LLVM  10.0.0svn
SimplifyCFG.cpp
Go to the documentation of this file.
1 //===- SimplifyCFG.cpp - Code to perform CFG simplification ---------------===//
2 //
3 // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
4 // See https://llvm.org/LICENSE.txt for license information.
5 // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
6 //
7 //===----------------------------------------------------------------------===//
8 //
9 // Peephole optimize the CFG.
10 //
11 //===----------------------------------------------------------------------===//
12 
13 #include "llvm/ADT/APInt.h"
14 #include "llvm/ADT/ArrayRef.h"
15 #include "llvm/ADT/DenseMap.h"
16 #include "llvm/ADT/Optional.h"
17 #include "llvm/ADT/STLExtras.h"
18 #include "llvm/ADT/SetOperations.h"
19 #include "llvm/ADT/SetVector.h"
20 #include "llvm/ADT/SmallPtrSet.h"
21 #include "llvm/ADT/SmallVector.h"
22 #include "llvm/ADT/Statistic.h"
23 #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"
71 #include <algorithm>
72 #include <cassert>
73 #include <climits>
74 #include <cstddef>
75 #include <cstdint>
76 #include <iterator>
77 #include <map>
78 #include <set>
79 #include <tuple>
80 #include <utility>
81 #include <vector>
82 
83 using namespace llvm;
84 using namespace PatternMatch;
85 
86 #define DEBUG_TYPE "simplifycfg"
87 
88 // Chosen as 2 so as to be cheap, but still to have enough power to fold
89 // a select, so the "clamp" idiom (of a min followed by a max) will be caught.
90 // To catch this, we need to fold a compare and a select, hence '2' being the
91 // minimum reasonable default.
93  "phi-node-folding-threshold", cl::Hidden, cl::init(2),
94  cl::desc(
95  "Control the amount of phi node folding to perform (default = 2)"));
96 
98  "two-entry-phi-node-folding-threshold", cl::Hidden, cl::init(4),
99  cl::desc("Control the maximal total instruction cost that we are willing "
100  "to speculatively execute to fold a 2-entry PHI node into a "
101  "select (default = 4)"));
102 
103 static cl::opt<bool> DupRet(
104  "simplifycfg-dup-ret", cl::Hidden, cl::init(false),
105  cl::desc("Duplicate return instructions into unconditional branches"));
106 
107 static cl::opt<bool>
108  SinkCommon("simplifycfg-sink-common", cl::Hidden, cl::init(true),
109  cl::desc("Sink common instructions down to the end block"));
110 
112  "simplifycfg-hoist-cond-stores", cl::Hidden, cl::init(true),
113  cl::desc("Hoist conditional stores if an unconditional store precedes"));
114 
116  "simplifycfg-merge-cond-stores", cl::Hidden, cl::init(true),
117  cl::desc("Hoist conditional stores even if an unconditional store does not "
118  "precede - hoist multiple conditional stores into a single "
119  "predicated store"));
120 
122  "simplifycfg-merge-cond-stores-aggressively", cl::Hidden, cl::init(false),
123  cl::desc("When merging conditional stores, do so even if the resultant "
124  "basic blocks are unlikely to be if-converted as a result"));
125 
127  "speculate-one-expensive-inst", cl::Hidden, cl::init(true),
128  cl::desc("Allow exactly one expensive instruction to be speculatively "
129  "executed"));
130 
132  "max-speculation-depth", cl::Hidden, cl::init(10),
133  cl::desc("Limit maximum recursion depth when calculating costs of "
134  "speculatively executed instructions"));
135 
136 STATISTIC(NumBitMaps, "Number of switch instructions turned into bitmaps");
137 STATISTIC(NumLinearMaps,
138  "Number of switch instructions turned into linear mapping");
139 STATISTIC(NumLookupTables,
140  "Number of switch instructions turned into lookup tables");
141 STATISTIC(
142  NumLookupTablesHoles,
143  "Number of switch instructions turned into lookup tables (holes checked)");
144 STATISTIC(NumTableCmpReuses, "Number of reused switch table lookup compares");
145 STATISTIC(NumSinkCommons,
146  "Number of common instructions sunk down to the end block");
147 STATISTIC(NumSpeculations, "Number of speculative executed instructions");
148 
149 namespace {
150 
151 // The first field contains the value that the switch produces when a certain
152 // case group is selected, and the second field is a vector containing the
153 // cases composing the case group.
154 using SwitchCaseResultVectorTy =
156 
157 // The first field contains the phi node that generates a result of the switch
158 // and the second field contains the value generated for a certain case in the
159 // switch for that PHI.
160 using SwitchCaseResultsTy = SmallVector<std::pair<PHINode *, Constant *>, 4>;
161 
162 /// ValueEqualityComparisonCase - Represents a case of a switch.
163 struct ValueEqualityComparisonCase {
165  BasicBlock *Dest;
166 
167  ValueEqualityComparisonCase(ConstantInt *Value, BasicBlock *Dest)
168  : Value(Value), Dest(Dest) {}
169 
170  bool operator<(ValueEqualityComparisonCase RHS) const {
171  // Comparing pointers is ok as we only rely on the order for uniquing.
172  return Value < RHS.Value;
173  }
174 
175  bool operator==(BasicBlock *RHSDest) const { return Dest == RHSDest; }
176 };
177 
178 class SimplifyCFGOpt {
179  const TargetTransformInfo &TTI;
180  const DataLayout &DL;
181  SmallPtrSetImpl<BasicBlock *> *LoopHeaders;
182  const SimplifyCFGOptions &Options;
183  bool Resimplify;
184 
185  Value *isValueEqualityComparison(Instruction *TI);
186  BasicBlock *GetValueEqualityComparisonCases(
187  Instruction *TI, std::vector<ValueEqualityComparisonCase> &Cases);
188  bool SimplifyEqualityComparisonWithOnlyPredecessor(Instruction *TI,
189  BasicBlock *Pred,
190  IRBuilder<> &Builder);
191  bool FoldValueComparisonIntoPredecessors(Instruction *TI,
192  IRBuilder<> &Builder);
193 
194  bool SimplifyReturn(ReturnInst *RI, IRBuilder<> &Builder);
195  bool SimplifyResume(ResumeInst *RI, IRBuilder<> &Builder);
196  bool SimplifySingleResume(ResumeInst *RI);
197  bool SimplifyCommonResume(ResumeInst *RI);
198  bool SimplifyCleanupReturn(CleanupReturnInst *RI);
199  bool SimplifyUnreachable(UnreachableInst *UI);
200  bool SimplifySwitch(SwitchInst *SI, IRBuilder<> &Builder);
201  bool SimplifyIndirectBr(IndirectBrInst *IBI);
202  bool SimplifyUncondBranch(BranchInst *BI, IRBuilder<> &Builder);
203  bool SimplifyCondBranch(BranchInst *BI, IRBuilder<> &Builder);
204 
205  bool tryToSimplifyUncondBranchWithICmpInIt(ICmpInst *ICI,
206  IRBuilder<> &Builder);
207 
208 public:
209  SimplifyCFGOpt(const TargetTransformInfo &TTI, const DataLayout &DL,
210  SmallPtrSetImpl<BasicBlock *> *LoopHeaders,
211  const SimplifyCFGOptions &Opts)
212  : TTI(TTI), DL(DL), LoopHeaders(LoopHeaders), Options(Opts) {}
213 
214  bool run(BasicBlock *BB);
215  bool simplifyOnce(BasicBlock *BB);
216 
217  // Helper to set Resimplify and return change indication.
218  bool requestResimplify() {
219  Resimplify = true;
220  return true;
221  }
222 };
223 
224 } // end anonymous namespace
225 
226 /// Return true if it is safe to merge these two
227 /// terminator instructions together.
228 static bool
230  SmallSetVector<BasicBlock *, 4> *FailBlocks = nullptr) {
231  if (SI1 == SI2)
232  return false; // Can't merge with self!
233 
234  // It is not safe to merge these two switch instructions if they have a common
235  // successor, and if that successor has a PHI node, and if *that* PHI node has
236  // conflicting incoming values from the two switch blocks.
237  BasicBlock *SI1BB = SI1->getParent();
238  BasicBlock *SI2BB = SI2->getParent();
239 
240  SmallPtrSet<BasicBlock *, 16> SI1Succs(succ_begin(SI1BB), succ_end(SI1BB));
241  bool Fail = false;
242  for (BasicBlock *Succ : successors(SI2BB))
243  if (SI1Succs.count(Succ))
244  for (BasicBlock::iterator BBI = Succ->begin(); isa<PHINode>(BBI); ++BBI) {
245  PHINode *PN = cast<PHINode>(BBI);
246  if (PN->getIncomingValueForBlock(SI1BB) !=
247  PN->getIncomingValueForBlock(SI2BB)) {
248  if (FailBlocks)
249  FailBlocks->insert(Succ);
250  Fail = true;
251  }
252  }
253 
254  return !Fail;
255 }
256 
257 /// Return true if it is safe and profitable to merge these two terminator
258 /// instructions together, where SI1 is an unconditional branch. PhiNodes will
259 /// store all PHI nodes in common successors.
260 static bool
262  Instruction *Cond,
263  SmallVectorImpl<PHINode *> &PhiNodes) {
264  if (SI1 == SI2)
265  return false; // Can't merge with self!
266  assert(SI1->isUnconditional() && SI2->isConditional());
267 
268  // We fold the unconditional branch if we can easily update all PHI nodes in
269  // common successors:
270  // 1> We have a constant incoming value for the conditional branch;
271  // 2> We have "Cond" as the incoming value for the unconditional branch;
272  // 3> SI2->getCondition() and Cond have same operands.
273  CmpInst *Ci2 = dyn_cast<CmpInst>(SI2->getCondition());
274  if (!Ci2)
275  return false;
276  if (!(Cond->getOperand(0) == Ci2->getOperand(0) &&
277  Cond->getOperand(1) == Ci2->getOperand(1)) &&
278  !(Cond->getOperand(0) == Ci2->getOperand(1) &&
279  Cond->getOperand(1) == Ci2->getOperand(0)))
280  return false;
281 
282  BasicBlock *SI1BB = SI1->getParent();
283  BasicBlock *SI2BB = SI2->getParent();
284  SmallPtrSet<BasicBlock *, 16> SI1Succs(succ_begin(SI1BB), succ_end(SI1BB));
285  for (BasicBlock *Succ : successors(SI2BB))
286  if (SI1Succs.count(Succ))
287  for (BasicBlock::iterator BBI = Succ->begin(); isa<PHINode>(BBI); ++BBI) {
288  PHINode *PN = cast<PHINode>(BBI);
289  if (PN->getIncomingValueForBlock(SI1BB) != Cond ||
290  !isa<ConstantInt>(PN->getIncomingValueForBlock(SI2BB)))
291  return false;
292  PhiNodes.push_back(PN);
293  }
294  return true;
295 }
296 
297 /// Update PHI nodes in Succ to indicate that there will now be entries in it
298 /// from the 'NewPred' block. The values that will be flowing into the PHI nodes
299 /// will be the same as those coming in from ExistPred, an existing predecessor
300 /// of Succ.
301 static void AddPredecessorToBlock(BasicBlock *Succ, BasicBlock *NewPred,
302  BasicBlock *ExistPred,
303  MemorySSAUpdater *MSSAU = nullptr) {
304  for (PHINode &PN : Succ->phis())
305  PN.addIncoming(PN.getIncomingValueForBlock(ExistPred), NewPred);
306  if (MSSAU)
307  if (auto *MPhi = MSSAU->getMemorySSA()->getMemoryAccess(Succ))
308  MPhi->addIncoming(MPhi->getIncomingValueForBlock(ExistPred), NewPred);
309 }
310 
311 /// Compute an abstract "cost" of speculating the given instruction,
312 /// which is assumed to be safe to speculate. TCC_Free means cheap,
313 /// TCC_Basic means less cheap, and TCC_Expensive means prohibitively
314 /// expensive.
315 static unsigned ComputeSpeculationCost(const User *I,
316  const TargetTransformInfo &TTI) {
318  "Instruction is not safe to speculatively execute!");
319  return TTI.getUserCost(I);
320 }
321 
322 /// If we have a merge point of an "if condition" as accepted above,
323 /// return true if the specified value dominates the block. We
324 /// don't handle the true generality of domination here, just a special case
325 /// which works well enough for us.
326 ///
327 /// If AggressiveInsts is non-null, and if V does not dominate BB, we check to
328 /// see if V (which must be an instruction) and its recursive operands
329 /// that do not dominate BB have a combined cost lower than CostRemaining and
330 /// are non-trapping. If both are true, the instruction is inserted into the
331 /// set and true is returned.
332 ///
333 /// The cost for most non-trapping instructions is defined as 1 except for
334 /// Select whose cost is 2.
335 ///
336 /// After this function returns, CostRemaining is decreased by the cost of
337 /// V plus its non-dominating operands. If that cost is greater than
338 /// CostRemaining, false is returned and CostRemaining is undefined.
340  SmallPtrSetImpl<Instruction *> &AggressiveInsts,
341  int &BudgetRemaining,
342  const TargetTransformInfo &TTI,
343  unsigned Depth = 0) {
344  // It is possible to hit a zero-cost cycle (phi/gep instructions for example),
345  // so limit the recursion depth.
346  // TODO: While this recursion limit does prevent pathological behavior, it
347  // would be better to track visited instructions to avoid cycles.
348  if (Depth == MaxSpeculationDepth)
349  return false;
350 
352  if (!I) {
353  // Non-instructions all dominate instructions, but not all constantexprs
354  // can be executed unconditionally.
355  if (ConstantExpr *C = dyn_cast<ConstantExpr>(V))
356  if (C->canTrap())
357  return false;
358  return true;
359  }
360  BasicBlock *PBB = I->getParent();
361 
362  // We don't want to allow weird loops that might have the "if condition" in
363  // the bottom of this block.
364  if (PBB == BB)
365  return false;
366 
367  // If this instruction is defined in a block that contains an unconditional
368  // branch to BB, then it must be in the 'conditional' part of the "if
369  // statement". If not, it definitely dominates the region.
371  if (!BI || BI->isConditional() || BI->getSuccessor(0) != BB)
372  return true;
373 
374  // If we have seen this instruction before, don't count it again.
375  if (AggressiveInsts.count(I))
376  return true;
377 
378  // Okay, it looks like the instruction IS in the "condition". Check to
379  // see if it's a cheap instruction to unconditionally compute, and if it
380  // only uses stuff defined outside of the condition. If so, hoist it out.
382  return false;
383 
384  BudgetRemaining -= ComputeSpeculationCost(I, TTI);
385 
386  // Allow exactly one instruction to be speculated regardless of its cost
387  // (as long as it is safe to do so).
388  // This is intended to flatten the CFG even if the instruction is a division
389  // or other expensive operation. The speculation of an expensive instruction
390  // is expected to be undone in CodeGenPrepare if the speculation has not
391  // enabled further IR optimizations.
392  if (BudgetRemaining < 0 &&
393  (!SpeculateOneExpensiveInst || !AggressiveInsts.empty() || Depth > 0))
394  return false;
395 
396  // Okay, we can only really hoist these out if their operands do
397  // not take us over the cost threshold.
398  for (User::op_iterator i = I->op_begin(), e = I->op_end(); i != e; ++i)
399  if (!DominatesMergePoint(*i, BB, AggressiveInsts, BudgetRemaining, TTI,
400  Depth + 1))
401  return false;
402  // Okay, it's safe to do this! Remember this instruction.
403  AggressiveInsts.insert(I);
404  return true;
405 }
406 
407 /// Extract ConstantInt from value, looking through IntToPtr
408 /// and PointerNullValue. Return NULL if value is not a constant int.
409 static ConstantInt *GetConstantInt(Value *V, const DataLayout &DL) {
410  // Normal constant int.
411  ConstantInt *CI = dyn_cast<ConstantInt>(V);
412  if (CI || !isa<Constant>(V) || !V->getType()->isPointerTy())
413  return CI;
414 
415  // This is some kind of pointer constant. Turn it into a pointer-sized
416  // ConstantInt if possible.
417  IntegerType *PtrTy = cast<IntegerType>(DL.getIntPtrType(V->getType()));
418 
419  // Null pointer means 0, see SelectionDAGBuilder::getValue(const Value*).
420  if (isa<ConstantPointerNull>(V))
421  return ConstantInt::get(PtrTy, 0);
422 
423  // IntToPtr const int.
424  if (ConstantExpr *CE = dyn_cast<ConstantExpr>(V))
425  if (CE->getOpcode() == Instruction::IntToPtr)
426  if (ConstantInt *CI = dyn_cast<ConstantInt>(CE->getOperand(0))) {
427  // The constant is very likely to have the right type already.
428  if (CI->getType() == PtrTy)
429  return CI;
430  else
431  return cast<ConstantInt>(
432  ConstantExpr::getIntegerCast(CI, PtrTy, /*isSigned=*/false));
433  }
434  return nullptr;
435 }
436 
437 namespace {
438 
439 /// Given a chain of or (||) or and (&&) comparison of a value against a
440 /// constant, this will try to recover the information required for a switch
441 /// structure.
442 /// It will depth-first traverse the chain of comparison, seeking for patterns
443 /// like %a == 12 or %a < 4 and combine them to produce a set of integer
444 /// representing the different cases for the switch.
445 /// Note that if the chain is composed of '||' it will build the set of elements
446 /// that matches the comparisons (i.e. any of this value validate the chain)
447 /// while for a chain of '&&' it will build the set elements that make the test
448 /// fail.
449 struct ConstantComparesGatherer {
450  const DataLayout &DL;
451 
452  /// Value found for the switch comparison
453  Value *CompValue = nullptr;
454 
455  /// Extra clause to be checked before the switch
456  Value *Extra = nullptr;
457 
458  /// Set of integers to match in switch
460 
461  /// Number of comparisons matched in the and/or chain
462  unsigned UsedICmps = 0;
463 
464  /// Construct and compute the result for the comparison instruction Cond
465  ConstantComparesGatherer(Instruction *Cond, const DataLayout &DL) : DL(DL) {
466  gather(Cond);
467  }
468 
469  ConstantComparesGatherer(const ConstantComparesGatherer &) = delete;
470  ConstantComparesGatherer &
471  operator=(const ConstantComparesGatherer &) = delete;
472 
473 private:
474  /// Try to set the current value used for the comparison, it succeeds only if
475  /// it wasn't set before or if the new value is the same as the old one
476  bool setValueOnce(Value *NewVal) {
477  if (CompValue && CompValue != NewVal)
478  return false;
479  CompValue = NewVal;
480  return (CompValue != nullptr);
481  }
482 
483  /// Try to match Instruction "I" as a comparison against a constant and
484  /// populates the array Vals with the set of values that match (or do not
485  /// match depending on isEQ).
486  /// Return false on failure. On success, the Value the comparison matched
487  /// against is placed in CompValue.
488  /// If CompValue is already set, the function is expected to fail if a match
489  /// is found but the value compared to is different.
490  bool matchInstruction(Instruction *I, bool isEQ) {
491  // If this is an icmp against a constant, handle this as one of the cases.
492  ICmpInst *ICI;
493  ConstantInt *C;
494  if (!((ICI = dyn_cast<ICmpInst>(I)) &&
495  (C = GetConstantInt(I->getOperand(1), DL)))) {
496  return false;
497  }
498 
499  Value *RHSVal;
500  const APInt *RHSC;
501 
502  // Pattern match a special case
503  // (x & ~2^z) == y --> x == y || x == y|2^z
504  // This undoes a transformation done by instcombine to fuse 2 compares.
505  if (ICI->getPredicate() == (isEQ ? ICmpInst::ICMP_EQ : ICmpInst::ICMP_NE)) {
506  // It's a little bit hard to see why the following transformations are
507  // correct. Here is a CVC3 program to verify them for 64-bit values:
508 
509  /*
510  ONE : BITVECTOR(64) = BVZEROEXTEND(0bin1, 63);
511  x : BITVECTOR(64);
512  y : BITVECTOR(64);
513  z : BITVECTOR(64);
514  mask : BITVECTOR(64) = BVSHL(ONE, z);
515  QUERY( (y & ~mask = y) =>
516  ((x & ~mask = y) <=> (x = y OR x = (y | mask)))
517  );
518  QUERY( (y | mask = y) =>
519  ((x | mask = y) <=> (x = y OR x = (y & ~mask)))
520  );
521  */
522 
523  // Please note that each pattern must be a dual implication (<--> or
524  // iff). One directional implication can create spurious matches. If the
525  // implication is only one-way, an unsatisfiable condition on the left
526  // side can imply a satisfiable condition on the right side. Dual
527  // implication ensures that satisfiable conditions are transformed to
528  // other satisfiable conditions and unsatisfiable conditions are
529  // transformed to other unsatisfiable conditions.
530 
531  // Here is a concrete example of a unsatisfiable condition on the left
532  // implying a satisfiable condition on the right:
533  //
534  // mask = (1 << z)
535  // (x & ~mask) == y --> (x == y || x == (y | mask))
536  //
537  // Substituting y = 3, z = 0 yields:
538  // (x & -2) == 3 --> (x == 3 || x == 2)
539 
540  // Pattern match a special case:
541  /*
542  QUERY( (y & ~mask = y) =>
543  ((x & ~mask = y) <=> (x = y OR x = (y | mask)))
544  );
545  */
546  if (match(ICI->getOperand(0),
547  m_And(m_Value(RHSVal), m_APInt(RHSC)))) {
548  APInt Mask = ~*RHSC;
549  if (Mask.isPowerOf2() && (C->getValue() & ~Mask) == C->getValue()) {
550  // If we already have a value for the switch, it has to match!
551  if (!setValueOnce(RHSVal))
552  return false;
553 
554  Vals.push_back(C);
555  Vals.push_back(
557  C->getValue() | Mask));
558  UsedICmps++;
559  return true;
560  }
561  }
562 
563  // Pattern match a special case:
564  /*
565  QUERY( (y | mask = y) =>
566  ((x | mask = y) <=> (x = y OR x = (y & ~mask)))
567  );
568  */
569  if (match(ICI->getOperand(0),
570  m_Or(m_Value(RHSVal), m_APInt(RHSC)))) {
571  APInt Mask = *RHSC;
572  if (Mask.isPowerOf2() && (C->getValue() | Mask) == C->getValue()) {
573  // If we already have a value for the switch, it has to match!
574  if (!setValueOnce(RHSVal))
575  return false;
576 
577  Vals.push_back(C);
579  C->getValue() & ~Mask));
580  UsedICmps++;
581  return true;
582  }
583  }
584 
585  // If we already have a value for the switch, it has to match!
586  if (!setValueOnce(ICI->getOperand(0)))
587  return false;
588 
589  UsedICmps++;
590  Vals.push_back(C);
591  return ICI->getOperand(0);
592  }
593 
594  // If we have "x ult 3", for example, then we can add 0,1,2 to the set.
596  ICI->getPredicate(), C->getValue());
597 
598  // Shift the range if the compare is fed by an add. This is the range
599  // compare idiom as emitted by instcombine.
600  Value *CandidateVal = I->getOperand(0);
601  if (match(I->getOperand(0), m_Add(m_Value(RHSVal), m_APInt(RHSC)))) {
602  Span = Span.subtract(*RHSC);
603  CandidateVal = RHSVal;
604  }
605 
606  // If this is an and/!= check, then we are looking to build the set of
607  // value that *don't* pass the and chain. I.e. to turn "x ugt 2" into
608  // x != 0 && x != 1.
609  if (!isEQ)
610  Span = Span.inverse();
611 
612  // If there are a ton of values, we don't want to make a ginormous switch.
613  if (Span.isSizeLargerThan(8) || Span.isEmptySet()) {
614  return false;
615  }
616 
617  // If we already have a value for the switch, it has to match!
618  if (!setValueOnce(CandidateVal))
619  return false;
620 
621  // Add all values from the range to the set
622  for (APInt Tmp = Span.getLower(); Tmp != Span.getUpper(); ++Tmp)
623  Vals.push_back(ConstantInt::get(I->getContext(), Tmp));
624 
625  UsedICmps++;
626  return true;
627  }
628 
629  /// Given a potentially 'or'd or 'and'd together collection of icmp
630  /// eq/ne/lt/gt instructions that compare a value against a constant, extract
631  /// the value being compared, and stick the list constants into the Vals
632  /// vector.
633  /// One "Extra" case is allowed to differ from the other.
634  void gather(Value *V) {
635  bool isEQ = (cast<Instruction>(V)->getOpcode() == Instruction::Or);
636 
637  // Keep a stack (SmallVector for efficiency) for depth-first traversal
639  SmallPtrSet<Value *, 8> Visited;
640 
641  // Initialize
642  Visited.insert(V);
643  DFT.push_back(V);
644 
645  while (!DFT.empty()) {
646  V = DFT.pop_back_val();
647 
648  if (Instruction *I = dyn_cast<Instruction>(V)) {
649  // If it is a || (or && depending on isEQ), process the operands.
650  if (I->getOpcode() == (isEQ ? Instruction::Or : Instruction::And)) {
651  if (Visited.insert(I->getOperand(1)).second)
652  DFT.push_back(I->getOperand(1));
653  if (Visited.insert(I->getOperand(0)).second)
654  DFT.push_back(I->getOperand(0));
655  continue;
656  }
657 
658  // Try to match the current instruction
659  if (matchInstruction(I, isEQ))
660  // Match succeed, continue the loop
661  continue;
662  }
663 
664  // One element of the sequence of || (or &&) could not be match as a
665  // comparison against the same value as the others.
666  // We allow only one "Extra" case to be checked before the switch
667  if (!Extra) {
668  Extra = V;
669  continue;
670  }
671  // Failed to parse a proper sequence, abort now
672  CompValue = nullptr;
673  break;
674  }
675  }
676 };
677 
678 } // end anonymous namespace
679 
681  MemorySSAUpdater *MSSAU = nullptr) {
682  Instruction *Cond = nullptr;
683  if (SwitchInst *SI = dyn_cast<SwitchInst>(TI)) {
684  Cond = dyn_cast<Instruction>(SI->getCondition());
685  } else if (BranchInst *BI = dyn_cast<BranchInst>(TI)) {
686  if (BI->isConditional())
687  Cond = dyn_cast<Instruction>(BI->getCondition());
688  } else if (IndirectBrInst *IBI = dyn_cast<IndirectBrInst>(TI)) {
689  Cond = dyn_cast<Instruction>(IBI->getAddress());
690  }
691 
692  TI->eraseFromParent();
693  if (Cond)
694  RecursivelyDeleteTriviallyDeadInstructions(Cond, nullptr, MSSAU);
695 }
696 
697 /// Return true if the specified terminator checks
698 /// to see if a value is equal to constant integer value.
699 Value *SimplifyCFGOpt::isValueEqualityComparison(Instruction *TI) {
700  Value *CV = nullptr;
701  if (SwitchInst *SI = dyn_cast<SwitchInst>(TI)) {
702  // Do not permit merging of large switch instructions into their
703  // predecessors unless there is only one predecessor.
704  if (!SI->getParent()->hasNPredecessorsOrMore(128 / SI->getNumSuccessors()))
705  CV = SI->getCondition();
706  } else if (BranchInst *BI = dyn_cast<BranchInst>(TI))
707  if (BI->isConditional() && BI->getCondition()->hasOneUse())
708  if (ICmpInst *ICI = dyn_cast<ICmpInst>(BI->getCondition())) {
709  if (ICI->isEquality() && GetConstantInt(ICI->getOperand(1), DL))
710  CV = ICI->getOperand(0);
711  }
712 
713  // Unwrap any lossless ptrtoint cast.
714  if (CV) {
715  if (PtrToIntInst *PTII = dyn_cast<PtrToIntInst>(CV)) {
716  Value *Ptr = PTII->getPointerOperand();
717  if (PTII->getType() == DL.getIntPtrType(Ptr->getType()))
718  CV = Ptr;
719  }
720  }
721  return CV;
722 }
723 
724 /// Given a value comparison instruction,
725 /// decode all of the 'cases' that it represents and return the 'default' block.
726 BasicBlock *SimplifyCFGOpt::GetValueEqualityComparisonCases(
727  Instruction *TI, std::vector<ValueEqualityComparisonCase> &Cases) {
728  if (SwitchInst *SI = dyn_cast<SwitchInst>(TI)) {
729  Cases.reserve(SI->getNumCases());
730  for (auto Case : SI->cases())
731  Cases.push_back(ValueEqualityComparisonCase(Case.getCaseValue(),
732  Case.getCaseSuccessor()));
733  return SI->getDefaultDest();
734  }
735 
736  BranchInst *BI = cast<BranchInst>(TI);
737  ICmpInst *ICI = cast<ICmpInst>(BI->getCondition());
738  BasicBlock *Succ = BI->getSuccessor(ICI->getPredicate() == ICmpInst::ICMP_NE);
739  Cases.push_back(ValueEqualityComparisonCase(
740  GetConstantInt(ICI->getOperand(1), DL), Succ));
741  return BI->getSuccessor(ICI->getPredicate() == ICmpInst::ICMP_EQ);
742 }
743 
744 /// Given a vector of bb/value pairs, remove any entries
745 /// in the list that match the specified block.
746 static void
748  std::vector<ValueEqualityComparisonCase> &Cases) {
749  Cases.erase(std::remove(Cases.begin(), Cases.end(), BB), Cases.end());
750 }
751 
752 /// Return true if there are any keys in C1 that exist in C2 as well.
753 static bool ValuesOverlap(std::vector<ValueEqualityComparisonCase> &C1,
754  std::vector<ValueEqualityComparisonCase> &C2) {
755  std::vector<ValueEqualityComparisonCase> *V1 = &C1, *V2 = &C2;
756 
757  // Make V1 be smaller than V2.
758  if (V1->size() > V2->size())
759  std::swap(V1, V2);
760 
761  if (V1->empty())
762  return false;
763  if (V1->size() == 1) {
764  // Just scan V2.
765  ConstantInt *TheVal = (*V1)[0].Value;
766  for (unsigned i = 0, e = V2->size(); i != e; ++i)
767  if (TheVal == (*V2)[i].Value)
768  return true;
769  }
770 
771  // Otherwise, just sort both lists and compare element by element.
772  array_pod_sort(V1->begin(), V1->end());
773  array_pod_sort(V2->begin(), V2->end());
774  unsigned i1 = 0, i2 = 0, e1 = V1->size(), e2 = V2->size();
775  while (i1 != e1 && i2 != e2) {
776  if ((*V1)[i1].Value == (*V2)[i2].Value)
777  return true;
778  if ((*V1)[i1].Value < (*V2)[i2].Value)
779  ++i1;
780  else
781  ++i2;
782  }
783  return false;
784 }
785 
786 // Set branch weights on SwitchInst. This sets the metadata if there is at
787 // least one non-zero weight.
789  // Check that there is at least one non-zero weight. Otherwise, pass
790  // nullptr to setMetadata which will erase the existing metadata.
791  MDNode *N = nullptr;
792  if (llvm::any_of(Weights, [](uint32_t W) { return W != 0; }))
793  N = MDBuilder(SI->getParent()->getContext()).createBranchWeights(Weights);
794  SI->setMetadata(LLVMContext::MD_prof, N);
795 }
796 
797 // Similar to the above, but for branch and select instructions that take
798 // exactly 2 weights.
799 static void setBranchWeights(Instruction *I, uint32_t TrueWeight,
800  uint32_t FalseWeight) {
801  assert(isa<BranchInst>(I) || isa<SelectInst>(I));
802  // Check that there is at least one non-zero weight. Otherwise, pass
803  // nullptr to setMetadata which will erase the existing metadata.
804  MDNode *N = nullptr;
805  if (TrueWeight || FalseWeight)
806  N = MDBuilder(I->getParent()->getContext())
807  .createBranchWeights(TrueWeight, FalseWeight);
808  I->setMetadata(LLVMContext::MD_prof, N);
809 }
810 
811 /// If TI is known to be a terminator instruction and its block is known to
812 /// only have a single predecessor block, check to see if that predecessor is
813 /// also a value comparison with the same value, and if that comparison
814 /// determines the outcome of this comparison. If so, simplify TI. This does a
815 /// very limited form of jump threading.
816 bool SimplifyCFGOpt::SimplifyEqualityComparisonWithOnlyPredecessor(
817  Instruction *TI, BasicBlock *Pred, IRBuilder<> &Builder) {
818  Value *PredVal = isValueEqualityComparison(Pred->getTerminator());
819  if (!PredVal)
820  return false; // Not a value comparison in predecessor.
821 
822  Value *ThisVal = isValueEqualityComparison(TI);
823  assert(ThisVal && "This isn't a value comparison!!");
824  if (ThisVal != PredVal)
825  return false; // Different predicates.
826 
827  // TODO: Preserve branch weight metadata, similarly to how
828  // FoldValueComparisonIntoPredecessors preserves it.
829 
830  // Find out information about when control will move from Pred to TI's block.
831  std::vector<ValueEqualityComparisonCase> PredCases;
832  BasicBlock *PredDef =
833  GetValueEqualityComparisonCases(Pred->getTerminator(), PredCases);
834  EliminateBlockCases(PredDef, PredCases); // Remove default from cases.
835 
836  // Find information about how control leaves this block.
837  std::vector<ValueEqualityComparisonCase> ThisCases;
838  BasicBlock *ThisDef = GetValueEqualityComparisonCases(TI, ThisCases);
839  EliminateBlockCases(ThisDef, ThisCases); // Remove default from cases.
840 
841  // If TI's block is the default block from Pred's comparison, potentially
842  // simplify TI based on this knowledge.
843  if (PredDef == TI->getParent()) {
844  // If we are here, we know that the value is none of those cases listed in
845  // PredCases. If there are any cases in ThisCases that are in PredCases, we
846  // can simplify TI.
847  if (!ValuesOverlap(PredCases, ThisCases))
848  return false;
849 
850  if (isa<BranchInst>(TI)) {
851  // Okay, one of the successors of this condbr is dead. Convert it to a
852  // uncond br.
853  assert(ThisCases.size() == 1 && "Branch can only have one case!");
854  // Insert the new branch.
855  Instruction *NI = Builder.CreateBr(ThisDef);
856  (void)NI;
857 
858  // Remove PHI node entries for the dead edge.
859  ThisCases[0].Dest->removePredecessor(TI->getParent());
860 
861  LLVM_DEBUG(dbgs() << "Threading pred instr: " << *Pred->getTerminator()
862  << "Through successor TI: " << *TI << "Leaving: " << *NI
863  << "\n");
864 
866  return true;
867  }
868 
869  SwitchInstProfUpdateWrapper SI = *cast<SwitchInst>(TI);
870  // Okay, TI has cases that are statically dead, prune them away.
871  SmallPtrSet<Constant *, 16> DeadCases;
872  for (unsigned i = 0, e = PredCases.size(); i != e; ++i)
873  DeadCases.insert(PredCases[i].Value);
874 
875  LLVM_DEBUG(dbgs() << "Threading pred instr: " << *Pred->getTerminator()
876  << "Through successor TI: " << *TI);
877 
878  for (SwitchInst::CaseIt i = SI->case_end(), e = SI->case_begin(); i != e;) {
879  --i;
880  if (DeadCases.count(i->getCaseValue())) {
881  i->getCaseSuccessor()->removePredecessor(TI->getParent());
882  SI.removeCase(i);
883  }
884  }
885  LLVM_DEBUG(dbgs() << "Leaving: " << *TI << "\n");
886  return true;
887  }
888 
889  // Otherwise, TI's block must correspond to some matched value. Find out
890  // which value (or set of values) this is.
891  ConstantInt *TIV = nullptr;
892  BasicBlock *TIBB = TI->getParent();
893  for (unsigned i = 0, e = PredCases.size(); i != e; ++i)
894  if (PredCases[i].Dest == TIBB) {
895  if (TIV)
896  return false; // Cannot handle multiple values coming to this block.
897  TIV = PredCases[i].Value;
898  }
899  assert(TIV && "No edge from pred to succ?");
900 
901  // Okay, we found the one constant that our value can be if we get into TI's
902  // BB. Find out which successor will unconditionally be branched to.
903  BasicBlock *TheRealDest = nullptr;
904  for (unsigned i = 0, e = ThisCases.size(); i != e; ++i)
905  if (ThisCases[i].Value == TIV) {
906  TheRealDest = ThisCases[i].Dest;
907  break;
908  }
909 
910  // If not handled by any explicit cases, it is handled by the default case.
911  if (!TheRealDest)
912  TheRealDest = ThisDef;
913 
914  // Remove PHI node entries for dead edges.
915  BasicBlock *CheckEdge = TheRealDest;
916  for (BasicBlock *Succ : successors(TIBB))
917  if (Succ != CheckEdge)
918  Succ->removePredecessor(TIBB);
919  else
920  CheckEdge = nullptr;
921 
922  // Insert the new branch.
923  Instruction *NI = Builder.CreateBr(TheRealDest);
924  (void)NI;
925 
926  LLVM_DEBUG(dbgs() << "Threading pred instr: " << *Pred->getTerminator()
927  << "Through successor TI: " << *TI << "Leaving: " << *NI
928  << "\n");
929 
931  return true;
932 }
933 
934 namespace {
935 
936 /// This class implements a stable ordering of constant
937 /// integers that does not depend on their address. This is important for
938 /// applications that sort ConstantInt's to ensure uniqueness.
939 struct ConstantIntOrdering {
940  bool operator()(const ConstantInt *LHS, const ConstantInt *RHS) const {
941  return LHS->getValue().ult(RHS->getValue());
942  }
943 };
944 
945 } // end anonymous namespace
946 
947 static int ConstantIntSortPredicate(ConstantInt *const *P1,
948  ConstantInt *const *P2) {
949  const ConstantInt *LHS = *P1;
950  const ConstantInt *RHS = *P2;
951  if (LHS == RHS)
952  return 0;
953  return LHS->getValue().ult(RHS->getValue()) ? 1 : -1;
954 }
955 
956 static inline bool HasBranchWeights(const Instruction *I) {
957  MDNode *ProfMD = I->getMetadata(LLVMContext::MD_prof);
958  if (ProfMD && ProfMD->getOperand(0))
959  if (MDString *MDS = dyn_cast<MDString>(ProfMD->getOperand(0)))
960  return MDS->getString().equals("branch_weights");
961 
962  return false;
963 }
964 
965 /// Get Weights of a given terminator, the default weight is at the front
966 /// of the vector. If TI is a conditional eq, we need to swap the branch-weight
967 /// metadata.
969  SmallVectorImpl<uint64_t> &Weights) {
970  MDNode *MD = TI->getMetadata(LLVMContext::MD_prof);
971  assert(MD);
972  for (unsigned i = 1, e = MD->getNumOperands(); i < e; ++i) {
973  ConstantInt *CI = mdconst::extract<ConstantInt>(MD->getOperand(i));
974  Weights.push_back(CI->getValue().getZExtValue());
975  }
976 
977  // If TI is a conditional eq, the default case is the false case,
978  // and the corresponding branch-weight data is at index 2. We swap the
979  // default weight to be the first entry.
980  if (BranchInst *BI = dyn_cast<BranchInst>(TI)) {
981  assert(Weights.size() == 2);
982  ICmpInst *ICI = cast<ICmpInst>(BI->getCondition());
983  if (ICI->getPredicate() == ICmpInst::ICMP_EQ)
984  std::swap(Weights.front(), Weights.back());
985  }
986 }
987 
988 /// Keep halving the weights until all can fit in uint32_t.
989 static void FitWeights(MutableArrayRef<uint64_t> Weights) {
990  uint64_t Max = *std::max_element(Weights.begin(), Weights.end());
991  if (Max > UINT_MAX) {
992  unsigned Offset = 32 - countLeadingZeros(Max);
993  for (uint64_t &I : Weights)
994  I >>= Offset;
995  }
996 }
997 
998 /// The specified terminator is a value equality comparison instruction
999 /// (either a switch or a branch on "X == c").
1000 /// See if any of the predecessors of the terminator block are value comparisons
1001 /// on the same value. If so, and if safe to do so, fold them together.
1002 bool SimplifyCFGOpt::FoldValueComparisonIntoPredecessors(Instruction *TI,
1003  IRBuilder<> &Builder) {
1004  BasicBlock *BB = TI->getParent();
1005  Value *CV = isValueEqualityComparison(TI); // CondVal
1006  assert(CV && "Not a comparison?");
1007  bool Changed = false;
1008 
1010  while (!Preds.empty()) {
1011  BasicBlock *Pred = Preds.pop_back_val();
1012 
1013  // See if the predecessor is a comparison with the same value.
1014  Instruction *PTI = Pred->getTerminator();
1015  Value *PCV = isValueEqualityComparison(PTI); // PredCondVal
1016 
1017  if (PCV == CV && TI != PTI) {
1018  SmallSetVector<BasicBlock*, 4> FailBlocks;
1019  if (!SafeToMergeTerminators(TI, PTI, &FailBlocks)) {
1020  for (auto *Succ : FailBlocks) {
1021  if (!SplitBlockPredecessors(Succ, TI->getParent(), ".fold.split"))
1022  return false;
1023  }
1024  }
1025 
1026  // Figure out which 'cases' to copy from SI to PSI.
1027  std::vector<ValueEqualityComparisonCase> BBCases;
1028  BasicBlock *BBDefault = GetValueEqualityComparisonCases(TI, BBCases);
1029 
1030  std::vector<ValueEqualityComparisonCase> PredCases;
1031  BasicBlock *PredDefault = GetValueEqualityComparisonCases(PTI, PredCases);
1032 
1033  // Based on whether the default edge from PTI goes to BB or not, fill in
1034  // PredCases and PredDefault with the new switch cases we would like to
1035  // build.
1036  SmallVector<BasicBlock *, 8> NewSuccessors;
1037 
1038  // Update the branch weight metadata along the way
1039  SmallVector<uint64_t, 8> Weights;
1040  bool PredHasWeights = HasBranchWeights(PTI);
1041  bool SuccHasWeights = HasBranchWeights(TI);
1042 
1043  if (PredHasWeights) {
1044  GetBranchWeights(PTI, Weights);
1045  // branch-weight metadata is inconsistent here.
1046  if (Weights.size() != 1 + PredCases.size())
1047  PredHasWeights = SuccHasWeights = false;
1048  } else if (SuccHasWeights)
1049  // If there are no predecessor weights but there are successor weights,
1050  // populate Weights with 1, which will later be scaled to the sum of
1051  // successor's weights
1052  Weights.assign(1 + PredCases.size(), 1);
1053 
1054  SmallVector<uint64_t, 8> SuccWeights;
1055  if (SuccHasWeights) {
1056  GetBranchWeights(TI, SuccWeights);
1057  // branch-weight metadata is inconsistent here.
1058  if (SuccWeights.size() != 1 + BBCases.size())
1059  PredHasWeights = SuccHasWeights = false;
1060  } else if (PredHasWeights)
1061  SuccWeights.assign(1 + BBCases.size(), 1);
1062 
1063  if (PredDefault == BB) {
1064  // If this is the default destination from PTI, only the edges in TI
1065  // that don't occur in PTI, or that branch to BB will be activated.
1066  std::set<ConstantInt *, ConstantIntOrdering> PTIHandled;
1067  for (unsigned i = 0, e = PredCases.size(); i != e; ++i)
1068  if (PredCases[i].Dest != BB)
1069  PTIHandled.insert(PredCases[i].Value);
1070  else {
1071  // The default destination is BB, we don't need explicit targets.
1072  std::swap(PredCases[i], PredCases.back());
1073 
1074  if (PredHasWeights || SuccHasWeights) {
1075  // Increase weight for the default case.
1076  Weights[0] += Weights[i + 1];
1077  std::swap(Weights[i + 1], Weights.back());
1078  Weights.pop_back();
1079  }
1080 
1081  PredCases.pop_back();
1082  --i;
1083  --e;
1084  }
1085 
1086  // Reconstruct the new switch statement we will be building.
1087  if (PredDefault != BBDefault) {
1088  PredDefault->removePredecessor(Pred);
1089  PredDefault = BBDefault;
1090  NewSuccessors.push_back(BBDefault);
1091  }
1092 
1093  unsigned CasesFromPred = Weights.size();
1094  uint64_t ValidTotalSuccWeight = 0;
1095  for (unsigned i = 0, e = BBCases.size(); i != e; ++i)
1096  if (!PTIHandled.count(BBCases[i].Value) &&
1097  BBCases[i].Dest != BBDefault) {
1098  PredCases.push_back(BBCases[i]);
1099  NewSuccessors.push_back(BBCases[i].Dest);
1100  if (SuccHasWeights || PredHasWeights) {
1101  // The default weight is at index 0, so weight for the ith case
1102  // should be at index i+1. Scale the cases from successor by
1103  // PredDefaultWeight (Weights[0]).
1104  Weights.push_back(Weights[0] * SuccWeights[i + 1]);
1105  ValidTotalSuccWeight += SuccWeights[i + 1];
1106  }
1107  }
1108 
1109  if (SuccHasWeights || PredHasWeights) {
1110  ValidTotalSuccWeight += SuccWeights[0];
1111  // Scale the cases from predecessor by ValidTotalSuccWeight.
1112  for (unsigned i = 1; i < CasesFromPred; ++i)
1113  Weights[i] *= ValidTotalSuccWeight;
1114  // Scale the default weight by SuccDefaultWeight (SuccWeights[0]).
1115  Weights[0] *= SuccWeights[0];
1116  }
1117  } else {
1118  // If this is not the default destination from PSI, only the edges
1119  // in SI that occur in PSI with a destination of BB will be
1120  // activated.
1121  std::set<ConstantInt *, ConstantIntOrdering> PTIHandled;
1122  std::map<ConstantInt *, uint64_t> WeightsForHandled;
1123  for (unsigned i = 0, e = PredCases.size(); i != e; ++i)
1124  if (PredCases[i].Dest == BB) {
1125  PTIHandled.insert(PredCases[i].Value);
1126 
1127  if (PredHasWeights || SuccHasWeights) {
1128  WeightsForHandled[PredCases[i].Value] = Weights[i + 1];
1129  std::swap(Weights[i + 1], Weights.back());
1130  Weights.pop_back();
1131  }
1132 
1133  std::swap(PredCases[i], PredCases.back());
1134  PredCases.pop_back();
1135  --i;
1136  --e;
1137  }
1138 
1139  // Okay, now we know which constants were sent to BB from the
1140  // predecessor. Figure out where they will all go now.
1141  for (unsigned i = 0, e = BBCases.size(); i != e; ++i)
1142  if (PTIHandled.count(BBCases[i].Value)) {
1143  // If this is one we are capable of getting...
1144  if (PredHasWeights || SuccHasWeights)
1145  Weights.push_back(WeightsForHandled[BBCases[i].Value]);
1146  PredCases.push_back(BBCases[i]);
1147  NewSuccessors.push_back(BBCases[i].Dest);
1148  PTIHandled.erase(
1149  BBCases[i].Value); // This constant is taken care of
1150  }
1151 
1152  // If there are any constants vectored to BB that TI doesn't handle,
1153  // they must go to the default destination of TI.
1154  for (ConstantInt *I : PTIHandled) {
1155  if (PredHasWeights || SuccHasWeights)
1156  Weights.push_back(WeightsForHandled[I]);
1157  PredCases.push_back(ValueEqualityComparisonCase(I, BBDefault));
1158  NewSuccessors.push_back(BBDefault);
1159  }
1160  }
1161 
1162  // Okay, at this point, we know which new successor Pred will get. Make
1163  // sure we update the number of entries in the PHI nodes for these
1164  // successors.
1165  for (BasicBlock *NewSuccessor : NewSuccessors)
1166  AddPredecessorToBlock(NewSuccessor, Pred, BB);
1167 
1168  Builder.SetInsertPoint(PTI);
1169  // Convert pointer to int before we switch.
1170  if (CV->getType()->isPointerTy()) {
1171  CV = Builder.CreatePtrToInt(CV, DL.getIntPtrType(CV->getType()),
1172  "magicptr");
1173  }
1174 
1175  // Now that the successors are updated, create the new Switch instruction.
1176  SwitchInst *NewSI =
1177  Builder.CreateSwitch(CV, PredDefault, PredCases.size());
1178  NewSI->setDebugLoc(PTI->getDebugLoc());
1179  for (ValueEqualityComparisonCase &V : PredCases)
1180  NewSI->addCase(V.Value, V.Dest);
1181 
1182  if (PredHasWeights || SuccHasWeights) {
1183  // Halve the weights if any of them cannot fit in an uint32_t
1184  FitWeights(Weights);
1185 
1186  SmallVector<uint32_t, 8> MDWeights(Weights.begin(), Weights.end());
1187 
1188  setBranchWeights(NewSI, MDWeights);
1189  }
1190 
1192 
1193  // Okay, last check. If BB is still a successor of PSI, then we must
1194  // have an infinite loop case. If so, add an infinitely looping block
1195  // to handle the case to preserve the behavior of the code.
1196  BasicBlock *InfLoopBlock = nullptr;
1197  for (unsigned i = 0, e = NewSI->getNumSuccessors(); i != e; ++i)
1198  if (NewSI->getSuccessor(i) == BB) {
1199  if (!InfLoopBlock) {
1200  // Insert it at the end of the function, because it's either code,
1201  // or it won't matter if it's hot. :)
1202  InfLoopBlock = BasicBlock::Create(BB->getContext(), "infloop",
1203  BB->getParent());
1204  BranchInst::Create(InfLoopBlock, InfLoopBlock);
1205  }
1206  NewSI->setSuccessor(i, InfLoopBlock);
1207  }
1208 
1209  Changed = true;
1210  }
1211  }
1212  return Changed;
1213 }
1214 
1215 // If we would need to insert a select that uses the value of this invoke
1216 // (comments in HoistThenElseCodeToIf explain why we would need to do this), we
1217 // can't hoist the invoke, as there is nowhere to put the select in this case.
1219  Instruction *I1, Instruction *I2) {
1220  for (BasicBlock *Succ : successors(BB1)) {
1221  for (const PHINode &PN : Succ->phis()) {
1222  Value *BB1V = PN.getIncomingValueForBlock(BB1);
1223  Value *BB2V = PN.getIncomingValueForBlock(BB2);
1224  if (BB1V != BB2V && (BB1V == I1 || BB2V == I2)) {
1225  return false;
1226  }
1227  }
1228  }
1229  return true;
1230 }
1231 
1233 
1234 /// Given a conditional branch that goes to BB1 and BB2, hoist any common code
1235 /// in the two blocks up into the branch block. The caller of this function
1236 /// guarantees that BI's block dominates BB1 and BB2.
1238  const TargetTransformInfo &TTI) {
1239  // This does very trivial matching, with limited scanning, to find identical
1240  // instructions in the two blocks. In particular, we don't want to get into
1241  // O(M*N) situations here where M and N are the sizes of BB1 and BB2. As
1242  // such, we currently just scan for obviously identical instructions in an
1243  // identical order.
1244  BasicBlock *BB1 = BI->getSuccessor(0); // The true destination.
1245  BasicBlock *BB2 = BI->getSuccessor(1); // The false destination
1246 
1247  BasicBlock::iterator BB1_Itr = BB1->begin();
1248  BasicBlock::iterator BB2_Itr = BB2->begin();
1249 
1250  Instruction *I1 = &*BB1_Itr++, *I2 = &*BB2_Itr++;
1251  // Skip debug info if it is not identical.
1254  if (!DBI1 || !DBI2 || !DBI1->isIdenticalToWhenDefined(DBI2)) {
1255  while (isa<DbgInfoIntrinsic>(I1))
1256  I1 = &*BB1_Itr++;
1257  while (isa<DbgInfoIntrinsic>(I2))
1258  I2 = &*BB2_Itr++;
1259  }
1260  // FIXME: Can we define a safety predicate for CallBr?
1261  if (isa<PHINode>(I1) || !I1->isIdenticalToWhenDefined(I2) ||
1262  (isa<InvokeInst>(I1) && !isSafeToHoistInvoke(BB1, BB2, I1, I2)) ||
1263  isa<CallBrInst>(I1))
1264  return false;
1265 
1266  BasicBlock *BIParent = BI->getParent();
1267 
1268  bool Changed = false;
1269  do {
1270  // If we are hoisting the terminator instruction, don't move one (making a
1271  // broken BB), instead clone it, and remove BI.
1272  if (I1->isTerminator())
1273  goto HoistTerminator;
1274 
1275  // If we're going to hoist a call, make sure that the two instructions we're
1276  // commoning/hoisting are both marked with musttail, or neither of them is
1277  // marked as such. Otherwise, we might end up in a situation where we hoist
1278  // from a block where the terminator is a `ret` to a block where the terminator
1279  // is a `br`, and `musttail` calls expect to be followed by a return.
1280  auto *C1 = dyn_cast<CallInst>(I1);
1281  auto *C2 = dyn_cast<CallInst>(I2);
1282  if (C1 && C2)
1283  if (C1->isMustTailCall() != C2->isMustTailCall())
1284  return Changed;
1285 
1286  if (!TTI.isProfitableToHoist(I1) || !TTI.isProfitableToHoist(I2))
1287  return Changed;
1288 
1289  if (isa<DbgInfoIntrinsic>(I1) || isa<DbgInfoIntrinsic>(I2)) {
1290  assert (isa<DbgInfoIntrinsic>(I1) && isa<DbgInfoIntrinsic>(I2));
1291  // The debug location is an integral part of a debug info intrinsic
1292  // and can't be separated from it or replaced. Instead of attempting
1293  // to merge locations, simply hoist both copies of the intrinsic.
1294  BIParent->getInstList().splice(BI->getIterator(),
1295  BB1->getInstList(), I1);
1296  BIParent->getInstList().splice(BI->getIterator(),
1297  BB2->getInstList(), I2);
1298  Changed = true;
1299  } else {
1300  // For a normal instruction, we just move one to right before the branch,
1301  // then replace all uses of the other with the first. Finally, we remove
1302  // the now redundant second instruction.
1303  BIParent->getInstList().splice(BI->getIterator(),
1304  BB1->getInstList(), I1);
1305  if (!I2->use_empty())
1306  I2->replaceAllUsesWith(I1);
1307  I1->andIRFlags(I2);
1308  unsigned KnownIDs[] = {LLVMContext::MD_tbaa,
1309  LLVMContext::MD_range,
1310  LLVMContext::MD_fpmath,
1311  LLVMContext::MD_invariant_load,
1312  LLVMContext::MD_nonnull,
1313  LLVMContext::MD_invariant_group,
1314  LLVMContext::MD_align,
1315  LLVMContext::MD_dereferenceable,
1316  LLVMContext::MD_dereferenceable_or_null,
1317  LLVMContext::MD_mem_parallel_loop_access,
1318  LLVMContext::MD_access_group,
1319  LLVMContext::MD_preserve_access_index};
1320  combineMetadata(I1, I2, KnownIDs, true);
1321 
1322  // I1 and I2 are being combined into a single instruction. Its debug
1323  // location is the merged locations of the original instructions.
1324  I1->applyMergedLocation(I1->getDebugLoc(), I2->getDebugLoc());
1325 
1326  I2->eraseFromParent();
1327  Changed = true;
1328  }
1329 
1330  I1 = &*BB1_Itr++;
1331  I2 = &*BB2_Itr++;
1332  // Skip debug info if it is not identical.
1335  if (!DBI1 || !DBI2 || !DBI1->isIdenticalToWhenDefined(DBI2)) {
1336  while (isa<DbgInfoIntrinsic>(I1))
1337  I1 = &*BB1_Itr++;
1338  while (isa<DbgInfoIntrinsic>(I2))
1339  I2 = &*BB2_Itr++;
1340  }
1341  } while (I1->isIdenticalToWhenDefined(I2));
1342 
1343  return true;
1344 
1345 HoistTerminator:
1346  // It may not be possible to hoist an invoke.
1347  // FIXME: Can we define a safety predicate for CallBr?
1348  if (isa<InvokeInst>(I1) && !isSafeToHoistInvoke(BB1, BB2, I1, I2))
1349  return Changed;
1350 
1351  // TODO: callbr hoisting currently disabled pending further study.
1352  if (isa<CallBrInst>(I1))
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  // Ensure terminator gets a debug location, even an unknown one, in case
1385  // it involves inlinable calls.
1386  NT->applyMergedLocation(I1->getDebugLoc(), I2->getDebugLoc());
1387 
1388  // PHIs created below will adopt NT's merged DebugLoc.
1389  IRBuilder<NoFolder> Builder(NT);
1390 
1391  // Hoisting one of the terminators from our successor is a great thing.
1392  // Unfortunately, the successors of the if/else blocks may have PHI nodes in
1393  // them. If they do, all PHI entries for BB1/BB2 must agree for all PHI
1394  // nodes, so we insert select instruction to compute the final result.
1395  std::map<std::pair<Value *, Value *>, SelectInst *> InsertedSelects;
1396  for (BasicBlock *Succ : successors(BB1)) {
1397  for (PHINode &PN : Succ->phis()) {
1398  Value *BB1V = PN.getIncomingValueForBlock(BB1);
1399  Value *BB2V = PN.getIncomingValueForBlock(BB2);
1400  if (BB1V == BB2V)
1401  continue;
1402 
1403  // These values do not agree. Insert a select instruction before NT
1404  // that determines the right value.
1405  SelectInst *&SI = InsertedSelects[std::make_pair(BB1V, BB2V)];
1406  if (!SI)
1407  SI = cast<SelectInst>(
1408  Builder.CreateSelect(BI->getCondition(), BB1V, BB2V,
1409  BB1V->getName() + "." + BB2V->getName(), BI));
1410 
1411  // Make the PHI node use the select for all incoming values for BB1/BB2
1412  for (unsigned i = 0, e = PN.getNumIncomingValues(); i != e; ++i)
1413  if (PN.getIncomingBlock(i) == BB1 || PN.getIncomingBlock(i) == BB2)
1414  PN.setIncomingValue(i, SI);
1415  }
1416  }
1417 
1418  // Update any PHI nodes in our new successors.
1419  for (BasicBlock *Succ : successors(BB1))
1420  AddPredecessorToBlock(Succ, BIParent, BB1);
1421 
1423  return true;
1424 }
1425 
1426 // Check lifetime markers.
1427 static bool isLifeTimeMarker(const Instruction *I) {
1428  if (auto II = dyn_cast<IntrinsicInst>(I)) {
1429  switch (II->getIntrinsicID()) {
1430  default:
1431  break;
1432  case Intrinsic::lifetime_start:
1433  case Intrinsic::lifetime_end:
1434  return true;
1435  }
1436  }
1437  return false;
1438 }
1439 
1440 // All instructions in Insts belong to different blocks that all unconditionally
1441 // branch to a common successor. Analyze each instruction and return true if it
1442 // would be possible to sink them into their successor, creating one common
1443 // instruction instead. For every value that would be required to be provided by
1444 // PHI node (because an operand varies in each input block), add to PHIOperands.
1447  DenseMap<Instruction *, SmallVector<Value *, 4>> &PHIOperands) {
1448  // Prune out obviously bad instructions to move. Each instruction must have
1449  // exactly zero or one use, and we check later that use is by a single, common
1450  // PHI instruction in the successor.
1451  bool HasUse = !Insts.front()->user_empty();
1452  for (auto *I : Insts) {
1453  // These instructions may change or break semantics if moved.
1454  if (isa<PHINode>(I) || I->isEHPad() || isa<AllocaInst>(I) ||
1455  I->getType()->isTokenTy())
1456  return false;
1457 
1458  // Conservatively return false if I is an inline-asm instruction. Sinking
1459  // and merging inline-asm instructions can potentially create arguments
1460  // that cannot satisfy the inline-asm constraints.
1461  if (const auto *C = dyn_cast<CallBase>(I))
1462  if (C->isInlineAsm())
1463  return false;
1464 
1465  // Each instruction must have zero or one use.
1466  if (HasUse && !I->hasOneUse())
1467  return false;
1468  if (!HasUse && !I->user_empty())
1469  return false;
1470  }
1471 
1472  const Instruction *I0 = Insts.front();
1473  for (auto *I : Insts)
1474  if (!I->isSameOperationAs(I0))
1475  return false;
1476 
1477  // All instructions in Insts are known to be the same opcode. If they have a
1478  // use, check that the only user is a PHI or in the same block as the
1479  // instruction, because if a user is in the same block as an instruction we're
1480  // contemplating sinking, it must already be determined to be sinkable.
1481  if (HasUse) {
1482  auto *PNUse = dyn_cast<PHINode>(*I0->user_begin());
1483  auto *Succ = I0->getParent()->getTerminator()->getSuccessor(0);
1484  if (!all_of(Insts, [&PNUse,&Succ](const Instruction *I) -> bool {
1485  auto *U = cast<Instruction>(*I->user_begin());
1486  return (PNUse &&
1487  PNUse->getParent() == Succ &&
1488  PNUse->getIncomingValueForBlock(I->getParent()) == I) ||
1489  U->getParent() == I->getParent();
1490  }))
1491  return false;
1492  }
1493 
1494  // Because SROA can't handle speculating stores of selects, try not to sink
1495  // loads, stores or lifetime markers of allocas when we'd have to create a
1496  // PHI for the address operand. Also, because it is likely that loads or
1497  // stores of allocas will disappear when Mem2Reg/SROA is run, don't sink
1498  // them.
1499  // This can cause code churn which can have unintended consequences down
1500  // the line - see https://llvm.org/bugs/show_bug.cgi?id=30244.
1501  // FIXME: This is a workaround for a deficiency in SROA - see
1502  // https://llvm.org/bugs/show_bug.cgi?id=30188
1503  if (isa<StoreInst>(I0) && any_of(Insts, [](const Instruction *I) {
1504  return isa<AllocaInst>(I->getOperand(1)->stripPointerCasts());
1505  }))
1506  return false;
1507  if (isa<LoadInst>(I0) && any_of(Insts, [](const Instruction *I) {
1508  return isa<AllocaInst>(I->getOperand(0)->stripPointerCasts());
1509  }))
1510  return false;
1511  if (isLifeTimeMarker(I0) && any_of(Insts, [](const Instruction *I) {
1512  return isa<AllocaInst>(I->getOperand(1)->stripPointerCasts());
1513  }))
1514  return false;
1515 
1516  for (unsigned OI = 0, OE = I0->getNumOperands(); OI != OE; ++OI) {
1517  if (I0->getOperand(OI)->getType()->isTokenTy())
1518  // Don't touch any operand of token type.
1519  return false;
1520 
1521  auto SameAsI0 = [&I0, OI](const Instruction *I) {
1522  assert(I->getNumOperands() == I0->getNumOperands());
1523  return I->getOperand(OI) == I0->getOperand(OI);
1524  };
1525  if (!all_of(Insts, SameAsI0)) {
1526  if (!canReplaceOperandWithVariable(I0, OI))
1527  // We can't create a PHI from this GEP.
1528  return false;
1529  // Don't create indirect calls! The called value is the final operand.
1530  if (isa<CallBase>(I0) && OI == OE - 1) {
1531  // FIXME: if the call was *already* indirect, we should do this.
1532  return false;
1533  }
1534  for (auto *I : Insts)
1535  PHIOperands[I].push_back(I->getOperand(OI));
1536  }
1537  }
1538  return true;
1539 }
1540 
1541 // Assuming canSinkLastInstruction(Blocks) has returned true, sink the last
1542 // instruction of every block in Blocks to their common successor, commoning
1543 // into one instruction.
1545  auto *BBEnd = Blocks[0]->getTerminator()->getSuccessor(0);
1546 
1547  // canSinkLastInstruction returning true guarantees that every block has at
1548  // least one non-terminator instruction.
1550  for (auto *BB : Blocks) {
1551  Instruction *I = BB->getTerminator();
1552  do {
1553  I = I->getPrevNode();
1554  } while (isa<DbgInfoIntrinsic>(I) && I != &BB->front());
1555  if (!isa<DbgInfoIntrinsic>(I))
1556  Insts.push_back(I);
1557  }
1558 
1559  // The only checking we need to do now is that all users of all instructions
1560  // are the same PHI node. canSinkLastInstruction should have checked this but
1561  // it is slightly over-aggressive - it gets confused by commutative instructions
1562  // so double-check it here.
1563  Instruction *I0 = Insts.front();
1564  if (!I0->user_empty()) {
1565  auto *PNUse = dyn_cast<PHINode>(*I0->user_begin());
1566  if (!all_of(Insts, [&PNUse](const Instruction *I) -> bool {
1567  auto *U = cast<Instruction>(*I->user_begin());
1568  return U == PNUse;
1569  }))
1570  return false;
1571  }
1572 
1573  // We don't need to do any more checking here; canSinkLastInstruction should
1574  // have done it all for us.
1575  SmallVector<Value*, 4> NewOperands;
1576  for (unsigned O = 0, E = I0->getNumOperands(); O != E; ++O) {
1577  // This check is different to that in canSinkLastInstruction. There, we
1578  // cared about the global view once simplifycfg (and instcombine) have
1579  // completed - it takes into account PHIs that become trivially
1580  // simplifiable. However here we need a more local view; if an operand
1581  // differs we create a PHI and rely on instcombine to clean up the very
1582  // small mess we may make.
1583  bool NeedPHI = any_of(Insts, [&I0, O](const Instruction *I) {
1584  return I->getOperand(O) != I0->getOperand(O);
1585  });
1586  if (!NeedPHI) {
1587  NewOperands.push_back(I0->getOperand(O));
1588  continue;
1589  }
1590 
1591  // Create a new PHI in the successor block and populate it.
1592  auto *Op = I0->getOperand(O);
1593  assert(!Op->getType()->isTokenTy() && "Can't PHI tokens!");
1594  auto *PN = PHINode::Create(Op->getType(), Insts.size(),
1595  Op->getName() + ".sink", &BBEnd->front());
1596  for (auto *I : Insts)
1597  PN->addIncoming(I->getOperand(O), I->getParent());
1598  NewOperands.push_back(PN);
1599  }
1600 
1601  // Arbitrarily use I0 as the new "common" instruction; remap its operands
1602  // and move it to the start of the successor block.
1603  for (unsigned O = 0, E = I0->getNumOperands(); O != E; ++O)
1604  I0->getOperandUse(O).set(NewOperands[O]);
1605  I0->moveBefore(&*BBEnd->getFirstInsertionPt());
1606 
1607  // Update metadata and IR flags, and merge debug locations.
1608  for (auto *I : Insts)
1609  if (I != I0) {
1610  // The debug location for the "common" instruction is the merged locations
1611  // of all the commoned instructions. We start with the original location
1612  // of the "common" instruction and iteratively merge each location in the
1613  // loop below.
1614  // This is an N-way merge, which will be inefficient if I0 is a CallInst.
1615  // However, as N-way merge for CallInst is rare, so we use simplified API
1616  // instead of using complex API for N-way merge.
1617  I0->applyMergedLocation(I0->getDebugLoc(), I->getDebugLoc());
1618  combineMetadataForCSE(I0, I, true);
1619  I0->andIRFlags(I);
1620  }
1621 
1622  if (!I0->user_empty()) {
1623  // canSinkLastInstruction checked that all instructions were used by
1624  // one and only one PHI node. Find that now, RAUW it to our common
1625  // instruction and nuke it.
1626  auto *PN = cast<PHINode>(*I0->user_begin());
1627  PN->replaceAllUsesWith(I0);
1628  PN->eraseFromParent();
1629  }
1630 
1631  // Finally nuke all instructions apart from the common instruction.
1632  for (auto *I : Insts)
1633  if (I != I0)
1634  I->eraseFromParent();
1635 
1636  return true;
1637 }
1638 
1639 namespace {
1640 
1641  // LockstepReverseIterator - Iterates through instructions
1642  // in a set of blocks in reverse order from the first non-terminator.
1643  // For example (assume all blocks have size n):
1644  // LockstepReverseIterator I([B1, B2, B3]);
1645  // *I-- = [B1[n], B2[n], B3[n]];
1646  // *I-- = [B1[n-1], B2[n-1], B3[n-1]];
1647  // *I-- = [B1[n-2], B2[n-2], B3[n-2]];
1648  // ...
1649  class LockstepReverseIterator {
1650  ArrayRef<BasicBlock*> Blocks;
1652  bool Fail;
1653 
1654  public:
1655  LockstepReverseIterator(ArrayRef<BasicBlock*> Blocks) : Blocks(Blocks) {
1656  reset();
1657  }
1658 
1659  void reset() {
1660  Fail = false;
1661  Insts.clear();
1662  for (auto *BB : Blocks) {
1663  Instruction *Inst = BB->getTerminator();
1664  for (Inst = Inst->getPrevNode(); Inst && isa<DbgInfoIntrinsic>(Inst);)
1665  Inst = Inst->getPrevNode();
1666  if (!Inst) {
1667  // Block wasn't big enough.
1668  Fail = true;
1669  return;
1670  }
1671  Insts.push_back(Inst);
1672  }
1673  }
1674 
1675  bool isValid() const {
1676  return !Fail;
1677  }
1678 
1679  void operator--() {
1680  if (Fail)
1681  return;
1682  for (auto *&Inst : Insts) {
1683  for (Inst = Inst->getPrevNode(); Inst && isa<DbgInfoIntrinsic>(Inst);)
1684  Inst = Inst->getPrevNode();
1685  // Already at beginning of block.
1686  if (!Inst) {
1687  Fail = true;
1688  return;
1689  }
1690  }
1691  }
1692 
1694  return Insts;
1695  }
1696  };
1697 
1698 } // end anonymous namespace
1699 
1700 /// Check whether BB's predecessors end with unconditional branches. If it is
1701 /// true, sink any common code from the predecessors to BB.
1702 /// We also allow one predecessor to end with conditional branch (but no more
1703 /// than one).
1705  // We support two situations:
1706  // (1) all incoming arcs are unconditional
1707  // (2) one incoming arc is conditional
1708  //
1709  // (2) is very common in switch defaults and
1710  // else-if patterns;
1711  //
1712  // if (a) f(1);
1713  // else if (b) f(2);
1714  //
1715  // produces:
1716  //
1717  // [if]
1718  // / \
1719  // [f(1)] [if]
1720  // | | \
1721  // | | |
1722  // | [f(2)]|
1723  // \ | /
1724  // [ end ]
1725  //
1726  // [end] has two unconditional predecessor arcs and one conditional. The
1727  // conditional refers to the implicit empty 'else' arc. This conditional
1728  // arc can also be caused by an empty default block in a switch.
1729  //
1730  // In this case, we attempt to sink code from all *unconditional* arcs.
1731  // If we can sink instructions from these arcs (determined during the scan
1732  // phase below) we insert a common successor for all unconditional arcs and
1733  // connect that to [end], to enable sinking:
1734  //
1735  // [if]
1736  // / \
1737  // [x(1)] [if]
1738  // | | \
1739  // | | \
1740  // | [x(2)] |
1741  // \ / |
1742  // [sink.split] |
1743  // \ /
1744  // [ end ]
1745  //
1746  SmallVector<BasicBlock*,4> UnconditionalPreds;
1747  Instruction *Cond = nullptr;
1748  for (auto *B : predecessors(BB)) {
1749  auto *T = B->getTerminator();
1750  if (isa<BranchInst>(T) && cast<BranchInst>(T)->isUnconditional())
1751  UnconditionalPreds.push_back(B);
1752  else if ((isa<BranchInst>(T) || isa<SwitchInst>(T)) && !Cond)
1753  Cond = T;
1754  else
1755  return false;
1756  }
1757  if (UnconditionalPreds.size() < 2)
1758  return false;
1759 
1760  bool Changed = false;
1761  // We take a two-step approach to tail sinking. First we scan from the end of
1762  // each block upwards in lockstep. If the n'th instruction from the end of each
1763  // block can be sunk, those instructions are added to ValuesToSink and we
1764  // carry on. If we can sink an instruction but need to PHI-merge some operands
1765  // (because they're not identical in each instruction) we add these to
1766  // PHIOperands.
1767  unsigned ScanIdx = 0;
1768  SmallPtrSet<Value*,4> InstructionsToSink;
1770  LockstepReverseIterator LRI(UnconditionalPreds);
1771  while (LRI.isValid() &&
1772  canSinkInstructions(*LRI, PHIOperands)) {
1773  LLVM_DEBUG(dbgs() << "SINK: instruction can be sunk: " << *(*LRI)[0]
1774  << "\n");
1775  InstructionsToSink.insert((*LRI).begin(), (*LRI).end());
1776  ++ScanIdx;
1777  --LRI;
1778  }
1779 
1780  auto ProfitableToSinkInstruction = [&](LockstepReverseIterator &LRI) {
1781  unsigned NumPHIdValues = 0;
1782  for (auto *I : *LRI)
1783  for (auto *V : PHIOperands[I])
1784  if (InstructionsToSink.count(V) == 0)
1785  ++NumPHIdValues;
1786  LLVM_DEBUG(dbgs() << "SINK: #phid values: " << NumPHIdValues << "\n");
1787  unsigned NumPHIInsts = NumPHIdValues / UnconditionalPreds.size();
1788  if ((NumPHIdValues % UnconditionalPreds.size()) != 0)
1789  NumPHIInsts++;
1790 
1791  return NumPHIInsts <= 1;
1792  };
1793 
1794  if (ScanIdx > 0 && Cond) {
1795  // Check if we would actually sink anything first! This mutates the CFG and
1796  // adds an extra block. The goal in doing this is to allow instructions that
1797  // couldn't be sunk before to be sunk - obviously, speculatable instructions
1798  // (such as trunc, add) can be sunk and predicated already. So we check that
1799  // we're going to sink at least one non-speculatable instruction.
1800  LRI.reset();
1801  unsigned Idx = 0;
1802  bool Profitable = false;
1803  while (ProfitableToSinkInstruction(LRI) && Idx < ScanIdx) {
1804  if (!isSafeToSpeculativelyExecute((*LRI)[0])) {
1805  Profitable = true;
1806  break;
1807  }
1808  --LRI;
1809  ++Idx;
1810  }
1811  if (!Profitable)
1812  return false;
1813 
1814  LLVM_DEBUG(dbgs() << "SINK: Splitting edge\n");
1815  // We have a conditional edge and we're going to sink some instructions.
1816  // Insert a new block postdominating all blocks we're going to sink from.
1817  if (!SplitBlockPredecessors(BB, UnconditionalPreds, ".sink.split"))
1818  // Edges couldn't be split.
1819  return false;
1820  Changed = true;
1821  }
1822 
1823  // Now that we've analyzed all potential sinking candidates, perform the
1824  // actual sink. We iteratively sink the last non-terminator of the source
1825  // blocks into their common successor unless doing so would require too
1826  // many PHI instructions to be generated (currently only one PHI is allowed
1827  // per sunk instruction).
1828  //
1829  // We can use InstructionsToSink to discount values needing PHI-merging that will
1830  // actually be sunk in a later iteration. This allows us to be more
1831  // aggressive in what we sink. This does allow a false positive where we
1832  // sink presuming a later value will also be sunk, but stop half way through
1833  // and never actually sink it which means we produce more PHIs than intended.
1834  // This is unlikely in practice though.
1835  for (unsigned SinkIdx = 0; SinkIdx != ScanIdx; ++SinkIdx) {
1836  LLVM_DEBUG(dbgs() << "SINK: Sink: "
1837  << *UnconditionalPreds[0]->getTerminator()->getPrevNode()
1838  << "\n");
1839 
1840  // Because we've sunk every instruction in turn, the current instruction to
1841  // sink is always at index 0.
1842  LRI.reset();
1843  if (!ProfitableToSinkInstruction(LRI)) {
1844  // Too many PHIs would be created.
1845  LLVM_DEBUG(
1846  dbgs() << "SINK: stopping here, too many PHIs would be created!\n");
1847  break;
1848  }
1849 
1850  if (!sinkLastInstruction(UnconditionalPreds))
1851  return Changed;
1852  NumSinkCommons++;
1853  Changed = true;
1854  }
1855  return Changed;
1856 }
1857 
1858 /// Determine if we can hoist sink a sole store instruction out of a
1859 /// conditional block.
1860 ///
1861 /// We are looking for code like the following:
1862 /// BrBB:
1863 /// store i32 %add, i32* %arrayidx2
1864 /// ... // No other stores or function calls (we could be calling a memory
1865 /// ... // function).
1866 /// %cmp = icmp ult %x, %y
1867 /// br i1 %cmp, label %EndBB, label %ThenBB
1868 /// ThenBB:
1869 /// store i32 %add5, i32* %arrayidx2
1870 /// br label EndBB
1871 /// EndBB:
1872 /// ...
1873 /// We are going to transform this into:
1874 /// BrBB:
1875 /// store i32 %add, i32* %arrayidx2
1876 /// ... //
1877 /// %cmp = icmp ult %x, %y
1878 /// %add.add5 = select i1 %cmp, i32 %add, %add5
1879 /// store i32 %add.add5, i32* %arrayidx2
1880 /// ...
1881 ///
1882 /// \return The pointer to the value of the previous store if the store can be
1883 /// hoisted into the predecessor block. 0 otherwise.
1885  BasicBlock *StoreBB, BasicBlock *EndBB) {
1886  StoreInst *StoreToHoist = dyn_cast<StoreInst>(I);
1887  if (!StoreToHoist)
1888  return nullptr;
1889 
1890  // Volatile or atomic.
1891  if (!StoreToHoist->isSimple())
1892  return nullptr;
1893 
1894  Value *StorePtr = StoreToHoist->getPointerOperand();
1895 
1896  // Look for a store to the same pointer in BrBB.
1897  unsigned MaxNumInstToLookAt = 9;
1898  for (Instruction &CurI : reverse(BrBB->instructionsWithoutDebug())) {
1899  if (!MaxNumInstToLookAt)
1900  break;
1901  --MaxNumInstToLookAt;
1902 
1903  // Could be calling an instruction that affects memory like free().
1904  if (CurI.mayHaveSideEffects() && !isa<StoreInst>(CurI))
1905  return nullptr;
1906 
1907  if (auto *SI = dyn_cast<StoreInst>(&CurI)) {
1908  // Found the previous store make sure it stores to the same location.
1909  if (SI->getPointerOperand() == StorePtr)
1910  // Found the previous store, return its value operand.
1911  return SI->getValueOperand();
1912  return nullptr; // Unknown store.
1913  }
1914  }
1915 
1916  return nullptr;
1917 }
1918 
1919 /// Speculate a conditional basic block flattening the CFG.
1920 ///
1921 /// Note that this is a very risky transform currently. Speculating
1922 /// instructions like this is most often not desirable. Instead, there is an MI
1923 /// pass which can do it with full awareness of the resource constraints.
1924 /// However, some cases are "obvious" and we should do directly. An example of
1925 /// this is speculating a single, reasonably cheap instruction.
1926 ///
1927 /// There is only one distinct advantage to flattening the CFG at the IR level:
1928 /// it makes very common but simplistic optimizations such as are common in
1929 /// instcombine and the DAG combiner more powerful by removing CFG edges and
1930 /// modeling their effects with easier to reason about SSA value graphs.
1931 ///
1932 ///
1933 /// An illustration of this transform is turning this IR:
1934 /// \code
1935 /// BB:
1936 /// %cmp = icmp ult %x, %y
1937 /// br i1 %cmp, label %EndBB, label %ThenBB
1938 /// ThenBB:
1939 /// %sub = sub %x, %y
1940 /// br label BB2
1941 /// EndBB:
1942 /// %phi = phi [ %sub, %ThenBB ], [ 0, %EndBB ]
1943 /// ...
1944 /// \endcode
1945 ///
1946 /// Into this IR:
1947 /// \code
1948 /// BB:
1949 /// %cmp = icmp ult %x, %y
1950 /// %sub = sub %x, %y
1951 /// %cond = select i1 %cmp, 0, %sub
1952 /// ...
1953 /// \endcode
1954 ///
1955 /// \returns true if the conditional block is removed.
1957  const TargetTransformInfo &TTI) {
1958  // Be conservative for now. FP select instruction can often be expensive.
1959  Value *BrCond = BI->getCondition();
1960  if (isa<FCmpInst>(BrCond))
1961  return false;
1962 
1963  BasicBlock *BB = BI->getParent();
1964  BasicBlock *EndBB = ThenBB->getTerminator()->getSuccessor(0);
1965 
1966  // If ThenBB is actually on the false edge of the conditional branch, remember
1967  // to swap the select operands later.
1968  bool Invert = false;
1969  if (ThenBB != BI->getSuccessor(0)) {
1970  assert(ThenBB == BI->getSuccessor(1) && "No edge from 'if' block?");
1971  Invert = true;
1972  }
1973  assert(EndBB == BI->getSuccessor(!Invert) && "No edge from to end block");
1974 
1975  // Keep a count of how many times instructions are used within ThenBB when
1976  // they are candidates for sinking into ThenBB. Specifically:
1977  // - They are defined in BB, and
1978  // - They have no side effects, and
1979  // - All of their uses are in ThenBB.
1980  SmallDenseMap<Instruction *, unsigned, 4> SinkCandidateUseCounts;
1981 
1982  SmallVector<Instruction *, 4> SpeculatedDbgIntrinsics;
1983 
1984  unsigned SpeculatedInstructions = 0;
1985  Value *SpeculatedStoreValue = nullptr;
1986  StoreInst *SpeculatedStore = nullptr;
1987  for (BasicBlock::iterator BBI = ThenBB->begin(),
1988  BBE = std::prev(ThenBB->end());
1989  BBI != BBE; ++BBI) {
1990  Instruction *I = &*BBI;
1991  // Skip debug info.
1992  if (isa<DbgInfoIntrinsic>(I)) {
1993  SpeculatedDbgIntrinsics.push_back(I);
1994  continue;
1995  }
1996 
1997  // Only speculatively execute a single instruction (not counting the
1998  // terminator) for now.
1999  ++SpeculatedInstructions;
2000  if (SpeculatedInstructions > 1)
2001  return false;
2002 
2003  // Don't hoist the instruction if it's unsafe or expensive.
2004  if (!isSafeToSpeculativelyExecute(I) &&
2005  !(HoistCondStores && (SpeculatedStoreValue = isSafeToSpeculateStore(
2006  I, BB, ThenBB, EndBB))))
2007  return false;
2008  if (!SpeculatedStoreValue &&
2009  ComputeSpeculationCost(I, TTI) >
2011  return false;
2012 
2013  // Store the store speculation candidate.
2014  if (SpeculatedStoreValue)
2015  SpeculatedStore = cast<StoreInst>(I);
2016 
2017  // Do not hoist the instruction if any of its operands are defined but not
2018  // used in BB. The transformation will prevent the operand from
2019  // being sunk into the use block.
2020  for (User::op_iterator i = I->op_begin(), e = I->op_end(); i != e; ++i) {
2021  Instruction *OpI = dyn_cast<Instruction>(*i);
2022  if (!OpI || OpI->getParent() != BB || OpI->mayHaveSideEffects())
2023  continue; // Not a candidate for sinking.
2024 
2025  ++SinkCandidateUseCounts[OpI];
2026  }
2027  }
2028 
2029  // Consider any sink candidates which are only used in ThenBB as costs for
2030  // speculation. Note, while we iterate over a DenseMap here, we are summing
2031  // and so iteration order isn't significant.
2033  I = SinkCandidateUseCounts.begin(),
2034  E = SinkCandidateUseCounts.end();
2035  I != E; ++I)
2036  if (I->first->hasNUses(I->second)) {
2037  ++SpeculatedInstructions;
2038  if (SpeculatedInstructions > 1)
2039  return false;
2040  }
2041 
2042  // Check that the PHI nodes can be converted to selects.
2043  bool HaveRewritablePHIs = false;
2044  for (PHINode &PN : EndBB->phis()) {
2045  Value *OrigV = PN.getIncomingValueForBlock(BB);
2046  Value *ThenV = PN.getIncomingValueForBlock(ThenBB);
2047 
2048  // FIXME: Try to remove some of the duplication with HoistThenElseCodeToIf.
2049  // Skip PHIs which are trivial.
2050  if (ThenV == OrigV)
2051  continue;
2052 
2053  // Don't convert to selects if we could remove undefined behavior instead.
2054  if (passingValueIsAlwaysUndefined(OrigV, &PN) ||
2055  passingValueIsAlwaysUndefined(ThenV, &PN))
2056  return false;
2057 
2058  HaveRewritablePHIs = true;
2059  ConstantExpr *OrigCE = dyn_cast<ConstantExpr>(OrigV);
2060  ConstantExpr *ThenCE = dyn_cast<ConstantExpr>(ThenV);
2061  if (!OrigCE && !ThenCE)
2062  continue; // Known safe and cheap.
2063 
2064  if ((ThenCE && !isSafeToSpeculativelyExecute(ThenCE)) ||
2065  (OrigCE && !isSafeToSpeculativelyExecute(OrigCE)))
2066  return false;
2067  unsigned OrigCost = OrigCE ? ComputeSpeculationCost(OrigCE, TTI) : 0;
2068  unsigned ThenCost = ThenCE ? ComputeSpeculationCost(ThenCE, TTI) : 0;
2069  unsigned MaxCost =
2071  if (OrigCost + ThenCost > MaxCost)
2072  return false;
2073 
2074  // Account for the cost of an unfolded ConstantExpr which could end up
2075  // getting expanded into Instructions.
2076  // FIXME: This doesn't account for how many operations are combined in the
2077  // constant expression.
2078  ++SpeculatedInstructions;
2079  if (SpeculatedInstructions > 1)
2080  return false;
2081  }
2082 
2083  // If there are no PHIs to process, bail early. This helps ensure idempotence
2084  // as well.
2085  if (!HaveRewritablePHIs && !(HoistCondStores && SpeculatedStoreValue))
2086  return false;
2087 
2088  // If we get here, we can hoist the instruction and if-convert.
2089  LLVM_DEBUG(dbgs() << "SPECULATIVELY EXECUTING BB" << *ThenBB << "\n";);
2090 
2091  // Insert a select of the value of the speculated store.
2092  if (SpeculatedStoreValue) {
2093  IRBuilder<NoFolder> Builder(BI);
2094  Value *TrueV = SpeculatedStore->getValueOperand();
2095  Value *FalseV = SpeculatedStoreValue;
2096  if (Invert)
2097  std::swap(TrueV, FalseV);
2098  Value *S = Builder.CreateSelect(
2099  BrCond, TrueV, FalseV, "spec.store.select", BI);
2100  SpeculatedStore->setOperand(0, S);
2101  SpeculatedStore->applyMergedLocation(BI->getDebugLoc(),
2102  SpeculatedStore->getDebugLoc());
2103  }
2104 
2105  // Metadata can be dependent on the condition we are hoisting above.
2106  // Conservatively strip all metadata on the instruction.
2107  for (auto &I : *ThenBB)
2109 
2110  // Hoist the instructions.
2111  BB->getInstList().splice(BI->getIterator(), ThenBB->getInstList(),
2112  ThenBB->begin(), std::prev(ThenBB->end()));
2113 
2114  // Insert selects and rewrite the PHI operands.
2115  IRBuilder<NoFolder> Builder(BI);
2116  for (PHINode &PN : EndBB->phis()) {
2117  unsigned OrigI = PN.getBasicBlockIndex(BB);
2118  unsigned ThenI = PN.getBasicBlockIndex(ThenBB);
2119  Value *OrigV = PN.getIncomingValue(OrigI);
2120  Value *ThenV = PN.getIncomingValue(ThenI);
2121 
2122  // Skip PHIs which are trivial.
2123  if (OrigV == ThenV)
2124  continue;
2125 
2126  // Create a select whose true value is the speculatively executed value and
2127  // false value is the preexisting value. Swap them if the branch
2128  // destinations were inverted.
2129  Value *TrueV = ThenV, *FalseV = OrigV;
2130  if (Invert)
2131  std::swap(TrueV, FalseV);
2132  Value *V = Builder.CreateSelect(
2133  BrCond, TrueV, FalseV, "spec.select", BI);
2134  PN.setIncomingValue(OrigI, V);
2135  PN.setIncomingValue(ThenI, V);
2136  }
2137 
2138  // Remove speculated dbg intrinsics.
2139  // FIXME: Is it possible to do this in a more elegant way? Moving/merging the
2140  // dbg value for the different flows and inserting it after the select.
2141  for (Instruction *I : SpeculatedDbgIntrinsics)
2142  I->eraseFromParent();
2143 
2144  ++NumSpeculations;
2145  return true;
2146 }
2147 
2148 /// Return true if we can thread a branch across this block.
2150  unsigned Size = 0;
2151 
2152  for (Instruction &I : BB->instructionsWithoutDebug()) {
2153  if (Size > 10)
2154  return false; // Don't clone large BB's.
2155  ++Size;
2156 
2157  // We can only support instructions that do not define values that are
2158  // live outside of the current basic block.
2159  for (User *U : I.users()) {
2160  Instruction *UI = cast<Instruction>(U);
2161  if (UI->getParent() != BB || isa<PHINode>(UI))
2162  return false;
2163  }
2164 
2165  // Looks ok, continue checking.
2166  }
2167 
2168  return true;
2169 }
2170 
2171 /// If we have a conditional branch on a PHI node value that is defined in the
2172 /// same block as the branch and if any PHI entries are constants, thread edges
2173 /// corresponding to that entry to be branches to their ultimate destination.
2174 static bool FoldCondBranchOnPHI(BranchInst *BI, const DataLayout &DL,
2175  AssumptionCache *AC) {
2176  BasicBlock *BB = BI->getParent();
2177  PHINode *PN = dyn_cast<PHINode>(BI->getCondition());
2178  // NOTE: we currently cannot transform this case if the PHI node is used
2179  // outside of the block.
2180  if (!PN || PN->getParent() != BB || !PN->hasOneUse())
2181  return false;
2182 
2183  // Degenerate case of a single entry PHI.
2184  if (PN->getNumIncomingValues() == 1) {
2186  return true;
2187  }
2188 
2189  // Now we know that this block has multiple preds and two succs.
2191  return false;
2192 
2193  // Can't fold blocks that contain noduplicate or convergent calls.
2194  if (any_of(*BB, [](const Instruction &I) {
2195  const CallInst *CI = dyn_cast<CallInst>(&I);
2196  return CI && (CI->cannotDuplicate() || CI->isConvergent());
2197  }))
2198  return false;
2199 
2200  // Okay, this is a simple enough basic block. See if any phi values are
2201  // constants.
2202  for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
2204  if (!CB || !CB->getType()->isIntegerTy(1))
2205  continue;
2206 
2207  // Okay, we now know that all edges from PredBB should be revectored to
2208  // branch to RealDest.
2209  BasicBlock *PredBB = PN->getIncomingBlock(i);
2210  BasicBlock *RealDest = BI->getSuccessor(!CB->getZExtValue());
2211 
2212  if (RealDest == BB)
2213  continue; // Skip self loops.
2214  // Skip if the predecessor's terminator is an indirect branch.
2215  if (isa<IndirectBrInst>(PredBB->getTerminator()))
2216  continue;
2217 
2218  // The dest block might have PHI nodes, other predecessors and other
2219  // difficult cases. Instead of being smart about this, just insert a new
2220  // block that jumps to the destination block, effectively splitting
2221  // the edge we are about to create.
2222  BasicBlock *EdgeBB =
2223  BasicBlock::Create(BB->getContext(), RealDest->getName() + ".critedge",
2224  RealDest->getParent(), RealDest);
2225  BranchInst *CritEdgeBranch = BranchInst::Create(RealDest, EdgeBB);
2226  CritEdgeBranch->setDebugLoc(BI->getDebugLoc());
2227 
2228  // Update PHI nodes.
2229  AddPredecessorToBlock(RealDest, EdgeBB, BB);
2230 
2231  // BB may have instructions that are being threaded over. Clone these
2232  // instructions into EdgeBB. We know that there will be no uses of the
2233  // cloned instructions outside of EdgeBB.
2234  BasicBlock::iterator InsertPt = EdgeBB->begin();
2235  DenseMap<Value *, Value *> TranslateMap; // Track translated values.
2236  for (BasicBlock::iterator BBI = BB->begin(); &*BBI != BI; ++BBI) {
2237  if (PHINode *PN = dyn_cast<PHINode>(BBI)) {
2238  TranslateMap[PN] = PN->getIncomingValueForBlock(PredBB);
2239  continue;
2240  }
2241  // Clone the instruction.
2242  Instruction *N = BBI->clone();
2243  if (BBI->hasName())
2244  N->setName(BBI->getName() + ".c");
2245 
2246  // Update operands due to translation.
2247  for (User::op_iterator i = N->op_begin(), e = N->op_end(); i != e; ++i) {
2248  DenseMap<Value *, Value *>::iterator PI = TranslateMap.find(*i);
2249  if (PI != TranslateMap.end())
2250  *i = PI->second;
2251  }
2252 
2253  // Check for trivial simplification.
2254  if (Value *V = SimplifyInstruction(N, {DL, nullptr, nullptr, AC})) {
2255  if (!BBI->use_empty())
2256  TranslateMap[&*BBI] = V;
2257  if (!N->mayHaveSideEffects()) {
2258  N->deleteValue(); // Instruction folded away, don't need actual inst
2259  N = nullptr;
2260  }
2261  } else {
2262  if (!BBI->use_empty())
2263  TranslateMap[&*BBI] = N;
2264  }
2265  // Insert the new instruction into its new home.
2266  if (N)
2267  EdgeBB->getInstList().insert(InsertPt, N);
2268 
2269  // Register the new instruction with the assumption cache if necessary.
2270  if (auto *II = dyn_cast_or_null<IntrinsicInst>(N))
2271  if (II->getIntrinsicID() == Intrinsic::assume)
2272  AC->registerAssumption(II);
2273  }
2274 
2275  // Loop over all of the edges from PredBB to BB, changing them to branch
2276  // to EdgeBB instead.
2277  Instruction *PredBBTI = PredBB->getTerminator();
2278  for (unsigned i = 0, e = PredBBTI->getNumSuccessors(); i != e; ++i)
2279  if (PredBBTI->getSuccessor(i) == BB) {
2280  BB->removePredecessor(PredBB);
2281  PredBBTI->setSuccessor(i, EdgeBB);
2282  }
2283 
2284  // Recurse, simplifying any other constants.
2285  return FoldCondBranchOnPHI(BI, DL, AC) || true;
2286  }
2287 
2288  return false;
2289 }
2290 
2291 /// Given a BB that starts with the specified two-entry PHI node,
2292 /// see if we can eliminate it.
2294  const DataLayout &DL) {
2295  // Ok, this is a two entry PHI node. Check to see if this is a simple "if
2296  // statement", which has a very simple dominance structure. Basically, we
2297  // are trying to find the condition that is being branched on, which
2298  // subsequently causes this merge to happen. We really want control
2299  // dependence information for this check, but simplifycfg can't keep it up
2300  // to date, and this catches most of the cases we care about anyway.
2301  BasicBlock *BB = PN->getParent();
2302  const Function *Fn = BB->getParent();
2303  if (Fn && Fn->hasFnAttribute(Attribute::OptForFuzzing))
2304  return false;
2305 
2306  BasicBlock *IfTrue, *IfFalse;
2307  Value *IfCond = GetIfCondition(BB, IfTrue, IfFalse);
2308  if (!IfCond ||
2309  // Don't bother if the branch will be constant folded trivially.
2310  isa<ConstantInt>(IfCond))
2311  return false;
2312 
2313  // Okay, we found that we can merge this two-entry phi node into a select.
2314  // Doing so would require us to fold *all* two entry phi nodes in this block.
2315  // At some point this becomes non-profitable (particularly if the target
2316  // doesn't support cmov's). Only do this transformation if there are two or
2317  // fewer PHI nodes in this block.
2318  unsigned NumPhis = 0;
2319  for (BasicBlock::iterator I = BB->begin(); isa<PHINode>(I); ++NumPhis, ++I)
2320  if (NumPhis > 2)
2321  return false;
2322 
2323  // Loop over the PHI's seeing if we can promote them all to select
2324  // instructions. While we are at it, keep track of the instructions
2325  // that need to be moved to the dominating block.
2326  SmallPtrSet<Instruction *, 4> AggressiveInsts;
2327  int BudgetRemaining =
2329 
2330  for (BasicBlock::iterator II = BB->begin(); isa<PHINode>(II);) {
2331  PHINode *PN = cast<PHINode>(II++);
2332  if (Value *V = SimplifyInstruction(PN, {DL, PN})) {
2333  PN->replaceAllUsesWith(V);
2334  PN->eraseFromParent();
2335  continue;
2336  }
2337 
2338  if (!DominatesMergePoint(PN->getIncomingValue(0), BB, AggressiveInsts,
2339  BudgetRemaining, TTI) ||
2340  !DominatesMergePoint(PN->getIncomingValue(1), BB, AggressiveInsts,
2341  BudgetRemaining, TTI))
2342  return false;
2343  }
2344 
2345  // If we folded the first phi, PN dangles at this point. Refresh it. If
2346  // we ran out of PHIs then we simplified them all.
2347  PN = dyn_cast<PHINode>(BB->begin());
2348  if (!PN)
2349  return true;
2350 
2351  // Return true if at least one of these is a 'not', and another is either
2352  // a 'not' too, or a constant.
2353  auto CanHoistNotFromBothValues = [](Value *V0, Value *V1) {
2354  if (!match(V0, m_Not(m_Value())))
2355  std::swap(V0, V1);
2356  auto Invertible = m_CombineOr(m_Not(m_Value()), m_AnyIntegralConstant());
2357  return match(V0, m_Not(m_Value())) && match(V1, Invertible);
2358  };
2359 
2360  // Don't fold i1 branches on PHIs which contain binary operators, unless one
2361  // of the incoming values is an 'not' and another one is freely invertible.
2362  // These can often be turned into switches and other things.
2363  if (PN->getType()->isIntegerTy(1) &&
2364  (isa<BinaryOperator>(PN->getIncomingValue(0)) ||
2365  isa<BinaryOperator>(PN->getIncomingValue(1)) ||
2366  isa<BinaryOperator>(IfCond)) &&
2367  !CanHoistNotFromBothValues(PN->getIncomingValue(0),
2368  PN->getIncomingValue(1)))
2369  return false;
2370 
2371  // If all PHI nodes are promotable, check to make sure that all instructions
2372  // in the predecessor blocks can be promoted as well. If not, we won't be able
2373  // to get rid of the control flow, so it's not worth promoting to select
2374  // instructions.
2375  BasicBlock *DomBlock = nullptr;
2376  BasicBlock *IfBlock1 = PN->getIncomingBlock(0);
2377  BasicBlock *IfBlock2 = PN->getIncomingBlock(1);
2378  if (cast<BranchInst>(IfBlock1->getTerminator())->isConditional()) {
2379  IfBlock1 = nullptr;
2380  } else {
2381  DomBlock = *pred_begin(IfBlock1);
2382  for (BasicBlock::iterator I = IfBlock1->begin(); !I->isTerminator(); ++I)
2383  if (!AggressiveInsts.count(&*I) && !isa<DbgInfoIntrinsic>(I)) {
2384  // This is not an aggressive instruction that we can promote.
2385  // Because of this, we won't be able to get rid of the control flow, so
2386  // the xform is not worth it.
2387  return false;
2388  }
2389  }
2390 
2391  if (cast<BranchInst>(IfBlock2->getTerminator())->isConditional()) {
2392  IfBlock2 = nullptr;
2393  } else {
2394  DomBlock = *pred_begin(IfBlock2);
2395  for (BasicBlock::iterator I = IfBlock2->begin(); !I->isTerminator(); ++I)
2396  if (!AggressiveInsts.count(&*I) && !isa<DbgInfoIntrinsic>(I)) {
2397  // This is not an aggressive instruction that we can promote.
2398  // Because of this, we won't be able to get rid of the control flow, so
2399  // the xform is not worth it.
2400  return false;
2401  }
2402  }
2403  assert(DomBlock && "Failed to find root DomBlock");
2404 
2405  LLVM_DEBUG(dbgs() << "FOUND IF CONDITION! " << *IfCond
2406  << " T: " << IfTrue->getName()
2407  << " F: " << IfFalse->getName() << "\n");
2408 
2409  // If we can still promote the PHI nodes after this gauntlet of tests,
2410  // do all of the PHI's now.
2411  Instruction *InsertPt = DomBlock->getTerminator();
2412  IRBuilder<NoFolder> Builder(InsertPt);
2413 
2414  // Move all 'aggressive' instructions, which are defined in the
2415  // conditional parts of the if's up to the dominating block.
2416  if (IfBlock1)
2417  hoistAllInstructionsInto(DomBlock, InsertPt, IfBlock1);
2418  if (IfBlock2)
2419  hoistAllInstructionsInto(DomBlock, InsertPt, IfBlock2);
2420 
2421  while (PHINode *PN = dyn_cast<PHINode>(BB->begin())) {
2422  // Change the PHI node into a select instruction.
2423  Value *TrueVal = PN->getIncomingValue(PN->getIncomingBlock(0) == IfFalse);
2424  Value *FalseVal = PN->getIncomingValue(PN->getIncomingBlock(0) == IfTrue);
2425 
2426  Value *Sel = Builder.CreateSelect(IfCond, TrueVal, FalseVal, "", InsertPt);
2427  PN->replaceAllUsesWith(Sel);
2428  Sel->takeName(PN);
2429  PN->eraseFromParent();
2430  }
2431 
2432  // At this point, IfBlock1 and IfBlock2 are both empty, so our if statement
2433  // has been flattened. Change DomBlock to jump directly to our new block to
2434  // avoid other simplifycfg's kicking in on the diamond.
2435  Instruction *OldTI = DomBlock->getTerminator();
2436  Builder.SetInsertPoint(OldTI);
2437  Builder.CreateBr(BB);
2438  OldTI->eraseFromParent();
2439  return true;
2440 }
2441 
2442 /// If we found a conditional branch that goes to two returning blocks,
2443 /// try to merge them together into one return,
2444 /// introducing a select if the return values disagree.
2446  IRBuilder<> &Builder) {
2447  assert(BI->isConditional() && "Must be a conditional branch");
2448  BasicBlock *TrueSucc = BI->getSuccessor(0);
2449  BasicBlock *FalseSucc = BI->getSuccessor(1);
2450  ReturnInst *TrueRet = cast<ReturnInst>(TrueSucc->getTerminator());
2451  ReturnInst *FalseRet = cast<ReturnInst>(FalseSucc->getTerminator());
2452 
2453  // Check to ensure both blocks are empty (just a return) or optionally empty
2454  // with PHI nodes. If there are other instructions, merging would cause extra
2455  // computation on one path or the other.
2456  if (!TrueSucc->getFirstNonPHIOrDbg()->isTerminator())
2457  return false;
2458  if (!FalseSucc->getFirstNonPHIOrDbg()->isTerminator())
2459  return false;
2460 
2461  Builder.SetInsertPoint(BI);
2462  // Okay, we found a branch that is going to two return nodes. If
2463  // there is no return value for this function, just change the
2464  // branch into a return.
2465  if (FalseRet->getNumOperands() == 0) {
2466  TrueSucc->removePredecessor(BI->getParent());
2467  FalseSucc->removePredecessor(BI->getParent());
2468  Builder.CreateRetVoid();
2470  return true;
2471  }
2472 
2473  // Otherwise, figure out what the true and false return values are
2474  // so we can insert a new select instruction.
2475  Value *TrueValue = TrueRet->getReturnValue();
2476  Value *FalseValue = FalseRet->getReturnValue();
2477 
2478  // Unwrap any PHI nodes in the return blocks.
2479  if (PHINode *TVPN = dyn_cast_or_null<PHINode>(TrueValue))
2480  if (TVPN->getParent() == TrueSucc)
2481  TrueValue = TVPN->getIncomingValueForBlock(BI->getParent());
2482  if (PHINode *FVPN = dyn_cast_or_null<PHINode>(FalseValue))
2483  if (FVPN->getParent() == FalseSucc)
2484  FalseValue = FVPN->getIncomingValueForBlock(BI->getParent());
2485 
2486  // In order for this transformation to be safe, we must be able to
2487  // unconditionally execute both operands to the return. This is
2488  // normally the case, but we could have a potentially-trapping
2489  // constant expression that prevents this transformation from being
2490  // safe.
2491  if (ConstantExpr *TCV = dyn_cast_or_null<ConstantExpr>(TrueValue))
2492  if (TCV->canTrap())
2493  return false;
2494  if (ConstantExpr *FCV = dyn_cast_or_null<ConstantExpr>(FalseValue))
2495  if (FCV->canTrap())
2496  return false;
2497 
2498  // Okay, we collected all the mapped values and checked them for sanity, and
2499  // defined to really do this transformation. First, update the CFG.
2500  TrueSucc->removePredecessor(BI->getParent());
2501  FalseSucc->removePredecessor(BI->getParent());
2502 
2503  // Insert select instructions where needed.
2504  Value *BrCond = BI->getCondition();
2505  if (TrueValue) {
2506  // Insert a select if the results differ.
2507  if (TrueValue == FalseValue || isa<UndefValue>(FalseValue)) {
2508  } else if (isa<UndefValue>(TrueValue)) {
2509  TrueValue = FalseValue;
2510  } else {
2511  TrueValue =
2512  Builder.CreateSelect(BrCond, TrueValue, FalseValue, "retval", BI);
2513  }
2514  }
2515 
2516  Value *RI =
2517  !TrueValue ? Builder.CreateRetVoid() : Builder.CreateRet(TrueValue);
2518 
2519  (void)RI;
2520 
2521  LLVM_DEBUG(dbgs() << "\nCHANGING BRANCH TO TWO RETURNS INTO SELECT:"
2522  << "\n " << *BI << "NewRet = " << *RI << "TRUEBLOCK: "
2523  << *TrueSucc << "FALSEBLOCK: " << *FalseSucc);
2524 
2526 
2527  return true;
2528 }
2529 
2530 /// Return true if the given instruction is available
2531 /// in its predecessor block. If yes, the instruction will be removed.
2533  if (!isa<BinaryOperator>(Inst) && !isa<CmpInst>(Inst))
2534  return false;
2535  for (Instruction &I : *PB) {
2536  Instruction *PBI = &I;
2537  // Check whether Inst and PBI generate the same value.
2538  if (Inst->isIdenticalTo(PBI)) {
2539  Inst->replaceAllUsesWith(PBI);
2540  Inst->eraseFromParent();
2541  return true;
2542  }
2543  }
2544  return false;
2545 }
2546 
2547 /// Return true if either PBI or BI has branch weight available, and store
2548 /// the weights in {Pred|Succ}{True|False}Weight. If one of PBI and BI does
2549 /// not have branch weight, use 1:1 as its weight.
2551  uint64_t &PredTrueWeight,
2552  uint64_t &PredFalseWeight,
2553  uint64_t &SuccTrueWeight,
2554  uint64_t &SuccFalseWeight) {
2555  bool PredHasWeights =
2556  PBI->extractProfMetadata(PredTrueWeight, PredFalseWeight);
2557  bool SuccHasWeights =
2558  BI->extractProfMetadata(SuccTrueWeight, SuccFalseWeight);
2559  if (PredHasWeights || SuccHasWeights) {
2560  if (!PredHasWeights)
2561  PredTrueWeight = PredFalseWeight = 1;
2562  if (!SuccHasWeights)
2563  SuccTrueWeight = SuccFalseWeight = 1;
2564  return true;
2565  } else {
2566  return false;
2567  }
2568 }
2569 
2570 /// If this basic block is simple enough, and if a predecessor branches to us
2571 /// and one of our successors, fold the block into the predecessor and use
2572 /// logical operations to pick the right destination.
2574  unsigned BonusInstThreshold) {
2575  BasicBlock *BB = BI->getParent();
2576 
2577  const unsigned PredCount = pred_size(BB);
2578 
2579  Instruction *Cond = nullptr;
2580  if (BI->isConditional())
2581  Cond = dyn_cast<Instruction>(BI->getCondition());
2582  else {
2583  // For unconditional branch, check for a simple CFG pattern, where
2584  // BB has a single predecessor and BB's successor is also its predecessor's
2585  // successor. If such pattern exists, check for CSE between BB and its
2586  // predecessor.
2587  if (BasicBlock *PB = BB->getSinglePredecessor())
2588  if (BranchInst *PBI = dyn_cast<BranchInst>(PB->getTerminator()))
2589  if (PBI->isConditional() &&
2590  (BI->getSuccessor(0) == PBI->getSuccessor(0) ||
2591  BI->getSuccessor(0) == PBI->getSuccessor(1))) {
2592  for (auto I = BB->instructionsWithoutDebug().begin(),
2593  E = BB->instructionsWithoutDebug().end();
2594  I != E;) {
2595  Instruction *Curr = &*I++;
2596  if (isa<CmpInst>(Curr)) {
2597  Cond = Curr;
2598  break;
2599  }
2600  // Quit if we can't remove this instruction.
2601  if (!tryCSEWithPredecessor(Curr, PB))
2602  return false;
2603  }
2604  }
2605 
2606  if (!Cond)
2607  return false;
2608  }
2609 
2610  if (!Cond || (!isa<CmpInst>(Cond) && !isa<BinaryOperator>(Cond)) ||
2611  Cond->getParent() != BB || !Cond->hasOneUse())
2612  return false;
2613 
2614  // Make sure the instruction after the condition is the cond branch.
2615  BasicBlock::iterator CondIt = ++Cond->getIterator();
2616 
2617  // Ignore dbg intrinsics.
2618  while (isa<DbgInfoIntrinsic>(CondIt))
2619  ++CondIt;
2620 
2621  if (&*CondIt != BI)
2622  return false;
2623 
2624  // Only allow this transformation if computing the condition doesn't involve
2625  // too many instructions and these involved instructions can be executed
2626  // unconditionally. We denote all involved instructions except the condition
2627  // as "bonus instructions", and only allow this transformation when the
2628  // number of the bonus instructions we'll need to create when cloning into
2629  // each predecessor does not exceed a certain threshold.
2630  unsigned NumBonusInsts = 0;
2631  for (auto I = BB->begin(); Cond != &*I; ++I) {
2632  // Ignore dbg intrinsics.
2633  if (isa<DbgInfoIntrinsic>(I))
2634  continue;
2635  if (!I->hasOneUse() || !isSafeToSpeculativelyExecute(&*I))
2636  return false;
2637  // I has only one use and can be executed unconditionally.
2639  if (User == nullptr || User->getParent() != BB)
2640  return false;
2641  // I is used in the same BB. Since BI uses Cond and doesn't have more slots
2642  // to use any other instruction, User must be an instruction between next(I)
2643  // and Cond.
2644 
2645  // Account for the cost of duplicating this instruction into each
2646  // predecessor.
2647  NumBonusInsts += PredCount;
2648  // Early exits once we reach the limit.
2649  if (NumBonusInsts > BonusInstThreshold)
2650  return false;
2651  }
2652 
2653  // Cond is known to be a compare or binary operator. Check to make sure that
2654  // neither operand is a potentially-trapping constant expression.
2655  if (ConstantExpr *CE = dyn_cast<ConstantExpr>(Cond->getOperand(0)))
2656  if (CE->canTrap())
2657  return false;
2658  if (ConstantExpr *CE = dyn_cast<ConstantExpr>(Cond->getOperand(1)))
2659  if (CE->canTrap())
2660  return false;
2661 
2662  // Finally, don't infinitely unroll conditional loops.
2663  BasicBlock *TrueDest = BI->getSuccessor(0);
2664  BasicBlock *FalseDest = (BI->isConditional()) ? BI->getSuccessor(1) : nullptr;
2665  if (TrueDest == BB || FalseDest == BB)
2666  return false;
2667 
2668  for (pred_iterator PI = pred_begin(BB), E = pred_end(BB); PI != E; ++PI) {
2669  BasicBlock *PredBlock = *PI;
2670  BranchInst *PBI = dyn_cast<BranchInst>(PredBlock->getTerminator());
2671 
2672  // Check that we have two conditional branches. If there is a PHI node in
2673  // the common successor, verify that the same value flows in from both
2674  // blocks.
2676  if (!PBI || PBI->isUnconditional() ||
2677  (BI->isConditional() && !SafeToMergeTerminators(BI, PBI)) ||
2678  (!BI->isConditional() &&
2679  !isProfitableToFoldUnconditional(BI, PBI, Cond, PHIs)))
2680  continue;
2681 
2682  // Determine if the two branches share a common destination.
2683  Instruction::BinaryOps Opc = Instruction::BinaryOpsEnd;
2684  bool InvertPredCond = false;
2685 
2686  if (BI->isConditional()) {
2687  if (PBI->getSuccessor(0) == TrueDest) {
2688  Opc = Instruction::Or;
2689  } else if (PBI->getSuccessor(1) == FalseDest) {
2690  Opc = Instruction::And;
2691  } else if (PBI->getSuccessor(0) == FalseDest) {
2692  Opc = Instruction::And;
2693  InvertPredCond = true;
2694  } else if (PBI->getSuccessor(1) == TrueDest) {
2695  Opc = Instruction::Or;
2696  InvertPredCond = true;
2697  } else {
2698  continue;
2699  }
2700  } else {
2701  if (PBI->getSuccessor(0) != TrueDest && PBI->getSuccessor(1) != TrueDest)
2702  continue;
2703  }
2704 
2705  LLVM_DEBUG(dbgs() << "FOLDING BRANCH TO COMMON DEST:\n" << *PBI << *BB);
2706  IRBuilder<> Builder(PBI);
2707 
2708  // If we need to invert the condition in the pred block to match, do so now.
2709  if (InvertPredCond) {
2710  Value *NewCond = PBI->getCondition();
2711 
2712  if (NewCond->hasOneUse() && isa<CmpInst>(NewCond)) {
2713  CmpInst *CI = cast<CmpInst>(NewCond);
2714  CI->setPredicate(CI->getInversePredicate());
2715  } else {
2716  NewCond =
2717  Builder.CreateNot(NewCond, PBI->getCondition()->getName() + ".not");
2718  }
2719 
2720  PBI->setCondition(NewCond);
2721  PBI->swapSuccessors();
2722  }
2723 
2724  // If we have bonus instructions, clone them into the predecessor block.
2725  // Note that there may be multiple predecessor blocks, so we cannot move
2726  // bonus instructions to a predecessor block.
2727  ValueToValueMapTy VMap; // maps original values to cloned values
2728  // We already make sure Cond is the last instruction before BI. Therefore,
2729  // all instructions before Cond other than DbgInfoIntrinsic are bonus
2730  // instructions.
2731  for (auto BonusInst = BB->begin(); Cond != &*BonusInst; ++BonusInst) {
2732  if (isa<DbgInfoIntrinsic>(BonusInst))
2733  continue;
2734  Instruction *NewBonusInst = BonusInst->clone();
2735  RemapInstruction(NewBonusInst, VMap,
2737  VMap[&*BonusInst] = NewBonusInst;
2738 
2739  // If we moved a load, we cannot any longer claim any knowledge about
2740  // its potential value. The previous information might have been valid
2741  // only given the branch precondition.
2742  // For an analogous reason, we must also drop all the metadata whose
2743  // semantics we don't understand.
2744  NewBonusInst->dropUnknownNonDebugMetadata();
2745 
2746  PredBlock->getInstList().insert(PBI->getIterator(), NewBonusInst);
2747  NewBonusInst->takeName(&*BonusInst);
2748  BonusInst->setName(BonusInst->getName() + ".old");
2749  }
2750 
2751  // Clone Cond into the predecessor basic block, and or/and the
2752  // two conditions together.
2753  Instruction *CondInPred = Cond->clone();
2754  RemapInstruction(CondInPred, VMap,
2756  PredBlock->getInstList().insert(PBI->getIterator(), CondInPred);
2757  CondInPred->takeName(Cond);
2758  Cond->setName(CondInPred->getName() + ".old");
2759 
2760  if (BI->isConditional()) {
2761  Instruction *NewCond = cast<Instruction>(
2762  Builder.CreateBinOp(Opc, PBI->getCondition(), CondInPred, "or.cond"));
2763  PBI->setCondition(NewCond);
2764 
2765  uint64_t PredTrueWeight, PredFalseWeight, SuccTrueWeight, SuccFalseWeight;
2766  bool HasWeights =
2767  extractPredSuccWeights(PBI, BI, PredTrueWeight, PredFalseWeight,
2768  SuccTrueWeight, SuccFalseWeight);
2769  SmallVector<uint64_t, 8> NewWeights;
2770 
2771  if (PBI->getSuccessor(0) == BB) {
2772  if (HasWeights) {
2773  // PBI: br i1 %x, BB, FalseDest
2774  // BI: br i1 %y, TrueDest, FalseDest
2775  // TrueWeight is TrueWeight for PBI * TrueWeight for BI.
2776  NewWeights.push_back(PredTrueWeight * SuccTrueWeight);
2777  // FalseWeight is FalseWeight for PBI * TotalWeight for BI +
2778  // TrueWeight for PBI * FalseWeight for BI.
2779  // We assume that total weights of a BranchInst can fit into 32 bits.
2780  // Therefore, we will not have overflow using 64-bit arithmetic.
2781  NewWeights.push_back(PredFalseWeight *
2782  (SuccFalseWeight + SuccTrueWeight) +
2783  PredTrueWeight * SuccFalseWeight);
2784  }
2785  AddPredecessorToBlock(TrueDest, PredBlock, BB, MSSAU);
2786  PBI->setSuccessor(0, TrueDest);
2787  }
2788  if (PBI->getSuccessor(1) == BB) {
2789  if (HasWeights) {
2790  // PBI: br i1 %x, TrueDest, BB
2791  // BI: br i1 %y, TrueDest, FalseDest
2792  // TrueWeight is TrueWeight for PBI * TotalWeight for BI +
2793  // FalseWeight for PBI * TrueWeight for BI.
2794  NewWeights.push_back(PredTrueWeight *
2795  (SuccFalseWeight + SuccTrueWeight) +
2796  PredFalseWeight * SuccTrueWeight);
2797  // FalseWeight is FalseWeight for PBI * FalseWeight for BI.
2798  NewWeights.push_back(PredFalseWeight * SuccFalseWeight);
2799  }
2800  AddPredecessorToBlock(FalseDest, PredBlock, BB, MSSAU);
2801  PBI->setSuccessor(1, FalseDest);
2802  }
2803  if (NewWeights.size() == 2) {
2804  // Halve the weights if any of them cannot fit in an uint32_t
2805  FitWeights(NewWeights);
2806 
2807  SmallVector<uint32_t, 8> MDWeights(NewWeights.begin(),
2808  NewWeights.end());
2809  setBranchWeights(PBI, MDWeights[0], MDWeights[1]);
2810  } else
2811  PBI->setMetadata(LLVMContext::MD_prof, nullptr);
2812  } else {
2813  // Update PHI nodes in the common successors.
2814  for (unsigned i = 0, e = PHIs.size(); i != e; ++i) {
2815  ConstantInt *PBI_C = cast<ConstantInt>(
2816  PHIs[i]->getIncomingValueForBlock(PBI->getParent()));
2817  assert(PBI_C->getType()->isIntegerTy(1));
2818  Instruction *MergedCond = nullptr;
2819  if (PBI->getSuccessor(0) == TrueDest) {
2820  // Create (PBI_Cond and PBI_C) or (!PBI_Cond and BI_Value)
2821  // PBI_C is true: PBI_Cond or (!PBI_Cond and BI_Value)
2822  // is false: !PBI_Cond and BI_Value
2823  Instruction *NotCond = cast<Instruction>(
2824  Builder.CreateNot(PBI->getCondition(), "not.cond"));
2825  MergedCond = cast<Instruction>(
2826  Builder.CreateBinOp(Instruction::And, NotCond, CondInPred,
2827  "and.cond"));
2828  if (PBI_C->isOne())
2829  MergedCond = cast<Instruction>(Builder.CreateBinOp(
2830  Instruction::Or, PBI->getCondition(), MergedCond, "or.cond"));
2831  } else {
2832  // Create (PBI_Cond and BI_Value) or (!PBI_Cond and PBI_C)
2833  // PBI_C is true: (PBI_Cond and BI_Value) or (!PBI_Cond)
2834  // is false: PBI_Cond and BI_Value
2835  MergedCond = cast<Instruction>(Builder.CreateBinOp(
2836  Instruction::And, PBI->getCondition(), CondInPred, "and.cond"));
2837  if (PBI_C->isOne()) {
2838  Instruction *NotCond = cast<Instruction>(
2839  Builder.CreateNot(PBI->getCondition(), "not.cond"));
2840  MergedCond = cast<Instruction>(Builder.CreateBinOp(
2841  Instruction::Or, NotCond, MergedCond, "or.cond"));
2842  }
2843  }
2844  // Update PHI Node.
2845  PHIs[i]->setIncomingValueForBlock(PBI->getParent(), MergedCond);
2846  }
2847 
2848  // PBI is changed to branch to TrueDest below. Remove itself from
2849  // potential phis from all other successors.
2850  if (MSSAU)
2851  MSSAU->changeCondBranchToUnconditionalTo(PBI, TrueDest);
2852 
2853  // Change PBI from Conditional to Unconditional.
2854  BranchInst *New_PBI = BranchInst::Create(TrueDest, PBI);
2855  EraseTerminatorAndDCECond(PBI, MSSAU);
2856  PBI = New_PBI;
2857  }
2858 
2859  // If BI was a loop latch, it may have had associated loop metadata.
2860  // We need to copy it to the new latch, that is, PBI.
2861  if (MDNode *LoopMD = BI->getMetadata(LLVMContext::MD_loop))
2862  PBI->setMetadata(LLVMContext::MD_loop, LoopMD);
2863 
2864  // TODO: If BB is reachable from all paths through PredBlock, then we
2865  // could replace PBI's branch probabilities with BI's.
2866 
2867  // Copy any debug value intrinsics into the end of PredBlock.
2868  for (Instruction &I : *BB)
2869  if (isa<DbgInfoIntrinsic>(I))
2870  I.clone()->insertBefore(PBI);
2871 
2872  return true;
2873  }
2874  return false;
2875 }
2876 
2877 // If there is only one store in BB1 and BB2, return it, otherwise return
2878 // nullptr.
2880  StoreInst *S = nullptr;
2881  for (auto *BB : {BB1, BB2}) {
2882  if (!BB)
2883  continue;
2884  for (auto &I : *BB)
2885  if (auto *SI = dyn_cast<StoreInst>(&I)) {
2886  if (S)
2887  // Multiple stores seen.
2888  return nullptr;
2889  else
2890  S = SI;
2891  }
2892  }
2893  return S;
2894 }
2895 
2897  Value *AlternativeV = nullptr) {
2898  // PHI is going to be a PHI node that allows the value V that is defined in
2899  // BB to be referenced in BB's only successor.
2900  //
2901  // If AlternativeV is nullptr, the only value we care about in PHI is V. It
2902  // doesn't matter to us what the other operand is (it'll never get used). We
2903  // could just create a new PHI with an undef incoming value, but that could
2904  // increase register pressure if EarlyCSE/InstCombine can't fold it with some
2905  // other PHI. So here we directly look for some PHI in BB's successor with V
2906  // as an incoming operand. If we find one, we use it, else we create a new
2907  // one.
2908  //
2909  // If AlternativeV is not nullptr, we care about both incoming values in PHI.
2910  // PHI must be exactly: phi <ty> [ %BB, %V ], [ %OtherBB, %AlternativeV]
2911  // where OtherBB is the single other predecessor of BB's only successor.
2912  PHINode *PHI = nullptr;
2913  BasicBlock *Succ = BB->getSingleSuccessor();
2914 
2915  for (auto I = Succ->begin(); isa<PHINode>(I); ++I)
2916  if (cast<PHINode>(I)->getIncomingValueForBlock(BB) == V) {
2917  PHI = cast<PHINode>(I);
2918  if (!AlternativeV)
2919  break;
2920 
2921  assert(Succ->hasNPredecessors(2));
2922  auto PredI = pred_begin(Succ);
2923  BasicBlock *OtherPredBB = *PredI == BB ? *++PredI : *PredI;
2924  if (PHI->getIncomingValueForBlock(OtherPredBB) == AlternativeV)
2925  break;
2926  PHI = nullptr;
2927  }
2928  if (PHI)
2929  return PHI;
2930 
2931  // If V is not an instruction defined in BB, just return it.
2932  if (!AlternativeV &&
2933  (!isa<Instruction>(V) || cast<Instruction>(V)->getParent() != BB))
2934  return V;
2935 
2936  PHI = PHINode::Create(V->getType(), 2, "simplifycfg.merge", &Succ->front());
2937  PHI->addIncoming(V, BB);
2938  for (BasicBlock *PredBB : predecessors(Succ))
2939  if (PredBB != BB)
2940  PHI->addIncoming(
2941  AlternativeV ? AlternativeV : UndefValue::get(V->getType()), PredBB);
2942  return PHI;
2943 }
2944 
2946  BasicBlock *QTB, BasicBlock *QFB,
2947  BasicBlock *PostBB, Value *Address,
2948  bool InvertPCond, bool InvertQCond,
2949  const DataLayout &DL,
2950  const TargetTransformInfo &TTI) {
2951  // For every pointer, there must be exactly two stores, one coming from
2952  // PTB or PFB, and the other from QTB or QFB. We don't support more than one
2953  // store (to any address) in PTB,PFB or QTB,QFB.
2954  // FIXME: We could relax this restriction with a bit more work and performance
2955  // testing.
2956  StoreInst *PStore = findUniqueStoreInBlocks(PTB, PFB);
2957  StoreInst *QStore = findUniqueStoreInBlocks(QTB, QFB);
2958  if (!PStore || !QStore)
2959  return false;
2960 
2961  // Now check the stores are compatible.
2962  if (!QStore->isUnordered() || !PStore->isUnordered())
2963  return false;
2964 
2965  // Check that sinking the store won't cause program behavior changes. Sinking
2966  // the store out of the Q blocks won't change any behavior as we're sinking
2967  // from a block to its unconditional successor. But we're moving a store from
2968  // the P blocks down through the middle block (QBI) and past both QFB and QTB.
2969  // So we need to check that there are no aliasing loads or stores in
2970  // QBI, QTB and QFB. We also need to check there are no conflicting memory
2971  // operations between PStore and the end of its parent block.
2972  //
2973  // The ideal way to do this is to query AliasAnalysis, but we don't
2974  // preserve AA currently so that is dangerous. Be super safe and just
2975  // check there are no other memory operations at all.
2976  for (auto &I : *QFB->getSinglePredecessor())
2977  if (I.mayReadOrWriteMemory())
2978  return false;
2979  for (auto &I : *QFB)
2980  if (&I != QStore && I.mayReadOrWriteMemory())
2981  return false;
2982  if (QTB)
2983  for (auto &I : *QTB)
2984  if (&I != QStore && I.mayReadOrWriteMemory())
2985  return false;
2986  for (auto I = BasicBlock::iterator(PStore), E = PStore->getParent()->end();
2987  I != E; ++I)
2988  if (&*I != PStore && I->mayReadOrWriteMemory())
2989  return false;
2990 
2991  // If we're not in aggressive mode, we only optimize if we have some
2992  // confidence that by optimizing we'll allow P and/or Q to be if-converted.
2993  auto IsWorthwhile = [&](BasicBlock *BB, ArrayRef<StoreInst *> FreeStores) {
2994  if (!BB)
2995  return true;
2996  // Heuristic: if the block can be if-converted/phi-folded and the
2997  // instructions inside are all cheap (arithmetic/GEPs), it's worthwhile to
2998  // thread this store.
2999  int BudgetRemaining =
3001  for (auto &I : BB->instructionsWithoutDebug()) {
3002  // Consider terminator instruction to be free.
3003  if (I.isTerminator())
3004  continue;
3005  // If this is one the stores that we want to speculate out of this BB,
3006  // then don't count it's cost, consider it to be free.
3007  if (auto *S = dyn_cast<StoreInst>(&I))
3008  if (llvm::find(FreeStores, S))
3009  continue;
3010  // Else, we have a white-list of instructions that we are ak speculating.
3011  if (!isa<BinaryOperator>(I) && !isa<GetElementPtrInst>(I))
3012  return false; // Not in white-list - not worthwhile folding.
3013  // And finally, if this is a non-free instruction that we are okay
3014  // speculating, ensure that we consider the speculation budget.
3015  BudgetRemaining -= TTI.getUserCost(&I);
3016  if (BudgetRemaining < 0)
3017  return false; // Eagerly refuse to fold as soon as we're out of budget.
3018  }
3019  assert(BudgetRemaining >= 0 &&
3020  "When we run out of budget we will eagerly return from within the "
3021  "per-instruction loop.");
3022  return true;
3023  };
3024 
3025  const SmallVector<StoreInst *, 2> FreeStores = {PStore, QStore};
3027  (!IsWorthwhile(PTB, FreeStores) || !IsWorthwhile(PFB, FreeStores) ||
3028  !IsWorthwhile(QTB, FreeStores) || !IsWorthwhile(QFB, FreeStores)))
3029  return false;
3030 
3031  // If PostBB has more than two predecessors, we need to split it so we can
3032  // sink the store.
3033  if (std::next(pred_begin(PostBB), 2) != pred_end(PostBB)) {
3034  // We know that QFB's only successor is PostBB. And QFB has a single
3035  // predecessor. If QTB exists, then its only successor is also PostBB.
3036  // If QTB does not exist, then QFB's only predecessor has a conditional
3037  // branch to QFB and PostBB.
3038  BasicBlock *TruePred = QTB ? QTB : QFB->getSinglePredecessor();
3039  BasicBlock *NewBB = SplitBlockPredecessors(PostBB, { QFB, TruePred},
3040  "condstore.split");
3041  if (!NewBB)
3042  return false;
3043  PostBB = NewBB;
3044  }
3045 
3046  // OK, we're going to sink the stores to PostBB. The store has to be
3047  // conditional though, so first create the predicate.
3048  Value *PCond = cast<BranchInst>(PFB->getSinglePredecessor()->getTerminator())
3049  ->getCondition();
3050  Value *QCond = cast<BranchInst>(QFB->getSinglePredecessor()->getTerminator())
3051  ->getCondition();
3052 
3054  PStore->getParent());
3055  Value *QPHI = ensureValueAvailableInSuccessor(QStore->getValueOperand(),
3056  QStore->getParent(), PPHI);
3057 
3058  IRBuilder<> QB(&*PostBB->getFirstInsertionPt());
3059 
3060  Value *PPred = PStore->getParent() == PTB ? PCond : QB.CreateNot(PCond);
3061  Value *QPred = QStore->getParent() == QTB ? QCond : QB.CreateNot(QCond);
3062 
3063  if (InvertPCond)
3064  PPred = QB.CreateNot(PPred);
3065  if (InvertQCond)
3066  QPred = QB.CreateNot(QPred);
3067  Value *CombinedPred = QB.CreateOr(PPred, QPred);
3068 
3069  auto *T =
3070  SplitBlockAndInsertIfThen(CombinedPred, &*QB.GetInsertPoint(), false);
3071  QB.SetInsertPoint(T);
3072  StoreInst *SI = cast<StoreInst>(QB.CreateStore(QPHI, Address));
3073  AAMDNodes AAMD;
3074  PStore->getAAMetadata(AAMD, /*Merge=*/false);
3075  PStore->getAAMetadata(AAMD, /*Merge=*/true);
3076  SI->setAAMetadata(AAMD);
3077  unsigned PAlignment = PStore->getAlignment();
3078  unsigned QAlignment = QStore->getAlignment();
3079  unsigned TypeAlignment =
3081  unsigned MinAlignment;
3082  unsigned MaxAlignment;
3083  std::tie(MinAlignment, MaxAlignment) = std::minmax(PAlignment, QAlignment);
3084  // Choose the minimum alignment. If we could prove both stores execute, we
3085  // could use biggest one. In this case, though, we only know that one of the
3086  // stores executes. And we don't know it's safe to take the alignment from a
3087  // store that doesn't execute.
3088  if (MinAlignment != 0) {
3089  // Choose the minimum of all non-zero alignments.
3090  SI->setAlignment(Align(MinAlignment));
3091  } else if (MaxAlignment != 0) {
3092  // Choose the minimal alignment between the non-zero alignment and the ABI
3093  // default alignment for the type of the stored value.
3094  SI->setAlignment(Align(std::min(MaxAlignment, TypeAlignment)));
3095  } else {
3096  // If both alignments are zero, use ABI default alignment for the type of
3097  // the stored value.
3098  SI->setAlignment(Align(TypeAlignment));
3099  }
3100 
3101  QStore->eraseFromParent();
3102  PStore->eraseFromParent();
3103 
3104  return true;
3105 }
3106 
3108  const DataLayout &DL,
3109  const TargetTransformInfo &TTI) {
3110  // The intention here is to find diamonds or triangles (see below) where each
3111  // conditional block contains a store to the same address. Both of these
3112  // stores are conditional, so they can't be unconditionally sunk. But it may
3113  // be profitable to speculatively sink the stores into one merged store at the
3114  // end, and predicate the merged store on the union of the two conditions of
3115  // PBI and QBI.
3116  //
3117  // This can reduce the number of stores executed if both of the conditions are
3118  // true, and can allow the blocks to become small enough to be if-converted.
3119  // This optimization will also chain, so that ladders of test-and-set
3120  // sequences can be if-converted away.
3121  //
3122  // We only deal with simple diamonds or triangles:
3123  //
3124  // PBI or PBI or a combination of the two
3125  // / \ | \
3126  // PTB PFB | PFB
3127  // \ / | /
3128  // QBI QBI
3129  // / \ | \
3130  // QTB QFB | QFB
3131  // \ / | /
3132  // PostBB PostBB
3133  //
3134  // We model triangles as a type of diamond with a nullptr "true" block.
3135  // Triangles are canonicalized so that the fallthrough edge is represented by
3136  // a true condition, as in the diagram above.
3137  BasicBlock *PTB = PBI->getSuccessor(0);
3138  BasicBlock *PFB = PBI->getSuccessor(1);
3139  BasicBlock *QTB = QBI->getSuccessor(0);
3140  BasicBlock *QFB = QBI->getSuccessor(1);
3141  BasicBlock *PostBB = QFB->getSingleSuccessor();
3142 
3143  // Make sure we have a good guess for PostBB. If QTB's only successor is
3144  // QFB, then QFB is a better PostBB.
3145  if (QTB->getSingleSuccessor() == QFB)
3146  PostBB = QFB;
3147 
3148  // If we couldn't find a good PostBB, stop.
3149  if (!PostBB)
3150  return false;
3151 
3152  bool InvertPCond = false, InvertQCond = false;
3153  // Canonicalize fallthroughs to the true branches.
3154  if (PFB == QBI->getParent()) {
3155  std::swap(PFB, PTB);
3156  InvertPCond = true;
3157  }
3158  if (QFB == PostBB) {
3159  std::swap(QFB, QTB);
3160  InvertQCond = true;
3161  }
3162 
3163  // From this point on we can assume PTB or QTB may be fallthroughs but PFB
3164  // and QFB may not. Model fallthroughs as a nullptr block.
3165  if (PTB == QBI->getParent())
3166  PTB = nullptr;
3167  if (QTB == PostBB)
3168  QTB = nullptr;
3169 
3170  // Legality bailouts. We must have at least the non-fallthrough blocks and
3171  // the post-dominating block, and the non-fallthroughs must only have one
3172  // predecessor.
3173  auto HasOnePredAndOneSucc = [](BasicBlock *BB, BasicBlock *P, BasicBlock *S) {
3174  return BB->getSinglePredecessor() == P && BB->getSingleSuccessor() == S;
3175  };
3176  if (!HasOnePredAndOneSucc(PFB, PBI->getParent(), QBI->getParent()) ||
3177  !HasOnePredAndOneSucc(QFB, QBI->getParent(), PostBB))
3178  return false;
3179  if ((PTB && !HasOnePredAndOneSucc(PTB, PBI->getParent(), QBI->getParent())) ||
3180  (QTB && !HasOnePredAndOneSucc(QTB, QBI->getParent(), PostBB)))
3181  return false;
3182  if (!QBI->getParent()->hasNUses(2))
3183  return false;
3184 
3185  // OK, this is a sequence of two diamonds or triangles.
3186  // Check if there are stores in PTB or PFB that are repeated in QTB or QFB.
3187  SmallPtrSet<Value *, 4> PStoreAddresses, QStoreAddresses;
3188  for (auto *BB : {PTB, PFB}) {
3189  if (!BB)
3190  continue;
3191  for (auto &I : *BB)
3192  if (StoreInst *SI = dyn_cast<StoreInst>(&I))
3193  PStoreAddresses.insert(SI->getPointerOperand());
3194  }
3195  for (auto *BB : {QTB, QFB}) {
3196  if (!BB)
3197  continue;
3198  for (auto &I : *BB)
3199  if (StoreInst *SI = dyn_cast<StoreInst>(&I))
3200  QStoreAddresses.insert(SI->getPointerOperand());
3201  }
3202 
3203  set_intersect(PStoreAddresses, QStoreAddresses);
3204  // set_intersect mutates PStoreAddresses in place. Rename it here to make it
3205  // clear what it contains.
3206  auto &CommonAddresses = PStoreAddresses;
3207 
3208  bool Changed = false;
3209  for (auto *Address : CommonAddresses)
3210  Changed |= mergeConditionalStoreToAddress(
3211  PTB, PFB, QTB, QFB, PostBB, Address, InvertPCond, InvertQCond, DL, TTI);
3212  return Changed;
3213 }
3214 
3215 /// If we have a conditional branch as a predecessor of another block,
3216 /// this function tries to simplify it. We know
3217 /// that PBI and BI are both conditional branches, and BI is in one of the
3218 /// successor blocks of PBI - PBI branches to BI.
3220  const DataLayout &DL,
3221  const TargetTransformInfo &TTI) {
3222  assert(PBI->isConditional() && BI->isConditional());
3223  BasicBlock *BB = BI->getParent();
3224 
3225  // If this block ends with a branch instruction, and if there is a
3226  // predecessor that ends on a branch of the same condition, make
3227  // this conditional branch redundant.
3228  if (PBI->getCondition() == BI->getCondition() &&
3229  PBI->getSuccessor(0) != PBI->getSuccessor(1)) {
3230  // Okay, the outcome of this conditional branch is statically
3231  // knowable. If this block had a single pred, handle specially.
3232  if (BB->getSinglePredecessor()) {
3233  // Turn this into a branch on constant.
3234  bool CondIsTrue = PBI->getSuccessor(0) == BB;
3235  BI->setCondition(
3236  ConstantInt::get(Type::getInt1Ty(BB->getContext()), CondIsTrue));
3237  return true; // Nuke the branch on constant.
3238  }
3239 
3240  // Otherwise, if there are multiple predecessors, insert a PHI that merges
3241  // in the constant and simplify the block result. Subsequent passes of
3242  // simplifycfg will thread the block.
3244  pred_iterator PB = pred_begin(BB), PE = pred_end(BB);
3245  PHINode *NewPN = PHINode::Create(
3246  Type::getInt1Ty(BB->getContext()), std::distance(PB, PE),
3247  BI->getCondition()->getName() + ".pr", &BB->front());
3248  // Okay, we're going to insert the PHI node. Since PBI is not the only
3249  // predecessor, compute the PHI'd conditional value for all of the preds.
3250  // Any predecessor where the condition is not computable we keep symbolic.
3251  for (pred_iterator PI = PB; PI != PE; ++PI) {
3252  BasicBlock *P = *PI;
3253  if ((PBI = dyn_cast<BranchInst>(P->getTerminator())) && PBI != BI &&
3254  PBI->isConditional() && PBI->getCondition() == BI->getCondition() &&
3255  PBI->getSuccessor(0) != PBI->getSuccessor(1)) {
3256  bool CondIsTrue = PBI->getSuccessor(0) == BB;
3257  NewPN->addIncoming(
3258  ConstantInt::get(Type::getInt1Ty(BB->getContext()), CondIsTrue),
3259  P);
3260  } else {
3261  NewPN->addIncoming(BI->getCondition(), P);
3262  }
3263  }
3264 
3265  BI->setCondition(NewPN);
3266  return true;
3267  }
3268  }
3269 
3270  if (auto *CE = dyn_cast<ConstantExpr>(BI->getCondition()))
3271  if (CE->canTrap())
3272  return false;
3273 
3274  // If both branches are conditional and both contain stores to the same
3275  // address, remove the stores from the conditionals and create a conditional
3276  // merged store at the end.
3277  if (MergeCondStores && mergeConditionalStores(PBI, BI, DL, TTI))
3278  return true;
3279 
3280  // If this is a conditional branch in an empty block, and if any
3281  // predecessors are a conditional branch to one of our destinations,
3282  // fold the conditions into logical ops and one cond br.
3283 
3284  // Ignore dbg intrinsics.
3285  if (&*BB->instructionsWithoutDebug().begin() != BI)
3286  return false;
3287 
3288  int PBIOp, BIOp;
3289  if (PBI->getSuccessor(0) == BI->getSuccessor(0)) {
3290  PBIOp = 0;
3291  BIOp = 0;
3292  } else if (PBI->getSuccessor(0) == BI->getSuccessor(1)) {
3293  PBIOp = 0;
3294  BIOp = 1;
3295  } else if (PBI->getSuccessor(1) == BI->getSuccessor(0)) {
3296  PBIOp = 1;
3297  BIOp = 0;
3298  } else if (PBI->getSuccessor(1) == BI->getSuccessor(1)) {
3299  PBIOp = 1;
3300  BIOp = 1;
3301  } else {
3302  return false;
3303  }
3304 
3305  // Check to make sure that the other destination of this branch
3306  // isn't BB itself. If so, this is an infinite loop that will
3307  // keep getting unwound.
3308  if (PBI->getSuccessor(PBIOp) == BB)
3309  return false;
3310 
3311  // Do not perform this transformation if it would require
3312  // insertion of a large number of select instructions. For targets
3313  // without predication/cmovs, this is a big pessimization.
3314 
3315  // Also do not perform this transformation if any phi node in the common
3316  // destination block can trap when reached by BB or PBB (PR17073). In that
3317  // case, it would be unsafe to hoist the operation into a select instruction.
3318 
3319  BasicBlock *CommonDest = PBI->getSuccessor(PBIOp);
3320  unsigned NumPhis = 0;
3321  for (BasicBlock::iterator II = CommonDest->begin(); isa<PHINode>(II);
3322  ++II, ++NumPhis) {
3323  if (NumPhis > 2) // Disable this xform.
3324  return false;
3325 
3326  PHINode *PN = cast<PHINode>(II);
3327  Value *BIV = PN->getIncomingValueForBlock(BB);
3328  if (ConstantExpr *CE = dyn_cast<ConstantExpr>(BIV))
3329  if (CE->canTrap())
3330  return false;
3331 
3332  unsigned PBBIdx = PN->getBasicBlockIndex(PBI->getParent());
3333  Value *PBIV = PN->getIncomingValue(PBBIdx);
3334  if (ConstantExpr *CE = dyn_cast<ConstantExpr>(PBIV))
3335  if (CE->canTrap())
3336  return false;
3337  }
3338 
3339  // Finally, if everything is ok, fold the branches to logical ops.
3340  BasicBlock *OtherDest = BI->getSuccessor(BIOp ^ 1);
3341 
3342  LLVM_DEBUG(dbgs() << "FOLDING BRs:" << *PBI->getParent()
3343  << "AND: " << *BI->getParent());
3344 
3345  // If OtherDest *is* BB, then BB is a basic block with a single conditional
3346  // branch in it, where one edge (OtherDest) goes back to itself but the other
3347  // exits. We don't *know* that the program avoids the infinite loop
3348  // (even though that seems likely). If we do this xform naively, we'll end up
3349  // recursively unpeeling the loop. Since we know that (after the xform is
3350  // done) that the block *is* infinite if reached, we just make it an obviously
3351  // infinite loop with no cond branch.
3352  if (OtherDest == BB) {
3353  // Insert it at the end of the function, because it's either code,
3354  // or it won't matter if it's hot. :)
3355  BasicBlock *InfLoopBlock =
3356  BasicBlock::Create(BB->getContext(), "infloop", BB->getParent());
3357  BranchInst::Create(InfLoopBlock, InfLoopBlock);
3358  OtherDest = InfLoopBlock;
3359  }
3360 
3361  LLVM_DEBUG(dbgs() << *PBI->getParent()->getParent());
3362 
3363  // BI may have other predecessors. Because of this, we leave
3364  // it alone, but modify PBI.
3365 
3366  // Make sure we get to CommonDest on True&True directions.
3367  Value *PBICond = PBI->getCondition();
3368  IRBuilder<NoFolder> Builder(PBI);
3369  if (PBIOp)
3370  PBICond = Builder.CreateNot(PBICond, PBICond->getName() + ".not");
3371 
3372  Value *BICond = BI->getCondition();
3373  if (BIOp)
3374  BICond = Builder.CreateNot(BICond, BICond->getName() + ".not");
3375 
3376  // Merge the conditions.
3377  Value *Cond = Builder.CreateOr(PBICond, BICond, "brmerge");
3378 
3379  // Modify PBI to branch on the new condition to the new dests.
3380  PBI->setCondition(Cond);
3381  PBI->setSuccessor(0, CommonDest);
3382  PBI->setSuccessor(1, OtherDest);
3383 
3384  // Update branch weight for PBI.
3385  uint64_t PredTrueWeight, PredFalseWeight, SuccTrueWeight, SuccFalseWeight;
3386  uint64_t PredCommon, PredOther, SuccCommon, SuccOther;
3387  bool HasWeights =
3388  extractPredSuccWeights(PBI, BI, PredTrueWeight, PredFalseWeight,
3389  SuccTrueWeight, SuccFalseWeight);
3390  if (HasWeights) {
3391  PredCommon = PBIOp ? PredFalseWeight : PredTrueWeight;
3392  PredOther = PBIOp ? PredTrueWeight : PredFalseWeight;
3393  SuccCommon = BIOp ? SuccFalseWeight : SuccTrueWeight;
3394  SuccOther = BIOp ? SuccTrueWeight : SuccFalseWeight;
3395  // The weight to CommonDest should be PredCommon * SuccTotal +
3396  // PredOther * SuccCommon.
3397  // The weight to OtherDest should be PredOther * SuccOther.
3398  uint64_t NewWeights[2] = {PredCommon * (SuccCommon + SuccOther) +
3399  PredOther * SuccCommon,
3400  PredOther * SuccOther};
3401  // Halve the weights if any of them cannot fit in an uint32_t
3402  FitWeights(NewWeights);
3403 
3404  setBranchWeights(PBI, NewWeights[0], NewWeights[1]);
3405  }
3406 
3407  // OtherDest may have phi nodes. If so, add an entry from PBI's
3408  // block that are identical to the entries for BI's block.
3409  AddPredecessorToBlock(OtherDest, PBI->getParent(), BB);
3410 
3411  // We know that the CommonDest already had an edge from PBI to
3412  // it. If it has PHIs though, the PHIs may have different
3413  // entries for BB and PBI's BB. If so, insert a select to make
3414  // them agree.
3415  for (PHINode &PN : CommonDest->phis()) {
3416  Value *BIV = PN.getIncomingValueForBlock(BB);
3417  unsigned PBBIdx = PN.getBasicBlockIndex(PBI->getParent());
3418  Value *PBIV = PN.getIncomingValue(PBBIdx);
3419  if (BIV != PBIV) {
3420  // Insert a select in PBI to pick the right value.
3421  SelectInst *NV = cast<SelectInst>(
3422  Builder.CreateSelect(PBICond, PBIV, BIV, PBIV->getName() + ".mux"));
3423  PN.setIncomingValue(PBBIdx, NV);
3424  // Although the select has the same condition as PBI, the original branch
3425  // weights for PBI do not apply to the new select because the select's
3426  // 'logical' edges are incoming edges of the phi that is eliminated, not
3427  // the outgoing edges of PBI.
3428  if (HasWeights) {
3429  uint64_t PredCommon = PBIOp ? PredFalseWeight : PredTrueWeight;
3430  uint64_t PredOther = PBIOp ? PredTrueWeight : PredFalseWeight;
3431  uint64_t SuccCommon = BIOp ? SuccFalseWeight : SuccTrueWeight;
3432  uint64_t SuccOther = BIOp ? SuccTrueWeight : SuccFalseWeight;
3433  // The weight to PredCommonDest should be PredCommon * SuccTotal.
3434  // The weight to PredOtherDest should be PredOther * SuccCommon.
3435  uint64_t NewWeights[2] = {PredCommon * (SuccCommon + SuccOther),
3436  PredOther * SuccCommon};
3437 
3438  FitWeights(NewWeights);
3439 
3440  setBranchWeights(NV, NewWeights[0], NewWeights[1]);
3441  }
3442  }
3443  }
3444 
3445  LLVM_DEBUG(dbgs() << "INTO: " << *PBI->getParent());
3446  LLVM_DEBUG(dbgs() << *PBI->getParent()->getParent());
3447 
3448  // This basic block is probably dead. We know it has at least
3449  // one fewer predecessor.
3450  return true;
3451 }
3452 
3453 // Simplifies a terminator by replacing it with a branch to TrueBB if Cond is
3454 // true or to FalseBB if Cond is false.
3455 // Takes care of updating the successors and removing the old terminator.
3456 // Also makes sure not to introduce new successors by assuming that edges to
3457 // non-successor TrueBBs and FalseBBs aren't reachable.
3458 static bool SimplifyTerminatorOnSelect(Instruction *OldTerm, Value *Cond,
3459  BasicBlock *TrueBB, BasicBlock *FalseBB,
3460  uint32_t TrueWeight,
3461  uint32_t FalseWeight) {
3462  // Remove any superfluous successor edges from the CFG.
3463  // First, figure out which successors to preserve.
3464  // If TrueBB and FalseBB are equal, only try to preserve one copy of that
3465  // successor.
3466  BasicBlock *KeepEdge1 = TrueBB;
3467  BasicBlock *KeepEdge2 = TrueBB != FalseBB ? FalseBB : nullptr;
3468 
3469  // Then remove the rest.
3470  for (BasicBlock *Succ : successors(OldTerm)) {
3471  // Make sure only to keep exactly one copy of each edge.
3472  if (Succ == KeepEdge1)
3473  KeepEdge1 = nullptr;
3474  else if (Succ == KeepEdge2)
3475  KeepEdge2 = nullptr;
3476  else
3477  Succ->removePredecessor(OldTerm->getParent(),
3478  /*KeepOneInputPHIs=*/true);
3479  }
3480 
3481  IRBuilder<> Builder(OldTerm);
3482  Builder.SetCurrentDebugLocation(OldTerm->getDebugLoc());
3483 
3484  // Insert an appropriate new terminator.
3485  if (!KeepEdge1 && !KeepEdge2) {
3486  if (TrueBB == FalseBB)
3487  // We were only looking for one successor, and it was present.
3488  // Create an unconditional branch to it.
3489  Builder.CreateBr(TrueBB);
3490  else {
3491  // We found both of the successors we were looking for.
3492  // Create a conditional branch sharing the condition of the select.
3493  BranchInst *NewBI = Builder.CreateCondBr(Cond, TrueBB, FalseBB);
3494  if (TrueWeight != FalseWeight)
3495  setBranchWeights(NewBI, TrueWeight, FalseWeight);
3496  }
3497  } else if (KeepEdge1 && (KeepEdge2 || TrueBB == FalseBB)) {
3498  // Neither of the selected blocks were successors, so this
3499  // terminator must be unreachable.
3500  new UnreachableInst(OldTerm->getContext(), OldTerm);
3501  } else {
3502  // One of the selected values was a successor, but the other wasn't.
3503  // Insert an unconditional branch to the one that was found;
3504  // the edge to the one that wasn't must be unreachable.
3505  if (!KeepEdge1)
3506  // Only TrueBB was found.
3507  Builder.CreateBr(TrueBB);
3508  else
3509  // Only FalseBB was found.
3510  Builder.CreateBr(FalseBB);
3511  }
3512 
3513  EraseTerminatorAndDCECond(OldTerm);
3514  return true;
3515 }
3516 
3517 // Replaces
3518 // (switch (select cond, X, Y)) on constant X, Y
3519 // with a branch - conditional if X and Y lead to distinct BBs,
3520 // unconditional otherwise.
3522  // Check for constant integer values in the select.
3523  ConstantInt *TrueVal = dyn_cast<ConstantInt>(Select->getTrueValue());
3524  ConstantInt *FalseVal = dyn_cast<ConstantInt>(Select->getFalseValue());
3525  if (!TrueVal || !FalseVal)
3526  return false;
3527 
3528  // Find the relevant condition and destinations.
3529  Value *Condition = Select->getCondition();
3530  BasicBlock *TrueBB = SI->findCaseValue(TrueVal)->getCaseSuccessor();
3531  BasicBlock *FalseBB = SI->findCaseValue(FalseVal)->getCaseSuccessor();
3532 
3533  // Get weight for TrueBB and FalseBB.
3534  uint32_t TrueWeight = 0, FalseWeight = 0;
3535  SmallVector<uint64_t, 8> Weights;
3536  bool HasWeights = HasBranchWeights(SI);
3537  if (HasWeights) {
3538  GetBranchWeights(SI, Weights);
3539  if (Weights.size() == 1 + SI->getNumCases()) {
3540  TrueWeight =
3541  (uint32_t)Weights[SI->findCaseValue(TrueVal)->getSuccessorIndex()];
3542  FalseWeight =
3543  (uint32_t)Weights[SI->findCaseValue(FalseVal)->getSuccessorIndex()];
3544  }
3545  }
3546 
3547  // Perform the actual simplification.
3548  return SimplifyTerminatorOnSelect(SI, Condition, TrueBB, FalseBB, TrueWeight,
3549  FalseWeight);
3550 }
3551 
3552 // Replaces
3553 // (indirectbr (select cond, blockaddress(@fn, BlockA),
3554 // blockaddress(@fn, BlockB)))
3555 // with
3556 // (br cond, BlockA, BlockB).
3558  // Check that both operands of the select are block addresses.
3561  if (!TBA || !FBA)
3562  return false;
3563 
3564  // Extract the actual blocks.
3565  BasicBlock *TrueBB = TBA->getBasicBlock();
3566  BasicBlock *FalseBB = FBA->getBasicBlock();
3567 
3568  // Perform the actual simplification.
3569  return SimplifyTerminatorOnSelect(IBI, SI->getCondition(), TrueBB, FalseBB, 0,
3570  0);
3571 }
3572 
3573 /// This is called when we find an icmp instruction
3574 /// (a seteq/setne with a constant) as the only instruction in a
3575 /// block that ends with an uncond branch. We are looking for a very specific
3576 /// pattern that occurs when "A == 1 || A == 2 || A == 3" gets simplified. In
3577 /// this case, we merge the first two "or's of icmp" into a switch, but then the
3578 /// default value goes to an uncond block with a seteq in it, we get something
3579 /// like:
3580 ///
3581 /// switch i8 %A, label %DEFAULT [ i8 1, label %end i8 2, label %end ]
3582 /// DEFAULT:
3583 /// %tmp = icmp eq i8 %A, 92
3584 /// br label %end
3585 /// end:
3586 /// ... = phi i1 [ true, %entry ], [ %tmp, %DEFAULT ], [ true, %entry ]
3587 ///
3588 /// We prefer to split the edge to 'end' so that there is a true/false entry to
3589 /// the PHI, merging the third icmp into the switch.
3590 bool SimplifyCFGOpt::tryToSimplifyUncondBranchWithICmpInIt(
3591  ICmpInst *ICI, IRBuilder<> &Builder) {
3592  BasicBlock *BB = ICI->getParent();
3593 
3594  // If the block has any PHIs in it or the icmp has multiple uses, it is too
3595  // complex.
3596  if (isa<PHINode>(BB->begin()) || !ICI->hasOneUse())
3597  return false;
3598 
3599  Value *V = ICI->getOperand(0);
3600  ConstantInt *Cst = cast<ConstantInt>(ICI->getOperand(1));
3601 
3602  // The pattern we're looking for is where our only predecessor is a switch on
3603  // 'V' and this block is the default case for the switch. In this case we can
3604  // fold the compared value into the switch to simplify things.
3605  BasicBlock *Pred = BB->getSinglePredecessor();
3606  if (!Pred || !isa<SwitchInst>(Pred->getTerminator()))
3607  return false;
3608 
3609  SwitchInst *SI = cast<SwitchInst>(Pred->getTerminator());
3610  if (SI->getCondition() != V)
3611  return false;
3612 
3613  // If BB is reachable on a non-default case, then we simply know the value of
3614  // V in this block. Substitute it and constant fold the icmp instruction
3615  // away.
3616  if (SI->getDefaultDest() != BB) {
3617  ConstantInt *VVal = SI->findCaseDest(BB);
3618  assert(VVal && "Should have a unique destination value");
3619  ICI->setOperand(0, VVal);
3620 
3621  if (Value *V = SimplifyInstruction(ICI, {DL, ICI})) {
3622  ICI->replaceAllUsesWith(V);
3623  ICI->eraseFromParent();
3624  }
3625  // BB is now empty, so it is likely to simplify away.
3626  return requestResimplify();
3627  }
3628 
3629  // Ok, the block is reachable from the default dest. If the constant we're
3630  // comparing exists in one of the other edges, then we can constant fold ICI
3631  // and zap it.
3632  if (SI->findCaseValue(Cst) != SI->case_default()) {
3633  Value *V;
3634  if (ICI->getPredicate() == ICmpInst::ICMP_EQ)
3635  V = ConstantInt::getFalse(BB->getContext());
3636  else
3637  V = ConstantInt::getTrue(BB->getContext());
3638 
3639  ICI->replaceAllUsesWith(V);
3640  ICI->eraseFromParent();
3641  // BB is now empty, so it is likely to simplify away.
3642  return requestResimplify();
3643  }
3644 
3645  // The use of the icmp has to be in the 'end' block, by the only PHI node in
3646  // the block.
3647  BasicBlock *SuccBlock = BB->getTerminator()->getSuccessor(0);
3648  PHINode *PHIUse = dyn_cast<PHINode>(ICI->user_back());
3649  if (PHIUse == nullptr || PHIUse != &SuccBlock->front() ||
3650  isa<PHINode>(++BasicBlock::iterator(PHIUse)))
3651  return false;
3652 
3653  // If the icmp is a SETEQ, then the default dest gets false, the new edge gets
3654  // true in the PHI.
3655  Constant *DefaultCst = ConstantInt::getTrue(BB->getContext());
3656  Constant *NewCst = ConstantInt::getFalse(BB->getContext());
3657 
3658  if (ICI->getPredicate() == ICmpInst::ICMP_EQ)
3659  std::swap(DefaultCst, NewCst);
3660 
3661  // Replace ICI (which is used by the PHI for the default value) with true or
3662  // false depending on if it is EQ or NE.
3663  ICI->replaceAllUsesWith(DefaultCst);
3664  ICI->eraseFromParent();
3665 
3666  // Okay, the switch goes to this block on a default value. Add an edge from
3667  // the switch to the merge point on the compared value.
3668  BasicBlock *NewBB =
3669  BasicBlock::Create(BB->getContext(), "switch.edge", BB->getParent(), BB);
3670  {
3671  SwitchInstProfUpdateWrapper SIW(*SI);
3672  auto W0 = SIW.getSuccessorWeight(0);
3674  if (W0) {
3675  NewW = ((uint64_t(*W0) + 1) >> 1);
3676  SIW.setSuccessorWeight(0, *NewW);
3677  }
3678  SIW.addCase(Cst, NewBB, NewW);
3679  }
3680 
3681  // NewBB branches to the phi block, add the uncond branch and the phi entry.
3682  Builder.SetInsertPoint(NewBB);
3683  Builder.SetCurrentDebugLocation(SI->getDebugLoc());
3684  Builder.CreateBr(SuccBlock);
3685  PHIUse->addIncoming(NewCst, NewBB);
3686  return true;
3687 }
3688 
3689 /// The specified branch is a conditional branch.
3690 /// Check to see if it is branching on an or/and chain of icmp instructions, and
3691 /// fold it into a switch instruction if so.
3693  const DataLayout &DL) {
3694  Instruction *Cond = dyn_cast<Instruction>(BI->getCondition());
3695  if (!Cond)
3696  return false;
3697 
3698  // Change br (X == 0 | X == 1), T, F into a switch instruction.
3699  // If this is a bunch of seteq's or'd together, or if it's a bunch of
3700  // 'setne's and'ed together, collect them.
3701 
3702  // Try to gather values from a chain of and/or to be turned into a switch
3703  ConstantComparesGatherer ConstantCompare(Cond, DL);
3704  // Unpack the result
3705  SmallVectorImpl<ConstantInt *> &Values = ConstantCompare.Vals;
3706  Value *CompVal = ConstantCompare.CompValue;
3707  unsigned UsedICmps = ConstantCompare.UsedICmps;
3708  Value *ExtraCase = ConstantCompare.Extra;
3709 
3710  // If we didn't have a multiply compared value, fail.
3711  if (!CompVal)
3712  return false;
3713 
3714  // Avoid turning single icmps into a switch.
3715  if (UsedICmps <= 1)
3716  return false;
3717 
3718  bool TrueWhenEqual = (Cond->getOpcode() == Instruction::Or);
3719 
3720  // There might be duplicate constants in the list, which the switch
3721  // instruction can't handle, remove them now.
3722  array_pod_sort(Values.begin(), Values.end(), ConstantIntSortPredicate);
3723  Values.erase(std::unique(Values.begin(), Values.end()), Values.end());
3724 
3725  // If Extra was used, we require at least two switch values to do the
3726  // transformation. A switch with one value is just a conditional branch.
3727  if (ExtraCase && Values.size() < 2)
3728  return false;
3729 
3730  // TODO: Preserve branch weight metadata, similarly to how
3731  // FoldValueComparisonIntoPredecessors preserves it.
3732 
3733  // Figure out which block is which destination.
3734  BasicBlock *DefaultBB = BI->getSuccessor(1);
3735  BasicBlock *EdgeBB = BI->getSuccessor(0);
3736  if (!TrueWhenEqual)
3737  std::swap(DefaultBB, EdgeBB);
3738 
3739  BasicBlock *BB = BI->getParent();
3740 
3741  // MSAN does not like undefs as branch condition which can be introduced
3742  // with "explicit branch".
3743  if (ExtraCase && BB->getParent()->hasFnAttribute(Attribute::SanitizeMemory))
3744  return false;
3745 
3746  LLVM_DEBUG(dbgs() << "Converting 'icmp' chain with " << Values.size()
3747  << " cases into SWITCH. BB is:\n"
3748  << *BB);
3749 
3750  // If there are any extra values that couldn't be folded into the switch
3751  // then we evaluate them with an explicit branch first. Split the block
3752  // right before the condbr to handle it.
3753  if (ExtraCase) {
3754  BasicBlock *NewBB =
3755  BB->splitBasicBlock(BI->getIterator(), "switch.early.test");
3756  // Remove the uncond branch added to the old block.
3757  Instruction *OldTI = BB->getTerminator();
3758  Builder.SetInsertPoint(OldTI);
3759 
3760  if (TrueWhenEqual)
3761  Builder.CreateCondBr(ExtraCase, EdgeBB, NewBB);
3762  else
3763  Builder.CreateCondBr(ExtraCase, NewBB, EdgeBB);
3764 
3765  OldTI->eraseFromParent();
3766 
3767  // If there are PHI nodes in EdgeBB, then we need to add a new entry to them
3768  // for the edge we just added.
3769  AddPredecessorToBlock(EdgeBB, BB, NewBB);
3770 
3771  LLVM_DEBUG(dbgs() << " ** 'icmp' chain unhandled condition: " << *ExtraCase
3772  << "\nEXTRABB = " << *BB);
3773  BB = NewBB;
3774  }
3775 
3776  Builder.SetInsertPoint(BI);
3777  // Convert pointer to int before we switch.
3778  if (CompVal->getType()->isPointerTy()) {
3779  CompVal = Builder.CreatePtrToInt(
3780  CompVal, DL.getIntPtrType(CompVal->getType()), "magicptr");
3781  }
3782 
3783  // Create the new switch instruction now.
3784  SwitchInst *New = Builder.CreateSwitch(CompVal, DefaultBB, Values.size());
3785 
3786  // Add all of the 'cases' to the switch instruction.
3787  for (unsigned i = 0, e = Values.size(); i != e; ++i)
3788  New->addCase(Values[i], EdgeBB);
3789 
3790  // We added edges from PI to the EdgeBB. As such, if there were any
3791  // PHI nodes in EdgeBB, they need entries to be added corresponding to
3792  // the number of edges added.
3793  for (BasicBlock::iterator BBI = EdgeBB->begin(); isa<PHINode>(BBI); ++BBI) {
3794  PHINode *PN = cast<PHINode>(BBI);
3795  Value *InVal = PN->getIncomingValueForBlock(BB);
3796  for (unsigned i = 0, e = Values.size() - 1; i != e; ++i)
3797  PN->addIncoming(InVal, BB);
3798  }
3799 
3800  // Erase the old branch instruction.
3802 
3803  LLVM_DEBUG(dbgs() << " ** 'icmp' chain result is:\n" << *BB << '\n');
3804  return true;
3805 }
3806 
3807 bool SimplifyCFGOpt::SimplifyResume(ResumeInst *RI, IRBuilder<> &Builder) {
3808  if (isa<PHINode>(RI->getValue()))
3809  return SimplifyCommonResume(RI);
3810  else if (isa<LandingPadInst>(RI->getParent()->getFirstNonPHI()) &&
3811  RI->getValue() == RI->getParent()->getFirstNonPHI())
3812  // The resume must unwind the exception that caused control to branch here.
3813  return SimplifySingleResume(RI);
3814 
3815  return false;
3816 }
3817 
3818 // Simplify resume that is shared by several landing pads (phi of landing pad).
3819 bool SimplifyCFGOpt::SimplifyCommonResume(ResumeInst *RI) {
3820  BasicBlock *BB = RI->getParent();
3821 
3822  // Check that there are no other instructions except for debug intrinsics
3823  // between the phi of landing pads (RI->getValue()) and resume instruction.
3824  BasicBlock::iterator I = cast<Instruction>(RI->getValue())->getIterator(),
3825  E = RI->getIterator();
3826  while (++I != E)
3827  if (!isa<DbgInfoIntrinsic>(I))
3828  return false;
3829 
3830  SmallSetVector<BasicBlock *, 4> TrivialUnwindBlocks;
3831  auto *PhiLPInst = cast<PHINode>(RI->getValue());
3832 
3833  // Check incoming blocks to see if any of them are trivial.
3834  for (unsigned Idx = 0, End = PhiLPInst->getNumIncomingValues(); Idx != End;
3835  Idx++) {
3836  auto *IncomingBB = PhiLPInst->getIncomingBlock(Idx);
3837  auto *IncomingValue = PhiLPInst->getIncomingValue(Idx);
3838 
3839  // If the block has other successors, we can not delete it because
3840  // it has other dependents.
3841  if (IncomingBB->getUniqueSuccessor() != BB)
3842  continue;
3843 
3844  auto *LandingPad = dyn_cast<LandingPadInst>(IncomingBB->getFirstNonPHI());
3845  // Not the landing pad that caused the control to branch here.
3846  if (IncomingValue != LandingPad)
3847  continue;
3848 
3849  bool isTrivial = true;
3850 
3851  I = IncomingBB->getFirstNonPHI()->getIterator();
3852  E = IncomingBB->getTerminator()->getIterator();
3853  while (++I != E)
3854  if (!isa<DbgInfoIntrinsic>(I)) {
3855  isTrivial = false;
3856  break;
3857  }
3858 
3859  if (isTrivial)
3860  TrivialUnwindBlocks.insert(IncomingBB);
3861  }
3862 
3863  // If no trivial unwind blocks, don't do any simplifications.
3864  if (TrivialUnwindBlocks.empty())
3865  return false;
3866 
3867  // Turn all invokes that unwind here into calls.
3868  for (auto *TrivialBB : TrivialUnwindBlocks) {
3869  // Blocks that will be simplified should be removed from the phi node.
3870  // Note there could be multiple edges to the resume block, and we need
3871  // to remove them all.
3872  while (PhiLPInst->getBasicBlockIndex(TrivialBB) != -1)
3873  BB->removePredecessor(TrivialBB, true);
3874 
3875  for (pred_iterator PI = pred_begin(TrivialBB), PE = pred_end(TrivialBB);
3876  PI != PE;) {
3877  BasicBlock *Pred = *PI++;
3878  removeUnwindEdge(Pred);
3879  }
3880 
3881  // In each SimplifyCFG run, only the current processed block can be erased.
3882  // Otherwise, it will break the iteration of SimplifyCFG pass. So instead
3883  // of erasing TrivialBB, we only remove the branch to the common resume
3884  // block so that we can later erase the resume block since it has no
3885  // predecessors.
3886  TrivialBB->getTerminator()->eraseFromParent();
3887  new UnreachableInst(RI->getContext(), TrivialBB);
3888  }
3889 
3890  // Delete the resume block if all its predecessors have been removed.
3891  if (pred_empty(BB))
3892  BB->eraseFromParent();
3893 
3894  return !TrivialUnwindBlocks.empty();
3895 }
3896 
3897 // Simplify resume that is only used by a single (non-phi) landing pad.
3898 bool SimplifyCFGOpt::SimplifySingleResume(ResumeInst *RI) {
3899  BasicBlock *BB = RI->getParent();
3900  auto *LPInst = cast<LandingPadInst>(BB->getFirstNonPHI());
3901  assert(RI->getValue() == LPInst &&
3902  "Resume must unwind the exception that caused control to here");
3903 
3904  // Check that there are no other instructions except for debug intrinsics.
3905  BasicBlock::iterator I = LPInst->getIterator(), E = RI->getIterator();
3906  while (++I != E)
3907  if (!isa<DbgInfoIntrinsic>(I))
3908  return false;
3909 
3910  // Turn all invokes that unwind here into calls and delete the basic block.
3911  for (pred_iterator PI = pred_begin(BB), PE = pred_end(BB); PI != PE;) {
3912  BasicBlock *Pred = *PI++;
3913  removeUnwindEdge(Pred);
3914  }
3915 
3916  // The landingpad is now unreachable. Zap it.
3917  if (LoopHeaders)
3918  LoopHeaders->erase(BB);
3919  BB->eraseFromParent();
3920  return true;
3921 }
3922 
3924  // If this is a trivial cleanup pad that executes no instructions, it can be
3925  // eliminated. If the cleanup pad continues to the caller, any predecessor
3926  // that is an EH pad will be updated to continue to the caller and any
3927  // predecessor that terminates with an invoke instruction will have its invoke
3928  // instruction converted to a call instruction. If the cleanup pad being
3929  // simplified does not continue to the caller, each predecessor will be
3930  // updated to continue to the unwind destination of the cleanup pad being
3931  // simplified.
3932  BasicBlock *BB = RI->getParent();
3933  CleanupPadInst *CPInst = RI->getCleanupPad();
3934  if (CPInst->getParent() != BB)
3935  // This isn't an empty cleanup.
3936  return false;
3937 
3938  // We cannot kill the pad if it has multiple uses. This typically arises
3939  // from unreachable basic blocks.
3940  if (!CPInst->hasOneUse())
3941  return false;
3942 
3943  // Check that there are no other instructions except for benign intrinsics.
3944  BasicBlock::iterator I = CPInst->getIterator(), E = RI->getIterator();
3945  while (++I != E) {
3946  auto *II = dyn_cast<IntrinsicInst>(I);
3947  if (!II)
3948  return false;
3949 
3950  Intrinsic::ID IntrinsicID = II->getIntrinsicID();
3951  switch (IntrinsicID) {
3952  case Intrinsic::dbg_declare:
3953  case Intrinsic::dbg_value:
3954  case Intrinsic::dbg_label:
3955  case Intrinsic::lifetime_end:
3956  break;
3957  default:
3958  return false;
3959  }
3960  }
3961 
3962  // If the cleanup return we are simplifying unwinds to the caller, this will
3963  // set UnwindDest to nullptr.
3964  BasicBlock *UnwindDest = RI->getUnwindDest();
3965  Instruction *DestEHPad = UnwindDest ? UnwindDest->getFirstNonPHI() : nullptr;
3966 
3967  // We're about to remove BB from the control flow. Before we do, sink any
3968  // PHINodes into the unwind destination. Doing this before changing the
3969  // control flow avoids some potentially slow checks, since we can currently
3970  // be certain that UnwindDest and BB have no common predecessors (since they
3971  // are both EH pads).
3972  if (UnwindDest) {
3973  // First, go through the PHI nodes in UnwindDest and update any nodes that
3974  // reference the block we are removing
3975  for (BasicBlock::iterator I = UnwindDest->begin(),
3976  IE = DestEHPad->getIterator();
3977  I != IE; ++I) {
3978  PHINode *DestPN = cast<PHINode>(I);
3979 
3980  int Idx = DestPN->getBasicBlockIndex(BB);
3981  // Since BB unwinds to UnwindDest, it has to be in the PHI node.
3982  assert(Idx != -1);
3983  // This PHI node has an incoming value that corresponds to a control
3984  // path through the cleanup pad we are removing. If the incoming
3985  // value is in the cleanup pad, it must be a PHINode (because we
3986  // verified above that the block is otherwise empty). Otherwise, the
3987  // value is either a constant or a value that dominates the cleanup
3988  // pad being removed.
3989  //
3990  // Because BB and UnwindDest are both EH pads, all of their
3991  // predecessors must unwind to these blocks, and since no instruction
3992  // can have multiple unwind destinations, there will be no overlap in
3993  // incoming blocks between SrcPN and DestPN.
3994  Value *SrcVal = DestPN->getIncomingValue(Idx);
3995  PHINode *SrcPN = dyn_cast<PHINode>(SrcVal);
3996 
3997  // Remove the entry for the block we are deleting.
3998  DestPN->removeIncomingValue(Idx, false);
3999 
4000  if (SrcPN && SrcPN->getParent() == BB) {
4001  // If the incoming value was a PHI node in the cleanup pad we are
4002  // removing, we need to merge that PHI node's incoming values into
4003  // DestPN.
4004  for (unsigned SrcIdx = 0, SrcE = SrcPN->getNumIncomingValues();
4005  SrcIdx != SrcE; ++SrcIdx) {
4006  DestPN->addIncoming(SrcPN->getIncomingValue(SrcIdx),
4007  SrcPN->getIncomingBlock(SrcIdx));
4008  }
4009  } else {
4010  // Otherwise, the incoming value came from above BB and
4011  // so we can just reuse it. We must associate all of BB's
4012  // predecessors with this value.
4013  for (auto *pred : predecessors(BB)) {
4014  DestPN->addIncoming(SrcVal, pred);
4015  }
4016  }
4017  }
4018 
4019  // Sink any remaining PHI nodes directly into UnwindDest.
4020  Instruction *InsertPt = DestEHPad;
4021  for (BasicBlock::iterator I = BB->begin(),
4022  IE = BB->getFirstNonPHI()->getIterator();
4023  I != IE;) {
4024  // The iterator must be incremented here because the instructions are
4025  // being moved to another block.
4026  PHINode *PN = cast<PHINode>(I++);
4027  if (PN->use_empty())
4028  // If the PHI node has no uses, just leave it. It will be erased
4029  // when we erase BB below.
4030  continue;
4031 
4032  // Otherwise, sink this PHI node into UnwindDest.
4033  // Any predecessors to UnwindDest which are not already represented
4034  // must be back edges which inherit the value from the path through
4035  // BB. In this case, the PHI value must reference itself.
4036  for (auto *pred : predecessors(UnwindDest))
4037  if (pred != BB)
4038  PN->addIncoming(PN, pred);
4039  PN->moveBefore(InsertPt);
4040  }
4041  }
4042 
4043  for (pred_iterator PI = pred_begin(BB), PE = pred_end(BB); PI != PE;) {
4044  // The iterator must be updated here because we are removing this pred.
4045  BasicBlock *PredBB = *PI++;
4046  if (UnwindDest == nullptr) {
4047  removeUnwindEdge(PredBB);
4048  } else {
4049  Instruction *TI = PredBB->getTerminator();
4050  TI->replaceUsesOfWith(BB, UnwindDest);
4051  }
4052  }
4053 
4054  // The cleanup pad is now unreachable. Zap it.
4055  BB->eraseFromParent();
4056  return true;
4057 }
4058 
4059 // Try to merge two cleanuppads together.
4061  // Skip any cleanuprets which unwind to caller, there is nothing to merge
4062  // with.
4063  BasicBlock *UnwindDest = RI->getUnwindDest();
4064  if (!UnwindDest)
4065  return false;
4066 
4067  // This cleanupret isn't the only predecessor of this cleanuppad, it wouldn't
4068  // be safe to merge without code duplication.
4069  if (UnwindDest->getSinglePredecessor() != RI->getParent())
4070  return false;
4071 
4072  // Verify that our cleanuppad's unwind destination is another cleanuppad.
4073  auto *SuccessorCleanupPad = dyn_cast<CleanupPadInst>(&UnwindDest->front());
4074  if (!SuccessorCleanupPad)
4075  return false;
4076 
4077  CleanupPadInst *PredecessorCleanupPad = RI->getCleanupPad();
4078  // Replace any uses of the successor cleanupad with the predecessor pad
4079  // The only cleanuppad uses should be this cleanupret, it's cleanupret and
4080  // funclet bundle operands.
4081  SuccessorCleanupPad->replaceAllUsesWith(PredecessorCleanupPad);
4082  // Remove the old cleanuppad.
4083  SuccessorCleanupPad->eraseFromParent();
4084  // Now, we simply replace the cleanupret with a branch to the unwind
4085  // destination.
4086  BranchInst::Create(UnwindDest, RI->getParent());
4087  RI->eraseFromParent();
4088 
4089  return true;
4090 }
4091 
4092 bool SimplifyCFGOpt::SimplifyCleanupReturn(CleanupReturnInst *RI) {
4093  // It is possible to transiantly have an undef cleanuppad operand because we
4094  // have deleted some, but not all, dead blocks.
4095  // Eventually, this block will be deleted.
4096  if (isa<UndefValue>(RI->getOperand(0)))
4097  return false;
4098 
4099  if (mergeCleanupPad(RI))
4100  return true;
4101 
4102  if (removeEmptyCleanup(RI))
4103  return true;
4104 
4105  return false;
4106 }
4107 
4108 bool SimplifyCFGOpt::SimplifyReturn(ReturnInst *RI, IRBuilder<> &Builder) {
4109  BasicBlock *BB = RI->getParent();
4110  if (!BB->getFirstNonPHIOrDbg()->isTerminator())
4111  return false;
4112 
4113  // Find predecessors that end with branches.
4114  SmallVector<BasicBlock *, 8> UncondBranchPreds;
4115  SmallVector<BranchInst *, 8> CondBranchPreds;
4116  for (pred_iterator PI = pred_begin(BB), E = pred_end(BB); PI != E; ++PI) {
4117  BasicBlock *P = *PI;
4118  Instruction *PTI = P->getTerminator();
4119  if (BranchInst *BI = dyn_cast<BranchInst>(PTI)) {
4120  if (BI->isUnconditional())
4121  UncondBranchPreds.push_back(P);
4122  else
4123  CondBranchPreds.push_back(BI);
4124  }
4125  }
4126 
4127  // If we found some, do the transformation!
4128  if (!UncondBranchPreds.empty() && DupRet) {
4129  while (!UncondBranchPreds.empty()) {
4130  BasicBlock *Pred = UncondBranchPreds.pop_back_val();
4131  LLVM_DEBUG(dbgs() << "FOLDING: " << *BB
4132  << "INTO UNCOND BRANCH PRED: " << *Pred);
4133  (void)FoldReturnIntoUncondBranch(RI, BB, Pred);
4134  }
4135 
4136  // If we eliminated all predecessors of the block, delete the block now.
4137  if (pred_empty(BB)) {
4138  // We know there are no successors, so just nuke the block.
4139  if (LoopHeaders)
4140  LoopHeaders->erase(BB);
4141  BB->eraseFromParent();
4142  }
4143 
4144  return true;
4145  }
4146 
4147  // Check out all of the conditional branches going to this return
4148  // instruction. If any of them just select between returns, change the
4149  // branch itself into a select/return pair.
4150  while (!CondBranchPreds.empty()) {
4151  BranchInst *BI = CondBranchPreds.pop_back_val();
4152 
4153  // Check to see if the non-BB successor is also a return block.
4154  if (isa<ReturnInst>(BI->getSuccessor(0)->getTerminator()) &&
4155  isa<ReturnInst>(BI->getSuccessor(1)->getTerminator()) &&
4156  SimplifyCondBranchToTwoReturns(BI, Builder))
4157  return true;
4158  }
4159  return false;
4160 }
4161 
4162 bool SimplifyCFGOpt::SimplifyUnreachable(UnreachableInst *UI) {
4163  BasicBlock *BB = UI->getParent();
4164 
4165  bool Changed = false;
4166 
4167  // If there are any instructions immediately before the unreachable that can
4168  // be removed, do so.
4169  while (UI->getIterator() != BB->begin()) {
4170  BasicBlock::iterator BBI = UI->getIterator();
4171  --BBI;
4172  // Do not delete instructions that can have side effects which might cause
4173  // the unreachable to not be reachable; specifically, calls and volatile
4174  // operations may have this effect.
4175  if (isa<CallInst>(BBI) && !isa<DbgInfoIntrinsic>(BBI))
4176  break;
4177 
4178  if (BBI->mayHaveSideEffects()) {
4179  if (auto *SI = dyn_cast<StoreInst>(BBI)) {
4180  if (SI->isVolatile())
4181  break;
4182  } else if (auto *LI = dyn_cast<LoadInst>(BBI)) {
4183  if (LI->isVolatile())
4184  break;
4185  } else if (auto *RMWI = dyn_cast<AtomicRMWInst>(BBI)) {
4186  if (RMWI->isVolatile())
4187  break;
4188  } else if (auto *CXI = dyn_cast<AtomicCmpXchgInst>(BBI)) {
4189  if (CXI->isVolatile())
4190  break;
4191  } else if (isa<CatchPadInst>(BBI)) {
4192  // A catchpad may invoke exception object constructors and such, which
4193  // in some languages can be arbitrary code, so be conservative by
4194  // default.
4195  // For CoreCLR, it just involves a type test, so can be removed.
4198  break;
4199  } else if (!isa<FenceInst>(BBI) && !isa<VAArgInst>(BBI) &&
4200  !isa<LandingPadInst>(BBI)) {
4201  break;
4202  }
4203  // Note that deleting LandingPad's here is in fact okay, although it
4204  // involves a bit of subtle reasoning. If this inst is a LandingPad,
4205  // all the predecessors of this block will be the unwind edges of Invokes,
4206  // and we can therefore guarantee this block will be erased.
4207  }
4208 
4209  // Delete this instruction (any uses are guaranteed to be dead)
4210  if (!BBI->use_empty())
4211  BBI->replaceAllUsesWith(UndefValue::get(BBI->getType()));
4212  BBI->eraseFromParent();
4213  Changed = true;
4214  }
4215 
4216  // If the unreachable instruction is the first in the block, take a gander
4217  // at all of the predecessors of this instruction, and simplify them.
4218  if (&BB->front() != UI)
4219  return Changed;
4220 
4222  for (unsigned i = 0, e = Preds.size(); i != e; ++i) {
4223  Instruction *TI = Preds[i]->getTerminator();
4224  IRBuilder<> Builder(TI);
4225  if (auto *BI = dyn_cast<BranchInst>(TI)) {
4226  if (BI->isUnconditional()) {
4227  assert(BI->getSuccessor(0) == BB && "Incorrect CFG");
4228  new UnreachableInst(TI->getContext(), TI);
4229  TI->eraseFromParent();
4230  Changed = true;
4231  } else {
4232  Value* Cond = BI->getCondition();
4233  if (BI->getSuccessor(0) == BB) {
4234  Builder.CreateAssumption(Builder.CreateNot(Cond));
4235  Builder.CreateBr(BI->getSuccessor(1));
4236  } else {
4237  assert(BI->getSuccessor(1) == BB && "Incorrect CFG");
4238  Builder.CreateAssumption(Cond);
4239  Builder.CreateBr(BI->getSuccessor(0));
4240  }
4242  Changed = true;
4243  }
4244  } else if (auto *SI = dyn_cast<SwitchInst>(TI)) {
4246  for (auto i = SU->case_begin(), e = SU->case_end(); i != e;) {
4247  if (i->getCaseSuccessor() != BB) {
4248  ++i;
4249  continue;
4250  }
4251  BB->removePredecessor(SU->getParent());
4252  i = SU.removeCase(i);
4253  e = SU->case_end();
4254  Changed = true;
4255  }
4256  } else if (auto *II = dyn_cast<InvokeInst>(TI)) {
4257  if (II->getUnwindDest() == BB) {
4258  removeUnwindEdge(TI->getParent());
4259  Changed = true;
4260  }
4261  } else if (auto *CSI = dyn_cast<CatchSwitchInst>(TI)) {
4262  if (CSI->getUnwindDest() == BB) {
4263  removeUnwindEdge(TI->getParent());
4264  Changed = true;
4265  continue;
4266  }
4267 
4268  for (CatchSwitchInst::handler_iterator I = CSI->handler_begin(),
4269  E = CSI->handler_end();
4270  I != E; ++I) {
4271  if (*I == BB) {
4272  CSI->removeHandler(I);
4273  --I;
4274  --E;
4275  Changed = true;
4276  }
4277  }
4278  if (CSI->getNumHandlers() == 0) {
4279  BasicBlock *CatchSwitchBB = CSI->getParent();
4280  if (CSI->hasUnwindDest()) {
4281  // Redirect preds to the unwind dest
4282  CatchSwitchBB->replaceAllUsesWith(CSI->getUnwindDest());
4283  } else {
4284  // Rewrite all preds to unwind to caller (or from invoke to call).
4285  SmallVector<BasicBlock *, 8> EHPreds(predecessors(CatchSwitchBB));
4286  for (BasicBlock *EHPred : EHPreds)
4287  removeUnwindEdge(EHPred);
4288  }
4289  // The catchswitch is no longer reachable.
4290  new UnreachableInst(CSI->getContext(), CSI);
4291  CSI->eraseFromParent();
4292  Changed = true;
4293  }
4294  } else if (isa<CleanupReturnInst>(TI)) {
4295  new UnreachableInst(TI->getContext(), TI);
4296  TI->eraseFromParent();
4297  Changed = true;
4298  }
4299  }
4300 
4301  // If this block is now dead, remove it.
4302  if (pred_empty(BB) && BB != &BB->getParent()->getEntryBlock()) {
4303  // We know there are no successors, so just nuke the block.
4304  if (LoopHeaders)
4305  LoopHeaders->erase(BB);
4306  BB->eraseFromParent();
4307  return true;
4308  }
4309 
4310  return Changed;
4311 }
4312 
4314  assert(Cases.size() >= 1);
4315 
4316  array_pod_sort(Cases.begin(), Cases.end(), ConstantIntSortPredicate);
4317  for (size_t I = 1, E = Cases.size(); I != E; ++I) {
4318  if (Cases[I - 1]->getValue() != Cases[I]->getValue() + 1)
4319  return false;
4320  }
4321  return true;
4322 }
4323 
4325  LLVM_DEBUG(dbgs() << "SimplifyCFG: switch default is dead.\n");
4326  BasicBlock *NewDefaultBlock =
4327  SplitBlockPredecessors(Switch->getDefaultDest(), Switch->getParent(), "");
4328  Switch->setDefaultDest(&*NewDefaultBlock);
4329  SplitBlock(&*NewDefaultBlock, &NewDefaultBlock->front());
4330  auto *NewTerminator = NewDefaultBlock->getTerminator();
4331  new UnreachableInst(Switch->getContext(), NewTerminator);
4332  EraseTerminatorAndDCECond(NewTerminator);
4333 }
4334 
4335 /// Turn a switch with two reachable destinations into an integer range
4336 /// comparison and branch.
4338  assert(SI->getNumCases() > 1 && "Degenerate switch?");
4339 
4340  bool HasDefault =
4341  !isa<UnreachableInst>(SI->getDefaultDest()->getFirstNonPHIOrDbg());
4342 
4343  // Partition the cases into two sets with different destinations.
4344  BasicBlock *DestA = HasDefault ? SI->getDefaultDest() : nullptr;
4345  BasicBlock *DestB = nullptr;
4348 
4349  for (auto Case : SI->cases()) {
4350  BasicBlock *Dest = Case.getCaseSuccessor();
4351  if (!DestA)
4352  DestA = Dest;
4353  if (Dest == DestA) {
4354  CasesA.push_back(Case.getCaseValue());
4355  continue;
4356  }
4357  if (!DestB)
4358  DestB = Dest;
4359  if (Dest == DestB) {
4360  CasesB.push_back(Case.getCaseValue());
4361  continue;
4362  }
4363  return false; // More than two destinations.
4364  }
4365 
4366  assert(DestA && DestB &&
4367  "Single-destination switch should have been folded.");
4368  assert(DestA != DestB);
4369  assert(DestB != SI->getDefaultDest());
4370  assert(!CasesB.empty() && "There must be non-default cases.");
4371  assert(!CasesA.empty() || HasDefault);
4372 
4373  // Figure out if one of the sets of cases form a contiguous range.
4374  SmallVectorImpl<ConstantInt *> *ContiguousCases = nullptr;
4375  BasicBlock *ContiguousDest = nullptr;
4376  BasicBlock *OtherDest = nullptr;
4377  if (!CasesA.empty() && CasesAreContiguous(CasesA)) {
4378  ContiguousCases = &CasesA;
4379  ContiguousDest = DestA;
4380  OtherDest = DestB;
4381  } else if (CasesAreContiguous(CasesB)) {
4382  ContiguousCases = &CasesB;
4383  ContiguousDest = DestB;
4384  OtherDest = DestA;
4385  } else
4386  return false;
4387 
4388  // Start building the compare and branch.
4389 
4390  Constant *Offset = ConstantExpr::getNeg(ContiguousCases->back());
4391  Constant *NumCases =
4392  ConstantInt::get(Offset->getType(), ContiguousCases->size());
4393 
4394  Value *Sub = SI->getCondition();
4395  if (!Offset->isNullValue())
4396  Sub = Builder.CreateAdd(Sub, Offset, Sub->getName() + ".off");
4397 
4398  Value *Cmp;
4399  // If NumCases overflowed, then all possible values jump to the successor.
4400  if (NumCases->isNullValue() && !ContiguousCases->empty())
4401  Cmp = ConstantInt::getTrue(SI->getContext());
4402  else
4403  Cmp = Builder.CreateICmpULT(Sub, NumCases, "switch");
4404  BranchInst *NewBI = Builder.CreateCondBr(Cmp, ContiguousDest, OtherDest);
4405 
4406  // Update weight for the newly-created conditional branch.
4407  if (HasBranchWeights(SI)) {
4408  SmallVector<uint64_t, 8> Weights;
4409  GetBranchWeights(SI, Weights);
4410  if (Weights.size() == 1 + SI->getNumCases()) {
4411  uint64_t TrueWeight = 0;
4412  uint64_t FalseWeight = 0;
4413  for (size_t I = 0, E = Weights.size(); I != E; ++I) {
4414  if (SI->getSuccessor(I) == ContiguousDest)
4415  TrueWeight += Weights[I];
4416  else
4417  FalseWeight += Weights[I];
4418  }
4419  while (TrueWeight > UINT32_MAX || FalseWeight > UINT32_MAX) {
4420  TrueWeight /= 2;
4421  FalseWeight /= 2;
4422  }
4423  setBranchWeights(NewBI, TrueWeight, FalseWeight);
4424  }
4425  }
4426 
4427  // Prune obsolete incoming values off the successors' PHI nodes.
4428  for (auto BBI = ContiguousDest->begin(); isa<PHINode>(BBI); ++BBI) {
4429  unsigned PreviousEdges = ContiguousCases->size();
4430  if (ContiguousDest == SI->getDefaultDest())
4431  ++PreviousEdges;
4432  for (unsigned I = 0, E = PreviousEdges - 1; I != E; ++I)
4433  cast<PHINode>(BBI)->removeIncomingValue(SI->getParent());
4434  }
4435  for (auto BBI = OtherDest->begin(); isa<PHINode>(BBI); ++BBI) {
4436  unsigned PreviousEdges = SI->getNumCases() - ContiguousCases->size();
4437  if (OtherDest == SI->getDefaultDest())
4438  ++PreviousEdges;
4439  for (unsigned I = 0, E = PreviousEdges - 1; I != E; ++I)
4440  cast<PHINode>(BBI)->removeIncomingValue(SI->getParent());
4441  }
4442 
4443  // Clean up the default block - it may have phis or other instructions before
4444  // the unreachable terminator.
4445  if (!HasDefault)
4447 
4448  // Drop the switch.
4449  SI->eraseFromParent();
4450 
4451  return true;
4452 }
4453 
4454 /// Compute masked bits for the condition of a switch
4455 /// and use it to remove dead cases.
4457  const DataLayout &DL) {
4458  Value *Cond = SI->getCondition();
4459  unsigned Bits = Cond->getType()->getIntegerBitWidth();
4460  KnownBits Known = computeKnownBits(Cond, DL, 0, AC, SI);
4461 
4462  // We can also eliminate cases by determining that their values are outside of
4463  // the limited range of the condition based on how many significant (non-sign)
4464  // bits are in the condition value.
4465  unsigned ExtraSignBits = ComputeNumSignBits(Cond, DL, 0, AC, SI) - 1;
4466  unsigned MaxSignificantBitsInCond = Bits - ExtraSignBits;
4467 
4468  // Gather dead cases.
4470  for (auto &Case : SI->cases()) {
4471  const APInt &CaseVal = Case.getCaseValue()->getValue();
4472  if (Known.Zero.intersects(CaseVal) || !Known.One.isSubsetOf(CaseVal) ||
4473  (CaseVal.getMinSignedBits() > MaxSignificantBitsInCond)) {
4474  DeadCases.push_back(Case.getCaseValue());
4475  LLVM_DEBUG(dbgs() << "SimplifyCFG: switch case " << CaseVal
4476  << " is dead.\n");
4477  }
4478  }
4479 
4480  // If we can prove that the cases must cover all possible values, the
4481  // default destination becomes dead and we can remove it. If we know some
4482  // of the bits in the value, we can use that to more precisely compute the
4483  // number of possible unique case values.
4484  bool HasDefault =
4485  !isa<UnreachableInst>(SI->getDefaultDest()->getFirstNonPHIOrDbg());
4486  const unsigned NumUnknownBits =
4487  Bits - (Known.Zero | Known.One).countPopulation();
4488  assert(NumUnknownBits <= Bits);
4489  if (HasDefault && DeadCases.empty() &&
4490  NumUnknownBits < 64 /* avoid overflow */ &&
4491  SI->getNumCases() == (1ULL << NumUnknownBits)) {
4493  return true;
4494  }
4495 
4496  if (DeadCases.empty())
4497  return false;
4498 
4499  SwitchInstProfUpdateWrapper SIW(*SI);
4500  for (ConstantInt *DeadCase : DeadCases) {
4501  SwitchInst::CaseIt CaseI = SI->findCaseValue(DeadCase);
4502  assert(CaseI != SI->case_default() &&
4503  "Case was not found. Probably mistake in DeadCases forming.");
4504  // Prune unused values from PHI nodes.
4505  CaseI->getCaseSuccessor()->removePredecessor(SI->getParent());
4506  SIW.removeCase(CaseI);
4507  }
4508 
4509  return true;
4510 }
4511 
4512 /// If BB would be eligible for simplification by
4513 /// TryToSimplifyUncondBranchFromEmptyBlock (i.e. it is empty and terminated
4514 /// by an unconditional branch), look at the phi node for BB in the successor
4515 /// block and see if the incoming value is equal to CaseValue. If so, return
4516 /// the phi node, and set PhiIndex to BB's index in the phi node.
4518  BasicBlock *BB, int *PhiIndex) {
4519  if (BB->getFirstNonPHIOrDbg() != BB->getTerminator())
4520  return nullptr; // BB must be empty to be a candidate for simplification.
4521  if (!BB->getSinglePredecessor())
4522  return nullptr; // BB must be dominated by the switch.
4523 
4525  if (!Branch || !Branch->isUnconditional())
4526  return nullptr; // Terminator must be unconditional branch.
4527 
4528  BasicBlock *Succ = Branch->getSuccessor(0);
4529 
4530  for (PHINode &PHI : Succ->phis()) {
4531  int Idx = PHI.getBasicBlockIndex(BB);
4532  assert(Idx >= 0 && "PHI has no entry for predecessor?");
4533 
4534  Value *InValue = PHI.getIncomingValue(Idx);
4535  if (InValue != CaseValue)
4536  continue;
4537 
4538  *PhiIndex = Idx;
4539  return &PHI;
4540  }
4541 
4542  return nullptr;
4543 }
4544 
4545 /// Try to forward the condition of a switch instruction to a phi node
4546 /// dominated by the switch, if that would mean that some of the destination
4547 /// blocks of the switch can be folded away. Return true if a change is made.
4549  using ForwardingNodesMap = DenseMap<PHINode *, SmallVector<int, 4>>;
4550 
4551  ForwardingNodesMap ForwardingNodes;
4552  BasicBlock *SwitchBlock = SI->getParent();
4553  bool Changed = false;
4554  for (auto &Case : SI->cases()) {
4555  ConstantInt *CaseValue = Case.getCaseValue();
4556  BasicBlock *CaseDest = Case.getCaseSuccessor();
4557 
4558  // Replace phi operands in successor blocks that are using the constant case
4559  // value rather than the switch condition variable:
4560  // switchbb:
4561  // switch i32 %x, label %default [
4562  // i32 17, label %succ
4563  // ...
4564  // succ:
4565  // %r = phi i32 ... [ 17, %switchbb ] ...
4566  // -->
4567  // %r = phi i32 ... [ %x, %switchbb ] ...
4568 
4569  for (PHINode &Phi : CaseDest->phis()) {
4570  // This only works if there is exactly 1 incoming edge from the switch to
4571  // a phi. If there is >1, that means multiple cases of the switch map to 1
4572  // value in the phi, and that phi value is not the switch condition. Thus,
4573  // this transform would not make sense (the phi would be invalid because
4574  // a phi can't have different incoming values from the same block).
4575  int SwitchBBIdx = Phi.getBasicBlockIndex(SwitchBlock);
4576  if (Phi.getIncomingValue(SwitchBBIdx) == CaseValue &&
4577  count(Phi.blocks(), SwitchBlock) == 1) {
4578  Phi.setIncomingValue(SwitchBBIdx, SI->getCondition());
4579  Changed = true;
4580  }
4581  }
4582 
4583  // Collect phi nodes that are indirectly using this switch's case constants.
4584  int PhiIdx;
4585  if (auto *Phi = FindPHIForConditionForwarding(CaseValue, CaseDest, &PhiIdx))
4586  ForwardingNodes[Phi].push_back(PhiIdx);
4587  }
4588 
4589  for (auto &ForwardingNode : ForwardingNodes) {
4590  PHINode *Phi = ForwardingNode.first;
4591  SmallVectorImpl<int> &Indexes = ForwardingNode.second;
4592  if (Indexes.size() < 2)
4593  continue;
4594 
4595  for (int Index : Indexes)
4596  Phi->setIncomingValue(Index, SI->getCondition());
4597  Changed = true;
4598  }
4599 
4600  return Changed;
4601 }
4602 
4603 /// Return true if the backend will be able to handle
4604 /// initializing an array of constants like C.
4606  if (C->isThreadDependent())
4607  return false;
4608  if (C->isDLLImportDependent())
4609  return false;
4610 
4611  if (!isa<ConstantFP>(C) && !isa<ConstantInt>(C) &&
4612  !isa<ConstantPointerNull>(C) && !isa<GlobalValue>(C) &&
4613  !isa<UndefValue>(C) && !isa<ConstantExpr>(C))
4614  return false;
4615 
4616  if (ConstantExpr *CE = dyn_cast<ConstantExpr>(C)) {
4617  if (!CE->isGEPWithNoNotionalOverIndexing())
4618  return false;
4619  if (!ValidLookupTableConstant(CE->getOperand(0), TTI))
4620  return false;
4621  }
4622 
4624  return false;
4625 
4626  return true;
4627 }
4628 
4629 /// If V is a Constant, return it. Otherwise, try to look up
4630 /// its constant value in ConstantPool, returning 0 if it's not there.
4631 static Constant *
4634  if (Constant *C = dyn_cast<Constant>(V))
4635  return C;
4636  return ConstantPool.lookup(V);
4637 }
4638 
4639 /// Try to fold instruction I into a constant. This works for
4640 /// simple instructions such as binary operations where both operands are
4641 /// constant or can be replaced by constants from the ConstantPool. Returns the
4642 /// resulting constant on success, 0 otherwise.
4643 static Constant *
4646  if (SelectInst *Select = dyn_cast<SelectInst>(I)) {
4647  Constant *A = LookupConstant(Select->getCondition(), ConstantPool);
4648  if (!A)
4649  return nullptr;
4650  if (A->isAllOnesValue())
4651  return LookupConstant(Select->getTrueValue(), ConstantPool);
4652  if (A->isNullValue())
4653  return LookupConstant(Select->getFalseValue(), ConstantPool);
4654  return nullptr;
4655  }
4656 
4658  for (unsigned N = 0, E = I->getNumOperands(); N != E; ++N) {
4660  COps.push_back(A);
4661  else
4662  return nullptr;
4663  }
4664 
4665  if (CmpInst *Cmp = dyn_cast<CmpInst>(I)) {
4666  return ConstantFoldCompareInstOperands(Cmp->getPredicate(), COps[0],
4667  COps[1], DL);
4668  }
4669 
4670  return ConstantFoldInstOperands(I, COps, DL);
4671 }
4672 
4673 /// Try to determine the resulting constant values in phi nodes
4674 /// at the common destination basic block, *CommonDest, for one of the case
4675 /// destionations CaseDest corresponding to value CaseVal (0 for the default
4676 /// case), of a switch instruction SI.
4677 static bool
4679  BasicBlock **CommonDest,
4680  SmallVectorImpl<std::pair<PHINode *, Constant *>> &Res,
4681  const DataLayout &DL, const TargetTransformInfo &TTI) {
4682  // The block from which we enter the common destination.
4683  BasicBlock *Pred = SI->getParent();
4684 
4685  // If CaseDest is empty except for some side-effect free instructions through
4686  // which we can constant-propagate the CaseVal, continue to its successor.
4688  ConstantPool.insert(std::make_pair(SI->getCondition(), CaseVal));
4689  for (Instruction &I :CaseDest->instructionsWithoutDebug()) {
4690  if (I.isTerminator()) {
4691  // If the terminator is a simple branch, continue to the next block.
4692  if (I.getNumSuccessors() != 1 || I.isExceptionalTerminator())
4693  return false;
4694  Pred = CaseDest;
4695  CaseDest = I.getSuccessor(0);
4696  } else if (Constant *C = ConstantFold(&I, DL, ConstantPool)) {
4697  // Instruction is side-effect free and constant.
4698 
4699  // If the instruction has uses outside this block or a phi node slot for
4700  // the block, it is not safe to bypass the instruction since it would then
4701  // no longer dominate all its uses.
4702  for (auto &Use : I.uses()) {
4703  User *User = Use.getUser();
4704  if (Instruction *I = dyn_cast<Instruction>(User))
4705  if (I->getParent() == CaseDest)
4706  continue;
4707  if (PHINode *Phi = dyn_cast<PHINode>(User))
4708  if (Phi->getIncomingBlock(Use) == CaseDest)
4709  continue;
4710  return false;
4711  }
4712 
4713  ConstantPool.insert(std::make_pair(&I, C));
4714  } else {
4715  break;
4716  }
4717  }
4718 
4719  // If we did not have a CommonDest before, use the current one.
4720  if (!*CommonDest)
4721  *CommonDest = CaseDest;
4722  // If the destination isn't the common one, abort.
4723  if (CaseDest != *CommonDest)
4724  return false;
4725 
4726  // Get the values for this case from phi nodes in the destination block.
4727  for (PHINode &PHI : (*CommonDest)->phis()) {
4728  int Idx = PHI.getBasicBlockIndex(Pred);
4729  if (Idx == -1)
4730  continue;
4731 
4732  Constant *ConstVal =
4733  LookupConstant(PHI.getIncomingValue(Idx), ConstantPool);
4734  if (!ConstVal)
4735  return false;
4736 
4737  // Be conservative about which kinds of constants we support.
4738  if (!ValidLookupTableConstant(ConstVal, TTI))
4739  return false;
4740 
4741  Res.push_back(std::make_pair(&PHI, ConstVal));
4742  }
4743 
4744  return Res.size() > 0;
4745 }
4746 
4747 // Helper function used to add CaseVal to the list of cases that generate
4748 // Result. Returns the updated number of cases that generate this result.
4749 static uintptr_t MapCaseToResult(ConstantInt *CaseVal,
4750  SwitchCaseResultVectorTy &UniqueResults,
4751  Constant *Result) {
4752  for (auto &I : UniqueResults) {
4753  if (I.first == Result) {
4754  I.second.push_back(CaseVal);
4755  return I.second.size();
4756  }
4757  }
4758  UniqueResults.push_back(
4759  std::make_pair(Result, SmallVector<ConstantInt *, 4>(1, CaseVal)));
4760  return 1;
4761 }
4762 
4763 // Helper function that initializes a map containing
4764 // results for the PHI node of the common destination block for a switch
4765 // instruction. Returns false if multiple PHI nodes have been found or if
4766 // there is not a common destination block for the switch.
4767 static bool
4769  SwitchCaseResultVectorTy &UniqueResults,
4770  Constant *&DefaultResult, const DataLayout &DL,
4771  const TargetTransformInfo &TTI,
4772  uintptr_t MaxUniqueResults, uintptr_t MaxCasesPerResult) {
4773  for (auto &I : SI->cases()) {
4774  ConstantInt *CaseVal = I.getCaseValue();
4775 
4776  // Resulting value at phi nodes for this case value.
4777  SwitchCaseResultsTy Results;
4778  if (!GetCaseResults(SI, CaseVal, I.getCaseSuccessor(), &CommonDest, Results,
4779  DL, TTI))
4780  return false;
4781 
4782  // Only one value per case is permitted.
4783  if (Results.size() > 1)
4784  return false;
4785 
4786  // Add the case->result mapping to UniqueResults.
4787  const uintptr_t NumCasesForResult =
4788  MapCaseToResult(CaseVal, UniqueResults, Results.begin()->second);
4789 
4790  // Early out if there are too many cases for this result.
4791  if (NumCasesForResult > MaxCasesPerResult)
4792  return false;
4793 
4794  // Early out if there are too many unique results.
4795  if (UniqueResults.size() > MaxUniqueResults)
4796  return false;
4797 
4798  // Check the PHI consistency.
4799  if (!PHI)
4800  PHI = Results[0].first;
4801  else if (PHI != Results[0].first)
4802  return false;
4803  }
4804  // Find the default result value.
4806  BasicBlock *DefaultDest = SI->getDefaultDest();
4807  GetCaseResults(SI, nullptr, SI->getDefaultDest(), &CommonDest, DefaultResults,
4808  DL, TTI);
4809  // If the default value is not found abort unless the default destination
4810  // is unreachable.
4811  DefaultResult =
4812  DefaultResults.size() == 1 ? DefaultResults.begin()->second : nullptr;
4813  if ((!DefaultResult &&
4814  !isa<UnreachableInst>(DefaultDest->getFirstNonPHIOrDbg())))
4815  return false;
4816 
4817  return true;
4818 }
4819 
4820 // Helper function that checks if it is possible to transform a switch with only
4821 // two cases (or two cases + default) that produces a result into a select.
4822 // Example:
4823 // switch (a) {
4824 // case 10: %0 = icmp eq i32 %a, 10
4825 // return 10; %1 = select i1 %0, i32 10, i32 4
4826 // case 20: ----> %2 = icmp eq i32 %a, 20
4827 // return 2; %3 = select i1 %2, i32 2, i32 %1
4828 // default:
4829 // return 4;
4830 // }
4831 static Value *ConvertTwoCaseSwitch(const SwitchCaseResultVectorTy &ResultVector,
4832  Constant *DefaultResult, Value *Condition,
4833  IRBuilder<> &Builder) {
4834  assert(ResultVector.size() == 2 &&
4835  "We should have exactly two unique results at this point");
4836  // If we are selecting between only two cases transform into a simple
4837  // select or a two-way select if default is possible.
4838  if (ResultVector[0].second.size() == 1 &&
4839  ResultVector[1].second.size() == 1) {
4840  ConstantInt *const FirstCase = ResultVector[0].second[0];
4841  ConstantInt *const SecondCase = ResultVector[1].second[0];
4842 
4843  bool DefaultCanTrigger = DefaultResult;
4844  Value *SelectValue = ResultVector[1].first;
4845  if (DefaultCanTrigger) {
4846  Value *const ValueCompare =
4847  Builder.CreateICmpEQ(Condition, SecondCase, "switch.selectcmp");
4848  SelectValue = Builder.CreateSelect(ValueCompare, ResultVector[1].first,
4849  DefaultResult, "switch.select");
4850  }
4851  Value *const ValueCompare =
4852  Builder.CreateICmpEQ(Condition, FirstCase, "switch.selectcmp");
4853  return Builder.CreateSelect(ValueCompare, ResultVector[0].first,
4854  SelectValue, "switch.select");
4855  }
4856 
4857  return nullptr;
4858 }
4859 
4860 // Helper function to cleanup a switch instruction that has been converted into
4861 // a select, fixing up PHI nodes and basic blocks.
4863  Value *SelectValue,
4864  IRBuilder<> &Builder) {
4865  BasicBlock *SelectBB = SI->getParent();
4866  while (PHI->getBasicBlockIndex(SelectBB) >= 0)
4867  PHI->removeIncomingValue(SelectBB);
4868  PHI->addIncoming(SelectValue, SelectBB);
4869 
4870  Builder.CreateBr(PHI->getParent());
4871 
4872  // Remove the switch.
4873  for (unsigned i = 0, e = SI->getNumSuccessors(); i < e; ++i) {
4874  BasicBlock *Succ = SI->getSuccessor(i);
4875 
4876  if (Succ == PHI->getParent())
4877  continue;
4878  Succ->removePredecessor(SelectBB);
4879  }
4880  SI->eraseFromParent();
4881 }
4882 
4883 /// If the switch is only used to initialize one or more
4884 /// phi nodes in a common successor block with only two different
4885 /// constant values, replace the switch with select.
4886 static bool switchToSelect(SwitchInst *SI, IRBuilder<> &Builder,
4887  const DataLayout &DL,
4888  const TargetTransformInfo &TTI) {
4889  Value *const Cond = SI->getCondition();
4890  PHINode *PHI = nullptr;
4891  BasicBlock *CommonDest = nullptr;
4892  Constant *DefaultResult;
4893  SwitchCaseResultVectorTy UniqueResults;
4894  // Collect all the cases that will deliver the same value from the switch.
4895  if (!InitializeUniqueCases(SI, PHI, CommonDest, UniqueResults, DefaultResult,
4896  DL, TTI, 2, 1))
4897  return false;
4898  // Selects choose between maximum two values.
4899  if (UniqueResults.size() != 2)
4900  return false;
4901  assert(PHI != nullptr && "PHI for value select not found");
4902 
4903  Builder.SetInsertPoint(SI);
4904  Value *SelectValue =
4905  ConvertTwoCaseSwitch(UniqueResults, DefaultResult, Cond, Builder);
4906  if (SelectValue) {
4907  RemoveSwitchAfterSelectConversion(SI, PHI, SelectValue, Builder);
4908  return true;
4909  }
4910  // The switch couldn't be converted into a select.
4911  return false;
4912 }
4913 
4914 namespace {
4915 
4916 /// This class represents a lookup table that can be used to replace a switch.
4917 class SwitchLookupTable {
4918 public:
4919  /// Create a lookup table to use as a switch replacement with the contents
4920  /// of Values, using DefaultValue to fill any holes in the table.
4921  SwitchLookupTable(
4922  Module &M, uint64_t TableSize, ConstantInt *Offset,
4923  const SmallVectorImpl<std::pair<ConstantInt *, Constant *>> &Values,
4924  Constant *DefaultValue, const DataLayout &DL, const StringRef &FuncName);
4925 
4926  /// Build instructions with Builder to retrieve the value at
4927  /// the position given by Index in the lookup table.
4928  Value *BuildLookup(Value *Index, IRBuilder<> &Builder);
4929 
4930  /// Return true if a table with TableSize elements of
4931  /// type ElementType would fit in a target-legal register.
4932  static bool WouldFitInRegister(const DataLayout &DL, uint64_t TableSize,
4933  Type *ElementType);
4934 
4935 private:
4936  // Depending on the contents of the table, it can be represented in
4937  // different ways.
4938  enum {
4939  // For tables where each element contains the same value, we just have to
4940  // store that single value and return it for each lookup.
4941  SingleValueKind,
4942 
4943  // For tables where there is a linear relationship between table index
4944  // and values. We calculate the result with a simple multiplication
4945  // and addition instead of a table lookup.
4946  LinearMapKind,
4947 
4948  // For small tables with integer elements, we can pack them into a bitmap
4949  // that fits into a target-legal register. Values are retrieved by
4950  // shift and mask operations.
4951  BitMapKind,
4952 
4953  // The table is stored as an array of values. Values are retrieved by load
4954  // instructions from the table.
4955  ArrayKind
4956  } Kind;
4957 
4958  // For SingleValueKind, this is the single value.
4959  Constant *SingleValue = nullptr;
4960 
4961  // For BitMapKind, this is the bitmap.
4962  ConstantInt *BitMap = nullptr;
4963  IntegerType *BitMapElementTy = nullptr;
4964 
4965  // For LinearMapKind, these are the constants used to derive the value.
4966  ConstantInt *LinearOffset = nullptr;
4967  ConstantInt *LinearMultiplier = nullptr;
4968 
4969  // For ArrayKind, this is the array.
4970  GlobalVariable *Array = nullptr;
4971 };
4972 
4973 } // end anonymous namespace
4974 
4975 SwitchLookupTable::SwitchLookupTable(
4976  Module &M, uint64_t TableSize, ConstantInt *Offset,
4977  const SmallVectorImpl<std::pair<ConstantInt *, Constant *>> &Values,
4978  Constant *DefaultValue, const DataLayout &DL, const StringRef &FuncName) {
4979  assert(Values.size() && "Can't build lookup table without values!");
4980  assert(TableSize >= Values.size() && "Can't fit values in table!");
4981 
4982  // If all values in the table are equal, this is that value.
4983  SingleValue = Values.begin()->second;
4984 
4985  Type *ValueType = Values.begin()->second->getType();
4986 
4987  // Build up the table contents.
4988  SmallVector<Constant *, 64> TableContents(TableSize);
4989  for (size_t I = 0, E = Values.size(); I != E; ++I) {
4990  ConstantInt *CaseVal = Values[I].first;
4991  Constant *CaseRes = Values[I].second;
4992  assert(CaseRes->getType() == ValueType);
4993 
4994  uint64_t Idx = (CaseVal->getValue() - Offset->getValue()).getLimitedValue();
4995  TableContents[Idx] = CaseRes;
4996 
4997  if (CaseRes != SingleValue)
4998  SingleValue = nullptr;
4999  }
5000 
5001  // Fill in any holes in the table with the default result.
5002  if (Values.size() < TableSize) {
5003  assert(DefaultValue &&
5004  "Need a default value to fill the lookup table holes.");
5005  assert(DefaultValue->getType() == ValueType);
5006  for (uint64_t I = 0; I < TableSize; ++I) {
5007  if (!TableContents[I])
5008  TableContents[I] = DefaultValue;
5009  }
5010 
5011  if (DefaultValue != SingleValue)
5012  SingleValue = nullptr;
5013  }
5014 
5015  // If each element in the table contains the same value, we only need to store
5016  // that single value.
5017  if (SingleValue) {
5018  Kind = SingleValueKind;
5019  return;
5020  }
5021 
5022  // Check if we can derive the value with a linear transformation from the
5023  // table index.
5024  if (isa<IntegerType>(ValueType)) {
5025  bool LinearMappingPossible = true;
5026  APInt PrevVal;
5027  APInt DistToPrev;
5028  assert(TableSize >= 2 && "Should be a SingleValue table.");
5029  // Check if there is the same distance between two consecutive values.
5030  for (uint64_t I = 0; I < TableSize; ++I) {
5031  ConstantInt *ConstVal = dyn_cast<ConstantInt>(TableContents[I]);
5032  if (!ConstVal) {
5033  // This is an undef. We could deal with it, but undefs in lookup tables
5034  // are very seldom. It's probably not worth the additional complexity.
5035  LinearMappingPossible = false;
5036  break;
5037  }
5038  const APInt &Val = ConstVal->getValue();
5039  if (I != 0) {
5040  APInt Dist = Val - PrevVal;
5041  if (I == 1) {
5042  DistToPrev = Dist;
5043  } else if (Dist != DistToPrev) {
5044  LinearMappingPossible = false;
5045  break;
5046  }
5047  }
5048  PrevVal = Val;
5049  }
5050  if (LinearMappingPossible) {
5051  LinearOffset = cast<ConstantInt>(TableContents[0]);
5052  LinearMultiplier = ConstantInt::get(M.getContext(), DistToPrev);
5053  Kind = LinearMapKind;
5054  ++NumLinearMaps;
5055  return;
5056  }
5057  }
5058 
5059  // If the type is integer and the table fits in a register, build a bitmap.
5060  if (WouldFitInRegister(DL, TableSize, ValueType)) {
5061  IntegerType *IT = cast<IntegerType>(ValueType);
5062  APInt TableInt(TableSize * IT->getBitWidth(), 0);
5063  for (uint64_t I = TableSize; I > 0; --I) {
5064  TableInt <<= IT->getBitWidth();
5065  // Insert values into the bitmap. Undef values are set to zero.
5066  if (!isa<UndefValue>(TableContents[I - 1])) {
5067  ConstantInt *Val = cast<ConstantInt>(TableContents[I - 1]);
5068  TableInt |= Val->getValue().zext(TableInt.getBitWidth());
5069  }
5070  }
5071  BitMap = ConstantInt::get(M.getContext(), TableInt);
5072  BitMapElementTy = IT;
5073  Kind = BitMapKind;
5074  ++NumBitMaps;
5075  return;
5076  }
5077 
5078  // Store the table in an array.
5079  ArrayType *ArrayTy = ArrayType::get(ValueType, TableSize);
5080  Constant *Initializer = ConstantArray::get(ArrayTy, TableContents);
5081 
5082  Array = new GlobalVariable(M, ArrayTy, /*isConstant=*/true,
5083  GlobalVariable::PrivateLinkage, Initializer,
5084  "switch.table." + FuncName);
5085  Array->setUnnamedAddr(GlobalValue::UnnamedAddr::Global);
5086  // Set the alignment to that of an array items. We will be only loading one
5087  // value out of it.
5088  Array->setAlignment(Align(DL.getPrefTypeAlignment(ValueType)));
5089  Kind = ArrayKind;
5090 }
5091 
5092 Value *SwitchLookupTable::BuildLookup(Value *Index, IRBuilder<> &Builder) {
5093  switch (Kind) {
5094  case SingleValueKind:
5095  return SingleValue;
5096  case LinearMapKind: {
5097  // Derive the result value from the input value.
5098  Value *Result = Builder.CreateIntCast(Index, LinearMultiplier->getType(),
5099  false, "switch.idx.cast");
5100  if (!LinearMultiplier->isOne())
5101  Result = Builder.CreateMul(Result, LinearMultiplier, "switch.idx.mult");
5102  if (!LinearOffset->isZero())
5103  Result = Builder.CreateAdd(Result, LinearOffset, "switch.offset");
5104  return Result;
5105  }
5106  case BitMapKind: {
5107  // Type of the bitmap (e.g. i59).
5108  IntegerType *MapTy = BitMap->getType();
5109 
5110  // Cast Index to the same type as the bitmap.
5111  // Note: The Index is <= the number of elements in the table, so
5112  // truncating it to the width of the bitmask is safe.
5113  Value *ShiftAmt = Builder.CreateZExtOrTrunc(Index, MapTy, "switch.cast");
5114 
5115  // Multiply the shift amount by the element width.
5116  ShiftAmt = Builder.CreateMul(
5117  ShiftAmt, ConstantInt::get(MapTy, BitMapElementTy->getBitWidth()),
5118  "switch.shiftamt");
5119 
5120  // Shift down.
5121  Value *DownShifted =
5122  Builder.CreateLShr(BitMap, ShiftAmt, "switch.downshift");
5123  // Mask off.
5124  return Builder.CreateTrunc(DownShifted, BitMapElementTy, "switch.masked");
5125  }
5126  case ArrayKind: {
5127  // Make sure the table index will not overflow when treated as signed.
5128  IntegerType *IT = cast<IntegerType>(Index->getType());
5129  uint64_t TableSize =
5130  Array->getInitializer()->getType()->getArrayNumElements();
5131  if (TableSize > (1ULL << (IT->getBitWidth() - 1)))
5132  Index = Builder.CreateZExt(
5133  Index, IntegerType::get(IT->getContext(), IT->getBitWidth() + 1),
5134  "switch.tableidx.zext");
5135 
5136  Value *GEPIndices[] = {Builder.getInt32(0), Index};
5137  Value *GEP = Builder.CreateInBoundsGEP(Array->getValueType(), Array,
5138  GEPIndices, "switch.gep");
5139  return Builder.CreateLoad(
5140  cast<ArrayType>(Array->getValueType())->getElementType(), GEP,
5141  "switch.load");
5142  }
5143  }
5144  llvm_unreachable("Unknown lookup table kind!");
5145 }
5146 
5147 bool SwitchLookupTable::WouldFitInRegister(const DataLayout &DL,
5148  uint64_t TableSize,
5149  Type *ElementType) {
5150  auto *IT = dyn_cast<IntegerType>(ElementType);
5151  if (!IT)
5152  return false;
5153  // FIXME: If the type is wider than it needs to be, e.g. i8 but all values
5154  // are <= 15, we could try to narrow the type.
5155 
5156  // Avoid overflow, fitsInLegalInteger uses unsigned int for the width.
5157  if (TableSize >= UINT_MAX / IT->getBitWidth())
5158  return false;
5159  return DL.fitsInLegalInteger(TableSize * IT->getBitWidth());
5160 }
5161 
5162 /// Determine whether a lookup table should be built for this switch, based on
5163 /// the number of cases, size of the table, and the types of the results.
5164 static bool
5165 ShouldBuildLookupTable(SwitchInst *SI, uint64_t TableSize,
5166  const TargetTransformInfo &TTI, const DataLayout &DL,
5167  const SmallDenseMap<PHINode *, Type *> &ResultTypes) {
5168  if (SI->getNumCases() > TableSize || TableSize >= UINT64_MAX / 10)
5169  return false; // TableSize overflowed, or mul below might overflow.
5170 
5171  bool AllTablesFitInRegister = true;
5172  bool HasIllegalType = false;
5173  for (const auto &I : ResultTypes) {
5174  Type *Ty = I.second;
5175 
5176  // Saturate this flag to true.
5177  HasIllegalType = HasIllegalType || !TTI.isTypeLegal(Ty);
5178 
5179  // Saturate this flag to false.
5180  AllTablesFitInRegister =
5181  AllTablesFitInRegister &&
5182  SwitchLookupTable::WouldFitInRegister(DL, TableSize, Ty);
5183 
5184  // If both flags saturate, we're done. NOTE: This *only* works with
5185  // saturating flags, and all flags have to saturate first due to the
5186  // non-deterministic behavior of iterating over a dense map.
5187  if (HasIllegalType && !AllTablesFitInRegister)
5188  break;
5189  }
5190 
5191  // If each table would fit in a register, we should build it anyway.
5192  if (AllTablesFitInRegister)
5193  return true;
5194 
5195  // Don't build a table that doesn't fit in-register if it has illegal types.
5196  if (HasIllegalType)
5197  return false;
5198 
5199  // The table density should be at least 40%. This is the same criterion as for
5200  // jump tables, see SelectionDAGBuilder::handleJTSwitchCase.
5201  // FIXME: Find the best cut-off.
5202  return SI->getNumCases() * 10 >= TableSize * 4;
5203 }
5204 
5205 /// Try to reuse the switch table index compare. Following pattern:
5206 /// \code
5207 /// if (idx < tablesize)
5208 /// r = table[idx]; // table does not contain default_value
5209 /// else
5210 /// r = default_value;
5211 /// if (r != default_value)
5212 /// ...
5213 /// \endcode
5214 /// Is optimized to:
5215 /// \code
5216 /// cond = idx < tablesize;
5217 /// if (cond)
5218 /// r = table[idx];
5219 /// else
5220 /// r = default_value;
5221 /// if (cond)
5222 /// ...
5223 /// \endcode
5224 /// Jump threading will then eliminate the second if(cond).
5225 static void reuseTableCompare(
5226  User *PhiUser, BasicBlock *PhiBlock, BranchInst *RangeCheckBranch,
5227  Constant *DefaultValue,
5228  const SmallVectorImpl<std::pair<ConstantInt *, Constant *>> &Values) {
5229  ICmpInst *CmpInst = dyn_cast<ICmpInst>(PhiUser);
5230  if (!CmpInst)
5231  return;
5232 
5233  // We require that the compare is in the same block as the phi so that jump
5234  // threading can do its work afterwards.
5235  if (CmpInst->getParent() != PhiBlock)
5236  return;
5237 
5238  Constant *CmpOp1 = dyn_cast<Constant>(CmpInst->getOperand(1));
5239  if (!CmpOp1)
5240  return;
5241 
5242  Value *RangeCmp = RangeCheckBranch->getCondition();
5243  Constant *TrueConst = ConstantInt::getTrue(RangeCmp->getType());
5244  Constant *FalseConst = ConstantInt::getFalse(RangeCmp->getType());
5245 
5246  // Check if the compare with the default value is constant true or false.
5247  Constant *DefaultConst = ConstantExpr::getICmp(CmpInst->getPredicate(),
5248  DefaultValue, CmpOp1, true);
5249  if (DefaultConst != TrueConst && DefaultConst != FalseConst)
5250  return;
5251 
5252  // Check if the compare with the case values is distinct from the default
5253  // compare result.
5254  for (auto ValuePair : Values) {
5255  Constant *CaseConst = ConstantExpr::getICmp(CmpInst->getPredicate(),
5256  ValuePair.second, CmpOp1, true);
5257  if (!CaseConst || CaseConst == DefaultConst || isa<UndefValue>(CaseConst))
5258  return;
5259  assert((CaseConst == TrueConst || CaseConst == FalseConst) &&
5260  "Expect true or false as compare result.");
5261  }
5262 
5263  // Check if the branch instruction dominates the phi node. It's a simple
5264  // dominance check, but sufficient for our needs.
5265  // Although this check is invariant in the calling loops, it's better to do it
5266  // at this late stage. Practically we do it at most once for a switch.
5267  BasicBlock *BranchBlock = RangeCheckBranch->getParent();
5268  for (auto PI = pred_begin(PhiBlock), E = pred_end(PhiBlock); PI != E; ++PI) {
5269  BasicBlock *Pred = *PI;
5270  if (Pred != BranchBlock && Pred->getUniquePredecessor() != BranchBlock)
5271  return;
5272  }
5273 
5274  if (DefaultConst == FalseConst) {
5275  // The compare yields the same result. We can replace it.
5276  CmpInst->replaceAllUsesWith(RangeCmp);
5277  ++NumTableCmpReuses;
5278  } else {
5279  // The compare yields the same result, just inverted. We can replace it.
5280  Value *InvertedTableCmp = BinaryOperator::CreateXor(
5281  RangeCmp, ConstantInt::get(RangeCmp->getType(), 1), "inverted.cmp",
5282  RangeCheckBranch);
5283  CmpInst->replaceAllUsesWith(InvertedTableCmp);
5284  ++NumTableCmpReuses;
5285  }
5286 }
5287 
5288 /// If the switch is only used to initialize one or more phi nodes in a common
5289 /// successor block with different constant values, replace the switch with
5290 /// lookup tables.
5292  const DataLayout &DL,
5293  const TargetTransformInfo &TTI) {
5294  assert(SI->getNumCases() > 1 && "Degenerate switch?");
5295 
5296  Function *Fn = SI->getParent()->getParent();
5297  // Only build lookup table when we have a target that supports it or the
5298  // attribute is not set.
5299  if (!TTI.shouldBuildLookupTables() ||
5300  (Fn->getFnAttribute("no-jump-tables").getValueAsString() == "true"))
5301  return false;
5302 
5303  // FIXME: If the switch is too sparse for a lookup table, perhaps we could
5304  // split off a dense part and build a lookup table for that.
5305 
5306  // FIXME: This creates arrays of GEPs to constant strings, which means each
5307  // GEP needs a runtime relocation in PIC code. We should just build one big
5308  // string and lookup indices into that.
5309 
5310  // Ignore switches with less than three cases. Lookup tables will not make
5311  // them faster, so we don't analyze them.
5312  if (SI->getNumCases() < 3)
5313  return false;
5314 
5315  // Figure out the corresponding result for each case value and phi node in the
5316  // common destination, as well as the min and max case values.
5317  assert(!SI->cases().empty());
5318  SwitchInst::CaseIt CI = SI->case_begin();
5319  ConstantInt *MinCaseVal = CI->getCaseValue();
5320  ConstantInt *MaxCaseVal = CI->getCaseValue();
5321 
5322  BasicBlock *CommonDest = nullptr;
5323 
5324  using ResultListTy = SmallVector<std::pair<ConstantInt *, Constant *>, 4>;
5326 
5327  SmallDenseMap<PHINode *, Constant *> DefaultResults;
5330 
5331  for (SwitchInst::CaseIt E = SI->case_end(); CI != E; ++CI) {
5332  ConstantInt *CaseVal = CI->getCaseValue();
5333  if (CaseVal->getValue().slt(MinCaseVal->getValue()))
5334  MinCaseVal = CaseVal;
5335  if (CaseVal->getValue().sgt(MaxCaseVal->getValue()))
5336  MaxCaseVal = CaseVal;
5337 
5338  // Resulting value at phi nodes for this case value.
5339  using ResultsTy = SmallVector<std::pair<PHINode *, Constant *>, 4>;
5340  ResultsTy Results;
5341  if (!GetCaseResults(SI, CaseVal, CI->getCaseSuccessor(), &CommonDest,
5342  Results, DL, TTI))
5343  return false;
5344 
5345  // Append the result from this case to the list for each phi.
5346  for (const auto &I : Results) {
5347  PHINode *PHI = I.first;
5348  Constant *Value = I.second;
5349  if (!ResultLists.count(PHI))
5350  PHIs.push_back(PHI);
5351  ResultLists[PHI].push_back(std::make_pair(CaseVal, Value));
5352  }
5353  }
5354 
5355  // Keep track of the result types.
5356  for (PHINode *PHI : PHIs) {
5357  ResultTypes[PHI] = ResultLists[PHI][0].second->getType();
5358  }
5359 
5360  uint64_t NumResults = ResultLists[PHIs[0]].size();
5361  APInt RangeSpread = MaxCaseVal->getValue() - MinCaseVal->getValue();
5362  uint64_t TableSize = RangeSpread.getLimitedValue() + 1;
5363  bool TableHasHoles = (NumResults < TableSize);
5364 
5365  // If the table has holes, we need a constant result for the default case
5366  // or a bitmask that fits in a register.
5367  SmallVector<std::pair<PHINode *, Constant *>, 4> DefaultResultsList;
5368  bool HasDefaultResults =
5369  GetCaseResults(SI, nullptr, SI->getDefaultDest(), &CommonDest,
5370  DefaultResultsList, DL, TTI);
5371 
5372  bool NeedMask = (TableHasHoles && !HasDefaultResults);
5373  if (NeedMask) {
5374  // As an extra penalty for the validity test we require more cases.
5375  if (SI->getNumCases() < 4) // FIXME: Find best threshold value (benchmark).
5376  return false;
5377  if (!DL.fitsInLegalInteger(TableSize))
5378  return false;
5379  }
5380 
5381  for (const auto &I : DefaultResultsList) {
5382  PHINode *PHI = I.first;
5383  Constant *Result = I.second;
5384  DefaultResults[PHI] = Result;
5385  }
5386 
5387  if (!ShouldBuildLookupTable(SI, TableSize, TTI, DL, ResultTypes))
5388  return false;
5389 
5390  // Create the BB that does the lookups.
5391  Module &Mod = *CommonDest->getParent()->getParent();
5392  BasicBlock *LookupBB = BasicBlock::Create(
5393  Mod.getContext(), "switch.lookup", CommonDest->getParent(), CommonDest);
5394 
5395  // Compute the table index value.
5396  Builder.SetInsertPoint(SI);
5397  Value *TableIndex;
5398  if (MinCaseVal->isNullValue())
5399  TableIndex = SI->getCondition();
5400  else
5401  TableIndex = Builder.CreateSub(SI->getCondition(), MinCaseVal,
5402  "switch.tableidx");
5403 
5404  // Compute the maximum table size representable by the integer type we are
5405  // switching upon.
5406  unsigned CaseSize = MinCaseVal->getType()->getPrimitiveSizeInBits();
5407  uint64_t MaxTableSize = CaseSize > 63 ? UINT64_MAX : 1ULL << CaseSize;
5408  assert(MaxTableSize >= TableSize &&
5409  "It is impossible for a switch to have more entries than the max "
5410  "representable value of its input integer type's size.");
5411 
5412  // If the default destination is unreachable, or if the lookup table covers
5413  // all values of the conditional variable, branch directly to the lookup table
5414  // BB. Otherwise, check that the condition is within the case range.
5415  const bool DefaultIsReachable =
5416  !isa<UnreachableInst>(SI->getDefaultDest()->getFirstNonPHIOrDbg());
5417  const bool GeneratingCoveredLookupTable = (MaxTableSize == TableSize);
5418  BranchInst *RangeCheckBranch = nullptr;
5419 
5420  if (!DefaultIsReachable || GeneratingCoveredLookupTable) {
5421  Builder.CreateBr(LookupBB);
5422  // Note: We call removeProdecessor later since we need to be able to get the
5423  // PHI value for the default case in case we're using a bit mask.
5424  } else {
5425  Value *Cmp = Builder.CreateICmpULT(
5426  TableIndex, ConstantInt::get(MinCaseVal->getType(), TableSize));
5427  RangeCheckBranch =
5428  Builder.CreateCondBr(Cmp, LookupBB, SI->getDefaultDest());
5429  }
5430 
5431  // Populate the BB that does the lookups.
5432  Builder.SetInsertPoint(LookupBB);
5433 
5434  if (NeedMask) {
5435  // Before doing the lookup, we do the hole check. The LookupBB is therefore
5436  // re-purposed to do the hole check, and we create a new LookupBB.
5437  BasicBlock *MaskBB = LookupBB;
5438  MaskBB->setName("switch.hole_check");
5439  LookupBB = BasicBlock::Create(Mod.getContext(), "switch.lookup",
5440  CommonDest->getParent(), CommonDest);
5441 
5442  // Make the mask's bitwidth at least 8-bit and a power-of-2 to avoid
5443  // unnecessary illegal types.
5444  uint64_t TableSizePowOf2 = NextPowerOf2(std::max(7ULL, TableSize - 1ULL));
5445  APInt MaskInt(TableSizePowOf2, 0);
5446  APInt One(TableSizePowOf2, 1);
5447  // Build bitmask; fill in a 1 bit for every case.
5448  const ResultListTy &ResultList = ResultLists[PHIs[0]];
5449  for (size_t I = 0, E = ResultList.size(); I != E; ++I) {
5450  uint64_t Idx = (ResultList[I].first->getValue() - MinCaseVal->getValue())
5451  .getLimitedValue();
5452  MaskInt |= One << Idx;
5453  }
5454  ConstantInt *TableMask = ConstantInt::get(Mod.getContext(), MaskInt);
5455 
5456  // Get the TableIndex'th bit of the bitmask.
5457  // If this bit is 0 (meaning hole) jump to the default destination,
5458  // else continue with table lookup.
5459  IntegerType *MapTy = TableMask->getType();
5460  Value *MaskIndex =
5461  Builder.CreateZExtOrTrunc(TableIndex, MapTy, "switch.maskindex");
5462  Value *Shifted = Builder.CreateLShr(TableMask, MaskIndex, "switch.shifted");
5463  Value *LoBit = Builder.CreateTrunc(
5464  Shifted, Type::getInt1Ty(Mod.getContext()), "switch.lobit");
5465  Builder.CreateCondBr(LoBit, LookupBB, SI->getDefaultDest());
5466 
5467  Builder.SetInsertPoint(LookupBB);
5468  AddPredecessorToBlock(SI->getDefaultDest(), MaskBB, SI->getParent());
5469  }
5470 
5471  if (!DefaultIsReachable || GeneratingCoveredLookupTable) {
5472  // We cached PHINodes in PHIs. To avoid accessing deleted PHINodes later,
5473  // do not delete PHINodes here.
5475  /*KeepOneInputPHIs=*/true);
5476  }
5477 
5478  bool ReturnedEarly = false;
5479  for (PHINode *PHI : PHIs) {
5480  const ResultListTy &ResultList = ResultLists[PHI];
5481 
5482  // If using a bitmask, use any value to fill the lookup table holes.
5483  Constant *DV = NeedMask ? ResultLists[PHI][0].second : DefaultResults[PHI];
5484  StringRef FuncName = Fn->getName();
5485  SwitchLookupTable Table(Mod, TableSize, MinCaseVal, ResultList, DV, DL,
5486  FuncName);
5487 
5488  Value *Result = Table.BuildLookup(TableIndex, Builder);
5489 
5490  // If the result is used to return immediately from the function, we want to
5491  // do that right here.
5492  if (PHI->hasOneUse() && isa<ReturnInst>(*PHI->user_begin()) &&
5493  PHI->user_back() == CommonDest->getFirstNonPHIOrDbg()) {
5494  Builder.CreateRet(Result);
5495  ReturnedEarly = true;
5496  break;
5497  }
5498 
5499  // Do a small peephole optimization: re-use the switch table compare if
5500  // possible.
5501  if (!TableHasHoles && HasDefaultResults && RangeCheckBranch) {
5502  BasicBlock *PhiBlock = PHI->getParent();
5503  // Search for compare instructions which use the phi.
5504  for (auto *User : PHI->users()) {
5505  reuseTableCompare(User, PhiBlock, RangeCheckBranch, DV, ResultList);
5506  }
5507  }
5508 
5509  PHI->addIncoming(Result, LookupBB);
5510  }
5511 
5512  if (!ReturnedEarly)
5513  Builder.CreateBr(CommonDest);
5514 
5515  // Remove the switch.
5516  for (unsigned i = 0, e = SI->getNumSuccessors(); i < e; ++i) {
5517  BasicBlock *Succ = SI->getSuccessor(i);
5518 
5519  if (Succ == SI->getDefaultDest())
5520  continue;
5521  Succ->removePredecessor(SI->getParent());
5522  }
5523  SI->eraseFromParent();
5524 
5525  ++NumLookupTables;
5526  if (NeedMask)
5527  ++NumLookupTablesHoles;
5528  return true;
5529 }
5530 
5531 static bool isSwitchDense(ArrayRef<int64_t> Values) {
5532  // See also SelectionDAGBuilder::isDense(), which this function was based on.
5533  uint64_t Diff = (uint64_t)Values.back() - (uint64_t)Values.front();
5534  uint64_t Range = Diff + 1;
5535  uint64_t NumCases = Values.size();
5536  // 40% is the default density for building a jump table in optsize/minsize mode.
5537  uint64_t MinDensity = 40;
5538 
5539  return NumCases * 100 >= Range * MinDensity;
5540 }
5541 
5542 /// Try to transform a switch that has "holes" in it to a contiguous sequence
5543 /// of cases.
5544 ///
5545 /// A switch such as: switch(i) {case 5: case 9: case 13: case 17:} can be
5546 /// range-reduced to: switch ((i-5) / 4) {case 0: case 1: case 2: case 3:}.
5547 ///
5548 /// This converts a sparse switch into a dense switch which allows better
5549 /// lowering and could also allow transforming into a lookup table.
5551  const DataLayout &DL,
5552  const TargetTransformInfo &TTI) {
5553  auto *CondTy = cast<IntegerType>(SI->getCondition()->getType());
5554  if (CondTy->getIntegerBitWidth() > 64 ||
5555  !DL.fitsInLegalInteger(CondTy->getIntegerBitWidth()))
5556  return false;
5557  // Only bother with this optimization if there are more than 3 switch cases;
5558  // SDAG will only bother creating jump tables for 4 or more cases.
5559  if (SI->getNumCases() < 4)
5560  return false;
5561 
5562  // This transform is agnostic to the signedness of the input or case values. We
5563  // can treat the case values as signed or unsigned. We can optimize more common
5564  // cases such as a sequence crossing zero {-4,0,4,8} if we interpret case values
5565  // as signed.
5566  SmallVector<int64_t,4> Values;
5567  for (auto &C : SI->cases())
5568  Values.push_back(C.getCaseValue()->getValue().getSExtValue());
5569