LLVM  6.0.0svn
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"
31 #include "llvm/IR/BasicBlock.h"
32 #include "llvm/IR/CFG.h"
33 #include "llvm/IR/CallSite.h"
34 #include "llvm/IR/Constant.h"
35 #include "llvm/IR/ConstantRange.h"
36 #include "llvm/IR/Constants.h"
37 #include "llvm/IR/DataLayout.h"
38 #include "llvm/IR/DebugInfo.h"
39 #include "llvm/IR/DerivedTypes.h"
40 #include "llvm/IR/GlobalValue.h"
41 #include "llvm/IR/GlobalVariable.h"
42 #include "llvm/IR/IRBuilder.h"
43 #include "llvm/IR/InstrTypes.h"
44 #include "llvm/IR/Instruction.h"
45 #include "llvm/IR/Instructions.h"
46 #include "llvm/IR/IntrinsicInst.h"
47 #include "llvm/IR/Intrinsics.h"
48 #include "llvm/IR/LLVMContext.h"
49 #include "llvm/IR/MDBuilder.h"
50 #include "llvm/IR/Metadata.h"
51 #include "llvm/IR/Module.h"
52 #include "llvm/IR/NoFolder.h"
53 #include "llvm/IR/Operator.h"
54 #include "llvm/IR/PatternMatch.h"
55 #include "llvm/IR/Type.h"
56 #include "llvm/IR/User.h"
57 #include "llvm/IR/Value.h"
58 #include "llvm/Support/Casting.h"
60 #include "llvm/Support/Debug.h"
62 #include "llvm/Support/KnownBits.h"
68 #include <algorithm>
69 #include <cassert>
70 #include <climits>
71 #include <cstddef>
72 #include <cstdint>
73 #include <iterator>
74 #include <map>
75 #include <set>
76 #include <utility>
77 #include <vector>
78 
79 using namespace llvm;
80 using namespace PatternMatch;
81 
82 #define DEBUG_TYPE "simplifycfg"
83 
84 // Chosen as 2 so as to be cheap, but still to have enough power to fold
85 // a select, so the "clamp" idiom (of a min followed by a max) will be caught.
86 // To catch this, we need to fold a compare and a select, hence '2' being the
87 // minimum reasonable default.
89  "phi-node-folding-threshold", cl::Hidden, cl::init(2),
90  cl::desc(
91  "Control the amount of phi node folding to perform (default = 2)"));
92 
93 static cl::opt<bool> DupRet(
94  "simplifycfg-dup-ret", cl::Hidden, cl::init(false),
95  cl::desc("Duplicate return instructions into unconditional branches"));
96 
97 static cl::opt<bool>
98  SinkCommon("simplifycfg-sink-common", cl::Hidden, cl::init(true),
99  cl::desc("Sink common instructions down to the end block"));
100 
102  "simplifycfg-hoist-cond-stores", cl::Hidden, cl::init(true),
103  cl::desc("Hoist conditional stores if an unconditional store precedes"));
104 
106  "simplifycfg-merge-cond-stores", cl::Hidden, cl::init(true),
107  cl::desc("Hoist conditional stores even if an unconditional store does not "
108  "precede - hoist multiple conditional stores into a single "
109  "predicated store"));
110 
112  "simplifycfg-merge-cond-stores-aggressively", cl::Hidden, cl::init(false),
113  cl::desc("When merging conditional stores, do so even if the resultant "
114  "basic blocks are unlikely to be if-converted as a result"));
115 
117  "speculate-one-expensive-inst", cl::Hidden, cl::init(true),
118  cl::desc("Allow exactly one expensive instruction to be speculatively "
119  "executed"));
120 
122  "max-speculation-depth", cl::Hidden, cl::init(10),
123  cl::desc("Limit maximum recursion depth when calculating costs of "
124  "speculatively executed instructions"));
125 
126 STATISTIC(NumBitMaps, "Number of switch instructions turned into bitmaps");
127 STATISTIC(NumLinearMaps,
128  "Number of switch instructions turned into linear mapping");
129 STATISTIC(NumLookupTables,
130  "Number of switch instructions turned into lookup tables");
131 STATISTIC(
132  NumLookupTablesHoles,
133  "Number of switch instructions turned into lookup tables (holes checked)");
134 STATISTIC(NumTableCmpReuses, "Number of reused switch table lookup compares");
135 STATISTIC(NumSinkCommons,
136  "Number of common instructions sunk down to the end block");
137 STATISTIC(NumSpeculations, "Number of speculative executed instructions");
138 
139 namespace {
140 
141 // The first field contains the value that the switch produces when a certain
142 // case group is selected, and the second field is a vector containing the
143 // cases composing the case group.
145  SwitchCaseResultVectorTy;
146 // The first field contains the phi node that generates a result of the switch
147 // and the second field contains the value generated for a certain case in the
148 // switch for that PHI.
149 typedef SmallVector<std::pair<PHINode *, Constant *>, 4> SwitchCaseResultsTy;
150 
151 /// ValueEqualityComparisonCase - Represents a case of a switch.
152 struct ValueEqualityComparisonCase {
154  BasicBlock *Dest;
155 
156  ValueEqualityComparisonCase(ConstantInt *Value, BasicBlock *Dest)
157  : Value(Value), Dest(Dest) {}
158 
159  bool operator<(ValueEqualityComparisonCase RHS) const {
160  // Comparing pointers is ok as we only rely on the order for uniquing.
161  return Value < RHS.Value;
162  }
163 
164  bool operator==(BasicBlock *RHSDest) const { return Dest == RHSDest; }
165 };
166 
167 class SimplifyCFGOpt {
168  const TargetTransformInfo &TTI;
169  const DataLayout &DL;
170  SmallPtrSetImpl<BasicBlock *> *LoopHeaders;
171  const SimplifyCFGOptions &Options;
172 
173  Value *isValueEqualityComparison(TerminatorInst *TI);
174  BasicBlock *GetValueEqualityComparisonCases(
175  TerminatorInst *TI, std::vector<ValueEqualityComparisonCase> &Cases);
176  bool SimplifyEqualityComparisonWithOnlyPredecessor(TerminatorInst *TI,
177  BasicBlock *Pred,
178  IRBuilder<> &Builder);
179  bool FoldValueComparisonIntoPredecessors(TerminatorInst *TI,
180  IRBuilder<> &Builder);
181 
182  bool SimplifyReturn(ReturnInst *RI, IRBuilder<> &Builder);
183  bool SimplifyResume(ResumeInst *RI, IRBuilder<> &Builder);
184  bool SimplifySingleResume(ResumeInst *RI);
185  bool SimplifyCommonResume(ResumeInst *RI);
186  bool SimplifyCleanupReturn(CleanupReturnInst *RI);
187  bool SimplifyUnreachable(UnreachableInst *UI);
188  bool SimplifySwitch(SwitchInst *SI, IRBuilder<> &Builder);
189  bool SimplifyIndirectBr(IndirectBrInst *IBI);
190  bool SimplifyUncondBranch(BranchInst *BI, IRBuilder<> &Builder);
191  bool SimplifyCondBranch(BranchInst *BI, IRBuilder<> &Builder);
192 
193 public:
194  SimplifyCFGOpt(const TargetTransformInfo &TTI, const DataLayout &DL,
195  SmallPtrSetImpl<BasicBlock *> *LoopHeaders,
196  const SimplifyCFGOptions &Opts)
197  : TTI(TTI), DL(DL), LoopHeaders(LoopHeaders), Options(Opts) {}
198 
199  bool run(BasicBlock *BB);
200 };
201 
202 } // end anonymous namespace
203 
204 /// Return true if it is safe to merge these two
205 /// terminator instructions together.
206 static bool
208  SmallSetVector<BasicBlock *, 4> *FailBlocks = nullptr) {
209  if (SI1 == SI2)
210  return false; // Can't merge with self!
211 
212  // It is not safe to merge these two switch instructions if they have a common
213  // successor, and if that successor has a PHI node, and if *that* PHI node has
214  // conflicting incoming values from the two switch blocks.
215  BasicBlock *SI1BB = SI1->getParent();
216  BasicBlock *SI2BB = SI2->getParent();
217 
218  SmallPtrSet<BasicBlock *, 16> SI1Succs(succ_begin(SI1BB), succ_end(SI1BB));
219  bool Fail = false;
220  for (BasicBlock *Succ : successors(SI2BB))
221  if (SI1Succs.count(Succ))
222  for (BasicBlock::iterator BBI = Succ->begin(); isa<PHINode>(BBI); ++BBI) {
223  PHINode *PN = cast<PHINode>(BBI);
224  if (PN->getIncomingValueForBlock(SI1BB) !=
225  PN->getIncomingValueForBlock(SI2BB)) {
226  if (FailBlocks)
227  FailBlocks->insert(Succ);
228  Fail = true;
229  }
230  }
231 
232  return !Fail;
233 }
234 
235 /// Return true if it is safe and profitable to merge these two terminator
236 /// instructions together, where SI1 is an unconditional branch. PhiNodes will
237 /// store all PHI nodes in common successors.
238 static bool
240  Instruction *Cond,
241  SmallVectorImpl<PHINode *> &PhiNodes) {
242  if (SI1 == SI2)
243  return false; // Can't merge with self!
244  assert(SI1->isUnconditional() && SI2->isConditional());
245 
246  // We fold the unconditional branch if we can easily update all PHI nodes in
247  // common successors:
248  // 1> We have a constant incoming value for the conditional branch;
249  // 2> We have "Cond" as the incoming value for the unconditional branch;
250  // 3> SI2->getCondition() and Cond have same operands.
251  CmpInst *Ci2 = dyn_cast<CmpInst>(SI2->getCondition());
252  if (!Ci2)
253  return false;
254  if (!(Cond->getOperand(0) == Ci2->getOperand(0) &&
255  Cond->getOperand(1) == Ci2->getOperand(1)) &&
256  !(Cond->getOperand(0) == Ci2->getOperand(1) &&
257  Cond->getOperand(1) == Ci2->getOperand(0)))
258  return false;
259 
260  BasicBlock *SI1BB = SI1->getParent();
261  BasicBlock *SI2BB = SI2->getParent();
262  SmallPtrSet<BasicBlock *, 16> SI1Succs(succ_begin(SI1BB), succ_end(SI1BB));
263  for (BasicBlock *Succ : successors(SI2BB))
264  if (SI1Succs.count(Succ))
265  for (BasicBlock::iterator BBI = Succ->begin(); isa<PHINode>(BBI); ++BBI) {
266  PHINode *PN = cast<PHINode>(BBI);
267  if (PN->getIncomingValueForBlock(SI1BB) != Cond ||
268  !isa<ConstantInt>(PN->getIncomingValueForBlock(SI2BB)))
269  return false;
270  PhiNodes.push_back(PN);
271  }
272  return true;
273 }
274 
275 /// Update PHI nodes in Succ to indicate that there will now be entries in it
276 /// from the 'NewPred' block. The values that will be flowing into the PHI nodes
277 /// will be the same as those coming in from ExistPred, an existing predecessor
278 /// of Succ.
279 static void AddPredecessorToBlock(BasicBlock *Succ, BasicBlock *NewPred,
280  BasicBlock *ExistPred) {
281  if (!isa<PHINode>(Succ->begin()))
282  return; // Quick exit if nothing to do
283 
284  PHINode *PN;
285  for (BasicBlock::iterator I = Succ->begin(); (PN = dyn_cast<PHINode>(I)); ++I)
286  PN->addIncoming(PN->getIncomingValueForBlock(ExistPred), NewPred);
287 }
288 
289 /// Compute an abstract "cost" of speculating the given instruction,
290 /// which is assumed to be safe to speculate. TCC_Free means cheap,
291 /// TCC_Basic means less cheap, and TCC_Expensive means prohibitively
292 /// expensive.
293 static unsigned ComputeSpeculationCost(const User *I,
294  const TargetTransformInfo &TTI) {
296  "Instruction is not safe to speculatively execute!");
297  return TTI.getUserCost(I);
298 }
299 
300 /// If we have a merge point of an "if condition" as accepted above,
301 /// return true if the specified value dominates the block. We
302 /// don't handle the true generality of domination here, just a special case
303 /// which works well enough for us.
304 ///
305 /// If AggressiveInsts is non-null, and if V does not dominate BB, we check to
306 /// see if V (which must be an instruction) and its recursive operands
307 /// that do not dominate BB have a combined cost lower than CostRemaining and
308 /// are non-trapping. If both are true, the instruction is inserted into the
309 /// set and true is returned.
310 ///
311 /// The cost for most non-trapping instructions is defined as 1 except for
312 /// Select whose cost is 2.
313 ///
314 /// After this function returns, CostRemaining is decreased by the cost of
315 /// V plus its non-dominating operands. If that cost is greater than
316 /// CostRemaining, false is returned and CostRemaining is undefined.
318  SmallPtrSetImpl<Instruction *> *AggressiveInsts,
319  unsigned &CostRemaining,
320  const TargetTransformInfo &TTI,
321  unsigned Depth = 0) {
322  // It is possible to hit a zero-cost cycle (phi/gep instructions for example),
323  // so limit the recursion depth.
324  // TODO: While this recursion limit does prevent pathological behavior, it
325  // would be better to track visited instructions to avoid cycles.
326  if (Depth == MaxSpeculationDepth)
327  return false;
328 
330  if (!I) {
331  // Non-instructions all dominate instructions, but not all constantexprs
332  // can be executed unconditionally.
333  if (ConstantExpr *C = dyn_cast<ConstantExpr>(V))
334  if (C->canTrap())
335  return false;
336  return true;
337  }
338  BasicBlock *PBB = I->getParent();
339 
340  // We don't want to allow weird loops that might have the "if condition" in
341  // the bottom of this block.
342  if (PBB == BB)
343  return false;
344 
345  // If this instruction is defined in a block that contains an unconditional
346  // branch to BB, then it must be in the 'conditional' part of the "if
347  // statement". If not, it definitely dominates the region.
349  if (!BI || BI->isConditional() || BI->getSuccessor(0) != BB)
350  return true;
351 
352  // If we aren't allowing aggressive promotion anymore, then don't consider
353  // instructions in the 'if region'.
354  if (!AggressiveInsts)
355  return false;
356 
357  // If we have seen this instruction before, don't count it again.
358  if (AggressiveInsts->count(I))
359  return true;
360 
361  // Okay, it looks like the instruction IS in the "condition". Check to
362  // see if it's a cheap instruction to unconditionally compute, and if it
363  // only uses stuff defined outside of the condition. If so, hoist it out.
365  return false;
366 
367  unsigned Cost = ComputeSpeculationCost(I, TTI);
368 
369  // Allow exactly one instruction to be speculated regardless of its cost
370  // (as long as it is safe to do so).
371  // This is intended to flatten the CFG even if the instruction is a division
372  // or other expensive operation. The speculation of an expensive instruction
373  // is expected to be undone in CodeGenPrepare if the speculation has not
374  // enabled further IR optimizations.
375  if (Cost > CostRemaining &&
376  (!SpeculateOneExpensiveInst || !AggressiveInsts->empty() || Depth > 0))
377  return false;
378 
379  // Avoid unsigned wrap.
380  CostRemaining = (Cost > CostRemaining) ? 0 : CostRemaining - Cost;
381 
382  // Okay, we can only really hoist these out if their operands do
383  // not take us over the cost threshold.
384  for (User::op_iterator i = I->op_begin(), e = I->op_end(); i != e; ++i)
385  if (!DominatesMergePoint(*i, BB, AggressiveInsts, CostRemaining, TTI,
386  Depth + 1))
387  return false;
388  // Okay, it's safe to do this! Remember this instruction.
389  AggressiveInsts->insert(I);
390  return true;
391 }
392 
393 /// Extract ConstantInt from value, looking through IntToPtr
394 /// and PointerNullValue. Return NULL if value is not a constant int.
395 static ConstantInt *GetConstantInt(Value *V, const DataLayout &DL) {
396  // Normal constant int.
397  ConstantInt *CI = dyn_cast<ConstantInt>(V);
398  if (CI || !isa<Constant>(V) || !V->getType()->isPointerTy())
399  return CI;
400 
401  // This is some kind of pointer constant. Turn it into a pointer-sized
402  // ConstantInt if possible.
403  IntegerType *PtrTy = cast<IntegerType>(DL.getIntPtrType(V->getType()));
404 
405  // Null pointer means 0, see SelectionDAGBuilder::getValue(const Value*).
406  if (isa<ConstantPointerNull>(V))
407  return ConstantInt::get(PtrTy, 0);
408 
409  // IntToPtr const int.
410  if (ConstantExpr *CE = dyn_cast<ConstantExpr>(V))
411  if (CE->getOpcode() == Instruction::IntToPtr)
412  if (ConstantInt *CI = dyn_cast<ConstantInt>(CE->getOperand(0))) {
413  // The constant is very likely to have the right type already.
414  if (CI->getType() == PtrTy)
415  return CI;
416  else
417  return cast<ConstantInt>(
418  ConstantExpr::getIntegerCast(CI, PtrTy, /*isSigned=*/false));
419  }
420  return nullptr;
421 }
422 
423 namespace {
424 
425 /// Given a chain of or (||) or and (&&) comparison of a value against a
426 /// constant, this will try to recover the information required for a switch
427 /// structure.
428 /// It will depth-first traverse the chain of comparison, seeking for patterns
429 /// like %a == 12 or %a < 4 and combine them to produce a set of integer
430 /// representing the different cases for the switch.
431 /// Note that if the chain is composed of '||' it will build the set of elements
432 /// that matches the comparisons (i.e. any of this value validate the chain)
433 /// while for a chain of '&&' it will build the set elements that make the test
434 /// fail.
435 struct ConstantComparesGatherer {
436  const DataLayout &DL;
437  Value *CompValue; /// Value found for the switch comparison
438  Value *Extra; /// Extra clause to be checked before the switch
439  SmallVector<ConstantInt *, 8> Vals; /// Set of integers to match in switch
440  unsigned UsedICmps; /// Number of comparisons matched in the and/or chain
441 
442  /// Construct and compute the result for the comparison instruction Cond
443  ConstantComparesGatherer(Instruction *Cond, const DataLayout &DL)
444  : DL(DL), CompValue(nullptr), Extra(nullptr), UsedICmps(0) {
445  gather(Cond);
446  }
447 
448  /// Prevent copy
449  ConstantComparesGatherer(const ConstantComparesGatherer &) = delete;
450  ConstantComparesGatherer &
451  operator=(const ConstantComparesGatherer &) = delete;
452 
453 private:
454  /// Try to set the current value used for the comparison, it succeeds only if
455  /// it wasn't set before or if the new value is the same as the old one
456  bool setValueOnce(Value *NewVal) {
457  if (CompValue && CompValue != NewVal)
458  return false;
459  CompValue = NewVal;
460  return (CompValue != nullptr);
461  }
462 
463  /// Try to match Instruction "I" as a comparison against a constant and
464  /// populates the array Vals with the set of values that match (or do not
465  /// match depending on isEQ).
466  /// Return false on failure. On success, the Value the comparison matched
467  /// against is placed in CompValue.
468  /// If CompValue is already set, the function is expected to fail if a match
469  /// is found but the value compared to is different.
470  bool matchInstruction(Instruction *I, bool isEQ) {
471  // If this is an icmp against a constant, handle this as one of the cases.
472  ICmpInst *ICI;
473  ConstantInt *C;
474  if (!((ICI = dyn_cast<ICmpInst>(I)) &&
475  (C = GetConstantInt(I->getOperand(1), DL)))) {
476  return false;
477  }
478 
479  Value *RHSVal;
480  const APInt *RHSC;
481 
482  // Pattern match a special case
483  // (x & ~2^z) == y --> x == y || x == y|2^z
484  // This undoes a transformation done by instcombine to fuse 2 compares.
485  if (ICI->getPredicate() == (isEQ ? ICmpInst::ICMP_EQ : ICmpInst::ICMP_NE)) {
486 
487  // It's a little bit hard to see why the following transformations are
488  // correct. Here is a CVC3 program to verify them for 64-bit values:
489 
490  /*
491  ONE : BITVECTOR(64) = BVZEROEXTEND(0bin1, 63);
492  x : BITVECTOR(64);
493  y : BITVECTOR(64);
494  z : BITVECTOR(64);
495  mask : BITVECTOR(64) = BVSHL(ONE, z);
496  QUERY( (y & ~mask = y) =>
497  ((x & ~mask = y) <=> (x = y OR x = (y | mask)))
498  );
499  QUERY( (y | mask = y) =>
500  ((x | mask = y) <=> (x = y OR x = (y & ~mask)))
501  );
502  */
503 
504  // Please note that each pattern must be a dual implication (<--> or
505  // iff). One directional implication can create spurious matches. If the
506  // implication is only one-way, an unsatisfiable condition on the left
507  // side can imply a satisfiable condition on the right side. Dual
508  // implication ensures that satisfiable conditions are transformed to
509  // other satisfiable conditions and unsatisfiable conditions are
510  // transformed to other unsatisfiable conditions.
511 
512  // Here is a concrete example of a unsatisfiable condition on the left
513  // implying a satisfiable condition on the right:
514  //
515  // mask = (1 << z)
516  // (x & ~mask) == y --> (x == y || x == (y | mask))
517  //
518  // Substituting y = 3, z = 0 yields:
519  // (x & -2) == 3 --> (x == 3 || x == 2)
520 
521  // Pattern match a special case:
522  /*
523  QUERY( (y & ~mask = y) =>
524  ((x & ~mask = y) <=> (x = y OR x = (y | mask)))
525  );
526  */
527  if (match(ICI->getOperand(0),
528  m_And(m_Value(RHSVal), m_APInt(RHSC)))) {
529  APInt Mask = ~*RHSC;
530  if (Mask.isPowerOf2() && (C->getValue() & ~Mask) == C->getValue()) {
531  // If we already have a value for the switch, it has to match!
532  if (!setValueOnce(RHSVal))
533  return false;
534 
535  Vals.push_back(C);
536  Vals.push_back(
538  C->getValue() | Mask));
539  UsedICmps++;
540  return true;
541  }
542  }
543 
544  // Pattern match a special case:
545  /*
546  QUERY( (y | mask = y) =>
547  ((x | mask = y) <=> (x = y OR x = (y & ~mask)))
548  );
549  */
550  if (match(ICI->getOperand(0),
551  m_Or(m_Value(RHSVal), m_APInt(RHSC)))) {
552  APInt Mask = *RHSC;
553  if (Mask.isPowerOf2() && (C->getValue() | Mask) == C->getValue()) {
554  // If we already have a value for the switch, it has to match!
555  if (!setValueOnce(RHSVal))
556  return false;
557 
558  Vals.push_back(C);
560  C->getValue() & ~Mask));
561  UsedICmps++;
562  return true;
563  }
564  }
565 
566  // If we already have a value for the switch, it has to match!
567  if (!setValueOnce(ICI->getOperand(0)))
568  return false;
569 
570  UsedICmps++;
571  Vals.push_back(C);
572  return ICI->getOperand(0);
573  }
574 
575  // If we have "x ult 3", for example, then we can add 0,1,2 to the set.
577  ICI->getPredicate(), C->getValue());
578 
579  // Shift the range if the compare is fed by an add. This is the range
580  // compare idiom as emitted by instcombine.
581  Value *CandidateVal = I->getOperand(0);
582  if (match(I->getOperand(0), m_Add(m_Value(RHSVal), m_APInt(RHSC)))) {
583  Span = Span.subtract(*RHSC);
584  CandidateVal = RHSVal;
585  }
586 
587  // If this is an and/!= check, then we are looking to build the set of
588  // value that *don't* pass the and chain. I.e. to turn "x ugt 2" into
589  // x != 0 && x != 1.
590  if (!isEQ)
591  Span = Span.inverse();
592 
593  // If there are a ton of values, we don't want to make a ginormous switch.
594  if (Span.isSizeLargerThan(8) || Span.isEmptySet()) {
595  return false;
596  }
597 
598  // If we already have a value for the switch, it has to match!
599  if (!setValueOnce(CandidateVal))
600  return false;
601 
602  // Add all values from the range to the set
603  for (APInt Tmp = Span.getLower(); Tmp != Span.getUpper(); ++Tmp)
604  Vals.push_back(ConstantInt::get(I->getContext(), Tmp));
605 
606  UsedICmps++;
607  return true;
608  }
609 
610  /// Given a potentially 'or'd or 'and'd together collection of icmp
611  /// eq/ne/lt/gt instructions that compare a value against a constant, extract
612  /// the value being compared, and stick the list constants into the Vals
613  /// vector.
614  /// One "Extra" case is allowed to differ from the other.
615  void gather(Value *V) {
617  bool isEQ = (I->getOpcode() == Instruction::Or);
618 
619  // Keep a stack (SmallVector for efficiency) for depth-first traversal
621  SmallPtrSet<Value *, 8> Visited;
622 
623  // Initialize
624  Visited.insert(V);
625  DFT.push_back(V);
626 
627  while (!DFT.empty()) {
628  V = DFT.pop_back_val();
629 
630  if (Instruction *I = dyn_cast<Instruction>(V)) {
631  // If it is a || (or && depending on isEQ), process the operands.
632  if (I->getOpcode() == (isEQ ? Instruction::Or : Instruction::And)) {
633  if (Visited.insert(I->getOperand(1)).second)
634  DFT.push_back(I->getOperand(1));
635  if (Visited.insert(I->getOperand(0)).second)
636  DFT.push_back(I->getOperand(0));
637  continue;
638  }
639 
640  // Try to match the current instruction
641  if (matchInstruction(I, isEQ))
642  // Match succeed, continue the loop
643  continue;
644  }
645 
646  // One element of the sequence of || (or &&) could not be match as a
647  // comparison against the same value as the others.
648  // We allow only one "Extra" case to be checked before the switch
649  if (!Extra) {
650  Extra = V;
651  continue;
652  }
653  // Failed to parse a proper sequence, abort now
654  CompValue = nullptr;
655  break;
656  }
657  }
658 };
659 
660 } // end anonymous namespace
661 
663  Instruction *Cond = nullptr;
664  if (SwitchInst *SI = dyn_cast<SwitchInst>(TI)) {
665  Cond = dyn_cast<Instruction>(SI->getCondition());
666  } else if (BranchInst *BI = dyn_cast<BranchInst>(TI)) {
667  if (BI->isConditional())
668  Cond = dyn_cast<Instruction>(BI->getCondition());
669  } else if (IndirectBrInst *IBI = dyn_cast<IndirectBrInst>(TI)) {
670  Cond = dyn_cast<Instruction>(IBI->getAddress());
671  }
672 
673  TI->eraseFromParent();
674  if (Cond)
676 }
677 
678 /// Return true if the specified terminator checks
679 /// to see if a value is equal to constant integer value.
680 Value *SimplifyCFGOpt::isValueEqualityComparison(TerminatorInst *TI) {
681  Value *CV = nullptr;
682  if (SwitchInst *SI = dyn_cast<SwitchInst>(TI)) {
683  // Do not permit merging of large switch instructions into their
684  // predecessors unless there is only one predecessor.
685  if (SI->getNumSuccessors() * std::distance(pred_begin(SI->getParent()),
686  pred_end(SI->getParent())) <=
687  128)
688  CV = SI->getCondition();
689  } else if (BranchInst *BI = dyn_cast<BranchInst>(TI))
690  if (BI->isConditional() && BI->getCondition()->hasOneUse())
691  if (ICmpInst *ICI = dyn_cast<ICmpInst>(BI->getCondition())) {
692  if (ICI->isEquality() && GetConstantInt(ICI->getOperand(1), DL))
693  CV = ICI->getOperand(0);
694  }
695 
696  // Unwrap any lossless ptrtoint cast.
697  if (CV) {
698  if (PtrToIntInst *PTII = dyn_cast<PtrToIntInst>(CV)) {
699  Value *Ptr = PTII->getPointerOperand();
700  if (PTII->getType() == DL.getIntPtrType(Ptr->getType()))
701  CV = Ptr;
702  }
703  }
704  return CV;
705 }
706 
707 /// Given a value comparison instruction,
708 /// decode all of the 'cases' that it represents and return the 'default' block.
709 BasicBlock *SimplifyCFGOpt::GetValueEqualityComparisonCases(
710  TerminatorInst *TI, std::vector<ValueEqualityComparisonCase> &Cases) {
711  if (SwitchInst *SI = dyn_cast<SwitchInst>(TI)) {
712  Cases.reserve(SI->getNumCases());
713  for (auto Case : SI->cases())
714  Cases.push_back(ValueEqualityComparisonCase(Case.getCaseValue(),
715  Case.getCaseSuccessor()));
716  return SI->getDefaultDest();
717  }
718 
719  BranchInst *BI = cast<BranchInst>(TI);
720  ICmpInst *ICI = cast<ICmpInst>(BI->getCondition());
721  BasicBlock *Succ = BI->getSuccessor(ICI->getPredicate() == ICmpInst::ICMP_NE);
722  Cases.push_back(ValueEqualityComparisonCase(
723  GetConstantInt(ICI->getOperand(1), DL), Succ));
724  return BI->getSuccessor(ICI->getPredicate() == ICmpInst::ICMP_EQ);
725 }
726 
727 /// Given a vector of bb/value pairs, remove any entries
728 /// in the list that match the specified block.
729 static void
731  std::vector<ValueEqualityComparisonCase> &Cases) {
732  Cases.erase(std::remove(Cases.begin(), Cases.end(), BB), Cases.end());
733 }
734 
735 /// Return true if there are any keys in C1 that exist in C2 as well.
736 static bool ValuesOverlap(std::vector<ValueEqualityComparisonCase> &C1,
737  std::vector<ValueEqualityComparisonCase> &C2) {
738  std::vector<ValueEqualityComparisonCase> *V1 = &C1, *V2 = &C2;
739 
740  // Make V1 be smaller than V2.
741  if (V1->size() > V2->size())
742  std::swap(V1, V2);
743 
744  if (V1->empty())
745  return false;
746  if (V1->size() == 1) {
747  // Just scan V2.
748  ConstantInt *TheVal = (*V1)[0].Value;
749  for (unsigned i = 0, e = V2->size(); i != e; ++i)
750  if (TheVal == (*V2)[i].Value)
751  return true;
752  }
753 
754  // Otherwise, just sort both lists and compare element by element.
755  array_pod_sort(V1->begin(), V1->end());
756  array_pod_sort(V2->begin(), V2->end());
757  unsigned i1 = 0, i2 = 0, e1 = V1->size(), e2 = V2->size();
758  while (i1 != e1 && i2 != e2) {
759  if ((*V1)[i1].Value == (*V2)[i2].Value)
760  return true;
761  if ((*V1)[i1].Value < (*V2)[i2].Value)
762  ++i1;
763  else
764  ++i2;
765  }
766  return false;
767 }
768 
769 /// If TI is known to be a terminator instruction and its block is known to
770 /// only have a single predecessor block, check to see if that predecessor is
771 /// also a value comparison with the same value, and if that comparison
772 /// determines the outcome of this comparison. If so, simplify TI. This does a
773 /// very limited form of jump threading.
774 bool SimplifyCFGOpt::SimplifyEqualityComparisonWithOnlyPredecessor(
775  TerminatorInst *TI, BasicBlock *Pred, IRBuilder<> &Builder) {
776  Value *PredVal = isValueEqualityComparison(Pred->getTerminator());
777  if (!PredVal)
778  return false; // Not a value comparison in predecessor.
779 
780  Value *ThisVal = isValueEqualityComparison(TI);
781  assert(ThisVal && "This isn't a value comparison!!");
782  if (ThisVal != PredVal)
783  return false; // Different predicates.
784 
785  // TODO: Preserve branch weight metadata, similarly to how
786  // FoldValueComparisonIntoPredecessors preserves it.
787 
788  // Find out information about when control will move from Pred to TI's block.
789  std::vector<ValueEqualityComparisonCase> PredCases;
790  BasicBlock *PredDef =
791  GetValueEqualityComparisonCases(Pred->getTerminator(), PredCases);
792  EliminateBlockCases(PredDef, PredCases); // Remove default from cases.
793 
794  // Find information about how control leaves this block.
795  std::vector<ValueEqualityComparisonCase> ThisCases;
796  BasicBlock *ThisDef = GetValueEqualityComparisonCases(TI, ThisCases);
797  EliminateBlockCases(ThisDef, ThisCases); // Remove default from cases.
798 
799  // If TI's block is the default block from Pred's comparison, potentially
800  // simplify TI based on this knowledge.
801  if (PredDef == TI->getParent()) {
802  // If we are here, we know that the value is none of those cases listed in
803  // PredCases. If there are any cases in ThisCases that are in PredCases, we
804  // can simplify TI.
805  if (!ValuesOverlap(PredCases, ThisCases))
806  return false;
807 
808  if (isa<BranchInst>(TI)) {
809  // Okay, one of the successors of this condbr is dead. Convert it to a
810  // uncond br.
811  assert(ThisCases.size() == 1 && "Branch can only have one case!");
812  // Insert the new branch.
813  Instruction *NI = Builder.CreateBr(ThisDef);
814  (void)NI;
815 
816  // Remove PHI node entries for the dead edge.
817  ThisCases[0].Dest->removePredecessor(TI->getParent());
818 
819  DEBUG(dbgs() << "Threading pred instr: " << *Pred->getTerminator()
820  << "Through successor TI: " << *TI << "Leaving: " << *NI
821  << "\n");
822 
824  return true;
825  }
826 
827  SwitchInst *SI = cast<SwitchInst>(TI);
828  // Okay, TI has cases that are statically dead, prune them away.
829  SmallPtrSet<Constant *, 16> DeadCases;
830  for (unsigned i = 0, e = PredCases.size(); i != e; ++i)
831  DeadCases.insert(PredCases[i].Value);
832 
833  DEBUG(dbgs() << "Threading pred instr: " << *Pred->getTerminator()
834  << "Through successor TI: " << *TI);
835 
836  // Collect branch weights into a vector.
837  SmallVector<uint32_t, 8> Weights;
839  bool HasWeight = MD && (MD->getNumOperands() == 2 + SI->getNumCases());
840  if (HasWeight)
841  for (unsigned MD_i = 1, MD_e = MD->getNumOperands(); MD_i < MD_e;
842  ++MD_i) {
843  ConstantInt *CI = mdconst::extract<ConstantInt>(MD->getOperand(MD_i));
844  Weights.push_back(CI->getValue().getZExtValue());
845  }
846  for (SwitchInst::CaseIt i = SI->case_end(), e = SI->case_begin(); i != e;) {
847  --i;
848  if (DeadCases.count(i->getCaseValue())) {
849  if (HasWeight) {
850  std::swap(Weights[i->getCaseIndex() + 1], Weights.back());
851  Weights.pop_back();
852  }
853  i->getCaseSuccessor()->removePredecessor(TI->getParent());
854  SI->removeCase(i);
855  }
856  }
857  if (HasWeight && Weights.size() >= 2)
859  MDBuilder(SI->getParent()->getContext())
860  .createBranchWeights(Weights));
861 
862  DEBUG(dbgs() << "Leaving: " << *TI << "\n");
863  return true;
864  }
865 
866  // Otherwise, TI's block must correspond to some matched value. Find out
867  // which value (or set of values) this is.
868  ConstantInt *TIV = nullptr;
869  BasicBlock *TIBB = TI->getParent();
870  for (unsigned i = 0, e = PredCases.size(); i != e; ++i)
871  if (PredCases[i].Dest == TIBB) {
872  if (TIV)
873  return false; // Cannot handle multiple values coming to this block.
874  TIV = PredCases[i].Value;
875  }
876  assert(TIV && "No edge from pred to succ?");
877 
878  // Okay, we found the one constant that our value can be if we get into TI's
879  // BB. Find out which successor will unconditionally be branched to.
880  BasicBlock *TheRealDest = nullptr;
881  for (unsigned i = 0, e = ThisCases.size(); i != e; ++i)
882  if (ThisCases[i].Value == TIV) {
883  TheRealDest = ThisCases[i].Dest;
884  break;
885  }
886 
887  // If not handled by any explicit cases, it is handled by the default case.
888  if (!TheRealDest)
889  TheRealDest = ThisDef;
890 
891  // Remove PHI node entries for dead edges.
892  BasicBlock *CheckEdge = TheRealDest;
893  for (BasicBlock *Succ : successors(TIBB))
894  if (Succ != CheckEdge)
895  Succ->removePredecessor(TIBB);
896  else
897  CheckEdge = nullptr;
898 
899  // Insert the new branch.
900  Instruction *NI = Builder.CreateBr(TheRealDest);
901  (void)NI;
902 
903  DEBUG(dbgs() << "Threading pred instr: " << *Pred->getTerminator()
904  << "Through successor TI: " << *TI << "Leaving: " << *NI
905  << "\n");
906 
908  return true;
909 }
910 
911 namespace {
912 
913 /// This class implements a stable ordering of constant
914 /// integers that does not depend on their address. This is important for
915 /// applications that sort ConstantInt's to ensure uniqueness.
916 struct ConstantIntOrdering {
917  bool operator()(const ConstantInt *LHS, const ConstantInt *RHS) const {
918  return LHS->getValue().ult(RHS->getValue());
919  }
920 };
921 
922 } // end anonymous namespace
923 
924 static int ConstantIntSortPredicate(ConstantInt *const *P1,
925  ConstantInt *const *P2) {
926  const ConstantInt *LHS = *P1;
927  const ConstantInt *RHS = *P2;
928  if (LHS == RHS)
929  return 0;
930  return LHS->getValue().ult(RHS->getValue()) ? 1 : -1;
931 }
932 
933 static inline bool HasBranchWeights(const Instruction *I) {
935  if (ProfMD && ProfMD->getOperand(0))
936  if (MDString *MDS = dyn_cast<MDString>(ProfMD->getOperand(0)))
937  return MDS->getString().equals("branch_weights");
938 
939  return false;
940 }
941 
942 /// Get Weights of a given TerminatorInst, the default weight is at the front
943 /// of the vector. If TI is a conditional eq, we need to swap the branch-weight
944 /// metadata.
946  SmallVectorImpl<uint64_t> &Weights) {
948  assert(MD);
949  for (unsigned i = 1, e = MD->getNumOperands(); i < e; ++i) {
950  ConstantInt *CI = mdconst::extract<ConstantInt>(MD->getOperand(i));
951  Weights.push_back(CI->getValue().getZExtValue());
952  }
953 
954  // If TI is a conditional eq, the default case is the false case,
955  // and the corresponding branch-weight data is at index 2. We swap the
956  // default weight to be the first entry.
957  if (BranchInst *BI = dyn_cast<BranchInst>(TI)) {
958  assert(Weights.size() == 2);
959  ICmpInst *ICI = cast<ICmpInst>(BI->getCondition());
960  if (ICI->getPredicate() == ICmpInst::ICMP_EQ)
961  std::swap(Weights.front(), Weights.back());
962  }
963 }
964 
965 /// Keep halving the weights until all can fit in uint32_t.
966 static void FitWeights(MutableArrayRef<uint64_t> Weights) {
967  uint64_t Max = *std::max_element(Weights.begin(), Weights.end());
968  if (Max > UINT_MAX) {
969  unsigned Offset = 32 - countLeadingZeros(Max);
970  for (uint64_t &I : Weights)
971  I >>= Offset;
972  }
973 }
974 
975 /// The specified terminator is a value equality comparison instruction
976 /// (either a switch or a branch on "X == c").
977 /// See if any of the predecessors of the terminator block are value comparisons
978 /// on the same value. If so, and if safe to do so, fold them together.
979 bool SimplifyCFGOpt::FoldValueComparisonIntoPredecessors(TerminatorInst *TI,
980  IRBuilder<> &Builder) {
981  BasicBlock *BB = TI->getParent();
982  Value *CV = isValueEqualityComparison(TI); // CondVal
983  assert(CV && "Not a comparison?");
984  bool Changed = false;
985 
987  while (!Preds.empty()) {
988  BasicBlock *Pred = Preds.pop_back_val();
989 
990  // See if the predecessor is a comparison with the same value.
991  TerminatorInst *PTI = Pred->getTerminator();
992  Value *PCV = isValueEqualityComparison(PTI); // PredCondVal
993 
994  if (PCV == CV && TI != PTI) {
996  if (!SafeToMergeTerminators(TI, PTI, &FailBlocks)) {
997  for (auto *Succ : FailBlocks) {
998  if (!SplitBlockPredecessors(Succ, TI->getParent(), ".fold.split"))
999  return false;
1000  }
1001  }
1002 
1003  // Figure out which 'cases' to copy from SI to PSI.
1004  std::vector<ValueEqualityComparisonCase> BBCases;
1005  BasicBlock *BBDefault = GetValueEqualityComparisonCases(TI, BBCases);
1006 
1007  std::vector<ValueEqualityComparisonCase> PredCases;
1008  BasicBlock *PredDefault = GetValueEqualityComparisonCases(PTI, PredCases);
1009 
1010  // Based on whether the default edge from PTI goes to BB or not, fill in
1011  // PredCases and PredDefault with the new switch cases we would like to
1012  // build.
1013  SmallVector<BasicBlock *, 8> NewSuccessors;
1014 
1015  // Update the branch weight metadata along the way
1016  SmallVector<uint64_t, 8> Weights;
1017  bool PredHasWeights = HasBranchWeights(PTI);
1018  bool SuccHasWeights = HasBranchWeights(TI);
1019 
1020  if (PredHasWeights) {
1021  GetBranchWeights(PTI, Weights);
1022  // branch-weight metadata is inconsistent here.
1023  if (Weights.size() != 1 + PredCases.size())
1024  PredHasWeights = SuccHasWeights = false;
1025  } else if (SuccHasWeights)
1026  // If there are no predecessor weights but there are successor weights,
1027  // populate Weights with 1, which will later be scaled to the sum of
1028  // successor's weights
1029  Weights.assign(1 + PredCases.size(), 1);
1030 
1031  SmallVector<uint64_t, 8> SuccWeights;
1032  if (SuccHasWeights) {
1033  GetBranchWeights(TI, SuccWeights);
1034  // branch-weight metadata is inconsistent here.
1035  if (SuccWeights.size() != 1 + BBCases.size())
1036  PredHasWeights = SuccHasWeights = false;
1037  } else if (PredHasWeights)
1038  SuccWeights.assign(1 + BBCases.size(), 1);
1039 
1040  if (PredDefault == BB) {
1041  // If this is the default destination from PTI, only the edges in TI
1042  // that don't occur in PTI, or that branch to BB will be activated.
1043  std::set<ConstantInt *, ConstantIntOrdering> PTIHandled;
1044  for (unsigned i = 0, e = PredCases.size(); i != e; ++i)
1045  if (PredCases[i].Dest != BB)
1046  PTIHandled.insert(PredCases[i].Value);
1047  else {
1048  // The default destination is BB, we don't need explicit targets.
1049  std::swap(PredCases[i], PredCases.back());
1050 
1051  if (PredHasWeights || SuccHasWeights) {
1052  // Increase weight for the default case.
1053  Weights[0] += Weights[i + 1];
1054  std::swap(Weights[i + 1], Weights.back());
1055  Weights.pop_back();
1056  }
1057 
1058  PredCases.pop_back();
1059  --i;
1060  --e;
1061  }
1062 
1063  // Reconstruct the new switch statement we will be building.
1064  if (PredDefault != BBDefault) {
1065  PredDefault->removePredecessor(Pred);
1066  PredDefault = BBDefault;
1067  NewSuccessors.push_back(BBDefault);
1068  }
1069 
1070  unsigned CasesFromPred = Weights.size();
1071  uint64_t ValidTotalSuccWeight = 0;
1072  for (unsigned i = 0, e = BBCases.size(); i != e; ++i)
1073  if (!PTIHandled.count(BBCases[i].Value) &&
1074  BBCases[i].Dest != BBDefault) {
1075  PredCases.push_back(BBCases[i]);
1076  NewSuccessors.push_back(BBCases[i].Dest);
1077  if (SuccHasWeights || PredHasWeights) {
1078  // The default weight is at index 0, so weight for the ith case
1079  // should be at index i+1. Scale the cases from successor by
1080  // PredDefaultWeight (Weights[0]).
1081  Weights.push_back(Weights[0] * SuccWeights[i + 1]);
1082  ValidTotalSuccWeight += SuccWeights[i + 1];
1083  }
1084  }
1085 
1086  if (SuccHasWeights || PredHasWeights) {
1087  ValidTotalSuccWeight += SuccWeights[0];
1088  // Scale the cases from predecessor by ValidTotalSuccWeight.
1089  for (unsigned i = 1; i < CasesFromPred; ++i)
1090  Weights[i] *= ValidTotalSuccWeight;
1091  // Scale the default weight by SuccDefaultWeight (SuccWeights[0]).
1092  Weights[0] *= SuccWeights[0];
1093  }
1094  } else {
1095  // If this is not the default destination from PSI, only the edges
1096  // in SI that occur in PSI with a destination of BB will be
1097  // activated.
1098  std::set<ConstantInt *, ConstantIntOrdering> PTIHandled;
1099  std::map<ConstantInt *, uint64_t> WeightsForHandled;
1100  for (unsigned i = 0, e = PredCases.size(); i != e; ++i)
1101  if (PredCases[i].Dest == BB) {
1102  PTIHandled.insert(PredCases[i].Value);
1103 
1104  if (PredHasWeights || SuccHasWeights) {
1105  WeightsForHandled[PredCases[i].Value] = Weights[i + 1];
1106  std::swap(Weights[i + 1], Weights.back());
1107  Weights.pop_back();
1108  }
1109 
1110  std::swap(PredCases[i], PredCases.back());
1111  PredCases.pop_back();
1112  --i;
1113  --e;
1114  }
1115 
1116  // Okay, now we know which constants were sent to BB from the
1117  // predecessor. Figure out where they will all go now.
1118  for (unsigned i = 0, e = BBCases.size(); i != e; ++i)
1119  if (PTIHandled.count(BBCases[i].Value)) {
1120  // If this is one we are capable of getting...
1121  if (PredHasWeights || SuccHasWeights)
1122  Weights.push_back(WeightsForHandled[BBCases[i].Value]);
1123  PredCases.push_back(BBCases[i]);
1124  NewSuccessors.push_back(BBCases[i].Dest);
1125  PTIHandled.erase(
1126  BBCases[i].Value); // This constant is taken care of
1127  }
1128 
1129  // If there are any constants vectored to BB that TI doesn't handle,
1130  // they must go to the default destination of TI.
1131  for (ConstantInt *I : PTIHandled) {
1132  if (PredHasWeights || SuccHasWeights)
1133  Weights.push_back(WeightsForHandled[I]);
1134  PredCases.push_back(ValueEqualityComparisonCase(I, BBDefault));
1135  NewSuccessors.push_back(BBDefault);
1136  }
1137  }
1138 
1139  // Okay, at this point, we know which new successor Pred will get. Make
1140  // sure we update the number of entries in the PHI nodes for these
1141  // successors.
1142  for (BasicBlock *NewSuccessor : NewSuccessors)
1143  AddPredecessorToBlock(NewSuccessor, Pred, BB);
1144 
1145  Builder.SetInsertPoint(PTI);
1146  // Convert pointer to int before we switch.
1147  if (CV->getType()->isPointerTy()) {
1148  CV = Builder.CreatePtrToInt(CV, DL.getIntPtrType(CV->getType()),
1149  "magicptr");
1150  }
1151 
1152  // Now that the successors are updated, create the new Switch instruction.
1153  SwitchInst *NewSI =
1154  Builder.CreateSwitch(CV, PredDefault, PredCases.size());
1155  NewSI->setDebugLoc(PTI->getDebugLoc());
1156  for (ValueEqualityComparisonCase &V : PredCases)
1157  NewSI->addCase(V.Value, V.Dest);
1158 
1159  if (PredHasWeights || SuccHasWeights) {
1160  // Halve the weights if any of them cannot fit in an uint32_t
1161  FitWeights(Weights);
1162 
1163  SmallVector<uint32_t, 8> MDWeights(Weights.begin(), Weights.end());
1164 
1165  NewSI->setMetadata(
1167  MDBuilder(BB->getContext()).createBranchWeights(MDWeights));
1168  }
1169 
1171 
1172  // Okay, last check. If BB is still a successor of PSI, then we must
1173  // have an infinite loop case. If so, add an infinitely looping block
1174  // to handle the case to preserve the behavior of the code.
1175  BasicBlock *InfLoopBlock = nullptr;
1176  for (unsigned i = 0, e = NewSI->getNumSuccessors(); i != e; ++i)
1177  if (NewSI->getSuccessor(i) == BB) {
1178  if (!InfLoopBlock) {
1179  // Insert it at the end of the function, because it's either code,
1180  // or it won't matter if it's hot. :)
1181  InfLoopBlock = BasicBlock::Create(BB->getContext(), "infloop",
1182  BB->getParent());
1183  BranchInst::Create(InfLoopBlock, InfLoopBlock);
1184  }
1185  NewSI->setSuccessor(i, InfLoopBlock);
1186  }
1187 
1188  Changed = true;
1189  }
1190  }
1191  return Changed;
1192 }
1193 
1194 // If we would need to insert a select that uses the value of this invoke
1195 // (comments in HoistThenElseCodeToIf explain why we would need to do this), we
1196 // can't hoist the invoke, as there is nowhere to put the select in this case.
1198  Instruction *I1, Instruction *I2) {
1199  for (BasicBlock *Succ : successors(BB1)) {
1200  PHINode *PN;
1201  for (BasicBlock::iterator BBI = Succ->begin();
1202  (PN = dyn_cast<PHINode>(BBI)); ++BBI) {
1203  Value *BB1V = PN->getIncomingValueForBlock(BB1);
1204  Value *BB2V = PN->getIncomingValueForBlock(BB2);
1205  if (BB1V != BB2V && (BB1V == I1 || BB2V == I2)) {
1206  return false;
1207  }
1208  }
1209  }
1210  return true;
1211 }
1212 
1214 
1215 /// Given a conditional branch that goes to BB1 and BB2, hoist any common code
1216 /// in the two blocks up into the branch block. The caller of this function
1217 /// guarantees that BI's block dominates BB1 and BB2.
1219  const TargetTransformInfo &TTI) {
1220  // This does very trivial matching, with limited scanning, to find identical
1221  // instructions in the two blocks. In particular, we don't want to get into
1222  // O(M*N) situations here where M and N are the sizes of BB1 and BB2. As
1223  // such, we currently just scan for obviously identical instructions in an
1224  // identical order.
1225  BasicBlock *BB1 = BI->getSuccessor(0); // The true destination.
1226  BasicBlock *BB2 = BI->getSuccessor(1); // The false destination
1227 
1228  BasicBlock::iterator BB1_Itr = BB1->begin();
1229  BasicBlock::iterator BB2_Itr = BB2->begin();
1230 
1231  Instruction *I1 = &*BB1_Itr++, *I2 = &*BB2_Itr++;
1232  // Skip debug info if it is not identical.
1235  if (!DBI1 || !DBI2 || !DBI1->isIdenticalToWhenDefined(DBI2)) {
1236  while (isa<DbgInfoIntrinsic>(I1))
1237  I1 = &*BB1_Itr++;
1238  while (isa<DbgInfoIntrinsic>(I2))
1239  I2 = &*BB2_Itr++;
1240  }
1241  if (isa<PHINode>(I1) || !I1->isIdenticalToWhenDefined(I2) ||
1242  (isa<InvokeInst>(I1) && !isSafeToHoistInvoke(BB1, BB2, I1, I2)))
1243  return false;
1244 
1245  BasicBlock *BIParent = BI->getParent();
1246 
1247  bool Changed = false;
1248  do {
1249  // If we are hoisting the terminator instruction, don't move one (making a
1250  // broken BB), instead clone it, and remove BI.
1251  if (isa<TerminatorInst>(I1))
1252  goto HoistTerminator;
1253 
1254  if (!TTI.isProfitableToHoist(I1) || !TTI.isProfitableToHoist(I2))
1255  return Changed;
1256 
1257  // For a normal instruction, we just move one to right before the branch,
1258  // then replace all uses of the other with the first. Finally, we remove
1259  // the now redundant second instruction.
1260  BIParent->getInstList().splice(BI->getIterator(), BB1->getInstList(), I1);
1261  if (!I2->use_empty())
1262  I2->replaceAllUsesWith(I1);
1263  I1->andIRFlags(I2);
1264  unsigned KnownIDs[] = {LLVMContext::MD_tbaa,
1274  combineMetadata(I1, I2, KnownIDs);
1275 
1276  // I1 and I2 are being combined into a single instruction. Its debug
1277  // location is the merged locations of the original instructions.
1278  I1->applyMergedLocation(I1->getDebugLoc(), I2->getDebugLoc());
1279 
1280  I2->eraseFromParent();
1281  Changed = true;
1282 
1283  I1 = &*BB1_Itr++;
1284  I2 = &*BB2_Itr++;
1285  // Skip debug info if it is not identical.
1288  if (!DBI1 || !DBI2 || !DBI1->isIdenticalToWhenDefined(DBI2)) {
1289  while (isa<DbgInfoIntrinsic>(I1))
1290  I1 = &*BB1_Itr++;
1291  while (isa<DbgInfoIntrinsic>(I2))
1292  I2 = &*BB2_Itr++;
1293  }
1294  } while (I1->isIdenticalToWhenDefined(I2));
1295 
1296  return true;
1297 
1298 HoistTerminator:
1299  // It may not be possible to hoist an invoke.
1300  if (isa<InvokeInst>(I1) && !isSafeToHoistInvoke(BB1, BB2, I1, I2))
1301  return Changed;
1302 
1303  for (BasicBlock *Succ : successors(BB1)) {
1304  PHINode *PN;
1305  for (BasicBlock::iterator BBI = Succ->begin();
1306  (PN = dyn_cast<PHINode>(BBI)); ++BBI) {
1307  Value *BB1V = PN->getIncomingValueForBlock(BB1);
1308  Value *BB2V = PN->getIncomingValueForBlock(BB2);
1309  if (BB1V == BB2V)
1310  continue;
1311 
1312  // Check for passingValueIsAlwaysUndefined here because we would rather
1313  // eliminate undefined control flow then converting it to a select.
1314  if (passingValueIsAlwaysUndefined(BB1V, PN) ||
1316  return Changed;
1317 
1318  if (isa<ConstantExpr>(BB1V) && !isSafeToSpeculativelyExecute(BB1V))
1319  return Changed;
1320  if (isa<ConstantExpr>(BB2V) && !isSafeToSpeculativelyExecute(BB2V))
1321  return Changed;
1322  }
1323  }
1324 
1325  // Okay, it is safe to hoist the terminator.
1326  Instruction *NT = I1->clone();
1327  BIParent->getInstList().insert(BI->getIterator(), NT);
1328  if (!NT->getType()->isVoidTy()) {
1329  I1->replaceAllUsesWith(NT);
1330  I2->replaceAllUsesWith(NT);
1331  NT->takeName(I1);
1332  }
1333 
1334  IRBuilder<NoFolder> Builder(NT);
1335  // Hoisting one of the terminators from our successor is a great thing.
1336  // Unfortunately, the successors of the if/else blocks may have PHI nodes in
1337  // them. If they do, all PHI entries for BB1/BB2 must agree for all PHI
1338  // nodes, so we insert select instruction to compute the final result.
1339  std::map<std::pair<Value *, Value *>, SelectInst *> InsertedSelects;
1340  for (BasicBlock *Succ : successors(BB1)) {
1341  PHINode *PN;
1342  for (BasicBlock::iterator BBI = Succ->begin();
1343  (PN = dyn_cast<PHINode>(BBI)); ++BBI) {
1344  Value *BB1V = PN->getIncomingValueForBlock(BB1);
1345  Value *BB2V = PN->getIncomingValueForBlock(BB2);
1346  if (BB1V == BB2V)
1347  continue;
1348 
1349  // These values do not agree. Insert a select instruction before NT
1350  // that determines the right value.
1351  SelectInst *&SI = InsertedSelects[std::make_pair(BB1V, BB2V)];
1352  if (!SI)
1353  SI = cast<SelectInst>(
1354  Builder.CreateSelect(BI->getCondition(), BB1V, BB2V,
1355  BB1V->getName() + "." + BB2V->getName(), BI));
1356 
1357  // Make the PHI node use the select for all incoming values for BB1/BB2
1358  for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i)
1359  if (PN->getIncomingBlock(i) == BB1 || PN->getIncomingBlock(i) == BB2)
1360  PN->setIncomingValue(i, SI);
1361  }
1362  }
1363 
1364  // Update any PHI nodes in our new successors.
1365  for (BasicBlock *Succ : successors(BB1))
1366  AddPredecessorToBlock(Succ, BIParent, BB1);
1367 
1369  return true;
1370 }
1371 
1372 // All instructions in Insts belong to different blocks that all unconditionally
1373 // branch to a common successor. Analyze each instruction and return true if it
1374 // would be possible to sink them into their successor, creating one common
1375 // instruction instead. For every value that would be required to be provided by
1376 // PHI node (because an operand varies in each input block), add to PHIOperands.
1379  DenseMap<Instruction *, SmallVector<Value *, 4>> &PHIOperands) {
1380  // Prune out obviously bad instructions to move. Any non-store instruction
1381  // must have exactly one use, and we check later that use is by a single,
1382  // common PHI instruction in the successor.
1383  for (auto *I : Insts) {
1384  // These instructions may change or break semantics if moved.
1385  if (isa<PHINode>(I) || I->isEHPad() || isa<AllocaInst>(I) ||
1386  I->getType()->isTokenTy())
1387  return false;
1388 
1389  // Conservatively return false if I is an inline-asm instruction. Sinking
1390  // and merging inline-asm instructions can potentially create arguments
1391  // that cannot satisfy the inline-asm constraints.
1392  if (const auto *C = dyn_cast<CallInst>(I))
1393  if (C->isInlineAsm())
1394  return false;
1395 
1396  // Everything must have only one use too, apart from stores which
1397  // have no uses.
1398  if (!isa<StoreInst>(I) && !I->hasOneUse())
1399  return false;
1400  }
1401 
1402  const Instruction *I0 = Insts.front();
1403  for (auto *I : Insts)
1404  if (!I->isSameOperationAs(I0))
1405  return false;
1406 
1407  // All instructions in Insts are known to be the same opcode. If they aren't
1408  // stores, check the only user of each is a PHI or in the same block as the
1409  // instruction, because if a user is in the same block as an instruction
1410  // we're contemplating sinking, it must already be determined to be sinkable.
1411  if (!isa<StoreInst>(I0)) {
1412  auto *PNUse = dyn_cast<PHINode>(*I0->user_begin());
1413  auto *Succ = I0->getParent()->getTerminator()->getSuccessor(0);
1414  if (!all_of(Insts, [&PNUse,&Succ](const Instruction *I) -> bool {
1415  auto *U = cast<Instruction>(*I->user_begin());
1416  return (PNUse &&
1417  PNUse->getParent() == Succ &&
1418  PNUse->getIncomingValueForBlock(I->getParent()) == I) ||
1419  U->getParent() == I->getParent();
1420  }))
1421  return false;
1422  }
1423 
1424  // Because SROA can't handle speculating stores of selects, try not
1425  // to sink loads or stores of allocas when we'd have to create a PHI for
1426  // the address operand. Also, because it is likely that loads or stores
1427  // of allocas will disappear when Mem2Reg/SROA is run, don't sink them.
1428  // This can cause code churn which can have unintended consequences down
1429  // the line - see https://llvm.org/bugs/show_bug.cgi?id=30244.
1430  // FIXME: This is a workaround for a deficiency in SROA - see
1431  // https://llvm.org/bugs/show_bug.cgi?id=30188
1432  if (isa<StoreInst>(I0) && any_of(Insts, [](const Instruction *I) {
1433  return isa<AllocaInst>(I->getOperand(1));
1434  }))
1435  return false;
1436  if (isa<LoadInst>(I0) && any_of(Insts, [](const Instruction *I) {
1437  return isa<AllocaInst>(I->getOperand(0));
1438  }))
1439  return false;
1440 
1441  for (unsigned OI = 0, OE = I0->getNumOperands(); OI != OE; ++OI) {
1442  if (I0->getOperand(OI)->getType()->isTokenTy())
1443  // Don't touch any operand of token type.
1444  return false;
1445 
1446  auto SameAsI0 = [&I0, OI](const Instruction *I) {
1447  assert(I->getNumOperands() == I0->getNumOperands());
1448  return I->getOperand(OI) == I0->getOperand(OI);
1449  };
1450  if (!all_of(Insts, SameAsI0)) {
1451  if (!canReplaceOperandWithVariable(I0, OI))
1452  // We can't create a PHI from this GEP.
1453  return false;
1454  // Don't create indirect calls! The called value is the final operand.
1455  if ((isa<CallInst>(I0) || isa<InvokeInst>(I0)) && OI == OE - 1) {
1456  // FIXME: if the call was *already* indirect, we should do this.
1457  return false;
1458  }
1459  for (auto *I : Insts)
1460  PHIOperands[I].push_back(I->getOperand(OI));
1461  }
1462  }
1463  return true;
1464 }
1465 
1466 // Assuming canSinkLastInstruction(Blocks) has returned true, sink the last
1467 // instruction of every block in Blocks to their common successor, commoning
1468 // into one instruction.
1470  auto *BBEnd = Blocks[0]->getTerminator()->getSuccessor(0);
1471 
1472  // canSinkLastInstruction returning true guarantees that every block has at
1473  // least one non-terminator instruction.
1475  for (auto *BB : Blocks) {
1476  Instruction *I = BB->getTerminator();
1477  do {
1478  I = I->getPrevNode();
1479  } while (isa<DbgInfoIntrinsic>(I) && I != &BB->front());
1480  if (!isa<DbgInfoIntrinsic>(I))
1481  Insts.push_back(I);
1482  }
1483 
1484  // The only checking we need to do now is that all users of all instructions
1485  // are the same PHI node. canSinkLastInstruction should have checked this but
1486  // it is slightly over-aggressive - it gets confused by commutative instructions
1487  // so double-check it here.
1488  Instruction *I0 = Insts.front();
1489  if (!isa<StoreInst>(I0)) {
1490  auto *PNUse = dyn_cast<PHINode>(*I0->user_begin());
1491  if (!all_of(Insts, [&PNUse](const Instruction *I) -> bool {
1492  auto *U = cast<Instruction>(*I->user_begin());
1493  return U == PNUse;
1494  }))
1495  return false;
1496  }
1497 
1498  // We don't need to do any more checking here; canSinkLastInstruction should
1499  // have done it all for us.
1500  SmallVector<Value*, 4> NewOperands;
1501  for (unsigned O = 0, E = I0->getNumOperands(); O != E; ++O) {
1502  // This check is different to that in canSinkLastInstruction. There, we
1503  // cared about the global view once simplifycfg (and instcombine) have
1504  // completed - it takes into account PHIs that become trivially
1505  // simplifiable. However here we need a more local view; if an operand
1506  // differs we create a PHI and rely on instcombine to clean up the very
1507  // small mess we may make.
1508  bool NeedPHI = any_of(Insts, [&I0, O](const Instruction *I) {
1509  return I->getOperand(O) != I0->getOperand(O);
1510  });
1511  if (!NeedPHI) {
1512  NewOperands.push_back(I0->getOperand(O));
1513  continue;
1514  }
1515 
1516  // Create a new PHI in the successor block and populate it.
1517  auto *Op = I0->getOperand(O);
1518  assert(!Op->getType()->isTokenTy() && "Can't PHI tokens!");
1519  auto *PN = PHINode::Create(Op->getType(), Insts.size(),
1520  Op->getName() + ".sink", &BBEnd->front());
1521  for (auto *I : Insts)
1522  PN->addIncoming(I->getOperand(O), I->getParent());
1523  NewOperands.push_back(PN);
1524  }
1525 
1526  // Arbitrarily use I0 as the new "common" instruction; remap its operands
1527  // and move it to the start of the successor block.
1528  for (unsigned O = 0, E = I0->getNumOperands(); O != E; ++O)
1529  I0->getOperandUse(O).set(NewOperands[O]);
1530  I0->moveBefore(&*BBEnd->getFirstInsertionPt());
1531 
1532  // Update metadata and IR flags, and merge debug locations.
1533  for (auto *I : Insts)
1534  if (I != I0) {
1535  // The debug location for the "common" instruction is the merged locations
1536  // of all the commoned instructions. We start with the original location
1537  // of the "common" instruction and iteratively merge each location in the
1538  // loop below.
1539  // This is an N-way merge, which will be inefficient if I0 is a CallInst.
1540  // However, as N-way merge for CallInst is rare, so we use simplified API
1541  // instead of using complex API for N-way merge.
1542  I0->applyMergedLocation(I0->getDebugLoc(), I->getDebugLoc());
1543  combineMetadataForCSE(I0, I);
1544  I0->andIRFlags(I);
1545  }
1546 
1547  if (!isa<StoreInst>(I0)) {
1548  // canSinkLastInstruction checked that all instructions were used by
1549  // one and only one PHI node. Find that now, RAUW it to our common
1550  // instruction and nuke it.
1551  assert(I0->hasOneUse());
1552  auto *PN = cast<PHINode>(*I0->user_begin());
1553  PN->replaceAllUsesWith(I0);
1554  PN->eraseFromParent();
1555  }
1556 
1557  // Finally nuke all instructions apart from the common instruction.
1558  for (auto *I : Insts)
1559  if (I != I0)
1560  I->eraseFromParent();
1561 
1562  return true;
1563 }
1564 
1565 namespace {
1566 
1567  // LockstepReverseIterator - Iterates through instructions
1568  // in a set of blocks in reverse order from the first non-terminator.
1569  // For example (assume all blocks have size n):
1570  // LockstepReverseIterator I([B1, B2, B3]);
1571  // *I-- = [B1[n], B2[n], B3[n]];
1572  // *I-- = [B1[n-1], B2[n-1], B3[n-1]];
1573  // *I-- = [B1[n-2], B2[n-2], B3[n-2]];
1574  // ...
1575  class LockstepReverseIterator {
1576  ArrayRef<BasicBlock*> Blocks;
1578  bool Fail;
1579  public:
1580  LockstepReverseIterator(ArrayRef<BasicBlock*> Blocks) :
1581  Blocks(Blocks) {
1582  reset();
1583  }
1584 
1585  void reset() {
1586  Fail = false;
1587  Insts.clear();
1588  for (auto *BB : Blocks) {
1589  Instruction *Inst = BB->getTerminator();
1590  for (Inst = Inst->getPrevNode(); Inst && isa<DbgInfoIntrinsic>(Inst);)
1591  Inst = Inst->getPrevNode();
1592  if (!Inst) {
1593  // Block wasn't big enough.
1594  Fail = true;
1595  return;
1596  }
1597  Insts.push_back(Inst);
1598  }
1599  }
1600 
1601  bool isValid() const {
1602  return !Fail;
1603  }
1604 
1605  void operator -- () {
1606  if (Fail)
1607  return;
1608  for (auto *&Inst : Insts) {
1609  for (Inst = Inst->getPrevNode(); Inst && isa<DbgInfoIntrinsic>(Inst);)
1610  Inst = Inst->getPrevNode();
1611  // Already at beginning of block.
1612  if (!Inst) {
1613  Fail = true;
1614  return;
1615  }
1616  }
1617  }
1618 
1620  return Insts;
1621  }
1622  };
1623 
1624 } // end anonymous namespace
1625 
1626 /// Given an unconditional branch that goes to BBEnd,
1627 /// check whether BBEnd has only two predecessors and the other predecessor
1628 /// ends with an unconditional branch. If it is true, sink any common code
1629 /// in the two predecessors to BBEnd.
1631  assert(BI1->isUnconditional());
1632  BasicBlock *BBEnd = BI1->getSuccessor(0);
1633 
1634  // We support two situations:
1635  // (1) all incoming arcs are unconditional
1636  // (2) one incoming arc is conditional
1637  //
1638  // (2) is very common in switch defaults and
1639  // else-if patterns;
1640  //
1641  // if (a) f(1);
1642  // else if (b) f(2);
1643  //
1644  // produces:
1645  //
1646  // [if]
1647  // / \
1648  // [f(1)] [if]
1649  // | | \
1650  // | | |
1651  // | [f(2)]|
1652  // \ | /
1653  // [ end ]
1654  //
1655  // [end] has two unconditional predecessor arcs and one conditional. The
1656  // conditional refers to the implicit empty 'else' arc. This conditional
1657  // arc can also be caused by an empty default block in a switch.
1658  //
1659  // In this case, we attempt to sink code from all *unconditional* arcs.
1660  // If we can sink instructions from these arcs (determined during the scan
1661  // phase below) we insert a common successor for all unconditional arcs and
1662  // connect that to [end], to enable sinking:
1663  //
1664  // [if]
1665  // / \
1666  // [x(1)] [if]
1667  // | | \
1668  // | | \
1669  // | [x(2)] |
1670  // \ / |
1671  // [sink.split] |
1672  // \ /
1673  // [ end ]
1674  //
1675  SmallVector<BasicBlock*,4> UnconditionalPreds;
1676  Instruction *Cond = nullptr;
1677  for (auto *B : predecessors(BBEnd)) {
1678  auto *T = B->getTerminator();
1679  if (isa<BranchInst>(T) && cast<BranchInst>(T)->isUnconditional())
1680  UnconditionalPreds.push_back(B);
1681  else if ((isa<BranchInst>(T) || isa<SwitchInst>(T)) && !Cond)
1682  Cond = T;
1683  else
1684  return false;
1685  }
1686  if (UnconditionalPreds.size() < 2)
1687  return false;
1688 
1689  bool Changed = false;
1690  // We take a two-step approach to tail sinking. First we scan from the end of
1691  // each block upwards in lockstep. If the n'th instruction from the end of each
1692  // block can be sunk, those instructions are added to ValuesToSink and we
1693  // carry on. If we can sink an instruction but need to PHI-merge some operands
1694  // (because they're not identical in each instruction) we add these to
1695  // PHIOperands.
1696  unsigned ScanIdx = 0;
1697  SmallPtrSet<Value*,4> InstructionsToSink;
1699  LockstepReverseIterator LRI(UnconditionalPreds);
1700  while (LRI.isValid() &&
1701  canSinkInstructions(*LRI, PHIOperands)) {
1702  DEBUG(dbgs() << "SINK: instruction can be sunk: " << *(*LRI)[0] << "\n");
1703  InstructionsToSink.insert((*LRI).begin(), (*LRI).end());
1704  ++ScanIdx;
1705  --LRI;
1706  }
1707 
1708  auto ProfitableToSinkInstruction = [&](LockstepReverseIterator &LRI) {
1709  unsigned NumPHIdValues = 0;
1710  for (auto *I : *LRI)
1711  for (auto *V : PHIOperands[I])
1712  if (InstructionsToSink.count(V) == 0)
1713  ++NumPHIdValues;
1714  DEBUG(dbgs() << "SINK: #phid values: " << NumPHIdValues << "\n");
1715  unsigned NumPHIInsts = NumPHIdValues / UnconditionalPreds.size();
1716  if ((NumPHIdValues % UnconditionalPreds.size()) != 0)
1717  NumPHIInsts++;
1718 
1719  return NumPHIInsts <= 1;
1720  };
1721 
1722  if (ScanIdx > 0 && Cond) {
1723  // Check if we would actually sink anything first! This mutates the CFG and
1724  // adds an extra block. The goal in doing this is to allow instructions that
1725  // couldn't be sunk before to be sunk - obviously, speculatable instructions
1726  // (such as trunc, add) can be sunk and predicated already. So we check that
1727  // we're going to sink at least one non-speculatable instruction.
1728  LRI.reset();
1729  unsigned Idx = 0;
1730  bool Profitable = false;
1731  while (ProfitableToSinkInstruction(LRI) && Idx < ScanIdx) {
1732  if (!isSafeToSpeculativelyExecute((*LRI)[0])) {
1733  Profitable = true;
1734  break;
1735  }
1736  --LRI;
1737  ++Idx;
1738  }
1739  if (!Profitable)
1740  return false;
1741 
1742  DEBUG(dbgs() << "SINK: Splitting edge\n");
1743  // We have a conditional edge and we're going to sink some instructions.
1744  // Insert a new block postdominating all blocks we're going to sink from.
1745  if (!SplitBlockPredecessors(BI1->getSuccessor(0), UnconditionalPreds,
1746  ".sink.split"))
1747  // Edges couldn't be split.
1748  return false;
1749  Changed = true;
1750  }
1751 
1752  // Now that we've analyzed all potential sinking candidates, perform the
1753  // actual sink. We iteratively sink the last non-terminator of the source
1754  // blocks into their common successor unless doing so would require too
1755  // many PHI instructions to be generated (currently only one PHI is allowed
1756  // per sunk instruction).
1757  //
1758  // We can use InstructionsToSink to discount values needing PHI-merging that will
1759  // actually be sunk in a later iteration. This allows us to be more
1760  // aggressive in what we sink. This does allow a false positive where we
1761  // sink presuming a later value will also be sunk, but stop half way through
1762  // and never actually sink it which means we produce more PHIs than intended.
1763  // This is unlikely in practice though.
1764  for (unsigned SinkIdx = 0; SinkIdx != ScanIdx; ++SinkIdx) {
1765  DEBUG(dbgs() << "SINK: Sink: "
1766  << *UnconditionalPreds[0]->getTerminator()->getPrevNode()
1767  << "\n");
1768 
1769  // Because we've sunk every instruction in turn, the current instruction to
1770  // sink is always at index 0.
1771  LRI.reset();
1772  if (!ProfitableToSinkInstruction(LRI)) {
1773  // Too many PHIs would be created.
1774  DEBUG(dbgs() << "SINK: stopping here, too many PHIs would be created!\n");
1775  break;
1776  }
1777 
1778  if (!sinkLastInstruction(UnconditionalPreds))
1779  return Changed;
1780  NumSinkCommons++;
1781  Changed = true;
1782  }
1783  return Changed;
1784 }
1785 
1786 /// \brief Determine if we can hoist sink a sole store instruction out of a
1787 /// conditional block.
1788 ///
1789 /// We are looking for code like the following:
1790 /// BrBB:
1791 /// store i32 %add, i32* %arrayidx2
1792 /// ... // No other stores or function calls (we could be calling a memory
1793 /// ... // function).
1794 /// %cmp = icmp ult %x, %y
1795 /// br i1 %cmp, label %EndBB, label %ThenBB
1796 /// ThenBB:
1797 /// store i32 %add5, i32* %arrayidx2
1798 /// br label EndBB
1799 /// EndBB:
1800 /// ...
1801 /// We are going to transform this into:
1802 /// BrBB:
1803 /// store i32 %add, i32* %arrayidx2
1804 /// ... //
1805 /// %cmp = icmp ult %x, %y
1806 /// %add.add5 = select i1 %cmp, i32 %add, %add5
1807 /// store i32 %add.add5, i32* %arrayidx2
1808 /// ...
1809 ///
1810 /// \return The pointer to the value of the previous store if the store can be
1811 /// hoisted into the predecessor block. 0 otherwise.
1813  BasicBlock *StoreBB, BasicBlock *EndBB) {
1814  StoreInst *StoreToHoist = dyn_cast<StoreInst>(I);
1815  if (!StoreToHoist)
1816  return nullptr;
1817 
1818  // Volatile or atomic.
1819  if (!StoreToHoist->isSimple())
1820  return nullptr;
1821 
1822  Value *StorePtr = StoreToHoist->getPointerOperand();
1823 
1824  // Look for a store to the same pointer in BrBB.
1825  unsigned MaxNumInstToLookAt = 9;
1826  for (Instruction &CurI : reverse(*BrBB)) {
1827  if (!MaxNumInstToLookAt)
1828  break;
1829  // Skip debug info.
1830  if (isa<DbgInfoIntrinsic>(CurI))
1831  continue;
1832  --MaxNumInstToLookAt;
1833 
1834  // Could be calling an instruction that affects memory like free().
1835  if (CurI.mayHaveSideEffects() && !isa<StoreInst>(CurI))
1836  return nullptr;
1837 
1838  if (auto *SI = dyn_cast<StoreInst>(&CurI)) {
1839  // Found the previous store make sure it stores to the same location.
1840  if (SI->getPointerOperand() == StorePtr)
1841  // Found the previous store, return its value operand.
1842  return SI->getValueOperand();
1843  return nullptr; // Unknown store.
1844  }
1845  }
1846 
1847  return nullptr;
1848 }
1849 
1850 /// \brief Speculate a conditional basic block flattening the CFG.
1851 ///
1852 /// Note that this is a very risky transform currently. Speculating
1853 /// instructions like this is most often not desirable. Instead, there is an MI
1854 /// pass which can do it with full awareness of the resource constraints.
1855 /// However, some cases are "obvious" and we should do directly. An example of
1856 /// this is speculating a single, reasonably cheap instruction.
1857 ///
1858 /// There is only one distinct advantage to flattening the CFG at the IR level:
1859 /// it makes very common but simplistic optimizations such as are common in
1860 /// instcombine and the DAG combiner more powerful by removing CFG edges and
1861 /// modeling their effects with easier to reason about SSA value graphs.
1862 ///
1863 ///
1864 /// An illustration of this transform is turning this IR:
1865 /// \code
1866 /// BB:
1867 /// %cmp = icmp ult %x, %y
1868 /// br i1 %cmp, label %EndBB, label %ThenBB
1869 /// ThenBB:
1870 /// %sub = sub %x, %y
1871 /// br label BB2
1872 /// EndBB:
1873 /// %phi = phi [ %sub, %ThenBB ], [ 0, %EndBB ]
1874 /// ...
1875 /// \endcode
1876 ///
1877 /// Into this IR:
1878 /// \code
1879 /// BB:
1880 /// %cmp = icmp ult %x, %y
1881 /// %sub = sub %x, %y
1882 /// %cond = select i1 %cmp, 0, %sub
1883 /// ...
1884 /// \endcode
1885 ///
1886 /// \returns true if the conditional block is removed.
1888  const TargetTransformInfo &TTI) {
1889  // Be conservative for now. FP select instruction can often be expensive.
1890  Value *BrCond = BI->getCondition();
1891  if (isa<FCmpInst>(BrCond))
1892  return false;
1893 
1894  BasicBlock *BB = BI->getParent();
1895  BasicBlock *EndBB = ThenBB->getTerminator()->getSuccessor(0);
1896 
1897  // If ThenBB is actually on the false edge of the conditional branch, remember
1898  // to swap the select operands later.
1899  bool Invert = false;
1900  if (ThenBB != BI->getSuccessor(0)) {
1901  assert(ThenBB == BI->getSuccessor(1) && "No edge from 'if' block?");
1902  Invert = true;
1903  }
1904  assert(EndBB == BI->getSuccessor(!Invert) && "No edge from to end block");
1905 
1906  // Keep a count of how many times instructions are used within CondBB when
1907  // they are candidates for sinking into CondBB. Specifically:
1908  // - They are defined in BB, and
1909  // - They have no side effects, and
1910  // - All of their uses are in CondBB.
1911  SmallDenseMap<Instruction *, unsigned, 4> SinkCandidateUseCounts;
1912 
1913  unsigned SpeculationCost = 0;
1914  Value *SpeculatedStoreValue = nullptr;
1915  StoreInst *SpeculatedStore = nullptr;
1916  for (BasicBlock::iterator BBI = ThenBB->begin(),
1917  BBE = std::prev(ThenBB->end());
1918  BBI != BBE; ++BBI) {
1919  Instruction *I = &*BBI;
1920  // Skip debug info.
1921  if (isa<DbgInfoIntrinsic>(I))
1922  continue;
1923 
1924  // Only speculatively execute a single instruction (not counting the
1925  // terminator) for now.
1926  ++SpeculationCost;
1927  if (SpeculationCost > 1)
1928  return false;
1929 
1930  // Don't hoist the instruction if it's unsafe or expensive.
1931  if (!isSafeToSpeculativelyExecute(I) &&
1932  !(HoistCondStores && (SpeculatedStoreValue = isSafeToSpeculateStore(
1933  I, BB, ThenBB, EndBB))))
1934  return false;
1935  if (!SpeculatedStoreValue &&
1936  ComputeSpeculationCost(I, TTI) >
1938  return false;
1939 
1940  // Store the store speculation candidate.
1941  if (SpeculatedStoreValue)
1942  SpeculatedStore = cast<StoreInst>(I);
1943 
1944  // Do not hoist the instruction if any of its operands are defined but not
1945  // used in BB. The transformation will prevent the operand from
1946  // being sunk into the use block.
1947  for (User::op_iterator i = I->op_begin(), e = I->op_end(); i != e; ++i) {
1948  Instruction *OpI = dyn_cast<Instruction>(*i);
1949  if (!OpI || OpI->getParent() != BB || OpI->mayHaveSideEffects())
1950  continue; // Not a candidate for sinking.
1951 
1952  ++SinkCandidateUseCounts[OpI];
1953  }
1954  }
1955 
1956  // Consider any sink candidates which are only used in CondBB as costs for
1957  // speculation. Note, while we iterate over a DenseMap here, we are summing
1958  // and so iteration order isn't significant.
1960  I = SinkCandidateUseCounts.begin(),
1961  E = SinkCandidateUseCounts.end();
1962  I != E; ++I)
1963  if (I->first->getNumUses() == I->second) {
1964  ++SpeculationCost;
1965  if (SpeculationCost > 1)
1966  return false;
1967  }
1968 
1969  // Check that the PHI nodes can be converted to selects.
1970  bool HaveRewritablePHIs = false;
1971  for (BasicBlock::iterator I = EndBB->begin();
1972  PHINode *PN = dyn_cast<PHINode>(I); ++I) {
1973  Value *OrigV = PN->getIncomingValueForBlock(BB);
1974  Value *ThenV = PN->getIncomingValueForBlock(ThenBB);
1975 
1976  // FIXME: Try to remove some of the duplication with HoistThenElseCodeToIf.
1977  // Skip PHIs which are trivial.
1978  if (ThenV == OrigV)
1979  continue;
1980 
1981  // Don't convert to selects if we could remove undefined behavior instead.
1982  if (passingValueIsAlwaysUndefined(OrigV, PN) ||
1983  passingValueIsAlwaysUndefined(ThenV, PN))
1984  return false;
1985 
1986  HaveRewritablePHIs = true;
1987  ConstantExpr *OrigCE = dyn_cast<ConstantExpr>(OrigV);
1988  ConstantExpr *ThenCE = dyn_cast<ConstantExpr>(ThenV);
1989  if (!OrigCE && !ThenCE)
1990  continue; // Known safe and cheap.
1991 
1992  if ((ThenCE && !isSafeToSpeculativelyExecute(ThenCE)) ||
1993  (OrigCE && !isSafeToSpeculativelyExecute(OrigCE)))
1994  return false;
1995  unsigned OrigCost = OrigCE ? ComputeSpeculationCost(OrigCE, TTI) : 0;
1996  unsigned ThenCost = ThenCE ? ComputeSpeculationCost(ThenCE, TTI) : 0;
1997  unsigned MaxCost =
1999  if (OrigCost + ThenCost > MaxCost)
2000  return false;
2001 
2002  // Account for the cost of an unfolded ConstantExpr which could end up
2003  // getting expanded into Instructions.
2004  // FIXME: This doesn't account for how many operations are combined in the
2005  // constant expression.
2006  ++SpeculationCost;
2007  if (SpeculationCost > 1)
2008  return false;
2009  }
2010 
2011  // If there are no PHIs to process, bail early. This helps ensure idempotence
2012  // as well.
2013  if (!HaveRewritablePHIs && !(HoistCondStores && SpeculatedStoreValue))
2014  return false;
2015 
2016  // If we get here, we can hoist the instruction and if-convert.
2017  DEBUG(dbgs() << "SPECULATIVELY EXECUTING BB" << *ThenBB << "\n";);
2018 
2019  // Insert a select of the value of the speculated store.
2020  if (SpeculatedStoreValue) {
2021  IRBuilder<NoFolder> Builder(BI);
2022  Value *TrueV = SpeculatedStore->getValueOperand();
2023  Value *FalseV = SpeculatedStoreValue;
2024  if (Invert)
2025  std::swap(TrueV, FalseV);
2026  Value *S = Builder.CreateSelect(
2027  BrCond, TrueV, FalseV, TrueV->getName() + "." + FalseV->getName(), BI);
2028  SpeculatedStore->setOperand(0, S);
2029  SpeculatedStore->applyMergedLocation(BI->getDebugLoc(),
2030  SpeculatedStore->getDebugLoc());
2031  }
2032 
2033  // Metadata can be dependent on the condition we are hoisting above.
2034  // Conservatively strip all metadata on the instruction.
2035  for (auto &I : *ThenBB)
2037 
2038  // Hoist the instructions.
2039  BB->getInstList().splice(BI->getIterator(), ThenBB->getInstList(),
2040  ThenBB->begin(), std::prev(ThenBB->end()));
2041 
2042  // Insert selects and rewrite the PHI operands.
2043  IRBuilder<NoFolder> Builder(BI);
2044  for (BasicBlock::iterator I = EndBB->begin();
2045  PHINode *PN = dyn_cast<PHINode>(I); ++I) {
2046  unsigned OrigI = PN->getBasicBlockIndex(BB);
2047  unsigned ThenI = PN->getBasicBlockIndex(ThenBB);
2048  Value *OrigV = PN->getIncomingValue(OrigI);
2049  Value *ThenV = PN->getIncomingValue(ThenI);
2050 
2051  // Skip PHIs which are trivial.
2052  if (OrigV == ThenV)
2053  continue;
2054 
2055  // Create a select whose true value is the speculatively executed value and
2056  // false value is the preexisting value. Swap them if the branch
2057  // destinations were inverted.
2058  Value *TrueV = ThenV, *FalseV = OrigV;
2059  if (Invert)
2060  std::swap(TrueV, FalseV);
2061  Value *V = Builder.CreateSelect(
2062  BrCond, TrueV, FalseV, TrueV->getName() + "." + FalseV->getName(), BI);
2063  PN->setIncomingValue(OrigI, V);
2064  PN->setIncomingValue(ThenI, V);
2065  }
2066 
2067  ++NumSpeculations;
2068  return true;
2069 }
2070 
2071 /// Return true if we can thread a branch across this block.
2073  BranchInst *BI = cast<BranchInst>(BB->getTerminator());
2074  unsigned Size = 0;
2075 
2076  for (BasicBlock::iterator BBI = BB->begin(); &*BBI != BI; ++BBI) {
2077  if (isa<DbgInfoIntrinsic>(BBI))
2078  continue;
2079  if (Size > 10)
2080  return false; // Don't clone large BB's.
2081  ++Size;
2082 
2083  // We can only support instructions that do not define values that are
2084  // live outside of the current basic block.
2085  for (User *U : BBI->users()) {
2086  Instruction *UI = cast<Instruction>(U);
2087  if (UI->getParent() != BB || isa<PHINode>(UI))
2088  return false;
2089  }
2090 
2091  // Looks ok, continue checking.
2092  }
2093 
2094  return true;
2095 }
2096 
2097 /// If we have a conditional branch on a PHI node value that is defined in the
2098 /// same block as the branch and if any PHI entries are constants, thread edges
2099 /// corresponding to that entry to be branches to their ultimate destination.
2100 static bool FoldCondBranchOnPHI(BranchInst *BI, const DataLayout &DL,
2101  AssumptionCache *AC) {
2102  BasicBlock *BB = BI->getParent();
2103  PHINode *PN = dyn_cast<PHINode>(BI->getCondition());
2104  // NOTE: we currently cannot transform this case if the PHI node is used
2105  // outside of the block.
2106  if (!PN || PN->getParent() != BB || !PN->hasOneUse())
2107  return false;
2108 
2109  // Degenerate case of a single entry PHI.
2110  if (PN->getNumIncomingValues() == 1) {
2112  return true;
2113  }
2114 
2115  // Now we know that this block has multiple preds and two succs.
2117  return false;
2118 
2119  // Can't fold blocks that contain noduplicate or convergent calls.
2120  if (any_of(*BB, [](const Instruction &I) {
2121  const CallInst *CI = dyn_cast<CallInst>(&I);
2122  return CI && (CI->cannotDuplicate() || CI->isConvergent());
2123  }))
2124  return false;
2125 
2126  // Okay, this is a simple enough basic block. See if any phi values are
2127  // constants.
2128  for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
2130  if (!CB || !CB->getType()->isIntegerTy(1))
2131  continue;
2132 
2133  // Okay, we now know that all edges from PredBB should be revectored to
2134  // branch to RealDest.
2135  BasicBlock *PredBB = PN->getIncomingBlock(i);
2136  BasicBlock *RealDest = BI->getSuccessor(!CB->getZExtValue());
2137 
2138  if (RealDest == BB)
2139  continue; // Skip self loops.
2140  // Skip if the predecessor's terminator is an indirect branch.
2141  if (isa<IndirectBrInst>(PredBB->getTerminator()))
2142  continue;
2143 
2144  // The dest block might have PHI nodes, other predecessors and other
2145  // difficult cases. Instead of being smart about this, just insert a new
2146  // block that jumps to the destination block, effectively splitting
2147  // the edge we are about to create.
2148  BasicBlock *EdgeBB =
2149  BasicBlock::Create(BB->getContext(), RealDest->getName() + ".critedge",
2150  RealDest->getParent(), RealDest);
2151  BranchInst::Create(RealDest, EdgeBB);
2152 
2153  // Update PHI nodes.
2154  AddPredecessorToBlock(RealDest, EdgeBB, BB);
2155 
2156  // BB may have instructions that are being threaded over. Clone these
2157  // instructions into EdgeBB. We know that there will be no uses of the
2158  // cloned instructions outside of EdgeBB.
2159  BasicBlock::iterator InsertPt = EdgeBB->begin();
2160  DenseMap<Value *, Value *> TranslateMap; // Track translated values.
2161  for (BasicBlock::iterator BBI = BB->begin(); &*BBI != BI; ++BBI) {
2162  if (PHINode *PN = dyn_cast<PHINode>(BBI)) {
2163  TranslateMap[PN] = PN->getIncomingValueForBlock(PredBB);
2164  continue;
2165  }
2166  // Clone the instruction.
2167  Instruction *N = BBI->clone();
2168  if (BBI->hasName())
2169  N->setName(BBI->getName() + ".c");
2170 
2171  // Update operands due to translation.
2172  for (User::op_iterator i = N->op_begin(), e = N->op_end(); i != e; ++i) {
2173  DenseMap<Value *, Value *>::iterator PI = TranslateMap.find(*i);
2174  if (PI != TranslateMap.end())
2175  *i = PI->second;
2176  }
2177 
2178  // Check for trivial simplification.
2179  if (Value *V = SimplifyInstruction(N, {DL, nullptr, nullptr, AC})) {
2180  if (!BBI->use_empty())
2181  TranslateMap[&*BBI] = V;
2182  if (!N->mayHaveSideEffects()) {
2183  N->deleteValue(); // Instruction folded away, don't need actual inst
2184  N = nullptr;
2185  }
2186  } else {
2187  if (!BBI->use_empty())
2188  TranslateMap[&*BBI] = N;
2189  }
2190  // Insert the new instruction into its new home.
2191  if (N)
2192  EdgeBB->getInstList().insert(InsertPt, N);
2193 
2194  // Register the new instruction with the assumption cache if necessary.
2195  if (auto *II = dyn_cast_or_null<IntrinsicInst>(N))
2196  if (II->getIntrinsicID() == Intrinsic::assume)
2197  AC->registerAssumption(II);
2198  }
2199 
2200  // Loop over all of the edges from PredBB to BB, changing them to branch
2201  // to EdgeBB instead.
2202  TerminatorInst *PredBBTI = PredBB->getTerminator();
2203  for (unsigned i = 0, e = PredBBTI->getNumSuccessors(); i != e; ++i)
2204  if (PredBBTI->getSuccessor(i) == BB) {
2205  BB->removePredecessor(PredBB);
2206  PredBBTI->setSuccessor(i, EdgeBB);
2207  }
2208 
2209  // Recurse, simplifying any other constants.
2210  return FoldCondBranchOnPHI(BI, DL, AC) | true;
2211  }
2212 
2213  return false;
2214 }
2215 
2216 /// Given a BB that starts with the specified two-entry PHI node,
2217 /// see if we can eliminate it.
2219  const DataLayout &DL) {
2220  // Ok, this is a two entry PHI node. Check to see if this is a simple "if
2221  // statement", which has a very simple dominance structure. Basically, we
2222  // are trying to find the condition that is being branched on, which
2223  // subsequently causes this merge to happen. We really want control
2224  // dependence information for this check, but simplifycfg can't keep it up
2225  // to date, and this catches most of the cases we care about anyway.
2226  BasicBlock *BB = PN->getParent();
2227  BasicBlock *IfTrue, *IfFalse;
2228  Value *IfCond = GetIfCondition(BB, IfTrue, IfFalse);
2229  if (!IfCond ||
2230  // Don't bother if the branch will be constant folded trivially.
2231  isa<ConstantInt>(IfCond))
2232  return false;
2233 
2234  // Okay, we found that we can merge this two-entry phi node into a select.
2235  // Doing so would require us to fold *all* two entry phi nodes in this block.
2236  // At some point this becomes non-profitable (particularly if the target
2237  // doesn't support cmov's). Only do this transformation if there are two or
2238  // fewer PHI nodes in this block.
2239  unsigned NumPhis = 0;
2240  for (BasicBlock::iterator I = BB->begin(); isa<PHINode>(I); ++NumPhis, ++I)
2241  if (NumPhis > 2)
2242  return false;
2243 
2244  // Loop over the PHI's seeing if we can promote them all to select
2245  // instructions. While we are at it, keep track of the instructions
2246  // that need to be moved to the dominating block.
2247  SmallPtrSet<Instruction *, 4> AggressiveInsts;
2248  unsigned MaxCostVal0 = PHINodeFoldingThreshold,
2249  MaxCostVal1 = PHINodeFoldingThreshold;
2250  MaxCostVal0 *= TargetTransformInfo::TCC_Basic;
2251  MaxCostVal1 *= TargetTransformInfo::TCC_Basic;
2252 
2253  for (BasicBlock::iterator II = BB->begin(); isa<PHINode>(II);) {
2254  PHINode *PN = cast<PHINode>(II++);
2255  if (Value *V = SimplifyInstruction(PN, {DL, PN})) {
2256  PN->replaceAllUsesWith(V);
2257  PN->eraseFromParent();
2258  continue;
2259  }
2260 
2261  if (!DominatesMergePoint(PN->getIncomingValue(0), BB, &AggressiveInsts,
2262  MaxCostVal0, TTI) ||
2263  !DominatesMergePoint(PN->getIncomingValue(1), BB, &AggressiveInsts,
2264  MaxCostVal1, TTI))
2265  return false;
2266  }
2267 
2268  // If we folded the first phi, PN dangles at this point. Refresh it. If
2269  // we ran out of PHIs then we simplified them all.
2270  PN = dyn_cast<PHINode>(BB->begin());
2271  if (!PN)
2272  return true;
2273 
2274  // Don't fold i1 branches on PHIs which contain binary operators. These can
2275  // often be turned into switches and other things.
2276  if (PN->getType()->isIntegerTy(1) &&
2277  (isa<BinaryOperator>(PN->getIncomingValue(0)) ||
2278  isa<BinaryOperator>(PN->getIncomingValue(1)) ||
2279  isa<BinaryOperator>(IfCond)))
2280  return false;
2281 
2282  // If all PHI nodes are promotable, check to make sure that all instructions
2283  // in the predecessor blocks can be promoted as well. If not, we won't be able
2284  // to get rid of the control flow, so it's not worth promoting to select
2285  // instructions.
2286  BasicBlock *DomBlock = nullptr;
2287  BasicBlock *IfBlock1 = PN->getIncomingBlock(0);
2288  BasicBlock *IfBlock2 = PN->getIncomingBlock(1);
2289  if (cast<BranchInst>(IfBlock1->getTerminator())->isConditional()) {
2290  IfBlock1 = nullptr;
2291  } else {
2292  DomBlock = *pred_begin(IfBlock1);
2293  for (BasicBlock::iterator I = IfBlock1->begin(); !isa<TerminatorInst>(I);
2294  ++I)
2295  if (!AggressiveInsts.count(&*I) && !isa<DbgInfoIntrinsic>(I)) {
2296  // This is not an aggressive instruction that we can promote.
2297  // Because of this, we won't be able to get rid of the control flow, so
2298  // the xform is not worth it.
2299  return false;
2300  }
2301  }
2302 
2303  if (cast<BranchInst>(IfBlock2->getTerminator())->isConditional()) {
2304  IfBlock2 = nullptr;
2305  } else {
2306  DomBlock = *pred_begin(IfBlock2);
2307  for (BasicBlock::iterator I = IfBlock2->begin(); !isa<TerminatorInst>(I);
2308  ++I)
2309  if (!AggressiveInsts.count(&*I) && !isa<DbgInfoIntrinsic>(I)) {
2310  // This is not an aggressive instruction that we can promote.
2311  // Because of this, we won't be able to get rid of the control flow, so
2312  // the xform is not worth it.
2313  return false;
2314  }
2315  }
2316 
2317  DEBUG(dbgs() << "FOUND IF CONDITION! " << *IfCond << " T: "
2318  << IfTrue->getName() << " F: " << IfFalse->getName() << "\n");
2319 
2320  // If we can still promote the PHI nodes after this gauntlet of tests,
2321  // do all of the PHI's now.
2322  Instruction *InsertPt = DomBlock->getTerminator();
2323  IRBuilder<NoFolder> Builder(InsertPt);
2324 
2325  // Move all 'aggressive' instructions, which are defined in the
2326  // conditional parts of the if's up to the dominating block.
2327  if (IfBlock1) {
2328  for (auto &I : *IfBlock1)
2330  DomBlock->getInstList().splice(InsertPt->getIterator(),
2331  IfBlock1->getInstList(), IfBlock1->begin(),
2332  IfBlock1->getTerminator()->getIterator());
2333  }
2334  if (IfBlock2) {
2335  for (auto &I : *IfBlock2)
2337  DomBlock->getInstList().splice(InsertPt->getIterator(),
2338  IfBlock2->getInstList(), IfBlock2->begin(),
2339  IfBlock2->getTerminator()->getIterator());
2340  }
2341 
2342  while (PHINode *PN = dyn_cast<PHINode>(BB->begin())) {
2343  // Change the PHI node into a select instruction.
2344  Value *TrueVal = PN->getIncomingValue(PN->getIncomingBlock(0) == IfFalse);
2345  Value *FalseVal = PN->getIncomingValue(PN->getIncomingBlock(0) == IfTrue);
2346 
2347  Value *Sel = Builder.CreateSelect(IfCond, TrueVal, FalseVal, "", InsertPt);
2348  PN->replaceAllUsesWith(Sel);
2349  Sel->takeName(PN);
2350  PN->eraseFromParent();
2351  }
2352 
2353  // At this point, IfBlock1 and IfBlock2 are both empty, so our if statement
2354  // has been flattened. Change DomBlock to jump directly to our new block to
2355  // avoid other simplifycfg's kicking in on the diamond.
2356  TerminatorInst *OldTI = DomBlock->getTerminator();
2357  Builder.SetInsertPoint(OldTI);
2358  Builder.CreateBr(BB);
2359  OldTI->eraseFromParent();
2360  return true;
2361 }
2362 
2363 /// If we found a conditional branch that goes to two returning blocks,
2364 /// try to merge them together into one return,
2365 /// introducing a select if the return values disagree.
2367  IRBuilder<> &Builder) {
2368  assert(BI->isConditional() && "Must be a conditional branch");
2369  BasicBlock *TrueSucc = BI->getSuccessor(0);
2370  BasicBlock *FalseSucc = BI->getSuccessor(1);
2371  ReturnInst *TrueRet = cast<ReturnInst>(TrueSucc->getTerminator());
2372  ReturnInst *FalseRet = cast<ReturnInst>(FalseSucc->getTerminator());
2373 
2374  // Check to ensure both blocks are empty (just a return) or optionally empty
2375  // with PHI nodes. If there are other instructions, merging would cause extra
2376  // computation on one path or the other.
2377  if (!TrueSucc->getFirstNonPHIOrDbg()->isTerminator())
2378  return false;
2379  if (!FalseSucc->getFirstNonPHIOrDbg()->isTerminator())
2380  return false;
2381 
2382  Builder.SetInsertPoint(BI);
2383  // Okay, we found a branch that is going to two return nodes. If
2384  // there is no return value for this function, just change the
2385  // branch into a return.
2386  if (FalseRet->getNumOperands() == 0) {
2387  TrueSucc->removePredecessor(BI->getParent());
2388  FalseSucc->removePredecessor(BI->getParent());
2389  Builder.CreateRetVoid();
2391  return true;
2392  }
2393 
2394  // Otherwise, figure out what the true and false return values are
2395  // so we can insert a new select instruction.
2396  Value *TrueValue = TrueRet->getReturnValue();
2397  Value *FalseValue = FalseRet->getReturnValue();
2398 
2399  // Unwrap any PHI nodes in the return blocks.
2400  if (PHINode *TVPN = dyn_cast_or_null<PHINode>(TrueValue))
2401  if (TVPN->getParent() == TrueSucc)
2402  TrueValue = TVPN->getIncomingValueForBlock(BI->getParent());
2403  if (PHINode *FVPN = dyn_cast_or_null<PHINode>(FalseValue))
2404  if (FVPN->getParent() == FalseSucc)
2405  FalseValue = FVPN->getIncomingValueForBlock(BI->getParent());
2406 
2407  // In order for this transformation to be safe, we must be able to
2408  // unconditionally execute both operands to the return. This is
2409  // normally the case, but we could have a potentially-trapping
2410  // constant expression that prevents this transformation from being
2411  // safe.
2412  if (ConstantExpr *TCV = dyn_cast_or_null<ConstantExpr>(TrueValue))
2413  if (TCV->canTrap())
2414  return false;
2415  if (ConstantExpr *FCV = dyn_cast_or_null<ConstantExpr>(FalseValue))
2416  if (FCV->canTrap())
2417  return false;
2418 
2419  // Okay, we collected all the mapped values and checked them for sanity, and
2420  // defined to really do this transformation. First, update the CFG.
2421  TrueSucc->removePredecessor(BI->getParent());
2422  FalseSucc->removePredecessor(BI->getParent());
2423 
2424  // Insert select instructions where needed.
2425  Value *BrCond = BI->getCondition();
2426  if (TrueValue) {
2427  // Insert a select if the results differ.
2428  if (TrueValue == FalseValue || isa<UndefValue>(FalseValue)) {
2429  } else if (isa<UndefValue>(TrueValue)) {
2430  TrueValue = FalseValue;
2431  } else {
2432  TrueValue =
2433  Builder.CreateSelect(BrCond, TrueValue, FalseValue, "retval", BI);
2434  }
2435  }
2436 
2437  Value *RI =
2438  !TrueValue ? Builder.CreateRetVoid() : Builder.CreateRet(TrueValue);
2439 
2440  (void)RI;
2441 
2442  DEBUG(dbgs() << "\nCHANGING BRANCH TO TWO RETURNS INTO SELECT:"
2443  << "\n " << *BI << "NewRet = " << *RI
2444  << "TRUEBLOCK: " << *TrueSucc << "FALSEBLOCK: " << *FalseSucc);
2445 
2447 
2448  return true;
2449 }
2450 
2451 /// Return true if the given instruction is available
2452 /// in its predecessor block. If yes, the instruction will be removed.
2454  if (!isa<BinaryOperator>(Inst) && !isa<CmpInst>(Inst))
2455  return false;
2456  for (Instruction &I : *PB) {
2457  Instruction *PBI = &I;
2458  // Check whether Inst and PBI generate the same value.
2459  if (Inst->isIdenticalTo(PBI)) {
2460  Inst->replaceAllUsesWith(PBI);
2461  Inst->eraseFromParent();
2462  return true;
2463  }
2464  }
2465  return false;
2466 }
2467 
2468 /// Return true if either PBI or BI has branch weight available, and store
2469 /// the weights in {Pred|Succ}{True|False}Weight. If one of PBI and BI does
2470 /// not have branch weight, use 1:1 as its weight.
2472  uint64_t &PredTrueWeight,
2473  uint64_t &PredFalseWeight,
2474  uint64_t &SuccTrueWeight,
2475  uint64_t &SuccFalseWeight) {
2476  bool PredHasWeights =
2477  PBI->extractProfMetadata(PredTrueWeight, PredFalseWeight);
2478  bool SuccHasWeights =
2479  BI->extractProfMetadata(SuccTrueWeight, SuccFalseWeight);
2480  if (PredHasWeights || SuccHasWeights) {
2481  if (!PredHasWeights)
2482  PredTrueWeight = PredFalseWeight = 1;
2483  if (!SuccHasWeights)
2484  SuccTrueWeight = SuccFalseWeight = 1;
2485  return true;
2486  } else {
2487  return false;
2488  }
2489 }
2490 
2491 /// If this basic block is simple enough, and if a predecessor branches to us
2492 /// and one of our successors, fold the block into the predecessor and use
2493 /// logical operations to pick the right destination.
2494 bool llvm::FoldBranchToCommonDest(BranchInst *BI, unsigned BonusInstThreshold) {
2495  BasicBlock *BB = BI->getParent();
2496 
2497  Instruction *Cond = nullptr;
2498  if (BI->isConditional())
2499  Cond = dyn_cast<Instruction>(BI->getCondition());
2500  else {
2501  // For unconditional branch, check for a simple CFG pattern, where
2502  // BB has a single predecessor and BB's successor is also its predecessor's
2503  // successor. If such pattern exists, check for CSE between BB and its
2504  // predecessor.
2505  if (BasicBlock *PB = BB->getSinglePredecessor())
2506  if (BranchInst *PBI = dyn_cast<BranchInst>(PB->getTerminator()))
2507  if (PBI->isConditional() &&
2508  (BI->getSuccessor(0) == PBI->getSuccessor(0) ||
2509  BI->getSuccessor(0) == PBI->getSuccessor(1))) {
2510  for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E;) {
2511  Instruction *Curr = &*I++;
2512  if (isa<CmpInst>(Curr)) {
2513  Cond = Curr;
2514  break;
2515  }
2516  // Quit if we can't remove this instruction.
2517  if (!checkCSEInPredecessor(Curr, PB))
2518  return false;
2519  }
2520  }
2521 
2522  if (!Cond)
2523  return false;
2524  }
2525 
2526  if (!Cond || (!isa<CmpInst>(Cond) && !isa<BinaryOperator>(Cond)) ||
2527  Cond->getParent() != BB || !Cond->hasOneUse())
2528  return false;
2529 
2530  // Make sure the instruction after the condition is the cond branch.
2531  BasicBlock::iterator CondIt = ++Cond->getIterator();
2532 
2533  // Ignore dbg intrinsics.
2534  while (isa<DbgInfoIntrinsic>(CondIt))
2535  ++CondIt;
2536 
2537  if (&*CondIt != BI)
2538  return false;
2539 
2540  // Only allow this transformation if computing the condition doesn't involve
2541  // too many instructions and these involved instructions can be executed
2542  // unconditionally. We denote all involved instructions except the condition
2543  // as "bonus instructions", and only allow this transformation when the
2544  // number of the bonus instructions does not exceed a certain threshold.
2545  unsigned NumBonusInsts = 0;
2546  for (auto I = BB->begin(); Cond != &*I; ++I) {
2547  // Ignore dbg intrinsics.
2548  if (isa<DbgInfoIntrinsic>(I))
2549  continue;
2550  if (!I->hasOneUse() || !isSafeToSpeculativelyExecute(&*I))
2551  return false;
2552  // I has only one use and can be executed unconditionally.
2554  if (User == nullptr || User->getParent() != BB)
2555  return false;
2556  // I is used in the same BB. Since BI uses Cond and doesn't have more slots
2557  // to use any other instruction, User must be an instruction between next(I)
2558  // and Cond.
2559  ++NumBonusInsts;
2560  // Early exits once we reach the limit.
2561  if (NumBonusInsts > BonusInstThreshold)
2562  return false;
2563  }
2564 
2565  // Cond is known to be a compare or binary operator. Check to make sure that
2566  // neither operand is a potentially-trapping constant expression.
2567  if (ConstantExpr *CE = dyn_cast<ConstantExpr>(Cond->getOperand(0)))
2568  if (CE->canTrap())
2569  return false;
2570  if (ConstantExpr *CE = dyn_cast<ConstantExpr>(Cond->getOperand(1)))
2571  if (CE->canTrap())
2572  return false;
2573 
2574  // Finally, don't infinitely unroll conditional loops.
2575  BasicBlock *TrueDest = BI->getSuccessor(0);
2576  BasicBlock *FalseDest = (BI->isConditional()) ? BI->getSuccessor(1) : nullptr;
2577  if (TrueDest == BB || FalseDest == BB)
2578  return false;
2579 
2580  for (pred_iterator PI = pred_begin(BB), E = pred_end(BB); PI != E; ++PI) {
2581  BasicBlock *PredBlock = *PI;
2582  BranchInst *PBI = dyn_cast<BranchInst>(PredBlock->getTerminator());
2583 
2584  // Check that we have two conditional branches. If there is a PHI node in
2585  // the common successor, verify that the same value flows in from both
2586  // blocks.
2588  if (!PBI || PBI->isUnconditional() ||
2589  (BI->isConditional() && !SafeToMergeTerminators(BI, PBI)) ||
2590  (!BI->isConditional() &&
2591  !isProfitableToFoldUnconditional(BI, PBI, Cond, PHIs)))
2592  continue;
2593 
2594  // Determine if the two branches share a common destination.
2595  Instruction::BinaryOps Opc = Instruction::BinaryOpsEnd;
2596  bool InvertPredCond = false;
2597 
2598  if (BI->isConditional()) {
2599  if (PBI->getSuccessor(0) == TrueDest) {
2600  Opc = Instruction::Or;
2601  } else if (PBI->getSuccessor(1) == FalseDest) {
2602  Opc = Instruction::And;
2603  } else if (PBI->getSuccessor(0) == FalseDest) {
2604  Opc = Instruction::And;
2605  InvertPredCond = true;
2606  } else if (PBI->getSuccessor(1) == TrueDest) {
2607  Opc = Instruction::Or;
2608  InvertPredCond = true;
2609  } else {
2610  continue;
2611  }
2612  } else {
2613  if (PBI->getSuccessor(0) != TrueDest && PBI->getSuccessor(1) != TrueDest)
2614  continue;
2615  }
2616 
2617  DEBUG(dbgs() << "FOLDING BRANCH TO COMMON DEST:\n" << *PBI << *BB);
2618  IRBuilder<> Builder(PBI);
2619 
2620  // If we need to invert the condition in the pred block to match, do so now.
2621  if (InvertPredCond) {
2622  Value *NewCond = PBI->getCondition();
2623 
2624  if (NewCond->hasOneUse() && isa<CmpInst>(NewCond)) {
2625  CmpInst *CI = cast<CmpInst>(NewCond);
2626  CI->setPredicate(CI->getInversePredicate());
2627  } else {
2628  NewCond =
2629  Builder.CreateNot(NewCond, PBI->getCondition()->getName() + ".not");
2630  }
2631 
2632  PBI->setCondition(NewCond);
2633  PBI->swapSuccessors();
2634  }
2635 
2636  // If we have bonus instructions, clone them into the predecessor block.
2637  // Note that there may be multiple predecessor blocks, so we cannot move
2638  // bonus instructions to a predecessor block.
2639  ValueToValueMapTy VMap; // maps original values to cloned values
2640  // We already make sure Cond is the last instruction before BI. Therefore,
2641  // all instructions before Cond other than DbgInfoIntrinsic are bonus
2642  // instructions.
2643  for (auto BonusInst = BB->begin(); Cond != &*BonusInst; ++BonusInst) {
2644  if (isa<DbgInfoIntrinsic>(BonusInst))
2645  continue;
2646  Instruction *NewBonusInst = BonusInst->clone();
2647  RemapInstruction(NewBonusInst, VMap,
2649  VMap[&*BonusInst] = NewBonusInst;
2650 
2651  // If we moved a load, we cannot any longer claim any knowledge about
2652  // its potential value. The previous information might have been valid
2653  // only given the branch precondition.
2654  // For an analogous reason, we must also drop all the metadata whose
2655  // semantics we don't understand.
2656  NewBonusInst->dropUnknownNonDebugMetadata();
2657 
2658  PredBlock->getInstList().insert(PBI->getIterator(), NewBonusInst);
2659  NewBonusInst->takeName(&*BonusInst);
2660  BonusInst->setName(BonusInst->getName() + ".old");
2661  }
2662 
2663  // Clone Cond into the predecessor basic block, and or/and the
2664  // two conditions together.
2665  Instruction *New = Cond->clone();
2666  RemapInstruction(New, VMap,
2668  PredBlock->getInstList().insert(PBI->getIterator(), New);
2669  New->takeName(Cond);
2670  Cond->setName(New->getName() + ".old");
2671 
2672  if (BI->isConditional()) {
2673  Instruction *NewCond = cast<Instruction>(
2674  Builder.CreateBinOp(Opc, PBI->getCondition(), New, "or.cond"));
2675  PBI->setCondition(NewCond);
2676 
2677  uint64_t PredTrueWeight, PredFalseWeight, SuccTrueWeight, SuccFalseWeight;
2678  bool HasWeights =
2679  extractPredSuccWeights(PBI, BI, PredTrueWeight, PredFalseWeight,
2680  SuccTrueWeight, SuccFalseWeight);
2681  SmallVector<uint64_t, 8> NewWeights;
2682 
2683  if (PBI->getSuccessor(0) == BB) {
2684  if (HasWeights) {
2685  // PBI: br i1 %x, BB, FalseDest
2686  // BI: br i1 %y, TrueDest, FalseDest
2687  // TrueWeight is TrueWeight for PBI * TrueWeight for BI.
2688  NewWeights.push_back(PredTrueWeight * SuccTrueWeight);
2689  // FalseWeight is FalseWeight for PBI * TotalWeight for BI +
2690  // TrueWeight for PBI * FalseWeight for BI.
2691  // We assume that total weights of a BranchInst can fit into 32 bits.
2692  // Therefore, we will not have overflow using 64-bit arithmetic.
2693  NewWeights.push_back(PredFalseWeight *
2694  (SuccFalseWeight + SuccTrueWeight) +
2695  PredTrueWeight * SuccFalseWeight);
2696  }
2697  AddPredecessorToBlock(TrueDest, PredBlock, BB);
2698  PBI->setSuccessor(0, TrueDest);
2699  }
2700  if (PBI->getSuccessor(1) == BB) {
2701  if (HasWeights) {
2702  // PBI: br i1 %x, TrueDest, BB
2703  // BI: br i1 %y, TrueDest, FalseDest
2704  // TrueWeight is TrueWeight for PBI * TotalWeight for BI +
2705  // FalseWeight for PBI * TrueWeight for BI.
2706  NewWeights.push_back(PredTrueWeight *
2707  (SuccFalseWeight + SuccTrueWeight) +
2708  PredFalseWeight * SuccTrueWeight);
2709  // FalseWeight is FalseWeight for PBI * FalseWeight for BI.
2710  NewWeights.push_back(PredFalseWeight * SuccFalseWeight);
2711  }
2712  AddPredecessorToBlock(FalseDest, PredBlock, BB);
2713  PBI->setSuccessor(1, FalseDest);
2714  }
2715  if (NewWeights.size() == 2) {
2716  // Halve the weights if any of them cannot fit in an uint32_t
2717  FitWeights(NewWeights);
2718 
2719  SmallVector<uint32_t, 8> MDWeights(NewWeights.begin(),
2720  NewWeights.end());
2721  PBI->setMetadata(
2723  MDBuilder(BI->getContext()).createBranchWeights(MDWeights));
2724  } else
2725  PBI->setMetadata(LLVMContext::MD_prof, nullptr);
2726  } else {
2727  // Update PHI nodes in the common successors.
2728  for (unsigned i = 0, e = PHIs.size(); i != e; ++i) {
2729  ConstantInt *PBI_C = cast<ConstantInt>(
2730  PHIs[i]->getIncomingValueForBlock(PBI->getParent()));
2731  assert(PBI_C->getType()->isIntegerTy(1));
2732  Instruction *MergedCond = nullptr;
2733  if (PBI->getSuccessor(0) == TrueDest) {
2734  // Create (PBI_Cond and PBI_C) or (!PBI_Cond and BI_Value)
2735  // PBI_C is true: PBI_Cond or (!PBI_Cond and BI_Value)
2736  // is false: !PBI_Cond and BI_Value
2737  Instruction *NotCond = cast<Instruction>(
2738  Builder.CreateNot(PBI->getCondition(), "not.cond"));
2739  MergedCond = cast<Instruction>(
2740  Builder.CreateBinOp(Instruction::And, NotCond, New, "and.cond"));
2741  if (PBI_C->isOne())
2742  MergedCond = cast<Instruction>(Builder.CreateBinOp(
2743  Instruction::Or, PBI->getCondition(), MergedCond, "or.cond"));
2744  } else {
2745  // Create (PBI_Cond and BI_Value) or (!PBI_Cond and PBI_C)
2746  // PBI_C is true: (PBI_Cond and BI_Value) or (!PBI_Cond)
2747  // is false: PBI_Cond and BI_Value
2748  MergedCond = cast<Instruction>(Builder.CreateBinOp(
2749  Instruction::And, PBI->getCondition(), New, "and.cond"));
2750  if (PBI_C->isOne()) {
2751  Instruction *NotCond = cast<Instruction>(
2752  Builder.CreateNot(PBI->getCondition(), "not.cond"));
2753  MergedCond = cast<Instruction>(Builder.CreateBinOp(
2754  Instruction::Or, NotCond, MergedCond, "or.cond"));
2755  }
2756  }
2757  // Update PHI Node.
2758  PHIs[i]->setIncomingValue(PHIs[i]->getBasicBlockIndex(PBI->getParent()),
2759  MergedCond);
2760  }
2761  // Change PBI from Conditional to Unconditional.
2762  BranchInst *New_PBI = BranchInst::Create(TrueDest, PBI);
2764  PBI = New_PBI;
2765  }
2766 
2767  // If BI was a loop latch, it may have had associated loop metadata.
2768  // We need to copy it to the new latch, that is, PBI.
2769  if (MDNode *LoopMD = BI->getMetadata(LLVMContext::MD_loop))
2770  PBI->setMetadata(LLVMContext::MD_loop, LoopMD);
2771 
2772  // TODO: If BB is reachable from all paths through PredBlock, then we
2773  // could replace PBI's branch probabilities with BI's.
2774 
2775  // Copy any debug value intrinsics into the end of PredBlock.
2776  for (Instruction &I : *BB)
2777  if (isa<DbgInfoIntrinsic>(I))
2778  I.clone()->insertBefore(PBI);
2779 
2780  return true;
2781  }
2782  return false;
2783 }
2784 
2785 // If there is only one store in BB1 and BB2, return it, otherwise return
2786 // nullptr.
2788  StoreInst *S = nullptr;
2789  for (auto *BB : {BB1, BB2}) {
2790  if (!BB)
2791  continue;
2792  for (auto &I : *BB)
2793  if (auto *SI = dyn_cast<StoreInst>(&I)) {
2794  if (S)
2795  // Multiple stores seen.
2796  return nullptr;
2797  else
2798  S = SI;
2799  }
2800  }
2801  return S;
2802 }
2803 
2805  Value *AlternativeV = nullptr) {
2806  // PHI is going to be a PHI node that allows the value V that is defined in
2807  // BB to be referenced in BB's only successor.
2808  //
2809  // If AlternativeV is nullptr, the only value we care about in PHI is V. It
2810  // doesn't matter to us what the other operand is (it'll never get used). We
2811  // could just create a new PHI with an undef incoming value, but that could
2812  // increase register pressure if EarlyCSE/InstCombine can't fold it with some
2813  // other PHI. So here we directly look for some PHI in BB's successor with V
2814  // as an incoming operand. If we find one, we use it, else we create a new
2815  // one.
2816  //
2817  // If AlternativeV is not nullptr, we care about both incoming values in PHI.
2818  // PHI must be exactly: phi <ty> [ %BB, %V ], [ %OtherBB, %AlternativeV]
2819  // where OtherBB is the single other predecessor of BB's only successor.
2820  PHINode *PHI = nullptr;
2821  BasicBlock *Succ = BB->getSingleSuccessor();
2822 
2823  for (auto I = Succ->begin(); isa<PHINode>(I); ++I)
2824  if (cast<PHINode>(I)->getIncomingValueForBlock(BB) == V) {
2825  PHI = cast<PHINode>(I);
2826  if (!AlternativeV)
2827  break;
2828 
2829  assert(std::distance(pred_begin(Succ), pred_end(Succ)) == 2);
2830  auto PredI = pred_begin(Succ);
2831  BasicBlock *OtherPredBB = *PredI == BB ? *++PredI : *PredI;
2832  if (PHI->getIncomingValueForBlock(OtherPredBB) == AlternativeV)
2833  break;
2834  PHI = nullptr;
2835  }
2836  if (PHI)
2837  return PHI;
2838 
2839  // If V is not an instruction defined in BB, just return it.
2840  if (!AlternativeV &&
2841  (!isa<Instruction>(V) || cast<Instruction>(V)->getParent() != BB))
2842  return V;
2843 
2844  PHI = PHINode::Create(V->getType(), 2, "simplifycfg.merge", &Succ->front());
2845  PHI->addIncoming(V, BB);
2846  for (BasicBlock *PredBB : predecessors(Succ))
2847  if (PredBB != BB)
2848  PHI->addIncoming(
2849  AlternativeV ? AlternativeV : UndefValue::get(V->getType()), PredBB);
2850  return PHI;
2851 }
2852 
2854  BasicBlock *QTB, BasicBlock *QFB,
2855  BasicBlock *PostBB, Value *Address,
2856  bool InvertPCond, bool InvertQCond,
2857  const DataLayout &DL) {
2858  auto IsaBitcastOfPointerType = [](const Instruction &I) {
2859  return Operator::getOpcode(&I) == Instruction::BitCast &&
2860  I.getType()->isPointerTy();
2861  };
2862 
2863  // If we're not in aggressive mode, we only optimize if we have some
2864  // confidence that by optimizing we'll allow P and/or Q to be if-converted.
2865  auto IsWorthwhile = [&](BasicBlock *BB) {
2866  if (!BB)
2867  return true;
2868  // Heuristic: if the block can be if-converted/phi-folded and the
2869  // instructions inside are all cheap (arithmetic/GEPs), it's worthwhile to
2870  // thread this store.
2871  unsigned N = 0;
2872  for (auto &I : *BB) {
2873  // Cheap instructions viable for folding.
2874  if (isa<BinaryOperator>(I) || isa<GetElementPtrInst>(I) ||
2875  isa<StoreInst>(I))
2876  ++N;
2877  // Free instructions.
2878  else if (isa<TerminatorInst>(I) || isa<DbgInfoIntrinsic>(I) ||
2879  IsaBitcastOfPointerType(I))
2880  continue;
2881  else
2882  return false;
2883  }
2884  return N <= PHINodeFoldingThreshold;
2885  };
2886 
2888  (!IsWorthwhile(PTB) || !IsWorthwhile(PFB) || !IsWorthwhile(QTB) ||
2889  !IsWorthwhile(QFB)))
2890  return false;
2891 
2892  // For every pointer, there must be exactly two stores, one coming from
2893  // PTB or PFB, and the other from QTB or QFB. We don't support more than one
2894  // store (to any address) in PTB,PFB or QTB,QFB.
2895  // FIXME: We could relax this restriction with a bit more work and performance
2896  // testing.
2897  StoreInst *PStore = findUniqueStoreInBlocks(PTB, PFB);
2898  StoreInst *QStore = findUniqueStoreInBlocks(QTB, QFB);
2899  if (!PStore || !QStore)
2900  return false;
2901 
2902  // Now check the stores are compatible.
2903  if (!QStore->isUnordered() || !PStore->isUnordered())
2904  return false;
2905 
2906  // Check that sinking the store won't cause program behavior changes. Sinking
2907  // the store out of the Q blocks won't change any behavior as we're sinking
2908  // from a block to its unconditional successor. But we're moving a store from
2909  // the P blocks down through the middle block (QBI) and past both QFB and QTB.
2910  // So we need to check that there are no aliasing loads or stores in
2911  // QBI, QTB and QFB. We also need to check there are no conflicting memory
2912  // operations between PStore and the end of its parent block.
2913  //
2914  // The ideal way to do this is to query AliasAnalysis, but we don't
2915  // preserve AA currently so that is dangerous. Be super safe and just
2916  // check there are no other memory operations at all.
2917  for (auto &I : *QFB->getSinglePredecessor())
2918  if (I.mayReadOrWriteMemory())
2919  return false;
2920  for (auto &I : *QFB)
2921  if (&I != QStore && I.mayReadOrWriteMemory())
2922  return false;
2923  if (QTB)
2924  for (auto &I : *QTB)
2925  if (&I != QStore && I.mayReadOrWriteMemory())
2926  return false;
2927  for (auto I = BasicBlock::iterator(PStore), E = PStore->getParent()->end();
2928  I != E; ++I)
2929  if (&*I != PStore && I->mayReadOrWriteMemory())
2930  return false;
2931 
2932  // OK, we're going to sink the stores to PostBB. The store has to be
2933  // conditional though, so first create the predicate.
2934  Value *PCond = cast<BranchInst>(PFB->getSinglePredecessor()->getTerminator())
2935  ->getCondition();
2936  Value *QCond = cast<BranchInst>(QFB->getSinglePredecessor()->getTerminator())
2937  ->getCondition();
2938 
2940  PStore->getParent());
2942  QStore->getParent(), PPHI);
2943 
2944  IRBuilder<> QB(&*PostBB->getFirstInsertionPt());
2945 
2946  Value *PPred = PStore->getParent() == PTB ? PCond : QB.CreateNot(PCond);
2947  Value *QPred = QStore->getParent() == QTB ? QCond : QB.CreateNot(QCond);
2948 
2949  if (InvertPCond)
2950  PPred = QB.CreateNot(PPred);
2951  if (InvertQCond)
2952  QPred = QB.CreateNot(QPred);
2953  Value *CombinedPred = QB.CreateOr(PPred, QPred);
2954 
2955  auto *T =
2956  SplitBlockAndInsertIfThen(CombinedPred, &*QB.GetInsertPoint(), false);
2957  QB.SetInsertPoint(T);
2958  StoreInst *SI = cast<StoreInst>(QB.CreateStore(QPHI, Address));
2959  AAMDNodes AAMD;
2960  PStore->getAAMetadata(AAMD, /*Merge=*/false);
2961  PStore->getAAMetadata(AAMD, /*Merge=*/true);
2962  SI->setAAMetadata(AAMD);
2963  unsigned PAlignment = PStore->getAlignment();
2964  unsigned QAlignment = QStore->getAlignment();
2965  unsigned TypeAlignment =
2967  unsigned MinAlignment;
2968  unsigned MaxAlignment;
2969  std::tie(MinAlignment, MaxAlignment) = std::minmax(PAlignment, QAlignment);
2970  // Choose the minimum alignment. If we could prove both stores execute, we
2971  // could use biggest one. In this case, though, we only know that one of the
2972  // stores executes. And we don't know it's safe to take the alignment from a
2973  // store that doesn't execute.
2974  if (MinAlignment != 0) {
2975  // Choose the minimum of all non-zero alignments.
2976  SI->setAlignment(MinAlignment);
2977  } else if (MaxAlignment != 0) {
2978  // Choose the minimal alignment between the non-zero alignment and the ABI
2979  // default alignment for the type of the stored value.
2980  SI->setAlignment(std::min(MaxAlignment, TypeAlignment));
2981  } else {
2982  // If both alignments are zero, use ABI default alignment for the type of
2983  // the stored value.
2984  SI->setAlignment(TypeAlignment);
2985  }
2986 
2987  QStore->eraseFromParent();
2988  PStore->eraseFromParent();
2989 
2990  return true;
2991 }
2992 
2994  const DataLayout &DL) {
2995  // The intention here is to find diamonds or triangles (see below) where each
2996  // conditional block contains a store to the same address. Both of these
2997  // stores are conditional, so they can't be unconditionally sunk. But it may
2998  // be profitable to speculatively sink the stores into one merged store at the
2999  // end, and predicate the merged store on the union of the two conditions of
3000  // PBI and QBI.
3001  //
3002  // This can reduce the number of stores executed if both of the conditions are
3003  // true, and can allow the blocks to become small enough to be if-converted.
3004  // This optimization will also chain, so that ladders of test-and-set
3005  // sequences can be if-converted away.
3006  //
3007  // We only deal with simple diamonds or triangles:
3008  //
3009  // PBI or PBI or a combination of the two
3010  // / \ | \
3011  // PTB PFB | PFB
3012  // \ / | /
3013  // QBI QBI
3014  // / \ | \
3015  // QTB QFB | QFB
3016  // \ / | /
3017  // PostBB PostBB
3018  //
3019  // We model triangles as a type of diamond with a nullptr "true" block.
3020  // Triangles are canonicalized so that the fallthrough edge is represented by
3021  // a true condition, as in the diagram above.
3022  //
3023  BasicBlock *PTB = PBI->getSuccessor(0);
3024  BasicBlock *PFB = PBI->getSuccessor(1);
3025  BasicBlock *QTB = QBI->getSuccessor(0);
3026  BasicBlock *QFB = QBI->getSuccessor(1);
3027  BasicBlock *PostBB = QFB->getSingleSuccessor();
3028 
3029  // Make sure we have a good guess for PostBB. If QTB's only successor is
3030  // QFB, then QFB is a better PostBB.
3031  if (QTB->getSingleSuccessor() == QFB)
3032  PostBB = QFB;
3033 
3034  // If we couldn't find a good PostBB, stop.
3035  if (!PostBB)
3036  return false;
3037 
3038  bool InvertPCond = false, InvertQCond = false;
3039  // Canonicalize fallthroughs to the true branches.
3040  if (PFB == QBI->getParent()) {
3041  std::swap(PFB, PTB);
3042  InvertPCond = true;
3043  }
3044  if (QFB == PostBB) {
3045  std::swap(QFB, QTB);
3046  InvertQCond = true;
3047  }
3048 
3049  // From this point on we can assume PTB or QTB may be fallthroughs but PFB
3050  // and QFB may not. Model fallthroughs as a nullptr block.
3051  if (PTB == QBI->getParent())
3052  PTB = nullptr;
3053  if (QTB == PostBB)
3054  QTB = nullptr;
3055 
3056  // Legality bailouts. We must have at least the non-fallthrough blocks and
3057  // the post-dominating block, and the non-fallthroughs must only have one
3058  // predecessor.
3059  auto HasOnePredAndOneSucc = [](BasicBlock *BB, BasicBlock *P, BasicBlock *S) {
3060  return BB->getSinglePredecessor() == P && BB->getSingleSuccessor() == S;
3061  };
3062  if (!HasOnePredAndOneSucc(PFB, PBI->getParent(), QBI->getParent()) ||
3063  !HasOnePredAndOneSucc(QFB, QBI->getParent(), PostBB))
3064  return false;
3065  if ((PTB && !HasOnePredAndOneSucc(PTB, PBI->getParent(), QBI->getParent())) ||
3066  (QTB && !HasOnePredAndOneSucc(QTB, QBI->getParent(), PostBB)))
3067  return false;
3068  if (!PostBB->hasNUses(2) || !QBI->getParent()->hasNUses(2))
3069  return false;
3070 
3071  // OK, this is a sequence of two diamonds or triangles.
3072  // Check if there are stores in PTB or PFB that are repeated in QTB or QFB.
3073  SmallPtrSet<Value *, 4> PStoreAddresses, QStoreAddresses;
3074  for (auto *BB : {PTB, PFB}) {
3075  if (!BB)
3076  continue;
3077  for (auto &I : *BB)
3078  if (StoreInst *SI = dyn_cast<StoreInst>(&I))
3079  PStoreAddresses.insert(SI->getPointerOperand());
3080  }
3081  for (auto *BB : {QTB, QFB}) {
3082  if (!BB)
3083  continue;
3084  for (auto &I : *BB)
3085  if (StoreInst *SI = dyn_cast<StoreInst>(&I))
3086  QStoreAddresses.insert(SI->getPointerOperand());
3087  }
3088 
3089  set_intersect(PStoreAddresses, QStoreAddresses);
3090  // set_intersect mutates PStoreAddresses in place. Rename it here to make it
3091  // clear what it contains.
3092  auto &CommonAddresses = PStoreAddresses;
3093 
3094  bool Changed = false;
3095  for (auto *Address : CommonAddresses)
3096  Changed |= mergeConditionalStoreToAddress(
3097  PTB, PFB, QTB, QFB, PostBB, Address, InvertPCond, InvertQCond, DL);
3098  return Changed;
3099 }
3100 
3101 /// If we have a conditional branch as a predecessor of another block,
3102 /// this function tries to simplify it. We know
3103 /// that PBI and BI are both conditional branches, and BI is in one of the
3104 /// successor blocks of PBI - PBI branches to BI.
3106  const DataLayout &DL) {
3107  assert(PBI->isConditional() && BI->isConditional());
3108  BasicBlock *BB = BI->getParent();
3109 
3110  // If this block ends with a branch instruction, and if there is a
3111  // predecessor that ends on a branch of the same condition, make
3112  // this conditional branch redundant.
3113  if (PBI->getCondition() == BI->getCondition() &&
3114  PBI->getSuccessor(0) != PBI->getSuccessor(1)) {
3115  // Okay, the outcome of this conditional branch is statically
3116  // knowable. If this block had a single pred, handle specially.
3117  if (BB->getSinglePredecessor()) {
3118  // Turn this into a branch on constant.
3119  bool CondIsTrue = PBI->getSuccessor(0) == BB;
3120  BI->setCondition(
3121  ConstantInt::get(Type::getInt1Ty(BB->getContext()), CondIsTrue));
3122  return true; // Nuke the branch on constant.
3123  }
3124 
3125  // Otherwise, if there are multiple predecessors, insert a PHI that merges
3126  // in the constant and simplify the block result. Subsequent passes of
3127  // simplifycfg will thread the block.
3129  pred_iterator PB = pred_begin(BB), PE = pred_end(BB);
3130  PHINode *NewPN = PHINode::Create(
3131  Type::getInt1Ty(BB->getContext()), std::distance(PB, PE),
3132  BI->getCondition()->getName() + ".pr", &BB->front());
3133  // Okay, we're going to insert the PHI node. Since PBI is not the only
3134  // predecessor, compute the PHI'd conditional value for all of the preds.
3135  // Any predecessor where the condition is not computable we keep symbolic.
3136  for (pred_iterator PI = PB; PI != PE; ++PI) {
3137  BasicBlock *P = *PI;
3138  if ((PBI = dyn_cast<BranchInst>(P->getTerminator())) && PBI != BI &&
3139  PBI->isConditional() && PBI->getCondition() == BI->getCondition() &&
3140  PBI->getSuccessor(0) != PBI->getSuccessor(1)) {
3141  bool CondIsTrue = PBI->getSuccessor(0) == BB;
3142  NewPN->addIncoming(
3143  ConstantInt::get(Type::getInt1Ty(BB->getContext()), CondIsTrue),
3144  P);
3145  } else {
3146  NewPN->addIncoming(BI->getCondition(), P);
3147  }
3148  }
3149 
3150  BI->setCondition(NewPN);
3151  return true;
3152  }
3153  }
3154 
3155  if (auto *CE = dyn_cast<ConstantExpr>(BI->getCondition()))
3156  if (CE->canTrap())
3157  return false;
3158 
3159  // If both branches are conditional and both contain stores to the same
3160  // address, remove the stores from the conditionals and create a conditional
3161  // merged store at the end.
3162  if (MergeCondStores && mergeConditionalStores(PBI, BI, DL))
3163  return true;
3164 
3165  // If this is a conditional branch in an empty block, and if any
3166  // predecessors are a conditional branch to one of our destinations,
3167  // fold the conditions into logical ops and one cond br.
3168  BasicBlock::iterator BBI = BB->begin();
3169  // Ignore dbg intrinsics.
3170  while (isa<DbgInfoIntrinsic>(BBI))
3171  ++BBI;
3172  if (&*BBI != BI)
3173  return false;
3174 
3175  int PBIOp, BIOp;
3176  if (PBI->getSuccessor(0) == BI->getSuccessor(0)) {
3177  PBIOp = 0;
3178  BIOp = 0;
3179  } else if (PBI->getSuccessor(0) == BI->getSuccessor(1)) {
3180  PBIOp = 0;
3181  BIOp = 1;
3182  } else if (PBI->getSuccessor(1) == BI->getSuccessor(0)) {
3183  PBIOp = 1;
3184  BIOp = 0;
3185  } else if (PBI->getSuccessor(1) == BI->getSuccessor(1)) {
3186  PBIOp = 1;
3187  BIOp = 1;
3188  } else {
3189  return false;
3190  }
3191 
3192  // Check to make sure that the other destination of this branch
3193  // isn't BB itself. If so, this is an infinite loop that will
3194  // keep getting unwound.
3195  if (PBI->getSuccessor(PBIOp) == BB)
3196  return false;
3197 
3198  // Do not perform this transformation if it would require
3199  // insertion of a large number of select instructions. For targets
3200  // without predication/cmovs, this is a big pessimization.
3201 
3202  // Also do not perform this transformation if any phi node in the common
3203  // destination block can trap when reached by BB or PBB (PR17073). In that
3204  // case, it would be unsafe to hoist the operation into a select instruction.
3205 
3206  BasicBlock *CommonDest = PBI->getSuccessor(PBIOp);
3207  unsigned NumPhis = 0;
3208  for (BasicBlock::iterator II = CommonDest->begin(); isa<PHINode>(II);
3209  ++II, ++NumPhis) {
3210  if (NumPhis > 2) // Disable this xform.
3211  return false;
3212 
3213  PHINode *PN = cast<PHINode>(II);
3214  Value *BIV = PN->getIncomingValueForBlock(BB);
3215  if (ConstantExpr *CE = dyn_cast<ConstantExpr>(BIV))
3216  if (CE->canTrap())
3217  return false;
3218 
3219  unsigned PBBIdx = PN->getBasicBlockIndex(PBI->getParent());
3220  Value *PBIV = PN->getIncomingValue(PBBIdx);
3221  if (ConstantExpr *CE = dyn_cast<ConstantExpr>(PBIV))
3222  if (CE->canTrap())
3223  return false;
3224  }
3225 
3226  // Finally, if everything is ok, fold the branches to logical ops.
3227  BasicBlock *OtherDest = BI->getSuccessor(BIOp ^ 1);
3228 
3229  DEBUG(dbgs() << "FOLDING BRs:" << *PBI->getParent()
3230  << "AND: " << *BI->getParent());
3231 
3232  // If OtherDest *is* BB, then BB is a basic block with a single conditional
3233  // branch in it, where one edge (OtherDest) goes back to itself but the other
3234  // exits. We don't *know* that the program avoids the infinite loop
3235  // (even though that seems likely). If we do this xform naively, we'll end up
3236  // recursively unpeeling the loop. Since we know that (after the xform is
3237  // done) that the block *is* infinite if reached, we just make it an obviously
3238  // infinite loop with no cond branch.
3239  if (OtherDest == BB) {
3240  // Insert it at the end of the function, because it's either code,
3241  // or it won't matter if it's hot. :)
3242  BasicBlock *InfLoopBlock =
3243  BasicBlock::Create(BB->getContext(), "infloop", BB->getParent());
3244  BranchInst::Create(InfLoopBlock, InfLoopBlock);
3245  OtherDest = InfLoopBlock;
3246  }
3247 
3248  DEBUG(dbgs() << *PBI->getParent()->getParent());
3249 
3250  // BI may have other predecessors. Because of this, we leave
3251  // it alone, but modify PBI.
3252 
3253  // Make sure we get to CommonDest on True&True directions.
3254  Value *PBICond = PBI->getCondition();
3255  IRBuilder<NoFolder> Builder(PBI);
3256  if (PBIOp)
3257  PBICond = Builder.CreateNot(PBICond, PBICond->getName() + ".not");
3258 
3259  Value *BICond = BI->getCondition();
3260  if (BIOp)
3261  BICond = Builder.CreateNot(BICond, BICond->getName() + ".not");
3262 
3263  // Merge the conditions.
3264  Value *Cond = Builder.CreateOr(PBICond, BICond, "brmerge");
3265 
3266  // Modify PBI to branch on the new condition to the new dests.
3267  PBI->setCondition(Cond);
3268  PBI->setSuccessor(0, CommonDest);
3269  PBI->setSuccessor(1, OtherDest);
3270 
3271  // Update branch weight for PBI.
3272  uint64_t PredTrueWeight, PredFalseWeight, SuccTrueWeight, SuccFalseWeight;
3273  uint64_t PredCommon, PredOther, SuccCommon, SuccOther;
3274  bool HasWeights =
3275  extractPredSuccWeights(PBI, BI, PredTrueWeight, PredFalseWeight,
3276  SuccTrueWeight, SuccFalseWeight);
3277  if (HasWeights) {
3278  PredCommon = PBIOp ? PredFalseWeight : PredTrueWeight;
3279  PredOther = PBIOp ? PredTrueWeight : PredFalseWeight;
3280  SuccCommon = BIOp ? SuccFalseWeight : SuccTrueWeight;
3281  SuccOther = BIOp ? SuccTrueWeight : SuccFalseWeight;
3282  // The weight to CommonDest should be PredCommon * SuccTotal +
3283  // PredOther * SuccCommon.
3284  // The weight to OtherDest should be PredOther * SuccOther.
3285  uint64_t NewWeights[2] = {PredCommon * (SuccCommon + SuccOther) +
3286  PredOther * SuccCommon,
3287  PredOther * SuccOther};
3288  // Halve the weights if any of them cannot fit in an uint32_t
3289  FitWeights(NewWeights);
3290 
3292  MDBuilder(BI->getContext())
3293  .createBranchWeights(NewWeights[0], NewWeights[1]));
3294  }
3295 
3296  // OtherDest may have phi nodes. If so, add an entry from PBI's
3297  // block that are identical to the entries for BI's block.
3298  AddPredecessorToBlock(OtherDest, PBI->getParent(), BB);
3299 
3300  // We know that the CommonDest already had an edge from PBI to
3301  // it. If it has PHIs though, the PHIs may have different
3302  // entries for BB and PBI's BB. If so, insert a select to make
3303  // them agree.
3304  PHINode *PN;
3305  for (BasicBlock::iterator II = CommonDest->begin();
3306  (PN = dyn_cast<PHINode>(II)); ++II) {
3307  Value *BIV = PN->getIncomingValueForBlock(BB);
3308  unsigned PBBIdx = PN->getBasicBlockIndex(PBI->getParent());
3309  Value *PBIV = PN->getIncomingValue(PBBIdx);
3310  if (BIV != PBIV) {
3311  // Insert a select in PBI to pick the right value.
3312  SelectInst *NV = cast<SelectInst>(
3313  Builder.CreateSelect(PBICond, PBIV, BIV, PBIV->getName() + ".mux"));
3314  PN->setIncomingValue(PBBIdx, NV);
3315  // Although the select has the same condition as PBI, the original branch
3316  // weights for PBI do not apply to the new select because the select's
3317  // 'logical' edges are incoming edges of the phi that is eliminated, not
3318  // the outgoing edges of PBI.
3319  if (HasWeights) {
3320  uint64_t PredCommon = PBIOp ? PredFalseWeight : PredTrueWeight;
3321  uint64_t PredOther = PBIOp ? PredTrueWeight : PredFalseWeight;
3322  uint64_t SuccCommon = BIOp ? SuccFalseWeight : SuccTrueWeight;
3323  uint64_t SuccOther = BIOp ? SuccTrueWeight : SuccFalseWeight;
3324  // The weight to PredCommonDest should be PredCommon * SuccTotal.
3325  // The weight to PredOtherDest should be PredOther * SuccCommon.
3326  uint64_t NewWeights[2] = {PredCommon * (SuccCommon + SuccOther),
3327  PredOther * SuccCommon};
3328 
3329  FitWeights(NewWeights);
3330 
3332  MDBuilder(BI->getContext())
3333  .createBranchWeights(NewWeights[0], NewWeights[1]));
3334  }
3335  }
3336  }
3337 
3338  DEBUG(dbgs() << "INTO: " << *PBI->getParent());
3339  DEBUG(dbgs() << *PBI->getParent()->getParent());
3340 
3341  // This basic block is probably dead. We know it has at least
3342  // one fewer predecessor.
3343  return true;
3344 }
3345 
3346 // Simplifies a terminator by replacing it with a branch to TrueBB if Cond is
3347 // true or to FalseBB if Cond is false.
3348 // Takes care of updating the successors and removing the old terminator.
3349 // Also makes sure not to introduce new successors by assuming that edges to
3350 // non-successor TrueBBs and FalseBBs aren't reachable.
3352  BasicBlock *TrueBB, BasicBlock *FalseBB,
3353  uint32_t TrueWeight,
3354  uint32_t FalseWeight) {
3355  // Remove any superfluous successor edges from the CFG.
3356  // First, figure out which successors to preserve.
3357  // If TrueBB and FalseBB are equal, only try to preserve one copy of that
3358  // successor.
3359  BasicBlock *KeepEdge1 = TrueBB;
3360  BasicBlock *KeepEdge2 = TrueBB != FalseBB ? FalseBB : nullptr;
3361 
3362  // Then remove the rest.
3363  for (BasicBlock *Succ : OldTerm->successors()) {
3364  // Make sure only to keep exactly one copy of each edge.
3365  if (Succ == KeepEdge1)
3366  KeepEdge1 = nullptr;
3367  else if (Succ == KeepEdge2)
3368  KeepEdge2 = nullptr;
3369  else
3370  Succ->removePredecessor(OldTerm->getParent(),
3371  /*DontDeleteUselessPHIs=*/true);
3372  }
3373 
3374  IRBuilder<> Builder(OldTerm);
3375  Builder.SetCurrentDebugLocation(OldTerm->getDebugLoc());
3376 
3377  // Insert an appropriate new terminator.
3378  if (!KeepEdge1 && !KeepEdge2) {
3379  if (TrueBB == FalseBB)
3380  // We were only looking for one successor, and it was present.
3381  // Create an unconditional branch to it.
3382  Builder.CreateBr(TrueBB);
3383  else {
3384  // We found both of the successors we were looking for.
3385  // Create a conditional branch sharing the condition of the select.
3386  BranchInst *NewBI = Builder.CreateCondBr(Cond, TrueBB, FalseBB);
3387  if (TrueWeight != FalseWeight)
3389  MDBuilder(OldTerm->getContext())
3390  .createBranchWeights(TrueWeight, FalseWeight));
3391  }
3392  } else if (KeepEdge1 && (KeepEdge2 || TrueBB == FalseBB)) {
3393  // Neither of the selected blocks were successors, so this
3394  // terminator must be unreachable.
3395  new UnreachableInst(OldTerm->getContext(), OldTerm);
3396  } else {
3397  // One of the selected values was a successor, but the other wasn't.
3398  // Insert an unconditional branch to the one that was found;
3399  // the edge to the one that wasn't must be unreachable.
3400  if (!KeepEdge1)
3401  // Only TrueBB was found.
3402  Builder.CreateBr(TrueBB);
3403  else
3404  // Only FalseBB was found.
3405  Builder.CreateBr(FalseBB);
3406  }
3407 
3409  return true;
3410 }
3411 
3412 // Replaces
3413 // (switch (select cond, X, Y)) on constant X, Y
3414 // with a branch - conditional if X and Y lead to distinct BBs,
3415 // unconditional otherwise.
3417  // Check for constant integer values in the select.
3418  ConstantInt *TrueVal = dyn_cast<ConstantInt>(Select->getTrueValue());
3419  ConstantInt *FalseVal = dyn_cast<ConstantInt>(Select->getFalseValue());
3420  if (!TrueVal || !FalseVal)
3421  return false;
3422 
3423  // Find the relevant condition and destinations.
3424  Value *Condition = Select->getCondition();
3425  BasicBlock *TrueBB = SI->findCaseValue(TrueVal)->getCaseSuccessor();
3426  BasicBlock *FalseBB = SI->findCaseValue(FalseVal)->getCaseSuccessor();
3427 
3428  // Get weight for TrueBB and FalseBB.
3429  uint32_t TrueWeight = 0, FalseWeight = 0;
3430  SmallVector<uint64_t, 8> Weights;
3431  bool HasWeights = HasBranchWeights(SI);
3432  if (HasWeights) {
3433  GetBranchWeights(SI, Weights);
3434  if (Weights.size() == 1 + SI->getNumCases()) {
3435  TrueWeight =
3436  (uint32_t)Weights[SI->findCaseValue(TrueVal)->getSuccessorIndex()];
3437  FalseWeight =
3438  (uint32_t)Weights[SI->findCaseValue(FalseVal)->getSuccessorIndex()];
3439  }
3440  }
3441 
3442  // Perform the actual simplification.
3443  return SimplifyTerminatorOnSelect(SI, Condition, TrueBB, FalseBB, TrueWeight,
3444  FalseWeight);
3445 }
3446 
3447 // Replaces
3448 // (indirectbr (select cond, blockaddress(@fn, BlockA),
3449 // blockaddress(@fn, BlockB)))
3450 // with
3451 // (br cond, BlockA, BlockB).
3453  // Check that both operands of the select are block addresses.
3456  if (!TBA || !FBA)
3457  return false;
3458 
3459  // Extract the actual blocks.
3460  BasicBlock *TrueBB = TBA->getBasicBlock();
3461  BasicBlock *FalseBB = FBA->getBasicBlock();
3462 
3463  // Perform the actual simplification.
3464  return SimplifyTerminatorOnSelect(IBI, SI->getCondition(), TrueBB, FalseBB, 0,
3465  0);
3466 }
3467 
3468 /// This is called when we find an icmp instruction
3469 /// (a seteq/setne with a constant) as the only instruction in a
3470 /// block that ends with an uncond branch. We are looking for a very specific
3471 /// pattern that occurs when "A == 1 || A == 2 || A == 3" gets simplified. In
3472 /// this case, we merge the first two "or's of icmp" into a switch, but then the
3473 /// default value goes to an uncond block with a seteq in it, we get something
3474 /// like:
3475 ///
3476 /// switch i8 %A, label %DEFAULT [ i8 1, label %end i8 2, label %end ]
3477 /// DEFAULT:
3478 /// %tmp = icmp eq i8 %A, 92
3479 /// br label %end
3480 /// end:
3481 /// ... = phi i1 [ true, %entry ], [ %tmp, %DEFAULT ], [ true, %entry ]
3482 ///
3483 /// We prefer to split the edge to 'end' so that there is a true/false entry to
3484 /// the PHI, merging the third icmp into the switch.
3486  ICmpInst *ICI, IRBuilder<> &Builder, const DataLayout &DL,
3487  const TargetTransformInfo &TTI, const SimplifyCFGOptions &Options) {
3488  BasicBlock *BB = ICI->getParent();
3489 
3490  // If the block has any PHIs in it or the icmp has multiple uses, it is too
3491  // complex.
3492  if (isa<PHINode>(BB->begin()) || !ICI->hasOneUse())
3493  return false;
3494 
3495  Value *V = ICI->getOperand(0);
3496  ConstantInt *Cst = cast<ConstantInt>(ICI->getOperand(1));
3497 
3498  // The pattern we're looking for is where our only predecessor is a switch on
3499  // 'V' and this block is the default case for the switch. In this case we can
3500  // fold the compared value into the switch to simplify things.
3501  BasicBlock *Pred = BB->getSinglePredecessor();
3502  if (!Pred || !isa<SwitchInst>(Pred->getTerminator()))
3503  return false;
3504 
3505  SwitchInst *SI = cast<SwitchInst>(Pred->getTerminator());
3506  if (SI->getCondition() != V)
3507  return false;
3508 
3509  // If BB is reachable on a non-default case, then we simply know the value of
3510  // V in this block. Substitute it and constant fold the icmp instruction
3511  // away.
3512  if (SI->getDefaultDest() != BB) {
3513  ConstantInt *VVal = SI->findCaseDest(BB);
3514  assert(VVal && "Should have a unique destination value");
3515  ICI->setOperand(0, VVal);
3516 
3517  if (Value *V = SimplifyInstruction(ICI, {DL, ICI})) {
3518  ICI->replaceAllUsesWith(V);
3519  ICI->eraseFromParent();
3520  }
3521  // BB is now empty, so it is likely to simplify away.
3522  return simplifyCFG(BB, TTI, Options) | true;
3523  }
3524 
3525  // Ok, the block is reachable from the default dest. If the constant we're
3526  // comparing exists in one of the other edges, then we can constant fold ICI
3527  // and zap it.
3528  if (SI->findCaseValue(Cst) != SI->case_default()) {
3529  Value *V;
3530  if (ICI->getPredicate() == ICmpInst::ICMP_EQ)
3531  V = ConstantInt::getFalse(BB->getContext());
3532  else
3533  V = ConstantInt::getTrue(BB->getContext());
3534 
3535  ICI->replaceAllUsesWith(V);
3536  ICI->eraseFromParent();
3537  // BB is now empty, so it is likely to simplify away.
3538  return simplifyCFG(BB, TTI, Options) | true;
3539  }
3540 
3541  // The use of the icmp has to be in the 'end' block, by the only PHI node in
3542  // the block.
3543  BasicBlock *SuccBlock = BB->getTerminator()->getSuccessor(0);
3544  PHINode *PHIUse = dyn_cast<PHINode>(ICI->user_back());
3545  if (PHIUse == nullptr || PHIUse != &SuccBlock->front() ||
3546  isa<PHINode>(++BasicBlock::iterator(PHIUse)))
3547  return false;
3548 
3549  // If the icmp is a SETEQ, then the default dest gets false, the new edge gets
3550  // true in the PHI.
3551  Constant *DefaultCst = ConstantInt::getTrue(BB->getContext());
3552  Constant *NewCst = ConstantInt::getFalse(BB->getContext());
3553 
3554  if (ICI->getPredicate() == ICmpInst::ICMP_EQ)
3555  std::swap(DefaultCst, NewCst);
3556 
3557  // Replace ICI (which is used by the PHI for the default value) with true or
3558  // false depending on if it is EQ or NE.
3559  ICI->replaceAllUsesWith(DefaultCst);
3560  ICI->eraseFromParent();
3561 
3562  // Okay, the switch goes to this block on a default value. Add an edge from
3563  // the switch to the merge point on the compared value.
3564  BasicBlock *NewBB =
3565  BasicBlock::Create(BB->getContext(), "switch.edge", BB->getParent(), BB);
3566  SmallVector<uint64_t, 8> Weights;
3567  bool HasWeights = HasBranchWeights(SI);
3568  if (HasWeights) {
3569  GetBranchWeights(SI, Weights);
3570  if (Weights.size() == 1 + SI->getNumCases()) {
3571  // Split weight for default case to case for "Cst".
3572  Weights[0] = (Weights[0] + 1) >> 1;
3573  Weights.push_back(Weights[0]);
3574 
3575  SmallVector<uint32_t, 8> MDWeights(Weights.begin(), Weights.end());
3576  SI->setMetadata(
3578  MDBuilder(SI->getContext()).createBranchWeights(MDWeights));
3579  }
3580  }
3581  SI->addCase(Cst, NewBB);
3582 
3583  // NewBB branches to the phi block, add the uncond branch and the phi entry.
3584  Builder.SetInsertPoint(NewBB);
3585  Builder.SetCurrentDebugLocation(SI->getDebugLoc());
3586  Builder.CreateBr(SuccBlock);
3587  PHIUse->addIncoming(NewCst, NewBB);
3588  return true;
3589 }
3590 
3591 /// The specified branch is a conditional branch.
3592 /// Check to see if it is branching on an or/and chain of icmp instructions, and
3593 /// fold it into a switch instruction if so.
3595  const DataLayout &DL) {
3596  Instruction *Cond = dyn_cast<Instruction>(BI->getCondition());
3597  if (!Cond)
3598  return false;
3599 
3600  // Change br (X == 0 | X == 1), T, F into a switch instruction.
3601  // If this is a bunch of seteq's or'd together, or if it's a bunch of
3602  // 'setne's and'ed together, collect them.
3603 
3604  // Try to gather values from a chain of and/or to be turned into a switch
3605  ConstantComparesGatherer ConstantCompare(Cond, DL);
3606  // Unpack the result
3607  SmallVectorImpl<ConstantInt *> &Values = ConstantCompare.Vals;
3608  Value *CompVal = ConstantCompare.CompValue;
3609  unsigned UsedICmps = ConstantCompare.UsedICmps;
3610  Value *ExtraCase = ConstantCompare.Extra;
3611 
3612  // If we didn't have a multiply compared value, fail.
3613  if (!CompVal)
3614  return false;
3615 
3616  // Avoid turning single icmps into a switch.
3617  if (UsedICmps <= 1)
3618  return false;
3619 
3620  bool TrueWhenEqual = (Cond->getOpcode() == Instruction::Or);
3621 
3622  // There might be duplicate constants in the list, which the switch
3623  // instruction can't handle, remove them now.
3624  array_pod_sort(Values.begin(), Values.end(), ConstantIntSortPredicate);
3625  Values.erase(std::unique(Values.begin(), Values.end()), Values.end());
3626 
3627  // If Extra was used, we require at least two switch values to do the
3628  // transformation. A switch with one value is just a conditional branch.
3629  if (ExtraCase && Values.size() < 2)
3630  return false;
3631 
3632  // TODO: Preserve branch weight metadata, similarly to how
3633  // FoldValueComparisonIntoPredecessors preserves it.
3634 
3635  // Figure out which block is which destination.
3636  BasicBlock *DefaultBB = BI->getSuccessor(1);
3637  BasicBlock *EdgeBB = BI->getSuccessor(0);
3638  if (!TrueWhenEqual)
3639  std::swap(DefaultBB, EdgeBB);
3640 
3641  BasicBlock *BB = BI->getParent();
3642 
3643  DEBUG(dbgs() << "Converting 'icmp' chain with " << Values.size()
3644  << " cases into SWITCH. BB is:\n"
3645  << *BB);
3646 
3647  // If there are any extra values that couldn't be folded into the switch
3648  // then we evaluate them with an explicit branch first. Split the block
3649  // right before the condbr to handle it.
3650  if (ExtraCase) {
3651  BasicBlock *NewBB =
3652  BB->splitBasicBlock(BI->getIterator(), "switch.early.test");
3653  // Remove the uncond branch added to the old block.
3654  TerminatorInst *OldTI = BB->getTerminator();
3655  Builder.SetInsertPoint(OldTI);
3656 
3657  if (TrueWhenEqual)
3658  Builder.CreateCondBr(ExtraCase, EdgeBB, NewBB);
3659  else
3660  Builder.CreateCondBr(ExtraCase, NewBB, EdgeBB);
3661 
3662  OldTI->eraseFromParent();
3663 
3664  // If there are PHI nodes in EdgeBB, then we need to add a new entry to them
3665  // for the edge we just added.
3666  AddPredecessorToBlock(EdgeBB, BB, NewBB);
3667 
3668  DEBUG(dbgs() << " ** 'icmp' chain unhandled condition: " << *ExtraCase
3669  << "\nEXTRABB = " << *BB);
3670  BB = NewBB;
3671  }
3672 
3673  Builder.SetInsertPoint(BI);
3674  // Convert pointer to int before we switch.
3675  if (CompVal->getType()->isPointerTy()) {
3676  CompVal = Builder.CreatePtrToInt(
3677  CompVal, DL.getIntPtrType(CompVal->getType()), "magicptr");
3678  }
3679 
3680  // Create the new switch instruction now.
3681  SwitchInst *New = Builder.CreateSwitch(CompVal, DefaultBB, Values.size());
3682 
3683  // Add all of the 'cases' to the switch instruction.
3684  for (unsigned i = 0, e = Values.size(); i != e; ++i)
3685  New->addCase(Values[i], EdgeBB);
3686 
3687  // We added edges from PI to the EdgeBB. As such, if there were any
3688  // PHI nodes in EdgeBB, they need entries to be added corresponding to
3689  // the number of edges added.
3690  for (BasicBlock::iterator BBI = EdgeBB->begin(); isa<PHINode>(BBI); ++BBI) {
3691  PHINode *PN = cast<PHINode>(BBI);
3692  Value *InVal = PN->getIncomingValueForBlock(BB);
3693  for (unsigned i = 0, e = Values.size() - 1; i != e; ++i)
3694  PN->addIncoming(InVal, BB);
3695  }
3696 
3697  // Erase the old branch instruction.
3699 
3700  DEBUG(dbgs() << " ** 'icmp' chain result is:\n" << *BB << '\n');
3701  return true;
3702 }
3703 
3704 bool SimplifyCFGOpt::SimplifyResume(ResumeInst *RI, IRBuilder<> &Builder) {
3705  if (isa<PHINode>(RI->getValue()))
3706  return SimplifyCommonResume(RI);
3707  else if (isa<LandingPadInst>(RI->getParent()->getFirstNonPHI()) &&
3708  RI->getValue() == RI->getParent()->getFirstNonPHI())
3709  // The resume must unwind the exception that caused control to branch here.
3710  return SimplifySingleResume(RI);
3711 
3712  return false;
3713 }
3714 
3715 // Simplify resume that is shared by several landing pads (phi of landing pad).
3716 bool SimplifyCFGOpt::SimplifyCommonResume(ResumeInst *RI) {
3717  BasicBlock *BB = RI->getParent();
3718 
3719  // Check that there are no other instructions except for debug intrinsics
3720  // between the phi of landing pads (RI->getValue()) and resume instruction.
3721  BasicBlock::iterator I = cast<Instruction>(RI->getValue())->getIterator(),
3722  E = RI->getIterator();
3723  while (++I != E)
3724  if (!isa<DbgInfoIntrinsic>(I))
3725  return false;
3726 
3727  SmallSetVector<BasicBlock *, 4> TrivialUnwindBlocks;
3728  auto *PhiLPInst = cast<PHINode>(RI->getValue());
3729 
3730  // Check incoming blocks to see if any of them are trivial.
3731  for (unsigned Idx = 0, End = PhiLPInst->getNumIncomingValues(); Idx != End;
3732  Idx++) {
3733  auto *IncomingBB = PhiLPInst->getIncomingBlock(Idx);
3734  auto *IncomingValue = PhiLPInst->getIncomingValue(Idx);
3735 
3736  // If the block has other successors, we can not delete it because
3737  // it has other dependents.
3738  if (IncomingBB->getUniqueSuccessor() != BB)
3739  continue;
3740 
3741  auto *LandingPad = dyn_cast<LandingPadInst>(IncomingBB->getFirstNonPHI());
3742  // Not the landing pad that caused the control to branch here.
3743  if (IncomingValue != LandingPad)
3744  continue;
3745 
3746  bool isTrivial = true;
3747 
3748  I = IncomingBB->getFirstNonPHI()->getIterator();
3749  E = IncomingBB->getTerminator()->getIterator();
3750  while (++I != E)
3751  if (!isa<DbgInfoIntrinsic>(I)) {
3752  isTrivial = false;
3753  break;
3754  }
3755 
3756  if (isTrivial)
3757  TrivialUnwindBlocks.insert(IncomingBB);
3758  }
3759 
3760  // If no trivial unwind blocks, don't do any simplifications.
3761  if (TrivialUnwindBlocks.empty())
3762  return false;
3763 
3764  // Turn all invokes that unwind here into calls.
3765  for (auto *TrivialBB : TrivialUnwindBlocks) {
3766  // Blocks that will be simplified should be removed from the phi node.
3767  // Note there could be multiple edges to the resume block, and we need
3768  // to remove them all.
3769  while (PhiLPInst->getBasicBlockIndex(TrivialBB) != -1)
3770  BB->removePredecessor(TrivialBB, true);
3771 
3772  for (pred_iterator PI = pred_begin(TrivialBB), PE = pred_end(TrivialBB);
3773  PI != PE;) {
3774  BasicBlock *Pred = *PI++;
3775  removeUnwindEdge(Pred);
3776  }
3777 
3778  // In each SimplifyCFG run, only the current processed block can be erased.
3779  // Otherwise, it will break the iteration of SimplifyCFG pass. So instead
3780  // of erasing TrivialBB, we only remove the branch to the common resume
3781  // block so that we can later erase the resume block since it has no
3782  // predecessors.
3783  TrivialBB->getTerminator()->eraseFromParent();
3784  new UnreachableInst(RI->getContext(), TrivialBB);
3785  }
3786 
3787  // Delete the resume block if all its predecessors have been removed.
3788  if (pred_empty(BB))
3789  BB->eraseFromParent();
3790 
3791  return !TrivialUnwindBlocks.empty();
3792 }
3793 
3794 // Simplify resume that is only used by a single (non-phi) landing pad.
3795 bool SimplifyCFGOpt::SimplifySingleResume(ResumeInst *RI) {
3796  BasicBlock *BB = RI->getParent();
3798  assert(RI->getValue() == LPInst &&
3799  "Resume must unwind the exception that caused control to here");
3800 
3801  // Check that there are no other instructions except for debug intrinsics.
3802  BasicBlock::iterator I = LPInst->getIterator(), E = RI->getIterator();
3803  while (++I != E)
3804  if (!isa<DbgInfoIntrinsic>(I))
3805  return false;
3806 
3807  // Turn all invokes that unwind here into calls and delete the basic block.
3808  for (pred_iterator PI = pred_begin(BB), PE = pred_end(BB); PI != PE;) {
3809  BasicBlock *Pred = *PI++;
3810  removeUnwindEdge(Pred);
3811  }
3812 
3813  // The landingpad is now unreachable. Zap it.
3814  BB->eraseFromParent();
3815  if (LoopHeaders)
3816  LoopHeaders->erase(BB);
3817  return true;
3818 }
3819 
3821  // If this is a trivial cleanup pad that executes no instructions, it can be
3822  // eliminated. If the cleanup pad continues to the caller, any predecessor
3823  // that is an EH pad will be updated to continue to the caller and any
3824  // predecessor that terminates with an invoke instruction will have its invoke
3825  // instruction converted to a call instruction. If the cleanup pad being
3826  // simplified does not continue to the caller, each predecessor will be
3827  // updated to continue to the unwind destination of the cleanup pad being
3828  // simplified.
3829  BasicBlock *BB = RI->getParent();
3830  CleanupPadInst *CPInst = RI->getCleanupPad();
3831  if (CPInst->getParent() != BB)
3832  // This isn't an empty cleanup.
3833  return false;
3834 
3835  // We cannot kill the pad if it has multiple uses. This typically arises
3836  // from unreachable basic blocks.
3837  if (!CPInst->hasOneUse())
3838  return false;
3839 
3840  // Check that there are no other instructions except for benign intrinsics.
3841  BasicBlock::iterator I = CPInst->getIterator(), E = RI->getIterator();
3842  while (++I != E) {
3843  auto *II = dyn_cast<IntrinsicInst>(I);
3844  if (!II)
3845  return false;
3846 
3847  Intrinsic::ID IntrinsicID = II->getIntrinsicID();
3848  switch (IntrinsicID) {
3849  case Intrinsic::dbg_declare:
3850  case Intrinsic::dbg_value:
3851  case Intrinsic::lifetime_end:
3852  break;
3853  default:
3854  return false;
3855  }
3856  }
3857 
3858  // If the cleanup return we are simplifying unwinds to the caller, this will
3859  // set UnwindDest to nullptr.
3860  BasicBlock *UnwindDest = RI->getUnwindDest();
3861  Instruction *DestEHPad = UnwindDest ? UnwindDest->getFirstNonPHI() : nullptr;
3862 
3863  // We're about to remove BB from the control flow. Before we do, sink any
3864  // PHINodes into the unwind destination. Doing this before changing the
3865  // control flow avoids some potentially slow checks, since we can currently
3866  // be certain that UnwindDest and BB have no common predecessors (since they
3867  // are both EH pads).
3868  if (UnwindDest) {
3869  // First, go through the PHI nodes in UnwindDest and update any nodes that
3870  // reference the block we are removing
3871  for (BasicBlock::iterator I = UnwindDest->begin(),
3872  IE = DestEHPad->getIterator();
3873  I != IE; ++I) {
3874  PHINode *DestPN = cast<PHINode>(I);
3875 
3876  int Idx = DestPN->getBasicBlockIndex(BB);
3877  // Since BB unwinds to UnwindDest, it has to be in the PHI node.
3878  assert(Idx != -1);
3879  // This PHI node has an incoming value that corresponds to a control
3880  // path through the cleanup pad we are removing. If the incoming
3881  // value is in the cleanup pad, it must be a PHINode (because we
3882  // verified above that the block is otherwise empty). Otherwise, the
3883  // value is either a constant or a value that dominates the cleanup
3884  // pad being removed.
3885  //
3886  // Because BB and UnwindDest are both EH pads, all of their
3887  // predecessors must unwind to these blocks, and since no instruction
3888  // can have multiple unwind destinations, there will be no overlap in
3889  // incoming blocks between SrcPN and DestPN.
3890  Value *SrcVal = DestPN->getIncomingValue(Idx);
3891  PHINode *SrcPN = dyn_cast<PHINode>(SrcVal);
3892 
3893  // Remove the entry for the block we are deleting.
3894  DestPN->removeIncomingValue(Idx, false);
3895 
3896  if (SrcPN && SrcPN->getParent() == BB) {
3897  // If the incoming value was a PHI node in the cleanup pad we are
3898  // removing, we need to merge that PHI node's incoming values into
3899  // DestPN.
3900  for (unsigned SrcIdx = 0, SrcE = SrcPN->getNumIncomingValues();
3901  SrcIdx != SrcE; ++SrcIdx) {
3902  DestPN->addIncoming(SrcPN->getIncomingValue(SrcIdx),
3903  SrcPN->getIncomingBlock(SrcIdx));
3904  }
3905  } else {
3906  // Otherwise, the incoming value came from above BB and
3907  // so we can just reuse it. We must associate all of BB's
3908  // predecessors with this value.
3909  for (auto *pred : predecessors(BB)) {
3910  DestPN->addIncoming(SrcVal, pred);
3911  }
3912  }
3913  }
3914 
3915  // Sink any remaining PHI nodes directly into UnwindDest.
3916  Instruction *InsertPt = DestEHPad;
3917  for (BasicBlock::iterator I = BB->begin(),
3918  IE = BB->getFirstNonPHI()->getIterator();
3919  I != IE;) {
3920  // The iterator must be incremented here because the instructions are
3921  // being moved to another block.
3922  PHINode *PN = cast<PHINode>(I++);
3923  if (PN->use_empty())
3924  // If the PHI node has no uses, just leave it. It will be erased
3925  // when we erase BB below.
3926  continue;
3927 
3928  // Otherwise, sink this PHI node into UnwindDest.
3929  // Any predecessors to UnwindDest which are not already represented
3930  // must be back edges which inherit the value from the path through
3931  // BB. In this case, the PHI value must reference itself.
3932  for (auto *pred : predecessors(UnwindDest))
3933  if (pred != BB)
3934  PN->addIncoming(PN, pred);
3935  PN->moveBefore(InsertPt);
3936  }
3937  }
3938 
3939  for (pred_iterator PI = pred_begin(BB), PE = pred_end(BB); PI != PE;) {
3940  // The iterator must be updated here because we are removing this pred.
3941  BasicBlock *PredBB = *PI++;
3942  if (UnwindDest == nullptr) {
3943  removeUnwindEdge(PredBB);
3944  } else {
3945  TerminatorInst *TI = PredBB->getTerminator();
3946  TI->replaceUsesOfWith(BB, UnwindDest);
3947  }
3948  }
3949 
3950  // The cleanup pad is now unreachable. Zap it.
3951  BB->eraseFromParent();
3952  return true;
3953 }
3954 
3955 // Try to merge two cleanuppads together.
3957  // Skip any cleanuprets which unwind to caller, there is nothing to merge
3958  // with.
3959  BasicBlock *UnwindDest = RI->getUnwindDest();
3960  if (!UnwindDest)
3961  return false;
3962 
3963  // This cleanupret isn't the only predecessor of this cleanuppad, it wouldn't
3964  // be safe to merge without code duplication.
3965  if (UnwindDest->getSinglePredecessor() != RI->getParent())
3966  return false;
3967 
3968  // Verify that our cleanuppad's unwind destination is another cleanuppad.
3969  auto *SuccessorCleanupPad = dyn_cast<CleanupPadInst>(&UnwindDest->front());
3970  if (!SuccessorCleanupPad)
3971  return false;
3972 
3973  CleanupPadInst *PredecessorCleanupPad = RI->getCleanupPad();
3974  // Replace any uses of the successor cleanupad with the predecessor pad
3975  // The only cleanuppad uses should be this cleanupret, it's cleanupret and
3976  // funclet bundle operands.
3977  SuccessorCleanupPad->replaceAllUsesWith(PredecessorCleanupPad);
3978  // Remove the old cleanuppad.
3979  SuccessorCleanupPad->eraseFromParent();
3980  // Now, we simply replace the cleanupret with a branch to the unwind
3981  // destination.
3982  BranchInst::Create(UnwindDest, RI->getParent());
3983  RI->eraseFromParent();
3984 
3985  return true;
3986 }
3987 
3988 bool SimplifyCFGOpt::SimplifyCleanupReturn(CleanupReturnInst *RI) {
3989  // It is possible to transiantly have an undef cleanuppad operand because we
3990  // have deleted some, but not all, dead blocks.
3991  // Eventually, this block will be deleted.
3992  if (isa<UndefValue>(RI->getOperand(0)))
3993  return false;
3994 
3995  if (mergeCleanupPad(RI))
3996  return true;
3997 
3998  if (removeEmptyCleanup(RI))
3999  return true;
4000 
4001  return false;
4002 }
4003 
4004 bool SimplifyCFGOpt::SimplifyReturn(ReturnInst *RI, IRBuilder<> &Builder) {
4005  BasicBlock *BB = RI->getParent();
4006  if (!BB->getFirstNonPHIOrDbg()->isTerminator())
4007  return false;
4008 
4009  // Find predecessors that end with branches.
4010  SmallVector<BasicBlock *, 8> UncondBranchPreds;
4011  SmallVector<BranchInst *, 8> CondBranchPreds;
4012  for (pred_iterator PI = pred_begin(BB), E = pred_end(BB); PI != E; ++PI) {
4013  BasicBlock *P = *PI;
4014  TerminatorInst *PTI = P->getTerminator();
4015  if (BranchInst *BI = dyn_cast<BranchInst>(PTI)) {
4016  if (BI->isUnconditional())
4017  UncondBranchPreds.push_back(P);
4018  else
4019  CondBranchPreds.push_back(BI);
4020  }
4021  }
4022 
4023  // If we found some, do the transformation!
4024  if (!UncondBranchPreds.empty() && DupRet) {
4025  while (!UncondBranchPreds.empty()) {
4026  BasicBlock *Pred = UncondBranchPreds.pop_back_val();
4027  DEBUG(dbgs() << "FOLDING: " << *BB
4028  << "INTO UNCOND BRANCH PRED: " << *Pred);
4029  (void)FoldReturnIntoUncondBranch(RI, BB, Pred);
4030  }
4031 
4032  // If we eliminated all predecessors of the block, delete the block now.
4033  if (pred_empty(BB)) {
4034  // We know there are no successors, so just nuke the block.
4035  BB->eraseFromParent();
4036  if (LoopHeaders)
4037  LoopHeaders->erase(BB);
4038  }
4039 
4040  return true;
4041  }
4042 
4043  // Check out all of the conditional branches going to this return
4044  // instruction. If any of them just select between returns, change the
4045  // branch itself into a select/return pair.
4046  while (!CondBranchPreds.empty()) {
4047  BranchInst *BI = CondBranchPreds.pop_back_val();
4048 
4049  // Check to see if the non-BB successor is also a return block.
4050  if (isa<ReturnInst>(BI->getSuccessor(0)->getTerminator()) &&
4051  isa<ReturnInst>(BI->getSuccessor(1)->getTerminator()) &&
4052  SimplifyCondBranchToTwoReturns(BI, Builder))
4053  return true;
4054  }
4055  return false;
4056 }
4057 
4058 bool SimplifyCFGOpt::SimplifyUnreachable(UnreachableInst *UI) {
4059  BasicBlock *BB = UI->getParent();
4060 
4061  bool Changed = false;
4062 
4063  // If there are any instructions immediately before the unreachable that can
4064  // be removed, do so.
4065  while (UI->getIterator() != BB->begin()) {
4066  BasicBlock::iterator BBI = UI->getIterator();
4067  --BBI;
4068  // Do not delete instructions that can have side effects which might cause
4069  // the unreachable to not be reachable; specifically, calls and volatile
4070  // operations may have this effect.
4071  if (isa<CallInst>(BBI) && !isa<DbgInfoIntrinsic>(BBI))
4072  break;
4073 
4074  if (BBI->mayHaveSideEffects()) {
4075  if (auto *SI = dyn_cast<StoreInst>(BBI)) {
4076  if (SI->isVolatile())
4077  break;
4078  } else if (auto *LI = dyn_cast<LoadInst>(BBI)) {
4079  if (LI->isVolatile())
4080  break;
4081  } else if (auto *RMWI = dyn_cast<AtomicRMWInst>(BBI)) {
4082  if (RMWI->isVolatile())
4083  break;
4084  } else if (auto *CXI = dyn_cast<AtomicCmpXchgInst>(BBI)) {
4085  if (CXI->isVolatile())
4086  break;
4087  } else if (isa<CatchPadInst>(BBI)) {
4088  // A catchpad may invoke exception object constructors and such, which
4089  // in some languages can be arbitrary code, so be conservative by
4090  // default.
4091  // For CoreCLR, it just involves a type test, so can be removed.
4094  break;
4095  } else if (!isa<FenceInst>(BBI) && !isa<VAArgInst>(BBI) &&
4096  !isa<LandingPadInst>(BBI)) {
4097  break;
4098  }
4099  // Note that deleting LandingPad's here is in fact okay, although it
4100  // involves a bit of subtle reasoning. If this inst is a LandingPad,
4101  // all the predecessors of this block will be the unwind edges of Invokes,
4102  // and we can therefore guarantee this block will be erased.
4103  }
4104 
4105  // Delete this instruction (any uses are guaranteed to be dead)
4106  if (!BBI->use_empty())
4107  BBI->replaceAllUsesWith(UndefValue::get(BBI->getType()));
4108  BBI->eraseFromParent();
4109  Changed = true;
4110  }
4111 
4112  // If the unreachable instruction is the first in the block, take a gander
4113  // at all of the predecessors of this instruction, and simplify them.
4114  if (&BB->front() != UI)
4115  return Changed;
4116 
4118  for (unsigned i = 0, e = Preds.size(); i != e; ++i) {
4119  TerminatorInst *TI = Preds[i]->getTerminator();
4120  IRBuilder<> Builder(TI);
4121  if (auto *BI = dyn_cast<BranchInst>(TI)) {
4122  if (BI->isUnconditional()) {
4123  if (BI->getSuccessor(0) == BB) {
4124  new UnreachableInst(TI->getContext(), TI);
4125  TI->eraseFromParent();
4126  Changed = true;
4127  }
4128  } else {
4129  if (BI->getSuccessor(0) == BB) {
4130  Builder.CreateBr(BI->getSuccessor(1));
4132  } else if (BI->getSuccessor(1) == BB) {
4133  Builder.CreateBr(BI->getSuccessor(0));
4135  Changed = true;
4136  }
4137  }
4138  } else if (auto *SI = dyn_cast<SwitchInst>(TI)) {
4139  for (auto i = SI->case_begin(), e = SI->case_end(); i != e;) {
4140  if (i->getCaseSuccessor() != BB) {
4141  ++i;
4142  continue;
4143  }
4144  BB->removePredecessor(SI->getParent());
4145  i = SI->removeCase(i);
4146  e = SI->case_end();
4147  Changed = true;
4148  }
4149  } else if (auto *II = dyn_cast<InvokeInst>(TI)) {
4150  if (II->getUnwindDest() == BB) {
4151  removeUnwindEdge(TI->getParent());
4152  Changed = true;
4153  }
4154  } else if (auto *CSI = dyn_cast<CatchSwitchInst>(TI)) {
4155  if (CSI->getUnwindDest() == BB) {
4156  removeUnwindEdge(TI->getParent());
4157  Changed = true;
4158  continue;
4159  }
4160 
4161  for (CatchSwitchInst::handler_iterator I = CSI->handler_begin(),
4162  E = CSI->handler_end();
4163  I != E; ++I) {
4164  if (*I == BB) {
4165  CSI->removeHandler(I);
4166  --I;
4167  --E;
4168  Changed = true;
4169  }
4170  }
4171  if (CSI->getNumHandlers() == 0) {
4172  BasicBlock *CatchSwitchBB = CSI->getParent();
4173  if (CSI->hasUnwindDest()) {
4174  // Redirect preds to the unwind dest
4175  CatchSwitchBB->replaceAllUsesWith(CSI->getUnwindDest());
4176  } else {
4177  // Rewrite all preds to unwind to caller (or from invoke to call).
4178  SmallVector<BasicBlock *, 8> EHPreds(predecessors(CatchSwitchBB));
4179  for (BasicBlock *EHPred : EHPreds)
4180  removeUnwindEdge(EHPred);
4181  }
4182  // The catchswitch is no longer reachable.
4183  new UnreachableInst(CSI->getContext(), CSI);
4184  CSI->eraseFromParent();
4185  Changed = true;
4186  }
4187  } else if (isa<CleanupReturnInst>(TI)) {
4188  new UnreachableInst(TI->getContext(), TI);
4189  TI->eraseFromParent();
4190  Changed = true;
4191  }
4192  }
4193 
4194  // If this block is now dead, remove it.
4195  if (pred_empty(BB) && BB != &BB->getParent()->getEntryBlock()) {
4196  // We know there are no successors, so just nuke the block.
4197  BB->eraseFromParent();
4198  if (LoopHeaders)
4199  LoopHeaders->erase(BB);
4200  return true;
4201  }
4202 
4203  return Changed;
4204 }
4205 
4207  assert(Cases.size() >= 1);
4208 
4209  array_pod_sort(Cases.begin(), Cases.end(), ConstantIntSortPredicate);
4210  for (size_t I = 1, E = Cases.size(); I != E; ++I) {
4211  if (Cases[I - 1]->getValue() != Cases[I]->getValue() + 1)
4212  return false;
4213  }
4214  return true;
4215 }
4216 
4217 /// Turn a switch with two reachable destinations into an integer range
4218 /// comparison and branch.
4220  assert(SI->getNumCases() > 1 && "Degenerate switch?");
4221 
4222  bool HasDefault =
4223  !isa<UnreachableInst>(SI->getDefaultDest()->getFirstNonPHIOrDbg());
4224 
4225  // Partition the cases into two sets with different destinations.
4226  BasicBlock *DestA = HasDefault ? SI->getDefaultDest() : nullptr;
4227  BasicBlock *DestB = nullptr;
4230 
4231  for (auto Case : SI->cases()) {
4232  BasicBlock *Dest = Case.getCaseSuccessor();
4233  if (!DestA)
4234  DestA = Dest;
4235  if (Dest == DestA) {
4236  CasesA.push_back(Case.getCaseValue());
4237  continue;
4238  }
4239  if (!DestB)
4240  DestB = Dest;
4241  if (Dest == DestB) {
4242  CasesB.push_back(Case.getCaseValue());
4243  continue;
4244  }
4245  return false; // More than two destinations.
4246  }
4247 
4248  assert(DestA && DestB &&
4249  "Single-destination switch should have been folded.");
4250  assert(DestA != DestB);
4251  assert(DestB != SI->getDefaultDest());
4252  assert(!CasesB.empty() && "There must be non-default cases.");
4253  assert(!CasesA.empty() || HasDefault);
4254 
4255  // Figure out if one of the sets of cases form a contiguous range.
4256  SmallVectorImpl<ConstantInt *> *ContiguousCases = nullptr;
4257  BasicBlock *ContiguousDest = nullptr;
4258  BasicBlock *OtherDest = nullptr;
4259  if (!CasesA.empty() && CasesAreContiguous(CasesA)) {
4260  ContiguousCases = &CasesA;
4261  ContiguousDest = DestA;
4262  OtherDest = DestB;
4263  } else if (CasesAreContiguous(CasesB)) {
4264  ContiguousCases = &CasesB;
4265  ContiguousDest = DestB;
4266  OtherDest = DestA;
4267  } else
4268  return false;
4269 
4270  // Start building the compare and branch.
4271 
4272  Constant *Offset = ConstantExpr::getNeg(ContiguousCases->back());
4273  Constant *NumCases =
4274  ConstantInt::get(Offset->getType(), ContiguousCases->size());
4275 
4276  Value *Sub = SI->getCondition();
4277  if (!Offset->isNullValue())
4278  Sub = Builder.CreateAdd(Sub, Offset, Sub->getName() + ".off");
4279 
4280  Value *Cmp;
4281  // If NumCases overflowed, then all possible values jump to the successor.
4282  if (NumCases->isNullValue() && !ContiguousCases->empty())
4283  Cmp = ConstantInt::getTrue(SI->getContext());
4284  else
4285  Cmp = Builder.CreateICmpULT(Sub, NumCases, "switch");
4286  BranchInst *NewBI = Builder.CreateCondBr(Cmp, ContiguousDest, OtherDest);
4287 
4288  // Update weight for the newly-created conditional branch.
4289  if (HasBranchWeights(SI)) {
4290  SmallVector<uint64_t, 8> Weights;
4291  GetBranchWeights(SI, Weights);
4292  if (Weights.size() == 1 + SI->getNumCases()) {
4293  uint64_t TrueWeight = 0;
4294  uint64_t FalseWeight = 0;
4295  for (size_t I = 0, E = Weights.size(); I != E; ++I) {
4296  if (SI->getSuccessor(I) == ContiguousDest)
4297  TrueWeight += Weights[I];
4298  else
4299  FalseWeight += Weights[I];
4300  }
4301  while (TrueWeight > UINT32_MAX || FalseWeight > UINT32_MAX) {
4302  TrueWeight /= 2;
4303  FalseWeight /= 2;
4304  }
4306  MDBuilder(SI->getContext())
4307  .createBranchWeights((uint32_t)TrueWeight,
4308  (uint32_t)FalseWeight));
4309  }
4310  }
4311 
4312  // Prune obsolete incoming values off the successors' PHI nodes.
4313  for (auto BBI = ContiguousDest->begin(); isa<PHINode>(BBI); ++BBI) {
4314  unsigned PreviousEdges = ContiguousCases->size();
4315  if (ContiguousDest == SI->getDefaultDest())
4316  ++PreviousEdges;
4317  for (unsigned I = 0, E = PreviousEdges - 1; I != E; ++I)
4318  cast<PHINode>(BBI)->removeIncomingValue(SI->getParent());
4319  }
4320  for (auto BBI = OtherDest->begin(); isa<PHINode>(BBI); ++BBI) {
4321  unsigned PreviousEdges = SI->getNumCases() - ContiguousCases->size();
4322  if (OtherDest == SI->getDefaultDest())
4323  ++PreviousEdges;
4324  for (unsigned I = 0, E = PreviousEdges - 1; I != E; ++I)
4325  cast<PHINode>(BBI)->removeIncomingValue(SI->getParent());
4326  }
4327 
4328  // Drop the switch.
4329  SI->eraseFromParent();
4330 
4331  return true;
4332 }
4333 
4334 /// Compute masked bits for the condition of a switch
4335 /// and use it to remove dead cases.
4337  const DataLayout &DL) {
4338  Value *Cond = SI->getCondition();
4339  unsigned Bits = Cond->getType()->getIntegerBitWidth();
4340  KnownBits Known = computeKnownBits(Cond, DL, 0, AC, SI);
4341 
4342  // We can also eliminate cases by determining that their values are outside of
4343  // the limited range of the condition based on how many significant (non-sign)
4344  // bits are in the condition value.
4345  unsigned ExtraSignBits = ComputeNumSignBits(Cond, DL, 0, AC, SI) - 1;
4346  unsigned MaxSignificantBitsInCond = Bits - ExtraSignBits;
4347 
4348  // Gather dead cases.
4350  for (auto &Case : SI->cases()) {
4351  const APInt &CaseVal = Case.getCaseValue()->getValue();
4352  if (Known.Zero.intersects(CaseVal) || !Known.One.isSubsetOf(CaseVal) ||
4353  (CaseVal.getMinSignedBits() > MaxSignificantBitsInCond)) {
4354  DeadCases.push_back(Case.getCaseValue());
4355  DEBUG(dbgs() << "SimplifyCFG: switch case " << CaseVal << " is dead.\n");
4356  }
4357  }
4358 
4359  // If we can prove that the cases must cover all possible values, the
4360  // default destination becomes dead and we can remove it. If we know some
4361  // of the bits in the value, we can use that to more precisely compute the
4362  // number of possible unique case values.
4363  bool HasDefault =
4364  !isa<UnreachableInst>(SI->getDefaultDest()->getFirstNonPHIOrDbg());
4365  const unsigned NumUnknownBits =
4366  Bits - (Known.Zero | Known.One).countPopulation();
4367  assert(NumUnknownBits <= Bits);
4368  if (HasDefault && DeadCases.empty() &&
4369  NumUnknownBits < 64 /* avoid overflow */ &&
4370  SI->getNumCases() == (1ULL << NumUnknownBits)) {
4371  DEBUG(dbgs() << "SimplifyCFG: switch default is dead.\n");
4372  BasicBlock *NewDefault =
4374  SI->setDefaultDest(&*NewDefault);
4375  SplitBlock(&*NewDefault, &NewDefault->front());
4376  auto *OldTI = NewDefault->getTerminator();
4377  new UnreachableInst(SI->getContext(), OldTI);
4379  return true;
4380  }
4381 
4382  SmallVector<uint64_t, 8> Weights;
4383  bool HasWeight = HasBranchWeights(SI);
4384  if (HasWeight) {
4385  GetBranchWeights(SI, Weights);
4386  HasWeight = (Weights.size() == 1 + SI->getNumCases());
4387  }
4388 
4389  // Remove dead cases from the switch.
4390  for (ConstantInt *DeadCase : DeadCases) {
4391  SwitchInst::CaseIt CaseI = SI->findCaseValue(DeadCase);
4392  assert(CaseI != SI->case_default() &&
4393  "Case was not found. Probably mistake in DeadCases forming.");
4394  if (HasWeight) {
4395  std::swap(Weights[CaseI->getCaseIndex() + 1], Weights.back());
4396  Weights.pop_back();
4397  }
4398 
4399  // Prune unused values from PHI nodes.
4400  CaseI->getCaseSuccessor()->removePredecessor(SI->getParent());
4401  SI->removeCase(CaseI);
4402  }
4403  if (HasWeight && Weights.size() >= 2) {
4404  SmallVector<uint32_t, 8> MDWeights(Weights.begin(), Weights.end());
4406  MDBuilder(SI->getParent()->getContext())
4407  .createBranchWeights(MDWeights));
4408  }
4409 
4410  return !DeadCases.empty();
4411 }
4412 
4413 /// If BB would be eligible for simplification by
4414 /// TryToSimplifyUncondBranchFromEmptyBlock (i.e. it is empty and terminated
4415 /// by an unconditional branch), look at the phi node for BB in the successor
4416 /// block and see if the incoming value is equal to CaseValue. If so, return
4417 /// the phi node, and set PhiIndex to BB's index in the phi node.
4419  BasicBlock *BB, int *PhiIndex) {
4420  if (BB->getFirstNonPHIOrDbg() != BB->getTerminator())
4421  return nullptr; // BB must be empty to be a candidate for simplification.
4422  if (!BB->getSinglePredecessor())
4423  return nullptr; // BB must be dominated by the switch.
4424 
4426  if (!Branch || !Branch->isUnconditional())
4427  return nullptr; // Terminator must be unconditional branch.
4428 
4429  BasicBlock *Succ = Branch->getSuccessor(0);
4430 
4431  BasicBlock::iterator I = Succ->begin();
4432  while (PHINode *PHI = dyn_cast<PHINode>(I++)) {
4433  int Idx = PHI->getBasicBlockIndex(BB);
4434  assert(Idx >= 0 && "PHI has no entry for predecessor?");
4435 
4436  Value *InValue = PHI->getIncomingValue(Idx);
4437  if (InValue != CaseValue)
4438  continue;
4439 
4440  *PhiIndex = Idx;
4441  return PHI;
4442  }
4443 
4444  return nullptr;
4445 }
4446 
4447 /// Try to forward the condition of a switch instruction to a phi node
4448 /// dominated by the switch, if that would mean that some of the destination
4449 /// blocks of the switch can be folded away. Return true if a change is made.
4451  typedef DenseMap<PHINode *, SmallVector<int, 4>> ForwardingNodesMap;
4452  ForwardingNodesMap ForwardingNodes;
4453 
4454  for (auto &Case : SI->cases()) {
4455  ConstantInt *CaseValue = Case.getCaseValue();
4456  BasicBlock *CaseDest = Case.getCaseSuccessor();
4457  int PhiIdx;
4458  if (auto *Phi = FindPHIForConditionForwarding(CaseValue, CaseDest, &PhiIdx))
4459  ForwardingNodes[Phi].push_back(PhiIdx);
4460  }
4461 
4462  bool Changed = false;
4463  for (auto &ForwardingNode : ForwardingNodes) {
4464  PHINode *Phi = ForwardingNode.first;
4465  SmallVectorImpl<int> &Indexes = ForwardingNode.second;
4466  if (Indexes.size() < 2)
4467  continue;
4468 
4469  for (int Index : Indexes)
4470  Phi->setIncomingValue(Index, SI->getCondition());
4471  Changed = true;
4472  }
4473 
4474  return Changed;
4475 }
4476 
4477 /// Return true if the backend will be able to handle
4478 /// initializing an array of constants like C.
4480  if (C->isThreadDependent())
4481  return false;
4482  if (C->isDLLImportDependent())
4483  return false;
4484 
4485  if (!isa<ConstantFP>(C) && !isa<ConstantInt>(C) &&
4486  !isa<ConstantPointerNull>(C) && !isa<GlobalValue>(C) &&
4487  !isa<UndefValue>(C) && !isa<ConstantExpr>(C))
4488  return false;
4489 
4490  if (ConstantExpr *CE = dyn_cast<ConstantExpr>(C)) {
4491  if (!CE->isGEPWithNoNotionalOverIndexing())
4492  return false;
4493  if (!ValidLookupTableConstant(CE->getOperand(0), TTI))
4494  return false;
4495  }
4496 
4498  return false;
4499 
4500  return true;
4501 }
4502 
4503 /// If V is a Constant, return it. Otherwise, try to look up
4504 /// its constant value in ConstantPool, returning 0 if it's not there.
4505 static Constant *
4508  if (Constant *C = dyn_cast<Constant>(V))
4509  return C;
4510  return ConstantPool.lookup(V);
4511 }
4512 
4513 /// Try to fold instruction I into a constant. This works for
4514 /// simple instructions such as binary operations where both operands are
4515 /// constant or can be replaced by constants from the ConstantPool. Returns the
4516 /// resulting constant on success, 0 otherwise.
4517 static Constant *
4520  if (SelectInst *Select = dyn_cast<SelectInst>(I)) {
4521  Constant *A = LookupConstant(Select->getCondition(), ConstantPool);
4522  if (!A)
4523  return nullptr;
4524  if (A->isAllOnesValue())
4525  return LookupConstant(Select->getTrueValue(), ConstantPool);
4526  if (A->isNullValue())
4527  return LookupConstant(Select->getFalseValue(), ConstantPool);
4528  return nullptr;
4529  }
4530 
4532  for (unsigned N = 0, E = I->getNumOperands(); N != E; ++N) {
4534  COps.push_back(A);
4535  else
4536  return nullptr;
4537  }
4538 
4539  if (CmpInst *Cmp = dyn_cast<CmpInst>(I)) {
4540  return ConstantFoldCompareInstOperands(Cmp->getPredicate(), COps[0],
4541  COps[1], DL);
4542  }
4543 
4544  return ConstantFoldInstOperands(I, COps, DL);
4545 }
4546 
4547 /// Try to determine the resulting constant values in phi nodes
4548 /// at the common destination basic block, *CommonDest, for one of the case
4549 /// destionations CaseDest corresponding to value CaseVal (0 for the default
4550 /// case), of a switch instruction SI.
4551 static bool
4553  BasicBlock **CommonDest,
4554  SmallVectorImpl<std::pair<PHINode *, Constant *>> &Res,
4555  const DataLayout &DL, const TargetTransformInfo &TTI) {
4556  // The block from which we enter the common destination.
4557  BasicBlock *Pred = SI->getParent();
4558 
4559  // If CaseDest is empty except for some side-effect free instructions through
4560  // which we can constant-propagate the CaseVal, continue to its successor.
4562  ConstantPool.insert(std::make_pair(SI->getCondition(), CaseVal));
4563  for (BasicBlock::iterator I = CaseDest->begin(), E = CaseDest->end(); I != E;
4564  ++I) {
4565  if (TerminatorInst *T = dyn_cast<TerminatorInst>(I)) {
4566  // If the terminator is a simple branch, continue to the next block.
4567  if (T->getNumSuccessors() != 1 || T->isExceptional())
4568  return false;
4569  Pred = CaseDest;
4570  CaseDest = T->getSuccessor(0);
4571  } else if (isa<DbgInfoIntrinsic>(I)) {
4572  // Skip debug intrinsic.
4573  continue;
4574  } else if (Constant *C = ConstantFold(&*I, DL, ConstantPool)) {
4575  // Instruction is side-effect free and constant.
4576 
4577  // If the instruction has uses outside this block or a phi node slot for
4578  // the block, it is not safe to bypass the instruction since it would then
4579  // no longer dominate all its uses.
4580  for (auto &Use : I->uses()) {
4581  User *User = Use.getUser();
4582  if (Instruction *I = dyn_cast<Instruction>(User))
4583  if (I->getParent() == CaseDest)
4584  continue;
4585  if (PHINode *Phi = dyn_cast<PHINode>(User))
4586  if (Phi->getIncomingBlock(Use) == CaseDest)
4587  continue;
4588  return false;
4589  }
4590 
4591  ConstantPool.insert(std::make_pair(&*I, C));
4592  } else {
4593  break;
4594  }
4595  }
4596 
4597  // If we did not have a CommonDest before, use the current one.
4598  if (!*CommonDest)
4599  *CommonDest = CaseDest;
4600  // If the destination isn't the common one, abort.
4601  if (CaseDest != *CommonDest)
4602  return false;
4603 
4604  // Get the values for this case from phi nodes in the destination block.
4605  BasicBlock::iterator I = (*CommonDest)->begin();
4606  while (PHINode *PHI = dyn_cast<PHINode>(I++)) {
4607  int Idx = PHI->getBasicBlockIndex(Pred);
4608  if (Idx == -1)
4609  continue;
4610 
4611  Constant *ConstVal =
4612  LookupConstant(PHI->getIncomingValue(Idx), ConstantPool);
4613  if (!ConstVal)
4614  return false;
4615 
4616  // Be conservative about which kinds of constants we support.
4617  if (!ValidLookupTableConstant(ConstVal, TTI))
4618  return false;
4619 
4620  Res.push_back(std::make_pair(PHI, ConstVal));
4621  }
4622 
4623  return Res.size() > 0;
4624 }
4625 
4626 // Helper function used to add CaseVal to the list of cases that generate
4627 // Result.
4628 static void MapCaseToResult(ConstantInt *CaseVal,
4629  SwitchCaseResultVectorTy &UniqueResults,
4630  Constant *Result) {
4631  for (auto &I : UniqueResults) {
4632  if (I.first == Result) {
4633  I.second.push_back(CaseVal);
4634  return;
4635  }
4636  }
4637  UniqueResults.push_back(
4638  std::make_pair(Result, SmallVector<ConstantInt *, 4>(1, CaseVal)));
4639 }
4640 
4641 // Helper function that initializes a map containing
4642 // results for the PHI node of the common destination block for a switch
4643 // instruction. Returns false if multiple PHI nodes have been found or if
4644 // there is not a common destination block for the switch.
4646  BasicBlock *&CommonDest,
4647  SwitchCaseResultVectorTy &UniqueResults,
4648  Constant *&DefaultResult,
4649  const DataLayout &DL,
4650  const TargetTransformInfo &TTI) {
4651  for (auto &I : SI->cases()) {
4652  ConstantInt *CaseVal = I.getCaseValue();
4653 
4654  // Resulting value at phi nodes for this case value.
4655  SwitchCaseResultsTy Results;
4656  if (!GetCaseResults(SI, CaseVal, I.getCaseSuccessor(), &CommonDest, Results,
4657  DL, TTI))
4658  return false;
4659 
4660  // Only one value per case is permitted
4661  if (Results.size() > 1)
4662  return false;
4663  MapCaseToResult(CaseVal, UniqueResults, Results.begin()->second);
4664 
4665  // Check the PHI consistency.
4666  if (!PHI)
4667  PHI = Results[0].first;
4668  else if (PHI != Results[0].first)
4669  return false;
4670  }
4671  // Find the default result value.
4673  BasicBlock *DefaultDest = SI->getDefaultDest();
4674  GetCaseResults(SI, nullptr, SI->getDefaultDest(), &CommonDest, DefaultResults,
4675  DL, TTI);
4676  // If the default value is not found abort unless the default destination
4677  // is unreachable.
4678  DefaultResult =
4679  DefaultResults.size() == 1 ? DefaultResults.begin()->second : nullptr;
4680  if ((!DefaultResult &&
4681  !isa<UnreachableInst>(DefaultDest->getFirstNonPHIOrDbg())))
4682  return false;
4683 
4684  return true;
4685 }
4686 
4687 // Helper function that checks if it is possible to transform a switch with only
4688 // two cases (or two cases + default) that produces a result into a select.
4689 // Example:
4690 // switch (a) {
4691 // case 10: %0 = icmp eq i32 %a, 10
4692 // return 10; %1 = select i1 %0, i32 10, i32 4
4693 // case 20: ----> %2 = icmp eq i32 %a, 20
4694 // return 2; %3 = select i1 %2, i32 2, i32 %1
4695 // default:
4696 // return 4;
4697 // }
4698 static Value *ConvertTwoCaseSwitch(const SwitchCaseResultVectorTy &ResultVector,
4699  Constant *DefaultResult, Value *Condition,
4700  IRBuilder<> &Builder) {
4701  assert(ResultVector.size() == 2 &&
4702  "We should have exactly two unique results at this point");
4703  // If we are selecting between only two cases transform into a simple
4704  // select or a two-way select if default is possible.
4705  if (ResultVector[0].second.size() == 1 &&
4706  ResultVector[1].second.size() == 1) {
4707  ConstantInt *const FirstCase = ResultVector[0].second[0];
4708  ConstantInt *const SecondCase = ResultVector[1].second[0];
4709 
4710  bool DefaultCanTrigger = DefaultResult;
4711  Value *SelectValue = ResultVector[1].first;
4712  if (DefaultCanTrigger) {
4713  Value *const ValueCompare =
4714  Builder.CreateICmpEQ(Condition, SecondCase, "switch.selectcmp");
4715  SelectValue = Builder.CreateSelect(ValueCompare, ResultVector[1].first,
4716  DefaultResult, "switch.select");
4717  }
4718  Value *const ValueCompare =
4719  Builder.CreateICmpEQ(Condition, FirstCase, "switch.selectcmp");
4720  return Builder.CreateSelect(ValueCompare, ResultVector[0].first,
4721  SelectValue, "switch.select");
4722  }
4723 
4724  return nullptr;
4725 }
4726 
4727 // Helper function to cleanup a switch instruction that has been converted into
4728 // a select, fixing up PHI nodes and basic blocks.
4730  Value *SelectValue,
4731  IRBuilder<> &Builder) {
4732  BasicBlock *SelectBB = SI->getParent();
4733  while (PHI->getBasicBlockIndex(SelectBB) >= 0)
4734  PHI->removeIncomingValue(SelectBB);
4735  PHI->addIncoming(SelectValue, SelectBB);
4736 
4737  Builder.CreateBr(PHI->getParent());
4738 
4739  // Remove the switch.
4740  for (unsigned i = 0, e = SI->getNumSuccessors(); i < e; ++i) {
4741  BasicBlock *Succ = SI->getSuccessor(i);
4742 
4743  if (Succ == PHI->getParent())
4744  continue;
4745  Succ->removePredecessor(SelectBB);
4746  }
4747  SI->eraseFromParent();
4748 }
4749 
4750 /// If the switch is only used to initialize one or more
4751 /// phi nodes in a common successor block with only two different
4752 /// constant values, replace the switch with select.
4753 static bool switchToSelect(SwitchInst *SI, IRBuilder<> &Builder,
4754  const DataLayout &DL,
4755  const TargetTransformInfo &TTI) {
4756  Value *const Cond = SI->getCondition();
4757  PHINode *PHI = nullptr;
4758  BasicBlock *CommonDest = nullptr;
4759  Constant *DefaultResult;
4760  SwitchCaseResultVectorTy UniqueResults;
4761  // Collect all the cases that will deliver the same value from the switch.
4762  if (!InitializeUniqueCases(SI, PHI, CommonDest, UniqueResults, DefaultResult,
4763  DL, TTI))
4764  return false;
4765  // Selects choose between maximum two values.
4766  if (UniqueResults.size() != 2)
4767  return false;
4768  assert(PHI != nullptr && "PHI for value select not found");
4769 
4770  Builder.SetInsertPoint(SI);
4771  Value *SelectValue =
4772  ConvertTwoCaseSwitch(UniqueResults, DefaultResult, Cond, Builder);
4773  if (SelectValue) {
4774  RemoveSwitchAfterSelectConversion(SI, PHI, SelectValue, Builder);
4775  return true;
4776  }
4777  // The switch couldn't be converted into a select.
4778  return false;
4779 }
4780 
4781 namespace {
4782 
4783 /// This class represents a lookup table that can be used to replace a switch.
4784 class SwitchLookupTable {
4785 public:
4786  /// Create a lookup table to use as a switch replacement with the contents
4787  /// of Values, using DefaultValue to fill any holes in the table.
4788  SwitchLookupTable(
4789  Module &M, uint64_t TableSize, ConstantInt *Offset,
4790  const SmallVectorImpl<std::pair<ConstantInt *, Constant *>> &Values,
4791  Constant *DefaultValue, const DataLayout &DL, const StringRef &FuncName);
4792 
4793  /// Build instructions with Builder to retrieve the value at
4794  /// the position given by Index in the lookup table.
4795  Value *BuildLookup(Value *Index, IRBuilder<> &Builder);
4796 
4797  /// Return true if a table with TableSize elements of
4798  /// type ElementType would fit in a target-legal register.
4799  static bool WouldFitInRegister(const DataLayout &DL, uint64_t TableSize,
4800  Type *ElementType);
4801 
4802 private:
4803  // Depending on the contents of the table, it can be represented in
4804  // different ways.
4805  enum {
4806  // For tables where each element contains the same value, we just have to
4807  // store that single value and return it for each lookup.
4808  SingleValueKind,
4809 
4810  // For tables where there is a linear relationship between table index
4811  // and values. We calculate the result with a simple multiplication
4812  // and addition instead of a table lookup.
4813  LinearMapKind,
4814 
4815  // For small tables with integer elements, we can pack them into a bitmap
4816  // that fits into a target-legal register. Values are retrieved by
4817  // shift and mask operations.
4818  BitMapKind,
4819 
4820  // The table is stored as an array of values. Values are retrieved by load
4821  // instructions from the table.
4822  ArrayKind
4823  } Kind;
4824 
4825  // For SingleValueKind, this is the single value.
4826  Constant *SingleValue;
4827 
4828  // For BitMapKind, this is the bitmap.
4829  ConstantInt *BitMap;
4830  IntegerType *BitMapElementTy;
4831 
4832  // For LinearMapKind, these are the constants used to derive the value.
4833  ConstantInt *LinearOffset;
4834  ConstantInt *LinearMultiplier;
4835 
4836  // For ArrayKind, this is the array.
4837  GlobalVariable *Array;
4838 };
4839 
4840 } // end anonymous namespace
4841 
4842 SwitchLookupTable::SwitchLookupTable(
4843  Module &M, uint64_t TableSize, ConstantInt *Offset,
4844  const SmallVectorImpl<std::pair<ConstantInt *, Constant *>> &Values,
4845  Constant *DefaultValue, const DataLayout &DL, const StringRef &FuncName)
4846  : SingleValue(nullptr), BitMap(nullptr), BitMapElementTy(nullptr),
4847  LinearOffset(nullptr), LinearMultiplier(nullptr), Array(nullptr) {
4848  assert(Values.size() && "Can't build lookup table without values!");
4849  assert(TableSize >= Values.size() && "Can't fit values in table!");
4850 
4851  // If all values in the table are equal, this is that value.
4852  SingleValue = Values.begin()->second;
4853 
4854  Type *ValueType = Values.begin()->second->getType();
4855 
4856  // Build up the table contents.
4857  SmallVector<Constant *, 64> TableContents(TableSize);
4858  for (size_t I = 0, E = Values.size(); I != E; ++I) {
4859  ConstantInt *CaseVal = Values[I].first;
4860  Constant *CaseRes = Values[I].second;
4861  assert(CaseRes->getType() == ValueType);
4862 
4863  uint64_t Idx = (CaseVal->getValue() - Offset->getValue()).getLimitedValue();
4864  TableContents[Idx] = CaseRes;
4865 
4866  if (CaseRes != SingleValue)
4867  SingleValue = nullptr;
4868  }
4869 
4870  // Fill in any holes in the table with the default result.
4871  if (Values.size() < TableSize) {
4872  assert(DefaultValue &&
4873  "Need a default value to fill the lookup table holes.");
4874  assert(DefaultValue->getType() == ValueType);
4875  for (uint64_t I = 0; I < TableSize; ++I) {
4876  if (!TableContents[I])
4877  TableContents[I] = DefaultValue;
4878  }
4879 
4880  if (DefaultValue != SingleValue)
4881  SingleValue = nullptr;
4882  }
4883 
4884  // If each element in the table contains the same value, we only need to store
4885  // that single value.
4886  if (SingleValue) {
4887  Kind = SingleValueKind;
4888  return;
4889  }
4890 
4891  // Check if we can derive the value with a linear transformation from the
4892  // table index.
4893  if (isa<IntegerType>(ValueType)) {
4894  bool LinearMappingPossible = true;
4895  APInt PrevVal;
4896  APInt DistToPrev;
4897  assert(TableSize >= 2 && "Should be a SingleValue table.");
4898  // Check if there is the same distance between two consecutive values.
4899  for (uint64_t I = 0; I < TableSize; ++I) {
4900  ConstantInt *ConstVal = dyn_cast<ConstantInt>(TableContents[I]);
4901  if (!ConstVal) {
4902  // This is an undef. We could deal with it, but undefs in lookup tables
4903  // are very seldom. It's probably not worth the additional complexity.
4904  LinearMappingPossible = false;
4905  break;
4906  }
4907  const APInt &Val = ConstVal->getValue();
4908  if (I != 0) {
4909  APInt Dist = Val - PrevVal;
4910  if (I == 1) {
4911  DistToPrev = Dist;
4912  } else if (Dist != DistToPrev) {
4913  LinearMappingPossible = false;
4914  break;
4915  }
4916  }
4917  PrevVal = Val;
4918  }
4919  if (LinearMappingPossible) {
4920  LinearOffset = cast<ConstantInt>(TableContents[0]);
4921  LinearMultiplier = ConstantInt::get(M.getContext(), DistToPrev);
4922  Kind = LinearMapKind;
4923  ++NumLinearMaps;
4924  return;
4925  }
4926  }
4927 
4928  // If the type is integer and the table fits in a register, build a bitmap.
4929  if (WouldFitInRegister(DL, TableSize, ValueType)) {
4930  IntegerType *IT = cast<IntegerType>(ValueType);
4931  APInt TableInt(TableSize * IT->getBitWidth(), 0);
4932  for (uint64_t I = TableSize; I > 0; --I) {
4933  TableInt <<= IT->getBitWidth();
4934  // Insert values into the bitmap. Undef values are set to zero.
4935  if (!isa<UndefValue>(TableContents[I - 1])) {
4936  ConstantInt *Val = cast<ConstantInt>(TableContents[I - 1]);
4937  TableInt |= Val->getValue().zext(TableInt.getBitWidth());
4938  }
4939  }
4940  BitMap = ConstantInt::get(M.getContext(), TableInt);
4941  BitMapElementTy = IT;
4942  Kind = BitMapKind;
4943  ++NumBitMaps;
4944  return;
4945  }
4946 
4947  // Store the table in an array.
4948  ArrayType *ArrayTy = ArrayType::get(ValueType, TableSize);
4949  Constant *Initializer = ConstantArray::get(ArrayTy, TableContents);
4950 
4951  Array = new GlobalVariable(M, ArrayTy, /*constant=*/true,
4952  GlobalVariable::PrivateLinkage, Initializer,
4953  "switch.table." + FuncName);
4954  Array->setUnnamedAddr(GlobalValue::UnnamedAddr::Global);
4955  Kind = ArrayKind;
4956 }
4957 
4958 Value *SwitchLookupTable::BuildLookup(Value *Index, IRBuilder<> &Builder) {
4959  switch (Kind) {
4960  case SingleValueKind:
4961  return SingleValue;
4962  case LinearMapKind: {
4963  // Derive the result value from the input value.
4964  Value *Result = Builder.CreateIntCast(Index, LinearMultiplier->getType(),
4965  false, "switch.idx.cast");
4966  if (!LinearMultiplier->isOne())
4967  Result = Builder.CreateMul(Result, LinearMultiplier, "switch.idx.mult");
4968  if (!LinearOffset->isZero())
4969  Result = Builder.CreateAdd(Result, LinearOffset, "switch.offset");
4970  return Result;
4971  }
4972  case BitMapKind: {
4973  // Type of the bitmap (e.g. i59).
4974  IntegerType *MapTy = BitMap->getType();
4975 
4976  // Cast Index to the same type as the bitmap.
4977  // Note: The Index is <= the number of elements in the table, so
4978  // truncating it to the width of the bitmask is safe.
4979  Value *ShiftAmt = Builder.CreateZExtOrTrunc(Index, MapTy, "switch.cast");
4980 
4981  // Multiply the shift amount by the element width.
4982  ShiftAmt = Builder.CreateMul(
4983  ShiftAmt, ConstantInt::get(MapTy, BitMapElementTy->getBitWidth()),
4984  "switch.shiftamt");
4985 
4986  // Shift down.
4987  Value *DownShifted =
4988  Builder.CreateLShr(BitMap, ShiftAmt, "switch.downshift");
4989  // Mask off.
4990  return Builder.CreateTrunc(DownShifted, BitMapElementTy, "switch.masked");
4991  }
4992  case ArrayKind: {
4993  // Make sure the table index will not overflow when treated as signed.
4994  IntegerType *IT = cast<IntegerType>(Index->getType());
4995  uint64_t TableSize =
4996  Array->getInitializer()->getType()->getArrayNumElements();
4997  if (TableSize > (1ULL << (IT->getBitWidth() - 1)))
4998  Index = Builder.CreateZExt(
4999  Index, IntegerType::get(IT->getContext(), IT->getBitWidth() + 1),
5000  "switch.tableidx.zext");
5001 
5002  Value *GEPIndices[] = {Builder.getInt32(0), Index};
5003  Value *GEP = Builder.CreateInBoundsGEP(Array->getValueType(), Array,
5004  GEPIndices, "switch.gep");
5005  return Builder.CreateLoad(GEP, "switch.load");
5006  }
5007  }
5008  llvm_unreachable("Unknown lookup table kind!");
5009 }
5010 
5011 bool SwitchLookupTable::WouldFitInRegister(const DataLayout &DL,
5012  uint64_t TableSize,
5013  Type *ElementType) {
5014  auto *IT = dyn_cast<IntegerType>(ElementType);
5015  if (!IT)
5016  return false;
5017  // FIXME: If the type is wider than it needs to be, e.g. i8 but all values
5018  // are <= 15, we could try to narrow the type.
5019 
5020  // Avoid overflow, fitsInLegalInteger uses unsigned int for the width.
5021  if (TableSize >= UINT_MAX / IT->getBitWidth())
5022  return false;
5023  return DL.fitsInLegalInteger(TableSize * IT->getBitWidth());
5024 }
5025 
5026 /// Determine whether a lookup table should be built for this switch, based on
5027 /// the number of cases, size of the table, and the types of the results.
5028 static bool
5029 ShouldBuildLookupTable(SwitchInst *SI, uint64_t TableSize,
5030  const TargetTransformInfo &TTI, const DataLayout &DL,
5031  const SmallDenseMap<PHINode *, Type *> &ResultTypes) {
5032  if (SI->getNumCases() > TableSize || TableSize >= UINT64_MAX / 10)
5033  return false; // TableSize overflowed, or mul below might overflow.
5034 
5035  bool AllTablesFitInRegister = true;
5036  bool HasIllegalType = false;
5037  for (const auto &I : ResultTypes) {
5038  Type *Ty = I.second;
5039 
5040  // Saturate this flag to true.
5041  HasIllegalType = HasIllegalType || !TTI.isTypeLegal(Ty);
5042 
5043  // Saturate this flag to false.
5044  AllTablesFitInRegister =
5045  AllTablesFitInRegister &&
5046  SwitchLookupTable::WouldFitInRegister(DL, TableSize, Ty);
5047 
5048  // If both flags saturate, we're done. NOTE: This *only* works with
5049  // saturating flags, and all flags have to saturate first due to the
5050  // non-deterministic behavior of iterating over a dense map.
5051  if (HasIllegalType && !AllTablesFitInRegister)
5052  break;
5053  }
5054 
5055  // If each table would fit in a register, we should build it anyway.
5056  if (AllTablesFitInRegister)
5057  return true;
5058 
5059  // Don't build a table that doesn't fit in-register if it has illegal types.
5060  if (HasIllegalType)
5061  return false;
5062 
5063  // The table density should be at least 40%. This is the same criterion as for
5064  // jump tables, see SelectionDAGBuilder::handleJTSwitchCase.
5065  // FIXME: Find the best cut-off.
5066  return SI->getNumCases() * 10 >= TableSize * 4;
5067 }
5068 
5069 /// Try to reuse the switch table index compare. Following pattern:
5070 /// \code
5071 /// if (idx < tablesize)
5072 /// r = table[idx]; // table does not contain default_value
5073 /// else
5074 /// r = default_value;
5075 /// if (r != default_value)
5076 /// ...
5077 /// \endcode
5078 /// Is optimized to:
5079 /// \code
5080 /// cond = idx < tablesize;
5081 /// if (cond)
5082 /// r = table[idx];
5083 /// else
5084 /// r = default_value;
5085 /// if (cond)
5086 /// ...
5087 /// \endcode
5088 /// Jump threading will then eliminate the second if(cond).
5089 static void reuseTableCompare(
5090  User *PhiUser, BasicBlock *PhiBlock, BranchInst *RangeCheckBranch,
5091  Constant *DefaultValue,
5092  const SmallVectorImpl<std::pair<ConstantInt *, Constant *>> &Values) {
5093 
5094  ICmpInst *CmpInst = dyn_cast<ICmpInst>(PhiUser);
5095  if (!CmpInst)
5096  return;
5097 
5098  // We require that the compare is in the same block as the phi so that jump
5099  // threading can do its work afterwards.
5100  if (CmpInst->getParent() != PhiBlock)
5101  return;
5102 
5103  Constant *CmpOp1 = dyn_cast<Constant>(CmpInst->getOperand(1));
5104  if (!CmpOp1)
5105  return;
5106 
5107  Value *RangeCmp = RangeCheckBranch->getCondition();
5108  Constant *TrueConst = ConstantInt::getTrue(RangeCmp->getType());
5109  Constant *FalseConst = ConstantInt::getFalse(RangeCmp->getType());
5110 
5111  // Check if the compare with the default value is constant true or false.
5112  Constant *DefaultConst = ConstantExpr::getICmp(CmpInst->getPredicate(),
5113  DefaultValue, CmpOp1, true);
5114  if (DefaultConst != TrueConst && DefaultConst != FalseConst)
5115  return;
5116 
5117  // Check if the compare with the case values is distinct from the default
5118  // compare result.
5119  for (auto ValuePair : Values) {
5120  Constant *CaseConst = ConstantExpr::getICmp(CmpInst->getPredicate(),
5121  ValuePair.second, CmpOp1, true);
5122  if (!CaseConst || CaseConst == DefaultConst)
5123  return;
5124  assert((CaseConst == TrueConst || CaseConst == FalseConst) &&
5125  "Expect true or false as compare result.");
5126  }
5127 
5128  // Check if the branch instruction dominates the phi node. It's a simple
5129  // dominance check, but sufficient for our needs.
5130  // Although this check is invariant in the calling loops, it's better to do it
5131  // at this late stage. Practically we do it at most once for a switch.
5132  BasicBlock *BranchBlock = RangeCheckBranch->getParent();
5133  for (auto PI = pred_begin(PhiBlock), E = pred_end(PhiBlock); PI != E; ++PI) {
5134  BasicBlock *Pred = *PI;
5135  if (Pred != BranchBlock && Pred->getUniquePredecessor() != BranchBlock)
5136  return;
5137  }
5138 
5139  if (DefaultConst == FalseConst) {
5140  // The compare yields the same result. We can replace it.
5141  CmpInst->replaceAllUsesWith(RangeCmp);
5142  ++NumTableCmpReuses;
5143  } else {
5144  // The compare yields the same result, just inverted. We can replace it.
5145  Value *InvertedTableCmp = BinaryOperator::CreateXor(
5146  RangeCmp, ConstantInt::get(RangeCmp->getType(), 1), "inverted.cmp",
5147  RangeCheckBranch);
5148  CmpInst->replaceAllUsesWith(InvertedTableCmp);
5149  ++NumTableCmpReuses;
5150  }
5151 }
5152 
5153 /// If the switch is only used to initialize one or more phi nodes in a common
5154 /// successor block with different constant values, replace the switch with
5155 /// lookup tables.
5157  const DataLayout &DL,
5158  const TargetTransformInfo &TTI) {
5159  assert(SI->getNumCases() > 1 && "Degenerate switch?");
5160 
5161  Function *Fn = SI->getParent()->getParent();
5162  // Only build lookup table when we have a target that supports it or the
5163  // attribute is not set.
5164  if (!TTI.shouldBuildLookupTables() ||
5165  (Fn->getFnAttribute("no-jump-tables").getValueAsString() == "true"))
5166  return false;
5167 
5168  // FIXME: If the switch is too sparse for a lookup table, perhaps we could
5169  // split off a dense part and build a lookup table for that.
5170 
5171  // FIXME: This creates arrays of GEPs to constant strings, which means each
5172  // GEP needs a runtime relocation in PIC code. We should just build one big
5173  // string and lookup indices into that.
5174 
5175  // Ignore switches with less than three cases. Lookup tables will not make
5176  // them faster, so we don't analyze them.
5177  if (SI->getNumCases() < 3)
5178  return false;
5179 
5180  // Figure out the corresponding result for each case value and phi node in the
5181  // common destination, as well as the min and max case values.
5182  assert(SI->case_begin() != SI->case_end());
5183  SwitchInst::CaseIt CI = SI->case_begin();
5184  ConstantInt *MinCaseVal = CI->getCaseValue();
5185  ConstantInt *MaxCaseVal = CI->getCaseValue();
5186 
5187  BasicBlock *CommonDest = nullptr;
5188  typedef SmallVector<std::pair<ConstantInt *, Constant *>, 4> ResultListTy;
5190  SmallDenseMap<PHINode *, Constant *> DefaultResults;
5193 
5194  for (SwitchInst::CaseIt E = SI->case_end(); CI != E; ++CI) {
5195  ConstantInt *CaseVal = CI->getCaseValue();
5196  if (CaseVal->getValue().slt(MinCaseVal->getValue()))
5197  MinCaseVal = CaseVal;
5198  if (CaseVal->getValue().sgt(MaxCaseVal->getValue()))
5199  MaxCaseVal = CaseVal;
5200 
5201  // Resulting value at phi nodes for this case value.
5202  typedef SmallVector<std::pair<PHINode *, Constant *>, 4> ResultsTy;
5203  ResultsTy Results;
5204  if (!GetCaseResults(SI, CaseVal, CI->getCaseSuccessor(), &CommonDest,
5205  Results, DL, TTI))
5206  return false;
5207 
5208  // Append the result from this case to the list for each phi.
5209  for (const auto &I : Results) {
5210  PHINode *PHI = I.first;
5211  Constant *Value = I.second;
5212  if (!ResultLists.count(PHI))
5213  PHIs.push_back(PHI);
5214  ResultLists[PHI].push_back(std::make_pair(CaseVal, Value));
5215  }
5216  }
5217 
5218  // Keep track of the result types.
5219  for (PHINode *PHI : PHIs) {
5220  ResultTypes[PHI] = ResultLists[PHI][0].second->getType();
5221  }
5222 
5223  uint64_t NumResults = ResultLists[PHIs[0]].size();
5224  APInt RangeSpread = MaxCaseVal->getValue() - MinCaseVal->getValue();
5225  uint64_t TableSize = RangeSpread.getLimitedValue() + 1;
5226  bool TableHasHoles = (NumResults < TableSize);
5227 
5228  // If the table has holes, we need a constant result for the default case
5229  // or a bitmask that fits in a register.
5230  SmallVector<std::pair<PHINode *, Constant *>, 4> DefaultResultsList;
5231  bool HasDefaultResults =
5232  GetCaseResults(SI, nullptr, SI->getDefaultDest(), &CommonDest,
5233  DefaultResultsList, DL, TTI);
5234 
5235  bool NeedMask = (TableHasHoles && !HasDefaultResults);
5236  if (NeedMask) {
5237  // As an extra penalty for the validity test we require more cases.
5238  if (SI->getNumCases() < 4) // FIXME: Find best threshold value (benchmark).
5239  return false;
5240  if (!DL.fitsInLegalInteger(TableSize))
5241  return false;
5242  }
5243 
5244  for (const auto &I : DefaultResultsList) {
5245  PHINode *PHI = I.first;
5246  Constant *Result = I.second;
5247  DefaultResults[PHI] = Result;
5248  }
5249 
5250  if (!ShouldBuildLookupTable(SI, TableSize, TTI, DL, ResultTypes))
5251  return false;
5252 
5253  // Create the BB that does the lookups.
5254  Module &Mod = *CommonDest->getParent()->getParent();
5255  BasicBlock *LookupBB = BasicBlock::Create(
5256  Mod.getContext(), "switch.lookup", CommonDest->getParent(), CommonDest);
5257 
5258  // Compute the table index value.
5259  Builder.SetInsertPoint(SI);
5260  Value *TableIndex;
5261  if (MinCaseVal->isNullValue())
5262  TableIndex = SI->getCondition();
5263  else
5264  TableIndex = Builder.CreateSub(SI->getCondition(), MinCaseVal,
5265  "switch.tableidx");
5266 
5267  // Compute the maximum table size representable by the integer type we are
5268  // switching upon.
5269  unsigned CaseSize = MinCaseVal->getType()->getPrimitiveSizeInBits();
5270  uint64_t MaxTableSize = CaseSize > 63 ? UINT64_MAX : 1ULL << CaseSize;
5271  assert(MaxTableSize >= TableSize &&
5272  "It is impossible for a switch to have more entries than the max "
5273  "representable value of its input integer type's size.");
5274 
5275  // If the default destination is unreachable, or if the lookup table covers
5276  // all values of the conditional variable, branch directly to the lookup table
5277  // BB. Otherwise, check that the condition is within the case range.
5278  const bool DefaultIsReachable =
5279  !isa<UnreachableInst>(SI->getDefaultDest()->getFirstNonPHIOrDbg());
5280  const bool GeneratingCoveredLookupTable = (MaxTableSize == TableSize);
5281  BranchInst *RangeCheckBranch = nullptr;
5282 
5283  if (!DefaultIsReachable || GeneratingCoveredLookupTable) {
5284  Builder.CreateBr(LookupBB);
5285  // Note: We call removeProdecessor later since we need to be able to get the
5286  // PHI value for the default case in case we're using a bit mask.
5287  } else {
5288  Value *Cmp = Builder.CreateICmpULT(
5289  TableIndex, ConstantInt::get(MinCaseVal->getType(), TableSize));
5290  RangeCheckBranch =
5291  Builder.CreateCondBr(Cmp, LookupBB, SI->getDefaultDest());
5292  }
5293 
5294  // Populate the BB that does the lookups.
5295  Builder.SetInsertPoint(LookupBB);
5296 
5297  if (NeedMask) {
5298  // Before doing the lookup, we do the hole check. The LookupBB is therefore
5299  // re-purposed to do the hole check, and we create a new LookupBB.
5300  BasicBlock *MaskBB = LookupBB;
5301  MaskBB->setName("switch.hole_check");
5302  LookupBB = BasicBlock::Create(Mod.getContext(), "switch.lookup",
5303  CommonDest->getParent(), CommonDest);
5304 
5305  // Make the mask's bitwidth at least 8-bit and a power-of-2 to avoid
5306  // unnecessary illegal types.
5307  uint64_t TableSizePowOf2 = NextPowerOf2(std::max(7ULL, TableSize - 1ULL));
5308  APInt MaskInt(TableSizePowOf2, 0);
5309  APInt One(TableSizePowOf2, 1);
5310  // Build bitmask; fill in a 1 bit for every case.
5311  const ResultListTy &ResultList = ResultLists[PHIs[0]];
5312  for (size_t I = 0, E = ResultList.size(); I != E; ++I) {
5313  uint64_t Idx = (ResultList[I].first->getValue() - MinCaseVal->getValue())
5314  .getLimitedValue();
5315  MaskInt |= One << Idx;
5316  }
5317  ConstantInt *TableMask = ConstantInt::get(Mod.getContext(), MaskInt);
5318 
5319  // Get the TableIndex'th bit of the bitmask.
5320  // If this bit is 0 (meaning hole) jump to the default destination,
5321  // else continue with table lookup.
5322  IntegerType *MapTy = TableMask->getType();
5323  Value *MaskIndex =
5324  Builder.CreateZExtOrTrunc(TableIndex, MapTy, "switch.maskindex");
5325  Value *Shifted = Builder.CreateLShr(TableMask, MaskIndex, "switch.shifted");
5326  Value *LoBit = Builder.CreateTrunc(
5327  Shifted, Type::getInt1Ty(Mod.getContext()), "switch.lobit");
5328  Builder.CreateCondBr(LoBit, LookupBB, SI->getDefaultDest());
5329 
5330  Builder.SetInsertPoint(LookupBB);
5331  AddPredecessorToBlock(SI->getDefaultDest(), MaskBB, SI->getParent());
5332  }
5333 
5334  if (!DefaultIsReachable || GeneratingCoveredLookupTable) {
5335  // We cached PHINodes in PHIs. To avoid accessing deleted PHINodes later,
5336  // do not delete PHINodes here.
5338  /*DontDeleteUselessPHIs=*/true);
5339  }
5340 
5341  bool ReturnedEarly = false;
5342  for (size_t I = 0, E = PHIs.size(); I != E; ++I) {
5343  PHINode *PHI = PHIs[I];
5344  const ResultListTy &ResultList = ResultLists[PHI];
5345 
5346  // If using a bitmask, use any value to fill the lookup table holes.
5347  Constant *DV = NeedMask ? ResultLists[PHI][0].second : DefaultResults[PHI];
5348  StringRef FuncName = Fn->getName();
5349  SwitchLookupTable Table(Mod, TableSize, MinCaseVal, ResultList, DV, DL,
5350  FuncName);
5351 
5352  Value *Result = Table.BuildLookup(TableIndex, Builder);
5353 
5354  // If the result is used to return immediately from the function, we want to
5355  // do that right here.
5356  if (PHI->hasOneUse() && isa<ReturnInst>(*PHI->user_begin()) &&
5357  PHI->user_back() == CommonDest->getFirstNonPHIOrDbg()) {
5358  Builder.CreateRet(Result);
5359  ReturnedEarly = true;
5360  break;
5361  }
5362 
5363  // Do a small peephole optimization: re-use the switch table compare if
5364  // possible.
5365  if (!TableHasHoles && HasDefaultResults && RangeCheckBranch) {
5366  BasicBlock *PhiBlock = PHI->getParent();
5367  // Search for compare instructions which use the phi.
5368  for (auto *User : PHI->users()) {
5369  reuseTableCompare(User, PhiBlock, RangeCheckBranch, DV, ResultList);
5370  }
5371  }
5372 
5373  PHI->addIncoming(Result, LookupBB);
5374  }
5375 
5376  if (!ReturnedEarly)
5377  Builder.CreateBr(CommonDest);
5378 
5379  // Remove the switch.
5380  for (unsigned i = 0, e = SI->getNumSuccessors(); i < e; ++i) {
5381  BasicBlock *Succ = SI->getSuccessor(i);
5382 
5383  if (Succ == SI->getDefaultDest())
5384  continue;
5385  Succ->removePredecessor(SI->getParent());
5386  }
5387  SI->eraseFromParent();
5388 
5389  ++NumLookupTables;
5390  if (NeedMask)
5391  ++NumLookupTablesHoles;
5392  return true;
5393 }
5394 
5395 static bool isSwitchDense(ArrayRef<int64_t> Values) {
5396  // See also SelectionDAGBuilder::isDense(), which this function was based on.
5397  uint64_t Diff = (uint64_t)Values.back() - (uint64_t)Values.front();
5398  uint64_t Range = Diff + 1;
5399  uint64_t NumCases = Values.size();
5400  // 40% is the default density for building a jump table in optsize/minsize mode.
5401  uint64_t MinDensity = 40;
5402 
5403  return NumCases * 100 >= Range * MinDensity;
5404 }
5405 
5406 /// Try to transform a switch that has "holes" in it to a contiguous sequence
5407 /// of cases.
5408 ///
5409 /// A switch such as: switch(i) {case 5: case 9: case 13: case 17:} can be
5410 /// range-reduced to: switch ((i-5) / 4) {case 0: case 1: case 2: case 3:}.
5411 ///
5412 /// This converts a sparse switch into a dense switch which allows better
5413 /// lowering and could also allow transforming into a lookup table.
5415  const DataLayout &DL,
5416  const TargetTransformInfo &TTI) {
5417  auto *CondTy = cast<IntegerType>(SI->getCondition()->getType());
5418  if (CondTy->getIntegerBitWidth() > 64 ||
5419  !DL.fitsInLegalInteger(CondTy->getIntegerBitWidth()))
5420  return false;
5421  // Only bother with this optimization if there are more than 3 switch cases;
5422  // SDAG will only bother creating jump tables for 4 or more cases.
5423  if (SI->getNumCases() < 4)
5424  return false;
5425 
5426  // This transform is agnostic to the signedness of the input or case values. We
5427  // can treat the case values as signed or unsigned. We can optimize more common
5428  // cases such as a sequence crossing zero {-4,0,4,8} if we interpret case values
5429  // as signed.
5430  SmallVector<int64_t,4> Values;
5431  for (auto &C : SI->cases())
5432  Values.push_back(C.getCaseValue()->getValue().getSExtValue());
5433  std::sort(Values.begin(), Values.end());
5434 
5435  // If the switch is already dense, there's nothing useful to do here.
5436  if (isSwitchDense(Values))
5437  return false;
5438 
5439  // First, transform the values such that they start at zero and ascend.
5440  int64_t Base = Values[0];
5441  for (auto &V : Values)
5442  V -= (uint64_t)(Base);
5443 
5444  // Now we have signed numbers that have been shifted so that, given enough
5445  // precision, there are no negative values. Since the rest of the transform
5446  // is bitwise only, we switch now to an unsigned representation.
5447  uint64_t GCD = 0;
5448  for (auto &V : Values)
5449  GCD = GreatestCommonDivisor64(GCD, (uint64_t)V);
5450 
5451  // This transform can be done speculatively because it is so cheap - it results
5452  // in a single rotate operation being inserted. This can only happen if the
5453  // factor extracted is a power of 2.
5454  // FIXME: If the GCD is an odd number we can multiply by the multiplicative
5455  // inverse of GCD and then perform this transform.
5456  // FIXME: It's possible that optimizing a switch on powers of two might also
5457  // be beneficial - flag values are often powers of two and we could use a CLZ
5458  // as the key function.
5459  if (GCD <= 1 || !isPowerOf2_64(GCD))
5460  // No common divisor found or too expensive to compute key function.
5461  return false;
5462 
5463  unsigned Shift = Log2_64(GCD);
5464  for (auto &V : Values)
5465  V = (int64_t)((uint64_t)V >> Shift);
5466 
5467  if (!isSwitchDense(Values))
5468  // Transform didn't create a dense switch.
5469  return false;
5470 
5471  // The obvious transform is to shift the switch condition right and emit a
5472  // check that the condition actually cleanly divided by GCD, i.e.
5473  // C & (1 << Shift - 1) == 0
5474  // inserting a new CFG edge to handle the case where it didn't divide cleanly.
5475  //
5476  // A cheaper way of doing this is a simple ROTR(C, Shift). This performs the
5477  // shift and puts the shifted-off bits in the uppermost bits. If any of these
5478  // are nonzero then the switch condition will be very large and will hit the
5479  // default case.
5480 
5481  auto *Ty = cast<IntegerType>(SI->getCondition()->getType());
5482  Builder.SetInsertPoint(SI);
5483  auto *ShiftC = ConstantInt::get(Ty, Shift);
5484  auto *Sub = Builder.CreateSub(SI->getCondition(), ConstantInt::get(Ty, Base));
5485  auto *LShr = Builder.CreateLShr(Sub, ShiftC);
5486  auto *Shl = Builder.CreateShl(Sub, Ty->getBitWidth() - Shift);
5487  auto *Rot = Builder.CreateOr(LShr, Shl);
5488  SI->replaceUsesOfWith(SI->getCondition(), Rot);
5489 
5490  for (auto Case : SI->cases()) {
5491  auto *Orig = Case.getCaseValue();
5492  auto Sub = Orig->getValue() - APInt(Ty->getBitWidth(), Base);
5493  Case.setValue(
5494  cast<ConstantInt>(ConstantInt::get(Ty, Sub.lshr(ShiftC->getValue()))));
5495  }
5496  return true;
5497 }
5498 
5499 bool SimplifyCFGOpt::SimplifySwitch(SwitchInst *SI, IRBuilder<> &Builder) {
5500  BasicBlock *BB = SI->getParent();
5501 
5502  if (isValueEqualityComparison(SI)) {
5503  // If we only have one predecessor, and if it is a branch on this value,
5504  // see if that predecessor totally determines the outcome of this switch.
5505  if (BasicBlock *OnlyPred = BB->getSinglePredecessor())
5506  if (SimplifyEqualityComparisonWithOnlyPredecessor(SI, OnlyPred, Builder))
5507  return simplifyCFG(BB, TTI, Options) | true;
5508 
5509  Value *Cond = SI->getCondition();
5510  if (SelectInst *Select = dyn_cast<SelectInst>(Cond))
5511  if (SimplifySwitchOnSelect(SI, Select))
5512  return simplifyCFG(BB, TTI, Options) | true;
5513 
5514  // If the block only contains the switch, see if we can fold the block
5515  // away into any preds.
5516  BasicBlock::iterator BBI = BB->begin();
5517  // Ignore dbg intrinsics.
5518  while (isa<DbgInfoIntrinsic>(BBI))
5519  ++BBI;
5520  if (SI == &*BBI)
5521  if (FoldValueComparisonIntoPredecessors(SI, Builder))
5522  return simplifyCFG(BB, TTI, Options) | true;
5523  }
5524 
5525  // Try to transform the switch into an icmp and a branch.
5526  if (TurnSwitchRangeIntoICmp(SI, Builder))
5527  return simplifyCFG(BB, TTI, Options) | true;
5528 
5529  // Remove unreachable cases.
5530  if (eliminateDeadSwitchCases(SI, Options.AC, DL))
5531  return simplifyCFG(BB, TTI, Options) | true;
5532 
5533  if (switchToSelect(SI, Builder, DL, TTI))
5534  return simplifyCFG(BB, TTI, Options) | true;
5535 
5537  return simplifyCFG(BB, TTI, Options) | true;
5538 
5539  // The conversion from switch to lookup tables results in difficult-to-analyze
5540  // code and makes pruning branches much harder. This is a problem if the
5541  // switch expression itself can still be restricted as a result of inlining or
5542  // CVP. Therefore, only apply this transformation during late stages of the
5543  // optimisation pipeline.
5544  if (Options.ConvertSwitchToLookupTable &&
5545  SwitchToLookupTable(SI, Builder, DL, TTI))
5546  return simplifyCFG(BB, TTI, Options) | true;
5547 
5548  if (ReduceSwitchRange(SI, Builder, DL, TTI))
5549  return simplifyCFG(BB, TTI, Options) | true;
5550 
5551  re