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  unsigned BonusInstThreshold;
171  AssumptionCache *AC;
172  SmallPtrSetImpl<BasicBlock *> *LoopHeaders;
173  // See comments in SimplifyCFGOpt::SimplifySwitch.
174  bool LateSimplifyCFG;
175  Value *isValueEqualityComparison(TerminatorInst *TI);
176  BasicBlock *GetValueEqualityComparisonCases(
177  TerminatorInst *TI, std::vector<ValueEqualityComparisonCase> &Cases);
178  bool SimplifyEqualityComparisonWithOnlyPredecessor(TerminatorInst *TI,
179  BasicBlock *Pred,
180  IRBuilder<> &Builder);
181  bool FoldValueComparisonIntoPredecessors(TerminatorInst *TI,
182  IRBuilder<> &Builder);
183 
184  bool SimplifyReturn(ReturnInst *RI, IRBuilder<> &Builder);
185  bool SimplifyResume(ResumeInst *RI, IRBuilder<> &Builder);
186  bool SimplifySingleResume(ResumeInst *RI);
187  bool SimplifyCommonResume(ResumeInst *RI);
188  bool SimplifyCleanupReturn(CleanupReturnInst *RI);
189  bool SimplifyUnreachable(UnreachableInst *UI);
190  bool SimplifySwitch(SwitchInst *SI, IRBuilder<> &Builder);
191  bool SimplifyIndirectBr(IndirectBrInst *IBI);
192  bool SimplifyUncondBranch(BranchInst *BI, IRBuilder<> &Builder);
193  bool SimplifyCondBranch(BranchInst *BI, IRBuilder<> &Builder);
194 
195 public:
196  SimplifyCFGOpt(const TargetTransformInfo &TTI, const DataLayout &DL,
197  unsigned BonusInstThreshold, AssumptionCache *AC,
198  SmallPtrSetImpl<BasicBlock *> *LoopHeaders,
199  bool LateSimplifyCFG)
200  : TTI(TTI), DL(DL), BonusInstThreshold(BonusInstThreshold), AC(AC),
201  LoopHeaders(LoopHeaders), LateSimplifyCFG(LateSimplifyCFG) {}
202 
203  bool run(BasicBlock *BB);
204 };
205 
206 } // end anonymous namespace
207 
208 /// Return true if it is safe to merge these two
209 /// terminator instructions together.
210 static bool
212  SmallSetVector<BasicBlock *, 4> *FailBlocks = nullptr) {
213  if (SI1 == SI2)
214  return false; // Can't merge with self!
215 
216  // It is not safe to merge these two switch instructions if they have a common
217  // successor, and if that successor has a PHI node, and if *that* PHI node has
218  // conflicting incoming values from the two switch blocks.
219  BasicBlock *SI1BB = SI1->getParent();
220  BasicBlock *SI2BB = SI2->getParent();
221 
222  SmallPtrSet<BasicBlock *, 16> SI1Succs(succ_begin(SI1BB), succ_end(SI1BB));
223  bool Fail = false;
224  for (BasicBlock *Succ : successors(SI2BB))
225  if (SI1Succs.count(Succ))
226  for (BasicBlock::iterator BBI = Succ->begin(); isa<PHINode>(BBI); ++BBI) {
227  PHINode *PN = cast<PHINode>(BBI);
228  if (PN->getIncomingValueForBlock(SI1BB) !=
229  PN->getIncomingValueForBlock(SI2BB)) {
230  if (FailBlocks)
231  FailBlocks->insert(Succ);
232  Fail = true;
233  }
234  }
235 
236  return !Fail;
237 }
238 
239 /// Return true if it is safe and profitable to merge these two terminator
240 /// instructions together, where SI1 is an unconditional branch. PhiNodes will
241 /// store all PHI nodes in common successors.
242 static bool
244  Instruction *Cond,
245  SmallVectorImpl<PHINode *> &PhiNodes) {
246  if (SI1 == SI2)
247  return false; // Can't merge with self!
248  assert(SI1->isUnconditional() && SI2->isConditional());
249 
250  // We fold the unconditional branch if we can easily update all PHI nodes in
251  // common successors:
252  // 1> We have a constant incoming value for the conditional branch;
253  // 2> We have "Cond" as the incoming value for the unconditional branch;
254  // 3> SI2->getCondition() and Cond have same operands.
255  CmpInst *Ci2 = dyn_cast<CmpInst>(SI2->getCondition());
256  if (!Ci2)
257  return false;
258  if (!(Cond->getOperand(0) == Ci2->getOperand(0) &&
259  Cond->getOperand(1) == Ci2->getOperand(1)) &&
260  !(Cond->getOperand(0) == Ci2->getOperand(1) &&
261  Cond->getOperand(1) == Ci2->getOperand(0)))
262  return false;
263 
264  BasicBlock *SI1BB = SI1->getParent();
265  BasicBlock *SI2BB = SI2->getParent();
266  SmallPtrSet<BasicBlock *, 16> SI1Succs(succ_begin(SI1BB), succ_end(SI1BB));
267  for (BasicBlock *Succ : successors(SI2BB))
268  if (SI1Succs.count(Succ))
269  for (BasicBlock::iterator BBI = Succ->begin(); isa<PHINode>(BBI); ++BBI) {
270  PHINode *PN = cast<PHINode>(BBI);
271  if (PN->getIncomingValueForBlock(SI1BB) != Cond ||
272  !isa<ConstantInt>(PN->getIncomingValueForBlock(SI2BB)))
273  return false;
274  PhiNodes.push_back(PN);
275  }
276  return true;
277 }
278 
279 /// Update PHI nodes in Succ to indicate that there will now be entries in it
280 /// from the 'NewPred' block. The values that will be flowing into the PHI nodes
281 /// will be the same as those coming in from ExistPred, an existing predecessor
282 /// of Succ.
283 static void AddPredecessorToBlock(BasicBlock *Succ, BasicBlock *NewPred,
284  BasicBlock *ExistPred) {
285  if (!isa<PHINode>(Succ->begin()))
286  return; // Quick exit if nothing to do
287 
288  PHINode *PN;
289  for (BasicBlock::iterator I = Succ->begin(); (PN = dyn_cast<PHINode>(I)); ++I)
290  PN->addIncoming(PN->getIncomingValueForBlock(ExistPred), NewPred);
291 }
292 
293 /// Compute an abstract "cost" of speculating the given instruction,
294 /// which is assumed to be safe to speculate. TCC_Free means cheap,
295 /// TCC_Basic means less cheap, and TCC_Expensive means prohibitively
296 /// expensive.
297 static unsigned ComputeSpeculationCost(const User *I,
298  const TargetTransformInfo &TTI) {
300  "Instruction is not safe to speculatively execute!");
301  return TTI.getUserCost(I);
302 }
303 
304 /// If we have a merge point of an "if condition" as accepted above,
305 /// return true if the specified value dominates the block. We
306 /// don't handle the true generality of domination here, just a special case
307 /// which works well enough for us.
308 ///
309 /// If AggressiveInsts is non-null, and if V does not dominate BB, we check to
310 /// see if V (which must be an instruction) and its recursive operands
311 /// that do not dominate BB have a combined cost lower than CostRemaining and
312 /// are non-trapping. If both are true, the instruction is inserted into the
313 /// set and true is returned.
314 ///
315 /// The cost for most non-trapping instructions is defined as 1 except for
316 /// Select whose cost is 2.
317 ///
318 /// After this function returns, CostRemaining is decreased by the cost of
319 /// V plus its non-dominating operands. If that cost is greater than
320 /// CostRemaining, false is returned and CostRemaining is undefined.
322  SmallPtrSetImpl<Instruction *> *AggressiveInsts,
323  unsigned &CostRemaining,
324  const TargetTransformInfo &TTI,
325  unsigned Depth = 0) {
326  // It is possible to hit a zero-cost cycle (phi/gep instructions for example),
327  // so limit the recursion depth.
328  // TODO: While this recursion limit does prevent pathological behavior, it
329  // would be better to track visited instructions to avoid cycles.
330  if (Depth == MaxSpeculationDepth)
331  return false;
332 
334  if (!I) {
335  // Non-instructions all dominate instructions, but not all constantexprs
336  // can be executed unconditionally.
337  if (ConstantExpr *C = dyn_cast<ConstantExpr>(V))
338  if (C->canTrap())
339  return false;
340  return true;
341  }
342  BasicBlock *PBB = I->getParent();
343 
344  // We don't want to allow weird loops that might have the "if condition" in
345  // the bottom of this block.
346  if (PBB == BB)
347  return false;
348 
349  // If this instruction is defined in a block that contains an unconditional
350  // branch to BB, then it must be in the 'conditional' part of the "if
351  // statement". If not, it definitely dominates the region.
353  if (!BI || BI->isConditional() || BI->getSuccessor(0) != BB)
354  return true;
355 
356  // If we aren't allowing aggressive promotion anymore, then don't consider
357  // instructions in the 'if region'.
358  if (!AggressiveInsts)
359  return false;
360 
361  // If we have seen this instruction before, don't count it again.
362  if (AggressiveInsts->count(I))
363  return true;
364 
365  // Okay, it looks like the instruction IS in the "condition". Check to
366  // see if it's a cheap instruction to unconditionally compute, and if it
367  // only uses stuff defined outside of the condition. If so, hoist it out.
369  return false;
370 
371  unsigned Cost = ComputeSpeculationCost(I, TTI);
372 
373  // Allow exactly one instruction to be speculated regardless of its cost
374  // (as long as it is safe to do so).
375  // This is intended to flatten the CFG even if the instruction is a division
376  // or other expensive operation. The speculation of an expensive instruction
377  // is expected to be undone in CodeGenPrepare if the speculation has not
378  // enabled further IR optimizations.
379  if (Cost > CostRemaining &&
380  (!SpeculateOneExpensiveInst || !AggressiveInsts->empty() || Depth > 0))
381  return false;
382 
383  // Avoid unsigned wrap.
384  CostRemaining = (Cost > CostRemaining) ? 0 : CostRemaining - Cost;
385 
386  // Okay, we can only really hoist these out if their operands do
387  // not take us over the cost threshold.
388  for (User::op_iterator i = I->op_begin(), e = I->op_end(); i != e; ++i)
389  if (!DominatesMergePoint(*i, BB, AggressiveInsts, CostRemaining, TTI,
390  Depth + 1))
391  return false;
392  // Okay, it's safe to do this! Remember this instruction.
393  AggressiveInsts->insert(I);
394  return true;
395 }
396 
397 /// Extract ConstantInt from value, looking through IntToPtr
398 /// and PointerNullValue. Return NULL if value is not a constant int.
399 static ConstantInt *GetConstantInt(Value *V, const DataLayout &DL) {
400  // Normal constant int.
401  ConstantInt *CI = dyn_cast<ConstantInt>(V);
402  if (CI || !isa<Constant>(V) || !V->getType()->isPointerTy())
403  return CI;
404 
405  // This is some kind of pointer constant. Turn it into a pointer-sized
406  // ConstantInt if possible.
407  IntegerType *PtrTy = cast<IntegerType>(DL.getIntPtrType(V->getType()));
408 
409  // Null pointer means 0, see SelectionDAGBuilder::getValue(const Value*).
410  if (isa<ConstantPointerNull>(V))
411  return ConstantInt::get(PtrTy, 0);
412 
413  // IntToPtr const int.
414  if (ConstantExpr *CE = dyn_cast<ConstantExpr>(V))
415  if (CE->getOpcode() == Instruction::IntToPtr)
416  if (ConstantInt *CI = dyn_cast<ConstantInt>(CE->getOperand(0))) {
417  // The constant is very likely to have the right type already.
418  if (CI->getType() == PtrTy)
419  return CI;
420  else
421  return cast<ConstantInt>(
422  ConstantExpr::getIntegerCast(CI, PtrTy, /*isSigned=*/false));
423  }
424  return nullptr;
425 }
426 
427 namespace {
428 
429 /// Given a chain of or (||) or and (&&) comparison of a value against a
430 /// constant, this will try to recover the information required for a switch
431 /// structure.
432 /// It will depth-first traverse the chain of comparison, seeking for patterns
433 /// like %a == 12 or %a < 4 and combine them to produce a set of integer
434 /// representing the different cases for the switch.
435 /// Note that if the chain is composed of '||' it will build the set of elements
436 /// that matches the comparisons (i.e. any of this value validate the chain)
437 /// while for a chain of '&&' it will build the set elements that make the test
438 /// fail.
439 struct ConstantComparesGatherer {
440  const DataLayout &DL;
441  Value *CompValue; /// Value found for the switch comparison
442  Value *Extra; /// Extra clause to be checked before the switch
443  SmallVector<ConstantInt *, 8> Vals; /// Set of integers to match in switch
444  unsigned UsedICmps; /// Number of comparisons matched in the and/or chain
445 
446  /// Construct and compute the result for the comparison instruction Cond
447  ConstantComparesGatherer(Instruction *Cond, const DataLayout &DL)
448  : DL(DL), CompValue(nullptr), Extra(nullptr), UsedICmps(0) {
449  gather(Cond);
450  }
451 
452  /// Prevent copy
453  ConstantComparesGatherer(const ConstantComparesGatherer &) = delete;
454  ConstantComparesGatherer &
455  operator=(const ConstantComparesGatherer &) = delete;
456 
457 private:
458  /// Try to set the current value used for the comparison, it succeeds only if
459  /// it wasn't set before or if the new value is the same as the old one
460  bool setValueOnce(Value *NewVal) {
461  if (CompValue && CompValue != NewVal)
462  return false;
463  CompValue = NewVal;
464  return (CompValue != nullptr);
465  }
466 
467  /// Try to match Instruction "I" as a comparison against a constant and
468  /// populates the array Vals with the set of values that match (or do not
469  /// match depending on isEQ).
470  /// Return false on failure. On success, the Value the comparison matched
471  /// against is placed in CompValue.
472  /// If CompValue is already set, the function is expected to fail if a match
473  /// is found but the value compared to is different.
474  bool matchInstruction(Instruction *I, bool isEQ) {
475  // If this is an icmp against a constant, handle this as one of the cases.
476  ICmpInst *ICI;
477  ConstantInt *C;
478  if (!((ICI = dyn_cast<ICmpInst>(I)) &&
479  (C = GetConstantInt(I->getOperand(1), DL)))) {
480  return false;
481  }
482 
483  Value *RHSVal;
484  const APInt *RHSC;
485 
486  // Pattern match a special case
487  // (x & ~2^z) == y --> x == y || x == y|2^z
488  // This undoes a transformation done by instcombine to fuse 2 compares.
489  if (ICI->getPredicate() == (isEQ ? ICmpInst::ICMP_EQ : ICmpInst::ICMP_NE)) {
490 
491  // It's a little bit hard to see why the following transformations are
492  // correct. Here is a CVC3 program to verify them for 64-bit values:
493 
494  /*
495  ONE : BITVECTOR(64) = BVZEROEXTEND(0bin1, 63);
496  x : BITVECTOR(64);
497  y : BITVECTOR(64);
498  z : BITVECTOR(64);
499  mask : BITVECTOR(64) = BVSHL(ONE, z);
500  QUERY( (y & ~mask = y) =>
501  ((x & ~mask = y) <=> (x = y OR x = (y | mask)))
502  );
503  QUERY( (y | mask = y) =>
504  ((x | mask = y) <=> (x = y OR x = (y & ~mask)))
505  );
506  */
507 
508  // Please note that each pattern must be a dual implication (<--> or
509  // iff). One directional implication can create spurious matches. If the
510  // implication is only one-way, an unsatisfiable condition on the left
511  // side can imply a satisfiable condition on the right side. Dual
512  // implication ensures that satisfiable conditions are transformed to
513  // other satisfiable conditions and unsatisfiable conditions are
514  // transformed to other unsatisfiable conditions.
515 
516  // Here is a concrete example of a unsatisfiable condition on the left
517  // implying a satisfiable condition on the right:
518  //
519  // mask = (1 << z)
520  // (x & ~mask) == y --> (x == y || x == (y | mask))
521  //
522  // Substituting y = 3, z = 0 yields:
523  // (x & -2) == 3 --> (x == 3 || x == 2)
524 
525  // Pattern match a special case:
526  /*
527  QUERY( (y & ~mask = y) =>
528  ((x & ~mask = y) <=> (x = y OR x = (y | mask)))
529  );
530  */
531  if (match(ICI->getOperand(0),
532  m_And(m_Value(RHSVal), m_APInt(RHSC)))) {
533  APInt Mask = ~*RHSC;
534  if (Mask.isPowerOf2() && (C->getValue() & ~Mask) == C->getValue()) {
535  // If we already have a value for the switch, it has to match!
536  if (!setValueOnce(RHSVal))
537  return false;
538 
539  Vals.push_back(C);
540  Vals.push_back(
542  C->getValue() | Mask));
543  UsedICmps++;
544  return true;
545  }
546  }
547 
548  // Pattern match a special case:
549  /*
550  QUERY( (y | mask = y) =>
551  ((x | mask = y) <=> (x = y OR x = (y & ~mask)))
552  );
553  */
554  if (match(ICI->getOperand(0),
555  m_Or(m_Value(RHSVal), m_APInt(RHSC)))) {
556  APInt Mask = *RHSC;
557  if (Mask.isPowerOf2() && (C->getValue() | Mask) == C->getValue()) {
558  // If we already have a value for the switch, it has to match!
559  if (!setValueOnce(RHSVal))
560  return false;
561 
562  Vals.push_back(C);
564  C->getValue() & ~Mask));
565  UsedICmps++;
566  return true;
567  }
568  }
569 
570  // If we already have a value for the switch, it has to match!
571  if (!setValueOnce(ICI->getOperand(0)))
572  return false;
573 
574  UsedICmps++;
575  Vals.push_back(C);
576  return ICI->getOperand(0);
577  }
578 
579  // If we have "x ult 3", for example, then we can add 0,1,2 to the set.
581  ICI->getPredicate(), C->getValue());
582 
583  // Shift the range if the compare is fed by an add. This is the range
584  // compare idiom as emitted by instcombine.
585  Value *CandidateVal = I->getOperand(0);
586  if (match(I->getOperand(0), m_Add(m_Value(RHSVal), m_APInt(RHSC)))) {
587  Span = Span.subtract(*RHSC);
588  CandidateVal = RHSVal;
589  }
590 
591  // If this is an and/!= check, then we are looking to build the set of
592  // value that *don't* pass the and chain. I.e. to turn "x ugt 2" into
593  // x != 0 && x != 1.
594  if (!isEQ)
595  Span = Span.inverse();
596 
597  // If there are a ton of values, we don't want to make a ginormous switch.
598  if (Span.isSizeLargerThan(8) || Span.isEmptySet()) {
599  return false;
600  }
601 
602  // If we already have a value for the switch, it has to match!
603  if (!setValueOnce(CandidateVal))
604  return false;
605 
606  // Add all values from the range to the set
607  for (APInt Tmp = Span.getLower(); Tmp != Span.getUpper(); ++Tmp)
608  Vals.push_back(ConstantInt::get(I->getContext(), Tmp));
609 
610  UsedICmps++;
611  return true;
612  }
613 
614  /// Given a potentially 'or'd or 'and'd together collection of icmp
615  /// eq/ne/lt/gt instructions that compare a value against a constant, extract
616  /// the value being compared, and stick the list constants into the Vals
617  /// vector.
618  /// One "Extra" case is allowed to differ from the other.
619  void gather(Value *V) {
621  bool isEQ = (I->getOpcode() == Instruction::Or);
622 
623  // Keep a stack (SmallVector for efficiency) for depth-first traversal
625  SmallPtrSet<Value *, 8> Visited;
626 
627  // Initialize
628  Visited.insert(V);
629  DFT.push_back(V);
630 
631  while (!DFT.empty()) {
632  V = DFT.pop_back_val();
633 
634  if (Instruction *I = dyn_cast<Instruction>(V)) {
635  // If it is a || (or && depending on isEQ), process the operands.
636  if (I->getOpcode() == (isEQ ? Instruction::Or : Instruction::And)) {
637  if (Visited.insert(I->getOperand(1)).second)
638  DFT.push_back(I->getOperand(1));
639  if (Visited.insert(I->getOperand(0)).second)
640  DFT.push_back(I->getOperand(0));
641  continue;
642  }
643 
644  // Try to match the current instruction
645  if (matchInstruction(I, isEQ))
646  // Match succeed, continue the loop
647  continue;
648  }
649 
650  // One element of the sequence of || (or &&) could not be match as a
651  // comparison against the same value as the others.
652  // We allow only one "Extra" case to be checked before the switch
653  if (!Extra) {
654  Extra = V;
655  continue;
656  }
657  // Failed to parse a proper sequence, abort now
658  CompValue = nullptr;
659  break;
660  }
661  }
662 };
663 
664 } // end anonymous namespace
665 
667  Instruction *Cond = nullptr;
668  if (SwitchInst *SI = dyn_cast<SwitchInst>(TI)) {
669  Cond = dyn_cast<Instruction>(SI->getCondition());
670  } else if (BranchInst *BI = dyn_cast<BranchInst>(TI)) {
671  if (BI->isConditional())
672  Cond = dyn_cast<Instruction>(BI->getCondition());
673  } else if (IndirectBrInst *IBI = dyn_cast<IndirectBrInst>(TI)) {
674  Cond = dyn_cast<Instruction>(IBI->getAddress());
675  }
676 
677  TI->eraseFromParent();
678  if (Cond)
680 }
681 
682 /// Return true if the specified terminator checks
683 /// to see if a value is equal to constant integer value.
684 Value *SimplifyCFGOpt::isValueEqualityComparison(TerminatorInst *TI) {
685  Value *CV = nullptr;
686  if (SwitchInst *SI = dyn_cast<SwitchInst>(TI)) {
687  // Do not permit merging of large switch instructions into their
688  // predecessors unless there is only one predecessor.
689  if (SI->getNumSuccessors() * std::distance(pred_begin(SI->getParent()),
690  pred_end(SI->getParent())) <=
691  128)
692  CV = SI->getCondition();
693  } else if (BranchInst *BI = dyn_cast<BranchInst>(TI))
694  if (BI->isConditional() && BI->getCondition()->hasOneUse())
695  if (ICmpInst *ICI = dyn_cast<ICmpInst>(BI->getCondition())) {
696  if (ICI->isEquality() && GetConstantInt(ICI->getOperand(1), DL))
697  CV = ICI->getOperand(0);
698  }
699 
700  // Unwrap any lossless ptrtoint cast.
701  if (CV) {
702  if (PtrToIntInst *PTII = dyn_cast<PtrToIntInst>(CV)) {
703  Value *Ptr = PTII->getPointerOperand();
704  if (PTII->getType() == DL.getIntPtrType(Ptr->getType()))
705  CV = Ptr;
706  }
707  }
708  return CV;
709 }
710 
711 /// Given a value comparison instruction,
712 /// decode all of the 'cases' that it represents and return the 'default' block.
713 BasicBlock *SimplifyCFGOpt::GetValueEqualityComparisonCases(
714  TerminatorInst *TI, std::vector<ValueEqualityComparisonCase> &Cases) {
715  if (SwitchInst *SI = dyn_cast<SwitchInst>(TI)) {
716  Cases.reserve(SI->getNumCases());
717  for (auto Case : SI->cases())
718  Cases.push_back(ValueEqualityComparisonCase(Case.getCaseValue(),
719  Case.getCaseSuccessor()));
720  return SI->getDefaultDest();
721  }
722 
723  BranchInst *BI = cast<BranchInst>(TI);
724  ICmpInst *ICI = cast<ICmpInst>(BI->getCondition());
725  BasicBlock *Succ = BI->getSuccessor(ICI->getPredicate() == ICmpInst::ICMP_NE);
726  Cases.push_back(ValueEqualityComparisonCase(
727  GetConstantInt(ICI->getOperand(1), DL), Succ));
728  return BI->getSuccessor(ICI->getPredicate() == ICmpInst::ICMP_EQ);
729 }
730 
731 /// Given a vector of bb/value pairs, remove any entries
732 /// in the list that match the specified block.
733 static void
735  std::vector<ValueEqualityComparisonCase> &Cases) {
736  Cases.erase(std::remove(Cases.begin(), Cases.end(), BB), Cases.end());
737 }
738 
739 /// Return true if there are any keys in C1 that exist in C2 as well.
740 static bool ValuesOverlap(std::vector<ValueEqualityComparisonCase> &C1,
741  std::vector<ValueEqualityComparisonCase> &C2) {
742  std::vector<ValueEqualityComparisonCase> *V1 = &C1, *V2 = &C2;
743 
744  // Make V1 be smaller than V2.
745  if (V1->size() > V2->size())
746  std::swap(V1, V2);
747 
748  if (V1->empty())
749  return false;
750  if (V1->size() == 1) {
751  // Just scan V2.
752  ConstantInt *TheVal = (*V1)[0].Value;
753  for (unsigned i = 0, e = V2->size(); i != e; ++i)
754  if (TheVal == (*V2)[i].Value)
755  return true;
756  }
757 
758  // Otherwise, just sort both lists and compare element by element.
759  array_pod_sort(V1->begin(), V1->end());
760  array_pod_sort(V2->begin(), V2->end());
761  unsigned i1 = 0, i2 = 0, e1 = V1->size(), e2 = V2->size();
762  while (i1 != e1 && i2 != e2) {
763  if ((*V1)[i1].Value == (*V2)[i2].Value)
764  return true;
765  if ((*V1)[i1].Value < (*V2)[i2].Value)
766  ++i1;
767  else
768  ++i2;
769  }
770  return false;
771 }
772 
773 /// If TI is known to be a terminator instruction and its block is known to
774 /// only have a single predecessor block, check to see if that predecessor is
775 /// also a value comparison with the same value, and if that comparison
776 /// determines the outcome of this comparison. If so, simplify TI. This does a
777 /// very limited form of jump threading.
778 bool SimplifyCFGOpt::SimplifyEqualityComparisonWithOnlyPredecessor(
779  TerminatorInst *TI, BasicBlock *Pred, IRBuilder<> &Builder) {
780  Value *PredVal = isValueEqualityComparison(Pred->getTerminator());
781  if (!PredVal)
782  return false; // Not a value comparison in predecessor.
783 
784  Value *ThisVal = isValueEqualityComparison(TI);
785  assert(ThisVal && "This isn't a value comparison!!");
786  if (ThisVal != PredVal)
787  return false; // Different predicates.
788 
789  // TODO: Preserve branch weight metadata, similarly to how
790  // FoldValueComparisonIntoPredecessors preserves it.
791 
792  // Find out information about when control will move from Pred to TI's block.
793  std::vector<ValueEqualityComparisonCase> PredCases;
794  BasicBlock *PredDef =
795  GetValueEqualityComparisonCases(Pred->getTerminator(), PredCases);
796  EliminateBlockCases(PredDef, PredCases); // Remove default from cases.
797 
798  // Find information about how control leaves this block.
799  std::vector<ValueEqualityComparisonCase> ThisCases;
800  BasicBlock *ThisDef = GetValueEqualityComparisonCases(TI, ThisCases);
801  EliminateBlockCases(ThisDef, ThisCases); // Remove default from cases.
802 
803  // If TI's block is the default block from Pred's comparison, potentially
804  // simplify TI based on this knowledge.
805  if (PredDef == TI->getParent()) {
806  // If we are here, we know that the value is none of those cases listed in
807  // PredCases. If there are any cases in ThisCases that are in PredCases, we
808  // can simplify TI.
809  if (!ValuesOverlap(PredCases, ThisCases))
810  return false;
811 
812  if (isa<BranchInst>(TI)) {
813  // Okay, one of the successors of this condbr is dead. Convert it to a
814  // uncond br.
815  assert(ThisCases.size() == 1 && "Branch can only have one case!");
816  // Insert the new branch.
817  Instruction *NI = Builder.CreateBr(ThisDef);
818  (void)NI;
819 
820  // Remove PHI node entries for the dead edge.
821  ThisCases[0].Dest->removePredecessor(TI->getParent());
822 
823  DEBUG(dbgs() << "Threading pred instr: " << *Pred->getTerminator()
824  << "Through successor TI: " << *TI << "Leaving: " << *NI
825  << "\n");
826 
828  return true;
829  }
830 
831  SwitchInst *SI = cast<SwitchInst>(TI);
832  // Okay, TI has cases that are statically dead, prune them away.
833  SmallPtrSet<Constant *, 16> DeadCases;
834  for (unsigned i = 0, e = PredCases.size(); i != e; ++i)
835  DeadCases.insert(PredCases[i].Value);
836 
837  DEBUG(dbgs() << "Threading pred instr: " << *Pred->getTerminator()
838  << "Through successor TI: " << *TI);
839 
840  // Collect branch weights into a vector.
841  SmallVector<uint32_t, 8> Weights;
843  bool HasWeight = MD && (MD->getNumOperands() == 2 + SI->getNumCases());
844  if (HasWeight)
845  for (unsigned MD_i = 1, MD_e = MD->getNumOperands(); MD_i < MD_e;
846  ++MD_i) {
847  ConstantInt *CI = mdconst::extract<ConstantInt>(MD->getOperand(MD_i));
848  Weights.push_back(CI->getValue().getZExtValue());
849  }
850  for (SwitchInst::CaseIt i = SI->case_end(), e = SI->case_begin(); i != e;) {
851  --i;
852  if (DeadCases.count(i->getCaseValue())) {
853  if (HasWeight) {
854  std::swap(Weights[i->getCaseIndex() + 1], Weights.back());
855  Weights.pop_back();
856  }
857  i->getCaseSuccessor()->removePredecessor(TI->getParent());
858  SI->removeCase(i);
859  }
860  }
861  if (HasWeight && Weights.size() >= 2)
863  MDBuilder(SI->getParent()->getContext())
864  .createBranchWeights(Weights));
865 
866  DEBUG(dbgs() << "Leaving: " << *TI << "\n");
867  return true;
868  }
869 
870  // Otherwise, TI's block must correspond to some matched value. Find out
871  // which value (or set of values) this is.
872  ConstantInt *TIV = nullptr;
873  BasicBlock *TIBB = TI->getParent();
874  for (unsigned i = 0, e = PredCases.size(); i != e; ++i)
875  if (PredCases[i].Dest == TIBB) {
876  if (TIV)
877  return false; // Cannot handle multiple values coming to this block.
878  TIV = PredCases[i].Value;
879  }
880  assert(TIV && "No edge from pred to succ?");
881 
882  // Okay, we found the one constant that our value can be if we get into TI's
883  // BB. Find out which successor will unconditionally be branched to.
884  BasicBlock *TheRealDest = nullptr;
885  for (unsigned i = 0, e = ThisCases.size(); i != e; ++i)
886  if (ThisCases[i].Value == TIV) {
887  TheRealDest = ThisCases[i].Dest;
888  break;
889  }
890 
891  // If not handled by any explicit cases, it is handled by the default case.
892  if (!TheRealDest)
893  TheRealDest = ThisDef;
894 
895  // Remove PHI node entries for dead edges.
896  BasicBlock *CheckEdge = TheRealDest;
897  for (BasicBlock *Succ : successors(TIBB))
898  if (Succ != CheckEdge)
899  Succ->removePredecessor(TIBB);
900  else
901  CheckEdge = nullptr;
902 
903  // Insert the new branch.
904  Instruction *NI = Builder.CreateBr(TheRealDest);
905  (void)NI;
906 
907  DEBUG(dbgs() << "Threading pred instr: " << *Pred->getTerminator()
908  << "Through successor TI: " << *TI << "Leaving: " << *NI
909  << "\n");
910 
912  return true;
913 }
914 
915 namespace {
916 
917 /// This class implements a stable ordering of constant
918 /// integers that does not depend on their address. This is important for
919 /// applications that sort ConstantInt's to ensure uniqueness.
920 struct ConstantIntOrdering {
921  bool operator()(const ConstantInt *LHS, const ConstantInt *RHS) const {
922  return LHS->getValue().ult(RHS->getValue());
923  }
924 };
925 
926 } // end anonymous namespace
927 
928 static int ConstantIntSortPredicate(ConstantInt *const *P1,
929  ConstantInt *const *P2) {
930  const ConstantInt *LHS = *P1;
931  const ConstantInt *RHS = *P2;
932  if (LHS == RHS)
933  return 0;
934  return LHS->getValue().ult(RHS->getValue()) ? 1 : -1;
935 }
936 
937 static inline bool HasBranchWeights(const Instruction *I) {
939  if (ProfMD && ProfMD->getOperand(0))
940  if (MDString *MDS = dyn_cast<MDString>(ProfMD->getOperand(0)))
941  return MDS->getString().equals("branch_weights");
942 
943  return false;
944 }
945 
946 /// Get Weights of a given TerminatorInst, the default weight is at the front
947 /// of the vector. If TI is a conditional eq, we need to swap the branch-weight
948 /// metadata.
950  SmallVectorImpl<uint64_t> &Weights) {
952  assert(MD);
953  for (unsigned i = 1, e = MD->getNumOperands(); i < e; ++i) {
954  ConstantInt *CI = mdconst::extract<ConstantInt>(MD->getOperand(i));
955  Weights.push_back(CI->getValue().getZExtValue());
956  }
957 
958  // If TI is a conditional eq, the default case is the false case,
959  // and the corresponding branch-weight data is at index 2. We swap the
960  // default weight to be the first entry.
961  if (BranchInst *BI = dyn_cast<BranchInst>(TI)) {
962  assert(Weights.size() == 2);
963  ICmpInst *ICI = cast<ICmpInst>(BI->getCondition());
964  if (ICI->getPredicate() == ICmpInst::ICMP_EQ)
965  std::swap(Weights.front(), Weights.back());
966  }
967 }
968 
969 /// Keep halving the weights until all can fit in uint32_t.
970 static void FitWeights(MutableArrayRef<uint64_t> Weights) {
971  uint64_t Max = *std::max_element(Weights.begin(), Weights.end());
972  if (Max > UINT_MAX) {
973  unsigned Offset = 32 - countLeadingZeros(Max);
974  for (uint64_t &I : Weights)
975  I >>= Offset;
976  }
977 }
978 
979 /// The specified terminator is a value equality comparison instruction
980 /// (either a switch or a branch on "X == c").
981 /// See if any of the predecessors of the terminator block are value comparisons
982 /// on the same value. If so, and if safe to do so, fold them together.
983 bool SimplifyCFGOpt::FoldValueComparisonIntoPredecessors(TerminatorInst *TI,
984  IRBuilder<> &Builder) {
985  BasicBlock *BB = TI->getParent();
986  Value *CV = isValueEqualityComparison(TI); // CondVal
987  assert(CV && "Not a comparison?");
988  bool Changed = false;
989 
991  while (!Preds.empty()) {
992  BasicBlock *Pred = Preds.pop_back_val();
993 
994  // See if the predecessor is a comparison with the same value.
995  TerminatorInst *PTI = Pred->getTerminator();
996  Value *PCV = isValueEqualityComparison(PTI); // PredCondVal
997 
998  if (PCV == CV && TI != PTI) {
1000  if (!SafeToMergeTerminators(TI, PTI, &FailBlocks)) {
1001  for (auto *Succ : FailBlocks) {
1002  if (!SplitBlockPredecessors(Succ, TI->getParent(), ".fold.split"))
1003  return false;
1004  }
1005  }
1006 
1007  // Figure out which 'cases' to copy from SI to PSI.
1008  std::vector<ValueEqualityComparisonCase> BBCases;
1009  BasicBlock *BBDefault = GetValueEqualityComparisonCases(TI, BBCases);
1010 
1011  std::vector<ValueEqualityComparisonCase> PredCases;
1012  BasicBlock *PredDefault = GetValueEqualityComparisonCases(PTI, PredCases);
1013 
1014  // Based on whether the default edge from PTI goes to BB or not, fill in
1015  // PredCases and PredDefault with the new switch cases we would like to
1016  // build.
1017  SmallVector<BasicBlock *, 8> NewSuccessors;
1018 
1019  // Update the branch weight metadata along the way
1020  SmallVector<uint64_t, 8> Weights;
1021  bool PredHasWeights = HasBranchWeights(PTI);
1022  bool SuccHasWeights = HasBranchWeights(TI);
1023 
1024  if (PredHasWeights) {
1025  GetBranchWeights(PTI, Weights);
1026  // branch-weight metadata is inconsistent here.
1027  if (Weights.size() != 1 + PredCases.size())
1028  PredHasWeights = SuccHasWeights = false;
1029  } else if (SuccHasWeights)
1030  // If there are no predecessor weights but there are successor weights,
1031  // populate Weights with 1, which will later be scaled to the sum of
1032  // successor's weights
1033  Weights.assign(1 + PredCases.size(), 1);
1034 
1035  SmallVector<uint64_t, 8> SuccWeights;
1036  if (SuccHasWeights) {
1037  GetBranchWeights(TI, SuccWeights);
1038  // branch-weight metadata is inconsistent here.
1039  if (SuccWeights.size() != 1 + BBCases.size())
1040  PredHasWeights = SuccHasWeights = false;
1041  } else if (PredHasWeights)
1042  SuccWeights.assign(1 + BBCases.size(), 1);
1043 
1044  if (PredDefault == BB) {
1045  // If this is the default destination from PTI, only the edges in TI
1046  // that don't occur in PTI, or that branch to BB will be activated.
1047  std::set<ConstantInt *, ConstantIntOrdering> PTIHandled;
1048  for (unsigned i = 0, e = PredCases.size(); i != e; ++i)
1049  if (PredCases[i].Dest != BB)
1050  PTIHandled.insert(PredCases[i].Value);
1051  else {
1052  // The default destination is BB, we don't need explicit targets.
1053  std::swap(PredCases[i], PredCases.back());
1054 
1055  if (PredHasWeights || SuccHasWeights) {
1056  // Increase weight for the default case.
1057  Weights[0] += Weights[i + 1];
1058  std::swap(Weights[i + 1], Weights.back());
1059  Weights.pop_back();
1060  }
1061 
1062  PredCases.pop_back();
1063  --i;
1064  --e;
1065  }
1066 
1067  // Reconstruct the new switch statement we will be building.
1068  if (PredDefault != BBDefault) {
1069  PredDefault->removePredecessor(Pred);
1070  PredDefault = BBDefault;
1071  NewSuccessors.push_back(BBDefault);
1072  }
1073 
1074  unsigned CasesFromPred = Weights.size();
1075  uint64_t ValidTotalSuccWeight = 0;
1076  for (unsigned i = 0, e = BBCases.size(); i != e; ++i)
1077  if (!PTIHandled.count(BBCases[i].Value) &&
1078  BBCases[i].Dest != BBDefault) {
1079  PredCases.push_back(BBCases[i]);
1080  NewSuccessors.push_back(BBCases[i].Dest);
1081  if (SuccHasWeights || PredHasWeights) {
1082  // The default weight is at index 0, so weight for the ith case
1083  // should be at index i+1. Scale the cases from successor by
1084  // PredDefaultWeight (Weights[0]).
1085  Weights.push_back(Weights[0] * SuccWeights[i + 1]);
1086  ValidTotalSuccWeight += SuccWeights[i + 1];
1087  }
1088  }
1089 
1090  if (SuccHasWeights || PredHasWeights) {
1091  ValidTotalSuccWeight += SuccWeights[0];
1092  // Scale the cases from predecessor by ValidTotalSuccWeight.
1093  for (unsigned i = 1; i < CasesFromPred; ++i)
1094  Weights[i] *= ValidTotalSuccWeight;
1095  // Scale the default weight by SuccDefaultWeight (SuccWeights[0]).
1096  Weights[0] *= SuccWeights[0];
1097  }
1098  } else {
1099  // If this is not the default destination from PSI, only the edges
1100  // in SI that occur in PSI with a destination of BB will be
1101  // activated.
1102  std::set<ConstantInt *, ConstantIntOrdering> PTIHandled;
1103  std::map<ConstantInt *, uint64_t> WeightsForHandled;
1104  for (unsigned i = 0, e = PredCases.size(); i != e; ++i)
1105  if (PredCases[i].Dest == BB) {
1106  PTIHandled.insert(PredCases[i].Value);
1107 
1108  if (PredHasWeights || SuccHasWeights) {
1109  WeightsForHandled[PredCases[i].Value] = Weights[i + 1];
1110  std::swap(Weights[i + 1], Weights.back());
1111  Weights.pop_back();
1112  }
1113 
1114  std::swap(PredCases[i], PredCases.back());
1115  PredCases.pop_back();
1116  --i;
1117  --e;
1118  }
1119 
1120  // Okay, now we know which constants were sent to BB from the
1121  // predecessor. Figure out where they will all go now.
1122  for (unsigned i = 0, e = BBCases.size(); i != e; ++i)
1123  if (PTIHandled.count(BBCases[i].Value)) {
1124  // If this is one we are capable of getting...
1125  if (PredHasWeights || SuccHasWeights)
1126  Weights.push_back(WeightsForHandled[BBCases[i].Value]);
1127  PredCases.push_back(BBCases[i]);
1128  NewSuccessors.push_back(BBCases[i].Dest);
1129  PTIHandled.erase(
1130  BBCases[i].Value); // This constant is taken care of
1131  }
1132 
1133  // If there are any constants vectored to BB that TI doesn't handle,
1134  // they must go to the default destination of TI.
1135  for (ConstantInt *I : PTIHandled) {
1136  if (PredHasWeights || SuccHasWeights)
1137  Weights.push_back(WeightsForHandled[I]);
1138  PredCases.push_back(ValueEqualityComparisonCase(I, BBDefault));
1139  NewSuccessors.push_back(BBDefault);
1140  }
1141  }
1142 
1143  // Okay, at this point, we know which new successor Pred will get. Make
1144  // sure we update the number of entries in the PHI nodes for these
1145  // successors.
1146  for (BasicBlock *NewSuccessor : NewSuccessors)
1147  AddPredecessorToBlock(NewSuccessor, Pred, BB);
1148 
1149  Builder.SetInsertPoint(PTI);
1150  // Convert pointer to int before we switch.
1151  if (CV->getType()->isPointerTy()) {
1152  CV = Builder.CreatePtrToInt(CV, DL.getIntPtrType(CV->getType()),
1153  "magicptr");
1154  }
1155 
1156  // Now that the successors are updated, create the new Switch instruction.
1157  SwitchInst *NewSI =
1158  Builder.CreateSwitch(CV, PredDefault, PredCases.size());
1159  NewSI->setDebugLoc(PTI->getDebugLoc());
1160  for (ValueEqualityComparisonCase &V : PredCases)
1161  NewSI->addCase(V.Value, V.Dest);
1162 
1163  if (PredHasWeights || SuccHasWeights) {
1164  // Halve the weights if any of them cannot fit in an uint32_t
1165  FitWeights(Weights);
1166 
1167  SmallVector<uint32_t, 8> MDWeights(Weights.begin(), Weights.end());
1168 
1169  NewSI->setMetadata(
1171  MDBuilder(BB->getContext()).createBranchWeights(MDWeights));
1172  }
1173 
1175 
1176  // Okay, last check. If BB is still a successor of PSI, then we must
1177  // have an infinite loop case. If so, add an infinitely looping block
1178  // to handle the case to preserve the behavior of the code.
1179  BasicBlock *InfLoopBlock = nullptr;
1180  for (unsigned i = 0, e = NewSI->getNumSuccessors(); i != e; ++i)
1181  if (NewSI->getSuccessor(i) == BB) {
1182  if (!InfLoopBlock) {
1183  // Insert it at the end of the function, because it's either code,
1184  // or it won't matter if it's hot. :)
1185  InfLoopBlock = BasicBlock::Create(BB->getContext(), "infloop",
1186  BB->getParent());
1187  BranchInst::Create(InfLoopBlock, InfLoopBlock);
1188  }
1189  NewSI->setSuccessor(i, InfLoopBlock);
1190  }
1191 
1192  Changed = true;
1193  }
1194  }
1195  return Changed;
1196 }
1197 
1198 // If we would need to insert a select that uses the value of this invoke
1199 // (comments in HoistThenElseCodeToIf explain why we would need to do this), we
1200 // can't hoist the invoke, as there is nowhere to put the select in this case.
1202  Instruction *I1, Instruction *I2) {
1203  for (BasicBlock *Succ : successors(BB1)) {
1204  PHINode *PN;
1205  for (BasicBlock::iterator BBI = Succ->begin();
1206  (PN = dyn_cast<PHINode>(BBI)); ++BBI) {
1207  Value *BB1V = PN->getIncomingValueForBlock(BB1);
1208  Value *BB2V = PN->getIncomingValueForBlock(BB2);
1209  if (BB1V != BB2V && (BB1V == I1 || BB2V == I2)) {
1210  return false;
1211  }
1212  }
1213  }
1214  return true;
1215 }
1216 
1218 
1219 /// Given a conditional branch that goes to BB1 and BB2, hoist any common code
1220 /// in the two blocks up into the branch block. The caller of this function
1221 /// guarantees that BI's block dominates BB1 and BB2.
1223  const TargetTransformInfo &TTI) {
1224  // This does very trivial matching, with limited scanning, to find identical
1225  // instructions in the two blocks. In particular, we don't want to get into
1226  // O(M*N) situations here where M and N are the sizes of BB1 and BB2. As
1227  // such, we currently just scan for obviously identical instructions in an
1228  // identical order.
1229  BasicBlock *BB1 = BI->getSuccessor(0); // The true destination.
1230  BasicBlock *BB2 = BI->getSuccessor(1); // The false destination
1231 
1232  BasicBlock::iterator BB1_Itr = BB1->begin();
1233  BasicBlock::iterator BB2_Itr = BB2->begin();
1234 
1235  Instruction *I1 = &*BB1_Itr++, *I2 = &*BB2_Itr++;
1236  // Skip debug info if it is not identical.
1239  if (!DBI1 || !DBI2 || !DBI1->isIdenticalToWhenDefined(DBI2)) {
1240  while (isa<DbgInfoIntrinsic>(I1))
1241  I1 = &*BB1_Itr++;
1242  while (isa<DbgInfoIntrinsic>(I2))
1243  I2 = &*BB2_Itr++;
1244  }
1245  if (isa<PHINode>(I1) || !I1->isIdenticalToWhenDefined(I2) ||
1246  (isa<InvokeInst>(I1) && !isSafeToHoistInvoke(BB1, BB2, I1, I2)))
1247  return false;
1248 
1249  BasicBlock *BIParent = BI->getParent();
1250 
1251  bool Changed = false;
1252  do {
1253  // If we are hoisting the terminator instruction, don't move one (making a
1254  // broken BB), instead clone it, and remove BI.
1255  if (isa<TerminatorInst>(I1))
1256  goto HoistTerminator;
1257 
1258  if (!TTI.isProfitableToHoist(I1) || !TTI.isProfitableToHoist(I2))
1259  return Changed;
1260 
1261  // For a normal instruction, we just move one to right before the branch,
1262  // then replace all uses of the other with the first. Finally, we remove
1263  // the now redundant second instruction.
1264  BIParent->getInstList().splice(BI->getIterator(), BB1->getInstList(), I1);
1265  if (!I2->use_empty())
1266  I2->replaceAllUsesWith(I1);
1267  I1->andIRFlags(I2);
1268  unsigned KnownIDs[] = {LLVMContext::MD_tbaa,
1278  combineMetadata(I1, I2, KnownIDs);
1279 
1280  // I1 and I2 are being combined into a single instruction. Its debug
1281  // location is the merged locations of the original instructions.
1282  if (!isa<CallInst>(I1))
1283  I1->setDebugLoc(
1284  DILocation::getMergedLocation(I1->getDebugLoc(), I2->getDebugLoc()));
1285 
1286  I2->eraseFromParent();
1287  Changed = true;
1288 
1289  I1 = &*BB1_Itr++;
1290  I2 = &*BB2_Itr++;
1291  // Skip debug info if it is not identical.
1294  if (!DBI1 || !DBI2 || !DBI1->isIdenticalToWhenDefined(DBI2)) {
1295  while (isa<DbgInfoIntrinsic>(I1))
1296  I1 = &*BB1_Itr++;
1297  while (isa<DbgInfoIntrinsic>(I2))
1298  I2 = &*BB2_Itr++;
1299  }
1300  } while (I1->isIdenticalToWhenDefined(I2));
1301 
1302  return true;
1303 
1304 HoistTerminator:
1305  // It may not be possible to hoist an invoke.
1306  if (isa<InvokeInst>(I1) && !isSafeToHoistInvoke(BB1, BB2, I1, I2))
1307  return Changed;
1308 
1309  for (BasicBlock *Succ : successors(BB1)) {
1310  PHINode *PN;
1311  for (BasicBlock::iterator BBI = Succ->begin();
1312  (PN = dyn_cast<PHINode>(BBI)); ++BBI) {
1313  Value *BB1V = PN->getIncomingValueForBlock(BB1);
1314  Value *BB2V = PN->getIncomingValueForBlock(BB2);
1315  if (BB1V == BB2V)
1316  continue;
1317 
1318  // Check for passingValueIsAlwaysUndefined here because we would rather
1319  // eliminate undefined control flow then converting it to a select.
1320  if (passingValueIsAlwaysUndefined(BB1V, PN) ||
1322  return Changed;
1323 
1324  if (isa<ConstantExpr>(BB1V) && !isSafeToSpeculativelyExecute(BB1V))
1325  return Changed;
1326  if (isa<ConstantExpr>(BB2V) && !isSafeToSpeculativelyExecute(BB2V))
1327  return Changed;
1328  }
1329  }
1330 
1331  // Okay, it is safe to hoist the terminator.
1332  Instruction *NT = I1->clone();
1333  BIParent->getInstList().insert(BI->getIterator(), NT);
1334  if (!NT->getType()->isVoidTy()) {
1335  I1->replaceAllUsesWith(NT);
1336  I2->replaceAllUsesWith(NT);
1337  NT->takeName(I1);
1338  }
1340  I1->getDebugLoc(), I2->getDebugLoc()));
1341 
1342  IRBuilder<NoFolder> Builder(NT);
1343  // Hoisting one of the terminators from our successor is a great thing.
1344  // Unfortunately, the successors of the if/else blocks may have PHI nodes in
1345  // them. If they do, all PHI entries for BB1/BB2 must agree for all PHI
1346  // nodes, so we insert select instruction to compute the final result.
1347  std::map<std::pair<Value *, Value *>, SelectInst *> InsertedSelects;
1348  for (BasicBlock *Succ : successors(BB1)) {
1349  PHINode *PN;
1350  for (BasicBlock::iterator BBI = Succ->begin();
1351  (PN = dyn_cast<PHINode>(BBI)); ++BBI) {
1352  Value *BB1V = PN->getIncomingValueForBlock(BB1);
1353  Value *BB2V = PN->getIncomingValueForBlock(BB2);
1354  if (BB1V == BB2V)
1355  continue;
1356 
1357  // These values do not agree. Insert a select instruction before NT
1358  // that determines the right value.
1359  SelectInst *&SI = InsertedSelects[std::make_pair(BB1V, BB2V)];
1360  if (!SI)
1361  SI = cast<SelectInst>(
1362  Builder.CreateSelect(BI->getCondition(), BB1V, BB2V,
1363  BB1V->getName() + "." + BB2V->getName(), BI));
1364 
1365  // Make the PHI node use the select for all incoming values for BB1/BB2
1366  for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i)
1367  if (PN->getIncomingBlock(i) == BB1 || PN->getIncomingBlock(i) == BB2)
1368  PN->setIncomingValue(i, SI);
1369  }
1370  }
1371 
1372  // Update any PHI nodes in our new successors.
1373  for (BasicBlock *Succ : successors(BB1))
1374  AddPredecessorToBlock(Succ, BIParent, BB1);
1375 
1377  return true;
1378 }
1379 
1380 // All instructions in Insts belong to different blocks that all unconditionally
1381 // branch to a common successor. Analyze each instruction and return true if it
1382 // would be possible to sink them into their successor, creating one common
1383 // instruction instead. For every value that would be required to be provided by
1384 // PHI node (because an operand varies in each input block), add to PHIOperands.
1387  DenseMap<Instruction *, SmallVector<Value *, 4>> &PHIOperands) {
1388  // Prune out obviously bad instructions to move. Any non-store instruction
1389  // must have exactly one use, and we check later that use is by a single,
1390  // common PHI instruction in the successor.
1391  for (auto *I : Insts) {
1392  // These instructions may change or break semantics if moved.
1393  if (isa<PHINode>(I) || I->isEHPad() || isa<AllocaInst>(I) ||
1394  I->getType()->isTokenTy())
1395  return false;
1396 
1397  // Conservatively return false if I is an inline-asm instruction. Sinking
1398  // and merging inline-asm instructions can potentially create arguments
1399  // that cannot satisfy the inline-asm constraints.
1400  if (const auto *C = dyn_cast<CallInst>(I))
1401  if (C->isInlineAsm())
1402  return false;
1403 
1404  // Everything must have only one use too, apart from stores which
1405  // have no uses.
1406  if (!isa<StoreInst>(I) && !I->hasOneUse())
1407  return false;
1408  }
1409 
1410  const Instruction *I0 = Insts.front();
1411  for (auto *I : Insts)
1412  if (!I->isSameOperationAs(I0))
1413  return false;
1414 
1415  // All instructions in Insts are known to be the same opcode. If they aren't
1416  // stores, check the only user of each is a PHI or in the same block as the
1417  // instruction, because if a user is in the same block as an instruction
1418  // we're contemplating sinking, it must already be determined to be sinkable.
1419  if (!isa<StoreInst>(I0)) {
1420  auto *PNUse = dyn_cast<PHINode>(*I0->user_begin());
1421  auto *Succ = I0->getParent()->getTerminator()->getSuccessor(0);
1422  if (!all_of(Insts, [&PNUse,&Succ](const Instruction *I) -> bool {
1423  auto *U = cast<Instruction>(*I->user_begin());
1424  return (PNUse &&
1425  PNUse->getParent() == Succ &&
1426  PNUse->getIncomingValueForBlock(I->getParent()) == I) ||
1427  U->getParent() == I->getParent();
1428  }))
1429  return false;
1430  }
1431 
1432  // Because SROA can't handle speculating stores of selects, try not
1433  // to sink loads or stores of allocas when we'd have to create a PHI for
1434  // the address operand. Also, because it is likely that loads or stores
1435  // of allocas will disappear when Mem2Reg/SROA is run, don't sink them.
1436  // This can cause code churn which can have unintended consequences down
1437  // the line - see https://llvm.org/bugs/show_bug.cgi?id=30244.
1438  // FIXME: This is a workaround for a deficiency in SROA - see
1439  // https://llvm.org/bugs/show_bug.cgi?id=30188
1440  if (isa<StoreInst>(I0) && any_of(Insts, [](const Instruction *I) {
1441  return isa<AllocaInst>(I->getOperand(1));
1442  }))
1443  return false;
1444  if (isa<LoadInst>(I0) && any_of(Insts, [](const Instruction *I) {
1445  return isa<AllocaInst>(I->getOperand(0));
1446  }))
1447  return false;
1448 
1449  for (unsigned OI = 0, OE = I0->getNumOperands(); OI != OE; ++OI) {
1450  if (I0->getOperand(OI)->getType()->isTokenTy())
1451  // Don't touch any operand of token type.
1452  return false;
1453 
1454  auto SameAsI0 = [&I0, OI](const Instruction *I) {
1455  assert(I->getNumOperands() == I0->getNumOperands());
1456  return I->getOperand(OI) == I0->getOperand(OI);
1457  };
1458  if (!all_of(Insts, SameAsI0)) {
1459  if (!canReplaceOperandWithVariable(I0, OI))
1460  // We can't create a PHI from this GEP.
1461  return false;
1462  // Don't create indirect calls! The called value is the final operand.
1463  if ((isa<CallInst>(I0) || isa<InvokeInst>(I0)) && OI == OE - 1) {
1464  // FIXME: if the call was *already* indirect, we should do this.
1465  return false;
1466  }
1467  for (auto *I : Insts)
1468  PHIOperands[I].push_back(I->getOperand(OI));
1469  }
1470  }
1471  return true;
1472 }
1473 
1474 // Assuming canSinkLastInstruction(Blocks) has returned true, sink the last
1475 // instruction of every block in Blocks to their common successor, commoning
1476 // into one instruction.
1478  auto *BBEnd = Blocks[0]->getTerminator()->getSuccessor(0);
1479 
1480  // canSinkLastInstruction returning true guarantees that every block has at
1481  // least one non-terminator instruction.
1483  for (auto *BB : Blocks) {
1484  Instruction *I = BB->getTerminator();
1485  do {
1486  I = I->getPrevNode();
1487  } while (isa<DbgInfoIntrinsic>(I) && I != &BB->front());
1488  if (!isa<DbgInfoIntrinsic>(I))
1489  Insts.push_back(I);
1490  }
1491 
1492  // The only checking we need to do now is that all users of all instructions
1493  // are the same PHI node. canSinkLastInstruction should have checked this but
1494  // it is slightly over-aggressive - it gets confused by commutative instructions
1495  // so double-check it here.
1496  Instruction *I0 = Insts.front();
1497  if (!isa<StoreInst>(I0)) {
1498  auto *PNUse = dyn_cast<PHINode>(*I0->user_begin());
1499  if (!all_of(Insts, [&PNUse](const Instruction *I) -> bool {
1500  auto *U = cast<Instruction>(*I->user_begin());
1501  return U == PNUse;
1502  }))
1503  return false;
1504  }
1505 
1506  // We don't need to do any more checking here; canSinkLastInstruction should
1507  // have done it all for us.
1508  SmallVector<Value*, 4> NewOperands;
1509  for (unsigned O = 0, E = I0->getNumOperands(); O != E; ++O) {
1510  // This check is different to that in canSinkLastInstruction. There, we
1511  // cared about the global view once simplifycfg (and instcombine) have
1512  // completed - it takes into account PHIs that become trivially
1513  // simplifiable. However here we need a more local view; if an operand
1514  // differs we create a PHI and rely on instcombine to clean up the very
1515  // small mess we may make.
1516  bool NeedPHI = any_of(Insts, [&I0, O](const Instruction *I) {
1517  return I->getOperand(O) != I0->getOperand(O);
1518  });
1519  if (!NeedPHI) {
1520  NewOperands.push_back(I0->getOperand(O));
1521  continue;
1522  }
1523 
1524  // Create a new PHI in the successor block and populate it.
1525  auto *Op = I0->getOperand(O);
1526  assert(!Op->getType()->isTokenTy() && "Can't PHI tokens!");
1527  auto *PN = PHINode::Create(Op->getType(), Insts.size(),
1528  Op->getName() + ".sink", &BBEnd->front());
1529  for (auto *I : Insts)
1530  PN->addIncoming(I->getOperand(O), I->getParent());
1531  NewOperands.push_back(PN);
1532  }
1533 
1534  // Arbitrarily use I0 as the new "common" instruction; remap its operands
1535  // and move it to the start of the successor block.
1536  for (unsigned O = 0, E = I0->getNumOperands(); O != E; ++O)
1537  I0->getOperandUse(O).set(NewOperands[O]);
1538  I0->moveBefore(&*BBEnd->getFirstInsertionPt());
1539 
1540  // The debug location for the "common" instruction is the merged locations of
1541  // all the commoned instructions. We start with the original location of the
1542  // "common" instruction and iteratively merge each location in the loop below.
1543  const DILocation *Loc = I0->getDebugLoc();
1544 
1545  // Update metadata and IR flags, and merge debug locations.
1546  for (auto *I : Insts)
1547  if (I != I0) {
1548  Loc = DILocation::getMergedLocation(Loc, I->getDebugLoc());
1549  combineMetadataForCSE(I0, I);
1550  I0->andIRFlags(I);
1551  }
1552  if (!isa<CallInst>(I0))
1553  I0->setDebugLoc(Loc);
1554 
1555  if (!isa<StoreInst>(I0)) {
1556  // canSinkLastInstruction checked that all instructions were used by
1557  // one and only one PHI node. Find that now, RAUW it to our common
1558  // instruction and nuke it.
1559  assert(I0->hasOneUse());
1560  auto *PN = cast<PHINode>(*I0->user_begin());
1561  PN->replaceAllUsesWith(I0);
1562  PN->eraseFromParent();
1563  }
1564 
1565  // Finally nuke all instructions apart from the common instruction.
1566  for (auto *I : Insts)
1567  if (I != I0)
1568  I->eraseFromParent();
1569 
1570  return true;
1571 }
1572 
1573 namespace {
1574 
1575  // LockstepReverseIterator - Iterates through instructions
1576  // in a set of blocks in reverse order from the first non-terminator.
1577  // For example (assume all blocks have size n):
1578  // LockstepReverseIterator I([B1, B2, B3]);
1579  // *I-- = [B1[n], B2[n], B3[n]];
1580  // *I-- = [B1[n-1], B2[n-1], B3[n-1]];
1581  // *I-- = [B1[n-2], B2[n-2], B3[n-2]];
1582  // ...
1583  class LockstepReverseIterator {
1584  ArrayRef<BasicBlock*> Blocks;
1586  bool Fail;
1587  public:
1588  LockstepReverseIterator(ArrayRef<BasicBlock*> Blocks) :
1589  Blocks(Blocks) {
1590  reset();
1591  }
1592 
1593  void reset() {
1594  Fail = false;
1595  Insts.clear();
1596  for (auto *BB : Blocks) {
1597  Instruction *Inst = BB->getTerminator();
1598  for (Inst = Inst->getPrevNode(); Inst && isa<DbgInfoIntrinsic>(Inst);)
1599  Inst = Inst->getPrevNode();
1600  if (!Inst) {
1601  // Block wasn't big enough.
1602  Fail = true;
1603  return;
1604  }
1605  Insts.push_back(Inst);
1606  }
1607  }
1608 
1609  bool isValid() const {
1610  return !Fail;
1611  }
1612 
1613  void operator -- () {
1614  if (Fail)
1615  return;
1616  for (auto *&Inst : Insts) {
1617  for (Inst = Inst->getPrevNode(); Inst && isa<DbgInfoIntrinsic>(Inst);)
1618  Inst = Inst->getPrevNode();
1619  // Already at beginning of block.
1620  if (!Inst) {
1621  Fail = true;
1622  return;
1623  }
1624  }
1625  }
1626 
1628  return Insts;
1629  }
1630  };
1631 
1632 } // end anonymous namespace
1633 
1634 /// Given an unconditional branch that goes to BBEnd,
1635 /// check whether BBEnd has only two predecessors and the other predecessor
1636 /// ends with an unconditional branch. If it is true, sink any common code
1637 /// in the two predecessors to BBEnd.
1639  assert(BI1->isUnconditional());
1640  BasicBlock *BBEnd = BI1->getSuccessor(0);
1641 
1642  // We support two situations:
1643  // (1) all incoming arcs are unconditional
1644  // (2) one incoming arc is conditional
1645  //
1646  // (2) is very common in switch defaults and
1647  // else-if patterns;
1648  //
1649  // if (a) f(1);
1650  // else if (b) f(2);
1651  //
1652  // produces:
1653  //
1654  // [if]
1655  // / \
1656  // [f(1)] [if]
1657  // | | \
1658  // | | |
1659  // | [f(2)]|
1660  // \ | /
1661  // [ end ]
1662  //
1663  // [end] has two unconditional predecessor arcs and one conditional. The
1664  // conditional refers to the implicit empty 'else' arc. This conditional
1665  // arc can also be caused by an empty default block in a switch.
1666  //
1667  // In this case, we attempt to sink code from all *unconditional* arcs.
1668  // If we can sink instructions from these arcs (determined during the scan
1669  // phase below) we insert a common successor for all unconditional arcs and
1670  // connect that to [end], to enable sinking:
1671  //
1672  // [if]
1673  // / \
1674  // [x(1)] [if]
1675  // | | \
1676  // | | \
1677  // | [x(2)] |
1678  // \ / |
1679  // [sink.split] |
1680  // \ /
1681  // [ end ]
1682  //
1683  SmallVector<BasicBlock*,4> UnconditionalPreds;
1684  Instruction *Cond = nullptr;
1685  for (auto *B : predecessors(BBEnd)) {
1686  auto *T = B->getTerminator();
1687  if (isa<BranchInst>(T) && cast<BranchInst>(T)->isUnconditional())
1688  UnconditionalPreds.push_back(B);
1689  else if ((isa<BranchInst>(T) || isa<SwitchInst>(T)) && !Cond)
1690  Cond = T;
1691  else
1692  return false;
1693  }
1694  if (UnconditionalPreds.size() < 2)
1695  return false;
1696 
1697  bool Changed = false;
1698  // We take a two-step approach to tail sinking. First we scan from the end of
1699  // each block upwards in lockstep. If the n'th instruction from the end of each
1700  // block can be sunk, those instructions are added to ValuesToSink and we
1701  // carry on. If we can sink an instruction but need to PHI-merge some operands
1702  // (because they're not identical in each instruction) we add these to
1703  // PHIOperands.
1704  unsigned ScanIdx = 0;
1705  SmallPtrSet<Value*,4> InstructionsToSink;
1707  LockstepReverseIterator LRI(UnconditionalPreds);
1708  while (LRI.isValid() &&
1709  canSinkInstructions(*LRI, PHIOperands)) {
1710  DEBUG(dbgs() << "SINK: instruction can be sunk: " << *(*LRI)[0] << "\n");
1711  InstructionsToSink.insert((*LRI).begin(), (*LRI).end());
1712  ++ScanIdx;
1713  --LRI;
1714  }
1715 
1716  auto ProfitableToSinkInstruction = [&](LockstepReverseIterator &LRI) {
1717  unsigned NumPHIdValues = 0;
1718  for (auto *I : *LRI)
1719  for (auto *V : PHIOperands[I])
1720  if (InstructionsToSink.count(V) == 0)
1721  ++NumPHIdValues;
1722  DEBUG(dbgs() << "SINK: #phid values: " << NumPHIdValues << "\n");
1723  unsigned NumPHIInsts = NumPHIdValues / UnconditionalPreds.size();
1724  if ((NumPHIdValues % UnconditionalPreds.size()) != 0)
1725  NumPHIInsts++;
1726 
1727  return NumPHIInsts <= 1;
1728  };
1729 
1730  if (ScanIdx > 0 && Cond) {
1731  // Check if we would actually sink anything first! This mutates the CFG and
1732  // adds an extra block. The goal in doing this is to allow instructions that
1733  // couldn't be sunk before to be sunk - obviously, speculatable instructions
1734  // (such as trunc, add) can be sunk and predicated already. So we check that
1735  // we're going to sink at least one non-speculatable instruction.
1736  LRI.reset();
1737  unsigned Idx = 0;
1738  bool Profitable = false;
1739  while (ProfitableToSinkInstruction(LRI) && Idx < ScanIdx) {
1740  if (!isSafeToSpeculativelyExecute((*LRI)[0])) {
1741  Profitable = true;
1742  break;
1743  }
1744  --LRI;
1745  ++Idx;
1746  }
1747  if (!Profitable)
1748  return false;
1749 
1750  DEBUG(dbgs() << "SINK: Splitting edge\n");
1751  // We have a conditional edge and we're going to sink some instructions.
1752  // Insert a new block postdominating all blocks we're going to sink from.
1753  if (!SplitBlockPredecessors(BI1->getSuccessor(0), UnconditionalPreds,
1754  ".sink.split"))
1755  // Edges couldn't be split.
1756  return false;
1757  Changed = true;
1758  }
1759 
1760  // Now that we've analyzed all potential sinking candidates, perform the
1761  // actual sink. We iteratively sink the last non-terminator of the source
1762  // blocks into their common successor unless doing so would require too
1763  // many PHI instructions to be generated (currently only one PHI is allowed
1764  // per sunk instruction).
1765  //
1766  // We can use InstructionsToSink to discount values needing PHI-merging that will
1767  // actually be sunk in a later iteration. This allows us to be more
1768  // aggressive in what we sink. This does allow a false positive where we
1769  // sink presuming a later value will also be sunk, but stop half way through
1770  // and never actually sink it which means we produce more PHIs than intended.
1771  // This is unlikely in practice though.
1772  for (unsigned SinkIdx = 0; SinkIdx != ScanIdx; ++SinkIdx) {
1773  DEBUG(dbgs() << "SINK: Sink: "
1774  << *UnconditionalPreds[0]->getTerminator()->getPrevNode()
1775  << "\n");
1776 
1777  // Because we've sunk every instruction in turn, the current instruction to
1778  // sink is always at index 0.
1779  LRI.reset();
1780  if (!ProfitableToSinkInstruction(LRI)) {
1781  // Too many PHIs would be created.
1782  DEBUG(dbgs() << "SINK: stopping here, too many PHIs would be created!\n");
1783  break;
1784  }
1785 
1786  if (!sinkLastInstruction(UnconditionalPreds))
1787  return Changed;
1788  NumSinkCommons++;
1789  Changed = true;
1790  }
1791  return Changed;
1792 }
1793 
1794 /// \brief Determine if we can hoist sink a sole store instruction out of a
1795 /// conditional block.
1796 ///
1797 /// We are looking for code like the following:
1798 /// BrBB:
1799 /// store i32 %add, i32* %arrayidx2
1800 /// ... // No other stores or function calls (we could be calling a memory
1801 /// ... // function).
1802 /// %cmp = icmp ult %x, %y
1803 /// br i1 %cmp, label %EndBB, label %ThenBB
1804 /// ThenBB:
1805 /// store i32 %add5, i32* %arrayidx2
1806 /// br label EndBB
1807 /// EndBB:
1808 /// ...
1809 /// We are going to transform this into:
1810 /// BrBB:
1811 /// store i32 %add, i32* %arrayidx2
1812 /// ... //
1813 /// %cmp = icmp ult %x, %y
1814 /// %add.add5 = select i1 %cmp, i32 %add, %add5
1815 /// store i32 %add.add5, i32* %arrayidx2
1816 /// ...
1817 ///
1818 /// \return The pointer to the value of the previous store if the store can be
1819 /// hoisted into the predecessor block. 0 otherwise.
1821  BasicBlock *StoreBB, BasicBlock *EndBB) {
1822  StoreInst *StoreToHoist = dyn_cast<StoreInst>(I);
1823  if (!StoreToHoist)
1824  return nullptr;
1825 
1826  // Volatile or atomic.
1827  if (!StoreToHoist->isSimple())
1828  return nullptr;
1829 
1830  Value *StorePtr = StoreToHoist->getPointerOperand();
1831 
1832  // Look for a store to the same pointer in BrBB.
1833  unsigned MaxNumInstToLookAt = 9;
1834  for (Instruction &CurI : reverse(*BrBB)) {
1835  if (!MaxNumInstToLookAt)
1836  break;
1837  // Skip debug info.
1838  if (isa<DbgInfoIntrinsic>(CurI))
1839  continue;
1840  --MaxNumInstToLookAt;
1841 
1842  // Could be calling an instruction that affects memory like free().
1843  if (CurI.mayHaveSideEffects() && !isa<StoreInst>(CurI))
1844  return nullptr;
1845 
1846  if (auto *SI = dyn_cast<StoreInst>(&CurI)) {
1847  // Found the previous store make sure it stores to the same location.
1848  if (SI->getPointerOperand() == StorePtr)
1849  // Found the previous store, return its value operand.
1850  return SI->getValueOperand();
1851  return nullptr; // Unknown store.
1852  }
1853  }
1854 
1855  return nullptr;
1856 }
1857 
1858 /// \brief Speculate a conditional basic block flattening the CFG.
1859 ///
1860 /// Note that this is a very risky transform currently. Speculating
1861 /// instructions like this is most often not desirable. Instead, there is an MI
1862 /// pass which can do it with full awareness of the resource constraints.
1863 /// However, some cases are "obvious" and we should do directly. An example of
1864 /// this is speculating a single, reasonably cheap instruction.
1865 ///
1866 /// There is only one distinct advantage to flattening the CFG at the IR level:
1867 /// it makes very common but simplistic optimizations such as are common in
1868 /// instcombine and the DAG combiner more powerful by removing CFG edges and
1869 /// modeling their effects with easier to reason about SSA value graphs.
1870 ///
1871 ///
1872 /// An illustration of this transform is turning this IR:
1873 /// \code
1874 /// BB:
1875 /// %cmp = icmp ult %x, %y
1876 /// br i1 %cmp, label %EndBB, label %ThenBB
1877 /// ThenBB:
1878 /// %sub = sub %x, %y
1879 /// br label BB2
1880 /// EndBB:
1881 /// %phi = phi [ %sub, %ThenBB ], [ 0, %EndBB ]
1882 /// ...
1883 /// \endcode
1884 ///
1885 /// Into this IR:
1886 /// \code
1887 /// BB:
1888 /// %cmp = icmp ult %x, %y
1889 /// %sub = sub %x, %y
1890 /// %cond = select i1 %cmp, 0, %sub
1891 /// ...
1892 /// \endcode
1893 ///
1894 /// \returns true if the conditional block is removed.
1896  const TargetTransformInfo &TTI) {
1897  // Be conservative for now. FP select instruction can often be expensive.
1898  Value *BrCond = BI->getCondition();
1899  if (isa<FCmpInst>(BrCond))
1900  return false;
1901 
1902  BasicBlock *BB = BI->getParent();
1903  BasicBlock *EndBB = ThenBB->getTerminator()->getSuccessor(0);
1904 
1905  // If ThenBB is actually on the false edge of the conditional branch, remember
1906  // to swap the select operands later.
1907  bool Invert = false;
1908  if (ThenBB != BI->getSuccessor(0)) {
1909  assert(ThenBB == BI->getSuccessor(1) && "No edge from 'if' block?");
1910  Invert = true;
1911  }
1912  assert(EndBB == BI->getSuccessor(!Invert) && "No edge from to end block");
1913 
1914  // Keep a count of how many times instructions are used within CondBB when
1915  // they are candidates for sinking into CondBB. Specifically:
1916  // - They are defined in BB, and
1917  // - They have no side effects, and
1918  // - All of their uses are in CondBB.
1919  SmallDenseMap<Instruction *, unsigned, 4> SinkCandidateUseCounts;
1920 
1921  unsigned SpeculationCost = 0;
1922  Value *SpeculatedStoreValue = nullptr;
1923  StoreInst *SpeculatedStore = nullptr;
1924  for (BasicBlock::iterator BBI = ThenBB->begin(),
1925  BBE = std::prev(ThenBB->end());
1926  BBI != BBE; ++BBI) {
1927  Instruction *I = &*BBI;
1928  // Skip debug info.
1929  if (isa<DbgInfoIntrinsic>(I))
1930  continue;
1931 
1932  // Only speculatively execute a single instruction (not counting the
1933  // terminator) for now.
1934  ++SpeculationCost;
1935  if (SpeculationCost > 1)
1936  return false;
1937 
1938  // Don't hoist the instruction if it's unsafe or expensive.
1939  if (!isSafeToSpeculativelyExecute(I) &&
1940  !(HoistCondStores && (SpeculatedStoreValue = isSafeToSpeculateStore(
1941  I, BB, ThenBB, EndBB))))
1942  return false;
1943  if (!SpeculatedStoreValue &&
1944  ComputeSpeculationCost(I, TTI) >
1946  return false;
1947 
1948  // Store the store speculation candidate.
1949  if (SpeculatedStoreValue)
1950  SpeculatedStore = cast<StoreInst>(I);
1951 
1952  // Do not hoist the instruction if any of its operands are defined but not
1953  // used in BB. The transformation will prevent the operand from
1954  // being sunk into the use block.
1955  for (User::op_iterator i = I->op_begin(), e = I->op_end(); i != e; ++i) {
1956  Instruction *OpI = dyn_cast<Instruction>(*i);
1957  if (!OpI || OpI->getParent() != BB || OpI->mayHaveSideEffects())
1958  continue; // Not a candidate for sinking.
1959 
1960  ++SinkCandidateUseCounts[OpI];
1961  }
1962  }
1963 
1964  // Consider any sink candidates which are only used in CondBB as costs for
1965  // speculation. Note, while we iterate over a DenseMap here, we are summing
1966  // and so iteration order isn't significant.
1968  I = SinkCandidateUseCounts.begin(),
1969  E = SinkCandidateUseCounts.end();
1970  I != E; ++I)
1971  if (I->first->getNumUses() == I->second) {
1972  ++SpeculationCost;
1973  if (SpeculationCost > 1)
1974  return false;
1975  }
1976 
1977  // Check that the PHI nodes can be converted to selects.
1978  bool HaveRewritablePHIs = false;
1979  for (BasicBlock::iterator I = EndBB->begin();
1980  PHINode *PN = dyn_cast<PHINode>(I); ++I) {
1981  Value *OrigV = PN->getIncomingValueForBlock(BB);
1982  Value *ThenV = PN->getIncomingValueForBlock(ThenBB);
1983 
1984  // FIXME: Try to remove some of the duplication with HoistThenElseCodeToIf.
1985  // Skip PHIs which are trivial.
1986  if (ThenV == OrigV)
1987  continue;
1988 
1989  // Don't convert to selects if we could remove undefined behavior instead.
1990  if (passingValueIsAlwaysUndefined(OrigV, PN) ||
1991  passingValueIsAlwaysUndefined(ThenV, PN))
1992  return false;
1993 
1994  HaveRewritablePHIs = true;
1995  ConstantExpr *OrigCE = dyn_cast<ConstantExpr>(OrigV);
1996  ConstantExpr *ThenCE = dyn_cast<ConstantExpr>(ThenV);
1997  if (!OrigCE && !ThenCE)
1998  continue; // Known safe and cheap.
1999 
2000  if ((ThenCE && !isSafeToSpeculativelyExecute(ThenCE)) ||
2001  (OrigCE && !isSafeToSpeculativelyExecute(OrigCE)))
2002  return false;
2003  unsigned OrigCost = OrigCE ? ComputeSpeculationCost(OrigCE, TTI) : 0;
2004  unsigned ThenCost = ThenCE ? ComputeSpeculationCost(ThenCE, TTI) : 0;
2005  unsigned MaxCost =
2007  if (OrigCost + ThenCost > MaxCost)
2008  return false;
2009 
2010  // Account for the cost of an unfolded ConstantExpr which could end up
2011  // getting expanded into Instructions.
2012  // FIXME: This doesn't account for how many operations are combined in the
2013  // constant expression.
2014  ++SpeculationCost;
2015  if (SpeculationCost > 1)
2016  return false;
2017  }
2018 
2019  // If there are no PHIs to process, bail early. This helps ensure idempotence
2020  // as well.
2021  if (!HaveRewritablePHIs && !(HoistCondStores && SpeculatedStoreValue))
2022  return false;
2023 
2024  // If we get here, we can hoist the instruction and if-convert.
2025  DEBUG(dbgs() << "SPECULATIVELY EXECUTING BB" << *ThenBB << "\n";);
2026 
2027  // Insert a select of the value of the speculated store.
2028  if (SpeculatedStoreValue) {
2029  IRBuilder<NoFolder> Builder(BI);
2030  Value *TrueV = SpeculatedStore->getValueOperand();
2031  Value *FalseV = SpeculatedStoreValue;
2032  if (Invert)
2033  std::swap(TrueV, FalseV);
2034  Value *S = Builder.CreateSelect(
2035  BrCond, TrueV, FalseV, TrueV->getName() + "." + FalseV->getName(), BI);
2036  SpeculatedStore->setOperand(0, S);
2037  SpeculatedStore->setDebugLoc(
2039  BI->getDebugLoc(), SpeculatedStore->getDebugLoc()));
2040  }
2041 
2042  // Metadata can be dependent on the condition we are hoisting above.
2043  // Conservatively strip all metadata on the instruction.
2044  for (auto &I : *ThenBB)
2046 
2047  // Hoist the instructions.
2048  BB->getInstList().splice(BI->getIterator(), ThenBB->getInstList(),
2049  ThenBB->begin(), std::prev(ThenBB->end()));
2050 
2051  // Insert selects and rewrite the PHI operands.
2052  IRBuilder<NoFolder> Builder(BI);
2053  for (BasicBlock::iterator I = EndBB->begin();
2054  PHINode *PN = dyn_cast<PHINode>(I); ++I) {
2055  unsigned OrigI = PN->getBasicBlockIndex(BB);
2056  unsigned ThenI = PN->getBasicBlockIndex(ThenBB);
2057  Value *OrigV = PN->getIncomingValue(OrigI);
2058  Value *ThenV = PN->getIncomingValue(ThenI);
2059 
2060  // Skip PHIs which are trivial.
2061  if (OrigV == ThenV)
2062  continue;
2063 
2064  // Create a select whose true value is the speculatively executed value and
2065  // false value is the preexisting value. Swap them if the branch
2066  // destinations were inverted.
2067  Value *TrueV = ThenV, *FalseV = OrigV;
2068  if (Invert)
2069  std::swap(TrueV, FalseV);
2070  Value *V = Builder.CreateSelect(
2071  BrCond, TrueV, FalseV, TrueV->getName() + "." + FalseV->getName(), BI);
2072  PN->setIncomingValue(OrigI, V);
2073  PN->setIncomingValue(ThenI, V);
2074  }
2075 
2076  ++NumSpeculations;
2077  return true;
2078 }
2079 
2080 /// Return true if we can thread a branch across this block.
2082  BranchInst *BI = cast<BranchInst>(BB->getTerminator());
2083  unsigned Size = 0;
2084 
2085  for (BasicBlock::iterator BBI = BB->begin(); &*BBI != BI; ++BBI) {
2086  if (isa<DbgInfoIntrinsic>(BBI))
2087  continue;
2088  if (Size > 10)
2089  return false; // Don't clone large BB's.
2090  ++Size;
2091 
2092  // We can only support instructions that do not define values that are
2093  // live outside of the current basic block.
2094  for (User *U : BBI->users()) {
2095  Instruction *UI = cast<Instruction>(U);
2096  if (UI->getParent() != BB || isa<PHINode>(UI))
2097  return false;
2098  }
2099 
2100  // Looks ok, continue checking.
2101  }
2102 
2103  return true;
2104 }
2105 
2106 /// If we have a conditional branch on a PHI node value that is defined in the
2107 /// same block as the branch and if any PHI entries are constants, thread edges
2108 /// corresponding to that entry to be branches to their ultimate destination.
2109 static bool FoldCondBranchOnPHI(BranchInst *BI, const DataLayout &DL,
2110  AssumptionCache *AC) {
2111  BasicBlock *BB = BI->getParent();
2112  PHINode *PN = dyn_cast<PHINode>(BI->getCondition());
2113  // NOTE: we currently cannot transform this case if the PHI node is used
2114  // outside of the block.
2115  if (!PN || PN->getParent() != BB || !PN->hasOneUse())
2116  return false;
2117 
2118  // Degenerate case of a single entry PHI.
2119  if (PN->getNumIncomingValues() == 1) {
2121  return true;
2122  }
2123 
2124  // Now we know that this block has multiple preds and two succs.
2126  return false;
2127 
2128  // Can't fold blocks that contain noduplicate or convergent calls.
2129  if (any_of(*BB, [](const Instruction &I) {
2130  const CallInst *CI = dyn_cast<CallInst>(&I);
2131  return CI && (CI->cannotDuplicate() || CI->isConvergent());
2132  }))
2133  return false;
2134 
2135  // Okay, this is a simple enough basic block. See if any phi values are
2136  // constants.
2137  for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
2139  if (!CB || !CB->getType()->isIntegerTy(1))
2140  continue;
2141 
2142  // Okay, we now know that all edges from PredBB should be revectored to
2143  // branch to RealDest.
2144  BasicBlock *PredBB = PN->getIncomingBlock(i);
2145  BasicBlock *RealDest = BI->getSuccessor(!CB->getZExtValue());
2146 
2147  if (RealDest == BB)
2148  continue; // Skip self loops.
2149  // Skip if the predecessor's terminator is an indirect branch.
2150  if (isa<IndirectBrInst>(PredBB->getTerminator()))
2151  continue;
2152 
2153  // The dest block might have PHI nodes, other predecessors and other
2154  // difficult cases. Instead of being smart about this, just insert a new
2155  // block that jumps to the destination block, effectively splitting
2156  // the edge we are about to create.
2157  BasicBlock *EdgeBB =
2158  BasicBlock::Create(BB->getContext(), RealDest->getName() + ".critedge",
2159  RealDest->getParent(), RealDest);
2160  BranchInst::Create(RealDest, EdgeBB);
2161 
2162  // Update PHI nodes.
2163  AddPredecessorToBlock(RealDest, EdgeBB, BB);
2164 
2165  // BB may have instructions that are being threaded over. Clone these
2166  // instructions into EdgeBB. We know that there will be no uses of the
2167  // cloned instructions outside of EdgeBB.
2168  BasicBlock::iterator InsertPt = EdgeBB->begin();
2169  DenseMap<Value *, Value *> TranslateMap; // Track translated values.
2170  for (BasicBlock::iterator BBI = BB->begin(); &*BBI != BI; ++BBI) {
2171  if (PHINode *PN = dyn_cast<PHINode>(BBI)) {
2172  TranslateMap[PN] = PN->getIncomingValueForBlock(PredBB);
2173  continue;
2174  }
2175  // Clone the instruction.
2176  Instruction *N = BBI->clone();
2177  if (BBI->hasName())
2178  N->setName(BBI->getName() + ".c");
2179 
2180  // Update operands due to translation.
2181  for (User::op_iterator i = N->op_begin(), e = N->op_end(); i != e; ++i) {
2182  DenseMap<Value *, Value *>::iterator PI = TranslateMap.find(*i);
2183  if (PI != TranslateMap.end())
2184  *i = PI->second;
2185  }
2186 
2187  // Check for trivial simplification.
2188  if (Value *V = SimplifyInstruction(N, {DL, nullptr, nullptr, AC})) {
2189  if (!BBI->use_empty())
2190  TranslateMap[&*BBI] = V;
2191  if (!N->mayHaveSideEffects()) {
2192  N->deleteValue(); // Instruction folded away, don't need actual inst
2193  N = nullptr;
2194  }
2195  } else {
2196  if (!BBI->use_empty())
2197  TranslateMap[&*BBI] = N;
2198  }
2199  // Insert the new instruction into its new home.
2200  if (N)
2201  EdgeBB->getInstList().insert(InsertPt, N);
2202 
2203  // Register the new instruction with the assumption cache if necessary.
2204  if (auto *II = dyn_cast_or_null<IntrinsicInst>(N))
2205  if (II->getIntrinsicID() == Intrinsic::assume)
2206  AC->registerAssumption(II);
2207  }
2208 
2209  // Loop over all of the edges from PredBB to BB, changing them to branch
2210  // to EdgeBB instead.
2211  TerminatorInst *PredBBTI = PredBB->getTerminator();
2212  for (unsigned i = 0, e = PredBBTI->getNumSuccessors(); i != e; ++i)
2213  if (PredBBTI->getSuccessor(i) == BB) {
2214  BB->removePredecessor(PredBB);
2215  PredBBTI->setSuccessor(i, EdgeBB);
2216  }
2217 
2218  // Recurse, simplifying any other constants.
2219  return FoldCondBranchOnPHI(BI, DL, AC) | true;
2220  }
2221 
2222  return false;
2223 }
2224 
2225 /// Given a BB that starts with the specified two-entry PHI node,
2226 /// see if we can eliminate it.
2228  const DataLayout &DL) {
2229  // Ok, this is a two entry PHI node. Check to see if this is a simple "if
2230  // statement", which has a very simple dominance structure. Basically, we
2231  // are trying to find the condition that is being branched on, which
2232  // subsequently causes this merge to happen. We really want control
2233  // dependence information for this check, but simplifycfg can't keep it up
2234  // to date, and this catches most of the cases we care about anyway.
2235  BasicBlock *BB = PN->getParent();
2236  BasicBlock *IfTrue, *IfFalse;
2237  Value *IfCond = GetIfCondition(BB, IfTrue, IfFalse);
2238  if (!IfCond ||
2239  // Don't bother if the branch will be constant folded trivially.
2240  isa<ConstantInt>(IfCond))
2241  return false;
2242 
2243  // Okay, we found that we can merge this two-entry phi node into a select.
2244  // Doing so would require us to fold *all* two entry phi nodes in this block.
2245  // At some point this becomes non-profitable (particularly if the target
2246  // doesn't support cmov's). Only do this transformation if there are two or
2247  // fewer PHI nodes in this block.
2248  unsigned NumPhis = 0;
2249  for (BasicBlock::iterator I = BB->begin(); isa<PHINode>(I); ++NumPhis, ++I)
2250  if (NumPhis > 2)
2251  return false;
2252 
2253  // Loop over the PHI's seeing if we can promote them all to select
2254  // instructions. While we are at it, keep track of the instructions
2255  // that need to be moved to the dominating block.
2256  SmallPtrSet<Instruction *, 4> AggressiveInsts;
2257  unsigned MaxCostVal0 = PHINodeFoldingThreshold,
2258  MaxCostVal1 = PHINodeFoldingThreshold;
2259  MaxCostVal0 *= TargetTransformInfo::TCC_Basic;
2260  MaxCostVal1 *= TargetTransformInfo::TCC_Basic;
2261 
2262  for (BasicBlock::iterator II = BB->begin(); isa<PHINode>(II);) {
2263  PHINode *PN = cast<PHINode>(II++);
2264  if (Value *V = SimplifyInstruction(PN, {DL, PN})) {
2265  PN->replaceAllUsesWith(V);
2266  PN->eraseFromParent();
2267  continue;
2268  }
2269 
2270  if (!DominatesMergePoint(PN->getIncomingValue(0), BB, &AggressiveInsts,
2271  MaxCostVal0, TTI) ||
2272  !DominatesMergePoint(PN->getIncomingValue(1), BB, &AggressiveInsts,
2273  MaxCostVal1, TTI))
2274  return false;
2275  }
2276 
2277  // If we folded the first phi, PN dangles at this point. Refresh it. If
2278  // we ran out of PHIs then we simplified them all.
2279  PN = dyn_cast<PHINode>(BB->begin());
2280  if (!PN)
2281  return true;
2282 
2283  // Don't fold i1 branches on PHIs which contain binary operators. These can
2284  // often be turned into switches and other things.
2285  if (PN->getType()->isIntegerTy(1) &&
2286  (isa<BinaryOperator>(PN->getIncomingValue(0)) ||
2287  isa<BinaryOperator>(PN->getIncomingValue(1)) ||
2288  isa<BinaryOperator>(IfCond)))
2289  return false;
2290 
2291  // If all PHI nodes are promotable, check to make sure that all instructions
2292  // in the predecessor blocks can be promoted as well. If not, we won't be able
2293  // to get rid of the control flow, so it's not worth promoting to select
2294  // instructions.
2295  BasicBlock *DomBlock = nullptr;
2296  BasicBlock *IfBlock1 = PN->getIncomingBlock(0);
2297  BasicBlock *IfBlock2 = PN->getIncomingBlock(1);
2298  if (cast<BranchInst>(IfBlock1->getTerminator())->isConditional()) {
2299  IfBlock1 = nullptr;
2300  } else {
2301  DomBlock = *pred_begin(IfBlock1);
2302  for (BasicBlock::iterator I = IfBlock1->begin(); !isa<TerminatorInst>(I);
2303  ++I)
2304  if (!AggressiveInsts.count(&*I) && !isa<DbgInfoIntrinsic>(I)) {
2305  // This is not an aggressive instruction that we can promote.
2306  // Because of this, we won't be able to get rid of the control flow, so
2307  // the xform is not worth it.
2308  return false;
2309  }
2310  }
2311 
2312  if (cast<BranchInst>(IfBlock2->getTerminator())->isConditional()) {
2313  IfBlock2 = nullptr;
2314  } else {
2315  DomBlock = *pred_begin(IfBlock2);
2316  for (BasicBlock::iterator I = IfBlock2->begin(); !isa<TerminatorInst>(I);
2317  ++I)
2318  if (!AggressiveInsts.count(&*I) && !isa<DbgInfoIntrinsic>(I)) {
2319  // This is not an aggressive instruction that we can promote.
2320  // Because of this, we won't be able to get rid of the control flow, so
2321  // the xform is not worth it.
2322  return false;
2323  }
2324  }
2325 
2326  DEBUG(dbgs() << "FOUND IF CONDITION! " << *IfCond << " T: "
2327  << IfTrue->getName() << " F: " << IfFalse->getName() << "\n");
2328 
2329  // If we can still promote the PHI nodes after this gauntlet of tests,
2330  // do all of the PHI's now.
2331  Instruction *InsertPt = DomBlock->getTerminator();
2332  IRBuilder<NoFolder> Builder(InsertPt);
2333 
2334  // Move all 'aggressive' instructions, which are defined in the
2335  // conditional parts of the if's up to the dominating block.
2336  if (IfBlock1) {
2337  for (auto &I : *IfBlock1)
2339  DomBlock->getInstList().splice(InsertPt->getIterator(),
2340  IfBlock1->getInstList(), IfBlock1->begin(),
2341  IfBlock1->getTerminator()->getIterator());
2342  }
2343  if (IfBlock2) {
2344  for (auto &I : *IfBlock2)
2346  DomBlock->getInstList().splice(InsertPt->getIterator(),
2347  IfBlock2->getInstList(), IfBlock2->begin(),
2348  IfBlock2->getTerminator()->getIterator());
2349  }
2350 
2351  while (PHINode *PN = dyn_cast<PHINode>(BB->begin())) {
2352  // Change the PHI node into a select instruction.
2353  Value *TrueVal = PN->getIncomingValue(PN->getIncomingBlock(0) == IfFalse);
2354  Value *FalseVal = PN->getIncomingValue(PN->getIncomingBlock(0) == IfTrue);
2355 
2356  Value *Sel = Builder.CreateSelect(IfCond, TrueVal, FalseVal, "", InsertPt);
2357  PN->replaceAllUsesWith(Sel);
2358  Sel->takeName(PN);
2359  PN->eraseFromParent();
2360  }
2361 
2362  // At this point, IfBlock1 and IfBlock2 are both empty, so our if statement
2363  // has been flattened. Change DomBlock to jump directly to our new block to
2364  // avoid other simplifycfg's kicking in on the diamond.
2365  TerminatorInst *OldTI = DomBlock->getTerminator();
2366  Builder.SetInsertPoint(OldTI);
2367  Builder.CreateBr(BB);
2368  OldTI->eraseFromParent();
2369  return true;
2370 }
2371 
2372 /// If we found a conditional branch that goes to two returning blocks,
2373 /// try to merge them together into one return,
2374 /// introducing a select if the return values disagree.
2376  IRBuilder<> &Builder) {
2377  assert(BI->isConditional() && "Must be a conditional branch");
2378  BasicBlock *TrueSucc = BI->getSuccessor(0);
2379  BasicBlock *FalseSucc = BI->getSuccessor(1);
2380  ReturnInst *TrueRet = cast<ReturnInst>(TrueSucc->getTerminator());
2381  ReturnInst *FalseRet = cast<ReturnInst>(FalseSucc->getTerminator());
2382 
2383  // Check to ensure both blocks are empty (just a return) or optionally empty
2384  // with PHI nodes. If there are other instructions, merging would cause extra
2385  // computation on one path or the other.
2386  if (!TrueSucc->getFirstNonPHIOrDbg()->isTerminator())
2387  return false;
2388  if (!FalseSucc->getFirstNonPHIOrDbg()->isTerminator())
2389  return false;
2390 
2391  Builder.SetInsertPoint(BI);
2392  // Okay, we found a branch that is going to two return nodes. If
2393  // there is no return value for this function, just change the
2394  // branch into a return.
2395  if (FalseRet->getNumOperands() == 0) {
2396  TrueSucc->removePredecessor(BI->getParent());
2397  FalseSucc->removePredecessor(BI->getParent());
2398  Builder.CreateRetVoid();
2400  return true;
2401  }
2402 
2403  // Otherwise, figure out what the true and false return values are
2404  // so we can insert a new select instruction.
2405  Value *TrueValue = TrueRet->getReturnValue();
2406  Value *FalseValue = FalseRet->getReturnValue();
2407 
2408  // Unwrap any PHI nodes in the return blocks.
2409  if (PHINode *TVPN = dyn_cast_or_null<PHINode>(TrueValue))
2410  if (TVPN->getParent() == TrueSucc)
2411  TrueValue = TVPN->getIncomingValueForBlock(BI->getParent());
2412  if (PHINode *FVPN = dyn_cast_or_null<PHINode>(FalseValue))
2413  if (FVPN->getParent() == FalseSucc)
2414  FalseValue = FVPN->getIncomingValueForBlock(BI->getParent());
2415 
2416  // In order for this transformation to be safe, we must be able to
2417  // unconditionally execute both operands to the return. This is
2418  // normally the case, but we could have a potentially-trapping
2419  // constant expression that prevents this transformation from being
2420  // safe.
2421  if (ConstantExpr *TCV = dyn_cast_or_null<ConstantExpr>(TrueValue))
2422  if (TCV->canTrap())
2423  return false;
2424  if (ConstantExpr *FCV = dyn_cast_or_null<ConstantExpr>(FalseValue))
2425  if (FCV->canTrap())
2426  return false;
2427 
2428  // Okay, we collected all the mapped values and checked them for sanity, and
2429  // defined to really do this transformation. First, update the CFG.
2430  TrueSucc->removePredecessor(BI->getParent());
2431  FalseSucc->removePredecessor(BI->getParent());
2432 
2433  // Insert select instructions where needed.
2434  Value *BrCond = BI->getCondition();
2435  if (TrueValue) {
2436  // Insert a select if the results differ.
2437  if (TrueValue == FalseValue || isa<UndefValue>(FalseValue)) {
2438  } else if (isa<UndefValue>(TrueValue)) {
2439  TrueValue = FalseValue;
2440  } else {
2441  TrueValue =
2442  Builder.CreateSelect(BrCond, TrueValue, FalseValue, "retval", BI);
2443  }
2444  }
2445 
2446  Value *RI =
2447  !TrueValue ? Builder.CreateRetVoid() : Builder.CreateRet(TrueValue);
2448 
2449  (void)RI;
2450 
2451  DEBUG(dbgs() << "\nCHANGING BRANCH TO TWO RETURNS INTO SELECT:"
2452  << "\n " << *BI << "NewRet = " << *RI
2453  << "TRUEBLOCK: " << *TrueSucc << "FALSEBLOCK: " << *FalseSucc);
2454 
2456 
2457  return true;
2458 }
2459 
2460 /// Return true if the given instruction is available
2461 /// in its predecessor block. If yes, the instruction will be removed.
2463  if (!isa<BinaryOperator>(Inst) && !isa<CmpInst>(Inst))
2464  return false;
2465  for (Instruction &I : *PB) {
2466  Instruction *PBI = &I;
2467  // Check whether Inst and PBI generate the same value.
2468  if (Inst->isIdenticalTo(PBI)) {
2469  Inst->replaceAllUsesWith(PBI);
2470  Inst->eraseFromParent();
2471  return true;
2472  }
2473  }
2474  return false;
2475 }
2476 
2477 /// Return true if either PBI or BI has branch weight available, and store
2478 /// the weights in {Pred|Succ}{True|False}Weight. If one of PBI and BI does
2479 /// not have branch weight, use 1:1 as its weight.
2481  uint64_t &PredTrueWeight,
2482  uint64_t &PredFalseWeight,
2483  uint64_t &SuccTrueWeight,
2484  uint64_t &SuccFalseWeight) {
2485  bool PredHasWeights =
2486  PBI->extractProfMetadata(PredTrueWeight, PredFalseWeight);
2487  bool SuccHasWeights =
2488  BI->extractProfMetadata(SuccTrueWeight, SuccFalseWeight);
2489  if (PredHasWeights || SuccHasWeights) {
2490  if (!PredHasWeights)
2491  PredTrueWeight = PredFalseWeight = 1;
2492  if (!SuccHasWeights)
2493  SuccTrueWeight = SuccFalseWeight = 1;
2494  return true;
2495  } else {
2496  return false;
2497  }
2498 }
2499 
2500 /// If this basic block is simple enough, and if a predecessor branches to us
2501 /// and one of our successors, fold the block into the predecessor and use
2502 /// logical operations to pick the right destination.
2503 bool llvm::FoldBranchToCommonDest(BranchInst *BI, unsigned BonusInstThreshold) {
2504  BasicBlock *BB = BI->getParent();
2505 
2506  Instruction *Cond = nullptr;
2507  if (BI->isConditional())
2508  Cond = dyn_cast<Instruction>(BI->getCondition());
2509  else {
2510  // For unconditional branch, check for a simple CFG pattern, where
2511  // BB has a single predecessor and BB's successor is also its predecessor's
2512  // successor. If such pattern exists, check for CSE between BB and its
2513  // predecessor.
2514  if (BasicBlock *PB = BB->getSinglePredecessor())
2515  if (BranchInst *PBI = dyn_cast<BranchInst>(PB->getTerminator()))
2516  if (PBI->isConditional() &&
2517  (BI->getSuccessor(0) == PBI->getSuccessor(0) ||
2518  BI->getSuccessor(0) == PBI->getSuccessor(1))) {
2519  for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E;) {
2520  Instruction *Curr = &*I++;
2521  if (isa<CmpInst>(Curr)) {
2522  Cond = Curr;
2523  break;
2524  }
2525  // Quit if we can't remove this instruction.
2526  if (!checkCSEInPredecessor(Curr, PB))
2527  return false;
2528  }
2529  }
2530 
2531  if (!Cond)
2532  return false;
2533  }
2534 
2535  if (!Cond || (!isa<CmpInst>(Cond) && !isa<BinaryOperator>(Cond)) ||
2536  Cond->getParent() != BB || !Cond->hasOneUse())
2537  return false;
2538 
2539  // Make sure the instruction after the condition is the cond branch.
2540  BasicBlock::iterator CondIt = ++Cond->getIterator();
2541 
2542  // Ignore dbg intrinsics.
2543  while (isa<DbgInfoIntrinsic>(CondIt))
2544  ++CondIt;
2545 
2546  if (&*CondIt != BI)
2547  return false;
2548 
2549  // Only allow this transformation if computing the condition doesn't involve
2550  // too many instructions and these involved instructions can be executed
2551  // unconditionally. We denote all involved instructions except the condition
2552  // as "bonus instructions", and only allow this transformation when the
2553  // number of the bonus instructions does not exceed a certain threshold.
2554  unsigned NumBonusInsts = 0;
2555  for (auto I = BB->begin(); Cond != &*I; ++I) {
2556  // Ignore dbg intrinsics.
2557  if (isa<DbgInfoIntrinsic>(I))
2558  continue;
2559  if (!I->hasOneUse() || !isSafeToSpeculativelyExecute(&*I))
2560  return false;
2561  // I has only one use and can be executed unconditionally.
2563  if (User == nullptr || User->getParent() != BB)
2564  return false;
2565  // I is used in the same BB. Since BI uses Cond and doesn't have more slots
2566  // to use any other instruction, User must be an instruction between next(I)
2567  // and Cond.
2568  ++NumBonusInsts;
2569  // Early exits once we reach the limit.
2570  if (NumBonusInsts > BonusInstThreshold)
2571  return false;
2572  }
2573 
2574  // Cond is known to be a compare or binary operator. Check to make sure that
2575  // neither operand is a potentially-trapping constant expression.
2576  if (ConstantExpr *CE = dyn_cast<ConstantExpr>(Cond->getOperand(0)))
2577  if (CE->canTrap())
2578  return false;
2579  if (ConstantExpr *CE = dyn_cast<ConstantExpr>(Cond->getOperand(1)))
2580  if (CE->canTrap())
2581  return false;
2582 
2583  // Finally, don't infinitely unroll conditional loops.
2584  BasicBlock *TrueDest = BI->getSuccessor(0);
2585  BasicBlock *FalseDest = (BI->isConditional()) ? BI->getSuccessor(1) : nullptr;
2586  if (TrueDest == BB || FalseDest == BB)
2587  return false;
2588 
2589  for (pred_iterator PI = pred_begin(BB), E = pred_end(BB); PI != E; ++PI) {
2590  BasicBlock *PredBlock = *PI;
2591  BranchInst *PBI = dyn_cast<BranchInst>(PredBlock->getTerminator());
2592 
2593  // Check that we have two conditional branches. If there is a PHI node in
2594  // the common successor, verify that the same value flows in from both
2595  // blocks.
2597  if (!PBI || PBI->isUnconditional() ||
2598  (BI->isConditional() && !SafeToMergeTerminators(BI, PBI)) ||
2599  (!BI->isConditional() &&
2600  !isProfitableToFoldUnconditional(BI, PBI, Cond, PHIs)))
2601  continue;
2602 
2603  // Determine if the two branches share a common destination.
2604  Instruction::BinaryOps Opc = Instruction::BinaryOpsEnd;
2605  bool InvertPredCond = false;
2606 
2607  if (BI->isConditional()) {
2608  if (PBI->getSuccessor(0) == TrueDest) {
2609  Opc = Instruction::Or;
2610  } else if (PBI->getSuccessor(1) == FalseDest) {
2611  Opc = Instruction::And;
2612  } else if (PBI->getSuccessor(0) == FalseDest) {
2613  Opc = Instruction::And;
2614  InvertPredCond = true;
2615  } else if (PBI->getSuccessor(1) == TrueDest) {
2616  Opc = Instruction::Or;
2617  InvertPredCond = true;
2618  } else {
2619  continue;
2620  }
2621  } else {
2622  if (PBI->getSuccessor(0) != TrueDest && PBI->getSuccessor(1) != TrueDest)
2623  continue;
2624  }
2625 
2626  DEBUG(dbgs() << "FOLDING BRANCH TO COMMON DEST:\n" << *PBI << *BB);
2627  IRBuilder<> Builder(PBI);
2628 
2629  // If we need to invert the condition in the pred block to match, do so now.
2630  if (InvertPredCond) {
2631  Value *NewCond = PBI->getCondition();
2632 
2633  if (NewCond->hasOneUse() && isa<CmpInst>(NewCond)) {
2634  CmpInst *CI = cast<CmpInst>(NewCond);
2635  CI->setPredicate(CI->getInversePredicate());
2636  } else {
2637  NewCond =
2638  Builder.CreateNot(NewCond, PBI->getCondition()->getName() + ".not");
2639  }
2640 
2641  PBI->setCondition(NewCond);
2642  PBI->swapSuccessors();
2643  }
2644 
2645  // If we have bonus instructions, clone them into the predecessor block.
2646  // Note that there may be multiple predecessor blocks, so we cannot move
2647  // bonus instructions to a predecessor block.
2648  ValueToValueMapTy VMap; // maps original values to cloned values
2649  // We already make sure Cond is the last instruction before BI. Therefore,
2650  // all instructions before Cond other than DbgInfoIntrinsic are bonus
2651  // instructions.
2652  for (auto BonusInst = BB->begin(); Cond != &*BonusInst; ++BonusInst) {
2653  if (isa<DbgInfoIntrinsic>(BonusInst))
2654  continue;
2655  Instruction *NewBonusInst = BonusInst->clone();
2656  RemapInstruction(NewBonusInst, VMap,
2658  VMap[&*BonusInst] = NewBonusInst;
2659 
2660  // If we moved a load, we cannot any longer claim any knowledge about
2661  // its potential value. The previous information might have been valid
2662  // only given the branch precondition.
2663  // For an analogous reason, we must also drop all the metadata whose
2664  // semantics we don't understand.
2665  NewBonusInst->dropUnknownNonDebugMetadata();
2666 
2667  PredBlock->getInstList().insert(PBI->getIterator(), NewBonusInst);
2668  NewBonusInst->takeName(&*BonusInst);
2669  BonusInst->setName(BonusInst->getName() + ".old");
2670  }
2671 
2672  // Clone Cond into the predecessor basic block, and or/and the
2673  // two conditions together.
2674  Instruction *New = Cond->clone();
2675  RemapInstruction(New, VMap,
2677  PredBlock->getInstList().insert(PBI->getIterator(), New);
2678  New->takeName(Cond);
2679  Cond->setName(New->getName() + ".old");
2680 
2681  if (BI->isConditional()) {
2682  Instruction *NewCond = cast<Instruction>(
2683  Builder.CreateBinOp(Opc, PBI->getCondition(), New, "or.cond"));
2684  PBI->setCondition(NewCond);
2685 
2686  uint64_t PredTrueWeight, PredFalseWeight, SuccTrueWeight, SuccFalseWeight;
2687  bool HasWeights =
2688  extractPredSuccWeights(PBI, BI, PredTrueWeight, PredFalseWeight,
2689  SuccTrueWeight, SuccFalseWeight);
2690  SmallVector<uint64_t, 8> NewWeights;
2691 
2692  if (PBI->getSuccessor(0) == BB) {
2693  if (HasWeights) {
2694  // PBI: br i1 %x, BB, FalseDest
2695  // BI: br i1 %y, TrueDest, FalseDest
2696  // TrueWeight is TrueWeight for PBI * TrueWeight for BI.
2697  NewWeights.push_back(PredTrueWeight * SuccTrueWeight);
2698  // FalseWeight is FalseWeight for PBI * TotalWeight for BI +
2699  // TrueWeight for PBI * FalseWeight for BI.
2700  // We assume that total weights of a BranchInst can fit into 32 bits.
2701  // Therefore, we will not have overflow using 64-bit arithmetic.
2702  NewWeights.push_back(PredFalseWeight *
2703  (SuccFalseWeight + SuccTrueWeight) +
2704  PredTrueWeight * SuccFalseWeight);
2705  }
2706  AddPredecessorToBlock(TrueDest, PredBlock, BB);
2707  PBI->setSuccessor(0, TrueDest);
2708  }
2709  if (PBI->getSuccessor(1) == BB) {
2710  if (HasWeights) {
2711  // PBI: br i1 %x, TrueDest, BB
2712  // BI: br i1 %y, TrueDest, FalseDest
2713  // TrueWeight is TrueWeight for PBI * TotalWeight for BI +
2714  // FalseWeight for PBI * TrueWeight for BI.
2715  NewWeights.push_back(PredTrueWeight *
2716  (SuccFalseWeight + SuccTrueWeight) +
2717  PredFalseWeight * SuccTrueWeight);
2718  // FalseWeight is FalseWeight for PBI * FalseWeight for BI.
2719  NewWeights.push_back(PredFalseWeight * SuccFalseWeight);
2720  }
2721  AddPredecessorToBlock(FalseDest, PredBlock, BB);
2722  PBI->setSuccessor(1, FalseDest);
2723  }
2724  if (NewWeights.size() == 2) {
2725  // Halve the weights if any of them cannot fit in an uint32_t
2726  FitWeights(NewWeights);
2727 
2728  SmallVector<uint32_t, 8> MDWeights(NewWeights.begin(),
2729  NewWeights.end());
2730  PBI->setMetadata(
2732  MDBuilder(BI->getContext()).createBranchWeights(MDWeights));
2733  } else
2734  PBI->setMetadata(LLVMContext::MD_prof, nullptr);
2735  } else {
2736  // Update PHI nodes in the common successors.
2737  for (unsigned i = 0, e = PHIs.size(); i != e; ++i) {
2738  ConstantInt *PBI_C = cast<ConstantInt>(
2739  PHIs[i]->getIncomingValueForBlock(PBI->getParent()));
2740  assert(PBI_C->getType()->isIntegerTy(1));
2741  Instruction *MergedCond = nullptr;
2742  if (PBI->getSuccessor(0) == TrueDest) {
2743  // Create (PBI_Cond and PBI_C) or (!PBI_Cond and BI_Value)
2744  // PBI_C is true: PBI_Cond or (!PBI_Cond and BI_Value)
2745  // is false: !PBI_Cond and BI_Value
2746  Instruction *NotCond = cast<Instruction>(
2747  Builder.CreateNot(PBI->getCondition(), "not.cond"));
2748  MergedCond = cast<Instruction>(
2749  Builder.CreateBinOp(Instruction::And, NotCond, New, "and.cond"));
2750  if (PBI_C->isOne())
2751  MergedCond = cast<Instruction>(Builder.CreateBinOp(
2752  Instruction::Or, PBI->getCondition(), MergedCond, "or.cond"));
2753  } else {
2754  // Create (PBI_Cond and BI_Value) or (!PBI_Cond and PBI_C)
2755  // PBI_C is true: (PBI_Cond and BI_Value) or (!PBI_Cond)
2756  // is false: PBI_Cond and BI_Value
2757  MergedCond = cast<Instruction>(Builder.CreateBinOp(
2758  Instruction::And, PBI->getCondition(), New, "and.cond"));
2759  if (PBI_C->isOne()) {
2760  Instruction *NotCond = cast<Instruction>(
2761  Builder.CreateNot(PBI->getCondition(), "not.cond"));
2762  MergedCond = cast<Instruction>(Builder.CreateBinOp(
2763  Instruction::Or, NotCond, MergedCond, "or.cond"));
2764  }
2765  }
2766  // Update PHI Node.
2767  PHIs[i]->setIncomingValue(PHIs[i]->getBasicBlockIndex(PBI->getParent()),
2768  MergedCond);
2769  }
2770  // Change PBI from Conditional to Unconditional.
2771  BranchInst *New_PBI = BranchInst::Create(TrueDest, PBI);
2773  PBI = New_PBI;
2774  }
2775 
2776  // If BI was a loop latch, it may have had associated loop metadata.
2777  // We need to copy it to the new latch, that is, PBI.
2778  if (MDNode *LoopMD = BI->getMetadata(LLVMContext::MD_loop))
2779  PBI->setMetadata(LLVMContext::MD_loop, LoopMD);
2780 
2781  // TODO: If BB is reachable from all paths through PredBlock, then we
2782  // could replace PBI's branch probabilities with BI's.
2783 
2784  // Copy any debug value intrinsics into the end of PredBlock.
2785  for (Instruction &I : *BB)
2786  if (isa<DbgInfoIntrinsic>(I))
2787  I.clone()->insertBefore(PBI);
2788 
2789  return true;
2790  }
2791  return false;
2792 }
2793 
2794 // If there is only one store in BB1 and BB2, return it, otherwise return
2795 // nullptr.
2797  StoreInst *S = nullptr;
2798  for (auto *BB : {BB1, BB2}) {
2799  if (!BB)
2800  continue;
2801  for (auto &I : *BB)
2802  if (auto *SI = dyn_cast<StoreInst>(&I)) {
2803  if (S)
2804  // Multiple stores seen.
2805  return nullptr;
2806  else
2807  S = SI;
2808  }
2809  }
2810  return S;
2811 }
2812 
2814  Value *AlternativeV = nullptr) {
2815  // PHI is going to be a PHI node that allows the value V that is defined in
2816  // BB to be referenced in BB's only successor.
2817  //
2818  // If AlternativeV is nullptr, the only value we care about in PHI is V. It
2819  // doesn't matter to us what the other operand is (it'll never get used). We
2820  // could just create a new PHI with an undef incoming value, but that could
2821  // increase register pressure if EarlyCSE/InstCombine can't fold it with some
2822  // other PHI. So here we directly look for some PHI in BB's successor with V
2823  // as an incoming operand. If we find one, we use it, else we create a new
2824  // one.
2825  //
2826  // If AlternativeV is not nullptr, we care about both incoming values in PHI.
2827  // PHI must be exactly: phi <ty> [ %BB, %V ], [ %OtherBB, %AlternativeV]
2828  // where OtherBB is the single other predecessor of BB's only successor.
2829  PHINode *PHI = nullptr;
2830  BasicBlock *Succ = BB->getSingleSuccessor();
2831 
2832  for (auto I = Succ->begin(); isa<PHINode>(I); ++I)
2833  if (cast<PHINode>(I)->getIncomingValueForBlock(BB) == V) {
2834  PHI = cast<PHINode>(I);
2835  if (!AlternativeV)
2836  break;
2837 
2838  assert(std::distance(pred_begin(Succ), pred_end(Succ)) == 2);
2839  auto PredI = pred_begin(Succ);
2840  BasicBlock *OtherPredBB = *PredI == BB ? *++PredI : *PredI;
2841  if (PHI->getIncomingValueForBlock(OtherPredBB) == AlternativeV)
2842  break;
2843  PHI = nullptr;
2844  }
2845  if (PHI)
2846  return PHI;
2847 
2848  // If V is not an instruction defined in BB, just return it.
2849  if (!AlternativeV &&
2850  (!isa<Instruction>(V) || cast<Instruction>(V)->getParent() != BB))
2851  return V;
2852 
2853  PHI = PHINode::Create(V->getType(), 2, "simplifycfg.merge", &Succ->front());
2854  PHI->addIncoming(V, BB);
2855  for (BasicBlock *PredBB : predecessors(Succ))
2856  if (PredBB != BB)
2857  PHI->addIncoming(
2858  AlternativeV ? AlternativeV : UndefValue::get(V->getType()), PredBB);
2859  return PHI;
2860 }
2861 
2863  BasicBlock *QTB, BasicBlock *QFB,
2864  BasicBlock *PostBB, Value *Address,
2865  bool InvertPCond, bool InvertQCond) {
2866  auto IsaBitcastOfPointerType = [](const Instruction &I) {
2867  return Operator::getOpcode(&I) == Instruction::BitCast &&
2868  I.getType()->isPointerTy();
2869  };
2870 
2871  // If we're not in aggressive mode, we only optimize if we have some
2872  // confidence that by optimizing we'll allow P and/or Q to be if-converted.
2873  auto IsWorthwhile = [&](BasicBlock *BB) {
2874  if (!BB)
2875  return true;
2876  // Heuristic: if the block can be if-converted/phi-folded and the
2877  // instructions inside are all cheap (arithmetic/GEPs), it's worthwhile to
2878  // thread this store.
2879  unsigned N = 0;
2880  for (auto &I : *BB) {
2881  // Cheap instructions viable for folding.
2882  if (isa<BinaryOperator>(I) || isa<GetElementPtrInst>(I) ||
2883  isa<StoreInst>(I))
2884  ++N;
2885  // Free instructions.
2886  else if (isa<TerminatorInst>(I) || isa<DbgInfoIntrinsic>(I) ||
2887  IsaBitcastOfPointerType(I))
2888  continue;
2889  else
2890  return false;
2891  }
2892  return N <= PHINodeFoldingThreshold;
2893  };
2894 
2896  (!IsWorthwhile(PTB) || !IsWorthwhile(PFB) || !IsWorthwhile(QTB) ||
2897  !IsWorthwhile(QFB)))
2898  return false;
2899 
2900  // For every pointer, there must be exactly two stores, one coming from
2901  // PTB or PFB, and the other from QTB or QFB. We don't support more than one
2902  // store (to any address) in PTB,PFB or QTB,QFB.
2903  // FIXME: We could relax this restriction with a bit more work and performance
2904  // testing.
2905  StoreInst *PStore = findUniqueStoreInBlocks(PTB, PFB);
2906  StoreInst *QStore = findUniqueStoreInBlocks(QTB, QFB);
2907  if (!PStore || !QStore)
2908  return false;
2909 
2910  // Now check the stores are compatible.
2911  if (!QStore->isUnordered() || !PStore->isUnordered())
2912  return false;
2913 
2914  // Check that sinking the store won't cause program behavior changes. Sinking
2915  // the store out of the Q blocks won't change any behavior as we're sinking
2916  // from a block to its unconditional successor. But we're moving a store from
2917  // the P blocks down through the middle block (QBI) and past both QFB and QTB.
2918  // So we need to check that there are no aliasing loads or stores in
2919  // QBI, QTB and QFB. We also need to check there are no conflicting memory
2920  // operations between PStore and the end of its parent block.
2921  //
2922  // The ideal way to do this is to query AliasAnalysis, but we don't
2923  // preserve AA currently so that is dangerous. Be super safe and just
2924  // check there are no other memory operations at all.
2925  for (auto &I : *QFB->getSinglePredecessor())
2926  if (I.mayReadOrWriteMemory())
2927  return false;
2928  for (auto &I : *QFB)
2929  if (&I != QStore && I.mayReadOrWriteMemory())
2930  return false;
2931  if (QTB)
2932  for (auto &I : *QTB)
2933  if (&I != QStore && I.mayReadOrWriteMemory())
2934  return false;
2935  for (auto I = BasicBlock::iterator(PStore), E = PStore->getParent()->end();
2936  I != E; ++I)
2937  if (&*I != PStore && I->mayReadOrWriteMemory())
2938  return false;
2939 
2940  // OK, we're going to sink the stores to PostBB. The store has to be
2941  // conditional though, so first create the predicate.
2942  Value *PCond = cast<BranchInst>(PFB->getSinglePredecessor()->getTerminator())
2943  ->getCondition();
2944  Value *QCond = cast<BranchInst>(QFB->getSinglePredecessor()->getTerminator())
2945  ->getCondition();
2946 
2948  PStore->getParent());
2950  QStore->getParent(), PPHI);
2951 
2952  IRBuilder<> QB(&*PostBB->getFirstInsertionPt());
2953 
2954  Value *PPred = PStore->getParent() == PTB ? PCond : QB.CreateNot(PCond);
2955  Value *QPred = QStore->getParent() == QTB ? QCond : QB.CreateNot(QCond);
2956 
2957  if (InvertPCond)
2958  PPred = QB.CreateNot(PPred);
2959  if (InvertQCond)
2960  QPred = QB.CreateNot(QPred);
2961  Value *CombinedPred = QB.CreateOr(PPred, QPred);
2962 
2963  auto *T =
2964  SplitBlockAndInsertIfThen(CombinedPred, &*QB.GetInsertPoint(), false);
2965  QB.SetInsertPoint(T);
2966  StoreInst *SI = cast<StoreInst>(QB.CreateStore(QPHI, Address));
2967  AAMDNodes AAMD;
2968  PStore->getAAMetadata(AAMD, /*Merge=*/false);
2969  PStore->getAAMetadata(AAMD, /*Merge=*/true);
2970  SI->setAAMetadata(AAMD);
2971 
2972  QStore->eraseFromParent();
2973  PStore->eraseFromParent();
2974 
2975  return true;
2976 }
2977 
2979  // The intention here is to find diamonds or triangles (see below) where each
2980  // conditional block contains a store to the same address. Both of these
2981  // stores are conditional, so they can't be unconditionally sunk. But it may
2982  // be profitable to speculatively sink the stores into one merged store at the
2983  // end, and predicate the merged store on the union of the two conditions of
2984  // PBI and QBI.
2985  //
2986  // This can reduce the number of stores executed if both of the conditions are
2987  // true, and can allow the blocks to become small enough to be if-converted.
2988  // This optimization will also chain, so that ladders of test-and-set
2989  // sequences can be if-converted away.
2990  //
2991  // We only deal with simple diamonds or triangles:
2992  //
2993  // PBI or PBI or a combination of the two
2994  // / \ | \
2995  // PTB PFB | PFB
2996  // \ / | /
2997  // QBI QBI
2998  // / \ | \
2999  // QTB QFB | QFB
3000  // \ / | /
3001  // PostBB PostBB
3002  //
3003  // We model triangles as a type of diamond with a nullptr "true" block.
3004  // Triangles are canonicalized so that the fallthrough edge is represented by
3005  // a true condition, as in the diagram above.
3006  //
3007  BasicBlock *PTB = PBI->getSuccessor(0);
3008  BasicBlock *PFB = PBI->getSuccessor(1);
3009  BasicBlock *QTB = QBI->getSuccessor(0);
3010  BasicBlock *QFB = QBI->getSuccessor(1);
3011  BasicBlock *PostBB = QFB->getSingleSuccessor();
3012 
3013  // Make sure we have a good guess for PostBB. If QTB's only successor is
3014  // QFB, then QFB is a better PostBB.
3015  if (QTB->getSingleSuccessor() == QFB)
3016  PostBB = QFB;
3017 
3018  // If we couldn't find a good PostBB, stop.
3019  if (!PostBB)
3020  return false;
3021 
3022  bool InvertPCond = false, InvertQCond = false;
3023  // Canonicalize fallthroughs to the true branches.
3024  if (PFB == QBI->getParent()) {
3025  std::swap(PFB, PTB);
3026  InvertPCond = true;
3027  }
3028  if (QFB == PostBB) {
3029  std::swap(QFB, QTB);
3030  InvertQCond = true;
3031  }
3032 
3033  // From this point on we can assume PTB or QTB may be fallthroughs but PFB
3034  // and QFB may not. Model fallthroughs as a nullptr block.
3035  if (PTB == QBI->getParent())
3036  PTB = nullptr;
3037  if (QTB == PostBB)
3038  QTB = nullptr;
3039 
3040  // Legality bailouts. We must have at least the non-fallthrough blocks and
3041  // the post-dominating block, and the non-fallthroughs must only have one
3042  // predecessor.
3043  auto HasOnePredAndOneSucc = [](BasicBlock *BB, BasicBlock *P, BasicBlock *S) {
3044  return BB->getSinglePredecessor() == P && BB->getSingleSuccessor() == S;
3045  };
3046  if (!HasOnePredAndOneSucc(PFB, PBI->getParent(), QBI->getParent()) ||
3047  !HasOnePredAndOneSucc(QFB, QBI->getParent(), PostBB))
3048  return false;
3049  if ((PTB && !HasOnePredAndOneSucc(PTB, PBI->getParent(), QBI->getParent())) ||
3050  (QTB && !HasOnePredAndOneSucc(QTB, QBI->getParent(), PostBB)))
3051  return false;
3052  if (!PostBB->hasNUses(2) || !QBI->getParent()->hasNUses(2))
3053  return false;
3054 
3055  // OK, this is a sequence of two diamonds or triangles.
3056  // Check if there are stores in PTB or PFB that are repeated in QTB or QFB.
3057  SmallPtrSet<Value *, 4> PStoreAddresses, QStoreAddresses;
3058  for (auto *BB : {PTB, PFB}) {
3059  if (!BB)
3060  continue;
3061  for (auto &I : *BB)
3062  if (StoreInst *SI = dyn_cast<StoreInst>(&I))
3063  PStoreAddresses.insert(SI->getPointerOperand());
3064  }
3065  for (auto *BB : {QTB, QFB}) {
3066  if (!BB)
3067  continue;
3068  for (auto &I : *BB)
3069  if (StoreInst *SI = dyn_cast<StoreInst>(&I))
3070  QStoreAddresses.insert(SI->getPointerOperand());
3071  }
3072 
3073  set_intersect(PStoreAddresses, QStoreAddresses);
3074  // set_intersect mutates PStoreAddresses in place. Rename it here to make it
3075  // clear what it contains.
3076  auto &CommonAddresses = PStoreAddresses;
3077 
3078  bool Changed = false;
3079  for (auto *Address : CommonAddresses)
3080  Changed |= mergeConditionalStoreToAddress(
3081  PTB, PFB, QTB, QFB, PostBB, Address, InvertPCond, InvertQCond);
3082  return Changed;
3083 }
3084 
3085 /// If we have a conditional branch as a predecessor of another block,
3086 /// this function tries to simplify it. We know
3087 /// that PBI and BI are both conditional branches, and BI is in one of the
3088 /// successor blocks of PBI - PBI branches to BI.
3090  const DataLayout &DL) {
3091  assert(PBI->isConditional() && BI->isConditional());
3092  BasicBlock *BB = BI->getParent();
3093 
3094  // If this block ends with a branch instruction, and if there is a
3095  // predecessor that ends on a branch of the same condition, make
3096  // this conditional branch redundant.
3097  if (PBI->getCondition() == BI->getCondition() &&
3098  PBI->getSuccessor(0) != PBI->getSuccessor(1)) {
3099  // Okay, the outcome of this conditional branch is statically
3100  // knowable. If this block had a single pred, handle specially.
3101  if (BB->getSinglePredecessor()) {
3102  // Turn this into a branch on constant.
3103  bool CondIsTrue = PBI->getSuccessor(0) == BB;
3104  BI->setCondition(
3105  ConstantInt::get(Type::getInt1Ty(BB->getContext()), CondIsTrue));
3106  return true; // Nuke the branch on constant.
3107  }
3108 
3109  // Otherwise, if there are multiple predecessors, insert a PHI that merges
3110  // in the constant and simplify the block result. Subsequent passes of
3111  // simplifycfg will thread the block.
3113  pred_iterator PB = pred_begin(BB), PE = pred_end(BB);
3114  PHINode *NewPN = PHINode::Create(
3115  Type::getInt1Ty(BB->getContext()), std::distance(PB, PE),
3116  BI->getCondition()->getName() + ".pr", &BB->front());
3117  // Okay, we're going to insert the PHI node. Since PBI is not the only
3118  // predecessor, compute the PHI'd conditional value for all of the preds.
3119  // Any predecessor where the condition is not computable we keep symbolic.
3120  for (pred_iterator PI = PB; PI != PE; ++PI) {
3121  BasicBlock *P = *PI;
3122  if ((PBI = dyn_cast<BranchInst>(P->getTerminator())) && PBI != BI &&
3123  PBI->isConditional() && PBI->getCondition() == BI->getCondition() &&
3124  PBI->getSuccessor(0) != PBI->getSuccessor(1)) {
3125  bool CondIsTrue = PBI->getSuccessor(0) == BB;
3126  NewPN->addIncoming(
3127  ConstantInt::get(Type::getInt1Ty(BB->getContext()), CondIsTrue),
3128  P);
3129  } else {
3130  NewPN->addIncoming(BI->getCondition(), P);
3131  }
3132  }
3133 
3134  BI->setCondition(NewPN);
3135  return true;
3136  }
3137  }
3138 
3139  if (auto *CE = dyn_cast<ConstantExpr>(BI->getCondition()))
3140  if (CE->canTrap())
3141  return false;
3142 
3143  // If both branches are conditional and both contain stores to the same
3144  // address, remove the stores from the conditionals and create a conditional
3145  // merged store at the end.
3146  if (MergeCondStores && mergeConditionalStores(PBI, BI))
3147  return true;
3148 
3149  // If this is a conditional branch in an empty block, and if any
3150  // predecessors are a conditional branch to one of our destinations,
3151  // fold the conditions into logical ops and one cond br.
3152  BasicBlock::iterator BBI = BB->begin();
3153  // Ignore dbg intrinsics.
3154  while (isa<DbgInfoIntrinsic>(BBI))
3155  ++BBI;
3156  if (&*BBI != BI)
3157  return false;
3158 
3159  int PBIOp, BIOp;
3160  if (PBI->getSuccessor(0) == BI->getSuccessor(0)) {
3161  PBIOp = 0;
3162  BIOp = 0;
3163  } else if (PBI->getSuccessor(0) == BI->getSuccessor(1)) {
3164  PBIOp = 0;
3165  BIOp = 1;
3166  } else if (PBI->getSuccessor(1) == BI->getSuccessor(0)) {
3167  PBIOp = 1;
3168  BIOp = 0;
3169  } else if (PBI->getSuccessor(1) == BI->getSuccessor(1)) {
3170  PBIOp = 1;
3171  BIOp = 1;
3172  } else {
3173  return false;
3174  }
3175 
3176  // Check to make sure that the other destination of this branch
3177  // isn't BB itself. If so, this is an infinite loop that will
3178  // keep getting unwound.
3179  if (PBI->getSuccessor(PBIOp) == BB)
3180  return false;
3181 
3182  // Do not perform this transformation if it would require
3183  // insertion of a large number of select instructions. For targets
3184  // without predication/cmovs, this is a big pessimization.
3185 
3186  // Also do not perform this transformation if any phi node in the common
3187  // destination block can trap when reached by BB or PBB (PR17073). In that
3188  // case, it would be unsafe to hoist the operation into a select instruction.
3189 
3190  BasicBlock *CommonDest = PBI->getSuccessor(PBIOp);
3191  unsigned NumPhis = 0;
3192  for (BasicBlock::iterator II = CommonDest->begin(); isa<PHINode>(II);
3193  ++II, ++NumPhis) {
3194  if (NumPhis > 2) // Disable this xform.
3195  return false;
3196 
3197  PHINode *PN = cast<PHINode>(II);
3198  Value *BIV = PN->getIncomingValueForBlock(BB);
3199  if (ConstantExpr *CE = dyn_cast<ConstantExpr>(BIV))
3200  if (CE->canTrap())
3201  return false;
3202 
3203  unsigned PBBIdx = PN->getBasicBlockIndex(PBI->getParent());
3204  Value *PBIV = PN->getIncomingValue(PBBIdx);
3205  if (ConstantExpr *CE = dyn_cast<ConstantExpr>(PBIV))
3206  if (CE->canTrap())
3207  return false;
3208  }
3209 
3210  // Finally, if everything is ok, fold the branches to logical ops.
3211  BasicBlock *OtherDest = BI->getSuccessor(BIOp ^ 1);
3212 
3213  DEBUG(dbgs() << "FOLDING BRs:" << *PBI->getParent()
3214  << "AND: " << *BI->getParent());
3215 
3216  // If OtherDest *is* BB, then BB is a basic block with a single conditional
3217  // branch in it, where one edge (OtherDest) goes back to itself but the other
3218  // exits. We don't *know* that the program avoids the infinite loop
3219  // (even though that seems likely). If we do this xform naively, we'll end up
3220  // recursively unpeeling the loop. Since we know that (after the xform is
3221  // done) that the block *is* infinite if reached, we just make it an obviously
3222  // infinite loop with no cond branch.
3223  if (OtherDest == BB) {
3224  // Insert it at the end of the function, because it's either code,
3225  // or it won't matter if it's hot. :)
3226  BasicBlock *InfLoopBlock =
3227  BasicBlock::Create(BB->getContext(), "infloop", BB->getParent());
3228  BranchInst::Create(InfLoopBlock, InfLoopBlock);
3229  OtherDest = InfLoopBlock;
3230  }
3231 
3232  DEBUG(dbgs() << *PBI->getParent()->getParent());
3233 
3234  // BI may have other predecessors. Because of this, we leave
3235  // it alone, but modify PBI.
3236 
3237  // Make sure we get to CommonDest on True&True directions.
3238  Value *PBICond = PBI->getCondition();
3239  IRBuilder<NoFolder> Builder(PBI);
3240  if (PBIOp)
3241  PBICond = Builder.CreateNot(PBICond, PBICond->getName() + ".not");
3242 
3243  Value *BICond = BI->getCondition();
3244  if (BIOp)
3245  BICond = Builder.CreateNot(BICond, BICond->getName() + ".not");
3246 
3247  // Merge the conditions.
3248  Value *Cond = Builder.CreateOr(PBICond, BICond, "brmerge");
3249 
3250  // Modify PBI to branch on the new condition to the new dests.
3251  PBI->setCondition(Cond);
3252  PBI->setSuccessor(0, CommonDest);
3253  PBI->setSuccessor(1, OtherDest);
3254 
3255  // Update branch weight for PBI.
3256  uint64_t PredTrueWeight, PredFalseWeight, SuccTrueWeight, SuccFalseWeight;
3257  uint64_t PredCommon, PredOther, SuccCommon, SuccOther;
3258  bool HasWeights =
3259  extractPredSuccWeights(PBI, BI, PredTrueWeight, PredFalseWeight,
3260  SuccTrueWeight, SuccFalseWeight);
3261  if (HasWeights) {
3262  PredCommon = PBIOp ? PredFalseWeight : PredTrueWeight;
3263  PredOther = PBIOp ? PredTrueWeight : PredFalseWeight;
3264  SuccCommon = BIOp ? SuccFalseWeight : SuccTrueWeight;
3265  SuccOther = BIOp ? SuccTrueWeight : SuccFalseWeight;
3266  // The weight to CommonDest should be PredCommon * SuccTotal +
3267  // PredOther * SuccCommon.
3268  // The weight to OtherDest should be PredOther * SuccOther.
3269  uint64_t NewWeights[2] = {PredCommon * (SuccCommon + SuccOther) +
3270  PredOther * SuccCommon,
3271  PredOther * SuccOther};
3272  // Halve the weights if any of them cannot fit in an uint32_t
3273  FitWeights(NewWeights);
3274 
3276  MDBuilder(BI->getContext())
3277  .createBranchWeights(NewWeights[0], NewWeights[1]));
3278  }
3279 
3280  // OtherDest may have phi nodes. If so, add an entry from PBI's
3281  // block that are identical to the entries for BI's block.
3282  AddPredecessorToBlock(OtherDest, PBI->getParent(), BB);
3283 
3284  // We know that the CommonDest already had an edge from PBI to
3285  // it. If it has PHIs though, the PHIs may have different
3286  // entries for BB and PBI's BB. If so, insert a select to make
3287  // them agree.
3288  PHINode *PN;
3289  for (BasicBlock::iterator II = CommonDest->begin();
3290  (PN = dyn_cast<PHINode>(II)); ++II) {
3291  Value *BIV = PN->getIncomingValueForBlock(BB);
3292  unsigned PBBIdx = PN->getBasicBlockIndex(PBI->getParent());
3293  Value *PBIV = PN->getIncomingValue(PBBIdx);
3294  if (BIV != PBIV) {
3295  // Insert a select in PBI to pick the right value.
3296  SelectInst *NV = cast<SelectInst>(
3297  Builder.CreateSelect(PBICond, PBIV, BIV, PBIV->getName() + ".mux"));
3298  PN->setIncomingValue(PBBIdx, NV);
3299  // Although the select has the same condition as PBI, the original branch
3300  // weights for PBI do not apply to the new select because the select's
3301  // 'logical' edges are incoming edges of the phi that is eliminated, not
3302  // the outgoing edges of PBI.
3303  if (HasWeights) {
3304  uint64_t PredCommon = PBIOp ? PredFalseWeight : PredTrueWeight;
3305  uint64_t PredOther = PBIOp ? PredTrueWeight : PredFalseWeight;
3306  uint64_t SuccCommon = BIOp ? SuccFalseWeight : SuccTrueWeight;
3307  uint64_t SuccOther = BIOp ? SuccTrueWeight : SuccFalseWeight;
3308  // The weight to PredCommonDest should be PredCommon * SuccTotal.
3309  // The weight to PredOtherDest should be PredOther * SuccCommon.
3310  uint64_t NewWeights[2] = {PredCommon * (SuccCommon + SuccOther),
3311  PredOther * SuccCommon};
3312 
3313  FitWeights(NewWeights);
3314 
3316  MDBuilder(BI->getContext())
3317  .createBranchWeights(NewWeights[0], NewWeights[1]));
3318  }
3319  }
3320  }
3321 
3322  DEBUG(dbgs() << "INTO: " << *PBI->getParent());
3323  DEBUG(dbgs() << *PBI->getParent()->getParent());
3324 
3325  // This basic block is probably dead. We know it has at least
3326  // one fewer predecessor.
3327  return true;
3328 }
3329 
3330 // Simplifies a terminator by replacing it with a branch to TrueBB if Cond is
3331 // true or to FalseBB if Cond is false.
3332 // Takes care of updating the successors and removing the old terminator.
3333 // Also makes sure not to introduce new successors by assuming that edges to
3334 // non-successor TrueBBs and FalseBBs aren't reachable.
3336  BasicBlock *TrueBB, BasicBlock *FalseBB,
3337  uint32_t TrueWeight,
3338  uint32_t FalseWeight) {
3339  // Remove any superfluous successor edges from the CFG.
3340  // First, figure out which successors to preserve.
3341  // If TrueBB and FalseBB are equal, only try to preserve one copy of that
3342  // successor.
3343  BasicBlock *KeepEdge1 = TrueBB;
3344  BasicBlock *KeepEdge2 = TrueBB != FalseBB ? FalseBB : nullptr;
3345 
3346  // Then remove the rest.
3347  for (BasicBlock *Succ : OldTerm->successors()) {
3348  // Make sure only to keep exactly one copy of each edge.
3349  if (Succ == KeepEdge1)
3350  KeepEdge1 = nullptr;
3351  else if (Succ == KeepEdge2)
3352  KeepEdge2 = nullptr;
3353  else
3354  Succ->removePredecessor(OldTerm->getParent(),
3355  /*DontDeleteUselessPHIs=*/true);
3356  }
3357 
3358  IRBuilder<> Builder(OldTerm);
3359  Builder.SetCurrentDebugLocation(OldTerm->getDebugLoc());
3360 
3361  // Insert an appropriate new terminator.
3362  if (!KeepEdge1 && !KeepEdge2) {
3363  if (TrueBB == FalseBB)
3364  // We were only looking for one successor, and it was present.
3365  // Create an unconditional branch to it.
3366  Builder.CreateBr(TrueBB);
3367  else {
3368  // We found both of the successors we were looking for.
3369  // Create a conditional branch sharing the condition of the select.
3370  BranchInst *NewBI = Builder.CreateCondBr(Cond, TrueBB, FalseBB);
3371  if (TrueWeight != FalseWeight)
3373  MDBuilder(OldTerm->getContext())
3374  .createBranchWeights(TrueWeight, FalseWeight));
3375  }
3376  } else if (KeepEdge1 && (KeepEdge2 || TrueBB == FalseBB)) {
3377  // Neither of the selected blocks were successors, so this
3378  // terminator must be unreachable.
3379  new UnreachableInst(OldTerm->getContext(), OldTerm);
3380  } else {
3381  // One of the selected values was a successor, but the other wasn't.
3382  // Insert an unconditional branch to the one that was found;
3383  // the edge to the one that wasn't must be unreachable.
3384  if (!KeepEdge1)
3385  // Only TrueBB was found.
3386  Builder.CreateBr(TrueBB);
3387  else
3388  // Only FalseBB was found.
3389  Builder.CreateBr(FalseBB);
3390  }
3391 
3393  return true;
3394 }
3395 
3396 // Replaces
3397 // (switch (select cond, X, Y)) on constant X, Y
3398 // with a branch - conditional if X and Y lead to distinct BBs,
3399 // unconditional otherwise.
3401  // Check for constant integer values in the select.
3402  ConstantInt *TrueVal = dyn_cast<ConstantInt>(Select->getTrueValue());
3403  ConstantInt *FalseVal = dyn_cast<ConstantInt>(Select->getFalseValue());
3404  if (!TrueVal || !FalseVal)
3405  return false;
3406 
3407  // Find the relevant condition and destinations.
3408  Value *Condition = Select->getCondition();
3409  BasicBlock *TrueBB = SI->findCaseValue(TrueVal)->getCaseSuccessor();
3410  BasicBlock *FalseBB = SI->findCaseValue(FalseVal)->getCaseSuccessor();
3411 
3412  // Get weight for TrueBB and FalseBB.
3413  uint32_t TrueWeight = 0, FalseWeight = 0;
3414  SmallVector<uint64_t, 8> Weights;
3415  bool HasWeights = HasBranchWeights(SI);
3416  if (HasWeights) {
3417  GetBranchWeights(SI, Weights);
3418  if (Weights.size() == 1 + SI->getNumCases()) {
3419  TrueWeight =
3420  (uint32_t)Weights[SI->findCaseValue(TrueVal)->getSuccessorIndex()];
3421  FalseWeight =
3422  (uint32_t)Weights[SI->findCaseValue(FalseVal)->getSuccessorIndex()];
3423  }
3424  }
3425 
3426  // Perform the actual simplification.
3427  return SimplifyTerminatorOnSelect(SI, Condition, TrueBB, FalseBB, TrueWeight,
3428  FalseWeight);
3429 }
3430 
3431 // Replaces
3432 // (indirectbr (select cond, blockaddress(@fn, BlockA),
3433 // blockaddress(@fn, BlockB)))
3434 // with
3435 // (br cond, BlockA, BlockB).
3437  // Check that both operands of the select are block addresses.
3440  if (!TBA || !FBA)
3441  return false;
3442 
3443  // Extract the actual blocks.
3444  BasicBlock *TrueBB = TBA->getBasicBlock();
3445  BasicBlock *FalseBB = FBA->getBasicBlock();
3446 
3447  // Perform the actual simplification.
3448  return SimplifyTerminatorOnSelect(IBI, SI->getCondition(), TrueBB, FalseBB, 0,
3449  0);
3450 }
3451 
3452 /// This is called when we find an icmp instruction
3453 /// (a seteq/setne with a constant) as the only instruction in a
3454 /// block that ends with an uncond branch. We are looking for a very specific
3455 /// pattern that occurs when "A == 1 || A == 2 || A == 3" gets simplified. In
3456 /// this case, we merge the first two "or's of icmp" into a switch, but then the
3457 /// default value goes to an uncond block with a seteq in it, we get something
3458 /// like:
3459 ///
3460 /// switch i8 %A, label %DEFAULT [ i8 1, label %end i8 2, label %end ]
3461 /// DEFAULT:
3462 /// %tmp = icmp eq i8 %A, 92
3463 /// br label %end
3464 /// end:
3465 /// ... = phi i1 [ true, %entry ], [ %tmp, %DEFAULT ], [ true, %entry ]
3466 ///
3467 /// We prefer to split the edge to 'end' so that there is a true/false entry to
3468 /// the PHI, merging the third icmp into the switch.
3470  ICmpInst *ICI, IRBuilder<> &Builder, const DataLayout &DL,
3471  const TargetTransformInfo &TTI, unsigned BonusInstThreshold,
3472  AssumptionCache *AC) {
3473  BasicBlock *BB = ICI->getParent();
3474 
3475  // If the block has any PHIs in it or the icmp has multiple uses, it is too
3476  // complex.
3477  if (isa<PHINode>(BB->begin()) || !ICI->hasOneUse())
3478  return false;
3479 
3480  Value *V = ICI->getOperand(0);
3481  ConstantInt *Cst = cast<ConstantInt>(ICI->getOperand(1));
3482 
3483  // The pattern we're looking for is where our only predecessor is a switch on
3484  // 'V' and this block is the default case for the switch. In this case we can
3485  // fold the compared value into the switch to simplify things.
3486  BasicBlock *Pred = BB->getSinglePredecessor();
3487  if (!Pred || !isa<SwitchInst>(Pred->getTerminator()))
3488  return false;
3489 
3490  SwitchInst *SI = cast<SwitchInst>(Pred->getTerminator());
3491  if (SI->getCondition() != V)
3492  return false;
3493 
3494  // If BB is reachable on a non-default case, then we simply know the value of
3495  // V in this block. Substitute it and constant fold the icmp instruction
3496  // away.
3497  if (SI->getDefaultDest() != BB) {
3498  ConstantInt *VVal = SI->findCaseDest(BB);
3499  assert(VVal && "Should have a unique destination value");
3500  ICI->setOperand(0, VVal);
3501 
3502  if (Value *V = SimplifyInstruction(ICI, {DL, ICI})) {
3503  ICI->replaceAllUsesWith(V);
3504  ICI->eraseFromParent();
3505  }
3506  // BB is now empty, so it is likely to simplify away.
3507  return SimplifyCFG(BB, TTI, BonusInstThreshold, AC) | true;
3508  }
3509 
3510  // Ok, the block is reachable from the default dest. If the constant we're
3511  // comparing exists in one of the other edges, then we can constant fold ICI
3512  // and zap it.
3513  if (SI->findCaseValue(Cst) != SI->case_default()) {
3514  Value *V;
3515  if (ICI->getPredicate() == ICmpInst::ICMP_EQ)
3516  V = ConstantInt::getFalse(BB->getContext());
3517  else
3518  V = ConstantInt::getTrue(BB->getContext());
3519 
3520  ICI->replaceAllUsesWith(V);
3521  ICI->eraseFromParent();
3522  // BB is now empty, so it is likely to simplify away.
3523  return SimplifyCFG(BB, TTI, BonusInstThreshold, AC) | true;
3524  }
3525 
3526  // The use of the icmp has to be in the 'end' block, by the only PHI node in
3527  // the block.
3528  BasicBlock *SuccBlock = BB->getTerminator()->getSuccessor(0);
3529  PHINode *PHIUse = dyn_cast<PHINode>(ICI->user_back());
3530  if (PHIUse == nullptr || PHIUse != &SuccBlock->front() ||
3531  isa<PHINode>(++BasicBlock::iterator(PHIUse)))
3532  return false;
3533 
3534  // If the icmp is a SETEQ, then the default dest gets false, the new edge gets
3535  // true in the PHI.
3536  Constant *DefaultCst = ConstantInt::getTrue(BB->getContext());
3537  Constant *NewCst = ConstantInt::getFalse(BB->getContext());
3538 
3539  if (ICI->getPredicate() == ICmpInst::ICMP_EQ)
3540  std::swap(DefaultCst, NewCst);
3541 
3542  // Replace ICI (which is used by the PHI for the default value) with true or
3543  // false depending on if it is EQ or NE.
3544  ICI->replaceAllUsesWith(DefaultCst);
3545  ICI->eraseFromParent();
3546 
3547  // Okay, the switch goes to this block on a default value. Add an edge from
3548  // the switch to the merge point on the compared value.
3549  BasicBlock *NewBB =
3550  BasicBlock::Create(BB->getContext(), "switch.edge", BB->getParent(), BB);
3551  SmallVector<uint64_t, 8> Weights;
3552  bool HasWeights = HasBranchWeights(SI);
3553  if (HasWeights) {
3554  GetBranchWeights(SI, Weights);
3555  if (Weights.size() == 1 + SI->getNumCases()) {
3556  // Split weight for default case to case for "Cst".
3557  Weights[0] = (Weights[0] + 1) >> 1;
3558  Weights.push_back(Weights[0]);
3559 
3560  SmallVector<uint32_t, 8> MDWeights(Weights.begin(), Weights.end());
3561  SI->setMetadata(
3563  MDBuilder(SI->getContext()).createBranchWeights(MDWeights));
3564  }
3565  }
3566  SI->addCase(Cst, NewBB);
3567 
3568  // NewBB branches to the phi block, add the uncond branch and the phi entry.
3569  Builder.SetInsertPoint(NewBB);
3570  Builder.SetCurrentDebugLocation(SI->getDebugLoc());
3571  Builder.CreateBr(SuccBlock);
3572  PHIUse->addIncoming(NewCst, NewBB);
3573  return true;
3574 }
3575 
3576 /// The specified branch is a conditional branch.
3577 /// Check to see if it is branching on an or/and chain of icmp instructions, and
3578 /// fold it into a switch instruction if so.
3580  const DataLayout &DL) {
3581  Instruction *Cond = dyn_cast<Instruction>(BI->getCondition());
3582  if (!Cond)
3583  return false;
3584 
3585  // Change br (X == 0 | X == 1), T, F into a switch instruction.
3586  // If this is a bunch of seteq's or'd together, or if it's a bunch of
3587  // 'setne's and'ed together, collect them.
3588 
3589  // Try to gather values from a chain of and/or to be turned into a switch
3590  ConstantComparesGatherer ConstantCompare(Cond, DL);
3591  // Unpack the result
3592  SmallVectorImpl<ConstantInt *> &Values = ConstantCompare.Vals;
3593  Value *CompVal = ConstantCompare.CompValue;
3594  unsigned UsedICmps = ConstantCompare.UsedICmps;
3595  Value *ExtraCase = ConstantCompare.Extra;
3596 
3597  // If we didn't have a multiply compared value, fail.
3598  if (!CompVal)
3599  return false;
3600 
3601  // Avoid turning single icmps into a switch.
3602  if (UsedICmps <= 1)
3603  return false;
3604 
3605  bool TrueWhenEqual = (Cond->getOpcode() == Instruction::Or);
3606 
3607  // There might be duplicate constants in the list, which the switch
3608  // instruction can't handle, remove them now.
3609  array_pod_sort(Values.begin(), Values.end(), ConstantIntSortPredicate);
3610  Values.erase(std::unique(Values.begin(), Values.end()), Values.end());
3611 
3612  // If Extra was used, we require at least two switch values to do the
3613  // transformation. A switch with one value is just a conditional branch.
3614  if (ExtraCase && Values.size() < 2)
3615  return false;
3616 
3617  // TODO: Preserve branch weight metadata, similarly to how
3618  // FoldValueComparisonIntoPredecessors preserves it.
3619 
3620  // Figure out which block is which destination.
3621  BasicBlock *DefaultBB = BI->getSuccessor(1);
3622  BasicBlock *EdgeBB = BI->getSuccessor(0);
3623  if (!TrueWhenEqual)
3624  std::swap(DefaultBB, EdgeBB);
3625 
3626  BasicBlock *BB = BI->getParent();
3627 
3628  DEBUG(dbgs() << "Converting 'icmp' chain with " << Values.size()
3629  << " cases into SWITCH. BB is:\n"
3630  << *BB);
3631 
3632  // If there are any extra values that couldn't be folded into the switch
3633  // then we evaluate them with an explicit branch first. Split the block
3634  // right before the condbr to handle it.
3635  if (ExtraCase) {
3636  BasicBlock *NewBB =
3637  BB->splitBasicBlock(BI->getIterator(), "switch.early.test");
3638  // Remove the uncond branch added to the old block.
3639  TerminatorInst *OldTI = BB->getTerminator();
3640  Builder.SetInsertPoint(OldTI);
3641 
3642  if (TrueWhenEqual)
3643  Builder.CreateCondBr(ExtraCase, EdgeBB, NewBB);
3644  else
3645  Builder.CreateCondBr(ExtraCase, NewBB, EdgeBB);
3646 
3647  OldTI->eraseFromParent();
3648 
3649  // If there are PHI nodes in EdgeBB, then we need to add a new entry to them
3650  // for the edge we just added.
3651  AddPredecessorToBlock(EdgeBB, BB, NewBB);
3652 
3653  DEBUG(dbgs() << " ** 'icmp' chain unhandled condition: " << *ExtraCase
3654  << "\nEXTRABB = " << *BB);
3655  BB = NewBB;
3656  }
3657 
3658  Builder.SetInsertPoint(BI);
3659  // Convert pointer to int before we switch.
3660  if (CompVal->getType()->isPointerTy()) {
3661  CompVal = Builder.CreatePtrToInt(
3662  CompVal, DL.getIntPtrType(CompVal->getType()), "magicptr");
3663  }
3664 
3665  // Create the new switch instruction now.
3666  SwitchInst *New = Builder.CreateSwitch(CompVal, DefaultBB, Values.size());
3667 
3668  // Add all of the 'cases' to the switch instruction.
3669  for (unsigned i = 0, e = Values.size(); i != e; ++i)
3670  New->addCase(Values[i], EdgeBB);
3671 
3672  // We added edges from PI to the EdgeBB. As such, if there were any
3673  // PHI nodes in EdgeBB, they need entries to be added corresponding to
3674  // the number of edges added.
3675  for (BasicBlock::iterator BBI = EdgeBB->begin(); isa<PHINode>(BBI); ++BBI) {
3676  PHINode *PN = cast<PHINode>(BBI);
3677  Value *InVal = PN->getIncomingValueForBlock(BB);
3678  for (unsigned i = 0, e = Values.size() - 1; i != e; ++i)
3679  PN->addIncoming(InVal, BB);
3680  }
3681 
3682  // Erase the old branch instruction.
3684 
3685  DEBUG(dbgs() << " ** 'icmp' chain result is:\n" << *BB << '\n');
3686  return true;
3687 }
3688 
3689 bool SimplifyCFGOpt::SimplifyResume(ResumeInst *RI, IRBuilder<> &Builder) {
3690  if (isa<PHINode>(RI->getValue()))
3691  return SimplifyCommonResume(RI);
3692  else if (isa<LandingPadInst>(RI->getParent()->getFirstNonPHI()) &&
3693  RI->getValue() == RI->getParent()->getFirstNonPHI())
3694  // The resume must unwind the exception that caused control to branch here.
3695  return SimplifySingleResume(RI);
3696 
3697  return false;
3698 }
3699 
3700 // Simplify resume that is shared by several landing pads (phi of landing pad).
3701 bool SimplifyCFGOpt::SimplifyCommonResume(ResumeInst *RI) {
3702  BasicBlock *BB = RI->getParent();
3703 
3704  // Check that there are no other instructions except for debug intrinsics
3705  // between the phi of landing pads (RI->getValue()) and resume instruction.
3706  BasicBlock::iterator I = cast<Instruction>(RI->getValue())->getIterator(),
3707  E = RI->getIterator();
3708  while (++I != E)
3709  if (!isa<DbgInfoIntrinsic>(I))
3710  return false;
3711 
3712  SmallSetVector<BasicBlock *, 4> TrivialUnwindBlocks;
3713  auto *PhiLPInst = cast<PHINode>(RI->getValue());
3714 
3715  // Check incoming blocks to see if any of them are trivial.
3716  for (unsigned Idx = 0, End = PhiLPInst->getNumIncomingValues(); Idx != End;
3717  Idx++) {
3718  auto *IncomingBB = PhiLPInst->getIncomingBlock(Idx);
3719  auto *IncomingValue = PhiLPInst->getIncomingValue(Idx);
3720 
3721  // If the block has other successors, we can not delete it because
3722  // it has other dependents.
3723  if (IncomingBB->getUniqueSuccessor() != BB)
3724  continue;
3725 
3726  auto *LandingPad = dyn_cast<LandingPadInst>(IncomingBB->getFirstNonPHI());
3727  // Not the landing pad that caused the control to branch here.
3728  if (IncomingValue != LandingPad)
3729  continue;
3730 
3731  bool isTrivial = true;
3732 
3733  I = IncomingBB->getFirstNonPHI()->getIterator();
3734  E = IncomingBB->getTerminator()->getIterator();
3735  while (++I != E)
3736  if (!isa<DbgInfoIntrinsic>(I)) {
3737  isTrivial = false;
3738  break;
3739  }
3740 
3741  if (isTrivial)
3742  TrivialUnwindBlocks.insert(IncomingBB);
3743  }
3744 
3745  // If no trivial unwind blocks, don't do any simplifications.
3746  if (TrivialUnwindBlocks.empty())
3747  return false;
3748 
3749  // Turn all invokes that unwind here into calls.
3750  for (auto *TrivialBB : TrivialUnwindBlocks) {
3751  // Blocks that will be simplified should be removed from the phi node.
3752  // Note there could be multiple edges to the resume block, and we need
3753  // to remove them all.
3754  while (PhiLPInst->getBasicBlockIndex(TrivialBB) != -1)
3755  BB->removePredecessor(TrivialBB, true);
3756 
3757  for (pred_iterator PI = pred_begin(TrivialBB), PE = pred_end(TrivialBB);
3758  PI != PE;) {
3759  BasicBlock *Pred = *PI++;
3760  removeUnwindEdge(Pred);
3761  }
3762 
3763  // In each SimplifyCFG run, only the current processed block can be erased.
3764  // Otherwise, it will break the iteration of SimplifyCFG pass. So instead
3765  // of erasing TrivialBB, we only remove the branch to the common resume
3766  // block so that we can later erase the resume block since it has no
3767  // predecessors.
3768  TrivialBB->getTerminator()->eraseFromParent();
3769  new UnreachableInst(RI->getContext(), TrivialBB);
3770  }
3771 
3772  // Delete the resume block if all its predecessors have been removed.
3773  if (pred_empty(BB))
3774  BB->eraseFromParent();
3775 
3776  return !TrivialUnwindBlocks.empty();
3777 }
3778 
3779 // Simplify resume that is only used by a single (non-phi) landing pad.
3780 bool SimplifyCFGOpt::SimplifySingleResume(ResumeInst *RI) {
3781  BasicBlock *BB = RI->getParent();
3783  assert(RI->getValue() == LPInst &&
3784  "Resume must unwind the exception that caused control to here");
3785 
3786  // Check that there are no other instructions except for debug intrinsics.
3787  BasicBlock::iterator I = LPInst->getIterator(), E = RI->getIterator();
3788  while (++I != E)
3789  if (!isa<DbgInfoIntrinsic>(I))
3790  return false;
3791 
3792  // Turn all invokes that unwind here into calls and delete the basic block.
3793  for (pred_iterator PI = pred_begin(BB), PE = pred_end(BB); PI != PE;) {
3794  BasicBlock *Pred = *PI++;
3795  removeUnwindEdge(Pred);
3796  }
3797 
3798  // The landingpad is now unreachable. Zap it.
3799  BB->eraseFromParent();
3800  if (LoopHeaders)
3801  LoopHeaders->erase(BB);
3802  return true;
3803 }
3804 
3806  // If this is a trivial cleanup pad that executes no instructions, it can be
3807  // eliminated. If the cleanup pad continues to the caller, any predecessor
3808  // that is an EH pad will be updated to continue to the caller and any
3809  // predecessor that terminates with an invoke instruction will have its invoke
3810  // instruction converted to a call instruction. If the cleanup pad being
3811  // simplified does not continue to the caller, each predecessor will be
3812  // updated to continue to the unwind destination of the cleanup pad being
3813  // simplified.
3814  BasicBlock *BB = RI->getParent();
3815  CleanupPadInst *CPInst = RI->getCleanupPad();
3816  if (CPInst->getParent() != BB)
3817  // This isn't an empty cleanup.
3818  return false;
3819 
3820  // We cannot kill the pad if it has multiple uses. This typically arises
3821  // from unreachable basic blocks.
3822  if (!CPInst->hasOneUse())
3823  return false;
3824 
3825  // Check that there are no other instructions except for benign intrinsics.
3826  BasicBlock::iterator I = CPInst->getIterator(), E = RI->getIterator();
3827  while (++I != E) {
3828  auto *II = dyn_cast<IntrinsicInst>(I);
3829  if (!II)
3830  return false;
3831 
3832  Intrinsic::ID IntrinsicID = II->getIntrinsicID();
3833  switch (IntrinsicID) {
3834  case Intrinsic::dbg_declare:
3835  case Intrinsic::dbg_value:
3836  case Intrinsic::lifetime_end:
3837  break;
3838  default:
3839  return false;
3840  }
3841  }
3842 
3843  // If the cleanup return we are simplifying unwinds to the caller, this will
3844  // set UnwindDest to nullptr.
3845  BasicBlock *UnwindDest = RI->getUnwindDest();
3846  Instruction *DestEHPad = UnwindDest ? UnwindDest->getFirstNonPHI() : nullptr;
3847 
3848  // We're about to remove BB from the control flow. Before we do, sink any
3849  // PHINodes into the unwind destination. Doing this before changing the
3850  // control flow avoids some potentially slow checks, since we can currently
3851  // be certain that UnwindDest and BB have no common predecessors (since they
3852  // are both EH pads).
3853  if (UnwindDest) {
3854  // First, go through the PHI nodes in UnwindDest and update any nodes that
3855  // reference the block we are removing
3856  for (BasicBlock::iterator I = UnwindDest->begin(),
3857  IE = DestEHPad->getIterator();
3858  I != IE; ++I) {
3859  PHINode *DestPN = cast<PHINode>(I);
3860 
3861  int Idx = DestPN->getBasicBlockIndex(BB);
3862  // Since BB unwinds to UnwindDest, it has to be in the PHI node.
3863  assert(Idx != -1);
3864  // This PHI node has an incoming value that corresponds to a control
3865  // path through the cleanup pad we are removing. If the incoming
3866  // value is in the cleanup pad, it must be a PHINode (because we
3867  // verified above that the block is otherwise empty). Otherwise, the
3868  // value is either a constant or a value that dominates the cleanup
3869  // pad being removed.
3870  //
3871  // Because BB and UnwindDest are both EH pads, all of their
3872  // predecessors must unwind to these blocks, and since no instruction
3873  // can have multiple unwind destinations, there will be no overlap in
3874  // incoming blocks between SrcPN and DestPN.
3875  Value *SrcVal = DestPN->getIncomingValue(Idx);
3876  PHINode *SrcPN = dyn_cast<PHINode>(SrcVal);
3877 
3878  // Remove the entry for the block we are deleting.
3879  DestPN->removeIncomingValue(Idx, false);
3880 
3881  if (SrcPN && SrcPN->getParent() == BB) {
3882  // If the incoming value was a PHI node in the cleanup pad we are
3883  // removing, we need to merge that PHI node's incoming values into
3884  // DestPN.
3885  for (unsigned SrcIdx = 0, SrcE = SrcPN->getNumIncomingValues();
3886  SrcIdx != SrcE; ++SrcIdx) {
3887  DestPN->addIncoming(SrcPN->getIncomingValue(SrcIdx),
3888  SrcPN->getIncomingBlock(SrcIdx));
3889  }
3890  } else {
3891  // Otherwise, the incoming value came from above BB and
3892  // so we can just reuse it. We must associate all of BB's
3893  // predecessors with this value.
3894  for (auto *pred : predecessors(BB)) {
3895  DestPN->addIncoming(SrcVal, pred);
3896  }
3897  }
3898  }
3899 
3900  // Sink any remaining PHI nodes directly into UnwindDest.
3901  Instruction *InsertPt = DestEHPad;
3902  for (BasicBlock::iterator I = BB->begin(),
3903  IE = BB->getFirstNonPHI()->getIterator();
3904  I != IE;) {
3905  // The iterator must be incremented here because the instructions are
3906  // being moved to another block.
3907  PHINode *PN = cast<PHINode>(I++);
3908  if (PN->use_empty())
3909  // If the PHI node has no uses, just leave it. It will be erased
3910  // when we erase BB below.
3911  continue;
3912 
3913  // Otherwise, sink this PHI node into UnwindDest.
3914  // Any predecessors to UnwindDest which are not already represented
3915  // must be back edges which inherit the value from the path through
3916  // BB. In this case, the PHI value must reference itself.
3917  for (auto *pred : predecessors(UnwindDest))
3918  if (pred != BB)
3919  PN->addIncoming(PN, pred);
3920  PN->moveBefore(InsertPt);
3921  }
3922  }
3923 
3924  for (pred_iterator PI = pred_begin(BB), PE = pred_end(BB); PI != PE;) {
3925  // The iterator must be updated here because we are removing this pred.
3926  BasicBlock *PredBB = *PI++;
3927  if (UnwindDest == nullptr) {
3928  removeUnwindEdge(PredBB);
3929  } else {
3930  TerminatorInst *TI = PredBB->getTerminator();
3931  TI->replaceUsesOfWith(BB, UnwindDest);
3932  }
3933  }
3934 
3935  // The cleanup pad is now unreachable. Zap it.
3936  BB->eraseFromParent();
3937  return true;
3938 }
3939 
3940 // Try to merge two cleanuppads together.
3942  // Skip any cleanuprets which unwind to caller, there is nothing to merge
3943  // with.
3944  BasicBlock *UnwindDest = RI->getUnwindDest();
3945  if (!UnwindDest)
3946  return false;
3947 
3948  // This cleanupret isn't the only predecessor of this cleanuppad, it wouldn't
3949  // be safe to merge without code duplication.
3950  if (UnwindDest->getSinglePredecessor() != RI->getParent())
3951  return false;
3952 
3953  // Verify that our cleanuppad's unwind destination is another cleanuppad.
3954  auto *SuccessorCleanupPad = dyn_cast<CleanupPadInst>(&UnwindDest->front());
3955  if (!SuccessorCleanupPad)
3956  return false;
3957 
3958  CleanupPadInst *PredecessorCleanupPad = RI->getCleanupPad();
3959  // Replace any uses of the successor cleanupad with the predecessor pad
3960  // The only cleanuppad uses should be this cleanupret, it's cleanupret and
3961  // funclet bundle operands.
3962  SuccessorCleanupPad->replaceAllUsesWith(PredecessorCleanupPad);
3963  // Remove the old cleanuppad.
3964  SuccessorCleanupPad->eraseFromParent();
3965  // Now, we simply replace the cleanupret with a branch to the unwind
3966  // destination.
3967  BranchInst::Create(UnwindDest, RI->getParent());
3968  RI->eraseFromParent();
3969 
3970  return true;
3971 }
3972 
3973 bool SimplifyCFGOpt::SimplifyCleanupReturn(CleanupReturnInst *RI) {
3974  // It is possible to transiantly have an undef cleanuppad operand because we
3975  // have deleted some, but not all, dead blocks.
3976  // Eventually, this block will be deleted.
3977  if (isa<UndefValue>(RI->getOperand(0)))
3978  return false;
3979 
3980  if (mergeCleanupPad(RI))
3981  return true;
3982 
3983  if (removeEmptyCleanup(RI))
3984  return true;
3985 
3986  return false;
3987 }
3988 
3989 bool SimplifyCFGOpt::SimplifyReturn(ReturnInst *RI, IRBuilder<> &Builder) {
3990  BasicBlock *BB = RI->getParent();
3991  if (!BB->getFirstNonPHIOrDbg()->isTerminator())
3992  return false;
3993 
3994  // Find predecessors that end with branches.
3995  SmallVector<BasicBlock *, 8> UncondBranchPreds;
3996  SmallVector<BranchInst *, 8> CondBranchPreds;
3997  for (pred_iterator PI = pred_begin(BB), E = pred_end(BB); PI != E; ++PI) {
3998  BasicBlock *P = *PI;
3999  TerminatorInst *PTI = P->getTerminator();
4000  if (BranchInst *BI = dyn_cast<BranchInst>(PTI)) {
4001  if (BI->isUnconditional())
4002  UncondBranchPreds.push_back(P);
4003  else
4004  CondBranchPreds.push_back(BI);
4005  }
4006  }
4007 
4008  // If we found some, do the transformation!
4009  if (!UncondBranchPreds.empty() && DupRet) {
4010  while (!UncondBranchPreds.empty()) {
4011  BasicBlock *Pred = UncondBranchPreds.pop_back_val();
4012  DEBUG(dbgs() << "FOLDING: " << *BB
4013  << "INTO UNCOND BRANCH PRED: " << *Pred);
4014  (void)FoldReturnIntoUncondBranch(RI, BB, Pred);
4015  }
4016 
4017  // If we eliminated all predecessors of the block, delete the block now.
4018  if (pred_empty(BB)) {
4019  // We know there are no successors, so just nuke the block.
4020  BB->eraseFromParent();
4021  if (LoopHeaders)
4022  LoopHeaders->erase(BB);
4023  }
4024 
4025  return true;
4026  }
4027 
4028  // Check out all of the conditional branches going to this return
4029  // instruction. If any of them just select between returns, change the
4030  // branch itself into a select/return pair.
4031  while (!CondBranchPreds.empty()) {
4032  BranchInst *BI = CondBranchPreds.pop_back_val();
4033 
4034  // Check to see if the non-BB successor is also a return block.
4035  if (isa<ReturnInst>(BI->getSuccessor(0)->getTerminator()) &&
4036  isa<ReturnInst>(BI->getSuccessor(1)->getTerminator()) &&
4037  SimplifyCondBranchToTwoReturns(BI, Builder))
4038  return true;
4039  }
4040  return false;
4041 }
4042 
4043 bool SimplifyCFGOpt::SimplifyUnreachable(UnreachableInst *UI) {
4044  BasicBlock *BB = UI->getParent();
4045 
4046  bool Changed = false;
4047 
4048  // If there are any instructions immediately before the unreachable that can
4049  // be removed, do so.
4050  while (UI->getIterator() != BB->begin()) {
4051  BasicBlock::iterator BBI = UI->getIterator();
4052  --BBI;
4053  // Do not delete instructions that can have side effects which might cause
4054  // the unreachable to not be reachable; specifically, calls and volatile
4055  // operations may have this effect.
4056  if (isa<CallInst>(BBI) && !isa<DbgInfoIntrinsic>(BBI))
4057  break;
4058 
4059  if (BBI->mayHaveSideEffects()) {
4060  if (auto *SI = dyn_cast<StoreInst>(BBI)) {
4061  if (SI->isVolatile())
4062  break;
4063  } else if (auto *LI = dyn_cast<LoadInst>(BBI)) {
4064  if (LI->isVolatile())
4065  break;
4066  } else if (auto *RMWI = dyn_cast<AtomicRMWInst>(BBI)) {
4067  if (RMWI->isVolatile())
4068  break;
4069  } else if (auto *CXI = dyn_cast<AtomicCmpXchgInst>(BBI)) {
4070  if (CXI->isVolatile())
4071  break;
4072  } else if (isa<CatchPadInst>(BBI)) {
4073  // A catchpad may invoke exception object constructors and such, which
4074  // in some languages can be arbitrary code, so be conservative by
4075  // default.
4076  // For CoreCLR, it just involves a type test, so can be removed.
4079  break;
4080  } else if (!isa<FenceInst>(BBI) && !isa<VAArgInst>(BBI) &&
4081  !isa<LandingPadInst>(BBI)) {
4082  break;
4083  }
4084  // Note that deleting LandingPad's here is in fact okay, although it
4085  // involves a bit of subtle reasoning. If this inst is a LandingPad,
4086  // all the predecessors of this block will be the unwind edges of Invokes,
4087  // and we can therefore guarantee this block will be erased.
4088  }
4089 
4090  // Delete this instruction (any uses are guaranteed to be dead)
4091  if (!BBI->use_empty())
4092  BBI->replaceAllUsesWith(UndefValue::get(BBI->getType()));
4093  BBI->eraseFromParent();
4094  Changed = true;
4095  }
4096 
4097  // If the unreachable instruction is the first in the block, take a gander
4098  // at all of the predecessors of this instruction, and simplify them.
4099  if (&BB->front() != UI)
4100  return Changed;
4101 
4103  for (unsigned i = 0, e = Preds.size(); i != e; ++i) {
4104  TerminatorInst *TI = Preds[i]->getTerminator();
4105  IRBuilder<> Builder(TI);
4106  if (auto *BI = dyn_cast<BranchInst>(TI)) {
4107  if (BI->isUnconditional()) {
4108  if (BI->getSuccessor(0) == BB) {
4109  new UnreachableInst(TI->getContext(), TI);
4110  TI->eraseFromParent();
4111  Changed = true;
4112  }
4113  } else {
4114  if (BI->getSuccessor(0) == BB) {
4115  Builder.CreateBr(BI->getSuccessor(1));
4117  } else if (BI->getSuccessor(1) == BB) {
4118  Builder.CreateBr(BI->getSuccessor(0));
4120  Changed = true;
4121  }
4122  }
4123  } else if (auto *SI = dyn_cast<SwitchInst>(TI)) {
4124  for (auto i = SI->case_begin(), e = SI->case_end(); i != e;) {
4125  if (i->getCaseSuccessor() != BB) {
4126  ++i;
4127  continue;
4128  }
4129  BB->removePredecessor(SI->getParent());
4130  i = SI->removeCase(i);
4131  e = SI->case_end();
4132  Changed = true;
4133  }
4134  } else if (auto *II = dyn_cast<InvokeInst>(TI)) {
4135  if (II->getUnwindDest() == BB) {
4136  removeUnwindEdge(TI->getParent());
4137  Changed = true;
4138  }
4139  } else if (auto *CSI = dyn_cast<CatchSwitchInst>(TI)) {
4140  if (CSI->getUnwindDest() == BB) {
4141  removeUnwindEdge(TI->getParent());
4142  Changed = true;
4143  continue;
4144  }
4145 
4146  for (CatchSwitchInst::handler_iterator I = CSI->handler_begin(),
4147  E = CSI->handler_end();
4148  I != E; ++I) {
4149  if (*I == BB) {
4150  CSI->removeHandler(I);
4151  --I;
4152  --E;
4153  Changed = true;
4154  }
4155  }
4156  if (CSI->getNumHandlers() == 0) {
4157  BasicBlock *CatchSwitchBB = CSI->getParent();
4158  if (CSI->hasUnwindDest()) {
4159  // Redirect preds to the unwind dest
4160  CatchSwitchBB->replaceAllUsesWith(CSI->getUnwindDest());
4161  } else {
4162  // Rewrite all preds to unwind to caller (or from invoke to call).
4163  SmallVector<BasicBlock *, 8> EHPreds(predecessors(CatchSwitchBB));
4164  for (BasicBlock *EHPred : EHPreds)
4165  removeUnwindEdge(EHPred);
4166  }
4167  // The catchswitch is no longer reachable.
4168  new UnreachableInst(CSI->getContext(), CSI);
4169  CSI->eraseFromParent();
4170  Changed = true;
4171  }
4172  } else if (isa<CleanupReturnInst>(TI)) {
4173  new UnreachableInst(TI->getContext(), TI);
4174  TI->eraseFromParent();
4175  Changed = true;
4176  }
4177  }
4178 
4179  // If this block is now dead, remove it.
4180  if (pred_empty(BB) && BB != &BB->getParent()->getEntryBlock()) {
4181  // We know there are no successors, so just nuke the block.
4182  BB->eraseFromParent();
4183  if (LoopHeaders)
4184  LoopHeaders->erase(BB);
4185  return true;
4186  }
4187 
4188  return Changed;
4189 }
4190 
4192  assert(Cases.size() >= 1);
4193 
4194  array_pod_sort(Cases.begin(), Cases.end(), ConstantIntSortPredicate);
4195  for (size_t I = 1, E = Cases.size(); I != E; ++I) {
4196  if (Cases[I - 1]->getValue() != Cases[I]->getValue() + 1)
4197  return false;
4198  }
4199  return true;
4200 }
4201 
4202 /// Turn a switch with two reachable destinations into an integer range
4203 /// comparison and branch.
4204 static bool TurnSwitchRangeIntoICmp(SwitchInst *SI, IRBuilder<> &Builder) {
4205  assert(SI->getNumCases() > 1 && "Degenerate switch?");
4206 
4207  bool HasDefault =
4208  !isa<UnreachableInst>(SI->getDefaultDest()->getFirstNonPHIOrDbg());
4209 
4210  // Partition the cases into two sets with different destinations.
4211  BasicBlock *DestA = HasDefault ? SI->getDefaultDest() : nullptr;
4212  BasicBlock *DestB = nullptr;
4215 
4216  for (auto Case : SI->cases()) {
4217  BasicBlock *Dest = Case.getCaseSuccessor();
4218  if (!DestA)
4219  DestA = Dest;
4220  if (Dest == DestA) {
4221  CasesA.push_back(Case.getCaseValue());
4222  continue;
4223  }
4224  if (!DestB)
4225  DestB = Dest;
4226  if (Dest == DestB) {
4227  CasesB.push_back(Case.getCaseValue());
4228  continue;
4229  }
4230  return false; // More than two destinations.
4231  }
4232 
4233  assert(DestA && DestB &&
4234  "Single-destination switch should have been folded.");
4235  assert(DestA != DestB);
4236  assert(DestB != SI->getDefaultDest());
4237  assert(!CasesB.empty() && "There must be non-default cases.");
4238  assert(!CasesA.empty() || HasDefault);
4239 
4240  // Figure out if one of the sets of cases form a contiguous range.
4241  SmallVectorImpl<ConstantInt *> *ContiguousCases = nullptr;
4242  BasicBlock *ContiguousDest = nullptr;
4243  BasicBlock *OtherDest = nullptr;
4244  if (!CasesA.empty() && CasesAreContiguous(CasesA)) {
4245  ContiguousCases = &CasesA;
4246  ContiguousDest = DestA;
4247  OtherDest = DestB;
4248  } else if (CasesAreContiguous(CasesB)) {
4249  ContiguousCases = &CasesB;
4250  ContiguousDest = DestB;
4251  OtherDest = DestA;
4252  } else
4253  return false;
4254 
4255  // Start building the compare and branch.
4256 
4257  Constant *Offset = ConstantExpr::getNeg(ContiguousCases->back());
4258  Constant *NumCases =
4259  ConstantInt::get(Offset->getType(), ContiguousCases->size());
4260 
4261  Value *Sub = SI->getCondition();
4262  if (!Offset->isNullValue())
4263  Sub = Builder.CreateAdd(Sub, Offset, Sub->getName() + ".off");
4264 
4265  Value *Cmp;
4266  // If NumCases overflowed, then all possible values jump to the successor.
4267  if (NumCases->isNullValue() && !ContiguousCases->empty())
4268  Cmp = ConstantInt::getTrue(SI->getContext());
4269  else
4270  Cmp = Builder.CreateICmpULT(Sub, NumCases, "switch");
4271  BranchInst *NewBI = Builder.CreateCondBr(Cmp, ContiguousDest, OtherDest);
4272 
4273  // Update weight for the newly-created conditional branch.
4274  if (HasBranchWeights(SI)) {
4275  SmallVector<uint64_t, 8> Weights;
4276  GetBranchWeights(SI, Weights);
4277  if (Weights.size() == 1 + SI->getNumCases()) {
4278  uint64_t TrueWeight = 0;
4279  uint64_t FalseWeight = 0;
4280  for (size_t I = 0, E = Weights.size(); I != E; ++I) {
4281  if (SI->getSuccessor(I) == ContiguousDest)
4282  TrueWeight += Weights[I];
4283  else
4284  FalseWeight += Weights[I];
4285  }
4286  while (TrueWeight > UINT32_MAX || FalseWeight > UINT32_MAX) {
4287  TrueWeight /= 2;
4288  FalseWeight /= 2;
4289  }
4291  MDBuilder(SI->getContext())
4292  .createBranchWeights((uint32_t)TrueWeight,
4293  (uint32_t)FalseWeight));
4294  }
4295  }
4296 
4297  // Prune obsolete incoming values off the successors' PHI nodes.
4298  for (auto BBI = ContiguousDest->begin(); isa<PHINode>(BBI); ++BBI) {
4299  unsigned PreviousEdges = ContiguousCases->size();
4300  if (ContiguousDest == SI->getDefaultDest())
4301  ++PreviousEdges;
4302  for (unsigned I = 0, E = PreviousEdges - 1; I != E; ++I)
4303  cast<PHINode>(BBI)->removeIncomingValue(SI->getParent());
4304  }
4305  for (auto BBI = OtherDest->begin(); isa<PHINode>(BBI); ++BBI) {
4306  unsigned PreviousEdges = SI->getNumCases() - ContiguousCases->size();
4307  if (OtherDest == SI->getDefaultDest())
4308  ++PreviousEdges;
4309  for (unsigned I = 0, E = PreviousEdges - 1; I != E; ++I)
4310  cast<PHINode>(BBI)->removeIncomingValue(SI->getParent());
4311  }
4312 
4313  // Drop the switch.
4314  SI->eraseFromParent();
4315 
4316  return true;
4317 }
4318 
4319 /// Compute masked bits for the condition of a switch
4320 /// and use it to remove dead cases.
4322  const DataLayout &DL) {
4323  Value *Cond = SI->getCondition();
4324  unsigned Bits = Cond->getType()->getIntegerBitWidth();
4325  KnownBits Known = computeKnownBits(Cond, DL, 0, AC, SI);
4326 
4327  // We can also eliminate cases by determining that their values are outside of
4328  // the limited range of the condition based on how many significant (non-sign)
4329  // bits are in the condition value.
4330  unsigned ExtraSignBits = ComputeNumSignBits(Cond, DL, 0, AC, SI) - 1;
4331  unsigned MaxSignificantBitsInCond = Bits - ExtraSignBits;
4332 
4333  // Gather dead cases.
4335  for (auto &Case : SI->cases()) {
4336  const APInt &CaseVal = Case.getCaseValue()->getValue();
4337  if (Known.Zero.intersects(CaseVal) || !Known.One.isSubsetOf(CaseVal) ||
4338  (CaseVal.getMinSignedBits() > MaxSignificantBitsInCond)) {
4339  DeadCases.push_back(Case.getCaseValue());
4340  DEBUG(dbgs() << "SimplifyCFG: switch case " << CaseVal << " is dead.\n");
4341  }
4342  }
4343 
4344  // If we can prove that the cases must cover all possible values, the
4345  // default destination becomes dead and we can remove it. If we know some
4346  // of the bits in the value, we can use that to more precisely compute the
4347  // number of possible unique case values.
4348  bool HasDefault =
4349  !isa<UnreachableInst>(SI->getDefaultDest()->getFirstNonPHIOrDbg());
4350  const unsigned NumUnknownBits =
4351  Bits - (Known.Zero | Known.One).countPopulation();
4352  assert(NumUnknownBits <= Bits);
4353  if (HasDefault && DeadCases.empty() &&
4354  NumUnknownBits < 64 /* avoid overflow */ &&
4355  SI->getNumCases() == (1ULL << NumUnknownBits)) {
4356  DEBUG(dbgs() << "SimplifyCFG: switch default is dead.\n");
4357  BasicBlock *NewDefault =
4359  SI->setDefaultDest(&*NewDefault);
4360  SplitBlock(&*NewDefault, &NewDefault->front());
4361  auto *OldTI = NewDefault->getTerminator();
4362  new UnreachableInst(SI->getContext(), OldTI);
4364  return true;
4365  }
4366 
4367  SmallVector<uint64_t, 8> Weights;
4368  bool HasWeight = HasBranchWeights(SI);
4369  if (HasWeight) {
4370  GetBranchWeights(SI, Weights);
4371  HasWeight = (Weights.size() == 1 + SI->getNumCases());
4372  }
4373 
4374  // Remove dead cases from the switch.
4375  for (ConstantInt *DeadCase : DeadCases) {
4376  SwitchInst::CaseIt CaseI = SI->findCaseValue(DeadCase);
4377  assert(CaseI != SI->case_default() &&
4378  "Case was not found. Probably mistake in DeadCases forming.");
4379  if (HasWeight) {
4380  std::swap(Weights[CaseI->getCaseIndex() + 1], Weights.back());
4381  Weights.pop_back();
4382  }
4383 
4384  // Prune unused values from PHI nodes.
4385  CaseI->getCaseSuccessor()->removePredecessor(SI->getParent());
4386  SI->removeCase(CaseI);
4387  }
4388  if (HasWeight && Weights.size() >= 2) {
4389  SmallVector<uint32_t, 8> MDWeights(Weights.begin(), Weights.end());
4391  MDBuilder(SI->getParent()->getContext())
4392  .createBranchWeights(MDWeights));
4393  }
4394 
4395  return !DeadCases.empty();
4396 }
4397 
4398 /// If BB would be eligible for simplification by
4399 /// TryToSimplifyUncondBranchFromEmptyBlock (i.e. it is empty and terminated
4400 /// by an unconditional branch), look at the phi node for BB in the successor
4401 /// block and see if the incoming value is equal to CaseValue. If so, return
4402 /// the phi node, and set PhiIndex to BB's index in the phi node.
4404  BasicBlock *BB, int *PhiIndex) {
4405  if (BB->getFirstNonPHIOrDbg() != BB->getTerminator())
4406  return nullptr; // BB must be empty to be a candidate for simplification.
4407  if (!BB->getSinglePredecessor())
4408  return nullptr; // BB must be dominated by the switch.
4409 
4411  if (!Branch || !Branch->isUnconditional())
4412  return nullptr; // Terminator must be unconditional branch.
4413 
4414  BasicBlock *Succ = Branch->getSuccessor(0);
4415 
4416  BasicBlock::iterator I = Succ->begin();
4417  while (PHINode *PHI = dyn_cast<PHINode>(I++)) {
4418  int Idx = PHI->getBasicBlockIndex(BB);
4419  assert(Idx >= 0 && "PHI has no entry for predecessor?");
4420 
4421  Value *InValue = PHI->getIncomingValue(Idx);
4422  if (InValue != CaseValue)
4423  continue;
4424 
4425  *PhiIndex = Idx;
4426  return PHI;
4427  }
4428 
4429  return nullptr;
4430 }
4431 
4432 /// Try to forward the condition of a switch instruction to a phi node
4433 /// dominated by the switch, if that would mean that some of the destination
4434 /// blocks of the switch can be folded away.
4435 /// Returns true if a change is made.
4437  typedef DenseMap<PHINode *, SmallVector<int, 4>> ForwardingNodesMap;
4438  ForwardingNodesMap ForwardingNodes;
4439 
4440  for (auto Case : SI->cases()) {
4441  ConstantInt *CaseValue = Case.getCaseValue();
4442  BasicBlock *CaseDest = Case.getCaseSuccessor();
4443 
4444  int PhiIndex;
4445  PHINode *PHI =
4446  FindPHIForConditionForwarding(CaseValue, CaseDest, &PhiIndex);
4447  if (!PHI)
4448  continue;
4449 
4450  ForwardingNodes[PHI].push_back(PhiIndex);
4451  }
4452 
4453  bool Changed = false;
4454 
4455  for (ForwardingNodesMap::iterator I = ForwardingNodes.begin(),
4456  E = ForwardingNodes.end();
4457  I != E; ++I) {
4458  PHINode *Phi = I->first;
4459  SmallVectorImpl<int> &Indexes = I->second;
4460 
4461  if (Indexes.size() < 2)
4462  continue;
4463 
4464  for (size_t I = 0, E = Indexes.size(); I != E; ++I)
4465  Phi->setIncomingValue(Indexes[I], SI->getCondition());
4466  Changed = true;
4467  }
4468 
4469  return Changed;
4470 }
4471 
4472 /// Return true if the backend will be able to handle
4473 /// initializing an array of constants like C.
4475  if (C->isThreadDependent())
4476  return false;
4477  if (C->isDLLImportDependent())
4478  return false;
4479 
4480  if (!isa<ConstantFP>(C) && !isa<ConstantInt>(C) &&
4481  !isa<ConstantPointerNull>(C) && !isa<GlobalValue>(C) &&
4482  !isa<UndefValue>(C) && !isa<ConstantExpr>(C))
4483  return false;
4484 
4485  if (ConstantExpr *CE = dyn_cast<ConstantExpr>(C)) {
4486  if (!CE->isGEPWithNoNotionalOverIndexing())
4487  return false;
4488  if (!ValidLookupTableConstant(CE->getOperand(0), TTI))
4489  return false;
4490  }
4491 
4493  return false;
4494 
4495  return true;
4496 }
4497 
4498 /// If V is a Constant, return it. Otherwise, try to look up
4499 /// its constant value in ConstantPool, returning 0 if it's not there.
4500 static Constant *
4503  if (Constant *C = dyn_cast<Constant>(V))
4504  return C;
4505  return ConstantPool.lookup(V);
4506 }
4507 
4508 /// Try to fold instruction I into a constant. This works for
4509 /// simple instructions such as binary operations where both operands are
4510 /// constant or can be replaced by constants from the ConstantPool. Returns the
4511 /// resulting constant on success, 0 otherwise.
4512 static Constant *
4515  if (SelectInst *Select = dyn_cast<SelectInst>(I)) {
4516  Constant *A = LookupConstant(Select->getCondition(), ConstantPool);
4517  if (!A)
4518  return nullptr;
4519  if (A->isAllOnesValue())
4520  return LookupConstant(Select->getTrueValue(), ConstantPool);
4521  if (A->isNullValue())
4522  return LookupConstant(Select->getFalseValue(), ConstantPool);
4523  return nullptr;
4524  }
4525 
4527  for (unsigned N = 0, E = I->getNumOperands(); N != E; ++N) {
4529  COps.push_back(A);
4530  else
4531  return nullptr;
4532  }
4533 
4534  if (CmpInst *Cmp = dyn_cast<CmpInst>(I)) {
4535  return ConstantFoldCompareInstOperands(Cmp->getPredicate(), COps[0],
4536  COps[1], DL);
4537  }
4538 
4539  return ConstantFoldInstOperands(I, COps, DL);
4540 }
4541 
4542 /// Try to determine the resulting constant values in phi nodes
4543 /// at the common destination basic block, *CommonDest, for one of the case
4544 /// destionations CaseDest corresponding to value CaseVal (0 for the default
4545 /// case), of a switch instruction SI.
4546 static bool
4548  BasicBlock **CommonDest,
4549  SmallVectorImpl<std::pair<PHINode *, Constant *>> &Res,
4550  const DataLayout &DL, const TargetTransformInfo &TTI) {
4551  // The block from which we enter the common destination.
4552  BasicBlock *Pred = SI->getParent();
4553 
4554  // If CaseDest is empty except for some side-effect free instructions through
4555  // which we can constant-propagate the CaseVal, continue to its successor.
4557  ConstantPool.insert(std::make_pair(SI->getCondition(), CaseVal));
4558  for (BasicBlock::iterator I = CaseDest->begin(), E = CaseDest->end(); I != E;
4559  ++I) {
4560  if (TerminatorInst *T = dyn_cast<TerminatorInst>(I)) {
4561  // If the terminator is a simple branch, continue to the next block.
4562  if (T->getNumSuccessors() != 1 || T->isExceptional())
4563  return false;
4564  Pred = CaseDest;
4565  CaseDest = T->getSuccessor(0);
4566  } else if (isa<DbgInfoIntrinsic>(I)) {
4567  // Skip debug intrinsic.
4568  continue;
4569  } else if (Constant *C = ConstantFold(&*I, DL, ConstantPool)) {
4570  // Instruction is side-effect free and constant.
4571 
4572  // If the instruction has uses outside this block or a phi node slot for
4573  // the block, it is not safe to bypass the instruction since it would then
4574  // no longer dominate all its uses.
4575  for (auto &Use : I->uses()) {
4576  User *User = Use.getUser();
4577  if (Instruction *I = dyn_cast<Instruction>(User))
4578  if (I->getParent() == CaseDest)
4579  continue;
4580  if (PHINode *Phi = dyn_cast<PHINode>(User))
4581  if (Phi->getIncomingBlock(Use) == CaseDest)
4582  continue;
4583  return false;
4584  }
4585 
4586  ConstantPool.insert(std::make_pair(&*I, C));
4587  } else {
4588  break;
4589  }
4590  }
4591 
4592  // If we did not have a CommonDest before, use the current one.
4593  if (!*CommonDest)
4594  *CommonDest = CaseDest;
4595  // If the destination isn't the common one, abort.
4596  if (CaseDest != *CommonDest)
4597  return false;
4598 
4599  // Get the values for this case from phi nodes in the destination block.
4600  BasicBlock::iterator I = (*CommonDest)->begin();
4601  while (PHINode *PHI = dyn_cast<PHINode>(I++)) {
4602  int Idx = PHI->getBasicBlockIndex(Pred);
4603  if (Idx == -1)
4604  continue;
4605 
4606  Constant *ConstVal =
4607  LookupConstant(PHI->getIncomingValue(Idx), ConstantPool);
4608  if (!ConstVal)
4609  return false;
4610 
4611  // Be conservative about which kinds of constants we support.
4612  if (!ValidLookupTableConstant(ConstVal, TTI))
4613  return false;
4614 
4615  Res.push_back(std::make_pair(PHI, ConstVal));
4616  }
4617 
4618  return Res.size() > 0;
4619 }
4620 
4621 // Helper function used to add CaseVal to the list of cases that generate
4622 // Result.
4623 static void MapCaseToResult(ConstantInt *CaseVal,
4624  SwitchCaseResultVectorTy &UniqueResults,
4625  Constant *Result) {
4626  for (auto &I : UniqueResults) {
4627  if (I.first == Result) {
4628  I.second.push_back(CaseVal);
4629  return;
4630  }
4631  }
4632  UniqueResults.push_back(
4633  std::make_pair(Result, SmallVector<ConstantInt *, 4>(1, CaseVal)));
4634 }
4635 
4636 // Helper function that initializes a map containing
4637 // results for the PHI node of the common destination block for a switch
4638 // instruction. Returns false if multiple PHI nodes have been found or if
4639 // there is not a common destination block for the switch.
4641  BasicBlock *&CommonDest,
4642  SwitchCaseResultVectorTy &UniqueResults,
4643  Constant *&DefaultResult,
4644  const DataLayout &DL,
4645  const TargetTransformInfo &TTI) {
4646  for (auto &I : SI->cases()) {
4647  ConstantInt *CaseVal = I.getCaseValue();
4648 
4649  // Resulting value at phi nodes for this case value.
4650  SwitchCaseResultsTy Results;
4651  if (!GetCaseResults(SI, CaseVal, I.getCaseSuccessor(), &CommonDest, Results,
4652  DL, TTI))
4653  return false;
4654 
4655  // Only one value per case is permitted
4656  if (Results.size() > 1)
4657  return false;
4658  MapCaseToResult(CaseVal, UniqueResults, Results.begin()->second);
4659 
4660  // Check the PHI consistency.
4661  if (!PHI)
4662  PHI = Results[0].first;
4663  else if (PHI != Results[0].first)
4664  return false;
4665  }
4666  // Find the default result value.
4668  BasicBlock *DefaultDest = SI->getDefaultDest();
4669  GetCaseResults(SI, nullptr, SI->getDefaultDest(), &CommonDest, DefaultResults,
4670  DL, TTI);
4671  // If the default value is not found abort unless the default destination
4672  // is unreachable.
4673  DefaultResult =
4674  DefaultResults.size() == 1 ? DefaultResults.begin()->second : nullptr;
4675  if ((!DefaultResult &&
4676  !isa<UnreachableInst>(DefaultDest->getFirstNonPHIOrDbg())))
4677  return false;
4678 
4679  return true;
4680 }
4681 
4682 // Helper function that checks if it is possible to transform a switch with only
4683 // two cases (or two cases + default) that produces a result into a select.
4684 // Example:
4685 // switch (a) {
4686 // case 10: %0 = icmp eq i32 %a, 10
4687 // return 10; %1 = select i1 %0, i32 10, i32 4
4688 // case 20: ----> %2 = icmp eq i32 %a, 20
4689 // return 2; %3 = select i1 %2, i32 2, i32 %1
4690 // default:
4691 // return 4;
4692 // }
4693 static Value *ConvertTwoCaseSwitch(const SwitchCaseResultVectorTy &ResultVector,
4694  Constant *DefaultResult, Value *Condition,
4695  IRBuilder<> &Builder) {
4696  assert(ResultVector.size() == 2 &&
4697  "We should have exactly two unique results at this point");
4698  // If we are selecting between only two cases transform into a simple
4699  // select or a two-way select if default is possible.
4700  if (ResultVector[0].second.size() == 1 &&
4701  ResultVector[1].second.size() == 1) {
4702  ConstantInt *const FirstCase = ResultVector[0].second[0];
4703  ConstantInt *const SecondCase = ResultVector[1].second[0];
4704 
4705  bool DefaultCanTrigger = DefaultResult;
4706  Value *SelectValue = ResultVector[1].first;
4707  if (DefaultCanTrigger) {
4708  Value *const ValueCompare =
4709  Builder.CreateICmpEQ(Condition, SecondCase, "switch.selectcmp");
4710  SelectValue = Builder.CreateSelect(ValueCompare, ResultVector[1].first,
4711  DefaultResult, "switch.select");
4712  }
4713  Value *const ValueCompare =
4714  Builder.CreateICmpEQ(Condition, FirstCase, "switch.selectcmp");
4715  return Builder.CreateSelect(ValueCompare, ResultVector[0].first,
4716  SelectValue, "switch.select");
4717  }
4718 
4719  return nullptr;
4720 }
4721 
4722 // Helper function to cleanup a switch instruction that has been converted into
4723 // a select, fixing up PHI nodes and basic blocks.
4725  Value *SelectValue,
4726  IRBuilder<> &Builder) {
4727  BasicBlock *SelectBB = SI->getParent();
4728  while (PHI->getBasicBlockIndex(SelectBB) >= 0)
4729  PHI->removeIncomingValue(SelectBB);
4730  PHI->addIncoming(SelectValue, SelectBB);
4731 
4732  Builder.CreateBr(PHI->getParent());
4733 
4734  // Remove the switch.
4735  for (unsigned i = 0, e = SI->getNumSuccessors(); i < e; ++i) {
4736  BasicBlock *Succ = SI->getSuccessor(i);
4737 
4738  if (Succ == PHI->getParent())
4739  continue;
4740  Succ->removePredecessor(SelectBB);
4741  }
4742  SI->eraseFromParent();
4743 }
4744 
4745 /// If the switch is only used to initialize one or more
4746 /// phi nodes in a common successor block with only two different
4747 /// constant values, replace the switch with select.
4748 static bool SwitchToSelect(SwitchInst *SI, IRBuilder<> &Builder,
4749  AssumptionCache *AC, const DataLayout &DL,
4750  const TargetTransformInfo &TTI) {
4751  Value *const Cond = SI->getCondition();
4752  PHINode *PHI = nullptr;
4753  BasicBlock *CommonDest = nullptr;
4754  Constant *DefaultResult;
4755  SwitchCaseResultVectorTy UniqueResults;
4756  // Collect all the cases that will deliver the same value from the switch.
4757  if (!InitializeUniqueCases(SI, PHI, CommonDest, UniqueResults, DefaultResult,
4758  DL, TTI))
4759  return false;
4760  // Selects choose between maximum two values.
4761  if (UniqueResults.size() != 2)
4762  return false;
4763  assert(PHI != nullptr && "PHI for value select not found");
4764 
4765  Builder.SetInsertPoint(SI);
4766  Value *SelectValue =
4767  ConvertTwoCaseSwitch(UniqueResults, DefaultResult, Cond, Builder);
4768  if (SelectValue) {
4769  RemoveSwitchAfterSelectConversion(SI, PHI, SelectValue, Builder);
4770  return true;
4771  }
4772  // The switch couldn't be converted into a select.
4773  return false;
4774 }
4775 
4776 namespace {
4777 
4778 /// This class represents a lookup table that can be used to replace a switch.
4779 class SwitchLookupTable {
4780 public:
4781  /// Create a lookup table to use as a switch replacement with the contents
4782  /// of Values, using DefaultValue to fill any holes in the table.
4783  SwitchLookupTable(
4784  Module &M, uint64_t TableSize, ConstantInt *Offset,
4785  const SmallVectorImpl<std::pair<ConstantInt *, Constant *>> &Values,
4786  Constant *DefaultValue, const DataLayout &DL, const StringRef &FuncName);
4787 
4788  /// Build instructions with Builder to retrieve the value at
4789  /// the position given by Index in the lookup table.
4790  Value *BuildLookup(Value *Index, IRBuilder<> &Builder);
4791 
4792  /// Return true if a table with TableSize elements of
4793  /// type ElementType would fit in a target-legal register.
4794  static bool WouldFitInRegister(const DataLayout &DL, uint64_t TableSize,
4795  Type *ElementType);
4796 
4797 private:
4798  // Depending on the contents of the table, it can be represented in
4799  // different ways.
4800  enum {
4801  // For tables where each element contains the same value, we just have to
4802  // store that single value and return it for each lookup.
4803  SingleValueKind,
4804 
4805  // For tables where there is a linear relationship between table index
4806  // and values. We calculate the result with a simple multiplication
4807  // and addition instead of a table lookup.
4808  LinearMapKind,
4809 
4810  // For small tables with integer elements, we can pack them into a bitmap
4811  // that fits into a target-legal register. Values are retrieved by
4812  // shift and mask operations.
4813  BitMapKind,
4814 
4815  // The table is stored as an array of values. Values are retrieved by load
4816  // instructions from the table.
4817  ArrayKind
4818  } Kind;
4819 
4820  // For SingleValueKind, this is the single value.
4821  Constant *SingleValue;
4822 
4823  // For BitMapKind, this is the bitmap.
4824  ConstantInt *BitMap;
4825  IntegerType *BitMapElementTy;
4826 
4827  // For LinearMapKind, these are the constants used to derive the value.
4828  ConstantInt *LinearOffset;
4829  ConstantInt *LinearMultiplier;
4830 
4831  // For ArrayKind, this is the array.
4832  GlobalVariable *Array;
4833 };
4834 
4835 } // end anonymous namespace
4836 
4837 SwitchLookupTable::SwitchLookupTable(
4838  Module &M, uint64_t TableSize, ConstantInt *Offset,
4839  const SmallVectorImpl<std::pair<ConstantInt *, Constant *>> &Values,
4840  Constant *DefaultValue, const DataLayout &DL, const StringRef &FuncName)
4841  : SingleValue(nullptr), BitMap(nullptr), BitMapElementTy(nullptr),
4842  LinearOffset(nullptr), LinearMultiplier(nullptr), Array(nullptr) {
4843  assert(Values.size() && "Can't build lookup table without values!");
4844  assert(TableSize >= Values.size() && "Can't fit values in table!");
4845 
4846  // If all values in the table are equal, this is that value.
4847  SingleValue = Values.begin()->second;
4848 
4849  Type *ValueType = Values.begin()->second->getType();
4850 
4851  // Build up the table contents.
4852  SmallVector<Constant *, 64> TableContents(TableSize);
4853  for (size_t I = 0, E = Values.size(); I != E; ++I) {
4854  ConstantInt *CaseVal = Values[I].first;
4855  Constant *CaseRes = Values[I].second;
4856  assert(CaseRes->getType() == ValueType);
4857 
4858  uint64_t Idx = (CaseVal->getValue() - Offset->getValue()).getLimitedValue();
4859  TableContents[Idx] = CaseRes;
4860 
4861  if (CaseRes != SingleValue)
4862  SingleValue = nullptr;
4863  }
4864 
4865  // Fill in any holes in the table with the default result.
4866  if (Values.size() < TableSize) {
4867  assert(DefaultValue &&
4868  "Need a default value to fill the lookup table holes.");
4869  assert(DefaultValue->getType() == ValueType);
4870  for (uint64_t I = 0; I < TableSize; ++I) {
4871  if (!TableContents[I])
4872  TableContents[I] = DefaultValue;
4873  }
4874 
4875  if (DefaultValue != SingleValue)
4876  SingleValue = nullptr;
4877  }
4878 
4879  // If each element in the table contains the same value, we only need to store
4880  // that single value.
4881  if (SingleValue) {
4882  Kind = SingleValueKind;
4883  return;
4884  }
4885 
4886  // Check if we can derive the value with a linear transformation from the
4887  // table index.
4888  if (isa<IntegerType>(ValueType)) {
4889  bool LinearMappingPossible = true;
4890  APInt PrevVal;
4891  APInt DistToPrev;
4892  assert(TableSize >= 2 && "Should be a SingleValue table.");
4893  // Check if there is the same distance between two consecutive values.
4894  for (uint64_t I = 0; I < TableSize; ++I) {
4895  ConstantInt *ConstVal = dyn_cast<ConstantInt>(TableContents[I]);
4896  if (!ConstVal) {
4897  // This is an undef. We could deal with it, but undefs in lookup tables
4898  // are very seldom. It's probably not worth the additional complexity.
4899  LinearMappingPossible = false;
4900  break;
4901  }
4902  const APInt &Val = ConstVal->getValue();
4903  if (I != 0) {
4904  APInt Dist = Val - PrevVal;
4905  if (I == 1) {
4906  DistToPrev = Dist;
4907  } else if (Dist != DistToPrev) {
4908  LinearMappingPossible = false;
4909  break;
4910  }
4911  }
4912  PrevVal = Val;
4913  }
4914  if (LinearMappingPossible) {
4915  LinearOffset = cast<ConstantInt>(TableContents[0]);
4916  LinearMultiplier = ConstantInt::get(M.getContext(), DistToPrev);
4917  Kind = LinearMapKind;
4918  ++NumLinearMaps;
4919  return;
4920  }
4921  }
4922 
4923  // If the type is integer and the table fits in a register, build a bitmap.
4924  if (WouldFitInRegister(DL, TableSize, ValueType)) {
4925  IntegerType *IT = cast<IntegerType>(ValueType);
4926  APInt TableInt(TableSize * IT->getBitWidth(), 0);
4927  for (uint64_t I = TableSize; I > 0; --I) {
4928  TableInt <<= IT->getBitWidth();
4929  // Insert values into the bitmap. Undef values are set to zero.
4930  if (!isa<UndefValue>(TableContents[I - 1])) {
4931  ConstantInt *Val = cast<ConstantInt>(TableContents[I - 1]);
4932  TableInt |= Val->getValue().zext(TableInt.getBitWidth());
4933  }
4934  }
4935  BitMap = ConstantInt::get(M.getContext(), TableInt);
4936  BitMapElementTy = IT;
4937  Kind = BitMapKind;
4938  ++NumBitMaps;
4939  return;
4940  }
4941 
4942  // Store the table in an array.
4943  ArrayType *ArrayTy = ArrayType::get(ValueType, TableSize);
4944  Constant *Initializer = ConstantArray::get(ArrayTy, TableContents);
4945 
4946  Array = new GlobalVariable(M, ArrayTy, /*constant=*/true,
4947  GlobalVariable::PrivateLinkage, Initializer,
4948  "switch.table." + FuncName);
4949  Array->setUnnamedAddr(GlobalValue::UnnamedAddr::Global);
4950  Kind = ArrayKind;
4951 }
4952 
4953 Value *SwitchLookupTable::BuildLookup(Value *Index, IRBuilder<> &Builder) {
4954  switch (Kind) {
4955  case SingleValueKind:
4956  return SingleValue;
4957  case LinearMapKind: {
4958  // Derive the result value from the input value.
4959  Value *Result = Builder.CreateIntCast(Index, LinearMultiplier->getType(),
4960  false, "switch.idx.cast");
4961  if (!LinearMultiplier->isOne())
4962  Result = Builder.CreateMul(Result, LinearMultiplier, "switch.idx.mult");
4963  if (!LinearOffset->isZero())
4964  Result = Builder.CreateAdd(Result, LinearOffset, "switch.offset");
4965  return Result;
4966  }
4967  case BitMapKind: {
4968  // Type of the bitmap (e.g. i59).
4969  IntegerType *MapTy = BitMap->getType();
4970 
4971  // Cast Index to the same type as the bitmap.
4972  // Note: The Index is <= the number of elements in the table, so
4973  // truncating it to the width of the bitmask is safe.
4974  Value *ShiftAmt = Builder.CreateZExtOrTrunc(Index, MapTy, "switch.cast");
4975 
4976  // Multiply the shift amount by the element width.
4977  ShiftAmt = Builder.CreateMul(
4978  ShiftAmt, ConstantInt::get(MapTy, BitMapElementTy->getBitWidth()),
4979  "switch.shiftamt");
4980 
4981  // Shift down.
4982  Value *DownShifted =
4983  Builder.CreateLShr(BitMap, ShiftAmt, "switch.downshift");
4984  // Mask off.
4985  return Builder.CreateTrunc(DownShifted, BitMapElementTy, "switch.masked");
4986  }
4987  case ArrayKind: {
4988  // Make sure the table index will not overflow when treated as signed.
4989  IntegerType *IT = cast<IntegerType>(Index->getType());
4990  uint64_t TableSize =
4991  Array->getInitializer()->getType()->getArrayNumElements();
4992  if (TableSize > (1ULL << (IT->getBitWidth() - 1)))
4993  Index = Builder.CreateZExt(
4994  Index, IntegerType::get(IT->getContext(), IT->getBitWidth() + 1),
4995  "switch.tableidx.zext");
4996 
4997  Value *GEPIndices[] = {Builder.getInt32(0), Index};
4998  Value *GEP = Builder.CreateInBoundsGEP(Array->getValueType(), Array,
4999  GEPIndices, "switch.gep");
5000  return Builder.CreateLoad(GEP, "switch.load");
5001  }
5002  }
5003  llvm_unreachable("Unknown lookup table kind!");
5004 }
5005 
5006 bool SwitchLookupTable::WouldFitInRegister(const DataLayout &DL,
5007  uint64_t TableSize,
5008  Type *ElementType) {
5009  auto *IT = dyn_cast<IntegerType>(ElementType);
5010  if (!IT)
5011  return false;
5012  // FIXME: If the type is wider than it needs to be, e.g. i8 but all values
5013  // are <= 15, we could try to narrow the type.
5014 
5015  // Avoid overflow, fitsInLegalInteger uses unsigned int for the width.
5016  if (TableSize >= UINT_MAX / IT->getBitWidth())
5017  return false;
5018  return DL.fitsInLegalInteger(TableSize * IT->getBitWidth());
5019 }
5020 
5021 /// Determine whether a lookup table should be built for this switch, based on
5022 /// the number of cases, size of the table, and the types of the results.
5023 static bool
5024 ShouldBuildLookupTable(SwitchInst *SI, uint64_t TableSize,
5025  const TargetTransformInfo &TTI, const DataLayout &DL,
5026  const SmallDenseMap<PHINode *, Type *> &ResultTypes) {
5027  if (SI->getNumCases() > TableSize || TableSize >= UINT64_MAX / 10)
5028  return false; // TableSize overflowed, or mul below might overflow.
5029 
5030  bool AllTablesFitInRegister = true;
5031  bool HasIllegalType = false;
5032  for (const auto &I : ResultTypes) {
5033  Type *Ty = I.second;
5034 
5035  // Saturate this flag to true.
5036  HasIllegalType = HasIllegalType || !TTI.isTypeLegal(Ty);
5037 
5038  // Saturate this flag to false.
5039  AllTablesFitInRegister =
5040  AllTablesFitInRegister &&
5041  SwitchLookupTable::WouldFitInRegister(DL, TableSize, Ty);
5042 
5043  // If both flags saturate, we're done. NOTE: This *only* works with
5044  // saturating flags, and all flags have to saturate first due to the
5045  // non-deterministic behavior of iterating over a dense map.
5046  if (HasIllegalType && !AllTablesFitInRegister)
5047  break;
5048  }
5049 
5050  // If each table would fit in a register, we should build it anyway.
5051  if (AllTablesFitInRegister)
5052  return true;
5053 
5054  // Don't build a table that doesn't fit in-register if it has illegal types.
5055  if (HasIllegalType)
5056  return false;
5057 
5058  // The table density should be at least 40%. This is the same criterion as for
5059  // jump tables, see SelectionDAGBuilder::handleJTSwitchCase.
5060  // FIXME: Find the best cut-off.
5061  return SI->getNumCases() * 10 >= TableSize * 4;
5062 }
5063 
5064 /// Try to reuse the switch table index compare. Following pattern:
5065 /// \code
5066 /// if (idx < tablesize)
5067 /// r = table[idx]; // table does not contain default_value
5068 /// else
5069 /// r = default_value;
5070 /// if (r != default_value)
5071 /// ...
5072 /// \endcode
5073 /// Is optimized to:
5074 /// \code
5075 /// cond = idx < tablesize;
5076 /// if (cond)
5077 /// r = table[idx];
5078 /// else
5079 /// r = default_value;
5080 /// if (cond)
5081 /// ...
5082 /// \endcode
5083 /// Jump threading will then eliminate the second if(cond).
5084 static void reuseTableCompare(
5085  User *PhiUser, BasicBlock *PhiBlock, BranchInst *RangeCheckBranch,
5086  Constant *DefaultValue,
5087  const SmallVectorImpl<std::pair<ConstantInt *, Constant *>> &Values) {
5088 
5089  ICmpInst *CmpInst = dyn_cast<ICmpInst>(PhiUser);
5090  if (!CmpInst)
5091  return;
5092 
5093  // We require that the compare is in the same block as the phi so that jump
5094  // threading can do its work afterwards.
5095  if (CmpInst->getParent() != PhiBlock)
5096  return;
5097 
5098  Constant *CmpOp1 = dyn_cast<Constant>(CmpInst->getOperand(1));
5099  if (!CmpOp1)
5100  return;
5101 
5102  Value *RangeCmp = RangeCheckBranch->getCondition();
5103  Constant *TrueConst = ConstantInt::getTrue(RangeCmp->getType());
5104  Constant *FalseConst = ConstantInt::getFalse(RangeCmp->getType());
5105 
5106  // Check if the compare with the default value is constant true or false.
5107  Constant *DefaultConst = ConstantExpr::getICmp(CmpInst->getPredicate(),
5108  DefaultValue, CmpOp1, true);
5109  if (DefaultConst != TrueConst && DefaultConst != FalseConst)
5110  return;
5111 
5112  // Check if the compare with the case values is distinct from the default
5113  // compare result.
5114  for (auto ValuePair : Values) {
5115  Constant *CaseConst = ConstantExpr::getICmp(CmpInst->getPredicate(),
5116  ValuePair.second, CmpOp1, true);
5117  if (!CaseConst || CaseConst == DefaultConst)
5118  return;
5119  assert((CaseConst == TrueConst || CaseConst == FalseConst) &&
5120  "Expect true or false as compare result.");
5121  }
5122 
5123  // Check if the branch instruction dominates the phi node. It's a simple
5124  // dominance check, but sufficient for our needs.
5125  // Although this check is invariant in the calling loops, it's better to do it
5126  // at this late stage. Practically we do it at most once for a switch.
5127  BasicBlock *BranchBlock = RangeCheckBranch->getParent();
5128  for (auto PI = pred_begin(PhiBlock), E = pred_end(PhiBlock); PI != E; ++PI) {
5129  BasicBlock *Pred = *PI;
5130  if (Pred != BranchBlock && Pred->getUniquePredecessor() != BranchBlock)
5131  return;
5132  }
5133 
5134  if (DefaultConst == FalseConst) {
5135  // The compare yields the same result. We can replace it.
5136  CmpInst->replaceAllUsesWith(RangeCmp);
5137  ++NumTableCmpReuses;
5138  } else {
5139  // The compare yields the same result, just inverted. We can replace it.
5140  Value *InvertedTableCmp = BinaryOperator::CreateXor(
5141  RangeCmp, ConstantInt::get(RangeCmp->getType(), 1), "inverted.cmp",
5142  RangeCheckBranch);
5143  CmpInst->replaceAllUsesWith(InvertedTableCmp);
5144  ++NumTableCmpReuses;
5145  }
5146 }
5147 
5148 /// If the switch is only used to initialize one or more phi nodes in a common
5149 /// successor block with different constant values, replace the switch with
5150 /// lookup tables.
5151 static bool SwitchToLookupTable(SwitchInst *SI, IRBuilder<> &Builder,
5152  const DataLayout &DL,
5153  const TargetTransformInfo &TTI) {
5154  assert(SI->getNumCases() > 1 && "Degenerate switch?");
5155 
5156  Function *Fn = SI->getParent()->getParent();
5157  // Only build lookup table when we have a target that supports it or the
5158  // attribute is not set.
5159  if (!TTI.shouldBuildLookupTables() ||
5160  (Fn->getFnAttribute("no-jump-tables").getValueAsString() == "true"))
5161  return false;
5162 
5163  // FIXME: If the switch is too sparse for a lookup table, perhaps we could
5164  // split off a dense part and build a lookup table for that.
5165 
5166  // FIXME: This creates arrays of GEPs to constant strings, which means each
5167  // GEP needs a runtime relocation in PIC code. We should just build one big
5168  // string and lookup indices into that.
5169 
5170  // Ignore switches with less than three cases. Lookup tables will not make
5171  // them
5172  // faster, so we don't analyze them.
5173  if (SI->getNumCases() < 3)
5174  return false;
5175 
5176  // Figure out the corresponding result for each case value and phi node in the
5177  // common destination, as well as the min and max case values.
5178  assert(SI->case_begin() != SI->case_end());
5179  SwitchInst::CaseIt CI = SI->case_begin();
5180  ConstantInt *MinCaseVal = CI->getCaseValue();
5181  ConstantInt *MaxCaseVal = CI->getCaseValue();
5182 
5183  BasicBlock *CommonDest = nullptr;
5184  typedef SmallVector<std::pair<ConstantInt *, Constant *>, 4> ResultListTy;
5186  SmallDenseMap<PHINode *, Constant *> DefaultResults;
5189 
5190  for (SwitchInst::CaseIt E = SI->case_end(); CI != E; ++CI) {
5191  ConstantInt *CaseVal = CI->getCaseValue();
5192  if (CaseVal->getValue().slt(MinCaseVal->getValue()))
5193  MinCaseVal = CaseVal;
5194  if (CaseVal->getValue().sgt(MaxCaseVal->getValue()))
5195  MaxCaseVal = CaseVal;
5196 
5197  // Resulting value at phi nodes for this case value.
5198  typedef SmallVector<std::pair<PHINode *, Constant *>, 4> ResultsTy;
5199  ResultsTy Results;
5200  if (!GetCaseResults(SI, CaseVal, CI->getCaseSuccessor(), &CommonDest,
5201  Results, DL, TTI))
5202  return false;
5203 
5204  // Append the result from this case to the list for each phi.
5205  for (const auto &I : Results) {
5206  PHINode *PHI = I.first;
5207  Constant *Value = I.second;
5208  if (!ResultLists.count(PHI))
5209  PHIs.push_back(PHI);
5210  ResultLists[PHI].push_back(std::make_pair(CaseVal, Value));
5211  }
5212  }
5213 
5214  // Keep track of the result types.
5215  for (PHINode *PHI : PHIs) {
5216  ResultTypes[PHI] = ResultLists[PHI][0].second->getType();
5217  }
5218 
5219  uint64_t NumResults = ResultLists[PHIs[0]].size();
5220  APInt RangeSpread = MaxCaseVal->getValue() - MinCaseVal->getValue();
5221  uint64_t TableSize = RangeSpread.getLimitedValue() + 1;
5222  bool TableHasHoles = (NumResults < TableSize);
5223 
5224  // If the table has holes, we need a constant result for the default case
5225  // or a bitmask that fits in a register.
5226  SmallVector<std::pair<PHINode *, Constant *>, 4> DefaultResultsList;
5227  bool HasDefaultResults =
5228  GetCaseResults(SI, nullptr, SI->getDefaultDest(), &CommonDest,
5229  DefaultResultsList, DL, TTI);
5230 
5231  bool NeedMask = (TableHasHoles && !HasDefaultResults);
5232  if (NeedMask) {
5233  // As an extra penalty for the validity test we require more cases.
5234  if (SI->getNumCases() < 4) // FIXME: Find best threshold value (benchmark).
5235  return false;
5236  if (!DL.fitsInLegalInteger(TableSize))
5237  return false;
5238  }
5239 
5240  for (const auto &I : DefaultResultsList) {
5241  PHINode *PHI = I.first;
5242  Constant *Result = I.second;
5243  DefaultResults[PHI] = Result;
5244  }
5245 
5246  if (!ShouldBuildLookupTable(SI, TableSize, TTI, DL, ResultTypes))
5247  return false;
5248 
5249  // Create the BB that does the lookups.
5250  Module &Mod = *CommonDest->getParent()->getParent();
5251  BasicBlock *LookupBB = BasicBlock::Create(
5252  Mod.getContext(), "switch.lookup", CommonDest->getParent(), CommonDest);
5253 
5254  // Compute the table index value.
5255  Builder.SetInsertPoint(SI);
5256  Value *TableIndex =
5257  Builder.CreateSub(SI->getCondition(), MinCaseVal, "switch.tableidx");
5258 
5259  // Compute the maximum table size representable by the integer type we are
5260  // switching upon.
5261  unsigned CaseSize = MinCaseVal->getType()->getPrimitiveSizeInBits();
5262  uint64_t MaxTableSize = CaseSize > 63 ? UINT64_MAX : 1ULL << CaseSize;
5263  assert(MaxTableSize >= TableSize &&
5264  "It is impossible for a switch to have more entries than the max "
5265  "representable value of its input integer type's size.");
5266 
5267  // If the default destination is unreachable, or if the lookup table covers
5268  // all values of the conditional variable, branch directly to the lookup table
5269  // BB. Otherwise, check that the condition is within the case range.
5270  const bool DefaultIsReachable =
5271  !isa<UnreachableInst>(SI->getDefaultDest()->getFirstNonPHIOrDbg());
5272  const bool GeneratingCoveredLookupTable = (MaxTableSize == TableSize);
5273  BranchInst *RangeCheckBranch = nullptr;
5274 
5275  if (!DefaultIsReachable || GeneratingCoveredLookupTable) {
5276  Builder.CreateBr(LookupBB);
5277  // Note: We call removeProdecessor later since we need to be able to get the
5278  // PHI value for the default case in case we're using a bit mask.
5279  } else {
5280  Value *Cmp = Builder.CreateICmpULT(
5281  TableIndex, ConstantInt::get(MinCaseVal->getType(), TableSize));
5282  RangeCheckBranch =
5283  Builder.CreateCondBr(Cmp, LookupBB, SI->getDefaultDest());
5284  }
5285 
5286  // Populate the BB that does the lookups.
5287  Builder.SetInsertPoint(LookupBB);
5288 
5289  if (NeedMask) {
5290  // Before doing the lookup we do the hole check.
5291  // The LookupBB is therefore re-purposed to do the hole check
5292  // and we create a new LookupBB.
5293  BasicBlock *MaskBB = LookupBB;
5294  MaskBB->setName("switch.hole_check");
5295  LookupBB = BasicBlock::Create(Mod.getContext(), "switch.lookup",
5296  CommonDest->getParent(), CommonDest);
5297 
5298  // Make the mask's bitwidth at least 8bit and a power-of-2 to avoid
5299  // unnecessary illegal types.
5300  uint64_t TableSizePowOf2 = NextPowerOf2(std::max(7ULL, TableSize - 1ULL));
5301  APInt MaskInt(TableSizePowOf2, 0);
5302  APInt One(TableSizePowOf2, 1);
5303  // Build bitmask; fill in a 1 bit for every case.
5304  const ResultListTy &ResultList = ResultLists[PHIs[0]];
5305  for (size_t I = 0, E = ResultList.size(); I != E; ++I) {
5306  uint64_t Idx = (ResultList[I].first->getValue() - MinCaseVal->getValue())
5307  .getLimitedValue();
5308  MaskInt |= One << Idx;
5309  }
5310  ConstantInt *TableMask = ConstantInt::get(Mod.getContext(), MaskInt);
5311 
5312  // Get the TableIndex'th bit of the bitmask.
5313  // If this bit is 0 (meaning hole) jump to the default destination,
5314  // else continue with table lookup.
5315  IntegerType *MapTy = TableMask->getType();
5316  Value *MaskIndex =
5317  Builder.CreateZExtOrTrunc(TableIndex, MapTy, "switch.maskindex");
5318  Value *Shifted = Builder.CreateLShr(TableMask, MaskIndex, "switch.shifted");
5319  Value *LoBit = Builder.CreateTrunc(
5320  Shifted, Type::getInt1Ty(Mod.getContext()), "switch.lobit");
5321  Builder.CreateCondBr(LoBit, LookupBB, SI->getDefaultDest());
5322 
5323  Builder.SetInsertPoint(LookupBB);
5324  AddPredecessorToBlock(SI->getDefaultDest(), MaskBB, SI->getParent());
5325  }
5326 
5327  if (!DefaultIsReachable || GeneratingCoveredLookupTable) {
5328  // We cached PHINodes in PHIs, to avoid accessing deleted PHINodes later,
5329  // do not delete PHINodes here.
5331  /*DontDeleteUselessPHIs=*/true);
5332  }
5333 
5334  bool ReturnedEarly = false;
5335  for (size_t I = 0, E = PHIs.size(); I != E; ++I) {
5336  PHINode *PHI = PHIs[I];
5337  const ResultListTy &ResultList = ResultLists[PHI];
5338 
5339  // If using a bitmask, use any value to fill the lookup table holes.
5340  Constant *DV = NeedMask ? ResultLists[PHI][0].second : DefaultResults[PHI];
5341  StringRef FuncName = Fn->getName();
5342  SwitchLookupTable Table(Mod, TableSize, MinCaseVal, ResultList, DV, DL,
5343  FuncName);
5344 
5345  Value *Result = Table.BuildLookup(TableIndex, Builder);
5346 
5347  // If the result is used to return immediately from the function, we want to
5348  // do that right here.
5349  if (PHI->hasOneUse() && isa<ReturnInst>(*PHI->user_begin()) &&
5350  PHI->user_back() == CommonDest->getFirstNonPHIOrDbg()) {
5351  Builder.CreateRet(Result);
5352  ReturnedEarly = true;
5353  break;
5354  }
5355 
5356  // Do a small peephole optimization: re-use the switch table compare if
5357  // possible.
5358  if (!TableHasHoles && HasDefaultResults && RangeCheckBranch) {
5359  BasicBlock *PhiBlock = PHI->getParent();
5360  // Search for compare instructions which use the phi.
5361  for (auto *User : PHI->users()) {
5362  reuseTableCompare(User, PhiBlock, RangeCheckBranch, DV, ResultList);
5363  }
5364  }
5365 
5366  PHI->addIncoming(Result, LookupBB);
5367  }
5368 
5369  if (!ReturnedEarly)
5370  Builder.CreateBr(CommonDest);
5371 
5372  // Remove the switch.
5373  for (unsigned i = 0, e = SI->getNumSuccessors(); i < e; ++i) {
5374  BasicBlock *Succ = SI->getSuccessor(i);
5375 
5376  if (Succ == SI->getDefaultDest())
5377  continue;
5378  Succ->removePredecessor(SI->getParent());
5379  }
5380  SI->eraseFromParent();
5381 
5382  ++NumLookupTables;
5383  if (NeedMask)
5384  ++NumLookupTablesHoles;
5385  return true;
5386 }
5387 
5388 static bool isSwitchDense(ArrayRef<int64_t> Values) {
5389  // See also SelectionDAGBuilder::isDense(), which this function was based on.
5390  uint64_t Diff = (uint64_t)Values.back() - (uint64_t)Values.front();
5391  uint64_t Range = Diff + 1;
5392  uint64_t NumCases = Values.size();
5393  // 40% is the default density for building a jump table in optsize/minsize mode.
5394  uint64_t MinDensity = 40;
5395 
5396  return NumCases * 100 >= Range * MinDensity;
5397 }
5398 
5399 // Try and transform a switch that has "holes" in it to a contiguous sequence
5400 // of cases.
5401 //
5402 // A switch such as: switch(i) {case 5: case 9: case 13: case 17:} can be
5403 // range-reduced to: switch ((i-5) / 4) {case 0: case 1: case 2: case 3:}.
5404 //
5405 // This converts a sparse switch into a dense switch which allows better
5406 // lowering and could also allow transforming into a lookup table.
5407 static bool ReduceSwitchRange(SwitchInst *SI, IRBuilder<> &Builder,
5408  const DataLayout &DL,
5409  const TargetTransformInfo &TTI) {
5410  auto *CondTy = cast<IntegerType>(SI->getCondition()->getType());
5411  if (CondTy->getIntegerBitWidth() > 64 ||
5412  !DL.fitsInLegalInteger(CondTy->getIntegerBitWidth()))
5413  return false;
5414  // Only bother with this optimization if there are more than 3 switch cases;
5415  // SDAG will only bother creating jump tables for 4 or more cases.
5416  if (SI->getNumCases() < 4)
5417  return false;
5418 
5419  // This transform is agnostic to the signedness of the input or case values. We
5420  // can treat the case values as signed or unsigned. We can optimize more common
5421  // cases such as a sequence crossing zero {-4,0,4,8} if we interpret case values
5422  // as signed.
5423  SmallVector<int64_t,4> Values;
5424  for (auto &C : SI->cases())
5425  Values.push_back(C.getCaseValue()->getValue().getSExtValue());
5426  std::sort(Values.begin(), Values.end());
5427 
5428  // If the switch is already dense, there's nothing useful to do here.
5429  if (isSwitchDense(Values))
5430  return false;
5431 
5432  // First, transform the values such that they start at zero and ascend.
5433  int64_t Base = Values[0];
5434  for (auto &V : Values)
5435  V -= Base;
5436 
5437  // Now we have signed numbers that have been shifted so that, given enough
5438  // precision, there are no negative values. Since the rest of the transform
5439  // is bitwise only, we switch now to an unsigned representation.
5440  uint64_t GCD = 0;
5441  for (auto &V : Values)
5442  GCD = GreatestCommonDivisor64(GCD, (uint64_t)V);
5443 
5444  // This transform can be done speculatively because it is so cheap - it results
5445  // in a single rotate operation being inserted. This can only happen if the
5446  // factor extracted is a power of 2.
5447  // FIXME: If the GCD is an odd number we can multiply by the multiplicative
5448  // inverse of GCD and then perform this transform.
5449  // FIXME: It's possible that optimizing a switch on powers of two might also
5450  // be beneficial - flag values are often powers of two and we could use a CLZ
5451  // as the key function.
5452  if (GCD <= 1 || !isPowerOf2_64(GCD))
5453  // No common divisor found or too expensive to compute key function.
5454  return false;
5455 
5456  unsigned Shift = Log2_64(GCD);
5457  for (auto &V : Values)
5458  V = (int64_t)((uint64_t)V >> Shift);
5459 
5460  if (!isSwitchDense(Values))
5461  // Transform didn't create a dense switch.
5462  return false;
5463 
5464  // The obvious transform is to shift the switch condition right and emit a
5465  // check that the condition actually cleanly divided by GCD, i.e.
5466  // C & (1 << Shift - 1) == 0
5467  // inserting a new CFG edge to handle the case where it didn't divide cleanly.
5468  //
5469  // A cheaper way of doing this is a simple ROTR(C, Shift). This performs the
5470  // shift and puts the shifted-off bits in the uppermost bits. If any of these
5471  // are nonzero then the switch condition will be very large and will hit the
5472  // default case.
5473 
5474  auto *Ty = cast<IntegerType>(SI->getCondition()->getType());
5475  Builder.SetInsertPoint(SI);
5476  auto *ShiftC = ConstantInt::get(Ty, Shift);
5477  auto *Sub = Builder.CreateSub(SI->getCondition(), ConstantInt::get(Ty, Base));
5478  auto *LShr = Builder.CreateLShr(Sub, ShiftC);
5479  auto *Shl = Builder.CreateShl(Sub, Ty->getBitWidth() - Shift);
5480  auto *Rot = Builder.CreateOr(LShr, Shl);
5481  SI->replaceUsesOfWith(SI->getCondition(), Rot);
5482 
5483  for (auto Case : SI->cases()) {
5484  auto *Orig = Case.getCaseValue();
5485  auto Sub = Orig->getValue() - APInt(Ty->getBitWidth(), Base);
5486  Case.setValue(
5487  cast<ConstantInt>(ConstantInt::get(Ty, Sub.lshr(ShiftC->getValue()))));
5488  }
5489  return true;
5490 }
5491 
5492 bool SimplifyCFGOpt::SimplifySwitch(SwitchInst *SI, IRBuilder<> &Builder) {
5493  BasicBlock *BB = SI->getParent();
5494 
5495  if (isValueEqualityComparison(SI)) {
5496  // If we only have one predecessor, and if it is a branch on this value,
5497  // see if that predecessor totally determines the outcome of this switch.
5498  if (BasicBlock *OnlyPred = BB->getSinglePredecessor())
5499  if (SimplifyEqualityComparisonWithOnlyPredecessor(SI, OnlyPred, Builder))
5500  return SimplifyCFG(BB, TTI, BonusInstThreshold, AC) | true;
5501 
5502  Value *Cond = SI->getCondition();
5503  if (SelectInst *Select = dyn_cast<SelectInst>(Cond))
5504  if (SimplifySwitchOnSelect(SI, Select))
5505  return SimplifyCFG(BB, TTI, BonusInstThreshold, AC) | true;
5506 
5507  // If the block only contains the switch, see if we can fold the block
5508  // away into any preds.
5509  BasicBlock::iterator BBI = BB->begin();
5510  // Ignore dbg intrinsics.
5511  while (isa<DbgInfoIntrinsic>(BBI))
5512  ++BBI;
5513  if (SI == &*BBI)
5514  if (FoldValueComparisonIntoPredecessors(SI, Builder))
5515  return SimplifyCFG(BB, TTI, BonusInstThreshold, AC) | true;
5516  }
5517 
5518  // Try to transform the switch into an icmp and a branch.
5519  if (TurnSwitchRangeIntoICmp(SI, Builder))
5520  return SimplifyCFG(BB, TTI, BonusInstThreshold, AC) | true;
5521 
5522  // Remove unreachable cases.
5523  if (EliminateDeadSwitchCases(SI, AC, DL))
5524  return SimplifyCFG(BB, TTI, BonusInstThreshold, AC) | true;
5525 
5526  if (SwitchToSelect(SI, Builder, AC, DL, TTI))
5527  return SimplifyCFG(BB, TTI, BonusInstThreshold, AC) | true;
5528 
5530  return SimplifyCFG(BB, TTI, BonusInstThreshold, AC) | true;
5531 
5532  // The conversion from switch to lookup tables results in difficult
5533  // to analyze code and makes pruning branches much harder.
5534  // This is a problem of the switch expression itself can still be
5535  // restricted as a result of inlining or CVP. There only apply this
5536  // transformation during late steps of the optimisation chain.
5537  if (LateSimplifyCFG && SwitchToLookupTable(SI, Builder, DL, TTI))
5538  return SimplifyCFG(BB, TTI, BonusInstThreshold, AC) | true;
5539 
5540  if (ReduceSwitchRange(SI, Builder, DL, TTI))
5541  return SimplifyCFG(BB, TTI, BonusInstThreshold, AC) | true;
5542 
5543  return false;
5544 }
5545 
5546 bool SimplifyCFGOpt::SimplifyIndirectBr(IndirectBrInst *IBI) {
5547  BasicBlock *BB = IBI->getParent();
5548  bool Changed = false;
5549 
5550  // Eliminate redundant destinations.
5552  for (unsigned i = 0, e = IBI->getNumDestinations(); i != e; ++i) {
5553  BasicBlock *Dest = IBI->getDestination(i);
5554  if (!Dest->hasAddressTaken() || !Succs.insert(Dest).second) {
5555  Dest->removePredecessor(BB);
5556  IBI->removeDestination(i);
5557  --i;
5558  --e;
5559  Changed = true;
5560  }
5561  }
5562 
5563  if (IBI->getNumDestinations() == 0) {
5564  // If the indirectbr has no successors, change it to unreachable.
5565  new UnreachableInst(IBI->getContext(), IBI);
5567  return true;
5568  }
5569 
5570  if (IBI->