LLVM  7.0.0svn
SimplifyCFG.cpp
Go to the documentation of this file.
1 //===- SimplifyCFG.cpp - Code to perform CFG simplification ---------------===//
2 //
3 // The LLVM Compiler Infrastructure
4 //
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
7 //
8 //===----------------------------------------------------------------------===//
9 //
10 // Peephole optimize the CFG.
11 //
12 //===----------------------------------------------------------------------===//
13 
14 #include "llvm/ADT/APInt.h"
15 #include "llvm/ADT/ArrayRef.h"
16 #include "llvm/ADT/DenseMap.h"
17 #include "llvm/ADT/Optional.h"
18 #include "llvm/ADT/STLExtras.h"
19 #include "llvm/ADT/SetOperations.h"
20 #include "llvm/ADT/SetVector.h"
21 #include "llvm/ADT/SmallPtrSet.h"
22 #include "llvm/ADT/SmallSet.h"
23 #include "llvm/ADT/SmallVector.h"
24 #include "llvm/ADT/Statistic.h"
25 #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 
178  Value *isValueEqualityComparison(TerminatorInst *TI);
179  BasicBlock *GetValueEqualityComparisonCases(
180  TerminatorInst *TI, std::vector<ValueEqualityComparisonCase> &Cases);
181  bool SimplifyEqualityComparisonWithOnlyPredecessor(TerminatorInst *TI,
182  BasicBlock *Pred,
183  IRBuilder<> &Builder);
184  bool FoldValueComparisonIntoPredecessors(TerminatorInst *TI,
185  IRBuilder<> &Builder);
186 
187  bool SimplifyReturn(ReturnInst *RI, IRBuilder<> &Builder);
188  bool SimplifyResume(ResumeInst *RI, IRBuilder<> &Builder);
189  bool SimplifySingleResume(ResumeInst *RI);
190  bool SimplifyCommonResume(ResumeInst *RI);
191  bool SimplifyCleanupReturn(CleanupReturnInst *RI);
192  bool SimplifyUnreachable(UnreachableInst *UI);
193  bool SimplifySwitch(SwitchInst *SI, IRBuilder<> &Builder);
194  bool SimplifyIndirectBr(IndirectBrInst *IBI);
195  bool SimplifyUncondBranch(BranchInst *BI, IRBuilder<> &Builder);
196  bool SimplifyCondBranch(BranchInst *BI, IRBuilder<> &Builder);
197 
198 public:
199  SimplifyCFGOpt(const TargetTransformInfo &TTI, const DataLayout &DL,
200  SmallPtrSetImpl<BasicBlock *> *LoopHeaders,
201  const SimplifyCFGOptions &Opts)
202  : TTI(TTI), DL(DL), LoopHeaders(LoopHeaders), Options(Opts) {}
203 
204  bool run(BasicBlock *BB);
205 };
206 
207 } // end anonymous namespace
208 
209 /// Return true if it is safe to merge these two
210 /// terminator instructions together.
211 static bool
213  SmallSetVector<BasicBlock *, 4> *FailBlocks = nullptr) {
214  if (SI1 == SI2)
215  return false; // Can't merge with self!
216 
217  // It is not safe to merge these two switch instructions if they have a common
218  // successor, and if that successor has a PHI node, and if *that* PHI node has
219  // conflicting incoming values from the two switch blocks.
220  BasicBlock *SI1BB = SI1->getParent();
221  BasicBlock *SI2BB = SI2->getParent();
222 
223  SmallPtrSet<BasicBlock *, 16> SI1Succs(succ_begin(SI1BB), succ_end(SI1BB));
224  bool Fail = false;
225  for (BasicBlock *Succ : successors(SI2BB))
226  if (SI1Succs.count(Succ))
227  for (BasicBlock::iterator BBI = Succ->begin(); isa<PHINode>(BBI); ++BBI) {
228  PHINode *PN = cast<PHINode>(BBI);
229  if (PN->getIncomingValueForBlock(SI1BB) !=
230  PN->getIncomingValueForBlock(SI2BB)) {
231  if (FailBlocks)
232  FailBlocks->insert(Succ);
233  Fail = true;
234  }
235  }
236 
237  return !Fail;
238 }
239 
240 /// Return true if it is safe and profitable to merge these two terminator
241 /// instructions together, where SI1 is an unconditional branch. PhiNodes will
242 /// store all PHI nodes in common successors.
243 static bool
245  Instruction *Cond,
246  SmallVectorImpl<PHINode *> &PhiNodes) {
247  if (SI1 == SI2)
248  return false; // Can't merge with self!
249  assert(SI1->isUnconditional() && SI2->isConditional());
250 
251  // We fold the unconditional branch if we can easily update all PHI nodes in
252  // common successors:
253  // 1> We have a constant incoming value for the conditional branch;
254  // 2> We have "Cond" as the incoming value for the unconditional branch;
255  // 3> SI2->getCondition() and Cond have same operands.
256  CmpInst *Ci2 = dyn_cast<CmpInst>(SI2->getCondition());
257  if (!Ci2)
258  return false;
259  if (!(Cond->getOperand(0) == Ci2->getOperand(0) &&
260  Cond->getOperand(1) == Ci2->getOperand(1)) &&
261  !(Cond->getOperand(0) == Ci2->getOperand(1) &&
262  Cond->getOperand(1) == Ci2->getOperand(0)))
263  return false;
264 
265  BasicBlock *SI1BB = SI1->getParent();
266  BasicBlock *SI2BB = SI2->getParent();
267  SmallPtrSet<BasicBlock *, 16> SI1Succs(succ_begin(SI1BB), succ_end(SI1BB));
268  for (BasicBlock *Succ : successors(SI2BB))
269  if (SI1Succs.count(Succ))
270  for (BasicBlock::iterator BBI = Succ->begin(); isa<PHINode>(BBI); ++BBI) {
271  PHINode *PN = cast<PHINode>(BBI);
272  if (PN->getIncomingValueForBlock(SI1BB) != Cond ||
273  !isa<ConstantInt>(PN->getIncomingValueForBlock(SI2BB)))
274  return false;
275  PhiNodes.push_back(PN);
276  }
277  return true;
278 }
279 
280 /// Update PHI nodes in Succ to indicate that there will now be entries in it
281 /// from the 'NewPred' block. The values that will be flowing into the PHI nodes
282 /// will be the same as those coming in from ExistPred, an existing predecessor
283 /// of Succ.
284 static void AddPredecessorToBlock(BasicBlock *Succ, BasicBlock *NewPred,
285  BasicBlock *ExistPred) {
286  for (PHINode &PN : Succ->phis())
287  PN.addIncoming(PN.getIncomingValueForBlock(ExistPred), NewPred);
288 }
289 
290 /// Compute an abstract "cost" of speculating the given instruction,
291 /// which is assumed to be safe to speculate. TCC_Free means cheap,
292 /// TCC_Basic means less cheap, and TCC_Expensive means prohibitively
293 /// expensive.
294 static unsigned ComputeSpeculationCost(const User *I,
295  const TargetTransformInfo &TTI) {
297  "Instruction is not safe to speculatively execute!");
298  return TTI.getUserCost(I);
299 }
300 
301 /// If we have a merge point of an "if condition" as accepted above,
302 /// return true if the specified value dominates the block. We
303 /// don't handle the true generality of domination here, just a special case
304 /// which works well enough for us.
305 ///
306 /// If AggressiveInsts is non-null, and if V does not dominate BB, we check to
307 /// see if V (which must be an instruction) and its recursive operands
308 /// that do not dominate BB have a combined cost lower than CostRemaining and
309 /// are non-trapping. If both are true, the instruction is inserted into the
310 /// set and true is returned.
311 ///
312 /// The cost for most non-trapping instructions is defined as 1 except for
313 /// Select whose cost is 2.
314 ///
315 /// After this function returns, CostRemaining is decreased by the cost of
316 /// V plus its non-dominating operands. If that cost is greater than
317 /// CostRemaining, false is returned and CostRemaining is undefined.
319  SmallPtrSetImpl<Instruction *> *AggressiveInsts,
320  unsigned &CostRemaining,
321  const TargetTransformInfo &TTI,
322  unsigned Depth = 0) {
323  // It is possible to hit a zero-cost cycle (phi/gep instructions for example),
324  // so limit the recursion depth.
325  // TODO: While this recursion limit does prevent pathological behavior, it
326  // would be better to track visited instructions to avoid cycles.
327  if (Depth == MaxSpeculationDepth)
328  return false;
329 
331  if (!I) {
332  // Non-instructions all dominate instructions, but not all constantexprs
333  // can be executed unconditionally.
334  if (ConstantExpr *C = dyn_cast<ConstantExpr>(V))
335  if (C->canTrap())
336  return false;
337  return true;
338  }
339  BasicBlock *PBB = I->getParent();
340 
341  // We don't want to allow weird loops that might have the "if condition" in
342  // the bottom of this block.
343  if (PBB == BB)
344  return false;
345 
346  // If this instruction is defined in a block that contains an unconditional
347  // branch to BB, then it must be in the 'conditional' part of the "if
348  // statement". If not, it definitely dominates the region.
350  if (!BI || BI->isConditional() || BI->getSuccessor(0) != BB)
351  return true;
352 
353  // If we aren't allowing aggressive promotion anymore, then don't consider
354  // instructions in the 'if region'.
355  if (!AggressiveInsts)
356  return false;
357 
358  // If we have seen this instruction before, don't count it again.
359  if (AggressiveInsts->count(I))
360  return true;
361 
362  // Okay, it looks like the instruction IS in the "condition". Check to
363  // see if it's a cheap instruction to unconditionally compute, and if it
364  // only uses stuff defined outside of the condition. If so, hoist it out.
366  return false;
367 
368  unsigned Cost = ComputeSpeculationCost(I, TTI);
369 
370  // Allow exactly one instruction to be speculated regardless of its cost
371  // (as long as it is safe to do so).
372  // This is intended to flatten the CFG even if the instruction is a division
373  // or other expensive operation. The speculation of an expensive instruction
374  // is expected to be undone in CodeGenPrepare if the speculation has not
375  // enabled further IR optimizations.
376  if (Cost > CostRemaining &&
377  (!SpeculateOneExpensiveInst || !AggressiveInsts->empty() || Depth > 0))
378  return false;
379 
380  // Avoid unsigned wrap.
381  CostRemaining = (Cost > CostRemaining) ? 0 : CostRemaining - Cost;
382 
383  // Okay, we can only really hoist these out if their operands do
384  // not take us over the cost threshold.
385  for (User::op_iterator i = I->op_begin(), e = I->op_end(); i != e; ++i)
386  if (!DominatesMergePoint(*i, BB, AggressiveInsts, CostRemaining, TTI,
387  Depth + 1))
388  return false;
389  // Okay, it's safe to do this! Remember this instruction.
390  AggressiveInsts->insert(I);
391  return true;
392 }
393 
394 /// Extract ConstantInt from value, looking through IntToPtr
395 /// and PointerNullValue. Return NULL if value is not a constant int.
396 static ConstantInt *GetConstantInt(Value *V, const DataLayout &DL) {
397  // Normal constant int.
398  ConstantInt *CI = dyn_cast<ConstantInt>(V);
399  if (CI || !isa<Constant>(V) || !V->getType()->isPointerTy())
400  return CI;
401 
402  // This is some kind of pointer constant. Turn it into a pointer-sized
403  // ConstantInt if possible.
404  IntegerType *PtrTy = cast<IntegerType>(DL.getIntPtrType(V->getType()));
405 
406  // Null pointer means 0, see SelectionDAGBuilder::getValue(const Value*).
407  if (isa<ConstantPointerNull>(V))
408  return ConstantInt::get(PtrTy, 0);
409 
410  // IntToPtr const int.
411  if (ConstantExpr *CE = dyn_cast<ConstantExpr>(V))
412  if (CE->getOpcode() == Instruction::IntToPtr)
413  if (ConstantInt *CI = dyn_cast<ConstantInt>(CE->getOperand(0))) {
414  // The constant is very likely to have the right type already.
415  if (CI->getType() == PtrTy)
416  return CI;
417  else
418  return cast<ConstantInt>(
419  ConstantExpr::getIntegerCast(CI, PtrTy, /*isSigned=*/false));
420  }
421  return nullptr;
422 }
423 
424 namespace {
425 
426 /// Given a chain of or (||) or and (&&) comparison of a value against a
427 /// constant, this will try to recover the information required for a switch
428 /// structure.
429 /// It will depth-first traverse the chain of comparison, seeking for patterns
430 /// like %a == 12 or %a < 4 and combine them to produce a set of integer
431 /// representing the different cases for the switch.
432 /// Note that if the chain is composed of '||' it will build the set of elements
433 /// that matches the comparisons (i.e. any of this value validate the chain)
434 /// while for a chain of '&&' it will build the set elements that make the test
435 /// fail.
436 struct ConstantComparesGatherer {
437  const DataLayout &DL;
438 
439  /// Value found for the switch comparison
440  Value *CompValue = nullptr;
441 
442  /// Extra clause to be checked before the switch
443  Value *Extra = nullptr;
444 
445  /// Set of integers to match in switch
447 
448  /// Number of comparisons matched in the and/or chain
449  unsigned UsedICmps = 0;
450 
451  /// Construct and compute the result for the comparison instruction Cond
452  ConstantComparesGatherer(Instruction *Cond, const DataLayout &DL) : DL(DL) {
453  gather(Cond);
454  }
455 
456  ConstantComparesGatherer(const ConstantComparesGatherer &) = delete;
457  ConstantComparesGatherer &
458  operator=(const ConstantComparesGatherer &) = delete;
459 
460 private:
461  /// Try to set the current value used for the comparison, it succeeds only if
462  /// it wasn't set before or if the new value is the same as the old one
463  bool setValueOnce(Value *NewVal) {
464  if (CompValue && CompValue != NewVal)
465  return false;
466  CompValue = NewVal;
467  return (CompValue != nullptr);
468  }
469 
470  /// Try to match Instruction "I" as a comparison against a constant and
471  /// populates the array Vals with the set of values that match (or do not
472  /// match depending on isEQ).
473  /// Return false on failure. On success, the Value the comparison matched
474  /// against is placed in CompValue.
475  /// If CompValue is already set, the function is expected to fail if a match
476  /// is found but the value compared to is different.
477  bool matchInstruction(Instruction *I, bool isEQ) {
478  // If this is an icmp against a constant, handle this as one of the cases.
479  ICmpInst *ICI;
480  ConstantInt *C;
481  if (!((ICI = dyn_cast<ICmpInst>(I)) &&
482  (C = GetConstantInt(I->getOperand(1), DL)))) {
483  return false;
484  }
485 
486  Value *RHSVal;
487  const APInt *RHSC;
488 
489  // Pattern match a special case
490  // (x & ~2^z) == y --> x == y || x == y|2^z
491  // This undoes a transformation done by instcombine to fuse 2 compares.
492  if (ICI->getPredicate() == (isEQ ? ICmpInst::ICMP_EQ : ICmpInst::ICMP_NE)) {
493  // It's a little bit hard to see why the following transformations are
494  // correct. Here is a CVC3 program to verify them for 64-bit values:
495 
496  /*
497  ONE : BITVECTOR(64) = BVZEROEXTEND(0bin1, 63);
498  x : BITVECTOR(64);
499  y : BITVECTOR(64);
500  z : BITVECTOR(64);
501  mask : BITVECTOR(64) = BVSHL(ONE, z);
502  QUERY( (y & ~mask = y) =>
503  ((x & ~mask = y) <=> (x = y OR x = (y | mask)))
504  );
505  QUERY( (y | mask = y) =>
506  ((x | mask = y) <=> (x = y OR x = (y & ~mask)))
507  );
508  */
509 
510  // Please note that each pattern must be a dual implication (<--> or
511  // iff). One directional implication can create spurious matches. If the
512  // implication is only one-way, an unsatisfiable condition on the left
513  // side can imply a satisfiable condition on the right side. Dual
514  // implication ensures that satisfiable conditions are transformed to
515  // other satisfiable conditions and unsatisfiable conditions are
516  // transformed to other unsatisfiable conditions.
517 
518  // Here is a concrete example of a unsatisfiable condition on the left
519  // implying a satisfiable condition on the right:
520  //
521  // mask = (1 << z)
522  // (x & ~mask) == y --> (x == y || x == (y | mask))
523  //
524  // Substituting y = 3, z = 0 yields:
525  // (x & -2) == 3 --> (x == 3 || x == 2)
526 
527  // Pattern match a special case:
528  /*
529  QUERY( (y & ~mask = y) =>
530  ((x & ~mask = y) <=> (x = y OR x = (y | mask)))
531  );
532  */
533  if (match(ICI->getOperand(0),
534  m_And(m_Value(RHSVal), m_APInt(RHSC)))) {
535  APInt Mask = ~*RHSC;
536  if (Mask.isPowerOf2() && (C->getValue() & ~Mask) == C->getValue()) {
537  // If we already have a value for the switch, it has to match!
538  if (!setValueOnce(RHSVal))
539  return false;
540 
541  Vals.push_back(C);
542  Vals.push_back(
544  C->getValue() | Mask));
545  UsedICmps++;
546  return true;
547  }
548  }
549 
550  // Pattern match a special case:
551  /*
552  QUERY( (y | mask = y) =>
553  ((x | mask = y) <=> (x = y OR x = (y & ~mask)))
554  );
555  */
556  if (match(ICI->getOperand(0),
557  m_Or(m_Value(RHSVal), m_APInt(RHSC)))) {
558  APInt Mask = *RHSC;
559  if (Mask.isPowerOf2() && (C->getValue() | Mask) == C->getValue()) {
560  // If we already have a value for the switch, it has to match!
561  if (!setValueOnce(RHSVal))
562  return false;
563 
564  Vals.push_back(C);
566  C->getValue() & ~Mask));
567  UsedICmps++;
568  return true;
569  }
570  }
571 
572  // If we already have a value for the switch, it has to match!
573  if (!setValueOnce(ICI->getOperand(0)))
574  return false;
575 
576  UsedICmps++;
577  Vals.push_back(C);
578  return ICI->getOperand(0);
579  }
580 
581  // If we have "x ult 3", for example, then we can add 0,1,2 to the set.
583  ICI->getPredicate(), C->getValue());
584 
585  // Shift the range if the compare is fed by an add. This is the range
586  // compare idiom as emitted by instcombine.
587  Value *CandidateVal = I->getOperand(0);
588  if (match(I->getOperand(0), m_Add(m_Value(RHSVal), m_APInt(RHSC)))) {
589  Span = Span.subtract(*RHSC);
590  CandidateVal = RHSVal;
591  }
592 
593  // If this is an and/!= check, then we are looking to build the set of
594  // value that *don't* pass the and chain. I.e. to turn "x ugt 2" into
595  // x != 0 && x != 1.
596  if (!isEQ)
597  Span = Span.inverse();
598 
599  // If there are a ton of values, we don't want to make a ginormous switch.
600  if (Span.isSizeLargerThan(8) || Span.isEmptySet()) {
601  return false;
602  }
603 
604  // If we already have a value for the switch, it has to match!
605  if (!setValueOnce(CandidateVal))
606  return false;
607 
608  // Add all values from the range to the set
609  for (APInt Tmp = Span.getLower(); Tmp != Span.getUpper(); ++Tmp)
610  Vals.push_back(ConstantInt::get(I->getContext(), Tmp));
611 
612  UsedICmps++;
613  return true;
614  }
615 
616  /// Given a potentially 'or'd or 'and'd together collection of icmp
617  /// eq/ne/lt/gt instructions that compare a value against a constant, extract
618  /// the value being compared, and stick the list constants into the Vals
619  /// vector.
620  /// One "Extra" case is allowed to differ from the other.
621  void gather(Value *V) {
623  bool isEQ = (I->getOpcode() == Instruction::Or);
624 
625  // Keep a stack (SmallVector for efficiency) for depth-first traversal
627  SmallPtrSet<Value *, 8> Visited;
628 
629  // Initialize
630  Visited.insert(V);
631  DFT.push_back(V);
632 
633  while (!DFT.empty()) {
634  V = DFT.pop_back_val();
635 
636  if (Instruction *I = dyn_cast<Instruction>(V)) {
637  // If it is a || (or && depending on isEQ), process the operands.
638  if (I->getOpcode() == (isEQ ? Instruction::Or : Instruction::And)) {
639  if (Visited.insert(I->getOperand(1)).second)
640  DFT.push_back(I->getOperand(1));
641  if (Visited.insert(I->getOperand(0)).second)
642  DFT.push_back(I->getOperand(0));
643  continue;
644  }
645 
646  // Try to match the current instruction
647  if (matchInstruction(I, isEQ))
648  // Match succeed, continue the loop
649  continue;
650  }
651 
652  // One element of the sequence of || (or &&) could not be match as a
653  // comparison against the same value as the others.
654  // We allow only one "Extra" case to be checked before the switch
655  if (!Extra) {
656  Extra = V;
657  continue;
658  }
659  // Failed to parse a proper sequence, abort now
660  CompValue = nullptr;
661  break;
662  }
663  }
664 };
665 
666 } // end anonymous namespace
667 
669  Instruction *Cond = nullptr;
670  if (SwitchInst *SI = dyn_cast<SwitchInst>(TI)) {
671  Cond = dyn_cast<Instruction>(SI->getCondition());
672  } else if (BranchInst *BI = dyn_cast<BranchInst>(TI)) {
673  if (BI->isConditional())
674  Cond = dyn_cast<Instruction>(BI->getCondition());
675  } else if (IndirectBrInst *IBI = dyn_cast<IndirectBrInst>(TI)) {
676  Cond = dyn_cast<Instruction>(IBI->getAddress());
677  }
678 
679  TI->eraseFromParent();
680  if (Cond)
682 }
683 
684 /// Return true if the specified terminator checks
685 /// to see if a value is equal to constant integer value.
686 Value *SimplifyCFGOpt::isValueEqualityComparison(TerminatorInst *TI) {
687  Value *CV = nullptr;
688  if (SwitchInst *SI = dyn_cast<SwitchInst>(TI)) {
689  // Do not permit merging of large switch instructions into their
690  // predecessors unless there is only one predecessor.
691  if (SI->getNumSuccessors() * std::distance(pred_begin(SI->getParent()),
692  pred_end(SI->getParent())) <=
693  128)
694  CV = SI->getCondition();
695  } else if (BranchInst *BI = dyn_cast<BranchInst>(TI))
696  if (BI->isConditional() && BI->getCondition()->hasOneUse())
697  if (ICmpInst *ICI = dyn_cast<ICmpInst>(BI->getCondition())) {
698  if (ICI->isEquality() && GetConstantInt(ICI->getOperand(1), DL))
699  CV = ICI->getOperand(0);
700  }
701 
702  // Unwrap any lossless ptrtoint cast.
703  if (CV) {
704  if (PtrToIntInst *PTII = dyn_cast<PtrToIntInst>(CV)) {
705  Value *Ptr = PTII->getPointerOperand();
706  if (PTII->getType() == DL.getIntPtrType(Ptr->getType()))
707  CV = Ptr;
708  }
709  }
710  return CV;
711 }
712 
713 /// Given a value comparison instruction,
714 /// decode all of the 'cases' that it represents and return the 'default' block.
715 BasicBlock *SimplifyCFGOpt::GetValueEqualityComparisonCases(
716  TerminatorInst *TI, std::vector<ValueEqualityComparisonCase> &Cases) {
717  if (SwitchInst *SI = dyn_cast<SwitchInst>(TI)) {
718  Cases.reserve(SI->getNumCases());
719  for (auto Case : SI->cases())
720  Cases.push_back(ValueEqualityComparisonCase(Case.getCaseValue(),
721  Case.getCaseSuccessor()));
722  return SI->getDefaultDest();
723  }
724 
725  BranchInst *BI = cast<BranchInst>(TI);
726  ICmpInst *ICI = cast<ICmpInst>(BI->getCondition());
727  BasicBlock *Succ = BI->getSuccessor(ICI->getPredicate() == ICmpInst::ICMP_NE);
728  Cases.push_back(ValueEqualityComparisonCase(
729  GetConstantInt(ICI->getOperand(1), DL), Succ));
730  return BI->getSuccessor(ICI->getPredicate() == ICmpInst::ICMP_EQ);
731 }
732 
733 /// Given a vector of bb/value pairs, remove any entries
734 /// in the list that match the specified block.
735 static void
737  std::vector<ValueEqualityComparisonCase> &Cases) {
738  Cases.erase(std::remove(Cases.begin(), Cases.end(), BB), Cases.end());
739 }
740 
741 /// Return true if there are any keys in C1 that exist in C2 as well.
742 static bool ValuesOverlap(std::vector<ValueEqualityComparisonCase> &C1,
743  std::vector<ValueEqualityComparisonCase> &C2) {
744  std::vector<ValueEqualityComparisonCase> *V1 = &C1, *V2 = &C2;
745 
746  // Make V1 be smaller than V2.
747  if (V1->size() > V2->size())
748  std::swap(V1, V2);
749 
750  if (V1->empty())
751  return false;
752  if (V1->size() == 1) {
753  // Just scan V2.
754  ConstantInt *TheVal = (*V1)[0].Value;
755  for (unsigned i = 0, e = V2->size(); i != e; ++i)
756  if (TheVal == (*V2)[i].Value)
757  return true;
758  }
759 
760  // Otherwise, just sort both lists and compare element by element.
761  array_pod_sort(V1->begin(), V1->end());
762  array_pod_sort(V2->begin(), V2->end());
763  unsigned i1 = 0, i2 = 0, e1 = V1->size(), e2 = V2->size();
764  while (i1 != e1 && i2 != e2) {
765  if ((*V1)[i1].Value == (*V2)[i2].Value)
766  return true;
767  if ((*V1)[i1].Value < (*V2)[i2].Value)
768  ++i1;
769  else
770  ++i2;
771  }
772  return false;
773 }
774 
775 // Set branch weights on SwitchInst. This sets the metadata if there is at
776 // least one non-zero weight.
778  // Check that there is at least one non-zero weight. Otherwise, pass
779  // nullptr to setMetadata which will erase the existing metadata.
780  MDNode *N = nullptr;
781  if (llvm::any_of(Weights, [](uint32_t W) { return W != 0; }))
782  N = MDBuilder(SI->getParent()->getContext()).createBranchWeights(Weights);
784 }
785 
786 // Similar to the above, but for branch and select instructions that take
787 // exactly 2 weights.
788 static void setBranchWeights(Instruction *I, uint32_t TrueWeight,
789  uint32_t FalseWeight) {
790  assert(isa<BranchInst>(I) || isa<SelectInst>(I));
791  // Check that there is at least one non-zero weight. Otherwise, pass
792  // nullptr to setMetadata which will erase the existing metadata.
793  MDNode *N = nullptr;
794  if (TrueWeight || FalseWeight)
795  N = MDBuilder(I->getParent()->getContext())
796  .createBranchWeights(TrueWeight, FalseWeight);
798 }
799 
800 /// If TI is known to be a terminator instruction and its block is known to
801 /// only have a single predecessor block, check to see if that predecessor is
802 /// also a value comparison with the same value, and if that comparison
803 /// determines the outcome of this comparison. If so, simplify TI. This does a
804 /// very limited form of jump threading.
805 bool SimplifyCFGOpt::SimplifyEqualityComparisonWithOnlyPredecessor(
806  TerminatorInst *TI, BasicBlock *Pred, IRBuilder<> &Builder) {
807  Value *PredVal = isValueEqualityComparison(Pred->getTerminator());
808  if (!PredVal)
809  return false; // Not a value comparison in predecessor.
810 
811  Value *ThisVal = isValueEqualityComparison(TI);
812  assert(ThisVal && "This isn't a value comparison!!");
813  if (ThisVal != PredVal)
814  return false; // Different predicates.
815 
816  // TODO: Preserve branch weight metadata, similarly to how
817  // FoldValueComparisonIntoPredecessors preserves it.
818 
819  // Find out information about when control will move from Pred to TI's block.
820  std::vector<ValueEqualityComparisonCase> PredCases;
821  BasicBlock *PredDef =
822  GetValueEqualityComparisonCases(Pred->getTerminator(), PredCases);
823  EliminateBlockCases(PredDef, PredCases); // Remove default from cases.
824 
825  // Find information about how control leaves this block.
826  std::vector<ValueEqualityComparisonCase> ThisCases;
827  BasicBlock *ThisDef = GetValueEqualityComparisonCases(TI, ThisCases);
828  EliminateBlockCases(ThisDef, ThisCases); // Remove default from cases.
829 
830  // If TI's block is the default block from Pred's comparison, potentially
831  // simplify TI based on this knowledge.
832  if (PredDef == TI->getParent()) {
833  // If we are here, we know that the value is none of those cases listed in
834  // PredCases. If there are any cases in ThisCases that are in PredCases, we
835  // can simplify TI.
836  if (!ValuesOverlap(PredCases, ThisCases))
837  return false;
838 
839  if (isa<BranchInst>(TI)) {
840  // Okay, one of the successors of this condbr is dead. Convert it to a
841  // uncond br.
842  assert(ThisCases.size() == 1 && "Branch can only have one case!");
843  // Insert the new branch.
844  Instruction *NI = Builder.CreateBr(ThisDef);
845  (void)NI;
846 
847  // Remove PHI node entries for the dead edge.
848  ThisCases[0].Dest->removePredecessor(TI->getParent());
849 
850  DEBUG(dbgs() << "Threading pred instr: " << *Pred->getTerminator()
851  << "Through successor TI: " << *TI << "Leaving: " << *NI
852  << "\n");
853 
855  return true;
856  }
857 
858  SwitchInst *SI = cast<SwitchInst>(TI);
859  // Okay, TI has cases that are statically dead, prune them away.
860  SmallPtrSet<Constant *, 16> DeadCases;
861  for (unsigned i = 0, e = PredCases.size(); i != e; ++i)
862  DeadCases.insert(PredCases[i].Value);
863 
864  DEBUG(dbgs() << "Threading pred instr: " << *Pred->getTerminator()
865  << "Through successor TI: " << *TI);
866 
867  // Collect branch weights into a vector.
868  SmallVector<uint32_t, 8> Weights;
870  bool HasWeight = MD && (MD->getNumOperands() == 2 + SI->getNumCases());
871  if (HasWeight)
872  for (unsigned MD_i = 1, MD_e = MD->getNumOperands(); MD_i < MD_e;
873  ++MD_i) {
874  ConstantInt *CI = mdconst::extract<ConstantInt>(MD->getOperand(MD_i));
875  Weights.push_back(CI->getValue().getZExtValue());
876  }
877  for (SwitchInst::CaseIt i = SI->case_end(), e = SI->case_begin(); i != e;) {
878  --i;
879  if (DeadCases.count(i->getCaseValue())) {
880  if (HasWeight) {
881  std::swap(Weights[i->getCaseIndex() + 1], Weights.back());
882  Weights.pop_back();
883  }
884  i->getCaseSuccessor()->removePredecessor(TI->getParent());
885  SI->removeCase(i);
886  }
887  }
888  if (HasWeight && Weights.size() >= 2)
889  setBranchWeights(SI, Weights);
890 
891  DEBUG(dbgs() << "Leaving: " << *TI << "\n");
892  return true;
893  }
894 
895  // Otherwise, TI's block must correspond to some matched value. Find out
896  // which value (or set of values) this is.
897  ConstantInt *TIV = nullptr;
898  BasicBlock *TIBB = TI->getParent();
899  for (unsigned i = 0, e = PredCases.size(); i != e; ++i)
900  if (PredCases[i].Dest == TIBB) {
901  if (TIV)
902  return false; // Cannot handle multiple values coming to this block.
903  TIV = PredCases[i].Value;
904  }
905  assert(TIV && "No edge from pred to succ?");
906 
907  // Okay, we found the one constant that our value can be if we get into TI's
908  // BB. Find out which successor will unconditionally be branched to.
909  BasicBlock *TheRealDest = nullptr;
910  for (unsigned i = 0, e = ThisCases.size(); i != e; ++i)
911  if (ThisCases[i].Value == TIV) {
912  TheRealDest = ThisCases[i].Dest;
913  break;
914  }
915 
916  // If not handled by any explicit cases, it is handled by the default case.
917  if (!TheRealDest)
918  TheRealDest = ThisDef;
919 
920  // Remove PHI node entries for dead edges.
921  BasicBlock *CheckEdge = TheRealDest;
922  for (BasicBlock *Succ : successors(TIBB))
923  if (Succ != CheckEdge)
924  Succ->removePredecessor(TIBB);
925  else
926  CheckEdge = nullptr;
927 
928  // Insert the new branch.
929  Instruction *NI = Builder.CreateBr(TheRealDest);
930  (void)NI;
931 
932  DEBUG(dbgs() << "Threading pred instr: " << *Pred->getTerminator()
933  << "Through successor TI: " << *TI << "Leaving: " << *NI
934  << "\n");
935 
937  return true;
938 }
939 
940 namespace {
941 
942 /// This class implements a stable ordering of constant
943 /// integers that does not depend on their address. This is important for
944 /// applications that sort ConstantInt's to ensure uniqueness.
945 struct ConstantIntOrdering {
946  bool operator()(const ConstantInt *LHS, const ConstantInt *RHS) const {
947  return LHS->getValue().ult(RHS->getValue());
948  }
949 };
950 
951 } // end anonymous namespace
952 
953 static int ConstantIntSortPredicate(ConstantInt *const *P1,
954  ConstantInt *const *P2) {
955  const ConstantInt *LHS = *P1;
956  const ConstantInt *RHS = *P2;
957  if (LHS == RHS)
958  return 0;
959  return LHS->getValue().ult(RHS->getValue()) ? 1 : -1;
960 }
961 
962 static inline bool HasBranchWeights(const Instruction *I) {
964  if (ProfMD && ProfMD->getOperand(0))
965  if (MDString *MDS = dyn_cast<MDString>(ProfMD->getOperand(0)))
966  return MDS->getString().equals("branch_weights");
967 
968  return false;
969 }
970 
971 /// Get Weights of a given TerminatorInst, the default weight is at the front
972 /// of the vector. If TI is a conditional eq, we need to swap the branch-weight
973 /// metadata.
975  SmallVectorImpl<uint64_t> &Weights) {
977  assert(MD);
978  for (unsigned i = 1, e = MD->getNumOperands(); i < e; ++i) {
979  ConstantInt *CI = mdconst::extract<ConstantInt>(MD->getOperand(i));
980  Weights.push_back(CI->getValue().getZExtValue());
981  }
982 
983  // If TI is a conditional eq, the default case is the false case,
984  // and the corresponding branch-weight data is at index 2. We swap the
985  // default weight to be the first entry.
986  if (BranchInst *BI = dyn_cast<BranchInst>(TI)) {
987  assert(Weights.size() == 2);
988  ICmpInst *ICI = cast<ICmpInst>(BI->getCondition());
989  if (ICI->getPredicate() == ICmpInst::ICMP_EQ)
990  std::swap(Weights.front(), Weights.back());
991  }
992 }
993 
994 /// Keep halving the weights until all can fit in uint32_t.
995 static void FitWeights(MutableArrayRef<uint64_t> Weights) {
996  uint64_t Max = *std::max_element(Weights.begin(), Weights.end());
997  if (Max > UINT_MAX) {
998  unsigned Offset = 32 - countLeadingZeros(Max);
999  for (uint64_t &I : Weights)
1000  I >>= Offset;
1001  }
1002 }
1003 
1004 /// The specified terminator is a value equality comparison instruction
1005 /// (either a switch or a branch on "X == c").
1006 /// See if any of the predecessors of the terminator block are value comparisons
1007 /// on the same value. If so, and if safe to do so, fold them together.
1008 bool SimplifyCFGOpt::FoldValueComparisonIntoPredecessors(TerminatorInst *TI,
1009  IRBuilder<> &Builder) {
1010  BasicBlock *BB = TI->getParent();
1011  Value *CV = isValueEqualityComparison(TI); // CondVal
1012  assert(CV && "Not a comparison?");
1013  bool Changed = false;
1014 
1016  while (!Preds.empty()) {
1017  BasicBlock *Pred = Preds.pop_back_val();
1018 
1019  // See if the predecessor is a comparison with the same value.
1020  TerminatorInst *PTI = Pred->getTerminator();
1021  Value *PCV = isValueEqualityComparison(PTI); // PredCondVal
1022 
1023  if (PCV == CV && TI != PTI) {
1024  SmallSetVector<BasicBlock*, 4> FailBlocks;
1025  if (!SafeToMergeTerminators(TI, PTI, &FailBlocks)) {
1026  for (auto *Succ : FailBlocks) {
1027  if (!SplitBlockPredecessors(Succ, TI->getParent(), ".fold.split"))
1028  return false;
1029  }
1030  }
1031 
1032  // Figure out which 'cases' to copy from SI to PSI.
1033  std::vector<ValueEqualityComparisonCase> BBCases;
1034  BasicBlock *BBDefault = GetValueEqualityComparisonCases(TI, BBCases);
1035 
1036  std::vector<ValueEqualityComparisonCase> PredCases;
1037  BasicBlock *PredDefault = GetValueEqualityComparisonCases(PTI, PredCases);
1038 
1039  // Based on whether the default edge from PTI goes to BB or not, fill in
1040  // PredCases and PredDefault with the new switch cases we would like to
1041  // build.
1042  SmallVector<BasicBlock *, 8> NewSuccessors;
1043 
1044  // Update the branch weight metadata along the way
1045  SmallVector<uint64_t, 8> Weights;
1046  bool PredHasWeights = HasBranchWeights(PTI);
1047  bool SuccHasWeights = HasBranchWeights(TI);
1048 
1049  if (PredHasWeights) {
1050  GetBranchWeights(PTI, Weights);
1051  // branch-weight metadata is inconsistent here.
1052  if (Weights.size() != 1 + PredCases.size())
1053  PredHasWeights = SuccHasWeights = false;
1054  } else if (SuccHasWeights)
1055  // If there are no predecessor weights but there are successor weights,
1056  // populate Weights with 1, which will later be scaled to the sum of
1057  // successor's weights
1058  Weights.assign(1 + PredCases.size(), 1);
1059 
1060  SmallVector<uint64_t, 8> SuccWeights;
1061  if (SuccHasWeights) {
1062  GetBranchWeights(TI, SuccWeights);
1063  // branch-weight metadata is inconsistent here.
1064  if (SuccWeights.size() != 1 + BBCases.size())
1065  PredHasWeights = SuccHasWeights = false;
1066  } else if (PredHasWeights)
1067  SuccWeights.assign(1 + BBCases.size(), 1);
1068 
1069  if (PredDefault == BB) {
1070  // If this is the default destination from PTI, only the edges in TI
1071  // that don't occur in PTI, or that branch to BB will be activated.
1072  std::set<ConstantInt *, ConstantIntOrdering> PTIHandled;
1073  for (unsigned i = 0, e = PredCases.size(); i != e; ++i)
1074  if (PredCases[i].Dest != BB)
1075  PTIHandled.insert(PredCases[i].Value);
1076  else {
1077  // The default destination is BB, we don't need explicit targets.
1078  std::swap(PredCases[i], PredCases.back());
1079 
1080  if (PredHasWeights || SuccHasWeights) {
1081  // Increase weight for the default case.
1082  Weights[0] += Weights[i + 1];
1083  std::swap(Weights[i + 1], Weights.back());
1084  Weights.pop_back();
1085  }
1086 
1087  PredCases.pop_back();
1088  --i;
1089  --e;
1090  }
1091 
1092  // Reconstruct the new switch statement we will be building.
1093  if (PredDefault != BBDefault) {
1094  PredDefault->removePredecessor(Pred);
1095  PredDefault = BBDefault;
1096  NewSuccessors.push_back(BBDefault);
1097  }
1098 
1099  unsigned CasesFromPred = Weights.size();
1100  uint64_t ValidTotalSuccWeight = 0;
1101  for (unsigned i = 0, e = BBCases.size(); i != e; ++i)
1102  if (!PTIHandled.count(BBCases[i].Value) &&
1103  BBCases[i].Dest != BBDefault) {
1104  PredCases.push_back(BBCases[i]);
1105  NewSuccessors.push_back(BBCases[i].Dest);
1106  if (SuccHasWeights || PredHasWeights) {
1107  // The default weight is at index 0, so weight for the ith case
1108  // should be at index i+1. Scale the cases from successor by
1109  // PredDefaultWeight (Weights[0]).
1110  Weights.push_back(Weights[0] * SuccWeights[i + 1]);
1111  ValidTotalSuccWeight += SuccWeights[i + 1];
1112  }
1113  }
1114 
1115  if (SuccHasWeights || PredHasWeights) {
1116  ValidTotalSuccWeight += SuccWeights[0];
1117  // Scale the cases from predecessor by ValidTotalSuccWeight.
1118  for (unsigned i = 1; i < CasesFromPred; ++i)
1119  Weights[i] *= ValidTotalSuccWeight;
1120  // Scale the default weight by SuccDefaultWeight (SuccWeights[0]).
1121  Weights[0] *= SuccWeights[0];
1122  }
1123  } else {
1124  // If this is not the default destination from PSI, only the edges
1125  // in SI that occur in PSI with a destination of BB will be
1126  // activated.
1127  std::set<ConstantInt *, ConstantIntOrdering> PTIHandled;
1128  std::map<ConstantInt *, uint64_t> WeightsForHandled;
1129  for (unsigned i = 0, e = PredCases.size(); i != e; ++i)
1130  if (PredCases[i].Dest == BB) {
1131  PTIHandled.insert(PredCases[i].Value);
1132 
1133  if (PredHasWeights || SuccHasWeights) {
1134  WeightsForHandled[PredCases[i].Value] = Weights[i + 1];
1135  std::swap(Weights[i + 1], Weights.back());
1136  Weights.pop_back();
1137  }
1138 
1139  std::swap(PredCases[i], PredCases.back());
1140  PredCases.pop_back();
1141  --i;
1142  --e;
1143  }
1144 
1145  // Okay, now we know which constants were sent to BB from the
1146  // predecessor. Figure out where they will all go now.
1147  for (unsigned i = 0, e = BBCases.size(); i != e; ++i)
1148  if (PTIHandled.count(BBCases[i].Value)) {
1149  // If this is one we are capable of getting...
1150  if (PredHasWeights || SuccHasWeights)
1151  Weights.push_back(WeightsForHandled[BBCases[i].Value]);
1152  PredCases.push_back(BBCases[i]);
1153  NewSuccessors.push_back(BBCases[i].Dest);
1154  PTIHandled.erase(
1155  BBCases[i].Value); // This constant is taken care of
1156  }
1157 
1158  // If there are any constants vectored to BB that TI doesn't handle,
1159  // they must go to the default destination of TI.
1160  for (ConstantInt *I : PTIHandled) {
1161  if (PredHasWeights || SuccHasWeights)
1162  Weights.push_back(WeightsForHandled[I]);
1163  PredCases.push_back(ValueEqualityComparisonCase(I, BBDefault));
1164  NewSuccessors.push_back(BBDefault);
1165  }
1166  }
1167 
1168  // Okay, at this point, we know which new successor Pred will get. Make
1169  // sure we update the number of entries in the PHI nodes for these
1170  // successors.
1171  for (BasicBlock *NewSuccessor : NewSuccessors)
1172  AddPredecessorToBlock(NewSuccessor, Pred, BB);
1173 
1174  Builder.SetInsertPoint(PTI);
1175  // Convert pointer to int before we switch.
1176  if (CV->getType()->isPointerTy()) {
1177  CV = Builder.CreatePtrToInt(CV, DL.getIntPtrType(CV->getType()),
1178  "magicptr");
1179  }
1180 
1181  // Now that the successors are updated, create the new Switch instruction.
1182  SwitchInst *NewSI =
1183  Builder.CreateSwitch(CV, PredDefault, PredCases.size());
1184  NewSI->setDebugLoc(PTI->getDebugLoc());
1185  for (ValueEqualityComparisonCase &V : PredCases)
1186  NewSI->addCase(V.Value, V.Dest);
1187 
1188  if (PredHasWeights || SuccHasWeights) {
1189  // Halve the weights if any of them cannot fit in an uint32_t
1190  FitWeights(Weights);
1191 
1192  SmallVector<uint32_t, 8> MDWeights(Weights.begin(), Weights.end());
1193 
1194  setBranchWeights(NewSI, MDWeights);
1195  }
1196 
1198 
1199  // Okay, last check. If BB is still a successor of PSI, then we must
1200  // have an infinite loop case. If so, add an infinitely looping block
1201  // to handle the case to preserve the behavior of the code.
1202  BasicBlock *InfLoopBlock = nullptr;
1203  for (unsigned i = 0, e = NewSI->getNumSuccessors(); i != e; ++i)
1204  if (NewSI->getSuccessor(i) == BB) {
1205  if (!InfLoopBlock) {
1206  // Insert it at the end of the function, because it's either code,
1207  // or it won't matter if it's hot. :)
1208  InfLoopBlock = BasicBlock::Create(BB->getContext(), "infloop",
1209  BB->getParent());
1210  BranchInst::Create(InfLoopBlock, InfLoopBlock);
1211  }
1212  NewSI->setSuccessor(i, InfLoopBlock);
1213  }
1214 
1215  Changed = true;
1216  }
1217  }
1218  return Changed;
1219 }
1220 
1221 // If we would need to insert a select that uses the value of this invoke
1222 // (comments in HoistThenElseCodeToIf explain why we would need to do this), we
1223 // can't hoist the invoke, as there is nowhere to put the select in this case.
1225  Instruction *I1, Instruction *I2) {
1226  for (BasicBlock *Succ : successors(BB1)) {
1227  for (const PHINode &PN : Succ->phis()) {
1228  Value *BB1V = PN.getIncomingValueForBlock(BB1);
1229  Value *BB2V = PN.getIncomingValueForBlock(BB2);
1230  if (BB1V != BB2V && (BB1V == I1 || BB2V == I2)) {
1231  return false;
1232  }
1233  }
1234  }
1235  return true;
1236 }
1237 
1239 
1240 /// Given a conditional branch that goes to BB1 and BB2, hoist any common code
1241 /// in the two blocks up into the branch block. The caller of this function
1242 /// guarantees that BI's block dominates BB1 and BB2.
1244  const TargetTransformInfo &TTI) {
1245  // This does very trivial matching, with limited scanning, to find identical
1246  // instructions in the two blocks. In particular, we don't want to get into
1247  // O(M*N) situations here where M and N are the sizes of BB1 and BB2. As
1248  // such, we currently just scan for obviously identical instructions in an
1249  // identical order.
1250  BasicBlock *BB1 = BI->getSuccessor(0); // The true destination.
1251  BasicBlock *BB2 = BI->getSuccessor(1); // The false destination
1252 
1253  BasicBlock::iterator BB1_Itr = BB1->begin();
1254  BasicBlock::iterator BB2_Itr = BB2->begin();
1255 
1256  Instruction *I1 = &*BB1_Itr++, *I2 = &*BB2_Itr++;
1257  // Skip debug info if it is not identical.
1260  if (!DBI1 || !DBI2 || !DBI1->isIdenticalToWhenDefined(DBI2)) {
1261  while (isa<DbgInfoIntrinsic>(I1))
1262  I1 = &*BB1_Itr++;
1263  while (isa<DbgInfoIntrinsic>(I2))
1264  I2 = &*BB2_Itr++;
1265  }
1266  if (isa<PHINode>(I1) || !I1->isIdenticalToWhenDefined(I2) ||
1267  (isa<InvokeInst>(I1) && !isSafeToHoistInvoke(BB1, BB2, I1, I2)))
1268  return false;
1269 
1270  BasicBlock *BIParent = BI->getParent();
1271 
1272  bool Changed = false;
1273  do {
1274  // If we are hoisting the terminator instruction, don't move one (making a
1275  // broken BB), instead clone it, and remove BI.
1276  if (isa<TerminatorInst>(I1))
1277  goto HoistTerminator;
1278 
1279  // If we're going to hoist a call, make sure that the two instructions we're
1280  // commoning/hoisting are both marked with musttail, or neither of them is
1281  // marked as such. Otherwise, we might end up in a situation where we hoist
1282  // from a block where the terminator is a `ret` to a block where the terminator
1283  // is a `br`, and `musttail` calls expect to be followed by a return.
1284  auto *C1 = dyn_cast<CallInst>(I1);
1285  auto *C2 = dyn_cast<CallInst>(I2);
1286  if (C1 && C2)
1287  if (C1->isMustTailCall() != C2->isMustTailCall())
1288  return Changed;
1289 
1290  if (!TTI.isProfitableToHoist(I1) || !TTI.isProfitableToHoist(I2))
1291  return Changed;
1292 
1293  // For a normal instruction, we just move one to right before the branch,
1294  // then replace all uses of the other with the first. Finally, we remove
1295  // the now redundant second instruction.
1296  BIParent->getInstList().splice(BI->getIterator(), BB1->getInstList(), I1);
1297  if (!I2->use_empty())
1298  I2->replaceAllUsesWith(I1);
1299  I1->andIRFlags(I2);
1300  unsigned KnownIDs[] = {LLVMContext::MD_tbaa,
1310  combineMetadata(I1, I2, KnownIDs);
1311 
1312  // I1 and I2 are being combined into a single instruction. Its debug
1313  // location is the merged locations of the original instructions.
1314  I1->applyMergedLocation(I1->getDebugLoc(), I2->getDebugLoc());
1315 
1316  I2->eraseFromParent();
1317  Changed = true;
1318 
1319  I1 = &*BB1_Itr++;
1320  I2 = &*BB2_Itr++;
1321  // Skip debug info if it is not identical.
1324  if (!DBI1 || !DBI2 || !DBI1->isIdenticalToWhenDefined(DBI2)) {
1325  while (isa<DbgInfoIntrinsic>(I1))
1326  I1 = &*BB1_Itr++;
1327  while (isa<DbgInfoIntrinsic>(I2))
1328  I2 = &*BB2_Itr++;
1329  }
1330  } while (I1->isIdenticalToWhenDefined(I2));
1331 
1332  return true;
1333 
1334 HoistTerminator:
1335  // It may not be possible to hoist an invoke.
1336  if (isa<InvokeInst>(I1) && !isSafeToHoistInvoke(BB1, BB2, I1, I2))
1337  return Changed;
1338 
1339  for (BasicBlock *Succ : successors(BB1)) {
1340  for (PHINode &PN : Succ->phis()) {
1341  Value *BB1V = PN.getIncomingValueForBlock(BB1);
1342  Value *BB2V = PN.getIncomingValueForBlock(BB2);
1343  if (BB1V == BB2V)
1344  continue;
1345 
1346  // Check for passingValueIsAlwaysUndefined here because we would rather
1347  // eliminate undefined control flow then converting it to a select.
1348  if (passingValueIsAlwaysUndefined(BB1V, &PN) ||
1349  passingValueIsAlwaysUndefined(BB2V, &PN))
1350  return Changed;
1351 
1352  if (isa<ConstantExpr>(BB1V) && !isSafeToSpeculativelyExecute(BB1V))
1353  return Changed;
1354  if (isa<ConstantExpr>(BB2V) && !isSafeToSpeculativelyExecute(BB2V))
1355  return Changed;
1356  }
1357  }
1358 
1359  // Okay, it is safe to hoist the terminator.
1360  Instruction *NT = I1->clone();
1361  BIParent->getInstList().insert(BI->getIterator(), NT);
1362  if (!NT->getType()->isVoidTy()) {
1363  I1->replaceAllUsesWith(NT);
1364  I2->replaceAllUsesWith(NT);
1365  NT->takeName(I1);
1366  }
1367 
1368  IRBuilder<NoFolder> Builder(NT);
1369  // Hoisting one of the terminators from our successor is a great thing.
1370  // Unfortunately, the successors of the if/else blocks may have PHI nodes in
1371  // them. If they do, all PHI entries for BB1/BB2 must agree for all PHI
1372  // nodes, so we insert select instruction to compute the final result.
1373  std::map<std::pair<Value *, Value *>, SelectInst *> InsertedSelects;
1374  for (BasicBlock *Succ : successors(BB1)) {
1375  for (PHINode &PN : Succ->phis()) {
1376  Value *BB1V = PN.getIncomingValueForBlock(BB1);
1377  Value *BB2V = PN.getIncomingValueForBlock(BB2);
1378  if (BB1V == BB2V)
1379  continue;
1380 
1381  // These values do not agree. Insert a select instruction before NT
1382  // that determines the right value.
1383  SelectInst *&SI = InsertedSelects[std::make_pair(BB1V, BB2V)];
1384  if (!SI)
1385  SI = cast<SelectInst>(
1386  Builder.CreateSelect(BI->getCondition(), BB1V, BB2V,
1387  BB1V->getName() + "." + BB2V->getName(), BI));
1388 
1389  // Make the PHI node use the select for all incoming values for BB1/BB2
1390  for (unsigned i = 0, e = PN.getNumIncomingValues(); i != e; ++i)
1391  if (PN.getIncomingBlock(i) == BB1 || PN.getIncomingBlock(i) == BB2)
1392  PN.setIncomingValue(i, SI);
1393  }
1394  }
1395 
1396  // Update any PHI nodes in our new successors.
1397  for (BasicBlock *Succ : successors(BB1))
1398  AddPredecessorToBlock(Succ, BIParent, BB1);
1399 
1401  return true;
1402 }
1403 
1404 // All instructions in Insts belong to different blocks that all unconditionally
1405 // branch to a common successor. Analyze each instruction and return true if it
1406 // would be possible to sink them into their successor, creating one common
1407 // instruction instead. For every value that would be required to be provided by
1408 // PHI node (because an operand varies in each input block), add to PHIOperands.
1411  DenseMap<Instruction *, SmallVector<Value *, 4>> &PHIOperands) {
1412  // Prune out obviously bad instructions to move. Any non-store instruction
1413  // must have exactly one use, and we check later that use is by a single,
1414  // common PHI instruction in the successor.
1415  for (auto *I : Insts) {
1416  // These instructions may change or break semantics if moved.
1417  if (isa<PHINode>(I) || I->isEHPad() || isa<AllocaInst>(I) ||
1418  I->getType()->isTokenTy())
1419  return false;
1420 
1421  // Conservatively return false if I is an inline-asm instruction. Sinking
1422  // and merging inline-asm instructions can potentially create arguments
1423  // that cannot satisfy the inline-asm constraints.
1424  if (const auto *C = dyn_cast<CallInst>(I))
1425  if (C->isInlineAsm())
1426  return false;
1427 
1428  // Everything must have only one use too, apart from stores which
1429  // have no uses.
1430  if (!isa<StoreInst>(I) && !I->hasOneUse())
1431  return false;
1432  }
1433 
1434  const Instruction *I0 = Insts.front();
1435  for (auto *I : Insts)
1436  if (!I->isSameOperationAs(I0))
1437  return false;
1438 
1439  // All instructions in Insts are known to be the same opcode. If they aren't
1440  // stores, check the only user of each is a PHI or in the same block as the
1441  // instruction, because if a user is in the same block as an instruction
1442  // we're contemplating sinking, it must already be determined to be sinkable.
1443  if (!isa<StoreInst>(I0)) {
1444  auto *PNUse = dyn_cast<PHINode>(*I0->user_begin());
1445  auto *Succ = I0->getParent()->getTerminator()->getSuccessor(0);
1446  if (!all_of(Insts, [&PNUse,&Succ](const Instruction *I) -> bool {
1447  auto *U = cast<Instruction>(*I->user_begin());
1448  return (PNUse &&
1449  PNUse->getParent() == Succ &&
1450  PNUse->getIncomingValueForBlock(I->getParent()) == I) ||
1451  U->getParent() == I->getParent();
1452  }))
1453  return false;
1454  }
1455 
1456  // Because SROA can't handle speculating stores of selects, try not
1457  // to sink loads or stores of allocas when we'd have to create a PHI for
1458  // the address operand. Also, because it is likely that loads or stores
1459  // of allocas will disappear when Mem2Reg/SROA is run, don't sink them.
1460  // This can cause code churn which can have unintended consequences down
1461  // the line - see https://llvm.org/bugs/show_bug.cgi?id=30244.
1462  // FIXME: This is a workaround for a deficiency in SROA - see
1463  // https://llvm.org/bugs/show_bug.cgi?id=30188
1464  if (isa<StoreInst>(I0) && any_of(Insts, [](const Instruction *I) {
1465  return isa<AllocaInst>(I->getOperand(1));
1466  }))
1467  return false;
1468  if (isa<LoadInst>(I0) && any_of(Insts, [](const Instruction *I) {
1469  return isa<AllocaInst>(I->getOperand(0));
1470  }))
1471  return false;
1472 
1473  for (unsigned OI = 0, OE = I0->getNumOperands(); OI != OE; ++OI) {
1474  if (I0->getOperand(OI)->getType()->isTokenTy())
1475  // Don't touch any operand of token type.
1476  return false;
1477 
1478  auto SameAsI0 = [&I0, OI](const Instruction *I) {
1479  assert(I->getNumOperands() == I0->getNumOperands());
1480  return I->getOperand(OI) == I0->getOperand(OI);
1481  };
1482  if (!all_of(Insts, SameAsI0)) {
1483  if (!canReplaceOperandWithVariable(I0, OI))
1484  // We can't create a PHI from this GEP.
1485  return false;
1486  // Don't create indirect calls! The called value is the final operand.
1487  if ((isa<CallInst>(I0) || isa<InvokeInst>(I0)) && OI == OE - 1) {
1488  // FIXME: if the call was *already* indirect, we should do this.
1489  return false;
1490  }
1491  for (auto *I : Insts)
1492  PHIOperands[I].push_back(I->getOperand(OI));
1493  }
1494  }
1495  return true;
1496 }
1497 
1498 // Assuming canSinkLastInstruction(Blocks) has returned true, sink the last
1499 // instruction of every block in Blocks to their common successor, commoning
1500 // into one instruction.
1502  auto *BBEnd = Blocks[0]->getTerminator()->getSuccessor(0);
1503 
1504  // canSinkLastInstruction returning true guarantees that every block has at
1505  // least one non-terminator instruction.
1507  for (auto *BB : Blocks) {
1508  Instruction *I = BB->getTerminator();
1509  do {
1510  I = I->getPrevNode();
1511  } while (isa<DbgInfoIntrinsic>(I) && I != &BB->front());
1512  if (!isa<DbgInfoIntrinsic>(I))
1513  Insts.push_back(I);
1514  }
1515 
1516  // The only checking we need to do now is that all users of all instructions
1517  // are the same PHI node. canSinkLastInstruction should have checked this but
1518  // it is slightly over-aggressive - it gets confused by commutative instructions
1519  // so double-check it here.
1520  Instruction *I0 = Insts.front();
1521  if (!isa<StoreInst>(I0)) {
1522  auto *PNUse = dyn_cast<PHINode>(*I0->user_begin());
1523  if (!all_of(Insts, [&PNUse](const Instruction *I) -> bool {
1524  auto *U = cast<Instruction>(*I->user_begin());
1525  return U == PNUse;
1526  }))
1527  return false;
1528  }
1529 
1530  // We don't need to do any more checking here; canSinkLastInstruction should
1531  // have done it all for us.
1532  SmallVector<Value*, 4> NewOperands;
1533  for (unsigned O = 0, E = I0->getNumOperands(); O != E; ++O) {
1534  // This check is different to that in canSinkLastInstruction. There, we
1535  // cared about the global view once simplifycfg (and instcombine) have
1536  // completed - it takes into account PHIs that become trivially
1537  // simplifiable. However here we need a more local view; if an operand
1538  // differs we create a PHI and rely on instcombine to clean up the very
1539  // small mess we may make.
1540  bool NeedPHI = any_of(Insts, [&I0, O](const Instruction *I) {
1541  return I->getOperand(O) != I0->getOperand(O);
1542  });
1543  if (!NeedPHI) {
1544  NewOperands.push_back(I0->getOperand(O));
1545  continue;
1546  }
1547 
1548  // Create a new PHI in the successor block and populate it.
1549  auto *Op = I0->getOperand(O);
1550  assert(!Op->getType()->isTokenTy() && "Can't PHI tokens!");
1551  auto *PN = PHINode::Create(Op->getType(), Insts.size(),
1552  Op->getName() + ".sink", &BBEnd->front());
1553  for (auto *I : Insts)
1554  PN->addIncoming(I->getOperand(O), I->getParent());
1555  NewOperands.push_back(PN);
1556  }
1557 
1558  // Arbitrarily use I0 as the new "common" instruction; remap its operands
1559  // and move it to the start of the successor block.
1560  for (unsigned O = 0, E = I0->getNumOperands(); O != E; ++O)
1561  I0->getOperandUse(O).set(NewOperands[O]);
1562  I0->moveBefore(&*BBEnd->getFirstInsertionPt());
1563 
1564  // Update metadata and IR flags, and merge debug locations.
1565  for (auto *I : Insts)
1566  if (I != I0) {
1567  // The debug location for the "common" instruction is the merged locations
1568  // of all the commoned instructions. We start with the original location
1569  // of the "common" instruction and iteratively merge each location in the
1570  // loop below.
1571  // This is an N-way merge, which will be inefficient if I0 is a CallInst.
1572  // However, as N-way merge for CallInst is rare, so we use simplified API
1573  // instead of using complex API for N-way merge.
1574  I0->applyMergedLocation(I0->getDebugLoc(), I->getDebugLoc());
1575  combineMetadataForCSE(I0, I);
1576  I0->andIRFlags(I);
1577  }
1578 
1579  if (!isa<StoreInst>(I0)) {
1580  // canSinkLastInstruction checked that all instructions were used by
1581  // one and only one PHI node. Find that now, RAUW it to our common
1582  // instruction and nuke it.
1583  assert(I0->hasOneUse());
1584  auto *PN = cast<PHINode>(*I0->user_begin());
1585  PN->replaceAllUsesWith(I0);
1586  PN->eraseFromParent();
1587  }
1588 
1589  // Finally nuke all instructions apart from the common instruction.
1590  for (auto *I : Insts)
1591  if (I != I0)
1592  I->eraseFromParent();
1593 
1594  return true;
1595 }
1596 
1597 namespace {
1598 
1599  // LockstepReverseIterator - Iterates through instructions
1600  // in a set of blocks in reverse order from the first non-terminator.
1601  // For example (assume all blocks have size n):
1602  // LockstepReverseIterator I([B1, B2, B3]);
1603  // *I-- = [B1[n], B2[n], B3[n]];
1604  // *I-- = [B1[n-1], B2[n-1], B3[n-1]];
1605  // *I-- = [B1[n-2], B2[n-2], B3[n-2]];
1606  // ...
1607  class LockstepReverseIterator {
1608  ArrayRef<BasicBlock*> Blocks;
1610  bool Fail;
1611 
1612  public:
1613  LockstepReverseIterator(ArrayRef<BasicBlock*> Blocks) : Blocks(Blocks) {
1614  reset();
1615  }
1616 
1617  void reset() {
1618  Fail = false;
1619  Insts.clear();
1620  for (auto *BB : Blocks) {
1621  Instruction *Inst = BB->getTerminator();
1622  for (Inst = Inst->getPrevNode(); Inst && isa<DbgInfoIntrinsic>(Inst);)
1623  Inst = Inst->getPrevNode();
1624  if (!Inst) {
1625  // Block wasn't big enough.
1626  Fail = true;
1627  return;
1628  }
1629  Insts.push_back(Inst);
1630  }
1631  }
1632 
1633  bool isValid() const {
1634  return !Fail;
1635  }
1636 
1637  void operator--() {
1638  if (Fail)
1639  return;
1640  for (auto *&Inst : Insts) {
1641  for (Inst = Inst->getPrevNode(); Inst && isa<DbgInfoIntrinsic>(Inst);)
1642  Inst = Inst->getPrevNode();
1643  // Already at beginning of block.
1644  if (!Inst) {
1645  Fail = true;
1646  return;
1647  }
1648  }
1649  }
1650 
1652  return Insts;
1653  }
1654  };
1655 
1656 } // end anonymous namespace
1657 
1658 /// Check whether BB's predecessors end with unconditional branches. If it is
1659 /// true, sink any common code from the predecessors to BB.
1660 /// We also allow one predecessor to end with conditional branch (but no more
1661 /// than one).
1663  // We support two situations:
1664  // (1) all incoming arcs are unconditional
1665  // (2) one incoming arc is conditional
1666  //
1667  // (2) is very common in switch defaults and
1668  // else-if patterns;
1669  //
1670  // if (a) f(1);
1671  // else if (b) f(2);
1672  //
1673  // produces:
1674  //
1675  // [if]
1676  // / \
1677  // [f(1)] [if]
1678  // | | \
1679  // | | |
1680  // | [f(2)]|
1681  // \ | /
1682  // [ end ]
1683  //
1684  // [end] has two unconditional predecessor arcs and one conditional. The
1685  // conditional refers to the implicit empty 'else' arc. This conditional
1686  // arc can also be caused by an empty default block in a switch.
1687  //
1688  // In this case, we attempt to sink code from all *unconditional* arcs.
1689  // If we can sink instructions from these arcs (determined during the scan
1690  // phase below) we insert a common successor for all unconditional arcs and
1691  // connect that to [end], to enable sinking:
1692  //
1693  // [if]
1694  // / \
1695  // [x(1)] [if]
1696  // | | \
1697  // | | \
1698  // | [x(2)] |
1699  // \ / |
1700  // [sink.split] |
1701  // \ /
1702  // [ end ]
1703  //
1704  SmallVector<BasicBlock*,4> UnconditionalPreds;
1705  Instruction *Cond = nullptr;
1706  for (auto *B : predecessors(BB)) {
1707  auto *T = B->getTerminator();
1708  if (isa<BranchInst>(T) && cast<BranchInst>(T)->isUnconditional())
1709  UnconditionalPreds.push_back(B);
1710  else if ((isa<BranchInst>(T) || isa<SwitchInst>(T)) && !Cond)
1711  Cond = T;
1712  else
1713  return false;
1714  }
1715  if (UnconditionalPreds.size() < 2)
1716  return false;
1717 
1718  bool Changed = false;
1719  // We take a two-step approach to tail sinking. First we scan from the end of
1720  // each block upwards in lockstep. If the n'th instruction from the end of each
1721  // block can be sunk, those instructions are added to ValuesToSink and we
1722  // carry on. If we can sink an instruction but need to PHI-merge some operands
1723  // (because they're not identical in each instruction) we add these to
1724  // PHIOperands.
1725  unsigned ScanIdx = 0;
1726  SmallPtrSet<Value*,4> InstructionsToSink;
1728  LockstepReverseIterator LRI(UnconditionalPreds);
1729  while (LRI.isValid() &&
1730  canSinkInstructions(*LRI, PHIOperands)) {
1731  DEBUG(dbgs() << "SINK: instruction can be sunk: " << *(*LRI)[0] << "\n");
1732  InstructionsToSink.insert((*LRI).begin(), (*LRI).end());
1733  ++ScanIdx;
1734  --LRI;
1735  }
1736 
1737  auto ProfitableToSinkInstruction = [&](LockstepReverseIterator &LRI) {
1738  unsigned NumPHIdValues = 0;
1739  for (auto *I : *LRI)
1740  for (auto *V : PHIOperands[I])
1741  if (InstructionsToSink.count(V) == 0)
1742  ++NumPHIdValues;
1743  DEBUG(dbgs() << "SINK: #phid values: " << NumPHIdValues << "\n");
1744  unsigned NumPHIInsts = NumPHIdValues / UnconditionalPreds.size();
1745  if ((NumPHIdValues % UnconditionalPreds.size()) != 0)
1746  NumPHIInsts++;
1747 
1748  return NumPHIInsts <= 1;
1749  };
1750 
1751  if (ScanIdx > 0 && Cond) {
1752  // Check if we would actually sink anything first! This mutates the CFG and
1753  // adds an extra block. The goal in doing this is to allow instructions that
1754  // couldn't be sunk before to be sunk - obviously, speculatable instructions
1755  // (such as trunc, add) can be sunk and predicated already. So we check that
1756  // we're going to sink at least one non-speculatable instruction.
1757  LRI.reset();
1758  unsigned Idx = 0;
1759  bool Profitable = false;
1760  while (ProfitableToSinkInstruction(LRI) && Idx < ScanIdx) {
1761  if (!isSafeToSpeculativelyExecute((*LRI)[0])) {
1762  Profitable = true;
1763  break;
1764  }
1765  --LRI;
1766  ++Idx;
1767  }
1768  if (!Profitable)
1769  return false;
1770 
1771  DEBUG(dbgs() << "SINK: Splitting edge\n");
1772  // We have a conditional edge and we're going to sink some instructions.
1773  // Insert a new block postdominating all blocks we're going to sink from.
1774  if (!SplitBlockPredecessors(BB, UnconditionalPreds, ".sink.split"))
1775  // Edges couldn't be split.
1776  return false;
1777  Changed = true;
1778  }
1779 
1780  // Now that we've analyzed all potential sinking candidates, perform the
1781  // actual sink. We iteratively sink the last non-terminator of the source
1782  // blocks into their common successor unless doing so would require too
1783  // many PHI instructions to be generated (currently only one PHI is allowed
1784  // per sunk instruction).
1785  //
1786  // We can use InstructionsToSink to discount values needing PHI-merging that will
1787  // actually be sunk in a later iteration. This allows us to be more
1788  // aggressive in what we sink. This does allow a false positive where we
1789  // sink presuming a later value will also be sunk, but stop half way through
1790  // and never actually sink it which means we produce more PHIs than intended.
1791  // This is unlikely in practice though.
1792  for (unsigned SinkIdx = 0; SinkIdx != ScanIdx; ++SinkIdx) {
1793  DEBUG(dbgs() << "SINK: Sink: "
1794  << *UnconditionalPreds[0]->getTerminator()->getPrevNode()
1795  << "\n");
1796 
1797  // Because we've sunk every instruction in turn, the current instruction to
1798  // sink is always at index 0.
1799  LRI.reset();
1800  if (!ProfitableToSinkInstruction(LRI)) {
1801  // Too many PHIs would be created.
1802  DEBUG(dbgs() << "SINK: stopping here, too many PHIs would be created!\n");
1803  break;
1804  }
1805 
1806  if (!sinkLastInstruction(UnconditionalPreds))
1807  return Changed;
1808  NumSinkCommons++;
1809  Changed = true;
1810  }
1811  return Changed;
1812 }
1813 
1814 /// \brief Determine if we can hoist sink a sole store instruction out of a
1815 /// conditional block.
1816 ///
1817 /// We are looking for code like the following:
1818 /// BrBB:
1819 /// store i32 %add, i32* %arrayidx2
1820 /// ... // No other stores or function calls (we could be calling a memory
1821 /// ... // function).
1822 /// %cmp = icmp ult %x, %y
1823 /// br i1 %cmp, label %EndBB, label %ThenBB
1824 /// ThenBB:
1825 /// store i32 %add5, i32* %arrayidx2
1826 /// br label EndBB
1827 /// EndBB:
1828 /// ...
1829 /// We are going to transform this into:
1830 /// BrBB:
1831 /// store i32 %add, i32* %arrayidx2
1832 /// ... //
1833 /// %cmp = icmp ult %x, %y
1834 /// %add.add5 = select i1 %cmp, i32 %add, %add5
1835 /// store i32 %add.add5, i32* %arrayidx2
1836 /// ...
1837 ///
1838 /// \return The pointer to the value of the previous store if the store can be
1839 /// hoisted into the predecessor block. 0 otherwise.
1841  BasicBlock *StoreBB, BasicBlock *EndBB) {
1842  StoreInst *StoreToHoist = dyn_cast<StoreInst>(I);
1843  if (!StoreToHoist)
1844  return nullptr;
1845 
1846  // Volatile or atomic.
1847  if (!StoreToHoist->isSimple())
1848  return nullptr;
1849 
1850  Value *StorePtr = StoreToHoist->getPointerOperand();
1851 
1852  // Look for a store to the same pointer in BrBB.
1853  unsigned MaxNumInstToLookAt = 9;
1854  for (Instruction &CurI : reverse(*BrBB)) {
1855  if (!MaxNumInstToLookAt)
1856  break;
1857  // Skip debug info.
1858  if (isa<DbgInfoIntrinsic>(CurI))
1859  continue;
1860  --MaxNumInstToLookAt;
1861 
1862  // Could be calling an instruction that affects memory like free().
1863  if (CurI.mayHaveSideEffects() && !isa<StoreInst>(CurI))
1864  return nullptr;
1865 
1866  if (auto *SI = dyn_cast<StoreInst>(&CurI)) {
1867  // Found the previous store make sure it stores to the same location.
1868  if (SI->getPointerOperand() == StorePtr)
1869  // Found the previous store, return its value operand.
1870  return SI->getValueOperand();
1871  return nullptr; // Unknown store.
1872  }
1873  }
1874 
1875  return nullptr;
1876 }
1877 
1878 /// \brief Speculate a conditional basic block flattening the CFG.
1879 ///
1880 /// Note that this is a very risky transform currently. Speculating
1881 /// instructions like this is most often not desirable. Instead, there is an MI
1882 /// pass which can do it with full awareness of the resource constraints.
1883 /// However, some cases are "obvious" and we should do directly. An example of
1884 /// this is speculating a single, reasonably cheap instruction.
1885 ///
1886 /// There is only one distinct advantage to flattening the CFG at the IR level:
1887 /// it makes very common but simplistic optimizations such as are common in
1888 /// instcombine and the DAG combiner more powerful by removing CFG edges and
1889 /// modeling their effects with easier to reason about SSA value graphs.
1890 ///
1891 ///
1892 /// An illustration of this transform is turning this IR:
1893 /// \code
1894 /// BB:
1895 /// %cmp = icmp ult %x, %y
1896 /// br i1 %cmp, label %EndBB, label %ThenBB
1897 /// ThenBB:
1898 /// %sub = sub %x, %y
1899 /// br label BB2
1900 /// EndBB:
1901 /// %phi = phi [ %sub, %ThenBB ], [ 0, %EndBB ]
1902 /// ...
1903 /// \endcode
1904 ///
1905 /// Into this IR:
1906 /// \code
1907 /// BB:
1908 /// %cmp = icmp ult %x, %y
1909 /// %sub = sub %x, %y
1910 /// %cond = select i1 %cmp, 0, %sub
1911 /// ...
1912 /// \endcode
1913 ///
1914 /// \returns true if the conditional block is removed.
1916  const TargetTransformInfo &TTI) {
1917  // Be conservative for now. FP select instruction can often be expensive.
1918  Value *BrCond = BI->getCondition();
1919  if (isa<FCmpInst>(BrCond))
1920  return false;
1921 
1922  BasicBlock *BB = BI->getParent();
1923  BasicBlock *EndBB = ThenBB->getTerminator()->getSuccessor(0);
1924 
1925  // If ThenBB is actually on the false edge of the conditional branch, remember
1926  // to swap the select operands later.
1927  bool Invert = false;
1928  if (ThenBB != BI->getSuccessor(0)) {
1929  assert(ThenBB == BI->getSuccessor(1) && "No edge from 'if' block?");
1930  Invert = true;
1931  }
1932  assert(EndBB == BI->getSuccessor(!Invert) && "No edge from to end block");
1933 
1934  // Keep a count of how many times instructions are used within CondBB when
1935  // they are candidates for sinking into CondBB. Specifically:
1936  // - They are defined in BB, and
1937  // - They have no side effects, and
1938  // - All of their uses are in CondBB.
1939  SmallDenseMap<Instruction *, unsigned, 4> SinkCandidateUseCounts;
1940 
1941  SmallVector<Instruction *, 4> SpeculatedDbgIntrinsics;
1942 
1943  unsigned SpeculationCost = 0;
1944  Value *SpeculatedStoreValue = nullptr;
1945  StoreInst *SpeculatedStore = nullptr;
1946  for (BasicBlock::iterator BBI = ThenBB->begin(),
1947  BBE = std::prev(ThenBB->end());
1948  BBI != BBE; ++BBI) {
1949  Instruction *I = &*BBI;
1950  // Skip debug info.
1951  if (isa<DbgInfoIntrinsic>(I)) {
1952  SpeculatedDbgIntrinsics.push_back(I);
1953  continue;
1954  }
1955 
1956  // Only speculatively execute a single instruction (not counting the
1957  // terminator) for now.
1958  ++SpeculationCost;
1959  if (SpeculationCost > 1)
1960  return false;
1961 
1962  // Don't hoist the instruction if it's unsafe or expensive.
1963  if (!isSafeToSpeculativelyExecute(I) &&
1964  !(HoistCondStores && (SpeculatedStoreValue = isSafeToSpeculateStore(
1965  I, BB, ThenBB, EndBB))))
1966  return false;
1967  if (!SpeculatedStoreValue &&
1968  ComputeSpeculationCost(I, TTI) >
1970  return false;
1971 
1972  // Store the store speculation candidate.
1973  if (SpeculatedStoreValue)
1974  SpeculatedStore = cast<StoreInst>(I);
1975 
1976  // Do not hoist the instruction if any of its operands are defined but not
1977  // used in BB. The transformation will prevent the operand from
1978  // being sunk into the use block.
1979  for (User::op_iterator i = I->op_begin(), e = I->op_end(); i != e; ++i) {
1980  Instruction *OpI = dyn_cast<Instruction>(*i);
1981  if (!OpI || OpI->getParent() != BB || OpI->mayHaveSideEffects())
1982  continue; // Not a candidate for sinking.
1983 
1984  ++SinkCandidateUseCounts[OpI];
1985  }
1986  }
1987 
1988  // Consider any sink candidates which are only used in CondBB as costs for
1989  // speculation. Note, while we iterate over a DenseMap here, we are summing
1990  // and so iteration order isn't significant.
1992  I = SinkCandidateUseCounts.begin(),
1993  E = SinkCandidateUseCounts.end();
1994  I != E; ++I)
1995  if (I->first->getNumUses() == I->second) {
1996  ++SpeculationCost;
1997  if (SpeculationCost > 1)
1998  return false;
1999  }
2000 
2001  // Check that the PHI nodes can be converted to selects.
2002  bool HaveRewritablePHIs = false;
2003  for (PHINode &PN : EndBB->phis()) {
2004  Value *OrigV = PN.getIncomingValueForBlock(BB);
2005  Value *ThenV = PN.getIncomingValueForBlock(ThenBB);
2006 
2007  // FIXME: Try to remove some of the duplication with HoistThenElseCodeToIf.
2008  // Skip PHIs which are trivial.
2009  if (ThenV == OrigV)
2010  continue;
2011 
2012  // Don't convert to selects if we could remove undefined behavior instead.
2013  if (passingValueIsAlwaysUndefined(OrigV, &PN) ||
2014  passingValueIsAlwaysUndefined(ThenV, &PN))
2015  return false;
2016 
2017  HaveRewritablePHIs = true;
2018  ConstantExpr *OrigCE = dyn_cast<ConstantExpr>(OrigV);
2019  ConstantExpr *ThenCE = dyn_cast<ConstantExpr>(ThenV);
2020  if (!OrigCE && !ThenCE)
2021  continue; // Known safe and cheap.
2022 
2023  if ((ThenCE && !isSafeToSpeculativelyExecute(ThenCE)) ||
2024  (OrigCE && !isSafeToSpeculativelyExecute(OrigCE)))
2025  return false;
2026  unsigned OrigCost = OrigCE ? ComputeSpeculationCost(OrigCE, TTI) : 0;
2027  unsigned ThenCost = ThenCE ? ComputeSpeculationCost(ThenCE, TTI) : 0;
2028  unsigned MaxCost =
2030  if (OrigCost + ThenCost > MaxCost)
2031  return false;
2032 
2033  // Account for the cost of an unfolded ConstantExpr which could end up
2034  // getting expanded into Instructions.
2035  // FIXME: This doesn't account for how many operations are combined in the
2036  // constant expression.
2037  ++SpeculationCost;
2038  if (SpeculationCost > 1)
2039  return false;
2040  }
2041 
2042  // If there are no PHIs to process, bail early. This helps ensure idempotence
2043  // as well.
2044  if (!HaveRewritablePHIs && !(HoistCondStores && SpeculatedStoreValue))
2045  return false;
2046 
2047  // If we get here, we can hoist the instruction and if-convert.
2048  DEBUG(dbgs() << "SPECULATIVELY EXECUTING BB" << *ThenBB << "\n";);
2049 
2050  // Insert a select of the value of the speculated store.
2051  if (SpeculatedStoreValue) {
2052  IRBuilder<NoFolder> Builder(BI);
2053  Value *TrueV = SpeculatedStore->getValueOperand();
2054  Value *FalseV = SpeculatedStoreValue;
2055  if (Invert)
2056  std::swap(TrueV, FalseV);
2057  Value *S = Builder.CreateSelect(
2058  BrCond, TrueV, FalseV, "spec.store.select", BI);
2059  SpeculatedStore->setOperand(0, S);
2060  SpeculatedStore->applyMergedLocation(BI->getDebugLoc(),
2061  SpeculatedStore->getDebugLoc());
2062  }
2063 
2064  // Metadata can be dependent on the condition we are hoisting above.
2065  // Conservatively strip all metadata on the instruction.
2066  for (auto &I : *ThenBB)
2068 
2069  // Hoist the instructions.
2070  BB->getInstList().splice(BI->getIterator(), ThenBB->getInstList(),
2071  ThenBB->begin(), std::prev(ThenBB->end()));
2072 
2073  // Insert selects and rewrite the PHI operands.
2074  IRBuilder<NoFolder> Builder(BI);
2075  for (PHINode &PN : EndBB->phis()) {
2076  unsigned OrigI = PN.getBasicBlockIndex(BB);
2077  unsigned ThenI = PN.getBasicBlockIndex(ThenBB);
2078  Value *OrigV = PN.getIncomingValue(OrigI);
2079  Value *ThenV = PN.getIncomingValue(ThenI);
2080 
2081  // Skip PHIs which are trivial.
2082  if (OrigV == ThenV)
2083  continue;
2084 
2085  // Create a select whose true value is the speculatively executed value and
2086  // false value is the preexisting value. Swap them if the branch
2087  // destinations were inverted.
2088  Value *TrueV = ThenV, *FalseV = OrigV;
2089  if (Invert)
2090  std::swap(TrueV, FalseV);
2091  Value *V = Builder.CreateSelect(
2092  BrCond, TrueV, FalseV, "spec.select", BI);
2093  PN.setIncomingValue(OrigI, V);
2094  PN.setIncomingValue(ThenI, V);
2095  }
2096 
2097  // Remove speculated dbg intrinsics.
2098  // FIXME: Is it possible to do this in a more elegant way? Moving/merging the
2099  // dbg value for the different flows and inserting it after the select.
2100  for (Instruction *I : SpeculatedDbgIntrinsics)
2101  I->eraseFromParent();
2102 
2103  ++NumSpeculations;
2104  return true;
2105 }
2106 
2107 /// Return true if we can thread a branch across this block.
2109  BranchInst *BI = cast<BranchInst>(BB->getTerminator());
2110  unsigned Size = 0;
2111 
2112  for (BasicBlock::iterator BBI = BB->begin(); &*BBI != BI; ++BBI) {
2113  if (isa<DbgInfoIntrinsic>(BBI))
2114  continue;
2115  if (Size > 10)
2116  return false; // Don't clone large BB's.
2117  ++Size;
2118 
2119  // We can only support instructions that do not define values that are
2120  // live outside of the current basic block.
2121  for (User *U : BBI->users()) {
2122  Instruction *UI = cast<Instruction>(U);
2123  if (UI->getParent() != BB || isa<PHINode>(UI))
2124  return false;
2125  }
2126 
2127  // Looks ok, continue checking.
2128  }
2129 
2130  return true;
2131 }
2132 
2133 /// If we have a conditional branch on a PHI node value that is defined in the
2134 /// same block as the branch and if any PHI entries are constants, thread edges
2135 /// corresponding to that entry to be branches to their ultimate destination.
2136 static bool FoldCondBranchOnPHI(BranchInst *BI, const DataLayout &DL,
2137  AssumptionCache *AC) {
2138  BasicBlock *BB = BI->getParent();
2139  PHINode *PN = dyn_cast<PHINode>(BI->getCondition());
2140  // NOTE: we currently cannot transform this case if the PHI node is used
2141  // outside of the block.
2142  if (!PN || PN->getParent() != BB || !PN->hasOneUse())
2143  return false;
2144 
2145  // Degenerate case of a single entry PHI.
2146  if (PN->getNumIncomingValues() == 1) {
2148  return true;
2149  }
2150 
2151  // Now we know that this block has multiple preds and two succs.
2153  return false;
2154 
2155  // Can't fold blocks that contain noduplicate or convergent calls.
2156  if (any_of(*BB, [](const Instruction &I) {
2157  const CallInst *CI = dyn_cast<CallInst>(&I);
2158  return CI && (CI->cannotDuplicate() || CI->isConvergent());
2159  }))
2160  return false;
2161 
2162  // Okay, this is a simple enough basic block. See if any phi values are
2163  // constants.
2164  for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
2166  if (!CB || !CB->getType()->isIntegerTy(1))
2167  continue;
2168 
2169  // Okay, we now know that all edges from PredBB should be revectored to
2170  // branch to RealDest.
2171  BasicBlock *PredBB = PN->getIncomingBlock(i);
2172  BasicBlock *RealDest = BI->getSuccessor(!CB->getZExtValue());
2173 
2174  if (RealDest == BB)
2175  continue; // Skip self loops.
2176  // Skip if the predecessor's terminator is an indirect branch.
2177  if (isa<IndirectBrInst>(PredBB->getTerminator()))
2178  continue;
2179 
2180  // The dest block might have PHI nodes, other predecessors and other
2181  // difficult cases. Instead of being smart about this, just insert a new
2182  // block that jumps to the destination block, effectively splitting
2183  // the edge we are about to create.
2184  BasicBlock *EdgeBB =
2185  BasicBlock::Create(BB->getContext(), RealDest->getName() + ".critedge",
2186  RealDest->getParent(), RealDest);
2187  BranchInst::Create(RealDest, EdgeBB);
2188 
2189  // Update PHI nodes.
2190  AddPredecessorToBlock(RealDest, EdgeBB, BB);
2191 
2192  // BB may have instructions that are being threaded over. Clone these
2193  // instructions into EdgeBB. We know that there will be no uses of the
2194  // cloned instructions outside of EdgeBB.
2195  BasicBlock::iterator InsertPt = EdgeBB->begin();
2196  DenseMap<Value *, Value *> TranslateMap; // Track translated values.
2197  for (BasicBlock::iterator BBI = BB->begin(); &*BBI != BI; ++BBI) {
2198  if (PHINode *PN = dyn_cast<PHINode>(BBI)) {
2199  TranslateMap[PN] = PN->getIncomingValueForBlock(PredBB);
2200  continue;
2201  }
2202  // Clone the instruction.
2203  Instruction *N = BBI->clone();
2204  if (BBI->hasName())
2205  N->setName(BBI->getName() + ".c");
2206 
2207  // Update operands due to translation.
2208  for (User::op_iterator i = N->op_begin(), e = N->op_end(); i != e; ++i) {
2209  DenseMap<Value *, Value *>::iterator PI = TranslateMap.find(*i);
2210  if (PI != TranslateMap.end())
2211  *i = PI->second;
2212  }
2213 
2214  // Check for trivial simplification.
2215  if (Value *V = SimplifyInstruction(N, {DL, nullptr, nullptr, AC})) {
2216  if (!BBI->use_empty())
2217  TranslateMap[&*BBI] = V;
2218  if (!N->mayHaveSideEffects()) {
2219  N->deleteValue(); // Instruction folded away, don't need actual inst
2220  N = nullptr;
2221  }
2222  } else {
2223  if (!BBI->use_empty())
2224  TranslateMap[&*BBI] = N;
2225  }
2226  // Insert the new instruction into its new home.
2227  if (N)
2228  EdgeBB->getInstList().insert(InsertPt, N);
2229 
2230  // Register the new instruction with the assumption cache if necessary.
2231  if (auto *II = dyn_cast_or_null<IntrinsicInst>(N))
2232  if (II->getIntrinsicID() == Intrinsic::assume)
2233  AC->registerAssumption(II);
2234  }
2235 
2236  // Loop over all of the edges from PredBB to BB, changing them to branch
2237  // to EdgeBB instead.
2238  TerminatorInst *PredBBTI = PredBB->getTerminator();
2239  for (unsigned i = 0, e = PredBBTI->getNumSuccessors(); i != e; ++i)
2240  if (PredBBTI->getSuccessor(i) == BB) {
2241  BB->removePredecessor(PredBB);
2242  PredBBTI->setSuccessor(i, EdgeBB);
2243  }
2244 
2245  // Recurse, simplifying any other constants.
2246  return FoldCondBranchOnPHI(BI, DL, AC) | true;
2247  }
2248 
2249  return false;
2250 }
2251 
2252 /// Given a BB that starts with the specified two-entry PHI node,
2253 /// see if we can eliminate it.
2255  const DataLayout &DL) {
2256  // Ok, this is a two entry PHI node. Check to see if this is a simple "if
2257  // statement", which has a very simple dominance structure. Basically, we
2258  // are trying to find the condition that is being branched on, which
2259  // subsequently causes this merge to happen. We really want control
2260  // dependence information for this check, but simplifycfg can't keep it up
2261  // to date, and this catches most of the cases we care about anyway.
2262  BasicBlock *BB = PN->getParent();
2263  BasicBlock *IfTrue, *IfFalse;
2264  Value *IfCond = GetIfCondition(BB, IfTrue, IfFalse);
2265  if (!IfCond ||
2266  // Don't bother if the branch will be constant folded trivially.
2267  isa<ConstantInt>(IfCond))
2268  return false;
2269 
2270  // Okay, we found that we can merge this two-entry phi node into a select.
2271  // Doing so would require us to fold *all* two entry phi nodes in this block.
2272  // At some point this becomes non-profitable (particularly if the target
2273  // doesn't support cmov's). Only do this transformation if there are two or
2274  // fewer PHI nodes in this block.
2275  unsigned NumPhis = 0;
2276  for (BasicBlock::iterator I = BB->begin(); isa<PHINode>(I); ++NumPhis, ++I)
2277  if (NumPhis > 2)
2278  return false;
2279 
2280  // Loop over the PHI's seeing if we can promote them all to select
2281  // instructions. While we are at it, keep track of the instructions
2282  // that need to be moved to the dominating block.
2283  SmallPtrSet<Instruction *, 4> AggressiveInsts;
2284  unsigned MaxCostVal0 = PHINodeFoldingThreshold,
2285  MaxCostVal1 = PHINodeFoldingThreshold;
2286  MaxCostVal0 *= TargetTransformInfo::TCC_Basic;
2287  MaxCostVal1 *= TargetTransformInfo::TCC_Basic;
2288 
2289  for (BasicBlock::iterator II = BB->begin(); isa<PHINode>(II);) {
2290  PHINode *PN = cast<PHINode>(II++);
2291  if (Value *V = SimplifyInstruction(PN, {DL, PN})) {
2292  PN->replaceAllUsesWith(V);
2293  PN->eraseFromParent();
2294  continue;
2295  }
2296 
2297  if (!DominatesMergePoint(PN->getIncomingValue(0), BB, &AggressiveInsts,
2298  MaxCostVal0, TTI) ||
2299  !DominatesMergePoint(PN->getIncomingValue(1), BB, &AggressiveInsts,
2300  MaxCostVal1, TTI))
2301  return false;
2302  }
2303 
2304  // If we folded the first phi, PN dangles at this point. Refresh it. If
2305  // we ran out of PHIs then we simplified them all.
2306  PN = dyn_cast<PHINode>(BB->begin());
2307  if (!PN)
2308  return true;
2309 
2310  // Don't fold i1 branches on PHIs which contain binary operators. These can
2311  // often be turned into switches and other things.
2312  if (PN->getType()->isIntegerTy(1) &&
2313  (isa<BinaryOperator>(PN->getIncomingValue(0)) ||
2314  isa<BinaryOperator>(PN->getIncomingValue(1)) ||
2315  isa<BinaryOperator>(IfCond)))
2316  return false;
2317 
2318  // If all PHI nodes are promotable, check to make sure that all instructions
2319  // in the predecessor blocks can be promoted as well. If not, we won't be able
2320  // to get rid of the control flow, so it's not worth promoting to select
2321  // instructions.
2322  BasicBlock *DomBlock = nullptr;
2323  BasicBlock *IfBlock1 = PN->getIncomingBlock(0);
2324  BasicBlock *IfBlock2 = PN->getIncomingBlock(1);
2325  if (cast<BranchInst>(IfBlock1->getTerminator())->isConditional()) {
2326  IfBlock1 = nullptr;
2327  } else {
2328  DomBlock = *pred_begin(IfBlock1);
2329  for (BasicBlock::iterator I = IfBlock1->begin(); !isa<TerminatorInst>(I);
2330  ++I)
2331  if (!AggressiveInsts.count(&*I) && !isa<DbgInfoIntrinsic>(I)) {
2332  // This is not an aggressive instruction that we can promote.
2333  // Because of this, we won't be able to get rid of the control flow, so
2334  // the xform is not worth it.
2335  return false;
2336  }
2337  }
2338 
2339  if (cast<BranchInst>(IfBlock2->getTerminator())->isConditional()) {
2340  IfBlock2 = nullptr;
2341  } else {
2342  DomBlock = *pred_begin(IfBlock2);
2343  for (BasicBlock::iterator I = IfBlock2->begin(); !isa<TerminatorInst>(I);
2344  ++I)
2345  if (!AggressiveInsts.count(&*I) && !isa<DbgInfoIntrinsic>(I)) {
2346  // This is not an aggressive instruction that we can promote.
2347  // Because of this, we won't be able to get rid of the control flow, so
2348  // the xform is not worth it.
2349  return false;
2350  }
2351  }
2352 
2353  DEBUG(dbgs() << "FOUND IF CONDITION! " << *IfCond << " T: "
2354  << IfTrue->getName() << " F: " << IfFalse->getName() << "\n");
2355 
2356  // If we can still promote the PHI nodes after this gauntlet of tests,
2357  // do all of the PHI's now.
2358  Instruction *InsertPt = DomBlock->getTerminator();
2359  IRBuilder<NoFolder> Builder(InsertPt);
2360 
2361  // Move all 'aggressive' instructions, which are defined in the
2362  // conditional parts of the if's up to the dominating block.
2363  if (IfBlock1) {
2364  for (auto &I : *IfBlock1)
2366  DomBlock->getInstList().splice(InsertPt->getIterator(),
2367  IfBlock1->getInstList(), IfBlock1->begin(),
2368  IfBlock1->getTerminator()->getIterator());
2369  }
2370  if (IfBlock2) {
2371  for (auto &I : *IfBlock2)
2373  DomBlock->getInstList().splice(InsertPt->getIterator(),
2374  IfBlock2->getInstList(), IfBlock2->begin(),
2375  IfBlock2->getTerminator()->getIterator());
2376  }
2377 
2378  while (PHINode *PN = dyn_cast<PHINode>(BB->begin())) {
2379  // Change the PHI node into a select instruction.
2380  Value *TrueVal = PN->getIncomingValue(PN->getIncomingBlock(0) == IfFalse);
2381  Value *FalseVal = PN->getIncomingValue(PN->getIncomingBlock(0) == IfTrue);
2382 
2383  Value *Sel = Builder.CreateSelect(IfCond, TrueVal, FalseVal, "", InsertPt);
2384  PN->replaceAllUsesWith(Sel);
2385  Sel->takeName(PN);
2386  PN->eraseFromParent();
2387  }
2388 
2389  // At this point, IfBlock1 and IfBlock2 are both empty, so our if statement
2390  // has been flattened. Change DomBlock to jump directly to our new block to
2391  // avoid other simplifycfg's kicking in on the diamond.
2392  TerminatorInst *OldTI = DomBlock->getTerminator();
2393  Builder.SetInsertPoint(OldTI);
2394  Builder.CreateBr(BB);
2395  OldTI->eraseFromParent();
2396  return true;
2397 }
2398 
2399 /// If we found a conditional branch that goes to two returning blocks,
2400 /// try to merge them together into one return,
2401 /// introducing a select if the return values disagree.
2403  IRBuilder<> &Builder) {
2404  assert(BI->isConditional() && "Must be a conditional branch");
2405  BasicBlock *TrueSucc = BI->getSuccessor(0);
2406  BasicBlock *FalseSucc = BI->getSuccessor(1);
2407  ReturnInst *TrueRet = cast<ReturnInst>(TrueSucc->getTerminator());
2408  ReturnInst *FalseRet = cast<ReturnInst>(FalseSucc->getTerminator());
2409 
2410  // Check to ensure both blocks are empty (just a return) or optionally empty
2411  // with PHI nodes. If there are other instructions, merging would cause extra
2412  // computation on one path or the other.
2413  if (!TrueSucc->getFirstNonPHIOrDbg()->isTerminator())
2414  return false;
2415  if (!FalseSucc->getFirstNonPHIOrDbg()->isTerminator())
2416  return false;
2417 
2418  Builder.SetInsertPoint(BI);
2419  // Okay, we found a branch that is going to two return nodes. If
2420  // there is no return value for this function, just change the
2421  // branch into a return.
2422  if (FalseRet->getNumOperands() == 0) {
2423  TrueSucc->removePredecessor(BI->getParent());
2424  FalseSucc->removePredecessor(BI->getParent());
2425  Builder.CreateRetVoid();
2427  return true;
2428  }
2429 
2430  // Otherwise, figure out what the true and false return values are
2431  // so we can insert a new select instruction.
2432  Value *TrueValue = TrueRet->getReturnValue();
2433  Value *FalseValue = FalseRet->getReturnValue();
2434 
2435  // Unwrap any PHI nodes in the return blocks.
2436  if (PHINode *TVPN = dyn_cast_or_null<PHINode>(TrueValue))
2437  if (TVPN->getParent() == TrueSucc)
2438  TrueValue = TVPN->getIncomingValueForBlock(BI->getParent());
2439  if (PHINode *FVPN = dyn_cast_or_null<PHINode>(FalseValue))
2440  if (FVPN->getParent() == FalseSucc)
2441  FalseValue = FVPN->getIncomingValueForBlock(BI->getParent());
2442 
2443  // In order for this transformation to be safe, we must be able to
2444  // unconditionally execute both operands to the return. This is
2445  // normally the case, but we could have a potentially-trapping
2446  // constant expression that prevents this transformation from being
2447  // safe.
2448  if (ConstantExpr *TCV = dyn_cast_or_null<ConstantExpr>(TrueValue))
2449  if (TCV->canTrap())
2450  return false;
2451  if (ConstantExpr *FCV = dyn_cast_or_null<ConstantExpr>(FalseValue))
2452  if (FCV->canTrap())
2453  return false;
2454 
2455  // Okay, we collected all the mapped values and checked them for sanity, and
2456  // defined to really do this transformation. First, update the CFG.
2457  TrueSucc->removePredecessor(BI->getParent());
2458  FalseSucc->removePredecessor(BI->getParent());
2459 
2460  // Insert select instructions where needed.
2461  Value *BrCond = BI->getCondition();
2462  if (TrueValue) {
2463  // Insert a select if the results differ.
2464  if (TrueValue == FalseValue || isa<UndefValue>(FalseValue)) {
2465  } else if (isa<UndefValue>(TrueValue)) {
2466  TrueValue = FalseValue;
2467  } else {
2468  TrueValue =
2469  Builder.CreateSelect(BrCond, TrueValue, FalseValue, "retval", BI);
2470  }
2471  }
2472 
2473  Value *RI =
2474  !TrueValue ? Builder.CreateRetVoid() : Builder.CreateRet(TrueValue);
2475 
2476  (void)RI;
2477 
2478  DEBUG(dbgs() << "\nCHANGING BRANCH TO TWO RETURNS INTO SELECT:"
2479  << "\n " << *BI << "NewRet = " << *RI
2480  << "TRUEBLOCK: " << *TrueSucc << "FALSEBLOCK: " << *FalseSucc);
2481 
2483 
2484  return true;
2485 }
2486 
2487 /// Return true if the given instruction is available
2488 /// in its predecessor block. If yes, the instruction will be removed.
2490  if (!isa<BinaryOperator>(Inst) && !isa<CmpInst>(Inst))
2491  return false;
2492  for (Instruction &I : *PB) {
2493  Instruction *PBI = &I;
2494  // Check whether Inst and PBI generate the same value.
2495  if (Inst->isIdenticalTo(PBI)) {
2496  Inst->replaceAllUsesWith(PBI);
2497  Inst->eraseFromParent();
2498  return true;
2499  }
2500  }
2501  return false;
2502 }
2503 
2504 /// Return true if either PBI or BI has branch weight available, and store
2505 /// the weights in {Pred|Succ}{True|False}Weight. If one of PBI and BI does
2506 /// not have branch weight, use 1:1 as its weight.
2508  uint64_t &PredTrueWeight,
2509  uint64_t &PredFalseWeight,
2510  uint64_t &SuccTrueWeight,
2511  uint64_t &SuccFalseWeight) {
2512  bool PredHasWeights =
2513  PBI->extractProfMetadata(PredTrueWeight, PredFalseWeight);
2514  bool SuccHasWeights =
2515  BI->extractProfMetadata(SuccTrueWeight, SuccFalseWeight);
2516  if (PredHasWeights || SuccHasWeights) {
2517  if (!PredHasWeights)
2518  PredTrueWeight = PredFalseWeight = 1;
2519  if (!SuccHasWeights)
2520  SuccTrueWeight = SuccFalseWeight = 1;
2521  return true;
2522  } else {
2523  return false;
2524  }
2525 }
2526 
2527 /// If this basic block is simple enough, and if a predecessor branches to us
2528 /// and one of our successors, fold the block into the predecessor and use
2529 /// logical operations to pick the right destination.
2530 bool llvm::FoldBranchToCommonDest(BranchInst *BI, unsigned BonusInstThreshold) {
2531  BasicBlock *BB = BI->getParent();
2532 
2533  Instruction *Cond = nullptr;
2534  if (BI->isConditional())
2535  Cond = dyn_cast<Instruction>(BI->getCondition());
2536  else {
2537  // For unconditional branch, check for a simple CFG pattern, where
2538  // BB has a single predecessor and BB's successor is also its predecessor's
2539  // successor. If such pattern exists, check for CSE between BB and its
2540  // predecessor.
2541  if (BasicBlock *PB = BB->getSinglePredecessor())
2542  if (BranchInst *PBI = dyn_cast<BranchInst>(PB->getTerminator()))
2543  if (PBI->isConditional() &&
2544  (BI->getSuccessor(0) == PBI->getSuccessor(0) ||
2545  BI->getSuccessor(0) == PBI->getSuccessor(1))) {
2546  for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E;) {
2547  Instruction *Curr = &*I++;
2548  if (isa<CmpInst>(Curr)) {
2549  Cond = Curr;
2550  break;
2551  }
2552  // Quit if we can't remove this instruction.
2553  if (!checkCSEInPredecessor(Curr, PB))
2554  return false;
2555  }
2556  }
2557 
2558  if (!Cond)
2559  return false;
2560  }
2561 
2562  if (!Cond || (!isa<CmpInst>(Cond) && !isa<BinaryOperator>(Cond)) ||
2563  Cond->getParent() != BB || !Cond->hasOneUse())
2564  return false;
2565 
2566  // Make sure the instruction after the condition is the cond branch.
2567  BasicBlock::iterator CondIt = ++Cond->getIterator();
2568 
2569  // Ignore dbg intrinsics.
2570  while (isa<DbgInfoIntrinsic>(CondIt))
2571  ++CondIt;
2572 
2573  if (&*CondIt != BI)
2574  return false;
2575 
2576  // Only allow this transformation if computing the condition doesn't involve
2577  // too many instructions and these involved instructions can be executed
2578  // unconditionally. We denote all involved instructions except the condition
2579  // as "bonus instructions", and only allow this transformation when the
2580  // number of the bonus instructions does not exceed a certain threshold.
2581  unsigned NumBonusInsts = 0;
2582  for (auto I = BB->begin(); Cond != &*I; ++I) {
2583  // Ignore dbg intrinsics.
2584  if (isa<DbgInfoIntrinsic>(I))
2585  continue;
2586  if (!I->hasOneUse() || !isSafeToSpeculativelyExecute(&*I))
2587  return false;
2588  // I has only one use and can be executed unconditionally.
2590  if (User == nullptr || User->getParent() != BB)
2591  return false;
2592  // I is used in the same BB. Since BI uses Cond and doesn't have more slots
2593  // to use any other instruction, User must be an instruction between next(I)
2594  // and Cond.
2595  ++NumBonusInsts;
2596  // Early exits once we reach the limit.
2597  if (NumBonusInsts > BonusInstThreshold)
2598  return false;
2599  }
2600 
2601  // Cond is known to be a compare or binary operator. Check to make sure that
2602  // neither operand is a potentially-trapping constant expression.
2603  if (ConstantExpr *CE = dyn_cast<ConstantExpr>(Cond->getOperand(0)))
2604  if (CE->canTrap())
2605  return false;
2606  if (ConstantExpr *CE = dyn_cast<ConstantExpr>(Cond->getOperand(1)))
2607  if (CE->canTrap())
2608  return false;
2609 
2610  // Finally, don't infinitely unroll conditional loops.
2611  BasicBlock *TrueDest = BI->getSuccessor(0);
2612  BasicBlock *FalseDest = (BI->isConditional()) ? BI->getSuccessor(1) : nullptr;
2613  if (TrueDest == BB || FalseDest == BB)
2614  return false;
2615 
2616  for (pred_iterator PI = pred_begin(BB), E = pred_end(BB); PI != E; ++PI) {
2617  BasicBlock *PredBlock = *PI;
2618  BranchInst *PBI = dyn_cast<BranchInst>(PredBlock->getTerminator());
2619 
2620  // Check that we have two conditional branches. If there is a PHI node in
2621  // the common successor, verify that the same value flows in from both
2622  // blocks.
2624  if (!PBI || PBI->isUnconditional() ||
2625  (BI->isConditional() && !SafeToMergeTerminators(BI, PBI)) ||
2626  (!BI->isConditional() &&
2627  !isProfitableToFoldUnconditional(BI, PBI, Cond, PHIs)))
2628  continue;
2629 
2630  // Determine if the two branches share a common destination.
2631  Instruction::BinaryOps Opc = Instruction::BinaryOpsEnd;
2632  bool InvertPredCond = false;
2633 
2634  if (BI->isConditional()) {
2635  if (PBI->getSuccessor(0) == TrueDest) {
2636  Opc = Instruction::Or;
2637  } else if (PBI->getSuccessor(1) == FalseDest) {
2638  Opc = Instruction::And;
2639  } else if (PBI->getSuccessor(0) == FalseDest) {
2640  Opc = Instruction::And;
2641  InvertPredCond = true;
2642  } else if (PBI->getSuccessor(1) == TrueDest) {
2643  Opc = Instruction::Or;
2644  InvertPredCond = true;
2645  } else {
2646  continue;
2647  }
2648  } else {
2649  if (PBI->getSuccessor(0) != TrueDest && PBI->getSuccessor(1) != TrueDest)
2650  continue;
2651  }
2652 
2653  DEBUG(dbgs() << "FOLDING BRANCH TO COMMON DEST:\n" << *PBI << *BB);
2654  IRBuilder<> Builder(PBI);
2655 
2656  // If we need to invert the condition in the pred block to match, do so now.
2657  if (InvertPredCond) {
2658  Value *NewCond = PBI->getCondition();
2659 
2660  if (NewCond->hasOneUse() && isa<CmpInst>(NewCond)) {
2661  CmpInst *CI = cast<CmpInst>(NewCond);
2662  CI->setPredicate(CI->getInversePredicate());
2663  } else {
2664  NewCond =
2665  Builder.CreateNot(NewCond, PBI->getCondition()->getName() + ".not");
2666  }
2667 
2668  PBI->setCondition(NewCond);
2669  PBI->swapSuccessors();
2670  }
2671 
2672  // If we have bonus instructions, clone them into the predecessor block.
2673  // Note that there may be multiple predecessor blocks, so we cannot move
2674  // bonus instructions to a predecessor block.
2675  ValueToValueMapTy VMap; // maps original values to cloned values
2676  // We already make sure Cond is the last instruction before BI. Therefore,
2677  // all instructions before Cond other than DbgInfoIntrinsic are bonus
2678  // instructions.
2679  for (auto BonusInst = BB->begin(); Cond != &*BonusInst; ++BonusInst) {
2680  if (isa<DbgInfoIntrinsic>(BonusInst))
2681  continue;
2682  Instruction *NewBonusInst = BonusInst->clone();
2683  RemapInstruction(NewBonusInst, VMap,
2685  VMap[&*BonusInst] = NewBonusInst;
2686 
2687  // If we moved a load, we cannot any longer claim any knowledge about
2688  // its potential value. The previous information might have been valid
2689  // only given the branch precondition.
2690  // For an analogous reason, we must also drop all the metadata whose
2691  // semantics we don't understand.
2692  NewBonusInst->dropUnknownNonDebugMetadata();
2693 
2694  PredBlock->getInstList().insert(PBI->getIterator(), NewBonusInst);
2695  NewBonusInst->takeName(&*BonusInst);
2696  BonusInst->setName(BonusInst->getName() + ".old");
2697  }
2698 
2699  // Clone Cond into the predecessor basic block, and or/and the
2700  // two conditions together.
2701  Instruction *New = Cond->clone();
2702  RemapInstruction(New, VMap,
2704  PredBlock->getInstList().insert(PBI->getIterator(), New);
2705  New->takeName(Cond);
2706  Cond->setName(New->getName() + ".old");
2707 
2708  if (BI->isConditional()) {
2709  Instruction *NewCond = cast<Instruction>(
2710  Builder.CreateBinOp(Opc, PBI->getCondition(), New, "or.cond"));
2711  PBI->setCondition(NewCond);
2712 
2713  uint64_t PredTrueWeight, PredFalseWeight, SuccTrueWeight, SuccFalseWeight;
2714  bool HasWeights =
2715  extractPredSuccWeights(PBI, BI, PredTrueWeight, PredFalseWeight,
2716  SuccTrueWeight, SuccFalseWeight);
2717  SmallVector<uint64_t, 8> NewWeights;
2718 
2719  if (PBI->getSuccessor(0) == BB) {
2720  if (HasWeights) {
2721  // PBI: br i1 %x, BB, FalseDest
2722  // BI: br i1 %y, TrueDest, FalseDest
2723  // TrueWeight is TrueWeight for PBI * TrueWeight for BI.
2724  NewWeights.push_back(PredTrueWeight * SuccTrueWeight);
2725  // FalseWeight is FalseWeight for PBI * TotalWeight for BI +
2726  // TrueWeight for PBI * FalseWeight for BI.
2727  // We assume that total weights of a BranchInst can fit into 32 bits.
2728  // Therefore, we will not have overflow using 64-bit arithmetic.
2729  NewWeights.push_back(PredFalseWeight *
2730  (SuccFalseWeight + SuccTrueWeight) +
2731  PredTrueWeight * SuccFalseWeight);
2732  }
2733  AddPredecessorToBlock(TrueDest, PredBlock, BB);
2734  PBI->setSuccessor(0, TrueDest);
2735  }
2736  if (PBI->getSuccessor(1) == BB) {
2737  if (HasWeights) {
2738  // PBI: br i1 %x, TrueDest, BB
2739  // BI: br i1 %y, TrueDest, FalseDest
2740  // TrueWeight is TrueWeight for PBI * TotalWeight for BI +
2741  // FalseWeight for PBI * TrueWeight for BI.
2742  NewWeights.push_back(PredTrueWeight *
2743  (SuccFalseWeight + SuccTrueWeight) +
2744  PredFalseWeight * SuccTrueWeight);
2745  // FalseWeight is FalseWeight for PBI * FalseWeight for BI.
2746  NewWeights.push_back(PredFalseWeight * SuccFalseWeight);
2747  }
2748  AddPredecessorToBlock(FalseDest, PredBlock, BB);
2749  PBI->setSuccessor(1, FalseDest);
2750  }
2751  if (NewWeights.size() == 2) {
2752  // Halve the weights if any of them cannot fit in an uint32_t
2753  FitWeights(NewWeights);
2754 
2755  SmallVector<uint32_t, 8> MDWeights(NewWeights.begin(),
2756  NewWeights.end());
2757  setBranchWeights(PBI, MDWeights[0], MDWeights[1]);
2758  } else
2759  PBI->setMetadata(LLVMContext::MD_prof, nullptr);
2760  } else {
2761  // Update PHI nodes in the common successors.
2762  for (unsigned i = 0, e = PHIs.size(); i != e; ++i) {
2763  ConstantInt *PBI_C = cast<ConstantInt>(
2764  PHIs[i]->getIncomingValueForBlock(PBI->getParent()));
2765  assert(PBI_C->getType()->isIntegerTy(1));
2766  Instruction *MergedCond = nullptr;
2767  if (PBI->getSuccessor(0) == TrueDest) {
2768  // Create (PBI_Cond and PBI_C) or (!PBI_Cond and BI_Value)
2769  // PBI_C is true: PBI_Cond or (!PBI_Cond and BI_Value)
2770  // is false: !PBI_Cond and BI_Value
2771  Instruction *NotCond = cast<Instruction>(
2772  Builder.CreateNot(PBI->getCondition(), "not.cond"));
2773  MergedCond = cast<Instruction>(
2774  Builder.CreateBinOp(Instruction::And, NotCond, New, "and.cond"));
2775  if (PBI_C->isOne())
2776  MergedCond = cast<Instruction>(Builder.CreateBinOp(
2777  Instruction::Or, PBI->getCondition(), MergedCond, "or.cond"));
2778  } else {
2779  // Create (PBI_Cond and BI_Value) or (!PBI_Cond and PBI_C)
2780  // PBI_C is true: (PBI_Cond and BI_Value) or (!PBI_Cond)
2781  // is false: PBI_Cond and BI_Value
2782  MergedCond = cast<Instruction>(Builder.CreateBinOp(
2783  Instruction::And, PBI->getCondition(), New, "and.cond"));
2784  if (PBI_C->isOne()) {
2785  Instruction *NotCond = cast<Instruction>(
2786  Builder.CreateNot(PBI->getCondition(), "not.cond"));
2787  MergedCond = cast<Instruction>(Builder.CreateBinOp(
2788  Instruction::Or, NotCond, MergedCond, "or.cond"));
2789  }
2790  }
2791  // Update PHI Node.
2792  PHIs[i]->setIncomingValue(PHIs[i]->getBasicBlockIndex(PBI->getParent()),
2793  MergedCond);
2794  }
2795  // Change PBI from Conditional to Unconditional.
2796  BranchInst *New_PBI = BranchInst::Create(TrueDest, PBI);
2798  PBI = New_PBI;
2799  }
2800 
2801  // If BI was a loop latch, it may have had associated loop metadata.
2802  // We need to copy it to the new latch, that is, PBI.
2803  if (MDNode *LoopMD = BI->getMetadata(LLVMContext::MD_loop))
2804  PBI->setMetadata(LLVMContext::MD_loop, LoopMD);
2805 
2806  // TODO: If BB is reachable from all paths through PredBlock, then we
2807  // could replace PBI's branch probabilities with BI's.
2808 
2809  // Copy any debug value intrinsics into the end of PredBlock.
2810  for (Instruction &I : *BB)
2811  if (isa<DbgInfoIntrinsic>(I))
2812  I.clone()->insertBefore(PBI);
2813 
2814  return true;
2815  }
2816  return false;
2817 }
2818 
2819 // If there is only one store in BB1 and BB2, return it, otherwise return
2820 // nullptr.
2822  StoreInst *S = nullptr;
2823  for (auto *BB : {BB1, BB2}) {
2824  if (!BB)
2825  continue;
2826  for (auto &I : *BB)
2827  if (auto *SI = dyn_cast<StoreInst>(&I)) {
2828  if (S)
2829  // Multiple stores seen.
2830  return nullptr;
2831  else
2832  S = SI;
2833  }
2834  }
2835  return S;
2836 }
2837 
2839  Value *AlternativeV = nullptr) {
2840  // PHI is going to be a PHI node that allows the value V that is defined in
2841  // BB to be referenced in BB's only successor.
2842  //
2843  // If AlternativeV is nullptr, the only value we care about in PHI is V. It
2844  // doesn't matter to us what the other operand is (it'll never get used). We
2845  // could just create a new PHI with an undef incoming value, but that could
2846  // increase register pressure if EarlyCSE/InstCombine can't fold it with some
2847  // other PHI. So here we directly look for some PHI in BB's successor with V
2848  // as an incoming operand. If we find one, we use it, else we create a new
2849  // one.
2850  //
2851  // If AlternativeV is not nullptr, we care about both incoming values in PHI.
2852  // PHI must be exactly: phi <ty> [ %BB, %V ], [ %OtherBB, %AlternativeV]
2853  // where OtherBB is the single other predecessor of BB's only successor.
2854  PHINode *PHI = nullptr;
2855  BasicBlock *Succ = BB->getSingleSuccessor();
2856 
2857  for (auto I = Succ->begin(); isa<PHINode>(I); ++I)
2858  if (cast<PHINode>(I)->getIncomingValueForBlock(BB) == V) {
2859  PHI = cast<PHINode>(I);
2860  if (!AlternativeV)
2861  break;
2862 
2863  assert(std::distance(pred_begin(Succ), pred_end(Succ)) == 2);
2864  auto PredI = pred_begin(Succ);
2865  BasicBlock *OtherPredBB = *PredI == BB ? *++PredI : *PredI;
2866  if (PHI->getIncomingValueForBlock(OtherPredBB) == AlternativeV)
2867  break;
2868  PHI = nullptr;
2869  }
2870  if (PHI)
2871  return PHI;
2872 
2873  // If V is not an instruction defined in BB, just return it.
2874  if (!AlternativeV &&
2875  (!isa<Instruction>(V) || cast<Instruction>(V)->getParent() != BB))
2876  return V;
2877 
2878  PHI = PHINode::Create(V->getType(), 2, "simplifycfg.merge", &Succ->front());
2879  PHI->addIncoming(V, BB);
2880  for (BasicBlock *PredBB : predecessors(Succ))
2881  if (PredBB != BB)
2882  PHI->addIncoming(
2883  AlternativeV ? AlternativeV : UndefValue::get(V->getType()), PredBB);
2884  return PHI;
2885 }
2886 
2888  BasicBlock *QTB, BasicBlock *QFB,
2889  BasicBlock *PostBB, Value *Address,
2890  bool InvertPCond, bool InvertQCond,
2891  const DataLayout &DL) {
2892  auto IsaBitcastOfPointerType = [](const Instruction &I) {
2893  return Operator::getOpcode(&I) == Instruction::BitCast &&
2894  I.getType()->isPointerTy();
2895  };
2896 
2897  // If we're not in aggressive mode, we only optimize if we have some
2898  // confidence that by optimizing we'll allow P and/or Q to be if-converted.
2899  auto IsWorthwhile = [&](BasicBlock *BB) {
2900  if (!BB)
2901  return true;
2902  // Heuristic: if the block can be if-converted/phi-folded and the
2903  // instructions inside are all cheap (arithmetic/GEPs), it's worthwhile to
2904  // thread this store.
2905  unsigned N = 0;
2906  for (auto &I : *BB) {
2907  // Cheap instructions viable for folding.
2908  if (isa<BinaryOperator>(I) || isa<GetElementPtrInst>(I) ||
2909  isa<StoreInst>(I))
2910  ++N;
2911  // Free instructions.
2912  else if (isa<TerminatorInst>(I) || isa<DbgInfoIntrinsic>(I) ||
2913  IsaBitcastOfPointerType(I))
2914  continue;
2915  else
2916  return false;
2917  }
2918  // The store we want to merge is counted in N, so add 1 to make sure
2919  // we're counting the instructions that would be left.
2920  return N <= (PHINodeFoldingThreshold + 1);
2921  };
2922 
2924  (!IsWorthwhile(PTB) || !IsWorthwhile(PFB) || !IsWorthwhile(QTB) ||
2925  !IsWorthwhile(QFB)))
2926  return false;
2927 
2928  // For every pointer, there must be exactly two stores, one coming from
2929  // PTB or PFB, and the other from QTB or QFB. We don't support more than one
2930  // store (to any address) in PTB,PFB or QTB,QFB.
2931  // FIXME: We could relax this restriction with a bit more work and performance
2932  // testing.
2933  StoreInst *PStore = findUniqueStoreInBlocks(PTB, PFB);
2934  StoreInst *QStore = findUniqueStoreInBlocks(QTB, QFB);
2935  if (!PStore || !QStore)
2936  return false;
2937 
2938  // Now check the stores are compatible.
2939  if (!QStore->isUnordered() || !PStore->isUnordered())
2940  return false;
2941 
2942  // Check that sinking the store won't cause program behavior changes. Sinking
2943  // the store out of the Q blocks won't change any behavior as we're sinking
2944  // from a block to its unconditional successor. But we're moving a store from
2945  // the P blocks down through the middle block (QBI) and past both QFB and QTB.
2946  // So we need to check that there are no aliasing loads or stores in
2947  // QBI, QTB and QFB. We also need to check there are no conflicting memory
2948  // operations between PStore and the end of its parent block.
2949  //
2950  // The ideal way to do this is to query AliasAnalysis, but we don't
2951  // preserve AA currently so that is dangerous. Be super safe and just
2952  // check there are no other memory operations at all.
2953  for (auto &I : *QFB->getSinglePredecessor())
2954  if (I.mayReadOrWriteMemory())
2955  return false;
2956  for (auto &I : *QFB)
2957  if (&I != QStore && I.mayReadOrWriteMemory())
2958  return false;
2959  if (QTB)
2960  for (auto &I : *QTB)
2961  if (&I != QStore && I.mayReadOrWriteMemory())
2962  return false;
2963  for (auto I = BasicBlock::iterator(PStore), E = PStore->getParent()->end();
2964  I != E; ++I)
2965  if (&*I != PStore && I->mayReadOrWriteMemory())
2966  return false;
2967 
2968  // OK, we're going to sink the stores to PostBB. The store has to be
2969  // conditional though, so first create the predicate.
2970  Value *PCond = cast<BranchInst>(PFB->getSinglePredecessor()->getTerminator())
2971  ->getCondition();
2972  Value *QCond = cast<BranchInst>(QFB->getSinglePredecessor()->getTerminator())
2973  ->getCondition();
2974 
2976  PStore->getParent());
2978  QStore->getParent(), PPHI);
2979 
2980  IRBuilder<> QB(&*PostBB->getFirstInsertionPt());
2981 
2982  Value *PPred = PStore->getParent() == PTB ? PCond : QB.CreateNot(PCond);
2983  Value *QPred = QStore->getParent() == QTB ? QCond : QB.CreateNot(QCond);
2984 
2985  if (InvertPCond)
2986  PPred = QB.CreateNot(PPred);
2987  if (InvertQCond)
2988  QPred = QB.CreateNot(QPred);
2989  Value *CombinedPred = QB.CreateOr(PPred, QPred);
2990 
2991  auto *T =
2992  SplitBlockAndInsertIfThen(CombinedPred, &*QB.GetInsertPoint(), false);
2993  QB.SetInsertPoint(T);
2994  StoreInst *SI = cast<StoreInst>(QB.CreateStore(QPHI, Address));
2995  AAMDNodes AAMD;
2996  PStore->getAAMetadata(AAMD, /*Merge=*/false);
2997  PStore->getAAMetadata(AAMD, /*Merge=*/true);
2998  SI->setAAMetadata(AAMD);
2999  unsigned PAlignment = PStore->getAlignment();
3000  unsigned QAlignment = QStore->getAlignment();
3001  unsigned TypeAlignment =
3003  unsigned MinAlignment;
3004  unsigned MaxAlignment;
3005  std::tie(MinAlignment, MaxAlignment) = std::minmax(PAlignment, QAlignment);
3006  // Choose the minimum alignment. If we could prove both stores execute, we
3007  // could use biggest one. In this case, though, we only know that one of the
3008  // stores executes. And we don't know it's safe to take the alignment from a
3009  // store that doesn't execute.
3010  if (MinAlignment != 0) {
3011  // Choose the minimum of all non-zero alignments.
3012  SI->setAlignment(MinAlignment);
3013  } else if (MaxAlignment != 0) {
3014  // Choose the minimal alignment between the non-zero alignment and the ABI
3015  // default alignment for the type of the stored value.
3016  SI->setAlignment(std::min(MaxAlignment, TypeAlignment));
3017  } else {
3018  // If both alignments are zero, use ABI default alignment for the type of
3019  // the stored value.
3020  SI->setAlignment(TypeAlignment);
3021  }
3022 
3023  QStore->eraseFromParent();
3024  PStore->eraseFromParent();
3025 
3026  return true;
3027 }
3028 
3030  const DataLayout &DL) {
3031  // The intention here is to find diamonds or triangles (see below) where each
3032  // conditional block contains a store to the same address. Both of these
3033  // stores are conditional, so they can't be unconditionally sunk. But it may
3034  // be profitable to speculatively sink the stores into one merged store at the
3035  // end, and predicate the merged store on the union of the two conditions of
3036  // PBI and QBI.
3037  //
3038  // This can reduce the number of stores executed if both of the conditions are
3039  // true, and can allow the blocks to become small enough to be if-converted.
3040  // This optimization will also chain, so that ladders of test-and-set
3041  // sequences can be if-converted away.
3042  //
3043  // We only deal with simple diamonds or triangles:
3044  //
3045  // PBI or PBI or a combination of the two
3046  // / \ | \
3047  // PTB PFB | PFB
3048  // \ / | /
3049  // QBI QBI
3050  // / \ | \
3051  // QTB QFB | QFB
3052  // \ / | /
3053  // PostBB PostBB
3054  //
3055  // We model triangles as a type of diamond with a nullptr "true" block.
3056  // Triangles are canonicalized so that the fallthrough edge is represented by
3057  // a true condition, as in the diagram above.
3058  BasicBlock *PTB = PBI->getSuccessor(0);
3059  BasicBlock *PFB = PBI->getSuccessor(1);
3060  BasicBlock *QTB = QBI->getSuccessor(0);
3061  BasicBlock *QFB = QBI->getSuccessor(1);
3062  BasicBlock *PostBB = QFB->getSingleSuccessor();
3063 
3064  // Make sure we have a good guess for PostBB. If QTB's only successor is
3065  // QFB, then QFB is a better PostBB.
3066  if (QTB->getSingleSuccessor() == QFB)
3067  PostBB = QFB;
3068 
3069  // If we couldn't find a good PostBB, stop.
3070  if (!PostBB)
3071  return false;
3072 
3073  bool InvertPCond = false, InvertQCond = false;
3074  // Canonicalize fallthroughs to the true branches.
3075  if (PFB == QBI->getParent()) {
3076  std::swap(PFB, PTB);
3077  InvertPCond = true;
3078  }
3079  if (QFB == PostBB) {
3080  std::swap(QFB, QTB);
3081  InvertQCond = true;
3082  }
3083 
3084  // From this point on we can assume PTB or QTB may be fallthroughs but PFB
3085  // and QFB may not. Model fallthroughs as a nullptr block.
3086  if (PTB == QBI->getParent())
3087  PTB = nullptr;
3088  if (QTB == PostBB)
3089  QTB = nullptr;
3090 
3091  // Legality bailouts. We must have at least the non-fallthrough blocks and
3092  // the post-dominating block, and the non-fallthroughs must only have one
3093  // predecessor.
3094  auto HasOnePredAndOneSucc = [](BasicBlock *BB, BasicBlock *P, BasicBlock *S) {
3095  return BB->getSinglePredecessor() == P && BB->getSingleSuccessor() == S;
3096  };
3097  if (!HasOnePredAndOneSucc(PFB, PBI->getParent(), QBI->getParent()) ||
3098  !HasOnePredAndOneSucc(QFB, QBI->getParent(), PostBB))
3099  return false;
3100  if ((PTB && !HasOnePredAndOneSucc(PTB, PBI->getParent(), QBI->getParent())) ||
3101  (QTB && !HasOnePredAndOneSucc(QTB, QBI->getParent(), PostBB)))
3102  return false;
3103  if (!PostBB->hasNUses(2) || !QBI->getParent()->hasNUses(2))
3104  return false;
3105 
3106  // OK, this is a sequence of two diamonds or triangles.
3107  // Check if there are stores in PTB or PFB that are repeated in QTB or QFB.
3108  SmallPtrSet<Value *, 4> PStoreAddresses, QStoreAddresses;
3109  for (auto *BB : {PTB, PFB}) {
3110  if (!BB)
3111  continue;
3112  for (auto &I : *BB)
3113  if (StoreInst *SI = dyn_cast<StoreInst>(&I))
3114  PStoreAddresses.insert(SI->getPointerOperand());
3115  }
3116  for (auto *BB : {QTB, QFB}) {
3117  if (!BB)
3118  continue;
3119  for (auto &I : *BB)
3120  if (StoreInst *SI = dyn_cast<StoreInst>(&I))
3121  QStoreAddresses.insert(SI->getPointerOperand());
3122  }
3123 
3124  set_intersect(PStoreAddresses, QStoreAddresses);
3125  // set_intersect mutates PStoreAddresses in place. Rename it here to make it
3126  // clear what it contains.
3127  auto &CommonAddresses = PStoreAddresses;
3128 
3129  bool Changed = false;
3130  for (auto *Address : CommonAddresses)
3131  Changed |= mergeConditionalStoreToAddress(
3132  PTB, PFB, QTB, QFB, PostBB, Address, InvertPCond, InvertQCond, DL);
3133  return Changed;
3134 }
3135 
3136 /// If we have a conditional branch as a predecessor of another block,
3137 /// this function tries to simplify it. We know
3138 /// that PBI and BI are both conditional branches, and BI is in one of the
3139 /// successor blocks of PBI - PBI branches to BI.
3141  const DataLayout &DL) {
3142  assert(PBI->isConditional() && BI->isConditional());
3143  BasicBlock *BB = BI->getParent();
3144 
3145  // If this block ends with a branch instruction, and if there is a
3146  // predecessor that ends on a branch of the same condition, make
3147  // this conditional branch redundant.
3148  if (PBI->getCondition() == BI->getCondition() &&
3149  PBI->getSuccessor(0) != PBI->getSuccessor(1)) {
3150  // Okay, the outcome of this conditional branch is statically
3151  // knowable. If this block had a single pred, handle specially.
3152  if (BB->getSinglePredecessor()) {
3153  // Turn this into a branch on constant.
3154  bool CondIsTrue = PBI->getSuccessor(0) == BB;
3155  BI->setCondition(
3156  ConstantInt::get(Type::getInt1Ty(BB->getContext()), CondIsTrue));
3157  return true; // Nuke the branch on constant.
3158  }
3159 
3160  // Otherwise, if there are multiple predecessors, insert a PHI that merges
3161  // in the constant and simplify the block result. Subsequent passes of
3162  // simplifycfg will thread the block.
3164  pred_iterator PB = pred_begin(BB), PE = pred_end(BB);
3165  PHINode *NewPN = PHINode::Create(
3166  Type::getInt1Ty(BB->getContext()), std::distance(PB, PE),
3167  BI->getCondition()->getName() + ".pr", &BB->front());
3168  // Okay, we're going to insert the PHI node. Since PBI is not the only
3169  // predecessor, compute the PHI'd conditional value for all of the preds.
3170  // Any predecessor where the condition is not computable we keep symbolic.
3171  for (pred_iterator PI = PB; PI != PE; ++PI) {
3172  BasicBlock *P = *PI;
3173  if ((PBI = dyn_cast<BranchInst>(P->getTerminator())) && PBI != BI &&
3174  PBI->isConditional() && PBI->getCondition() == BI->getCondition() &&
3175  PBI->getSuccessor(0) != PBI->getSuccessor(1)) {
3176  bool CondIsTrue = PBI->getSuccessor(0) == BB;
3177  NewPN->addIncoming(
3178  ConstantInt::get(Type::getInt1Ty(BB->getContext()), CondIsTrue),
3179  P);
3180  } else {
3181  NewPN->addIncoming(BI->getCondition(), P);
3182  }
3183  }
3184 
3185  BI->setCondition(NewPN);
3186  return true;
3187  }
3188  }
3189 
3190  if (auto *CE = dyn_cast<ConstantExpr>(BI->getCondition()))
3191  if (CE->canTrap())
3192  return false;
3193 
3194  // If both branches are conditional and both contain stores to the same
3195  // address, remove the stores from the conditionals and create a conditional
3196  // merged store at the end.
3197  if (MergeCondStores && mergeConditionalStores(PBI, BI, DL))
3198  return true;
3199 
3200  // If this is a conditional branch in an empty block, and if any
3201  // predecessors are a conditional branch to one of our destinations,
3202  // fold the conditions into logical ops and one cond br.
3203  BasicBlock::iterator BBI = BB->begin();
3204  // Ignore dbg intrinsics.
3205  while (isa<DbgInfoIntrinsic>(BBI))
3206  ++BBI;
3207  if (&*BBI != BI)
3208  return false;
3209 
3210  int PBIOp, BIOp;
3211  if (PBI->getSuccessor(0) == BI->getSuccessor(0)) {
3212  PBIOp = 0;
3213  BIOp = 0;
3214  } else if (PBI->getSuccessor(0) == BI->getSuccessor(1)) {
3215  PBIOp = 0;
3216  BIOp = 1;
3217  } else if (PBI->getSuccessor(1) == BI->getSuccessor(0)) {
3218  PBIOp = 1;
3219  BIOp = 0;
3220  } else if (PBI->getSuccessor(1) == BI->getSuccessor(1)) {
3221  PBIOp = 1;
3222  BIOp = 1;
3223  } else {
3224  return false;
3225  }
3226 
3227  // Check to make sure that the other destination of this branch
3228  // isn't BB itself. If so, this is an infinite loop that will
3229  // keep getting unwound.
3230  if (PBI->getSuccessor(PBIOp) == BB)
3231  return false;
3232 
3233  // Do not perform this transformation if it would require
3234  // insertion of a large number of select instructions. For targets
3235  // without predication/cmovs, this is a big pessimization.
3236 
3237  // Also do not perform this transformation if any phi node in the common
3238  // destination block can trap when reached by BB or PBB (PR17073). In that
3239  // case, it would be unsafe to hoist the operation into a select instruction.
3240 
3241  BasicBlock *CommonDest = PBI->getSuccessor(PBIOp);
3242  unsigned NumPhis = 0;
3243  for (BasicBlock::iterator II = CommonDest->begin(); isa<PHINode>(II);
3244  ++II, ++NumPhis) {
3245  if (NumPhis > 2) // Disable this xform.
3246  return false;
3247 
3248  PHINode *PN = cast<PHINode>(II);
3249  Value *BIV = PN->getIncomingValueForBlock(BB);
3250  if (ConstantExpr *CE = dyn_cast<ConstantExpr>(BIV))
3251  if (CE->canTrap())
3252  return false;
3253 
3254  unsigned PBBIdx = PN->getBasicBlockIndex(PBI->getParent());
3255  Value *PBIV = PN->getIncomingValue(PBBIdx);
3256  if (ConstantExpr *CE = dyn_cast<ConstantExpr>(PBIV))
3257  if (CE->canTrap())
3258  return false;
3259  }
3260 
3261  // Finally, if everything is ok, fold the branches to logical ops.
3262  BasicBlock *OtherDest = BI->getSuccessor(BIOp ^ 1);
3263 
3264  DEBUG(dbgs() << "FOLDING BRs:" << *PBI->getParent()
3265  << "AND: " << *BI->getParent());
3266 
3267  // If OtherDest *is* BB, then BB is a basic block with a single conditional
3268  // branch in it, where one edge (OtherDest) goes back to itself but the other
3269  // exits. We don't *know* that the program avoids the infinite loop
3270  // (even though that seems likely). If we do this xform naively, we'll end up
3271  // recursively unpeeling the loop. Since we know that (after the xform is
3272  // done) that the block *is* infinite if reached, we just make it an obviously
3273  // infinite loop with no cond branch.
3274  if (OtherDest == BB) {
3275  // Insert it at the end of the function, because it's either code,
3276  // or it won't matter if it's hot. :)
3277  BasicBlock *InfLoopBlock =
3278  BasicBlock::Create(BB->getContext(), "infloop", BB->getParent());
3279  BranchInst::Create(InfLoopBlock, InfLoopBlock);
3280  OtherDest = InfLoopBlock;
3281  }
3282 
3283  DEBUG(dbgs() << *PBI->getParent()->getParent());
3284 
3285  // BI may have other predecessors. Because of this, we leave
3286  // it alone, but modify PBI.
3287 
3288  // Make sure we get to CommonDest on True&True directions.
3289  Value *PBICond = PBI->getCondition();
3290  IRBuilder<NoFolder> Builder(PBI);
3291  if (PBIOp)
3292  PBICond = Builder.CreateNot(PBICond, PBICond->getName() + ".not");
3293 
3294  Value *BICond = BI->getCondition();
3295  if (BIOp)
3296  BICond = Builder.CreateNot(BICond, BICond->getName() + ".not");
3297 
3298  // Merge the conditions.
3299  Value *Cond = Builder.CreateOr(PBICond, BICond, "brmerge");
3300 
3301  // Modify PBI to branch on the new condition to the new dests.
3302  PBI->setCondition(Cond);
3303  PBI->setSuccessor(0, CommonDest);
3304  PBI->setSuccessor(1, OtherDest);
3305 
3306  // Update branch weight for PBI.
3307  uint64_t PredTrueWeight, PredFalseWeight, SuccTrueWeight, SuccFalseWeight;
3308  uint64_t PredCommon, PredOther, SuccCommon, SuccOther;
3309  bool HasWeights =
3310  extractPredSuccWeights(PBI, BI, PredTrueWeight, PredFalseWeight,
3311  SuccTrueWeight, SuccFalseWeight);
3312  if (HasWeights) {
3313  PredCommon = PBIOp ? PredFalseWeight : PredTrueWeight;
3314  PredOther = PBIOp ? PredTrueWeight : PredFalseWeight;
3315  SuccCommon = BIOp ? SuccFalseWeight : SuccTrueWeight;
3316  SuccOther = BIOp ? SuccTrueWeight : SuccFalseWeight;
3317  // The weight to CommonDest should be PredCommon * SuccTotal +
3318  // PredOther * SuccCommon.
3319  // The weight to OtherDest should be PredOther * SuccOther.
3320  uint64_t NewWeights[2] = {PredCommon * (SuccCommon + SuccOther) +
3321  PredOther * SuccCommon,
3322  PredOther * SuccOther};
3323  // Halve the weights if any of them cannot fit in an uint32_t
3324  FitWeights(NewWeights);
3325 
3326  setBranchWeights(PBI, NewWeights[0], NewWeights[1]);
3327  }
3328 
3329  // OtherDest may have phi nodes. If so, add an entry from PBI's
3330  // block that are identical to the entries for BI's block.
3331  AddPredecessorToBlock(OtherDest, PBI->getParent(), BB);
3332 
3333  // We know that the CommonDest already had an edge from PBI to
3334  // it. If it has PHIs though, the PHIs may have different
3335  // entries for BB and PBI's BB. If so, insert a select to make
3336  // them agree.
3337  for (PHINode &PN : CommonDest->phis()) {
3338  Value *BIV = PN.getIncomingValueForBlock(BB);
3339  unsigned PBBIdx = PN.getBasicBlockIndex(PBI->getParent());
3340  Value *PBIV = PN.getIncomingValue(PBBIdx);
3341  if (BIV != PBIV) {
3342  // Insert a select in PBI to pick the right value.
3343  SelectInst *NV = cast<SelectInst>(
3344  Builder.CreateSelect(PBICond, PBIV, BIV, PBIV->getName() + ".mux"));
3345  PN.setIncomingValue(PBBIdx, NV);
3346  // Although the select has the same condition as PBI, the original branch
3347  // weights for PBI do not apply to the new select because the select's
3348  // 'logical' edges are incoming edges of the phi that is eliminated, not
3349  // the outgoing edges of PBI.
3350  if (HasWeights) {
3351  uint64_t PredCommon = PBIOp ? PredFalseWeight : PredTrueWeight;
3352  uint64_t PredOther = PBIOp ? PredTrueWeight : PredFalseWeight;
3353  uint64_t SuccCommon = BIOp ? SuccFalseWeight : SuccTrueWeight;
3354  uint64_t SuccOther = BIOp ? SuccTrueWeight : SuccFalseWeight;
3355  // The weight to PredCommonDest should be PredCommon * SuccTotal.
3356  // The weight to PredOtherDest should be PredOther * SuccCommon.
3357  uint64_t NewWeights[2] = {PredCommon * (SuccCommon + SuccOther),
3358  PredOther * SuccCommon};
3359 
3360  FitWeights(NewWeights);
3361 
3362  setBranchWeights(NV, NewWeights[0], NewWeights[1]);
3363  }
3364  }
3365  }
3366 
3367  DEBUG(dbgs() << "INTO: " << *PBI->getParent());
3368  DEBUG(dbgs() << *PBI->getParent()->getParent());
3369 
3370  // This basic block is probably dead. We know it has at least
3371  // one fewer predecessor.
3372  return true;
3373 }
3374 
3375 // Simplifies a terminator by replacing it with a branch to TrueBB if Cond is
3376 // true or to FalseBB if Cond is false.
3377 // Takes care of updating the successors and removing the old terminator.
3378 // Also makes sure not to introduce new successors by assuming that edges to
3379 // non-successor TrueBBs and FalseBBs aren't reachable.
3381  BasicBlock *TrueBB, BasicBlock *FalseBB,
3382  uint32_t TrueWeight,
3383  uint32_t FalseWeight) {
3384  // Remove any superfluous successor edges from the CFG.
3385  // First, figure out which successors to preserve.
3386  // If TrueBB and FalseBB are equal, only try to preserve one copy of that
3387  // successor.
3388  BasicBlock *KeepEdge1 = TrueBB;
3389  BasicBlock *KeepEdge2 = TrueBB != FalseBB ? FalseBB : nullptr;
3390 
3391  // Then remove the rest.
3392  for (BasicBlock *Succ : OldTerm->successors()) {
3393  // Make sure only to keep exactly one copy of each edge.
3394  if (Succ == KeepEdge1)
3395  KeepEdge1 = nullptr;
3396  else if (Succ == KeepEdge2)
3397  KeepEdge2 = nullptr;
3398  else
3399  Succ->removePredecessor(OldTerm->getParent(),
3400  /*DontDeleteUselessPHIs=*/true);
3401  }
3402 
3403  IRBuilder<> Builder(OldTerm);
3404  Builder.SetCurrentDebugLocation(OldTerm->getDebugLoc());
3405 
3406  // Insert an appropriate new terminator.
3407  if (!KeepEdge1 && !KeepEdge2) {
3408  if (TrueBB == FalseBB)
3409  // We were only looking for one successor, and it was present.
3410  // Create an unconditional branch to it.
3411  Builder.CreateBr(TrueBB);
3412  else {
3413  // We found both of the successors we were looking for.
3414  // Create a conditional branch sharing the condition of the select.
3415  BranchInst *NewBI = Builder.CreateCondBr(Cond, TrueBB, FalseBB);
3416  if (TrueWeight != FalseWeight)
3417  setBranchWeights(NewBI, TrueWeight, FalseWeight);
3418  }
3419  } else if (KeepEdge1 && (KeepEdge2 || TrueBB == FalseBB)) {
3420  // Neither of the selected blocks were successors, so this
3421  // terminator must be unreachable.
3422  new UnreachableInst(OldTerm->getContext(), OldTerm);
3423  } else {
3424  // One of the selected values was a successor, but the other wasn't.
3425  // Insert an unconditional branch to the one that was found;
3426  // the edge to the one that wasn't must be unreachable.
3427  if (!KeepEdge1)
3428  // Only TrueBB was found.
3429  Builder.CreateBr(TrueBB);
3430  else
3431  // Only FalseBB was found.
3432  Builder.CreateBr(FalseBB);
3433  }
3434 
3436  return true;
3437 }
3438 
3439 // Replaces
3440 // (switch (select cond, X, Y)) on constant X, Y
3441 // with a branch - conditional if X and Y lead to distinct BBs,
3442 // unconditional otherwise.
3444  // Check for constant integer values in the select.
3445  ConstantInt *TrueVal = dyn_cast<ConstantInt>(Select->getTrueValue());
3446  ConstantInt *FalseVal = dyn_cast<ConstantInt>(Select->getFalseValue());
3447  if (!TrueVal || !FalseVal)
3448  return false;
3449 
3450  // Find the relevant condition and destinations.
3451  Value *Condition = Select->getCondition();
3452  BasicBlock *TrueBB = SI->findCaseValue(TrueVal)->getCaseSuccessor();
3453  BasicBlock *FalseBB = SI->findCaseValue(FalseVal)->getCaseSuccessor();
3454 
3455  // Get weight for TrueBB and FalseBB.
3456  uint32_t TrueWeight = 0, FalseWeight = 0;
3457  SmallVector<uint64_t, 8> Weights;
3458  bool HasWeights = HasBranchWeights(SI);
3459  if (HasWeights) {
3460  GetBranchWeights(SI, Weights);
3461  if (Weights.size() == 1 + SI->getNumCases()) {
3462  TrueWeight =
3463  (uint32_t)Weights[SI->findCaseValue(TrueVal)->getSuccessorIndex()];
3464  FalseWeight =
3465  (uint32_t)Weights[SI->findCaseValue(FalseVal)->getSuccessorIndex()];
3466  }
3467  }
3468 
3469  // Perform the actual simplification.
3470  return SimplifyTerminatorOnSelect(SI, Condition, TrueBB, FalseBB, TrueWeight,
3471  FalseWeight);
3472 }
3473 
3474 // Replaces
3475 // (indirectbr (select cond, blockaddress(@fn, BlockA),
3476 // blockaddress(@fn, BlockB)))
3477 // with
3478 // (br cond, BlockA, BlockB).
3480  // Check that both operands of the select are block addresses.
3483  if (!TBA || !FBA)
3484  return false;
3485 
3486  // Extract the actual blocks.
3487  BasicBlock *TrueBB = TBA->getBasicBlock();
3488  BasicBlock *FalseBB = FBA->getBasicBlock();
3489 
3490  // Perform the actual simplification.
3491  return SimplifyTerminatorOnSelect(IBI, SI->getCondition(), TrueBB, FalseBB, 0,
3492  0);
3493 }
3494 
3495 /// This is called when we find an icmp instruction
3496 /// (a seteq/setne with a constant) as the only instruction in a
3497 /// block that ends with an uncond branch. We are looking for a very specific
3498 /// pattern that occurs when "A == 1 || A == 2 || A == 3" gets simplified. In
3499 /// this case, we merge the first two "or's of icmp" into a switch, but then the
3500 /// default value goes to an uncond block with a seteq in it, we get something
3501 /// like:
3502 ///
3503 /// switch i8 %A, label %DEFAULT [ i8 1, label %end i8 2, label %end ]
3504 /// DEFAULT:
3505 /// %tmp = icmp eq i8 %A, 92
3506 /// br label %end
3507 /// end:
3508 /// ... = phi i1 [ true, %entry ], [ %tmp, %DEFAULT ], [ true, %entry ]
3509 ///
3510 /// We prefer to split the edge to 'end' so that there is a true/false entry to
3511 /// the PHI, merging the third icmp into the switch.
3513  ICmpInst *ICI, IRBuilder<> &Builder, const DataLayout &DL,
3514  const TargetTransformInfo &TTI, const SimplifyCFGOptions &Options) {
3515  BasicBlock *BB = ICI->getParent();
3516 
3517  // If the block has any PHIs in it or the icmp has multiple uses, it is too
3518  // complex.
3519  if (isa<PHINode>(BB->begin()) || !ICI->hasOneUse())
3520  return false;
3521 
3522  Value *V = ICI->getOperand(0);
3523  ConstantInt *Cst = cast<ConstantInt>(ICI->getOperand(1));
3524 
3525  // The pattern we're looking for is where our only predecessor is a switch on
3526  // 'V' and this block is the default case for the switch. In this case we can
3527  // fold the compared value into the switch to simplify things.
3528  BasicBlock *Pred = BB->getSinglePredecessor();
3529  if (!Pred || !isa<SwitchInst>(Pred->getTerminator()))
3530  return false;
3531 
3532  SwitchInst *SI = cast<SwitchInst>(Pred->getTerminator());
3533  if (SI->getCondition() != V)
3534  return false;
3535 
3536  // If BB is reachable on a non-default case, then we simply know the value of
3537  // V in this block. Substitute it and constant fold the icmp instruction
3538  // away.
3539  if (SI->getDefaultDest() != BB) {
3540  ConstantInt *VVal = SI->findCaseDest(BB);
3541  assert(VVal && "Should have a unique destination value");
3542  ICI->setOperand(0, VVal);
3543 
3544  if (Value *V = SimplifyInstruction(ICI, {DL, ICI})) {
3545  ICI->replaceAllUsesWith(V);
3546  ICI->eraseFromParent();
3547  }
3548  // BB is now empty, so it is likely to simplify away.
3549  return simplifyCFG(BB, TTI, Options) | true;
3550  }
3551 
3552  // Ok, the block is reachable from the default dest. If the constant we're
3553  // comparing exists in one of the other edges, then we can constant fold ICI
3554  // and zap it.
3555  if (SI->findCaseValue(Cst) != SI->case_default()) {
3556  Value *V;
3557  if (ICI->getPredicate() == ICmpInst::ICMP_EQ)
3558  V = ConstantInt::getFalse(BB->getContext());
3559  else
3560  V = ConstantInt::getTrue(BB->getContext());
3561 
3562  ICI->replaceAllUsesWith(V);
3563  ICI->eraseFromParent();
3564  // BB is now empty, so it is likely to simplify away.
3565  return simplifyCFG(BB, TTI, Options) | true;
3566  }
3567 
3568  // The use of the icmp has to be in the 'end' block, by the only PHI node in
3569  // the block.
3570  BasicBlock *SuccBlock = BB->getTerminator()->getSuccessor(0);
3571  PHINode *PHIUse = dyn_cast<PHINode>(ICI->user_back());
3572  if (PHIUse == nullptr || PHIUse != &SuccBlock->front() ||
3573  isa<PHINode>(++BasicBlock::iterator(PHIUse)))
3574  return false;
3575 
3576  // If the icmp is a SETEQ, then the default dest gets false, the new edge gets
3577  // true in the PHI.
3578  Constant *DefaultCst = ConstantInt::getTrue(BB->getContext());
3579  Constant *NewCst = ConstantInt::getFalse(BB->getContext());
3580 
3581  if (ICI->getPredicate() == ICmpInst::ICMP_EQ)
3582  std::swap(DefaultCst, NewCst);
3583 
3584  // Replace ICI (which is used by the PHI for the default value) with true or
3585  // false depending on if it is EQ or NE.
3586  ICI->replaceAllUsesWith(DefaultCst);
3587  ICI->eraseFromParent();
3588 
3589  // Okay, the switch goes to this block on a default value. Add an edge from
3590  // the switch to the merge point on the compared value.
3591  BasicBlock *NewBB =
3592  BasicBlock::Create(BB->getContext(), "switch.edge", BB->getParent(), BB);
3593  SmallVector<uint64_t, 8> Weights;
3594  bool HasWeights = HasBranchWeights(SI);
3595  if (HasWeights) {
3596  GetBranchWeights(SI, Weights);
3597  if (Weights.size() == 1 + SI->getNumCases()) {
3598  // Split weight for default case to case for "Cst".
3599  Weights[0] = (Weights[0] + 1) >> 1;
3600  Weights.push_back(Weights[0]);
3601 
3602  SmallVector<uint32_t, 8> MDWeights(Weights.begin(), Weights.end());
3603  setBranchWeights(SI, MDWeights);
3604  }
3605  }
3606  SI->addCase(Cst, NewBB);
3607 
3608  // NewBB branches to the phi block, add the uncond branch and the phi entry.
3609  Builder.SetInsertPoint(NewBB);
3610  Builder.SetCurrentDebugLocation(SI->getDebugLoc());
3611  Builder.CreateBr(SuccBlock);
3612  PHIUse->addIncoming(NewCst, NewBB);
3613  return true;
3614 }
3615 
3616 /// The specified branch is a conditional branch.
3617 /// Check to see if it is branching on an or/and chain of icmp instructions, and
3618 /// fold it into a switch instruction if so.
3620  const DataLayout &DL) {
3621  Instruction *Cond = dyn_cast<Instruction>(BI->getCondition());
3622  if (!Cond)
3623  return false;
3624 
3625  // Change br (X == 0 | X == 1), T, F into a switch instruction.
3626  // If this is a bunch of seteq's or'd together, or if it's a bunch of
3627  // 'setne's and'ed together, collect them.
3628 
3629  // Try to gather values from a chain of and/or to be turned into a switch
3630  ConstantComparesGatherer ConstantCompare(Cond, DL);
3631  // Unpack the result
3632  SmallVectorImpl<ConstantInt *> &Values = ConstantCompare.Vals;
3633  Value *CompVal = ConstantCompare.CompValue;
3634  unsigned UsedICmps = ConstantCompare.UsedICmps;
3635  Value *ExtraCase = ConstantCompare.Extra;
3636 
3637  // If we didn't have a multiply compared value, fail.
3638  if (!CompVal)
3639  return false;
3640 
3641  // Avoid turning single icmps into a switch.
3642  if (UsedICmps <= 1)
3643  return false;
3644 
3645  bool TrueWhenEqual = (Cond->getOpcode() == Instruction::Or);
3646 
3647  // There might be duplicate constants in the list, which the switch
3648  // instruction can't handle, remove them now.
3649  array_pod_sort(Values.begin(), Values.end(), ConstantIntSortPredicate);
3650  Values.erase(std::unique(Values.begin(), Values.end()), Values.end());
3651 
3652  // If Extra was used, we require at least two switch values to do the
3653  // transformation. A switch with one value is just a conditional branch.
3654  if (ExtraCase && Values.size() < 2)
3655  return false;
3656 
3657  // TODO: Preserve branch weight metadata, similarly to how
3658  // FoldValueComparisonIntoPredecessors preserves it.
3659 
3660  // Figure out which block is which destination.
3661  BasicBlock *DefaultBB = BI->getSuccessor(1);
3662  BasicBlock *EdgeBB = BI->getSuccessor(0);
3663  if (!TrueWhenEqual)
3664  std::swap(DefaultBB, EdgeBB);
3665 
3666  BasicBlock *BB = BI->getParent();
3667 
3668  DEBUG(dbgs() << "Converting 'icmp' chain with " << Values.size()
3669  << " cases into SWITCH. BB is:\n"
3670  << *BB);
3671 
3672  // If there are any extra values that couldn't be folded into the switch
3673  // then we evaluate them with an explicit branch first. Split the block
3674  // right before the condbr to handle it.
3675  if (ExtraCase) {
3676  BasicBlock *NewBB =
3677  BB->splitBasicBlock(BI->getIterator(), "switch.early.test");
3678  // Remove the uncond branch added to the old block.
3679  TerminatorInst *OldTI = BB->getTerminator();
3680  Builder.SetInsertPoint(OldTI);
3681 
3682  if (TrueWhenEqual)
3683  Builder.CreateCondBr(ExtraCase, EdgeBB, NewBB);
3684  else
3685  Builder.CreateCondBr(ExtraCase, NewBB, EdgeBB);
3686 
3687  OldTI->eraseFromParent();
3688 
3689  // If there are PHI nodes in EdgeBB, then we need to add a new entry to them
3690  // for the edge we just added.
3691  AddPredecessorToBlock(EdgeBB, BB, NewBB);
3692 
3693  DEBUG(dbgs() << " ** 'icmp' chain unhandled condition: " << *ExtraCase
3694  << "\nEXTRABB = " << *BB);
3695  BB = NewBB;
3696  }
3697 
3698  Builder.SetInsertPoint(BI);
3699  // Convert pointer to int before we switch.
3700  if (CompVal->getType()->isPointerTy()) {
3701  CompVal = Builder.CreatePtrToInt(
3702  CompVal, DL.getIntPtrType(CompVal->getType()), "magicptr");
3703  }
3704 
3705  // Create the new switch instruction now.
3706  SwitchInst *New = Builder.CreateSwitch(CompVal, DefaultBB, Values.size());
3707 
3708  // Add all of the 'cases' to the switch instruction.
3709  for (unsigned i = 0, e = Values.size(); i != e; ++i)
3710  New->addCase(Values[i], EdgeBB);
3711 
3712  // We added edges from PI to the EdgeBB. As such, if there were any
3713  // PHI nodes in EdgeBB, they need entries to be added corresponding to
3714  // the number of edges added.
3715  for (BasicBlock::iterator BBI = EdgeBB->begin(); isa<PHINode>(BBI); ++BBI) {
3716  PHINode *PN = cast<PHINode>(BBI);
3717  Value *InVal = PN->getIncomingValueForBlock(BB);
3718  for (unsigned i = 0, e = Values.size() - 1; i != e; ++i)
3719  PN->addIncoming(InVal, BB);
3720  }
3721 
3722  // Erase the old branch instruction.
3724 
3725  DEBUG(dbgs() << " ** 'icmp' chain result is:\n" << *BB << '\n');
3726  return true;
3727 }
3728 
3729 bool SimplifyCFGOpt::SimplifyResume(ResumeInst *RI, IRBuilder<> &Builder) {
3730  if (isa<PHINode>(RI->getValue()))
3731  return SimplifyCommonResume(RI);
3732  else if (isa<LandingPadInst>(RI->getParent()->getFirstNonPHI()) &&
3733  RI->getValue() == RI->getParent()->getFirstNonPHI())
3734  // The resume must unwind the exception that caused control to branch here.
3735  return SimplifySingleResume(RI);
3736 
3737  return false;
3738 }
3739 
3740 // Simplify resume that is shared by several landing pads (phi of landing pad).
3741 bool SimplifyCFGOpt::SimplifyCommonResume(ResumeInst *RI) {
3742  BasicBlock *BB = RI->getParent();
3743 
3744  // Check that there are no other instructions except for debug intrinsics
3745  // between the phi of landing pads (RI->getValue()) and resume instruction.
3746  BasicBlock::iterator I = cast<Instruction>(RI->getValue())->getIterator(),
3747  E = RI->getIterator();
3748  while (++I != E)
3749  if (!isa<DbgInfoIntrinsic>(I))
3750  return false;
3751 
3752  SmallSetVector<BasicBlock *, 4> TrivialUnwindBlocks;
3753  auto *PhiLPInst = cast<PHINode>(RI->getValue());
3754 
3755  // Check incoming blocks to see if any of them are trivial.
3756  for (unsigned Idx = 0, End = PhiLPInst->getNumIncomingValues(); Idx != End;
3757  Idx++) {
3758  auto *IncomingBB = PhiLPInst->getIncomingBlock(Idx);
3759  auto *IncomingValue = PhiLPInst->getIncomingValue(Idx);
3760 
3761  // If the block has other successors, we can not delete it because
3762  // it has other dependents.
3763  if (IncomingBB->getUniqueSuccessor() != BB)
3764  continue;
3765 
3766  auto *LandingPad = dyn_cast<LandingPadInst>(IncomingBB->getFirstNonPHI());
3767  // Not the landing pad that caused the control to branch here.
3768  if (IncomingValue != LandingPad)
3769  continue;
3770 
3771  bool isTrivial = true;
3772 
3773  I = IncomingBB->getFirstNonPHI()->getIterator();
3774  E = IncomingBB->getTerminator()->getIterator();
3775  while (++I != E)
3776  if (!isa<DbgInfoIntrinsic>(I)) {
3777  isTrivial = false;
3778  break;
3779  }
3780 
3781  if (isTrivial)
3782  TrivialUnwindBlocks.insert(IncomingBB);
3783  }
3784 
3785  // If no trivial unwind blocks, don't do any simplifications.
3786  if (TrivialUnwindBlocks.empty())
3787  return false;
3788 
3789  // Turn all invokes that unwind here into calls.
3790  for (auto *TrivialBB : TrivialUnwindBlocks) {
3791  // Blocks that will be simplified should be removed from the phi node.
3792  // Note there could be multiple edges to the resume block, and we need
3793  // to remove them all.
3794  while (PhiLPInst->getBasicBlockIndex(TrivialBB) != -1)
3795  BB->removePredecessor(TrivialBB, true);
3796 
3797  for (pred_iterator PI = pred_begin(TrivialBB), PE = pred_end(TrivialBB);
3798  PI != PE;) {
3799  BasicBlock *Pred = *PI++;
3800  removeUnwindEdge(Pred);
3801  }
3802 
3803  // In each SimplifyCFG run, only the current processed block can be erased.
3804  // Otherwise, it will break the iteration of SimplifyCFG pass. So instead
3805  // of erasing TrivialBB, we only remove the branch to the common resume
3806  // block so that we can later erase the resume block since it has no
3807  // predecessors.
3808  TrivialBB->getTerminator()->eraseFromParent();
3809  new UnreachableInst(RI->getContext(), TrivialBB);
3810  }
3811 
3812  // Delete the resume block if all its predecessors have been removed.
3813  if (pred_empty(BB))
3814  BB->eraseFromParent();
3815 
3816  return !TrivialUnwindBlocks.empty();
3817 }
3818 
3819 // Simplify resume that is only used by a single (non-phi) landing pad.
3820 bool SimplifyCFGOpt::SimplifySingleResume(ResumeInst *RI) {
3821  BasicBlock *BB = RI->getParent();
3823  assert(RI->getValue() == LPInst &&
3824  "Resume must unwind the exception that caused control to here");
3825 
3826  // Check that there are no other instructions except for debug intrinsics.
3827  BasicBlock::iterator I = LPInst->getIterator(), E = RI->getIterator();
3828  while (++I != E)
3829  if (!isa<DbgInfoIntrinsic>(I))
3830  return false;
3831 
3832  // Turn all invokes that unwind here into calls and delete the basic block.
3833  for (pred_iterator PI = pred_begin(BB), PE = pred_end(BB); PI != PE;) {
3834  BasicBlock *Pred = *PI++;
3835  removeUnwindEdge(Pred);
3836  }
3837 
3838  // The landingpad is now unreachable. Zap it.
3839  BB->eraseFromParent();
3840  if (LoopHeaders)
3841  LoopHeaders->erase(BB);
3842  return true;
3843 }
3844 
3846  // If this is a trivial cleanup pad that executes no instructions, it can be
3847  // eliminated. If the cleanup pad continues to the caller, any predecessor
3848  // that is an EH pad will be updated to continue to the caller and any
3849  // predecessor that terminates with an invoke instruction will have its invoke
3850  // instruction converted to a call instruction. If the cleanup pad being
3851  // simplified does not continue to the caller, each predecessor will be
3852  // updated to continue to the unwind destination of the cleanup pad being
3853  // simplified.
3854  BasicBlock *BB = RI->getParent();
3855  CleanupPadInst *CPInst = RI->getCleanupPad();
3856  if (CPInst->getParent() != BB)
3857  // This isn't an empty cleanup.
3858  return false;
3859 
3860  // We cannot kill the pad if it has multiple uses. This typically arises
3861  // from unreachable basic blocks.
3862  if (!CPInst->hasOneUse())
3863  return false;
3864 
3865  // Check that there are no other instructions except for benign intrinsics.
3866  BasicBlock::iterator I = CPInst->getIterator(), E = RI->getIterator();
3867  while (++I != E) {
3868  auto *II = dyn_cast<IntrinsicInst>(I);
3869  if (!II)
3870  return false;
3871 
3872  Intrinsic::ID IntrinsicID = II->getIntrinsicID();
3873  switch (IntrinsicID) {
3874  case Intrinsic::dbg_declare:
3875  case Intrinsic::dbg_value:
3876  case Intrinsic::lifetime_end:
3877  break;
3878  default:
3879  return false;
3880  }
3881  }
3882 
3883  // If the cleanup return we are simplifying unwinds to the caller, this will
3884  // set UnwindDest to nullptr.
3885  BasicBlock *UnwindDest = RI->getUnwindDest();
3886  Instruction *DestEHPad = UnwindDest ? UnwindDest->getFirstNonPHI() : nullptr;
3887 
3888  // We're about to remove BB from the control flow. Before we do, sink any
3889  // PHINodes into the unwind destination. Doing this before changing the
3890  // control flow avoids some potentially slow checks, since we can currently
3891  // be certain that UnwindDest and BB have no common predecessors (since they
3892  // are both EH pads).
3893  if (UnwindDest) {
3894  // First, go through the PHI nodes in UnwindDest and update any nodes that
3895  // reference the block we are removing
3896  for (BasicBlock::iterator I = UnwindDest->begin(),
3897  IE = DestEHPad->getIterator();
3898  I != IE; ++I) {
3899  PHINode *DestPN = cast<PHINode>(I);
3900 
3901  int Idx = DestPN->getBasicBlockIndex(BB);
3902  // Since BB unwinds to UnwindDest, it has to be in the PHI node.
3903  assert(Idx != -1);
3904  // This PHI node has an incoming value that corresponds to a control
3905  // path through the cleanup pad we are removing. If the incoming
3906  // value is in the cleanup pad, it must be a PHINode (because we
3907  // verified above that the block is otherwise empty). Otherwise, the
3908  // value is either a constant or a value that dominates the cleanup
3909  // pad being removed.
3910  //
3911  // Because BB and UnwindDest are both EH pads, all of their
3912  // predecessors must unwind to these blocks, and since no instruction
3913  // can have multiple unwind destinations, there will be no overlap in
3914  // incoming blocks between SrcPN and DestPN.
3915  Value *SrcVal = DestPN->getIncomingValue(Idx);
3916  PHINode *SrcPN = dyn_cast<PHINode>(SrcVal);
3917 
3918  // Remove the entry for the block we are deleting.
3919  DestPN->removeIncomingValue(Idx, false);
3920 
3921  if (SrcPN && SrcPN->getParent() == BB) {
3922  // If the incoming value was a PHI node in the cleanup pad we are
3923  // removing, we need to merge that PHI node's incoming values into
3924  // DestPN.
3925  for (unsigned SrcIdx = 0, SrcE = SrcPN->getNumIncomingValues();
3926  SrcIdx != SrcE; ++SrcIdx) {
3927  DestPN->addIncoming(SrcPN->getIncomingValue(SrcIdx),
3928  SrcPN->getIncomingBlock(SrcIdx));
3929  }
3930  } else {
3931  // Otherwise, the incoming value came from above BB and
3932  // so we can just reuse it. We must associate all of BB's
3933  // predecessors with this value.
3934  for (auto *pred : predecessors(BB)) {
3935  DestPN->addIncoming(SrcVal, pred);
3936  }
3937  }
3938  }
3939 
3940  // Sink any remaining PHI nodes directly into UnwindDest.
3941  Instruction *InsertPt = DestEHPad;
3942  for (BasicBlock::iterator I = BB->begin(),
3943  IE = BB->getFirstNonPHI()->getIterator();
3944  I != IE;) {
3945  // The iterator must be incremented here because the instructions are
3946  // being moved to another block.
3947  PHINode *PN = cast<PHINode>(I++);
3948  if (PN->use_empty())
3949  // If the PHI node has no uses, just leave it. It will be erased
3950  // when we erase BB below.
3951  continue;
3952 
3953  // Otherwise, sink this PHI node into UnwindDest.
3954  // Any predecessors to UnwindDest which are not already represented
3955  // must be back edges which inherit the value from the path through
3956  // BB. In this case, the PHI value must reference itself.
3957  for (auto *pred : predecessors(UnwindDest))
3958  if (pred != BB)
3959  PN->addIncoming(PN, pred);
3960  PN->moveBefore(InsertPt);
3961  }
3962  }
3963 
3964  for (pred_iterator PI = pred_begin(BB), PE = pred_end(BB); PI != PE;) {
3965  // The iterator must be updated here because we are removing this pred.
3966  BasicBlock *PredBB = *PI++;
3967  if (UnwindDest == nullptr) {
3968  removeUnwindEdge(PredBB);
3969  } else {
3970  TerminatorInst *TI = PredBB->getTerminator();
3971  TI->replaceUsesOfWith(BB, UnwindDest);
3972  }
3973  }
3974 
3975  // The cleanup pad is now unreachable. Zap it.
3976  BB->eraseFromParent();
3977  return true;
3978 }
3979 
3980 // Try to merge two cleanuppads together.
3982  // Skip any cleanuprets which unwind to caller, there is nothing to merge
3983  // with.
3984  BasicBlock *UnwindDest = RI->getUnwindDest();
3985  if (!UnwindDest)
3986  return false;
3987 
3988  // This cleanupret isn't the only predecessor of this cleanuppad, it wouldn't
3989  // be safe to merge without code duplication.
3990  if (UnwindDest->getSinglePredecessor() != RI->getParent())
3991  return false;
3992 
3993  // Verify that our cleanuppad's unwind destination is another cleanuppad.
3994  auto *SuccessorCleanupPad = dyn_cast<CleanupPadInst>(&UnwindDest->front());
3995  if (!SuccessorCleanupPad)
3996  return false;
3997 
3998  CleanupPadInst *PredecessorCleanupPad = RI->getCleanupPad();
3999  // Replace any uses of the successor cleanupad with the predecessor pad
4000  // The only cleanuppad uses should be this cleanupret, it's cleanupret and
4001  // funclet bundle operands.
4002  SuccessorCleanupPad->replaceAllUsesWith(PredecessorCleanupPad);
4003  // Remove the old cleanuppad.
4004  SuccessorCleanupPad->eraseFromParent();
4005  // Now, we simply replace the cleanupret with a branch to the unwind
4006  // destination.
4007  BranchInst::Create(UnwindDest, RI->getParent());
4008  RI->eraseFromParent();
4009 
4010  return true;
4011 }
4012 
4013 bool SimplifyCFGOpt::SimplifyCleanupReturn(CleanupReturnInst *RI) {
4014  // It is possible to transiantly have an undef cleanuppad operand because we
4015  // have deleted some, but not all, dead blocks.
4016  // Eventually, this block will be deleted.
4017  if (isa<UndefValue>(RI->getOperand(0)))
4018  return false;
4019 
4020  if (mergeCleanupPad(RI))
4021  return true;
4022 
4023  if (removeEmptyCleanup(RI))
4024  return true;
4025 
4026  return false;
4027 }
4028 
4029 bool SimplifyCFGOpt::SimplifyReturn(ReturnInst *RI, IRBuilder<> &Builder) {
4030  BasicBlock *BB = RI->getParent();
4031  if (!BB->getFirstNonPHIOrDbg()->isTerminator())
4032  return false;
4033 
4034  // Find predecessors that end with branches.
4035  SmallVector<BasicBlock *, 8> UncondBranchPreds;
4036  SmallVector<BranchInst *, 8> CondBranchPreds;
4037  for (pred_iterator PI = pred_begin(BB), E = pred_end(BB); PI != E; ++PI) {
4038  BasicBlock *P = *PI;
4039  TerminatorInst *PTI = P->getTerminator();
4040  if (BranchInst *BI = dyn_cast<BranchInst>(PTI)) {
4041  if (BI->isUnconditional())
4042  UncondBranchPreds.push_back(P);
4043  else
4044  CondBranchPreds.push_back(BI);
4045  }
4046  }
4047 
4048  // If we found some, do the transformation!
4049  if (!UncondBranchPreds.empty() && DupRet) {
4050  while (!UncondBranchPreds.empty()) {
4051  BasicBlock *Pred = UncondBranchPreds.pop_back_val();
4052  DEBUG(dbgs() << "FOLDING: " << *BB
4053  << "INTO UNCOND BRANCH PRED: " << *Pred);
4054  (void)FoldReturnIntoUncondBranch(RI, BB, Pred);
4055  }
4056 
4057  // If we eliminated all predecessors of the block, delete the block now.
4058  if (pred_empty(BB)) {
4059  // We know there are no successors, so just nuke the block.
4060  BB->eraseFromParent();
4061  if (LoopHeaders)
4062  LoopHeaders->erase(BB);
4063  }
4064 
4065  return true;
4066  }
4067 
4068  // Check out all of the conditional branches going to this return
4069  // instruction. If any of them just select between returns, change the
4070  // branch itself into a select/return pair.
4071  while (!CondBranchPreds.empty()) {
4072  BranchInst *BI = CondBranchPreds.pop_back_val();
4073 
4074  // Check to see if the non-BB successor is also a return block.
4075  if (isa<ReturnInst>(BI->getSuccessor(0)->getTerminator()) &&
4076  isa<ReturnInst>(BI->getSuccessor(1)->getTerminator()) &&
4077  SimplifyCondBranchToTwoReturns(BI, Builder))
4078  return true;
4079  }
4080  return false;
4081 }
4082 
4083 bool SimplifyCFGOpt::SimplifyUnreachable(UnreachableInst *UI) {
4084  BasicBlock *BB = UI->getParent();
4085 
4086  bool Changed = false;
4087 
4088  // If there are any instructions immediately before the unreachable that can
4089  // be removed, do so.
4090  while (UI->getIterator() != BB->begin()) {
4091  BasicBlock::iterator BBI = UI->getIterator();
4092  --BBI;
4093  // Do not delete instructions that can have side effects which might cause
4094  // the unreachable to not be reachable; specifically, calls and volatile
4095  // operations may have this effect.
4096  if (isa<CallInst>(BBI) && !isa<DbgInfoIntrinsic>(BBI))
4097  break;
4098 
4099  if (BBI->mayHaveSideEffects()) {
4100  if (auto *SI = dyn_cast<StoreInst>(BBI)) {
4101  if (SI->isVolatile())
4102  break;
4103  } else if (auto *LI = dyn_cast<LoadInst>(BBI)) {
4104  if (LI->isVolatile())
4105  break;
4106  } else if (auto *RMWI = dyn_cast<AtomicRMWInst>(BBI)) {
4107  if (RMWI->isVolatile())
4108  break;
4109  } else if (auto *CXI = dyn_cast<AtomicCmpXchgInst>(BBI)) {
4110  if (CXI->isVolatile())
4111  break;
4112  } else if (isa<CatchPadInst>(BBI)) {
4113  // A catchpad may invoke exception object constructors and such, which
4114  // in some languages can be arbitrary code, so be conservative by
4115  // default.
4116  // For CoreCLR, it just involves a type test, so can be removed.
4119  break;
4120  } else if (!isa<FenceInst>(BBI) && !isa<VAArgInst>(BBI) &&
4121  !isa<LandingPadInst>(BBI)) {
4122  break;
4123  }
4124  // Note that deleting LandingPad's here is in fact okay, although it
4125  // involves a bit of subtle reasoning. If this inst is a LandingPad,
4126  // all the predecessors of this block will be the unwind edges of Invokes,
4127  // and we can therefore guarantee this block will be erased.
4128  }
4129 
4130  // Delete this instruction (any uses are guaranteed to be dead)
4131  if (!BBI->use_empty())
4132  BBI->replaceAllUsesWith(UndefValue::get(BBI->getType()));
4133  BBI->eraseFromParent();
4134  Changed = true;
4135  }
4136 
4137  // If the unreachable instruction is the first in the block, take a gander
4138  // at all of the predecessors of this instruction, and simplify them.
4139  if (&BB->front() != UI)
4140  return Changed;
4141 
4143  for (unsigned i = 0, e = Preds.size(); i != e; ++i) {
4144  TerminatorInst *TI = Preds[i]->getTerminator();
4145  IRBuilder<> Builder(TI);
4146  if (auto *BI = dyn_cast<BranchInst>(TI)) {
4147  if (BI->isUnconditional()) {
4148  if (BI->getSuccessor(0) == BB) {
4149  new UnreachableInst(TI->getContext(), TI);
4150  TI->eraseFromParent();
4151  Changed = true;
4152  }
4153  } else {
4154  if (BI->getSuccessor(0) == BB) {
4155  Builder.CreateBr(BI->getSuccessor(1));
4157  } else if (BI->getSuccessor(1) == BB) {
4158  Builder.CreateBr(BI->getSuccessor(0));
4160  Changed = true;
4161  }
4162  }
4163  } else if (auto *SI = dyn_cast<SwitchInst>(TI)) {
4164  for (auto i = SI->case_begin(), e = SI->case_end(); i != e;) {
4165  if (i->getCaseSuccessor() != BB) {
4166  ++i;
4167  continue;
4168  }
4169  BB->removePredecessor(SI->getParent());
4170  i = SI->removeCase(i);
4171  e = SI->case_end();
4172  Changed = true;
4173  }
4174  } else if (auto *II = dyn_cast<InvokeInst>(TI)) {
4175  if (II->getUnwindDest() == BB) {
4176  removeUnwindEdge(TI->getParent());
4177  Changed = true;
4178  }
4179  } else if (auto *CSI = dyn_cast<CatchSwitchInst>(TI)) {
4180  if (CSI->getUnwindDest() == BB) {
4181  removeUnwindEdge(TI->getParent());
4182  Changed = true;
4183  continue;
4184  }
4185 
4186  for (CatchSwitchInst::handler_iterator I = CSI->handler_begin(),
4187  E = CSI->handler_end();
4188  I != E; ++I) {
4189  if (*I == BB) {
4190  CSI->removeHandler(I);
4191  --I;
4192  --E;
4193  Changed = true;
4194  }
4195  }
4196  if (CSI->getNumHandlers() == 0) {
4197  BasicBlock *CatchSwitchBB = CSI->getParent();
4198  if (CSI->hasUnwindDest()) {
4199  // Redirect preds to the unwind dest
4200  CatchSwitchBB->replaceAllUsesWith(CSI->getUnwindDest());
4201  } else {
4202  // Rewrite all preds to unwind to caller (or from invoke to call).
4203  SmallVector<BasicBlock *, 8> EHPreds(predecessors(CatchSwitchBB));
4204  for (BasicBlock *EHPred : EHPreds)
4205  removeUnwindEdge(EHPred);
4206  }
4207  // The catchswitch is no longer reachable.
4208  new UnreachableInst(CSI->getContext(), CSI);
4209  CSI->eraseFromParent();
4210  Changed = true;
4211  }
4212  } else if (isa<CleanupReturnInst>(TI)) {
4213  new UnreachableInst(TI->getContext(), TI);
4214  TI->eraseFromParent();
4215  Changed = true;
4216  }
4217  }
4218 
4219  // If this block is now dead, remove it.
4220  if (pred_empty(BB) && BB != &BB->getParent()->getEntryBlock()) {
4221  // We know there are no successors, so just nuke the block.
4222  BB->eraseFromParent();
4223  if (LoopHeaders)
4224  LoopHeaders->erase(BB);
4225  return true;
4226  }
4227 
4228  return Changed;
4229 }
4230 
4232  assert(Cases.size() >= 1);
4233 
4234  array_pod_sort(Cases.begin(), Cases.end(), ConstantIntSortPredicate);
4235  for (size_t I = 1, E = Cases.size(); I != E; ++I) {
4236  if (Cases[I - 1]->getValue() != Cases[I]->getValue() + 1)
4237  return false;
4238  }
4239  return true;
4240 }
4241 
4242 /// Turn a switch with two reachable destinations into an integer range
4243 /// comparison and branch.
4245  assert(SI->getNumCases() > 1 && "Degenerate switch?");
4246 
4247  bool HasDefault =
4248  !isa<UnreachableInst>(SI->getDefaultDest()->getFirstNonPHIOrDbg());
4249 
4250  // Partition the cases into two sets with different destinations.
4251  BasicBlock *DestA = HasDefault ? SI->getDefaultDest() : nullptr;
4252  BasicBlock *DestB = nullptr;
4255 
4256  for (auto Case : SI->cases()) {
4257  BasicBlock *Dest = Case.getCaseSuccessor();
4258  if (!DestA)
4259  DestA = Dest;
4260  if (Dest == DestA) {
4261  CasesA.push_back(Case.getCaseValue());
4262  continue;
4263  }
4264  if (!DestB)
4265  DestB = Dest;
4266  if (Dest == DestB) {
4267  CasesB.push_back(Case.getCaseValue());
4268  continue;
4269  }
4270  return false; // More than two destinations.
4271  }
4272 
4273  assert(DestA && DestB &&
4274  "Single-destination switch should have been folded.");
4275  assert(DestA != DestB);
4276  assert(DestB != SI->getDefaultDest());
4277  assert(!CasesB.empty() && "There must be non-default cases.");
4278  assert(!CasesA.empty() || HasDefault);
4279 
4280  // Figure out if one of the sets of cases form a contiguous range.
4281  SmallVectorImpl<ConstantInt *> *ContiguousCases = nullptr;
4282  BasicBlock *ContiguousDest = nullptr;
4283  BasicBlock *OtherDest = nullptr;
4284  if (!CasesA.empty() && CasesAreContiguous(CasesA)) {
4285  ContiguousCases = &CasesA;
4286  ContiguousDest = DestA;
4287  OtherDest = DestB;
4288  } else if (CasesAreContiguous(CasesB)) {
4289  ContiguousCases = &CasesB;
4290  ContiguousDest = DestB;
4291  OtherDest = DestA;
4292  } else
4293  return false;
4294 
4295  // Start building the compare and branch.
4296 
4297  Constant *Offset = ConstantExpr::getNeg(ContiguousCases->back());
4298  Constant *NumCases =
4299  ConstantInt::get(Offset->getType(), ContiguousCases->size());
4300 
4301  Value *Sub = SI->getCondition();
4302  if (!Offset->isNullValue())
4303  Sub = Builder.CreateAdd(Sub, Offset, Sub->getName() + ".off");
4304 
4305  Value *Cmp;
4306  // If NumCases overflowed, then all possible values jump to the successor.
4307  if (NumCases->isNullValue() && !ContiguousCases->empty())
4308  Cmp = ConstantInt::getTrue(SI->getContext());
4309  else
4310  Cmp = Builder.CreateICmpULT(Sub, NumCases, "switch");
4311  BranchInst *NewBI = Builder.CreateCondBr(Cmp, ContiguousDest, OtherDest);
4312 
4313  // Update weight for the newly-created conditional branch.
4314  if (HasBranchWeights(SI)) {
4315  SmallVector<uint64_t, 8> Weights;
4316  GetBranchWeights(SI, Weights);
4317  if (Weights.size() == 1 + SI->getNumCases()) {
4318  uint64_t TrueWeight = 0;
4319  uint64_t FalseWeight = 0;
4320  for (size_t I = 0, E = Weights.size(); I != E; ++I) {
4321  if (SI->getSuccessor(I) == ContiguousDest)
4322  TrueWeight += Weights[I];
4323  else
4324  FalseWeight += Weights[I];
4325  }
4326  while (TrueWeight > UINT32_MAX || FalseWeight > UINT32_MAX) {
4327  TrueWeight /= 2;
4328  FalseWeight /= 2;
4329  }
4330  setBranchWeights(NewBI, TrueWeight, FalseWeight);
4331  }
4332  }
4333 
4334  // Prune obsolete incoming values off the successors' PHI nodes.
4335  for (auto BBI = ContiguousDest->begin(); isa<PHINode>(BBI); ++BBI) {
4336  unsigned PreviousEdges = ContiguousCases->size();
4337  if (ContiguousDest == SI->getDefaultDest())
4338  ++PreviousEdges;
4339  for (unsigned I = 0, E = PreviousEdges - 1; I != E; ++I)
4340  cast<PHINode>(BBI)->removeIncomingValue(SI->getParent());
4341  }
4342  for (auto BBI = OtherDest->begin(); isa<PHINode>(BBI); ++BBI) {
4343  unsigned PreviousEdges = SI->getNumCases() - ContiguousCases->size();
4344  if (OtherDest == SI->getDefaultDest())
4345  ++PreviousEdges;
4346  for (unsigned I = 0, E = PreviousEdges - 1; I != E; ++I)
4347  cast<PHINode>(BBI)->removeIncomingValue(SI->getParent());
4348  }
4349 
4350  // Drop the switch.
4351  SI->eraseFromParent();
4352 
4353  return true;
4354 }
4355 
4356 /// Compute masked bits for the condition of a switch
4357 /// and use it to remove dead cases.
4359  const DataLayout &DL) {
4360  Value *Cond = SI->getCondition();
4361  unsigned Bits = Cond->getType()->getIntegerBitWidth();
4362  KnownBits Known = computeKnownBits(Cond, DL, 0, AC, SI);
4363 
4364  // We can also eliminate cases by determining that their values are outside of
4365  // the limited range of the condition based on how many significant (non-sign)
4366  // bits are in the condition value.
4367  unsigned ExtraSignBits = ComputeNumSignBits(Cond, DL, 0, AC, SI) - 1;
4368  unsigned MaxSignificantBitsInCond = Bits - ExtraSignBits;
4369 
4370  // Gather dead cases.
4372  for (auto &Case : SI->cases()) {
4373  const APInt &CaseVal = Case.getCaseValue()->getValue();
4374  if (Known.Zero.intersects(CaseVal) || !Known.One.isSubsetOf(CaseVal) ||
4375  (CaseVal.getMinSignedBits() > MaxSignificantBitsInCond)) {
4376  DeadCases.push_back(Case.getCaseValue());
4377  DEBUG(dbgs() << "SimplifyCFG: switch case " << CaseVal << " is dead.\n");
4378  }
4379  }
4380 
4381  // If we can prove that the cases must cover all possible values, the
4382  // default destination becomes dead and we can remove it. If we know some
4383  // of the bits in the value, we can use that to more precisely compute the
4384  // number of possible unique case values.
4385  bool HasDefault =
4386  !isa<UnreachableInst>(SI->getDefaultDest()->getFirstNonPHIOrDbg());
4387  const unsigned NumUnknownBits =
4388  Bits - (Known.Zero | Known.One).countPopulation();
4389  assert(NumUnknownBits <= Bits);
4390  if (HasDefault && DeadCases.empty() &&
4391  NumUnknownBits < 64 /* avoid overflow */ &&
4392  SI->getNumCases() == (1ULL << NumUnknownBits)) {
4393  DEBUG(dbgs() << "SimplifyCFG: switch default is dead.\n");
4394  BasicBlock *NewDefault =
4396  SI->setDefaultDest(&*NewDefault);
4397  SplitBlock(&*NewDefault, &NewDefault->front());
4398  auto *OldTI = NewDefault->getTerminator();
4399  new UnreachableInst(SI->getContext(), OldTI);
4401  return true;
4402  }
4403 
4404  SmallVector<uint64_t, 8> Weights;
4405  bool HasWeight = HasBranchWeights(SI);
4406  if (HasWeight) {
4407  GetBranchWeights(SI, Weights);
4408  HasWeight = (Weights.size() == 1 + SI->getNumCases());
4409  }
4410 
4411  // Remove dead cases from the switch.
4412  for (ConstantInt *DeadCase : DeadCases) {
4413  SwitchInst::CaseIt CaseI = SI->findCaseValue(DeadCase);
4414  assert(CaseI != SI->case_default() &&
4415  "Case was not found. Probably mistake in DeadCases forming.");
4416  if (HasWeight) {
4417  std::swap(Weights[CaseI->getCaseIndex() + 1], Weights.back());
4418  Weights.pop_back();
4419  }
4420 
4421  // Prune unused values from PHI nodes.
4422  CaseI->getCaseSuccessor()->removePredecessor(SI->getParent());
4423  SI->removeCase(CaseI);
4424  }
4425  if (HasWeight && Weights.size() >= 2) {
4426  SmallVector<uint32_t, 8> MDWeights(Weights.begin(), Weights.end());
4427  setBranchWeights(SI, MDWeights);
4428  }
4429 
4430  return !DeadCases.empty();
4431 }
4432 
4433 /// If BB would be eligible for simplification by
4434 /// TryToSimplifyUncondBranchFromEmptyBlock (i.e. it is empty and terminated
4435 /// by an unconditional branch), look at the phi node for BB in the successor
4436 /// block and see if the incoming value is equal to CaseValue. If so, return
4437 /// the phi node, and set PhiIndex to BB's index in the phi node.
4439  BasicBlock *BB, int *PhiIndex) {
4440  if (BB->getFirstNonPHIOrDbg() != BB->getTerminator())
4441  return nullptr; // BB must be empty to be a candidate for simplification.
4442  if (!BB->getSinglePredecessor())
4443  return nullptr; // BB must be dominated by the switch.
4444 
4446  if (!Branch || !Branch->isUnconditional())
4447  return nullptr; // Terminator must be unconditional branch.
4448 
4449  BasicBlock *Succ = Branch->getSuccessor(0);
4450 
4451  for (PHINode &PHI : Succ->phis()) {
4452  int Idx = PHI.getBasicBlockIndex(BB);
4453  assert(Idx >= 0 && "PHI has no entry for predecessor?");
4454 
4455  Value *InValue = PHI.getIncomingValue(Idx);
4456  if (InValue != CaseValue)
4457  continue;
4458 
4459  *PhiIndex = Idx;
4460  return &PHI;
4461  }
4462 
4463  return nullptr;
4464 }
4465 
4466 /// Try to forward the condition of a switch instruction to a phi node
4467 /// dominated by the switch, if that would mean that some of the destination
4468 /// blocks of the switch can be folded away. Return true if a change is made.
4470  using ForwardingNodesMap = DenseMap<PHINode *, SmallVector<int, 4>>;
4471 
4472  ForwardingNodesMap ForwardingNodes;
4473  BasicBlock *SwitchBlock = SI->getParent();
4474  bool Changed = false;
4475  for (auto &Case : SI->cases()) {
4476  ConstantInt *CaseValue = Case.getCaseValue();
4477  BasicBlock *CaseDest = Case.getCaseSuccessor();
4478 
4479  // Replace phi operands in successor blocks that are using the constant case
4480  // value rather than the switch condition variable:
4481  // switchbb:
4482  // switch i32 %x, label %default [
4483  // i32 17, label %succ
4484  // ...
4485  // succ:
4486  // %r = phi i32 ... [ 17, %switchbb ] ...
4487  // -->
4488  // %r = phi i32 ... [ %x, %switchbb ] ...
4489 
4490  for (PHINode &Phi : CaseDest->phis()) {
4491  // This only works if there is exactly 1 incoming edge from the switch to
4492  // a phi. If there is >1, that means multiple cases of the switch map to 1
4493  // value in the phi, and that phi value is not the switch condition. Thus,
4494  // this transform would not make sense (the phi would be invalid because
4495  // a phi can't have different incoming values from the same block).
4496  int SwitchBBIdx = Phi.getBasicBlockIndex(SwitchBlock);
4497  if (Phi.getIncomingValue(SwitchBBIdx) == CaseValue &&
4498  count(Phi.blocks(), SwitchBlock) == 1) {
4499  Phi.setIncomingValue(SwitchBBIdx, SI->getCondition());
4500  Changed = true;
4501  }
4502  }
4503 
4504  // Collect phi nodes that are indirectly using this switch's case constants.
4505  int PhiIdx;
4506  if (auto *Phi = FindPHIForConditionForwarding(CaseValue, CaseDest, &PhiIdx))
4507  ForwardingNodes[Phi].push_back(PhiIdx);
4508  }
4509 
4510  for (auto &ForwardingNode : ForwardingNodes) {
4511  PHINode *Phi = ForwardingNode.first;
4512  SmallVectorImpl<int> &Indexes = ForwardingNode.second;
4513  if (Indexes.size() < 2)
4514  continue;
4515 
4516  for (int Index : Indexes)
4517  Phi->setIncomingValue(Index, SI->getCondition());
4518  Changed = true;
4519  }
4520 
4521  return Changed;
4522 }
4523 
4524 /// Return true if the backend will be able to handle
4525 /// initializing an array of constants like C.
4527  if (C->isThreadDependent())
4528  return false;
4529  if (C->isDLLImportDependent())
4530  return false;
4531 
4532  if (!isa<ConstantFP>(C) && !isa<ConstantInt>(C) &&
4533  !isa<ConstantPointerNull>(C) && !isa<GlobalValue>(C) &&
4534  !isa<UndefValue>(C) && !isa<ConstantExpr>(C))
4535  return false;
4536 
4537  if (ConstantExpr *CE = dyn_cast<ConstantExpr>(C)) {
4538  if (!CE->isGEPWithNoNotionalOverIndexing())
4539  return false;
4540  if (!ValidLookupTableConstant(CE->getOperand(0), TTI))
4541  return false;
4542  }
4543 
4545  return false;
4546 
4547  return true;
4548 }
4549 
4550 /// If V is a Constant, return it. Otherwise, try to look up
4551 /// its constant value in ConstantPool, returning 0 if it's not there.
4552 static Constant *
4555  if (Constant *C = dyn_cast<Constant>(V))
4556  return C;
4557  return ConstantPool.lookup(V);
4558 }
4559 
4560 /// Try to fold instruction I into a constant. This works for
4561 /// simple instructions such as binary operations where both operands are
4562 /// constant or can be replaced by constants from the ConstantPool. Returns the
4563 /// resulting constant on success, 0 otherwise.
4564 static Constant *
4567  if (SelectInst *Select = dyn_cast<SelectInst>(I)) {
4568  Constant *A = LookupConstant(Select->getCondition(), ConstantPool);
4569  if (!A)
4570  return nullptr;
4571  if (A->isAllOnesValue())
4572  return LookupConstant(Select->getTrueValue(), ConstantPool);
4573  if (A->isNullValue())
4574  return LookupConstant(Select->getFalseValue(), ConstantPool);
4575  return nullptr;
4576  }
4577 
4579  for (unsigned N = 0, E = I->getNumOperands(); N != E; ++N) {
4581  COps.push_back(A);
4582  else
4583  return nullptr;
4584  }
4585 
4586  if (CmpInst *Cmp = dyn_cast<CmpInst>(I)) {
4587  return ConstantFoldCompareInstOperands(Cmp->getPredicate(), COps[0],
4588  COps[1], DL);
4589  }
4590 
4591  return ConstantFoldInstOperands(I, COps, DL);
4592 }
4593 
4594 /// Try to determine the resulting constant values in phi nodes
4595 /// at the common destination basic block, *CommonDest, for one of the case
4596 /// destionations CaseDest corresponding to value CaseVal (0 for the default
4597 /// case), of a switch instruction SI.
4598 static bool
4600  BasicBlock **CommonDest,
4601  SmallVectorImpl<std::pair<PHINode *, Constant *>> &Res,
4602  const DataLayout &DL, const TargetTransformInfo &TTI) {
4603  // The block from which we enter the common destination.
4604  BasicBlock *Pred = SI->getParent();
4605 
4606  // If CaseDest is empty except for some side-effect free instructions through
4607  // which we can constant-propagate the CaseVal, continue to its successor.
4609  ConstantPool.insert(std::make_pair(SI->getCondition(), CaseVal));
4610  for (BasicBlock::iterator I = CaseDest->begin(), E = CaseDest->end(); I != E;
4611  ++I) {
4612  if (TerminatorInst *T = dyn_cast<TerminatorInst>(I)) {
4613  // If the terminator is a simple branch, continue to the next block.
4614  if (T->getNumSuccessors() != 1 || T->isExceptional())
4615  return false;
4616  Pred = CaseDest;
4617  CaseDest = T->getSuccessor(0);
4618  } else if (isa<DbgInfoIntrinsic>(I)) {
4619  // Skip debug intrinsic.
4620  continue;
4621  } else if (Constant *C = ConstantFold(&*I, DL, ConstantPool)) {
4622  // Instruction is side-effect free and constant.
4623 
4624  // If the instruction has uses outside this block or a phi node slot for
4625  // the block, it is not safe to bypass the instruction since it would then
4626  // no longer dominate all its uses.
4627  for (auto &Use : I->uses()) {
4628  User *User = Use.getUser();
4629  if (Instruction *I = dyn_cast<Instruction>(User))
4630  if (I->getParent() == CaseDest)
4631  continue;
4632  if (PHINode *Phi = dyn_cast<PHINode>(User))
4633  if (Phi->getIncomingBlock(Use) == CaseDest)
4634  continue;
4635  return false;
4636  }
4637 
4638  ConstantPool.insert(std::make_pair(&*I, C));
4639  } else {
4640  break;
4641  }
4642  }
4643 
4644  // If we did not have a CommonDest before, use the current one.
4645  if (!*CommonDest)
4646  *CommonDest = CaseDest;
4647  // If the destination isn't the common one, abort.
4648  if (CaseDest != *CommonDest)
4649  return false;
4650 
4651  // Get the values for this case from phi nodes in the destination block.
4652  for (PHINode &PHI : (*CommonDest)->phis()) {
4653  int Idx = PHI.getBasicBlockIndex(Pred);
4654  if (Idx == -1)
4655  continue;
4656 
4657  Constant *ConstVal =
4658  LookupConstant(PHI.getIncomingValue(Idx), ConstantPool);
4659  if (!ConstVal)
4660  return false;
4661 
4662  // Be conservative about which kinds of constants we support.
4663  if (!ValidLookupTableConstant(ConstVal, TTI))
4664  return false;
4665 
4666  Res.push_back(std::make_pair(&PHI, ConstVal));
4667  }
4668 
4669  return Res.size() > 0;
4670 }
4671 
4672 // Helper function used to add CaseVal to the list of cases that generate
4673 // Result. Returns the updated number of cases that generate this result.
4674 static uintptr_t MapCaseToResult(ConstantInt *CaseVal,
4675  SwitchCaseResultVectorTy &UniqueResults,
4676  Constant *Result) {
4677  for (auto &I : UniqueResults) {
4678  if (I.first == Result) {
4679  I.second.push_back(CaseVal);
4680  return I.second.size();
4681  }
4682  }
4683  UniqueResults.push_back(
4684  std::make_pair(Result, SmallVector<ConstantInt *, 4>(1, CaseVal)));
4685  return 1;
4686 }
4687 
4688 // Helper function that initializes a map containing
4689 // results for the PHI node of the common destination block for a switch
4690 // instruction. Returns false if multiple PHI nodes have been found or if
4691 // there is not a common destination block for the switch.
4692 static bool
4694  SwitchCaseResultVectorTy &UniqueResults,
4695  Constant *&DefaultResult, const DataLayout &DL,
4696  const TargetTransformInfo &TTI,
4697  uintptr_t MaxUniqueResults, uintptr_t MaxCasesPerResult) {
4698  for (auto &I : SI->cases()) {
4699  ConstantInt *CaseVal = I.getCaseValue();
4700 
4701  // Resulting value at phi nodes for this case value.
4702  SwitchCaseResultsTy Results;
4703  if (!GetCaseResults(SI, CaseVal, I.getCaseSuccessor(), &CommonDest, Results,
4704  DL, TTI))
4705  return false;
4706 
4707  // Only one value per case is permitted.
4708  if (Results.size() > 1)
4709  return false;
4710 
4711  // Add the case->result mapping to UniqueResults.
4712  const uintptr_t NumCasesForResult =
4713  MapCaseToResult(CaseVal, UniqueResults, Results.begin()->second);
4714 
4715  // Early out if there are too many cases for this result.
4716  if (NumCasesForResult > MaxCasesPerResult)
4717  return false;
4718 
4719  // Early out if there are too many unique results.
4720  if (UniqueResults.size() > MaxUniqueResults)
4721  return false;
4722 
4723  // Check the PHI consistency.
4724  if (!PHI)
4725  PHI = Results[0].first;
4726  else if (PHI != Results[0].first)
4727  return false;
4728  }
4729  // Find the default result value.
4731  BasicBlock *DefaultDest = SI->getDefaultDest();
4732  GetCaseResults(SI, nullptr, SI->getDefaultDest(), &CommonDest, DefaultResults,
4733  DL, TTI);
4734  // If the default value is not found abort unless the default destination
4735  // is unreachable.
4736  DefaultResult =
4737  DefaultResults.size() == 1 ? DefaultResults.begin()->second : nullptr;
4738  if ((!DefaultResult &&
4739  !isa<UnreachableInst>(DefaultDest->getFirstNonPHIOrDbg())))
4740  return false;
4741 
4742  return true;
4743 }
4744 
4745 // Helper function that checks if it is possible to transform a switch with only
4746 // two cases (or two cases + default) that produces a result into a select.
4747 // Example:
4748 // switch (a) {
4749 // case 10: %0 = icmp eq i32 %a, 10
4750 // return 10; %1 = select i1 %0, i32 10, i32 4
4751 // case 20: ----> %2 = icmp eq i32 %a, 20
4752 // return 2; %3 = select i1 %2, i32 2, i32 %1
4753 // default:
4754 // return 4;
4755 // }
4756 static Value *ConvertTwoCaseSwitch(const SwitchCaseResultVectorTy &ResultVector,
4757  Constant *DefaultResult, Value *Condition,
4758  IRBuilder<> &Builder) {
4759  assert(ResultVector.size() == 2 &&
4760  "We should have exactly two unique results at this point");
4761  // If we are selecting between only two cases transform into a simple
4762  // select or a two-way select if default is possible.
4763  if (ResultVector[0].second.size() == 1 &&
4764  ResultVector[1].second.size() == 1) {
4765  ConstantInt *const FirstCase = ResultVector[0].second[0];
4766  ConstantInt *const SecondCase = ResultVector[1].second[0];
4767 
4768  bool DefaultCanTrigger = DefaultResult;
4769  Value *SelectValue = ResultVector[1].first;
4770  if (DefaultCanTrigger) {
4771  Value *const ValueCompare =
4772  Builder.CreateICmpEQ(Condition, SecondCase, "switch.selectcmp");
4773  SelectValue = Builder.CreateSelect(ValueCompare, ResultVector[1].first,
4774  DefaultResult, "switch.select");
4775  }
4776  Value *const ValueCompare =
4777  Builder.CreateICmpEQ(Condition, FirstCase, "switch.selectcmp");
4778  return Builder.CreateSelect(ValueCompare, ResultVector[0].first,
4779  SelectValue, "switch.select");
4780  }
4781 
4782  return nullptr;
4783 }
4784 
4785 // Helper function to cleanup a switch instruction that has been converted into
4786 // a select, fixing up PHI nodes and basic blocks.
4788  Value *SelectValue,
4789  IRBuilder<> &Builder) {
4790  BasicBlock *SelectBB = SI->getParent();
4791  while (PHI->getBasicBlockIndex(SelectBB) >= 0)
4792  PHI->removeIncomingValue(SelectBB);
4793  PHI->addIncoming(SelectValue, SelectBB);
4794 
4795  Builder.CreateBr(PHI->getParent());
4796 
4797  // Remove the switch.
4798  for (unsigned i = 0, e = SI->getNumSuccessors(); i < e; ++i) {
4799  BasicBlock *Succ = SI->getSuccessor(i);
4800 
4801  if (Succ == PHI->getParent())
4802  continue;
4803  Succ->removePredecessor(SelectBB);
4804  }
4805  SI->eraseFromParent();
4806 }
4807 
4808 /// If the switch is only used to initialize one or more
4809 /// phi nodes in a common successor block with only two different
4810 /// constant values, replace the switch with select.
4811 static bool switchToSelect(SwitchInst *SI, IRBuilder<> &Builder,
4812  const DataLayout &DL,
4813  const TargetTransformInfo &TTI) {
4814  Value *const Cond = SI->getCondition();
4815  PHINode *PHI = nullptr;
4816  BasicBlock *CommonDest = nullptr;
4817  Constant *DefaultResult;
4818  SwitchCaseResultVectorTy UniqueResults;
4819  // Collect all the cases that will deliver the same value from the switch.
4820  if (!InitializeUniqueCases(SI, PHI, CommonDest, UniqueResults, DefaultResult,
4821  DL, TTI, 2, 1))
4822  return false;
4823  // Selects choose between maximum two values.
4824  if (UniqueResults.size() != 2)
4825  return false;
4826  assert(PHI != nullptr && "PHI for value select not found");
4827 
4828  Builder.SetInsertPoint(SI);
4829  Value *SelectValue =
4830  ConvertTwoCaseSwitch(UniqueResults, DefaultResult, Cond, Builder);
4831  if (SelectValue) {
4832  RemoveSwitchAfterSelectConversion(SI, PHI, SelectValue, Builder);
4833  return true;
4834  }
4835  // The switch couldn't be converted into a select.
4836  return false;
4837 }
4838 
4839 namespace {
4840 
4841 /// This class represents a lookup table that can be used to replace a switch.
4842 class SwitchLookupTable {
4843 public:
4844  /// Create a lookup table to use as a switch replacement with the contents
4845  /// of Values, using DefaultValue to fill any holes in the table.
4846  SwitchLookupTable(
4847  Module &M, uint64_t TableSize, ConstantInt *Offset,
4848  const SmallVectorImpl<std::pair<ConstantInt *, Constant *>> &Values,
4849  Constant *DefaultValue, const DataLayout &DL, const StringRef &FuncName);
4850 
4851  /// Build instructions with Builder to retrieve the value at
4852  /// the position given by Index in the lookup table.
4853  Value *BuildLookup(Value *Index, IRBuilder<> &Builder);
4854 
4855  /// Return true if a table with TableSize elements of
4856  /// type ElementType would fit in a target-legal register.
4857  static bool WouldFitInRegister(const DataLayout &DL, uint64_t TableSize,
4858  Type *ElementType);
4859 
4860 private:
4861  // Depending on the contents of the table, it can be represented in
4862  // different ways.
4863  enum {
4864  // For tables where each element contains the same value, we just have to
4865  // store that single value and return it for each lookup.
4866  SingleValueKind,
4867 
4868  // For tables where there is a linear relationship between table index
4869  // and values. We calculate the result with a simple multiplication
4870  // and addition instead of a table lookup.
4871  LinearMapKind,
4872 
4873  // For small tables with integer elements, we can pack them into a bitmap
4874  // that fits into a target-legal register. Values are retrieved by
4875  // shift and mask operations.
4876  BitMapKind,
4877 
4878  // The table is stored as an array of values. Values are retrieved by load
4879  // instructions from the table.
4880  ArrayKind
4881  } Kind;
4882 
4883  // For SingleValueKind, this is the single value.
4884  Constant *SingleValue = nullptr;
4885 
4886  // For BitMapKind, this is the bitmap.
4887  ConstantInt *BitMap = nullptr;
4888  IntegerType *BitMapElementTy = nullptr;
4889 
4890  // For LinearMapKind, these are the constants used to derive the value.
4891  ConstantInt *LinearOffset = nullptr;
4892  ConstantInt *LinearMultiplier = nullptr;
4893 
4894  // For ArrayKind, this is the array.
4895  GlobalVariable *Array = nullptr;
4896 };
4897 
4898 } // end anonymous namespace
4899 
4900 SwitchLookupTable::SwitchLookupTable(
4901  Module &M, uint64_t TableSize, ConstantInt *Offset,
4902  const SmallVectorImpl<std::pair<ConstantInt *, Constant *>> &Values,
4903  Constant *DefaultValue, const DataLayout &DL, const StringRef &FuncName) {
4904  assert(Values.size() && "Can't build lookup table without values!");
4905  assert(TableSize >= Values.size() && "Can't fit values in table!");
4906 
4907  // If all values in the table are equal, this is that value.
4908  SingleValue = Values.begin()->second;
4909 
4910  Type *ValueType = Values.begin()->second->getType();
4911 
4912  // Build up the table contents.
4913  SmallVector<Constant *, 64> TableContents(TableSize);
4914  for (size_t I = 0, E = Values.size(); I != E; ++I) {
4915  ConstantInt *CaseVal = Values[I].first;
4916  Constant *CaseRes = Values[I].second;
4917  assert(CaseRes->getType() == ValueType);
4918 
4919  uint64_t Idx = (CaseVal->getValue() - Offset->getValue()).getLimitedValue();
4920  TableContents[Idx] = CaseRes;
4921 
4922  if (CaseRes != SingleValue)
4923  SingleValue = nullptr;
4924  }
4925 
4926  // Fill in any holes in the table with the default result.
4927  if (Values.size() < TableSize) {
4928  assert(DefaultValue &&
4929  "Need a default value to fill the lookup table holes.");
4930  assert(DefaultValue->getType() == ValueType);
4931  for (uint64_t I = 0; I < TableSize; ++I) {
4932  if (!TableContents[I])
4933  TableContents[I] = DefaultValue;
4934  }
4935 
4936  if (DefaultValue != SingleValue)
4937  SingleValue = nullptr;
4938  }
4939 
4940  // If each element in the table contains the same value, we only need to store
4941  // that single value.
4942  if (SingleValue) {
4943  Kind = SingleValueKind;
4944  return;
4945  }
4946 
4947  // Check if we can derive the value with a linear transformation from the
4948  // table index.
4949  if (isa<IntegerType>(ValueType)) {
4950  bool LinearMappingPossible = true;
4951  APInt PrevVal;
4952  APInt DistToPrev;
4953  assert(TableSize >= 2 && "Should be a SingleValue table.");
4954  // Check if there is the same distance between two consecutive values.
4955  for (uint64_t I = 0; I < TableSize; ++I) {
4956  ConstantInt *ConstVal = dyn_cast<ConstantInt>(TableContents[I]);
4957  if (!ConstVal) {
4958  // This is an undef. We could deal with it, but undefs in lookup tables
4959  // are very seldom. It's probably not worth the additional complexity.
4960  LinearMappingPossible = false;
4961  break;
4962  }
4963  const APInt &Val = ConstVal->getValue();
4964  if (I != 0) {
4965  APInt Dist = Val - PrevVal;
4966  if (I == 1) {
4967  DistToPrev = Dist;
4968  } else if (Dist != DistToPrev) {
4969  LinearMappingPossible = false;
4970  break;
4971  }
4972  }
4973  PrevVal = Val;
4974  }
4975  if (LinearMappingPossible) {
4976  LinearOffset = cast<ConstantInt>(TableContents[0]);
4977  LinearMultiplier = ConstantInt::get(M.getContext(), DistToPrev);
4978  Kind = LinearMapKind;
4979  ++NumLinearMaps;
4980  return;
4981  }
4982  }
4983 
4984  // If the type is integer and the table fits in a register, build a bitmap.
4985  if (WouldFitInRegister(DL, TableSize, ValueType)) {
4986  IntegerType *IT = cast<IntegerType>(ValueType);
4987  APInt TableInt(TableSize * IT->getBitWidth(), 0);
4988  for (uint64_t I = TableSize; I > 0; --I) {
4989  TableInt <<= IT->getBitWidth();
4990  // Insert values into the bitmap. Undef values are set to zero.
4991  if (!isa<UndefValue>(TableContents[I - 1])) {
4992  ConstantInt *Val = cast<ConstantInt>(TableContents[I - 1]);
4993  TableInt |= Val->getValue().zext(TableInt.getBitWidth());
4994  }
4995  }
4996  BitMap = ConstantInt::get(M.getContext(), TableInt);
4997  BitMapElementTy = IT;
4998  Kind = BitMapKind;
4999  ++NumBitMaps;
5000  return;
5001  }
5002 
5003  // Store the table in an array.
5004  ArrayType *ArrayTy = ArrayType::get(ValueType, TableSize);
5005  Constant *Initializer = ConstantArray::get(ArrayTy, TableContents);
5006 
5007  Array = new GlobalVariable(M, ArrayTy, /*constant=*/true,
5008  GlobalVariable::PrivateLinkage, Initializer,
5009  "switch.table." + FuncName);
5010  Array->setUnnamedAddr(GlobalValue::UnnamedAddr::Global);
5011  Kind = ArrayKind;
5012 }
5013 
5014 Value *SwitchLookupTable::BuildLookup(Value *Index, IRBuilder<> &Builder) {
5015  switch (Kind) {
5016  case SingleValueKind:
5017  return SingleValue;
5018  case LinearMapKind: {
5019  // Derive the result value from the input value.
5020  Value *Result = Builder.CreateIntCast(Index, LinearMultiplier->getType(),
5021  false, "switch.idx.cast");
5022  if (!LinearMultiplier->isOne())
5023  Result = Builder.CreateMul(Result, LinearMultiplier, "switch.idx.mult");
5024  if (!LinearOffset->isZero())
5025  Result = Builder.CreateAdd(Result, LinearOffset, "switch.offset");
5026  return Result;
5027  }
5028  case BitMapKind: {
5029  // Type of the bitmap (e.g. i59).
5030  IntegerType *MapTy = BitMap->getType();
5031 
5032  // Cast Index to the same type as the bitmap.
5033  // Note: The Index is <= the number of elements in the table, so
5034  // truncating it to the width of the bitmask is safe.
5035  Value *ShiftAmt = Builder.CreateZExtOrTrunc(Index, MapTy, "switch.cast");
5036 
5037  // Multiply the shift amount by the element width.
5038  ShiftAmt = Builder.CreateMul(
5039  ShiftAmt, ConstantInt::get(MapTy, BitMapElementTy->getBitWidth()),
5040  "switch.shiftamt");
5041 
5042  // Shift down.
5043  Value *DownShifted =
5044  Builder.CreateLShr(BitMap, ShiftAmt, "switch.downshift");
5045  // Mask off.
5046  return Builder.CreateTrunc(DownShifted, BitMapElementTy, "switch.masked");
5047  }
5048  case ArrayKind: {
5049  // Make sure the table index will not overflow when treated as signed.
5050  IntegerType *IT = cast<IntegerType>(Index->getType());
5051  uint64_t TableSize =
5052  Array->getInitializer()->getType()->getArrayNumElements();
5053  if (TableSize > (1ULL << (IT->getBitWidth() - 1)))
5054  Index = Builder.CreateZExt(
5055  Index, IntegerType::get(IT->getContext(), IT->getBitWidth() + 1),
5056  "switch.tableidx.zext");
5057 
5058  Value *GEPIndices[] = {Builder.getInt32(0), Index};
5059  Value *GEP = Builder.CreateInBoundsGEP(Array->getValueType(), Array,
5060  GEPIndices, "switch.gep");
5061  return Builder.CreateLoad(GEP, "switch.load");
5062  }
5063  }
5064  llvm_unreachable("Unknown lookup table kind!");
5065 }
5066 
5067 bool SwitchLookupTable::WouldFitInRegister(const DataLayout &DL,
5068  uint64_t TableSize,
5069  Type *ElementType) {
5070  auto *IT = dyn_cast<IntegerType>(ElementType);
5071  if (!IT)
5072  return false;
5073  // FIXME: If the type is wider than it needs to be, e.g. i8 but all values
5074  // are <= 15, we could try to narrow the type.
5075 
5076  // Avoid overflow, fitsInLegalInteger uses unsigned int for the width.
5077  if (TableSize >= UINT_MAX / IT->getBitWidth())
5078  return false;
5079  return DL.fitsInLegalInteger(TableSize * IT->getBitWidth());
5080 }
5081 
5082 /// Determine whether a lookup table should be built for this switch, based on
5083 /// the number of cases, size of the table, and the types of the results.
5084 static bool
5085 ShouldBuildLookupTable(SwitchInst *SI, uint64_t TableSize,
5086  const TargetTransformInfo &TTI, const DataLayout &DL,
5087  const SmallDenseMap<PHINode *, Type *> &ResultTypes) {
5088  if (SI->getNumCases() > TableSize || TableSize >= UINT64_MAX / 10)
5089  return false; // TableSize overflowed, or mul below might overflow.
5090 
5091  bool AllTablesFitInRegister = true;
5092  bool HasIllegalType = false;
5093  for (const auto &I : ResultTypes) {
5094  Type *Ty = I.second;
5095 
5096  // Saturate this flag to true.
5097  HasIllegalType = HasIllegalType || !TTI.isTypeLegal(Ty);
5098 
5099  // Saturate this flag to false.
5100  AllTablesFitInRegister =
5101  AllTablesFitInRegister &&
5102  SwitchLookupTable::WouldFitInRegister(DL, TableSize, Ty);
5103 
5104  // If both flags saturate, we're done. NOTE: This *only* works with
5105  // saturating flags, and all flags have to saturate first due to the
5106  // non-deterministic behavior of iterating over a dense map.
5107  if (HasIllegalType && !AllTablesFitInRegister)
5108  break;
5109  }
5110 
5111  // If each table would fit in a register, we should build it anyway.
5112  if (AllTablesFitInRegister)
5113  return true;
5114 
5115  // Don't build a table that doesn't fit in-register if it has illegal types.
5116  if (HasIllegalType)
5117  return false;
5118 
5119  // The table density should be at least 40%. This is the same criterion as for
5120  // jump tables, see SelectionDAGBuilder::handleJTSwitchCase.
5121  // FIXME: Find the best cut-off.
5122  return SI->getNumCases() * 10 >= TableSize * 4;
5123 }
5124 
5125 /// Try to reuse the switch table index compare. Following pattern:
5126 /// \code
5127 /// if (idx < tablesize)
5128 /// r = table[idx]; // table does not contain default_value
5129 /// else
5130 /// r = default_value;
5131 /// if (r != default_value)
5132 /// ...
5133 /// \endcode
5134 /// Is optimized to:
5135 /// \code
5136 /// cond = idx < tablesize;
5137 /// if (cond)
5138 /// r = table[idx];
5139 /// else
5140 /// r = default_value;
5141 /// if (cond)
5142 /// ...
5143 /// \endcode
5144 /// Jump threading will then eliminate the second if(cond).
5145 static void reuseTableCompare(
5146  User *PhiUser, BasicBlock *PhiBlock, BranchInst *RangeCheckBranch,
5147  Constant *DefaultValue,
5148  const SmallVectorImpl<std::pair<ConstantInt *, Constant *>> &Values) {
5149  ICmpInst *CmpInst = dyn_cast<ICmpInst>(PhiUser);
5150  if (!CmpInst)
5151  return;
5152 
5153  // We require that the compare is in the same block as the phi so that jump
5154  // threading can do its work afterwards.
5155  if (CmpInst->getParent() != PhiBlock)
5156  return;
5157 
5158  Constant *CmpOp1 = dyn_cast<Constant>(CmpInst->getOperand(1));
5159  if (!CmpOp1)
5160  return;
5161 
5162  Value *RangeCmp = RangeCheckBranch->getCondition();
5163  Constant *TrueConst = ConstantInt::getTrue(RangeCmp->getType());
5164  Constant *FalseConst = ConstantInt::getFalse(RangeCmp->getType());
5165 
5166  // Check if the compare with the default value is constant true or false.
5167  Constant *DefaultConst = ConstantExpr::getICmp(CmpInst->getPredicate(),
5168  DefaultValue, CmpOp1, true);
5169  if (DefaultConst != TrueConst && DefaultConst != FalseConst)
5170  return;
5171 
5172  // Check if the compare with the case values is distinct from the default
5173  // compare result.
5174  for (auto ValuePair : Values) {
5175  Constant *CaseConst = ConstantExpr::getICmp(CmpInst->getPredicate(),
5176  ValuePair.second, CmpOp1, true);
5177  if (!CaseConst || CaseConst == DefaultConst || isa<UndefValue>(CaseConst))
5178  return;
5179  assert((CaseConst == TrueConst || CaseConst == FalseConst) &&
5180  "Expect true or false as compare result.");
5181  }
5182 
5183  // Check if the branch instruction dominates the phi node. It's a simple
5184  // dominance check, but sufficient for our needs.
5185  // Although this check is invariant in the calling loops, it's better to do it
5186  // at this late stage. Practically we do it at most once for a switch.
5187  BasicBlock *BranchBlock = RangeCheckBranch->getParent();
5188  for (auto PI = pred_begin(PhiBlock), E = pred_end(PhiBlock); PI != E; ++PI) {
5189  BasicBlock *Pred = *PI;
5190  if (Pred != BranchBlock && Pred->getUniquePredecessor() != BranchBlock)
5191  return;
5192  }
5193 
5194  if (DefaultConst == FalseConst) {
5195  // The compare yields the same result. We can replace it.
5196  CmpInst->replaceAllUsesWith(RangeCmp);
5197  ++NumTableCmpReuses;
5198  } else {
5199  // The compare yields the same result, just inverted. We can replace it.
5200  Value *InvertedTableCmp = BinaryOperator::CreateXor(
5201  RangeCmp, ConstantInt::get(RangeCmp->getType(), 1), "inverted.cmp",
5202  RangeCheckBranch);
5203  CmpInst->replaceAllUsesWith(InvertedTableCmp);
5204  ++NumTableCmpReuses;
5205  }
5206 }
5207 
5208 /// If the switch is only used to initialize one or more phi nodes in a common
5209 /// successor block with different constant values, replace the switch with
5210 /// lookup tables.
5212  const DataLayout &DL,
5213  const TargetTransformInfo &TTI) {
5214  assert(SI->getNumCases() > 1 && "Degenerate switch?");
5215 
5216  Function *Fn = SI->getParent()->getParent();
5217  // Only build lookup table when we have a target that supports it or the
5218  // attribute is not set.
5219  if (!TTI.shouldBuildLookupTables() ||
5220  (Fn->getFnAttribute("no-jump-tables").getValueAsString() == "true"))
5221  return false;
5222 
5223  // FIXME: If the switch is too sparse for a lookup table, perhaps we could
5224  // split off a dense part and build a lookup table for that.
5225 
5226  // FIXME: This creates arrays of GEPs to constant strings, which means each
5227  // GEP needs a runtime relocation in PIC code. We should just build one big
5228  // string and lookup indices into that.
5229 
5230  // Ignore switches with less than three cases. Lookup tables will not make
5231  // them faster, so we don't analyze them.
5232  if (SI->getNumCases() < 3)
5233  return false;
5234 
5235  // Figure out the corresponding result for each case value and phi node in the
5236  // common destination, as well as the min and max case values.
5237  assert(SI->case_begin() != SI->case_end());
5238  SwitchInst::CaseIt CI = SI->case_begin();
5239  ConstantInt *MinCaseVal = CI->getCaseValue();
5240  ConstantInt *MaxCaseVal = CI->getCaseValue();
5241 
5242  BasicBlock *CommonDest = nullptr;
5243 
5244  using ResultListTy = SmallVector<std::pair<ConstantInt *, Constant *>, 4>;
5246 
5247  SmallDenseMap<PHINode *, Constant *> DefaultResults;
5250 
5251  for (SwitchInst::CaseIt E = SI->case_end(); CI != E; ++CI) {
5252  ConstantInt *CaseVal = CI->getCaseValue();
5253  if (CaseVal->getValue().slt(MinCaseVal->getValue()))
5254  MinCaseVal = CaseVal;
5255  if (CaseVal->getValue().sgt(MaxCaseVal->getValue()))
5256  MaxCaseVal = CaseVal;
5257 
5258  // Resulting value at phi nodes for this case value.
5259  using ResultsTy = SmallVector<std::pair<PHINode *, Constant *>, 4>;
5260  ResultsTy Results;
5261  if (!GetCaseResults(SI, CaseVal, CI->getCaseSuccessor(), &CommonDest,
5262  Results, DL, TTI))
5263  return false;
5264 
5265  // Append the result from this case to the list for each phi.
5266  for (const auto &I : Results) {
5267  PHINode *PHI = I.first;
5268  Constant *Value = I.second;
5269  if (!ResultLists.count(PHI))
5270  PHIs.push_back(PHI);
5271  ResultLists[PHI].push_back(std::make_pair(CaseVal, Value));
5272  }
5273  }
5274 
5275  // Keep track of the result types.
5276  for (PHINode *PHI : PHIs) {
5277  ResultTypes[PHI] = ResultLists[PHI][0].second->getType();
5278  }
5279 
5280  uint64_t NumResults = ResultLists[PHIs[0]].size();
5281  APInt RangeSpread = MaxCaseVal->getValue() - MinCaseVal->getValue();
5282  uint64_t TableSize = RangeSpread.getLimitedValue() + 1;
5283  bool TableHasHoles = (NumResults < TableSize);
5284 
5285  // If the table has holes, we need a constant result for the default case
5286  // or a bitmask that fits in a register.
5287  SmallVector<std::pair<PHINode *, Constant *>, 4> DefaultResultsList;
5288  bool HasDefaultResults =
5289  GetCaseResults(SI, nullptr, SI->getDefaultDest(), &CommonDest,
5290  DefaultResultsList, DL, TTI);
5291 
5292  bool NeedMask = (TableHasHoles && !HasDefaultResults);
5293  if (NeedMask) {
5294  // As an extra penalty for the validity test we require more cases.
5295  if (SI->getNumCases() < 4) // FIXME: Find best threshold value (benchmark).
5296  return false;
5297  if (!DL.fitsInLegalInteger(TableSize))
5298  return false;
5299  }
5300 
5301  for (const auto &I : DefaultResultsList) {
5302  PHINode *PHI = I.first;
5303  Constant *Result = I.second;
5304  DefaultResults[PHI] = Result;
5305  }
5306 
5307  if (!ShouldBuildLookupTable(SI, TableSize, TTI, DL, ResultTypes))
5308  return false;
5309 
5310  // Create the BB that does the lookups.
5311  Module &Mod = *CommonDest->getParent()->getParent();
5312  BasicBlock *LookupBB = BasicBlock::Create(
5313  Mod.getContext(), "switch.lookup", CommonDest->getParent(), CommonDest);
5314 
5315  // Compute the table index value.
5316  Builder.SetInsertPoint(SI);
5317  Value *TableIndex;
5318  if (MinCaseVal->isNullValue())
5319  TableIndex = SI->getCondition();
5320  else
5321  TableIndex = Builder.CreateSub(SI->getCondition(), MinCaseVal,
5322  "switch.tableidx");
5323 
5324  // Compute the maximum table size representable by the integer type we are
5325  // switching upon.
5326  unsigned CaseSize = MinCaseVal->getType()->getPrimitiveSizeInBits();
5327  uint64_t MaxTableSize = CaseSize > 63 ? UINT64_MAX : 1ULL << CaseSize;
5328  assert(MaxTableSize >= TableSize &&
5329  "It is impossible for a switch to have more entries than the max "
5330  "representable value of its input integer type's size.");
5331 
5332  // If the default destination is unreachable, or if the lookup table covers
5333  // all values of the conditional variable, branch directly to the lookup table
5334  // BB. Otherwise, check that the condition is within the case range.
5335  const bool DefaultIsReachable =
5336  !isa<UnreachableInst>(SI->getDefaultDest()->getFirstNonPHIOrDbg());
5337  const bool GeneratingCoveredLookupTable = (MaxTableSize == TableSize);
5338  BranchInst *RangeCheckBranch = nullptr;
5339 
5340  if (!DefaultIsReachable || GeneratingCoveredLookupTable) {
5341  Builder.CreateBr(LookupBB);
5342  // Note: We call removeProdecessor later since we need to be able to get the
5343  // PHI value for the default case in case we're using a bit mask.
5344  } else {
5345  Value *Cmp = Builder.CreateICmpULT(
5346  TableIndex, ConstantInt::get(MinCaseVal->getType(), TableSize));
5347  RangeCheckBranch =
5348  Builder.CreateCondBr(Cmp, LookupBB, SI->getDefaultDest());
5349  }
5350 
5351  // Populate the BB that does the lookups.
5352  Builder.SetInsertPoint(LookupBB);
5353 
5354  if (NeedMask) {
5355  // Before doing the lookup, we do the hole check. The LookupBB is therefore
5356  // re-purposed to do the hole check, and we create a new LookupBB.
5357  BasicBlock *MaskBB = LookupBB;
5358  MaskBB->setName("switch.hole_check");
5359  LookupBB = BasicBlock::Create(Mod.getContext(), "switch.lookup",
5360  CommonDest->getParent(), CommonDest);
5361 
5362  // Make the mask's bitwidth at least 8-bit and a power-of-2 to avoid
5363  // unnecessary illegal types.
5364  uint64_t TableSizePowOf2 = NextPowerOf2(std::max(7ULL, TableSize - 1ULL));
5365  APInt MaskInt(TableSizePowOf2, 0);
5366  APInt One(TableSizePowOf2, 1);
5367  // Build bitmask; fill in a 1 bit for every case.
5368  const ResultListTy &ResultList = ResultLists[PHIs[0]];
5369  for (size_t I = 0, E = ResultList.size(); I != E; ++I) {
5370  uint64_t Idx = (ResultList[I].first->getValue() - MinCaseVal->getValue())
5371  .getLimitedValue();
5372  MaskInt |= One << Idx;
5373  }
5374  ConstantInt *TableMask = ConstantInt::get(Mod.getContext(), MaskInt);
5375 
5376  // Get the TableIndex'th bit of the bitmask.
5377  // If this bit is 0 (meaning hole) jump to the default destination,
5378  // else continue with table lookup.
5379  IntegerType *MapTy = TableMask->getType();
5380  Value *MaskIndex =
5381  Builder.CreateZExtOrTrunc(TableIndex, MapTy, "switch.maskindex");
5382  Value *Shifted = Builder.CreateLShr(TableMask, MaskIndex, "switch.shifted");
5383  Value *LoBit = Builder.CreateTrunc(
5384  Shifted, Type::getInt1Ty(Mod.getContext()), "switch.lobit");
5385  Builder.CreateCondBr(LoBit, LookupBB, SI->getDefaultDest());
5386 
5387  Builder.SetInsertPoint(LookupBB);
5388  AddPredecessorToBlock(SI->getDefaultDest(), MaskBB, SI->getParent());
5389  }
5390 
5391  if (!DefaultIsReachable || GeneratingCoveredLookupTable) {
5392  // We cached PHINodes in PHIs. To avoid accessing deleted PHINodes later,
5393  // do not delete PHINodes here.
5395  /*DontDeleteUselessPHIs=*/true);
5396  }
5397 
5398  bool ReturnedEarly = false;
5399  for (size_t I = 0, E = PHIs.size(); I != E; ++I) {
5400  PHINode *PHI = PHIs[I];
5401  const ResultListTy &ResultList = ResultLists[PHI];
5402 
5403  // If using a bitmask, use any value to fill the lookup table holes.
5404  Constant *DV = NeedMask ? ResultLists[PHI][0].second : DefaultResults[PHI];
5405  StringRef FuncName = Fn->getName();
5406  SwitchLookupTable Table(Mod, TableSize, MinCaseVal, ResultList, DV, DL,
5407  FuncName);
5408 
5409  Value *Result = Table.BuildLookup(TableIndex, Builder);
5410 
5411  // If the result is used to return immediately from the function, we want to
5412  // do that right here.
5413  if (PHI->hasOneUse() && isa<ReturnInst>(*PHI->user_begin()) &&
5414  PHI->user_back() == CommonDest->getFirstNonPHIOrDbg()) {
5415  Builder.CreateRet(Result);
5416  ReturnedEarly = true;
5417  break;
5418  }
5419 
5420  // Do a small peephole optimization: re-use the switch table compare if
5421  // possible.
5422  if (!TableHasHoles && HasDefaultResults && RangeCheckBranch) {
5423  BasicBlock *PhiBlock = PHI->getParent();
5424  // Search for compare instructions which use the phi.
5425  for (auto *User : PHI->users()) {
5426  reuseTableCompare(User, PhiBlock, RangeCheckBranch, DV, ResultList);
5427  }
5428  }
5429 
5430  PHI->addIncoming(Result, LookupBB);
5431  }
5432 
5433  if (!ReturnedEarly)
5434  Builder.CreateBr(CommonDest);
5435 
5436  // Remove the switch.
5437  for (unsigned i = 0, e = SI->getNumSuccessors(); i < e; ++i) {
5438  BasicBlock *Succ = SI->getSuccessor(i);
5439 
5440  if (Succ == SI->getDefaultDest())
5441  continue;
5442  Succ->removePredecessor(SI->getParent());
5443  }
5444  SI->eraseFromParent();
5445 
5446  ++NumLookupTables;
5447  if (NeedMask)
5448  ++NumLookupTablesHoles;
5449  return true;
5450 }
5451 
5452 static bool isSwitchDense(ArrayRef<int64_t> Values) {
5453  // See also SelectionDAGBuilder::isDense(), which this function was based on.
5454  uint64_t Diff = (uint64_t)Values.back() - (uint64_t)Values.front();
5455  uint64_t Range = Diff + 1;
5456  uint64_t NumCases = Values.size();
5457  // 40% is the default density for building a jump table in optsize/minsize mode.
5458  uint64_t MinDensity = 40;
5459 
5460  return NumCases * 100 >= Range * MinDensity;
5461 }
5462 
5463 /// Try to transform a switch that has "holes" in it to a contiguous sequence
5464 /// of cases.
5465 ///
5466 /// A switch such as: switch(i) {case 5: case 9: case 13: case 17:} can be
5467 /// range-reduced to: switch ((i-5) / 4) {case 0: case 1: case 2: case 3:}.
5468 ///
5469 /// This converts a sparse switch into a dense switch which allows better
5470 /// lowering and could also allow transforming into a lookup table.
5472  const DataLayout &DL,
5473  const TargetTransformInfo &TTI) {
5474  auto *CondTy = cast<IntegerType>(SI->getCondition()->getType());
5475  if (CondTy->getIntegerBitWidth() > 64 ||
5476  !DL.fitsInLegalInteger(CondTy->getIntegerBitWidth()))
5477  return false;
5478  // Only bother with this optimization if there are more than 3 switch cases;
5479  // SDAG will only bother creating jump tables for 4 or more cases.
5480  if (SI->getNumCases() < 4)
5481  return false;
5482 
5483  // This transform is agnostic to the signedness of the input or case values. We
5484  // can treat the case values as signed or unsigned. We can optimize more common
5485  // cases such as a sequence crossing zero {-4,0,4,8} if we interpret case values
5486  // as signed.
5487  SmallVector<int64_t,4> Values;
5488  for (auto &C : SI->cases())
5489  Values.push_back(C.getCaseValue()->getValue().getSExtValue());
5490  std::sort(Values.begin(), Values.end());
5491 
5492  // If the switch is already dense, there's nothing useful to do here.
5493  if (isSwitchDense(Values))
5494  return false;
5495 
5496  // First, transform the values such that they start at zero and ascend.
5497  int64_t Base = Values[0];
5498  for (auto &V : Values)
5499  V -= (uint64_t)(Base);
5500 
5501  // Now we have signed numbers that have been shifted so that, given enough
5502  // precision, there are no negative values. Since the rest of the transform
5503  // is bitwise only, we switch now to an unsigned representation.
5504  uint64_t GCD = 0;
5505  for (auto &V : Values)
5506  GCD = GreatestCommonDivisor64(GCD, (uint64_t)V);
5507 
5508  // This transform can be done speculatively because it is so cheap - it results
5509  // in a single rotate operation being inserted. This can only happen if the
5510  // factor extracted is a power of 2.
5511  // FIXME: If the GCD is an odd number we can multiply by the multiplicative
5512  // inverse of GCD and then perform this transform.
5513  // FIXME: It's possible that optimizing a switch on powers of two might also
5514  // be beneficial - flag values are often powers of two and we could use a CLZ
5515  // as the key function.
5516  if (GCD <= 1 || !isPowerOf2_64(GCD))
5517  // No common divisor found or too expensive to compute key function.
5518  return false;
5519 
5520  unsigned Shift = Log2_64(GCD);
5521  for (auto &V : Values)
5522  V = (int64_t)((uint64_t)V >> Shift);
5523 
5524  if (!isSwitchDense(Values))
5525  // Transform didn't create a dense switch.
5526  return false;
5527 
5528  // The obvious transform is to shift the switch condition right and emit a
5529  // check that the condition actually cleanly divided by GCD, i.e.
5530  // C & (1 << Shift - 1) == 0
5531  // inserting a new CFG edge to handle the case where it didn't divide cleanly.
5532  //
5533  // A cheaper way of doing this is a simple ROTR(C, Shift). This performs the
5534  // shift and puts the shifted-off bits in the uppermost bits. If any of these
5535  // are nonzero then the switch condition will be very large and will hit the
5536  // default case.
5537 
5538  auto *Ty = cast<IntegerType>(SI->getCondition()->getType());
5539  Builder.SetInsertPoint(SI);
5540  auto *ShiftC = ConstantInt::get(Ty, Shift);
5541  auto *Sub = Builder.CreateSub(SI->getCondition(), ConstantInt::get(Ty, Base));
5542  auto *LShr = Builder.CreateLShr(Sub, ShiftC);
5543  auto *Shl = Builder.CreateShl(Sub, Ty->getBitWidth() - Shift);
5544  auto *Rot = Builder.CreateOr(LShr, Shl);
5545  SI->replaceUsesOfWith(SI->getCondition(), Rot);
5546 
5547  for (auto Case : SI->cases()) {
5548  auto *Orig = Case.getCaseValue();
5549  auto Sub = Orig->getValue() - APInt(Ty->getBitWidth(), Base);
5550  Case.setValue(
5551  cast<ConstantInt>(ConstantInt::get(Ty, Sub.lshr(ShiftC->getValue()))));
5552  }
5553  return true;
5554 }
5555 
5556 bool SimplifyCFGOpt::SimplifySwitch(SwitchInst *SI, IRBuilder<> &Builder) {
5557  BasicBlock *BB = SI->getParent();
5558 
5559  if (isValueEqualityComparison(SI)) {
5560  // If we only have one predecessor, and if it is a branch on this value,
5561  // see if that predecessor totally determines the outcome of this switch.
5562  if (BasicBlock *OnlyPred = BB->getSinglePredecessor())
5563  if (SimplifyEqualityComparisonWithOnlyPredecessor(SI, OnlyPred, Builder))
5564  return simplifyCFG(BB, TTI, Options) | true;
5565 
5566  Value *Cond = SI->getCondition();
5567  if (SelectInst *Select = dyn_cast<SelectInst>(Cond))
5568  if (SimplifySwitchOnSelect(SI, Select))
5569  return simplifyCFG(BB, TTI, Options) | true;
5570 
5571  // If the block only contains the switch, see if we can fold the block
5572  // away into any preds.
5573  BasicBlock::iterator BBI = BB->begin();
5574  // Ignore dbg intrinsics.
5575  while (isa<DbgInfoIntrinsic>(BBI))
5576  ++BBI;
5577  if (SI == &*BBI)
5578  if (FoldValueComparisonIntoPredecessors(SI, Builder))
5579  return