the switch condition will be very large and will hit the
5575  // default case.
5576 
5577  auto *Ty = cast<IntegerType>(SI->getCondition()->getType());
5578  Builder.SetInsertPoint(SI);
5579  auto *ShiftC = ConstantInt::get(Ty, Shift);
5580  auto *Sub = Builder.CreateSub(SI->getCondition(), ConstantInt::get(Ty, Base));
5581  auto *LShr = Builder.CreateLShr(Sub, ShiftC);
5582  auto *Shl = Builder.CreateShl(Sub, Ty->getBitWidth() - Shift);
5583  auto *Rot = Builder.CreateOr(LShr, Shl);
5584  SI->replaceUsesOfWith(SI->getCondition(), Rot);
5585 
5586  for (auto Case : SI->cases()) {
5587  auto *Orig = Case.getCaseValue();
5588  auto Sub = Orig->getValue() - APInt(Ty->getBitWidth(), Base);
5589  Case.setValue(
5590  cast<ConstantInt>(ConstantInt::get(Ty, Sub.lshr(ShiftC->getValue()))));
5591  }
5592  return true;
5593 }
5594 
5595 bool SimplifyCFGOpt::SimplifySwitch(SwitchInst *