LLVM  7.0.0svn
SimplifyCFG.cpp
Go to the documentation of this file.
1 //===- SimplifyCFG.cpp - Code to perform CFG simplification ---------------===//
2 //
3 // The LLVM Compiler Infrastructure
4 //
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
7 //
8 //===----------------------------------------------------------------------===//
9 //
10 // Peephole optimize the CFG.
11 //
12 //===----------------------------------------------------------------------===//
13 
14 #include "llvm/ADT/APInt.h"
15 #include "llvm/ADT/ArrayRef.h"
16 #include "llvm/ADT/DenseMap.h"
17 #include "llvm/ADT/Optional.h"
18 #include "llvm/ADT/STLExtras.h"
19 #include "llvm/ADT/SetOperations.h"
20 #include "llvm/ADT/SetVector.h"
21 #include "llvm/ADT/SmallPtrSet.h"
22 #include "llvm/ADT/SmallSet.h"
23 #include "llvm/ADT/SmallVector.h"
24 #include "llvm/ADT/Statistic.h"
25 #include "llvm/ADT/StringRef.h"
33 #include "llvm/IR/Attributes.h"
34 #include "llvm/IR/BasicBlock.h"
35 #include "llvm/IR/CFG.h"
36 #include "llvm/IR/CallSite.h"
37 #include "llvm/IR/Constant.h"
38 #include "llvm/IR/ConstantRange.h"
39 #include "llvm/IR/Constants.h"
40 #include "llvm/IR/DataLayout.h"
41 #include "llvm/IR/DerivedTypes.h"
42 #include "llvm/IR/Function.h"
43 #include "llvm/IR/GlobalValue.h"
44 #include "llvm/IR/GlobalVariable.h"
45 #include "llvm/IR/IRBuilder.h"
46 #include "llvm/IR/InstrTypes.h"
47 #include "llvm/IR/Instruction.h"
48 #include "llvm/IR/Instructions.h"
49 #include "llvm/IR/IntrinsicInst.h"
50 #include "llvm/IR/Intrinsics.h"
51 #include "llvm/IR/LLVMContext.h"
52 #include "llvm/IR/MDBuilder.h"
53 #include "llvm/IR/Metadata.h"
54 #include "llvm/IR/Module.h"
55 #include "llvm/IR/NoFolder.h"
56 #include "llvm/IR/Operator.h"
57 #include "llvm/IR/PatternMatch.h"
58 #include "llvm/IR/Type.h"
59 #include "llvm/IR/Use.h"
60 #include "llvm/IR/User.h"
61 #include "llvm/IR/Value.h"
62 #include "llvm/Support/Casting.h"
64 #include "llvm/Support/Debug.h"
66 #include "llvm/Support/KnownBits.h"
71 #include <algorithm>
72 #include <cassert>
73 #include <climits>
74 #include <cstddef>
75 #include <cstdint>
76 #include <iterator>
77 #include <map>
78 #include <set>
79 #include <tuple>
80 #include <utility>
81 #include <vector>
82 
83 using namespace llvm;
84 using namespace PatternMatch;
85 
86 #define DEBUG_TYPE "simplifycfg"
87 
88 // Chosen as 2 so as to be cheap, but still to have enough power to fold
89 // a select, so the "clamp" idiom (of a min followed by a max) will be caught.
90 // To catch this, we need to fold a compare and a select, hence '2' being the
91 // minimum reasonable default.
93  "phi-node-folding-threshold", cl::Hidden, cl::init(2),
94  cl::desc(
95  "Control the amount of phi node folding to perform (default = 2)"));
96 
97 static cl::opt<bool> DupRet(
98  "simplifycfg-dup-ret", cl::Hidden, cl::init(false),
99  cl::desc("Duplicate return instructions into unconditional branches"));
100 
101 static cl::opt<bool>
102  SinkCommon("simplifycfg-sink-common", cl::Hidden, cl::init(true),
103  cl::desc("Sink common instructions down to the end block"));
104 
106  "simplifycfg-hoist-cond-stores", cl::Hidden, cl::init(true),
107  cl::desc("Hoist conditional stores if an unconditional store precedes"));
108 
110  "simplifycfg-merge-cond-stores", cl::Hidden, cl::init(true),
111  cl::desc("Hoist conditional stores even if an unconditional store does not "
112  "precede - hoist multiple conditional stores into a single "
113  "predicated store"));
114 
116  "simplifycfg-merge-cond-stores-aggressively", cl::Hidden, cl::init(false),
117  cl::desc("When merging conditional stores, do so even if the resultant "
118  "basic blocks are unlikely to be if-converted as a result"));
119 
121  "speculate-one-expensive-inst", cl::Hidden, cl::init(true),
122  cl::desc("Allow exactly one expensive instruction to be speculatively "
123  "executed"));
124 
126  "max-speculation-depth", cl::Hidden, cl::init(10),
127  cl::desc("Limit maximum recursion depth when calculating costs of "
128  "speculatively executed instructions"));
129 
130 STATISTIC(NumBitMaps, "Number of switch instructions turned into bitmaps");
131 STATISTIC(NumLinearMaps,
132  "Number of switch instructions turned into linear mapping");
133 STATISTIC(NumLookupTables,
134  "Number of switch instructions turned into lookup tables");
135 STATISTIC(
136  NumLookupTablesHoles,
137  "Number of switch instructions turned into lookup tables (holes checked)");
138 STATISTIC(NumTableCmpReuses, "Number of reused switch table lookup compares");
139 STATISTIC(NumSinkCommons,
140  "Number of common instructions sunk down to the end block");
141 STATISTIC(NumSpeculations, "Number of speculative executed instructions");
142 
143 namespace {
144 
145 // The first field contains the value that the switch produces when a certain
146 // case group is selected, and the second field is a vector containing the
147 // cases composing the case group.
148 using SwitchCaseResultVectorTy =
150 
151 // The first field contains the phi node that generates a result of the switch
152 // and the second field contains the value generated for a certain case in the
153 // switch for that PHI.
154 using SwitchCaseResultsTy = SmallVector<std::pair<PHINode *, Constant *>, 4>;
155 
156 /// ValueEqualityComparisonCase - Represents a case of a switch.
157 struct ValueEqualityComparisonCase {
159  BasicBlock *Dest;
160 
161  ValueEqualityComparisonCase(ConstantInt *Value, BasicBlock *Dest)
162  : Value(Value), Dest(Dest) {}
163 
164  bool operator<(ValueEqualityComparisonCase RHS) const {
165  // Comparing pointers is ok as we only rely on the order for uniquing.
166  return Value < RHS.Value;
167  }
168 
169  bool operator==(BasicBlock *RHSDest) const { return Dest == RHSDest; }
170 };
171 
172 class SimplifyCFGOpt {
173  const TargetTransformInfo &TTI;
174  const DataLayout &DL;
175  SmallPtrSetImpl<BasicBlock *> *LoopHeaders;
176  const SimplifyCFGOptions &Options;
177 
178  Value *isValueEqualityComparison(TerminatorInst *TI);
179  BasicBlock *GetValueEqualityComparisonCases(
180  TerminatorInst *TI, std::vector<ValueEqualityComparisonCase> &Cases);
181  bool SimplifyEqualityComparisonWithOnlyPredecessor(TerminatorInst *TI,
182  BasicBlock *Pred,
183  IRBuilder<> &Builder);
184  bool FoldValueComparisonIntoPredecessors(TerminatorInst *TI,
185  IRBuilder<> &Builder);
186 
187  bool SimplifyReturn(ReturnInst *RI, IRBuilder<> &Builder);
188  bool SimplifyResume(ResumeInst *RI, IRBuilder<> &Builder);
189  bool SimplifySingleResume(ResumeInst *RI);
190  bool SimplifyCommonResume(ResumeInst *RI);
191  bool SimplifyCleanupReturn(CleanupReturnInst *RI);
192  bool SimplifyUnreachable(UnreachableInst *UI);
193  bool SimplifySwitch(SwitchInst *SI, IRBuilder<> &Builder);
194  bool SimplifyIndirectBr(IndirectBrInst *IBI);
195  bool SimplifyUncondBranch(BranchInst *BI, IRBuilder<> &Builder);
196  bool SimplifyCondBranch(BranchInst *BI, IRBuilder<> &Builder);
197 
198 public:
199  SimplifyCFGOpt(const TargetTransformInfo &TTI, const DataLayout &DL,
200  SmallPtrSetImpl<BasicBlock *> *LoopHeaders,
201  const SimplifyCFGOptions &Opts)
202  : TTI(TTI), DL(DL), LoopHeaders(LoopHeaders), Options(Opts) {}
203 
204  bool run(BasicBlock *BB);
205 };
206 
207 } // end anonymous namespace
208 
209 /// Return true if it is safe to merge these two
210 /// terminator instructions together.
211 static bool
213  SmallSetVector<BasicBlock *, 4> *FailBlocks = nullptr) {
214  if (SI1 == SI2)
215  return false; // Can't merge with self!
216 
217  // It is not safe to merge these two switch instructions if they have a common
218  // successor, and if that successor has a PHI node, and if *that* PHI node has
219  // conflicting incoming values from the two switch blocks.
220  BasicBlock *SI1BB = SI1->getParent();
221  BasicBlock *SI2BB = SI2->getParent();
222 
223  SmallPtrSet<BasicBlock *, 16> SI1Succs(succ_begin(SI1BB), succ_end(SI1BB));
224  bool Fail = false;
225  for (BasicBlock *Succ : successors(SI2BB))
226  if (SI1Succs.count(Succ))
227  for (BasicBlock::iterator BBI = Succ->begin(); isa<PHINode>(BBI); ++BBI) {
228  PHINode *PN = cast<PHINode>(BBI);
229  if (PN->getIncomingValueForBlock(SI1BB) !=
230  PN->getIncomingValueForBlock(SI2BB)) {
231  if (FailBlocks)
232  FailBlocks->insert(Succ);
233  Fail = true;
234  }
235  }
236 
237  return !Fail;
238 }
239 
240 /// Return true if it is safe and profitable to merge these two terminator
241 /// instructions together, where SI1 is an unconditional branch. PhiNodes will
242 /// store all PHI nodes in common successors.
243 static bool
245  Instruction *Cond,
246  SmallVectorImpl<PHINode *> &PhiNodes) {
247  if (SI1 == SI2)
248  return false; // Can't merge with self!
249  assert(SI1->isUnconditional() && SI2->isConditional());
250 
251  // We fold the unconditional branch if we can easily update all PHI nodes in
252  // common successors:
253  // 1> We have a constant incoming value for the conditional branch;
254  // 2> We have "Cond" as the incoming value for the unconditional branch;
255  // 3> SI2->getCondition() and Cond have same operands.
256  CmpInst *Ci2 = dyn_cast<CmpInst>(SI2->getCondition());
257  if (!Ci2)
258  return false;
259  if (!(Cond->getOperand(0) == Ci2->getOperand(0) &&
260  Cond->getOperand(1) == Ci2->getOperand(1)) &&
261  !(Cond->getOperand(0) == Ci2->getOperand(1) &&
262  Cond->getOperand(1) == Ci2->getOperand(0)))
263  return false;
264 
265  BasicBlock *SI1BB = SI1->getParent();
266  BasicBlock *SI2BB = SI2->getParent();
267  SmallPtrSet<BasicBlock *, 16> SI1Succs(succ_begin(SI1BB), succ_end(SI1BB));
268  for (BasicBlock *Succ : successors(SI2BB))
269  if (SI1Succs.count(Succ))
270  for (BasicBlock::iterator BBI = Succ->begin(); isa<PHINode>(BBI); ++BBI) {
271  PHINode *PN = cast<PHINode>(BBI);
272  if (PN->getIncomingValueForBlock(SI1BB) != Cond ||
273  !isa<ConstantInt>(PN->getIncomingValueForBlock(SI2BB)))
274  return false;
275  PhiNodes.push_back(PN);
276  }
277  return true;
278 }
279 
280 /// Update PHI nodes in Succ to indicate that there will now be entries in it
281 /// from the 'NewPred' block. The values that will be flowing into the PHI nodes
282 /// will be the same as those coming in from ExistPred, an existing predecessor
283 /// of Succ.
284 static void AddPredecessorToBlock(BasicBlock *Succ, BasicBlock *NewPred,
285  BasicBlock *ExistPred) {
286  for (PHINode &PN : Succ->phis())
287  PN.addIncoming(PN.getIncomingValueForBlock(ExistPred), NewPred);
288 }
289 
290 /// Compute an abstract "cost" of speculating the given instruction,
291 /// which is assumed to be safe to speculate. TCC_Free means cheap,
292 /// TCC_Basic means less cheap, and TCC_Expensive means prohibitively
293 /// expensive.
294 static unsigned ComputeSpeculationCost(const User *I,
295  const TargetTransformInfo &TTI) {
297  "Instruction is not safe to speculatively execute!");
298  return TTI.getUserCost(I);
299 }
300 
301 /// If we have a merge point of an "if condition" as accepted above,
302 /// return true if the specified value dominates the block. We
303 /// don't handle the true generality of domination here, just a special case
304 /// which works well enough for us.
305 ///
306 /// If AggressiveInsts is non-null, and if V does not dominate BB, we check to
307 /// see if V (which must be an instruction) and its recursive operands
308 /// that do not dominate BB have a combined cost lower than CostRemaining and
309 /// are non-trapping. If both are true, the instruction is inserted into the
310 /// set and true is returned.
311 ///
312 /// The cost for most non-trapping instructions is defined as 1 except for
313 /// Select whose cost is 2.
314 ///
315 /// After this function returns, CostRemaining is decreased by the cost of
316 /// V plus its non-dominating operands. If that cost is greater than
317 /// CostRemaining, false is returned and CostRemaining is undefined.
319  SmallPtrSetImpl<Instruction *> *AggressiveInsts,
320  unsigned &CostRemaining,
321  const TargetTransformInfo &TTI,
322  unsigned Depth = 0) {
323  // It is possible to hit a zero-cost cycle (phi/gep instructions for example),
324  // so limit the recursion depth.
325  // TODO: While this recursion limit does prevent pathological behavior, it
326  // would be better to track visited instructions to avoid cycles.
327  if (Depth == MaxSpeculationDepth)
328  return false;
329 
331  if (!I) {
332  // Non-instructions all dominate instructions, but not all constantexprs
333  // can be executed unconditionally.
334  if (ConstantExpr *C = dyn_cast<ConstantExpr>(V))
335  if (C->canTrap())
336  return false;
337  return true;
338  }
339  BasicBlock *PBB = I->getParent();
340 
341  // We don't want to allow weird loops that might have the "if condition" in
342  // the bottom of this block.
343  if (PBB == BB)
344  return false;
345 
346  // If this instruction is defined in a block that contains an unconditional
347  // branch to BB, then it must be in the 'conditional' part of the "if
348  // statement". If not, it definitely dominates the region.
350  if (!BI || BI->isConditional() || BI->getSuccessor(0) != BB)
351  return true;
352 
353  // If we aren't allowing aggressive promotion anymore, then don't consider
354  // instructions in the 'if region'.
355  if (!AggressiveInsts)
356  return false;
357 
358  // If we have seen this instruction before, don't count it again.
359  if (AggressiveInsts->count(I))
360  return true;
361 
362  // Okay, it looks like the instruction IS in the "condition". Check to
363  // see if it's a cheap instruction to unconditionally compute, and if it
364  // only uses stuff defined outside of the condition. If so, hoist it out.
366  return false;
367 
368  unsigned Cost = ComputeSpeculationCost(I, TTI);
369 
370  // Allow exactly one instruction to be speculated regardless of its cost
371  // (as long as it is safe to do so).
372  // This is intended to flatten the CFG even if the instruction is a division
373  // or other expensive operation. The speculation of an expensive instruction
374  // is expected to be undone in CodeGenPrepare if the speculation has not
375  // enabled further IR optimizations.
376  if (Cost > CostRemaining &&
377  (!SpeculateOneExpensiveInst || !AggressiveInsts->empty() || Depth > 0))
378  return false;
379 
380  // Avoid unsigned wrap.
381  CostRemaining = (Cost > CostRemaining) ? 0 : CostRemaining - Cost;
382 
383  // Okay, we can only really hoist these out if their operands do
384  // not take us over the cost threshold.
385  for (User::op_iterator i = I->op_begin(), e = I->op_end(); i != e; ++i)
386  if (!DominatesMergePoint(*i, BB, AggressiveInsts, CostRemaining, TTI,
387  Depth + 1))
388  return false;
389  // Okay, it's safe to do this! Remember this instruction.
390  AggressiveInsts->insert(I);
391  return true;
392 }
393 
394 /// Extract ConstantInt from value, looking through IntToPtr
395 /// and PointerNullValue. Return NULL if value is not a constant int.
396 static ConstantInt *GetConstantInt(Value *V, const DataLayout &DL) {
397  // Normal constant int.
398  ConstantInt *CI = dyn_cast<ConstantInt>(V);
399  if (CI || !isa<Constant>(V) || !V->getType()->isPointerTy())
400  return CI;
401 
402  // This is some kind of pointer constant. Turn it into a pointer-sized
403  // ConstantInt if possible.
404  IntegerType *PtrTy = cast<IntegerType>(DL.getIntPtrType(V->getType()));
405 
406  // Null pointer means 0, see SelectionDAGBuilder::getValue(const Value*).
407  if (isa<ConstantPointerNull>(V))
408  return ConstantInt::get(PtrTy, 0);
409 
410  // IntToPtr const int.
411  if (ConstantExpr *CE = dyn_cast<ConstantExpr>(V))
412  if (CE->getOpcode() == Instruction::IntToPtr)
413  if (ConstantInt *CI = dyn_cast<ConstantInt>(CE->getOperand(0))) {
414  // The constant is very likely to have the right type already.
415  if (CI->getType() == PtrTy)
416  return CI;
417  else
418  return cast<ConstantInt>(
419  ConstantExpr::getIntegerCast(CI, PtrTy, /*isSigned=*/false));
420  }
421  return nullptr;
422 }
423 
424 namespace {
425 
426 /// Given a chain of or (||) or and (&&) comparison of a value against a
427 /// constant, this will try to recover the information required for a switch
428 /// structure.
429 /// It will depth-first traverse the chain of comparison, seeking for patterns
430 /// like %a == 12 or %a < 4 and combine them to produce a set of integer
431 /// representing the different cases for the switch.
432 /// Note that if the chain is composed of '||' it will build the set of elements
433 /// that matches the comparisons (i.e. any of this value validate the chain)
434 /// while for a chain of '&&' it will build the set elements that make the test
435 /// fail.
436 struct ConstantComparesGatherer {
437  const DataLayout &DL;
438 
439  /// Value found for the switch comparison
440  Value *CompValue = nullptr;
441 
442  /// Extra clause to be checked before the switch
443  Value *Extra = nullptr;
444 
445  /// Set of integers to match in switch
447 
448  /// Number of comparisons matched in the and/or chain
449  unsigned UsedICmps = 0;
450 
451  /// Construct and compute the result for the comparison instruction Cond
452  ConstantComparesGatherer(Instruction *Cond, const DataLayout &DL) : DL(DL) {
453  gather(Cond);
454  }
455 
456  ConstantComparesGatherer(const ConstantComparesGatherer &) = delete;
457  ConstantComparesGatherer &
458  operator=(const ConstantComparesGatherer &) = delete;
459 
460 private:
461  /// Try to set the current value used for the comparison, it succeeds only if
462  /// it wasn't set before or if the new value is the same as the old one
463  bool setValueOnce(Value *NewVal) {
464  if (CompValue && CompValue != NewVal)
465  return false;
466  CompValue = NewVal;
467  return (CompValue != nullptr);
468  }
469 
470  /// Try to match Instruction "I" as a comparison against a constant and
471  /// populates the array Vals with the set of values that match (or do not
472  /// match depending on isEQ).
473  /// Return false on failure. On success, the Value the comparison matched
474  /// against is placed in CompValue.
475  /// If CompValue is already set, the function is expected to fail if a match
476  /// is found but the value compared to is different.
477  bool matchInstruction(Instruction *I, bool isEQ) {
478  // If this is an icmp against a constant, handle this as one of the cases.
479  ICmpInst *ICI;
480  ConstantInt *C;
481  if (!((ICI = dyn_cast<ICmpInst>(I)) &&
482  (C = GetConstantInt(I->getOperand(1), DL)))) {
483  return false;
484  }
485 
486  Value *RHSVal;
487  const APInt *RHSC;
488 
489  // Pattern match a special case
490  // (x & ~2^z) == y --> x == y || x == y|2^z
491  // This undoes a transformation done by instcombine to fuse 2 compares.
492  if (ICI->getPredicate() == (isEQ ? ICmpInst::ICMP_EQ : ICmpInst::ICMP_NE)) {
493  // It's a little bit hard to see why the following transformations are
494  // correct. Here is a CVC3 program to verify them for 64-bit values:
495 
496  /*
497  ONE : BITVECTOR(64) = BVZEROEXTEND(0bin1, 63);
498  x : BITVECTOR(64);
499  y : BITVECTOR(64);
500  z : BITVECTOR(64);
501  mask : BITVECTOR(64) = BVSHL(ONE, z);
502  QUERY( (y & ~mask = y) =>
503  ((x & ~mask = y) <=> (x = y OR x = (y | mask)))
504  );
505  QUERY( (y | mask = y) =>
506  ((x | mask = y) <=> (x = y OR x = (y & ~mask)))
507  );
508  */
509 
510  // Please note that each pattern must be a dual implication (<--> or
511  // iff). One directional implication can create spurious matches. If the
512  // implication is only one-way, an unsatisfiable condition on the left
513  // side can imply a satisfiable condition on the right side. Dual
514  // implication ensures that satisfiable conditions are transformed to
515  // other satisfiable conditions and unsatisfiable conditions are
516  // transformed to other unsatisfiable conditions.
517 
518  // Here is a concrete example of a unsatisfiable condition on the left
519  // implying a satisfiable condition on the right:
520  //
521  // mask = (1 << z)
522  // (x & ~mask) == y --> (x == y || x == (y | mask))
523  //
524  // Substituting y = 3, z = 0 yields:
525  // (x & -2) == 3 --> (x == 3 || x == 2)
526 
527  // Pattern match a special case:
528  /*
529  QUERY( (y & ~mask = y) =>
530  ((x & ~mask = y) <=> (x = y OR x = (y | mask)))
531  );
532  */
533  if (match(ICI->getOperand(0),
534  m_And(m_Value(RHSVal), m_APInt(RHSC)))) {
535  APInt Mask = ~*RHSC;
536  if (Mask.isPowerOf2() && (C->getValue() & ~Mask) == C->getValue()) {
537  // If we already have a value for the switch, it has to match!
538  if (!setValueOnce(RHSVal))
539  return false;
540 
541  Vals.push_back(C);
542  Vals.push_back(
544  C->getValue() | Mask));
545  UsedICmps++;
546  return true;
547  }
548  }
549 
550  // Pattern match a special case:
551  /*
552  QUERY( (y | mask = y) =>
553  ((x | mask = y) <=> (x = y OR x = (y & ~mask)))
554  );
555  */
556  if (match(ICI->getOperand(0),
557  m_Or(m_Value(RHSVal), m_APInt(RHSC)))) {
558  APInt Mask = *RHSC;
559  if (Mask.isPowerOf2() && (C->getValue() | Mask) == C->getValue()) {
560  // If we already have a value for the switch, it has to match!
561  if (!setValueOnce(RHSVal))
562  return false;
563 
564  Vals.push_back(C);
566  C->getValue() & ~Mask));
567  UsedICmps++;
568  return true;
569  }
570  }
571 
572  // If we already have a value for the switch, it has to match!
573  if (!setValueOnce(ICI->getOperand(0)))
574  return false;
575 
576  UsedICmps++;
577  Vals.push_back(C);
578  return ICI->getOperand(0);
579  }
580 
581  // If we have "x ult 3", for example, then we can add 0,1,2 to the set.
583  ICI->getPredicate(), C->getValue());
584 
585  // Shift the range if the compare is fed by an add. This is the range
586  // compare idiom as emitted by instcombine.
587  Value *CandidateVal = I->getOperand(0);
588  if (match(I->getOperand(0), m_Add(m_Value(RHSVal), m_APInt(RHSC)))) {
589  Span = Span.subtract(*RHSC);
590  CandidateVal = RHSVal;
591  }
592 
593  // If this is an and/!= check, then we are looking to build the set of
594  // value that *don't* pass the and chain. I.e. to turn "x ugt 2" into
595  // x != 0 && x != 1.
596  if (!isEQ)
597  Span = Span.inverse();
598 
599  // If there are a ton of values, we don't want to make a ginormous switch.
600  if (Span.isSizeLargerThan(8) || Span.isEmptySet()) {
601  return false;
602  }
603 
604  // If we already have a value for the switch, it has to match!
605  if (!setValueOnce(CandidateVal))
606  return false;
607 
608  // Add all values from the range to the set
609  for (APInt Tmp = Span.getLower(); Tmp != Span.getUpper(); ++Tmp)
610  Vals.push_back(ConstantInt::get(I->getContext(), Tmp));
611 
612  UsedICmps++;
613  return true;
614  }
615 
616  /// Given a potentially 'or'd or 'and'd together collection of icmp
617  /// eq/ne/lt/gt instructions that compare a value against a constant, extract
618  /// the value being compared, and stick the list constants into the Vals
619  /// vector.
620  /// One "Extra" case is allowed to differ from the other.
621  void gather(Value *V) {
623  bool isEQ = (I->getOpcode() == Instruction::Or);
624 
625  // Keep a stack (SmallVector for efficiency) for depth-first traversal
627  SmallPtrSet<Value *, 8> Visited;
628 
629  // Initialize
630  Visited.insert(V);
631  DFT.push_back(V);
632 
633  while (!DFT.empty()) {
634  V = DFT.pop_back_val();
635 
636  if (Instruction *I = dyn_cast<Instruction>(V)) {
637  // If it is a || (or && depending on isEQ), process the operands.
638  if (I->getOpcode() == (isEQ ? Instruction::Or : Instruction::And)) {
639  if (Visited.insert(I->getOperand(1)).second)
640  DFT.push_back(I->getOperand(1));
641  if (Visited.insert(I->getOperand(0)).second)
642  DFT.push_back(I->getOperand(0));
643  continue;
644  }
645 
646  // Try to match the current instruction
647  if (matchInstruction(I, isEQ))
648  // Match succeed, continue the loop
649  continue;
650  }
651 
652  // One element of the sequence of || (or &&) could not be match as a
653  // comparison against the same value as the others.
654  // We allow only one "Extra" case to be checked before the switch
655  if (!Extra) {
656  Extra = V;
657  continue;
658  }
659  // Failed to parse a proper sequence, abort now
660  CompValue = nullptr;
661  break;
662  }
663  }
664 };
665 
666 } // end anonymous namespace
667 
669  Instruction *Cond = nullptr;
670  if (SwitchInst *SI = dyn_cast<SwitchInst>(TI)) {
671  Cond = dyn_cast<Instruction>(SI->getCondition());
672  } else if (BranchInst *BI = dyn_cast<BranchInst>(TI)) {
673  if (BI->isConditional())
674  Cond = dyn_cast<Instruction>(BI->getCondition());
675  } else if (IndirectBrInst *IBI = dyn_cast<IndirectBrInst>(TI)) {
676  Cond = dyn_cast<Instruction>(IBI->getAddress());
677  }
678 
679  TI->eraseFromParent();
680  if (Cond)
682 }
683 
684 /// Return true if the specified terminator checks
685 /// to see if a value is equal to constant integer value.
686 Value *SimplifyCFGOpt::isValueEqualityComparison(TerminatorInst *TI) {
687  Value *CV = nullptr;
688  if (SwitchInst *SI = dyn_cast<SwitchInst>(TI)) {
689  // Do not permit merging of large switch instructions into their
690  // predecessors unless there is only one predecessor.
691  if (SI->getNumSuccessors() * pred_size(SI->getParent()) <= 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 // Set branch weights on SwitchInst. This sets the metadata if there is at
774 // least one non-zero weight.
776  // Check that there is at least one non-zero weight. Otherwise, pass
777  // nullptr to setMetadata which will erase the existing metadata.
778  MDNode *N = nullptr;
779  if (llvm::any_of(Weights, [](uint32_t W) { return W != 0; }))
780  N = MDBuilder(SI->getParent()->getContext()).createBranchWeights(Weights);
782 }
783 
784 // Similar to the above, but for branch and select instructions that take
785 // exactly 2 weights.
786 static void setBranchWeights(Instruction *I, uint32_t TrueWeight,
787  uint32_t FalseWeight) {
788  assert(isa<BranchInst>(I) || isa<SelectInst>(I));
789  // Check that there is at least one non-zero weight. Otherwise, pass
790  // nullptr to setMetadata which will erase the existing metadata.
791  MDNode *N = nullptr;
792  if (TrueWeight || FalseWeight)
793  N = MDBuilder(I->getParent()->getContext())
794  .createBranchWeights(TrueWeight, FalseWeight);
796 }
797 
798 /// If TI is known to be a terminator instruction and its block is known to
799 /// only have a single predecessor block, check to see if that predecessor is
800 /// also a value comparison with the same value, and if that comparison
801 /// determines the outcome of this comparison. If so, simplify TI. This does a
802 /// very limited form of jump threading.
803 bool SimplifyCFGOpt::SimplifyEqualityComparisonWithOnlyPredecessor(
804  TerminatorInst *TI, BasicBlock *Pred, IRBuilder<> &Builder) {
805  Value *PredVal = isValueEqualityComparison(Pred->getTerminator());
806  if (!PredVal)
807  return false; // Not a value comparison in predecessor.
808 
809  Value *ThisVal = isValueEqualityComparison(TI);
810  assert(ThisVal && "This isn't a value comparison!!");
811  if (ThisVal != PredVal)
812  return false; // Different predicates.
813 
814  // TODO: Preserve branch weight metadata, similarly to how
815  // FoldValueComparisonIntoPredecessors preserves it.
816 
817  // Find out information about when control will move from Pred to TI's block.
818  std::vector<ValueEqualityComparisonCase> PredCases;
819  BasicBlock *PredDef =
820  GetValueEqualityComparisonCases(Pred->getTerminator(), PredCases);
821  EliminateBlockCases(PredDef, PredCases); // Remove default from cases.
822 
823  // Find information about how control leaves this block.
824  std::vector<ValueEqualityComparisonCase> ThisCases;
825  BasicBlock *ThisDef = GetValueEqualityComparisonCases(TI, ThisCases);
826  EliminateBlockCases(ThisDef, ThisCases); // Remove default from cases.
827 
828  // If TI's block is the default block from Pred's comparison, potentially
829  // simplify TI based on this knowledge.
830  if (PredDef == TI->getParent()) {
831  // If we are here, we know that the value is none of those cases listed in
832  // PredCases. If there are any cases in ThisCases that are in PredCases, we
833  // can simplify TI.
834  if (!ValuesOverlap(PredCases, ThisCases))
835  return false;
836 
837  if (isa<BranchInst>(TI)) {
838  // Okay, one of the successors of this condbr is dead. Convert it to a
839  // uncond br.
840  assert(ThisCases.size() == 1 && "Branch can only have one case!");
841  // Insert the new branch.
842  Instruction *NI = Builder.CreateBr(ThisDef);
843  (void)NI;
844 
845  // Remove PHI node entries for the dead edge.
846  ThisCases[0].Dest->removePredecessor(TI->getParent());
847 
848  LLVM_DEBUG(dbgs() << "Threading pred instr: " << *Pred->getTerminator()
849  << "Through successor TI: " << *TI << "Leaving: " << *NI
850  << "\n");
851 
853  return true;
854  }
855 
856  SwitchInst *SI = cast<SwitchInst>(TI);
857  // Okay, TI has cases that are statically dead, prune them away.
858  SmallPtrSet<Constant *, 16> DeadCases;
859  for (unsigned i = 0, e = PredCases.size(); i != e; ++i)
860  DeadCases.insert(PredCases[i].Value);
861 
862  LLVM_DEBUG(dbgs() << "Threading pred instr: " << *Pred->getTerminator()
863  << "Through successor TI: " << *TI);
864 
865  // Collect branch weights into a vector.
866  SmallVector<uint32_t, 8> Weights;
868  bool HasWeight = MD && (MD->getNumOperands() == 2 + SI->getNumCases());
869  if (HasWeight)
870  for (unsigned MD_i = 1, MD_e = MD->getNumOperands(); MD_i < MD_e;
871  ++MD_i) {
872  ConstantInt *CI = mdconst::extract<ConstantInt>(MD->getOperand(MD_i));
873  Weights.push_back(CI->getValue().getZExtValue());
874  }
875  for (SwitchInst::CaseIt i = SI->case_end(), e = SI->case_begin(); i != e;) {
876  --i;
877  if (DeadCases.count(i->getCaseValue())) {
878  if (HasWeight) {
879  std::swap(Weights[i->getCaseIndex() + 1], Weights.back());
880  Weights.pop_back();
881  }
882  i->getCaseSuccessor()->removePredecessor(TI->getParent());
883  SI->removeCase(i);
884  }
885  }
886  if (HasWeight && Weights.size() >= 2)
887  setBranchWeights(SI, Weights);
888 
889  LLVM_DEBUG(dbgs() << "Leaving: " << *TI << "\n");
890  return true;
891  }
892 
893  // Otherwise, TI's block must correspond to some matched value. Find out
894  // which value (or set of values) this is.
895  ConstantInt *TIV = nullptr;
896  BasicBlock *TIBB = TI->getParent();
897  for (unsigned i = 0, e = PredCases.size(); i != e; ++i)
898  if (PredCases[i].Dest == TIBB) {
899  if (TIV)
900  return false; // Cannot handle multiple values coming to this block.
901  TIV = PredCases[i].Value;
902  }
903  assert(TIV && "No edge from pred to succ?");
904 
905  // Okay, we found the one constant that our value can be if we get into TI's
906  // BB. Find out which successor will unconditionally be branched to.
907  BasicBlock *TheRealDest = nullptr;
908  for (unsigned i = 0, e = ThisCases.size(); i != e; ++i)
909  if (ThisCases[i].Value == TIV) {
910  TheRealDest = ThisCases[i].Dest;
911  break;
912  }
913 
914  // If not handled by any explicit cases, it is handled by the default case.
915  if (!TheRealDest)
916  TheRealDest = ThisDef;
917 
918  // Remove PHI node entries for dead edges.
919  BasicBlock *CheckEdge = TheRealDest;
920  for (BasicBlock *Succ : successors(TIBB))
921  if (Succ != CheckEdge)
922  Succ->removePredecessor(TIBB);
923  else
924  CheckEdge = nullptr;
925 
926  // Insert the new branch.
927  Instruction *NI = Builder.CreateBr(TheRealDest);
928  (void)NI;
929 
930  LLVM_DEBUG(dbgs() << "Threading pred instr: " << *Pred->getTerminator()
931  << "Through successor TI: " << *TI << "Leaving: " << *NI
932  << "\n");
933 
935  return true;
936 }
937 
938 namespace {
939 
940 /// This class implements a stable ordering of constant
941 /// integers that does not depend on their address. This is important for
942 /// applications that sort ConstantInt's to ensure uniqueness.
943 struct ConstantIntOrdering {
944  bool operator()(const ConstantInt *LHS, const ConstantInt *RHS) const {
945  return LHS->getValue().ult(RHS->getValue());
946  }
947 };
948 
949 } // end anonymous namespace
950 
951 static int ConstantIntSortPredicate(ConstantInt *const *P1,
952  ConstantInt *const *P2) {
953  const ConstantInt *LHS = *P1;
954  const ConstantInt *RHS = *P2;
955  if (LHS == RHS)
956  return 0;
957  return LHS->getValue().ult(RHS->getValue()) ? 1 : -1;
958 }
959 
960 static inline bool HasBranchWeights(const Instruction *I) {
962  if (ProfMD && ProfMD->getOperand(0))
963  if (MDString *MDS = dyn_cast<MDString>(ProfMD->getOperand(0)))
964  return MDS->getString().equals("branch_weights");
965 
966  return false;
967 }
968 
969 /// Get Weights of a given TerminatorInst, the default weight is at the front
970 /// of the vector. If TI is a conditional eq, we need to swap the branch-weight
971 /// metadata.
973  SmallVectorImpl<uint64_t> &Weights) {
975  assert(MD);
976  for (unsigned i = 1, e = MD->getNumOperands(); i < e; ++i) {
977  ConstantInt *CI = mdconst::extract<ConstantInt>(MD->getOperand(i));
978  Weights.push_back(CI->getValue().getZExtValue());
979  }
980 
981  // If TI is a conditional eq, the default case is the false case,
982  // and the corresponding branch-weight data is at index 2. We swap the
983  // default weight to be the first entry.
984  if (BranchInst *BI = dyn_cast<BranchInst>(TI)) {
985  assert(Weights.size() == 2);
986  ICmpInst *ICI = cast<ICmpInst>(BI->getCondition());
987  if (ICI->getPredicate() == ICmpInst::ICMP_EQ)
988  std::swap(Weights.front(), Weights.back());
989  }
990 }
991 
992 /// Keep halving the weights until all can fit in uint32_t.
993 static void FitWeights(MutableArrayRef<uint64_t> Weights) {
994  uint64_t Max = *std::max_element(Weights.begin(), Weights.end());
995  if (Max > UINT_MAX) {
996  unsigned Offset = 32 - countLeadingZeros(Max);
997  for (uint64_t &I : Weights)
998  I >>= Offset;
999  }
1000 }
1001 
1002 /// The specified terminator is a value equality comparison instruction
1003 /// (either a switch or a branch on "X == c").
1004 /// See if any of the predecessors of the terminator block are value comparisons
1005 /// on the same value. If so, and if safe to do so, fold them together.
1006 bool SimplifyCFGOpt::FoldValueComparisonIntoPredecessors(TerminatorInst *TI,
1007  IRBuilder<> &Builder) {
1008  BasicBlock *BB = TI->getParent();
1009  Value *CV = isValueEqualityComparison(TI); // CondVal
1010  assert(CV && "Not a comparison?");
1011  bool Changed = false;
1012 
1014  while (!Preds.empty()) {
1015  BasicBlock *Pred = Preds.pop_back_val();
1016 
1017  // See if the predecessor is a comparison with the same value.
1018  TerminatorInst *PTI = Pred->getTerminator();
1019  Value *PCV = isValueEqualityComparison(PTI); // PredCondVal
1020 
1021  if (PCV == CV && TI != PTI) {
1022  SmallSetVector<BasicBlock*, 4> FailBlocks;
1023  if (!SafeToMergeTerminators(TI, PTI, &FailBlocks)) {
1024  for (auto *Succ : FailBlocks) {
1025  if (!SplitBlockPredecessors(Succ, TI->getParent(), ".fold.split"))
1026  return false;
1027  }
1028  }
1029 
1030  // Figure out which 'cases' to copy from SI to PSI.
1031  std::vector<ValueEqualityComparisonCase> BBCases;
1032  BasicBlock *BBDefault = GetValueEqualityComparisonCases(TI, BBCases);
1033 
1034  std::vector<ValueEqualityComparisonCase> PredCases;
1035  BasicBlock *PredDefault = GetValueEqualityComparisonCases(PTI, PredCases);
1036 
1037  // Based on whether the default edge from PTI goes to BB or not, fill in
1038  // PredCases and PredDefault with the new switch cases we would like to
1039  // build.
1040  SmallVector<BasicBlock *, 8> NewSuccessors;
1041 
1042  // Update the branch weight metadata along the way
1043  SmallVector<uint64_t, 8> Weights;
1044  bool PredHasWeights = HasBranchWeights(PTI);
1045  bool SuccHasWeights = HasBranchWeights(TI);
1046 
1047  if (PredHasWeights) {
1048  GetBranchWeights(PTI, Weights);
1049  // branch-weight metadata is inconsistent here.
1050  if (Weights.size() != 1 + PredCases.size())
1051  PredHasWeights = SuccHasWeights = false;
1052  } else if (SuccHasWeights)
1053  // If there are no predecessor weights but there are successor weights,
1054  // populate Weights with 1, which will later be scaled to the sum of
1055  // successor's weights
1056  Weights.assign(1 + PredCases.size(), 1);
1057 
1058  SmallVector<uint64_t, 8> SuccWeights;
1059  if (SuccHasWeights) {
1060  GetBranchWeights(TI, SuccWeights);
1061  // branch-weight metadata is inconsistent here.
1062  if (SuccWeights.size() != 1 + BBCases.size())
1063  PredHasWeights = SuccHasWeights = false;
1064  } else if (PredHasWeights)
1065  SuccWeights.assign(1 + BBCases.size(), 1);
1066 
1067  if (PredDefault == BB) {
1068  // If this is the default destination from PTI, only the edges in TI
1069  // that don't occur in PTI, or that branch to BB will be activated.
1070  std::set<ConstantInt *, ConstantIntOrdering> PTIHandled;
1071  for (unsigned i = 0, e = PredCases.size(); i != e; ++i)
1072  if (PredCases[i].Dest != BB)
1073  PTIHandled.insert(PredCases[i].Value);
1074  else {
1075  // The default destination is BB, we don't need explicit targets.
1076  std::swap(PredCases[i], PredCases.back());
1077 
1078  if (PredHasWeights || SuccHasWeights) {
1079  // Increase weight for the default case.
1080  Weights[0] += Weights[i + 1];
1081  std::swap(Weights[i + 1], Weights.back());
1082  Weights.pop_back();
1083  }
1084 
1085  PredCases.pop_back();
1086  --i;
1087  --e;
1088  }
1089 
1090  // Reconstruct the new switch statement we will be building.
1091  if (PredDefault != BBDefault) {
1092  PredDefault->removePredecessor(Pred);
1093  PredDefault = BBDefault;
1094  NewSuccessors.push_back(BBDefault);
1095  }
1096 
1097  unsigned CasesFromPred = Weights.size();
1098  uint64_t ValidTotalSuccWeight = 0;
1099  for (unsigned i = 0, e = BBCases.size(); i != e; ++i)
1100  if (!PTIHandled.count(BBCases[i].Value) &&
1101  BBCases[i].Dest != BBDefault) {
1102  PredCases.push_back(BBCases[i]);
1103  NewSuccessors.push_back(BBCases[i].Dest);
1104  if (SuccHasWeights || PredHasWeights) {
1105  // The default weight is at index 0, so weight for the ith case
1106  // should be at index i+1. Scale the cases from successor by
1107  // PredDefaultWeight (Weights[0]).
1108  Weights.push_back(Weights[0] * SuccWeights[i + 1]);
1109  ValidTotalSuccWeight += SuccWeights[i + 1];
1110  }
1111  }
1112 
1113  if (SuccHasWeights || PredHasWeights) {
1114  ValidTotalSuccWeight += SuccWeights[0];
1115  // Scale the cases from predecessor by ValidTotalSuccWeight.
1116  for (unsigned i = 1; i < CasesFromPred; ++i)
1117  Weights[i] *= ValidTotalSuccWeight;
1118  // Scale the default weight by SuccDefaultWeight (SuccWeights[0]).
1119  Weights[0] *= SuccWeights[0];
1120  }
1121  } else {
1122  // If this is not the default destination from PSI, only the edges
1123  // in SI that occur in PSI with a destination of BB will be
1124  // activated.
1125  std::set<ConstantInt *, ConstantIntOrdering> PTIHandled;
1126  std::map<ConstantInt *, uint64_t> WeightsForHandled;
1127  for (unsigned i = 0, e = PredCases.size(); i != e; ++i)
1128  if (PredCases[i].Dest == BB) {
1129  PTIHandled.insert(PredCases[i].Value);
1130 
1131  if (PredHasWeights || SuccHasWeights) {
1132  WeightsForHandled[PredCases[i].Value] = Weights[i + 1];
1133  std::swap(Weights[i + 1], Weights.back());
1134  Weights.pop_back();
1135  }
1136 
1137  std::swap(PredCases[i], PredCases.back());
1138  PredCases.pop_back();
1139  --i;
1140  --e;
1141  }
1142 
1143  // Okay, now we know which constants were sent to BB from the
1144  // predecessor. Figure out where they will all go now.
1145  for (unsigned i = 0, e = BBCases.size(); i != e; ++i)
1146  if (PTIHandled.count(BBCases[i].Value)) {
1147  // If this is one we are capable of getting...
1148  if (PredHasWeights || SuccHasWeights)
1149  Weights.push_back(WeightsForHandled[BBCases[i].Value]);
1150  PredCases.push_back(BBCases[i]);
1151  NewSuccessors.push_back(BBCases[i].Dest);
1152  PTIHandled.erase(
1153  BBCases[i].Value); // This constant is taken care of
1154  }
1155 
1156  // If there are any constants vectored to BB that TI doesn't handle,
1157  // they must go to the default destination of TI.
1158  for (ConstantInt *I : PTIHandled) {
1159  if (PredHasWeights || SuccHasWeights)
1160  Weights.push_back(WeightsForHandled[I]);
1161  PredCases.push_back(ValueEqualityComparisonCase(I, BBDefault));
1162  NewSuccessors.push_back(BBDefault);
1163  }
1164  }
1165 
1166  // Okay, at this point, we know which new successor Pred will get. Make
1167  // sure we update the number of entries in the PHI nodes for these
1168  // successors.
1169  for (BasicBlock *NewSuccessor : NewSuccessors)
1170  AddPredecessorToBlock(NewSuccessor, Pred, BB);
1171 
1172  Builder.SetInsertPoint(PTI);
1173  // Convert pointer to int before we switch.
1174  if (CV->getType()->isPointerTy()) {
1175  CV = Builder.CreatePtrToInt(CV, DL.getIntPtrType(CV->getType()),
1176  "magicptr");
1177  }
1178 
1179  // Now that the successors are updated, create the new Switch instruction.
1180  SwitchInst *NewSI =
1181  Builder.CreateSwitch(CV, PredDefault, PredCases.size());
1182  NewSI->setDebugLoc(PTI->getDebugLoc());
1183  for (ValueEqualityComparisonCase &V : PredCases)
1184  NewSI->addCase(V.Value, V.Dest);
1185 
1186  if (PredHasWeights || SuccHasWeights) {
1187  // Halve the weights if any of them cannot fit in an uint32_t
1188  FitWeights(Weights);
1189 
1190  SmallVector<uint32_t, 8> MDWeights(Weights.begin(), Weights.end());
1191 
1192  setBranchWeights(NewSI, MDWeights);
1193  }
1194 
1196 
1197  // Okay, last check. If BB is still a successor of PSI, then we must
1198  // have an infinite loop case. If so, add an infinitely looping block
1199  // to handle the case to preserve the behavior of the code.
1200  BasicBlock *InfLoopBlock = nullptr;
1201  for (unsigned i = 0, e = NewSI->getNumSuccessors(); i != e; ++i)
1202  if (NewSI->getSuccessor(i) == BB) {
1203  if (!InfLoopBlock) {
1204  // Insert it at the end of the function, because it's either code,
1205  // or it won't matter if it's hot. :)
1206  InfLoopBlock = BasicBlock::Create(BB->getContext(), "infloop",
1207  BB->getParent());
1208  BranchInst::Create(InfLoopBlock, InfLoopBlock);
1209  }
1210  NewSI->setSuccessor(i, InfLoopBlock);
1211  }
1212 
1213  Changed = true;
1214  }
1215  }
1216  return Changed;
1217 }
1218 
1219 // If we would need to insert a select that uses the value of this invoke
1220 // (comments in HoistThenElseCodeToIf explain why we would need to do this), we
1221 // can't hoist the invoke, as there is nowhere to put the select in this case.
1223  Instruction *I1, Instruction *I2) {
1224  for (BasicBlock *Succ : successors(BB1)) {
1225  for (const PHINode &PN : Succ->phis()) {
1226  Value *BB1V = PN.getIncomingValueForBlock(BB1);
1227  Value *BB2V = PN.getIncomingValueForBlock(BB2);
1228  if (BB1V != BB2V && (BB1V == I1 || BB2V == I2)) {
1229  return false;
1230  }
1231  }
1232  }
1233  return true;
1234 }
1235 
1237 
1238 /// Given a conditional branch that goes to BB1 and BB2, hoist any common code
1239 /// in the two blocks up into the branch block. The caller of this function
1240 /// guarantees that BI's block dominates BB1 and BB2.
1242  const TargetTransformInfo &TTI) {
1243  // This does very trivial matching, with limited scanning, to find identical
1244  // instructions in the two blocks. In particular, we don't want to get into
1245  // O(M*N) situations here where M and N are the sizes of BB1 and BB2. As
1246  // such, we currently just scan for obviously identical instructions in an
1247  // identical order.
1248  BasicBlock *BB1 = BI->getSuccessor(0); // The true destination.
1249  BasicBlock *BB2 = BI->getSuccessor(1); // The false destination
1250 
1251  BasicBlock::iterator BB1_Itr = BB1->begin();
1252  BasicBlock::iterator BB2_Itr = BB2->begin();
1253 
1254  Instruction *I1 = &*BB1_Itr++, *I2 = &*BB2_Itr++;
1255  // Skip debug info if it is not identical.
1258  if (!DBI1 || !DBI2 || !DBI1->isIdenticalToWhenDefined(DBI2)) {
1259  while (isa<DbgInfoIntrinsic>(I1))
1260  I1 = &*BB1_Itr++;
1261  while (isa<DbgInfoIntrinsic>(I2))
1262  I2 = &*BB2_Itr++;
1263  }
1264  if (isa<PHINode>(I1) || !I1->isIdenticalToWhenDefined(I2) ||
1265  (isa<InvokeInst>(I1) && !isSafeToHoistInvoke(BB1, BB2, I1, I2)))
1266  return false;
1267 
1268  BasicBlock *BIParent = BI->getParent();
1269 
1270  bool Changed = false;
1271  do {
1272  // If we are hoisting the terminator instruction, don't move one (making a
1273  // broken BB), instead clone it, and remove BI.
1274  if (isa<TerminatorInst>(I1))
1275  goto HoistTerminator;
1276 
1277  // If we're going to hoist a call, make sure that the two instructions we're
1278  // commoning/hoisting are both marked with musttail, or neither of them is
1279  // marked as such. Otherwise, we might end up in a situation where we hoist
1280  // from a block where the terminator is a `ret` to a block where the terminator
1281  // is a `br`, and `musttail` calls expect to be followed by a return.
1282  auto *C1 = dyn_cast<CallInst>(I1);
1283  auto *C2 = dyn_cast<CallInst>(I2);
1284  if (C1 && C2)
1285  if (C1->isMustTailCall() != C2->isMustTailCall())
1286  return Changed;
1287 
1288  if (!TTI.isProfitableToHoist(I1) || !TTI.isProfitableToHoist(I2))
1289  return Changed;
1290 
1291  if (isa<DbgInfoIntrinsic>(I1) || isa<DbgInfoIntrinsic>(I2)) {
1292  assert (isa<DbgInfoIntrinsic>(I1) && isa<DbgInfoIntrinsic>(I2));
1293  // The debug location is an integral part of a debug info intrinsic
1294  // and can't be separated from it or replaced. Instead of attempting
1295  // to merge locations, simply hoist both copies of the intrinsic.
1296  BIParent->getInstList().splice(BI->getIterator(),
1297  BB1->getInstList(), I1);
1298  BIParent->getInstList().splice(BI->getIterator(),
1299  BB2->getInstList(), I2);
1300  Changed = true;
1301  } else {
1302  // For a normal instruction, we just move one to right before the branch,
1303  // then replace all uses of the other with the first. Finally, we remove
1304  // the now redundant second instruction.
1305  BIParent->getInstList().splice(BI->getIterator(),
1306  BB1->getInstList(), I1);
1307  if (!I2->use_empty())
1308  I2->replaceAllUsesWith(I1);
1309  I1->andIRFlags(I2);
1310  unsigned KnownIDs[] = {LLVMContext::MD_tbaa,
1320  combineMetadata(I1, I2, KnownIDs);
1321 
1322  // I1 and I2 are being combined into a single instruction. Its debug
1323  // location is the merged locations of the original instructions.
1324  I1->applyMergedLocation(I1->getDebugLoc(), I2->getDebugLoc());
1325 
1326  I2->eraseFromParent();
1327  Changed = true;
1328  }
1329 
1330  I1 = &*BB1_Itr++;
1331  I2 = &*BB2_Itr++;
1332  // Skip debug info if it is not identical.
1335  if (!DBI1 || !DBI2 || !DBI1->isIdenticalToWhenDefined(DBI2)) {
1336  while (isa<DbgInfoIntrinsic>(I1))
1337  I1 = &*BB1_Itr++;
1338  while (isa<DbgInfoIntrinsic>(I2))
1339  I2 = &*BB2_Itr++;
1340  }
1341  } while (I1->isIdenticalToWhenDefined(I2));
1342 
1343  return true;
1344 
1345 HoistTerminator:
1346  // It may not be possible to hoist an invoke.
1347  if (isa<InvokeInst>(I1) && !isSafeToHoistInvoke(BB1, BB2, I1, I2))
1348  return Changed;
1349 
1350  for (BasicBlock *Succ : successors(BB1)) {
1351  for (PHINode &PN : Succ->phis()) {
1352  Value *BB1V = PN.getIncomingValueForBlock(BB1);
1353  Value *BB2V = PN.getIncomingValueForBlock(BB2);
1354  if (BB1V == BB2V)
1355  continue;
1356 
1357  // Check for passingValueIsAlwaysUndefined here because we would rather
1358  // eliminate undefined control flow then converting it to a select.
1359  if (passingValueIsAlwaysUndefined(BB1V, &PN) ||
1360  passingValueIsAlwaysUndefined(BB2V, &PN))
1361  return Changed;
1362 
1363  if (isa<ConstantExpr>(BB1V) && !isSafeToSpeculativelyExecute(BB1V))
1364  return Changed;
1365  if (isa<ConstantExpr>(BB2V) && !isSafeToSpeculativelyExecute(BB2V))
1366  return Changed;
1367  }
1368  }
1369 
1370  // Okay, it is safe to hoist the terminator.
1371  Instruction *NT = I1->clone();
1372  BIParent->getInstList().insert(BI->getIterator(), NT);
1373  if (!NT->getType()->isVoidTy()) {
1374  I1->replaceAllUsesWith(NT);
1375  I2->replaceAllUsesWith(NT);
1376  NT->takeName(I1);
1377  }
1378 
1379  IRBuilder<NoFolder> Builder(NT);
1380  // Hoisting one of the terminators from our successor is a great thing.
1381  // Unfortunately, the successors of the if/else blocks may have PHI nodes in
1382  // them. If they do, all PHI entries for BB1/BB2 must agree for all PHI
1383  // nodes, so we insert select instruction to compute the final result.
1384  std::map<std::pair<Value *, Value *>, SelectInst *> InsertedSelects;
1385  for (BasicBlock *Succ : successors(BB1)) {
1386  for (PHINode &PN : Succ->phis()) {
1387  Value *BB1V = PN.getIncomingValueForBlock(BB1);
1388  Value *BB2V = PN.getIncomingValueForBlock(BB2);
1389  if (BB1V == BB2V)
1390  continue;
1391 
1392  // These values do not agree. Insert a select instruction before NT
1393  // that determines the right value.
1394  SelectInst *&SI = InsertedSelects[std::make_pair(BB1V, BB2V)];
1395  if (!SI)
1396  SI = cast<SelectInst>(
1397  Builder.CreateSelect(BI->getCondition(), BB1V, BB2V,
1398  BB1V->getName() + "." + BB2V->getName(), BI));
1399 
1400  // Make the PHI node use the select for all incoming values for BB1/BB2
1401  for (unsigned i = 0, e = PN.getNumIncomingValues(); i != e; ++i)
1402  if (PN.getIncomingBlock(i) == BB1 || PN.getIncomingBlock(i) == BB2)
1403  PN.setIncomingValue(i, SI);
1404  }
1405  }
1406 
1407  // Update any PHI nodes in our new successors.
1408  for (BasicBlock *Succ : successors(BB1))
1409  AddPredecessorToBlock(Succ, BIParent, BB1);
1410 
1412  return true;
1413 }
1414 
1415 // All instructions in Insts belong to different blocks that all unconditionally
1416 // branch to a common successor. Analyze each instruction and return true if it
1417 // would be possible to sink them into their successor, creating one common
1418 // instruction instead. For every value that would be required to be provided by
1419 // PHI node (because an operand varies in each input block), add to PHIOperands.
1422  DenseMap<Instruction *, SmallVector<Value *, 4>> &PHIOperands) {
1423  // Prune out obviously bad instructions to move. Any non-store instruction
1424  // must have exactly one use, and we check later that use is by a single,
1425  // common PHI instruction in the successor.
1426  for (auto *I : Insts) {
1427  // These instructions may change or break semantics if moved.
1428  if (isa<PHINode>(I) || I->isEHPad() || isa<AllocaInst>(I) ||
1429  I->getType()->isTokenTy())
1430  return false;
1431 
1432  // Conservatively return false if I is an inline-asm instruction. Sinking
1433  // and merging inline-asm instructions can potentially create arguments
1434  // that cannot satisfy the inline-asm constraints.
1435  if (const auto *C = dyn_cast<CallInst>(I))
1436  if (C->isInlineAsm())
1437  return false;
1438 
1439  // Everything must have only one use too, apart from stores which
1440  // have no uses.
1441  if (!isa<StoreInst>(I) && !I->hasOneUse())
1442  return false;
1443  }
1444 
1445  const Instruction *I0 = Insts.front();
1446  for (auto *I : Insts)
1447  if (!I->isSameOperationAs(I0))
1448  return false;
1449 
1450  // All instructions in Insts are known to be the same opcode. If they aren't
1451  // stores, check the only user of each is a PHI or in the same block as the
1452  // instruction, because if a user is in the same block as an instruction
1453  // we're contemplating sinking, it must already be determined to be sinkable.
1454  if (!isa<StoreInst>(I0)) {
1455  auto *PNUse = dyn_cast<PHINode>(*I0->user_begin());
1456  auto *Succ = I0->getParent()->getTerminator()->getSuccessor(0);
1457  if (!all_of(Insts, [&PNUse,&Succ](const Instruction *I) -> bool {
1458  auto *U = cast<Instruction>(*I->user_begin());
1459  return (PNUse &&
1460  PNUse->getParent() == Succ &&
1461  PNUse->getIncomingValueForBlock(I->getParent()) == I) ||
1462  U->getParent() == I->getParent();
1463  }))
1464  return false;
1465  }
1466 
1467  // Because SROA can't handle speculating stores of selects, try not
1468  // to sink loads or stores of allocas when we'd have to create a PHI for
1469  // the address operand. Also, because it is likely that loads or stores
1470  // of allocas will disappear when Mem2Reg/SROA is run, don't sink them.
1471  // This can cause code churn which can have unintended consequences down
1472  // the line - see https://llvm.org/bugs/show_bug.cgi?id=30244.
1473  // FIXME: This is a workaround for a deficiency in SROA - see
1474  // https://llvm.org/bugs/show_bug.cgi?id=30188
1475  if (isa<StoreInst>(I0) && any_of(Insts, [](const Instruction *I) {
1476  return isa<AllocaInst>(I->getOperand(1));
1477  }))
1478  return false;
1479  if (isa<LoadInst>(I0) && any_of(Insts, [](const Instruction *I) {
1480  return isa<AllocaInst>(I->getOperand(0));
1481  }))
1482  return false;
1483 
1484  for (unsigned OI = 0, OE = I0->getNumOperands(); OI != OE; ++OI) {
1485  if (I0->getOperand(OI)->getType()->isTokenTy())
1486  // Don't touch any operand of token type.
1487  return false;
1488 
1489  auto SameAsI0 = [&I0, OI](const Instruction *I) {
1490  assert(I->getNumOperands() == I0->getNumOperands());
1491  return I->getOperand(OI) == I0->getOperand(OI);
1492  };
1493  if (!all_of(Insts, SameAsI0)) {
1494  if (!canReplaceOperandWithVariable(I0, OI))
1495  // We can't create a PHI from this GEP.
1496  return false;
1497  // Don't create indirect calls! The called value is the final operand.
1498  if ((isa<CallInst>(I0) || isa<InvokeInst>(I0)) && OI == OE - 1) {
1499  // FIXME: if the call was *already* indirect, we should do this.
1500  return false;
1501  }
1502  for (auto *I : Insts)
1503  PHIOperands[I].push_back(I->getOperand(OI));
1504  }
1505  }
1506  return true;
1507 }
1508 
1509 // Assuming canSinkLastInstruction(Blocks) has returned true, sink the last
1510 // instruction of every block in Blocks to their common successor, commoning
1511 // into one instruction.
1513  auto *BBEnd = Blocks[0]->getTerminator()->getSuccessor(0);
1514 
1515  // canSinkLastInstruction returning true guarantees that every block has at
1516  // least one non-terminator instruction.
1518  for (auto *BB : Blocks) {
1519  Instruction *I = BB->getTerminator();
1520  do {
1521  I = I->getPrevNode();
1522  } while (isa<DbgInfoIntrinsic>(I) && I != &BB->front());
1523  if (!isa<DbgInfoIntrinsic>(I))
1524  Insts.push_back(I);
1525  }
1526 
1527  // The only checking we need to do now is that all users of all instructions
1528  // are the same PHI node. canSinkLastInstruction should have checked this but
1529  // it is slightly over-aggressive - it gets confused by commutative instructions
1530  // so double-check it here.
1531  Instruction *I0 = Insts.front();
1532  if (!isa<StoreInst>(I0)) {
1533  auto *PNUse = dyn_cast<PHINode>(*I0->user_begin());
1534  if (!all_of(Insts, [&PNUse](const Instruction *I) -> bool {
1535  auto *U = cast<Instruction>(*I->user_begin());
1536  return U == PNUse;
1537  }))
1538  return false;
1539  }
1540 
1541  // We don't need to do any more checking here; canSinkLastInstruction should
1542  // have done it all for us.
1543  SmallVector<Value*, 4> NewOperands;
1544  for (unsigned O = 0, E = I0->getNumOperands(); O != E; ++O) {
1545  // This check is different to that in canSinkLastInstruction. There, we
1546  // cared about the global view once simplifycfg (and instcombine) have
1547  // completed - it takes into account PHIs that become trivially
1548  // simplifiable. However here we need a more local view; if an operand
1549  // differs we create a PHI and rely on instcombine to clean up the very
1550  // small mess we may make.
1551  bool NeedPHI = any_of(Insts, [&I0, O](const Instruction *I) {
1552  return I->getOperand(O) != I0->getOperand(O);
1553  });
1554  if (!NeedPHI) {
1555  NewOperands.push_back(I0->getOperand(O));
1556  continue;
1557  }
1558 
1559  // Create a new PHI in the successor block and populate it.
1560  auto *Op = I0->getOperand(O);
1561  assert(!Op->getType()->isTokenTy() && "Can't PHI tokens!");
1562  auto *PN = PHINode::Create(Op->getType(), Insts.size(),
1563  Op->getName() + ".sink", &BBEnd->front());
1564  for (auto *I : Insts)
1565  PN->addIncoming(I->getOperand(O), I->getParent());
1566  NewOperands.push_back(PN);
1567  }
1568 
1569  // Arbitrarily use I0 as the new "common" instruction; remap its operands
1570  // and move it to the start of the successor block.
1571  for (unsigned O = 0, E = I0->getNumOperands(); O != E; ++O)
1572  I0->getOperandUse(O).set(NewOperands[O]);
1573  I0->moveBefore(&*BBEnd->getFirstInsertionPt());
1574 
1575  // Update metadata and IR flags, and merge debug locations.
1576  for (auto *I : Insts)
1577  if (I != I0) {
1578  // The debug location for the "common" instruction is the merged locations
1579  // of all the commoned instructions. We start with the original location
1580  // of the "common" instruction and iteratively merge each location in the
1581  // loop below.
1582  // This is an N-way merge, which will be inefficient if I0 is a CallInst.
1583  // However, as N-way merge for CallInst is rare, so we use simplified API
1584  // instead of using complex API for N-way merge.
1585  I0->applyMergedLocation(I0->getDebugLoc(), I->getDebugLoc());
1586  combineMetadataForCSE(I0, I);
1587  I0->andIRFlags(I);
1588  }
1589 
1590  if (!isa<StoreInst>(I0)) {
1591  // canSinkLastInstruction checked that all instructions were used by
1592  // one and only one PHI node. Find that now, RAUW it to our common
1593  // instruction and nuke it.
1594  assert(I0->hasOneUse());
1595  auto *PN = cast<PHINode>(*I0->user_begin());
1596  PN->replaceAllUsesWith(I0);
1597  PN->eraseFromParent();
1598  }
1599 
1600  // Finally nuke all instructions apart from the common instruction.
1601  for (auto *I : Insts)
1602  if (I != I0)
1603  I->eraseFromParent();
1604 
1605  return true;
1606 }
1607 
1608 namespace {
1609 
1610  // LockstepReverseIterator - Iterates through instructions
1611  // in a set of blocks in reverse order from the first non-terminator.
1612  // For example (assume all blocks have size n):
1613  // LockstepReverseIterator I([B1, B2, B3]);
1614  // *I-- = [B1[n], B2[n], B3[n]];
1615  // *I-- = [B1[n-1], B2[n-1], B3[n-1]];
1616  // *I-- = [B1[n-2], B2[n-2], B3[n-2]];
1617  // ...
1618  class LockstepReverseIterator {
1619  ArrayRef<BasicBlock*> Blocks;
1621  bool Fail;
1622 
1623  public:
1624  LockstepReverseIterator(ArrayRef<BasicBlock*> Blocks) : Blocks(Blocks) {
1625  reset();
1626  }
1627 
1628  void reset() {
1629  Fail = false;
1630  Insts.clear();
1631  for (auto *BB : Blocks) {
1632  Instruction *Inst = BB->getTerminator();
1633  for (Inst = Inst->getPrevNode(); Inst && isa<DbgInfoIntrinsic>(Inst);)
1634  Inst = Inst->getPrevNode();
1635  if (!Inst) {
1636  // Block wasn't big enough.
1637  Fail = true;
1638  return;
1639  }
1640  Insts.push_back(Inst);
1641  }
1642  }
1643 
1644  bool isValid() const {
1645  return !Fail;
1646  }
1647 
1648  void operator--() {
1649  if (Fail)
1650  return;
1651  for (auto *&Inst : Insts) {
1652  for (Inst = Inst->getPrevNode(); Inst && isa<DbgInfoIntrinsic>(Inst);)
1653  Inst = Inst->getPrevNode();
1654  // Already at beginning of block.
1655  if (!Inst) {
1656  Fail = true;
1657  return;
1658  }
1659  }
1660  }
1661 
1663  return Insts;
1664  }
1665  };
1666 
1667 } // end anonymous namespace
1668 
1669 /// Check whether BB's predecessors end with unconditional branches. If it is
1670 /// true, sink any common code from the predecessors to BB.
1671 /// We also allow one predecessor to end with conditional branch (but no more
1672 /// than one).
1674  // We support two situations:
1675  // (1) all incoming arcs are unconditional
1676  // (2) one incoming arc is conditional
1677  //
1678  // (2) is very common in switch defaults and
1679  // else-if patterns;
1680  //
1681  // if (a) f(1);
1682  // else if (b) f(2);
1683  //
1684  // produces:
1685  //
1686  // [if]
1687  // / \
1688  // [f(1)] [if]
1689  // | | \
1690  // | | |
1691  // | [f(2)]|
1692  // \ | /
1693  // [ end ]
1694  //
1695  // [end] has two unconditional predecessor arcs and one conditional. The
1696  // conditional refers to the implicit empty 'else' arc. This conditional
1697  // arc can also be caused by an empty default block in a switch.
1698  //
1699  // In this case, we attempt to sink code from all *unconditional* arcs.
1700  // If we can sink instructions from these arcs (determined during the scan
1701  // phase below) we insert a common successor for all unconditional arcs and
1702  // connect that to [end], to enable sinking:
1703  //
1704  // [if]
1705  // / \
1706  // [x(1)] [if]
1707  // | | \
1708  // | | \
1709  // | [x(2)] |
1710  // \ / |
1711  // [sink.split] |
1712  // \ /
1713  // [ end ]
1714  //
1715  SmallVector<BasicBlock*,4> UnconditionalPreds;
1716  Instruction *Cond = nullptr;
1717  for (auto *B : predecessors(BB)) {
1718  auto *T = B->getTerminator();
1719  if (isa<BranchInst>(T) && cast<BranchInst>(T)->isUnconditional())
1720  UnconditionalPreds.push_back(B);
1721  else if ((isa<BranchInst>(T) || isa<SwitchInst>(T)) && !Cond)
1722  Cond = T;
1723  else
1724  return false;
1725  }
1726  if (UnconditionalPreds.size() < 2)
1727  return false;
1728 
1729  bool Changed = false;
1730  // We take a two-step approach to tail sinking. First we scan from the end of
1731  // each block upwards in lockstep. If the n'th instruction from the end of each
1732  // block can be sunk, those instructions are added to ValuesToSink and we
1733  // carry on. If we can sink an instruction but need to PHI-merge some operands
1734  // (because they're not identical in each instruction) we add these to
1735  // PHIOperands.
1736  unsigned ScanIdx = 0;
1737  SmallPtrSet<Value*,4> InstructionsToSink;
1739  LockstepReverseIterator LRI(UnconditionalPreds);
1740  while (LRI.isValid() &&
1741  canSinkInstructions(*LRI, PHIOperands)) {
1742  LLVM_DEBUG(dbgs() << "SINK: instruction can be sunk: " << *(*LRI)[0]
1743  << "\n");
1744  InstructionsToSink.insert((*LRI).begin(), (*LRI).end());
1745  ++ScanIdx;
1746  --LRI;
1747  }
1748 
1749  auto ProfitableToSinkInstruction = [&](LockstepReverseIterator &LRI) {
1750  unsigned NumPHIdValues = 0;
1751  for (auto *I : *LRI)
1752  for (auto *V : PHIOperands[I])
1753  if (InstructionsToSink.count(V) == 0)
1754  ++NumPHIdValues;
1755  LLVM_DEBUG(dbgs() << "SINK: #phid values: " << NumPHIdValues << "\n");
1756  unsigned NumPHIInsts = NumPHIdValues / UnconditionalPreds.size();
1757  if ((NumPHIdValues % UnconditionalPreds.size()) != 0)
1758  NumPHIInsts++;
1759 
1760  return NumPHIInsts <= 1;
1761  };
1762 
1763  if (ScanIdx > 0 && Cond) {
1764  // Check if we would actually sink anything first! This mutates the CFG and
1765  // adds an extra block. The goal in doing this is to allow instructions that
1766  // couldn't be sunk before to be sunk - obviously, speculatable instructions
1767  // (such as trunc, add) can be sunk and predicated already. So we check that
1768  // we're going to sink at least one non-speculatable instruction.
1769  LRI.reset();
1770  unsigned Idx = 0;
1771  bool Profitable = false;
1772  while (ProfitableToSinkInstruction(LRI) && Idx < ScanIdx) {
1773  if (!isSafeToSpeculativelyExecute((*LRI)[0])) {
1774  Profitable = true;
1775  break;
1776  }
1777  --LRI;
1778  ++Idx;
1779  }
1780  if (!Profitable)
1781  return false;
1782 
1783  LLVM_DEBUG(dbgs() << "SINK: Splitting edge\n");
1784  // We have a conditional edge and we're going to sink some instructions.
1785  // Insert a new block postdominating all blocks we're going to sink from.
1786  if (!SplitBlockPredecessors(BB, UnconditionalPreds, ".sink.split"))
1787  // Edges couldn't be split.
1788  return false;
1789  Changed = true;
1790  }
1791 
1792  // Now that we've analyzed all potential sinking candidates, perform the
1793  // actual sink. We iteratively sink the last non-terminator of the source
1794  // blocks into their common successor unless doing so would require too
1795  // many PHI instructions to be generated (currently only one PHI is allowed
1796  // per sunk instruction).
1797  //
1798  // We can use InstructionsToSink to discount values needing PHI-merging that will
1799  // actually be sunk in a later iteration. This allows us to be more
1800  // aggressive in what we sink. This does allow a false positive where we
1801  // sink presuming a later value will also be sunk, but stop half way through
1802  // and never actually sink it which means we produce more PHIs than intended.
1803  // This is unlikely in practice though.
1804  for (unsigned SinkIdx = 0; SinkIdx != ScanIdx; ++SinkIdx) {
1805  LLVM_DEBUG(dbgs() << "SINK: Sink: "
1806  << *UnconditionalPreds[0]->getTerminator()->getPrevNode()
1807  << "\n");
1808 
1809  // Because we've sunk every instruction in turn, the current instruction to
1810  // sink is always at index 0.
1811  LRI.reset();
1812  if (!ProfitableToSinkInstruction(LRI)) {
1813  // Too many PHIs would be created.
1814  LLVM_DEBUG(
1815  dbgs() << "SINK: stopping here, too many PHIs would be created!\n");
1816  break;
1817  }
1818 
1819  if (!sinkLastInstruction(UnconditionalPreds))
1820  return Changed;
1821  NumSinkCommons++;
1822  Changed = true;
1823  }
1824  return Changed;
1825 }
1826 
1827 /// Determine if we can hoist sink a sole store instruction out of a
1828 /// conditional block.
1829 ///
1830 /// We are looking for code like the following:
1831 /// BrBB:
1832 /// store i32 %add, i32* %arrayidx2
1833 /// ... // No other stores or function calls (we could be calling a memory
1834 /// ... // function).
1835 /// %cmp = icmp ult %x, %y
1836 /// br i1 %cmp, label %EndBB, label %ThenBB
1837 /// ThenBB:
1838 /// store i32 %add5, i32* %arrayidx2
1839 /// br label EndBB
1840 /// EndBB:
1841 /// ...
1842 /// We are going to transform this into:
1843 /// BrBB:
1844 /// store i32 %add, i32* %arrayidx2
1845 /// ... //
1846 /// %cmp = icmp ult %x, %y
1847 /// %add.add5 = select i1 %cmp, i32 %add, %add5
1848 /// store i32 %add.add5, i32* %arrayidx2
1849 /// ...
1850 ///
1851 /// \return The pointer to the value of the previous store if the store can be
1852 /// hoisted into the predecessor block. 0 otherwise.
1854  BasicBlock *StoreBB, BasicBlock *EndBB) {
1855  StoreInst *StoreToHoist = dyn_cast<StoreInst>(I);
1856  if (!StoreToHoist)
1857  return nullptr;
1858 
1859  // Volatile or atomic.
1860  if (!StoreToHoist->isSimple())
1861  return nullptr;
1862 
1863  Value *StorePtr = StoreToHoist->getPointerOperand();
1864 
1865  // Look for a store to the same pointer in BrBB.
1866  unsigned MaxNumInstToLookAt = 9;
1867  for (Instruction &CurI : reverse(BrBB->instructionsWithoutDebug())) {
1868  if (!MaxNumInstToLookAt)
1869  break;
1870  --MaxNumInstToLookAt;
1871 
1872  // Could be calling an instruction that affects memory like free().
1873  if (CurI.mayHaveSideEffects() && !isa<StoreInst>(CurI))
1874  return nullptr;
1875 
1876  if (auto *SI = dyn_cast<StoreInst>(&CurI)) {
1877  // Found the previous store make sure it stores to the same location.
1878  if (SI->getPointerOperand() == StorePtr)
1879  // Found the previous store, return its value operand.
1880  return SI->getValueOperand();
1881  return nullptr; // Unknown store.
1882  }
1883  }
1884 
1885  return nullptr;
1886 }
1887 
1888 /// Speculate a conditional basic block flattening the CFG.
1889 ///
1890 /// Note that this is a very risky transform currently. Speculating
1891 /// instructions like this is most often not desirable. Instead, there is an MI
1892 /// pass which can do it with full awareness of the resource constraints.
1893 /// However, some cases are "obvious" and we should do directly. An example of
1894 /// this is speculating a single, reasonably cheap instruction.
1895 ///
1896 /// There is only one distinct advantage to flattening the CFG at the IR level:
1897 /// it makes very common but simplistic optimizations such as are common in
1898 /// instcombine and the DAG combiner more powerful by removing CFG edges and
1899 /// modeling their effects with easier to reason about SSA value graphs.
1900 ///
1901 ///
1902 /// An illustration of this transform is turning this IR:
1903 /// \code
1904 /// BB:
1905 /// %cmp = icmp ult %x, %y
1906 /// br i1 %cmp, label %EndBB, label %ThenBB
1907 /// ThenBB:
1908 /// %sub = sub %x, %y
1909 /// br label BB2
1910 /// EndBB:
1911 /// %phi = phi [ %sub, %ThenBB ], [ 0, %EndBB ]
1912 /// ...
1913 /// \endcode
1914 ///
1915 /// Into this IR:
1916 /// \code
1917 /// BB:
1918 /// %cmp = icmp ult %x, %y
1919 /// %sub = sub %x, %y
1920 /// %cond = select i1 %cmp, 0, %sub
1921 /// ...
1922 /// \endcode
1923 ///
1924 /// \returns true if the conditional block is removed.
1926  const TargetTransformInfo &TTI) {
1927  // Be conservative for now. FP select instruction can often be expensive.
1928  Value *BrCond = BI->getCondition();
1929  if (isa<FCmpInst>(BrCond))
1930  return false;
1931 
1932  BasicBlock *BB = BI->getParent();
1933  BasicBlock *EndBB = ThenBB->getTerminator()->getSuccessor(0);
1934 
1935  // If ThenBB is actually on the false edge of the conditional branch, remember
1936  // to swap the select operands later.
1937  bool Invert = false;
1938  if (ThenBB != BI->getSuccessor(0)) {
1939  assert(ThenBB == BI->getSuccessor(1) && "No edge from 'if' block?");
1940  Invert = true;
1941  }
1942  assert(EndBB == BI->getSuccessor(!Invert) && "No edge from to end block");
1943 
1944  // Keep a count of how many times instructions are used within CondBB when
1945  // they are candidates for sinking into CondBB. Specifically:
1946  // - They are defined in BB, and
1947  // - They have no side effects, and
1948  // - All of their uses are in CondBB.
1949  SmallDenseMap<Instruction *, unsigned, 4> SinkCandidateUseCounts;
1950 
1951  SmallVector<Instruction *, 4> SpeculatedDbgIntrinsics;
1952 
1953  unsigned SpeculationCost = 0;
1954  Value *SpeculatedStoreValue = nullptr;
1955  StoreInst *SpeculatedStore = nullptr;
1956  for (BasicBlock::iterator BBI = ThenBB->begin(),
1957  BBE = std::prev(ThenBB->end());
1958  BBI != BBE; ++BBI) {
1959  Instruction *I = &*BBI;
1960  // Skip debug info.
1961  if (isa<DbgInfoIntrinsic>(I)) {
1962  SpeculatedDbgIntrinsics.push_back(I);
1963  continue;
1964  }
1965 
1966  // Only speculatively execute a single instruction (not counting the
1967  // terminator) for now.
1968  ++SpeculationCost;
1969  if (SpeculationCost > 1)
1970  return false;
1971 
1972  // Don't hoist the instruction if it's unsafe or expensive.
1973  if (!isSafeToSpeculativelyExecute(I) &&
1974  !(HoistCondStores && (SpeculatedStoreValue = isSafeToSpeculateStore(
1975  I, BB, ThenBB, EndBB))))
1976  return false;
1977  if (!SpeculatedStoreValue &&
1978  ComputeSpeculationCost(I, TTI) >
1980  return false;
1981 
1982  // Store the store speculation candidate.
1983  if (SpeculatedStoreValue)
1984  SpeculatedStore = cast<StoreInst>(I);
1985 
1986  // Do not hoist the instruction if any of its operands are defined but not
1987  // used in BB. The transformation will prevent the operand from
1988  // being sunk into the use block.
1989  for (User::op_iterator i = I->op_begin(), e = I->op_end(); i != e; ++i) {
1990  Instruction *OpI = dyn_cast<Instruction>(*i);
1991  if (!OpI || OpI->getParent() != BB || OpI->mayHaveSideEffects())
1992  continue; // Not a candidate for sinking.
1993 
1994  ++SinkCandidateUseCounts[OpI];
1995  }
1996  }
1997 
1998  // Consider any sink candidates which are only used in CondBB as costs for
1999  // speculation. Note, while we iterate over a DenseMap here, we are summing
2000  // and so iteration order isn't significant.
2002  I = SinkCandidateUseCounts.begin(),
2003  E = SinkCandidateUseCounts.end();
2004  I != E; ++I)
2005  if (I->first->getNumUses() == I->second) {
2006  ++SpeculationCost;
2007  if (SpeculationCost > 1)
2008  return false;
2009  }
2010 
2011  // Check that the PHI nodes can be converted to selects.
2012  bool HaveRewritablePHIs = false;
2013  for (PHINode &PN : EndBB->phis()) {
2014  Value *OrigV = PN.getIncomingValueForBlock(BB);
2015  Value *ThenV = PN.getIncomingValueForBlock(ThenBB);
2016 
2017  // FIXME: Try to remove some of the duplication with HoistThenElseCodeToIf.
2018  // Skip PHIs which are trivial.
2019  if (ThenV == OrigV)
2020  continue;
2021 
2022  // Don't convert to selects if we could remove undefined behavior instead.
2023  if (passingValueIsAlwaysUndefined(OrigV, &PN) ||
2024  passingValueIsAlwaysUndefined(ThenV, &PN))
2025  return false;
2026 
2027  HaveRewritablePHIs = true;
2028  ConstantExpr *OrigCE = dyn_cast<ConstantExpr>(OrigV);
2029  ConstantExpr *ThenCE = dyn_cast<ConstantExpr>(ThenV);
2030  if (!OrigCE && !ThenCE)
2031  continue; // Known safe and cheap.
2032 
2033  if ((ThenCE && !isSafeToSpeculativelyExecute(ThenCE)) ||
2034  (OrigCE && !isSafeToSpeculativelyExecute(OrigCE)))
2035  return false;
2036  unsigned OrigCost = OrigCE ? ComputeSpeculationCost(OrigCE, TTI) : 0;
2037  unsigned ThenCost = ThenCE ? ComputeSpeculationCost(ThenCE, TTI) : 0;
2038  unsigned MaxCost =
2040  if (OrigCost + ThenCost > MaxCost)
2041  return false;
2042 
2043  // Account for the cost of an unfolded ConstantExpr which could end up
2044  // getting expanded into Instructions.
2045  // FIXME: This doesn't account for how many operations are combined in the
2046  // constant expression.
2047  ++SpeculationCost;
2048  if (SpeculationCost > 1)
2049  return false;
2050  }
2051 
2052  // If there are no PHIs to process, bail early. This helps ensure idempotence
2053  // as well.
2054  if (!HaveRewritablePHIs && !(HoistCondStores && SpeculatedStoreValue))
2055  return false;
2056 
2057  // If we get here, we can hoist the instruction and if-convert.
2058  LLVM_DEBUG(dbgs() << "SPECULATIVELY EXECUTING BB" << *ThenBB << "\n";);
2059 
2060  // Insert a select of the value of the speculated store.
2061  if (SpeculatedStoreValue) {
2062  IRBuilder<NoFolder> Builder(BI);
2063  Value *TrueV = SpeculatedStore->getValueOperand();
2064  Value *FalseV = SpeculatedStoreValue;
2065  if (Invert)
2066  std::swap(TrueV, FalseV);
2067  Value *S = Builder.CreateSelect(
2068  BrCond, TrueV, FalseV, "spec.store.select", BI);
2069  SpeculatedStore->setOperand(0, S);
2070  SpeculatedStore->applyMergedLocation(BI->getDebugLoc(),
2071  SpeculatedStore->getDebugLoc());
2072  }
2073 
2074  // Metadata can be dependent on the condition we are hoisting above.
2075  // Conservatively strip all metadata on the instruction.
2076  for (auto &I : *ThenBB)
2078 
2079  // Hoist the instructions.
2080  BB->getInstList().splice(BI->getIterator(), ThenBB->getInstList(),
2081  ThenBB->begin(), std::prev(ThenBB->end()));
2082 
2083  // Insert selects and rewrite the PHI operands.
2084  IRBuilder<NoFolder> Builder(BI);
2085  for (PHINode &PN : EndBB->phis()) {
2086  unsigned OrigI = PN.getBasicBlockIndex(BB);
2087  unsigned ThenI = PN.getBasicBlockIndex(ThenBB);
2088  Value *OrigV = PN.getIncomingValue(OrigI);
2089  Value *ThenV = PN.getIncomingValue(ThenI);
2090 
2091  // Skip PHIs which are trivial.
2092  if (OrigV == ThenV)
2093  continue;
2094 
2095  // Create a select whose true value is the speculatively executed value and
2096  // false value is the preexisting value. Swap them if the branch
2097  // destinations were inverted.
2098  Value *TrueV = ThenV, *FalseV = OrigV;
2099  if (Invert)
2100  std::swap(TrueV, FalseV);
2101  Value *V = Builder.CreateSelect(
2102  BrCond, TrueV, FalseV, "spec.select", BI);
2103  PN.setIncomingValue(OrigI, V);
2104  PN.setIncomingValue(ThenI, V);
2105  }
2106 
2107  // Remove speculated dbg intrinsics.
2108  // FIXME: Is it possible to do this in a more elegant way? Moving/merging the
2109  // dbg value for the different flows and inserting it after the select.
2110  for (Instruction *I : SpeculatedDbgIntrinsics)
2111  I->eraseFromParent();
2112 
2113  ++NumSpeculations;
2114  return true;
2115 }
2116 
2117 /// Return true if we can thread a branch across this block.
2119  unsigned Size = 0;
2120 
2121  for (Instruction &I : BB->instructionsWithoutDebug()) {
2122  if (Size > 10)
2123  return false; // Don't clone large BB's.
2124  ++Size;
2125 
2126  // We can only support instructions that do not define values that are
2127  // live outside of the current basic block.
2128  for (User *U : I.users()) {
2129  Instruction *UI = cast<Instruction>(U);
2130  if (UI->getParent() != BB || isa<PHINode>(UI))
2131  return false;
2132  }
2133 
2134  // Looks ok, continue checking.
2135  }
2136 
2137  return true;
2138 }
2139 
2140 /// If we have a conditional branch on a PHI node value that is defined in the
2141 /// same block as the branch and if any PHI entries are constants, thread edges
2142 /// corresponding to that entry to be branches to their ultimate destination.
2143 static bool FoldCondBranchOnPHI(BranchInst *BI, const DataLayout &DL,
2144  AssumptionCache *AC) {
2145  BasicBlock *BB = BI->getParent();
2146  PHINode *PN = dyn_cast<PHINode>(BI->getCondition());
2147  // NOTE: we currently cannot transform this case if the PHI node is used
2148  // outside of the block.
2149  if (!PN || PN->getParent() != BB || !PN->hasOneUse())
2150  return false;
2151 
2152  // Degenerate case of a single entry PHI.
2153  if (PN->getNumIncomingValues() == 1) {
2155  return true;
2156  }
2157 
2158  // Now we know that this block has multiple preds and two succs.
2160  return false;
2161 
2162  // Can't fold blocks that contain noduplicate or convergent calls.
2163  if (any_of(*BB, [](const Instruction &I) {
2164  const CallInst *CI = dyn_cast<CallInst>(&I);
2165  return CI && (CI->cannotDuplicate() || CI->isConvergent());
2166  }))
2167  return false;
2168 
2169  // Okay, this is a simple enough basic block. See if any phi values are
2170  // constants.
2171  for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
2173  if (!CB || !CB->getType()->isIntegerTy(1))
2174  continue;
2175 
2176  // Okay, we now know that all edges from PredBB should be revectored to
2177  // branch to RealDest.
2178  BasicBlock *PredBB = PN->getIncomingBlock(i);
2179  BasicBlock *RealDest = BI->getSuccessor(!CB->getZExtValue());
2180 
2181  if (RealDest == BB)
2182  continue; // Skip self loops.
2183  // Skip if the predecessor's terminator is an indirect branch.
2184  if (isa<IndirectBrInst>(PredBB->getTerminator()))
2185  continue;
2186 
2187  // The dest block might have PHI nodes, other predecessors and other
2188  // difficult cases. Instead of being smart about this, just insert a new
2189  // block that jumps to the destination block, effectively splitting
2190  // the edge we are about to create.
2191  BasicBlock *EdgeBB =
2192  BasicBlock::Create(BB->getContext(), RealDest->getName() + ".critedge",
2193  RealDest->getParent(), RealDest);
2194  BranchInst::Create(RealDest, EdgeBB);
2195 
2196  // Update PHI nodes.
2197  AddPredecessorToBlock(RealDest, EdgeBB, BB);
2198 
2199  // BB may have instructions that are being threaded over. Clone these
2200  // instructions into EdgeBB. We know that there will be no uses of the
2201  // cloned instructions outside of EdgeBB.
2202  BasicBlock::iterator InsertPt = EdgeBB->begin();
2203  DenseMap<Value *, Value *> TranslateMap; // Track translated values.
2204  for (BasicBlock::iterator BBI = BB->begin(); &*BBI != BI; ++BBI) {
2205  if (PHINode *PN = dyn_cast<PHINode>(BBI)) {
2206  TranslateMap[PN] = PN->getIncomingValueForBlock(PredBB);
2207  continue;
2208  }
2209  // Clone the instruction.
2210  Instruction *N = BBI->clone();
2211  if (BBI->hasName())
2212  N->setName(BBI->getName() + ".c");
2213 
2214  // Update operands due to translation.
2215  for (User::op_iterator i = N->op_begin(), e = N->op_end(); i != e; ++i) {
2216  DenseMap<Value *, Value *>::iterator PI = TranslateMap.find(*i);
2217  if (PI != TranslateMap.end())
2218  *i = PI->second;
2219  }
2220 
2221  // Check for trivial simplification.
2222  if (Value *V = SimplifyInstruction(N, {DL, nullptr, nullptr, AC})) {
2223  if (!BBI->use_empty())
2224  TranslateMap[&*BBI] = V;
2225  if (!N->mayHaveSideEffects()) {
2226  N->deleteValue(); // Instruction folded away, don't need actual inst
2227  N = nullptr;
2228  }
2229  } else {
2230  if (!BBI->use_empty())
2231  TranslateMap[&*BBI] = N;
2232  }
2233  // Insert the new instruction into its new home.
2234  if (N)
2235  EdgeBB->getInstList().insert(InsertPt, N);
2236 
2237  // Register the new instruction with the assumption cache if necessary.
2238  if (auto *II = dyn_cast_or_null<IntrinsicInst>(N))
2239  if (II->getIntrinsicID() == Intrinsic::assume)
2240  AC->registerAssumption(II);
2241  }
2242 
2243  // Loop over all of the edges from PredBB to BB, changing them to branch
2244  // to EdgeBB instead.
2245  TerminatorInst *PredBBTI = PredBB->getTerminator();
2246  for (unsigned i = 0, e = PredBBTI->getNumSuccessors(); i != e; ++i)
2247  if (PredBBTI->getSuccessor(i) == BB) {
2248  BB->removePredecessor(PredBB);
2249  PredBBTI->setSuccessor(i, EdgeBB);
2250  }
2251 
2252  // Recurse, simplifying any other constants.
2253  return FoldCondBranchOnPHI(BI, DL, AC) | true;
2254  }
2255 
2256  return false;
2257 }
2258 
2259 /// Given a BB that starts with the specified two-entry PHI node,
2260 /// see if we can eliminate it.
2262  const DataLayout &DL) {
2263  // Ok, this is a two entry PHI node. Check to see if this is a simple "if
2264  // statement", which has a very simple dominance structure. Basically, we
2265  // are trying to find the condition that is being branched on, which
2266  // subsequently causes this merge to happen. We really want control
2267  // dependence information for this check, but simplifycfg can't keep it up
2268  // to date, and this catches most of the cases we care about anyway.
2269  BasicBlock *BB = PN->getParent();
2270  const Function *Fn = BB->getParent();
2271  if (Fn && Fn->hasFnAttribute(Attribute::OptForFuzzing))
2272  return false;
2273 
2274  BasicBlock *IfTrue, *IfFalse;
2275  Value *IfCond = GetIfCondition(BB, IfTrue, IfFalse);
2276  if (!IfCond ||
2277  // Don't bother if the branch will be constant folded trivially.
2278  isa<ConstantInt>(IfCond))
2279  return false;
2280 
2281  // Okay, we found that we can merge this two-entry phi node into a select.
2282  // Doing so would require us to fold *all* two entry phi nodes in this block.
2283  // At some point this becomes non-profitable (particularly if the target
2284  // doesn't support cmov's). Only do this transformation if there are two or
2285  // fewer PHI nodes in this block.
2286  unsigned NumPhis = 0;
2287  for (BasicBlock::iterator I = BB->begin(); isa<PHINode>(I); ++NumPhis, ++I)
2288  if (NumPhis > 2)
2289  return false;
2290 
2291  // Loop over the PHI's seeing if we can promote them all to select
2292  // instructions. While we are at it, keep track of the instructions
2293  // that need to be moved to the dominating block.
2294  SmallPtrSet<Instruction *, 4> AggressiveInsts;
2295  unsigned MaxCostVal0 = PHINodeFoldingThreshold,
2296  MaxCostVal1 = PHINodeFoldingThreshold;
2297  MaxCostVal0 *= TargetTransformInfo::TCC_Basic;
2298  MaxCostVal1 *= TargetTransformInfo::TCC_Basic;
2299 
2300  for (BasicBlock::iterator II = BB->begin(); isa<PHINode>(II);) {
2301  PHINode *PN = cast<PHINode>(II++);
2302  if (Value *V = SimplifyInstruction(PN, {DL, PN})) {
2303  PN->replaceAllUsesWith(V);
2304  PN->eraseFromParent();
2305  continue;
2306  }
2307 
2308  if (!DominatesMergePoint(PN->getIncomingValue(0), BB, &AggressiveInsts,
2309  MaxCostVal0, TTI) ||
2310  !DominatesMergePoint(PN->getIncomingValue(1), BB, &AggressiveInsts,
2311  MaxCostVal1, TTI))
2312  return false;
2313  }
2314 
2315  // If we folded the first phi, PN dangles at this point. Refresh it. If
2316  // we ran out of PHIs then we simplified them all.
2317  PN = dyn_cast<PHINode>(BB->begin());
2318  if (!PN)
2319  return true;
2320 
2321  // Don't fold i1 branches on PHIs which contain binary operators. These can
2322  // often be turned into switches and other things.
2323  if (PN->getType()->isIntegerTy(1) &&
2324  (isa<BinaryOperator>(PN->getIncomingValue(0)) ||
2325  isa<BinaryOperator>(PN->getIncomingValue(1)) ||
2326  isa<BinaryOperator>(IfCond)))
2327  return false;
2328 
2329  // If all PHI nodes are promotable, check to make sure that all instructions
2330  // in the predecessor blocks can be promoted as well. If not, we won't be able
2331  // to get rid of the control flow, so it's not worth promoting to select
2332  // instructions.
2333  BasicBlock *DomBlock = nullptr;
2334  BasicBlock *IfBlock1 = PN->getIncomingBlock(0);
2335  BasicBlock *IfBlock2 = PN->getIncomingBlock(1);
2336  if (cast<BranchInst>(IfBlock1->getTerminator())->isConditional()) {
2337  IfBlock1 = nullptr;
2338  } else {
2339  DomBlock = *pred_begin(IfBlock1);
2340  for (BasicBlock::iterator I = IfBlock1->begin(); !isa<TerminatorInst>(I);
2341  ++I)
2342  if (!AggressiveInsts.count(&*I) && !isa<DbgInfoIntrinsic>(I)) {
2343  // This is not an aggressive instruction that we can promote.
2344  // Because of this, we won't be able to get rid of the control flow, so
2345  // the xform is not worth it.
2346  return false;
2347  }
2348  }
2349 
2350  if (cast<BranchInst>(IfBlock2->getTerminator())->isConditional()) {
2351  IfBlock2 = nullptr;
2352  } else {
2353  DomBlock = *pred_begin(IfBlock2);
2354  for (BasicBlock::iterator I = IfBlock2->begin(); !isa<TerminatorInst>(I);
2355  ++I)
2356  if (!AggressiveInsts.count(&*I) && !isa<DbgInfoIntrinsic>(I)) {
2357  // This is not an aggressive instruction that we can promote.
2358  // Because of this, we won't be able to get rid of the control flow, so
2359  // the xform is not worth it.
2360  return false;
2361  }
2362  }
2363 
2364  LLVM_DEBUG(dbgs() << "FOUND IF CONDITION! " << *IfCond
2365  << " T: " << IfTrue->getName()
2366  << " F: " << IfFalse->getName() << "\n");
2367 
2368  // If we can still promote the PHI nodes after this gauntlet of tests,
2369  // do all of the PHI's now.
2370  Instruction *InsertPt = DomBlock->getTerminator();
2371  IRBuilder<NoFolder> Builder(InsertPt);
2372 
2373  // Move all 'aggressive' instructions, which are defined in the
2374  // conditional parts of the if's up to the dominating block.
2375  if (IfBlock1) {
2376  for (auto &I : *IfBlock1)
2378  DomBlock->getInstList().splice(InsertPt->getIterator(),
2379  IfBlock1->getInstList(), IfBlock1->begin(),
2380  IfBlock1->getTerminator()->getIterator());
2381  }
2382  if (IfBlock2) {
2383  for (auto &I : *IfBlock2)
2385  DomBlock->getInstList().splice(InsertPt->getIterator(),
2386  IfBlock2->getInstList(), IfBlock2->begin(),
2387  IfBlock2->getTerminator()->getIterator());
2388  }
2389 
2390  while (PHINode *PN = dyn_cast<PHINode>(BB->begin())) {
2391  // Change the PHI node into a select instruction.
2392  Value *TrueVal = PN->getIncomingValue(PN->getIncomingBlock(0) == IfFalse);
2393  Value *FalseVal = PN->getIncomingValue(PN->getIncomingBlock(0) == IfTrue);
2394 
2395  Value *Sel = Builder.CreateSelect(IfCond, TrueVal, FalseVal, "", InsertPt);
2396  PN->replaceAllUsesWith(Sel);
2397  Sel->takeName(PN);
2398  PN->eraseFromParent();
2399  }
2400 
2401  // At this point, IfBlock1 and IfBlock2 are both empty, so our if statement
2402  // has been flattened. Change DomBlock to jump directly to our new block to
2403  // avoid other simplifycfg's kicking in on the diamond.
2404  TerminatorInst *OldTI = DomBlock->getTerminator();
2405  Builder.SetInsertPoint(OldTI);
2406  Builder.CreateBr(BB);
2407  OldTI->eraseFromParent();
2408  return true;
2409 }
2410 
2411 /// If we found a conditional branch that goes to two returning blocks,
2412 /// try to merge them together into one return,
2413 /// introducing a select if the return values disagree.
2415  IRBuilder<> &Builder) {
2416  assert(BI->isConditional() && "Must be a conditional branch");
2417  BasicBlock *TrueSucc = BI->getSuccessor(0);
2418  BasicBlock *FalseSucc = BI->getSuccessor(1);
2419  ReturnInst *TrueRet = cast<ReturnInst>(TrueSucc->getTerminator());
2420  ReturnInst *FalseRet = cast<ReturnInst>(FalseSucc->getTerminator());
2421 
2422  // Check to ensure both blocks are empty (just a return) or optionally empty
2423  // with PHI nodes. If there are other instructions, merging would cause extra
2424  // computation on one path or the other.
2425  if (!TrueSucc->getFirstNonPHIOrDbg()->isTerminator())
2426  return false;
2427  if (!FalseSucc->getFirstNonPHIOrDbg()->isTerminator())
2428  return false;
2429 
2430  Builder.SetInsertPoint(BI);
2431  // Okay, we found a branch that is going to two return nodes. If
2432  // there is no return value for this function, just change the
2433  // branch into a return.
2434  if (FalseRet->getNumOperands() == 0) {
2435  TrueSucc->removePredecessor(BI->getParent());
2436  FalseSucc->removePredecessor(BI->getParent());
2437  Builder.CreateRetVoid();
2439  return true;
2440  }
2441 
2442  // Otherwise, figure out what the true and false return values are
2443  // so we can insert a new select instruction.
2444  Value *TrueValue = TrueRet->getReturnValue();
2445  Value *FalseValue = FalseRet->getReturnValue();
2446 
2447  // Unwrap any PHI nodes in the return blocks.
2448  if (PHINode *TVPN = dyn_cast_or_null<PHINode>(TrueValue))
2449  if (TVPN->getParent() == TrueSucc)
2450  TrueValue = TVPN->getIncomingValueForBlock(BI->getParent());
2451  if (PHINode *FVPN = dyn_cast_or_null<PHINode>(FalseValue))
2452  if (FVPN->getParent() == FalseSucc)
2453  FalseValue = FVPN->getIncomingValueForBlock(BI->getParent());
2454 
2455  // In order for this transformation to be safe, we must be able to
2456  // unconditionally execute both operands to the return. This is
2457  // normally the case, but we could have a potentially-trapping
2458  // constant expression that prevents this transformation from being
2459  // safe.
2460  if (ConstantExpr *TCV = dyn_cast_or_null<ConstantExpr>(TrueValue))
2461  if (TCV->canTrap())
2462  return false;
2463  if (ConstantExpr *FCV = dyn_cast_or_null<ConstantExpr>(FalseValue))
2464  if (FCV->canTrap())
2465  return false;
2466 
2467  // Okay, we collected all the mapped values and checked them for sanity, and
2468  // defined to really do this transformation. First, update the CFG.
2469  TrueSucc->removePredecessor(BI->getParent());
2470  FalseSucc->removePredecessor(BI->getParent());
2471 
2472  // Insert select instructions where needed.
2473  Value *BrCond = BI->getCondition();
2474  if (TrueValue) {
2475  // Insert a select if the results differ.
2476  if (TrueValue == FalseValue || isa<UndefValue>(FalseValue)) {
2477  } else if (isa<UndefValue>(TrueValue)) {
2478  TrueValue = FalseValue;
2479  } else {
2480  TrueValue =
2481  Builder.CreateSelect(BrCond, TrueValue, FalseValue, "retval", BI);
2482  }
2483  }
2484 
2485  Value *RI =
2486  !TrueValue ? Builder.CreateRetVoid() : Builder.CreateRet(TrueValue);
2487 
2488  (void)RI;
2489 
2490  LLVM_DEBUG(dbgs() << "\nCHANGING BRANCH TO TWO RETURNS INTO SELECT:"
2491  << "\n " << *BI << "NewRet = " << *RI << "TRUEBLOCK: "
2492  << *TrueSucc << "FALSEBLOCK: " << *FalseSucc);
2493 
2495 
2496  return true;
2497 }
2498 
2499 /// Return true if the given instruction is available
2500 /// in its predecessor block. If yes, the instruction will be removed.
2502  if (!isa<BinaryOperator>(Inst) && !isa<CmpInst>(Inst))
2503  return false;
2504  for (Instruction &I : *PB) {
2505  Instruction *PBI = &I;
2506  // Check whether Inst and PBI generate the same value.
2507  if (Inst->isIdenticalTo(PBI)) {
2508  Inst->replaceAllUsesWith(PBI);
2509  Inst->eraseFromParent();
2510  return true;
2511  }
2512  }
2513  return false;
2514 }
2515 
2516 /// Return true if either PBI or BI has branch weight available, and store
2517 /// the weights in {Pred|Succ}{True|False}Weight. If one of PBI and BI does
2518 /// not have branch weight, use 1:1 as its weight.
2520  uint64_t &PredTrueWeight,
2521  uint64_t &PredFalseWeight,
2522  uint64_t &SuccTrueWeight,
2523  uint64_t &SuccFalseWeight) {
2524  bool PredHasWeights =
2525  PBI->extractProfMetadata(PredTrueWeight, PredFalseWeight);
2526  bool SuccHasWeights =
2527  BI->extractProfMetadata(SuccTrueWeight, SuccFalseWeight);
2528  if (PredHasWeights || SuccHasWeights) {
2529  if (!PredHasWeights)
2530  PredTrueWeight = PredFalseWeight = 1;
2531  if (!SuccHasWeights)
2532  SuccTrueWeight = SuccFalseWeight = 1;
2533  return true;
2534  } else {
2535  return false;
2536  }
2537 }
2538 
2539 /// If this basic block is simple enough, and if a predecessor branches to us
2540 /// and one of our successors, fold the block into the predecessor and use
2541 /// logical operations to pick the right destination.
2542 bool llvm::FoldBranchToCommonDest(BranchInst *BI, unsigned BonusInstThreshold) {
2543  BasicBlock *BB = BI->getParent();
2544 
2545  Instruction *Cond = nullptr;
2546  if (BI->isConditional())
2547  Cond = dyn_cast<Instruction>(BI->getCondition());
2548  else {
2549  // For unconditional branch, check for a simple CFG pattern, where
2550  // BB has a single predecessor and BB's successor is also its predecessor's
2551  // successor. If such pattern exists, check for CSE between BB and its
2552  // predecessor.
2553  if (BasicBlock *PB = BB->getSinglePredecessor())
2554  if (BranchInst *PBI = dyn_cast<BranchInst>(PB->getTerminator()))
2555  if (PBI->isConditional() &&
2556  (BI->getSuccessor(0) == PBI->getSuccessor(0) ||
2557  BI->getSuccessor(0) == PBI->getSuccessor(1))) {
2558  for (auto I = BB->instructionsWithoutDebug().begin(),
2559  E = BB->instructionsWithoutDebug().end();
2560  I != E;) {
2561  Instruction *Curr = &*I++;
2562  if (isa<CmpInst>(Curr)) {
2563  Cond = Curr;
2564  break;
2565  }
2566  // Quit if we can't remove this instruction.
2567  if (!tryCSEWithPredecessor(Curr, PB))
2568  return false;
2569  }
2570  }
2571 
2572  if (!Cond)
2573  return false;
2574  }
2575 
2576  if (!Cond || (!isa<CmpInst>(Cond) && !isa<BinaryOperator>(Cond)) ||
2577  Cond->getParent() != BB || !Cond->hasOneUse())
2578  return false;
2579 
2580  // Make sure the instruction after the condition is the cond branch.
2581  BasicBlock::iterator CondIt = ++Cond->getIterator();
2582 
2583  // Ignore dbg intrinsics.
2584  while (isa<DbgInfoIntrinsic>(CondIt))
2585  ++CondIt;
2586 
2587  if (&*CondIt != BI)
2588  return false;
2589 
2590  // Only allow this transformation if computing the condition doesn't involve
2591  // too many instructions and these involved instructions can be executed
2592  // unconditionally. We denote all involved instructions except the condition
2593  // as "bonus instructions", and only allow this transformation when the
2594  // number of the bonus instructions does not exceed a certain threshold.
2595  unsigned NumBonusInsts = 0;
2596  for (auto I = BB->begin(); Cond != &*I; ++I) {
2597  // Ignore dbg intrinsics.
2598  if (isa<DbgInfoIntrinsic>(I))
2599  continue;
2600  if (!I->hasOneUse() || !isSafeToSpeculativelyExecute(&*I))
2601  return false;
2602  // I has only one use and can be executed unconditionally.
2604  if (User == nullptr || User->getParent() != BB)
2605  return false;
2606  // I is used in the same BB. Since BI uses Cond and doesn't have more slots
2607  // to use any other instruction, User must be an instruction between next(I)
2608  // and Cond.
2609  ++NumBonusInsts;
2610  // Early exits once we reach the limit.
2611  if (NumBonusInsts > BonusInstThreshold)
2612  return false;
2613  }
2614 
2615  // Cond is known to be a compare or binary operator. Check to make sure that
2616  // neither operand is a potentially-trapping constant expression.
2617  if (ConstantExpr *CE = dyn_cast<ConstantExpr>(Cond->getOperand(0)))
2618  if (CE->canTrap())
2619  return false;
2620  if (ConstantExpr *CE = dyn_cast<ConstantExpr>(Cond->getOperand(1)))
2621  if (CE->canTrap())
2622  return false;
2623 
2624  // Finally, don't infinitely unroll conditional loops.
2625  BasicBlock *TrueDest = BI->getSuccessor(0);
2626  BasicBlock *FalseDest = (BI->isConditional()) ? BI->getSuccessor(1) : nullptr;
2627  if (TrueDest == BB || FalseDest == BB)
2628  return false;
2629 
2630  for (pred_iterator PI = pred_begin(BB), E = pred_end(BB); PI != E; ++PI) {
2631  BasicBlock *PredBlock = *PI;
2632  BranchInst *PBI = dyn_cast<BranchInst>(PredBlock->getTerminator());
2633 
2634  // Check that we have two conditional branches. If there is a PHI node in
2635  // the common successor, verify that the same value flows in from both
2636  // blocks.
2638  if (!PBI || PBI->isUnconditional() ||
2639  (BI->isConditional() && !SafeToMergeTerminators(BI, PBI)) ||
2640  (!BI->isConditional() &&
2641  !isProfitableToFoldUnconditional(BI, PBI, Cond, PHIs)))
2642  continue;
2643 
2644  // Determine if the two branches share a common destination.
2645  Instruction::BinaryOps Opc = Instruction::BinaryOpsEnd;
2646  bool InvertPredCond = false;
2647 
2648  if (BI->isConditional()) {
2649  if (PBI->getSuccessor(0) == TrueDest) {
2650  Opc = Instruction::Or;
2651  } else if (PBI->getSuccessor(1) == FalseDest) {
2652  Opc = Instruction::And;
2653  } else if (PBI->getSuccessor(0) == FalseDest) {
2654  Opc = Instruction::And;
2655  InvertPredCond = true;
2656  } else if (PBI->getSuccessor(1) == TrueDest) {
2657  Opc = Instruction::Or;
2658  InvertPredCond = true;
2659  } else {
2660  continue;
2661  }
2662  } else {
2663  if (PBI->getSuccessor(0) != TrueDest && PBI->getSuccessor(1) != TrueDest)
2664  continue;
2665  }
2666 
2667  LLVM_DEBUG(dbgs() << "FOLDING BRANCH TO COMMON DEST:\n" << *PBI << *BB);
2668  IRBuilder<> Builder(PBI);
2669 
2670  // If we need to invert the condition in the pred block to match, do so now.
2671  if (InvertPredCond) {
2672  Value *NewCond = PBI->getCondition();
2673 
2674  if (NewCond->hasOneUse() && isa<CmpInst>(NewCond)) {
2675  CmpInst *CI = cast<CmpInst>(NewCond);
2676  CI->setPredicate(CI->getInversePredicate());
2677  } else {
2678  NewCond =
2679  Builder.CreateNot(NewCond, PBI->getCondition()->getName() + ".not");
2680  }
2681 
2682  PBI->setCondition(NewCond);
2683  PBI->swapSuccessors();
2684  }
2685 
2686  // If we have bonus instructions, clone them into the predecessor block.
2687  // Note that there may be multiple predecessor blocks, so we cannot move
2688  // bonus instructions to a predecessor block.
2689  ValueToValueMapTy VMap; // maps original values to cloned values
2690  // We already make sure Cond is the last instruction before BI. Therefore,
2691  // all instructions before Cond other than DbgInfoIntrinsic are bonus
2692  // instructions.
2693  for (auto BonusInst = BB->begin(); Cond != &*BonusInst; ++BonusInst) {
2694  if (isa<DbgInfoIntrinsic>(BonusInst))
2695  continue;
2696  Instruction *NewBonusInst = BonusInst->clone();
2697  RemapInstruction(NewBonusInst, VMap,
2699  VMap[&*BonusInst] = NewBonusInst;
2700 
2701  // If we moved a load, we cannot any longer claim any knowledge about
2702  // its potential value. The previous information might have been valid
2703  // only given the branch precondition.
2704  // For an analogous reason, we must also drop all the metadata whose
2705  // semantics we don't understand.
2706  NewBonusInst->dropUnknownNonDebugMetadata();
2707 
2708  PredBlock->getInstList().insert(PBI->getIterator(), NewBonusInst);
2709  NewBonusInst->takeName(&*BonusInst);
2710  BonusInst->setName(BonusInst->getName() + ".old");
2711  }
2712 
2713  // Clone Cond into the predecessor basic block, and or/and the
2714  // two conditions together.
2715  Instruction *New = Cond->clone();
2716  RemapInstruction(New, VMap,
2718  PredBlock->getInstList().insert(PBI->getIterator(), New);
2719  New->takeName(Cond);
2720  Cond->setName(New->getName() + ".old");
2721 
2722  if (BI->isConditional()) {
2723  Instruction *NewCond = cast<Instruction>(
2724  Builder.CreateBinOp(Opc, PBI->getCondition(), New, "or.cond"));
2725  PBI->setCondition(NewCond);
2726 
2727  uint64_t PredTrueWeight, PredFalseWeight, SuccTrueWeight, SuccFalseWeight;
2728  bool HasWeights =
2729  extractPredSuccWeights(PBI, BI, PredTrueWeight, PredFalseWeight,
2730  SuccTrueWeight, SuccFalseWeight);
2731  SmallVector<uint64_t, 8> NewWeights;
2732 
2733  if (PBI->getSuccessor(0) == BB) {
2734  if (HasWeights) {
2735  // PBI: br i1 %x, BB, FalseDest
2736  // BI: br i1 %y, TrueDest, FalseDest
2737  // TrueWeight is TrueWeight for PBI * TrueWeight for BI.
2738  NewWeights.push_back(PredTrueWeight * SuccTrueWeight);
2739  // FalseWeight is FalseWeight for PBI * TotalWeight for BI +
2740  // TrueWeight for PBI * FalseWeight for BI.
2741  // We assume that total weights of a BranchInst can fit into 32 bits.
2742  // Therefore, we will not have overflow using 64-bit arithmetic.
2743  NewWeights.push_back(PredFalseWeight *
2744  (SuccFalseWeight + SuccTrueWeight) +
2745  PredTrueWeight * SuccFalseWeight);
2746  }
2747  AddPredecessorToBlock(TrueDest, PredBlock, BB);
2748  PBI->setSuccessor(0, TrueDest);
2749  }
2750  if (PBI->getSuccessor(1) == BB) {
2751  if (HasWeights) {
2752  // PBI: br i1 %x, TrueDest, BB
2753  // BI: br i1 %y, TrueDest, FalseDest
2754  // TrueWeight is TrueWeight for PBI * TotalWeight for BI +
2755  // FalseWeight for PBI * TrueWeight for BI.
2756  NewWeights.push_back(PredTrueWeight *
2757  (SuccFalseWeight + SuccTrueWeight) +
2758  PredFalseWeight * SuccTrueWeight);
2759  // FalseWeight is FalseWeight for PBI * FalseWeight for BI.
2760  NewWeights.push_back(PredFalseWeight * SuccFalseWeight);
2761  }
2762  AddPredecessorToBlock(FalseDest, PredBlock, BB);
2763  PBI->setSuccessor(1, FalseDest);
2764  }
2765  if (NewWeights.size() == 2) {
2766  // Halve the weights if any of them cannot fit in an uint32_t
2767  FitWeights(NewWeights);
2768 
2769  SmallVector<uint32_t, 8> MDWeights(NewWeights.begin(),
2770  NewWeights.end());
2771  setBranchWeights(PBI, MDWeights[0], MDWeights[1]);
2772  } else
2773  PBI->setMetadata(LLVMContext::MD_prof, nullptr);
2774  } else {
2775  // Update PHI nodes in the common successors.
2776  for (unsigned i = 0, e = PHIs.size(); i != e; ++i) {
2777  ConstantInt *PBI_C = cast<ConstantInt>(
2778  PHIs[i]->getIncomingValueForBlock(PBI->getParent()));
2779  assert(PBI_C->getType()->isIntegerTy(1));
2780  Instruction *MergedCond = nullptr;
2781  if (PBI->getSuccessor(0) == TrueDest) {
2782  // Create (PBI_Cond and PBI_C) or (!PBI_Cond and BI_Value)
2783  // PBI_C is true: PBI_Cond or (!PBI_Cond and BI_Value)
2784  // is false: !PBI_Cond and BI_Value
2785  Instruction *NotCond = cast<Instruction>(
2786  Builder.CreateNot(PBI->getCondition(), "not.cond"));
2787  MergedCond = cast<Instruction>(
2788  Builder.CreateBinOp(Instruction::And, NotCond, New, "and.cond"));
2789  if (PBI_C->isOne())
2790  MergedCond = cast<Instruction>(Builder.CreateBinOp(
2791  Instruction::Or, PBI->getCondition(), MergedCond, "or.cond"));
2792  } else {
2793  // Create (PBI_Cond and BI_Value) or (!PBI_Cond and PBI_C)
2794  // PBI_C is true: (PBI_Cond and BI_Value) or (!PBI_Cond)
2795  // is false: PBI_Cond and BI_Value
2796  MergedCond = cast<Instruction>(Builder.CreateBinOp(
2797  Instruction::And, PBI->getCondition(), New, "and.cond"));
2798  if (PBI_C->isOne()) {
2799  Instruction *NotCond = cast<Instruction>(
2800  Builder.CreateNot(PBI->getCondition(), "not.cond"));
2801  MergedCond = cast<Instruction>(Builder.CreateBinOp(
2802  Instruction::Or, NotCond, MergedCond, "or.cond"));
2803  }
2804  }
2805  // Update PHI Node.
2806  PHIs[i]->setIncomingValue(PHIs[i]->getBasicBlockIndex(PBI->getParent()),
2807  MergedCond);
2808  }
2809  // Change PBI from Conditional to Unconditional.
2810  BranchInst *New_PBI = BranchInst::Create(TrueDest, PBI);
2812  PBI = New_PBI;
2813  }
2814 
2815  // If BI was a loop latch, it may have had associated loop metadata.
2816  // We need to copy it to the new latch, that is, PBI.
2817  if (MDNode *LoopMD = BI->getMetadata(LLVMContext::MD_loop))
2818  PBI->setMetadata(LLVMContext::MD_loop, LoopMD);
2819 
2820  // TODO: If BB is reachable from all paths through PredBlock, then we
2821  // could replace PBI's branch probabilities with BI's.
2822 
2823  // Copy any debug value intrinsics into the end of PredBlock.
2824  for (Instruction &I : *BB)
2825  if (isa<DbgInfoIntrinsic>(I))
2826  I.clone()->insertBefore(PBI);
2827 
2828  return true;
2829  }
2830  return false;
2831 }
2832 
2833 // If there is only one store in BB1 and BB2, return it, otherwise return
2834 // nullptr.
2836  StoreInst *S = nullptr;
2837  for (auto *BB : {BB1, BB2}) {
2838  if (!BB)
2839  continue;
2840  for (auto &I : *BB)
2841  if (auto *SI = dyn_cast<StoreInst>(&I)) {
2842  if (S)
2843  // Multiple stores seen.
2844  return nullptr;
2845  else
2846  S = SI;
2847  }
2848  }
2849  return S;
2850 }
2851 
2853  Value *AlternativeV = nullptr) {
2854  // PHI is going to be a PHI node that allows the value V that is defined in
2855  // BB to be referenced in BB's only successor.
2856  //
2857  // If AlternativeV is nullptr, the only value we care about in PHI is V. It
2858  // doesn't matter to us what the other operand is (it'll never get used). We
2859  // could just create a new PHI with an undef incoming value, but that could
2860  // increase register pressure if EarlyCSE/InstCombine can't fold it with some
2861  // other PHI. So here we directly look for some PHI in BB's successor with V
2862  // as an incoming operand. If we find one, we use it, else we create a new
2863  // one.
2864  //
2865  // If AlternativeV is not nullptr, we care about both incoming values in PHI.
2866  // PHI must be exactly: phi <ty> [ %BB, %V ], [ %OtherBB, %AlternativeV]
2867  // where OtherBB is the single other predecessor of BB's only successor.
2868  PHINode *PHI = nullptr;
2869  BasicBlock *Succ = BB->getSingleSuccessor();
2870 
2871  for (auto I = Succ->begin(); isa<PHINode>(I); ++I)
2872  if (cast<PHINode>(I)->getIncomingValueForBlock(BB) == V) {
2873  PHI = cast<PHINode>(I);
2874  if (!AlternativeV)
2875  break;
2876 
2877  assert(pred_size(Succ) == 2);
2878  auto PredI = pred_begin(Succ);
2879  BasicBlock *OtherPredBB = *PredI == BB ? *++PredI : *PredI;
2880  if (PHI->getIncomingValueForBlock(OtherPredBB) == AlternativeV)
2881  break;
2882  PHI = nullptr;
2883  }
2884  if (PHI)
2885  return PHI;
2886 
2887  // If V is not an instruction defined in BB, just return it.
2888  if (!AlternativeV &&
2889  (!isa<Instruction>(V) || cast<Instruction>(V)->getParent() != BB))
2890  return V;
2891 
2892  PHI = PHINode::Create(V->getType(), 2, "simplifycfg.merge", &Succ->front());
2893  PHI->addIncoming(V, BB);
2894  for (BasicBlock *PredBB : predecessors(Succ))
2895  if (PredBB != BB)
2896  PHI->addIncoming(
2897  AlternativeV ? AlternativeV : UndefValue::get(V->getType()), PredBB);
2898  return PHI;
2899 }
2900 
2902  BasicBlock *QTB, BasicBlock *QFB,
2903  BasicBlock *PostBB, Value *Address,
2904  bool InvertPCond, bool InvertQCond,
2905  const DataLayout &DL) {
2906  auto IsaBitcastOfPointerType = [](const Instruction &I) {
2907  return Operator::getOpcode(&I) == Instruction::BitCast &&
2908  I.getType()->isPointerTy();
2909  };
2910 
2911  // If we're not in aggressive mode, we only optimize if we have some
2912  // confidence that by optimizing we'll allow P and/or Q to be if-converted.
2913  auto IsWorthwhile = [&](BasicBlock *BB) {
2914  if (!BB)
2915  return true;
2916  // Heuristic: if the block can be if-converted/phi-folded and the
2917  // instructions inside are all cheap (arithmetic/GEPs), it's worthwhile to
2918  // thread this store.
2919  unsigned N = 0;
2920  for (auto &I : BB->instructionsWithoutDebug()) {
2921  // Cheap instructions viable for folding.
2922  if (isa<BinaryOperator>(I) || isa<GetElementPtrInst>(I) ||
2923  isa<StoreInst>(I))
2924  ++N;
2925  // Free instructions.
2926  else if (isa<TerminatorInst>(I) || IsaBitcastOfPointerType(I))
2927  continue;
2928  else
2929  return false;
2930  }
2931  // The store we want to merge is counted in N, so add 1 to make sure
2932  // we're counting the instructions that would be left.
2933  return N <= (PHINodeFoldingThreshold + 1);
2934  };
2935 
2937  (!IsWorthwhile(PTB) || !IsWorthwhile(PFB) || !IsWorthwhile(QTB) ||
2938  !IsWorthwhile(QFB)))
2939  return false;
2940 
2941  // For every pointer, there must be exactly two stores, one coming from
2942  // PTB or PFB, and the other from QTB or QFB. We don't support more than one
2943  // store (to any address) in PTB,PFB or QTB,QFB.
2944  // FIXME: We could relax this restriction with a bit more work and performance
2945  // testing.
2946  StoreInst *PStore = findUniqueStoreInBlocks(PTB, PFB);
2947  StoreInst *QStore = findUniqueStoreInBlocks(QTB, QFB);
2948  if (!PStore || !QStore)
2949  return false;
2950 
2951  // Now check the stores are compatible.
2952  if (!QStore->isUnordered() || !PStore->isUnordered())
2953  return false;
2954 
2955  // Check that sinking the store won't cause program behavior changes. Sinking
2956  // the store out of the Q blocks won't change any behavior as we're sinking
2957  // from a block to its unconditional successor. But we're moving a store from
2958  // the P blocks down through the middle block (QBI) and past both QFB and QTB.
2959  // So we need to check that there are no aliasing loads or stores in
2960  // QBI, QTB and QFB. We also need to check there are no conflicting memory
2961  // operations between PStore and the end of its parent block.
2962  //
2963  // The ideal way to do this is to query AliasAnalysis, but we don't
2964  // preserve AA currently so that is dangerous. Be super safe and just
2965  // check there are no other memory operations at all.
2966  for (auto &I : *QFB->getSinglePredecessor())
2967  if (I.mayReadOrWriteMemory())
2968  return false;
2969  for (auto &I : *QFB)
2970  if (&I != QStore && I.mayReadOrWriteMemory())
2971  return false;
2972  if (QTB)
2973  for (auto &I : *QTB)
2974  if (&I != QStore && I.mayReadOrWriteMemory())
2975  return false;
2976  for (auto I = BasicBlock::iterator(PStore), E = PStore->getParent()->end();
2977  I != E; ++I)
2978  if (&*I != PStore && I->mayReadOrWriteMemory())
2979  return false;
2980 
2981  // If PostBB has more than two predecessors, we need to split it so we can
2982  // sink the store.
2983  if (std::next(pred_begin(PostBB), 2) != pred_end(PostBB)) {
2984  // We know that QFB's only successor is PostBB. And QFB has a single
2985  // predecessor. If QTB exists, then its only successor is also PostBB.
2986  // If QTB does not exist, then QFB's only predecessor has a conditional
2987  // branch to QFB and PostBB.
2988  BasicBlock *TruePred = QTB ? QTB : QFB->getSinglePredecessor();
2989  BasicBlock *NewBB = SplitBlockPredecessors(PostBB, { QFB, TruePred},
2990  "condstore.split");
2991  if (!NewBB)
2992  return false;
2993  PostBB = NewBB;
2994  }
2995 
2996  // OK, we're going to sink the stores to PostBB. The store has to be
2997  // conditional though, so first create the predicate.
2998  Value *PCond = cast<BranchInst>(PFB->getSinglePredecessor()->getTerminator())
2999  ->getCondition();
3000  Value *QCond = cast<BranchInst>(QFB->getSinglePredecessor()->getTerminator())
3001  ->getCondition();
3002 
3004  PStore->getParent());
3006  QStore->getParent(), PPHI);
3007 
3008  IRBuilder<> QB(&*PostBB->getFirstInsertionPt());
3009 
3010  Value *PPred = PStore->getParent() == PTB ? PCond : QB.CreateNot(PCond);
3011  Value *QPred = QStore->getParent() == QTB ? QCond : QB.CreateNot(QCond);
3012 
3013  if (InvertPCond)
3014  PPred = QB.CreateNot(PPred);
3015  if (InvertQCond)
3016  QPred = QB.CreateNot(QPred);
3017  Value *CombinedPred = QB.CreateOr(PPred, QPred);
3018 
3019  auto *T =
3020  SplitBlockAndInsertIfThen(CombinedPred, &*QB.GetInsertPoint(), false);
3021  QB.SetInsertPoint(T);
3022  StoreInst *SI = cast<StoreInst>(QB.CreateStore(QPHI, Address));
3023  AAMDNodes AAMD;
3024  PStore->getAAMetadata(AAMD, /*Merge=*/false);
3025  PStore->getAAMetadata(AAMD, /*Merge=*/true);
3026  SI->setAAMetadata(AAMD);
3027  unsigned PAlignment = PStore->getAlignment();
3028  unsigned QAlignment = QStore->getAlignment();
3029  unsigned TypeAlignment =
3031  unsigned MinAlignment;
3032  unsigned MaxAlignment;
3033  std::tie(MinAlignment, MaxAlignment) = std::minmax(PAlignment, QAlignment);
3034  // Choose the minimum alignment. If we could prove both stores execute, we
3035  // could use biggest one. In this case, though, we only know that one of the
3036  // stores executes. And we don't know it's safe to take the alignment from a
3037  // store that doesn't execute.
3038  if (MinAlignment != 0) {
3039  // Choose the minimum of all non-zero alignments.
3040  SI->setAlignment(MinAlignment);
3041  } else if (MaxAlignment != 0) {
3042  // Choose the minimal alignment between the non-zero alignment and the ABI
3043  // default alignment for the type of the stored value.
3044  SI->setAlignment(std::min(MaxAlignment, TypeAlignment));
3045  } else {
3046  // If both alignments are zero, use ABI default alignment for the type of
3047  // the stored value.
3048  SI->setAlignment(TypeAlignment);
3049  }
3050 
3051  QStore->eraseFromParent();
3052  PStore->eraseFromParent();
3053 
3054  return true;
3055 }
3056 
3058  const DataLayout &DL) {
3059  // The intention here is to find diamonds or triangles (see below) where each
3060  // conditional block contains a store to the same address. Both of these
3061  // stores are conditional, so they can't be unconditionally sunk. But it may
3062  // be profitable to speculatively sink the stores into one merged store at the
3063  // end, and predicate the merged store on the union of the two conditions of
3064  // PBI and QBI.
3065  //
3066  // This can reduce the number of stores executed if both of the conditions are
3067  // true, and can allow the blocks to become small enough to be if-converted.
3068  // This optimization will also chain, so that ladders of test-and-set
3069  // sequences can be if-converted away.
3070  //
3071  // We only deal with simple diamonds or triangles:
3072  //
3073  // PBI or PBI or a combination of the two
3074  // / \ | \
3075  // PTB PFB | PFB
3076  // \ / | /
3077  // QBI QBI
3078  // / \ | \
3079  // QTB QFB | QFB
3080  // \ / | /
3081  // PostBB PostBB
3082  //
3083  // We model triangles as a type of diamond with a nullptr "true" block.
3084  // Triangles are canonicalized so that the fallthrough edge is represented by
3085  // a true condition, as in the diagram above.
3086  BasicBlock *PTB = PBI->getSuccessor(0);
3087  BasicBlock *PFB = PBI->getSuccessor(1);
3088  BasicBlock *QTB = QBI->getSuccessor(0);
3089  BasicBlock *QFB = QBI->getSuccessor(1);
3090  BasicBlock *PostBB = QFB->getSingleSuccessor();
3091 
3092  // Make sure we have a good guess for PostBB. If QTB's only successor is
3093  // QFB, then QFB is a better PostBB.
3094  if (QTB->getSingleSuccessor() == QFB)
3095  PostBB = QFB;
3096 
3097  // If we couldn't find a good PostBB, stop.
3098  if (!PostBB)
3099  return false;
3100 
3101  bool InvertPCond = false, InvertQCond = false;
3102  // Canonicalize fallthroughs to the true branches.
3103  if (PFB == QBI->getParent()) {
3104  std::swap(PFB, PTB);
3105  InvertPCond = true;
3106  }
3107  if (QFB == PostBB) {
3108  std::swap(QFB, QTB);
3109  InvertQCond = true;
3110  }
3111 
3112  // From this point on we can assume PTB or QTB may be fallthroughs but PFB
3113  // and QFB may not. Model fallthroughs as a nullptr block.
3114  if (PTB == QBI->getParent())
3115  PTB = nullptr;
3116  if (QTB == PostBB)
3117  QTB = nullptr;
3118 
3119  // Legality bailouts. We must have at least the non-fallthrough blocks and
3120  // the post-dominating block, and the non-fallthroughs must only have one
3121  // predecessor.
3122  auto HasOnePredAndOneSucc = [](BasicBlock *BB, BasicBlock *P, BasicBlock *S) {
3123  return BB->getSinglePredecessor() == P && BB->getSingleSuccessor() == S;
3124  };
3125  if (!HasOnePredAndOneSucc(PFB, PBI->getParent(), QBI->getParent()) ||
3126  !HasOnePredAndOneSucc(QFB, QBI->getParent(), PostBB))
3127  return false;
3128  if ((PTB && !HasOnePredAndOneSucc(PTB, PBI->getParent(), QBI->getParent())) ||
3129  (QTB && !HasOnePredAndOneSucc(QTB, QBI->getParent(), PostBB)))
3130  return false;
3131  if (!QBI->getParent()->hasNUses(2))
3132  return false;
3133 
3134  // OK, this is a sequence of two diamonds or triangles.
3135  // Check if there are stores in PTB or PFB that are repeated in QTB or QFB.
3136  SmallPtrSet<Value *, 4> PStoreAddresses, QStoreAddresses;
3137  for (auto *BB : {PTB, PFB}) {
3138  if (!BB)
3139  continue;
3140  for (auto &I : *BB)
3141  if (StoreInst *SI = dyn_cast<StoreInst>(&I))
3142  PStoreAddresses.insert(SI->getPointerOperand());
3143  }
3144  for (auto *BB : {QTB, QFB}) {
3145  if (!BB)
3146  continue;
3147  for (auto &I : *BB)
3148  if (StoreInst *SI = dyn_cast<StoreInst>(&I))
3149  QStoreAddresses.insert(SI->getPointerOperand());
3150  }
3151 
3152  set_intersect(PStoreAddresses, QStoreAddresses);
3153  // set_intersect mutates PStoreAddresses in place. Rename it here to make it
3154  // clear what it contains.
3155  auto &CommonAddresses = PStoreAddresses;
3156 
3157  bool Changed = false;
3158  for (auto *Address : CommonAddresses)
3159  Changed |= mergeConditionalStoreToAddress(
3160  PTB, PFB, QTB, QFB, PostBB, Address, InvertPCond, InvertQCond, DL);
3161  return Changed;
3162 }
3163 
3164 /// If we have a conditional branch as a predecessor of another block,
3165 /// this function tries to simplify it. We know
3166 /// that PBI and BI are both conditional branches, and BI is in one of the
3167 /// successor blocks of PBI - PBI branches to BI.
3169  const DataLayout &DL) {
3170  assert(PBI->isConditional() && BI->isConditional());
3171  BasicBlock *BB = BI->getParent();
3172 
3173  // If this block ends with a branch instruction, and if there is a
3174  // predecessor that ends on a branch of the same condition, make
3175  // this conditional branch redundant.
3176  if (PBI->getCondition() == BI->getCondition() &&
3177  PBI->getSuccessor(0) != PBI->getSuccessor(1)) {
3178  // Okay, the outcome of this conditional branch is statically
3179  // knowable. If this block had a single pred, handle specially.
3180  if (BB->getSinglePredecessor()) {
3181  // Turn this into a branch on constant.
3182  bool CondIsTrue = PBI->getSuccessor(0) == BB;
3183  BI->setCondition(
3184  ConstantInt::get(Type::getInt1Ty(BB->getContext()), CondIsTrue));
3185  return true; // Nuke the branch on constant.
3186  }
3187 
3188  // Otherwise, if there are multiple predecessors, insert a PHI that merges
3189  // in the constant and simplify the block result. Subsequent passes of
3190  // simplifycfg will thread the block.
3192  pred_iterator PB = pred_begin(BB), PE = pred_end(BB);
3193  PHINode *NewPN = PHINode::Create(
3194  Type::getInt1Ty(BB->getContext()), std::distance(PB, PE),
3195  BI->getCondition()->getName() + ".pr", &BB->front());
3196  // Okay, we're going to insert the PHI node. Since PBI is not the only
3197  // predecessor, compute the PHI'd conditional value for all of the preds.
3198  // Any predecessor where the condition is not computable we keep symbolic.
3199  for (pred_iterator PI = PB; PI != PE; ++PI) {
3200  BasicBlock *P = *PI;
3201  if ((PBI = dyn_cast<BranchInst>(P->getTerminator())) && PBI != BI &&
3202  PBI->isConditional() && PBI->getCondition() == BI->getCondition() &&
3203  PBI->getSuccessor(0) != PBI->getSuccessor(1)) {
3204  bool CondIsTrue = PBI->getSuccessor(0) == BB;
3205  NewPN->addIncoming(
3206  ConstantInt::get(Type::getInt1Ty(BB->getContext()), CondIsTrue),
3207  P);
3208  } else {
3209  NewPN->addIncoming(BI->getCondition(), P);
3210  }
3211  }
3212 
3213  BI->setCondition(NewPN);
3214  return true;
3215  }
3216  }
3217 
3218  if (auto *CE = dyn_cast<ConstantExpr>(BI->getCondition()))
3219  if (CE->canTrap())
3220  return false;
3221 
3222  // If both branches are conditional and both contain stores to the same
3223  // address, remove the stores from the conditionals and create a conditional
3224  // merged store at the end.
3225  if (MergeCondStores && mergeConditionalStores(PBI, BI, DL))
3226  return true;
3227 
3228  // If this is a conditional branch in an empty block, and if any
3229  // predecessors are a conditional branch to one of our destinations,
3230  // fold the conditions into logical ops and one cond br.
3231 
3232  // Ignore dbg intrinsics.
3233  if (&*BB->instructionsWithoutDebug().begin() != BI)
3234  return false;
3235 
3236  int PBIOp, BIOp;
3237  if (PBI->getSuccessor(0) == BI->getSuccessor(0)) {
3238  PBIOp = 0;
3239  BIOp = 0;
3240  } else if (PBI->getSuccessor(0) == BI->getSuccessor(1)) {
3241  PBIOp = 0;
3242  BIOp = 1;
3243  } else if (PBI->getSuccessor(1) == BI->getSuccessor(0)) {
3244  PBIOp = 1;
3245  BIOp = 0;
3246  } else if (PBI->getSuccessor(1) == BI->getSuccessor(1)) {
3247  PBIOp = 1;
3248  BIOp = 1;
3249  } else {
3250  return false;
3251  }
3252 
3253  // Check to make sure that the other destination of this branch
3254  // isn't BB itself. If so, this is an infinite loop that will
3255  // keep getting unwound.
3256  if (PBI->getSuccessor(PBIOp) == BB)
3257  return false;
3258 
3259  // Do not perform this transformation if it would require
3260  // insertion of a large number of select instructions. For targets
3261  // without predication/cmovs, this is a big pessimization.
3262 
3263  // Also do not perform this transformation if any phi node in the common
3264  // destination block can trap when reached by BB or PBB (PR17073). In that
3265  // case, it would be unsafe to hoist the operation into a select instruction.
3266 
3267  BasicBlock *CommonDest = PBI->getSuccessor(PBIOp);
3268  unsigned NumPhis = 0;
3269  for (BasicBlock::iterator II = CommonDest->begin(); isa<PHINode>(II);
3270  ++II, ++NumPhis) {
3271  if (NumPhis > 2) // Disable this xform.
3272  return false;
3273 
3274  PHINode *PN = cast<PHINode>(II);
3275  Value *BIV = PN->getIncomingValueForBlock(BB);
3276  if (ConstantExpr *CE = dyn_cast<ConstantExpr>(BIV))
3277  if (CE->canTrap())
3278  return false;
3279 
3280  unsigned PBBIdx = PN->getBasicBlockIndex(PBI->getParent());
3281  Value *PBIV = PN->getIncomingValue(PBBIdx);
3282  if (ConstantExpr *CE = dyn_cast<ConstantExpr>(PBIV))
3283  if (CE->canTrap())
3284  return false;
3285  }
3286 
3287  // Finally, if everything is ok, fold the branches to logical ops.
3288  BasicBlock *OtherDest = BI->getSuccessor(BIOp ^ 1);
3289 
3290  LLVM_DEBUG(dbgs() << "FOLDING BRs:" << *PBI->getParent()
3291  << "AND: " << *BI->getParent());
3292 
3293  // If OtherDest *is* BB, then BB is a basic block with a single conditional
3294  // branch in it, where one edge (OtherDest) goes back to itself but the other
3295  // exits. We don't *know* that the program avoids the infinite loop
3296  // (even though that seems likely). If we do this xform naively, we'll end up
3297  // recursively unpeeling the loop. Since we know that (after the xform is
3298  // done) that the block *is* infinite if reached, we just make it an obviously
3299  // infinite loop with no cond branch.
3300  if (OtherDest == BB) {
3301  // Insert it at the end of the function, because it's either code,
3302  // or it won't matter if it's hot. :)
3303  BasicBlock *InfLoopBlock =
3304  BasicBlock::Create(BB->getContext(), "infloop", BB->getParent());
3305  BranchInst::Create(InfLoopBlock, InfLoopBlock);
3306  OtherDest = InfLoopBlock;
3307  }
3308 
3309  LLVM_DEBUG(dbgs() << *PBI->getParent()->getParent());
3310 
3311  // BI may have other predecessors. Because of this, we leave
3312  // it alone, but modify PBI.
3313 
3314  // Make sure we get to CommonDest on True&True directions.
3315  Value *PBICond = PBI->getCondition();
3316  IRBuilder<NoFolder> Builder(PBI);
3317  if (PBIOp)
3318  PBICond = Builder.CreateNot(PBICond, PBICond->getName() + ".not");
3319 
3320  Value *BICond = BI->getCondition();
3321  if (BIOp)
3322  BICond = Builder.CreateNot(BICond, BICond->getName() + ".not");
3323 
3324  // Merge the conditions.
3325  Value *Cond = Builder.CreateOr(PBICond, BICond, "brmerge");
3326 
3327  // Modify PBI to branch on the new condition to the new dests.
3328  PBI->setCondition(Cond);
3329  PBI->setSuccessor(0, CommonDest);
3330  PBI->setSuccessor(1, OtherDest);
3331 
3332  // Update branch weight for PBI.
3333  uint64_t PredTrueWeight, PredFalseWeight, SuccTrueWeight, SuccFalseWeight;
3334  uint64_t PredCommon, PredOther, SuccCommon, SuccOther;
3335  bool HasWeights =
3336  extractPredSuccWeights(PBI, BI, PredTrueWeight, PredFalseWeight,
3337  SuccTrueWeight, SuccFalseWeight);
3338  if (HasWeights) {
3339  PredCommon = PBIOp ? PredFalseWeight : PredTrueWeight;
3340  PredOther = PBIOp ? PredTrueWeight : PredFalseWeight;
3341  SuccCommon = BIOp ? SuccFalseWeight : SuccTrueWeight;
3342  SuccOther = BIOp ? SuccTrueWeight : SuccFalseWeight;
3343  // The weight to CommonDest should be PredCommon * SuccTotal +
3344  // PredOther * SuccCommon.
3345  // The weight to OtherDest should be PredOther * SuccOther.
3346  uint64_t NewWeights[2] = {PredCommon * (SuccCommon + SuccOther) +
3347  PredOther * SuccCommon,
3348  PredOther * SuccOther};
3349  // Halve the weights if any of them cannot fit in an uint32_t
3350  FitWeights(NewWeights);
3351 
3352  setBranchWeights(PBI, NewWeights[0], NewWeights[1]);
3353  }
3354 
3355  // OtherDest may have phi nodes. If so, add an entry from PBI's
3356  // block that are identical to the entries for BI's block.
3357  AddPredecessorToBlock(OtherDest, PBI->getParent(), BB);
3358 
3359  // We know that the CommonDest already had an edge from PBI to
3360  // it. If it has PHIs though, the PHIs may have different
3361  // entries for BB and PBI's BB. If so, insert a select to make
3362  // them agree.
3363  for (PHINode &PN : CommonDest->phis()) {
3364  Value *BIV = PN.getIncomingValueForBlock(BB);
3365  unsigned PBBIdx = PN.getBasicBlockIndex(PBI->getParent());
3366  Value *PBIV = PN.getIncomingValue(PBBIdx);
3367  if (BIV != PBIV) {
3368  // Insert a select in PBI to pick the right value.
3369  SelectInst *NV = cast<SelectInst>(
3370  Builder.CreateSelect(PBICond, PBIV, BIV, PBIV->getName() + ".mux"));
3371  PN.setIncomingValue(PBBIdx, NV);
3372  // Although the select has the same condition as PBI, the original branch
3373  // weights for PBI do not apply to the new select because the select's
3374  // 'logical' edges are incoming edges of the phi that is eliminated, not
3375  // the outgoing edges of PBI.
3376  if (HasWeights) {
3377  uint64_t PredCommon = PBIOp ? PredFalseWeight : PredTrueWeight;
3378  uint64_t PredOther = PBIOp ? PredTrueWeight : PredFalseWeight;
3379  uint64_t SuccCommon = BIOp ? SuccFalseWeight : SuccTrueWeight;
3380  uint64_t SuccOther = BIOp ? SuccTrueWeight : SuccFalseWeight;
3381  // The weight to PredCommonDest should be PredCommon * SuccTotal.
3382  // The weight to PredOtherDest should be PredOther * SuccCommon.
3383  uint64_t NewWeights[2] = {PredCommon * (SuccCommon + SuccOther),
3384  PredOther * SuccCommon};
3385 
3386  FitWeights(NewWeights);
3387 
3388  setBranchWeights(NV, NewWeights[0], NewWeights[1]);
3389  }
3390  }
3391  }
3392 
3393  LLVM_DEBUG(dbgs() << "INTO: " << *PBI->getParent());
3394  LLVM_DEBUG(dbgs() << *PBI->getParent()->getParent());
3395 
3396  // This basic block is probably dead. We know it has at least
3397  // one fewer predecessor.
3398  return true;
3399 }
3400 
3401 // Simplifies a terminator by replacing it with a branch to TrueBB if Cond is
3402 // true or to FalseBB if Cond is false.
3403 // Takes care of updating the successors and removing the old terminator.
3404 // Also makes sure not to introduce new successors by assuming that edges to
3405 // non-successor TrueBBs and FalseBBs aren't reachable.
3407  BasicBlock *TrueBB, BasicBlock *FalseBB,
3408  uint32_t TrueWeight,
3409  uint32_t FalseWeight) {
3410  // Remove any superfluous successor edges from the CFG.
3411  // First, figure out which successors to preserve.
3412  // If TrueBB and FalseBB are equal, only try to preserve one copy of that
3413  // successor.
3414  BasicBlock *KeepEdge1 = TrueBB;
3415  BasicBlock *KeepEdge2 = TrueBB != FalseBB ? FalseBB : nullptr;
3416 
3417  // Then remove the rest.
3418  for (BasicBlock *Succ : OldTerm->successors()) {
3419  // Make sure only to keep exactly one copy of each edge.
3420  if (Succ == KeepEdge1)
3421  KeepEdge1 = nullptr;
3422  else if (Succ == KeepEdge2)
3423  KeepEdge2 = nullptr;
3424  else
3425  Succ->removePredecessor(OldTerm->getParent(),
3426  /*DontDeleteUselessPHIs=*/true);
3427  }
3428 
3429  IRBuilder<> Builder(OldTerm);
3430  Builder.SetCurrentDebugLocation(OldTerm->getDebugLoc());
3431 
3432  // Insert an appropriate new terminator.
3433  if (!KeepEdge1 && !KeepEdge2) {
3434  if (TrueBB == FalseBB)
3435  // We were only looking for one successor, and it was present.
3436  // Create an unconditional branch to it.
3437  Builder.CreateBr(TrueBB);
3438  else {
3439  // We found both of the successors we were looking for.
3440  // Create a conditional branch sharing the condition of the select.
3441  BranchInst *NewBI = Builder.CreateCondBr(Cond, TrueBB, FalseBB);
3442  if (TrueWeight != FalseWeight)
3443  setBranchWeights(NewBI, TrueWeight, FalseWeight);
3444  }
3445  } else if (KeepEdge1 && (KeepEdge2 || TrueBB == FalseBB)) {
3446  // Neither of the selected blocks were successors, so this
3447  // terminator must be unreachable.
3448  new UnreachableInst(OldTerm->getContext(), OldTerm);
3449  } else {
3450  // One of the selected values was a successor, but the other wasn't.
3451  // Insert an unconditional branch to the one that was found;
3452  // the edge to the one that wasn't must be unreachable.
3453  if (!KeepEdge1)
3454  // Only TrueBB was found.
3455  Builder.CreateBr(TrueBB);
3456  else
3457  // Only FalseBB was found.
3458  Builder.CreateBr(FalseBB);
3459  }
3460 
3462  return true;
3463 }
3464 
3465 // Replaces
3466 // (switch (select cond, X, Y)) on constant X, Y
3467 // with a branch - conditional if X and Y lead to distinct BBs,
3468 // unconditional otherwise.
3470  // Check for constant integer values in the select.
3471  ConstantInt *TrueVal = dyn_cast<ConstantInt>(Select->getTrueValue());
3472  ConstantInt *FalseVal = dyn_cast<ConstantInt>(Select->getFalseValue());
3473  if (!TrueVal || !FalseVal)
3474  return false;
3475 
3476  // Find the relevant condition and destinations.
3477  Value *Condition = Select->getCondition();
3478  BasicBlock *TrueBB = SI->findCaseValue(TrueVal)->getCaseSuccessor();
3479  BasicBlock *FalseBB = SI->findCaseValue(FalseVal)->getCaseSuccessor();
3480 
3481  // Get weight for TrueBB and FalseBB.
3482  uint32_t TrueWeight = 0, FalseWeight = 0;
3483  SmallVector<uint64_t, 8> Weights;
3484  bool HasWeights = HasBranchWeights(SI);
3485  if (HasWeights) {
3486  GetBranchWeights(SI, Weights);
3487  if (Weights.size() == 1 + SI->getNumCases()) {
3488  TrueWeight =
3489  (uint32_t)Weights[SI->findCaseValue(TrueVal)->getSuccessorIndex()];
3490  FalseWeight =
3491  (uint32_t)Weights[SI->findCaseValue(FalseVal)->getSuccessorIndex()];
3492  }
3493  }
3494 
3495  // Perform the actual simplification.
3496  return SimplifyTerminatorOnSelect(SI, Condition, TrueBB, FalseBB, TrueWeight,
3497  FalseWeight);
3498 }
3499 
3500 // Replaces
3501 // (indirectbr (select cond, blockaddress(@fn, BlockA),
3502 // blockaddress(@fn, BlockB)))
3503 // with
3504 // (br cond, BlockA, BlockB).
3506  // Check that both operands of the select are block addresses.
3509  if (!TBA || !FBA)
3510  return false;
3511 
3512  // Extract the actual blocks.
3513  BasicBlock *TrueBB = TBA->getBasicBlock();
3514  BasicBlock *FalseBB = FBA->getBasicBlock();
3515 
3516  // Perform the actual simplification.
3517  return SimplifyTerminatorOnSelect(IBI, SI->getCondition(), TrueBB, FalseBB, 0,
3518  0);
3519 }
3520 
3521 /// This is called when we find an icmp instruction
3522 /// (a seteq/setne with a constant) as the only instruction in a
3523 /// block that ends with an uncond branch. We are looking for a very specific
3524 /// pattern that occurs when "A == 1 || A == 2 || A == 3" gets simplified. In
3525 /// this case, we merge the first two "or's of icmp" into a switch, but then the
3526 /// default value goes to an uncond block with a seteq in it, we get something
3527 /// like:
3528 ///
3529 /// switch i8 %A, label %DEFAULT [ i8 1, label %end i8 2, label %end ]
3530 /// DEFAULT:
3531 /// %tmp = icmp eq i8 %A, 92
3532 /// br label %end
3533 /// end:
3534 /// ... = phi i1 [ true, %entry ], [ %tmp, %DEFAULT ], [ true, %entry ]
3535 ///
3536 /// We prefer to split the edge to 'end' so that there is a true/false entry to
3537 /// the PHI, merging the third icmp into the switch.
3539  ICmpInst *ICI, IRBuilder<> &Builder, const DataLayout &DL,
3540  const TargetTransformInfo &TTI, const SimplifyCFGOptions &Options) {
3541  BasicBlock *BB = ICI->getParent();
3542 
3543  // If the block has any PHIs in it or the icmp has multiple uses, it is too
3544  // complex.
3545  if (isa<PHINode>(BB->begin()) || !ICI->hasOneUse())
3546  return false;
3547 
3548  Value *V = ICI->getOperand(0);
3549  ConstantInt *Cst = cast<ConstantInt>(ICI->getOperand(1));
3550 
3551  // The pattern we're looking for is where our only predecessor is a switch on
3552  // 'V' and this block is the default case for the switch. In this case we can
3553  // fold the compared value into the switch to simplify things.
3554  BasicBlock *Pred = BB->getSinglePredecessor();
3555  if (!Pred || !isa<SwitchInst>(Pred->getTerminator()))
3556  return false;
3557 
3558  SwitchInst *SI = cast<SwitchInst>(Pred->getTerminator());
3559  if (SI->getCondition() != V)
3560  return false;
3561 
3562  // If BB is reachable on a non-default case, then we simply know the value of
3563  // V in this block. Substitute it and constant fold the icmp instruction
3564  // away.
3565  if (SI->getDefaultDest() != BB) {
3566  ConstantInt *VVal = SI->findCaseDest(BB);
3567  assert(VVal && "Should have a unique destination value");
3568  ICI->setOperand(0, VVal);
3569 
3570  if (Value *V = SimplifyInstruction(ICI, {DL, ICI})) {
3571  ICI->replaceAllUsesWith(V);
3572  ICI->eraseFromParent();
3573  }
3574  // BB is now empty, so it is likely to simplify away.
3575  return simplifyCFG(BB, TTI, Options) | true;
3576  }
3577 
3578  // Ok, the block is reachable from the default dest. If the constant we're
3579  // comparing exists in one of the other edges, then we can constant fold ICI
3580  // and zap it.
3581  if (SI->findCaseValue(Cst) != SI->case_default()) {
3582  Value *V;
3583  if (ICI->getPredicate() == ICmpInst::ICMP_EQ)
3584  V = ConstantInt::getFalse(BB->getContext());
3585  else
3586  V = ConstantInt::getTrue(BB->getContext());
3587 
3588  ICI->replaceAllUsesWith(V);
3589  ICI->eraseFromParent();
3590  // BB is now empty, so it is likely to simplify away.
3591  return simplifyCFG(BB, TTI, Options) | true;
3592  }
3593 
3594  // The use of the icmp has to be in the 'end' block, by the only PHI node in
3595  // the block.
3596  BasicBlock *SuccBlock = BB->getTerminator()->getSuccessor(0);
3597  PHINode *PHIUse = dyn_cast<PHINode>(ICI->user_back());
3598  if (PHIUse == nullptr || PHIUse != &SuccBlock->front() ||
3599  isa<PHINode>(++BasicBlock::iterator(PHIUse)))
3600  return false;
3601 
3602  // If the icmp is a SETEQ, then the default dest gets false, the new edge gets
3603  // true in the PHI.
3604  Constant *DefaultCst = ConstantInt::getTrue(BB->getContext());
3605  Constant *NewCst = ConstantInt::getFalse(BB->getContext());
3606 
3607  if (ICI->getPredicate() == ICmpInst::ICMP_EQ)
3608  std::swap(DefaultCst, NewCst);
3609 
3610  // Replace ICI (which is used by the PHI for the default value) with true or
3611  // false depending on if it is EQ or NE.
3612  ICI->replaceAllUsesWith(DefaultCst);
3613  ICI->eraseFromParent();
3614 
3615  // Okay, the switch goes to this block on a default value. Add an edge from
3616  // the switch to the merge point on the compared value.
3617  BasicBlock *NewBB =
3618  BasicBlock::Create(BB->getContext(), "switch.edge", BB->getParent(), BB);
3619  SmallVector<uint64_t, 8> Weights;
3620  bool HasWeights = HasBranchWeights(SI);
3621  if (HasWeights) {
3622  GetBranchWeights(SI, Weights);
3623  if (Weights.size() == 1 + SI->getNumCases()) {
3624  // Split weight for default case to case for "Cst".
3625  Weights[0] = (Weights[0] + 1) >> 1;
3626  Weights.push_back(Weights[0]);
3627 
3628  SmallVector<uint32_t, 8> MDWeights(Weights.begin(), Weights.end());
3629  setBranchWeights(SI, MDWeights);
3630  }
3631  }
3632  SI->addCase(Cst, NewBB);
3633 
3634  // NewBB branches to the phi block, add the uncond branch and the phi entry.
3635  Builder.SetInsertPoint(NewBB);
3636  Builder.SetCurrentDebugLocation(SI->getDebugLoc());
3637  Builder.CreateBr(SuccBlock);
3638  PHIUse->addIncoming(NewCst, NewBB);
3639  return true;
3640 }
3641 
3642 /// The specified branch is a conditional branch.
3643 /// Check to see if it is branching on an or/and chain of icmp instructions, and
3644 /// fold it into a switch instruction if so.
3646  const DataLayout &DL) {
3647  Instruction *Cond = dyn_cast<Instruction>(BI->getCondition());
3648  if (!Cond)
3649  return false;
3650 
3651  // Change br (X == 0 | X == 1), T, F into a switch instruction.
3652  // If this is a bunch of seteq's or'd together, or if it's a bunch of
3653  // 'setne's and'ed together, collect them.
3654 
3655  // Try to gather values from a chain of and/or to be turned into a switch
3656  ConstantComparesGatherer ConstantCompare(Cond, DL);
3657  // Unpack the result
3658  SmallVectorImpl<ConstantInt *> &Values = ConstantCompare.Vals;
3659  Value *CompVal = ConstantCompare.CompValue;
3660  unsigned UsedICmps = ConstantCompare.UsedICmps;
3661  Value *ExtraCase = ConstantCompare.Extra;
3662 
3663  // If we didn't have a multiply compared value, fail.
3664  if (!CompVal)
3665  return false;
3666 
3667  // Avoid turning single icmps into a switch.
3668  if (UsedICmps <= 1)
3669  return false;
3670 
3671  bool TrueWhenEqual = (Cond->getOpcode() == Instruction::Or);
3672 
3673  // There might be duplicate constants in the list, which the switch
3674  // instruction can't handle, remove them now.
3675  array_pod_sort(Values.begin(), Values.end(), ConstantIntSortPredicate);
3676  Values.erase(std::unique(Values.begin(), Values.end()), Values.end());
3677 
3678  // If Extra was used, we require at least two switch values to do the
3679  // transformation. A switch with one value is just a conditional branch.
3680  if (ExtraCase && Values.size() < 2)
3681  return false;
3682 
3683  // TODO: Preserve branch weight metadata, similarly to how
3684  // FoldValueComparisonIntoPredecessors preserves it.
3685 
3686  // Figure out which block is which destination.
3687  BasicBlock *DefaultBB = BI->getSuccessor(1);
3688  BasicBlock *EdgeBB = BI->getSuccessor(0);
3689  if (!TrueWhenEqual)
3690  std::swap(DefaultBB, EdgeBB);
3691 
3692  BasicBlock *BB = BI->getParent();
3693 
3694  LLVM_DEBUG(dbgs() << "Converting 'icmp' chain with " << Values.size()
3695  << " cases into SWITCH. BB is:\n"
3696  << *BB);
3697 
3698  // If there are any extra values that couldn't be folded into the switch
3699  // then we evaluate them with an explicit branch first. Split the block
3700  // right before the condbr to handle it.
3701  if (ExtraCase) {
3702  BasicBlock *NewBB =
3703  BB->splitBasicBlock(BI->getIterator(), "switch.early.test");
3704  // Remove the uncond branch added to the old block.
3705  TerminatorInst *OldTI = BB->getTerminator();
3706  Builder.SetInsertPoint(OldTI);
3707 
3708  if (TrueWhenEqual)
3709  Builder.CreateCondBr(ExtraCase, EdgeBB, NewBB);
3710  else
3711  Builder.CreateCondBr(ExtraCase, NewBB, EdgeBB);
3712 
3713  OldTI->eraseFromParent();
3714 
3715  // If there are PHI nodes in EdgeBB, then we need to add a new entry to them
3716  // for the edge we just added.
3717  AddPredecessorToBlock(EdgeBB, BB, NewBB);
3718 
3719  LLVM_DEBUG(dbgs() << " ** 'icmp' chain unhandled condition: " << *ExtraCase
3720  << "\nEXTRABB = " << *BB);
3721  BB = NewBB;
3722  }
3723 
3724  Builder.SetInsertPoint(BI);
3725  // Convert pointer to int before we switch.
3726  if (CompVal->getType()->isPointerTy()) {
3727  CompVal = Builder.CreatePtrToInt(
3728  CompVal, DL.getIntPtrType(CompVal->getType()), "magicptr");
3729  }
3730 
3731  // Create the new switch instruction now.
3732  SwitchInst *New = Builder.CreateSwitch(CompVal, DefaultBB, Values.size());
3733 
3734  // Add all of the 'cases' to the switch instruction.
3735  for (unsigned i = 0, e = Values.size(); i != e; ++i)
3736  New->addCase(Values[i], EdgeBB);
3737 
3738  // We added edges from PI to the EdgeBB. As such, if there were any
3739  // PHI nodes in EdgeBB, they need entries to be added corresponding to
3740  // the number of edges added.
3741  for (BasicBlock::iterator BBI = EdgeBB->begin(); isa<PHINode>(BBI); ++BBI) {
3742  PHINode *PN = cast<PHINode>(BBI);
3743  Value *InVal = PN->getIncomingValueForBlock(BB);
3744  for (unsigned i = 0, e = Values.size() - 1; i != e; ++i)
3745  PN->addIncoming(InVal, BB);
3746  }
3747 
3748  // Erase the old branch instruction.
3750 
3751  LLVM_DEBUG(dbgs() << " ** 'icmp' chain result is:\n" << *BB << '\n');
3752  return true;
3753 }
3754 
3755 bool SimplifyCFGOpt::SimplifyResume(ResumeInst *RI, IRBuilder<> &Builder) {
3756  if (isa<PHINode>(RI->getValue()))
3757  return SimplifyCommonResume(RI);
3758  else if (isa<LandingPadInst>(RI->getParent()->getFirstNonPHI()) &&
3759  RI->getValue() == RI->getParent()->getFirstNonPHI())
3760  // The resume must unwind the exception that caused control to branch here.
3761  return SimplifySingleResume(RI);
3762 
3763  return false;
3764 }
3765 
3766 // Simplify resume that is shared by several landing pads (phi of landing pad).
3767 bool SimplifyCFGOpt::SimplifyCommonResume(ResumeInst *RI) {
3768  BasicBlock *BB = RI->getParent();
3769 
3770  // Check that there are no other instructions except for debug intrinsics
3771  // between the phi of landing pads (RI->getValue()) and resume instruction.
3772  BasicBlock::iterator I = cast<Instruction>(RI->getValue())->getIterator(),
3773  E = RI->getIterator();
3774  while (++I != E)
3775  if (!isa<DbgInfoIntrinsic>(I))
3776  return false;
3777 
3778  SmallSetVector<BasicBlock *, 4> TrivialUnwindBlocks;
3779  auto *PhiLPInst = cast<PHINode>(RI->getValue());
3780 
3781  // Check incoming blocks to see if any of them are trivial.
3782  for (unsigned Idx = 0, End = PhiLPInst->getNumIncomingValues(); Idx != End;
3783  Idx++) {
3784  auto *IncomingBB = PhiLPInst->getIncomingBlock(Idx);
3785  auto *IncomingValue = PhiLPInst->getIncomingValue(Idx);
3786 
3787  // If the block has other successors, we can not delete it because
3788  // it has other dependents.
3789  if (IncomingBB->getUniqueSuccessor() != BB)
3790  continue;
3791 
3792  auto *LandingPad = dyn_cast<LandingPadInst>(IncomingBB->getFirstNonPHI());
3793  // Not the landing pad that caused the control to branch here.
3794  if (IncomingValue != LandingPad)
3795  continue;
3796 
3797  bool isTrivial = true;
3798 
3799  I = IncomingBB->getFirstNonPHI()->getIterator();
3800  E = IncomingBB->getTerminator()->getIterator();
3801  while (++I != E)
3802  if (!isa<DbgInfoIntrinsic>(I)) {
3803  isTrivial = false;
3804  break;
3805  }
3806 
3807  if (isTrivial)
3808  TrivialUnwindBlocks.insert(IncomingBB);
3809  }
3810 
3811  // If no trivial unwind blocks, don't do any simplifications.
3812  if (TrivialUnwindBlocks.empty())
3813  return false;
3814 
3815  // Turn all invokes that unwind here into calls.
3816  for (auto *TrivialBB : TrivialUnwindBlocks) {
3817  // Blocks that will be simplified should be removed from the phi node.
3818  // Note there could be multiple edges to the resume block, and we need
3819  // to remove them all.
3820  while (PhiLPInst->getBasicBlockIndex(TrivialBB) != -1)
3821  BB->removePredecessor(TrivialBB, true);
3822 
3823  for (pred_iterator PI = pred_begin(TrivialBB), PE = pred_end(TrivialBB);
3824  PI != PE;) {
3825  BasicBlock *Pred = *PI++;
3826  removeUnwindEdge(Pred);
3827  }
3828 
3829  // In each SimplifyCFG run, only the current processed block can be erased.
3830  // Otherwise, it will break the iteration of SimplifyCFG pass. So instead
3831  // of erasing TrivialBB, we only remove the branch to the common resume
3832  // block so that we can later erase the resume block since it has no
3833  // predecessors.
3834  TrivialBB->getTerminator()->eraseFromParent();
3835  new UnreachableInst(RI->getContext(), TrivialBB);
3836  }
3837 
3838  // Delete the resume block if all its predecessors have been removed.
3839  if (pred_empty(BB))
3840  BB->eraseFromParent();
3841 
3842  return !TrivialUnwindBlocks.empty();
3843 }
3844 
3845 // Simplify resume that is only used by a single (non-phi) landing pad.
3846 bool SimplifyCFGOpt::SimplifySingleResume(ResumeInst *RI) {
3847  BasicBlock *BB = RI->getParent();
3849  assert(RI->getValue() == LPInst &&
3850  "Resume must unwind the exception that caused control to here");
3851 
3852  // Check that there are no other instructions except for debug intrinsics.
3853  BasicBlock::iterator I = LPInst->getIterator(), E = RI->getIterator();
3854  while (++I != E)
3855  if (!isa<DbgInfoIntrinsic>(I))
3856  return false;
3857 
3858  // Turn all invokes that unwind here into calls and delete the basic block.
3859  for (pred_iterator PI = pred_begin(BB), PE = pred_end(BB); PI != PE;) {
3860  BasicBlock *Pred = *PI++;
3861  removeUnwindEdge(Pred);
3862  }
3863 
3864  // The landingpad is now unreachable. Zap it.
3865  BB->eraseFromParent();
3866  if (LoopHeaders)
3867  LoopHeaders->erase(BB);
3868  return true;
3869 }
3870 
3872  // If this is a trivial cleanup pad that executes no instructions, it can be
3873  // eliminated. If the cleanup pad continues to the caller, any predecessor
3874  // that is an EH pad will be updated to continue to the caller and any
3875  // predecessor that terminates with an invoke instruction will have its invoke
3876  // instruction converted to a call instruction. If the cleanup pad being
3877  // simplified does not continue to the caller, each predecessor will be
3878  // updated to continue to the unwind destination of the cleanup pad being
3879  // simplified.
3880  BasicBlock *BB = RI->getParent();
3881  CleanupPadInst *CPInst = RI->getCleanupPad();
3882  if (CPInst->getParent() != BB)
3883  // This isn't an empty cleanup.
3884  return false;
3885 
3886  // We cannot kill the pad if it has multiple uses. This typically arises
3887  // from unreachable basic blocks.
3888  if (!CPInst->hasOneUse())
3889  return false;
3890 
3891  // Check that there are no other instructions except for benign intrinsics.
3892  BasicBlock::iterator I = CPInst->getIterator(), E = RI->getIterator();
3893  while (++I != E) {
3894  auto *II = dyn_cast<IntrinsicInst>(I);
3895  if (!II)
3896  return false;
3897 
3898  Intrinsic::ID IntrinsicID = II->getIntrinsicID();
3899  switch (IntrinsicID) {
3900  case Intrinsic::dbg_declare:
3901  case Intrinsic::dbg_value:
3902  case Intrinsic::dbg_label:
3903  case Intrinsic::lifetime_end:
3904  break;
3905  default:
3906  return false;
3907  }
3908  }
3909 
3910  // If the cleanup return we are simplifying unwinds to the caller, this will
3911  // set UnwindDest to nullptr.
3912  BasicBlock *UnwindDest = RI->getUnwindDest();
3913  Instruction *DestEHPad = UnwindDest ? UnwindDest->getFirstNonPHI() : nullptr;
3914 
3915  // We're about to remove BB from the control flow. Before we do, sink any
3916  // PHINodes into the unwind destination. Doing this before changing the
3917  // control flow avoids some potentially slow checks, since we can currently
3918  // be certain that UnwindDest and BB have no common predecessors (since they
3919  // are both EH pads).
3920  if (UnwindDest) {
3921  // First, go through the PHI nodes in UnwindDest and update any nodes that
3922  // reference the block we are removing
3923  for (BasicBlock::iterator I = UnwindDest->begin(),
3924  IE = DestEHPad->getIterator();
3925  I != IE; ++I) {
3926  PHINode *DestPN = cast<PHINode>(I);
3927 
3928  int Idx = DestPN->getBasicBlockIndex(BB);
3929  // Since BB unwinds to UnwindDest, it has to be in the PHI node.
3930  assert(Idx != -1);
3931  // This PHI node has an incoming value that corresponds to a control
3932  // path through the cleanup pad we are removing. If the incoming
3933  // value is in the cleanup pad, it must be a PHINode (because we
3934  // verified above that the block is otherwise empty). Otherwise, the
3935  // value is either a constant or a value that dominates the cleanup
3936  // pad being removed.
3937  //
3938  // Because BB and UnwindDest are both EH pads, all of their
3939  // predecessors must unwind to these blocks, and since no instruction
3940  // can have multiple unwind destinations, there will be no overlap in
3941  // incoming blocks between SrcPN and DestPN.
3942  Value *SrcVal = DestPN->getIncomingValue(Idx);
3943  PHINode *SrcPN = dyn_cast<PHINode>(SrcVal);
3944 
3945  // Remove the entry for the block we are deleting.
3946  DestPN->removeIncomingValue(Idx, false);
3947 
3948  if (SrcPN && SrcPN->getParent() == BB) {
3949  // If the incoming value was a PHI node in the cleanup pad we are
3950  // removing, we need to merge that PHI node's incoming values into
3951  // DestPN.
3952  for (unsigned SrcIdx = 0, SrcE = SrcPN->getNumIncomingValues();
3953  SrcIdx != SrcE; ++SrcIdx) {
3954  DestPN->addIncoming(SrcPN->getIncomingValue(SrcIdx),
3955  SrcPN->getIncomingBlock(SrcIdx));
3956  }
3957  } else {
3958  // Otherwise, the incoming value came from above BB and
3959  // so we can just reuse it. We must associate all of BB's
3960  // predecessors with this value.
3961  for (auto *pred : predecessors(BB)) {
3962  DestPN->addIncoming(SrcVal, pred);
3963  }
3964  }
3965  }
3966 
3967  // Sink any remaining PHI nodes directly into UnwindDest.
3968  Instruction *InsertPt = DestEHPad;
3969  for (BasicBlock::iterator I = BB->begin(),
3970  IE = BB->getFirstNonPHI()->getIterator();
3971  I != IE;) {
3972  // The iterator must be incremented here because the instructions are
3973  // being moved to another block.
3974  PHINode *PN = cast<PHINode>(I++);
3975  if (PN->use_empty())
3976  // If the PHI node has no uses, just leave it. It will be erased
3977  // when we erase BB below.
3978  continue;
3979 
3980  // Otherwise, sink this PHI node into UnwindDest.
3981  // Any predecessors to UnwindDest which are not already represented
3982  // must be back edges which inherit the value from the path through
3983  // BB. In this case, the PHI value must reference itself.
3984  for (auto *pred : predecessors(UnwindDest))
3985  if (pred != BB)
3986  PN->addIncoming(PN, pred);
3987  PN->moveBefore(InsertPt);
3988  }
3989  }
3990 
3991  for (pred_iterator PI = pred_begin(BB), PE = pred_end(BB); PI != PE;) {
3992  // The iterator must be updated here because we are removing this pred.
3993  BasicBlock *PredBB = *PI++;
3994  if (UnwindDest == nullptr) {
3995  removeUnwindEdge(PredBB);
3996  } else {
3997  TerminatorInst *TI = PredBB->getTerminator();
3998  TI->replaceUsesOfWith(BB, UnwindDest);
3999  }
4000  }
4001 
4002  // The cleanup pad is now unreachable. Zap it.
4003  BB->eraseFromParent();
4004  return true;
4005 }
4006 
4007 // Try to merge two cleanuppads together.
4009  // Skip any cleanuprets which unwind to caller, there is nothing to merge
4010  // with.
4011  BasicBlock *UnwindDest = RI->getUnwindDest();
4012  if (!UnwindDest)
4013  return false;
4014 
4015  // This cleanupret isn't the only predecessor of this cleanuppad, it wouldn't
4016  // be safe to merge without code duplication.
4017  if (UnwindDest->getSinglePredecessor() != RI->getParent())
4018  return false;
4019 
4020  // Verify that our cleanuppad's unwind destination is another cleanuppad.
4021  auto *SuccessorCleanupPad = dyn_cast<CleanupPadInst>(&UnwindDest->front());
4022  if (!SuccessorCleanupPad)
4023  return false;
4024 
4025  CleanupPadInst *PredecessorCleanupPad = RI->getCleanupPad();
4026  // Replace any uses of the successor cleanupad with the predecessor pad
4027  // The only cleanuppad uses should be this cleanupret, it's cleanupret and
4028  // funclet bundle operands.
4029  SuccessorCleanupPad->replaceAllUsesWith(PredecessorCleanupPad);
4030  // Remove the old cleanuppad.
4031  SuccessorCleanupPad->eraseFromParent();
4032  // Now, we simply replace the cleanupret with a branch to the unwind
4033  // destination.
4034  BranchInst::Create(UnwindDest, RI->getParent());
4035  RI->eraseFromParent();
4036 
4037  return true;
4038 }
4039 
4040 bool SimplifyCFGOpt::SimplifyCleanupReturn(CleanupReturnInst *RI) {
4041  // It is possible to transiantly have an undef cleanuppad operand because we
4042  // have deleted some, but not all, dead blocks.
4043  // Eventually, this block will be deleted.
4044  if (isa<UndefValue>(RI->getOperand(0)))
4045  return false;
4046 
4047  if (mergeCleanupPad(RI))
4048  return true;
4049 
4050  if (removeEmptyCleanup(RI))
4051  return true;
4052 
4053  return false;
4054 }
4055 
4056 bool SimplifyCFGOpt::SimplifyReturn(ReturnInst *RI, IRBuilder<> &Builder) {
4057  BasicBlock *BB = RI->getParent();
4058  if (!BB->getFirstNonPHIOrDbg()->isTerminator())
4059  return false;
4060 
4061  // Find predecessors that end with branches.
4062  SmallVector<BasicBlock *, 8> UncondBranchPreds;
4063  SmallVector<BranchInst *, 8> CondBranchPreds;
4064  for (pred_iterator PI = pred_begin(BB), E = pred_end(BB); PI != E; ++PI) {
4065  BasicBlock *P = *PI;
4066  TerminatorInst *PTI = P->getTerminator();
4067  if (BranchInst *BI = dyn_cast<BranchInst>(PTI)) {
4068  if (BI->isUnconditional())
4069  UncondBranchPreds.push_back(P);
4070  else
4071  CondBranchPreds.push_back(BI);
4072  }
4073  }
4074 
4075  // If we found some, do the transformation!
4076  if (!UncondBranchPreds.empty() && DupRet) {
4077  while (!UncondBranchPreds.empty()) {
4078  BasicBlock *Pred = UncondBranchPreds.pop_back_val();
4079  LLVM_DEBUG(dbgs() << "FOLDING: " << *BB
4080  << "INTO UNCOND BRANCH PRED: " << *Pred);
4081  (void)FoldReturnIntoUncondBranch(RI, BB, Pred);
4082  }
4083 
4084  // If we eliminated all predecessors of the block, delete the block now.
4085  if (pred_empty(BB)) {
4086  // We know there are no successors, so just nuke the block.
4087  BB->eraseFromParent();
4088  if (LoopHeaders)
4089  LoopHeaders->erase(BB);
4090  }
4091 
4092  return true;
4093  }
4094 
4095  // Check out all of the conditional branches going to this return
4096  // instruction. If any of them just select between returns, change the
4097  // branch itself into a select/return pair.
4098  while (!CondBranchPreds.empty()) {
4099  BranchInst *BI = CondBranchPreds.pop_back_val();
4100 
4101  // Check to see if the non-BB successor is also a return block.
4102  if (isa<ReturnInst>(BI->getSuccessor(0)->getTerminator()) &&
4103  isa<ReturnInst>(BI->getSuccessor(1)->getTerminator()) &&
4104  SimplifyCondBranchToTwoReturns(BI, Builder))
4105  return true;
4106  }
4107  return false;
4108 }
4109 
4110 bool SimplifyCFGOpt::SimplifyUnreachable(UnreachableInst *UI) {
4111  BasicBlock *BB = UI->getParent();
4112 
4113  bool Changed = false;
4114 
4115  // If there are any instructions immediately before the unreachable that can
4116  // be removed, do so.
4117  while (UI->getIterator() != BB->begin()) {
4118  BasicBlock::iterator BBI = UI->getIterator();
4119  --BBI;
4120  // Do not delete instructions that can have side effects which might cause
4121  // the unreachable to not be reachable; specifically, calls and volatile
4122  // operations may have this effect.
4123  if (isa<CallInst>(BBI) && !isa<DbgInfoIntrinsic>(BBI))
4124  break;
4125 
4126  if (BBI->mayHaveSideEffects()) {
4127  if (auto *SI = dyn_cast<StoreInst>(BBI)) {
4128  if (SI->isVolatile())
4129  break;
4130  } else if (auto *LI = dyn_cast<LoadInst>(BBI)) {
4131  if (LI->isVolatile())
4132  break;
4133  } else if (auto *RMWI = dyn_cast<AtomicRMWInst>(BBI)) {
4134  if (RMWI->isVolatile())
4135  break;
4136  } else if (auto *CXI = dyn_cast<AtomicCmpXchgInst>(BBI)) {
4137  if (CXI->isVolatile())
4138  break;
4139  } else if (isa<CatchPadInst>(BBI)) {
4140  // A catchpad may invoke exception object constructors and such, which
4141  // in some languages can be arbitrary code, so be conservative by
4142  // default.
4143  // For CoreCLR, it just involves a type test, so can be removed.
4146  break;
4147  } else if (!isa<FenceInst>(BBI) && !isa<VAArgInst>(BBI) &&
4148  !isa<LandingPadInst>(BBI)) {
4149  break;
4150  }
4151  // Note that deleting LandingPad's here is in fact okay, although it
4152  // involves a bit of subtle reasoning. If this inst is a LandingPad,
4153  // all the predecessors of this block will be the unwind edges of Invokes,
4154  // and we can therefore guarantee this block will be erased.
4155  }
4156 
4157  // Delete this instruction (any uses are guaranteed to be dead)
4158  if (!BBI->use_empty())
4159  BBI->replaceAllUsesWith(UndefValue::get(BBI->getType()));
4160  BBI->eraseFromParent();
4161  Changed = true;
4162  }
4163 
4164  // If the unreachable instruction is the first in the block, take a gander
4165  // at all of the predecessors of this instruction, and simplify them.
4166  if (&BB->front() != UI)
4167  return Changed;
4168 
4170  for (unsigned i = 0, e = Preds.size(); i != e; ++i) {
4171  TerminatorInst *TI = Preds[i]->getTerminator();
4172  IRBuilder<> Builder(TI);
4173  if (auto *BI = dyn_cast<BranchInst>(TI)) {
4174  if (BI->isUnconditional()) {
4175  if (BI->getSuccessor(0) == BB) {
4176  new UnreachableInst(TI->getContext(), TI);
4177  TI->eraseFromParent();
4178  Changed = true;
4179  }
4180  } else {
4181  if (BI->getSuccessor(0) == BB) {
4182  Builder.CreateBr(BI->getSuccessor(1));
4184  } else if (BI->getSuccessor(1) == BB) {
4185  Builder.CreateBr(BI->getSuccessor(0));
4187  Changed = true;
4188  }
4189  }
4190  } else if (auto *SI = dyn_cast<SwitchInst>(TI)) {
4191  for (auto i = SI->case_begin(), e = SI->case_end(); i != e;) {
4192  if (i->getCaseSuccessor() != BB) {
4193  ++i;
4194  continue;
4195  }
4196  BB->removePredecessor(SI->getParent());
4197  i = SI->removeCase(i);
4198  e = SI->case_end();
4199  Changed = true;
4200  }
4201  } else if (auto *II = dyn_cast<InvokeInst>(TI)) {
4202  if (II->getUnwindDest() == BB) {
4203  removeUnwindEdge(TI->getParent());
4204  Changed = true;
4205  }
4206  } else if (auto *CSI = dyn_cast<CatchSwitchInst>(TI)) {
4207  if (CSI->getUnwindDest() == BB) {
4208  removeUnwindEdge(TI->getParent());
4209  Changed = true;
4210  continue;
4211  }
4212 
4213  for (CatchSwitchInst::handler_iterator I = CSI->handler_begin(),
4214  E = CSI->handler_end();
4215  I != E; ++I) {
4216  if (*I == BB) {
4217  CSI->removeHandler(I);
4218  --I;
4219  --E;
4220  Changed = true;
4221  }
4222  }
4223  if (CSI->getNumHandlers() == 0) {
4224  BasicBlock *CatchSwitchBB = CSI->getParent();
4225  if (CSI->hasUnwindDest()) {
4226  // Redirect preds to the unwind dest
4227  CatchSwitchBB->replaceAllUsesWith(CSI->getUnwindDest());
4228  } else {
4229  // Rewrite all preds to unwind to caller (or from invoke to call).
4230  SmallVector<BasicBlock *, 8> EHPreds(predecessors(CatchSwitchBB));
4231  for (BasicBlock *EHPred : EHPreds)
4232  removeUnwindEdge(EHPred);
4233  }
4234  // The catchswitch is no longer reachable.
4235  new UnreachableInst(CSI->getContext(), CSI);
4236  CSI->eraseFromParent();
4237  Changed = true;
4238  }
4239  } else if (isa<CleanupReturnInst>(TI)) {
4240  new UnreachableInst(TI->getContext(), TI);
4241  TI->eraseFromParent();
4242  Changed = true;
4243  }
4244  }
4245 
4246  // If this block is now dead, remove it.
4247  if (pred_empty(BB) && BB != &BB->getParent()->getEntryBlock()) {
4248  // We know there are no successors, so just nuke the block.
4249  BB->eraseFromParent();
4250  if (LoopHeaders)
4251  LoopHeaders->erase(BB);
4252  return true;
4253  }
4254 
4255  return Changed;
4256 }
4257 
4259  assert(Cases.size() >= 1);
4260 
4261  array_pod_sort(Cases.begin(), Cases.end(), ConstantIntSortPredicate);
4262  for (size_t I = 1, E = Cases.size(); I != E; ++I) {
4263  if (Cases[I - 1]->getValue() != Cases[I]->getValue() + 1)
4264  return false;
4265  }
4266  return true;
4267 }
4268 
4269 /// Turn a switch with two reachable destinations into an integer range
4270 /// comparison and branch.
4272  assert(SI->getNumCases() > 1 && "Degenerate switch?");
4273 
4274  bool HasDefault =
4275  !isa<UnreachableInst>(SI->getDefaultDest()->getFirstNonPHIOrDbg());
4276 
4277  // Partition the cases into two sets with different destinations.
4278  BasicBlock *DestA = HasDefault ? SI->getDefaultDest() : nullptr;
4279  BasicBlock *DestB = nullptr;
4282 
4283  for (auto Case : SI->cases()) {
4284  BasicBlock *Dest = Case.getCaseSuccessor();
4285  if (!DestA)
4286  DestA = Dest;
4287  if (Dest == DestA) {
4288  CasesA.push_back(Case.getCaseValue());
4289  continue;
4290  }
4291  if (!DestB)
4292  DestB = Dest;
4293  if (Dest == DestB) {
4294  CasesB.push_back(Case.getCaseValue());
4295  continue;
4296  }
4297  return false; // More than two destinations.
4298  }
4299 
4300  assert(DestA && DestB &&
4301  "Single-destination switch should have been folded.");
4302  assert(DestA != DestB);
4303  assert(DestB != SI->getDefaultDest());
4304  assert(!CasesB.empty() && "There must be non-default cases.");
4305  assert(!CasesA.empty() || HasDefault);
4306 
4307  // Figure out if one of the sets of cases form a contiguous range.
4308  SmallVectorImpl<ConstantInt *> *ContiguousCases = nullptr;
4309  BasicBlock *ContiguousDest = nullptr;
4310  BasicBlock *OtherDest = nullptr;
4311  if (!CasesA.empty() && CasesAreContiguous(CasesA)) {
4312  ContiguousCases = &CasesA;
4313  ContiguousDest = DestA;
4314  OtherDest = DestB;
4315  } else if (CasesAreContiguous(CasesB)) {
4316  ContiguousCases = &CasesB;
4317  ContiguousDest = DestB;
4318  OtherDest = DestA;
4319  } else
4320  return false;
4321 
4322  // Start building the compare and branch.
4323 
4324  Constant *Offset = ConstantExpr::getNeg(ContiguousCases->back());
4325  Constant *NumCases =
4326  ConstantInt::get(Offset->getType(), ContiguousCases->size());
4327 
4328  Value *Sub = SI->getCondition();
4329  if (!Offset->isNullValue())
4330  Sub = Builder.CreateAdd(Sub, Offset, Sub->getName() + ".off");
4331 
4332  Value *Cmp;
4333  // If NumCases overflowed, then all possible values jump to the successor.
4334  if (NumCases->isNullValue() && !ContiguousCases->empty())
4335  Cmp = ConstantInt::getTrue(SI->getContext());
4336  else
4337  Cmp = Builder.CreateICmpULT(Sub, NumCases, "switch");
4338  BranchInst *NewBI = Builder.CreateCondBr(Cmp, ContiguousDest, OtherDest);
4339 
4340  // Update weight for the newly-created conditional branch.
4341  if (HasBranchWeights(SI)) {
4342  SmallVector<uint64_t, 8> Weights;
4343  GetBranchWeights(SI, Weights);
4344  if (Weights.size() == 1 + SI->getNumCases()) {
4345  uint64_t TrueWeight = 0;
4346  uint64_t FalseWeight = 0;
4347  for (size_t I = 0, E = Weights.size(); I != E; ++I) {
4348  if (SI->getSuccessor(I) == ContiguousDest)
4349  TrueWeight += Weights[I];
4350  else
4351  FalseWeight += Weights[I];
4352  }
4353  while (TrueWeight > UINT32_MAX || FalseWeight > UINT32_MAX) {
4354  TrueWeight /= 2;
4355  FalseWeight /= 2;
4356  }
4357  setBranchWeights(NewBI, TrueWeight, FalseWeight);
4358  }
4359  }
4360 
4361  // Prune obsolete incoming values off the successors' PHI nodes.
4362  for (auto BBI = ContiguousDest->begin(); isa<PHINode>(BBI); ++BBI) {
4363  unsigned PreviousEdges = ContiguousCases->size();
4364  if (ContiguousDest == SI->getDefaultDest())
4365  ++PreviousEdges;
4366  for (unsigned I = 0, E = PreviousEdges - 1; I != E; ++I)
4367  cast<PHINode>(BBI)->removeIncomingValue(SI->getParent());
4368  }
4369  for (auto BBI = OtherDest->begin(); isa<PHINode>(BBI); ++BBI) {
4370  unsigned PreviousEdges = SI->getNumCases() - ContiguousCases->size();
4371  if (OtherDest == SI->getDefaultDest())
4372  ++PreviousEdges;
4373  for (unsigned I = 0, E = PreviousEdges - 1; I != E; ++I)
4374  cast<PHINode>(BBI)->removeIncomingValue(SI->getParent());
4375  }
4376 
4377  // Drop the switch.
4378  SI->eraseFromParent();
4379 
4380  return true;
4381 }
4382 
4383 /// Compute masked bits for the condition of a switch
4384 /// and use it to remove dead cases.
4386  const DataLayout &DL) {
4387  Value *Cond = SI->getCondition();
4388  unsigned Bits = Cond->getType()->getIntegerBitWidth();
4389  KnownBits Known = computeKnownBits(Cond, DL, 0, AC, SI);
4390 
4391  // We can also eliminate cases by determining that their values are outside of
4392  // the limited range of the condition based on how many significant (non-sign)
4393  // bits are in the condition value.
4394  unsigned ExtraSignBits = ComputeNumSignBits(Cond, DL, 0, AC, SI) - 1;
4395  unsigned MaxSignificantBitsInCond = Bits - ExtraSignBits;
4396 
4397  // Gather dead cases.
4399  for (auto &Case : SI->cases()) {
4400  const APInt &CaseVal = Case.getCaseValue()->getValue();
4401  if (Known.Zero.intersects(CaseVal) || !Known.One.isSubsetOf(CaseVal) ||
4402  (CaseVal.getMinSignedBits() > MaxSignificantBitsInCond)) {
4403  DeadCases.push_back(Case.getCaseValue());
4404  LLVM_DEBUG(dbgs() << "SimplifyCFG: switch case " << CaseVal
4405  << " is dead.\n");
4406  }
4407  }
4408 
4409  // If we can prove that the cases must cover all possible values, the
4410  // default destination becomes dead and we can remove it. If we know some
4411  // of the bits in the value, we can use that to more precisely compute the
4412  // number of possible unique case values.
4413  bool HasDefault =
4414  !isa<UnreachableInst>(SI->getDefaultDest()->getFirstNonPHIOrDbg());
4415  const unsigned NumUnknownBits =
4416  Bits - (Known.Zero | Known.One).countPopulation();
4417  assert(NumUnknownBits <= Bits);
4418  if (HasDefault && DeadCases.empty() &&
4419  NumUnknownBits < 64 /* avoid overflow */ &&
4420  SI->getNumCases() == (1ULL << NumUnknownBits)) {
4421  LLVM_DEBUG(dbgs() << "SimplifyCFG: switch default is dead.\n");
4422  BasicBlock *NewDefault =
4424  SI->setDefaultDest(&*NewDefault);
4425  SplitBlock(&*NewDefault, &NewDefault->front());
4426  auto *OldTI = NewDefault->getTerminator();
4427  new UnreachableInst(SI->getContext(), OldTI);
4429  return true;
4430  }
4431 
4432  SmallVector<uint64_t, 8> Weights;
4433  bool HasWeight = HasBranchWeights(SI);
4434  if (HasWeight) {
4435  GetBranchWeights(SI, Weights);
4436  HasWeight = (Weights.size() == 1 + SI->getNumCases());
4437  }
4438 
4439  // Remove dead cases from the switch.
4440  for (ConstantInt *DeadCase : DeadCases) {
4441  SwitchInst::CaseIt CaseI = SI->findCaseValue(DeadCase);
4442  assert(CaseI != SI->case_default() &&
4443  "Case was not found. Probably mistake in DeadCases forming.");
4444  if (HasWeight) {
4445  std::swap(Weights[CaseI->getCaseIndex() + 1], Weights.back());
4446  Weights.pop_back();
4447  }
4448 
4449  // Prune unused values from PHI nodes.
4450  CaseI->getCaseSuccessor()->removePredecessor(SI->getParent());
4451  SI->removeCase(CaseI);
4452  }
4453  if (HasWeight && Weights.size() >= 2) {
4454  SmallVector<uint32_t, 8> MDWeights(Weights.begin(), Weights.end());
4455  setBranchWeights(SI, MDWeights);
4456  }
4457 
4458  return !DeadCases.empty();
4459 }
4460 
4461 /// If BB would be eligible for simplification by
4462 /// TryToSimplifyUncondBranchFromEmptyBlock (i.e. it is empty and terminated
4463 /// by an unconditional branch), look at the phi node for BB in the successor
4464 /// block and see if the incoming value is equal to CaseValue. If so, return
4465 /// the phi node, and set PhiIndex to BB's index in the phi node.
4467  BasicBlock *BB, int *PhiIndex) {
4468  if (BB->getFirstNonPHIOrDbg() != BB->getTerminator())
4469  return nullptr; // BB must be empty to be a candidate for simplification.
4470  if (!BB->getSinglePredecessor())
4471  return nullptr; // BB must be dominated by the switch.
4472 
4474  if (!Branch || !Branch->isUnconditional())
4475  return nullptr; // Terminator must be unconditional branch.
4476 
4477  BasicBlock *Succ = Branch->getSuccessor(0);
4478 
4479  for (PHINode &PHI : Succ->phis()) {
4480  int Idx = PHI.getBasicBlockIndex(BB);
4481  assert(Idx >= 0 && "PHI has no entry for predecessor?");
4482 
4483  Value *InValue = PHI.getIncomingValue(Idx);
4484  if (InValue != CaseValue)
4485  continue;
4486 
4487  *PhiIndex = Idx;
4488  return &PHI;
4489  }
4490 
4491  return nullptr;
4492 }
4493 
4494 /// Try to forward the condition of a switch instruction to a phi node
4495 /// dominated by the switch, if that would mean that some of the destination
4496 /// blocks of the switch can be folded away. Return true if a change is made.
4498  using ForwardingNodesMap = DenseMap<PHINode *, SmallVector<int, 4>>;
4499 
4500  ForwardingNodesMap ForwardingNodes;
4501  BasicBlock *SwitchBlock = SI->getParent();
4502  bool Changed = false;
4503  for (auto &Case : SI->cases()) {
4504  ConstantInt *CaseValue = Case.getCaseValue();
4505  BasicBlock *CaseDest = Case.getCaseSuccessor();
4506 
4507  // Replace phi operands in successor blocks that are using the constant case
4508  // value rather than the switch condition variable:
4509  // switchbb:
4510  // switch i32 %x, label %default [
4511  // i32 17, label %succ
4512  // ...
4513  // succ:
4514  // %r = phi i32 ... [ 17, %switchbb ] ...
4515  // -->
4516  // %r = phi i32 ... [ %x, %switchbb ] ...
4517 
4518  for (PHINode &Phi : CaseDest->phis()) {
4519  // This only works if there is exactly 1 incoming edge from the switch to
4520  // a phi. If there is >1, that means multiple cases of the switch map to 1
4521  // value in the phi, and that phi value is not the switch condition. Thus,
4522  // this transform would not make sense (the phi would be invalid because
4523  // a phi can't have different incoming values from the same block).
4524  int SwitchBBIdx = Phi.getBasicBlockIndex(SwitchBlock);
4525  if (Phi.getIncomingValue(SwitchBBIdx) == CaseValue &&
4526  count(Phi.blocks(), SwitchBlock) == 1) {
4527  Phi.setIncomingValue(SwitchBBIdx, SI->getCondition());
4528  Changed = true;
4529  }
4530  }
4531 
4532  // Collect phi nodes that are indirectly using this switch's case constants.
4533  int PhiIdx;
4534  if (auto *Phi = FindPHIForConditionForwarding(CaseValue, CaseDest, &PhiIdx))
4535  ForwardingNodes[Phi].push_back(PhiIdx);
4536  }
4537 
4538  for (auto &ForwardingNode : ForwardingNodes) {
4539  PHINode *Phi = ForwardingNode.first;
4540  SmallVectorImpl<int> &Indexes = ForwardingNode.second;
4541  if (Indexes.size() < 2)
4542  continue;
4543 
4544  for (int Index : Indexes)
4545  Phi->setIncomingValue(Index, SI->getCondition());
4546  Changed = true;
4547  }
4548 
4549  return Changed;
4550 }
4551 
4552 /// Return true if the backend will be able to handle
4553 /// initializing an array of constants like C.
4555  if (C->isThreadDependent())
4556  return false;
4557  if (C->isDLLImportDependent())
4558  return false;
4559 
4560  if (!isa<ConstantFP>(C) && !isa<ConstantInt>(C) &&
4561  !isa<ConstantPointerNull>(C) && !isa<GlobalValue>(C) &&
4562  !isa<UndefValue>(C) && !isa<ConstantExpr>(C))
4563  return false;
4564 
4565  if (ConstantExpr *CE = dyn_cast<ConstantExpr>(C)) {
4566  if (!CE->isGEPWithNoNotionalOverIndexing())
4567  return false;
4568  if (!ValidLookupTableConstant(CE->getOperand(0), TTI))
4569  return false;
4570  }
4571 
4573  return false;
4574 
4575  return true;
4576 }
4577 
4578 /// If V is a Constant, return it. Otherwise, try to look up
4579 /// its constant value in ConstantPool, returning 0 if it's not there.
4580 static Constant *
4583  if (Constant *C = dyn_cast<Constant>(V))
4584  return C;
4585  return ConstantPool.lookup(V);
4586 }
4587 
4588 /// Try to fold instruction I into a constant. This works for
4589 /// simple instructions such as binary operations where both operands are
4590 /// constant or can be replaced by constants from the ConstantPool. Returns the
4591 /// resulting constant on success, 0 otherwise.
4592 static Constant *
4595  if (SelectInst *Select = dyn_cast<SelectInst>(I)) {
4596  Constant *A = LookupConstant(Select->getCondition(), ConstantPool);
4597  if (!A)
4598  return nullptr;
4599  if (A->isAllOnesValue())
4600  return LookupConstant(Select->getTrueValue(), ConstantPool);
4601  if (A->isNullValue())
4602  return LookupConstant(Select->getFalseValue(), ConstantPool);
4603  return nullptr;
4604  }
4605 
4607  for (unsigned N = 0, E = I->getNumOperands(); N != E; ++N) {
4609  COps.push_back(A);
4610  else
4611  return nullptr;
4612  }
4613 
4614  if (CmpInst *Cmp = dyn_cast<CmpInst>(I)) {
4615  return ConstantFoldCompareInstOperands(Cmp->getPredicate(), COps[0],
4616  COps[1], DL);
4617  }
4618 
4619  return ConstantFoldInstOperands(I, COps, DL);
4620 }
4621 
4622 /// Try to determine the resulting constant values in phi nodes
4623 /// at the common destination basic block, *CommonDest, for one of the case
4624 /// destionations CaseDest corresponding to value CaseVal (0 for the default
4625 /// case), of a switch instruction SI.
4626 static bool
4628  BasicBlock **CommonDest,
4629  SmallVectorImpl<std::pair<PHINode *, Constant *>> &Res,
4630  const DataLayout &DL, const TargetTransformInfo &TTI) {
4631  // The block from which we enter the common destination.
4632  BasicBlock *Pred = SI->getParent();
4633 
4634  // If CaseDest is empty except for some side-effect free instructions through
4635  // which we can constant-propagate the CaseVal, continue to its successor.
4637  ConstantPool.insert(std::make_pair(SI->getCondition(), CaseVal));
4638  for (Instruction &I :CaseDest->instructionsWithoutDebug()) {
4639  if (TerminatorInst *T = dyn_cast<TerminatorInst>(&I)) {
4640  // If the terminator is a simple branch, continue to the next block.
4641  if (T->getNumSuccessors() != 1 || T->isExceptional())
4642  return false;
4643  Pred = CaseDest;
4644  CaseDest = T->getSuccessor(0);
4645  } else if (Constant *C = ConstantFold(&I, DL, ConstantPool)) {
4646  // Instruction is side-effect free and constant.
4647 
4648  // If the instruction has uses outside this block or a phi node slot for
4649  // the block, it is not safe to bypass the instruction since it would then
4650  // no longer dominate all its uses.
4651  for (auto &Use : I.uses()) {
4652  User *User = Use.getUser();
4653  if (Instruction *I = dyn_cast<Instruction>(User))
4654  if (I->getParent() == CaseDest)
4655  continue;
4656  if (PHINode *Phi = dyn_cast<PHINode>(User))
4657  if (Phi->getIncomingBlock(Use) == CaseDest)
4658  continue;
4659  return false;
4660  }
4661 
4662  ConstantPool.insert(std::make_pair(&I, C));
4663  } else {
4664  break;
4665  }
4666  }
4667 
4668  // If we did not have a CommonDest before, use the current one.
4669  if (!*CommonDest)
4670  *CommonDest = CaseDest;
4671  // If the destination isn't the common one, abort.
4672  if (CaseDest != *CommonDest)
4673  return false;
4674 
4675  // Get the values for this case from phi nodes in the destination block.
4676  for (PHINode &PHI : (*CommonDest)->phis()) {
4677  int Idx = PHI.getBasicBlockIndex(Pred);
4678  if (Idx == -1)
4679  continue;
4680 
4681  Constant *ConstVal =
4682  LookupConstant(PHI.getIncomingValue(Idx), ConstantPool);
4683  if (!ConstVal)
4684  return false;
4685 
4686  // Be conservative about which kinds of constants we support.
4687  if (!ValidLookupTableConstant(ConstVal, TTI))
4688  return false;
4689 
4690  Res.push_back(std::make_pair(&PHI, ConstVal));
4691  }
4692 
4693  return Res.size() > 0;
4694 }
4695 
4696 // Helper function used to add CaseVal to the list of cases that generate
4697 // Result. Returns the updated number of cases that generate this result.
4698 static uintptr_t MapCaseToResult(ConstantInt *CaseVal,
4699  SwitchCaseResultVectorTy &UniqueResults,
4700  Constant *Result) {
4701  for (auto &I : UniqueResults) {
4702  if (I.first == Result) {
4703  I.second.push_back(CaseVal);
4704  return I.second.size();
4705  }
4706  }
4707  UniqueResults.push_back(
4708  std::make_pair(Result, SmallVector<ConstantInt *, 4>(1, CaseVal)));
4709  return 1;
4710 }
4711 
4712 // Helper function that initializes a map containing
4713 // results for the PHI node of the common destination block for a switch
4714 // instruction. Returns false if multiple PHI nodes have been found or if
4715 // there is not a common destination block for the switch.
4716 static bool
4718  SwitchCaseResultVectorTy &UniqueResults,
4719  Constant *&DefaultResult, const DataLayout &DL,
4720  const TargetTransformInfo &TTI,
4721  uintptr_t MaxUniqueResults, uintptr_t MaxCasesPerResult) {
4722  for (auto &I : SI->cases()) {
4723  ConstantInt *CaseVal = I.getCaseValue();
4724 
4725  // Resulting value at phi nodes for this case value.
4726  SwitchCaseResultsTy Results;
4727  if (!GetCaseResults(SI, CaseVal, I.getCaseSuccessor(), &CommonDest, Results,
4728  DL, TTI))
4729  return false;
4730 
4731  // Only one value per case is permitted.
4732  if (Results.size() > 1)
4733  return false;
4734 
4735  // Add the case->result mapping to UniqueResults.
4736  const uintptr_t NumCasesForResult =
4737  MapCaseToResult(CaseVal, UniqueResults, Results.begin()->second);
4738 
4739  // Early out if there are too many cases for this result.
4740  if (NumCasesForResult > MaxCasesPerResult)
4741  return false;
4742 
4743  // Early out if there are too many unique results.
4744  if (UniqueResults.size() > MaxUniqueResults)
4745  return false;
4746 
4747  // Check the PHI consistency.
4748  if (!PHI)
4749  PHI = Results[0].first;
4750  else if (PHI != Results[0].first)
4751  return false;
4752  }
4753  // Find the default result value.
4755  BasicBlock *DefaultDest = SI->getDefaultDest();
4756  GetCaseResults(SI, nullptr, SI->getDefaultDest(), &CommonDest, DefaultResults,
4757  DL, TTI);
4758  // If the default value is not found abort unless the default destination
4759  // is unreachable.
4760  DefaultResult =
4761  DefaultResults.size() == 1 ? DefaultResults.begin()->second : nullptr;
4762  if ((!DefaultResult &&
4763  !isa<UnreachableInst>(DefaultDest->getFirstNonPHIOrDbg())))
4764  return false;
4765 
4766  return true;
4767 }
4768 
4769 // Helper function that checks if it is possible to transform a switch with only
4770 // two cases (or two cases + default) that produces a result into a select.
4771 // Example:
4772 // switch (a) {
4773 // case 10: %0 = icmp eq i32 %a, 10
4774 // return 10; %1 = select i1 %0, i32 10, i32 4
4775 // case 20: ----> %2 = icmp eq i32 %a, 20
4776 // return 2; %3 = select i1 %2, i32 2, i32 %1
4777 // default:
4778 // return 4;
4779 // }
4780 static Value *ConvertTwoCaseSwitch(const SwitchCaseResultVectorTy &ResultVector,
4781  Constant *DefaultResult, Value *Condition,
4782  IRBuilder<> &Builder) {
4783  assert(ResultVector.size() == 2 &&
4784  "We should have exactly two unique results at this point");
4785  // If we are selecting between only two cases transform into a simple
4786  // select or a two-way select if default is possible.
4787  if (ResultVector[0].second.size() == 1 &&
4788  ResultVector[1].second.size() == 1) {
4789  ConstantInt *const FirstCase = ResultVector[0].second[0];
4790  ConstantInt *const SecondCase = ResultVector[1].second[0];
4791 
4792  bool DefaultCanTrigger = DefaultResult;
4793  Value *SelectValue = ResultVector[1].first;
4794  if (DefaultCanTrigger) {
4795  Value *const ValueCompare =
4796  Builder.CreateICmpEQ(Condition, SecondCase, "switch.selectcmp");
4797  SelectValue = Builder.CreateSelect(ValueCompare, ResultVector[1].first,
4798  DefaultResult, "switch.select");
4799  }
4800  Value *const ValueCompare =
4801  Builder.CreateICmpEQ(Condition, FirstCase, "switch.selectcmp");
4802  return Builder.CreateSelect(ValueCompare, ResultVector[0].first,
4803  SelectValue, "switch.select");
4804  }
4805 
4806  return nullptr;
4807 }
4808 
4809 // Helper function to cleanup a switch instruction that has been converted into
4810 // a select, fixing up PHI nodes and basic blocks.
4812  Value *SelectValue,
4813  IRBuilder<> &Builder) {
4814  BasicBlock *SelectBB = SI->getParent();
4815  while (PHI->getBasicBlockIndex(SelectBB) >= 0)
4816  PHI->removeIncomingValue(SelectBB);
4817  PHI->addIncoming(SelectValue, SelectBB);
4818 
4819  Builder.CreateBr(PHI->getParent());
4820 
4821  // Remove the switch.
4822  for (unsigned i = 0, e = SI->getNumSuccessors(); i < e; ++i) {
4823  BasicBlock *Succ = SI->getSuccessor(i);
4824 
4825  if (Succ == PHI->getParent())
4826  continue;
4827  Succ->removePredecessor(SelectBB);
4828  }
4829  SI->eraseFromParent();
4830 }
4831 
4832 /// If the switch is only used to initialize one or more
4833 /// phi nodes in a common successor block with only two different
4834 /// constant values, replace the switch with select.
4835 static bool switchToSelect(SwitchInst *SI, IRBuilder<> &Builder,
4836  const DataLayout &DL,
4837  const TargetTransformInfo &TTI) {
4838  Value *const Cond = SI->getCondition();
4839  PHINode *PHI = nullptr;
4840  BasicBlock *CommonDest = nullptr;
4841  Constant *DefaultResult;
4842  SwitchCaseResultVectorTy UniqueResults;
4843  // Collect all the cases that will deliver the same value from the switch.
4844  if (!InitializeUniqueCases(SI, PHI, CommonDest, UniqueResults, DefaultResult,
4845  DL, TTI, 2, 1))
4846  return false;
4847  // Selects choose between maximum two values.
4848  if (UniqueResults.size() != 2)
4849  return false;
4850  assert(PHI != nullptr && "PHI for value select not found");
4851 
4852  Builder.SetInsertPoint(SI);
4853  Value *SelectValue =
4854  ConvertTwoCaseSwitch(UniqueResults, DefaultResult, Cond, Builder);
4855  if (SelectValue) {
4856  RemoveSwitchAfterSelectConversion(SI, PHI, SelectValue, Builder);
4857  return true;
4858  }
4859  // The switch couldn't be converted into a select.
4860  return false;
4861 }
4862 
4863 namespace {
4864 
4865 /// This class represents a lookup table that can be used to replace a switch.
4866 class SwitchLookupTable {
4867 public:
4868  /// Create a lookup table to use as a switch replacement with the contents
4869  /// of Values, using DefaultValue to fill any holes in the table.
4870  SwitchLookupTable(
4871  Module &M, uint64_t TableSize, ConstantInt *Offset,
4872  const SmallVectorImpl<std::pair<ConstantInt *, Constant *>> &Values,
4873  Constant *DefaultValue, const DataLayout &DL, const StringRef &FuncName);
4874 
4875  /// Build instructions with Builder to retrieve the value at
4876  /// the position given by Index in the lookup table.
4877  Value *BuildLookup(Value *Index, IRBuilder<> &Builder);
4878 
4879  /// Return true if a table with TableSize elements of
4880  /// type ElementType would fit in a target-legal register.
4881  static bool WouldFitInRegister(const DataLayout &DL, uint64_t TableSize,
4882  Type *ElementType);
4883 
4884 private:
4885  // Depending on the contents of the table, it can be represented in
4886  // different ways.
4887  enum {
4888  // For tables where each element contains the same value, we just have to
4889  // store that single value and return it for each lookup.
4890  SingleValueKind,
4891 
4892  // For tables where there is a linear relationship between table index
4893  // and values. We calculate the result with a simple multiplication
4894  // and addition instead of a table lookup.
4895  LinearMapKind,
4896 
4897  // For small tables with integer elements, we can pack them into a bitmap
4898  // that fits into a target-legal register. Values are retrieved by
4899  // shift and mask operations.
4900  BitMapKind,
4901 
4902  // The table is stored as an array of values. Values are retrieved by load
4903  // instructions from the table.
4904  ArrayKind
4905  } Kind;
4906 
4907  // For SingleValueKind, this is the single value.
4908  Constant *SingleValue = nullptr;
4909 
4910  // For BitMapKind, this is the bitmap.
4911  ConstantInt *BitMap = nullptr;
4912  IntegerType *BitMapElementTy = nullptr;
4913 
4914  // For LinearMapKind, these are the constants used to derive the value.
4915  ConstantInt *LinearOffset = nullptr;
4916  ConstantInt *LinearMultiplier = nullptr;
4917 
4918  // For ArrayKind, this is the array.
4919  GlobalVariable *Array = nullptr;
4920 };
4921 
4922 } // end anonymous namespace
4923 
4924 SwitchLookupTable::SwitchLookupTable(
4925  Module &M, uint64_t TableSize, ConstantInt *Offset,
4926  const SmallVectorImpl<std::pair<ConstantInt *, Constant *>> &Values,
4927  Constant *DefaultValue, const DataLayout &DL, const StringRef &FuncName) {
4928  assert(Values.size() && "Can't build lookup table without values!");
4929  assert(TableSize >= Values.size() && "Can't fit values in table!");
4930 
4931  // If all values in the table are equal, this is that value.
4932  SingleValue = Values.begin()->second;
4933 
4934  Type *ValueType = Values.begin()->second->getType();
4935 
4936  // Build up the table contents.
4937  SmallVector<Constant *, 64> TableContents(TableSize);
4938  for (size_t I = 0, E = Values.size(); I != E; ++I) {
4939  ConstantInt *CaseVal = Values[I].first;
4940  Constant *CaseRes = Values[I].second;
4941  assert(CaseRes->getType() == ValueType);
4942 
4943  uint64_t Idx = (CaseVal->getValue() - Offset->getValue()).getLimitedValue();
4944  TableContents[Idx] = CaseRes;
4945 
4946  if (CaseRes != SingleValue)
4947  SingleValue = nullptr;
4948  }
4949 
4950  // Fill in any holes in the table with the default result.
4951  if (Values.size() < TableSize) {
4952  assert(DefaultValue &&
4953  "Need a default value to fill the lookup table holes.");
4954  assert(DefaultValue->getType() == ValueType);
4955  for (uint64_t I = 0; I < TableSize; ++I) {
4956  if (!TableContents[I])
4957  TableContents[I] = DefaultValue;
4958  }
4959 
4960  if (DefaultValue != SingleValue)
4961  SingleValue = nullptr;
4962  }
4963 
4964  // If each element in the table contains the same value, we only need to store
4965  // that single value.
4966  if (SingleValue) {
4967  Kind = SingleValueKind;
4968  return;
4969  }
4970 
4971  // Check if we can derive the value with a linear transformation from the
4972  // table index.
4973  if (isa<IntegerType>(ValueType)) {
4974  bool LinearMappingPossible = true;
4975  APInt PrevVal;
4976  APInt DistToPrev;
4977  assert(TableSize >= 2 && "Should be a SingleValue table.");
4978  // Check if there is the same distance between two consecutive values.
4979  for (uint64_t I = 0; I < TableSize; ++I) {
4980  ConstantInt *ConstVal = dyn_cast<ConstantInt>(TableContents[I]);
4981  if (!ConstVal) {
4982  // This is an undef. We could deal with it, but undefs in lookup tables
4983  // are very seldom. It's probably not worth the additional complexity.
4984  LinearMappingPossible = false;
4985  break;
4986  }
4987  const APInt &Val = ConstVal->getValue();
4988  if (I != 0) {
4989  APInt Dist = Val - PrevVal;
4990  if (I == 1) {
4991  DistToPrev = Dist;
4992  } else if (Dist != DistToPrev) {
4993  LinearMappingPossible = false;
4994  break;
4995  }
4996  }
4997  PrevVal = Val;
4998  }
4999  if (LinearMappingPossible) {
5000  LinearOffset = cast<ConstantInt>(TableContents[0]);
5001  LinearMultiplier = ConstantInt::get(M.getContext(), DistToPrev);
5002  Kind = LinearMapKind;
5003  ++NumLinearMaps;
5004  return;
5005  }
5006  }
5007 
5008  // If the type is integer and the table fits in a register, build a bitmap.
5009  if (WouldFitInRegister(DL, TableSize, ValueType)) {
5010  IntegerType *IT = cast<IntegerType>(ValueType);
5011  APInt TableInt(TableSize * IT->getBitWidth(), 0);
5012  for (uint64_t I = TableSize; I > 0; --I) {
5013  TableInt <<= IT->getBitWidth();
5014  // Insert values into the bitmap. Undef values are set to zero.
5015  if (!isa<UndefValue>(TableContents[I - 1])) {
5016  ConstantInt *Val = cast<ConstantInt>(TableContents[I - 1]);
5017  TableInt |= Val->getValue().zext(TableInt.getBitWidth());
5018  }
5019  }
5020  BitMap = ConstantInt::get(M.getContext(), TableInt);
5021  BitMapElementTy = IT;
5022  Kind = BitMapKind;
5023  ++NumBitMaps;
5024  return;
5025  }
5026 
5027  // Store the table in an array.
5028  ArrayType *ArrayTy = ArrayType::get(ValueType, TableSize);
5029  Constant *Initializer = ConstantArray::get(ArrayTy, TableContents);
5030 
5031  Array = new GlobalVariable(M, ArrayTy, /*constant=*/true,
5032  GlobalVariable::PrivateLinkage, Initializer,
5033  "switch.table." + FuncName);
5034  Array->setUnnamedAddr(GlobalValue::UnnamedAddr::Global);
5035  Kind = ArrayKind;
5036 }
5037 
5038 Value *SwitchLookupTable::BuildLookup(Value *Index, IRBuilder<> &Builder) {
5039  switch (Kind) {
5040  case SingleValueKind:
5041  return SingleValue;
5042  case LinearMapKind: {
5043  // Derive the result value from the input value.
5044  Value *Result = Builder.CreateIntCast(Index, LinearMultiplier->getType(),
5045  false, "switch.idx.cast");
5046  if (!LinearMultiplier->isOne())
5047  Result = Builder.CreateMul(Result, LinearMultiplier, "switch.idx.mult");
5048  if (!LinearOffset->isZero())
5049  Result = Builder.CreateAdd(Result, LinearOffset, "switch.offset");
5050  return Result;
5051  }
5052  case BitMapKind: {
5053  // Type of the bitmap (e.g. i59).
5054  IntegerType *MapTy = BitMap->getType();
5055 
5056  // Cast Index to the same type as the bitmap.
5057  // Note: The Index is <= the number of elements in the table, so
5058  // truncating it to the width of the bitmask is safe.
5059  Value *ShiftAmt = Builder.CreateZExtOrTrunc(Index, MapTy, "switch.cast");
5060 
5061  // Multiply the shift amount by the element width.
5062  ShiftAmt = Builder.CreateMul(
5063  ShiftAmt, ConstantInt::get(MapTy, BitMapElementTy->getBitWidth()),
5064  "switch.shiftamt");
5065 
5066  // Shift down.
5067  Value *DownShifted =
5068  Builder.CreateLShr(BitMap, ShiftAmt, "switch.downshift");
5069  // Mask off.
5070  return Builder.CreateTrunc(DownShifted, BitMapElementTy, "switch.masked");
5071  }
5072  case ArrayKind: {
5073  // Make sure the table index will not overflow when treated as signed.
5074  IntegerType *IT = cast<IntegerType>(Index->getType());
5075  uint64_t TableSize =
5076  Array->getInitializer()->getType()->getArrayNumElements();
5077  if (TableSize > (1ULL << (IT->getBitWidth() - 1)))
5078  Index = Builder.CreateZExt(
5079  Index, IntegerType::get(IT->getContext(), IT->getBitWidth() + 1),
5080  "switch.tableidx.zext");
5081 
5082  Value *GEPIndices[] = {Builder.getInt32(0), Index};
5083  Value *GEP = Builder.CreateInBoundsGEP(Array->getValueType(), Array,
5084  GEPIndices, "switch.gep");
5085  return Builder.CreateLoad(GEP, "switch.load");
5086  }
5087  }
5088  llvm_unreachable("Unknown lookup table kind!");
5089 }
5090 
5091 bool SwitchLookupTable::WouldFitInRegister(const DataLayout &DL,
5092  uint64_t TableSize,
5093  Type *ElementType) {
5094  auto *IT = dyn_cast<IntegerType>(ElementType);
5095  if (!IT)
5096  return false;
5097  // FIXME: If the type is wider than it needs to be, e.g. i8 but all values
5098  // are <= 15, we could try to narrow the type.
5099 
5100  // Avoid overflow, fitsInLegalInteger uses unsigned int for the width.
5101  if (TableSize >= UINT_MAX / IT->getBitWidth())
5102  return false;
5103  return DL.fitsInLegalInteger(TableSize * IT->getBitWidth());
5104 }
5105 
5106 /// Determine whether a lookup table should be built for this switch, based on
5107 /// the number of cases, size of the table, and the types of the results.
5108 static bool
5109 ShouldBuildLookupTable(SwitchInst *SI, uint64_t TableSize,
5110  const TargetTransformInfo &TTI, const DataLayout &DL,
5111  const SmallDenseMap<PHINode *, Type *> &ResultTypes) {
5112  if (SI->getNumCases() > TableSize || TableSize >= UINT64_MAX / 10)
5113  return false; // TableSize overflowed, or mul below might overflow.
5114 
5115  bool AllTablesFitInRegister = true;
5116  bool HasIllegalType = false;
5117  for (const auto &I : ResultTypes) {
5118  Type *Ty = I.second;
5119 
5120  // Saturate this flag to true.
5121  HasIllegalType = HasIllegalType || !TTI.isTypeLegal(Ty);
5122 
5123  // Saturate this flag to false.
5124  AllTablesFitInRegister =
5125  AllTablesFitInRegister &&
5126  SwitchLookupTable::WouldFitInRegister(DL, TableSize, Ty);
5127 
5128  // If both flags saturate, we're done. NOTE: This *only* works with
5129  // saturating flags, and all flags have to saturate first due to the
5130  // non-deterministic behavior of iterating over a dense map.
5131  if (HasIllegalType && !AllTablesFitInRegister)
5132  break;
5133  }
5134 
5135  // If each table would fit in a register, we should build it anyway.
5136  if (AllTablesFitInRegister)
5137  return true;
5138 
5139  // Don't build a table that doesn't fit in-register if it has illegal types.
5140  if (HasIllegalType)
5141  return false;
5142 
5143  // The table density should be at least 40%. This is the same criterion as for
5144  // jump tables, see SelectionDAGBuilder::handleJTSwitchCase.
5145  // FIXME: Find the best cut-off.
5146  return SI->getNumCases() * 10 >= TableSize * 4;
5147 }
5148 
5149 /// Try to reuse the switch table index compare. Following pattern:
5150 /// \code
5151 /// if (idx < tablesize)
5152 /// r = table[idx]; // table does not contain default_value
5153 /// else
5154 /// r = default_value;
5155 /// if (r != default_value)
5156 /// ...
5157 /// \endcode
5158 /// Is optimized to:
5159 /// \code
5160 /// cond = idx < tablesize;
5161 /// if (cond)
5162 /// r = table[idx];
5163 /// else
5164 /// r = default_value;
5165 /// if (cond)
5166 /// ...
5167 /// \endcode
5168 /// Jump threading will then eliminate the second if(cond).
5169 static void reuseTableCompare(
5170  User *PhiUser, BasicBlock *PhiBlock, BranchInst *RangeCheckBranch,
5171  Constant *DefaultValue,
5172  const SmallVectorImpl<std::pair<ConstantInt *, Constant *>> &Values) {
5173  ICmpInst *CmpInst = dyn_cast<ICmpInst>(PhiUser);
5174  if (!CmpInst)
5175  return;
5176 
5177  // We require that the compare is in the same block as the phi so that jump
5178  // threading can do its work afterwards.
5179  if (CmpInst->getParent() != PhiBlock)
5180  return;
5181 
5182  Constant *CmpOp1 = dyn_cast<Constant>(CmpInst->getOperand(1));
5183  if (!CmpOp1)
5184  return;
5185 
5186  Value *RangeCmp = RangeCheckBranch->getCondition();
5187  Constant *TrueConst = ConstantInt::getTrue(RangeCmp->getType());
5188  Constant *FalseConst = ConstantInt::getFalse(RangeCmp->getType());
5189 
5190  // Check if the compare with the default value is constant true or false.
5191  Constant *DefaultConst = ConstantExpr::getICmp(CmpInst->getPredicate(),
5192  DefaultValue, CmpOp1, true);
5193  if (DefaultConst != TrueConst && DefaultConst != FalseConst)
5194  return;
5195 
5196  // Check if the compare with the case values is distinct from the default
5197  // compare result.
5198  for (auto ValuePair : Values) {
5199  Constant *CaseConst = ConstantExpr::getICmp(CmpInst->getPredicate(),
5200  ValuePair.second, CmpOp1, true);
5201  if (!CaseConst || CaseConst == DefaultConst || isa<UndefValue>(CaseConst))
5202  return;
5203  assert((CaseConst == TrueConst || CaseConst == FalseConst) &&
5204  "Expect true or false as compare result.");
5205  }
5206 
5207  // Check if the branch instruction dominates the phi node. It's a simple
5208  // dominance check, but sufficient for our needs.
5209  // Although this check is invariant in the calling loops, it's better to do it
5210  // at this late stage. Practically we do it at most once for a switch.
5211  BasicBlock *BranchBlock = RangeCheckBranch->getParent();
5212  for (auto PI = pred_begin(PhiBlock), E = pred_end(PhiBlock); PI != E; ++PI) {
5213  BasicBlock *Pred = *PI;
5214  if (Pred != BranchBlock && Pred->getUniquePredecessor() != BranchBlock)
5215  return;
5216  }
5217 
5218  if (DefaultConst == FalseConst) {
5219  // The compare yields the same result. We can replace it.
5220  CmpInst->replaceAllUsesWith(RangeCmp);
5221  ++NumTableCmpReuses;
5222  } else {
5223  // The compare yields the same result, just inverted. We can replace it.
5224  Value *InvertedTableCmp = BinaryOperator::CreateXor(
5225  RangeCmp, ConstantInt::get(RangeCmp->getType(), 1), "inverted.cmp",
5226  RangeCheckBranch);
5227  CmpInst->replaceAllUsesWith(InvertedTableCmp);
5228  ++NumTableCmpReuses;
5229  }
5230 }
5231 
5232 /// If the switch is only used to initialize one or more phi nodes in a common
5233 /// successor block with different constant values, replace the switch with
5234 /// lookup tables.
5236  const DataLayout &DL,
5237  const TargetTransformInfo &TTI) {
5238  assert(SI->getNumCases() > 1 && "Degenerate switch?");
5239 
5240  Function *Fn = SI->getParent()->getParent();
5241  // Only build lookup table when we have a target that supports it or the
5242  // attribute is not set.
5243  if (!TTI.shouldBuildLookupTables() ||
5244  (Fn->getFnAttribute("no-jump-tables").getValueAsString() == "true"))
5245  return false;
5246 
5247  // FIXME: If the switch is too sparse for a lookup table, perhaps we could
5248  // split off a dense part and build a lookup table for that.
5249 
5250  // FIXME: This creates arrays of GEPs to constant strings, which means each
5251  // GEP needs a runtime relocation in PIC code. We should just build one big
5252  // string and lookup indices into that.
5253 
5254  // Ignore switches with less than three cases. Lookup tables will not make
5255  // them faster, so we don't analyze them.
5256  if (SI->getNumCases() < 3)
5257  return false;
5258 
5259  // Figure out the corresponding result for each case value and phi node in the
5260  // common destination, as well as the min and max case values.
5261  assert(SI->case_begin() != SI->case_end());
5262  SwitchInst::CaseIt CI = SI->case_begin();
5263  ConstantInt *MinCaseVal = CI->getCaseValue();
5264  ConstantInt *MaxCaseVal = CI->getCaseValue();
5265 
5266  BasicBlock *CommonDest = nullptr;
5267 
5268  using ResultListTy = SmallVector<std::pair<ConstantInt *, Constant *>, 4>;
5270 
5271  SmallDenseMap<PHINode *, Constant *> DefaultResults;
5274 
5275  for (SwitchInst::CaseIt E = SI->case_end(); CI != E; ++CI) {
5276  ConstantInt *CaseVal = CI->getCaseValue();
5277  if (CaseVal->getValue().slt(MinCaseVal->getValue()))
5278  MinCaseVal = CaseVal;
5279  if (CaseVal->getValue().sgt(MaxCaseVal->getValue()))
5280  MaxCaseVal = CaseVal;
5281 
5282  // Resulting value at phi nodes for this case value.
5283  using ResultsTy = SmallVector<std::pair<PHINode *, Constant *>, 4>;
5284  ResultsTy Results;
5285  if (!GetCaseResults(SI, CaseVal, CI->getCaseSuccessor(), &CommonDest,
5286  Results, DL, TTI))
5287  return false;
5288 
5289  // Append the result from this case to the list for each phi.
5290  for (const auto &I : Results) {
5291  PHINode *PHI = I.first;
5292  Constant *Value = I.second;
5293  if (!ResultLists.count(PHI))
5294  PHIs.push_back(PHI);
5295  ResultLists[PHI].push_back(std::make_pair(CaseVal, Value));
5296  }
5297  }
5298 
5299  // Keep track of the result types.
5300  for (PHINode *PHI : PHIs) {
5301  ResultTypes[PHI] = ResultLists[PHI][0].second->getType();
5302  }
5303 
5304  uint64_t NumResults = ResultLists[PHIs[0]].size();
5305  APInt RangeSpread = MaxCaseVal->getValue() - MinCaseVal->getValue();
5306  uint64_t TableSize = RangeSpread.getLimitedValue() + 1;
5307  bool TableHasHoles = (NumResults < TableSize);
5308 
5309  // If the table has holes, we need a constant result for the default case
5310  // or a bitmask that fits in a register.
5311  SmallVector<std::pair<PHINode *, Constant *>, 4> DefaultResultsList;
5312  bool HasDefaultResults =
5313  GetCaseResults(SI, nullptr, SI->getDefaultDest(), &CommonDest,
5314  DefaultResultsList, DL, TTI);
5315 
5316  bool NeedMask = (TableHasHoles && !HasDefaultResults);
5317  if (NeedMask) {
5318  // As an extra penalty for the validity test we require more cases.
5319  if (SI->getNumCases() < 4) // FIXME: Find best threshold value (benchmark).
5320  return false;
5321  if (!DL.fitsInLegalInteger(TableSize))
5322  return false;
5323  }
5324 
5325  for (const auto &I : DefaultResultsList) {
5326  PHINode *PHI = I.first;
5327  Constant *Result = I.second;
5328  DefaultResults[PHI] = Result;
5329  }
5330 
5331  if (!ShouldBuildLookupTable(SI, TableSize, TTI, DL, ResultTypes))
5332  return false;
5333 
5334  // Create the BB that does the lookups.
5335  Module &Mod = *CommonDest->getParent()->getParent();
5336  BasicBlock *LookupBB = BasicBlock::Create(
5337  Mod.getContext(), "switch.lookup", CommonDest->getParent(), CommonDest);
5338 
5339  // Compute the table index value.
5340  Builder.SetInsertPoint(SI);
5341  Value *TableIndex;
5342  if (MinCaseVal->isNullValue())
5343  TableIndex = SI->getCondition();
5344  else
5345  TableIndex = Builder.CreateSub(SI->getCondition(), MinCaseVal,
5346  "switch.tableidx");
5347 
5348  // Compute the maximum table size representable by the integer type we are
5349  // switching upon.
5350  unsigned CaseSize = MinCaseVal->getType()->getPrimitiveSizeInBits();
5351  uint64_t MaxTableSize = CaseSize > 63 ? UINT64_MAX : 1ULL << CaseSize;
5352  assert(MaxTableSize >= TableSize &&
5353  "It is impossible for a switch to have more entries than the max "
5354  "representable value of its input integer type's size.");
5355 
5356  // If the default destination is unreachable, or if the lookup table covers
5357  // all values of the conditional variable, branch directly to the lookup table
5358  // BB. Otherwise, check that the condition is within the case range.
5359  const bool DefaultIsReachable =
5360  !isa<UnreachableInst>(SI->getDefaultDest()->getFirstNonPHIOrDbg());
5361  const bool GeneratingCoveredLookupTable = (MaxTableSize == TableSize);
5362  BranchInst *RangeCheckBranch = nullptr;
5363 
5364  if (!DefaultIsReachable || GeneratingCoveredLookupTable) {
5365  Builder.CreateBr(LookupBB);
5366  // Note: We call removeProdecessor later since we need to be able to get the
5367  // PHI value for the default case in case we're using a bit mask.
5368  } else {
5369  Value *Cmp = Builder.CreateICmpULT(
5370  TableIndex, ConstantInt::get(MinCaseVal->getType(), TableSize));
5371  RangeCheckBranch =
5372  Builder.CreateCondBr(Cmp, LookupBB, SI->getDefaultDest());
5373  }
5374 
5375  // Populate the BB that does the lookups.
5376  Builder.SetInsertPoint(LookupBB);
5377 
5378  if (NeedMask) {
5379  // Before doing the lookup, we do the hole check. The LookupBB is therefore
5380  // re-purposed to do the hole check, and we create a new LookupBB.
5381  BasicBlock *MaskBB = LookupBB;
5382  MaskBB->setName("switch.hole_check");
5383  LookupBB = BasicBlock::Create(Mod.getContext(), "switch.lookup",
5384  CommonDest->getParent(), CommonDest);
5385 
5386  // Make the mask's bitwidth at least 8-bit and a power-of-2 to avoid
5387  // unnecessary illegal types.
5388  uint64_t TableSizePowOf2 = NextPowerOf2(std::max(7ULL, TableSize - 1ULL));
5389  APInt MaskInt(TableSizePowOf2, 0);
5390  APInt One(TableSizePowOf2, 1);
5391  // Build bitmask; fill in a 1 bit for every case.
5392  const ResultListTy &ResultList = ResultLists[PHIs[0]];
5393  for (size_t I = 0, E = ResultList.size(); I != E; ++I) {
5394  uint64_t Idx = (ResultList[I].first->getValue() - MinCaseVal->getValue())
5395  .getLimitedValue();
5396  MaskInt |= One << Idx;
5397  }
5398  ConstantInt *TableMask = ConstantInt::get(Mod.getContext(), MaskInt);
5399 
5400  // Get the TableIndex'th bit of the bitmask.
5401  // If this bit is 0 (meaning hole) jump to the default destination,
5402  // else continue with table lookup.
5403  IntegerType *MapTy = TableMask->getType();
5404  Value *MaskIndex =
5405  Builder.CreateZExtOrTrunc(TableIndex, MapTy, "switch.maskindex");
5406  Value *Shifted = Builder.CreateLShr(TableMask, MaskIndex, "switch.shifted");
5407  Value *LoBit = Builder.CreateTrunc(
5408  Shifted, Type::getInt1Ty(Mod.getContext()), "switch.lobit");
5409  Builder.CreateCondBr(LoBit, LookupBB, SI->getDefaultDest());
5410 
5411  Builder.SetInsertPoint(LookupBB);
5412  AddPredecessorToBlock(SI->getDefaultDest(), MaskBB, SI->getParent());
5413  }
5414 
5415  if (!DefaultIsReachable || GeneratingCoveredLookupTable) {
5416  // We cached PHINodes in PHIs. To avoid accessing deleted PHINodes later,
5417  // do not delete PHINodes here.
5419  /*DontDeleteUselessPHIs=*/true);
5420  }
5421 
5422  bool ReturnedEarly = false;
5423  for (PHINode *PHI : PHIs) {
5424  const ResultListTy &ResultList = ResultLists[PHI];
5425 
5426  // If using a bitmask, use any value to fill the lookup table holes.
5427  Constant *DV = NeedMask ? ResultLists[PHI][0].second : DefaultResults[PHI];
5428  StringRef FuncName = Fn->getName();
5429  SwitchLookupTable Table(Mod, TableSize, MinCaseVal, ResultList, DV, DL,
5430  FuncName);
5431 
5432  Value *Result = Table.BuildLookup(TableIndex, Builder);
5433 
5434  // If the result is used to return immediately from the function, we want to
5435  // do that right here.
5436  if (PHI->hasOneUse() && isa<ReturnInst>(*PHI->user_begin()) &&
5437  PHI->user_back() == CommonDest->getFirstNonPHIOrDbg()) {
5438  Builder.CreateRet(Result);
5439  ReturnedEarly = true;
5440  break;
5441  }
5442 
5443  // Do a small peephole optimization: re-use the switch table compare if
5444  // possible.
5445  if (!TableHasHoles && HasDefaultResults && RangeCheckBranch) {
5446  BasicBlock *PhiBlock = PHI->getParent();
5447  // Search for compare instructions which use the phi.
5448  for (auto *User : PHI->users()) {
5449  reuseTableCompare(User, PhiBlock, RangeCheckBranch, DV, ResultList);
5450  }
5451  }
5452 
5453  PHI->addIncoming(Result, LookupBB);
5454  }
5455 
5456  if (!ReturnedEarly)
5457  Builder.CreateBr(CommonDest);
5458 
5459  // Remove the switch.
5460  for (unsigned i = 0, e = SI->getNumSuccessors(); i < e; ++i) {
5461  BasicBlock *Succ = SI->getSuccessor(i);
5462 
5463  if (Succ == SI->getDefaultDest())
5464  continue;
5465  Succ->removePredecessor(SI->getParent());
5466  }
5467  SI->eraseFromParent();
5468 
5469  ++NumLookupTables;
5470  if (NeedMask)
5471  ++NumLookupTablesHoles;
5472  return true;
5473 }
5474 
5475 static bool isSwitchDense(ArrayRef<int64_t> Values) {
5476  // See also SelectionDAGBuilder::isDense(), which this function was based on.
5477  uint64_t Diff = (uint64_t)Values.back() - (uint64_t)Values.front();
5478  uint64_t Range = Diff + 1;
5479  uint64_t NumCases = Values.size();
5480  // 40% is the default density for building a jump table in optsize/minsize mode.
5481  uint64_t MinDensity = 40;
5482 
5483  return NumCases * 100 >= Range * MinDensity;
5484 }
5485 
5486 /// Try to transform a switch that has "holes" in it to a contiguous sequence
5487 /// of cases.
5488 ///
5489 /// A switch such as: switch(i) {case 5: case 9: case 13: case 17:} can be
5490 /// range-reduced to: switch ((i-5) / 4) {case 0: case 1: case 2: case 3:}.
5491 ///
5492 /// This converts a sparse switch into a dense switch which allows better
5493 /// lowering and could also allow transforming into a lookup table.
5495  const DataLayout &DL,
5496  const TargetTransformInfo &TTI) {
5497  auto *CondTy = cast<IntegerType>(SI->getCondition()->getType());
5498  if (CondTy->getIntegerBitWidth() > 64 ||
5499  !DL.fitsInLegalInteger(CondTy->getIntegerBitWidth()))
5500  return false;
5501  // Only bother with this optimization if there are more than 3 switch cases;
5502  // SDAG will only bother creating jump tables for 4 or more cases.
5503  if (SI->getNumCases() < 4)
5504  return false;
5505 
5506  // This transform is agnostic to the signedness of the input or case values. We
5507  // can treat the case values as signed or unsigned. We can optimize more common
5508  // cases such as a sequence crossing zero {-4,0,4,8} if we interpret case values
5509  // as signed.
5510  SmallVector<int64_t,4> Values;
5511  for (auto &C : SI->cases())
5512  Values.push_back(C.getCaseValue()->getValue().getSExtValue());
5513  llvm::sort(Values.begin(), Values.end());
5514 
5515  // If the switch is already dense, there's nothing useful to do here.
5516  if (isSwitchDense(Values))
5517  return false;
5518 
5519  // First, transform the values such that they start at zero and ascend.
5520  int64_t Base = Values[0];
5521  for (auto &V : Values)
5522  V -= (uint64_t)(Base);
5523 
5524  // Now we have signed numbers that have been shifted so that, given enough
5525  // precision, there are no negative values. Since the rest of the transform
5526  // is bitwise only, we switch now to an unsigned representation.
5527  uint64_t GCD = 0;
5528  for (auto &V : Values)
5529  GCD = GreatestCommonDivisor64(GCD, (uint64_t)V);
5530 
5531  // This transform can be done speculatively because it is so cheap - it results
5532  // in a single rotate operation being inserted. This can only happen if the
5533  // factor extracted is a power of 2.
5534  // FIXME: If the GCD is an odd number we can multiply by the multiplicative
5535  // inverse of GCD and then perform this transform.
5536  // FIXME: It's possible that optimizing a switch on powers of two might also
5537  // be beneficial - flag values are often powers of two and we could use a CLZ
5538  // as the key function.
5539  if (GCD <= 1 || !isPowerOf2_64(GCD))
5540  // No common divisor found or too expensive to compute key function.
5541  return false;
5542 
5543  unsigned Shift = Log2_64(GCD);
5544  for (auto &V : Values)
5545  V = (int64_t)((uint64_t)V >> Shift);
5546 
5547  if (!isSwitchDense(Values))
5548  // Transform didn't create a dense switch.
5549  return false;
5550 
5551  // The obvious transform is to shift the switch condition right and emit a
5552  // check that the condition actually cleanly divided by GCD, i.e.
5553  // C & (1 << Shift - 1) == 0
5554  // inserting a new CFG edge to handle the case where it didn't divide cleanly.
5555  //
5556  // A cheaper way of doing this is a simple ROTR(C, Shift). This performs the
5557  // shift and puts the shifted-off bits in the uppermost bits. If any of these
5558  // are nonzero then the switch condition will be very large and will hit the
5559  // default case.
5560 
5561  auto *Ty = cast<IntegerType>(SI->getCondition()->getType());
5562  Builder.SetInsertPoint(SI);
5563  auto *ShiftC = ConstantInt::get(Ty, Shift);
5564  auto *Sub = Builder.CreateSub(SI->getCondition(), ConstantInt::get(Ty, Base));
5565  auto *LShr = Builder.CreateLShr(Sub, ShiftC);
5566  auto *Shl = Builder.CreateShl(Sub, Ty->getBitWidth() - Shift);
5567  auto *Rot = Builder.CreateOr(LShr, Shl);
5568  SI->replaceUsesOfWith(SI->getCondition(), Rot);
5569 
5570  for (auto Case : SI->cases()) {
5571  auto *Orig = Case.getCaseValue();
5572  auto Sub = Orig->getValue() - APInt(Ty->getBitWidth(), Base);
5573  Case.setValue(
5574  cast<ConstantInt>(ConstantInt::get(Ty, Sub.lshr(ShiftC->getValue()))));
5575  }
5576  return true;
5577 }
5578 
5579 bool SimplifyCFGOpt::SimplifySwitch(SwitchInst *SI, IRBuilder<> &Builder) {
5580  BasicBlock *BB = SI->getParent();
5581 
5582  if (isValueEqualityComparison(SI)) {
5583  // If we only have one predecessor, and if it is a branch on this value,
5584  // see if that predecessor totally determines the outcome of this switch.
5585  if (BasicBlock *OnlyPred = BB->