LLVM  10.0.0svn
SimplifyCFG.cpp
Go to the documentation of this file.
1 //===- SimplifyCFG.cpp - Code to perform CFG simplification ---------------===//
2 //
3 // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
4 // See https://llvm.org/LICENSE.txt for license information.
5 // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
6 //
7 //===----------------------------------------------------------------------===//
8 //
9 // Peephole optimize the CFG.
10 //
11 //===----------------------------------------------------------------------===//
12 
13 #include "llvm/ADT/APInt.h"
14 #include "llvm/ADT/ArrayRef.h"
15 #include "llvm/ADT/DenseMap.h"
16 #include "llvm/ADT/Optional.h"
17 #include "llvm/ADT/STLExtras.h"
18 #include "llvm/ADT/SetOperations.h"
19 #include "llvm/ADT/SetVector.h"
20 #include "llvm/ADT/SmallPtrSet.h"
21 #include "llvm/ADT/SmallVector.h"
22 #include "llvm/ADT/Statistic.h"
23 #include "llvm/ADT/StringRef.h"
32 #include "llvm/IR/Attributes.h"
33 #include "llvm/IR/BasicBlock.h"
34 #include "llvm/IR/CFG.h"
35 #include "llvm/IR/CallSite.h"
36 #include "llvm/IR/Constant.h"
37 #include "llvm/IR/ConstantRange.h"
38 #include "llvm/IR/Constants.h"
39 #include "llvm/IR/DataLayout.h"
40 #include "llvm/IR/DerivedTypes.h"
41 #include "llvm/IR/Function.h"
42 #include "llvm/IR/GlobalValue.h"
43 #include "llvm/IR/GlobalVariable.h"
44 #include "llvm/IR/IRBuilder.h"
45 #include "llvm/IR/InstrTypes.h"
46 #include "llvm/IR/Instruction.h"
47 #include "llvm/IR/Instructions.h"
48 #include "llvm/IR/IntrinsicInst.h"
49 #include "llvm/IR/Intrinsics.h"
50 #include "llvm/IR/LLVMContext.h"
51 #include "llvm/IR/MDBuilder.h"
52 #include "llvm/IR/Metadata.h"
53 #include "llvm/IR/Module.h"
54 #include "llvm/IR/NoFolder.h"
55 #include "llvm/IR/Operator.h"
56 #include "llvm/IR/PatternMatch.h"
57 #include "llvm/IR/Type.h"
58 #include "llvm/IR/Use.h"
59 #include "llvm/IR/User.h"
60 #include "llvm/IR/Value.h"
61 #include "llvm/Support/Casting.h"
63 #include "llvm/Support/Debug.h"
65 #include "llvm/Support/KnownBits.h"
71 #include <algorithm>
72 #include <cassert>
73 #include <climits>
74 #include <cstddef>
75 #include <cstdint>
76 #include <iterator>
77 #include <map>
78 #include <set>
79 #include <tuple>
80 #include <utility>
81 #include <vector>
82 
83 using namespace llvm;
84 using namespace PatternMatch;
85 
86 #define DEBUG_TYPE "simplifycfg"
87 
88 // Chosen as 2 so as to be cheap, but still to have enough power to fold
89 // a select, so the "clamp" idiom (of a min followed by a max) will be caught.
90 // To catch this, we need to fold a compare and a select, hence '2' being the
91 // minimum reasonable default.
93  "phi-node-folding-threshold", cl::Hidden, cl::init(2),
94  cl::desc(
95  "Control the amount of phi node folding to perform (default = 2)"));
96 
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);
792  SI->setMetadata(LLVMContext::MD_prof, N);
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);
806  I->setMetadata(LLVMContext::MD_prof, N);
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) {
955  MDNode *ProfMD = I->getMetadata(LLVMContext::MD_prof);
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) {
968  MDNode *MD = TI->getMetadata(LLVMContext::MD_prof);
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,
1307  LLVMContext::MD_range,
1308  LLVMContext::MD_fpmath,
1309  LLVMContext::MD_invariant_load,
1310  LLVMContext::MD_nonnull,
1311  LLVMContext::MD_invariant_group,
1312  LLVMContext::MD_align,
1313  LLVMContext::MD_dereferenceable,
1314  LLVMContext::MD_dereferenceable_or_null,
1315  LLVMContext::MD_mem_parallel_loop_access,
1316  LLVMContext::MD_access_group,
1317  LLVMContext::MD_preserve_access_index};
1318  combineMetadata(I1, I2, KnownIDs, true);
1319 
1320  // I1 and I2 are being combined into a single instruction. Its debug
1321  // location is the merged locations of the original instructions.
1322  I1->applyMergedLocation(I1->getDebugLoc(), I2->getDebugLoc());
1323 
1324  I2->eraseFromParent();
1325  Changed = true;
1326  }
1327 
1328  I1 = &*BB1_Itr++;
1329  I2 = &*BB2_Itr++;
1330  // Skip debug info if it is not identical.
1333  if (!DBI1 || !DBI2 || !DBI1->isIdenticalToWhenDefined(DBI2)) {
1334  while (isa<DbgInfoIntrinsic>(I1))
1335  I1 = &*BB1_Itr++;
1336  while (isa<DbgInfoIntrinsic>(I2))
1337  I2 = &*BB2_Itr++;
1338  }
1339  } while (I1->isIdenticalToWhenDefined(I2));
1340 
1341  return true;
1342 
1343 HoistTerminator:
1344  // It may not be possible to hoist an invoke.
1345  // FIXME: Can we define a safety predicate for CallBr?
1346  if (isa<InvokeInst>(I1) && !isSafeToHoistInvoke(BB1, BB2, I1, I2))
1347  return Changed;
1348 
1349  // TODO: callbr hoisting currently disabled pending further study.
1350  if (isa<CallBrInst>(I1))
1351  return Changed;
1352 
1353  for (BasicBlock *Succ : successors(BB1)) {
1354  for (PHINode &PN : Succ->phis()) {
1355  Value *BB1V = PN.getIncomingValueForBlock(BB1);
1356  Value *BB2V = PN.getIncomingValueForBlock(BB2);
1357  if (BB1V == BB2V)
1358  continue;
1359 
1360  // Check for passingValueIsAlwaysUndefined here because we would rather
1361  // eliminate undefined control flow then converting it to a select.
1362  if (passingValueIsAlwaysUndefined(BB1V, &PN) ||
1363  passingValueIsAlwaysUndefined(BB2V, &PN))
1364  return Changed;
1365 
1366  if (isa<ConstantExpr>(BB1V) && !isSafeToSpeculativelyExecute(BB1V))
1367  return Changed;
1368  if (isa<ConstantExpr>(BB2V) && !isSafeToSpeculativelyExecute(BB2V))
1369  return Changed;
1370  }
1371  }
1372 
1373  // Okay, it is safe to hoist the terminator.
1374  Instruction *NT = I1->clone();
1375  BIParent->getInstList().insert(BI->getIterator(), NT);
1376  if (!NT->getType()->isVoidTy()) {
1377  I1->replaceAllUsesWith(NT);
1378  I2->replaceAllUsesWith(NT);
1379  NT->takeName(I1);
1380  }
1381 
1382  // Ensure terminator gets a debug location, even an unknown one, in case
1383  // it involves inlinable calls.
1384  NT->applyMergedLocation(I1->getDebugLoc(), I2->getDebugLoc());
1385 
1386  // PHIs created below will adopt NT's merged DebugLoc.
1387  IRBuilder<NoFolder> Builder(NT);
1388 
1389  // Hoisting one of the terminators from our successor is a great thing.
1390  // Unfortunately, the successors of the if/else blocks may have PHI nodes in
1391  // them. If they do, all PHI entries for BB1/BB2 must agree for all PHI
1392  // nodes, so we insert select instruction to compute the final result.
1393  std::map<std::pair<Value *, Value *>, SelectInst *> InsertedSelects;
1394  for (BasicBlock *Succ : successors(BB1)) {
1395  for (PHINode &PN : Succ->phis()) {
1396  Value *BB1V = PN.getIncomingValueForBlock(BB1);
1397  Value *BB2V = PN.getIncomingValueForBlock(BB2);
1398  if (BB1V == BB2V)
1399  continue;
1400 
1401  // These values do not agree. Insert a select instruction before NT
1402  // that determines the right value.
1403  SelectInst *&SI = InsertedSelects[std::make_pair(BB1V, BB2V)];
1404  if (!SI)
1405  SI = cast<SelectInst>(
1406  Builder.CreateSelect(BI->getCondition(), BB1V, BB2V,
1407  BB1V->getName() + "." + BB2V->getName(), BI));
1408 
1409  // Make the PHI node use the select for all incoming values for BB1/BB2
1410  for (unsigned i = 0, e = PN.getNumIncomingValues(); i != e; ++i)
1411  if (PN.getIncomingBlock(i) == BB1 || PN.getIncomingBlock(i) == BB2)
1412  PN.setIncomingValue(i, SI);
1413  }
1414  }
1415 
1416  // Update any PHI nodes in our new successors.
1417  for (BasicBlock *Succ : successors(BB1))
1418  AddPredecessorToBlock(Succ, BIParent, BB1);
1419 
1421  return true;
1422 }
1423 
1424 // All instructions in Insts belong to different blocks that all unconditionally
1425 // branch to a common successor. Analyze each instruction and return true if it
1426 // would be possible to sink them into their successor, creating one common
1427 // instruction instead. For every value that would be required to be provided by
1428 // PHI node (because an operand varies in each input block), add to PHIOperands.
1431  DenseMap<Instruction *, SmallVector<Value *, 4>> &PHIOperands) {
1432  // Prune out obviously bad instructions to move. Each instruction must have
1433  // exactly zero or one use, and we check later that use is by a single, common
1434  // PHI instruction in the successor.
1435  bool HasUse = !Insts.front()->user_empty();
1436  for (auto *I : Insts) {
1437  // These instructions may change or break semantics if moved.
1438  if (isa<PHINode>(I) || I->isEHPad() || isa<AllocaInst>(I) ||
1439  I->getType()->isTokenTy())
1440  return false;
1441 
1442  // Conservatively return false if I is an inline-asm instruction. Sinking
1443  // and merging inline-asm instructions can potentially create arguments
1444  // that cannot satisfy the inline-asm constraints.
1445  if (const auto *C = dyn_cast<CallBase>(I))
1446  if (C->isInlineAsm())
1447  return false;
1448 
1449  // Each instruction must have zero or one use.
1450  if (HasUse && !I->hasOneUse())
1451  return false;
1452  if (!HasUse && !I->user_empty())
1453  return false;
1454  }
1455 
1456  const Instruction *I0 = Insts.front();
1457  for (auto *I : Insts)
1458  if (!I->isSameOperationAs(I0))
1459  return false;
1460 
1461  // All instructions in Insts are known to be the same opcode. If they have a
1462  // use, check that the only user is a PHI or in the same block as the
1463  // instruction, because if a user is in the same block as an instruction we're
1464  // contemplating sinking, it must already be determined to be sinkable.
1465  if (HasUse) {
1466  auto *PNUse = dyn_cast<PHINode>(*I0->user_begin());
1467  auto *Succ = I0->getParent()->getTerminator()->getSuccessor(0);
1468  if (!all_of(Insts, [&PNUse,&Succ](const Instruction *I) -> bool {
1469  auto *U = cast<Instruction>(*I->user_begin());
1470  return (PNUse &&
1471  PNUse->getParent() == Succ &&
1472  PNUse->getIncomingValueForBlock(I->getParent()) == I) ||
1473  U->getParent() == I->getParent();
1474  }))
1475  return false;
1476  }
1477 
1478  // Because SROA can't handle speculating stores of selects, try not
1479  // to sink loads or stores of allocas when we'd have to create a PHI for
1480  // the address operand. Also, because it is likely that loads or stores
1481  // of allocas will disappear when Mem2Reg/SROA is run, don't sink them.
1482  // This can cause code churn which can have unintended consequences down
1483  // the line - see https://llvm.org/bugs/show_bug.cgi?id=30244.
1484  // FIXME: This is a workaround for a deficiency in SROA - see
1485  // https://llvm.org/bugs/show_bug.cgi?id=30188
1486  if (isa<StoreInst>(I0) && any_of(Insts, [](const Instruction *I) {
1487  return isa<AllocaInst>(I->getOperand(1));
1488  }))
1489  return false;
1490  if (isa<LoadInst>(I0) && any_of(Insts, [](const Instruction *I) {
1491  return isa<AllocaInst>(I->getOperand(0));
1492  }))
1493  return false;
1494 
1495  for (unsigned OI = 0, OE = I0->getNumOperands(); OI != OE; ++OI) {
1496  if (I0->getOperand(OI)->getType()->isTokenTy())
1497  // Don't touch any operand of token type.
1498  return false;
1499 
1500  auto SameAsI0 = [&I0, OI](const Instruction *I) {
1501  assert(I->getNumOperands() == I0->getNumOperands());
1502  return I->getOperand(OI) == I0->getOperand(OI);
1503  };
1504  if (!all_of(Insts, SameAsI0)) {
1505  if (!canReplaceOperandWithVariable(I0, OI))
1506  // We can't create a PHI from this GEP.
1507  return false;
1508  // Don't create indirect calls! The called value is the final operand.
1509  if (isa<CallBase>(I0) && OI == OE - 1) {
1510  // FIXME: if the call was *already* indirect, we should do this.
1511  return false;
1512  }
1513  for (auto *I : Insts)
1514  PHIOperands[I].push_back(I->getOperand(OI));
1515  }
1516  }
1517  return true;
1518 }
1519 
1520 // Assuming canSinkLastInstruction(Blocks) has returned true, sink the last
1521 // instruction of every block in Blocks to their common successor, commoning
1522 // into one instruction.
1524  auto *BBEnd = Blocks[0]->getTerminator()->getSuccessor(0);
1525 
1526  // canSinkLastInstruction returning true guarantees that every block has at
1527  // least one non-terminator instruction.
1529  for (auto *BB : Blocks) {
1530  Instruction *I = BB->getTerminator();
1531  do {
1532  I = I->getPrevNode();
1533  } while (isa<DbgInfoIntrinsic>(I) && I != &BB->front());
1534  if (!isa<DbgInfoIntrinsic>(I))
1535  Insts.push_back(I);
1536  }
1537 
1538  // The only checking we need to do now is that all users of all instructions
1539  // are the same PHI node. canSinkLastInstruction should have checked this but
1540  // it is slightly over-aggressive - it gets confused by commutative instructions
1541  // so double-check it here.
1542  Instruction *I0 = Insts.front();
1543  if (!I0->user_empty()) {
1544  auto *PNUse = dyn_cast<PHINode>(*I0->user_begin());
1545  if (!all_of(Insts, [&PNUse](const Instruction *I) -> bool {
1546  auto *U = cast<Instruction>(*I->user_begin());
1547  return U == PNUse;
1548  }))
1549  return false;
1550  }
1551 
1552  // We don't need to do any more checking here; canSinkLastInstruction should
1553  // have done it all for us.
1554  SmallVector<Value*, 4> NewOperands;
1555  for (unsigned O = 0, E = I0->getNumOperands(); O != E; ++O) {
1556  // This check is different to that in canSinkLastInstruction. There, we
1557  // cared about the global view once simplifycfg (and instcombine) have
1558  // completed - it takes into account PHIs that become trivially
1559  // simplifiable. However here we need a more local view; if an operand
1560  // differs we create a PHI and rely on instcombine to clean up the very
1561  // small mess we may make.
1562  bool NeedPHI = any_of(Insts, [&I0, O](const Instruction *I) {
1563  return I->getOperand(O) != I0->getOperand(O);
1564  });
1565  if (!NeedPHI) {
1566  NewOperands.push_back(I0->getOperand(O));
1567  continue;
1568  }
1569 
1570  // Create a new PHI in the successor block and populate it.
1571  auto *Op = I0->getOperand(O);
1572  assert(!Op->getType()->isTokenTy() && "Can't PHI tokens!");
1573  auto *PN = PHINode::Create(Op->getType(), Insts.size(),
1574  Op->getName() + ".sink", &BBEnd->front());
1575  for (auto *I : Insts)
1576  PN->addIncoming(I->getOperand(O), I->getParent());
1577  NewOperands.push_back(PN);
1578  }
1579 
1580  // Arbitrarily use I0 as the new "common" instruction; remap its operands
1581  // and move it to the start of the successor block.
1582  for (unsigned O = 0, E = I0->getNumOperands(); O != E; ++O)
1583  I0->getOperandUse(O).set(NewOperands[O]);
1584  I0->moveBefore(&*BBEnd->getFirstInsertionPt());
1585 
1586  // Update metadata and IR flags, and merge debug locations.
1587  for (auto *I : Insts)
1588  if (I != I0) {
1589  // The debug location for the "common" instruction is the merged locations
1590  // of all the commoned instructions. We start with the original location
1591  // of the "common" instruction and iteratively merge each location in the
1592  // loop below.
1593  // This is an N-way merge, which will be inefficient if I0 is a CallInst.
1594  // However, as N-way merge for CallInst is rare, so we use simplified API
1595  // instead of using complex API for N-way merge.
1596  I0->applyMergedLocation(I0->getDebugLoc(), I->getDebugLoc());
1597  combineMetadataForCSE(I0, I, true);
1598  I0->andIRFlags(I);
1599  }
1600 
1601  if (!I0->user_empty()) {
1602  // canSinkLastInstruction checked that all instructions were used by
1603  // one and only one PHI node. Find that now, RAUW it to our common
1604  // instruction and nuke it.
1605  auto *PN = cast<PHINode>(*I0->user_begin());
1606  PN->replaceAllUsesWith(I0);
1607  PN->eraseFromParent();
1608  }
1609 
1610  // Finally nuke all instructions apart from the common instruction.
1611  for (auto *I : Insts)
1612  if (I != I0)
1613  I->eraseFromParent();
1614 
1615  return true;
1616 }
1617 
1618 namespace {
1619 
1620  // LockstepReverseIterator - Iterates through instructions
1621  // in a set of blocks in reverse order from the first non-terminator.
1622  // For example (assume all blocks have size n):
1623  // LockstepReverseIterator I([B1, B2, B3]);
1624  // *I-- = [B1[n], B2[n], B3[n]];
1625  // *I-- = [B1[n-1], B2[n-1], B3[n-1]];
1626  // *I-- = [B1[n-2], B2[n-2], B3[n-2]];
1627  // ...
1628  class LockstepReverseIterator {
1629  ArrayRef<BasicBlock*> Blocks;
1631  bool Fail;
1632 
1633  public:
1634  LockstepReverseIterator(ArrayRef<BasicBlock*> Blocks) : Blocks(Blocks) {
1635  reset();
1636  }
1637 
1638  void reset() {
1639  Fail = false;
1640  Insts.clear();
1641  for (auto *BB : Blocks) {
1642  Instruction *Inst = BB->getTerminator();
1643  for (Inst = Inst->getPrevNode(); Inst && isa<DbgInfoIntrinsic>(Inst);)
1644  Inst = Inst->getPrevNode();
1645  if (!Inst) {
1646  // Block wasn't big enough.
1647  Fail = true;
1648  return;
1649  }
1650  Insts.push_back(Inst);
1651  }
1652  }
1653 
1654  bool isValid() const {
1655  return !Fail;
1656  }
1657 
1658  void operator--() {
1659  if (Fail)
1660  return;
1661  for (auto *&Inst : Insts) {
1662  for (Inst = Inst->getPrevNode(); Inst && isa<DbgInfoIntrinsic>(Inst);)
1663  Inst = Inst->getPrevNode();
1664  // Already at beginning of block.
1665  if (!Inst) {
1666  Fail = true;
1667  return;
1668  }
1669  }
1670  }
1671 
1673  return Insts;
1674  }
1675  };
1676 
1677 } // end anonymous namespace
1678 
1679 /// Check whether BB's predecessors end with unconditional branches. If it is
1680 /// true, sink any common code from the predecessors to BB.
1681 /// We also allow one predecessor to end with conditional branch (but no more
1682 /// than one).
1684  // We support two situations:
1685  // (1) all incoming arcs are unconditional
1686  // (2) one incoming arc is conditional
1687  //
1688  // (2) is very common in switch defaults and
1689  // else-if patterns;
1690  //
1691  // if (a) f(1);
1692  // else if (b) f(2);
1693  //
1694  // produces:
1695  //
1696  // [if]
1697  // / \
1698  // [f(1)] [if]
1699  // | | \
1700  // | | |
1701  // | [f(2)]|
1702  // \ | /
1703  // [ end ]
1704  //
1705  // [end] has two unconditional predecessor arcs and one conditional. The
1706  // conditional refers to the implicit empty 'else' arc. This conditional
1707  // arc can also be caused by an empty default block in a switch.
1708  //
1709  // In this case, we attempt to sink code from all *unconditional* arcs.
1710  // If we can sink instructions from these arcs (determined during the scan
1711  // phase below) we insert a common successor for all unconditional arcs and
1712  // connect that to [end], to enable sinking:
1713  //
1714  // [if]
1715  // / \
1716  // [x(1)] [if]
1717  // | | \
1718  // | | \
1719  // | [x(2)] |
1720  // \ / |
1721  // [sink.split] |
1722  // \ /
1723  // [ end ]
1724  //
1725  SmallVector<BasicBlock*,4> UnconditionalPreds;
1726  Instruction *Cond = nullptr;
1727  for (auto *B : predecessors(BB)) {
1728  auto *T = B->getTerminator();
1729  if (isa<BranchInst>(T) && cast<BranchInst>(T)->isUnconditional())
1730  UnconditionalPreds.push_back(B);
1731  else if ((isa<BranchInst>(T) || isa<SwitchInst>(T)) && !Cond)
1732  Cond = T;
1733  else
1734  return false;
1735  }
1736  if (UnconditionalPreds.size() < 2)
1737  return false;
1738 
1739  bool Changed = false;
1740  // We take a two-step approach to tail sinking. First we scan from the end of
1741  // each block upwards in lockstep. If the n'th instruction from the end of each
1742  // block can be sunk, those instructions are added to ValuesToSink and we
1743  // carry on. If we can sink an instruction but need to PHI-merge some operands
1744  // (because they're not identical in each instruction) we add these to
1745  // PHIOperands.
1746  unsigned ScanIdx = 0;
1747  SmallPtrSet<Value*,4> InstructionsToSink;
1749  LockstepReverseIterator LRI(UnconditionalPreds);
1750  while (LRI.isValid() &&
1751  canSinkInstructions(*LRI, PHIOperands)) {
1752  LLVM_DEBUG(dbgs() << "SINK: instruction can be sunk: " << *(*LRI)[0]
1753  << "\n");
1754  InstructionsToSink.insert((*LRI).begin(), (*LRI).end());
1755  ++ScanIdx;
1756  --LRI;
1757  }
1758 
1759  auto ProfitableToSinkInstruction = [&](LockstepReverseIterator &LRI) {
1760  unsigned NumPHIdValues = 0;
1761  for (auto *I : *LRI)
1762  for (auto *V : PHIOperands[I])
1763  if (InstructionsToSink.count(V) == 0)
1764  ++NumPHIdValues;
1765  LLVM_DEBUG(dbgs() << "SINK: #phid values: " << NumPHIdValues << "\n");
1766  unsigned NumPHIInsts = NumPHIdValues / UnconditionalPreds.size();
1767  if ((NumPHIdValues % UnconditionalPreds.size()) != 0)
1768  NumPHIInsts++;
1769 
1770  return NumPHIInsts <= 1;
1771  };
1772 
1773  if (ScanIdx > 0 && Cond) {
1774  // Check if we would actually sink anything first! This mutates the CFG and
1775  // adds an extra block. The goal in doing this is to allow instructions that
1776  // couldn't be sunk before to be sunk - obviously, speculatable instructions
1777  // (such as trunc, add) can be sunk and predicated already. So we check that
1778  // we're going to sink at least one non-speculatable instruction.
1779  LRI.reset();
1780  unsigned Idx = 0;
1781  bool Profitable = false;
1782  while (ProfitableToSinkInstruction(LRI) && Idx < ScanIdx) {
1783  if (!isSafeToSpeculativelyExecute((*LRI)[0])) {
1784  Profitable = true;
1785  break;
1786  }
1787  --LRI;
1788  ++Idx;
1789  }
1790  if (!Profitable)
1791  return false;
1792 
1793  LLVM_DEBUG(dbgs() << "SINK: Splitting edge\n");
1794  // We have a conditional edge and we're going to sink some instructions.
1795  // Insert a new block postdominating all blocks we're going to sink from.
1796  if (!SplitBlockPredecessors(BB, UnconditionalPreds, ".sink.split"))
1797  // Edges couldn't be split.
1798  return false;
1799  Changed = true;
1800  }
1801 
1802  // Now that we've analyzed all potential sinking candidates, perform the
1803  // actual sink. We iteratively sink the last non-terminator of the source
1804  // blocks into their common successor unless doing so would require too
1805  // many PHI instructions to be generated (currently only one PHI is allowed
1806  // per sunk instruction).
1807  //
1808  // We can use InstructionsToSink to discount values needing PHI-merging that will
1809  // actually be sunk in a later iteration. This allows us to be more
1810  // aggressive in what we sink. This does allow a false positive where we
1811  // sink presuming a later value will also be sunk, but stop half way through
1812  // and never actually sink it which means we produce more PHIs than intended.
1813  // This is unlikely in practice though.
1814  for (unsigned SinkIdx = 0; SinkIdx != ScanIdx; ++SinkIdx) {
1815  LLVM_DEBUG(dbgs() << "SINK: Sink: "
1816  << *UnconditionalPreds[0]->getTerminator()->getPrevNode()
1817  << "\n");
1818 
1819  // Because we've sunk every instruction in turn, the current instruction to
1820  // sink is always at index 0.
1821  LRI.reset();
1822  if (!ProfitableToSinkInstruction(LRI)) {
1823  // Too many PHIs would be created.
1824  LLVM_DEBUG(
1825  dbgs() << "SINK: stopping here, too many PHIs would be created!\n");
1826  break;
1827  }
1828 
1829  if (!sinkLastInstruction(UnconditionalPreds))
1830  return Changed;
1831  NumSinkCommons++;
1832  Changed = true;
1833  }
1834  return Changed;
1835 }
1836 
1837 /// Determine if we can hoist sink a sole store instruction out of a
1838 /// conditional block.
1839 ///
1840 /// We are looking for code like the following:
1841 /// BrBB:
1842 /// store i32 %add, i32* %arrayidx2
1843 /// ... // No other stores or function calls (we could be calling a memory
1844 /// ... // function).
1845 /// %cmp = icmp ult %x, %y
1846 /// br i1 %cmp, label %EndBB, label %ThenBB
1847 /// ThenBB:
1848 /// store i32 %add5, i32* %arrayidx2
1849 /// br label EndBB
1850 /// EndBB:
1851 /// ...
1852 /// We are going to transform this into:
1853 /// BrBB:
1854 /// store i32 %add, i32* %arrayidx2
1855 /// ... //
1856 /// %cmp = icmp ult %x, %y
1857 /// %add.add5 = select i1 %cmp, i32 %add, %add5
1858 /// store i32 %add.add5, i32* %arrayidx2
1859 /// ...
1860 ///
1861 /// \return The pointer to the value of the previous store if the store can be
1862 /// hoisted into the predecessor block. 0 otherwise.
1864  BasicBlock *StoreBB, BasicBlock *EndBB) {
1865  StoreInst *StoreToHoist = dyn_cast<StoreInst>(I);
1866  if (!StoreToHoist)
1867  return nullptr;
1868 
1869  // Volatile or atomic.
1870  if (!StoreToHoist->isSimple())
1871  return nullptr;
1872 
1873  Value *StorePtr = StoreToHoist->getPointerOperand();
1874 
1875  // Look for a store to the same pointer in BrBB.
1876  unsigned MaxNumInstToLookAt = 9;
1877  for (Instruction &CurI : reverse(BrBB->instructionsWithoutDebug())) {
1878  if (!MaxNumInstToLookAt)
1879  break;
1880  --MaxNumInstToLookAt;
1881 
1882  // Could be calling an instruction that affects memory like free().
1883  if (CurI.mayHaveSideEffects() && !isa<StoreInst>(CurI))
1884  return nullptr;
1885 
1886  if (auto *SI = dyn_cast<StoreInst>(&CurI)) {
1887  // Found the previous store make sure it stores to the same location.
1888  if (SI->getPointerOperand() == StorePtr)
1889  // Found the previous store, return its value operand.
1890  return SI->getValueOperand();
1891  return nullptr; // Unknown store.
1892  }
1893  }
1894 
1895  return nullptr;
1896 }
1897 
1898 /// Speculate a conditional basic block flattening the CFG.
1899 ///
1900 /// Note that this is a very risky transform currently. Speculating
1901 /// instructions like this is most often not desirable. Instead, there is an MI
1902 /// pass which can do it with full awareness of the resource constraints.
1903 /// However, some cases are "obvious" and we should do directly. An example of
1904 /// this is speculating a single, reasonably cheap instruction.
1905 ///
1906 /// There is only one distinct advantage to flattening the CFG at the IR level:
1907 /// it makes very common but simplistic optimizations such as are common in
1908 /// instcombine and the DAG combiner more powerful by removing CFG edges and
1909 /// modeling their effects with easier to reason about SSA value graphs.
1910 ///
1911 ///
1912 /// An illustration of this transform is turning this IR:
1913 /// \code
1914 /// BB:
1915 /// %cmp = icmp ult %x, %y
1916 /// br i1 %cmp, label %EndBB, label %ThenBB
1917 /// ThenBB:
1918 /// %sub = sub %x, %y
1919 /// br label BB2
1920 /// EndBB:
1921 /// %phi = phi [ %sub, %ThenBB ], [ 0, %EndBB ]
1922 /// ...
1923 /// \endcode
1924 ///
1925 /// Into this IR:
1926 /// \code
1927 /// BB:
1928 /// %cmp = icmp ult %x, %y
1929 /// %sub = sub %x, %y
1930 /// %cond = select i1 %cmp, 0, %sub
1931 /// ...
1932 /// \endcode
1933 ///
1934 /// \returns true if the conditional block is removed.
1936  const TargetTransformInfo &TTI) {
1937  // Be conservative for now. FP select instruction can often be expensive.
1938  Value *BrCond = BI->getCondition();
1939  if (isa<FCmpInst>(BrCond))
1940  return false;
1941 
1942  BasicBlock *BB = BI->getParent();
1943  BasicBlock *EndBB = ThenBB->getTerminator()->getSuccessor(0);
1944 
1945  // If ThenBB is actually on the false edge of the conditional branch, remember
1946  // to swap the select operands later.
1947  bool Invert = false;
1948  if (ThenBB != BI->getSuccessor(0)) {
1949  assert(ThenBB == BI->getSuccessor(1) && "No edge from 'if' block?");
1950  Invert = true;
1951  }
1952  assert(EndBB == BI->getSuccessor(!Invert) && "No edge from to end block");
1953 
1954  // Keep a count of how many times instructions are used within ThenBB when
1955  // they are candidates for sinking into ThenBB. Specifically:
1956  // - They are defined in BB, and
1957  // - They have no side effects, and
1958  // - All of their uses are in ThenBB.
1959  SmallDenseMap<Instruction *, unsigned, 4> SinkCandidateUseCounts;
1960 
1961  SmallVector<Instruction *, 4> SpeculatedDbgIntrinsics;
1962 
1963  unsigned SpeculationCost = 0;
1964  Value *SpeculatedStoreValue = nullptr;
1965  StoreInst *SpeculatedStore = nullptr;
1966  for (BasicBlock::iterator BBI = ThenBB->begin(),
1967  BBE = std::prev(ThenBB->end());
1968  BBI != BBE; ++BBI) {
1969  Instruction *I = &*BBI;
1970  // Skip debug info.
1971  if (isa<DbgInfoIntrinsic>(I)) {
1972  SpeculatedDbgIntrinsics.push_back(I);
1973  continue;
1974  }
1975 
1976  // Only speculatively execute a single instruction (not counting the
1977  // terminator) for now.
1978  ++SpeculationCost;
1979  if (SpeculationCost > 1)
1980  return false;
1981 
1982  // Don't hoist the instruction if it's unsafe or expensive.
1983  if (!isSafeToSpeculativelyExecute(I) &&
1984  !(HoistCondStores && (SpeculatedStoreValue = isSafeToSpeculateStore(
1985  I, BB, ThenBB, EndBB))))
1986  return false;
1987  if (!SpeculatedStoreValue &&
1988  ComputeSpeculationCost(I, TTI) >
1990  return false;
1991 
1992  // Store the store speculation candidate.
1993  if (SpeculatedStoreValue)
1994  SpeculatedStore = cast<StoreInst>(I);
1995 
1996  // Do not hoist the instruction if any of its operands are defined but not
1997  // used in BB. The transformation will prevent the operand from
1998  // being sunk into the use block.
1999  for (User::op_iterator i = I->op_begin(), e = I->op_end(); i != e; ++i) {
2000  Instruction *OpI = dyn_cast<Instruction>(*i);
2001  if (!OpI || OpI->getParent() != BB || OpI->mayHaveSideEffects())
2002  continue; // Not a candidate for sinking.
2003 
2004  ++SinkCandidateUseCounts[OpI];
2005  }
2006  }
2007 
2008  // Consider any sink candidates which are only used in ThenBB as costs for
2009  // speculation. Note, while we iterate over a DenseMap here, we are summing
2010  // and so iteration order isn't significant.
2012  I = SinkCandidateUseCounts.begin(),
2013  E = SinkCandidateUseCounts.end();
2014  I != E; ++I)
2015  if (I->first->hasNUses(I->second)) {
2016  ++SpeculationCost;
2017  if (SpeculationCost > 1)
2018  return false;
2019  }
2020 
2021  // Check that the PHI nodes can be converted to selects.
2022  bool HaveRewritablePHIs = false;
2023  for (PHINode &PN : EndBB->phis()) {
2024  Value *OrigV = PN.getIncomingValueForBlock(BB);
2025  Value *ThenV = PN.getIncomingValueForBlock(ThenBB);
2026 
2027  // FIXME: Try to remove some of the duplication with HoistThenElseCodeToIf.
2028  // Skip PHIs which are trivial.
2029  if (ThenV == OrigV)
2030  continue;
2031 
2032  // Don't convert to selects if we could remove undefined behavior instead.
2033  if (passingValueIsAlwaysUndefined(OrigV, &PN) ||
2034  passingValueIsAlwaysUndefined(ThenV, &PN))
2035  return false;
2036 
2037  HaveRewritablePHIs = true;
2038  ConstantExpr *OrigCE = dyn_cast<ConstantExpr>(OrigV);
2039  ConstantExpr *ThenCE = dyn_cast<ConstantExpr>(ThenV);
2040  if (!OrigCE && !ThenCE)
2041  continue; // Known safe and cheap.
2042 
2043  if ((ThenCE && !isSafeToSpeculativelyExecute(ThenCE)) ||
2044  (OrigCE && !isSafeToSpeculativelyExecute(OrigCE)))
2045  return false;
2046  unsigned OrigCost = OrigCE ? ComputeSpeculationCost(OrigCE, TTI) : 0;
2047  unsigned ThenCost = ThenCE ? ComputeSpeculationCost(ThenCE, TTI) : 0;
2048  unsigned MaxCost =
2050  if (OrigCost + ThenCost > MaxCost)
2051  return false;
2052 
2053  // Account for the cost of an unfolded ConstantExpr which could end up
2054  // getting expanded into Instructions.
2055  // FIXME: This doesn't account for how many operations are combined in the
2056  // constant expression.
2057  ++SpeculationCost;
2058  if (SpeculationCost > 1)
2059  return false;
2060  }
2061 
2062  // If there are no PHIs to process, bail early. This helps ensure idempotence
2063  // as well.
2064  if (!HaveRewritablePHIs && !(HoistCondStores && SpeculatedStoreValue))
2065  return false;
2066 
2067  // If we get here, we can hoist the instruction and if-convert.
2068  LLVM_DEBUG(dbgs() << "SPECULATIVELY EXECUTING BB" << *ThenBB << "\n";);
2069 
2070  // Insert a select of the value of the speculated store.
2071  if (SpeculatedStoreValue) {
2072  IRBuilder<NoFolder> Builder(BI);
2073  Value *TrueV = SpeculatedStore->getValueOperand();
2074  Value *FalseV = SpeculatedStoreValue;
2075  if (Invert)
2076  std::swap(TrueV, FalseV);
2077  Value *S = Builder.CreateSelect(
2078  BrCond, TrueV, FalseV, "spec.store.select", BI);
2079  SpeculatedStore->setOperand(0, S);
2080  SpeculatedStore->applyMergedLocation(BI->getDebugLoc(),
2081  SpeculatedStore->getDebugLoc());
2082  }
2083 
2084  // Metadata can be dependent on the condition we are hoisting above.
2085  // Conservatively strip all metadata on the instruction.
2086  for (auto &I : *ThenBB)
2088 
2089  // Hoist the instructions.
2090  BB->getInstList().splice(BI->getIterator(), ThenBB->getInstList(),
2091  ThenBB->begin(), std::prev(ThenBB->end()));
2092 
2093  // Insert selects and rewrite the PHI operands.
2094  IRBuilder<NoFolder> Builder(BI);
2095  for (PHINode &PN : EndBB->phis()) {
2096  unsigned OrigI = PN.getBasicBlockIndex(BB);
2097  unsigned ThenI = PN.getBasicBlockIndex(ThenBB);
2098  Value *OrigV = PN.getIncomingValue(OrigI);
2099  Value *ThenV = PN.getIncomingValue(ThenI);
2100 
2101  // Skip PHIs which are trivial.
2102  if (OrigV == ThenV)
2103  continue;
2104 
2105  // Create a select whose true value is the speculatively executed value and
2106  // false value is the preexisting value. Swap them if the branch
2107  // destinations were inverted.
2108  Value *TrueV = ThenV, *FalseV = OrigV;
2109  if (Invert)
2110  std::swap(TrueV, FalseV);
2111  Value *V = Builder.CreateSelect(
2112  BrCond, TrueV, FalseV, "spec.select", BI);
2113  PN.setIncomingValue(OrigI, V);
2114  PN.setIncomingValue(ThenI, V);
2115  }
2116 
2117  // Remove speculated dbg intrinsics.
2118  // FIXME: Is it possible to do this in a more elegant way? Moving/merging the
2119  // dbg value for the different flows and inserting it after the select.
2120  for (Instruction *I : SpeculatedDbgIntrinsics)
2121  I->eraseFromParent();
2122 
2123  ++NumSpeculations;
2124  return true;
2125 }
2126 
2127 /// Return true if we can thread a branch across this block.
2129  unsigned Size = 0;
2130 
2131  for (Instruction &I : BB->instructionsWithoutDebug()) {
2132  if (Size > 10)
2133  return false; // Don't clone large BB's.
2134  ++Size;
2135 
2136  // We can only support instructions that do not define values that are
2137  // live outside of the current basic block.
2138  for (User *U : I.users()) {
2139  Instruction *UI = cast<Instruction>(U);
2140  if (UI->getParent() != BB || isa<PHINode>(UI))
2141  return false;
2142  }
2143 
2144  // Looks ok, continue checking.
2145  }
2146 
2147  return true;
2148 }
2149 
2150 /// If we have a conditional branch on a PHI node value that is defined in the
2151 /// same block as the branch and if any PHI entries are constants, thread edges
2152 /// corresponding to that entry to be branches to their ultimate destination.
2153 static bool FoldCondBranchOnPHI(BranchInst *BI, const DataLayout &DL,
2154  AssumptionCache *AC) {
2155  BasicBlock *BB = BI->getParent();
2156  PHINode *PN = dyn_cast<PHINode>(BI->getCondition());
2157  // NOTE: we currently cannot transform this case if the PHI node is used
2158  // outside of the block.
2159  if (!PN || PN->getParent() != BB || !PN->hasOneUse())
2160  return false;
2161 
2162  // Degenerate case of a single entry PHI.
2163  if (PN->getNumIncomingValues() == 1) {
2165  return true;
2166  }
2167 
2168  // Now we know that this block has multiple preds and two succs.
2170  return false;
2171 
2172  // Can't fold blocks that contain noduplicate or convergent calls.
2173  if (any_of(*BB, [](const Instruction &I) {
2174  const CallInst *CI = dyn_cast<CallInst>(&I);
2175  return CI && (CI->cannotDuplicate() || CI->isConvergent());
2176  }))
2177  return false;
2178 
2179  // Okay, this is a simple enough basic block. See if any phi values are
2180  // constants.
2181  for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
2183  if (!CB || !CB->getType()->isIntegerTy(1))
2184  continue;
2185 
2186  // Okay, we now know that all edges from PredBB should be revectored to
2187  // branch to RealDest.
2188  BasicBlock *PredBB = PN->getIncomingBlock(i);
2189  BasicBlock *RealDest = BI->getSuccessor(!CB->getZExtValue());
2190 
2191  if (RealDest == BB)
2192  continue; // Skip self loops.
2193  // Skip if the predecessor's terminator is an indirect branch.
2194  if (isa<IndirectBrInst>(PredBB->getTerminator()))
2195  continue;
2196 
2197  // The dest block might have PHI nodes, other predecessors and other
2198  // difficult cases. Instead of being smart about this, just insert a new
2199  // block that jumps to the destination block, effectively splitting
2200  // the edge we are about to create.
2201  BasicBlock *EdgeBB =
2202  BasicBlock::Create(BB->getContext(), RealDest->getName() + ".critedge",
2203  RealDest->getParent(), RealDest);
2204  BranchInst *CritEdgeBranch = BranchInst::Create(RealDest, EdgeBB);
2205  CritEdgeBranch->setDebugLoc(BI->getDebugLoc());
2206 
2207  // Update PHI nodes.
2208  AddPredecessorToBlock(RealDest, EdgeBB, BB);
2209 
2210  // BB may have instructions that are being threaded over. Clone these
2211  // instructions into EdgeBB. We know that there will be no uses of the
2212  // cloned instructions outside of EdgeBB.
2213  BasicBlock::iterator InsertPt = EdgeBB->begin();
2214  DenseMap<Value *, Value *> TranslateMap; // Track translated values.
2215  for (BasicBlock::iterator BBI = BB->begin(); &*BBI != BI; ++BBI) {
2216  if (PHINode *PN = dyn_cast<PHINode>(BBI)) {
2217  TranslateMap[PN] = PN->getIncomingValueForBlock(PredBB);
2218  continue;
2219  }
2220  // Clone the instruction.
2221  Instruction *N = BBI->clone();
2222  if (BBI->hasName())
2223  N->setName(BBI->getName() + ".c");
2224 
2225  // Update operands due to translation.
2226  for (User::op_iterator i = N->op_begin(), e = N->op_end(); i != e; ++i) {
2227  DenseMap<Value *, Value *>::iterator PI = TranslateMap.find(*i);
2228  if (PI != TranslateMap.end())
2229  *i = PI->second;
2230  }
2231 
2232  // Check for trivial simplification.
2233  if (Value *V = SimplifyInstruction(N, {DL, nullptr, nullptr, AC})) {
2234  if (!BBI->use_empty())
2235  TranslateMap[&*BBI] = V;
2236  if (!N->mayHaveSideEffects()) {
2237  N->deleteValue(); // Instruction folded away, don't need actual inst
2238  N = nullptr;
2239  }
2240  } else {
2241  if (!BBI->use_empty())
2242  TranslateMap[&*BBI] = N;
2243  }
2244  // Insert the new instruction into its new home.
2245  if (N)
2246  EdgeBB->getInstList().insert(InsertPt, N);
2247 
2248  // Register the new instruction with the assumption cache if necessary.
2249  if (auto *II = dyn_cast_or_null<IntrinsicInst>(N))
2250  if (II->getIntrinsicID() == Intrinsic::assume)
2251  AC->registerAssumption(II);
2252  }
2253 
2254  // Loop over all of the edges from PredBB to BB, changing them to branch
2255  // to EdgeBB instead.
2256  Instruction *PredBBTI = PredBB->getTerminator();
2257  for (unsigned i = 0, e = PredBBTI->getNumSuccessors(); i != e; ++i)
2258  if (PredBBTI->getSuccessor(i) == BB) {
2259  BB->removePredecessor(PredBB);
2260  PredBBTI->setSuccessor(i, EdgeBB);
2261  }
2262 
2263  // Recurse, simplifying any other constants.
2264  return FoldCondBranchOnPHI(BI, DL, AC) || true;
2265  }
2266 
2267  return false;
2268 }
2269 
2270 /// Given a BB that starts with the specified two-entry PHI node,
2271 /// see if we can eliminate it.
2273  const DataLayout &DL) {
2274  // Ok, this is a two entry PHI node. Check to see if this is a simple "if
2275  // statement", which has a very simple dominance structure. Basically, we
2276  // are trying to find the condition that is being branched on, which
2277  // subsequently causes this merge to happen. We really want control
2278  // dependence information for this check, but simplifycfg can't keep it up
2279  // to date, and this catches most of the cases we care about anyway.
2280  BasicBlock *BB = PN->getParent();
2281  const Function *Fn = BB->getParent();
2282  if (Fn && Fn->hasFnAttribute(Attribute::OptForFuzzing))
2283  return false;
2284 
2285  BasicBlock *IfTrue, *IfFalse;
2286  Value *IfCond = GetIfCondition(BB, IfTrue, IfFalse);
2287  if (!IfCond ||
2288  // Don't bother if the branch will be constant folded trivially.
2289  isa<ConstantInt>(IfCond))
2290  return false;
2291 
2292  // Okay, we found that we can merge this two-entry phi node into a select.
2293  // Doing so would require us to fold *all* two entry phi nodes in this block.
2294  // At some point this becomes non-profitable (particularly if the target
2295  // doesn't support cmov's). Only do this transformation if there are two or
2296  // fewer PHI nodes in this block.
2297  unsigned NumPhis = 0;
2298  for (BasicBlock::iterator I = BB->begin(); isa<PHINode>(I); ++NumPhis, ++I)
2299  if (NumPhis > 2)
2300  return false;
2301 
2302  // Loop over the PHI's seeing if we can promote them all to select
2303  // instructions. While we are at it, keep track of the instructions
2304  // that need to be moved to the dominating block.
2305  SmallPtrSet<Instruction *, 4> AggressiveInsts;
2306  unsigned MaxCostVal0 = PHINodeFoldingThreshold,
2307  MaxCostVal1 = PHINodeFoldingThreshold;
2308  MaxCostVal0 *= TargetTransformInfo::TCC_Basic;
2309  MaxCostVal1 *= TargetTransformInfo::TCC_Basic;
2310 
2311  for (BasicBlock::iterator II = BB->begin(); isa<PHINode>(II);) {
2312  PHINode *PN = cast<PHINode>(II++);
2313  if (Value *V = SimplifyInstruction(PN, {DL, PN})) {
2314  PN->replaceAllUsesWith(V);
2315  PN->eraseFromParent();
2316  continue;
2317  }
2318 
2319  if (!DominatesMergePoint(PN->getIncomingValue(0), BB, AggressiveInsts,
2320  MaxCostVal0, TTI) ||
2321  !DominatesMergePoint(PN->getIncomingValue(1), BB, AggressiveInsts,
2322  MaxCostVal1, TTI))
2323  return false;
2324  }
2325 
2326  // If we folded the first phi, PN dangles at this point. Refresh it. If
2327  // we ran out of PHIs then we simplified them all.
2328  PN = dyn_cast<PHINode>(BB->begin());
2329  if (!PN)
2330  return true;
2331 
2332  // Don't fold i1 branches on PHIs which contain binary operators. These can
2333  // often be turned into switches and other things.
2334  if (PN->getType()->isIntegerTy(1) &&
2335  (isa<BinaryOperator>(PN->getIncomingValue(0)) ||
2336  isa<BinaryOperator>(PN->getIncomingValue(1)) ||
2337  isa<BinaryOperator>(IfCond)))
2338  return false;
2339 
2340  // If all PHI nodes are promotable, check to make sure that all instructions
2341  // in the predecessor blocks can be promoted as well. If not, we won't be able
2342  // to get rid of the control flow, so it's not worth promoting to select
2343  // instructions.
2344  BasicBlock *DomBlock = nullptr;
2345  BasicBlock *IfBlock1 = PN->getIncomingBlock(0);
2346  BasicBlock *IfBlock2 = PN->getIncomingBlock(1);
2347  if (cast<BranchInst>(IfBlock1->getTerminator())->isConditional()) {
2348  IfBlock1 = nullptr;
2349  } else {
2350  DomBlock = *pred_begin(IfBlock1);
2351  for (BasicBlock::iterator I = IfBlock1->begin(); !I->isTerminator(); ++I)
2352  if (!AggressiveInsts.count(&*I) && !isa<DbgInfoIntrinsic>(I)) {
2353  // This is not an aggressive instruction that we can promote.
2354  // Because of this, we won't be able to get rid of the control flow, so
2355  // the xform is not worth it.
2356  return false;
2357  }
2358  }
2359 
2360  if (cast<BranchInst>(IfBlock2->getTerminator())->isConditional()) {
2361  IfBlock2 = nullptr;
2362  } else {
2363  DomBlock = *pred_begin(IfBlock2);
2364  for (BasicBlock::iterator I = IfBlock2->begin(); !I->isTerminator(); ++I)
2365  if (!AggressiveInsts.count(&*I) && !isa<DbgInfoIntrinsic>(I)) {
2366  // This is not an aggressive instruction that we can promote.
2367  // Because of this, we won't be able to get rid of the control flow, so
2368  // the xform is not worth it.
2369  return false;
2370  }
2371  }
2372 
2373  LLVM_DEBUG(dbgs() << "FOUND IF CONDITION! " << *IfCond
2374  << " T: " << IfTrue->getName()
2375  << " F: " << IfFalse->getName() << "\n");
2376 
2377  // If we can still promote the PHI nodes after this gauntlet of tests,
2378  // do all of the PHI's now.
2379  Instruction *InsertPt = DomBlock->getTerminator();
2380  IRBuilder<NoFolder> Builder(InsertPt);
2381 
2382  // Move all 'aggressive' instructions, which are defined in the
2383  // conditional parts of the if's up to the dominating block.
2384  if (IfBlock1)
2385  hoistAllInstructionsInto(DomBlock, InsertPt, IfBlock1);
2386  if (IfBlock2)
2387  hoistAllInstructionsInto(DomBlock, InsertPt, IfBlock2);
2388 
2389  while (PHINode *PN = dyn_cast<PHINode>(BB->begin())) {
2390  // Change the PHI node into a select instruction.
2391  Value *TrueVal = PN->getIncomingValue(PN->getIncomingBlock(0) == IfFalse);
2392  Value *FalseVal = PN->getIncomingValue(PN->getIncomingBlock(0) == IfTrue);
2393 
2394  Value *Sel = Builder.CreateSelect(IfCond, TrueVal, FalseVal, "", InsertPt);
2395  PN->replaceAllUsesWith(Sel);
2396  Sel->takeName(PN);
2397  PN->eraseFromParent();
2398  }
2399 
2400  // At this point, IfBlock1 and IfBlock2 are both empty, so our if statement
2401  // has been flattened. Change DomBlock to jump directly to our new block to
2402  // avoid other simplifycfg's kicking in on the diamond.
2403  Instruction *OldTI = DomBlock->getTerminator();
2404  Builder.SetInsertPoint(OldTI);
2405  Builder.CreateBr(BB);
2406  OldTI->eraseFromParent();
2407  return true;
2408 }
2409 
2410 /// If we found a conditional branch that goes to two returning blocks,
2411 /// try to merge them together into one return,
2412 /// introducing a select if the return values disagree.
2414  IRBuilder<> &Builder) {
2415  assert(BI->isConditional() && "Must be a conditional branch");
2416  BasicBlock *TrueSucc = BI->getSuccessor(0);
2417  BasicBlock *FalseSucc = BI->getSuccessor(1);
2418  ReturnInst *TrueRet = cast<ReturnInst>(TrueSucc->getTerminator());
2419  ReturnInst *FalseRet = cast<ReturnInst>(FalseSucc->getTerminator());
2420 
2421  // Check to ensure both blocks are empty (just a return) or optionally empty
2422  // with PHI nodes. If there are other instructions, merging would cause extra
2423  // computation on one path or the other.
2424  if (!TrueSucc->getFirstNonPHIOrDbg()->isTerminator())
2425  return false;
2426  if (!FalseSucc->getFirstNonPHIOrDbg()->isTerminator())
2427  return false;
2428 
2429  Builder.SetInsertPoint(BI);
2430  // Okay, we found a branch that is going to two return nodes. If
2431  // there is no return value for this function, just change the
2432  // branch into a return.
2433  if (FalseRet->getNumOperands() == 0) {
2434  TrueSucc->removePredecessor(BI->getParent());
2435  FalseSucc->removePredecessor(BI->getParent());
2436  Builder.CreateRetVoid();
2438  return true;
2439  }
2440 
2441  // Otherwise, figure out what the true and false return values are
2442  // so we can insert a new select instruction.
2443  Value *TrueValue = TrueRet->getReturnValue();
2444  Value *FalseValue = FalseRet->getReturnValue();
2445 
2446  // Unwrap any PHI nodes in the return blocks.
2447  if (PHINode *TVPN = dyn_cast_or_null<PHINode>(TrueValue))
2448  if (TVPN->getParent() == TrueSucc)
2449  TrueValue = TVPN->getIncomingValueForBlock(BI->getParent());
2450  if (PHINode *FVPN = dyn_cast_or_null<PHINode>(FalseValue))
2451  if (FVPN->getParent() == FalseSucc)
2452  FalseValue = FVPN->getIncomingValueForBlock(BI->getParent());
2453 
2454  // In order for this transformation to be safe, we must be able to
2455  // unconditionally execute both operands to the return. This is
2456  // normally the case, but we could have a potentially-trapping
2457  // constant expression that prevents this transformation from being
2458  // safe.
2459  if (ConstantExpr *TCV = dyn_cast_or_null<ConstantExpr>(TrueValue))
2460  if (TCV->canTrap())
2461  return false;
2462  if (ConstantExpr *FCV = dyn_cast_or_null<ConstantExpr>(FalseValue))
2463  if (FCV->canTrap())
2464  return false;
2465 
2466  // Okay, we collected all the mapped values and checked them for sanity, and
2467  // defined to really do this transformation. First, update the CFG.
2468  TrueSucc->removePredecessor(BI->getParent());
2469  FalseSucc->removePredecessor(BI->getParent());
2470 
2471  // Insert select instructions where needed.
2472  Value *BrCond = BI->getCondition();
2473  if (TrueValue) {
2474  // Insert a select if the results differ.
2475  if (TrueValue == FalseValue || isa<UndefValue>(FalseValue)) {
2476  } else if (isa<UndefValue>(TrueValue)) {
2477  TrueValue = FalseValue;
2478  } else {
2479  TrueValue =
2480  Builder.CreateSelect(BrCond, TrueValue, FalseValue, "retval", BI);
2481  }
2482  }
2483 
2484  Value *RI =
2485  !TrueValue ? Builder.CreateRetVoid() : Builder.CreateRet(TrueValue);
2486 
2487  (void)RI;
2488 
2489  LLVM_DEBUG(dbgs() << "\nCHANGING BRANCH TO TWO RETURNS INTO SELECT:"
2490  << "\n " << *BI << "NewRet = " << *RI << "TRUEBLOCK: "
2491  << *TrueSucc << "FALSEBLOCK: " << *FalseSucc);
2492 
2494 
2495  return true;
2496 }
2497 
2498 /// Return true if the given instruction is available
2499 /// in its predecessor block. If yes, the instruction will be removed.
2501  if (!isa<BinaryOperator>(Inst) && !isa<CmpInst>(Inst))
2502  return false;
2503  for (Instruction &I : *PB) {
2504  Instruction *PBI = &I;
2505  // Check whether Inst and PBI generate the same value.
2506  if (Inst->isIdenticalTo(PBI)) {
2507  Inst->replaceAllUsesWith(PBI);
2508  Inst->eraseFromParent();
2509  return true;
2510  }
2511  }
2512  return false;
2513 }
2514 
2515 /// Return true if either PBI or BI has branch weight available, and store
2516 /// the weights in {Pred|Succ}{True|False}Weight. If one of PBI and BI does
2517 /// not have branch weight, use 1:1 as its weight.
2519  uint64_t &PredTrueWeight,
2520  uint64_t &PredFalseWeight,
2521  uint64_t &SuccTrueWeight,
2522  uint64_t &SuccFalseWeight) {
2523  bool PredHasWeights =
2524  PBI->extractProfMetadata(PredTrueWeight, PredFalseWeight);
2525  bool SuccHasWeights =
2526  BI->extractProfMetadata(SuccTrueWeight, SuccFalseWeight);
2527  if (PredHasWeights || SuccHasWeights) {
2528  if (!PredHasWeights)
2529  PredTrueWeight = PredFalseWeight = 1;
2530  if (!SuccHasWeights)
2531  SuccTrueWeight = SuccFalseWeight = 1;
2532  return true;
2533  } else {
2534  return false;
2535  }
2536 }
2537 
2538 /// If this basic block is simple enough, and if a predecessor branches to us
2539 /// and one of our successors, fold the block into the predecessor and use
2540 /// logical operations to pick the right destination.
2542  unsigned BonusInstThreshold) {
2543  BasicBlock *BB = BI->getParent();
2544 
2545  const unsigned PredCount = pred_size(BB);
2546 
2547  Instruction *Cond = nullptr;
2548  if (BI->isConditional())
2549  Cond = dyn_cast<Instruction>(BI->getCondition());
2550  else {
2551  // For unconditional branch, check for a simple CFG pattern, where
2552  // BB has a single predecessor and BB's successor is also its predecessor's
2553  // successor. If such pattern exists, check for CSE between BB and its
2554  // predecessor.
2555  if (BasicBlock *PB = BB->getSinglePredecessor())
2556  if (BranchInst *PBI = dyn_cast<BranchInst>(PB->getTerminator()))
2557  if (PBI->isConditional() &&
2558  (BI->getSuccessor(0) == PBI->getSuccessor(0) ||
2559  BI->getSuccessor(0) == PBI->getSuccessor(1))) {
2560  for (auto I = BB->instructionsWithoutDebug().begin(),
2561  E = BB->instructionsWithoutDebug().end();
2562  I != E;) {
2563  Instruction *Curr = &*I++;
2564  if (isa<CmpInst>(Curr)) {
2565  Cond = Curr;
2566  break;
2567  }
2568  // Quit if we can't remove this instruction.
2569  if (!tryCSEWithPredecessor(Curr, PB))
2570  return false;
2571  }
2572  }
2573 
2574  if (!Cond)
2575  return false;
2576  }
2577 
2578  if (!Cond || (!isa<CmpInst>(Cond) && !isa<BinaryOperator>(Cond)) ||
2579  Cond->getParent() != BB || !Cond->hasOneUse())
2580  return false;
2581 
2582  // Make sure the instruction after the condition is the cond branch.
2583  BasicBlock::iterator CondIt = ++Cond->getIterator();
2584 
2585  // Ignore dbg intrinsics.
2586  while (isa<DbgInfoIntrinsic>(CondIt))
2587  ++CondIt;
2588 
2589  if (&*CondIt != BI)
2590  return false;
2591 
2592  // Only allow this transformation if computing the condition doesn't involve
2593  // too many instructions and these involved instructions can be executed
2594  // unconditionally. We denote all involved instructions except the condition
2595  // as "bonus instructions", and only allow this transformation when the
2596  // number of the bonus instructions we'll need to create when cloning into
2597  // each predecessor does not exceed a certain threshold.
2598  unsigned NumBonusInsts = 0;
2599  for (auto I = BB->begin(); Cond != &*I; ++I) {
2600  // Ignore dbg intrinsics.
2601  if (isa<DbgInfoIntrinsic>(I))
2602  continue;
2603  if (!I->hasOneUse() || !isSafeToSpeculativelyExecute(&*I))
2604  return false;
2605  // I has only one use and can be executed unconditionally.
2607  if (User == nullptr || User->getParent() != BB)
2608  return false;
2609  // I is used in the same BB. Since BI uses Cond and doesn't have more slots
2610  // to use any other instruction, User must be an instruction between next(I)
2611  // and Cond.
2612 
2613  // Account for the cost of duplicating this instruction into each
2614  // predecessor.
2615  NumBonusInsts += PredCount;
2616  // Early exits once we reach the limit.
2617  if (NumBonusInsts > BonusInstThreshold)
2618  return false;
2619  }
2620 
2621  // Cond is known to be a compare or binary operator. Check to make sure that
2622  // neither operand is a potentially-trapping constant expression.
2623  if (ConstantExpr *CE = dyn_cast<ConstantExpr>(Cond->getOperand(0)))
2624  if (CE->canTrap())
2625  return false;
2626  if (ConstantExpr *CE = dyn_cast<ConstantExpr>(Cond->getOperand(1)))
2627  if (CE->canTrap())
2628  return false;
2629 
2630  // Finally, don't infinitely unroll conditional loops.
2631  BasicBlock *TrueDest = BI->getSuccessor(0);
2632  BasicBlock *FalseDest = (BI->isConditional()) ? BI->getSuccessor(1) : nullptr;
2633  if (TrueDest == BB || FalseDest == BB)
2634  return false;
2635 
2636  for (pred_iterator PI = pred_begin(BB), E = pred_end(BB); PI != E; ++PI) {
2637  BasicBlock *PredBlock = *PI;
2638  BranchInst *PBI = dyn_cast<BranchInst>(PredBlock->getTerminator());
2639 
2640  // Check that we have two conditional branches. If there is a PHI node in
2641  // the common successor, verify that the same value flows in from both
2642  // blocks.
2644  if (!PBI || PBI->isUnconditional() ||
2645  (BI->isConditional() && !SafeToMergeTerminators(BI, PBI)) ||
2646  (!BI->isConditional() &&
2647  !isProfitableToFoldUnconditional(BI, PBI, Cond, PHIs)))
2648  continue;
2649 
2650  // Determine if the two branches share a common destination.
2651  Instruction::BinaryOps Opc = Instruction::BinaryOpsEnd;
2652  bool InvertPredCond = false;
2653 
2654  if (BI->isConditional()) {
2655  if (PBI->getSuccessor(0) == TrueDest) {
2656  Opc = Instruction::Or;
2657  } else if (PBI->getSuccessor(1) == FalseDest) {
2658  Opc = Instruction::And;
2659  } else if (PBI->getSuccessor(0) == FalseDest) {
2660  Opc = Instruction::And;
2661  InvertPredCond = true;
2662  } else if (PBI->getSuccessor(1) == TrueDest) {
2663  Opc = Instruction::Or;
2664  InvertPredCond = true;
2665  } else {
2666  continue;
2667  }
2668  } else {
2669  if (PBI->getSuccessor(0) != TrueDest && PBI->getSuccessor(1) != TrueDest)
2670  continue;
2671  }
2672 
2673  LLVM_DEBUG(dbgs() << "FOLDING BRANCH TO COMMON DEST:\n" << *PBI << *BB);
2674  IRBuilder<> Builder(PBI);
2675 
2676  // If we need to invert the condition in the pred block to match, do so now.
2677  if (InvertPredCond) {
2678  Value *NewCond = PBI->getCondition();
2679 
2680  if (NewCond->hasOneUse() && isa<CmpInst>(NewCond)) {
2681  CmpInst *CI = cast<CmpInst>(NewCond);
2682  CI->setPredicate(CI->getInversePredicate());
2683  } else {
2684  NewCond =
2685  Builder.CreateNot(NewCond, PBI->getCondition()->getName() + ".not");
2686  }
2687 
2688  PBI->setCondition(NewCond);
2689  PBI->swapSuccessors();
2690  }
2691 
2692  // If we have bonus instructions, clone them into the predecessor block.
2693  // Note that there may be multiple predecessor blocks, so we cannot move
2694  // bonus instructions to a predecessor block.
2695  ValueToValueMapTy VMap; // maps original values to cloned values
2696  // We already make sure Cond is the last instruction before BI. Therefore,
2697  // all instructions before Cond other than DbgInfoIntrinsic are bonus
2698  // instructions.
2699  for (auto BonusInst = BB->begin(); Cond != &*BonusInst; ++BonusInst) {
2700  if (isa<DbgInfoIntrinsic>(BonusInst))
2701  continue;
2702  Instruction *NewBonusInst = BonusInst->clone();
2703  RemapInstruction(NewBonusInst, VMap,
2705  VMap[&*BonusInst] = NewBonusInst;
2706 
2707  // If we moved a load, we cannot any longer claim any knowledge about
2708  // its potential value. The previous information might have been valid
2709  // only given the branch precondition.
2710  // For an analogous reason, we must also drop all the metadata whose
2711  // semantics we don't understand.
2712  NewBonusInst->dropUnknownNonDebugMetadata();
2713 
2714  PredBlock->getInstList().insert(PBI->getIterator(), NewBonusInst);
2715  NewBonusInst->takeName(&*BonusInst);
2716  BonusInst->setName(BonusInst->getName() + ".old");
2717  }
2718 
2719  // Clone Cond into the predecessor basic block, and or/and the
2720  // two conditions together.
2721  Instruction *CondInPred = Cond->clone();
2722  RemapInstruction(CondInPred, VMap,
2724  PredBlock->getInstList().insert(PBI->getIterator(), CondInPred);
2725  CondInPred->takeName(Cond);
2726  Cond->setName(CondInPred->getName() + ".old");
2727 
2728  if (BI->isConditional()) {
2729  Instruction *NewCond = cast<Instruction>(
2730  Builder.CreateBinOp(Opc, PBI->getCondition(), CondInPred, "or.cond"));
2731  PBI->setCondition(NewCond);
2732 
2733  uint64_t PredTrueWeight, PredFalseWeight, SuccTrueWeight, SuccFalseWeight;
2734  bool HasWeights =
2735  extractPredSuccWeights(PBI, BI, PredTrueWeight, PredFalseWeight,
2736  SuccTrueWeight, SuccFalseWeight);
2737  SmallVector<uint64_t, 8> NewWeights;
2738 
2739  if (PBI->getSuccessor(0) == BB) {
2740  if (HasWeights) {
2741  // PBI: br i1 %x, BB, FalseDest
2742  // BI: br i1 %y, TrueDest, FalseDest
2743  // TrueWeight is TrueWeight for PBI * TrueWeight for BI.
2744  NewWeights.push_back(PredTrueWeight * SuccTrueWeight);
2745  // FalseWeight is FalseWeight for PBI * TotalWeight for BI +
2746  // TrueWeight for PBI * FalseWeight for BI.
2747  // We assume that total weights of a BranchInst can fit into 32 bits.
2748  // Therefore, we will not have overflow using 64-bit arithmetic.
2749  NewWeights.push_back(PredFalseWeight *
2750  (SuccFalseWeight + SuccTrueWeight) +
2751  PredTrueWeight * SuccFalseWeight);
2752  }
2753  AddPredecessorToBlock(TrueDest, PredBlock, BB, MSSAU);
2754  PBI->setSuccessor(0, TrueDest);
2755  }
2756  if (PBI->getSuccessor(1) == BB) {
2757  if (HasWeights) {
2758  // PBI: br i1 %x, TrueDest, BB
2759  // BI: br i1 %y, TrueDest, FalseDest
2760  // TrueWeight is TrueWeight for PBI * TotalWeight for BI +
2761  // FalseWeight for PBI * TrueWeight for BI.
2762  NewWeights.push_back(PredTrueWeight *
2763  (SuccFalseWeight + SuccTrueWeight) +
2764  PredFalseWeight * SuccTrueWeight);
2765  // FalseWeight is FalseWeight for PBI * FalseWeight for BI.
2766  NewWeights.push_back(PredFalseWeight * SuccFalseWeight);
2767  }
2768  AddPredecessorToBlock(FalseDest, PredBlock, BB, MSSAU);
2769  PBI->setSuccessor(1, FalseDest);
2770  }
2771  if (NewWeights.size() == 2) {
2772  // Halve the weights if any of them cannot fit in an uint32_t
2773  FitWeights(NewWeights);
2774 
2775  SmallVector<uint32_t, 8> MDWeights(NewWeights.begin(),
2776  NewWeights.end());
2777  setBranchWeights(PBI, MDWeights[0], MDWeights[1]);
2778  } else
2779  PBI->setMetadata(LLVMContext::MD_prof, nullptr);
2780  } else {
2781  // Update PHI nodes in the common successors.
2782  for (unsigned i = 0, e = PHIs.size(); i != e; ++i) {
2783  ConstantInt *PBI_C = cast<ConstantInt>(
2784  PHIs[i]->getIncomingValueForBlock(PBI->getParent()));
2785  assert(PBI_C->getType()->isIntegerTy(1));
2786  Instruction *MergedCond = nullptr;
2787  if (PBI->getSuccessor(0) == TrueDest) {
2788  // Create (PBI_Cond and PBI_C) or (!PBI_Cond and BI_Value)
2789  // PBI_C is true: PBI_Cond or (!PBI_Cond and BI_Value)
2790  // is false: !PBI_Cond and BI_Value
2791  Instruction *NotCond = cast<Instruction>(
2792  Builder.CreateNot(PBI->getCondition(), "not.cond"));
2793  MergedCond = cast<Instruction>(
2794  Builder.CreateBinOp(Instruction::And, NotCond, CondInPred,
2795  "and.cond"));
2796  if (PBI_C->isOne())
2797  MergedCond = cast<Instruction>(Builder.CreateBinOp(
2798  Instruction::Or, PBI->getCondition(), MergedCond, "or.cond"));
2799  } else {
2800  // Create (PBI_Cond and BI_Value) or (!PBI_Cond and PBI_C)
2801  // PBI_C is true: (PBI_Cond and BI_Value) or (!PBI_Cond)
2802  // is false: PBI_Cond and BI_Value
2803  MergedCond = cast<Instruction>(Builder.CreateBinOp(
2804  Instruction::And, PBI->getCondition(), CondInPred, "and.cond"));
2805  if (PBI_C->isOne()) {
2806  Instruction *NotCond = cast<Instruction>(
2807  Builder.CreateNot(PBI->getCondition(), "not.cond"));
2808  MergedCond = cast<Instruction>(Builder.CreateBinOp(
2809  Instruction::Or, NotCond, MergedCond, "or.cond"));
2810  }
2811  }
2812  // Update PHI Node.
2813  PHIs[i]->setIncomingValueForBlock(PBI->getParent(), MergedCond);
2814  }
2815 
2816  // PBI is changed to branch to TrueDest below. Remove itself from
2817  // potential phis from all other successors.
2818  if (MSSAU)
2819  MSSAU->changeCondBranchToUnconditionalTo(PBI, TrueDest);
2820 
2821  // Change PBI from Conditional to Unconditional.
2822  BranchInst *New_PBI = BranchInst::Create(TrueDest, PBI);
2823  EraseTerminatorAndDCECond(PBI, MSSAU);
2824  PBI = New_PBI;
2825  }
2826 
2827  // If BI was a loop latch, it may have had associated loop metadata.
2828  // We need to copy it to the new latch, that is, PBI.
2829  if (MDNode *LoopMD = BI->getMetadata(LLVMContext::MD_loop))
2830  PBI->setMetadata(LLVMContext::MD_loop, LoopMD);
2831 
2832  // TODO: If BB is reachable from all paths through PredBlock, then we
2833  // could replace PBI's branch probabilities with BI's.
2834 
2835  // Copy any debug value intrinsics into the end of PredBlock.
2836  for (Instruction &I : *BB)
2837  if (isa<DbgInfoIntrinsic>(I))
2838  I.clone()->insertBefore(PBI);
2839 
2840  return true;
2841  }
2842  return false;
2843 }
2844 
2845 // If there is only one store in BB1 and BB2, return it, otherwise return
2846 // nullptr.
2848  StoreInst *S = nullptr;
2849  for (auto *BB : {BB1, BB2}) {
2850  if (!BB)
2851  continue;
2852  for (auto &I : *BB)
2853  if (auto *SI = dyn_cast<StoreInst>(&I)) {
2854  if (S)
2855  // Multiple stores seen.
2856  return nullptr;
2857  else
2858  S = SI;
2859  }
2860  }
2861  return S;
2862 }
2863 
2865  Value *AlternativeV = nullptr) {
2866  // PHI is going to be a PHI node that allows the value V that is defined in
2867  // BB to be referenced in BB's only successor.
2868  //
2869  // If AlternativeV is nullptr, the only value we care about in PHI is V. It
2870  // doesn't matter to us what the other operand is (it'll never get used). We
2871  // could just create a new PHI with an undef incoming value, but that could
2872  // increase register pressure if EarlyCSE/InstCombine can't fold it with some
2873  // other PHI. So here we directly look for some PHI in BB's successor with V
2874  // as an incoming operand. If we find one, we use it, else we create a new
2875  // one.
2876  //
2877  // If AlternativeV is not nullptr, we care about both incoming values in PHI.
2878  // PHI must be exactly: phi <ty> [ %BB, %V ], [ %OtherBB, %AlternativeV]
2879  // where OtherBB is the single other predecessor of BB's only successor.
2880  PHINode *PHI = nullptr;
2881  BasicBlock *Succ = BB->getSingleSuccessor();
2882 
2883  for (auto I = Succ->begin(); isa<PHINode>(I); ++I)
2884  if (cast<PHINode>(I)->getIncomingValueForBlock(BB) == V) {
2885  PHI = cast<PHINode>(I);
2886  if (!AlternativeV)
2887  break;
2888 
2889  assert(Succ->hasNPredecessors(2));
2890  auto PredI = pred_begin(Succ);
2891  BasicBlock *OtherPredBB = *PredI == BB ? *++PredI : *PredI;
2892  if (PHI->getIncomingValueForBlock(OtherPredBB) == AlternativeV)
2893  break;
2894  PHI = nullptr;
2895  }
2896  if (PHI)
2897  return PHI;
2898 
2899  // If V is not an instruction defined in BB, just return it.
2900  if (!AlternativeV &&
2901  (!isa<Instruction>(V) || cast<Instruction>(V)->getParent() != BB))
2902  return V;
2903 
2904  PHI = PHINode::Create(V->getType(), 2, "simplifycfg.merge", &Succ->front());
2905  PHI->addIncoming(V, BB);
2906  for (BasicBlock *PredBB : predecessors(Succ))
2907  if (PredBB != BB)
2908  PHI->addIncoming(
2909  AlternativeV ? AlternativeV : UndefValue::get(V->getType()), PredBB);
2910  return PHI;
2911 }
2912 
2914  BasicBlock *QTB, BasicBlock *QFB,
2915  BasicBlock *PostBB, Value *Address,
2916  bool InvertPCond, bool InvertQCond,
2917  const DataLayout &DL) {
2918  auto IsaBitcastOfPointerType = [](const Instruction &I) {
2919  return Operator::getOpcode(&I) == Instruction::BitCast &&
2920  I.getType()->isPointerTy();
2921  };
2922 
2923  // If we're not in aggressive mode, we only optimize if we have some
2924  // confidence that by optimizing we'll allow P and/or Q to be if-converted.
2925  auto IsWorthwhile = [&](BasicBlock *BB) {
2926  if (!BB)
2927  return true;
2928  // Heuristic: if the block can be if-converted/phi-folded and the
2929  // instructions inside are all cheap (arithmetic/GEPs), it's worthwhile to
2930  // thread this store.
2931  unsigned N = 0;
2932  for (auto &I : BB->instructionsWithoutDebug()) {
2933  // Cheap instructions viable for folding.
2934  if (isa<BinaryOperator>(I) || isa<GetElementPtrInst>(I) ||
2935  isa<StoreInst>(I))
2936  ++N;
2937  // Free instructions.
2938  else if (I.isTerminator() || IsaBitcastOfPointerType(I))
2939  continue;
2940  else
2941  return false;
2942  }
2943  // The store we want to merge is counted in N, so add 1 to make sure
2944  // we're counting the instructions that would be left.
2945  return N <= (PHINodeFoldingThreshold + 1);
2946  };
2947 
2949  (!IsWorthwhile(PTB) || !IsWorthwhile(PFB) || !IsWorthwhile(QTB) ||
2950  !IsWorthwhile(QFB)))
2951  return false;
2952 
2953  // For every pointer, there must be exactly two stores, one coming from
2954  // PTB or PFB, and the other from QTB or QFB. We don't support more than one
2955  // store (to any address) in PTB,PFB or QTB,QFB.
2956  // FIXME: We could relax this restriction with a bit more work and performance
2957  // testing.
2958  StoreInst *PStore = findUniqueStoreInBlocks(PTB, PFB);
2959  StoreInst *QStore = findUniqueStoreInBlocks(QTB, QFB);
2960  if (!PStore || !QStore)
2961  return false;
2962 
2963  // Now check the stores are compatible.
2964  if (!QStore->isUnordered() || !PStore->isUnordered())
2965  return false;
2966 
2967  // Check that sinking the store won't cause program behavior changes. Sinking
2968  // the store out of the Q blocks won't change any behavior as we're sinking
2969  // from a block to its unconditional successor. But we're moving a store from
2970  // the P blocks down through the middle block (QBI) and past both QFB and QTB.
2971  // So we need to check that there are no aliasing loads or stores in
2972  // QBI, QTB and QFB. We also need to check there are no conflicting memory
2973  // operations between PStore and the end of its parent block.
2974  //
2975  // The ideal way to do this is to query AliasAnalysis, but we don't
2976  // preserve AA currently so that is dangerous. Be super safe and just
2977  // check there are no other memory operations at all.
2978  for (auto &I : *QFB->getSinglePredecessor())
2979  if (I.mayReadOrWriteMemory())
2980  return false;
2981  for (auto &I : *QFB)
2982  if (&I != QStore && I.mayReadOrWriteMemory())
2983  return false;
2984  if (QTB)
2985  for (auto &I : *QTB)
2986  if (&I != QStore && I.mayReadOrWriteMemory())
2987  return false;
2988  for (auto I = BasicBlock::iterator(PStore), E = PStore->getParent()->end();
2989  I != E; ++I)
2990  if (&*I != PStore && I->mayReadOrWriteMemory())
2991  return false;
2992 
2993  // If PostBB has more than two predecessors, we need to split it so we can
2994  // sink the store.
2995  if (std::next(pred_begin(PostBB), 2) != pred_end(PostBB)) {
2996  // We know that QFB's only successor is PostBB. And QFB has a single
2997  // predecessor. If QTB exists, then its only successor is also PostBB.
2998  // If QTB does not exist, then QFB's only predecessor has a conditional
2999  // branch to QFB and PostBB.
3000  BasicBlock *TruePred = QTB ? QTB : QFB->getSinglePredecessor();
3001  BasicBlock *NewBB = SplitBlockPredecessors(PostBB, { QFB, TruePred},
3002  "condstore.split");
3003  if (!NewBB)
3004  return false;
3005  PostBB = NewBB;
3006  }
3007 
3008  // OK, we're going to sink the stores to PostBB. The store has to be
3009  // conditional though, so first create the predicate.
3010  Value *PCond = cast<BranchInst>(PFB->getSinglePredecessor()->getTerminator())
3011  ->getCondition();
3012  Value *QCond = cast<BranchInst>(QFB->getSinglePredecessor()->getTerminator())
3013  ->getCondition();
3014 
3016  PStore->getParent());
3018  QStore->getParent(), PPHI);
3019 
3020  IRBuilder<> QB(&*PostBB->getFirstInsertionPt());
3021 
3022  Value *PPred = PStore->getParent() == PTB ? PCond : QB.CreateNot(PCond);
3023  Value *QPred = QStore->getParent() == QTB ? QCond : QB.CreateNot(QCond);
3024 
3025  if (InvertPCond)
3026  PPred = QB.CreateNot(PPred);
3027  if (InvertQCond)
3028  QPred = QB.CreateNot(QPred);
3029  Value *CombinedPred = QB.CreateOr(PPred, QPred);
3030 
3031  auto *T =
3032  SplitBlockAndInsertIfThen(CombinedPred, &*QB.GetInsertPoint(), false);
3033  QB.SetInsertPoint(T);
3034  StoreInst *SI = cast<StoreInst>(QB.CreateStore(QPHI, Address));
3035  AAMDNodes AAMD;
3036  PStore->getAAMetadata(AAMD, /*Merge=*/false);
3037  PStore->getAAMetadata(AAMD, /*Merge=*/true);
3038  SI->setAAMetadata(AAMD);
3039  unsigned PAlignment = PStore->getAlignment();
3040  unsigned QAlignment = QStore->getAlignment();
3041  unsigned TypeAlignment =
3043  unsigned MinAlignment;
3044  unsigned MaxAlignment;
3045  std::tie(MinAlignment, MaxAlignment) = std::minmax(PAlignment, QAlignment);
3046  // Choose the minimum alignment. If we could prove both stores execute, we
3047  // could use biggest one. In this case, though, we only know that one of the
3048  // stores executes. And we don't know it's safe to take the alignment from a
3049  // store that doesn't execute.
3050  if (MinAlignment != 0) {
3051  // Choose the minimum of all non-zero alignments.
3052  SI->setAlignment(MinAlignment);
3053  } else if (MaxAlignment != 0) {
3054  // Choose the minimal alignment between the non-zero alignment and the ABI
3055  // default alignment for the type of the stored value.
3056  SI->setAlignment(std::min(MaxAlignment, TypeAlignment));
3057  } else {
3058  // If both alignments are zero, use ABI default alignment for the type of
3059  // the stored value.
3060  SI->setAlignment(TypeAlignment);
3061  }
3062 
3063  QStore->eraseFromParent();
3064  PStore->eraseFromParent();
3065 
3066  return true;
3067 }
3068 
3070  const DataLayout &DL) {
3071  // The intention here is to find diamonds or triangles (see below) where each
3072  // conditional block contains a store to the same address. Both of these
3073  // stores are conditional, so they can't be unconditionally sunk. But it may
3074  // be profitable to speculatively sink the stores into one merged store at the
3075  // end, and predicate the merged store on the union of the two conditions of
3076  // PBI and QBI.
3077  //
3078  // This can reduce the number of stores executed if both of the conditions are
3079  // true, and can allow the blocks to become small enough to be if-converted.
3080  // This optimization will also chain, so that ladders of test-and-set
3081  // sequences can be if-converted away.
3082  //
3083  // We only deal with simple diamonds or triangles:
3084  //
3085  // PBI or PBI or a combination of the two
3086  // / \ | \
3087  // PTB PFB | PFB
3088  // \ / | /
3089  // QBI QBI
3090  // / \ | \
3091  // QTB QFB | QFB
3092  // \ / | /
3093  // PostBB PostBB
3094  //
3095  // We model triangles as a type of diamond with a nullptr "true" block.
3096  // Triangles are canonicalized so that the fallthrough edge is represented by
3097  // a true condition, as in the diagram above.
3098  BasicBlock *PTB = PBI->getSuccessor(0);
3099  BasicBlock *PFB = PBI->getSuccessor(1);
3100  BasicBlock *QTB = QBI->getSuccessor(0);
3101  BasicBlock *QFB = QBI->getSuccessor(1);
3102  BasicBlock *PostBB = QFB->getSingleSuccessor();
3103 
3104  // Make sure we have a good guess for PostBB. If QTB's only successor is
3105  // QFB, then QFB is a better PostBB.
3106  if (QTB->getSingleSuccessor() == QFB)
3107  PostBB = QFB;
3108 
3109  // If we couldn't find a good PostBB, stop.
3110  if (!PostBB)
3111  return false;
3112 
3113  bool InvertPCond = false, InvertQCond = false;
3114  // Canonicalize fallthroughs to the true branches.
3115  if (PFB == QBI->getParent()) {
3116  std::swap(PFB, PTB);
3117  InvertPCond = true;
3118  }
3119  if (QFB == PostBB) {
3120  std::swap(QFB, QTB);
3121  InvertQCond = true;
3122  }
3123 
3124  // From this point on we can assume PTB or QTB may be fallthroughs but PFB
3125  // and QFB may not. Model fallthroughs as a nullptr block.
3126  if (PTB == QBI->getParent())
3127  PTB = nullptr;
3128  if (QTB == PostBB)
3129  QTB = nullptr;
3130 
3131  // Legality bailouts. We must have at least the non-fallthrough blocks and
3132  // the post-dominating block, and the non-fallthroughs must only have one
3133  // predecessor.
3134  auto HasOnePredAndOneSucc = [](BasicBlock *BB, BasicBlock *P, BasicBlock *S) {
3135  return BB->getSinglePredecessor() == P && BB->getSingleSuccessor() == S;
3136  };
3137  if (!HasOnePredAndOneSucc(PFB, PBI->getParent(), QBI->getParent()) ||
3138  !HasOnePredAndOneSucc(QFB, QBI->getParent(), PostBB))
3139  return false;
3140  if ((PTB && !HasOnePredAndOneSucc(PTB, PBI->getParent(), QBI->getParent())) ||
3141  (QTB && !HasOnePredAndOneSucc(QTB, QBI->getParent(), PostBB)))
3142  return false;
3143  if (!QBI->getParent()->hasNUses(2))
3144  return false;
3145 
3146  // OK, this is a sequence of two diamonds or triangles.
3147  // Check if there are stores in PTB or PFB that are repeated in QTB or QFB.
3148  SmallPtrSet<Value *, 4> PStoreAddresses, QStoreAddresses;
3149  for (auto *BB : {PTB, PFB}) {
3150  if (!BB)
3151  continue;
3152  for (auto &I : *BB)
3153  if (StoreInst *SI = dyn_cast<StoreInst>(&I))
3154  PStoreAddresses.insert(SI->getPointerOperand());
3155  }
3156  for (auto *BB : {QTB, QFB}) {
3157  if (!BB)
3158  continue;
3159  for (auto &I : *BB)
3160  if (StoreInst *SI = dyn_cast<StoreInst>(&I))
3161  QStoreAddresses.insert(SI->getPointerOperand());
3162  }
3163 
3164  set_intersect(PStoreAddresses, QStoreAddresses);
3165  // set_intersect mutates PStoreAddresses in place. Rename it here to make it
3166  // clear what it contains.
3167  auto &CommonAddresses = PStoreAddresses;
3168 
3169  bool Changed = false;
3170  for (auto *Address : CommonAddresses)
3171  Changed |= mergeConditionalStoreToAddress(
3172  PTB, PFB, QTB, QFB, PostBB, Address, InvertPCond, InvertQCond, DL);
3173  return Changed;
3174 }
3175 
3176 /// If we have a conditional branch as a predecessor of another block,
3177 /// this function tries to simplify it. We know
3178 /// that PBI and BI are both conditional branches, and BI is in one of the
3179 /// successor blocks of PBI - PBI branches to BI.
3181  const DataLayout &DL) {
3182  assert(PBI->isConditional() && BI->isConditional());
3183  BasicBlock *BB = BI->getParent();
3184 
3185  // If this block ends with a branch instruction, and if there is a
3186  // predecessor that ends on a branch of the same condition, make
3187  // this conditional branch redundant.
3188  if (PBI->getCondition() == BI->getCondition() &&
3189  PBI->getSuccessor(0) != PBI->getSuccessor(1)) {
3190  // Okay, the outcome of this conditional branch is statically
3191  // knowable. If this block had a single pred, handle specially.
3192  if (BB->getSinglePredecessor()) {
3193  // Turn this into a branch on constant.
3194  bool CondIsTrue = PBI->getSuccessor(0) == BB;
3195  BI->setCondition(
3196  ConstantInt::get(Type::getInt1Ty(BB->getContext()), CondIsTrue));
3197  return true; // Nuke the branch on constant.
3198  }
3199 
3200  // Otherwise, if there are multiple predecessors, insert a PHI that merges
3201  // in the constant and simplify the block result. Subsequent passes of
3202  // simplifycfg will thread the block.
3204  pred_iterator PB = pred_begin(BB), PE = pred_end(BB);
3205  PHINode *NewPN = PHINode::Create(
3206  Type::getInt1Ty(BB->getContext()), std::distance(PB, PE),
3207  BI->getCondition()->getName() + ".pr", &BB->front());
3208  // Okay, we're going to insert the PHI node. Since PBI is not the only
3209  // predecessor, compute the PHI'd conditional value for all of the preds.
3210  // Any predecessor where the condition is not computable we keep symbolic.
3211  for (pred_iterator PI = PB; PI != PE; ++PI) {
3212  BasicBlock *P = *PI;
3213  if ((PBI = dyn_cast<BranchInst>(P->getTerminator())) && PBI != BI &&
3214  PBI->isConditional() && PBI->getCondition() == BI->getCondition() &&
3215  PBI->getSuccessor(0) != PBI->getSuccessor(1)) {
3216  bool CondIsTrue = PBI->getSuccessor(0) == BB;
3217  NewPN->addIncoming(
3218  ConstantInt::get(Type::getInt1Ty(BB->getContext()), CondIsTrue),
3219  P);
3220  } else {
3221  NewPN->addIncoming(BI->getCondition(), P);
3222  }
3223  }
3224 
3225  BI->setCondition(NewPN);
3226  return true;
3227  }
3228  }
3229 
3230  if (auto *CE = dyn_cast<ConstantExpr>(BI->getCondition()))
3231  if (CE->canTrap())
3232  return false;
3233 
3234  // If both branches are conditional and both contain stores to the same
3235  // address, remove the stores from the conditionals and create a conditional
3236  // merged store at the end.
3237  if (MergeCondStores && mergeConditionalStores(PBI, BI, DL))
3238  return true;
3239 
3240  // If this is a conditional branch in an empty block, and if any
3241  // predecessors are a conditional branch to one of our destinations,
3242  // fold the conditions into logical ops and one cond br.
3243 
3244  // Ignore dbg intrinsics.
3245  if (&*BB->instructionsWithoutDebug().begin() != BI)
3246  return false;
3247 
3248  int PBIOp, BIOp;
3249  if (PBI->getSuccessor(0) == BI->getSuccessor(0)) {
3250  PBIOp = 0;
3251  BIOp = 0;
3252  } else if (PBI->getSuccessor(0) == BI->getSuccessor(1)) {
3253  PBIOp = 0;
3254  BIOp = 1;
3255  } else if (PBI->getSuccessor(1) == BI->getSuccessor(0)) {
3256  PBIOp = 1;
3257  BIOp = 0;
3258  } else if (PBI->getSuccessor(1) == BI->getSuccessor(1)) {
3259  PBIOp = 1;
3260  BIOp = 1;
3261  } else {
3262  return false;
3263  }
3264 
3265  // Check to make sure that the other destination of this branch
3266  // isn't BB itself. If so, this is an infinite loop that will
3267  // keep getting unwound.
3268  if (PBI->getSuccessor(PBIOp) == BB)
3269  return false;
3270 
3271  // Do not perform this transformation if it would require
3272  // insertion of a large number of select instructions. For targets
3273  // without predication/cmovs, this is a big pessimization.
3274 
3275  // Also do not perform this transformation if any phi node in the common
3276  // destination block can trap when reached by BB or PBB (PR17073). In that
3277  // case, it would be unsafe to hoist the operation into a select instruction.
3278 
3279  BasicBlock *CommonDest = PBI->getSuccessor(PBIOp);
3280  unsigned NumPhis = 0;
3281  for (BasicBlock::iterator II = CommonDest->begin(); isa<PHINode>(II);
3282  ++II, ++NumPhis) {
3283  if (NumPhis > 2) // Disable this xform.
3284  return false;
3285 
3286  PHINode *PN = cast<PHINode>(II);
3287  Value *BIV = PN->getIncomingValueForBlock(BB);
3288  if (ConstantExpr *CE = dyn_cast<ConstantExpr>(BIV))
3289  if (CE->canTrap())
3290  return false;
3291 
3292  unsigned PBBIdx = PN->getBasicBlockIndex(PBI->getParent());
3293  Value *PBIV = PN->getIncomingValue(PBBIdx);
3294  if (ConstantExpr *CE = dyn_cast<ConstantExpr>(PBIV))
3295  if (CE->canTrap())
3296  return false;
3297  }
3298 
3299  // Finally, if everything is ok, fold the branches to logical ops.
3300  BasicBlock *OtherDest = BI->getSuccessor(BIOp ^ 1);
3301 
3302  LLVM_DEBUG(dbgs() << "FOLDING BRs:" << *PBI->getParent()
3303  << "AND: " << *BI->getParent());
3304 
3305  // If OtherDest *is* BB, then BB is a basic block with a single conditional
3306  // branch in it, where one edge (OtherDest) goes back to itself but the other
3307  // exits. We don't *know* that the program avoids the infinite loop
3308  // (even though that seems likely). If we do this xform naively, we'll end up
3309  // recursively unpeeling the loop. Since we know that (after the xform is
3310  // done) that the block *is* infinite if reached, we just make it an obviously
3311  // infinite loop with no cond branch.
3312  if (OtherDest == BB) {
3313  // Insert it at the end of the function, because it's either code,
3314  // or it won't matter if it's hot. :)
3315  BasicBlock *InfLoopBlock =
3316  BasicBlock::Create(BB->getContext(), "infloop", BB->getParent());
3317  BranchInst::Create(InfLoopBlock, InfLoopBlock);
3318  OtherDest = InfLoopBlock;
3319  }
3320 
3321  LLVM_DEBUG(dbgs() << *PBI->getParent()->getParent());
3322 
3323  // BI may have other predecessors. Because of this, we leave
3324  // it alone, but modify PBI.
3325 
3326  // Make sure we get to CommonDest on True&True directions.
3327  Value *PBICond = PBI->getCondition();
3328  IRBuilder<NoFolder> Builder(PBI);
3329  if (PBIOp)
3330  PBICond = Builder.CreateNot(PBICond, PBICond->getName() + ".not");
3331 
3332  Value *BICond = BI->getCondition();
3333  if (BIOp)
3334  BICond = Builder.CreateNot(BICond, BICond->getName() + ".not");
3335 
3336  // Merge the conditions.
3337  Value *Cond = Builder.CreateOr(PBICond, BICond, "brmerge");
3338 
3339  // Modify PBI to branch on the new condition to the new dests.
3340  PBI->setCondition(Cond);
3341  PBI->setSuccessor(0, CommonDest);
3342  PBI->setSuccessor(1, OtherDest);
3343 
3344  // Update branch weight for PBI.
3345  uint64_t PredTrueWeight, PredFalseWeight, SuccTrueWeight, SuccFalseWeight;
3346  uint64_t PredCommon, PredOther, SuccCommon, SuccOther;
3347  bool HasWeights =
3348  extractPredSuccWeights(PBI, BI, PredTrueWeight, PredFalseWeight,
3349  SuccTrueWeight, SuccFalseWeight);
3350  if (HasWeights) {
3351  PredCommon = PBIOp ? PredFalseWeight : PredTrueWeight;
3352  PredOther = PBIOp ? PredTrueWeight : PredFalseWeight;
3353  SuccCommon = BIOp ? SuccFalseWeight : SuccTrueWeight;
3354  SuccOther = BIOp ? SuccTrueWeight : SuccFalseWeight;
3355  // The weight to CommonDest should be PredCommon * SuccTotal +
3356  // PredOther * SuccCommon.
3357  // The weight to OtherDest should be PredOther * SuccOther.
3358  uint64_t NewWeights[2] = {PredCommon * (SuccCommon + SuccOther) +
3359  PredOther * SuccCommon,
3360  PredOther * SuccOther};
3361  // Halve the weights if any of them cannot fit in an uint32_t
3362  FitWeights(NewWeights);
3363 
3364  setBranchWeights(PBI, NewWeights[0], NewWeights[1]);
3365  }
3366 
3367  // OtherDest may have phi nodes. If so, add an entry from PBI's
3368  // block that are identical to the entries for BI's block.
3369  AddPredecessorToBlock(OtherDest, PBI->getParent(), BB);
3370 
3371  // We know that the CommonDest already had an edge from PBI to
3372  // it. If it has PHIs though, the PHIs may have different
3373  // entries for BB and PBI's BB. If so, insert a select to make
3374  // them agree.
3375  for (PHINode &PN : CommonDest->phis()) {
3376  Value *BIV = PN.getIncomingValueForBlock(BB);
3377  unsigned PBBIdx = PN.getBasicBlockIndex(PBI->getParent());
3378  Value *PBIV = PN.getIncomingValue(PBBIdx);
3379  if (BIV != PBIV) {
3380  // Insert a select in PBI to pick the right value.
3381  SelectInst *NV = cast<SelectInst>(
3382  Builder.CreateSelect(PBICond, PBIV, BIV, PBIV->getName() + ".mux"));
3383  PN.setIncomingValue(PBBIdx, NV);
3384  // Although the select has the same condition as PBI, the original branch
3385  // weights for PBI do not apply to the new select because the select's
3386  // 'logical' edges are incoming edges of the phi that is eliminated, not
3387  // the outgoing edges of PBI.
3388  if (HasWeights) {
3389  uint64_t PredCommon = PBIOp ? PredFalseWeight : PredTrueWeight;
3390  uint64_t PredOther = PBIOp ? PredTrueWeight : PredFalseWeight;
3391  uint64_t SuccCommon = BIOp ? SuccFalseWeight : SuccTrueWeight;
3392  uint64_t SuccOther = BIOp ? SuccTrueWeight : SuccFalseWeight;
3393  // The weight to PredCommonDest should be PredCommon * SuccTotal.
3394  // The weight to PredOtherDest should be PredOther * SuccCommon.
3395  uint64_t NewWeights[2] = {PredCommon * (SuccCommon + SuccOther),
3396  PredOther * SuccCommon};
3397 
3398  FitWeights(NewWeights);
3399 
3400  setBranchWeights(NV, NewWeights[0], NewWeights[1]);
3401  }
3402  }
3403  }
3404 
3405  LLVM_DEBUG(dbgs() << "INTO: " << *PBI->getParent());
3406  LLVM_DEBUG(dbgs() << *PBI->getParent()->getParent());
3407 
3408  // This basic block is probably dead. We know it has at least
3409  // one fewer predecessor.
3410  return true;
3411 }
3412 
3413 // Simplifies a terminator by replacing it with a branch to TrueBB if Cond is
3414 // true or to FalseBB if Cond is false.
3415 // Takes care of updating the successors and removing the old terminator.
3416 // Also makes sure not to introduce new successors by assuming that edges to
3417 // non-successor TrueBBs and FalseBBs aren't reachable.
3418 static bool SimplifyTerminatorOnSelect(Instruction *OldTerm, Value *Cond,
3419  BasicBlock *TrueBB, BasicBlock *FalseBB,
3420  uint32_t TrueWeight,
3421  uint32_t FalseWeight) {
3422  // Remove any superfluous successor edges from the CFG.
3423  // First, figure out which successors to preserve.
3424  // If TrueBB and FalseBB are equal, only try to preserve one copy of that
3425  // successor.
3426  BasicBlock *KeepEdge1 = TrueBB;
3427  BasicBlock *KeepEdge2 = TrueBB != FalseBB ? FalseBB : nullptr;
3428 
3429  // Then remove the rest.
3430  for (BasicBlock *Succ : successors(OldTerm)) {
3431  // Make sure only to keep exactly one copy of each edge.
3432  if (Succ == KeepEdge1)
3433  KeepEdge1 = nullptr;
3434  else if (Succ == KeepEdge2)
3435  KeepEdge2 = nullptr;
3436  else
3437  Succ->removePredecessor(OldTerm->getParent(),
3438  /*KeepOneInputPHIs=*/true);
3439  }
3440 
3441  IRBuilder<> Builder(OldTerm);
3442  Builder.SetCurrentDebugLocation(OldTerm->getDebugLoc());
3443 
3444  // Insert an appropriate new terminator.
3445  if (!KeepEdge1 && !KeepEdge2) {
3446  if (TrueBB == FalseBB)
3447  // We were only looking for one successor, and it was present.
3448  // Create an unconditional branch to it.
3449  Builder.CreateBr(TrueBB);
3450  else {
3451  // We found both of the successors we were looking for.
3452  // Create a conditional branch sharing the condition of the select.
3453  BranchInst *NewBI = Builder.CreateCondBr(Cond, TrueBB, FalseBB);
3454  if (TrueWeight != FalseWeight)
3455  setBranchWeights(NewBI, TrueWeight, FalseWeight);
3456  }
3457  } else if (KeepEdge1 && (KeepEdge2 || TrueBB == FalseBB)) {
3458  // Neither of the selected blocks were successors, so this
3459  // terminator must be unreachable.
3460  new UnreachableInst(OldTerm->getContext(), OldTerm);
3461  } else {
3462  // One of the selected values was a successor, but the other wasn't.
3463  // Insert an unconditional branch to the one that was found;
3464  // the edge to the one that wasn't must be unreachable.
3465  if (!KeepEdge1)
3466  // Only TrueBB was found.
3467  Builder.CreateBr(TrueBB);
3468  else
3469  // Only FalseBB was found.
3470  Builder.CreateBr(FalseBB);
3471  }
3472 
3473  EraseTerminatorAndDCECond(OldTerm);
3474  return true;
3475 }
3476 
3477 // Replaces
3478 // (switch (select cond, X, Y)) on constant X, Y
3479 // with a branch - conditional if X and Y lead to distinct BBs,
3480 // unconditional otherwise.
3482  // Check for constant integer values in the select.
3483  ConstantInt *TrueVal = dyn_cast<ConstantInt>(Select->getTrueValue());
3484  ConstantInt *FalseVal = dyn_cast<ConstantInt>(Select->getFalseValue());
3485  if (!TrueVal || !FalseVal)
3486  return false;
3487 
3488  // Find the relevant condition and destinations.
3489  Value *Condition = Select->getCondition();
3490  BasicBlock *TrueBB = SI->findCaseValue(TrueVal)->getCaseSuccessor();
3491  BasicBlock *FalseBB = SI->findCaseValue(FalseVal)->getCaseSuccessor();
3492 
3493  // Get weight for TrueBB and FalseBB.
3494  uint32_t TrueWeight = 0, FalseWeight = 0;
3495  SmallVector<uint64_t, 8> Weights;
3496  bool HasWeights = HasBranchWeights(SI);
3497  if (HasWeights) {
3498  GetBranchWeights(SI, Weights);
3499  if (Weights.size() == 1 + SI->getNumCases()) {
3500  TrueWeight =
3501  (uint32_t)Weights[SI->findCaseValue(TrueVal)->getSuccessorIndex()];
3502  FalseWeight =
3503  (uint32_t)Weights[SI->findCaseValue(FalseVal)->getSuccessorIndex()];
3504  }
3505  }
3506 
3507  // Perform the actual simplification.
3508  return SimplifyTerminatorOnSelect(SI, Condition, TrueBB, FalseBB, TrueWeight,
3509  FalseWeight);
3510 }
3511 
3512 // Replaces
3513 // (indirectbr (select cond, blockaddress(@fn, BlockA),
3514 // blockaddress(@fn, BlockB)))
3515 // with
3516 // (br cond, BlockA, BlockB).
3518  // Check that both operands of the select are block addresses.
3521  if (!TBA || !FBA)
3522  return false;
3523 
3524  // Extract the actual blocks.
3525  BasicBlock *TrueBB = TBA->getBasicBlock();
3526  BasicBlock *FalseBB = FBA->getBasicBlock();
3527 
3528  // Perform the actual simplification.
3529  return SimplifyTerminatorOnSelect(IBI, SI->getCondition(), TrueBB, FalseBB, 0,
3530  0);
3531 }
3532 
3533 /// This is called when we find an icmp instruction
3534 /// (a seteq/setne with a constant) as the only instruction in a
3535 /// block that ends with an uncond branch. We are looking for a very specific
3536 /// pattern that occurs when "A == 1 || A == 2 || A == 3" gets simplified. In
3537 /// this case, we merge the first two "or's of icmp" into a switch, but then the
3538 /// default value goes to an uncond block with a seteq in it, we get something
3539 /// like:
3540 ///
3541 /// switch i8 %A, label %DEFAULT [ i8 1, label %end i8 2, label %end ]
3542 /// DEFAULT:
3543 /// %tmp = icmp eq i8 %A, 92
3544 /// br label %end
3545 /// end:
3546 /// ... = phi i1 [ true, %entry ], [ %tmp, %DEFAULT ], [ true, %entry ]
3547 ///
3548 /// We prefer to split the edge to 'end' so that there is a true/false entry to
3549 /// the PHI, merging the third icmp into the switch.
3550 bool SimplifyCFGOpt::tryToSimplifyUncondBranchWithICmpInIt(
3551  ICmpInst *ICI, IRBuilder<> &Builder) {
3552  BasicBlock *BB = ICI->getParent();
3553 
3554  // If the block has any PHIs in it or the icmp has multiple uses, it is too
3555  // complex.
3556  if (isa<PHINode>(BB->begin()) || !ICI->hasOneUse())
3557  return false;
3558 
3559  Value *V = ICI->getOperand(0);
3560  ConstantInt *Cst = cast<ConstantInt>(ICI->getOperand(1));
3561 
3562  // The pattern we're looking for is where our only predecessor is a switch on
3563  // 'V' and this block is the default case for the switch. In this case we can
3564  // fold the compared value into the switch to simplify things.
3565  BasicBlock *Pred = BB->getSinglePredecessor();
3566  if (!Pred || !isa<SwitchInst>(Pred->getTerminator()))
3567  return false;
3568 
3569  SwitchInst *SI = cast<SwitchInst>(Pred->getTerminator());
3570  if (SI->getCondition() != V)
3571  return false;
3572 
3573  // If BB is reachable on a non-default case, then we simply know the value of
3574  // V in this block. Substitute it and constant fold the icmp instruction
3575  // away.
3576  if (SI->getDefaultDest() != BB) {
3577  ConstantInt *VVal = SI->findCaseDest(BB);
3578  assert(VVal && "Should have a unique destination value");
3579  ICI->setOperand(0, VVal);
3580 
3581  if (Value *V = SimplifyInstruction(ICI, {DL, ICI})) {
3582  ICI->replaceAllUsesWith(V);
3583  ICI->eraseFromParent();
3584  }
3585  // BB is now empty, so it is likely to simplify away.
3586  return requestResimplify();
3587  }
3588 
3589  // Ok, the block is reachable from the default dest. If the constant we're
3590  // comparing exists in one of the other edges, then we can constant fold ICI
3591  // and zap it.
3592  if (SI->findCaseValue(Cst) != SI->case_default()) {
3593  Value *V;
3594  if (ICI->getPredicate() == ICmpInst::ICMP_EQ)
3595  V = ConstantInt::getFalse(BB->getContext());
3596  else
3597  V = ConstantInt::getTrue(BB->getContext());
3598 
3599  ICI->replaceAllUsesWith(V);
3600  ICI->eraseFromParent();
3601  // BB is now empty, so it is likely to simplify away.
3602  return requestResimplify();
3603  }
3604 
3605  // The use of the icmp has to be in the 'end' block, by the only PHI node in
3606  // the block.
3607  BasicBlock *SuccBlock = BB->getTerminator()->getSuccessor(0);
3608  PHINode *PHIUse = dyn_cast<PHINode>(ICI->user_back());
3609  if (PHIUse == nullptr || PHIUse != &SuccBlock->front() ||
3610  isa<PHINode>(++BasicBlock::iterator(PHIUse)))
3611  return false;
3612 
3613  // If the icmp is a SETEQ, then the default dest gets false, the new edge gets
3614  // true in the PHI.
3615  Constant *DefaultCst = ConstantInt::getTrue(BB->getContext());
3616  Constant *NewCst = ConstantInt::getFalse(BB->getContext());
3617 
3618  if (ICI->getPredicate() == ICmpInst::ICMP_EQ)
3619  std::swap(DefaultCst, NewCst);
3620 
3621  // Replace ICI (which is used by the PHI for the default value) with true or
3622  // false depending on if it is EQ or NE.
3623  ICI->replaceAllUsesWith(DefaultCst);
3624  ICI->eraseFromParent();
3625 
3626  // Okay, the switch goes to this block on a default value. Add an edge from
3627  // the switch to the merge point on the compared value.
3628  BasicBlock *NewBB =
3629  BasicBlock::Create(BB->getContext(), "switch.edge", BB->getParent(), BB);
3630  {
3631  SwitchInstProfUpdateWrapper SIW(*SI);
3632  auto W0 = SIW.getSuccessorWeight(0);
3634  if (W0) {
3635  NewW = ((uint64_t(*W0) + 1) >> 1);
3636  SIW.setSuccessorWeight(0, *NewW);
3637  }
3638  SIW.addCase(Cst, NewBB, NewW);
3639  }
3640 
3641  // NewBB branches to the phi block, add the uncond branch and the phi entry.
3642  Builder.SetInsertPoint(NewBB);
3643  Builder.SetCurrentDebugLocation(SI->getDebugLoc());
3644  Builder.CreateBr(SuccBlock);
3645  PHIUse->addIncoming(NewCst, NewBB);
3646  return true;
3647 }
3648 
3649 /// The specified branch is a conditional branch.
3650 /// Check to see if it is branching on an or/and chain of icmp instructions, and
3651 /// fold it into a switch instruction if so.
3653  const DataLayout &DL) {
3654  Instruction *Cond = dyn_cast<Instruction>(BI->getCondition());
3655  if (!Cond)
3656  return false;
3657 
3658  // Change br (X == 0 | X == 1), T, F into a switch instruction.
3659  // If this is a bunch of seteq's or'd together, or if it's a bunch of
3660  // 'setne's and'ed together, collect them.
3661 
3662  // Try to gather values from a chain of and/or to be turned into a switch
3663  ConstantComparesGatherer ConstantCompare(Cond, DL);
3664  // Unpack the result
3665  SmallVectorImpl<ConstantInt *> &Values = ConstantCompare.Vals;
3666  Value *CompVal = ConstantCompare.CompValue;
3667  unsigned UsedICmps = ConstantCompare.UsedICmps;
3668  Value *ExtraCase = ConstantCompare.Extra;
3669 
3670  // If we didn't have a multiply compared value, fail.
3671  if (!CompVal)
3672  return false;
3673 
3674  // Avoid turning single icmps into a switch.
3675  if (UsedICmps <= 1)
3676  return false;
3677 
3678  bool TrueWhenEqual = (Cond->getOpcode() == Instruction::Or);
3679 
3680  // There might be duplicate constants in the list, which the switch
3681  // instruction can't handle, remove them now.
3682  array_pod_sort(Values.begin(), Values.end(), ConstantIntSortPredicate);
3683  Values.erase(std::unique(Values.begin(), Values.end()), Values.end());
3684 
3685  // If Extra was used, we require at least two switch values to do the
3686  // transformation. A switch with one value is just a conditional branch.
3687  if (ExtraCase && Values.size() < 2)
3688  return false;
3689 
3690  // TODO: Preserve branch weight metadata, similarly to how
3691  // FoldValueComparisonIntoPredecessors preserves it.
3692 
3693  // Figure out which block is which destination.
3694  BasicBlock *DefaultBB = BI->getSuccessor(1);
3695  BasicBlock *EdgeBB = BI->getSuccessor(0);
3696  if (!TrueWhenEqual)
3697  std::swap(DefaultBB, EdgeBB);
3698 
3699  BasicBlock *BB = BI->getParent();
3700 
3701  LLVM_DEBUG(dbgs() << "Converting 'icmp' chain with " << Values.size()
3702  << " cases into SWITCH. BB is:\n"
3703  << *BB);
3704 
3705  // If there are any extra values that couldn't be folded into the switch
3706  // then we evaluate them with an explicit branch first. Split the block
3707  // right before the condbr to handle it.
3708  if (ExtraCase) {
3709  BasicBlock *NewBB =
3710  BB->splitBasicBlock(BI->getIterator(), "switch.early.test");
3711  // Remove the uncond branch added to the old block.
3712  Instruction *OldTI = BB->getTerminator();
3713  Builder.SetInsertPoint(OldTI);
3714 
3715  if (TrueWhenEqual)
3716  Builder.CreateCondBr(ExtraCase, EdgeBB, NewBB);
3717  else
3718  Builder.CreateCondBr(ExtraCase, NewBB, EdgeBB);
3719 
3720  OldTI->eraseFromParent();
3721 
3722  // If there are PHI nodes in EdgeBB, then we need to add a new entry to them
3723  // for the edge we just added.
3724  AddPredecessorToBlock(EdgeBB, BB, NewBB);
3725 
3726  LLVM_DEBUG(dbgs() << " ** 'icmp' chain unhandled condition: " << *ExtraCase
3727  << "\nEXTRABB = " << *BB);
3728  BB = NewBB;
3729  }
3730 
3731  Builder.SetInsertPoint(BI);
3732  // Convert pointer to int before we switch.
3733  if (CompVal->getType()->isPointerTy()) {
3734  CompVal = Builder.CreatePtrToInt(
3735  CompVal, DL.getIntPtrType(CompVal->getType()), "magicptr");
3736  }
3737 
3738  // Create the new switch instruction now.
3739  SwitchInst *New = Builder.CreateSwitch(CompVal, DefaultBB, Values.size());
3740 
3741  // Add all of the 'cases' to the switch instruction.
3742  for (unsigned i = 0, e = Values.size(); i != e; ++i)
3743  New->addCase(Values[i], EdgeBB);
3744 
3745  // We added edges from PI to the EdgeBB. As such, if there were any
3746  // PHI nodes in EdgeBB, they need entries to be added corresponding to
3747  // the number of edges added.
3748  for (BasicBlock::iterator BBI = EdgeBB->begin(); isa<PHINode>(BBI); ++BBI) {
3749  PHINode *PN = cast<PHINode>(BBI);
3750  Value *InVal = PN->getIncomingValueForBlock(BB);
3751  for (unsigned i = 0, e = Values.size() - 1; i != e; ++i)
3752  PN->addIncoming(InVal, BB);
3753  }
3754 
3755  // Erase the old branch instruction.
3757 
3758  LLVM_DEBUG(dbgs() << " ** 'icmp' chain result is:\n" << *BB << '\n');
3759  return true;
3760 }
3761 
3762 bool SimplifyCFGOpt::SimplifyResume(ResumeInst *RI, IRBuilder<> &Builder) {
3763  if (isa<PHINode>(RI->getValue()))
3764  return SimplifyCommonResume(RI);
3765  else if (isa<LandingPadInst>(RI->getParent()->getFirstNonPHI()) &&
3766  RI->getValue() == RI->getParent()->getFirstNonPHI())
3767  // The resume must unwind the exception that caused control to branch here.
3768  return SimplifySingleResume(RI);
3769 
3770  return false;
3771 }
3772 
3773 // Simplify resume that is shared by several landing pads (phi of landing pad).
3774 bool SimplifyCFGOpt::SimplifyCommonResume(ResumeInst *RI) {
3775  BasicBlock *BB = RI->getParent();
3776 
3777  // Check that there are no other instructions except for debug intrinsics
3778  // between the phi of landing pads (RI->getValue()) and resume instruction.
3779  BasicBlock::iterator I = cast<Instruction>(RI->getValue())->getIterator(),
3780  E = RI->getIterator();
3781  while (++I != E)
3782  if (!isa<DbgInfoIntrinsic>(I))
3783  return false;
3784 
3785  SmallSetVector<BasicBlock *, 4> TrivialUnwindBlocks;
3786  auto *PhiLPInst = cast<PHINode>(RI->getValue());
3787 
3788  // Check incoming blocks to see if any of them are trivial.
3789  for (unsigned Idx = 0, End = PhiLPInst->getNumIncomingValues(); Idx != End;
3790  Idx++) {
3791  auto *IncomingBB = PhiLPInst->getIncomingBlock(Idx);
3792  auto *IncomingValue = PhiLPInst->getIncomingValue(Idx);
3793 
3794  // If the block has other successors, we can not delete it because
3795  // it has other dependents.
3796  if (IncomingBB->getUniqueSuccessor() != BB)
3797  continue;
3798 
3799  auto *LandingPad = dyn_cast<LandingPadInst>(IncomingBB->getFirstNonPHI());
3800  // Not the landing pad that caused the control to branch here.
3801  if (IncomingValue != LandingPad)
3802  continue;
3803 
3804  bool isTrivial = true;
3805 
3806  I = IncomingBB->getFirstNonPHI()->getIterator();
3807  E = IncomingBB->getTerminator()->getIterator();
3808  while (++I != E)
3809  if (!isa<DbgInfoIntrinsic>(I)) {
3810  isTrivial = false;
3811  break;
3812  }
3813 
3814  if (isTrivial)
3815  TrivialUnwindBlocks.insert(IncomingBB);
3816  }
3817 
3818  // If no trivial unwind blocks, don't do any simplifications.
3819  if (TrivialUnwindBlocks.empty())
3820  return false;
3821 
3822  // Turn all invokes that unwind here into calls.
3823  for (auto *TrivialBB : TrivialUnwindBlocks) {
3824  // Blocks that will be simplified should be removed from the phi node.
3825  // Note there could be multiple edges to the resume block, and we need
3826  // to remove them all.
3827  while (PhiLPInst->getBasicBlockIndex(TrivialBB) != -1)
3828  BB->removePredecessor(TrivialBB, true);
3829 
3830  for (pred_iterator PI = pred_begin(TrivialBB), PE = pred_end(TrivialBB);
3831  PI != PE;) {
3832  BasicBlock *Pred = *PI++;
3833  removeUnwindEdge(Pred);
3834  }
3835 
3836  // In each SimplifyCFG run, only the current processed block can be erased.
3837  // Otherwise, it will break the iteration of SimplifyCFG pass. So instead
3838  // of erasing TrivialBB, we only remove the branch to the common resume
3839  // block so that we can later erase the resume block since it has no
3840  // predecessors.
3841  TrivialBB->getTerminator()->eraseFromParent();
3842  new UnreachableInst(RI->getContext(), TrivialBB);
3843  }
3844 
3845  // Delete the resume block if all its predecessors have been removed.
3846  if (pred_empty(BB))
3847  BB->eraseFromParent();
3848 
3849  return !TrivialUnwindBlocks.empty();
3850 }
3851 
3852 // Simplify resume that is only used by a single (non-phi) landing pad.
3853 bool SimplifyCFGOpt::SimplifySingleResume(ResumeInst *RI) {
3854  BasicBlock *BB = RI->getParent();
3856  assert(RI->getValue() == LPInst &&
3857  "Resume must unwind the exception that caused control to here");
3858 
3859  // Check that there are no other instructions except for debug intrinsics.
3860  BasicBlock::iterator I = LPInst->getIterator(), E = RI->getIterator();
3861  while (++I != E)
3862  if (!isa<DbgInfoIntrinsic>(I))
3863  return false;
3864 
3865  // Turn all invokes that unwind here into calls and delete the basic block.
3866  for (pred_iterator PI = pred_begin(BB), PE = pred_end(BB); PI != PE;) {
3867  BasicBlock *Pred = *PI++;
3868  removeUnwindEdge(Pred);
3869  }
3870 
3871  // The landingpad is now unreachable. Zap it.
3872  if (LoopHeaders)
3873  LoopHeaders->erase(BB);
3874  BB->eraseFromParent();
3875  return true;
3876 }
3877 
3879  // If this is a trivial cleanup pad that executes no instructions, it can be
3880  // eliminated. If the cleanup pad continues to the caller, any predecessor
3881  // that is an EH pad will be updated to continue to the caller and any
3882  // predecessor that terminates with an invoke instruction will have its invoke
3883  // instruction converted to a call instruction. If the cleanup pad being
3884  // simplified does not continue to the caller, each predecessor will be
3885  // updated to continue to the unwind destination of the cleanup pad being
3886  // simplified.
3887  BasicBlock *BB = RI->getParent();
3888  CleanupPadInst *CPInst = RI->getCleanupPad();
3889  if (CPInst->getParent() != BB)
3890  // This isn't an empty cleanup.
3891  return false;
3892 
3893  // We cannot kill the pad if it has multiple uses. This typically arises
3894  // from unreachable basic blocks.
3895  if (!CPInst->hasOneUse())
3896  return false;
3897 
3898  // Check that there are no other instructions except for benign intrinsics.
3899  BasicBlock::iterator I = CPInst->getIterator(), E = RI->getIterator();
3900  while (++I != E) {
3901  auto *II = dyn_cast<IntrinsicInst>(I);
3902  if (!II)
3903  return false;
3904 
3905  Intrinsic::ID IntrinsicID = II->getIntrinsicID();
3906  switch (IntrinsicID) {
3907  case Intrinsic::dbg_declare:
3908  case Intrinsic::dbg_value:
3909  case Intrinsic::dbg_label:
3910  case Intrinsic::lifetime_end:
3911  break;
3912  default:
3913  return false;
3914  }
3915  }
3916 
3917  // If the cleanup return we are simplifying unwinds to the caller, this will
3918  // set UnwindDest to nullptr.
3919  BasicBlock *UnwindDest = RI->getUnwindDest();
3920  Instruction *DestEHPad = UnwindDest ? UnwindDest->getFirstNonPHI() : nullptr;
3921 
3922  // We're about to remove BB from the control flow. Before we do, sink any
3923  // PHINodes into the unwind destination. Doing this before changing the
3924  // control flow avoids some potentially slow checks, since we can currently
3925  // be certain that UnwindDest and BB have no common predecessors (since they
3926  // are both EH pads).
3927  if (UnwindDest) {
3928  // First, go through the PHI nodes in UnwindDest and update any nodes that
3929  // reference the block we are removing
3930  for (BasicBlock::iterator I = UnwindDest->begin(),
3931  IE = DestEHPad->getIterator();
3932  I != IE; ++I) {
3933  PHINode *DestPN = cast<PHINode>(I);
3934 
3935  int Idx = DestPN->getBasicBlockIndex(BB);
3936  // Since BB unwinds to UnwindDest, it has to be in the PHI node.
3937  assert(Idx != -1);
3938  // This PHI node has an incoming value that corresponds to a control
3939  // path through the cleanup pad we are removing. If the incoming
3940  // value is in the cleanup pad, it must be a PHINode (because we
3941  // verified above that the block is otherwise empty). Otherwise, the
3942  // value is either a constant or a value that dominates the cleanup
3943  // pad being removed.
3944  //
3945  // Because BB and UnwindDest are both EH pads, all of their
3946  // predecessors must unwind to these blocks, and since no instruction
3947  // can have multiple unwind destinations, there will be no overlap in
3948  // incoming blocks between SrcPN and DestPN.
3949  Value *SrcVal = DestPN->getIncomingValue(Idx);
3950  PHINode *SrcPN = dyn_cast<PHINode>(SrcVal);
3951 
3952  // Remove the entry for the block we are deleting.
3953  DestPN->removeIncomingValue(Idx, false);
3954 
3955  if (SrcPN && SrcPN->getParent() == BB) {
3956  // If the incoming value was a PHI node in the cleanup pad we are
3957  // removing, we need to merge that PHI node's incoming values into
3958  // DestPN.
3959  for (unsigned SrcIdx = 0, SrcE = SrcPN->getNumIncomingValues();
3960  SrcIdx != SrcE; ++SrcIdx) {
3961  DestPN->addIncoming(SrcPN->getIncomingValue(SrcIdx),
3962  SrcPN->getIncomingBlock(SrcIdx));
3963  }
3964  } else {
3965  // Otherwise, the incoming value came from above BB and
3966  // so we can just reuse it. We must associate all of BB's
3967  // predecessors with this value.
3968  for (auto *pred : predecessors(BB)) {
3969  DestPN->addIncoming(SrcVal, pred);
3970  }
3971  }
3972  }
3973 
3974  // Sink any remaining PHI nodes directly into UnwindDest.
3975  Instruction *InsertPt = DestEHPad;
3976  for (BasicBlock::iterator I = BB->begin(),
3977  IE = BB->getFirstNonPHI()->getIterator();
3978  I != IE;) {
3979  // The iterator must be incremented here because the instructions are
3980  // being moved to another block.
3981  PHINode *PN = cast<PHINode>(I++);
3982  if (PN->use_empty())
3983  // If the PHI node has no uses, just leave it. It will be erased
3984  // when we erase BB below.
3985  continue;
3986 
3987  // Otherwise, sink this PHI node into UnwindDest.
3988  // Any predecessors to UnwindDest which are not already represented
3989  // must be back edges which inherit the value from the path through
3990  // BB. In this case, the PHI value must reference itself.
3991  for (auto *pred : predecessors(UnwindDest))
3992  if (pred != BB)
3993  PN->addIncoming(PN, pred);
3994  PN->moveBefore(InsertPt);
3995  }
3996  }
3997 
3998  for (pred_iterator PI = pred_begin(BB), PE = pred_end(BB); PI != PE;) {
3999  // The iterator must be updated here because we are removing this pred.
4000  BasicBlock *PredBB = *PI++;
4001  if (UnwindDest == nullptr) {
4002  removeUnwindEdge(PredBB);
4003  } else {
4004  Instruction *TI = PredBB->getTerminator();
4005  TI->replaceUsesOfWith(BB, UnwindDest);
4006  }
4007  }
4008 
4009  // The cleanup pad is now unreachable. Zap it.
4010  BB->eraseFromParent();
4011  return true;
4012 }
4013 
4014 // Try to merge two cleanuppads together.
4016  // Skip any cleanuprets which unwind to caller, there is nothing to merge
4017  // with.
4018  BasicBlock *UnwindDest = RI->getUnwindDest();
4019  if (!UnwindDest)
4020  return false;
4021 
4022  // This cleanupret isn't the only predecessor of this cleanuppad, it wouldn't
4023  // be safe to merge without code duplication.
4024  if (UnwindDest->getSinglePredecessor() != RI->getParent())
4025  return false;
4026 
4027  // Verify that our cleanuppad's unwind destination is another cleanuppad.
4028  auto *SuccessorCleanupPad = dyn_cast<CleanupPadInst>(&UnwindDest->front());
4029  if (!SuccessorCleanupPad)
4030  return false;
4031 
4032  CleanupPadInst *PredecessorCleanupPad = RI->getCleanupPad();
4033  // Replace any uses of the successor cleanupad with the predecessor pad
4034  // The only cleanuppad uses should be this cleanupret, it's cleanupret and
4035  // funclet bundle operands.
4036  SuccessorCleanupPad->replaceAllUsesWith(PredecessorCleanupPad);
4037  // Remove the old cleanuppad.
4038  SuccessorCleanupPad->eraseFromParent();
4039  // Now, we simply replace the cleanupret with a branch to the unwind
4040  // destination.
4041  BranchInst::Create(UnwindDest, RI->getParent());
4042  RI->eraseFromParent();
4043 
4044  return true;
4045 }
4046 
4047 bool SimplifyCFGOpt::SimplifyCleanupReturn(CleanupReturnInst *RI) {
4048  // It is possible to transiantly have an undef cleanuppad operand because we
4049  // have deleted some, but not all, dead blocks.
4050  // Eventually, this block will be deleted.
4051  if (isa<UndefValue>(RI->getOperand(0)))
4052  return false;
4053 
4054  if (mergeCleanupPad(RI))
4055  return true;
4056 
4057  if (removeEmptyCleanup(RI))
4058  return true;
4059 
4060  return false;
4061 }
4062 
4063 bool SimplifyCFGOpt::SimplifyReturn(ReturnInst *RI, IRBuilder<> &Builder) {
4064  BasicBlock *BB = RI->getParent();
4065  if (!BB->getFirstNonPHIOrDbg()->isTerminator())
4066  return false;
4067 
4068  // Find predecessors that end with branches.
4069  SmallVector<BasicBlock *, 8> UncondBranchPreds;
4070  SmallVector<BranchInst *, 8> CondBranchPreds;
4071  for (pred_iterator PI = pred_begin(BB), E = pred_end(BB); PI != E; ++PI) {
4072  BasicBlock *P = *PI;
4073  Instruction *PTI = P->getTerminator();
4074  if (BranchInst *BI = dyn_cast<BranchInst>(PTI)) {
4075  if (BI->isUnconditional())
4076  UncondBranchPreds.push_back(P);
4077  else
4078  CondBranchPreds.push_back(BI);
4079  }
4080  }
4081 
4082  // If we found some, do the transformation!
4083  if (!UncondBranchPreds.empty() && DupRet) {
4084  while (!UncondBranchPreds.empty()) {
4085  BasicBlock *Pred = UncondBranchPreds.pop_back_val();
4086  LLVM_DEBUG(dbgs() << "FOLDING: " << *BB
4087  << "INTO UNCOND BRANCH PRED: " << *Pred);
4088  (void)FoldReturnIntoUncondBranch(RI, BB, Pred);
4089  }
4090 
4091  // If we eliminated all predecessors of the block, delete the block now.
4092  if (pred_empty(BB)) {
4093  // We know there are no successors, so just nuke the block.
4094  if (LoopHeaders)
4095  LoopHeaders->erase(BB);
4096  BB->eraseFromParent();
4097  }
4098 
4099  return true;
4100  }
4101 
4102  // Check out all of the conditional branches going to this return
4103  // instruction. If any of them just select between returns, change the
4104  // branch itself into a select/return pair.
4105  while (!CondBranchPreds.empty()) {
4106  BranchInst *BI = CondBranchPreds.pop_back_val();
4107 
4108  // Check to see if the non-BB successor is also a return block.
4109  if (isa<ReturnInst>(BI->getSuccessor(0)->getTerminator()) &&
4110  isa<ReturnInst>(BI->getSuccessor(1)->getTerminator()) &&
4111  SimplifyCondBranchToTwoReturns(BI, Builder))
4112  return true;
4113  }
4114  return false;
4115 }
4116 
4117 bool SimplifyCFGOpt::SimplifyUnreachable(UnreachableInst *UI) {
4118  BasicBlock *BB = UI->getParent();
4119 
4120  bool Changed = false;
4121 
4122  // If there are any instructions immediately before the unreachable that can
4123  // be removed, do so.
4124  while (UI->getIterator() != BB->begin()) {
4125  BasicBlock::iterator BBI = UI->getIterator();
4126  --BBI;
4127  // Do not delete instructions that can have side effects which might cause
4128  // the unreachable to not be reachable; specifically, calls and volatile
4129  // operations may have this effect.
4130  if (isa<CallInst>(BBI) && !isa<DbgInfoIntrinsic>(BBI))
4131  break;
4132 
4133  if (BBI->mayHaveSideEffects()) {
4134  if (auto *SI = dyn_cast<StoreInst>(BBI)) {
4135  if (SI->isVolatile())
4136  break;
4137  } else if (auto *LI = dyn_cast<LoadInst>(BBI)) {
4138  if (LI->isVolatile())
4139  break;
4140  } else if (auto *RMWI = dyn_cast<AtomicRMWInst>(BBI)) {
4141  if (RMWI->isVolatile())
4142  break;
4143  } else if (auto *CXI = dyn_cast<AtomicCmpXchgInst>(BBI)) {
4144  if (CXI->isVolatile())
4145  break;
4146  } else if (isa<CatchPadInst>(BBI)) {
4147  // A catchpad may invoke exception object constructors and such, which
4148  // in some languages can be arbitrary code, so be conservative by
4149  // default.
4150  // For CoreCLR, it just involves a type test, so can be removed.
4153  break;
4154  } else if (!isa<FenceInst>(BBI) && !isa<VAArgInst>(BBI) &&
4155  !isa<LandingPadInst>(BBI)) {
4156  break;
4157  }
4158  // Note that deleting LandingPad's here is in fact okay, although it
4159  // involves a bit of subtle reasoning. If this inst is a LandingPad,
4160  // all the predecessors of this block will be the unwind edges of Invokes,
4161  // and we can therefore guarantee this block will be erased.
4162  }
4163 
4164  // Delete this instruction (any uses are guaranteed to be dead)
4165  if (!BBI->use_empty())
4166  BBI->replaceAllUsesWith(UndefValue::get(BBI->getType()));
4167  BBI->eraseFromParent();
4168  Changed = true;
4169  }
4170 
4171  // If the unreachable instruction is the first in the block, take a gander
4172  // at all of the predecessors of this instruction, and simplify them.
4173  if (&BB->front() != UI)
4174  return Changed;
4175 
4177  for (unsigned i = 0, e = Preds.size(); i != e; ++i) {
4178  Instruction *TI = Preds[i]->getTerminator();
4179  IRBuilder<> Builder(TI);
4180  if (auto *BI = dyn_cast<BranchInst>(TI)) {
4181  if (BI->isUnconditional()) {
4182  assert(BI->getSuccessor(0) == BB && "Incorrect CFG");
4183  new UnreachableInst(TI->getContext(), TI);
4184  TI->eraseFromParent();
4185  Changed = true;
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));
4191  } else {
4192  assert(BI->getSuccessor(1) == BB && "Incorrect CFG");
4193  Builder.CreateAssumption(Cond);
4194  Builder.CreateBr(BI->getSuccessor(0));
4195  }
4197  Changed = true;
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 
4280  LLVM_DEBUG(dbgs() << "SimplifyCFG: switch default is dead.\n");
4281  BasicBlock *NewDefaultBlock =
4282  SplitBlockPredecessors(Switch->getDefaultDest(), Switch->getParent(), "");
4283  Switch->setDefaultDest(&*NewDefaultBlock);
4284  SplitBlock(&*NewDefaultBlock, &NewDefaultBlock->front());
4285  auto *NewTerminator = NewDefaultBlock->getTerminator();
4286  new UnreachableInst(Switch->getContext(), NewTerminator);
4287  EraseTerminatorAndDCECond(NewTerminator);
4288 }
4289 
4290 /// Turn a switch with two reachable destinations into an integer range
4291 /// comparison and branch.
4293  assert(SI->getNumCases() > 1 && "Degenerate switch?");
4294 
4295  bool HasDefault =
4296  !isa<UnreachableInst>(SI->getDefaultDest()->getFirstNonPHIOrDbg());
4297 
4298  // Partition the cases into two sets with different destinations.
4299  BasicBlock *DestA = HasDefault ? SI->getDefaultDest() : nullptr;
4300  BasicBlock *DestB = nullptr;
4303 
4304  for (auto Case : SI->cases()) {
4305  BasicBlock *Dest = Case.getCaseSuccessor();
4306  if (!DestA)
4307  DestA = Dest;
4308  if (Dest == DestA) {
4309  CasesA.push_back(Case.getCaseValue());
4310  continue;
4311  }
4312  if (!DestB)
4313  DestB = Dest;
4314  if (Dest == DestB) {
4315  CasesB.push_back(Case.getCaseValue());
4316  continue;
4317  }
4318  return false; // More than two destinations.
4319  }
4320 
4321  assert(DestA && DestB &&
4322  "Single-destination switch should have been folded.");
4323  assert(DestA != DestB);
4324  assert(DestB != SI->getDefaultDest());
4325  assert(!CasesB.empty() && "There must be non-default cases.");
4326  assert(!CasesA.empty() || HasDefault);
4327 
4328  // Figure out if one of the sets of cases form a contiguous range.
4329  SmallVectorImpl<ConstantInt *> *ContiguousCases = nullptr;
4330  BasicBlock *ContiguousDest = nullptr;
4331  BasicBlock *OtherDest = nullptr;
4332  if (!CasesA.empty() && CasesAreContiguous(CasesA)) {
4333  ContiguousCases = &CasesA;
4334  ContiguousDest = DestA;
4335  OtherDest = DestB;
4336  } else if (CasesAreContiguous(CasesB)) {
4337  ContiguousCases = &CasesB;
4338  ContiguousDest = DestB;
4339  OtherDest = DestA;
4340  } else
4341  return false;
4342 
4343  // Start building the compare and branch.
4344 
4345  Constant *Offset = ConstantExpr::getNeg(ContiguousCases->back());
4346  Constant *NumCases =
4347  ConstantInt::get(Offset->getType(), ContiguousCases->size());
4348 
4349  Value *Sub = SI->getCondition();
4350  if (!Offset->isNullValue())
4351  Sub = Builder.CreateAdd(Sub, Offset, Sub->getName() + ".off");
4352 
4353  Value *Cmp;
4354  // If NumCases overflowed, then all possible values jump to the successor.
4355  if (NumCases->isNullValue() && !ContiguousCases->empty())
4356  Cmp = ConstantInt::getTrue(SI->getContext());
4357  else
4358  Cmp = Builder.CreateICmpULT(Sub, NumCases, "switch");
4359  BranchInst *NewBI = Builder.CreateCondBr(Cmp, ContiguousDest, OtherDest);
4360 
4361  // Update weight for the newly-created conditional branch.
4362  if (HasBranchWeights(SI)) {
4363  SmallVector<uint64_t, 8> Weights;
4364  GetBranchWeights(SI, Weights);
4365  if (Weights.size() == 1 + SI->getNumCases()) {
4366  uint64_t TrueWeight = 0;
4367  uint64_t FalseWeight = 0;
4368  for (size_t I = 0, E = Weights.size(); I != E; ++I) {
4369  if (SI->getSuccessor(I) == ContiguousDest)
4370  TrueWeight += Weights[I];
4371  else
4372  FalseWeight += Weights[I];
4373  }
4374  while (TrueWeight > UINT32_MAX || FalseWeight > UINT32_MAX) {
4375  TrueWeight /= 2;
4376  FalseWeight /= 2;
4377  }
4378  setBranchWeights(NewBI, TrueWeight, FalseWeight);
4379  }
4380  }
4381 
4382  // Prune obsolete incoming values off the successors' PHI nodes.
4383  for (auto BBI = ContiguousDest->begin(); isa<PHINode>(BBI); ++BBI) {
4384  unsigned PreviousEdges = ContiguousCases->size();
4385  if (ContiguousDest == SI->getDefaultDest())
4386  ++PreviousEdges;
4387  for (unsigned I = 0, E = PreviousEdges - 1; I != E; ++I)
4388  cast<PHINode>(BBI)->removeIncomingValue(SI->getParent());
4389  }
4390  for (auto BBI = OtherDest->begin(); isa<PHINode>(BBI); ++BBI) {
4391  unsigned PreviousEdges = SI->getNumCases() - ContiguousCases->size();
4392  if (OtherDest == SI->getDefaultDest())
4393  ++PreviousEdges;
4394  for (unsigned I = 0, E = PreviousEdges - 1; I != E; ++I)
4395  cast<PHINode>(BBI)->removeIncomingValue(SI->getParent());
4396  }
4397 
4398  // Clean up the default block - it may have phis or other instructions before
4399  // the unreachable terminator.
4400  if (!HasDefault)
4402 
4403  // Drop the switch.
4404  SI->eraseFromParent();
4405 
4406  return true;
4407 }
4408 
4409 /// Compute masked bits for the condition of a switch
4410 /// and use it to remove dead cases.
4412  const DataLayout &DL) {
4413  Value *Cond = SI->getCondition();
4414  unsigned Bits = Cond->getType()->getIntegerBitWidth();
4415  KnownBits Known = computeKnownBits(Cond, DL, 0, AC, SI);
4416 
4417  // We can also eliminate cases by determining that their values are outside of
4418  // the limited range of the condition based on how many significant (non-sign)
4419  // bits are in the condition value.
4420  unsigned ExtraSignBits = ComputeNumSignBits(Cond, DL, 0, AC, SI) - 1;
4421  unsigned MaxSignificantBitsInCond = Bits - ExtraSignBits;
4422 
4423  // Gather dead cases.
4425  for (auto &Case : SI->cases()) {
4426  const APInt &CaseVal = Case.getCaseValue()->getValue();
4427  if (Known.Zero.intersects(CaseVal) || !Known.One.isSubsetOf(CaseVal) ||
4428  (CaseVal.getMinSignedBits() > MaxSignificantBitsInCond)) {
4429  DeadCases.push_back(Case.getCaseValue());
4430  LLVM_DEBUG(dbgs() << "SimplifyCFG: switch case " << CaseVal
4431  << " is dead.\n");
4432  }
4433  }
4434 
4435  // If we can prove that the cases must cover all possible values, the
4436  // default destination becomes dead and we can remove it. If we know some
4437  // of the bits in the value, we can use that to more precisely compute the
4438  // number of possible unique case values.
4439  bool HasDefault =
4440  !isa<UnreachableInst>(SI->getDefaultDest()->getFirstNonPHIOrDbg());
4441  const unsigned NumUnknownBits =
4442  Bits - (Known.Zero | Known.One).countPopulation();
4443  assert(NumUnknownBits <= Bits);
4444  if (HasDefault && DeadCases.empty() &&
4445  NumUnknownBits < 64 /* avoid overflow */ &&
4446  SI->getNumCases() == (1ULL << NumUnknownBits)) {
4448  return true;
4449  }
4450 
4451  if (DeadCases.empty())
4452  return false;
4453 
4454  SwitchInstProfUpdateWrapper SIW(*SI);
4455  for (ConstantInt *DeadCase : DeadCases) {
4456  SwitchInst::CaseIt CaseI = SI->findCaseValue(DeadCase);
4457  assert(CaseI != SI->case_default() &&
4458  "Case was not found. Probably mistake in DeadCases forming.");
4459  // Prune unused values from PHI nodes.
4460  CaseI->getCaseSuccessor()->removePredecessor(SI->getParent());
4461  SIW.removeCase(CaseI);
4462  }
4463 
4464  return true;
4465 }
4466 
4467 /// If BB would be eligible for simplification by
4468 /// TryToSimplifyUncondBranchFromEmptyBlock (i.e. it is empty and terminated
4469 /// by an unconditional branch), look at the phi node for BB in the successor
4470 /// block and see if the incoming value is equal to CaseValue. If so, return
4471 /// the phi node, and set PhiIndex to BB's index in the phi node.
4473  BasicBlock *BB, int *PhiIndex) {
4474  if (BB->getFirstNonPHIOrDbg() != BB->getTerminator())
4475  return nullptr; // BB must be empty to be a candidate for simplification.
4476  if (!BB->getSinglePredecessor())
4477  return nullptr; // BB must be dominated by the switch.
4478 
4480  if (!Branch || !Branch->isUnconditional())
4481  return nullptr; // Terminator must be unconditional branch.
4482 
4483  BasicBlock *Succ = Branch->getSuccessor(0);
4484 
4485  for (PHINode &PHI : Succ->phis()) {
4486  int Idx = PHI.getBasicBlockIndex(BB);
4487  assert(Idx >= 0 && "PHI has no entry for predecessor?");
4488 
4489  Value *InValue = PHI.getIncomingValue(Idx);
4490  if (InValue != CaseValue)
4491  continue;
4492 
4493  *PhiIndex = Idx;
4494  return &PHI;
4495  }
4496 
4497  return nullptr;
4498 }
4499 
4500 /// Try to forward the condition of a switch instruction to a phi node
4501 /// dominated by the switch, if that would mean that some of the destination
4502 /// blocks of the switch can be folded away. Return true if a change is made.
4504  using ForwardingNodesMap = DenseMap<PHINode *, SmallVector<int, 4>>;
4505 
4506  ForwardingNodesMap ForwardingNodes;
4507  BasicBlock *SwitchBlock = SI->getParent();
4508  bool Changed = false;
4509  for (auto &Case : SI->cases()) {
4510  ConstantInt *CaseValue = Case.getCaseValue();
4511  BasicBlock *CaseDest = Case.getCaseSuccessor();
4512 
4513  // Replace phi operands in successor blocks that are using the constant case
4514  // value rather than the switch condition variable:
4515  // switchbb:
4516  // switch i32 %x, label %default [
4517  // i32 17, label %succ
4518  // ...
4519  // succ:
4520  // %r = phi i32 ... [ 17, %switchbb ] ...
4521  // -->
4522  // %r = phi i32 ... [ %x, %switchbb ] ...
4523 
4524  for (PHINode &Phi : CaseDest->phis()) {
4525  // This only works if there is exactly 1 incoming edge from the switch to
4526  // a phi. If there is >1, that means multiple cases of the switch map to 1
4527  // value in the phi, and that phi value is not the switch condition. Thus,
4528  // this transform would not make sense (the phi would be invalid because
4529  // a phi can't have different incoming values from the same block).
4530  int SwitchBBIdx = Phi.getBasicBlockIndex(SwitchBlock);
4531  if (Phi.getIncomingValue(SwitchBBIdx) == CaseValue &&
4532  count(Phi.blocks(), SwitchBlock) == 1) {
4533  Phi.setIncomingValue(SwitchBBIdx, SI->getCondition());
4534  Changed = true;
4535  }
4536  }
4537 
4538  // Collect phi nodes that are indirectly using this switch's case constants.
4539  int PhiIdx;
4540  if (auto *Phi = FindPHIForConditionForwarding(CaseValue, CaseDest, &PhiIdx))
4541  ForwardingNodes[Phi].push_back(PhiIdx);
4542  }
4543 
4544  for (auto &ForwardingNode : ForwardingNodes) {
4545  PHINode *Phi = ForwardingNode.first;
4546  SmallVectorImpl<int> &Indexes = ForwardingNode.second;
4547  if (Indexes.size() < 2)
4548  continue;
4549 
4550  for (int Index : Indexes)
4551  Phi->setIncomingValue(Index, SI->getCondition());
4552  Changed = true;
4553  }
4554 
4555  return Changed;
4556 }
4557 
4558 /// Return true if the backend will be able to handle
4559 /// initializing an array of constants like C.
4561  if (C->isThreadDependent())
4562  return false;
4563  if (C->isDLLImportDependent())
4564  return false;
4565 
4566  if (!isa<ConstantFP>(C) && !isa<ConstantInt>(C) &&
4567  !isa<ConstantPointerNull>(C) && !isa<GlobalValue>(C) &&
4568  !isa<UndefValue>(C) && !isa<ConstantExpr>(C))
4569  return false;
4570 
4571  if (ConstantExpr *CE = dyn_cast<ConstantExpr>(C)) {
4572  if (!CE->isGEPWithNoNotionalOverIndexing())
4573  return false;
4574  if (!ValidLookupTableConstant(CE->getOperand(0), TTI))
4575  return false;
4576  }
4577 
4579  return false;
4580 
4581  return true;
4582 }
4583 
4584 /// If V is a Constant, return it. Otherwise, try to look up
4585 /// its constant value in ConstantPool, returning 0 if it's not there.
4586 static Constant *
4589  if (Constant *C = dyn_cast<Constant>(V))
4590  return C;
4591  return ConstantPool.lookup(V);
4592 }
4593 
4594 /// Try to fold instruction I into a constant. This works for
4595 /// simple instructions such as binary operations where both operands are
4596 /// constant or can be replaced by constants from the ConstantPool. Returns the
4597 /// resulting constant on success, 0 otherwise.
4598 static Constant *
4601  if (SelectInst *Select = dyn_cast<SelectInst>(I)) {
4602  Constant *A = LookupConstant(Select->getCondition(), ConstantPool);
4603  if (!A)
4604  return nullptr;
4605  if (A->isAllOnesValue())
4606  return LookupConstant(Select->getTrueValue(), ConstantPool);
4607  if (A->isNullValue())
4608  return LookupConstant(Select->getFalseValue(), ConstantPool);
4609  return nullptr;
4610  }
4611 
4613  for (unsigned N = 0, E = I->getNumOperands(); N != E; ++N) {
4615  COps.push_back(A);
4616  else
4617  return nullptr;
4618  }
4619 
4620  if (CmpInst *Cmp = dyn_cast<CmpInst>(I)) {
4621  return ConstantFoldCompareInstOperands(Cmp->getPredicate(), COps[0],
4622  COps[1], DL);
4623  }
4624 
4625  return ConstantFoldInstOperands(I, COps, DL);
4626 }
4627 
4628 /// Try to determine the resulting constant values in phi nodes
4629 /// at the common destination basic block, *CommonDest, for one of the case
4630 /// destionations CaseDest corresponding to value CaseVal (0 for the default
4631 /// case), of a switch instruction SI.
4632 static bool
4634  BasicBlock **CommonDest,
4635  SmallVectorImpl<std::pair<PHINode *, Constant *>> &Res,
4636  const DataLayout &DL, const TargetTransformInfo &TTI) {
4637  // The block from which we enter the common destination.
4638  BasicBlock *Pred = SI->getParent();
4639 
4640  // If CaseDest is empty except for some side-effect free instructions through
4641  // which we can constant-propagate the CaseVal, continue to its successor.
4643  ConstantPool.insert(std::make_pair(SI->getCondition(), CaseVal));
4644  for (Instruction &I :CaseDest->instructionsWithoutDebug()) {
4645  if (I.isTerminator()) {
4646  // If the terminator is a simple branch, continue to the next block.
4647  if (I.getNumSuccessors() != 1 || I.isExceptionalTerminator())
4648  return false;
4649  Pred = CaseDest;
4650  CaseDest = I.getSuccessor(0);
4651  } else if (Constant *C = ConstantFold(&I, DL, ConstantPool)) {
4652  // Instruction is side-effect free and constant.
4653 
4654  // If the instruction has uses outside this block or a phi node slot for
4655  // the block, it is not safe to bypass the instruction since it would then
4656  // no longer dominate all its uses.
4657  for (auto &Use : I.uses()) {
4658  User *User = Use.getUser();
4659  if (Instruction *I = dyn_cast<Instruction>(User))
4660  if (I->getParent() == CaseDest)
4661  continue;
4662  if (PHINode *Phi = dyn_cast<PHINode>(User))
4663  if (Phi->getIncomingBlock(Use) == CaseDest)
4664  continue;
4665  return false;
4666  }
4667 
4668  ConstantPool.insert(std::make_pair(&I, C));
4669  } else {
4670  break;
4671  }
4672  }
4673 
4674  // If we did not have a CommonDest before, use the current one.
4675  if (!*CommonDest)
4676  *CommonDest = CaseDest;
4677  // If the destination isn't the common one, abort.
4678  if (CaseDest != *CommonDest)
4679  return false;
4680 
4681  // Get the values for this case from phi nodes in the destination block.
4682  for (PHINode &PHI : (*CommonDest)->phis()) {
4683  int Idx = PHI.getBasicBlockIndex(Pred);
4684  if (Idx == -1)
4685  continue;
4686 
4687  Constant *ConstVal =
4688  LookupConstant(PHI.getIncomingValue(Idx), ConstantPool);
4689  if (!ConstVal)
4690  return false;
4691 
4692  // Be conservative about which kinds of constants we support.
4693  if (!ValidLookupTableConstant(ConstVal, TTI))
4694  return false;
4695 
4696  Res.push_back(std::make_pair(&PHI, ConstVal));
4697  }
4698 
4699  return Res.size() > 0;
4700 }
4701 
4702 // Helper function used to add CaseVal to the list of cases that generate
4703 // Result. Returns the updated number of cases that generate this result.
4704 static uintptr_t MapCaseToResult(ConstantInt *CaseVal,
4705  SwitchCaseResultVectorTy &UniqueResults,
4706  Constant *Result) {
4707  for (auto &I : UniqueResults) {
4708  if (I.first == Result) {
4709  I.second.push_back(CaseVal);
4710  return I.second.size();
4711  }
4712  }
4713  UniqueResults.push_back(
4714  std::make_pair(Result, SmallVector<ConstantInt *, 4>(1, CaseVal)));
4715  return 1;
4716 }
4717 
4718 // Helper function that initializes a map containing
4719 // results for the PHI node of the common destination block for a switch
4720 // instruction. Returns false if multiple PHI nodes have been found or if
4721 // there is not a common destination block for the switch.
4722 static bool
4724  SwitchCaseResultVectorTy &UniqueResults,
4725  Constant *&DefaultResult, const DataLayout &DL,
4726  const TargetTransformInfo &TTI,
4727  uintptr_t MaxUniqueResults, uintptr_t MaxCasesPerResult) {
4728  for (auto &I : SI->cases()) {
4729  ConstantInt *CaseVal = I.getCaseValue();
4730 
4731  // Resulting value at phi nodes for this case value.
4732  SwitchCaseResultsTy Results;
4733  if (!GetCaseResults(SI, CaseVal, I.getCaseSuccessor(), &CommonDest, Results,
4734  DL, TTI))
4735  return false;
4736 
4737  // Only one value per case is permitted.
4738  if (Results.size() > 1)
4739  return false;
4740 
4741  // Add the case->result mapping to UniqueResults.
4742  const uintptr_t NumCasesForResult =
4743  MapCaseToResult(CaseVal, UniqueResults, Results.begin()->second);
4744 
4745  // Early out if there are too many cases for this result.
4746  if (NumCasesForResult > MaxCasesPerResult)
4747  return false;
4748 
4749  // Early out if there are too many unique results.
4750  if (UniqueResults.size() > MaxUniqueResults)
4751  return false;
4752 
4753  // Check the PHI consistency.
4754  if (!PHI)
4755  PHI = Results[0].first;
4756  else if (PHI != Results[0].first)
4757  return false;
4758  }
4759  // Find the default result value.
4761  BasicBlock *DefaultDest = SI->getDefaultDest();
4762  GetCaseResults(SI, nullptr, SI->getDefaultDest(), &CommonDest, DefaultResults,
4763  DL, TTI);
4764  // If the default value is not found abort unless the default destination
4765  // is unreachable.
4766  DefaultResult =
4767  DefaultResults.size() == 1 ? DefaultResults.begin()->second : nullptr;
4768  if ((!DefaultResult &&
4769  !isa<UnreachableInst>(DefaultDest->getFirstNonPHIOrDbg())))
4770  return false;
4771 
4772  return true;
4773 }
4774 
4775 // Helper function that checks if it is possible to transform a switch with only
4776 // two cases (or two cases + default) that produces a result into a select.
4777 // Example:
4778 // switch (a) {
4779 // case 10: %0 = icmp eq i32 %a, 10
4780 // return 10; %1 = select i1 %0, i32 10, i32 4
4781 // case 20: ----> %2 = icmp eq i32 %a, 20
4782 // return 2; %3 = select i1 %2, i32 2, i32 %1
4783 // default:
4784 // return 4;
4785 // }
4786 static Value *ConvertTwoCaseSwitch(const SwitchCaseResultVectorTy &ResultVector,
4787  Constant *DefaultResult, Value *Condition,
4788  IRBuilder<> &Builder) {
4789  assert(ResultVector.size() == 2 &&
4790  "We should have exactly two unique results at this point");
4791  // If we are selecting between only two cases transform into a simple
4792  // select or a two-way select if default is possible.
4793  if (ResultVector[0].second.size() == 1 &&
4794  ResultVector[1].second.size() == 1) {
4795  ConstantInt *const FirstCase = ResultVector[0].second[0];
4796  ConstantInt *const SecondCase = ResultVector[1].second[0];
4797 
4798  bool DefaultCanTrigger = DefaultResult;
4799  Value *SelectValue = ResultVector[1].first;
4800  if (DefaultCanTrigger) {
4801  Value *const ValueCompare =
4802  Builder.CreateICmpEQ(Condition, SecondCase, "switch.selectcmp");
4803  SelectValue = Builder.CreateSelect(ValueCompare, ResultVector[1].first,
4804  DefaultResult, "switch.select");
4805  }
4806  Value *const ValueCompare =
4807  Builder.CreateICmpEQ(Condition, FirstCase, "switch.selectcmp");
4808  return Builder.CreateSelect(ValueCompare, ResultVector[0].first,
4809  SelectValue, "switch.select");
4810  }
4811 
4812  return nullptr;
4813 }
4814 
4815 // Helper function to cleanup a switch instruction that has been converted into
4816 // a select, fixing up PHI nodes and basic blocks.
4818  Value *SelectValue,
4819  IRBuilder<> &Builder) {
4820  BasicBlock *SelectBB = SI->getParent();
4821  while (PHI->getBasicBlockIndex(SelectBB) >= 0)
4822  PHI->removeIncomingValue(SelectBB);
4823  PHI->addIncoming(SelectValue, SelectBB);
4824 
4825  Builder.CreateBr(PHI->getParent());
4826 
4827  // Remove the switch.
4828  for (unsigned i = 0, e = SI->getNumSuccessors(); i < e; ++i) {
4829  BasicBlock *Succ = SI->getSuccessor(i);
4830 
4831  if (Succ == PHI->getParent())
4832  continue;
4833  Succ->removePredecessor(SelectBB);
4834  }
4835  SI->eraseFromParent();
4836 }
4837 
4838 /// If the switch is only used to initialize one or more
4839 /// phi nodes in a common successor block with only two different
4840 /// constant values, replace the switch with select.
4841 static bool switchToSelect(SwitchInst *SI, IRBuilder<> &Builder,
4842  const DataLayout &DL,
4843  const TargetTransformInfo &TTI) {
4844  Value *const Cond = SI->getCondition();
4845  PHINode *PHI = nullptr;
4846  BasicBlock *CommonDest = nullptr;
4847  Constant *DefaultResult;
4848  SwitchCaseResultVectorTy UniqueResults;
4849  // Collect all the cases that will deliver the same value from the switch.
4850  if (!InitializeUniqueCases(SI, PHI, CommonDest, UniqueResults, DefaultResult,
4851  DL, TTI, 2, 1))
4852  return false;
4853  // Selects choose between maximum two values.
4854  if (UniqueResults.size() != 2)
4855  return false;
4856  assert(PHI != nullptr && "PHI for value select not found");
4857 
4858  Builder.SetInsertPoint(SI);
4859  Value *SelectValue =
4860  ConvertTwoCaseSwitch(UniqueResults, DefaultResult, Cond, Builder);
4861  if (SelectValue) {
4862  RemoveSwitchAfterSelectConversion(SI, PHI, SelectValue, Builder);
4863  return true;
4864  }
4865  // The switch couldn't be converted into a select.
4866  return false;
4867 }
4868 
4869 namespace {
4870 
4871 /// This class represents a lookup table that can be used to replace a switch.
4872 class SwitchLookupTable {
4873 public:
4874  /// Create a lookup table to use as a switch replacement with the contents
4875  /// of Values, using DefaultValue to fill any holes in the table.
4876  SwitchLookupTable(
4877  Module &M, uint64_t TableSize, ConstantInt *Offset,
4878  const SmallVectorImpl<std::pair<ConstantInt *, Constant *>> &Values,
4879  Constant *DefaultValue, const DataLayout &DL, const StringRef &FuncName);
4880 
4881  /// Build instructions with Builder to retrieve the value at
4882  /// the position given by Index in the lookup table.
4883  Value *BuildLookup(Value *Index, IRBuilder<> &Builder);
4884 
4885  /// Return true if a table with TableSize elements of
4886  /// type ElementType would fit in a target-legal register.
4887  static bool WouldFitInRegister(const DataLayout &DL, uint64_t TableSize,
4888  Type *ElementType);
4889 
4890 private:
4891  // Depending on the contents of the table, it can be represented in
4892  // different ways.
4893  enum {
4894  // For tables where each element contains the same value, we just have to
4895  // store that single value and return it for each lookup.
4896  SingleValueKind,
4897 
4898  // For tables where there is a linear relationship between table index
4899  // and values. We calculate the result with a simple multiplication
4900  // and addition instead of a table lookup.
4901  LinearMapKind,
4902 
4903  // For small tables with integer elements, we can pack them into a bitmap
4904  // that fits into a target-legal register. Values are retrieved by
4905  // shift and mask operations.
4906  BitMapKind,
4907 
4908  // The table is stored as an array of values. Values are retrieved by load
4909  // instructions from the table.
4910  ArrayKind
4911  } Kind;
4912 
4913  // For SingleValueKind, this is the single value.
4914  Constant *SingleValue = nullptr;
4915 
4916  // For BitMapKind, this is the bitmap.
4917  ConstantInt *BitMap = nullptr;
4918  IntegerType *BitMapElementTy = nullptr;
4919 
4920  // For LinearMapKind, these are the constants used to derive the value.
4921  ConstantInt *LinearOffset = nullptr;
4922  ConstantInt *LinearMultiplier = nullptr;
4923 
4924  // For ArrayKind, this is the array.
4925  GlobalVariable *Array = nullptr;
4926 };
4927 
4928 } // end anonymous namespace
4929 
4930 SwitchLookupTable::SwitchLookupTable(
4931  Module &M, uint64_t TableSize, ConstantInt *Offset,
4932  const SmallVectorImpl<std::pair<ConstantInt *, Constant *>> &Values,
4933  Constant *DefaultValue, const DataLayout &DL, const StringRef &FuncName) {
4934  assert(Values.size() && "Can't build lookup table without values!");
4935  assert(TableSize >= Values.size() && "Can't fit values in table!");
4936 
4937  // If all values in the table are equal, this is that value.
4938  SingleValue = Values.begin()->second;
4939 
4940  Type *ValueType = Values.begin()->second->getType();
4941 
4942  // Build up the table contents.
4943  SmallVector<Constant *, 64> TableContents(TableSize);
4944  for (size_t I = 0, E = Values.size(); I != E; ++I) {
4945  ConstantInt *CaseVal = Values[I].first;
4946  Constant *CaseRes = Values[I].second;
4947  assert(CaseRes->getType() == ValueType);
4948 
4949  uint64_t Idx = (CaseVal->getValue() - Offset->getValue()).getLimitedValue();
4950  TableContents[Idx] = CaseRes;
4951 
4952  if (CaseRes != SingleValue)
4953  SingleValue = nullptr;
4954  }
4955 
4956  // Fill in any holes in the table with the default result.
4957  if (Values.size() < TableSize) {
4958  assert(DefaultValue &&
4959  "Need a default value to fill the lookup table holes.");
4960  assert(DefaultValue->getType() == ValueType);
4961  for (uint64_t I = 0; I < TableSize; ++I) {
4962  if (!TableContents[I])
4963  TableContents[I] = DefaultValue;
4964  }
4965 
4966  if (DefaultValue != SingleValue)
4967  SingleValue = nullptr;
4968  }
4969 
4970  // If each element in the table contains the same value, we only need to store
4971  // that single value.
4972  if (SingleValue) {
4973  Kind = SingleValueKind;
4974  return;
4975  }
4976 
4977  // Check if we can derive the value with a linear transformation from the
4978  // table index.
4979  if (isa<IntegerType>(ValueType)) {
4980  bool LinearMappingPossible = true;
4981  APInt PrevVal;
4982  APInt DistToPrev;
4983  assert(TableSize >= 2 && "Should be a SingleValue table.");
4984  // Check if there is the same distance between two consecutive values.
4985  for (uint64_t I = 0; I < TableSize; ++I) {
4986  ConstantInt *ConstVal = dyn_cast<ConstantInt>(TableContents[I]);
4987  if (!ConstVal) {
4988  // This is an undef. We could deal with it, but undefs in lookup tables
4989  // are very seldom. It's probably not worth the additional complexity.
4990  LinearMappingPossible = false;
4991  break;
4992  }
4993  const APInt &Val = ConstVal->getValue();
4994  if (I != 0) {
4995  APInt Dist = Val - PrevVal;
4996  if (I == 1) {
4997  DistToPrev = Dist;
4998  } else if (Dist != DistToPrev) {
4999  LinearMappingPossible = false;
5000  break;
5001  }
5002  }
5003  PrevVal = Val;
5004  }
5005  if (LinearMappingPossible) {
5006  LinearOffset = cast<ConstantInt>(TableContents[0]);
5007  LinearMultiplier = ConstantInt::get(M.getContext(), DistToPrev);
5008  Kind = LinearMapKind;
5009  ++NumLinearMaps;
5010  return;
5011  }
5012  }
5013 
5014  // If the type is integer and the table fits in a register, build a bitmap.
5015  if (WouldFitInRegister(DL, TableSize, ValueType)) {
5016  IntegerType *IT = cast<IntegerType>(ValueType);
5017  APInt TableInt(TableSize * IT->getBitWidth(), 0);
5018  for (uint64_t I = TableSize; I > 0; --I) {
5019  TableInt <<= IT->getBitWidth();
5020  // Insert values into the bitmap. Undef values are set to zero.
5021  if (!isa<UndefValue>(TableContents[I - 1])) {
5022  ConstantInt *Val = cast<ConstantInt>(TableContents[I - 1]);
5023  TableInt |= Val->getValue().zext(TableInt.getBitWidth());
5024  }
5025  }
5026  BitMap = ConstantInt::get(M.getContext(), TableInt);
5027  BitMapElementTy = IT;
5028  Kind = BitMapKind;
5029  ++NumBitMaps;
5030  return;
5031  }
5032 
5033  // Store the table in an array.
5034  ArrayType *ArrayTy = ArrayType::get(ValueType, TableSize);
5035  Constant *Initializer = ConstantArray::get(ArrayTy, TableContents);
5036 
5037  Array = new GlobalVariable(M, ArrayTy, /*isConstant=*/true,
5038  GlobalVariable::PrivateLinkage, Initializer,
5039  "switch.table." + FuncName);
5040  Array->setUnnamedAddr(GlobalValue::UnnamedAddr::Global);
5041  // Set the alignment to that of an array items. We will be only loading one
5042  // value out of it.
5043  Array->setAlignment(DL.getPrefTypeAlignment(ValueType));
5044  Kind = ArrayKind;
5045 }
5046 
5047 Value *SwitchLookupTable::BuildLookup(Value *Index, IRBuilder<> &Builder) {
5048  switch (Kind) {
5049  case SingleValueKind:
5050  return SingleValue;
5051  case LinearMapKind: {
5052  // Derive the result value from the input value.
5053  Value *Result = Builder.CreateIntCast(Index, LinearMultiplier->getType(),
5054  false, "switch.idx.cast");
5055  if (!LinearMultiplier->isOne())
5056  Result = Builder.CreateMul(Result, LinearMultiplier, "switch.idx.mult");
5057  if (!LinearOffset->isZero())
5058  Result = Builder.CreateAdd(Result, LinearOffset, "switch.offset");
5059  return Result;
5060  }
5061  case BitMapKind: {
5062  // Type of the bitmap (e.g. i59).
5063  IntegerType *MapTy = BitMap->getType();
5064 
5065  // Cast Index to the same type as the bitmap.
5066  // Note: The Index is <= the number of elements in the table, so
5067  // truncating it to the width of the bitmask is safe.
5068  Value *ShiftAmt = Builder.CreateZExtOrTrunc(Index, MapTy, "switch.cast");
5069 
5070  // Multiply the shift amount by the element width.
5071  ShiftAmt = Builder.CreateMul(
5072  ShiftAmt, ConstantInt::get(MapTy, BitMapElementTy->getBitWidth()),
5073  "switch.shiftamt");
5074 
5075  // Shift down.
5076  Value *DownShifted =
5077  Builder.CreateLShr(BitMap, ShiftAmt, "switch.downshift");
5078  // Mask off.
5079  return Builder.CreateTrunc(DownShifted, BitMapElementTy, "switch.masked");
5080  }
5081  case ArrayKind: {
5082  // Make sure the table index will not overflow when treated as signed.
5083  IntegerType *IT = cast<IntegerType>(Index->getType());
5084  uint64_t TableSize =
5085  Array->getInitializer()->getType()->getArrayNumElements();
5086  if (TableSize > (1ULL << (IT->getBitWidth() - 1)))
5087  Index = Builder.CreateZExt(
5088  Index, IntegerType::get(IT->getContext(), IT->getBitWidth() + 1),
5089  "switch.tableidx.zext");
5090 
5091  Value *GEPIndices[] = {Builder.getInt32(0), Index};
5092  Value *GEP = Builder.CreateInBoundsGEP(Array->getValueType(), Array,
5093  GEPIndices, "switch.gep");
5094  return Builder.CreateLoad(
5095  cast<ArrayType>(Array->getValueType())->getElementType(), GEP,
5096  "switch.load");
5097  }
5098  }
5099  llvm_unreachable("Unknown lookup table kind!");
5100 }
5101 
5102 bool SwitchLookupTable::WouldFitInRegister(const DataLayout &DL,
5103  uint64_t TableSize,
5104  Type *ElementType) {
5105  auto *IT = dyn_cast<IntegerType>(ElementType);
5106  if (!IT)
5107  return false;
5108  // FIXME: If the type is wider than it needs to be, e.g. i8 but all values
5109  // are <= 15, we could try to narrow the type.
5110 
5111  // Avoid overflow, fitsInLegalInteger uses unsigned int for the width.
5112  if (TableSize >= UINT_MAX / IT->getBitWidth())
5113  return false;
5114  return DL.fitsInLegalInteger(TableSize * IT->getBitWidth());
5115 }
5116 
5117 /// Determine whether a lookup table should be built for this switch, based on
5118 /// the number of cases, size of the table, and the types of the results.
5119 static bool
5120 ShouldBuildLookupTable(SwitchInst *SI, uint64_t TableSize,
5121  const TargetTransformInfo &TTI, const DataLayout &DL,
5122  const SmallDenseMap<PHINode *, Type *> &ResultTypes) {
5123  if (SI->getNumCases() > TableSize || TableSize >= UINT64_MAX / 10)
5124  return false; // TableSize overflowed, or mul below might overflow.
5125 
5126  bool AllTablesFitInRegister = true;
5127  bool HasIllegalType = false;
5128  for (const auto &I : ResultTypes) {
5129  Type *Ty = I.second;
5130 
5131  // Saturate this flag to true.
5132  HasIllegalType = HasIllegalType || !TTI.isTypeLegal(Ty);
5133 
5134  // Saturate this flag to false.
5135  AllTablesFitInRegister =
5136  AllTablesFitInRegister &&
5137  SwitchLookupTable::WouldFitInRegister(DL, TableSize, Ty);
5138 
5139  // If both flags saturate, we're done. NOTE: This *only* works with
5140  // saturating flags, and all flags have to saturate first due to the
5141  // non-deterministic behavior of iterating over a dense map.
5142  if (HasIllegalType && !AllTablesFitInRegister)
5143  break;
5144  }
5145 
5146  // If each table would fit in a register, we should build it anyway.
5147  if (AllTablesFitInRegister)
5148  return true;
5149 
5150  // Don't build a table that doesn't fit in-register if it has illegal types.
5151  if (HasIllegalType)
5152  return false;
5153 
5154  // The table density should be at least 40%. This is the same criterion as for
5155  // jump tables, see SelectionDAGBuilder::handleJTSwitchCase.
5156  // FIXME: Find the best cut-off.
5157  return SI->getNumCases() * 10 >= TableSize * 4;
5158 }
5159 
5160 /// Try to reuse the switch table index compare. Following pattern:
5161 /// \code
5162 /// if (idx < tablesize)
5163 /// r = table[idx]; // table does not contain default_value
5164 /// else
5165 /// r = default_value;
5166 /// if (r != default_value)
5167 /// ...
5168 /// \endcode
5169 /// Is optimized to:
5170 /// \code
5171 /// cond = idx < tablesize;
5172 /// if (cond)
5173 /// r = table[idx];
5174 /// else
5175 /// r = default_value;
5176 /// if (cond)
5177 /// ...
5178 /// \endcode
5179 /// Jump threading will then eliminate the second if(cond).
5180 static void reuseTableCompare(
5181  User *PhiUser, BasicBlock *PhiBlock, BranchInst *RangeCheckBranch,
5182  Constant *DefaultValue,
5183  const SmallVectorImpl<std::pair<ConstantInt *, Constant *>> &Values) {
5184  ICmpInst *CmpInst = dyn_cast<ICmpInst>(PhiUser);
5185  if (!CmpInst)
5186  return;
5187 
5188  // We require that the compare is in the same block as the phi so that jump
5189  // threading can do its work afterwards.
5190  if (CmpInst->getParent() != PhiBlock)
5191  return;
5192 
5193  Constant *CmpOp1 = dyn_cast<Constant>(CmpInst->getOperand(1));
5194  if (!CmpOp1)
5195  return;
5196 
5197  Value *RangeCmp = RangeCheckBranch->getCondition();
5198  Constant *TrueConst = ConstantInt::getTrue(RangeCmp->getType());
5199  Constant *FalseConst = ConstantInt::getFalse(RangeCmp->getType());
5200 
5201  // Check if the compare with the default value is constant true or false.
5202  Constant *DefaultConst = ConstantExpr::getICmp(CmpInst->getPredicate(),
5203  DefaultValue, CmpOp1, true);
5204  if (DefaultConst != TrueConst && DefaultConst != FalseConst)
5205  return;
5206 
5207  // Check if the compare with the case values is distinct from the default
5208  // compare result.
5209  for (auto ValuePair : Values) {
5210  Constant *CaseConst = ConstantExpr::getICmp(CmpInst->getPredicate(),
5211  ValuePair.second, CmpOp1, true);
5212  if (!CaseConst || CaseConst == DefaultConst || isa<UndefValue>(CaseConst))
5213  return;
5214  assert((CaseConst == TrueConst || CaseConst == FalseConst) &&
5215  "Expect true or false as compare result.");
5216  }
5217 
5218  // Check if the branch instruction dominates the phi node. It's a simple
5219  // dominance check, but sufficient for our needs.
5220  // Although this check is invariant in the calling loops, it's better to do it
5221  // at this late stage. Practically we do it at most once for a switch.
5222  BasicBlock *BranchBlock = RangeCheckBranch->getParent();
5223  for (auto PI = pred_begin(PhiBlock), E = pred_end(PhiBlock); PI != E; ++PI) {
5224  BasicBlock *Pred = *PI;
5225  if (Pred != BranchBlock && Pred->getUniquePredecessor() != BranchBlock)
5226  return;
5227  }
5228 
5229  if (DefaultConst == FalseConst) {
5230  // The compare yields the same result. We can replace it.
5231  CmpInst->replaceAllUsesWith(RangeCmp);
5232  ++NumTableCmpReuses;
5233  } else {
5234  // The compare yields the same result, just inverted. We can replace it.
5235  Value *InvertedTableCmp = BinaryOperator::CreateXor(
5236  RangeCmp, ConstantInt::get(RangeCmp->getType(), 1), "inverted.cmp",
5237  RangeCheckBranch);
5238  CmpInst->replaceAllUsesWith(InvertedTableCmp);
5239  ++NumTableCmpReuses;
5240  }
5241 }
5242 
5243 /// If the switch is only used to initialize one or more phi nodes in a common
5244 /// successor block with different constant values, replace the switch with
5245 /// lookup tables.
5247  const DataLayout &DL,
5248  const TargetTransformInfo &TTI) {
5249  assert(SI->getNumCases() > 1 && "Degenerate switch?");
5250 
5251  Function *Fn = SI->getParent()->getParent();
5252  // Only build lookup table when we have a target that supports it or the
5253  // attribute is not set.
5254  if (!TTI.shouldBuildLookupTables() ||
5255  (Fn->getFnAttribute("no-jump-tables").getValueAsString() == "true"))
5256  return false;
5257 
5258  // FIXME: If the switch is too sparse for a lookup table, perhaps we could
5259  // split off a dense part and build a lookup table for that.
5260 
5261  // FIXME: This creates arrays of GEPs to constant strings, which means each
5262  // GEP needs a runtime relocation in PIC code. We should just build one big
5263  // string and lookup indices into that.
5264 
5265  // Ignore switches with less than three cases. Lookup tables will not make
5266  // them faster, so we don't analyze them.
5267  if (SI->getNumCases() < 3)
5268  return false;
5269 
5270  // Figure out the corresponding result for each case value and phi node in the
5271  // common destination, as well as the min and max case values.
5272  assert(!empty(SI->cases()));
5273  SwitchInst::CaseIt CI = SI->case_begin();
5274  ConstantInt *MinCaseVal = CI->getCaseValue();
5275  ConstantInt *MaxCaseVal = CI->getCaseValue();
5276 
5277  BasicBlock *CommonDest = nullptr;
5278 
5279  using ResultListTy = SmallVector<std::pair<ConstantInt *, Constant *>, 4>;
5281 
5282  SmallDenseMap<PHINode *, Constant *> DefaultResults;
5285 
5286  for (SwitchInst::CaseIt E = SI->case_end(); CI != E; ++CI) {
5287  ConstantInt *CaseVal = CI->getCaseValue();
5288  if (CaseVal->getValue().slt(MinCaseVal->getValue()))
5289  MinCaseVal = CaseVal;
5290  if (CaseVal->getValue().sgt(MaxCaseVal->getValue()))
5291  MaxCaseVal = CaseVal;
5292 
5293  // Resulting value at phi nodes for this case value.
5294  using ResultsTy = SmallVector<std::pair<PHINode *, Constant *>, 4>;
5295  ResultsTy Results;
5296  if (!GetCaseResults(SI, CaseVal, CI->getCaseSuccessor(), &CommonDest,
5297  Results, DL, TTI))
5298  return false;
5299 
5300  // Append the result from this case to the list for each phi.
5301  for (const auto &I : Results) {
5302  PHINode *PHI = I.first;
5303  Constant *Value = I.second;
5304  if (!ResultLists.count(PHI))
5305  PHIs.push_back(PHI);
5306  ResultLists[PHI].push_back(std::make_pair(CaseVal, Value));
5307  }
5308  }
5309 
5310  // Keep track of the result types.
5311  for (PHINode *PHI : PHIs) {
5312  ResultTypes[PHI] = ResultLists[PHI][0].second->getType();
5313  }
5314 
5315  uint64_t NumResults = ResultLists[PHIs[0]].size();
5316  APInt RangeSpread = MaxCaseVal->getValue() - MinCaseVal->getValue();
5317  uint64_t TableSize = RangeSpread.getLimitedValue() + 1;
5318  bool TableHasHoles = (NumResults < TableSize);
5319 
5320  // If the table has holes, we need a constant result for the default case
5321  // or a bitmask that fits in a register.
5322  SmallVector<std::pair<PHINode *, Constant *>, 4> DefaultResultsList;
5323  bool HasDefaultResults =
5324  GetCaseResults(SI, nullptr, SI->getDefaultDest(), &CommonDest,
5325  DefaultResultsList, DL, TTI);
5326 
5327  bool NeedMask = (TableHasHoles && !HasDefaultResults);
5328  if (NeedMask) {
5329  // As an extra penalty for the validity test we require more cases.
5330  if (SI->getNumCases() < 4) // FIXME: Find best threshold value (benchmark).
5331  return false;
5332  if (!DL.fitsInLegalInteger(TableSize))
5333  return false;
5334  }
5335 
5336  for (const auto &I : DefaultResultsList) {
5337  PHINode *PHI = I.first;
5338  Constant *Result = I.second;
5339  DefaultResults[PHI] = Result;
5340  }
5341 
5342  if (!ShouldBuildLookupTable(SI, TableSize, TTI, DL, ResultTypes))
5343  return false;
5344 
5345  // Create the BB that does the lookups.
5346  Module &Mod = *CommonDest->getParent()->getParent();
5347  BasicBlock *LookupBB = BasicBlock::Create(
5348  Mod.getContext(), "switch.lookup", CommonDest->getParent(), CommonDest);
5349 
5350  // Compute the table index value.
5351  Builder.SetInsertPoint(SI);
5352  Value *TableIndex;
5353  if (MinCaseVal->isNullValue())
5354  TableIndex = SI->getCondition();
5355  else
5356  TableIndex = Builder.CreateSub(SI->getCondition(), MinCaseVal,
5357  "switch.tableidx");
5358 
5359  // Compute the maximum table size representable by the integer type we are
5360  // switching upon.
5361  unsigned CaseSize = MinCaseVal->getType()->getPrimitiveSizeInBits();
5362  uint64_t MaxTableSize = CaseSize > 63 ? UINT64_MAX : 1ULL << CaseSize;
5363  assert(MaxTableSize >= TableSize &&
5364  "It is impossible for a switch to have more entries than the max "
5365  "representable value of its input integer type's size.");
5366 
5367  // If the default destination is unreachable, or if the lookup table covers
5368  // all values of the conditional variable, branch directly to the lookup table
5369  // BB. Otherwise, check that the condition is within the case range.
5370  const bool DefaultIsReachable =
5371  !isa<UnreachableInst>(SI->getDefaultDest()->getFirstNonPHIOrDbg());
5372  const bool GeneratingCoveredLookupTable = (MaxTableSize == TableSize);
5373  BranchInst *RangeCheckBranch = nullptr;
5374 
5375  if (!DefaultIsReachable || GeneratingCoveredLookupTable) {
5376  Builder.CreateBr(LookupBB);
5377  // Note: We call removeProdecessor later since we need to be able to get the
5378  // PHI value for the default case in case we're using a bit mask.
5379  } else {
5380  Value *Cmp = Builder.CreateICmpULT(
5381  TableIndex, ConstantInt::get(MinCaseVal->getType(), TableSize));
5382  RangeCheckBranch =
5383  Builder.CreateCondBr(Cmp, LookupBB, SI->getDefaultDest());
5384  }
5385 
5386  // Populate the BB that does the lookups.
5387  Builder.SetInsertPoint(LookupBB);
5388 
5389  if (NeedMask) {
5390  // Before doing the lookup, we do the hole check. The LookupBB is therefore
5391  // re-purposed to do the hole check, and we create a new LookupBB.
5392  BasicBlock *MaskBB = LookupBB;
5393  MaskBB->setName("switch.hole_check");
5394  LookupBB = BasicBlock::Create(Mod.getContext(), "switch.lookup",
5395  CommonDest->getParent(), CommonDest);
5396 
5397  // Make the mask's bitwidth at least 8-bit and a power-of-2 to avoid
5398  // unnecessary illegal types.
5399  uint64_t TableSizePowOf2 = NextPowerOf2(std::max(7ULL, TableSize - 1ULL));
5400  APInt MaskInt(TableSizePowOf2, 0);
5401  APInt One(TableSizePowOf2, 1);
5402  // Build bitmask; fill in a 1 bit for every case.
5403  const ResultListTy &ResultList = ResultLists[PHIs[0]];
5404  for (size_t I = 0, E = ResultList.size(); I != E; ++I) {
5405  uint64_t Idx = (ResultList[I].first->getValue() - MinCaseVal->getValue())
5406  .getLimitedValue();
5407  MaskInt |= One << Idx;
5408  }
5409  ConstantInt *TableMask = ConstantInt::get(Mod.getContext(), MaskInt);
5410 
5411  // Get the TableIndex'th bit of the bitmask.
5412  // If this bit is 0 (meaning hole) jump to the default destination,
5413  // else continue with table lookup.
5414  IntegerType *MapTy = TableMask->getType();
5415  Value *MaskIndex =
5416  Builder.CreateZExtOrTrunc(TableIndex, MapTy, "switch.maskindex");
5417  Value *Shifted = Builder.CreateLShr(TableMask, MaskIndex, "switch.shifted");
5418  Value *LoBit = Builder.CreateTrunc(
5419  Shifted, Type::getInt1Ty(Mod.getContext()), "switch.lobit");
5420  Builder.CreateCondBr(LoBit, LookupBB, SI->getDefaultDest());
5421 
5422  Builder.SetInsertPoint(LookupBB);
5423  AddPredecessorToBlock(SI->getDefaultDest(), MaskBB, SI->getParent());
5424  }
5425 
5426  if (!DefaultIsReachable || GeneratingCoveredLookupTable) {
5427  // We cached PHINodes in PHIs. To avoid accessing deleted PHINodes later,
5428  // do not delete PHINodes here.
5430  /*KeepOneInputPHIs=*/true);
5431  }
5432 
5433  bool ReturnedEarly = false;
5434  for (PHINode *PHI : PHIs) {
5435  const ResultListTy &ResultList = ResultLists[PHI];
5436 
5437  // If using a bitmask, use any value to fill the lookup table holes.
5438  Constant *DV = NeedMask ? ResultLists[PHI][0].second : DefaultResults[PHI];
5439  StringRef FuncName = Fn->getName();
5440  SwitchLookupTable Table(Mod, TableSize, MinCaseVal, ResultList, DV, DL,
5441  FuncName);
5442 
5443  Value *Result = Table.BuildLookup(TableIndex, Builder);
5444 
5445  // If the result is used to return immediately from the function, we want to
5446  // do that right here.
5447  if (PHI->hasOneUse() && isa<ReturnInst>(*PHI->user_begin()) &&
5448  PHI->user_back() == CommonDest->getFirstNonPHIOrDbg()) {
5449  Builder.CreateRet(Result);
5450  ReturnedEarly = true;
5451  break;
5452  }
5453 
5454  // Do a small peephole optimization: re-use the switch table compare if
5455  // possible.
5456  if (!TableHasHoles && HasDefaultResults && RangeCheckBranch) {
5457  BasicBlock *PhiBlock = PHI->getParent();
5458  // Search for compare instructions which use the phi.
5459  for (auto *User : PHI->users()) {
5460  reuseTableCompare(User, PhiBlock, RangeCheckBranch, DV, ResultList);
5461  }
5462  }
5463 
5464  PHI->addIncoming(Result, LookupBB);
5465  }
5466 
5467  if (!ReturnedEarly)
5468  Builder.CreateBr(CommonDest);
5469 
5470  // Remove the switch.
5471  for (unsigned i = 0, e = SI->getNumSuccessors(); i < e; ++i) {
5472  BasicBlock *Succ = SI->getSuccessor(i);
5473 
5474  if (Succ == SI->getDefaultDest())
5475  continue;
5476  Succ->removePredecessor(SI->getParent());
5477  }
5478  SI->eraseFromParent();
5479 
5480  ++NumLookupTables;
5481  if (NeedMask)
5482  ++NumLookupTablesHoles;
5483  return true;
5484 }
5485 
5486 static bool isSwitchDense(ArrayRef<int64_t> Values) {
5487  // See also SelectionDAGBuilder::isDense(), which this function was based on.
5488  uint64_t Diff = (uint64_t)Values.back() - (uint64_t)Values.front();
5489  uint64_t Range = Diff + 1;
5490  uint64_t NumCases = Values.size();
5491  // 40% is the default density for building a jump table in optsize/minsize mode.
5492  uint64_t MinDensity = 40;
5493 
5494  return NumCases * 100 >= Range * MinDensity;
5495 }
5496 
5497 /// Try to transform a switch that has "holes" in it to a contiguous sequence
5498 /// of cases.
5499 ///
5500 /// A switch such as: switch(i) {case 5: case 9: case 13: case 17:} can be
5501 /// range-reduced to: switch ((i-5) / 4) {case 0: case 1: case 2: case 3:}.
5502 ///
5503 /// This converts a sparse switch into a dense switch which allows better
5504 /// lowering and could also allow transforming into a lookup table.
5506  const DataLayout &DL,
5507  const TargetTransformInfo &TTI) {
5508  auto *CondTy = cast<IntegerType>(SI->getCondition()->getType());
5509  if (CondTy->getIntegerBitWidth() > 64 ||
5510  !DL.fitsInLegalInteger(CondTy->getIntegerBitWidth()))
5511  return false;
5512  // Only bother with this optimization if there are more than 3 switch cases;
5513  // SDAG will only bother creating jump tables for 4 or more cases.
5514  if (SI->getNumCases() < 4)
5515  return false;
5516 
5517  // This transform is agnostic to the signedness of the input or case values. We
5518  // can treat the case values as signed or unsigned. We can optimize more common
5519  // cases such as a sequence crossing zero {-4,0,4,8} if we interpret case values
5520  // as signed.
5521  SmallVector<int64_t,4> Values;
5522  for (auto &C : SI->cases())
5523  Values.push_back(C.getCaseValue()->getValue().getSExtValue());
5524  llvm::sort(Values);
5525 
5526  // If the switch is already dense, there's nothing useful to do here.
5527  if (isSwitchDense(Values))
5528  return false;
5529 
5530  // First, transform the values such that they start at zero and ascend.
5531  int64_t Base = Values[0];
5532  for (auto &V : Values)
5533  V -= (uint64_t)(Base);
5534 
5535  // Now we have signed numbers that have been shifted so that, given enough
5536  // precision, there are no negative values. Since the rest of the transform
5537  // is bitwise only, we switch now to an unsigned representation.
5538 
5539  // This transform can be done speculatively because it is so cheap - it
5540  // results in a single rotate operation being inserted.
5541  // FIXME: It's possible that optimizing a switch on powers of two might also
5542  // be beneficial - flag values are often powers of two and we could use a CLZ
5543  // as the key function.
5544 
5545  // countTrailingZeros(0) returns 64. As Values is guaranteed to have more than
5546  // one element and LLVM disallows duplicate cases, Shift is guaranteed to be
5547  // less than 64.
5548  unsigned Shift = 64;
5549  for (auto &V : Values)
5550  Shift = std::min(Shift, countTrailingZeros((uint64_t)V));
5551  assert(Shift < 64);
5552  if (Shift > 0)
5553  for (auto &V : Values)
5554  V = (int64_t)((uint64_t)V >> Shift);
5555 
5556  if (!isSwitchDense(Values))
5557  // Transform didn't create a dense switch.
5558  return false;
5559 
5560  // The obvious transform is to shift the switch condition right and emit a
5561  // check that the condition actually cleanly divided by GCD, i.e.
5562  // C & (1 << Shift - 1) == 0
5563  // inserting a new CFG edge to handle the case where it didn't divide cleanly.
5564  //
5565  // A cheaper way of doing this is a simple ROTR(C, Shift). This performs the
5566  // shift and puts the shifted-off bits in the uppermost bits. If any of these
5567  // are nonzero then the switch condition will be very large and will hit the
5568  // default case.
5569 
5570  auto *Ty = cast<IntegerType>(SI->getCondition()->getType());
5571  Builder.SetInsertPoint(SI);
5572  auto *ShiftC = ConstantInt::get(Ty, Shift);
5573  auto *Sub = Builder.CreateSub(SI->getCondition(), ConstantInt::get(Ty, Base));
5574  auto *LShr = Builder.CreateLShr(Sub, ShiftC);
5575  auto *Shl = Builder.CreateShl(Sub, Ty->getBitWidth() - Shift);
5576  auto *Rot = Builder.CreateOr(LShr, Shl);
5577  SI->replaceUsesOfWith(SI->getCondition(), Rot);
5578 
5579  for (auto Case : SI->cases()) {
5580  auto *Orig = Case.getCaseValue();
5581  auto Sub = Orig->getValue() - APInt(Ty->getBitWidth(), Base);
5582  Case.setValue(
5583  cast<ConstantInt>(ConstantInt::get(Ty, Sub.lshr(ShiftC->getValue()))));
5584  }
5585  return true;
5586 }
5587 
5588 bool SimplifyCFGOpt::SimplifySwitch(SwitchInst *SI, IRBuilder<> &Builder) {
5589  BasicBlock *BB = SI->getParent();
5590 
5591  if (isValueEqualityComparison(SI)) {
5592  // If we only have one predecessor, and if it is a branch on this value,
5593  // see if that predecessor totally determines the outcome of this switch.
5594  if (BasicBlock *OnlyPred = BB->getSinglePredecessor())
5595  if (SimplifyEqualityComparisonWithOnlyPredecessor(SI, OnlyPred, Builder))
5596  return requestResimplify();
5597 
5598  Value *Cond = SI->getCondition();
5599  if (SelectInst *Select = dyn_cast<SelectInst>(Cond))
5600  if (SimplifySwitchOnSelect(SI, Select))
5601  return requestResimplify();
5602 
5603  // If the block only contains the switch, see if we can fold the block
5604  // away into any preds.
5605  if (SI == &*BB->instructionsWithoutDebug().begin())
5606  if (FoldValueComparisonI