LLVM  9.0.0svn
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 
97 static cl::opt<bool> DupRet(
98  "simplifycfg-dup-ret", cl::Hidden, cl::init(false),
99  cl::desc("Duplicate return instructions into unconditional branches"));
100 
101 static cl::opt<bool>
102  SinkCommon("simplifycfg-sink-common", cl::Hidden, cl::init(true),
103  cl::desc("Sink common instructions down to the end block"));
104 
106  "simplifycfg-hoist-cond-stores", cl::Hidden, cl::init(true),
107  cl::desc("Hoist conditional stores if an unconditional store precedes"));
108 
110  "simplifycfg-merge-cond-stores", cl::Hidden, cl::init(true),
111  cl::desc("Hoist conditional stores even if an unconditional store does not "
112  "precede - hoist multiple conditional stores into a single "
113  "predicated store"));
114 
116  "simplifycfg-merge-cond-stores-aggressively", cl::Hidden, cl::init(false),
117  cl::desc("When merging conditional stores, do so even if the resultant "
118  "basic blocks are unlikely to be if-converted as a result"));
119 
121  "speculate-one-expensive-inst", cl::Hidden, cl::init(true),
122  cl::desc("Allow exactly one expensive instruction to be speculatively "
123  "executed"));
124 
126  "max-speculation-depth", cl::Hidden, cl::init(10),
127  cl::desc("Limit maximum recursion depth when calculating costs of "
128  "speculatively executed instructions"));
129 
130 STATISTIC(NumBitMaps, "Number of switch instructions turned into bitmaps");
131 STATISTIC(NumLinearMaps,
132  "Number of switch instructions turned into linear mapping");
133 STATISTIC(NumLookupTables,
134  "Number of switch instructions turned into lookup tables");
135 STATISTIC(
136  NumLookupTablesHoles,
137  "Number of switch instructions turned into lookup tables (holes checked)");
138 STATISTIC(NumTableCmpReuses, "Number of reused switch table lookup compares");
139 STATISTIC(NumSinkCommons,
140  "Number of common instructions sunk down to the end block");
141 STATISTIC(NumSpeculations, "Number of speculative executed instructions");
142 
143 namespace {
144 
145 // The first field contains the value that the switch produces when a certain
146 // case group is selected, and the second field is a vector containing the
147 // cases composing the case group.
148 using SwitchCaseResultVectorTy =
150 
151 // The first field contains the phi node that generates a result of the switch
152 // and the second field contains the value generated for a certain case in the
153 // switch for that PHI.
154 using SwitchCaseResultsTy = SmallVector<std::pair<PHINode *, Constant *>, 4>;
155 
156 /// ValueEqualityComparisonCase - Represents a case of a switch.
157 struct ValueEqualityComparisonCase {
159  BasicBlock *Dest;
160 
161  ValueEqualityComparisonCase(ConstantInt *Value, BasicBlock *Dest)
162  : Value(Value), Dest(Dest) {}
163 
164  bool operator<(ValueEqualityComparisonCase RHS) const {
165  // Comparing pointers is ok as we only rely on the order for uniquing.
166  return Value < RHS.Value;
167  }
168 
169  bool operator==(BasicBlock *RHSDest) const { return Dest == RHSDest; }
170 };
171 
172 class SimplifyCFGOpt {
173  const TargetTransformInfo &TTI;
174  const DataLayout &DL;
175  SmallPtrSetImpl<BasicBlock *> *LoopHeaders;
176  const SimplifyCFGOptions &Options;
177  bool Resimplify;
178 
179  Value *isValueEqualityComparison(Instruction *TI);
180  BasicBlock *GetValueEqualityComparisonCases(
181  Instruction *TI, std::vector<ValueEqualityComparisonCase> &Cases);
182  bool SimplifyEqualityComparisonWithOnlyPredecessor(Instruction *TI,
183  BasicBlock *Pred,
184  IRBuilder<> &Builder);
185  bool FoldValueComparisonIntoPredecessors(Instruction *TI,
186  IRBuilder<> &Builder);
187 
188  bool SimplifyReturn(ReturnInst *RI, IRBuilder<> &Builder);
189  bool SimplifyResume(ResumeInst *RI, IRBuilder<> &Builder);
190  bool SimplifySingleResume(ResumeInst *RI);
191  bool SimplifyCommonResume(ResumeInst *RI);
192  bool SimplifyCleanupReturn(CleanupReturnInst *RI);
193  bool SimplifyUnreachable(UnreachableInst *UI);
194  bool SimplifySwitch(SwitchInst *SI, IRBuilder<> &Builder);
195  bool SimplifyIndirectBr(IndirectBrInst *IBI);
196  bool SimplifyUncondBranch(BranchInst *BI, IRBuilder<> &Builder);
197  bool SimplifyCondBranch(BranchInst *BI, IRBuilder<> &Builder);
198 
199  bool tryToSimplifyUncondBranchWithICmpInIt(ICmpInst *ICI,
200  IRBuilder<> &Builder);
201 
202 public:
203  SimplifyCFGOpt(const TargetTransformInfo &TTI, const DataLayout &DL,
204  SmallPtrSetImpl<BasicBlock *> *LoopHeaders,
205  const SimplifyCFGOptions &Opts)
206  : TTI(TTI), DL(DL), LoopHeaders(LoopHeaders), Options(Opts) {}
207 
208  bool run(BasicBlock *BB);
209  bool simplifyOnce(BasicBlock *BB);
210 
211  // Helper to set Resimplify and return change indication.
212  bool requestResimplify() {
213  Resimplify = true;
214  return true;
215  }
216 };
217 
218 } // end anonymous namespace
219 
220 /// Return true if it is safe to merge these two
221 /// terminator instructions together.
222 static bool
224  SmallSetVector<BasicBlock *, 4> *FailBlocks = nullptr) {
225  if (SI1 == SI2)
226  return false; // Can't merge with self!
227 
228  // It is not safe to merge these two switch instructions if they have a common
229  // successor, and if that successor has a PHI node, and if *that* PHI node has
230  // conflicting incoming values from the two switch blocks.
231  BasicBlock *SI1BB = SI1->getParent();
232  BasicBlock *SI2BB = SI2->getParent();
233 
234  SmallPtrSet<BasicBlock *, 16> SI1Succs(succ_begin(SI1BB), succ_end(SI1BB));
235  bool Fail = false;
236  for (BasicBlock *Succ : successors(SI2BB))
237  if (SI1Succs.count(Succ))
238  for (BasicBlock::iterator BBI = Succ->begin(); isa<PHINode>(BBI); ++BBI) {
239  PHINode *PN = cast<PHINode>(BBI);
240  if (PN->getIncomingValueForBlock(SI1BB) !=
241  PN->getIncomingValueForBlock(SI2BB)) {
242  if (FailBlocks)
243  FailBlocks->insert(Succ);
244  Fail = true;
245  }
246  }
247 
248  return !Fail;
249 }
250 
251 /// Return true if it is safe and profitable to merge these two terminator
252 /// instructions together, where SI1 is an unconditional branch. PhiNodes will
253 /// store all PHI nodes in common successors.
254 static bool
256  Instruction *Cond,
257  SmallVectorImpl<PHINode *> &PhiNodes) {
258  if (SI1 == SI2)
259  return false; // Can't merge with self!
260  assert(SI1->isUnconditional() && SI2->isConditional());
261 
262  // We fold the unconditional branch if we can easily update all PHI nodes in
263  // common successors:
264  // 1> We have a constant incoming value for the conditional branch;
265  // 2> We have "Cond" as the incoming value for the unconditional branch;
266  // 3> SI2->getCondition() and Cond have same operands.
267  CmpInst *Ci2 = dyn_cast<CmpInst>(SI2->getCondition());
268  if (!Ci2)
269  return false;
270  if (!(Cond->getOperand(0) == Ci2->getOperand(0) &&
271  Cond->getOperand(1) == Ci2->getOperand(1)) &&
272  !(Cond->getOperand(0) == Ci2->getOperand(1) &&
273  Cond->getOperand(1) == Ci2->getOperand(0)))
274  return false;
275 
276  BasicBlock *SI1BB = SI1->getParent();
277  BasicBlock *SI2BB = SI2->getParent();
278  SmallPtrSet<BasicBlock *, 16> SI1Succs(succ_begin(SI1BB), succ_end(SI1BB));
279  for (BasicBlock *Succ : successors(SI2BB))
280  if (SI1Succs.count(Succ))
281  for (BasicBlock::iterator BBI = Succ->begin(); isa<PHINode>(BBI); ++BBI) {
282  PHINode *PN = cast<PHINode>(BBI);
283  if (PN->getIncomingValueForBlock(SI1BB) != Cond ||
284  !isa<ConstantInt>(PN->getIncomingValueForBlock(SI2BB)))
285  return false;
286  PhiNodes.push_back(PN);
287  }
288  return true;
289 }
290 
291 /// Update PHI nodes in Succ to indicate that there will now be entries in it
292 /// from the 'NewPred' block. The values that will be flowing into the PHI nodes
293 /// will be the same as those coming in from ExistPred, an existing predecessor
294 /// of Succ.
295 static void AddPredecessorToBlock(BasicBlock *Succ, BasicBlock *NewPred,
296  BasicBlock *ExistPred,
297  MemorySSAUpdater *MSSAU = nullptr) {
298  for (PHINode &PN : Succ->phis())
299  PN.addIncoming(PN.getIncomingValueForBlock(ExistPred), NewPred);
300  if (MSSAU)
301  if (auto *MPhi = MSSAU->getMemorySSA()->getMemoryAccess(Succ))
302  MPhi->addIncoming(MPhi->getIncomingValueForBlock(ExistPred), NewPred);
303 }
304 
305 /// Compute an abstract "cost" of speculating the given instruction,
306 /// which is assumed to be safe to speculate. TCC_Free means cheap,
307 /// TCC_Basic means less cheap, and TCC_Expensive means prohibitively
308 /// expensive.
309 static unsigned ComputeSpeculationCost(const User *I,
310  const TargetTransformInfo &TTI) {
312  "Instruction is not safe to speculatively execute!");
313  return TTI.getUserCost(I);
314 }
315 
316 /// If we have a merge point of an "if condition" as accepted above,
317 /// return true if the specified value dominates the block. We
318 /// don't handle the true generality of domination here, just a special case
319 /// which works well enough for us.
320 ///
321 /// If AggressiveInsts is non-null, and if V does not dominate BB, we check to
322 /// see if V (which must be an instruction) and its recursive operands
323 /// that do not dominate BB have a combined cost lower than CostRemaining and
324 /// are non-trapping. If both are true, the instruction is inserted into the
325 /// set and true is returned.
326 ///
327 /// The cost for most non-trapping instructions is defined as 1 except for
328 /// Select whose cost is 2.
329 ///
330 /// After this function returns, CostRemaining is decreased by the cost of
331 /// V plus its non-dominating operands. If that cost is greater than
332 /// CostRemaining, false is returned and CostRemaining is undefined.
334  SmallPtrSetImpl<Instruction *> &AggressiveInsts,
335  unsigned &CostRemaining,
336  const TargetTransformInfo &TTI,
337  unsigned Depth = 0) {
338  // It is possible to hit a zero-cost cycle (phi/gep instructions for example),
339  // so limit the recursion depth.
340  // TODO: While this recursion limit does prevent pathological behavior, it
341  // would be better to track visited instructions to avoid cycles.
342  if (Depth == MaxSpeculationDepth)
343  return false;
344 
346  if (!I) {
347  // Non-instructions all dominate instructions, but not all constantexprs
348  // can be executed unconditionally.
349  if (ConstantExpr *C = dyn_cast<ConstantExpr>(V))
350  if (C->canTrap())
351  return false;
352  return true;
353  }
354  BasicBlock *PBB = I->getParent();
355 
356  // We don't want to allow weird loops that might have the "if condition" in
357  // the bottom of this block.
358  if (PBB == BB)
359  return false;
360 
361  // If this instruction is defined in a block that contains an unconditional
362  // branch to BB, then it must be in the 'conditional' part of the "if
363  // statement". If not, it definitely dominates the region.
365  if (!BI || BI->isConditional() || BI->getSuccessor(0) != BB)
366  return true;
367 
368  // If we have seen this instruction before, don't count it again.
369  if (AggressiveInsts.count(I))
370  return true;
371 
372  // Okay, it looks like the instruction IS in the "condition". Check to
373  // see if it's a cheap instruction to unconditionally compute, and if it
374  // only uses stuff defined outside of the condition. If so, hoist it out.
376  return false;
377 
378  unsigned Cost = ComputeSpeculationCost(I, TTI);
379 
380  // Allow exactly one instruction to be speculated regardless of its cost
381  // (as long as it is safe to do so).
382  // This is intended to flatten the CFG even if the instruction is a division
383  // or other expensive operation. The speculation of an expensive instruction
384  // is expected to be undone in CodeGenPrepare if the speculation has not
385  // enabled further IR optimizations.
386  if (Cost > CostRemaining &&
387  (!SpeculateOneExpensiveInst || !AggressiveInsts.empty() || Depth > 0))
388  return false;
389 
390  // Avoid unsigned wrap.
391  CostRemaining = (Cost > CostRemaining) ? 0 : CostRemaining - Cost;
392 
393  // Okay, we can only really hoist these out if their operands do
394  // not take us over the cost threshold.
395  for (User::op_iterator i = I->op_begin(), e = I->op_end(); i != e; ++i)
396  if (!DominatesMergePoint(*i, BB, AggressiveInsts, CostRemaining, TTI,
397  Depth + 1))
398  return false;
399  // Okay, it's safe to do this! Remember this instruction.
400  AggressiveInsts.insert(I);
401  return true;
402 }
403 
404 /// Extract ConstantInt from value, looking through IntToPtr
405 /// and PointerNullValue. Return NULL if value is not a constant int.
406 static ConstantInt *GetConstantInt(Value *V, const DataLayout &DL) {
407  // Normal constant int.
408  ConstantInt *CI = dyn_cast<ConstantInt>(V);
409  if (CI || !isa<Constant>(V) || !V->getType()->isPointerTy())
410  return CI;
411 
412  // This is some kind of pointer constant. Turn it into a pointer-sized
413  // ConstantInt if possible.
414  IntegerType *PtrTy = cast<IntegerType>(DL.getIntPtrType(V->getType()));
415 
416  // Null pointer means 0, see SelectionDAGBuilder::getValue(const Value*).
417  if (isa<ConstantPointerNull>(V))
418  return ConstantInt::get(PtrTy, 0);
419 
420  // IntToPtr const int.
421  if (ConstantExpr *CE = dyn_cast<ConstantExpr>(V))
422  if (CE->getOpcode() == Instruction::IntToPtr)
423  if (ConstantInt *CI = dyn_cast<ConstantInt>(CE->getOperand(0))) {
424  // The constant is very likely to have the right type already.
425  if (CI->getType() == PtrTy)
426  return CI;
427  else
428  return cast<ConstantInt>(
429  ConstantExpr::getIntegerCast(CI, PtrTy, /*isSigned=*/false));
430  }
431  return nullptr;
432 }
433 
434 namespace {
435 
436 /// Given a chain of or (||) or and (&&) comparison of a value against a
437 /// constant, this will try to recover the information required for a switch
438 /// structure.
439 /// It will depth-first traverse the chain of comparison, seeking for patterns
440 /// like %a == 12 or %a < 4 and combine them to produce a set of integer
441 /// representing the different cases for the switch.
442 /// Note that if the chain is composed of '||' it will build the set of elements
443 /// that matches the comparisons (i.e. any of this value validate the chain)
444 /// while for a chain of '&&' it will build the set elements that make the test
445 /// fail.
446 struct ConstantComparesGatherer {
447  const DataLayout &DL;
448 
449  /// Value found for the switch comparison
450  Value *CompValue = nullptr;
451 
452  /// Extra clause to be checked before the switch
453  Value *Extra = nullptr;
454 
455  /// Set of integers to match in switch
457 
458  /// Number of comparisons matched in the and/or chain
459  unsigned UsedICmps = 0;
460 
461  /// Construct and compute the result for the comparison instruction Cond
462  ConstantComparesGatherer(Instruction *Cond, const DataLayout &DL) : DL(DL) {
463  gather(Cond);
464  }
465 
466  ConstantComparesGatherer(const ConstantComparesGatherer &) = delete;
467  ConstantComparesGatherer &
468  operator=(const ConstantComparesGatherer &) = delete;
469 
470 private:
471  /// Try to set the current value used for the comparison, it succeeds only if
472  /// it wasn't set before or if the new value is the same as the old one
473  bool setValueOnce(Value *NewVal) {
474  if (CompValue && CompValue != NewVal)
475  return false;
476  CompValue = NewVal;
477  return (CompValue != nullptr);
478  }
479 
480  /// Try to match Instruction "I" as a comparison against a constant and
481  /// populates the array Vals with the set of values that match (or do not
482  /// match depending on isEQ).
483  /// Return false on failure. On success, the Value the comparison matched
484  /// against is placed in CompValue.
485  /// If CompValue is already set, the function is expected to fail if a match
486  /// is found but the value compared to is different.
487  bool matchInstruction(Instruction *I, bool isEQ) {
488  // If this is an icmp against a constant, handle this as one of the cases.
489  ICmpInst *ICI;
490  ConstantInt *C;
491  if (!((ICI = dyn_cast<ICmpInst>(I)) &&
492  (C = GetConstantInt(I->getOperand(1), DL)))) {
493  return false;
494  }
495 
496  Value *RHSVal;
497  const APInt *RHSC;
498 
499  // Pattern match a special case
500  // (x & ~2^z) == y --> x == y || x == y|2^z
501  // This undoes a transformation done by instcombine to fuse 2 compares.
502  if (ICI->getPredicate() == (isEQ ? ICmpInst::ICMP_EQ : ICmpInst::ICMP_NE)) {
503  // It's a little bit hard to see why the following transformations are
504  // correct. Here is a CVC3 program to verify them for 64-bit values:
505 
506  /*
507  ONE : BITVECTOR(64) = BVZEROEXTEND(0bin1, 63);
508  x : BITVECTOR(64);
509  y : BITVECTOR(64);
510  z : BITVECTOR(64);
511  mask : BITVECTOR(64) = BVSHL(ONE, z);
512  QUERY( (y & ~mask = y) =>
513  ((x & ~mask = y) <=> (x = y OR x = (y | mask)))
514  );
515  QUERY( (y | mask = y) =>
516  ((x | mask = y) <=> (x = y OR x = (y & ~mask)))
517  );
518  */
519 
520  // Please note that each pattern must be a dual implication (<--> or
521  // iff). One directional implication can create spurious matches. If the
522  // implication is only one-way, an unsatisfiable condition on the left
523  // side can imply a satisfiable condition on the right side. Dual
524  // implication ensures that satisfiable conditions are transformed to
525  // other satisfiable conditions and unsatisfiable conditions are
526  // transformed to other unsatisfiable conditions.
527 
528  // Here is a concrete example of a unsatisfiable condition on the left
529  // implying a satisfiable condition on the right:
530  //
531  // mask = (1 << z)
532  // (x & ~mask) == y --> (x == y || x == (y | mask))
533  //
534  // Substituting y = 3, z = 0 yields:
535  // (x & -2) == 3 --> (x == 3 || x == 2)
536 
537  // Pattern match a special case:
538  /*
539  QUERY( (y & ~mask = y) =>
540  ((x & ~mask = y) <=> (x = y OR x = (y | mask)))
541  );
542  */
543  if (match(ICI->getOperand(0),
544  m_And(m_Value(RHSVal), m_APInt(RHSC)))) {
545  APInt Mask = ~*RHSC;
546  if (Mask.isPowerOf2() && (C->getValue() & ~Mask) == C->getValue()) {
547  // If we already have a value for the switch, it has to match!
548  if (!setValueOnce(RHSVal))
549  return false;
550 
551  Vals.push_back(C);
552  Vals.push_back(
554  C->getValue() | Mask));
555  UsedICmps++;
556  return true;
557  }
558  }
559 
560  // Pattern match a special case:
561  /*
562  QUERY( (y | mask = y) =>
563  ((x | mask = y) <=> (x = y OR x = (y & ~mask)))
564  );
565  */
566  if (match(ICI->getOperand(0),
567  m_Or(m_Value(RHSVal), m_APInt(RHSC)))) {
568  APInt Mask = *RHSC;
569  if (Mask.isPowerOf2() && (C->getValue() | Mask) == C->getValue()) {
570  // If we already have a value for the switch, it has to match!
571  if (!setValueOnce(RHSVal))
572  return false;
573 
574  Vals.push_back(C);
576  C->getValue() & ~Mask));
577  UsedICmps++;
578  return true;
579  }
580  }
581 
582  // If we already have a value for the switch, it has to match!
583  if (!setValueOnce(ICI->getOperand(0)))
584  return false;
585 
586  UsedICmps++;
587  Vals.push_back(C);
588  return ICI->getOperand(0);
589  }
590 
591  // If we have "x ult 3", for example, then we can add 0,1,2 to the set.
593  ICI->getPredicate(), C->getValue());
594 
595  // Shift the range if the compare is fed by an add. This is the range
596  // compare idiom as emitted by instcombine.
597  Value *CandidateVal = I->getOperand(0);
598  if (match(I->getOperand(0), m_Add(m_Value(RHSVal), m_APInt(RHSC)))) {
599  Span = Span.subtract(*RHSC);
600  CandidateVal = RHSVal;
601  }
602 
603  // If this is an and/!= check, then we are looking to build the set of
604  // value that *don't* pass the and chain. I.e. to turn "x ugt 2" into
605  // x != 0 && x != 1.
606  if (!isEQ)
607  Span = Span.inverse();
608 
609  // If there are a ton of values, we don't want to make a ginormous switch.
610  if (Span.isSizeLargerThan(8) || Span.isEmptySet()) {
611  return false;
612  }
613 
614  // If we already have a value for the switch, it has to match!
615  if (!setValueOnce(CandidateVal))
616  return false;
617 
618  // Add all values from the range to the set
619  for (APInt Tmp = Span.getLower(); Tmp != Span.getUpper(); ++Tmp)
620  Vals.push_back(ConstantInt::get(I->getContext(), Tmp));
621 
622  UsedICmps++;
623  return true;
624  }
625 
626  /// Given a potentially 'or'd or 'and'd together collection of icmp
627  /// eq/ne/lt/gt instructions that compare a value against a constant, extract
628  /// the value being compared, and stick the list constants into the Vals
629  /// vector.
630  /// One "Extra" case is allowed to differ from the other.
631  void gather(Value *V) {
633  bool isEQ = (I->getOpcode() == Instruction::Or);
634 
635  // Keep a stack (SmallVector for efficiency) for depth-first traversal
637  SmallPtrSet<Value *, 8> Visited;
638 
639  // Initialize
640  Visited.insert(V);
641  DFT.push_back(V);
642 
643  while (!DFT.empty()) {
644  V = DFT.pop_back_val();
645 
646  if (Instruction *I = dyn_cast<Instruction>(V)) {
647  // If it is a || (or && depending on isEQ), process the operands.
648  if (I->getOpcode() == (isEQ ? Instruction::Or : Instruction::And)) {
649  if (Visited.insert(I->getOperand(1)).second)
650  DFT.push_back(I->getOperand(1));
651  if (Visited.insert(I->getOperand(0)).second)
652  DFT.push_back(I->getOperand(0));
653  continue;
654  }
655 
656  // Try to match the current instruction
657  if (matchInstruction(I, isEQ))
658  // Match succeed, continue the loop
659  continue;
660  }
661 
662  // One element of the sequence of || (or &&) could not be match as a
663  // comparison against the same value as the others.
664  // We allow only one "Extra" case to be checked before the switch
665  if (!Extra) {
666  Extra = V;
667  continue;
668  }
669  // Failed to parse a proper sequence, abort now
670  CompValue = nullptr;
671  break;
672  }
673  }
674 };
675 
676 } // end anonymous namespace
677 
679  MemorySSAUpdater *MSSAU = nullptr) {
680  Instruction *Cond = nullptr;
681  if (SwitchInst *SI = dyn_cast<SwitchInst>(TI)) {
682  Cond = dyn_cast<Instruction>(SI->getCondition());
683  } else if (BranchInst *BI = dyn_cast<BranchInst>(TI)) {
684  if (BI->isConditional())
685  Cond = dyn_cast<Instruction>(BI->getCondition());
686  } else if (IndirectBrInst *IBI = dyn_cast<IndirectBrInst>(TI)) {
687  Cond = dyn_cast<Instruction>(IBI->getAddress());
688  }
689 
690  TI->eraseFromParent();
691  if (Cond)
692  RecursivelyDeleteTriviallyDeadInstructions(Cond, nullptr, MSSAU);
693 }
694 
695 /// Return true if the specified terminator checks
696 /// to see if a value is equal to constant integer value.
697 Value *SimplifyCFGOpt::isValueEqualityComparison(Instruction *TI) {
698  Value *CV = nullptr;
699  if (SwitchInst *SI = dyn_cast<SwitchInst>(TI)) {
700  // Do not permit merging of large switch instructions into their
701  // predecessors unless there is only one predecessor.
702  if (!SI->getParent()->hasNPredecessorsOrMore(128 / SI->getNumSuccessors()))
703  CV = SI->getCondition();
704  } else if (BranchInst *BI = dyn_cast<BranchInst>(TI))
705  if (BI->isConditional() && BI->getCondition()->hasOneUse())
706  if (ICmpInst *ICI = dyn_cast<ICmpInst>(BI->getCondition())) {
707  if (ICI->isEquality() && GetConstantInt(ICI->getOperand(1), DL))
708  CV = ICI->getOperand(0);
709  }
710 
711  // Unwrap any lossless ptrtoint cast.
712  if (CV) {
713  if (PtrToIntInst *PTII = dyn_cast<PtrToIntInst>(CV)) {
714  Value *Ptr = PTII->getPointerOperand();
715  if (PTII->getType() == DL.getIntPtrType(Ptr->getType()))
716  CV = Ptr;
717  }
718  }
719  return CV;
720 }
721 
722 /// Given a value comparison instruction,
723 /// decode all of the 'cases' that it represents and return the 'default' block.
724 BasicBlock *SimplifyCFGOpt::GetValueEqualityComparisonCases(
725  Instruction *TI, std::vector<ValueEqualityComparisonCase> &Cases) {
726  if (SwitchInst *SI = dyn_cast<SwitchInst>(TI)) {
727  Cases.reserve(SI->getNumCases());
728  for (auto Case : SI->cases())
729  Cases.push_back(ValueEqualityComparisonCase(Case.getCaseValue(),
730  Case.getCaseSuccessor()));
731  return SI->getDefaultDest();
732  }
733 
734  BranchInst *BI = cast<BranchInst>(TI);
735  ICmpInst *ICI = cast<ICmpInst>(BI->getCondition());
736  BasicBlock *Succ = BI->getSuccessor(ICI->getPredicate() == ICmpInst::ICMP_NE);
737  Cases.push_back(ValueEqualityComparisonCase(
738  GetConstantInt(ICI->getOperand(1), DL), Succ));
739  return BI->getSuccessor(ICI->getPredicate() == ICmpInst::ICMP_EQ);
740 }
741 
742 /// Given a vector of bb/value pairs, remove any entries
743 /// in the list that match the specified block.
744 static void
746  std::vector<ValueEqualityComparisonCase> &Cases) {
747  Cases.erase(std::remove(Cases.begin(), Cases.end(), BB), Cases.end());
748 }
749 
750 /// Return true if there are any keys in C1 that exist in C2 as well.
751 static bool ValuesOverlap(std::vector<ValueEqualityComparisonCase> &C1,
752  std::vector<ValueEqualityComparisonCase> &C2) {
753  std::vector<ValueEqualityComparisonCase> *V1 = &C1, *V2 = &C2;
754 
755  // Make V1 be smaller than V2.
756  if (V1->size() > V2->size())
757  std::swap(V1, V2);
758 
759  if (V1->empty())
760  return false;
761  if (V1->size() == 1) {
762  // Just scan V2.
763  ConstantInt *TheVal = (*V1)[0].Value;
764  for (unsigned i = 0, e = V2->size(); i != e; ++i)
765  if (TheVal == (*V2)[i].Value)
766  return true;
767  }
768 
769  // Otherwise, just sort both lists and compare element by element.
770  array_pod_sort(V1->begin(), V1->end());
771  array_pod_sort(V2->begin(), V2->end());
772  unsigned i1 = 0, i2 = 0, e1 = V1->size(), e2 = V2->size();
773  while (i1 != e1 && i2 != e2) {
774  if ((*V1)[i1].Value == (*V2)[i2].Value)
775  return true;
776  if ((*V1)[i1].Value < (*V2)[i2].Value)
777  ++i1;
778  else
779  ++i2;
780  }
781  return false;
782 }
783 
784 // Set branch weights on SwitchInst. This sets the metadata if there is at
785 // least one non-zero weight.
787  // Check that there is at least one non-zero weight. Otherwise, pass
788  // nullptr to setMetadata which will erase the existing metadata.
789  MDNode *N = nullptr;
790  if (llvm::any_of(Weights, [](uint32_t W) { return W != 0; }))
791  N = MDBuilder(SI->getParent()->getContext()).createBranchWeights(Weights);
793 }
794 
795 // Similar to the above, but for branch and select instructions that take
796 // exactly 2 weights.
797 static void setBranchWeights(Instruction *I, uint32_t TrueWeight,
798  uint32_t FalseWeight) {
799  assert(isa<BranchInst>(I) || isa<SelectInst>(I));
800  // Check that there is at least one non-zero weight. Otherwise, pass
801  // nullptr to setMetadata which will erase the existing metadata.
802  MDNode *N = nullptr;
803  if (TrueWeight || FalseWeight)
804  N = MDBuilder(I->getParent()->getContext())
805  .createBranchWeights(TrueWeight, FalseWeight);
807 }
808 
809 /// If TI is known to be a terminator instruction and its block is known to
810 /// only have a single predecessor block, check to see if that predecessor is
811 /// also a value comparison with the same value, and if that comparison
812 /// determines the outcome of this comparison. If so, simplify TI. This does a
813 /// very limited form of jump threading.
814 bool SimplifyCFGOpt::SimplifyEqualityComparisonWithOnlyPredecessor(
815  Instruction *TI, BasicBlock *Pred, IRBuilder<> &Builder) {
816  Value *PredVal = isValueEqualityComparison(Pred->getTerminator());
817  if (!PredVal)
818  return false; // Not a value comparison in predecessor.
819 
820  Value *ThisVal = isValueEqualityComparison(TI);
821  assert(ThisVal && "This isn't a value comparison!!");
822  if (ThisVal != PredVal)
823  return false; // Different predicates.
824 
825  // TODO: Preserve branch weight metadata, similarly to how
826  // FoldValueComparisonIntoPredecessors preserves it.
827 
828  // Find out information about when control will move from Pred to TI's block.
829  std::vector<ValueEqualityComparisonCase> PredCases;
830  BasicBlock *PredDef =
831  GetValueEqualityComparisonCases(Pred->getTerminator(), PredCases);
832  EliminateBlockCases(PredDef, PredCases); // Remove default from cases.
833 
834  // Find information about how control leaves this block.
835  std::vector<ValueEqualityComparisonCase> ThisCases;
836  BasicBlock *ThisDef = GetValueEqualityComparisonCases(TI, ThisCases);
837  EliminateBlockCases(ThisDef, ThisCases); // Remove default from cases.
838 
839  // If TI's block is the default block from Pred's comparison, potentially
840  // simplify TI based on this knowledge.
841  if (PredDef == TI->getParent()) {
842  // If we are here, we know that the value is none of those cases listed in
843  // PredCases. If there are any cases in ThisCases that are in PredCases, we
844  // can simplify TI.
845  if (!ValuesOverlap(PredCases, ThisCases))
846  return false;
847 
848  if (isa<BranchInst>(TI)) {
849  // Okay, one of the successors of this condbr is dead. Convert it to a
850  // uncond br.
851  assert(ThisCases.size() == 1 && "Branch can only have one case!");
852  // Insert the new branch.
853  Instruction *NI = Builder.CreateBr(ThisDef);
854  (void)NI;
855 
856  // Remove PHI node entries for the dead edge.
857  ThisCases[0].Dest->removePredecessor(TI->getParent());
858 
859  LLVM_DEBUG(dbgs() << "Threading pred instr: " << *Pred->getTerminator()
860  << "Through successor TI: " << *TI << "Leaving: " << *NI
861  << "\n");
862 
864  return true;
865  }
866 
867  SwitchInstProfUpdateWrapper SI = *cast<SwitchInst>(TI);
868  // Okay, TI has cases that are statically dead, prune them away.
869  SmallPtrSet<Constant *, 16> DeadCases;
870  for (unsigned i = 0, e = PredCases.size(); i != e; ++i)
871  DeadCases.insert(PredCases[i].Value);
872 
873  LLVM_DEBUG(dbgs() << "Threading pred instr: " << *Pred->getTerminator()
874  << "Through successor TI: " << *TI);
875 
876  for (SwitchInst::CaseIt i = SI->case_end(), e = SI->case_begin(); i != e;) {
877  --i;
878  if (DeadCases.count(i->getCaseValue())) {
879  i->getCaseSuccessor()->removePredecessor(TI->getParent());
880  SI.removeCase(i);
881  }
882  }
883  LLVM_DEBUG(dbgs() << "Leaving: " << *TI << "\n");
884  return true;
885  }
886 
887  // Otherwise, TI's block must correspond to some matched value. Find out
888  // which value (or set of values) this is.
889  ConstantInt *TIV = nullptr;
890  BasicBlock *TIBB = TI->getParent();
891  for (unsigned i = 0, e = PredCases.size(); i != e; ++i)
892  if (PredCases[i].Dest == TIBB) {
893  if (TIV)
894  return false; // Cannot handle multiple values coming to this block.
895  TIV = PredCases[i].Value;
896  }
897  assert(TIV && "No edge from pred to succ?");
898 
899  // Okay, we found the one constant that our value can be if we get into TI's
900  // BB. Find out which successor will unconditionally be branched to.
901  BasicBlock *TheRealDest = nullptr;
902  for (unsigned i = 0, e = ThisCases.size(); i != e; ++i)
903  if (ThisCases[i].Value == TIV) {
904  TheRealDest = ThisCases[i].Dest;
905  break;
906  }
907 
908  // If not handled by any explicit cases, it is handled by the default case.
909  if (!TheRealDest)
910  TheRealDest = ThisDef;
911 
912  // Remove PHI node entries for dead edges.
913  BasicBlock *CheckEdge = TheRealDest;
914  for (BasicBlock *Succ : successors(TIBB))
915  if (Succ != CheckEdge)
916  Succ->removePredecessor(TIBB);
917  else
918  CheckEdge = nullptr;
919 
920  // Insert the new branch.
921  Instruction *NI = Builder.CreateBr(TheRealDest);
922  (void)NI;
923 
924  LLVM_DEBUG(dbgs() << "Threading pred instr: " << *Pred->getTerminator()
925  << "Through successor TI: " << *TI << "Leaving: " << *NI
926  << "\n");
927 
929  return true;
930 }
931 
932 namespace {
933 
934 /// This class implements a stable ordering of constant
935 /// integers that does not depend on their address. This is important for
936 /// applications that sort ConstantInt's to ensure uniqueness.
937 struct ConstantIntOrdering {
938  bool operator()(const ConstantInt *LHS, const ConstantInt *RHS) const {
939  return LHS->getValue().ult(RHS->getValue());
940  }
941 };
942 
943 } // end anonymous namespace
944 
945 static int ConstantIntSortPredicate(ConstantInt *const *P1,
946  ConstantInt *const *P2) {
947  const ConstantInt *LHS = *P1;
948  const ConstantInt *RHS = *P2;
949  if (LHS == RHS)
950  return 0;
951  return LHS->getValue().ult(RHS->getValue()) ? 1 : -1;
952 }
953 
954 static inline bool HasBranchWeights(const Instruction *I) {
956  if (ProfMD && ProfMD->getOperand(0))
957  if (MDString *MDS = dyn_cast<MDString>(ProfMD->getOperand(0)))
958  return MDS->getString().equals("branch_weights");
959 
960  return false;
961 }
962 
963 /// Get Weights of a given terminator, the default weight is at the front
964 /// of the vector. If TI is a conditional eq, we need to swap the branch-weight
965 /// metadata.
967  SmallVectorImpl<uint64_t> &Weights) {
969  assert(MD);
970  for (unsigned i = 1, e = MD->getNumOperands(); i < e; ++i) {
971  ConstantInt *CI = mdconst::extract<ConstantInt>(MD->getOperand(i));
972  Weights.push_back(CI->getValue().getZExtValue());
973  }
974 
975  // If TI is a conditional eq, the default case is the false case,
976  // and the corresponding branch-weight data is at index 2. We swap the
977  // default weight to be the first entry.
978  if (BranchInst *BI = dyn_cast<BranchInst>(TI)) {
979  assert(Weights.size() == 2);
980  ICmpInst *ICI = cast<ICmpInst>(BI->getCondition());
981  if (ICI->getPredicate() == ICmpInst::ICMP_EQ)
982  std::swap(Weights.front(), Weights.back());
983  }
984 }
985 
986 /// Keep halving the weights until all can fit in uint32_t.
987 static void FitWeights(MutableArrayRef<uint64_t> Weights) {
988  uint64_t Max = *std::max_element(Weights.begin(), Weights.end());
989  if (Max > UINT_MAX) {
990  unsigned Offset = 32 - countLeadingZeros(Max);
991  for (uint64_t &I : Weights)
992  I >>= Offset;
993  }
994 }
995 
996 /// The specified terminator is a value equality comparison instruction
997 /// (either a switch or a branch on "X == c").
998 /// See if any of the predecessors of the terminator block are value comparisons
999 /// on the same value. If so, and if safe to do so, fold them together.
1000 bool SimplifyCFGOpt::FoldValueComparisonIntoPredecessors(Instruction *TI,
1001  IRBuilder<> &Builder) {
1002  BasicBlock *BB = TI->getParent();
1003  Value *CV = isValueEqualityComparison(TI); // CondVal
1004  assert(CV && "Not a comparison?");
1005  bool Changed = false;
1006 
1008  while (!Preds.empty()) {
1009  BasicBlock *Pred = Preds.pop_back_val();
1010 
1011  // See if the predecessor is a comparison with the same value.
1012  Instruction *PTI = Pred->getTerminator();
1013  Value *PCV = isValueEqualityComparison(PTI); // PredCondVal
1014 
1015  if (PCV == CV && TI != PTI) {
1016  SmallSetVector<BasicBlock*, 4> FailBlocks;
1017  if (!SafeToMergeTerminators(TI, PTI, &FailBlocks)) {
1018  for (auto *Succ : FailBlocks) {
1019  if (!SplitBlockPredecessors(Succ, TI->getParent(), ".fold.split"))
1020  return false;
1021  }
1022  }
1023 
1024  // Figure out which 'cases' to copy from SI to PSI.
1025  std::vector<ValueEqualityComparisonCase> BBCases;
1026  BasicBlock *BBDefault = GetValueEqualityComparisonCases(TI, BBCases);
1027 
1028  std::vector<ValueEqualityComparisonCase> PredCases;
1029  BasicBlock *PredDefault = GetValueEqualityComparisonCases(PTI, PredCases);
1030 
1031  // Based on whether the default edge from PTI goes to BB or not, fill in
1032  // PredCases and PredDefault with the new switch cases we would like to
1033  // build.
1034  SmallVector<BasicBlock *, 8> NewSuccessors;
1035 
1036  // Update the branch weight metadata along the way
1037  SmallVector<uint64_t, 8> Weights;
1038  bool PredHasWeights = HasBranchWeights(PTI);
1039  bool SuccHasWeights = HasBranchWeights(TI);
1040 
1041  if (PredHasWeights) {
1042  GetBranchWeights(PTI, Weights);
1043  // branch-weight metadata is inconsistent here.
1044  if (Weights.size() != 1 + PredCases.size())
1045  PredHasWeights = SuccHasWeights = false;
1046  } else if (SuccHasWeights)
1047  // If there are no predecessor weights but there are successor weights,
1048  // populate Weights with 1, which will later be scaled to the sum of
1049  // successor's weights
1050  Weights.assign(1 + PredCases.size(), 1);
1051 
1052  SmallVector<uint64_t, 8> SuccWeights;
1053  if (SuccHasWeights) {
1054  GetBranchWeights(TI, SuccWeights);
1055  // branch-weight metadata is inconsistent here.
1056  if (SuccWeights.size() != 1 + BBCases.size())
1057  PredHasWeights = SuccHasWeights = false;
1058  } else if (PredHasWeights)
1059  SuccWeights.assign(1 + BBCases.size(), 1);
1060 
1061  if (PredDefault == BB) {
1062  // If this is the default destination from PTI, only the edges in TI
1063  // that don't occur in PTI, or that branch to BB will be activated.
1064  std::set<ConstantInt *, ConstantIntOrdering> PTIHandled;
1065  for (unsigned i = 0, e = PredCases.size(); i != e; ++i)
1066  if (PredCases[i].Dest != BB)
1067  PTIHandled.insert(PredCases[i].Value);
1068  else {
1069  // The default destination is BB, we don't need explicit targets.
1070  std::swap(PredCases[i], PredCases.back());
1071 
1072  if (PredHasWeights || SuccHasWeights) {
1073  // Increase weight for the default case.
1074  Weights[0] += Weights[i + 1];
1075  std::swap(Weights[i + 1], Weights.back());
1076  Weights.pop_back();
1077  }
1078 
1079  PredCases.pop_back();
1080  --i;
1081  --e;
1082  }
1083 
1084  // Reconstruct the new switch statement we will be building.
1085  if (PredDefault != BBDefault) {
1086  PredDefault->removePredecessor(Pred);
1087  PredDefault = BBDefault;
1088  NewSuccessors.push_back(BBDefault);
1089  }
1090 
1091  unsigned CasesFromPred = Weights.size();
1092  uint64_t ValidTotalSuccWeight = 0;
1093  for (unsigned i = 0, e = BBCases.size(); i != e; ++i)
1094  if (!PTIHandled.count(BBCases[i].Value) &&
1095  BBCases[i].Dest != BBDefault) {
1096  PredCases.push_back(BBCases[i]);
1097  NewSuccessors.push_back(BBCases[i].Dest);
1098  if (SuccHasWeights || PredHasWeights) {
1099  // The default weight is at index 0, so weight for the ith case
1100  // should be at index i+1. Scale the cases from successor by
1101  // PredDefaultWeight (Weights[0]).
1102  Weights.push_back(Weights[0] * SuccWeights[i + 1]);
1103  ValidTotalSuccWeight += SuccWeights[i + 1];
1104  }
1105  }
1106 
1107  if (SuccHasWeights || PredHasWeights) {
1108  ValidTotalSuccWeight += SuccWeights[0];
1109  // Scale the cases from predecessor by ValidTotalSuccWeight.
1110  for (unsigned i = 1; i < CasesFromPred; ++i)
1111  Weights[i] *= ValidTotalSuccWeight;
1112  // Scale the default weight by SuccDefaultWeight (SuccWeights[0]).
1113  Weights[0] *= SuccWeights[0];
1114  }
1115  } else {
1116  // If this is not the default destination from PSI, only the edges
1117  // in SI that occur in PSI with a destination of BB will be
1118  // activated.
1119  std::set<ConstantInt *, ConstantIntOrdering> PTIHandled;
1120  std::map<ConstantInt *, uint64_t> WeightsForHandled;
1121  for (unsigned i = 0, e = PredCases.size(); i != e; ++i)
1122  if (PredCases[i].Dest == BB) {
1123  PTIHandled.insert(PredCases[i].Value);
1124 
1125  if (PredHasWeights || SuccHasWeights) {
1126  WeightsForHandled[PredCases[i].Value] = Weights[i + 1];
1127  std::swap(Weights[i + 1], Weights.back());
1128  Weights.pop_back();
1129  }
1130 
1131  std::swap(PredCases[i], PredCases.back());
1132  PredCases.pop_back();
1133  --i;
1134  --e;
1135  }
1136 
1137  // Okay, now we know which constants were sent to BB from the
1138  // predecessor. Figure out where they will all go now.
1139  for (unsigned i = 0, e = BBCases.size(); i != e; ++i)
1140  if (PTIHandled.count(BBCases[i].Value)) {
1141  // If this is one we are capable of getting...
1142  if (PredHasWeights || SuccHasWeights)
1143  Weights.push_back(WeightsForHandled[BBCases[i].Value]);
1144  PredCases.push_back(BBCases[i]);
1145  NewSuccessors.push_back(BBCases[i].Dest);
1146  PTIHandled.erase(
1147  BBCases[i].Value); // This constant is taken care of
1148  }
1149 
1150  // If there are any constants vectored to BB that TI doesn't handle,
1151  // they must go to the default destination of TI.
1152  for (ConstantInt *I : PTIHandled) {
1153  if (PredHasWeights || SuccHasWeights)
1154  Weights.push_back(WeightsForHandled[I]);
1155  PredCases.push_back(ValueEqualityComparisonCase(I, BBDefault));
1156  NewSuccessors.push_back(BBDefault);
1157  }
1158  }
1159 
1160  // Okay, at this point, we know which new successor Pred will get. Make
1161  // sure we update the number of entries in the PHI nodes for these
1162  // successors.
1163  for (BasicBlock *NewSuccessor : NewSuccessors)
1164  AddPredecessorToBlock(NewSuccessor, Pred, BB);
1165 
1166  Builder.SetInsertPoint(PTI);
1167  // Convert pointer to int before we switch.
1168  if (CV->getType()->isPointerTy()) {
1169  CV = Builder.CreatePtrToInt(CV, DL.getIntPtrType(CV->getType()),
1170  "magicptr");
1171  }
1172 
1173  // Now that the successors are updated, create the new Switch instruction.
1174  SwitchInst *NewSI =
1175  Builder.CreateSwitch(CV, PredDefault, PredCases.size());
1176  NewSI->setDebugLoc(PTI->getDebugLoc());
1177  for (ValueEqualityComparisonCase &V : PredCases)
1178  NewSI->addCase(V.Value, V.Dest);
1179 
1180  if (PredHasWeights || SuccHasWeights) {
1181  // Halve the weights if any of them cannot fit in an uint32_t
1182  FitWeights(Weights);
1183 
1184  SmallVector<uint32_t, 8> MDWeights(Weights.begin(), Weights.end());
1185 
1186  setBranchWeights(NewSI, MDWeights);
1187  }
1188 
1190 
1191  // Okay, last check. If BB is still a successor of PSI, then we must
1192  // have an infinite loop case. If so, add an infinitely looping block
1193  // to handle the case to preserve the behavior of the code.
1194  BasicBlock *InfLoopBlock = nullptr;
1195  for (unsigned i = 0, e = NewSI->getNumSuccessors(); i != e; ++i)
1196  if (NewSI->getSuccessor(i) == BB) {
1197  if (!InfLoopBlock) {
1198  // Insert it at the end of the function, because it's either code,
1199  // or it won't matter if it's hot. :)
1200  InfLoopBlock = BasicBlock::Create(BB->getContext(), "infloop",
1201  BB->getParent());
1202  BranchInst::Create(InfLoopBlock, InfLoopBlock);
1203  }
1204  NewSI->setSuccessor(i, InfLoopBlock);
1205  }
1206 
1207  Changed = true;
1208  }
1209  }
1210  return Changed;
1211 }
1212 
1213 // If we would need to insert a select that uses the value of this invoke
1214 // (comments in HoistThenElseCodeToIf explain why we would need to do this), we
1215 // can't hoist the invoke, as there is nowhere to put the select in this case.
1217  Instruction *I1, Instruction *I2) {
1218  for (BasicBlock *Succ : successors(BB1)) {
1219  for (const PHINode &PN : Succ->phis()) {
1220  Value *BB1V = PN.getIncomingValueForBlock(BB1);
1221  Value *BB2V = PN.getIncomingValueForBlock(BB2);
1222  if (BB1V != BB2V && (BB1V == I1 || BB2V == I2)) {
1223  return false;
1224  }
1225  }
1226  }
1227  return true;
1228 }
1229 
1231 
1232 /// Given a conditional branch that goes to BB1 and BB2, hoist any common code
1233 /// in the two blocks up into the branch block. The caller of this function
1234 /// guarantees that BI's block dominates BB1 and BB2.
1236  const TargetTransformInfo &TTI) {
1237  // This does very trivial matching, with limited scanning, to find identical
1238  // instructions in the two blocks. In particular, we don't want to get into
1239  // O(M*N) situations here where M and N are the sizes of BB1 and BB2. As
1240  // such, we currently just scan for obviously identical instructions in an
1241  // identical order.
1242  BasicBlock *BB1 = BI->getSuccessor(0); // The true destination.
1243  BasicBlock *BB2 = BI->getSuccessor(1); // The false destination
1244 
1245  BasicBlock::iterator BB1_Itr = BB1->begin();
1246  BasicBlock::iterator BB2_Itr = BB2->begin();
1247 
1248  Instruction *I1 = &*BB1_Itr++, *I2 = &*BB2_Itr++;
1249  // Skip debug info if it is not identical.
1252  if (!DBI1 || !DBI2 || !DBI1->isIdenticalToWhenDefined(DBI2)) {
1253  while (isa<DbgInfoIntrinsic>(I1))
1254  I1 = &*BB1_Itr++;
1255  while (isa<DbgInfoIntrinsic>(I2))
1256  I2 = &*BB2_Itr++;
1257  }
1258  // FIXME: Can we define a safety predicate for CallBr?
1259  if (isa<PHINode>(I1) || !I1->isIdenticalToWhenDefined(I2) ||
1260  (isa<InvokeInst>(I1) && !isSafeToHoistInvoke(BB1, BB2, I1, I2)) ||
1261  isa<CallBrInst>(I1))
1262  return false;
1263 
1264  BasicBlock *BIParent = BI->getParent();
1265 
1266  bool Changed = false;
1267  do {
1268  // If we are hoisting the terminator instruction, don't move one (making a
1269  // broken BB), instead clone it, and remove BI.
1270  if (I1->isTerminator())
1271  goto HoistTerminator;
1272 
1273  // If we're going to hoist a call, make sure that the two instructions we're
1274  // commoning/hoisting are both marked with musttail, or neither of them is
1275  // marked as such. Otherwise, we might end up in a situation where we hoist
1276  // from a block where the terminator is a `ret` to a block where the terminator
1277  // is a `br`, and `musttail` calls expect to be followed by a return.
1278  auto *C1 = dyn_cast<CallInst>(I1);
1279  auto *C2 = dyn_cast<CallInst>(I2);
1280  if (C1 && C2)
1281  if (C1->isMustTailCall() != C2->isMustTailCall())
1282  return Changed;
1283 
1284  if (!TTI.isProfitableToHoist(I1) || !TTI.isProfitableToHoist(I2))
1285  return Changed;
1286 
1287  if (isa<DbgInfoIntrinsic>(I1) || isa<DbgInfoIntrinsic>(I2)) {
1288  assert (isa<DbgInfoIntrinsic>(I1) && isa<DbgInfoIntrinsic>(I2));
1289  // The debug location is an integral part of a debug info intrinsic
1290  // and can't be separated from it or replaced. Instead of attempting
1291  // to merge locations, simply hoist both copies of the intrinsic.
1292  BIParent->getInstList().splice(BI->getIterator(),
1293  BB1->getInstList(), I1);
1294  BIParent->getInstList().splice(BI->getIterator(),
1295  BB2->getInstList(), I2);
1296  Changed = true;
1297  } else {
1298  // For a normal instruction, we just move one to right before the branch,
1299  // then replace all uses of the other with the first. Finally, we remove
1300  // the now redundant second instruction.
1301  BIParent->getInstList().splice(BI->getIterator(),
1302  BB1->getInstList(), I1);
1303  if (!I2->use_empty())
1304  I2->replaceAllUsesWith(I1);
1305  I1->andIRFlags(I2);
1306  unsigned KnownIDs[] = {LLVMContext::MD_tbaa,
1317  combineMetadata(I1, I2, KnownIDs, true);
1318 
1319  // I1 and I2 are being combined into a single instruction. Its debug
1320  // location is the merged locations of the original instructions.
1321  I1->applyMergedLocation(I1->getDebugLoc(), I2->getDebugLoc());
1322 
1323  I2->eraseFromParent();
1324  Changed = true;
1325  }
1326 
1327  I1 = &*BB1_Itr++;
1328  I2 = &*BB2_Itr++;
1329  // Skip debug info if it is not identical.
1332  if (!DBI1 || !DBI2 || !DBI1->isIdenticalToWhenDefined(DBI2)) {
1333  while (isa<DbgInfoIntrinsic>(I1))
1334  I1 = &*BB1_Itr++;
1335  while (isa<DbgInfoIntrinsic>(I2))
1336  I2 = &*BB2_Itr++;
1337  }
1338  } while (I1->isIdenticalToWhenDefined(I2));
1339 
1340  return true;
1341 
1342 HoistTerminator:
1343  // It may not be possible to hoist an invoke.
1344  // FIXME: Can we define a safety predicate for CallBr?
1345  if (isa<InvokeInst>(I1) && !isSafeToHoistInvoke(BB1, BB2, I1, I2))
1346  return Changed;
1347 
1348  // TODO: callbr hoisting currently disabled pending further study.
1349  if (isa<CallBrInst>(I1))
1350  return Changed;
1351 
1352  for (BasicBlock *Succ : successors(BB1)) {
1353  for (PHINode &PN : Succ->phis()) {
1354  Value *BB1V = PN.getIncomingValueForBlock(BB1);
1355  Value *BB2V = PN.getIncomingValueForBlock(BB2);
1356  if (BB1V == BB2V)
1357  continue;
1358 
1359  // Check for passingValueIsAlwaysUndefined here because we would rather
1360  // eliminate undefined control flow then converting it to a select.
1361  if (passingValueIsAlwaysUndefined(BB1V, &PN) ||
1362  passingValueIsAlwaysUndefined(BB2V, &PN))
1363  return Changed;
1364 
1365  if (isa<ConstantExpr>(BB1V) && !isSafeToSpeculativelyExecute(BB1V))
1366  return Changed;
1367  if (isa<ConstantExpr>(BB2V) && !isSafeToSpeculativelyExecute(BB2V))
1368  return Changed;
1369  }
1370  }
1371 
1372  // Okay, it is safe to hoist the terminator.
1373  Instruction *NT = I1->clone();
1374  BIParent->getInstList().insert(BI->getIterator(), NT);
1375  if (!NT->getType()->isVoidTy()) {
1376  I1->replaceAllUsesWith(NT);
1377  I2->replaceAllUsesWith(NT);
1378  NT->takeName(I1);
1379  }
1380 
1381  // Ensure terminator gets a debug location, even an unknown one, in case
1382  // it involves inlinable calls.
1383  NT->applyMergedLocation(I1->getDebugLoc(), I2->getDebugLoc());
1384 
1385  // PHIs created below will adopt NT's merged DebugLoc.
1386  IRBuilder<NoFolder> Builder(NT);
1387 
1388  // Hoisting one of the terminators from our successor is a great thing.
1389  // Unfortunately, the successors of the if/else blocks may have PHI nodes in
1390  // them. If they do, all PHI entries for BB1/BB2 must agree for all PHI
1391  // nodes, so we insert select instruction to compute the final result.
1392  std::map<std::pair<Value *, Value *>, SelectInst *> InsertedSelects;
1393  for (BasicBlock *Succ : successors(BB1)) {
1394  for (PHINode &PN : Succ->phis()) {
1395  Value *BB1V = PN.getIncomingValueForBlock(BB1);
1396  Value *BB2V = PN.getIncomingValueForBlock(BB2);
1397  if (BB1V == BB2V)
1398  continue;
1399 
1400  // These values do not agree. Insert a select instruction before NT
1401  // that determines the right value.
1402  SelectInst *&SI = InsertedSelects[std::make_pair(BB1V, BB2V)];
1403  if (!SI)
1404  SI = cast<SelectInst>(
1405  Builder.CreateSelect(BI->getCondition(), BB1V, BB2V,
1406  BB1V->getName() + "." + BB2V->getName(), BI));
1407 
1408  // Make the PHI node use the select for all incoming values for BB1/BB2
1409  for (unsigned i = 0, e = PN.getNumIncomingValues(); i != e; ++i)
1410  if (PN.getIncomingBlock(i) == BB1 || PN.getIncomingBlock(i) == BB2)
1411  PN.setIncomingValue(i, SI);
1412  }
1413  }
1414 
1415  // Update any PHI nodes in our new successors.
1416  for (BasicBlock *Succ : successors(BB1))
1417  AddPredecessorToBlock(Succ, BIParent, BB1);
1418 
1420  return true;
1421 }
1422 
1423 // All instructions in Insts belong to different blocks that all unconditionally
1424 // branch to a common successor. Analyze each instruction and return true if it
1425 // would be possible to sink them into their successor, creating one common
1426 // instruction instead. For every value that would be required to be provided by
1427 // PHI node (because an operand varies in each input block), add to PHIOperands.
1430  DenseMap<Instruction *, SmallVector<Value *, 4>> &PHIOperands) {
1431  // Prune out obviously bad instructions to move. Each instruction must have
1432  // exactly zero or one use, and we check later that use is by a single, common
1433  // PHI instruction in the successor.
1434  bool HasUse = !Insts.front()->user_empty();
1435  for (auto *I : Insts) {
1436  // These instructions may change or break semantics if moved.
1437  if (isa<PHINode>(I) || I->isEHPad() || isa<AllocaInst>(I) ||
1438  I->getType()->isTokenTy())
1439  return false;
1440 
1441  // Conservatively return false if I is an inline-asm instruction. Sinking
1442  // and merging inline-asm instructions can potentially create arguments
1443  // that cannot satisfy the inline-asm constraints.
1444  if (const auto *C = dyn_cast<CallBase>(I))
1445  if (C->isInlineAsm())
1446  return false;
1447 
1448  // Each instruction must have zero or one use.
1449  if (HasUse && !I->hasOneUse())
1450  return false;
1451  if (!HasUse && !I->user_empty())
1452  return false;
1453  }
1454 
1455  const Instruction *I0 = Insts.front();
1456  for (auto *I : Insts)
1457  if (!I->isSameOperationAs(I0))
1458  return false;
1459 
1460  // All instructions in Insts are known to be the same opcode. If they have a
1461  // use, check that the only user is a PHI or in the same block as the
1462  // instruction, because if a user is in the same block as an instruction we're
1463  // contemplating sinking, it must already be determined to be sinkable.
1464  if (HasUse) {
1465  auto *PNUse = dyn_cast<PHINode>(*I0->user_begin());
1466  auto *Succ = I0->getParent()->getTerminator()->getSuccessor(0);
1467  if (!all_of(Insts, [&PNUse,&Succ](const Instruction *I) -> bool {
1468  auto *U = cast<Instruction>(*I->user_begin());
1469  return (PNUse &&
1470  PNUse->getParent() == Succ &&
1471  PNUse->getIncomingValueForBlock(I->getParent()) == I) ||
1472  U->getParent() == I->getParent();
1473  }))
1474  return false;
1475  }
1476 
1477  // Because SROA can't handle speculating stores of selects, try not
1478  // to sink loads or stores of allocas when we'd have to create a PHI for
1479  // the address operand. Also, because it is likely that loads or stores
1480  // of allocas will disappear when Mem2Reg/SROA is run, don't sink them.
1481  // This can cause code churn which can have unintended consequences down
1482  // the line - see https://llvm.org/bugs/show_bug.cgi?id=30244.
1483  // FIXME: This is a workaround for a deficiency in SROA - see
1484  // https://llvm.org/bugs/show_bug.cgi?id=30188
1485  if (isa<StoreInst>(I0) && any_of(Insts, [](const Instruction *I) {
1486  return isa<AllocaInst>(I->getOperand(1));
1487  }))
1488  return false;
1489  if (isa<LoadInst>(I0) && any_of(Insts, [](const Instruction *I) {
1490  return isa<AllocaInst>(I->getOperand(0));
1491  }))
1492  return false;
1493 
1494  for (unsigned OI = 0, OE = I0->getNumOperands(); OI != OE; ++OI) {
1495  if (I0->getOperand(OI)->getType()->isTokenTy())
1496  // Don't touch any operand of token type.
1497  return false;
1498 
1499  auto SameAsI0 = [&I0, OI](const Instruction *I) {
1500  assert(I->getNumOperands() == I0->getNumOperands());
1501  return I->getOperand(OI) == I0->getOperand(OI);
1502  };
1503  if (!all_of(Insts, SameAsI0)) {
1504  if (!canReplaceOperandWithVariable(I0, OI))
1505  // We can't create a PHI from this GEP.
1506  return false;
1507  // Don't create indirect calls! The called value is the final operand.
1508  if (isa<CallBase>(I0) && OI == OE - 1) {
1509  // FIXME: if the call was *already* indirect, we should do this.
1510  return false;
1511  }
1512  for (auto *I : Insts)
1513  PHIOperands[I].push_back(I->getOperand(OI));
1514  }
1515  }
1516  return true;
1517 }
1518 
1519 // Assuming canSinkLastInstruction(Blocks) has returned true, sink the last
1520 // instruction of every block in Blocks to their common successor, commoning
1521 // into one instruction.
1523  auto *BBEnd = Blocks[0]->getTerminator()->getSuccessor(0);
1524 
1525  // canSinkLastInstruction returning true guarantees that every block has at
1526  // least one non-terminator instruction.
1528  for (auto *BB : Blocks) {
1529  Instruction *I = BB->getTerminator();
1530  do {
1531  I = I->getPrevNode();
1532  } while (isa<DbgInfoIntrinsic>(I) && I != &BB->front());
1533  if (!isa<DbgInfoIntrinsic>(I))
1534  Insts.push_back(I);
1535  }
1536 
1537  // The only checking we need to do now is that all users of all instructions
1538  // are the same PHI node. canSinkLastInstruction should have checked this but
1539  // it is slightly over-aggressive - it gets confused by commutative instructions
1540  // so double-check it here.
1541  Instruction *I0 = Insts.front();
1542  if (!I0->user_empty()) {
1543  auto *PNUse = dyn_cast<PHINode>(*I0->user_begin());
1544  if (!all_of(Insts, [&PNUse](const Instruction *I) -> bool {
1545  auto *U = cast<Instruction>(*I->user_begin());
1546  return U == PNUse;
1547  }))
1548  return false;
1549  }
1550 
1551  // We don't need to do any more checking here; canSinkLastInstruction should
1552  // have done it all for us.
1553  SmallVector<Value*, 4> NewOperands;
1554  for (unsigned O = 0, E = I0->getNumOperands(); O != E; ++O) {
1555  // This check is different to that in canSinkLastInstruction. There, we
1556  // cared about the global view once simplifycfg (and instcombine) have
1557  // completed - it takes into account PHIs that become trivially
1558  // simplifiable. However here we need a more local view; if an operand
1559  // differs we create a PHI and rely on instcombine to clean up the very
1560  // small mess we may make.
1561  bool NeedPHI = any_of(Insts, [&I0, O](const Instruction *I) {
1562  return I->getOperand(O) != I0->getOperand(O);
1563  });
1564  if (!NeedPHI) {
1565  NewOperands.push_back(I0->getOperand(O));
1566  continue;
1567  }
1568 
1569  // Create a new PHI in the successor block and populate it.
1570  auto *Op = I0->getOperand(O);
1571  assert(!Op->getType()->isTokenTy() && "Can't PHI tokens!");
1572  auto *PN = PHINode::Create(Op->getType(), Insts.size(),
1573  Op->getName() + ".sink", &BBEnd->front());
1574  for (auto *I : Insts)
1575  PN->addIncoming(I->getOperand(O), I->getParent());
1576  NewOperands.push_back(PN);
1577  }
1578 
1579  // Arbitrarily use I0 as the new "common" instruction; remap its operands
1580  // and move it to the start of the successor block.
1581  for (unsigned O = 0, E = I0->getNumOperands(); O != E; ++O)
1582  I0->getOperandUse(O).set(NewOperands[O]);
1583  I0->moveBefore(&*BBEnd->getFirstInsertionPt());
1584 
1585  // Update metadata and IR flags, and merge debug locations.
1586  for (auto *I : Insts)
1587  if (I != I0) {
1588  // The debug location for the "common" instruction is the merged locations
1589  // of all the commoned instructions. We start with the original location
1590  // of the "common" instruction and iteratively merge each location in the
1591  // loop below.
1592  // This is an N-way merge, which will be inefficient if I0 is a CallInst.
1593  // However, as N-way merge for CallInst is rare, so we use simplified API
1594  // instead of using complex API for N-way merge.
1595  I0->applyMergedLocation(I0->getDebugLoc(), I->getDebugLoc());
1596  combineMetadataForCSE(I0, I, true);
1597  I0->andIRFlags(I);
1598  }
1599 
1600  if (!I0->user_empty()) {
1601  // canSinkLastInstruction checked that all instructions were used by
1602  // one and only one PHI node. Find that now, RAUW it to our common
1603  // instruction and nuke it.
1604  auto *PN = cast<PHINode>(*I0->user_begin());
1605  PN->replaceAllUsesWith(I0);
1606  PN->eraseFromParent();
1607  }
1608 
1609  // Finally nuke all instructions apart from the common instruction.
1610  for (auto *I : Insts)
1611  if (I != I0)
1612  I->eraseFromParent();
1613 
1614  return true;
1615 }
1616 
1617 namespace {
1618 
1619  // LockstepReverseIterator - Iterates through instructions
1620  // in a set of blocks in reverse order from the first non-terminator.
1621  // For example (assume all blocks have size n):
1622  // LockstepReverseIterator I([B1, B2, B3]);
1623  // *I-- = [B1[n], B2[n], B3[n]];
1624  // *I-- = [B1[n-1], B2[n-1], B3[n-1]];
1625  // *I-- = [B1[n-2], B2[n-2], B3[n-2]];
1626  // ...
1627  class LockstepReverseIterator {
1628  ArrayRef<BasicBlock*> Blocks;
1630  bool Fail;
1631 
1632  public:
1633  LockstepReverseIterator(ArrayRef<BasicBlock*> Blocks) : Blocks(Blocks) {
1634  reset();
1635  }
1636 
1637  void reset() {
1638  Fail = false;
1639  Insts.clear();
1640  for (auto *BB : Blocks) {
1641  Instruction *Inst = BB->getTerminator();
1642  for (Inst = Inst->getPrevNode(); Inst && isa<DbgInfoIntrinsic>(Inst);)
1643  Inst = Inst->getPrevNode();
1644  if (!Inst) {
1645  // Block wasn't big enough.
1646  Fail = true;
1647  return;
1648  }
1649  Insts.push_back(Inst);
1650  }
1651  }
1652 
1653  bool isValid() const {
1654  return !Fail;
1655  }
1656 
1657  void operator--() {
1658  if (Fail)
1659  return;
1660  for (auto *&Inst : Insts) {
1661  for (Inst = Inst->getPrevNode(); Inst && isa<DbgInfoIntrinsic>(Inst);)
1662  Inst = Inst->getPrevNode();
1663  // Already at beginning of block.
1664  if (!Inst) {
1665  Fail = true;
1666  return;
1667  }
1668  }
1669  }
1670 
1672  return Insts;
1673  }
1674  };
1675 
1676 } // end anonymous namespace
1677 
1678 /// Check whether BB's predecessors end with unconditional branches. If it is
1679 /// true, sink any common code from the predecessors to BB.
1680 /// We also allow one predecessor to end with conditional branch (but no more
1681 /// than one).
1683  // We support two situations:
1684  // (1) all incoming arcs are unconditional
1685  // (2) one incoming arc is conditional
1686  //
1687  // (2) is very common in switch defaults and
1688  // else-if patterns;
1689  //
1690  // if (a) f(1);
1691  // else if (b) f(2);
1692  //
1693  // produces:
1694  //
1695  // [if]
1696  // / \
1697  // [f(1)] [if]
1698  // | | \
1699  // | | |
1700  // | [f(2)]|
1701  // \ | /
1702  // [ end ]
1703  //
1704  // [end] has two unconditional predecessor arcs and one conditional. The
1705  // conditional refers to the implicit empty 'else' arc. This conditional
1706  // arc can also be caused by an empty default block in a switch.
1707  //
1708  // In this case, we attempt to sink code from all *unconditional* arcs.
1709  // If we can sink instructions from these arcs (determined during the scan
1710  // phase below) we insert a common successor for all unconditional arcs and
1711  // connect that to [end], to enable sinking:
1712  //
1713  // [if]
1714  // / \
1715  // [x(1)] [if]
1716  // | | \
1717  // | | \
1718  // | [x(2)] |
1719  // \ / |
1720  // [sink.split] |
1721  // \ /
1722  // [ end ]
1723  //
1724  SmallVector<BasicBlock*,4> UnconditionalPreds;
1725  Instruction *Cond = nullptr;
1726  for (auto *B : predecessors(BB)) {
1727  auto *T = B->getTerminator();
1728  if (isa<BranchInst>(T) && cast<BranchInst>(T)->isUnconditional())
1729  UnconditionalPreds.push_back(B);
1730  else if ((isa<BranchInst>(T) || isa<SwitchInst>(T)) && !Cond)
1731  Cond = T;
1732  else
1733  return false;
1734  }
1735  if (UnconditionalPreds.size() < 2)
1736  return false;
1737 
1738  bool Changed = false;
1739  // We take a two-step approach to tail sinking. First we scan from the end of
1740  // each block upwards in lockstep. If the n'th instruction from the end of each
1741  // block can be sunk, those instructions are added to ValuesToSink and we
1742  // carry on. If we can sink an instruction but need to PHI-merge some operands
1743  // (because they're not identical in each instruction) we add these to
1744  // PHIOperands.
1745  unsigned ScanIdx = 0;
1746  SmallPtrSet<Value*,4> InstructionsToSink;
1748  LockstepReverseIterator LRI(UnconditionalPreds);
1749  while (LRI.isValid() &&
1750  canSinkInstructions(*LRI, PHIOperands)) {
1751  LLVM_DEBUG(dbgs() << "SINK: instruction can be sunk: " << *(*LRI)[0]
1752  << "\n");
1753  InstructionsToSink.insert((*LRI).begin(), (*LRI).end());
1754  ++ScanIdx;
1755  --LRI;
1756  }
1757 
1758  auto ProfitableToSinkInstruction = [&](LockstepReverseIterator &LRI) {
1759  unsigned NumPHIdValues = 0;
1760  for (auto *I : *LRI)
1761  for (auto *V : PHIOperands[I])
1762  if (InstructionsToSink.count(V) == 0)
1763  ++NumPHIdValues;
1764  LLVM_DEBUG(dbgs() << "SINK: #phid values: " << NumPHIdValues << "\n");
1765  unsigned NumPHIInsts = NumPHIdValues / UnconditionalPreds.size();
1766  if ((NumPHIdValues % UnconditionalPreds.size()) != 0)
1767  NumPHIInsts++;
1768 
1769  return NumPHIInsts <= 1;
1770  };
1771 
1772  if (ScanIdx > 0 && Cond) {
1773  // Check if we would actually sink anything first! This mutates the CFG and
1774  // adds an extra block. The goal in doing this is to allow instructions that
1775  // couldn't be sunk before to be sunk - obviously, speculatable instructions
1776  // (such as trunc, add) can be sunk and predicated already. So we check that
1777  // we're going to sink at least one non-speculatable instruction.
1778  LRI.reset();
1779  unsigned Idx = 0;
1780  bool Profitable = false;
1781  while (ProfitableToSinkInstruction(LRI) && Idx < ScanIdx) {
1782  if (!isSafeToSpeculativelyExecute((*LRI)[0])) {
1783  Profitable = true;
1784  break;
1785  }
1786  --LRI;
1787  ++Idx;
1788  }
1789  if (!Profitable)
1790  return false;
1791 
1792  LLVM_DEBUG(dbgs() << "SINK: Splitting edge\n");
1793  // We have a conditional edge and we're going to sink some instructions.
1794  // Insert a new block postdominating all blocks we're going to sink from.
1795  if (!SplitBlockPredecessors(BB, UnconditionalPreds, ".sink.split"))
1796  // Edges couldn't be split.
1797  return false;
1798  Changed = true;
1799  }
1800 
1801  // Now that we've analyzed all potential sinking candidates, perform the
1802  // actual sink. We iteratively sink the last non-terminator of the source
1803  // blocks into their common successor unless doing so would require too
1804  // many PHI instructions to be generated (currently only one PHI is allowed
1805  // per sunk instruction).
1806  //
1807  // We can use InstructionsToSink to discount values needing PHI-merging that will
1808  // actually be sunk in a later iteration. This allows us to be more
1809  // aggressive in what we sink. This does allow a false positive where we
1810  // sink presuming a later value will also be sunk, but stop half way through
1811  // and never actually sink it which means we produce more PHIs than intended.
1812  // This is unlikely in practice though.
1813  for (unsigned SinkIdx = 0; SinkIdx != ScanIdx; ++SinkIdx) {
1814  LLVM_DEBUG(dbgs() << "SINK: Sink: "
1815  << *UnconditionalPreds[0]->getTerminator()->getPrevNode()
1816  << "\n");
1817 
1818  // Because we've sunk every instruction in turn, the current instruction to
1819  // sink is always at index 0.
1820  LRI.reset();
1821  if (!ProfitableToSinkInstruction(LRI)) {
1822  // Too many PHIs would be created.
1823  LLVM_DEBUG(
1824  dbgs() << "SINK: stopping here, too many PHIs would be created!\n");
1825  break;
1826  }
1827 
1828  if (!sinkLastInstruction(UnconditionalPreds))
1829  return Changed;
1830  NumSinkCommons++;
1831  Changed = true;
1832  }
1833  return Changed;
1834 }
1835 
1836 /// Determine if we can hoist sink a sole store instruction out of a
1837 /// conditional block.
1838 ///
1839 /// We are looking for code like the following:
1840 /// BrBB:
1841 /// store i32 %add, i32* %arrayidx2
1842 /// ... // No other stores or function calls (we could be calling a memory
1843 /// ... // function).
1844 /// %cmp = icmp ult %x, %y
1845 /// br i1 %cmp, label %EndBB, label %ThenBB
1846 /// ThenBB:
1847 /// store i32 %add5, i32* %arrayidx2
1848 /// br label EndBB
1849 /// EndBB:
1850 /// ...
1851 /// We are going to transform this into:
1852 /// BrBB:
1853 /// store i32 %add, i32* %arrayidx2
1854 /// ... //
1855 /// %cmp = icmp ult %x, %y
1856 /// %add.add5 = select i1 %cmp, i32 %add, %add5
1857 /// store i32 %add.add5, i32* %arrayidx2
1858 /// ...
1859 ///
1860 /// \return The pointer to the value of the previous store if the store can be
1861 /// hoisted into the predecessor block. 0 otherwise.
1863  BasicBlock *StoreBB, BasicBlock *EndBB) {
1864  StoreInst *StoreToHoist = dyn_cast<StoreInst>(I);
1865  if (!StoreToHoist)
1866  return nullptr;
1867 
1868  // Volatile or atomic.
1869  if (!StoreToHoist->isSimple())
1870  return nullptr;
1871 
1872  Value *StorePtr = StoreToHoist->getPointerOperand();
1873 
1874  // Look for a store to the same pointer in BrBB.
1875  unsigned MaxNumInstToLookAt = 9;
1876  for (Instruction &CurI : reverse(BrBB->instructionsWithoutDebug())) {
1877  if (!MaxNumInstToLookAt)
1878  break;
1879  --MaxNumInstToLookAt;
1880 
1881  // Could be calling an instruction that affects memory like free().
1882  if (CurI.mayHaveSideEffects() && !isa<StoreInst>(CurI))
1883  return nullptr;
1884 
1885  if (auto *SI = dyn_cast<StoreInst>(&CurI)) {
1886  // Found the previous store make sure it stores to the same location.
1887  if (SI->getPointerOperand() == StorePtr)
1888  // Found the previous store, return its value operand.
1889  return SI->getValueOperand();
1890  return nullptr; // Unknown store.
1891  }
1892  }
1893 
1894  return nullptr;
1895 }
1896 
1897 /// Speculate a conditional basic block flattening the CFG.
1898 ///
1899 /// Note that this is a very risky transform currently. Speculating
1900 /// instructions like this is most often not desirable. Instead, there is an MI
1901 /// pass which can do it with full awareness of the resource constraints.
1902 /// However, some cases are "obvious" and we should do directly. An example of
1903 /// this is speculating a single, reasonably cheap instruction.
1904 ///
1905 /// There is only one distinct advantage to flattening the CFG at the IR level:
1906 /// it makes very common but simplistic optimizations such as are common in
1907 /// instcombine and the DAG combiner more powerful by removing CFG edges and
1908 /// modeling their effects with easier to reason about SSA value graphs.
1909 ///
1910 ///
1911 /// An illustration of this transform is turning this IR:
1912 /// \code
1913 /// BB:
1914 /// %cmp = icmp ult %x, %y
1915 /// br i1 %cmp, label %EndBB, label %ThenBB
1916 /// ThenBB:
1917 /// %sub = sub %x, %y
1918 /// br label BB2
1919 /// EndBB:
1920 /// %phi = phi [ %sub, %ThenBB ], [ 0, %EndBB ]
1921 /// ...
1922 /// \endcode
1923 ///
1924 /// Into this IR:
1925 /// \code
1926 /// BB:
1927 /// %cmp = icmp ult %x, %y
1928 /// %sub = sub %x, %y
1929 /// %cond = select i1 %cmp, 0, %sub
1930 /// ...
1931 /// \endcode
1932 ///
1933 /// \returns true if the conditional block is removed.
1935  const TargetTransformInfo &TTI) {
1936  // Be conservative for now. FP select instruction can often be expensive.
1937  Value *BrCond = BI->getCondition();
1938  if (isa<FCmpInst>(BrCond))
1939  return false;
1940 
1941  BasicBlock *BB = BI->getParent();
1942  BasicBlock *EndBB = ThenBB->getTerminator()->getSuccessor(0);
1943 
1944  // If ThenBB is actually on the false edge of the conditional branch, remember
1945  // to swap the select operands later.
1946  bool Invert = false;
1947  if (ThenBB != BI->getSuccessor(0)) {
1948  assert(ThenBB == BI->getSuccessor(1) && "No edge from 'if' block?");
1949  Invert = true;
1950  }
1951  assert(EndBB == BI->getSuccessor(!Invert) && "No edge from to end block");
1952 
1953  // Keep a count of how many times instructions are used within ThenBB when
1954  // they are candidates for sinking into ThenBB. Specifically:
1955  // - They are defined in BB, and
1956  // - They have no side effects, and
1957  // - All of their uses are in ThenBB.
1958  SmallDenseMap<Instruction *, unsigned, 4> SinkCandidateUseCounts;
1959 
1960  SmallVector<Instruction *, 4> SpeculatedDbgIntrinsics;
1961 
1962  unsigned SpeculationCost = 0;
1963  Value *SpeculatedStoreValue = nullptr;
1964  StoreInst *SpeculatedStore = nullptr;
1965  for (BasicBlock::iterator BBI = ThenBB->begin(),
1966  BBE = std::prev(ThenBB->end());
1967  BBI != BBE; ++BBI) {
1968  Instruction *I = &*BBI;
1969  // Skip debug info.
1970  if (isa<DbgInfoIntrinsic>(I)) {
1971  SpeculatedDbgIntrinsics.push_back(I);
1972  continue;
1973  }
1974 
1975  // Only speculatively execute a single instruction (not counting the
1976  // terminator) for now.
1977  ++SpeculationCost;
1978  if (SpeculationCost > 1)
1979  return false;
1980 
1981  // Don't hoist the instruction if it's unsafe or expensive.
1982  if (!isSafeToSpeculativelyExecute(I) &&
1983  !(HoistCondStores && (SpeculatedStoreValue = isSafeToSpeculateStore(
1984  I, BB, ThenBB, EndBB))))
1985  return false;
1986  if (!SpeculatedStoreValue &&
1987  ComputeSpeculationCost(I, TTI) >
1989  return false;
1990 
1991  // Store the store speculation candidate.
1992  if (SpeculatedStoreValue)
1993  SpeculatedStore = cast<StoreInst>(I);
1994 
1995  // Do not hoist the instruction if any of its operands are defined but not
1996  // used in BB. The transformation will prevent the operand from
1997  // being sunk into the use block.
1998  for (User::op_iterator i = I->op_begin(), e = I->op_end(); i != e; ++i) {
1999  Instruction *OpI = dyn_cast<Instruction>(*i);
2000  if (!OpI || OpI->getParent() != BB || OpI->mayHaveSideEffects())
2001  continue; // Not a candidate for sinking.
2002 
2003  ++SinkCandidateUseCounts[OpI];
2004  }
2005  }
2006 
2007  // Consider any sink candidates which are only used in ThenBB as costs for
2008  // speculation. Note, while we iterate over a DenseMap here, we are summing
2009  // and so iteration order isn't significant.
2011  I = SinkCandidateUseCounts.begin(),
2012  E = SinkCandidateUseCounts.end();
2013  I != E; ++I)
2014  if (I->first->hasNUses(I->second)) {
2015  ++SpeculationCost;
2016  if (SpeculationCost > 1)
2017  return false;
2018  }
2019 
2020  // Check that the PHI nodes can be converted to selects.
2021  bool HaveRewritablePHIs = false;
2022  for (PHINode &PN : EndBB->phis()) {
2023  Value *OrigV = PN.getIncomingValueForBlock(BB);
2024  Value *ThenV = PN.getIncomingValueForBlock(ThenBB);
2025 
2026  // FIXME: Try to remove some of the duplication with HoistThenElseCodeToIf.
2027  // Skip PHIs which are trivial.
2028  if (ThenV == OrigV)
2029  continue;
2030 
2031  // Don't convert to selects if we could remove undefined behavior instead.
2032  if (passingValueIsAlwaysUndefined(OrigV, &PN) ||
2033  passingValueIsAlwaysUndefined(ThenV, &PN))
2034  return false;
2035 
2036  HaveRewritablePHIs = true;
2037  ConstantExpr *OrigCE = dyn_cast<ConstantExpr>(OrigV);
2038  ConstantExpr *ThenCE = dyn_cast<ConstantExpr>(ThenV);
2039  if (!OrigCE && !ThenCE)
2040  continue; // Known safe and cheap.
2041 
2042  if ((ThenCE && !isSafeToSpeculativelyExecute(ThenCE)) ||
2043  (OrigCE && !isSafeToSpeculativelyExecute(OrigCE)))
2044  return false;
2045  unsigned OrigCost = OrigCE ? ComputeSpeculationCost(OrigCE, TTI) : 0;
2046  unsigned ThenCost = ThenCE ? ComputeSpeculationCost(ThenCE, TTI) : 0;
2047  unsigned MaxCost =
2049  if (OrigCost + ThenCost > MaxCost)
2050  return false;
2051 
2052  // Account for the cost of an unfolded ConstantExpr which could end up
2053  // getting expanded into Instructions.
2054  // FIXME: This doesn't account for how many operations are combined in the
2055  // constant expression.
2056  ++SpeculationCost;
2057  if (SpeculationCost > 1)
2058  return false;
2059  }
2060 
2061  // If there are no PHIs to process, bail early. This helps ensure idempotence
2062  // as well.
2063  if (!HaveRewritablePHIs && !(HoistCondStores && SpeculatedStoreValue))
2064  return false;
2065 
2066  // If we get here, we can hoist the instruction and if-convert.
2067  LLVM_DEBUG(dbgs() << "SPECULATIVELY EXECUTING BB" << *ThenBB << "\n";);
2068 
2069  // Insert a select of the value of the speculated store.
2070  if (SpeculatedStoreValue) {
2071  IRBuilder<NoFolder> Builder(BI);
2072  Value *TrueV = SpeculatedStore->getValueOperand();
2073  Value *FalseV = SpeculatedStoreValue;
2074  if (Invert)
2075  std::swap(TrueV, FalseV);
2076  Value *S = Builder.CreateSelect(
2077  BrCond, TrueV, FalseV, "spec.store.select", BI);
2078  SpeculatedStore->setOperand(0, S);
2079  SpeculatedStore->applyMergedLocation(BI->getDebugLoc(),
2080  SpeculatedStore->getDebugLoc());
2081  }
2082 
2083  // Metadata can be dependent on the condition we are hoisting above.
2084  // Conservatively strip all metadata on the instruction.
2085  for (auto &I : *ThenBB)
2087 
2088  // Hoist the instructions.
2089  BB->getInstList().splice(BI->getIterator(), ThenBB->getInstList(),
2090  ThenBB->begin(), std::prev(ThenBB->end()));
2091 
2092  // Insert selects and rewrite the PHI operands.
2093  IRBuilder<NoFolder> Builder(BI);
2094  for (PHINode &PN : EndBB->phis()) {
2095  unsigned OrigI = PN.getBasicBlockIndex(BB);
2096  unsigned ThenI = PN.getBasicBlockIndex(ThenBB);
2097  Value *OrigV = PN.getIncomingValue(OrigI);
2098  Value *ThenV = PN.getIncomingValue(ThenI);
2099 
2100  // Skip PHIs which are trivial.
2101  if (OrigV == ThenV)
2102  continue;
2103 
2104  // Create a select whose true value is the speculatively executed value and
2105  // false value is the preexisting value. Swap them if the branch
2106  // destinations were inverted.
2107  Value *TrueV = ThenV, *FalseV = OrigV;
2108  if (Invert)
2109  std::swap(TrueV, FalseV);
2110  Value *V = Builder.CreateSelect(
2111  BrCond, TrueV, FalseV, "spec.select", BI);
2112  PN.setIncomingValue(OrigI, V);
2113  PN.setIncomingValue(ThenI, V);
2114  }
2115 
2116  // Remove speculated dbg intrinsics.
2117  // FIXME: Is it possible to do this in a more elegant way? Moving/merging the
2118  // dbg value for the different flows and inserting it after the select.
2119  for (Instruction *I : SpeculatedDbgIntrinsics)
2120  I->eraseFromParent();
2121 
2122  ++NumSpeculations;
2123  return true;
2124 }
2125 
2126 /// Return true if we can thread a branch across this block.
2128  unsigned Size = 0;
2129 
2130  for (Instruction &I : BB->instructionsWithoutDebug()) {
2131  if (Size > 10)
2132  return false; // Don't clone large BB's.
2133  ++Size;
2134 
2135  // We can only support instructions that do not define values that are
2136  // live outside of the current basic block.
2137  for (User *U : I.users()) {
2138  Instruction *UI = cast<Instruction>(U);
2139  if (UI->getParent() != BB || isa<PHINode>(UI))
2140  return false;
2141  }
2142 
2143  // Looks ok, continue checking.
2144  }
2145 
2146  return true;
2147 }
2148 
2149 /// If we have a conditional branch on a PHI node value that is defined in the
2150 /// same block as the branch and if any PHI entries are constants, thread edges
2151 /// corresponding to that entry to be branches to their ultimate destination.
2152 static bool FoldCondBranchOnPHI(BranchInst *BI, const DataLayout &DL,
2153  AssumptionCache *AC) {
2154  BasicBlock *BB = BI->getParent();
2155  PHINode *PN = dyn_cast<PHINode>(BI->getCondition());
2156  // NOTE: we currently cannot transform this case if the PHI node is used
2157  // outside of the block.
2158  if (!PN || PN->getParent() != BB || !PN->hasOneUse())
2159  return false;
2160 
2161  // Degenerate case of a single entry PHI.
2162  if (PN->getNumIncomingValues() == 1) {
2164  return true;
2165  }
2166 
2167  // Now we know that this block has multiple preds and two succs.
2169  return false;
2170 
2171  // Can't fold blocks that contain noduplicate or convergent calls.
2172  if (any_of(*BB, [](const Instruction &I) {
2173  const CallInst *CI = dyn_cast<CallInst>(&I);
2174  return CI && (CI->cannotDuplicate() || CI->isConvergent());
2175  }))
2176  return false;
2177 
2178  // Okay, this is a simple enough basic block. See if any phi values are
2179  // constants.
2180  for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
2182  if (!CB || !CB->getType()->isIntegerTy(1))
2183  continue;
2184 
2185  // Okay, we now know that all edges from PredBB should be revectored to
2186  // branch to RealDest.
2187  BasicBlock *PredBB = PN->getIncomingBlock(i);
2188  BasicBlock *RealDest = BI->getSuccessor(!CB->getZExtValue());
2189 
2190  if (RealDest == BB)
2191  continue; // Skip self loops.
2192  // Skip if the predecessor's terminator is an indirect branch.
2193  if (isa<IndirectBrInst>(PredBB->getTerminator()))
2194  continue;
2195 
2196  // The dest block might have PHI nodes, other predecessors and other
2197  // difficult cases. Instead of being smart about this, just insert a new
2198  // block that jumps to the destination block, effectively splitting
2199  // the edge we are about to create.
2200  BasicBlock *EdgeBB =
2201  BasicBlock::Create(BB->getContext(), RealDest->getName() + ".critedge",
2202  RealDest->getParent(), RealDest);
2203  BranchInst *CritEdgeBranch = BranchInst::Create(RealDest, EdgeBB);
2204  CritEdgeBranch->setDebugLoc(BI->getDebugLoc());
2205 
2206  // Update PHI nodes.
2207  AddPredecessorToBlock(RealDest, EdgeBB, BB);
2208 
2209  // BB may have instructions that are being threaded over. Clone these
2210  // instructions into EdgeBB. We know that there will be no uses of the
2211  // cloned instructions outside of EdgeBB.
2212  BasicBlock::iterator InsertPt = EdgeBB->begin();
2213  DenseMap<Value *, Value *> TranslateMap; // Track translated values.
2214  for (BasicBlock::iterator BBI = BB->begin(); &*BBI != BI; ++BBI) {
2215  if (PHINode *PN = dyn_cast<PHINode>(BBI)) {
2216  TranslateMap[PN] = PN->getIncomingValueForBlock(PredBB);
2217  continue;
2218  }
2219  // Clone the instruction.
2220  Instruction *N = BBI->clone();
2221  if (BBI->hasName())
2222  N->setName(BBI->getName() + ".c");
2223 
2224  // Update operands due to translation.
2225  for (User::op_iterator i = N->op_begin(), e = N->op_end(); i != e; ++i) {
2226  DenseMap<Value *, Value *>::iterator PI = TranslateMap.find(*i);
2227  if (PI != TranslateMap.end())
2228  *i = PI->second;
2229  }
2230 
2231  // Check for trivial simplification.
2232  if (Value *V = SimplifyInstruction(N, {DL, nullptr, nullptr, AC})) {
2233  if (!BBI->use_empty())
2234  TranslateMap[&*BBI] = V;
2235  if (!N->mayHaveSideEffects()) {
2236  N->deleteValue(); // Instruction folded away, don't need actual inst
2237  N = nullptr;
2238  }
2239  } else {
2240  if (!BBI->use_empty())
2241  TranslateMap[&*BBI] = N;
2242  }
2243  // Insert the new instruction into its new home.
2244  if (N)
2245  EdgeBB->getInstList().insert(InsertPt, N);
2246 
2247  // Register the new instruction with the assumption cache if necessary.
2248  if (auto *II = dyn_cast_or_null<IntrinsicInst>(N))
2249  if (II->getIntrinsicID() == Intrinsic::assume)
2250  AC->registerAssumption(II);
2251  }
2252 
2253  // Loop over all of the edges from PredBB to BB, changing them to branch
2254  // to EdgeBB instead.
2255  Instruction *PredBBTI = PredBB->getTerminator();
2256  for (unsigned i = 0, e = PredBBTI->getNumSuccessors(); i != e; ++i)
2257  if (PredBBTI->getSuccessor(i) == BB) {
2258  BB->removePredecessor(PredBB);
2259  PredBBTI->setSuccessor(i, EdgeBB);
2260  }
2261 
2262  // Recurse, simplifying any other constants.
2263  return FoldCondBranchOnPHI(BI, DL, AC) || true;
2264  }
2265 
2266  return false;
2267 }
2268 
2269 /// Given a BB that starts with the specified two-entry PHI node,
2270 /// see if we can eliminate it.
2272  const DataLayout &DL) {
2273  // Ok, this is a two entry PHI node. Check to see if this is a simple "if
2274  // statement", which has a very simple dominance structure. Basically, we
2275  // are trying to find the condition that is being branched on, which
2276  // subsequently causes this merge to happen. We really want control
2277  // dependence information for this check, but simplifycfg can't keep it up
2278  // to date, and this catches most of the cases we care about anyway.
2279  BasicBlock *BB = PN->getParent();
2280  const Function *Fn = BB->getParent();
2281  if (Fn && Fn->hasFnAttribute(Attribute::OptForFuzzing))
2282  return false;
2283 
2284  BasicBlock *IfTrue, *IfFalse;
2285  Value *IfCond = GetIfCondition(BB, IfTrue, IfFalse);
2286  if (!IfCond ||
2287  // Don't bother if the branch will be constant folded trivially.
2288  isa<ConstantInt>(IfCond))
2289  return false;
2290 
2291  // Okay, we found that we can merge this two-entry phi node into a select.
2292  // Doing so would require us to fold *all* two entry phi nodes in this block.
2293  // At some point this becomes non-profitable (particularly if the target
2294  // doesn't support cmov's). Only do this transformation if there are two or
2295  // fewer PHI nodes in this block.
2296  unsigned NumPhis = 0;
2297  for (BasicBlock::iterator I = BB->begin(); isa<PHINode>(I); ++NumPhis, ++I)
2298  if (NumPhis > 2)
2299  return false;
2300 
2301  // Loop over the PHI's seeing if we can promote them all to select
2302  // instructions. While we are at it, keep track of the instructions
2303  // that need to be moved to the dominating block.
2304  SmallPtrSet<Instruction *, 4> AggressiveInsts;
2305  unsigned MaxCostVal0 = PHINodeFoldingThreshold,
2306  MaxCostVal1 = PHINodeFoldingThreshold;
2307  MaxCostVal0 *= TargetTransformInfo::TCC_Basic;
2308  MaxCostVal1 *= TargetTransformInfo::TCC_Basic;
2309 
2310  for (BasicBlock::iterator II = BB->begin(); isa<PHINode>(II);) {
2311  PHINode *PN = cast<PHINode>(II++);
2312  if (Value *V = SimplifyInstruction(PN, {DL, PN})) {
2313  PN->replaceAllUsesWith(V);
2314  PN->eraseFromParent();
2315  continue;
2316  }
2317 
2318  if (!DominatesMergePoint(PN->getIncomingValue(0), BB, AggressiveInsts,
2319  MaxCostVal0, TTI) ||
2320  !DominatesMergePoint(PN->getIncomingValue(1), BB, AggressiveInsts,
2321  MaxCostVal1, TTI))
2322  return false;
2323  }
2324 
2325  // If we folded the first phi, PN dangles at this point. Refresh it. If
2326  // we ran out of PHIs then we simplified them all.
2327  PN = dyn_cast<PHINode>(BB->begin());
2328  if (!PN)
2329  return true;
2330 
2331  // Don't fold i1 branches on PHIs which contain binary operators. These can
2332  // often be turned into switches and other things.
2333  if (PN->getType()->isIntegerTy(1) &&
2334  (isa<BinaryOperator>(PN->getIncomingValue(0)) ||
2335  isa<BinaryOperator>(PN->getIncomingValue(1)) ||
2336  isa<BinaryOperator>(IfCond)))
2337  return false;
2338 
2339  // If all PHI nodes are promotable, check to make sure that all instructions
2340  // in the predecessor blocks can be promoted as well. If not, we won't be able
2341  // to get rid of the control flow, so it's not worth promoting to select
2342  // instructions.
2343  BasicBlock *DomBlock = nullptr;
2344  BasicBlock *IfBlock1 = PN->getIncomingBlock(0);
2345  BasicBlock *IfBlock2 = PN->getIncomingBlock(1);
2346  if (cast<BranchInst>(IfBlock1->getTerminator())->isConditional()) {
2347  IfBlock1 = nullptr;
2348  } else {
2349  DomBlock = *pred_begin(IfBlock1);
2350  for (BasicBlock::iterator I = IfBlock1->begin(); !I->isTerminator(); ++I)
2351  if (!AggressiveInsts.count(&*I) && !isa<DbgInfoIntrinsic>(I)) {
2352  // This is not an aggressive instruction that we can promote.
2353  // Because of this, we won't be able to get rid of the control flow, so
2354  // the xform is not worth it.
2355  return false;
2356  }
2357  }
2358 
2359  if (cast<BranchInst>(IfBlock2->getTerminator())->isConditional()) {
2360  IfBlock2 = nullptr;
2361  } else {
2362  DomBlock = *pred_begin(IfBlock2);
2363  for (BasicBlock::iterator I = IfBlock2->begin(); !I->isTerminator(); ++I)
2364  if (!AggressiveInsts.count(&*I) && !isa<DbgInfoIntrinsic>(I)) {
2365  // This is not an aggressive instruction that we can promote.
2366  // Because of this, we won't be able to get rid of the control flow, so
2367  // the xform is not worth it.
2368  return false;
2369  }
2370  }
2371 
2372  LLVM_DEBUG(dbgs() << "FOUND IF CONDITION! " << *IfCond
2373  << " T: " << IfTrue->getName()
2374  << " F: " << IfFalse->getName() << "\n");
2375 
2376  // If we can still promote the PHI nodes after this gauntlet of tests,
2377  // do all of the PHI's now.
2378  Instruction *InsertPt = DomBlock->getTerminator();
2379  IRBuilder<NoFolder> Builder(InsertPt);
2380 
2381  // Move all 'aggressive' instructions, which are defined in the
2382  // conditional parts of the if's up to the dominating block.
2383  if (IfBlock1)
2384  hoistAllInstructionsInto(DomBlock, InsertPt, IfBlock1);
2385  if (IfBlock2)
2386  hoistAllInstructionsInto(DomBlock, InsertPt, IfBlock2);
2387 
2388  while (PHINode *PN = dyn_cast<PHINode>(BB->begin())) {
2389  // Change the PHI node into a select instruction.
2390  Value *TrueVal = PN->getIncomingValue(PN->getIncomingBlock(0) == IfFalse);
2391  Value *FalseVal = PN->getIncomingValue(PN->getIncomingBlock(0) == IfTrue);
2392 
2393  Value *Sel = Builder.CreateSelect(IfCond, TrueVal, FalseVal, "", InsertPt);
2394  PN->replaceAllUsesWith(Sel);
2395  Sel->takeName(PN);
2396  PN->eraseFromParent();
2397  }
2398 
2399  // At this point, IfBlock1 and IfBlock2 are both empty, so our if statement
2400  // has been flattened. Change DomBlock to jump directly to our new block to
2401  // avoid other simplifycfg's kicking in on the diamond.
2402  Instruction *OldTI = DomBlock->getTerminator();
2403  Builder.SetInsertPoint(OldTI);
2404  Builder.CreateBr(BB);
2405  OldTI->eraseFromParent();
2406  return true;
2407 }
2408 
2409 /// If we found a conditional branch that goes to two returning blocks,
2410 /// try to merge them together into one return,
2411 /// introducing a select if the return values disagree.
2413  IRBuilder<> &Builder) {
2414  assert(BI->isConditional() && "Must be a conditional branch");
2415  BasicBlock *TrueSucc = BI->getSuccessor(0);
2416  BasicBlock *FalseSucc = BI->getSuccessor(1);
2417  ReturnInst *TrueRet = cast<ReturnInst>(TrueSucc->getTerminator());
2418  ReturnInst *FalseRet = cast<ReturnInst>(FalseSucc->getTerminator());
2419 
2420  // Check to ensure both blocks are empty (just a return) or optionally empty
2421  // with PHI nodes. If there are other instructions, merging would cause extra
2422  // computation on one path or the other.
2423  if (!TrueSucc->getFirstNonPHIOrDbg()->isTerminator())
2424  return false;
2425  if (!FalseSucc->getFirstNonPHIOrDbg()->isTerminator())
2426  return false;
2427 
2428  Builder.SetInsertPoint(BI);
2429  // Okay, we found a branch that is going to two return nodes. If
2430  // there is no return value for this function, just change the
2431  // branch into a return.
2432  if (FalseRet->getNumOperands() == 0) {
2433  TrueSucc->removePredecessor(BI->getParent());
2434  FalseSucc->removePredecessor(BI->getParent());
2435  Builder.CreateRetVoid();
2437  return true;
2438  }
2439 
2440  // Otherwise, figure out what the true and false return values are
2441  // so we can insert a new select instruction.
2442  Value *TrueValue = TrueRet->getReturnValue();
2443  Value *FalseValue = FalseRet->getReturnValue();
2444 
2445  // Unwrap any PHI nodes in the return blocks.
2446  if (PHINode *TVPN = dyn_cast_or_null<PHINode>(TrueValue))
2447  if (TVPN->getParent() == TrueSucc)
2448  TrueValue = TVPN->getIncomingValueForBlock(BI->getParent());
2449  if (PHINode *FVPN = dyn_cast_or_null<PHINode>(FalseValue))
2450  if (FVPN->getParent() == FalseSucc)
2451  FalseValue = FVPN->getIncomingValueForBlock(BI->getParent());
2452 
2453  // In order for this transformation to be safe, we must be able to
2454  // unconditionally execute both operands to the return. This is
2455  // normally the case, but we could have a potentially-trapping
2456  // constant expression that prevents this transformation from being
2457  // safe.
2458  if (ConstantExpr *TCV = dyn_cast_or_null<ConstantExpr>(TrueValue))
2459  if (TCV->canTrap())
2460  return false;
2461  if (ConstantExpr *FCV = dyn_cast_or_null<ConstantExpr>(FalseValue))
2462  if (FCV->canTrap())
2463  return false;
2464 
2465  // Okay, we collected all the mapped values and checked them for sanity, and
2466  // defined to really do this transformation. First, update the CFG.
2467  TrueSucc->removePredecessor(BI->getParent());
2468  FalseSucc->removePredecessor(BI->getParent());
2469 
2470  // Insert select instructions where needed.
2471  Value *BrCond = BI->getCondition();
2472  if (TrueValue) {
2473  // Insert a select if the results differ.
2474  if (TrueValue == FalseValue || isa<UndefValue>(FalseValue)) {
2475  } else if (isa<UndefValue>(TrueValue)) {
2476  TrueValue = FalseValue;
2477  } else {
2478  TrueValue =
2479  Builder.CreateSelect(BrCond, TrueValue, FalseValue, "retval", BI);
2480  }
2481  }
2482 
2483  Value *RI =
2484  !TrueValue ? Builder.CreateRetVoid() : Builder.CreateRet(TrueValue);
2485 
2486  (void)RI;
2487 
2488  LLVM_DEBUG(dbgs() << "\nCHANGING BRANCH TO TWO RETURNS INTO SELECT:"
2489  << "\n " << *BI << "NewRet = " << *RI << "TRUEBLOCK: "
2490  << *TrueSucc << "FALSEBLOCK: " << *FalseSucc);
2491 
2493 
2494  return true;
2495 }
2496 
2497 /// Return true if the given instruction is available
2498 /// in its predecessor block. If yes, the instruction will be removed.
2500  if (!isa<BinaryOperator>(Inst) && !isa<CmpInst>(Inst))
2501  return false;
2502  for (Instruction &I : *PB) {
2503  Instruction *PBI = &I;
2504  // Check whether Inst and PBI generate the same value.
2505  if (Inst->isIdenticalTo(PBI)) {
2506  Inst->replaceAllUsesWith(PBI);
2507  Inst->eraseFromParent();
2508  return true;
2509  }
2510  }
2511  return false;
2512 }
2513 
2514 /// Return true if either PBI or BI has branch weight available, and store
2515 /// the weights in {Pred|Succ}{True|False}Weight. If one of PBI and BI does
2516 /// not have branch weight, use 1:1 as its weight.
2518  uint64_t &PredTrueWeight,
2519  uint64_t &PredFalseWeight,
2520  uint64_t &SuccTrueWeight,
2521  uint64_t &SuccFalseWeight) {
2522  bool PredHasWeights =
2523  PBI->extractProfMetadata(PredTrueWeight, PredFalseWeight);
2524  bool SuccHasWeights =
2525  BI->extractProfMetadata(SuccTrueWeight, SuccFalseWeight);
2526  if (PredHasWeights || SuccHasWeights) {
2527  if (!PredHasWeights)
2528  PredTrueWeight = PredFalseWeight = 1;
2529  if (!SuccHasWeights)
2530  SuccTrueWeight = SuccFalseWeight = 1;
2531  return true;
2532  } else {
2533  return false;
2534  }
2535 }
2536 
2537 /// If this basic block is simple enough, and if a predecessor branches to us
2538 /// and one of our successors, fold the block into the predecessor and use
2539 /// logical operations to pick the right destination.
2541  unsigned BonusInstThreshold) {
2542  BasicBlock *BB = BI->getParent();
2543 
2544  const unsigned PredCount = pred_size(BB);
2545 
2546  Instruction *Cond = nullptr;
2547  if (BI->isConditional())
2548  Cond = dyn_cast<Instruction>(BI->getCondition());
2549  else {
2550  // For unconditional branch, check for a simple CFG pattern, where
2551  // BB has a single predecessor and BB's successor is also its predecessor's
2552  // successor. If such pattern exists, check for CSE between BB and its
2553  // predecessor.
2554  if (BasicBlock *PB = BB->getSinglePredecessor())
2555  if (BranchInst *PBI = dyn_cast<BranchInst>(PB->getTerminator()))
2556  if (PBI->isConditional() &&
2557  (BI->getSuccessor(0) == PBI->getSuccessor(0) ||
2558  BI->getSuccessor(0) == PBI->getSuccessor(1))) {
2559  for (auto I = BB->instructionsWithoutDebug().begin(),
2560  E = BB->instructionsWithoutDebug().end();
2561  I != E;) {
2562  Instruction *Curr = &*I++;
2563  if (isa<CmpInst>(Curr)) {
2564  Cond = Curr;
2565  break;
2566  }
2567  // Quit if we can't remove this instruction.
2568  if (!tryCSEWithPredecessor(Curr, PB))
2569  return false;
2570  }
2571  }
2572 
2573  if (!Cond)
2574  return false;
2575  }
2576 
2577  if (!Cond || (!isa<CmpInst>(Cond) && !isa<BinaryOperator>(Cond)) ||
2578  Cond->getParent() != BB || !Cond->hasOneUse())
2579  return false;
2580 
2581  // Make sure the instruction after the condition is the cond branch.
2582  BasicBlock::iterator CondIt = ++Cond->getIterator();
2583 
2584  // Ignore dbg intrinsics.
2585  while (isa<DbgInfoIntrinsic>(CondIt))
2586  ++CondIt;
2587 
2588  if (&*CondIt != BI)
2589  return false;
2590 
2591  // Only allow this transformation if computing the condition doesn't involve
2592  // too many instructions and these involved instructions can be executed
2593  // unconditionally. We denote all involved instructions except the condition
2594  // as "bonus instructions", and only allow this transformation when the
2595  // number of the bonus instructions we'll need to create when cloning into
2596  // each predecessor does not exceed a certain threshold.
2597  unsigned NumBonusInsts = 0;
2598  for (auto I = BB->begin(); Cond != &*I; ++I) {
2599  // Ignore dbg intrinsics.
2600  if (isa<DbgInfoIntrinsic>(I))
2601  continue;
2602  if (!I->hasOneUse() || !isSafeToSpeculativelyExecute(&*I))
2603  return false;
2604  // I has only one use and can be executed unconditionally.
2606  if (User == nullptr || User->getParent() != BB)
2607  return false;
2608  // I is used in the same BB. Since BI uses Cond and doesn't have more slots
2609  // to use any other instruction, User must be an instruction between next(I)
2610  // and Cond.
2611 
2612  // Account for the cost of duplicating this instruction into each
2613  // predecessor.
2614  NumBonusInsts += PredCount;
2615  // Early exits once we reach the limit.
2616  if (NumBonusInsts > BonusInstThreshold)
2617  return false;
2618  }
2619 
2620  // Cond is known to be a compare or binary operator. Check to make sure that
2621  // neither operand is a potentially-trapping constant expression.
2622  if (ConstantExpr *CE = dyn_cast<ConstantExpr>(Cond->getOperand(0)))
2623  if (CE->canTrap())
2624  return false;
2625  if (ConstantExpr *CE = dyn_cast<ConstantExpr>(Cond->getOperand(1)))
2626  if (CE->canTrap())
2627  return false;
2628 
2629  // Finally, don't infinitely unroll conditional loops.
2630  BasicBlock *TrueDest = BI->getSuccessor(0);
2631  BasicBlock *FalseDest = (BI->isConditional()) ? BI->getSuccessor(1) : nullptr;
2632  if (TrueDest == BB || FalseDest == BB)
2633  return false;
2634 
2635  for (pred_iterator PI = pred_begin(BB), E = pred_end(BB); PI != E; ++PI) {
2636  BasicBlock *PredBlock = *PI;
2637  BranchInst *PBI = dyn_cast<BranchInst>(PredBlock->getTerminator());
2638 
2639  // Check that we have two conditional branches. If there is a PHI node in
2640  // the common successor, verify that the same value flows in from both
2641  // blocks.
2643  if (!PBI || PBI->isUnconditional() ||
2644  (BI->isConditional() && !SafeToMergeTerminators(BI, PBI)) ||
2645  (!BI->isConditional() &&
2646  !isProfitableToFoldUnconditional(BI, PBI, Cond, PHIs)))
2647  continue;
2648 
2649  // Determine if the two branches share a common destination.
2650  Instruction::BinaryOps Opc = Instruction::BinaryOpsEnd;
2651  bool InvertPredCond = false;
2652 
2653  if (BI->isConditional()) {
2654  if (PBI->getSuccessor(0) == TrueDest) {
2655  Opc = Instruction::Or;
2656  } else if (PBI->getSuccessor(1) == FalseDest) {
2657  Opc = Instruction::And;
2658  } else if (PBI->getSuccessor(0) == FalseDest) {
2659  Opc = Instruction::And;
2660  InvertPredCond = true;
2661  } else if (PBI->getSuccessor(1) == TrueDest) {
2662  Opc = Instruction::Or;
2663  InvertPredCond = true;
2664  } else {
2665  continue;
2666  }
2667  } else {
2668  if (PBI->getSuccessor(0) != TrueDest && PBI->getSuccessor(1) != TrueDest)
2669  continue;
2670  }
2671 
2672  LLVM_DEBUG(dbgs() << "FOLDING BRANCH TO COMMON DEST:\n" << *PBI << *BB);
2673  IRBuilder<> Builder(PBI);
2674 
2675  // If we need to invert the condition in the pred block to match, do so now.
2676  if (InvertPredCond) {
2677  Value *NewCond = PBI->getCondition();
2678 
2679  if (NewCond->hasOneUse() && isa<CmpInst>(NewCond)) {
2680  CmpInst *CI = cast<CmpInst>(NewCond);
2681  CI->setPredicate(CI->getInversePredicate());
2682  } else {
2683  NewCond =
2684  Builder.CreateNot(NewCond, PBI->getCondition()->getName() + ".not");
2685  }
2686 
2687  PBI->setCondition(NewCond);
2688  PBI->swapSuccessors();
2689  }
2690 
2691  // If we have bonus instructions, clone them into the predecessor block.
2692  // Note that there may be multiple predecessor blocks, so we cannot move
2693  // bonus instructions to a predecessor block.
2694  ValueToValueMapTy VMap; // maps original values to cloned values
2695  // We already make sure Cond is the last instruction before BI. Therefore,
2696  // all instructions before Cond other than DbgInfoIntrinsic are bonus
2697  // instructions.
2698  for (auto BonusInst = BB->begin(); Cond != &*BonusInst; ++BonusInst) {
2699  if (isa<DbgInfoIntrinsic>(BonusInst))
2700  continue;
2701  Instruction *NewBonusInst = BonusInst->clone();
2702  RemapInstruction(NewBonusInst, VMap,
2704  VMap[&*BonusInst] = NewBonusInst;
2705 
2706  // If we moved a load, we cannot any longer claim any knowledge about
2707  // its potential value. The previous information might have been valid
2708  // only given the branch precondition.
2709  // For an analogous reason, we must also drop all the metadata whose
2710  // semantics we don't understand.
2711  NewBonusInst->dropUnknownNonDebugMetadata();
2712 
2713  PredBlock->getInstList().insert(PBI->getIterator(), NewBonusInst);
2714  NewBonusInst->takeName(&*BonusInst);
2715  BonusInst->setName(BonusInst->getName() + ".old");
2716  }
2717 
2718  // Clone Cond into the predecessor basic block, and or/and the
2719  // two conditions together.
2720  Instruction *CondInPred = Cond->clone();
2721  RemapInstruction(CondInPred, VMap,
2723  PredBlock->getInstList().insert(PBI->getIterator(), CondInPred);
2724  CondInPred->takeName(Cond);
2725  Cond->setName(CondInPred->getName() + ".old");
2726 
2727  if (BI->isConditional()) {
2728  Instruction *NewCond = cast<Instruction>(
2729  Builder.CreateBinOp(Opc, PBI->getCondition(), CondInPred, "or.cond"));
2730  PBI->setCondition(NewCond);
2731 
2732  uint64_t PredTrueWeight, PredFalseWeight, SuccTrueWeight, SuccFalseWeight;
2733  bool HasWeights =
2734  extractPredSuccWeights(PBI, BI, PredTrueWeight, PredFalseWeight,
2735  SuccTrueWeight, SuccFalseWeight);
2736  SmallVector<uint64_t, 8> NewWeights;
2737 
2738  if (PBI->getSuccessor(0) == BB) {
2739  if (HasWeights) {
2740  // PBI: br i1 %x, BB, FalseDest
2741  // BI: br i1 %y, TrueDest, FalseDest
2742  // TrueWeight is TrueWeight for PBI * TrueWeight for BI.
2743  NewWeights.push_back(PredTrueWeight * SuccTrueWeight);
2744  // FalseWeight is FalseWeight for PBI * TotalWeight for BI +
2745  // TrueWeight for PBI * FalseWeight for BI.
2746  // We assume that total weights of a BranchInst can fit into 32 bits.
2747  // Therefore, we will not have overflow using 64-bit arithmetic.
2748  NewWeights.push_back(PredFalseWeight *
2749  (SuccFalseWeight + SuccTrueWeight) +
2750  PredTrueWeight * SuccFalseWeight);
2751  }
2752  AddPredecessorToBlock(TrueDest, PredBlock, BB, MSSAU);
2753  PBI->setSuccessor(0, TrueDest);
2754  }
2755  if (PBI->getSuccessor(1) == BB) {
2756  if (HasWeights) {
2757  // PBI: br i1 %x, TrueDest, BB
2758  // BI: br i1 %y, TrueDest, FalseDest
2759  // TrueWeight is TrueWeight for PBI * TotalWeight for BI +
2760  // FalseWeight for PBI * TrueWeight for BI.
2761  NewWeights.push_back(PredTrueWeight *
2762  (SuccFalseWeight + SuccTrueWeight) +
2763  PredFalseWeight * SuccTrueWeight);
2764  // FalseWeight is FalseWeight for PBI * FalseWeight for BI.
2765  NewWeights.push_back(PredFalseWeight * SuccFalseWeight);
2766  }
2767  AddPredecessorToBlock(FalseDest, PredBlock, BB, MSSAU);
2768  PBI->setSuccessor(1, FalseDest);
2769  }
2770  if (NewWeights.size() == 2) {
2771  // Halve the weights if any of them cannot fit in an uint32_t
2772  FitWeights(NewWeights);
2773 
2774  SmallVector<uint32_t, 8> MDWeights(NewWeights.begin(),
2775  NewWeights.end());
2776  setBranchWeights(PBI, MDWeights[0], MDWeights[1]);
2777  } else
2778  PBI->setMetadata(LLVMContext::MD_prof, nullptr);
2779  } else {
2780  // Update PHI nodes in the common successors.
2781  for (unsigned i = 0, e = PHIs.size(); i != e; ++i) {
2782  ConstantInt *PBI_C = cast<ConstantInt>(
2783  PHIs[i]->getIncomingValueForBlock(PBI->getParent()));
2784  assert(PBI_C->getType()->isIntegerTy(1));
2785  Instruction *MergedCond = nullptr;
2786  if (PBI->getSuccessor(0) == TrueDest) {
2787  // Create (PBI_Cond and PBI_C) or (!PBI_Cond and BI_Value)
2788  // PBI_C is true: PBI_Cond or (!PBI_Cond and BI_Value)
2789  // is false: !PBI_Cond and BI_Value
2790  Instruction *NotCond = cast<Instruction>(
2791  Builder.CreateNot(PBI->getCondition(), "not.cond"));
2792  MergedCond = cast<Instruction>(
2793  Builder.CreateBinOp(Instruction::And, NotCond, CondInPred,
2794  "and.cond"));
2795  if (PBI_C->isOne())
2796  MergedCond = cast<Instruction>(Builder.CreateBinOp(
2797  Instruction::Or, PBI->getCondition(), MergedCond, "or.cond"));
2798  } else {
2799  // Create (PBI_Cond and BI_Value) or (!PBI_Cond and PBI_C)
2800  // PBI_C is true: (PBI_Cond and BI_Value) or (!PBI_Cond)
2801  // is false: PBI_Cond and BI_Value
2802  MergedCond = cast<Instruction>(Builder.CreateBinOp(
2803  Instruction::And, PBI->getCondition(), CondInPred, "and.cond"));
2804  if (PBI_C->isOne()) {
2805  Instruction *NotCond = cast<Instruction>(
2806  Builder.CreateNot(PBI->getCondition(), "not.cond"));
2807  MergedCond = cast<Instruction>(Builder.CreateBinOp(
2808  Instruction::Or, NotCond, MergedCond, "or.cond"));
2809  }
2810  }
2811  // Update PHI Node.
2812  PHIs[i]->setIncomingValueForBlock(PBI->getParent(), MergedCond);
2813  }
2814 
2815  // PBI is changed to branch to TrueDest below. Remove itself from
2816  // potential phis from all other successors.
2817  if (MSSAU)
2818  MSSAU->changeCondBranchToUnconditionalTo(PBI, TrueDest);
2819 
2820  // Change PBI from Conditional to Unconditional.
2821  BranchInst *New_PBI = BranchInst::Create(TrueDest, PBI);
2822  EraseTerminatorAndDCECond(PBI, MSSAU);
2823  PBI = New_PBI;
2824  }
2825 
2826  // If BI was a loop latch, it may have had associated loop metadata.
2827  // We need to copy it to the new latch, that is, PBI.
2828  if (MDNode *LoopMD = BI->getMetadata(LLVMContext::MD_loop))
2829  PBI->setMetadata(LLVMContext::MD_loop, LoopMD);
2830 
2831  // TODO: If BB is reachable from all paths through PredBlock, then we
2832  // could replace PBI's branch probabilities with BI's.
2833 
2834  // Copy any debug value intrinsics into the end of PredBlock.
2835  for (Instruction &I : *BB)
2836  if (isa<DbgInfoIntrinsic>(I))
2837  I.clone()->insertBefore(PBI);
2838 
2839  return true;
2840  }
2841  return false;
2842 }
2843 
2844 // If there is only one store in BB1 and BB2, return it, otherwise return
2845 // nullptr.
2847  StoreInst *S = nullptr;
2848  for (auto *BB : {BB1, BB2}) {
2849  if (!BB)
2850  continue;
2851  for (auto &I : *BB)
2852  if (auto *SI = dyn_cast<StoreInst>(&I)) {
2853  if (S)
2854  // Multiple stores seen.
2855  return nullptr;
2856  else
2857  S = SI;
2858  }
2859  }
2860  return S;
2861 }
2862 
2864  Value *AlternativeV = nullptr) {
2865  // PHI is going to be a PHI node that allows the value V that is defined in
2866  // BB to be referenced in BB's only successor.
2867  //
2868  // If AlternativeV is nullptr, the only value we care about in PHI is V. It
2869  // doesn't matter to us what the other operand is (it'll never get used). We
2870  // could just create a new PHI with an undef incoming value, but that could
2871  // increase register pressure if EarlyCSE/InstCombine can't fold it with some
2872  // other PHI. So here we directly look for some PHI in BB's successor with V
2873  // as an incoming operand. If we find one, we use it, else we create a new
2874  // one.
2875  //
2876  // If AlternativeV is not nullptr, we care about both incoming values in PHI.
2877  // PHI must be exactly: phi <ty> [ %BB, %V ], [ %OtherBB, %AlternativeV]
2878  // where OtherBB is the single other predecessor of BB's only successor.
2879  PHINode *PHI = nullptr;
2880  BasicBlock *Succ = BB->getSingleSuccessor();
2881 
2882  for (auto I = Succ->begin(); isa<PHINode>(I); ++I)
2883  if (cast<PHINode>(I)->getIncomingValueForBlock(BB) == V) {
2884  PHI = cast<PHINode>(I);
2885  if (!AlternativeV)
2886  break;
2887 
2888  assert(Succ->hasNPredecessors(2));
2889  auto PredI = pred_begin(Succ);
2890  BasicBlock *OtherPredBB = *PredI == BB ? *++PredI : *PredI;
2891  if (PHI->getIncomingValueForBlock(OtherPredBB) == AlternativeV)
2892  break;
2893  PHI = nullptr;
2894  }
2895  if (PHI)
2896  return PHI;
2897 
2898  // If V is not an instruction defined in BB, just return it.
2899  if (!AlternativeV &&
2900  (!isa<Instruction>(V) || cast<Instruction>(V)->getParent() != BB))
2901  return V;
2902 
2903  PHI = PHINode::Create(V->getType(), 2, "simplifycfg.merge", &Succ->front());
2904  PHI->addIncoming(V, BB);
2905  for (BasicBlock *PredBB : predecessors(Succ))
2906  if (PredBB != BB)
2907  PHI->addIncoming(
2908  AlternativeV ? AlternativeV : UndefValue::get(V->getType()), PredBB);
2909  return PHI;
2910 }
2911 
2913  BasicBlock *QTB, BasicBlock *QFB,
2914  BasicBlock *PostBB, Value *Address,
2915  bool InvertPCond, bool InvertQCond,
2916  const DataLayout &DL) {
2917  auto IsaBitcastOfPointerType = [](const Instruction &I) {
2918  return Operator::getOpcode(&I) == Instruction::BitCast &&
2919  I.getType()->isPointerTy();
2920  };
2921 
2922  // If we're not in aggressive mode, we only optimize if we have some
2923  // confidence that by optimizing we'll allow P and/or Q to be if-converted.
2924  auto IsWorthwhile = [&](BasicBlock *BB) {
2925  if (!BB)
2926  return true;
2927  // Heuristic: if the block can be if-converted/phi-folded and the
2928  // instructions inside are all cheap (arithmetic/GEPs), it's worthwhile to
2929  // thread this store.
2930  unsigned N = 0;
2931  for (auto &I : BB->instructionsWithoutDebug()) {
2932  // Cheap instructions viable for folding.
2933  if (isa<BinaryOperator>(I) || isa<GetElementPtrInst>(I) ||
2934  isa<StoreInst>(I))
2935  ++N;
2936  // Free instructions.
2937  else if (I.isTerminator() || IsaBitcastOfPointerType(I))
2938  continue;
2939  else
2940  return false;
2941  }
2942  // The store we want to merge is counted in N, so add 1 to make sure
2943  // we're counting the instructions that would be left.
2944  return N <= (PHINodeFoldingThreshold + 1);
2945  };
2946 
2948  (!IsWorthwhile(PTB) || !IsWorthwhile(PFB) || !IsWorthwhile(QTB) ||
2949  !IsWorthwhile(QFB)))
2950  return false;
2951 
2952  // For every pointer, there must be exactly two stores, one coming from
2953  // PTB or PFB, and the other from QTB or QFB. We don't support more than one
2954  // store (to any address) in PTB,PFB or QTB,QFB.
2955  // FIXME: We could relax this restriction with a bit more work and performance
2956  // testing.
2957  StoreInst *PStore = findUniqueStoreInBlocks(PTB, PFB);
2958  StoreInst *QStore = findUniqueStoreInBlocks(QTB, QFB);
2959  if (!PStore || !QStore)
2960  return false;
2961 
2962  // Now check the stores are compatible.
2963  if (!QStore->isUnordered() || !PStore->isUnordered())
2964  return false;
2965 
2966  // Check that sinking the store won't cause program behavior changes. Sinking
2967  // the store out of the Q blocks won't change any behavior as we're sinking
2968  // from a block to its unconditional successor. But we're moving a store from
2969  // the P blocks down through the middle block (QBI) and past both QFB and QTB.
2970  // So we need to check that there are no aliasing loads or stores in
2971  // QBI, QTB and QFB. We also need to check there are no conflicting memory
2972  // operations between PStore and the end of its parent block.
2973  //
2974  // The ideal way to do this is to query AliasAnalysis, but we don't
2975  // preserve AA currently so that is dangerous. Be super safe and just
2976  // check there are no other memory operations at all.
2977  for (auto &I : *QFB->getSinglePredecessor())
2978  if (I.mayReadOrWriteMemory())
2979  return false;
2980  for (auto &I : *QFB)
2981  if (&I != QStore && I.mayReadOrWriteMemory())
2982  return false;
2983  if (QTB)
2984  for (auto &I : *QTB)
2985  if (&I != QStore && I.mayReadOrWriteMemory())
2986  return false;
2987  for (auto I = BasicBlock::iterator(PStore), E = PStore->getParent()->end();
2988  I != E; ++I)
2989  if (&*I != PStore && I->mayReadOrWriteMemory())
2990  return false;
2991 
2992  // If PostBB has more than two predecessors, we need to split it so we can
2993  // sink the store.
2994  if (std::next(pred_begin(PostBB), 2) != pred_end(PostBB)) {
2995  // We know that QFB's only successor is PostBB. And QFB has a single
2996  // predecessor. If QTB exists, then its only successor is also PostBB.
2997  // If QTB does not exist, then QFB's only predecessor has a conditional
2998  // branch to QFB and PostBB.
2999  BasicBlock *TruePred = QTB ? QTB : QFB->getSinglePredecessor();
3000  BasicBlock *NewBB = SplitBlockPredecessors(PostBB, { QFB, TruePred},
3001  "condstore.split");
3002  if (!NewBB)
3003  return false;
3004  PostBB = NewBB;
3005  }
3006 
3007  // OK, we're going to sink the stores to PostBB. The store has to be
3008  // conditional though, so first create the predicate.
3009  Value *PCond = cast<BranchInst>(PFB->getSinglePredecessor()->getTerminator())
3010  ->getCondition();
3011  Value *QCond = cast<BranchInst>(QFB->getSinglePredecessor()->getTerminator())
3012  ->getCondition();
3013 
3015  PStore->getParent());
3017  QStore->getParent(), PPHI);
3018 
3019  IRBuilder<> QB(&*PostBB->getFirstInsertionPt());
3020 
3021  Value *PPred = PStore->getParent() == PTB ? PCond : QB.CreateNot(PCond);
3022  Value *QPred = QStore->getParent() == QTB ? QCond : QB.CreateNot(QCond);
3023 
3024  if (InvertPCond)
3025  PPred = QB.CreateNot(PPred);
3026  if (InvertQCond)
3027  QPred = QB.CreateNot(QPred);
3028  Value *CombinedPred = QB.CreateOr(PPred, QPred);
3029 
3030  auto *T =
3031  SplitBlockAndInsertIfThen(CombinedPred, &*QB.GetInsertPoint(), false);
3032  QB.SetInsertPoint(T);
3033  StoreInst *SI = cast<StoreInst>(QB.CreateStore(QPHI, Address));
3034  AAMDNodes AAMD;
3035  PStore->getAAMetadata(AAMD, /*Merge=*/false);
3036  PStore->getAAMetadata(AAMD, /*Merge=*/true);
3037  SI->setAAMetadata(AAMD);
3038  unsigned PAlignment = PStore->getAlignment();
3039  unsigned QAlignment = QStore->getAlignment();
3040  unsigned TypeAlignment =
3042  unsigned MinAlignment;
3043  unsigned MaxAlignment;
3044  std::tie(MinAlignment, MaxAlignment) = std::minmax(PAlignment, QAlignment);
3045  // Choose the minimum alignment. If we could prove both stores execute, we
3046  // could use biggest one. In this case, though, we only know that one of the
3047  // stores executes. And we don't know it's safe to take the alignment from a
3048  // store that doesn't execute.
3049  if (MinAlignment != 0) {
3050  // Choose the minimum of all non-zero alignments.
3051  SI->setAlignment(MinAlignment);
3052  } else if (MaxAlignment != 0) {
3053  // Choose the minimal alignment between the non-zero alignment and the ABI
3054  // default alignment for the type of the stored value.
3055  SI->setAlignment(std::min(MaxAlignment, TypeAlignment));
3056  } else {
3057  // If both alignments are zero, use ABI default alignment for the type of
3058  // the stored value.
3059  SI->setAlignment(TypeAlignment);
3060  }
3061 
3062  QStore->eraseFromParent();
3063  PStore->eraseFromParent();
3064 
3065  return true;
3066 }
3067 
3069  const DataLayout &DL) {
3070  // The intention here is to find diamonds or triangles (see below) where each
3071  // conditional block contains a store to the same address. Both of these
3072  // stores are conditional, so they can't be unconditionally sunk. But it may
3073  // be profitable to speculatively sink the stores into one merged store at the
3074  // end, and predicate the merged store on the union of the two conditions of
3075  // PBI and QBI.
3076  //
3077  // This can reduce the number of stores executed if both of the conditions are
3078  // true, and can allow the blocks to become small enough to be if-converted.
3079  // This optimization will also chain, so that ladders of test-and-set
3080  // sequences can be if-converted away.
3081  //
3082  // We only deal with simple diamonds or triangles:
3083  //
3084  // PBI or PBI or a combination of the two
3085  // / \ | \
3086  // PTB PFB | PFB
3087  // \ / | /
3088  // QBI QBI
3089  // / \ | \
3090  // QTB QFB | QFB
3091  // \ / | /
3092  // PostBB PostBB
3093  //
3094  // We model triangles as a type of diamond with a nullptr "true" block.
3095  // Triangles are canonicalized so that the fallthrough edge is represented by
3096  // a true condition, as in the diagram above.
3097  BasicBlock *PTB = PBI->getSuccessor(0);
3098  BasicBlock *PFB = PBI->getSuccessor(1);
3099  BasicBlock *QTB = QBI->getSuccessor(0);
3100  BasicBlock *QFB = QBI->getSuccessor(1);
3101  BasicBlock *PostBB = QFB->getSingleSuccessor();
3102 
3103  // Make sure we have a good guess for PostBB. If QTB's only successor is
3104  // QFB, then QFB is a better PostBB.
3105  if (QTB->getSingleSuccessor() == QFB)
3106  PostBB = QFB;
3107 
3108  // If we couldn't find a good PostBB, stop.
3109  if (!PostBB)
3110  return false;
3111 
3112  bool InvertPCond = false, InvertQCond = false;
3113  // Canonicalize fallthroughs to the true branches.
3114  if (PFB == QBI->getParent()) {
3115  std::swap(PFB, PTB);
3116  InvertPCond = true;
3117  }
3118  if (QFB == PostBB) {
3119  std::swap(QFB, QTB);
3120  InvertQCond = true;
3121  }
3122 
3123  // From this point on we can assume PTB or QTB may be fallthroughs but PFB
3124  // and QFB may not. Model fallthroughs as a nullptr block.
3125  if (PTB == QBI->getParent())
3126  PTB = nullptr;
3127  if (QTB == PostBB)
3128  QTB = nullptr;
3129 
3130  // Legality bailouts. We must have at least the non-fallthrough blocks and
3131  // the post-dominating block, and the non-fallthroughs must only have one
3132  // predecessor.
3133  auto HasOnePredAndOneSucc = [](BasicBlock *BB, BasicBlock *P, BasicBlock *S) {
3134  return BB->getSinglePredecessor() == P && BB->getSingleSuccessor() == S;
3135  };
3136  if (!HasOnePredAndOneSucc(PFB, PBI->getParent(), QBI->getParent()) ||
3137  !HasOnePredAndOneSucc(QFB, QBI->getParent(), PostBB))
3138  return false;
3139  if ((PTB && !HasOnePredAndOneSucc(PTB, PBI->getParent(), QBI->getParent())) ||
3140  (QTB && !HasOnePredAndOneSucc(QTB, QBI->getParent(), PostBB)))
3141  return false;
3142  if (!QBI->getParent()->hasNUses(2))
3143  return false;
3144 
3145  // OK, this is a sequence of two diamonds or triangles.
3146  // Check if there are stores in PTB or PFB that are repeated in QTB or QFB.
3147  SmallPtrSet<Value *, 4> PStoreAddresses, QStoreAddresses;
3148  for (auto *BB : {PTB, PFB}) {
3149  if (!BB)
3150  continue;
3151  for (auto &I : *BB)
3152  if (StoreInst *SI = dyn_cast<StoreInst>(&I))
3153  PStoreAddresses.insert(SI->getPointerOperand());
3154  }
3155  for (auto *BB : {QTB, QFB}) {
3156  if (!BB)
3157  continue;
3158  for (auto &I : *BB)
3159  if (StoreInst *SI = dyn_cast<StoreInst>(&I))
3160  QStoreAddresses.insert(SI->getPointerOperand());
3161  }
3162 
3163  set_intersect(PStoreAddresses, QStoreAddresses);
3164  // set_intersect mutates PStoreAddresses in place. Rename it here to make it
3165  // clear what it contains.
3166  auto &CommonAddresses = PStoreAddresses;
3167 
3168  bool Changed = false;
3169  for (auto *Address : CommonAddresses)
3170  Changed |= mergeConditionalStoreToAddress(
3171  PTB, PFB, QTB, QFB, PostBB, Address, InvertPCond, InvertQCond, DL);
3172  return Changed;
3173 }
3174 
3175 /// If we have a conditional branch as a predecessor of another block,
3176 /// this function tries to simplify it. We know
3177 /// that PBI and BI are both conditional branches, and BI is in one of the
3178 /// successor blocks of PBI - PBI branches to BI.
3180  const DataLayout &DL) {
3181  assert(PBI->isConditional() && BI->isConditional());
3182  BasicBlock *BB = BI->getParent();
3183 
3184  // If this block ends with a branch instruction, and if there is a
3185  // predecessor that ends on a branch of the same condition, make
3186  // this conditional branch redundant.
3187  if (PBI->getCondition() == BI->getCondition() &&
3188  PBI->getSuccessor(0) != PBI->getSuccessor(1)) {
3189  // Okay, the outcome of this conditional branch is statically
3190  // knowable. If this block had a single pred, handle specially.
3191  if (BB->getSinglePredecessor()) {
3192  // Turn this into a branch on constant.
3193  bool CondIsTrue = PBI->getSuccessor(0) == BB;
3194  BI->setCondition(
3195  ConstantInt::get(Type::getInt1Ty(BB->getContext()), CondIsTrue));
3196  return true; // Nuke the branch on constant.
3197  }
3198 
3199  // Otherwise, if there are multiple predecessors, insert a PHI that merges
3200  // in the constant and simplify the block result. Subsequent passes of
3201  // simplifycfg will thread the block.
3203  pred_iterator PB = pred_begin(BB), PE = pred_end(BB);
3204  PHINode *NewPN = PHINode::Create(
3205  Type::getInt1Ty(BB->getContext()), std::distance(PB, PE),
3206  BI->getCondition()->getName() + ".pr", &BB->front());
3207  // Okay, we're going to insert the PHI node. Since PBI is not the only
3208  // predecessor, compute the PHI'd conditional value for all of the preds.
3209  // Any predecessor where the condition is not computable we keep symbolic.
3210  for (pred_iterator PI = PB; PI != PE; ++PI) {
3211  BasicBlock *P = *PI;
3212  if ((PBI = dyn_cast<BranchInst>(P->getTerminator())) && PBI != BI &&
3213  PBI->isConditional() && PBI->getCondition() == BI->getCondition() &&
3214  PBI->getSuccessor(0) != PBI->getSuccessor(1)) {
3215  bool CondIsTrue = PBI->getSuccessor(0) == BB;
3216  NewPN->addIncoming(
3217  ConstantInt::get(Type::getInt1Ty(BB->getContext()), CondIsTrue),
3218  P);
3219  } else {
3220  NewPN->addIncoming(BI->getCondition(), P);
3221  }
3222  }
3223 
3224  BI->setCondition(NewPN);
3225  return true;
3226  }
3227  }
3228 
3229  if (auto *CE = dyn_cast<ConstantExpr>(BI->getCondition()))
3230  if (CE->canTrap())
3231  return false;
3232 
3233  // If both branches are conditional and both contain stores to the same
3234  // address, remove the stores from the conditionals and create a conditional
3235  // merged store at the end.
3236  if (MergeCondStores && mergeConditionalStores(PBI, BI, DL))
3237  return true;
3238 
3239  // If this is a conditional branch in an empty block, and if any
3240  // predecessors are a conditional branch to one of our destinations,
3241  // fold the conditions into logical ops and one cond br.
3242 
3243  // Ignore dbg intrinsics.
3244  if (&*BB->instructionsWithoutDebug().begin() != BI)
3245  return false;
3246 
3247  int PBIOp, BIOp;
3248  if (PBI->getSuccessor(0) == BI->getSuccessor(0)) {
3249  PBIOp = 0;
3250  BIOp = 0;
3251  } else if (PBI->getSuccessor(0) == BI->getSuccessor(1)) {
3252  PBIOp = 0;
3253  BIOp = 1;
3254  } else if (PBI->getSuccessor(1) == BI->getSuccessor(0)) {
3255  PBIOp = 1;
3256  BIOp = 0;
3257  } else if (PBI->getSuccessor(1) == BI->getSuccessor(1)) {
3258  PBIOp = 1;
3259  BIOp = 1;
3260  } else {
3261  return false;
3262  }
3263 
3264  // Check to make sure that the other destination of this branch
3265  // isn't BB itself. If so, this is an infinite loop that will
3266  // keep getting unwound.
3267  if (PBI->getSuccessor(PBIOp) == BB)
3268  return false;
3269 
3270  // Do not perform this transformation if it would require
3271  // insertion of a large number of select instructions. For targets
3272  // without predication/cmovs, this is a big pessimization.
3273 
3274  // Also do not perform this transformation if any phi node in the common
3275  // destination block can trap when reached by BB or PBB (PR17073). In that
3276  // case, it would be unsafe to hoist the operation into a select instruction.
3277 
3278  BasicBlock *CommonDest = PBI->getSuccessor(PBIOp);
3279  unsigned NumPhis = 0;
3280  for (BasicBlock::iterator II = CommonDest->begin(); isa<PHINode>(II);
3281  ++II, ++NumPhis) {
3282  if (NumPhis > 2) // Disable this xform.
3283  return false;
3284 
3285  PHINode *PN = cast<PHINode>(II);
3286  Value *BIV = PN->getIncomingValueForBlock(BB);
3287  if (ConstantExpr *CE = dyn_cast<ConstantExpr>(BIV))
3288  if (CE->canTrap())
3289  return false;
3290 
3291  unsigned PBBIdx = PN->getBasicBlockIndex(PBI->getParent());
3292  Value *PBIV = PN->getIncomingValue(PBBIdx);
3293  if (ConstantExpr *CE = dyn_cast<ConstantExpr>(PBIV))
3294  if (CE->canTrap())
3295  return false;
3296  }
3297 
3298  // Finally, if everything is ok, fold the branches to logical ops.
3299  BasicBlock *OtherDest = BI->getSuccessor(BIOp ^ 1);
3300 
3301  LLVM_DEBUG(dbgs() << "FOLDING BRs:" << *PBI->getParent()
3302  << "AND: " << *BI->getParent());
3303 
3304  // If OtherDest *is* BB, then BB is a basic block with a single conditional
3305  // branch in it, where one edge (OtherDest) goes back to itself but the other
3306  // exits. We don't *know* that the program avoids the infinite loop
3307  // (even though that seems likely). If we do this xform naively, we'll end up
3308  // recursively unpeeling the loop. Since we know that (after the xform is
3309  // done) that the block *is* infinite if reached, we just make it an obviously
3310  // infinite loop with no cond branch.
3311  if (OtherDest == BB) {
3312  // Insert it at the end of the function, because it's either code,
3313  // or it won't matter if it's hot. :)
3314  BasicBlock *InfLoopBlock =
3315  BasicBlock::Create(BB->getContext(), "infloop", BB->getParent());
3316  BranchInst::Create(InfLoopBlock, InfLoopBlock);
3317  OtherDest = InfLoopBlock;
3318  }
3319 
3320  LLVM_DEBUG(dbgs() << *PBI->getParent()->getParent());
3321 
3322  // BI may have other predecessors. Because of this, we leave
3323  // it alone, but modify PBI.
3324 
3325  // Make sure we get to CommonDest on True&True directions.
3326  Value *PBICond = PBI->getCondition();
3327  IRBuilder<NoFolder> Builder(PBI);
3328  if (PBIOp)
3329  PBICond = Builder.CreateNot(PBICond, PBICond->getName() + ".not");
3330 
3331  Value *BICond = BI->getCondition();
3332  if (BIOp)
3333  BICond = Builder.CreateNot(BICond, BICond->getName() + ".not");
3334 
3335  // Merge the conditions.
3336  Value *Cond = Builder.CreateOr(PBICond, BICond, "brmerge");
3337 
3338  // Modify PBI to branch on the new condition to the new dests.
3339  PBI->setCondition(Cond);
3340  PBI->setSuccessor(0, CommonDest);
3341  PBI->setSuccessor(1, OtherDest);
3342 
3343  // Update branch weight for PBI.
3344  uint64_t PredTrueWeight, PredFalseWeight, SuccTrueWeight, SuccFalseWeight;
3345  uint64_t PredCommon, PredOther, SuccCommon, SuccOther;
3346  bool HasWeights =
3347  extractPredSuccWeights(PBI, BI, PredTrueWeight, PredFalseWeight,
3348  SuccTrueWeight, SuccFalseWeight);
3349  if (HasWeights) {
3350  PredCommon = PBIOp ? PredFalseWeight : PredTrueWeight;
3351  PredOther = PBIOp ? PredTrueWeight : PredFalseWeight;
3352  SuccCommon = BIOp ? SuccFalseWeight : SuccTrueWeight;
3353  SuccOther = BIOp ? SuccTrueWeight : SuccFalseWeight;
3354  // The weight to CommonDest should be PredCommon * SuccTotal +
3355  // PredOther * SuccCommon.
3356  // The weight to OtherDest should be PredOther * SuccOther.
3357  uint64_t NewWeights[2] = {PredCommon * (SuccCommon + SuccOther) +
3358  PredOther * SuccCommon,
3359  PredOther * SuccOther};
3360  // Halve the weights if any of them cannot fit in an uint32_t
3361  FitWeights(NewWeights);
3362 
3363  setBranchWeights(PBI, NewWeights[0], NewWeights[1]);
3364  }
3365 
3366  // OtherDest may have phi nodes. If so, add an entry from PBI's
3367  // block that are identical to the entries for BI's block.
3368  AddPredecessorToBlock(OtherDest, PBI->getParent(), BB);
3369 
3370  // We know that the CommonDest already had an edge from PBI to
3371  // it. If it has PHIs though, the PHIs may have different
3372  // entries for BB and PBI's BB. If so, insert a select to make
3373  // them agree.
3374  for (PHINode &PN : CommonDest->phis()) {
3375  Value *BIV = PN.getIncomingValueForBlock(BB);
3376  unsigned PBBIdx = PN.getBasicBlockIndex(PBI->getParent());
3377  Value *PBIV = PN.getIncomingValue(PBBIdx);
3378  if (BIV != PBIV) {
3379  // Insert a select in PBI to pick the right value.
3380  SelectInst *NV = cast<SelectInst>(
3381  Builder.CreateSelect(PBICond, PBIV, BIV, PBIV->getName() + ".mux"));
3382  PN.setIncomingValue(PBBIdx, NV);
3383  // Although the select has the same condition as PBI, the original branch
3384  // weights for PBI do not apply to the new select because the select's
3385  // 'logical' edges are incoming edges of the phi that is eliminated, not
3386  // the outgoing edges of PBI.
3387  if (HasWeights) {
3388  uint64_t PredCommon = PBIOp ? PredFalseWeight : PredTrueWeight;
3389  uint64_t PredOther = PBIOp ? PredTrueWeight : PredFalseWeight;
3390  uint64_t SuccCommon = BIOp ? SuccFalseWeight : SuccTrueWeight;
3391  uint64_t SuccOther = BIOp ? SuccTrueWeight : SuccFalseWeight;
3392  // The weight to PredCommonDest should be PredCommon * SuccTotal.
3393  // The weight to PredOtherDest should be PredOther * SuccCommon.
3394  uint64_t NewWeights[2] = {PredCommon * (SuccCommon + SuccOther),
3395  PredOther * SuccCommon};
3396 
3397  FitWeights(NewWeights);
3398 
3399  setBranchWeights(NV, NewWeights[0], NewWeights[1]);
3400  }
3401  }
3402  }
3403 
3404  LLVM_DEBUG(dbgs() << "INTO: " << *PBI->getParent());
3405  LLVM_DEBUG(dbgs() << *PBI->getParent()->getParent());
3406 
3407  // This basic block is probably dead. We know it has at least
3408  // one fewer predecessor.
3409  return true;
3410 }
3411 
3412 // Simplifies a terminator by replacing it with a branch to TrueBB if Cond is
3413 // true or to FalseBB if Cond is false.
3414 // Takes care of updating the successors and removing the old terminator.
3415 // Also makes sure not to introduce new successors by assuming that edges to
3416 // non-successor TrueBBs and FalseBBs aren't reachable.
3417 static bool SimplifyTerminatorOnSelect(Instruction *OldTerm, Value *Cond,
3418  BasicBlock *TrueBB, BasicBlock *FalseBB,
3419  uint32_t TrueWeight,
3420  uint32_t FalseWeight) {
3421  // Remove any superfluous successor edges from the CFG.
3422  // First, figure out which successors to preserve.
3423  // If TrueBB and FalseBB are equal, only try to preserve one copy of that
3424  // successor.
3425  BasicBlock *KeepEdge1 = TrueBB;
3426  BasicBlock *KeepEdge2 = TrueBB != FalseBB ? FalseBB : nullptr;
3427 
3428  // Then remove the rest.
3429  for (BasicBlock *Succ : successors(OldTerm)) {
3430  // Make sure only to keep exactly one copy of each edge.
3431  if (Succ == KeepEdge1)
3432  KeepEdge1 = nullptr;
3433  else if (Succ == KeepEdge2)
3434  KeepEdge2 = nullptr;
3435  else
3436  Succ->removePredecessor(OldTerm->getParent(),
3437  /*KeepOneInputPHIs=*/true);
3438  }
3439 
3440  IRBuilder<> Builder(OldTerm);
3441  Builder.SetCurrentDebugLocation(OldTerm->getDebugLoc());
3442 
3443  // Insert an appropriate new terminator.
3444  if (!KeepEdge1 && !KeepEdge2) {
3445  if (TrueBB == FalseBB)
3446  // We were only looking for one successor, and it was present.
3447  // Create an unconditional branch to it.
3448  Builder.CreateBr(TrueBB);
3449  else {
3450  // We found both of the successors we were looking for.
3451  // Create a conditional branch sharing the condition of the select.
3452  BranchInst *NewBI = Builder.CreateCondBr(Cond, TrueBB, FalseBB);
3453  if (TrueWeight != FalseWeight)
3454  setBranchWeights(NewBI, TrueWeight, FalseWeight);
3455  }
3456  } else if (KeepEdge1 && (KeepEdge2 || TrueBB == FalseBB)) {
3457  // Neither of the selected blocks were successors, so this
3458  // terminator must be unreachable.
3459  new UnreachableInst(OldTerm->getContext(), OldTerm);
3460  } else {
3461  // One of the selected values was a successor, but the other wasn't.
3462  // Insert an unconditional branch to the one that was found;
3463  // the edge to the one that wasn't must be unreachable.
3464  if (!KeepEdge1)
3465  // Only TrueBB was found.
3466  Builder.CreateBr(TrueBB);
3467  else
3468  // Only FalseBB was found.
3469  Builder.CreateBr(FalseBB);
3470  }
3471 
3472  EraseTerminatorAndDCECond(OldTerm);
3473  return true;
3474 }
3475 
3476 // Replaces
3477 // (switch (select cond, X, Y)) on constant X, Y
3478 // with a branch - conditional if X and Y lead to distinct BBs,
3479 // unconditional otherwise.
3481  // Check for constant integer values in the select.
3482  ConstantInt *TrueVal = dyn_cast<ConstantInt>(Select->getTrueValue());
3483  ConstantInt *FalseVal = dyn_cast<ConstantInt>(Select->getFalseValue());
3484  if (!TrueVal || !FalseVal)
3485  return false;
3486 
3487  // Find the relevant condition and destinations.
3488  Value *Condition = Select->getCondition();
3489  BasicBlock *TrueBB = SI->findCaseValue(TrueVal)->getCaseSuccessor();
3490  BasicBlock *FalseBB = SI->findCaseValue(FalseVal)->getCaseSuccessor();
3491 
3492  // Get weight for TrueBB and FalseBB.
3493  uint32_t TrueWeight = 0, FalseWeight = 0;
3494  SmallVector<uint64_t, 8> Weights;
3495  bool HasWeights = HasBranchWeights(SI);
3496  if (HasWeights) {
3497  GetBranchWeights(SI, Weights);
3498  if (Weights.size() == 1 + SI->getNumCases()) {
3499  TrueWeight =
3500  (uint32_t)Weights[SI->findCaseValue(TrueVal)->getSuccessorIndex()];
3501  FalseWeight =
3502  (uint32_t)Weights[SI->findCaseValue(FalseVal)->getSuccessorIndex()];
3503  }
3504  }
3505 
3506  // Perform the actual simplification.
3507  return SimplifyTerminatorOnSelect(SI, Condition, TrueBB, FalseBB, TrueWeight,
3508  FalseWeight);
3509 }
3510 
3511 // Replaces
3512 // (indirectbr (select cond, blockaddress(@fn, BlockA),
3513 // blockaddress(@fn, BlockB)))
3514 // with
3515 // (br cond, BlockA, BlockB).
3517  // Check that both operands of the select are block addresses.
3520  if (!TBA || !FBA)
3521  return false;
3522 
3523  // Extract the actual blocks.
3524  BasicBlock *TrueBB = TBA->getBasicBlock();
3525  BasicBlock *FalseBB = FBA->getBasicBlock();
3526 
3527  // Perform the actual simplification.
3528  return SimplifyTerminatorOnSelect(IBI, SI->getCondition(), TrueBB, FalseBB, 0,
3529  0);
3530 }
3531 
3532 /// This is called when we find an icmp instruction
3533 /// (a seteq/setne with a constant) as the only instruction in a
3534 /// block that ends with an uncond branch. We are looking for a very specific
3535 /// pattern that occurs when "A == 1 || A == 2 || A == 3" gets simplified. In
3536 /// this case, we merge the first two "or's of icmp" into a switch, but then the
3537 /// default value goes to an uncond block with a seteq in it, we get something
3538 /// like:
3539 ///
3540 /// switch i8 %A, label %DEFAULT [ i8 1, label %end i8 2, label %end ]
3541 /// DEFAULT:
3542 /// %tmp = icmp eq i8 %A, 92
3543 /// br label %end
3544 /// end:
3545 /// ... = phi i1 [ true, %entry ], [ %tmp, %DEFAULT ], [ true, %entry ]
3546 ///
3547 /// We prefer to split the edge to 'end' so that there is a true/false entry to
3548 /// the PHI, merging the third icmp into the switch.
3549 bool SimplifyCFGOpt::tryToSimplifyUncondBranchWithICmpInIt(
3550  ICmpInst *ICI, IRBuilder<> &Builder) {
3551  BasicBlock *BB = ICI->getParent();
3552 
3553  // If the block has any PHIs in it or the icmp has multiple uses, it is too
3554  // complex.
3555  if (isa<PHINode>(BB->begin()) || !ICI->hasOneUse())
3556  return false;
3557 
3558  Value *V = ICI->getOperand(0);
3559  ConstantInt *Cst = cast<ConstantInt>(ICI->getOperand(1));
3560 
3561  // The pattern we're looking for is where our only predecessor is a switch on
3562  // 'V' and this block is the default case for the switch. In this case we can
3563  // fold the compared value into the switch to simplify things.
3564  BasicBlock *Pred = BB->getSinglePredecessor();
3565  if (!Pred || !isa<SwitchInst>(Pred->getTerminator()))
3566  return false;
3567 
3568  SwitchInst *SI = cast<SwitchInst>(Pred->getTerminator());
3569  if (SI->getCondition() != V)
3570  return false;
3571 
3572  // If BB is reachable on a non-default case, then we simply know the value of
3573  // V in this block. Substitute it and constant fold the icmp instruction
3574  // away.
3575  if (SI->getDefaultDest() != BB) {
3576  ConstantInt *VVal = SI->findCaseDest(BB);
3577  assert(VVal && "Should have a unique destination value");
3578  ICI->setOperand(0, VVal);
3579 
3580  if (Value *V = SimplifyInstruction(ICI, {DL, ICI})) {
3581  ICI->replaceAllUsesWith(V);
3582  ICI->eraseFromParent();
3583  }
3584  // BB is now empty, so it is likely to simplify away.
3585  return requestResimplify();
3586  }
3587 
3588  // Ok, the block is reachable from the default dest. If the constant we're
3589  // comparing exists in one of the other edges, then we can constant fold ICI
3590  // and zap it.
3591  if (SI->findCaseValue(Cst) != SI->case_default()) {
3592  Value *V;
3593  if (ICI->getPredicate() == ICmpInst::ICMP_EQ)
3594  V = ConstantInt::getFalse(BB->getContext());
3595  else
3596  V = ConstantInt::getTrue(BB->getContext());
3597 
3598  ICI->replaceAllUsesWith(V);
3599  ICI->eraseFromParent();
3600  // BB is now empty, so it is likely to simplify away.
3601  return requestResimplify();
3602  }
3603 
3604  // The use of the icmp has to be in the 'end' block, by the only PHI node in
3605  // the block.
3606  BasicBlock *SuccBlock = BB->getTerminator()->getSuccessor(0);
3607  PHINode *PHIUse = dyn_cast<PHINode>(ICI->user_back());
3608  if (PHIUse == nullptr || PHIUse != &SuccBlock->front() ||
3609  isa<PHINode>(++BasicBlock::iterator(PHIUse)))
3610  return false;
3611 
3612  // If the icmp is a SETEQ, then the default dest gets false, the new edge gets
3613  // true in the PHI.
3614  Constant *DefaultCst = ConstantInt::getTrue(BB->getContext());
3615  Constant *NewCst = ConstantInt::getFalse(BB->getContext());
3616 
3617  if (ICI->getPredicate() == ICmpInst::ICMP_EQ)
3618  std::swap(DefaultCst, NewCst);
3619 
3620  // Replace ICI (which is used by the PHI for the default value) with true or
3621  // false depending on if it is EQ or NE.
3622  ICI->replaceAllUsesWith(DefaultCst);
3623  ICI->eraseFromParent();
3624 
3625  // Okay, the switch goes to this block on a default value. Add an edge from
3626  // the switch to the merge point on the compared value.
3627  BasicBlock *NewBB =
3628  BasicBlock::Create(BB->getContext(), "switch.edge", BB->getParent(), BB);
3629  {
3630  SwitchInstProfUpdateWrapper SIW(*SI);
3631  auto W0 = SIW.getSuccessorWeight(0);
3633  if (W0) {
3634  NewW = ((uint64_t(*W0) + 1) >> 1);
3635  SIW.setSuccessorWeight(0, *NewW);
3636  }
3637  SIW.addCase(Cst, NewBB, NewW);
3638  }
3639 
3640  // NewBB branches to the phi block, add the uncond branch and the phi entry.
3641  Builder.SetInsertPoint(NewBB);
3642  Builder.SetCurrentDebugLocation(SI->getDebugLoc());
3643  Builder.CreateBr(SuccBlock);
3644  PHIUse->addIncoming(NewCst, NewBB);
3645  return true;
3646 }
3647 
3648 /// The specified branch is a conditional branch.
3649 /// Check to see if it is branching on an or/and chain of icmp instructions, and
3650 /// fold it into a switch instruction if so.
3652  const DataLayout &DL) {
3653  Instruction *Cond = dyn_cast<Instruction>(BI->getCondition());
3654  if (!Cond)
3655  return false;
3656 
3657  // Change br (X == 0 | X == 1), T, F into a switch instruction.
3658  // If this is a bunch of seteq's or'd together, or if it's a bunch of
3659  // 'setne's and'ed together, collect them.
3660 
3661  // Try to gather values from a chain of and/or to be turned into a switch
3662  ConstantComparesGatherer ConstantCompare(Cond, DL);
3663  // Unpack the result
3664  SmallVectorImpl<ConstantInt *> &Values = ConstantCompare.Vals;
3665  Value *CompVal = ConstantCompare.CompValue;
3666  unsigned UsedICmps = ConstantCompare.UsedICmps;
3667  Value *ExtraCase = ConstantCompare.Extra;
3668 
3669  // If we didn't have a multiply compared value, fail.
3670  if (!CompVal)
3671  return false;
3672 
3673  // Avoid turning single icmps into a switch.
3674  if (UsedICmps <= 1)
3675  return false;
3676 
3677  bool TrueWhenEqual = (Cond->getOpcode() == Instruction::Or);
3678 
3679  // There might be duplicate constants in the list, which the switch
3680  // instruction can't handle, remove them now.
3681  array_pod_sort(Values.begin(), Values.end(), ConstantIntSortPredicate);
3682  Values.erase(std::unique(Values.begin(), Values.end()), Values.end());
3683 
3684  // If Extra was used, we require at least two switch values to do the
3685  // transformation. A switch with one value is just a conditional branch.
3686  if (ExtraCase && Values.size() < 2)
3687  return false;
3688 
3689  // TODO: Preserve branch weight metadata, similarly to how
3690  // FoldValueComparisonIntoPredecessors preserves it.
3691 
3692  // Figure out which block is which destination.
3693  BasicBlock *DefaultBB = BI->getSuccessor(1);
3694  BasicBlock *EdgeBB = BI->getSuccessor(0);
3695  if (!TrueWhenEqual)
3696  std::swap(DefaultBB, EdgeBB);
3697 
3698  BasicBlock *BB = BI->getParent();
3699 
3700  LLVM_DEBUG(dbgs() << "Converting 'icmp' chain with " << Values.size()
3701  << " cases into SWITCH. BB is:\n"
3702  << *BB);
3703 
3704  // If there are any extra values that couldn't be folded into the switch
3705  // then we evaluate them with an explicit branch first. Split the block
3706  // right before the condbr to handle it.
3707  if (ExtraCase) {
3708  BasicBlock *NewBB =
3709  BB->splitBasicBlock(BI->getIterator(), "switch.early.test");
3710  // Remove the uncond branch added to the old block.
3711  Instruction *OldTI = BB->getTerminator();
3712  Builder.SetInsertPoint(OldTI);
3713 
3714  if (TrueWhenEqual)
3715  Builder.CreateCondBr(ExtraCase, EdgeBB, NewBB);
3716  else
3717  Builder.CreateCondBr(ExtraCase, NewBB, EdgeBB);
3718 
3719  OldTI->eraseFromParent();
3720 
3721  // If there are PHI nodes in EdgeBB, then we need to add a new entry to them
3722  // for the edge we just added.
3723  AddPredecessorToBlock(EdgeBB, BB, NewBB);
3724 
3725  LLVM_DEBUG(dbgs() << " ** 'icmp' chain unhandled condition: " << *ExtraCase
3726  << "\nEXTRABB = " << *BB);
3727  BB = NewBB;
3728  }
3729 
3730  Builder.SetInsertPoint(BI);
3731  // Convert pointer to int before we switch.
3732  if (CompVal->getType()->isPointerTy()) {
3733  CompVal = Builder.CreatePtrToInt(
3734  CompVal, DL.getIntPtrType(CompVal->getType()), "magicptr");
3735  }
3736 
3737  // Create the new switch instruction now.
3738  SwitchInst *New = Builder.CreateSwitch(CompVal, DefaultBB, Values.size());
3739 
3740  // Add all of the 'cases' to the switch instruction.
3741  for (unsigned i = 0, e = Values.size(); i != e; ++i)
3742  New->addCase(Values[i], EdgeBB);
3743 
3744  // We added edges from PI to the EdgeBB. As such, if there were any
3745  // PHI nodes in EdgeBB, they need entries to be added corresponding to
3746  // the number of edges added.
3747  for (BasicBlock::iterator BBI = EdgeBB->begin(); isa<PHINode>(BBI); ++BBI) {
3748  PHINode *PN = cast<PHINode>(BBI);
3749  Value *InVal = PN->getIncomingValueForBlock(BB);
3750  for (unsigned i = 0, e = Values.size() - 1; i != e; ++i)
3751  PN->addIncoming(InVal, BB);
3752  }
3753 
3754  // Erase the old branch instruction.
3756 
3757  LLVM_DEBUG(dbgs() << " ** 'icmp' chain result is:\n" << *BB << '\n');
3758  return true;
3759 }
3760 
3761 bool SimplifyCFGOpt::SimplifyResume(ResumeInst *RI, IRBuilder<> &Builder) {
3762  if (isa<PHINode>(RI->getValue()))
3763  return SimplifyCommonResume(RI);
3764  else if (isa<LandingPadInst>(RI->getParent()->getFirstNonPHI()) &&
3765  RI->getValue() == RI->getParent()->getFirstNonPHI())
3766  // The resume must unwind the exception that caused control to branch here.
3767  return SimplifySingleResume(RI);
3768 
3769  return false;
3770 }
3771 
3772 // Simplify resume that is shared by several landing pads (phi of landing pad).
3773 bool SimplifyCFGOpt::SimplifyCommonResume(ResumeInst *RI) {
3774  BasicBlock *BB = RI->getParent();
3775 
3776  // Check that there are no other instructions except for debug intrinsics
3777  // between the phi of landing pads (RI->getValue()) and resume instruction.
3778  BasicBlock::iterator I = cast<Instruction>(RI->getValue())->getIterator(),
3779  E = RI->getIterator();
3780  while (++I != E)
3781  if (!isa<DbgInfoIntrinsic>(I))
3782  return false;
3783 
3784  SmallSetVector<BasicBlock *, 4> TrivialUnwindBlocks;
3785  auto *PhiLPInst = cast<PHINode>(RI->getValue());
3786 
3787  // Check incoming blocks to see if any of them are trivial.
3788  for (unsigned Idx = 0, End = PhiLPInst->getNumIncomingValues(); Idx != End;
3789  Idx++) {
3790  auto *IncomingBB = PhiLPInst->getIncomingBlock(Idx);
3791  auto *IncomingValue = PhiLPInst->getIncomingValue(Idx);
3792 
3793  // If the block has other successors, we can not delete it because
3794  // it has other dependents.
3795  if (IncomingBB->getUniqueSuccessor() != BB)
3796  continue;
3797 
3798  auto *LandingPad = dyn_cast<LandingPadInst>(IncomingBB->getFirstNonPHI());
3799  // Not the landing pad that caused the control to branch here.
3800  if (IncomingValue != LandingPad)
3801  continue;
3802 
3803  bool isTrivial = true;
3804 
3805  I = IncomingBB->getFirstNonPHI()->getIterator();
3806  E = IncomingBB->getTerminator()->getIterator();
3807  while (++I != E)
3808  if (!isa<DbgInfoIntrinsic>(I)) {
3809  isTrivial = false;
3810  break;
3811  }
3812 
3813  if (isTrivial)
3814  TrivialUnwindBlocks.insert(IncomingBB);
3815  }
3816 
3817  // If no trivial unwind blocks, don't do any simplifications.
3818  if (TrivialUnwindBlocks.empty())
3819  return false;
3820 
3821  // Turn all invokes that unwind here into calls.
3822  for (auto *TrivialBB : TrivialUnwindBlocks) {
3823  // Blocks that will be simplified should be removed from the phi node.
3824  // Note there could be multiple edges to the resume block, and we need
3825  // to remove them all.
3826  while (PhiLPInst->getBasicBlockIndex(TrivialBB) != -1)
3827  BB->removePredecessor(TrivialBB, true);
3828 
3829  for (pred_iterator PI = pred_begin(TrivialBB), PE = pred_end(TrivialBB);
3830  PI != PE;) {
3831  BasicBlock *Pred = *PI++;
3832  removeUnwindEdge(Pred);
3833  }
3834 
3835  // In each SimplifyCFG run, only the current processed block can be erased.
3836  // Otherwise, it will break the iteration of SimplifyCFG pass. So instead
3837  // of erasing TrivialBB, we only remove the branch to the common resume
3838  // block so that we can later erase the resume block since it has no
3839  // predecessors.
3840  TrivialBB->getTerminator()->eraseFromParent();
3841  new UnreachableInst(RI->getContext(), TrivialBB);
3842  }
3843 
3844  // Delete the resume block if all its predecessors have been removed.
3845  if (pred_empty(BB))
3846  BB->eraseFromParent();
3847 
3848  return !TrivialUnwindBlocks.empty();
3849 }
3850 
3851 // Simplify resume that is only used by a single (non-phi) landing pad.
3852 bool SimplifyCFGOpt::SimplifySingleResume(ResumeInst *RI) {
3853  BasicBlock *BB = RI->getParent();
3855  assert(RI->getValue() == LPInst &&
3856  "Resume must unwind the exception that caused control to here");
3857 
3858  // Check that there are no other instructions except for debug intrinsics.
3859  BasicBlock::iterator I = LPInst->getIterator(), E = RI->getIterator();
3860  while (++I != E)
3861  if (!isa<DbgInfoIntrinsic>(I))
3862  return false;
3863 
3864  // Turn all invokes that unwind here into calls and delete the basic block.
3865  for (pred_iterator PI = pred_begin(BB), PE = pred_end(BB); PI != PE;) {
3866  BasicBlock *Pred = *PI++;
3867  removeUnwindEdge(Pred);
3868  }
3869 
3870  // The landingpad is now unreachable. Zap it.
3871  if (LoopHeaders)
3872  LoopHeaders->erase(BB);
3873  BB->eraseFromParent();
3874  return true;
3875 }
3876 
3878  // If this is a trivial cleanup pad that executes no instructions, it can be
3879  // eliminated. If the cleanup pad continues to the caller, any predecessor
3880  // that is an EH pad will be updated to continue to the caller and any
3881  // predecessor that terminates with an invoke instruction will have its invoke
3882  // instruction converted to a call instruction. If the cleanup pad being
3883  // simplified does not continue to the caller, each predecessor will be
3884  // updated to continue to the unwind destination of the cleanup pad being
3885  // simplified.
3886  BasicBlock *BB = RI->getParent();
3887  CleanupPadInst *CPInst = RI->getCleanupPad();
3888  if (CPInst->getParent() != BB)
3889  // This isn't an empty cleanup.
3890  return false;
3891 
3892  // We cannot kill the pad if it has multiple uses. This typically arises
3893  // from unreachable basic blocks.
3894  if (!CPInst->hasOneUse())
3895  return false;
3896 
3897  // Check that there are no other instructions except for benign intrinsics.
3898  BasicBlock::iterator I = CPInst->getIterator(), E = RI->getIterator();
3899  while (++I != E) {
3900  auto *II = dyn_cast<IntrinsicInst>(I);
3901  if (!II)
3902  return false;
3903 
3904  Intrinsic::ID IntrinsicID = II->getIntrinsicID();
3905  switch (IntrinsicID) {
3906  case Intrinsic::dbg_declare:
3907  case Intrinsic::dbg_value:
3908  case Intrinsic::dbg_label:
3909  case Intrinsic::lifetime_end:
3910  break;
3911  default:
3912  return false;
3913  }
3914  }
3915 
3916  // If the cleanup return we are simplifying unwinds to the caller, this will
3917  // set UnwindDest to nullptr.
3918  BasicBlock *UnwindDest = RI->getUnwindDest();
3919  Instruction *DestEHPad = UnwindDest ? UnwindDest->getFirstNonPHI() : nullptr;
3920 
3921  // We're about to remove BB from the control flow. Before we do, sink any
3922  // PHINodes into the unwind destination. Doing this before changing the
3923  // control flow avoids some potentially slow checks, since we can currently
3924  // be certain that UnwindDest and BB have no common predecessors (since they
3925  // are both EH pads).
3926  if (UnwindDest) {
3927  // First, go through the PHI nodes in UnwindDest and update any nodes that
3928  // reference the block we are removing
3929  for (BasicBlock::iterator I = UnwindDest->begin(),
3930  IE = DestEHPad->getIterator();
3931  I != IE; ++I) {
3932  PHINode *DestPN = cast<PHINode>(I);
3933 
3934  int Idx = DestPN->getBasicBlockIndex(BB);
3935  // Since BB unwinds to UnwindDest, it has to be in the PHI node.
3936  assert(Idx != -1);
3937  // This PHI node has an incoming value that corresponds to a control
3938  // path through the cleanup pad we are removing. If the incoming
3939  // value is in the cleanup pad, it must be a PHINode (because we
3940  // verified above that the block is otherwise empty). Otherwise, the
3941  // value is either a constant or a value that dominates the cleanup
3942  // pad being removed.
3943  //
3944  // Because BB and UnwindDest are both EH pads, all of their
3945  // predecessors must unwind to these blocks, and since no instruction
3946  // can have multiple unwind destinations, there will be no overlap in
3947  // incoming blocks between SrcPN and DestPN.
3948  Value *SrcVal = DestPN->getIncomingValue(Idx);
3949  PHINode *SrcPN = dyn_cast<PHINode>(SrcVal);
3950 
3951  // Remove the entry for the block we are deleting.
3952  DestPN->removeIncomingValue(Idx, false);
3953 
3954  if (SrcPN && SrcPN->getParent() == BB) {
3955  // If the incoming value was a PHI node in the cleanup pad we are
3956  // removing, we need to merge that PHI node's incoming values into
3957  // DestPN.
3958  for (unsigned SrcIdx = 0, SrcE = SrcPN->getNumIncomingValues();
3959  SrcIdx != SrcE; ++SrcIdx) {
3960  DestPN->addIncoming(SrcPN->getIncomingValue(SrcIdx),
3961  SrcPN->getIncomingBlock(SrcIdx));
3962  }
3963  } else {
3964  // Otherwise, the incoming value came from above BB and
3965  // so we can just reuse it. We must associate all of BB's
3966  // predecessors with this value.
3967  for (auto *pred : predecessors(BB)) {
3968  DestPN->addIncoming(SrcVal, pred);
3969  }
3970  }
3971  }
3972 
3973  // Sink any remaining PHI nodes directly into UnwindDest.
3974  Instruction *InsertPt = DestEHPad;
3975  for (BasicBlock::iterator I = BB->begin(),
3976  IE = BB->getFirstNonPHI()->getIterator();
3977  I != IE;) {
3978  // The iterator must be incremented here because the instructions are
3979  // being moved to another block.
3980  PHINode *PN = cast<PHINode>(I++);
3981  if (PN->use_empty())
3982  // If the PHI node has no uses, just leave it. It will be erased
3983  // when we erase BB below.
3984  continue;
3985 
3986  // Otherwise, sink this PHI node into UnwindDest.
3987  // Any predecessors to UnwindDest which are not already represented
3988  // must be back edges which inherit the value from the path through
3989  // BB. In this case, the PHI value must reference itself.
3990  for (auto *pred : predecessors(UnwindDest))
3991  if (pred != BB)
3992  PN->addIncoming(PN, pred);
3993  PN->moveBefore(InsertPt);
3994  }
3995  }
3996 
3997  for (pred_iterator PI = pred_begin(BB), PE = pred_end(BB); PI != PE;) {
3998  // The iterator must be updated here because we are removing this pred.
3999  BasicBlock *PredBB = *PI++;
4000  if (UnwindDest == nullptr) {
4001  removeUnwindEdge(PredBB);
4002  } else {
4003  Instruction *TI = PredBB->getTerminator();
4004  TI->replaceUsesOfWith(BB, UnwindDest);
4005  }
4006  }
4007 
4008  // The cleanup pad is now unreachable. Zap it.
4009  BB->eraseFromParent();
4010  return true;
4011 }
4012 
4013 // Try to merge two cleanuppads together.
4015  // Skip any cleanuprets which unwind to caller, there is nothing to merge
4016  // with.
4017  BasicBlock *UnwindDest = RI->getUnwindDest();
4018  if (!UnwindDest)
4019  return false;
4020 
4021  // This cleanupret isn't the only predecessor of this cleanuppad, it wouldn't
4022  // be safe to merge without code duplication.
4023  if (UnwindDest->getSinglePredecessor() != RI->getParent())
4024  return false;
4025 
4026  // Verify that our cleanuppad's unwind destination is another cleanuppad.
4027  auto *SuccessorCleanupPad = dyn_cast<CleanupPadInst>(&UnwindDest->front());
4028  if (!SuccessorCleanupPad)
4029  return false;
4030 
4031  CleanupPadInst *PredecessorCleanupPad = RI->getCleanupPad();
4032  // Replace any uses of the successor cleanupad with the predecessor pad
4033  // The only cleanuppad uses should be this cleanupret, it's cleanupret and
4034  // funclet bundle operands.
4035  SuccessorCleanupPad->replaceAllUsesWith(PredecessorCleanupPad);
4036  // Remove the old cleanuppad.
4037  SuccessorCleanupPad->eraseFromParent();
4038  // Now, we simply replace the cleanupret with a branch to the unwind
4039  // destination.
4040  BranchInst::Create(UnwindDest, RI->getParent());
4041  RI->eraseFromParent();
4042 
4043  return true;
4044 }
4045 
4046 bool SimplifyCFGOpt::SimplifyCleanupReturn(CleanupReturnInst *RI) {
4047  // It is possible to transiantly have an undef cleanuppad operand because we
4048  // have deleted some, but not all, dead blocks.
4049  // Eventually, this block will be deleted.
4050  if (isa<UndefValue>(RI->getOperand(0)))
4051  return false;
4052 
4053  if (mergeCleanupPad(RI))
4054  return true;
4055 
4056  if (removeEmptyCleanup(RI))
4057  return true;
4058 
4059  return false;
4060 }
4061 
4062 bool SimplifyCFGOpt::SimplifyReturn(ReturnInst *RI, IRBuilder<> &Builder) {
4063  BasicBlock *BB = RI->getParent();
4064  if (!BB->getFirstNonPHIOrDbg()->isTerminator())
4065  return false;
4066 
4067  // Find predecessors that end with branches.
4068  SmallVector<BasicBlock *, 8> UncondBranchPreds;
4069  SmallVector<BranchInst *, 8> CondBranchPreds;
4070  for (pred_iterator PI = pred_begin(BB), E = pred_end(BB); PI != E; ++PI) {
4071  BasicBlock *P = *PI;
4072  Instruction *PTI = P->getTerminator();
4073  if (BranchInst *BI = dyn_cast<BranchInst>(PTI)) {
4074  if (BI->isUnconditional())
4075  UncondBranchPreds.push_back(P);
4076  else
4077  CondBranchPreds.push_back(BI);
4078  }
4079  }
4080 
4081  // If we found some, do the transformation!
4082  if (!UncondBranchPreds.empty() && DupRet) {
4083  while (!UncondBranchPreds.empty()) {
4084  BasicBlock *Pred = UncondBranchPreds.pop_back_val();
4085  LLVM_DEBUG(dbgs() << "FOLDING: " << *BB
4086  << "INTO UNCOND BRANCH PRED: " << *Pred);
4087  (void)FoldReturnIntoUncondBranch(RI, BB, Pred);
4088  }
4089 
4090  // If we eliminated all predecessors of the block, delete the block now.
4091  if (pred_empty(BB)) {
4092  // We know there are no successors, so just nuke the block.
4093  if (LoopHeaders)
4094  LoopHeaders->erase(BB);
4095  BB->eraseFromParent();
4096  }
4097 
4098  return true;
4099  }
4100 
4101  // Check out all of the conditional branches going to this return
4102  // instruction. If any of them just select between returns, change the
4103  // branch itself into a select/return pair.
4104  while (!CondBranchPreds.empty()) {
4105  BranchInst *BI = CondBranchPreds.pop_back_val();
4106 
4107  // Check to see if the non-BB successor is also a return block.
4108  if (isa<ReturnInst>(BI->getSuccessor(0)->getTerminator()) &&
4109  isa<ReturnInst>(BI->getSuccessor(1)->getTerminator()) &&
4110  SimplifyCondBranchToTwoReturns(BI, Builder))
4111  return true;
4112  }
4113  return false;
4114 }
4115 
4116 bool SimplifyCFGOpt::SimplifyUnreachable(UnreachableInst *UI) {
4117  BasicBlock *BB = UI->getParent();
4118 
4119  bool Changed = false;
4120 
4121  // If there are any instructions immediately before the unreachable that can
4122  // be removed, do so.
4123  while (UI->getIterator() != BB->begin()) {
4124  BasicBlock::iterator BBI = UI->getIterator();
4125  --BBI;
4126  // Do not delete instructions that can have side effects which might cause
4127  // the unreachable to not be reachable; specifically, calls and volatile
4128  // operations may have this effect.
4129  if (isa<CallInst>(BBI) && !isa<DbgInfoIntrinsic>(BBI))
4130  break;
4131 
4132  if (BBI->mayHaveSideEffects()) {
4133  if (auto *SI = dyn_cast<StoreInst>(BBI)) {
4134  if (SI->isVolatile())
4135  break;
4136  } else if (auto *LI = dyn_cast<LoadInst>(BBI)) {
4137  if (LI->isVolatile())
4138  break;
4139  } else if (auto *RMWI = dyn_cast<AtomicRMWInst>(BBI)) {
4140  if (RMWI->isVolatile())
4141  break;
4142  } else if (auto *CXI = dyn_cast<AtomicCmpXchgInst>(BBI)) {
4143  if (CXI->isVolatile())
4144  break;
4145  } else if (isa<CatchPadInst>(BBI)) {
4146  // A catchpad may invoke exception object constructors and such, which
4147  // in some languages can be arbitrary code, so be conservative by
4148  // default.
4149  // For CoreCLR, it just involves a type test, so can be removed.
4152  break;
4153  } else if (!isa<FenceInst>(BBI) && !isa<VAArgInst>(BBI) &&
4154  !isa<LandingPadInst>(BBI)) {
4155  break;
4156  }
4157  // Note that deleting LandingPad's here is in fact okay, although it
4158  // involves a bit of subtle reasoning. If this inst is a LandingPad,
4159  // all the predecessors of this block will be the unwind edges of Invokes,
4160  // and we can therefore guarantee this block will be erased.
4161  }
4162 
4163  // Delete this instruction (any uses are guaranteed to be dead)
4164  if (!BBI->use_empty())
4165  BBI->replaceAllUsesWith(UndefValue::get(BBI->getType()));
4166  BBI->eraseFromParent();
4167  Changed = true;
4168  }
4169 
4170  // If the unreachable instruction is the first in the block, take a gander
4171  // at all of the predecessors of this instruction, and simplify them.
4172  if (&BB->front() != UI)
4173  return Changed;
4174 
4176  for (unsigned i = 0, e = Preds.size(); i != e; ++i) {
4177  Instruction *TI = Preds[i]->getTerminator();
4178  IRBuilder<> Builder(TI);
4179  if (auto *BI = dyn_cast<BranchInst>(TI)) {
4180  if (BI->isUnconditional()) {
4181  if (BI->getSuccessor(0) == BB) {
4182  new UnreachableInst(TI->getContext(), TI);
4183  TI->eraseFromParent();
4184  Changed = true;
4185  }
4186  } else {
4187  Value* Cond = BI->getCondition();
4188  if (BI->getSuccessor(0) == BB) {
4189  Builder.CreateAssumption(Builder.CreateNot(Cond));
4190  Builder.CreateBr(BI->getSuccessor(1));
4192  } else if (BI->getSuccessor(1) == BB) {
4193  Builder.CreateAssumption(Cond);
4194  Builder.CreateBr(BI->getSuccessor(0));
4196  Changed = true;
4197  }
4198  }
4199  } else if (auto *SI = dyn_cast<SwitchInst>(TI)) {
4201  for (auto i = SU->case_begin(), e = SU->case_end(); i != e;) {
4202  if (i->getCaseSuccessor() != BB) {
4203  ++i;
4204  continue;
4205  }
4206  BB->removePredecessor(SU->getParent());
4207  i = SU.removeCase(i);
4208  e = SU->case_end();
4209  Changed = true;
4210  }
4211  } else if (auto *II = dyn_cast<InvokeInst>(TI)) {
4212  if (II->getUnwindDest() == BB) {
4213  removeUnwindEdge(TI->getParent());
4214  Changed = true;
4215  }
4216  } else if (auto *CSI = dyn_cast<CatchSwitchInst>(TI)) {
4217  if (CSI->getUnwindDest() == BB) {
4218  removeUnwindEdge(TI->getParent());
4219  Changed = true;
4220  continue;
4221  }
4222 
4223  for (CatchSwitchInst::handler_iterator I = CSI->handler_begin(),
4224  E = CSI->handler_end();
4225  I != E; ++I) {
4226  if (*I == BB) {
4227  CSI->removeHandler(I);
4228  --I;
4229  --E;
4230  Changed = true;
4231  }
4232  }
4233  if (CSI->getNumHandlers() == 0) {
4234  BasicBlock *CatchSwitchBB = CSI->getParent();
4235  if (CSI->hasUnwindDest()) {
4236  // Redirect preds to the unwind dest
4237  CatchSwitchBB->replaceAllUsesWith(CSI->getUnwindDest());
4238  } else {
4239  // Rewrite all preds to unwind to caller (or from invoke to call).
4240  SmallVector<BasicBlock *, 8> EHPreds(predecessors(CatchSwitchBB));
4241  for (BasicBlock *EHPred : EHPreds)
4242  removeUnwindEdge(EHPred);
4243  }
4244  // The catchswitch is no longer reachable.
4245  new UnreachableInst(CSI->getContext(), CSI);
4246  CSI->eraseFromParent();
4247  Changed = true;
4248  }
4249  } else if (isa<CleanupReturnInst>(TI)) {
4250  new UnreachableInst(TI->getContext(), TI);
4251  TI->eraseFromParent();
4252  Changed = true;
4253  }
4254  }
4255 
4256  // If this block is now dead, remove it.
4257  if (pred_empty(BB) && BB != &BB->getParent()->getEntryBlock()) {
4258  // We know there are no successors, so just nuke the block.
4259  if (LoopHeaders)
4260  LoopHeaders->erase(BB);
4261  BB->eraseFromParent();
4262  return true;
4263  }
4264 
4265  return Changed;
4266 }
4267 
4269  assert(Cases.size() >= 1);
4270 
4271  array_pod_sort(Cases.begin(), Cases.end(), ConstantIntSortPredicate);
4272  for (size_t I = 1, E = Cases.size(); I != E; ++I) {
4273  if (Cases[I - 1]->getValue() != Cases[I]->getValue() + 1)
4274  return false;
4275  }
4276  return true;
4277 }
4278 
4279 /// Turn a switch with two reachable destinations into an integer range
4280 /// comparison and branch.
4282  assert(SI->getNumCases() > 1 && "Degenerate switch?");
4283 
4284  bool HasDefault =
4285  !isa<UnreachableInst>(SI->getDefaultDest()->getFirstNonPHIOrDbg());
4286 
4287  // Partition the cases into two sets with different destinations.
4288  BasicBlock *DestA = HasDefault ? SI->getDefaultDest() : nullptr;
4289  BasicBlock *DestB = nullptr;
4292 
4293  for (auto Case : SI->cases()) {
4294  BasicBlock *Dest = Case.getCaseSuccessor();
4295  if (!DestA)
4296  DestA = Dest;
4297  if (Dest == DestA) {
4298  CasesA.push_back(Case.getCaseValue());
4299  continue;
4300  }
4301  if (!DestB)
4302  DestB = Dest;
4303  if (Dest == DestB) {
4304  CasesB.push_back(Case.getCaseValue());
4305  continue;
4306  }
4307  return false; // More than two destinations.
4308  }
4309 
4310  assert(DestA && DestB &&
4311  "Single-destination switch should have been folded.");
4312  assert(DestA != DestB);
4313  assert(DestB != SI->getDefaultDest());
4314  assert(!CasesB.empty() && "There must be non-default cases.");
4315  assert(!CasesA.empty() || HasDefault);
4316 
4317  // Figure out if one of the sets of cases form a contiguous range.
4318  SmallVectorImpl<ConstantInt *> *ContiguousCases = nullptr;
4319  BasicBlock *ContiguousDest = nullptr;
4320  BasicBlock *OtherDest = nullptr;
4321  if (!CasesA.empty() && CasesAreContiguous(CasesA)) {
4322  ContiguousCases = &CasesA;
4323  ContiguousDest = DestA;
4324  OtherDest = DestB;
4325  } else if (CasesAreContiguous(CasesB)) {
4326  ContiguousCases = &CasesB;
4327  ContiguousDest = DestB;
4328  OtherDest = DestA;
4329  } else
4330  return false;
4331 
4332  // Start building the compare and branch.
4333 
4334  Constant *Offset = ConstantExpr::getNeg(ContiguousCases->back());
4335  Constant *NumCases =
4336  ConstantInt::get(Offset->getType(), ContiguousCases->size());
4337 
4338  Value *Sub = SI->getCondition();
4339  if (!Offset->isNullValue())
4340  Sub = Builder.CreateAdd(Sub, Offset, Sub->getName() + ".off");
4341 
4342  Value *Cmp;
4343  // If NumCases overflowed, then all possible values jump to the successor.
4344  if (NumCases->isNullValue() && !ContiguousCases->empty())
4345  Cmp = ConstantInt::getTrue(SI->getContext());
4346  else
4347  Cmp = Builder.CreateICmpULT(Sub, NumCases, "switch");
4348  BranchInst *NewBI = Builder.CreateCondBr(Cmp, ContiguousDest, OtherDest);
4349 
4350  // Update weight for the newly-created conditional branch.
4351  if (HasBranchWeights(SI)) {
4352  SmallVector<uint64_t, 8> Weights;
4353  GetBranchWeights(SI, Weights);
4354  if (Weights.size() == 1 + SI->getNumCases()) {
4355  uint64_t TrueWeight = 0;
4356  uint64_t FalseWeight = 0;
4357  for (size_t I = 0, E = Weights.size(); I != E; ++I) {
4358  if (SI->getSuccessor(I) == ContiguousDest)
4359  TrueWeight += Weights[I];
4360  else
4361  FalseWeight += Weights[I];
4362  }
4363  while (TrueWeight > UINT32_MAX || FalseWeight > UINT32_MAX) {
4364  TrueWeight /= 2;
4365  FalseWeight /= 2;
4366  }
4367  setBranchWeights(NewBI, TrueWeight, FalseWeight);
4368  }
4369  }
4370 
4371  // Prune obsolete incoming values off the successors' PHI nodes.
4372  for (auto BBI = ContiguousDest->begin(); isa<PHINode>(BBI); ++BBI) {
4373  unsigned PreviousEdges = ContiguousCases->size();
4374  if (ContiguousDest == SI->getDefaultDest())
4375  ++PreviousEdges;
4376  for (unsigned I = 0, E = PreviousEdges - 1; I != E; ++I)
4377  cast<PHINode>(BBI)->removeIncomingValue(SI->getParent());
4378  }
4379  for (auto BBI = OtherDest->begin(); isa<PHINode>(BBI); ++BBI) {
4380  unsigned PreviousEdges = SI->getNumCases() - ContiguousCases->size();
4381  if (OtherDest == SI->getDefaultDest())
4382  ++PreviousEdges;
4383  for (unsigned I = 0, E = PreviousEdges - 1; I != E; ++I)
4384  cast<PHINode>(BBI)->removeIncomingValue(SI->getParent());
4385  }
4386 
4387  // Drop the switch.
4388  SI->eraseFromParent();
4389 
4390  return true;
4391 }
4392 
4393 /// Compute masked bits for the condition of a switch
4394 /// and use it to remove dead cases.
4396  const DataLayout &DL) {
4397  Value *Cond = SI->getCondition();
4398  unsigned Bits = Cond->getType()->getIntegerBitWidth();
4399  KnownBits Known = computeKnownBits(Cond, DL, 0, AC, SI);
4400 
4401  // We can also eliminate cases by determining that their values are outside of
4402  // the limited range of the condition based on how many significant (non-sign)
4403  // bits are in the condition value.
4404  unsigned ExtraSignBits = ComputeNumSignBits(Cond, DL, 0, AC, SI) - 1;
4405  unsigned MaxSignificantBitsInCond = Bits - ExtraSignBits;
4406 
4407  // Gather dead cases.
4409  for (auto &Case : SI->cases()) {
4410  const APInt &CaseVal = Case.getCaseValue()->getValue();
4411  if (Known.Zero.intersects(CaseVal) || !Known.One.isSubsetOf(CaseVal) ||
4412  (CaseVal.getMinSignedBits() > MaxSignificantBitsInCond)) {
4413  DeadCases.push_back(Case.getCaseValue());
4414  LLVM_DEBUG(dbgs() << "SimplifyCFG: switch case " << CaseVal
4415  << " is dead.\n");
4416  }
4417  }
4418 
4419  // If we can prove that the cases must cover all possible values, the
4420  // default destination becomes dead and we can remove it. If we know some
4421  // of the bits in the value, we can use that to more precisely compute the
4422  // number of possible unique case values.
4423  bool HasDefault =
4424  !isa<UnreachableInst>(SI->getDefaultDest()->getFirstNonPHIOrDbg());
4425  const unsigned NumUnknownBits =
4426  Bits - (Known.Zero | Known.One).countPopulation();
4427  assert(NumUnknownBits <= Bits);
4428  if (HasDefault && DeadCases.empty() &&
4429  NumUnknownBits < 64 /* avoid overflow */ &&
4430  SI->getNumCases() == (1ULL << NumUnknownBits)) {
4431  LLVM_DEBUG(dbgs() << "SimplifyCFG: switch default is dead.\n");
4432  BasicBlock *NewDefault =
4434  SI->setDefaultDest(&*NewDefault);
4435  SplitBlock(&*NewDefault, &NewDefault->front());
4436  auto *OldTI = NewDefault->getTerminator();
4437  new UnreachableInst(SI->getContext(), OldTI);
4439  return true;
4440  }
4441 
4442  if (DeadCases.empty())
4443  return false;
4444 
4445  SwitchInstProfUpdateWrapper SIW(*SI);
4446  for (ConstantInt *DeadCase : DeadCases) {
4447  SwitchInst::CaseIt CaseI = SI->findCaseValue(DeadCase);
4448  assert(CaseI != SI->case_default() &&
4449  "Case was not found. Probably mistake in DeadCases forming.");
4450  // Prune unused values from PHI nodes.
4451  CaseI->getCaseSuccessor()->removePredecessor(SI->getParent());
4452  SIW.removeCase(CaseI);
4453  }
4454 
4455  return true;
4456 }
4457 
4458 /// If BB would be eligible for simplification by
4459 /// TryToSimplifyUncondBranchFromEmptyBlock (i.e. it is empty and terminated
4460 /// by an unconditional branch), look at the phi node for BB in the successor
4461 /// block and see if the incoming value is equal to CaseValue. If so, return
4462 /// the phi node, and set PhiIndex to BB's index in the phi node.
4464  BasicBlock *BB, int *PhiIndex) {
4465  if (BB->getFirstNonPHIOrDbg() != BB->getTerminator())
4466  return nullptr; // BB must be empty to be a candidate for simplification.
4467  if (!BB->getSinglePredecessor())
4468  return nullptr; // BB must be dominated by the switch.
4469 
4471  if (!Branch || !Branch->isUnconditional())
4472  return nullptr; // Terminator must be unconditional branch.
4473 
4474  BasicBlock *Succ = Branch->getSuccessor(0);
4475 
4476  for (PHINode &PHI : Succ->phis()) {
4477  int Idx = PHI.getBasicBlockIndex(BB);
4478  assert(Idx >= 0 && "PHI has no entry for predecessor?");
4479 
4480  Value *InValue = PHI.getIncomingValue(Idx);
4481  if (InValue != CaseValue)
4482  continue;
4483 
4484  *PhiIndex = Idx;
4485  return &PHI;
4486  }
4487 
4488  return nullptr;
4489 }
4490 
4491 /// Try to forward the condition of a switch instruction to a phi node
4492 /// dominated by the switch, if that would mean that some of the destination
4493 /// blocks of the switch can be folded away. Return true if a change is made.
4495  using ForwardingNodesMap = DenseMap<PHINode *, SmallVector<int, 4>>;
4496 
4497  ForwardingNodesMap ForwardingNodes;
4498  BasicBlock *SwitchBlock = SI->getParent();
4499  bool Changed = false;
4500  for (auto &Case : SI->cases()) {
4501  ConstantInt *CaseValue = Case.getCaseValue();
4502  BasicBlock *CaseDest = Case.getCaseSuccessor();
4503 
4504  // Replace phi operands in successor blocks that are using the constant case
4505  // value rather than the switch condition variable:
4506  // switchbb:
4507  // switch i32 %x, label %default [
4508  // i32 17, label %succ
4509  // ...
4510  // succ:
4511  // %r = phi i32 ... [ 17, %switchbb ] ...
4512  // -->
4513  // %r = phi i32 ... [ %x, %switchbb ] ...
4514 
4515  for (PHINode &Phi : CaseDest->phis()) {
4516  // This only works if there is exactly 1 incoming edge from the switch to
4517  // a phi. If there is >1, that means multiple cases of the switch map to 1
4518  // value in the phi, and that phi value is not the switch condition. Thus,
4519  // this transform would not make sense (the phi would be invalid because
4520  // a phi can't have different incoming values from the same block).
4521  int SwitchBBIdx = Phi.getBasicBlockIndex(SwitchBlock);
4522  if (Phi.getIncomingValue(SwitchBBIdx) == CaseValue &&
4523  count(Phi.blocks(), SwitchBlock) == 1) {
4524  Phi.setIncomingValue(SwitchBBIdx, SI->getCondition());
4525  Changed = true;
4526  }
4527  }
4528 
4529  // Collect phi nodes that are indirectly using this switch's case constants.
4530  int PhiIdx;
4531  if (auto *Phi = FindPHIForConditionForwarding(CaseValue, CaseDest, &PhiIdx))
4532  ForwardingNodes[Phi].push_back(PhiIdx);
4533  }
4534 
4535  for (auto &ForwardingNode : ForwardingNodes) {
4536  PHINode *Phi = ForwardingNode.first;
4537  SmallVectorImpl<int> &Indexes = ForwardingNode.second;
4538  if (Indexes.size() < 2)
4539  continue;
4540 
4541  for (int Index : Indexes)
4542  Phi->setIncomingValue(Index, SI->getCondition());
4543  Changed = true;
4544  }
4545 
4546  return Changed;
4547 }
4548 
4549 /// Return true if the backend will be able to handle
4550 /// initializing an array of constants like C.
4552  if (C->isThreadDependent())
4553  return false;
4554  if (C->isDLLImportDependent())
4555  return false;
4556 
4557  if (!isa<ConstantFP>(C) && !isa<ConstantInt>(C) &&
4558  !isa<ConstantPointerNull>(C) && !isa<GlobalValue>(C) &&
4559  !isa<UndefValue>(C) && !isa<ConstantExpr>(C))
4560  return false;
4561 
4562  if (ConstantExpr *CE = dyn_cast<ConstantExpr>(C)) {
4563  if (!CE->isGEPWithNoNotionalOverIndexing())
4564  return false;
4565  if (!ValidLookupTableConstant(CE->getOperand(0), TTI))
4566  return false;
4567  }
4568 
4570  return false;
4571 
4572  return true;
4573 }
4574 
4575 /// If V is a Constant, return it. Otherwise, try to look up
4576 /// its constant value in ConstantPool, returning 0 if it's not there.
4577 static Constant *
4580  if (Constant *C = dyn_cast<Constant>(V))
4581  return C;
4582  return ConstantPool.lookup(V);
4583 }
4584 
4585 /// Try to fold instruction I into a constant. This works for
4586 /// simple instructions such as binary operations where both operands are
4587 /// constant or can be replaced by constants from the ConstantPool. Returns the
4588 /// resulting constant on success, 0 otherwise.
4589 static Constant *
4592  if (SelectInst *Select = dyn_cast<SelectInst>(I)) {
4593  Constant *A = LookupConstant(Select->getCondition(), ConstantPool);
4594  if (!A)
4595  return nullptr;
4596  if (A->isAllOnesValue())
4597  return LookupConstant(Select->getTrueValue(), ConstantPool);
4598  if (A->isNullValue())
4599  return LookupConstant(Select->getFalseValue(), ConstantPool);
4600  return nullptr;
4601  }
4602 
4604  for (unsigned N = 0, E = I->getNumOperands(); N != E; ++N) {
4606  COps.push_back(A);
4607  else
4608  return nullptr;
4609  }
4610 
4611  if (CmpInst *Cmp = dyn_cast<CmpInst>(I)) {
4612  return ConstantFoldCompareInstOperands(Cmp->getPredicate(), COps[0],
4613  COps[1], DL);
4614  }
4615 
4616  return ConstantFoldInstOperands(I, COps, DL);
4617 }
4618 
4619 /// Try to determine the resulting constant values in phi nodes
4620 /// at the common destination basic block, *CommonDest, for one of the case
4621 /// destionations CaseDest corresponding to value CaseVal (0 for the default
4622 /// case), of a switch instruction SI.
4623 static bool
4625  BasicBlock **CommonDest,
4626  SmallVectorImpl<std::pair<PHINode *, Constant *>> &Res,
4627  const DataLayout &DL, const TargetTransformInfo &TTI) {
4628  // The block from which we enter the common destination.
4629  BasicBlock *Pred = SI->getParent();
4630 
4631  // If CaseDest is empty except for some side-effect free instructions through
4632  // which we can constant-propagate the CaseVal, continue to its successor.
4634  ConstantPool.insert(std::make_pair(SI->getCondition(), CaseVal));
4635  for (Instruction &I :CaseDest->instructionsWithoutDebug()) {
4636  if (I.isTerminator()) {
4637  // If the terminator is a simple branch, continue to the next block.
4638  if (I.getNumSuccessors() != 1 || I.isExceptionalTerminator())
4639  return false;
4640  Pred = CaseDest;
4641  CaseDest = I.getSuccessor(0);
4642  } else if (Constant *C = ConstantFold(&I, DL, ConstantPool)) {
4643  // Instruction is side-effect free and constant.
4644 
4645  // If the instruction has uses outside this block or a phi node slot for
4646  // the block, it is not safe to bypass the instruction since it would then
4647  // no longer dominate all its uses.
4648  for (auto &Use : I.uses()) {
4649  User *User = Use.getUser();
4650  if (Instruction *I = dyn_cast<Instruction>(User))
4651  if (I->getParent() == CaseDest)
4652  continue;
4653  if (PHINode *Phi = dyn_cast<PHINode>(User))
4654  if (Phi->getIncomingBlock(Use) == CaseDest)
4655  continue;
4656  return false;
4657  }
4658 
4659  ConstantPool.insert(std::make_pair(&I, C));
4660  } else {
4661  break;
4662  }
4663  }
4664 
4665  // If we did not have a CommonDest before, use the current one.
4666  if (!*CommonDest)
4667  *CommonDest = CaseDest;
4668  // If the destination isn't the common one, abort.
4669  if (CaseDest != *CommonDest)
4670  return false;
4671 
4672  // Get the values for this case from phi nodes in the destination block.
4673  for (PHINode &PHI : (*CommonDest)->phis()) {
4674  int Idx = PHI.getBasicBlockIndex(Pred);
4675  if (Idx == -1)
4676  continue;
4677 
4678  Constant *ConstVal =
4679  LookupConstant(PHI.getIncomingValue(Idx), ConstantPool);
4680  if (!ConstVal)
4681  return false;
4682 
4683  // Be conservative about which kinds of constants we support.
4684  if (!ValidLookupTableConstant(ConstVal, TTI))
4685  return false;
4686 
4687  Res.push_back(std::make_pair(&PHI, ConstVal));
4688  }
4689 
4690  return Res.size() > 0;
4691 }
4692 
4693 // Helper function used to add CaseVal to the list of cases that generate
4694 // Result. Returns the updated number of cases that generate this result.
4695 static uintptr_t MapCaseToResult(ConstantInt *CaseVal,
4696  SwitchCaseResultVectorTy &UniqueResults,
4697  Constant *Result) {
4698  for (auto &I : UniqueResults) {
4699  if (I.first == Result) {
4700  I.second.push_back(CaseVal);
4701  return I.second.size();
4702  }
4703  }
4704  UniqueResults.push_back(
4705  std::make_pair(Result, SmallVector<ConstantInt *, 4>(1, CaseVal)));
4706  return 1;
4707 }
4708 
4709 // Helper function that initializes a map containing
4710 // results for the PHI node of the common destination block for a switch
4711 // instruction. Returns false if multiple PHI nodes have been found or if
4712 // there is not a common destination block for the switch.
4713 static bool
4715  SwitchCaseResultVectorTy &UniqueResults,
4716  Constant *&DefaultResult, const DataLayout &DL,
4717  const TargetTransformInfo &TTI,
4718  uintptr_t MaxUniqueResults, uintptr_t MaxCasesPerResult) {
4719  for (auto &I : SI->cases()) {
4720  ConstantInt *CaseVal = I.getCaseValue();
4721 
4722  // Resulting value at phi nodes for this case value.
4723  SwitchCaseResultsTy Results;
4724  if (!GetCaseResults(SI, CaseVal, I.getCaseSuccessor(), &CommonDest, Results,
4725  DL, TTI))
4726  return false;
4727 
4728  // Only one value per case is permitted.
4729  if (Results.size() > 1)
4730  return false;
4731 
4732  // Add the case->result mapping to UniqueResults.
4733  const uintptr_t NumCasesForResult =
4734  MapCaseToResult(CaseVal, UniqueResults, Results.begin()->second);
4735 
4736  // Early out if there are too many cases for this result.
4737  if (NumCasesForResult > MaxCasesPerResult)
4738  return false;
4739 
4740  // Early out if there are too many unique results.
4741  if (UniqueResults.size() > MaxUniqueResults)
4742  return false;
4743 
4744  // Check the PHI consistency.
4745  if (!PHI)
4746  PHI = Results[0].first;
4747  else if (PHI != Results[0].first)
4748  return false;
4749  }
4750  // Find the default result value.
4752  BasicBlock *DefaultDest = SI->getDefaultDest();
4753  GetCaseResults(SI, nullptr, SI->getDefaultDest(), &CommonDest, DefaultResults,
4754  DL, TTI);
4755  // If the default value is not found abort unless the default destination
4756  // is unreachable.
4757  DefaultResult =
4758  DefaultResults.size() == 1 ? DefaultResults.begin()->second : nullptr;
4759  if ((!DefaultResult &&
4760  !isa<UnreachableInst>(DefaultDest->getFirstNonPHIOrDbg())))
4761  return false;
4762 
4763  return true;
4764 }
4765 
4766 // Helper function that checks if it is possible to transform a switch with only
4767 // two cases (or two cases + default) that produces a result into a select.
4768 // Example:
4769 // switch (a) {
4770 // case 10: %0 = icmp eq i32 %a, 10
4771 // return 10; %1 = select i1 %0, i32 10, i32 4
4772 // case 20: ----> %2 = icmp eq i32 %a, 20
4773 // return 2; %3 = select i1 %2, i32 2, i32 %1
4774 // default:
4775 // return 4;
4776 // }
4777 static Value *ConvertTwoCaseSwitch(const SwitchCaseResultVectorTy &ResultVector,
4778  Constant *DefaultResult, Value *Condition,
4779  IRBuilder<> &Builder) {
4780  assert(ResultVector.size() == 2 &&
4781  "We should have exactly two unique results at this point");
4782  // If we are selecting between only two cases transform into a simple
4783  // select or a two-way select if default is possible.
4784  if (ResultVector[0].second.size() == 1 &&
4785  ResultVector[1].second.size() == 1) {
4786  ConstantInt *const FirstCase = ResultVector[0].second[0];
4787  ConstantInt *const SecondCase = ResultVector[1].second[0];
4788 
4789  bool DefaultCanTrigger = DefaultResult;
4790  Value *SelectValue = ResultVector[1].first;
4791  if (DefaultCanTrigger) {
4792  Value *const ValueCompare =
4793  Builder.CreateICmpEQ(Condition, SecondCase, "switch.selectcmp");
4794  SelectValue = Builder.CreateSelect(ValueCompare, ResultVector[1].first,
4795  DefaultResult, "switch.select");
4796  }
4797  Value *const ValueCompare =
4798  Builder.CreateICmpEQ(Condition, FirstCase, "switch.selectcmp");
4799  return Builder.CreateSelect(ValueCompare, ResultVector[0].first,
4800  SelectValue, "switch.select");
4801  }
4802 
4803  return nullptr;
4804 }
4805 
4806 // Helper function to cleanup a switch instruction that has been converted into
4807 // a select, fixing up PHI nodes and basic blocks.
4809  Value *SelectValue,
4810  IRBuilder<> &Builder) {
4811  BasicBlock *SelectBB = SI->getParent();
4812  while (PHI->getBasicBlockIndex(SelectBB) >= 0)
4813  PHI->removeIncomingValue(SelectBB);
4814  PHI->addIncoming(SelectValue, SelectBB);
4815 
4816  Builder.CreateBr(PHI->getParent());
4817 
4818  // Remove the switch.
4819  for (unsigned i = 0, e = SI->getNumSuccessors(); i < e; ++i) {
4820  BasicBlock *Succ = SI->getSuccessor(i);
4821 
4822  if (Succ == PHI->getParent())
4823  continue;
4824  Succ->removePredecessor(SelectBB);
4825  }
4826  SI->eraseFromParent();
4827 }
4828 
4829 /// If the switch is only used to initialize one or more
4830 /// phi nodes in a common successor block with only two different
4831 /// constant values, replace the switch with select.
4832 static bool switchToSelect(SwitchInst *SI, IRBuilder<> &Builder,
4833  const DataLayout &DL,
4834  const TargetTransformInfo &TTI) {
4835  Value *const Cond = SI->getCondition();
4836  PHINode *PHI = nullptr;
4837  BasicBlock *CommonDest = nullptr;
4838  Constant *DefaultResult;
4839  SwitchCaseResultVectorTy UniqueResults;
4840  // Collect all the cases that will deliver the same value from the switch.
4841  if (!InitializeUniqueCases(SI, PHI, CommonDest, UniqueResults, DefaultResult,
4842  DL, TTI, 2, 1))
4843  return false;
4844  // Selects choose between maximum two values.
4845  if (UniqueResults.size() != 2)
4846  return false;
4847  assert(PHI != nullptr && "PHI for value select not found");
4848 
4849  Builder.SetInsertPoint(SI);
4850  Value *SelectValue =
4851  ConvertTwoCaseSwitch(UniqueResults, DefaultResult, Cond, Builder);
4852  if (SelectValue) {
4853  RemoveSwitchAfterSelectConversion(SI, PHI, SelectValue, Builder);
4854  return true;
4855  }
4856  // The switch couldn't be converted into a select.
4857  return false;
4858 }
4859 
4860 namespace {
4861 
4862 /// This class represents a lookup table that can be used to replace a switch.
4863 class SwitchLookupTable {
4864 public:
4865  /// Create a lookup table to use as a switch replacement with the contents
4866  /// of Values, using DefaultValue to fill any holes in the table.
4867  SwitchLookupTable(
4868  Module &M, uint64_t TableSize, ConstantInt *Offset,
4869  const SmallVectorImpl<std::pair<ConstantInt *, Constant *>> &Values,
4870  Constant *DefaultValue, const DataLayout &DL, const StringRef &FuncName);
4871 
4872  /// Build instructions with Builder to retrieve the value at
4873  /// the position given by Index in the lookup table.
4874  Value *BuildLookup(Value *Index, IRBuilder<> &Builder);
4875 
4876  /// Return true if a table with TableSize elements of
4877  /// type ElementType would fit in a target-legal register.
4878  static bool WouldFitInRegister(const DataLayout &DL, uint64_t TableSize,
4879  Type *ElementType);
4880 
4881 private:
4882  // Depending on the contents of the table, it can be represented in
4883  // different ways.
4884  enum {
4885  // For tables where each element contains the same value, we just have to
4886  // store that single value and return it for each lookup.
4887  SingleValueKind,
4888 
4889  // For tables where there is a linear relationship between table index
4890  // and values. We calculate the result with a simple multiplication
4891  // and addition instead of a table lookup.
4892  LinearMapKind,
4893 
4894  // For small tables with integer elements, we can pack them into a bitmap
4895  // that fits into a target-legal register. Values are retrieved by
4896  // shift and mask operations.
4897  BitMapKind,
4898 
4899  // The table is stored as an array of values. Values are retrieved by load
4900  // instructions from the table.
4901  ArrayKind
4902  } Kind;
4903 
4904  // For SingleValueKind, this is the single value.
4905  Constant *SingleValue = nullptr;
4906 
4907  // For BitMapKind, this is the bitmap.
4908  ConstantInt *BitMap = nullptr;
4909  IntegerType *BitMapElementTy = nullptr;
4910 
4911  // For LinearMapKind, these are the constants used to derive the value.
4912  ConstantInt *LinearOffset = nullptr;
4913  ConstantInt *LinearMultiplier = nullptr;
4914 
4915  // For ArrayKind, this is the array.
4916  GlobalVariable *Array = nullptr;
4917 };
4918 
4919 } // end anonymous namespace
4920 
4921 SwitchLookupTable::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  assert(Values.size() && "Can't build lookup table without values!");
4926  assert(TableSize >= Values.size() && "Can't fit values in table!");
4927 
4928  // If all values in the table are equal, this is that value.
4929  SingleValue = Values.begin()->second;
4930 
4931  Type *ValueType = Values.begin()->second->getType();
4932 
4933  // Build up the table contents.
4934  SmallVector<Constant *, 64> TableContents(TableSize);
4935  for (size_t I = 0, E = Values.size(); I != E; ++I) {
4936  ConstantInt *CaseVal = Values[I].first;
4937  Constant *CaseRes = Values[I].second;
4938  assert(CaseRes->getType() == ValueType);
4939 
4940  uint64_t Idx = (CaseVal->getValue() - Offset->getValue()).getLimitedValue();
4941  TableContents[Idx] = CaseRes;
4942 
4943  if (CaseRes != SingleValue)
4944  SingleValue = nullptr;
4945  }
4946 
4947  // Fill in any holes in the table with the default result.
4948  if (Values.size() < TableSize) {
4949  assert(DefaultValue &&
4950  "Need a default value to fill the lookup table holes.");
4951  assert(DefaultValue->getType() == ValueType);
4952  for (uint64_t I = 0; I < TableSize; ++I) {
4953  if (!TableContents[I])
4954  TableContents[I] = DefaultValue;
4955  }
4956 
4957  if (DefaultValue != SingleValue)
4958  SingleValue = nullptr;
4959  }
4960 
4961  // If each element in the table contains the same value, we only need to store
4962  // that single value.
4963  if (SingleValue) {
4964  Kind = SingleValueKind;
4965  return;
4966  }
4967 
4968  // Check if we can derive the value with a linear transformation from the
4969  // table index.
4970  if (isa<IntegerType>(ValueType)) {
4971  bool LinearMappingPossible = true;
4972  APInt PrevVal;
4973  APInt DistToPrev;
4974  assert(TableSize >= 2 && "Should be a SingleValue table.");
4975  // Check if there is the same distance between two consecutive values.
4976  for (uint64_t I = 0; I < TableSize; ++I) {
4977  ConstantInt *ConstVal = dyn_cast<ConstantInt>(TableContents[I]);
4978  if (!ConstVal) {
4979  // This is an undef. We could deal with it, but undefs in lookup tables
4980  // are very seldom. It's probably not worth the additional complexity.
4981  LinearMappingPossible = false;
4982  break;
4983  }
4984  const APInt &Val = ConstVal->getValue();
4985  if (I != 0) {
4986  APInt Dist = Val - PrevVal;
4987  if (I == 1) {
4988  DistToPrev = Dist;
4989  } else if (Dist != DistToPrev) {
4990  LinearMappingPossible = false;
4991  break;
4992  }
4993  }
4994  PrevVal = Val;
4995  }
4996  if (LinearMappingPossible) {
4997  LinearOffset = cast<ConstantInt>(TableContents[0]);
4998  LinearMultiplier = ConstantInt::get(M.getContext(), DistToPrev);
4999  Kind = LinearMapKind;
5000  ++NumLinearMaps;
5001  return;
5002  }
5003  }
5004 
5005  // If the type is integer and the table fits in a register, build a bitmap.
5006  if (WouldFitInRegister(DL, TableSize, ValueType)) {
5007  IntegerType *IT = cast<IntegerType>(ValueType);
5008  APInt TableInt(TableSize * IT->getBitWidth(), 0);
5009  for (uint64_t I = TableSize; I > 0; --I) {
5010  TableInt <<= IT->getBitWidth();
5011  // Insert values into the bitmap. Undef values are set to zero.
5012  if (!isa<UndefValue>(TableContents[I - 1])) {
5013  ConstantInt *Val = cast<ConstantInt>(TableContents[I - 1]);
5014  TableInt |= Val->getValue().zext(TableInt.getBitWidth());
5015  }
5016  }
5017  BitMap = ConstantInt::get(M.getContext(), TableInt);
5018  BitMapElementTy = IT;
5019  Kind = BitMapKind;
5020  ++NumBitMaps;
5021  return;
5022  }
5023 
5024  // Store the table in an array.
5025  ArrayType *ArrayTy = ArrayType::get(ValueType, TableSize);
5026  Constant *Initializer = ConstantArray::get(ArrayTy, TableContents);
5027 
5028  Array = new GlobalVariable(M, ArrayTy, /*constant=*/true,
5029  GlobalVariable::PrivateLinkage, Initializer,
5030  "switch.table." + FuncName);
5031  Array->setUnnamedAddr(GlobalValue::UnnamedAddr::Global);
5032  // Set the alignment to that of an array items. We will be only loading one
5033  // value out of it.
5034  Array->setAlignment(DL.getPrefTypeAlignment(ValueType));
5035  Kind = ArrayKind;
5036 }
5037 
5038 Value *SwitchLookupTable::BuildLookup(Value *Index, IRBuilder<> &Builder) {
5039  switch (Kind) {
5040  case SingleValueKind:
5041  return SingleValue;
5042  case LinearMapKind: {
5043  // Derive the result value from the input value.
5044  Value *Result = Builder.CreateIntCast(Index, LinearMultiplier->getType(),
5045  false, "switch.idx.cast");
5046  if (!LinearMultiplier->isOne())
5047  Result = Builder.CreateMul(Result, LinearMultiplier, "switch.idx.mult");
5048  if (!LinearOffset->isZero())
5049  Result = Builder.CreateAdd(Result, LinearOffset, "switch.offset");
5050  return Result;
5051  }
5052  case BitMapKind: {
5053  // Type of the bitmap (e.g. i59).
5054  IntegerType *MapTy = BitMap->getType();
5055 
5056  // Cast Index to the same type as the bitmap.
5057  // Note: The Index is <= the number of elements in the table, so
5058  // truncating it to the width of the bitmask is safe.
5059  Value *ShiftAmt = Builder.CreateZExtOrTrunc(Index, MapTy, "switch.cast");
5060 
5061  // Multiply the shift amount by the element width.
5062  ShiftAmt = Builder.CreateMul(
5063  ShiftAmt, ConstantInt::get(MapTy, BitMapElementTy->getBitWidth()),
5064  "switch.shiftamt");
5065 
5066  // Shift down.
5067  Value *DownShifted =
5068  Builder.CreateLShr(BitMap, ShiftAmt, "switch.downshift");
5069  // Mask off.
5070  return Builder.CreateTrunc(DownShifted, BitMapElementTy, "switch.masked");
5071  }
5072  case ArrayKind: {
5073  // Make sure the table index will not overflow when treated as signed.
5074  IntegerType *IT = cast<IntegerType>(Index->getType());
5075  uint64_t TableSize =
5076  Array->getInitializer()->getType()->getArrayNumElements();
5077  if (TableSize > (1ULL << (IT->getBitWidth() - 1)))
5078  Index = Builder.CreateZExt(
5079  Index, IntegerType::get(IT->getContext(), IT->getBitWidth() + 1),
5080  "switch.tableidx.zext");
5081 
5082  Value *GEPIndices[] = {Builder.getInt32(0), Index};
5083  Value *GEP = Builder.CreateInBoundsGEP(Array->getValueType(), Array,
5084  GEPIndices, "switch.gep");
5085  return Builder.CreateLoad(
5086  cast<ArrayType>(Array->getValueType())->getElementType(), GEP,
5087  "switch.load");
5088  }
5089  }
5090  llvm_unreachable("Unknown lookup table kind!");
5091 }
5092 
5093 bool SwitchLookupTable::WouldFitInRegister(const DataLayout &DL,
5094  uint64_t TableSize,
5095  Type *ElementType) {
5096  auto *IT = dyn_cast<IntegerType>(ElementType);
5097  if (!IT)
5098  return false;
5099  // FIXME: If the type is wider than it needs to be, e.g. i8 but all values
5100  // are <= 15, we could try to narrow the type.
5101 
5102  // Avoid overflow, fitsInLegalInteger uses unsigned int for the width.
5103  if (TableSize >= UINT_MAX / IT->getBitWidth())
5104  return false;
5105  return DL.fitsInLegalInteger(TableSize * IT->getBitWidth());
5106 }
5107 
5108 /// Determine whether a lookup table should be built for this switch, based on
5109 /// the number of cases, size of the table, and the types of the results.
5110 static bool
5111 ShouldBuildLookupTable(SwitchInst *SI, uint64_t TableSize,
5112  const TargetTransformInfo &TTI, const DataLayout &DL,
5113  const SmallDenseMap<PHINode *, Type *> &ResultTypes) {
5114  if (SI->getNumCases() > TableSize || TableSize >= UINT64_MAX / 10)
5115  return false; // TableSize overflowed, or mul below might overflow.
5116 
5117  bool AllTablesFitInRegister = true;
5118  bool HasIllegalType = false;
5119  for (const auto &I : ResultTypes) {
5120  Type *Ty = I.second;
5121 
5122  // Saturate this flag to true.
5123  HasIllegalType = HasIllegalType || !TTI.isTypeLegal(Ty);
5124 
5125  // Saturate this flag to false.
5126  AllTablesFitInRegister =
5127  AllTablesFitInRegister &&
5128  SwitchLookupTable::WouldFitInRegister(DL, TableSize, Ty);
5129 
5130  // If both flags saturate, we're done. NOTE: This *only* works with
5131  // saturating flags, and all flags have to saturate first due to the
5132  // non-deterministic behavior of iterating over a dense map.
5133  if (HasIllegalType && !AllTablesFitInRegister)
5134  break;
5135  }
5136 
5137  // If each table would fit in a register, we should build it anyway.
5138  if (AllTablesFitInRegister)
5139  return true;
5140 
5141  // Don't build a table that doesn't fit in-register if it has illegal types.
5142  if (HasIllegalType)
5143  return false;
5144 
5145  // The table density should be at least 40%. This is the same criterion as for
5146  // jump tables, see SelectionDAGBuilder::handleJTSwitchCase.
5147  // FIXME: Find the best cut-off.
5148  return SI->getNumCases() * 10 >= TableSize * 4;
5149 }
5150 
5151 /// Try to reuse the switch table index compare. Following pattern:
5152 /// \code
5153 /// if (idx < tablesize)
5154 /// r = table[idx]; // table does not contain default_value
5155 /// else
5156 /// r = default_value;
5157 /// if (r != default_value)
5158 /// ...
5159 /// \endcode
5160 /// Is optimized to:
5161 /// \code
5162 /// cond = idx < tablesize;
5163 /// if (cond)
5164 /// r = table[idx];
5165 /// else
5166 /// r = default_value;
5167 /// if (cond)
5168 /// ...
5169 /// \endcode
5170 /// Jump threading will then eliminate the second if(cond).
5171 static void reuseTableCompare(
5172  User *PhiUser, BasicBlock *PhiBlock, BranchInst *RangeCheckBranch,
5173  Constant *DefaultValue,
5174  const SmallVectorImpl<std::pair<ConstantInt *, Constant *>> &Values) {
5175  ICmpInst *CmpInst = dyn_cast<ICmpInst>(PhiUser);
5176  if (!CmpInst)
5177  return;
5178 
5179  // We require that the compare is in the same block as the phi so that jump
5180  // threading can do its work afterwards.
5181  if (CmpInst->getParent() != PhiBlock)
5182  return;
5183 
5184  Constant *CmpOp1 = dyn_cast<Constant>(CmpInst->getOperand(1));
5185  if (!CmpOp1)
5186  return;
5187 
5188  Value *RangeCmp = RangeCheckBranch->getCondition();
5189  Constant *TrueConst = ConstantInt::getTrue(RangeCmp->getType());
5190  Constant *FalseConst = ConstantInt::getFalse(RangeCmp->getType());
5191 
5192  // Check if the compare with the default value is constant true or false.
5193  Constant *DefaultConst = ConstantExpr::getICmp(CmpInst->getPredicate(),
5194  DefaultValue, CmpOp1, true);
5195  if (DefaultConst != TrueConst && DefaultConst != FalseConst)
5196  return;
5197 
5198  // Check if the compare with the case values is distinct from the default
5199  // compare result.
5200  for (auto ValuePair : Values) {
5201  Constant *CaseConst = ConstantExpr::getICmp(CmpInst->getPredicate(),
5202  ValuePair.second, CmpOp1, true);
5203  if (!CaseConst || CaseConst == DefaultConst || isa<UndefValue>(CaseConst))
5204  return;
5205  assert((CaseConst == TrueConst || CaseConst == FalseConst) &&
5206  "Expect true or false as compare result.");
5207  }
5208 
5209  // Check if the branch instruction dominates the phi node. It's a simple
5210  // dominance check, but sufficient for our needs.
5211  // Although this check is invariant in the calling loops, it's better to do it
5212  // at this late stage. Practically we do it at most once for a switch.
5213  BasicBlock *BranchBlock = RangeCheckBranch->getParent();
5214  for (auto PI = pred_begin(PhiBlock), E = pred_end(PhiBlock); PI != E; ++PI) {
5215  BasicBlock *Pred = *PI;
5216  if (Pred != BranchBlock && Pred->getUniquePredecessor() != BranchBlock)
5217  return;
5218  }
5219 
5220  if (DefaultConst == FalseConst) {
5221  // The compare yields the same result. We can replace it.
5222  CmpInst->replaceAllUsesWith(RangeCmp);
5223  ++NumTableCmpReuses;
5224  } else {
5225  // The compare yields the same result, just inverted. We can replace it.
5226  Value *InvertedTableCmp = BinaryOperator::CreateXor(
5227  RangeCmp, ConstantInt::get(RangeCmp->getType(), 1), "inverted.cmp",
5228  RangeCheckBranch);
5229  CmpInst->replaceAllUsesWith(InvertedTableCmp);
5230  ++NumTableCmpReuses;
5231  }
5232 }
5233 
5234 /// If the switch is only used to initialize one or more phi nodes in a common
5235 /// successor block with different constant values, replace the switch with
5236 /// lookup tables.
5238  const DataLayout &DL,
5239  const TargetTransformInfo &TTI) {
5240  assert(SI->getNumCases() > 1 && "Degenerate switch?");
5241 
5242  Function *Fn = SI->getParent()->getParent();
5243  // Only build lookup table when we have a target that supports it or the
5244  // attribute is not set.
5245  if (!TTI.shouldBuildLookupTables() ||
5246  (Fn->getFnAttribute("no-jump-tables").getValueAsString() == "true"))
5247  return false;
5248 
5249  // FIXME: If the switch is too sparse for a lookup table, perhaps we could
5250  // split off a dense part and build a lookup table for that.
5251 
5252  // FIXME: This creates arrays of GEPs to constant strings, which means each
5253  // GEP needs a runtime relocation in PIC code. We should just build one big
5254  // string and lookup indices into that.
5255 
5256  // Ignore switches with less than three cases. Lookup tables will not make
5257  // them faster, so we don't analyze them.
5258  if (SI->getNumCases() < 3)
5259  return false;
5260 
5261  // Figure out the corresponding result for each case value and phi node in the
5262  // common destination, as well as the min and max case values.
5263  assert(!empty(SI->cases()));
5264  SwitchInst::CaseIt CI = SI->case_begin();
5265  ConstantInt *MinCaseVal = CI->getCaseValue();
5266  ConstantInt *MaxCaseVal = CI->getCaseValue();
5267 
5268  BasicBlock *CommonDest = nullptr;
5269 
5270  using ResultListTy = SmallVector<std::pair<ConstantInt *, Constant *>, 4>;
5272 
5273  SmallDenseMap<PHINode *, Constant *> DefaultResults;
5276 
5277  for (SwitchInst::CaseIt E = SI->case_end(); CI != E; ++CI) {
5278  ConstantInt *CaseVal = CI->getCaseValue();
5279  if (CaseVal->getValue().slt(MinCaseVal->getValue()))
5280  MinCaseVal = CaseVal;
5281  if (CaseVal->getValue().sgt(MaxCaseVal->getValue()))
5282  MaxCaseVal = CaseVal;
5283 
5284  // Resulting value at phi nodes for this case value.
5285  using ResultsTy = SmallVector<std::pair<PHINode *, Constant *>, 4>;
5286  ResultsTy Results;
5287  if (!GetCaseResults(SI, CaseVal, CI->getCaseSuccessor(), &CommonDest,
5288  Results, DL, TTI))
5289  return false;
5290 
5291  // Append the result from this case to the list for each phi.
5292  for (const auto &I : Results) {
5293  PHINode *PHI = I.first;
5294  Constant *Value = I.second;
5295  if (!ResultLists.count(PHI))
5296  PHIs.push_back(PHI);
5297  ResultLists[PHI].push_back(std::make_pair(CaseVal, Value));
5298  }
5299  }
5300 
5301  // Keep track of the result types.
5302  for (PHINode *PHI : PHIs) {
5303  ResultTypes[PHI] = ResultLists[PHI][0].second->getType();
5304  }
5305 
5306  uint64_t NumResults = ResultLists[PHIs[0]].size();
5307  APInt RangeSpread = MaxCaseVal->getValue() - MinCaseVal->getValue();
5308  uint64_t TableSize = RangeSpread.getLimitedValue() + 1;
5309  bool TableHasHoles = (NumResults < TableSize);
5310 
5311  // If the table has holes, we need a constant result for the default case
5312  // or a bitmask that fits in a register.
5313  SmallVector<std::pair<PHINode *, Constant *>, 4> DefaultResultsList;
5314  bool HasDefaultResults =
5315  GetCaseResults(SI, nullptr, SI->getDefaultDest(), &CommonDest,
5316  DefaultResultsList, DL, TTI);
5317 
5318  bool NeedMask = (TableHasHoles && !HasDefaultResults);
5319  if (NeedMask) {
5320  // As an extra penalty for the validity test we require more cases.
5321  if (SI->getNumCases() < 4) // FIXME: Find best threshold value (benchmark).
5322  return false;
5323  if (!DL.fitsInLegalInteger(TableSize))
5324  return false;
5325  }
5326 
5327  for (const auto &I : DefaultResultsList) {
5328  PHINode *PHI = I.first;
5329  Constant *Result = I.second;
5330  DefaultResults[PHI] = Result;
5331  }
5332 
5333  if (!ShouldBuildLookupTable(SI, TableSize, TTI, DL, ResultTypes))
5334  return false;
5335 
5336  // Create the BB that does the lookups.
5337  Module &Mod = *CommonDest->getParent()->getParent();
5338  BasicBlock *LookupBB = BasicBlock::Create(
5339  Mod.getContext(), "switch.lookup", CommonDest->getParent(), CommonDest);
5340 
5341  // Compute the table index value.
5342  Builder.SetInsertPoint(SI);
5343  Value *TableIndex;
5344  if (MinCaseVal->isNullValue())
5345  TableIndex = SI->getCondition();
5346  else
5347  TableIndex = Builder.CreateSub(SI->getCondition(), MinCaseVal,
5348  "switch.tableidx");
5349 
5350  // Compute the maximum table size representable by the integer type we are
5351  // switching upon.
5352  unsigned CaseSize = MinCaseVal->getType()->getPrimitiveSizeInBits();
5353  uint64_t MaxTableSize = CaseSize > 63 ? UINT64_MAX : 1ULL << CaseSize;
5354  assert(MaxTableSize >= TableSize &&
5355  "It is impossible for a switch to have more entries than the max "
5356  "representable value of its input integer type's size.");
5357 
5358  // If the default destination is unreachable, or if the lookup table covers
5359  // all values of the conditional variable, branch directly to the lookup table
5360  // BB. Otherwise, check that the condition is within the case range.
5361  const bool DefaultIsReachable =
5362  !isa<UnreachableInst>(SI->getDefaultDest()->getFirstNonPHIOrDbg());
5363  const bool GeneratingCoveredLookupTable = (MaxTableSize == TableSize);
5364  BranchInst *RangeCheckBranch = nullptr;
5365 
5366  if (!DefaultIsReachable || GeneratingCoveredLookupTable) {
5367  Builder.CreateBr(LookupBB);
5368  // Note: We call removeProdecessor later since we need to be able to get the
5369  // PHI value for the default case in case we're using a bit mask.
5370  } else {
5371  Value *Cmp = Builder.CreateICmpULT(
5372  TableIndex, ConstantInt::get(MinCaseVal->getType(), TableSize));
5373  RangeCheckBranch =
5374  Builder.CreateCondBr(Cmp, LookupBB, SI->getDefaultDest());
5375  }
5376 
5377  // Populate the BB that does the lookups.
5378  Builder.SetInsertPoint(LookupBB);
5379 
5380  if (NeedMask) {
5381  // Before doing the lookup, we do the hole check. The LookupBB is therefore
5382  // re-purposed to do the hole check, and we create a new LookupBB.
5383  BasicBlock *MaskBB = LookupBB;
5384  MaskBB->setName("switch.hole_check");
5385  LookupBB = BasicBlock::Create(Mod.getContext(), "switch.lookup",
5386  CommonDest->getParent(), CommonDest);
5387 
5388  // Make the mask's bitwidth at least 8-bit and a power-of-2 to avoid
5389  // unnecessary illegal types.
5390  uint64_t TableSizePowOf2 = NextPowerOf2(std::max(7ULL, TableSize - 1ULL));
5391  APInt MaskInt(TableSizePowOf2, 0);
5392  APInt One(TableSizePowOf2, 1);
5393  // Build bitmask; fill in a 1 bit for every case.
5394  const ResultListTy &ResultList = ResultLists[PHIs[0]];
5395  for (size_t I = 0, E = ResultList.size(); I != E; ++I) {
5396  uint64_t Idx = (ResultList[I].first->getValue() - MinCaseVal->getValue())
5397  .getLimitedValue();
5398  MaskInt |= One << Idx;
5399  }
5400  ConstantInt *TableMask = ConstantInt::get(Mod.getContext(), MaskInt);
5401 
5402  // Get the TableIndex'th bit of the bitmask.
5403  // If this bit is 0 (meaning hole) jump to the default destination,
5404  // else continue with table lookup.
5405  IntegerType *MapTy = TableMask->getType();
5406  Value *MaskIndex =
5407  Builder.CreateZExtOrTrunc(TableIndex, MapTy, "switch.maskindex");
5408  Value *Shifted = Builder.CreateLShr(TableMask, MaskIndex, "switch.shifted");
5409  Value *LoBit = Builder.CreateTrunc(
5410  Shifted, Type::getInt1Ty(Mod.getContext()), "switch.lobit");
5411  Builder.CreateCondBr(LoBit, LookupBB, SI->getDefaultDest());
5412 
5413  Builder.SetInsertPoint(LookupBB);
5414  AddPredecessorToBlock(SI->getDefaultDest(), MaskBB, SI->getParent());
5415  }
5416 
5417  if (!DefaultIsReachable || GeneratingCoveredLookupTable) {
5418  // We cached PHINodes in PHIs. To avoid accessing deleted PHINodes later,
5419  // do not delete PHINodes here.
5421  /*KeepOneInputPHIs=*/true);
5422  }
5423 
5424  bool ReturnedEarly = false;
5425  for (PHINode *PHI : PHIs) {
5426  const ResultListTy &ResultList = ResultLists[PHI];
5427 
5428  // If using a bitmask, use any value to fill the lookup table holes.
5429  Constant *DV = NeedMask ? ResultLists[PHI][0].second : DefaultResults[PHI];
5430  StringRef FuncName = Fn->getName();
5431  SwitchLookupTable Table(Mod, TableSize, MinCaseVal, ResultList, DV, DL,
5432  FuncName);
5433 
5434  Value *Result = Table.BuildLookup(TableIndex, Builder);
5435 
5436  // If the result is used to return immediately from the function, we want to
5437  // do that right here.
5438  if (PHI->hasOneUse() && isa<ReturnInst>(*PHI->user_begin()) &&
5439  PHI->user_back() == CommonDest->getFirstNonPHIOrDbg()) {
5440  Builder.CreateRet(Result);
5441  ReturnedEarly = true;
5442  break;
5443  }
5444 
5445  // Do a small peephole optimization: re-use the switch table compare if
5446  // possible.
5447  if (!TableHasHoles && HasDefaultResults && RangeCheckBranch) {
5448  BasicBlock *PhiBlock = PHI->getParent();
5449  // Search for compare instructions which use the phi.
5450  for (auto *User : PHI->users()) {
5451  reuseTableCompare(User, PhiBlock, RangeCheckBranch, DV, ResultList);
5452  }
5453  }
5454 
5455  PHI->addIncoming(Result, LookupBB);
5456  }
5457 
5458  if (!ReturnedEarly)
5459  Builder.CreateBr(CommonDest);
5460 
5461  // Remove the switch.
5462  for (unsigned i = 0, e = SI->getNumSuccessors(); i < e; ++i) {
5463  BasicBlock *Succ = SI->getSuccessor(i);
5464 
5465  if (Succ == SI->getDefaultDest())
5466  continue;
5467  Succ->removePredecessor(SI->getParent());
5468  }
5469  SI->eraseFromParent();
5470 
5471  ++NumLookupTables;
5472  if (NeedMask)
5473  ++NumLookupTablesHoles;
5474  return true;
5475 }
5476 
5477 static bool isSwitchDense(ArrayRef<int64_t> Values) {
5478  // See also SelectionDAGBuilder::isDense(), which this function was based on.
5479  uint64_t Diff = (uint64_t)Values.back() - (uint64_t)Values.front();
5480  uint64_t Range = Diff + 1;
5481  uint64_t NumCases = Values.size();
5482  // 40% is the default density for building a jump table in optsize/minsize mode.
5483  uint64_t MinDensity = 40;
5484 
5485  return NumCases * 100 >= Range * MinDensity;
5486 }
5487 
5488 /// Try to transform a switch that has "holes" in it to a contiguous sequence
5489 /// of cases.
5490 ///
5491 /// A switch such as: switch(i) {case 5: case 9: case 13: case 17:} can be
5492 /// range-reduced to: switch ((i-5) / 4) {case 0: case 1: case 2: case 3:}.
5493 ///
5494 /// This converts a sparse switch into a dense switch which allows better
5495 /// lowering and could also allow transforming into a lookup table.
5497  const DataLayout &DL,
5498  const TargetTransformInfo &TTI) {
5499  auto *CondTy = cast<IntegerType>(SI->getCondition()->getType());
5500  if (CondTy->getIntegerBitWidth() > 64 ||
5501  !DL.fitsInLegalInteger(CondTy->getIntegerBitWidth()))
5502  return false;
5503  // Only bother with this optimization if there are more than 3 switch cases;
5504  // SDAG will only bother creating jump tables for 4 or more cases.
5505  if (SI->getNumCases() < 4)
5506  return false;
5507 
5508  // This transform is agnostic to the signedness of the input or case values. We
5509  // can treat the case values as signed or unsigned. We can optimize more common
5510  // cases such as a sequence crossing zero {-4,0,4,8} if we interpret case values
5511  // as signed.
5512  SmallVector<int64_t,4> Values;
5513  for (auto &C : SI->cases())
5514  Values.push_back(C.getCaseValue()->getValue().getSExtValue());
5515  llvm::sort(Values);
5516 
5517  // If the switch is already dense, there's nothing useful to do here.
5518  if (isSwitchDense(Values))
5519  return false;
5520 
5521  // First, transform the values such that they start at zero and ascend.
5522  int64_t Base = Values[0];
5523  for (auto &V : Values)
5524  V -= (uint64_t)(Base);
5525 
5526  // Now we have signed numbers that have been shifted so that, given enough
5527  // precision, there are no negative values. Since the rest of the transform
5528  // is bitwise only, we switch now to an unsigned representation.
5529 
5530  // This transform can be done speculatively because it is so cheap - it
5531  // results in a single rotate operation being inserted.
5532  // FIXME: It's possible that optimizing a switch on powers of two might also
5533  // be beneficial - flag values are often powers of two and we could use a CLZ
5534  // as the key function.
5535 
5536  // countTrailingZeros(0) returns 64. As Values is guaranteed to have more than
5537  // one element and LLVM disallows duplicate cases, Shift is guaranteed to be
5538  // less than 64.
5539  unsigned Shift = 64;
5540  for (auto &V : Values)
5541  Shift = std::min(Shift, countTrailingZeros((uint64_t)V));
5542  assert(Shift < 64);
5543  if (Shift > 0)
5544  for (auto &V : Values)
5545  V = (int64_t)((uint64_t)V >> Shift);
5546 
5547  if (!isSwitchDense(Values))
5548  // Transform didn't create a dense switch.
5549  return false;
5550 
5551  // The obvious transform is to shift the switch condition right and emit a
5552  // check that the condition actually cleanly divided by GCD, i.e.
5553  // C & (1 << Shift - 1) == 0
5554  // inserting a new CFG edge to handle the case where it didn't divide cleanly.
5555  //
5556  // A cheaper way of doing this is a simple ROTR(C, Shift). This performs the
5557  // shift and puts the shifted-off bits in the uppermost bits. If any of these
5558  // are nonzero then the switch condition will be very large and will hit the
5559  // default case.
5560 
5561  auto *Ty = cast<IntegerType>(SI->getCondition()->getType());
5562  Builder.SetInsertPoint(SI);
5563  auto *ShiftC = ConstantInt::get(Ty, Shift);
5564  auto *Sub = Builder.CreateSub(SI->getCondition(), ConstantInt::get(Ty, Base));
5565  auto *LShr = Builder.CreateLShr(Sub, ShiftC);
5566  auto *Shl = Builder.CreateShl(Sub, Ty->getBitWidth() - Shift);
5567  auto *Rot = Builder.CreateOr(LShr, Shl);
5568  SI->replaceUsesOfWith(SI->getCondition(), Rot);
5569 
5570  for (auto Case : SI->cases()) {
5571  auto *Orig = Case.getCaseValue();
5572  auto Sub = Orig->getValue() - APInt(Ty->getBitWidth(), Base);
5573  Case.setValue(
5574  cast<ConstantInt>(ConstantInt::get(Ty, Sub.lshr(ShiftC->getValue()))));
5575  }
5576  return true;
5577 }
5578 
5579 bool SimplifyCFGOpt::SimplifySwitch(SwitchInst *SI, IRBuilder<> &Builder) {
5580  BasicBlock *BB = SI->getParent();
5581 
5582  if (isValueEqualityComparison(SI)) {
5583  // If we only have one predecessor, and if it is a branch on this value,
5584  // see if that predecessor totally determines the outcome of this switch.
5585  if (BasicBlock *OnlyPred = BB->getSinglePredecessor())
5586  if (SimplifyEqualityComparisonWithOnlyPredecessor(SI, OnlyPred, Builder))
5587  return requestResimplify();
5588 
5589  Value *Cond = SI->getCondition();
5590  if (SelectInst *Select = dyn_cast<SelectInst>(Cond))
5591  if (SimplifySwitchOnSelect(SI, Select))
5592  return requestResimplify();
5593 
5594  // If the block only contains the switch, see if we can fold the block
5595  // away into any preds.</